Lecture Notes in Economics and Mathematical Systems Founding Editors: M. Beckmann H.P. Künzi Managing Editors: Prof. Dr...

Author:
Markus Bouziane

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Lecture Notes in Economics and Mathematical Systems Founding Editors: M. Beckmann H.P. Künzi Managing Editors: Prof. Dr. G. Fandel Fachbereich Wirtschaftswissenschaften Fernuniversität Hagen Feithstr. 140/AVZ II, 58084 Hagen, Germany Prof. Dr. W. Trockel Institut für Mathematische Wirtschaftsforschung (IMW) Universität Bielefeld Universitätsstr. 25, 33615 Bielefeld, Germany Editorial Board: A. Basile, A. Drexl, H. Dawid, K. Inderfurth, W. Kürsten

607

Markus Bouziane

Pricing Interest-Rate Derivatives A Fourier-Transform Based Approach

123

Dr. Markus Bouziane Landesbank Baden-Württemberg Am Hauptbahnhof 2 70173 Stuttgart Germany [email protected]

ISBN 978-3-540-77065-7

e-ISBN 978-3-540-77066-4

DOI 10.1007/978-3-540-77066-4 Lecture Notes in Economics and Mathematical Systems ISSN 0075-8442 Library of Congress Control Number: 2008920679 © 2008 Springer-Verlag Berlin Heidelberg This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Production: LE-TEX Jelonek, Schmidt & Vöckler GbR, Leipzig Cover design: WMX Design GmbH, Heidelberg Printed on acid-free paper 987654321 springer.com

To Sabine

Foreword

In a hypothetical conversation between a trader in interest-rate derivatives and a quantitative analyst, Brigo and Mercurio (2001) let the trader answer about the pros and cons of short rate models: ”... we should be careful in thinking market models are the ﬁnal and complete solution to all problems in interest rate models ... and who knows, maybe short rate models will come back one day...” In his dissertation Dr. Markus Bouziane contributes to this comeback of short rate models. Using Fourier Transform methods he develops a modular framework for the pricing of interest-rate derivatives within the class of exponential-aﬃne jump-diﬀusions. Based on a technique introduced by Lewis (2001) for equity options, the payoﬀs and the stochastic dynamics of interestrate derivatives are transformed separately. This not only simpliﬁes the application of the residue calculus but improves the eﬃciency of numerical evaluation schemes considerably. Dr. Bouziane introduces a reﬁned Fractional Inverse Fast Fourier Transformation algorithm which is able to calculate thousands of prices within seconds for a given strike range. The potential of this method is demonstrated for several one- and two-dimensional models. As a result the application of jump-enhanced short rate models for interestrate derivatives is on the agenda again. I hope, Dr. Bouziane’s monograph will stimulate further research in this direction.

T¨ ubingen, November 2007

Rainer Sch¨obel

Acknowledgements

This book is based on my Ph.D. thesis titled ”Pricing Interest-Rate Derivatives with Fourier Transform Techniques” accepted at the Eberhard Karls University of T¨ ubingen, Germany. Writing the dissertation, I am indebted to many people which contributed academic and personal development. Since any list would be insuﬃcient, I mention only those who bear in my opinion the closest relation to this work. First of all, I would like to thank my academic teacher and supervisor Prof. Dr.-Ing. Rainer Sch¨obel. He gave me valuable advice and support throughout the completion of my thesis. Furthermore, I would also express my gratitude to Prof. Dr. Joachim Grammig for being the co-referent of this thesis. Further thanks go to my colleagues from the faculty of Economics and Business Administration, especially Svenja Hager, Robert Frontczak, Wolfgang Kispert, Stefan Rostek and Martin Weiss for fruitful discussions and a pleasant working atmosphere. I very much enjoyed my time at the faculty. Financial support from the Stiftung Landesbank Baden-W¨ urttemberg is gratefully acknowledged. My deepest gratitude goes to my wife Sabine, my parents Ursula and Laredj Bouziane, and Norbert Gutbrod for their enduring support and encouragement.

T¨ ubingen, November 2007

Markus Bouziane

Contents

List of Abbreviations and Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . XV List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .XIX List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .XXI 1

2

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1

1.1 Motivation and Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Structure of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1 4

A General Multi-Factor Model of the Term Structure of Interest Rates and the Principles of Characteristic Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 An Extended Jump-Diﬀusion Term-Structure Model . . . . . . . . .

7 7

2.2 Technical Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.3 The Risk-Neutral Pricing Approach . . . . . . . . . . . . . . . . . . . . . . . . 13 2.3.1 Arbitrage and the Equivalent Martingale Measure . . . . . 15 2.3.2 Derivation of the Risk-Neutral Coeﬃcients . . . . . . . . . . . . 16 2.4 The Characteristic Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 3

Theoretical Prices of European Interest-Rate Derivatives . . 31 3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 3.2 Derivatives with Unconditional Payoﬀ Functions . . . . . . . . . . . . . 32 3.3 Derivatives with Conditional Payoﬀ Functions . . . . . . . . . . . . . . . 38

XII

4

Contents

Three Fourier Transform-Based Pricing Approaches . . . . . . . 45 4.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 4.2 Heston Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 4.3 Carr-Madan Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 4.4 Lewis Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

5

Payoﬀ Transformations and the Pricing of European Interest-Rate Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 5.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 5.2 Unconditional Payoﬀ Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 5.2.1 General Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 5.2.2 Pricing Unconditional Interest-Rate Contracts . . . . . . . . 79 5.3 Conditional Payoﬀ Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 5.3.1 General Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 5.3.2 Pricing of Zero-Bond Options and Interest-Rate Caps and Floors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 5.3.3 Pricing of Coupon-Bond Options and Yield-Based Swaptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

6

Numerical Computation of Model Prices . . . . . . . . . . . . . . . . . . 95 6.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 6.2 Contracts with Unconditional Exercise Rights . . . . . . . . . . . . . . . 96 6.3 Contracts with Conditional Exercise Rights . . . . . . . . . . . . . . . . . 97 6.3.1 Calculating Option Prices with the IFFT . . . . . . . . . . . . . 97 6.3.2 Reﬁnement of the IFFT Pricing Algorithm . . . . . . . . . . . 101 6.3.3 Determination of the Optimal Parameters for the Numerical Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

7

Jump Speciﬁcations for Aﬃne Term-Structure Models . . . . . 111 7.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 7.2 Exponentially Distributed Jumps . . . . . . . . . . . . . . . . . . . . . . . . . . 115 7.3 Normally Distributed Jumps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 7.4 Gamma Distributed Jumps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

8

Jump-Enhanced One-Factor Interest-Rate Models . . . . . . . . . 125 8.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 8.2 The Ornstein-Uhlenbeck Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

Contents

XIII

8.2.1 Derivation of the Characteristic Function . . . . . . . . . . . . . 126 8.2.2 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 8.3 The Square-Root Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 8.3.1 Derivation of the Characteristic Function . . . . . . . . . . . . . 136 8.3.2 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 9

Jump-Enhanced Two-Factor Interest-Rate Models . . . . . . . . . 145 9.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 9.2 The Additive OU-SR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 9.2.1 Derivation of the Characteristic Function . . . . . . . . . . . . . 146 9.2.2 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 9.3 The Fong-Vasicek Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 9.3.1 Derivation of the Characteristic Function . . . . . . . . . . . . . 159 9.3.2 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

10 Non-Aﬃne Term-Structure Models and Short-Rate Models with Stochastic Jump Intensity . . . . . . . . . . . . . . . . . . . . 171 10.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 10.2 Quadratic Gaussian Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 10.3 Stochastic Jump Intensity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 11 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175 A

Derivation of the Complex-Valued Coeﬃcients for the Characteristic Function in the Square-Root Model . . . . . . . . . 179

B

Derivation of the Complex-Valued Coeﬃcients for the Characteristic Function in the Fong-Vasicek Model . . . . . . . . 183

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

List of Abbreviations and Symbols

δ(x) Γ (x) ı ιM

Dirac delta function Gamma function √ imaginary unit, −1 diag[IM ]

F x [. . .], F −1 [. . .]

Fourier Transformation w.r.t. x and inverse Transformation operator

1A C

indicator function for the event A the set of complex-valued numbers

E[. . .], VAR[. . .] P, Q

expectation and variance operator real-world and equivalent martingale measure

R Ft

the set of real-valued numbers information set available up to time t

diag[. . .]

operator returning the diagonal elements of a quadratic matrix

FFT[. . .]

Fast Fourier Transformation operator

FRFT[. . . , ζ]

Fractional Fourier Transformation operator with parameter ζ

IFFT[. . .] Res[. . .]

inverse Fast Fourier Transformation operator residue operator

Re[z], Im[z]

real and imaginary part of the complex-valued variable z

RMSE RMSEa

root mean-squared error approximate root mean-squared error

tr[. . .]

trace operator

XVI

List of Abbreviations and Symbols

ψ(. . .), φ(. . .)

characteristic function and its logarithm

λ

vector of jump intensities governing the Poisson vector process

ΛΣ (xt ) , Λλ

vectors containing risk-compensating factors for the diﬀusion and jump parts, respectively

µ0 , µ1

constant coeﬃcients determining the drift component of xt

ν(J)

matrix of jump-size distributions for the ran-

Σ 0 , Σ1

dom variables J constant coeﬃcients determining the volatility

J

component of xt matrix of random jump sizes of xt

jn N(λt)

nth row of the matrix J vector of independent Poisson processes acting

Wt

with an intensity λ vector of independent Brownian motions

xt ξ(xt , t, T )

general stochastic vector process state-price kernel

a(z, τ ), b(z, τ )

complex-valued coeﬃcient functions of the general characteristic function

AD(. . .) ARCr (. . .)

Arrow-Debreu price level-based, average-rate contract

CAPr (. . .), CAPY (. . .)

level- and yield-based cap contract

CBC(. . .), CBP (. . .) F LRr (. . .), F LRY (. . .)

coupon-bond call and put option level- and yield-based ﬂoor contract

F RAr (. . .), F RAY (. . .) g0 , g 1

level- and yield-based forward rate agreement constant coeﬃcients determining the charac-

IM

teristic payoﬀ part of xt identity matrix of rank M

K N om

strike value of an option contract Nominal Value

P (. . .), CB(. . .) p(. . .)

zero bond, coupon bond probability density function

pEx (J, η)

density function of an exponentially distributed random variable J with mean and volatility η

List of Abbreviations and Symbols

pGa (J, η, p)

XVII

density function of a gamma distributed random variable J with mean ηp and volatility √ η p

pN o (J, µJ , σJ )

density function of a normally distributed random variable J with mean µJ and volatility σJ

SW Ar (. . .), SW AY (. . .) SW PY (. . .)

level- and yield-based receiver swap yield-based swaption contract

t, τ, T

calendar time, time to maturity and time of

U ARCr (. . .)

contract expiry level-based, unconditional average-rate contract

w0A (z), wA 1 (z)

complex-valued coeﬃcient functions determining the short rate for an average-rate contract

w0 , w1 (m) xt

constant coeﬃcients determining the short rate mth element of the vector xt

Y (. . .) z

simple yield to maturity Fourier-transform variable with real part zr

ZBC(. . .), ZBP (. . .)

and imaginary part zi , respectively zero-bond call and put option

ITM, ATM, OTM

in the money, at the money, and out of the money

ODE, PDE and SDE

ordinary, partial, and stochastic diﬀerential equation

OU

Ornstein-Uhlenbeck (process)

SR

Square-Root (process)

List of Tables

4.1

Idealized call option payoﬀ functions . . . . . . . . . . . . . . . . . . . . . . . . 48

8.1

Values of zero-bond call options for the jump-enhanced OU model, where the underlying zero-bond contract has a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

8.2

Values of short-rate caps for the jump-enhanced OU model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . . . . . . . . 134

8.3

Values of average-rate caps for the jump-enhanced OU model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . . . . . . . . 135

8.4

Values of zero-bond call options for the jump-enhanced SR model, where the underlying zero-bond contract has a nominal

8.5

value of 100 units. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 Values of short-rate (average-rate) caps for the jump-enhanced SR model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . 143

9.1

9.2

Values of zero-bond call options for the jump-enhanced OU-SR model, where the underlying zero-bond contract has a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 Values of zero-bond call options for the jump-enhanced OU-SR model, where the underlying zero-bond contract has a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

9.3

Values of short-rate caps for the jump-enhanced OU-SR model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . 155

9.4

Values of short-rate caps for the jump-enhanced OU-SR model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . 156

XX

List of Tables

9.5

Values of average-rate caps for the jump-enhanced OU-SR

9.6

model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . 157 Values of average-rate caps for the jump-enhanced OU-SR

9.7

model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . 158 Values of zero-bond call options for the jump-enhanced Fong-Vasicek model, where the underlying zero-bond contract has a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

9.8

Values of short-rate caps for the jump-enhanced Fong-Vasicek

9.9

model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . 169 Values of average-rate caps for the jump-enhanced Fong-Vasicek model, with a nominal value of 100 units. . . . . . . . 170

List of Figures

2.1

Diﬀerent contours of the Fourier transform in equation (2.26) for a strike of 90 units. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

4.1

Clockwise performed integral path for the derivation of ˜ 2 (xt , t, T ) in equation (4.27) on the real line. . . . . . . . . . . . . . . . 66 Π

5.1

Closed contour integral path for the derivation of P (xt , t, T ) in equation (5.2). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

5.2

Closed contour integral path for the discounted expectation of g (xT ). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

5.3

Closed contour integral path for the derivation of the put-call parity in equation (5.27). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

6.1

Absolute errors of zero-bond call prices for varying values of ω and zi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

6.2

Logarithmic RMSEs of zero-bond call options. . . . . . . . . . . . . . . . 106

6.3

Diﬀerences of the logarithmic RMSEa and the exact RMSE of zero-bond call options. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

6.4

Search for the optimal parameter couple (ω ∗ , zi∗ ). . . . . . . . . . . . . . 108

7.1

Possible combinations of basic diﬀusion processes and jump

7.2

parts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 The density function pEx (J, η) for varying η of an

7.3

exponentially distributed random variable. . . . . . . . . . . . . . . . . . . . 116 The density function pN o (J, µJ , σJ ) for ﬁxed µJ = 0 and varying σJ of a normally distributed random variable. . . . . . . . . . 118

XXII

7.4

List of Figures

The density function pGa (J, η, p) for ﬁxed η = 0.005 and varying p of a gamma distributed random variable. . . . . . . . . . . . 122

8.1

Probability densities for a short rate governed by a Vasicek diﬀusion model enhanced with an exponentially distributed jump component. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

8.2

Probability densities for a short rate governed by a Vasicek diﬀusion model enhanced with a gamma distributed jump component. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

8.3

8.4

Probability densities for a short rate governed by a Vasicek diﬀusion model enhanced with a normally distributed jump component. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 Probability densities for a short rate governed by a CIR diﬀusion model enhanced with an exponentially distributed jump component. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

8.5

Probability densities for a short rate governed by a CIR diﬀusion model enhanced with a gamma distributed jump component. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

9.1

Diﬀerences of the OU-SR model density function and the sum of the particular one-factor pendants for diﬀerent weighting factors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

9.2

Probability densities for a short rate governed by an OU-SR diﬀusion model enhanced with either a gamma or normally distributed jump component for the OU process. . . . . . . . . . . . . . 150

9.3

9.4

Probability densities for a short rate governed by an OU-SR diﬀusion model enhanced with a gamma distributed jump component for the SR process. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 Probability density functions of the Fong-Vasicek pure diﬀusion model for diﬀerent values of the correlation parameter ρ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164

9.5

9.6

Probability densities for a short rate governed by the Fong-Vasicek diﬀusion model enhanced with a gamma distributed jump component for the short-rate process. . . . . . . . . 165 Probability densities for a short rate governed by the Fong-Vasicek diﬀusion model enhanced with a gamma distributed jump component for the volatility process. . . . . . . . . 166

1 Introduction

1.1 Motivation and Objectives In the last few years the demand for sophisticated term-structure models, capable of reﬂecting the market behavior more realistically, e.g. models which can reproduce the feature of market shocks, has dramatically increased. For example, according to the results of their empirical study, Brown and Dybvig (1986) and A¨ıt-Sahalia (1996) question among others the use of pure diﬀusion models, such as the popular interest-rate models of Vasicek (1977) and Cox, Ingersoll and Ross (1985b), to describe the behavior of interest rates. Moreover, recent studies support the assumption of jump components in the term structure of interest rates. In the study of Hamilton (1996), Fed Funds rates on a daily base are analyzed. The author ﬁnds that settlement days and quarter-ends induce statistically signiﬁcant jumps in the term structure of interest rates. Das (2002) analyzed Fed Funds rates on daily bases over the period 1988-1997. As a result of this study, the proposed jump models show a substantially better ﬁt of the empirical data compared to the pure diﬀusion model. Durham (2005) also examined Fed Funds rates for the period 19882005. The model-generated yields of zero-bond prices are then calibrated to the Fed Funds Rate and one- and three-month U.S. Treasury bill rates. The author concludes that the so-called jump-diﬀusion models produce more accurate estimates of the interest-rate curves than the pure diﬀusion model1 . 1

Additional studies examining the empirical performance of jump-diﬀusion models are given in, e.g. Lin and Yeh (1999), Zhou (2001), Wilkens (2005), and Chan (2005).

2

1 Introduction

Thus, the ability of a term-structure model to reproduce these discount rate shocks, based e.g. on the adjustment of the discount rate by the European Central Bank, on an economic crisis, and quarter-end eﬀects, is highly appreciated. Accordingly, jump-diﬀusion interest-rate models were developed to cover this issue. Ahn and Thompson (1988) introduced one of the ﬁrst jumpdiﬀusion models for the term structure of interest rates. In their study, the interest-rate dynamics are derived within an equilibrium framework similar to the one used in Cox, Ingersoll and Ross (1985b) and particular approximate closed-form zero-bond prices are obtained. Das and Foresi (1996) also derived zero-bond prices for a jump-enhanced Vasicek (1977) model. The authors apply an exponentially distributed jump component, where the absolute value of the jump sign is drawn from a Bernoulli distribution. An alternative jump speciﬁcation for the mean-reverting normally distributed short rate is given in Baz and Das (1996)2 . In their approach, the jump-size distribution is given by a normal distribution and approximate zero-bond prices are derived. An empirical test of a Square-Root interest-rate model enhanced with uniformly distributed jumps is given in Zhou (2001). The author ﬁts the particular jump-diﬀusion model to weekly three-month Treasury bill yields. In Durham (2005), the author states an alternative approximation technique for zero-bond prices when the short rate follows the same dynamics as in Baz and Das (1996). Additionally, a bimodal normally distributed jump version of the Vasicek (1977) model together with a jump-enhanced two-factor model is presented3 . In addition, for derivatives research purposes, an important feature such interest-rate models should exhibit is the ability to generate analytical solutions for the derivatives contracts to be priced. If this can be accomplished, the interest-rate instrument can be examined in depth, e.g. doing some sensitivity analysis. However, dealing with jump components, we often have to rely on time-consuming Monte-Carlo methods in order to price interest-rate derivatives. Thus, more ambitious pricing approaches are needed. Recently, integral transformations have been found to be reliable in deriving semi closed-form 2 3

The same model speciﬁcation is used in Das (2002). The approximation technique is also discussed in depth in Durham (2006) for the bimodal normally and exponentially distributed jump extension of a Vasicek (1977) short-rate model.

1.1 Motivation and Objectives

3

solutions of derivatives contracts under more complicated stochastic dynamics. The term semi closed-form solutions in this case refers to closed-form solutions in the image space, according to the particular transformation rule. Especially the subclass of Fourier Transformations have been proven to be useful for pricing problems in ﬁnancial disciplines4 . Basically, the main advantage of this transform technique consists in providing distribution independent pricing formulae. However, even semi closed-form pricing formulae are hard to obtain, dealing with jump-size distributions such as the normal and the gamma. Accordingly, it is our objective to derive an eﬃcient and accurate pricing tool for interest-rate derivatives within a Fourier-transform pricing approach, which is generally applicable to exponential-aﬃne jump-diﬀusion models. This objective can be achieved within four steps. Firstly, we want a ﬂexible shortrate process, which is able to integrate both diﬀusion and jump components. Thus, we extend the exponential-aﬃne model presented in Duﬃe and Kan (1996) by introducing jump components. The second step is to reﬁne the concept of a modular option pricing as proposed in Zhu (2000) by applying the pricing methodology explained in Lewis (2000) and Lewis (2001)5 . Therefore, we want to formulate a distribution-independent pricing framework, where the particular interest-rate contract price can be clearly separated into stochastic and payoﬀ speciﬁc parts. Apart from the pricing theory, we also need a tool to obtain numerical values of the contracts to be priced. A very popular strategy to price derivatives is the Monte-Carlo approach. However, being generally applicable, this numerical pricing approach suﬀers from its time-consuming calculations and its poor convergence to true solutions. The third objective of this thesis is to develop an algorithm, which appropriately computes option prices in the Lewis (2001) pricing approach. In contrast to the Fast Fourier Transformation (FFT), as used in Carr and Madan (1999) for the pricing of 4

Heston (1993) is the seminal paper on this topic, where semi closed-form solutions for options on equities in a stochastic volatility model are derived for the ﬁrst time. Among others, we mention the inﬂuential work of Bakshi and Madan (2000) and Duﬃe, Pan and Singleton (2000) in deriving option prices using Fourier

5

Transformations. Even though this pricing method is mentioned for the ﬁrst time in Lewis (2000), we henceforth refer to Lewis (2001) as the source, because of the detailed discussion and derivation of the pricing methodology.

4

1 Introduction

equity options, we base our computations on the Inverse Fast Fourier Transform (IFFT). Consequently, we introduce in this thesis a new, IFFT-based pricing algorithm, which is able to calculate thousands of option prices within fractions of a second and is a straightforward application to option pricing in the Lewis (2001) framework. The last step is then to examine density functions and contract prices of some popular interest-rate diﬀusion models enhanced with three diﬀerent jump candidates.

1.2 Structure of the Thesis This thesis is organized as follows. We start in chapter two with the formulation of a general term-structure model, which is governed by a multivariate jump-diﬀusion process. After introducing some general concepts in stochastic calculus we demonstrate how the relevant risk-neutral coeﬃcients of the instantaneous interest-rate process can be obtained. Afterwards, we discuss the technique of performing a Fourier Transformation and its inverse and state the system of ordinary diﬀerential equations the general characteristic function has to solve. In chapter three we discuss a representative collection of some interest-rate derivative contracts which can be solved within the Fourierbased pricing mechanism. We distinguish between contracts with conditional and unconditional exercise rights, because of the diﬀerent pricing procedure. Subsequently, in chapter four we discuss three Fourier-based pricing approaches. We begin our summary with the pricing technique using Fouriertransformed Arrow-Debreu state prices. Since this type of valuation was ﬁrst applied by Heston (1993) and further discussed by Bakshi and Madan (2000), we henceforth refer to this approach as the Heston transform approach. Subsequently, we discuss the pricing procedure introduced by Carr and Madan (1999). In this thesis the authors exploit the Fourier Transformation applied not only to the state price densities but to the entire option price. They introduce a valuation approach where theoretical option prices can be subsequently recovered applying a highly eﬃcient algorithm, namely the Fast Fourier Transform, hereafter denoted as FFT. Finally, we discuss the valuation methodology applied by Lewis (2001). This approach features several advantages. Firstly, its composition is highly modularized. Secondly, employing Cauchy’s residue theorem, the approach can be consistently used both for

1.2 Structure of the Thesis

5

interest-rate derivatives with unconditional and conditional exercise rights. Fortunately, this methodology enables the application of an reﬁned IFFT algorithm which we implement in our pricing procedure. In chapter ﬁve, we derive the particular Fourier Transformations of payoﬀ functions needed in pricing the contract forms previously presented. Additionally, we derive in case of a one-factor term-structure model the Fourier representation of a swaption and a coupon-bond option, respectively. Chapter six gives an outline of the numerical algorithm used for pricing purposes. Again, we distinguish between the computation of derivatives with conditional and unconditional exercise rights. Subsequently, we present a further reﬁnement of the pricing algorithm for option contracts by the application of the Fractional Fourier Transformation according to the article of Bailey and Swarztrauber (1994). The last part of the chapter discusses the issue of ﬁnding the optimal parameter constellation of the numerical algorithm. In chapter seven we brieﬂy discuss three diﬀerent jump-size speciﬁcations and derive their general jump transforms. In chapters eight and nine we examine both jump-enhanced one-factor and two-factor interest-rate models and focus on the impact of diﬀerent jump speciﬁcations. The particular one-factor models we enhance with jump components are the prominent interest-rate models introduced in Vasicek (1977) and Cox, Ingersoll and Ross (1985b). For the class of two-factor models we exemplarily discuss an additive model used in Sch¨ obel and Zhu (2000) and a subordinated model according to Fong and Vasicek (1991a). To our knowledge, in case of the Fong and Vasicek (1991a) model, option prices are presented for the ﬁrst time. In chapter ten, we give a perspective of model extensions for which the pricing procedure is also capable in deriving numerical solutions. The ﬁrst extension is to consider a special model class of non-aﬃne interest-rate models. Another extension of our interest-rate model is to consider stochastic jump intensities. Since it ﬁts into the exponential-aﬃne model setup of Duﬃe, Pan and Singleton (2000), the implementation in our pricing procedure presents no greater diﬃculties. However, due to the non-existence of closed-form solutions in any case, we brieﬂy discuss these extensions. In the last chapter, we review the results of our study and give some concluding remarks.

2 A General Multi-Factor Model of the Term Structure of Interest Rates and the Principles of Characteristic Functions

2.1 An Extended Jump-Diﬀusion Term-Structure Model The evolution of the yield curve can be described in various ways. For instance, it is possible to use such quantities as zero-bond prices, instantaneous forward rates and short interest rates, respectively, to build the term structure of interest rates. If the transformation law from one quantity to the other is known, the choice of the independent variable is just a matter of convenience. In this thesis, we attempt to model the dynamics of the instantaneous interest rate, denoted hereafter by r(xt ), in order to construct our derivatives pricing framework. This instantaneous interest rate r(xt ) is also often referred to as the short-term interest rate or short rate, respectively, and characterizes the risk-free rate for borrowing or lending money over the inﬁnitesimal time period [t, t + dt]. Since we model the dynamics in a continuous trading environment, the relevant processes are described via stochastic diﬀerential equations. The economy we consider has the trading interval [0, T ]. The uncertainty under the physical probability measure is completely speciﬁed by the ﬁltered probability space (Ω, F, P). In this formulation Ω denotes the complete set of all possible outcome elements ω ∈ Ω. The information available in the economy is contained within the ﬁltration (F)t≥0 , such that the level of uncertainty is resolved over the trading interval with respect to the information ﬁltration. The last term, completing the probability space, is called the real-world probability measure P on (Ω, F), since it reﬂects the real-world probability law of the data.

8

2 A Multi-Factor Model and Characteristic Functions

We model the dynamic behavior of the term structure in the spirit of Duﬃe and Kan (1996) and Duﬃe, Pan and Singleton (2000), to preserve an exponential-aﬃne structure of the characteristic function. However, we extend the framework in Duﬃe and Kan (1996) to allow for N diﬀerent trigger processes6 , which oﬀers more ﬂexibility. The term structure is then modeled by a multi-factor structural Markov model of M factors, represented by a random vector xt , which solves the multivariate stochastic diﬀerential equation, (1) dxt dx(2) t .. P (2.1) dxt = = µP (xt ) dt + Σ(xt ) dWP t + J dN(λ t). . (M−1) dxt (M) dxt The coeﬃcient vector µP (xt ) has the aﬃne structure P µP (xt ) = µP 0 + µ1 xt

(2.2)

P M M×M and the variance-covariance matrix Σ(xt )Σ(xt ) with (µP 0 , µ1 ) ∈ R ×R suﬃces the relation

Σ(xt )Σ(xt ) = Σ0 + Σ1 xt ,

(2.3)

where Σ0 ∈ RM×M is a matrix and Σ1 ∈ RM×M×M is a third order tensor. The vector WP t in equation (2.1) represents M orthogonal Wiener processes. Thus, we have7

P EP ( dWP t dW t ) = IM dt

with IM as the M × M identity matrix. As mentioned above, we extend the ordinary diﬀusion model8 with N independent Poisson processes, condensed in the vector N(λP t). This vector process acts with constant and positive intensities9 λP . We allow for every 6

Chacko and Das (2002) model also the term structure with help of diﬀerent

7

Poisson processes. However, their approach consider a subordinated short rate. If not indicated otherwise, we subsequently use the shorthand notation E[ · ] for

8

the expression E[ · |Ft ]. This would be the original model approach presented in Duﬃe and Kan (1996). This exponential-aﬃne model can be easily extended to stochastic jump intensi-

9

P ties of the form λP (xt ) = λP 0 + λ 1 xt . See Chapter 10.

2.1 An Extended Jump-Diﬀusion Term-Structure Model

9

particular factor in xt an amount of N diﬀerent jumps drawn from a jump amplitude matrix J ∈ RM×N . Hence, the distribution functions of the particular jump amplitudes are given within the matrix ν(J). Finally, all jump amplitudes in J are independent of the state of the vector xt 10 . To preserve the exponential-aﬃne structure of any derivatives contract based on r(xt ) and xt , respectively, all random sources, the Brownian motions P WP t , intensities λ and jump amplitudes J are mutually independent. As a direct consequence of the independence of J and xt , there is no chance to generate an arbitrage opportunity according to available information before the particular jump occurs. Hence, given a jump time t∗ , we have formally J ∈ Ft∗− . Therefore, if a jump occurs at time t∗ , nobody is able to predict the exact jump amplitude and cannot gain an arbitrarily large proﬁt with certainty. In this thesis, the choice of jump amplitudes in J can draw on three different types of distribution. These are: •

Exponentially distributed jumps.

• •

Normally distributed jumps. Gamma distributed jumps.

These jump distributions and the resulting jump transforms, which are used in our pricing mechanism, are covered in Chapter 7. Basically, we prefer to model the term structure in terms of the instantaneous short interest rate r (xt )11 , because in this framework all fundamental quantities are properly deﬁned as the expectation of some functionals on the underlying process r (xt ). Accordingly, we are able to construct an arbitragefree economy and simultaneously guarantee a consistent pricing methodol10

From a technical point of view, it is either possible to introduce a dependence on xt for the jump intensity together with independent random jump amplitudes or a dependence on xt for the jump amplitude together with constant jump intensities.

11

See Zhou (2001), p. 4. Other approaches are possible, e.g. the direct approach as used in Sch¨ obel (1987) and Briys, Crouhy and Sch¨ obel (1991) or modeling the forward-rate process as done in Heath, Jarrow and Morton (1992).

10

2 A Multi-Factor Model and Characteristic Functions

ogy12 . The drawback of this approach is that we might not be able to explain perfectly the entire term structure extracted from observed bond market prices and therefore must content ourselves with a best ﬁt scenario. The literature distinguishes between two approaches in modeling the short interest rate in a multidimensional framework. Firstly, we can identify a strategy, which we call henceforth the subordinated modeling approach. Here, the short rate is modeled as (1)

(2)

(M)

r (xt ) = w0 + w1 xt (xt , . . . , xt

).

Consequently, the other M − 1 stochastic factors are subordinated loadings, containing e.g. a stochastic volatility and/or a stochastic mean13 . Apart from (1) the stochastic variable xt , we also consider the deterministic parameters w0 and w1 in modeling the short rate. Indeed, there are other factors, which can possibly have some other economic meaning worth to be included in the interest-rate model. The second method in modeling short rates, which we call the additive modeling approach, is to represent rt as a weighted sum over xt , formally given by r (xt ) = w0 + w xt , 12

13

This means that all derivative prices are based on the same price of risk. See Culot (2003), Section 2.1. In Brennan and Schwartz (1979), Brennan and Schwartz (1980), and Brennan and Schwartz (1982) the short-rate process is subordinated by a stochastic long-term rate. Beaglehole and Tenney (1991) discuss a two-factor interest-rate model with a stochastic long-term mean component and Fong and Vasicek (1991a) introduce a short-rate model with stochastic volatility. A model where the short rate depends on a stochastic inﬂation factor is modeled in Pennacchi (1991). Kellerhals (2001) analyzes an interest-rate model with a stochastic market price of risk component. In Balduzzi, Das, Foresi and Sundaram (1996), the authors present a short-rate model with a stochastic mean and volatility component.

2.2 Technical Preliminaries

11

where w is a M × 1 vector containing separate weights for the corresponding factor loadings in xt 14 . However, this model approach possibly entails diﬃculties in explaining the economic meaning of the variables xt 15 .

2.2 Technical Preliminaries Before we proceed any further, we have to discuss some general results and principles of stochastic analysis, which are commonly used in ﬁnancial engineering, namely the prominent Itˆ o’s Lemma and the equally famous FeynmanKac Theorem. These two principles play a major role in diﬀusion theory and are well connected. Since we consider discontinuous jumps in our model setup, we have to use extended versions of these two results. At ﬁrst we have to state some regularity conditions on the jump-diﬀusion process, in order to guarantee their application.

Deﬁnition 2.2.1 (Regularity Conditions for Jump-Diﬀusion Processes). If the vector process xt represents a multivariate jump-diﬀusion, the parameter coeﬃcients µ(xt ), Σ(xt ) have to satisfy the following technical conditions16 for all t ≥ 0 •

µ(xat ) − µ(xbt ) ≤ A1 xat − xbt

•

Σ(xat )) − Σ(xbt ) ≤ A2 xat − xbt

•

µ(xat ) ≤ A1 (1 + xat )

•

Σ(xat )) ≤ A2 (1 + xat )

where xat , xbt ∈ RM are two vectors containing diﬀerent realizations of xt and the constants A1 , A2 < ∞ denote some scalar barriers. Additionally, we need 14

Langetieg (1980) models the short rate as an additive process consisting of two correlated Ornstein-Uhlenbeck processes. In Beaglehole and Tenney (1991) an additive, multivariate quadratic Gaussian interest-rate model is given. Longstaﬀ and Schwartz (1992) and Chen and Scott (1992) model the interest-rate process

15 16

as the sum of two uncorrelated Square-Root processes. A comprehensive discussion on this topic is given in Piazzesi (2003). The ﬁrst two conditions are known as the Lipschitz conditions, the latter two represent the growth or polynomial growth conditions. See, for example, Karlin and Taylor (1981).

12

2 A Multi-Factor Model and Characteristic Functions

for the jump components the integral

R

ecJmn dν(Jmn ) to be well deﬁned for

every Jmn ∈ J and some constant c ∈ C.

If the conditions posed above are met, we are able to apply both Itˆo’s Lemma and the Feynman-Kac Theorem. We start with Itˆ o’s Lemma. This lemma enables us to determine the stochastic process driving some function f (xt , t, T ), depending on time t and a stochastic (vector) variable, e.g. the process xt given in equation (2.1). The variables t and xt , respectively, are hereafter denoted as the independent variables. The coeﬃcients µ(xt ) and λ used in this section have no superscripts, because the principles introduced here hold in general.

Theorem 2.2.2 (Itˆ o Formula for Jump-Diﬀusion Processes17 ). Assume the function f (xt , t, T ) is at least twice diﬀerentiable in xt and once diﬀerentiable in t. Then the canonical decomposition of the stochastic diﬀerential equation for f (xt , t, T ) is given by ∂f (xt , t, T ) ∂f (xt , t, T ) + µ(xt ) df (xt , t, T ) = ∂t ∂xt

2 1 ∂ f (xt , t, T ) + tr Σ(xt )Σ(xt ) dt 2 ∂xt ∂xt (2.4) ∂f (xt , t, T ) Σ(xt ) dWt + ∂xt + (f (xt , J, t, T ) − f (xt , t, T )) dN(λt), where the function f (xt , J, t, T ) contains all jump components with elements (f (xt , J, t, T ))n = f (xt + jn , t, T ) and jn ∈ RM contains as mth element Jmn of the amplitude matrix J.

Another key result which we use extensively is the Feynman-Kac theorem. This theorem provides us with a tool to determine the system of partial diﬀerential equations (PDEs), given an expectation. 17

See, Kushner (1967), p. 15, for the jump-extended version of Itˆ o’s lemma.

2.3 The Risk-Neutral Pricing Approach

13

Theorem 2.2.3 (Feynman-Kac). If the restrictions in deﬁnition 2.2.1 hold, we have the expectation f (xt , t, T ) = E e

−

T

h(xs ,s) ds t

f (xT , T, T ) ,

(2.5)

solving the partial diﬀerential equation

2 ∂f (xt , t, T ) 1 ∂f (xt , t) ∂ f (xt , t, T ) + µ(xt ) + tr Σ(xt )Σ(xt ) ∂t ∂xt 2 ∂xt ∂xt

(2.6)

+ EJ [f (xt , J, t, T ) − f (xt , t, T )] λ = h(xt , t)f (xt , t, T ), with boundary condition18 f (xT , T, T ) = G (xT )

(2.7)

and f (xt , J, t, T ) as deﬁned in theorem 2.2.2.

In diﬀusion theory, the function h(xt , t) is commonly addressed to as the killing rate of the expectation19 and can be interpreted as some short rate. Since we use equivalently as killing rate a short rate characterized by the time constant coeﬃcients w0 and w we set the relation h(xt , t) = r (xt ) . As we will see, these two principles are the fundamental tools in obtaining the solutions for our upcoming valuation problems, especially in calculating the general characteristic function of a stochastic process, which is discussed in the next sections.

2.3 The Risk-Neutral Pricing Approach So far, the stochastic behavior of the state vector xt was assumed to be modeled under the real-world probability measure P. This probability measure depends on the investor’s assessment of the market and therefore cannot be 18 19

The operator EJ [ · ] denotes the expectation with respect to the jump sizes J. See, for example, Øksendal (2003), p. 145.

14

2 A Multi-Factor Model and Characteristic Functions

used in calculating unique derivatives prices20 . However, for valuation purposes we need to derive contract prices under the condition of an arbitrage-free market21 , which will be shown in this section. According to the seminal papers of Harrison and Kreps (1979) and Harrison and Pliska (1981), it is a well known and rigorously proved fact, if one can ﬁnd at least one equivalent martingale measure with respect to P, then the observed market is arbitrage-free and therefore a derivatives pricing framework can be established. Thus, we establish the link between this equivalent martingale measure Q, also known as the risk-neutral probability measure22 , and the probability measure P in this section. Since we are dealing with M stochastic factors, primarily integrated in the short rate r (xt ), which are all non-tradable goods, we are confronted with an incomplete market. In contrast to other model frameworks in which factors represent prices of tradable goods, we encounter a somewhat more diﬃcult situation to end up in a consistent arbitrage-free pricing approach23 . Foremost, we need to introduce for every source of uncertainty a market price of risk reﬂecting the risk aversion of the market. The common procedure in this case is to choose a particular equivalent martingale measure, sometimes also called the pricing measure which determines the appropriate numeraire to be applied24 . Having chosen the numeraire, which has the function of a denominator of the expected contingent claim and determines the martingale condition for the expectation, we afterwards have to extract yields for diﬀerent maturities of zero-bond prices. In the next step the model prices of zero bonds 20 21

See, for example, Musiela and Rutkowski (2005), p. 10. The arbitrage-free approach is also known as the partial equilibrium approach. Including preferences of investors, i.e. working with utility functions would be a general equilibrium approach. Sch¨ obel (1995) gives a detailed overview of both

22

approaches. The terminology can be justiﬁed, since in a risk-neutral world, where all market participants act under a risk-neutral utility behavior, the probability measures P

23

and Q coincide. See, for example, Duﬃe (2001), p. 108. This statement holds only for tradable goods modeled by pure diﬀusion processes. Otherwise, due to the jump uncertainty one has again to implement some variable

24

compensating jump risk. See Merton (1976). This can be for example the money market account or zero-coupon bond prices. See Dai and Singleton (2003), pp. 635-637.

2.3 The Risk-Neutral Pricing Approach

15

are calibrated with respect to this empirical yield curve. In the calibration process for these parameters, two separate approaches can be utilized25 . In the ﬁrst approach one computes the particular model parameters under the P measure together with the diﬀerent market prices of risk. The other method would be to calibrate the model onto the parameters under the objective measure Q. A problem which is common to all model frameworks, where the instantaneous interest rate r(xt ) is used to describe the term structure of interest rates is that in general the given yield curve is not matched perfectly. Hence, we rather want an arbitrage-free model, which might not be able to explain perfectly all observed yields, but to state a model with an internally consistent stochastic environment. In the upcoming subsections, we will ﬁrst give an outline how the riskneutral measure is deﬁned and how the particular coeﬃcients under this probability measure Q can be derived for our aﬃne term-structure model. Due to the jump-diﬀusion framework, we also focus on the topic that our martingale measure should consider for discontinuous price shocks. 2.3.1 Arbitrage and the Equivalent Martingale Measure Before we start with the formulation of our option-pricing methodology, we need to ensure the existence of an arbitrage-free pricing system. A very useful insight for this delicate matter is given in the above mentioned work of Harrison and Kreps (1979) and Harrison and Pliska (1981). Using measure theory, they judge the market to be arbitrage free enabling the consistent calculation of derivative prices if at least one equivalent martingale measure can be found, corresponding to the physical measure P. Hence, using the money market account as numeraire in order to derive Q, the price of a derivative contract would be just the discounted expectation of its terminal payoﬀ G (xT )26 . So our ﬁrst step is to deﬁne the relevant conditions for an equivalent martingale measure. 25 26

See Duﬃe, Pan and Singleton (2000), p. 1354. See, for example, Geman, Karoui and Rochet (1995) and Dai and Singleton (2003), p. 635.

16

2 A Multi-Factor Model and Characteristic Functions

Deﬁnition 2.3.1 (Equivalent Probability Measure). Two probability measures P and Q are equivalent, if for any event A, P(A) > 0 if and only if Q(A) > 0. According to deﬁnition 2.3.1, the equivalent probability measure Q must only agree on the same null sets given by P. The next property we need, in order to obtain the probability measure Q, is the martingale property.

Deﬁnition 2.3.2 (Martingale Property). A stochastic process f (xt , t) is a martingale under the probability measure Q if and only if the equality f (xt , t, T ) = EQ [f (xT , T, T )]

(2.8)

holds for any t ≤ T .

This last deﬁnition ensures the fair game ability of our interest-rate market. Combining deﬁnitions 2.3.1 and 2.3.2 lead us to the equivalent martingale measure Q with respect to P. Thus, to be a fair game, respectively a martingale, the probability measure Q transforms the probability law for xt , leaving the null sets of P untouched. In the next subsection we show the transition of the probability law from the real-world measure P to the risk-neutral measure Q. 2.3.2 Derivation of the Risk-Neutral Coeﬃcients Having found the formal conditions of an equivalent martingale measure, we now want to derive the transformation rule from measure P to Q. This rule, also called the Radon-Nikodym derivative ξ(xt , t, T ), is represented by dQ ξ(xT , T, T ) . (2.9) = dP Ft ξ(xt , t, T ) In order to derive the risk-neutral coeﬃcients, we adopt the corresponding pricing-kernel methodology. Doing this, the pricing kernel or Radon-Nikodym derivative ξ(xt , t, T ), belongs itself to the class of exponential-aﬃne functions of xt 27 . The principle of risk-neutrality implies for the state-price kernel an 27

See, for example, Dai and Singleton (2003), p. 642.

2.3 The Risk-Neutral Pricing Approach

17

expected discount rate equal to the instantaneous risk-free rate r (xt ). Thus, we need the equation P

E

dξ(xt , t, T ) = −r (xt ) dt, ξ(xt , t, T )

(2.10)

to hold. Using this type of state-price kernel, we have the discounted expectation of an interest-rate derivatives price to fulﬁll the deﬁnition of a martingale as described in theorem 2.3.2. Consequently, ensuring the expectation made above holds and considering the systematic risk factors, we choose the speciﬁc form of ξ(xt , t, T ) to satisfy dξ(xt , t, T ) = −r (xt ) dt − ΛΣ (xt ) dWP − Λλ dN(λP t) − λP dt . (2.11) ξ(xt , t, T ) The vectors ΛΣ (xt ) and Λλ compensate the sources of risk under the riskneutral measure Q for the vector of Brownian motions and the vector of Poisson processes, respectively. The vector ΛΣ (xt ) is characterized by the two relations28

ΛΣ (xt ) ΛΣ (xt ) = l0 + l1 xt Σ (xt ) ΛΣ (xt ) = s0 + s1 xt with l0 ∈ R, l1 , s0 ∈ RM , and s1 ∈ RM×M . Deﬁning ΛΣ (xt ) like this, we ensure the exponential-aﬃne structure in the pricing kernel ξ(xt , t, T ). In contrast to the constant, N -dimensional vector Λλ , we need to establish in ΛΣ (xt ) a dependence on the state vector xt because of a possibly non(m) zero matrix Σ1 29 . Thus, if a particular factor xt has a constant volatility coeﬃcient, meaning its volatility does not depend on any element in xt , there is either no dependence on xt for the respective element in the the vector ΛΣ (xt ) and vice versa. Since λP is the vector of expected arrival rates, we have with

EP dN(λP t) − λP dt = 0N ,

a P-martingale, representing a vector of compensated Poisson processes30 . 28

29

30

Compare, for example, with Duﬃe, Pan and Singleton (2000), Culot (2003), and Dai and Singleton (2003). Dealing with a Square-Root process, we cannot set the particular market price of risk to a constant value, see Cox, Ingersoll and Ross (1985b), Section 5. A compensated Poisson process can be roughly seen as a discontinuous equivalent of a Brownian motion. See, for example, Karatzas and Shreve (1991), p. 12.

18

2 A Multi-Factor Model and Characteristic Functions

As a consequence of this incomplete market, the vectors ΛΣ (xt ) and Λλ are not uniquely deﬁned. Therefore, the pricing kernel itself is not uniquely deﬁned either and we have to determine these risk price vectors with a calibration of yields generated by the model to the empirical yield curve as mentioned earlier. We assume this calibration to depend on the yields of traded zero-coupon bonds P (xt , t, T ) with diﬀerent times to maturities31 . Suppressing unnecessary notations for convenience and applying Itˆo’s Lemma, we get the following SDE for the P-dynamics of a zero-coupon bond dP (xt , t, T ) = µP dt + σ P dWP + JP dN(λP t)

(2.12)

with drift, diﬀusion and jump components32 ∂P (xt , t, T ) ∂P (xt , t, T ) + µP (xt ) ∂t ∂xt

2 ∂ P (xt , t, T ) 1 + tr Σ(xt )Σ(xt ) , 2 ∂xt ∂xt ∂P (xt , t, T ) , σ P = Σ(xt ) ∂xt

µP =

JP = P(xt , J, t, T ) − P (xt , t, T ) .

(2.13)

(2.14) (2.15)

On the other hand, we impose the martingale condition for traded contracts, which is due to the chosen numeraire, T P (xt , t, T ) =EQ e =EP

−

r(xs ) ds t

P (xT , T, T )

ξ(xT , T, T ) P (xT , T, T ) . ξ(xt , t, T )

(2.16)

Multiplying this last equation with ξ(xt , t, T ), which is known at time t and therefore a certain quantity, we consequently have ξ(xt , t, T )P (xt , t, T ) to be a martingale and the inﬁnitesimal increment d (ξ(xt , t, T )P (xt , t, T )) to be a local martingale33 . According to Theorem 2.2.2 we have 31

32

33

Since coupon bonds are commonly traded, zero-bond values can be synthetically generated by coupon stripping. P(xt , J, t, T ) has the equivalent deﬁnition as f (xt , J, t, T ) with all calculations made with respect to P (xt , t, T ). See Theorem 2.2.2. The existence of a local martingale under the new measure Q is suﬃcient for the no-arbitrage condition. See Delbaen and Schachermayer (1995) and Øksendal (2003) Section 12.1., respectively.

2.3 The Risk-Neutral Pricing Approach

19

d(ξ(xt , t, T )P (xt , t, T )) = ξ(xt , t, T ) dP (xt , t, T ) + P (xt , t, T ) dξ(xt , t, T ) + dP (xt , t, T ) dξ(xt , t, T ) = ξ(xt , t, T )µP dt + ξ(xt , t, T )σP dWP + ξ(xt , t, T )JP N(λP )

(2.17)

− P (xt , t, T ) ξ(xt , t, T )r (xt ) dt

− P (xt , t, T ) ξ(xt , t, T )ΛΣ (xt ) dWP − P (xt , t, T ) ξ(xt , t, T )Λλ dN(λP t) − λP dt P

− ξ(xt , t, T )σP ΛΣ (xt ) dt − ξ(xt , t, T )JP Iλ N Λλ dt. In the last equation, we used for the inﬁnitesimal time increments the relation dt dt = 0, and for the vector of uncorrelated Brownian motions

dWP dWP = IM dt. Similarly, the corresponding expression for the vector of independent Poisson processes is P

dN(λP t) dN(λP t) = Iλ N dt, P

where Iλ N represents a matrix consisting of the diagonal elements P = λP , diag Iλ N and zeros otherwise. In the next step, we divide for notational ease all coefﬁcients of the zero-bond SDE (2.12) by P (xt , t, T ). Hence, we use hereafter the normalized coeﬃcients, µP , P (xt , t, T ) σP ˜P = σ , P (xt , t, T ) JP ˜P = J . P (xt , t, T ) µ ˜P =

Combining condition (2.16) and equation (2.17), and keeping in mind that under P-dynamics, the Brownian motions and the compensated Poisson processes in equation (2.11) are martingales, we get for the expectation

20

2 A Multi-Factor Model and Characteristic Functions

EP

d (ξ(xt , t, T )P (xt , t, T )) ˜ P λP dt =µ ˜P dt + EJ J ξ(xt , t, T )P (xt , t, T ) ˜ P ΛΣ (xt ) dt − r (xt ) dt − σ P ˜ P Iλ Λλ dt ≡ 0. − EJ J N

(2.18)

If we now solve equation (2.18) for the modiﬁed drift coeﬃcient µ ˜ P , subsequently eliminating all dt terms, we eventually end up with the relation P ˜ P Iλ Λλ − λP , ˜ P ΛΣ (xt ) + EJ J µ ˜P = r (xt ) + σ (2.19) N which means that the rate of return of a zero bond must be equal to the risk free short rate plus some terms reﬂecting the particular risk premiums of the diﬀerent sources of uncertainty. We are now ready to identify the corresponding formal expressions under Q-dynamics of the coeﬃcient parameters µP and λP . Comparing equation (2.13) with (2.19) lead us to the fundamental partial diﬀerential equation for zero-bond prices34 ∂P (xt , t, T ) ∂P (xt , t, T ) P + µ − Σ(xt )ΛΣ (xt ) ∂t ∂x t

1 ∂ 2 P (xt , t, T ) + tr Σ(xt )Σ(xt ) 2 ∂xt ∂xt P + EJ [JP ] λP − Iλ N Λλ = r (xt ) P (xt , t, T ) .

(2.20)

According to equation (2.20), together with Itˆ o’s Lemma, and the FeynmanKac representation, we are able to express the risk-neutral parameters as Q µQ = µP − Σ(xt )ΛΣ (xt ) = µQ 0 + µ1 x t , Q

P

λ =λ −

P Iλ N Λλ .

(2.21) (2.22)

Since the jump intensities λQ have to be positive, we need Λλ small enough to ensure the positiveness of the jump intensities under the risk-neutral measure Q given the intensity vector λP . The constant coeﬃcients in the variancecovariance matrix (2.3) remain unchanged under the new measure Q. This 34

Once the risk-neutral coeﬃcients for the interest-rate process are determined, equation (2.20) can be used to price any European contingent claim by exchanging the terminal condition and replacing P (xt , t, T ) with the particular function representing the price of the derivative security to be calculated.

2.4 The Characteristic Function

21

phenomenon is often referred to as the diﬀusion invariance principle, although this terminology is not completely correct. We want to emphasize that the variations of the Brownian motions only coincide under both measures P and Q, if the variance-covariance matrix exclusively exhibits constant coeﬃcients35 . Otherwise, we are implicitly dealing with a diﬀerent time-dependent variance-covariance matrix, since the vector xt experiences a drift correction and therefore aﬀects the relation given in equation (2.3). Consequently, the probability transformation law of the process xt from P to Q does not only contain a drift compensation. Moreover, besides the jump intensity correction, the very shape of the probability density itself can be changed, due to the implicitly altered variations of the diﬀusion terms. Hence, calibrating the theoretical term-structure model to zero-bond yields, whether estimating the parameters of the left or the right sides of equations 2.21 and 2.22, results in the following SDE governing the particular factors under risk-neutral dynamics Q dxt = µQ (xt ) dt + Σ(xt ) dWQ t + J dN λ t ,

(2.23)

which we use in the subsequent sections as starting point for our calculations.

2.4 The Characteristic Function In this section, we ﬁrst give a brief overview of the abilities of characteristic functions and show afterwards how the characteristic function of an exponential-aﬃne process, as given in equation (2.1), can be derived. We generalize the principle of building characteristic functions for some scalar process g(xt ), which is essential for our derivatives pricing technique. Since characteristic functions play a major part in our derivation of semi closed-form solutions for interest-rate derivatives, we discuss also some of their fundamental properties. Before we introduce the characteristic function itself, we ﬁrst need to state a deﬁnition of Fourier Transformations of some deterministic variable x36 . 35 36

In this case, we would deal with the matrix Σ(x)Σ(x) = Σ0 . In the literature, there seems to exist various deﬁnitions for this type of transformation. Thus, we want to clarify the issue by giving a straightforward deﬁnition

22

2 A Multi-Factor Model and Characteristic Functions

This concept belongs to the ﬁeld of integral transformations37 and is a widely used tool in engineering disciplines, especially in signal processing. Deﬁnition 2.4.1 (General one-dimensional Fourier Transformation and its Inversion). We deﬁne the Fourier Transformation F x [ · ] of some function f (x) with respect to the independent variable x as ∞ eızx f (x) dx = fˆ(z),

F [f (x)] = x

(2.24)

−∞

where z ∈ C denotes the transform variable in Fourier space, satisfying the restriction Im(z) ∈ (χ, χ) with χ and χ denoting some lower and upper bound√ aries guaranteeing the existence of the Fourier Transformation, ı = −1 as the standard imaginary unit, and fˆ(z) as the shorthand notation for the Fourier Transformation of f (x) with respect to its argument x. Accordingly, the inverse transformation operator F −1 [ · ] is then deﬁned by F

−1

1 [fˆ(z)] = 2π

∞

e−ızx fˆ(z) dz = f (x).

(2.25)

−∞

Due to the exponential character of the Fourier Transformation, we need to establish in equation (2.25) a normalization factor of 2π. The terminology general one-dimensional Fourier Transformation, in contrast to an ordinary one-dimensional Fourier Transformation, is used because we do not limit the transformation variable z to be on the real line38 . Thus, we allow z to be complex-valued, which makes equation (2.24) and (2.25) a line integral, performed parallel to the real line. Note that both the transform and its inverse in this section. In ﬁnancial studies our deﬁnition according to equation (2.24) of a Fourier Transformation seems to be commonly accepted. See, for example, Carr and Madan (1999), Bakshi and Madan (2000) and Raible (2000). On the other hand in engineering sciences, the opposite deﬁnition of a Fourier Transformation 37

and its inverse operation does exist. See, for example, Duﬀy (2004). Other popular integral transformations are e.g. the Laplace transformation or the z-transformation. A comprehensive discussion of the Laplace Transformation is

38

given in Doetsch (1967). Hence, the equivalent expression complex Fourier Transformation is sometimes used in the literature.

2.4 The Characteristic Function

23

operation have to take place on the same strip going through Im(z), in order to reconstruct the original function f (x). The advantage in performing this general Fourier Transformation is the possibility to derive image functions in cases where the ordinary transform approach would fail, e.g. for functions which are unbounded39 . However, in these cases, the general approach enables us to derive solutions for their Fourier Transformations. For example, if we want to compute the Fourier Transformation of a function40 G(x) = max(ex − K, 0), the ordinary transformation approach appears to be useless, since F x [G(x)] → ∞. Performing a general transformation, in this case within the strip Im(z) ∈ (1, ∞), we get41

K 1+ız , (2.26) ız(1 + ız) where Im(z) can be ﬁxed at every value within the above mentioned strip to derive the original function by applying the inverse Fourier Transformation. F x [G(x)] =

The diﬀerent contours in Fourier space of the transformed payoﬀ function given in equation (2.26) are depicted in Figure 2.1. Having derived the fundamental technique to compute Fourier Transformations, which is an essential part in this thesis, we go further and have a look at Fourier Transformations of density functions of stochastic variables, which are commonly known as characteristic functions. Deﬁnition 2.4.2 (Scalar Characteristic Functions). We deﬁne the scalar (m)

characteristic function ψ x (xt , z, w0 , w, t, T ) as the expected value of the ter(m) minal condition G (xT ) = eızxT , given the state xt at time t ≤ T . This can be expressed more formally as 39

40

This is the case for most payoﬀ structures of option contracts, e.g. plain vanilla call or put options. This function represents, for instance, the payoﬀ function of a plain vanilla call option in an asset pricing environment, where x is the natural logarithm of the

41

underlying asset price. In Section 5.3, Fourier Transformations are derived in detail for diﬀerent types of payoﬀ functions.

24

2 A Multi-Factor Model and Characteristic Functions

2 1.5

Re(f(z))

1 0.5 0 −0.5 −1 2 1.5

Im(z)

1

−4

−6

−2

2

0

4

6

Re(z)

Fig. 2.1. Diﬀerent contours of the Fourier transform in equation (2.26) for a strike of 90 units.

ψx

(m)

(xt , z, w0 , w, t, T ) =E e

−

=

T t

(m)

r(xs ) ds+ızxT

(2.27)

e

(m)

ızxT

p(xt , xT , w0 , w, t, T ) dxT ,

RM

for all m = 1, . . . , M . In the last equality of equation (2.27), the function p(xt , xT , w0 , w, t, T ) represents the (discounted) transition probability density, starting with an initial state xt and ending up in time T at xT . The continuous discounting is conducted with respect to r (xt∗ ) for t > t∗ ≥ T . Obviously, if the stochastic process consists only of one variable xt , the characteristic function ψ x (xt , z, 0, 0, t, T ) is then just the Fourier Transformation of the particular transition density function p(xt , xT , 0, 0, t, T ). Although the transform operation in equation (2.27) is performed with respect to the (m)

terminal state of one single random variable xT , we have to consider the state of the vector xt as an argument of the characteristic function. In fact,

2.4 The Characteristic Function

25

since we are looking at the overall expectation, equation (2.27) is generally built as the M -dimensional integral over the entire state vector xT 42 . Therefore, we are also able to apply the deﬁnition presented above of building a characteristic function for the more general case g (xT ) = g0 + g xT

(2.28)

with g0 ∈ R and g ∈ RM . This implies, as long as g (xT ) is a linear combination of the elements in xT that only one single transformation variable z necessary. Hence, if we are able to build the characteristic function for the scalar g (xT )43 , there is only a one-dimensional integral for the inverse operation to be performed, independent of the number of state variables included in g (xT ). Note, this powerful result will be used in our multi-factor framework. Equipped with these deﬁnitions we state next some general and important properties of Fourier Transformations on which we rely in our thesis.

Proposition 2.4.3 (Important Properties of Characteristic Functions and Fourier Transformations). Let α, β, x, y ∈ R, and f (x), g(y) some real-valued functions with Fourier transforms fˆ(z), gˆ(z) and Fourier Transformation variable z ∈ C. Then the following relations hold: 1. Linearity: F x [αf (x) + βg(x)] = αfˆ(z) + βˆ g (z). 2. Diﬀerentiation:

Fx

dα f (x) = (ız)α fˆ(z). dxα

3. Convolution: F x [f (x) ∗ g(x)] = fˆ(z)ˆ g (z). 4. Symmetry: ∞ πf (x) =

e

−ızx

0 42

If

(m) xt

0 fˆ(z) dz =

e−ızx fˆ(z) dz.

−∞

would be no subordinated process and independent from all other

state variables, equation (2.27) could still utilize the joint density function 43

p(xt , xT , w0 , w, t, T ) due to the possible discount factor including r(xt ). For example, calculating the general characteristic function for the short rate r (xt ) itself, we set g (xT ) = r (xT ).

26

2 A Multi-Factor Model and Characteristic Functions

5. Relation of the Moment-Generating and the Characteristic Function: dα ψ x (xt , z, 0, 0M , t, T ) . E [xα ] = (−ı)α dz α z=0 Taking a second glance at Figure 2.1, we are able to justify the symmetry of the Fourier Transformation (2.26) of a real-valued function, mentioned in Proposition 2.4.3. Furthermore, one can clearly identify the dampening property of the characteristic function which is essential in developing a numerical algorithm to compute derivative prices. In the following, we show how the characteristic function for a scalar function g (xT ) is derived within the exponential-aﬃne framework. Following Bakshi and Madan (2000), we interpret the characteristic function as a hypothetical contingent claim. Taking more elaborated payoﬀ structures into account, we have to extend the list of permissible arguments for the characteristic function. The more general representation of the characteristic function, which we use hereafter is ψ g(x) (xt , z, w0 , w, g0 , g, t, T ) with the complex-valued payoﬀ representation at maturity T , ψ g(x) (xt , z, w0 , w, g0 , g, T, T ) = eızg(xT ) .

(2.29)

As discussed in the last section, we have to consider that all contingent claims need to be priced under the risk-neutral probability measure Q. Hence, all prices are derived as discounted expectations. Consequently, the discounted expectation of the general form of the terminal condition can be represented as T ψ g(x) (xt , z, w0 , w, g0 , g, t, T ) = EQ e

−

r(xs ) ds+ızg(xT ) t

.

(2.30)

However, we need to compute discounted expectations, e.g. for vanilla zerobond calls, or undiscounted expectations, e.g. in the case of futures instruments. Hence, for futures-style contracts, w0 equals zero and w is a zero valued vector44 . In calculating European derivative prices, we rather need the general characteristic function ψ g(x) (xt , z, w0 , w, g0 , g, t, T ) than the special case of the 44

The characteristic marking to market for standardized futures-style contracts results in the non-existence of a discount factor in the pricing formula and the relevant PDE, respectively, of such a contract under the risk-neutral measure Q.

2.4 The Characteristic Function

27

characteristic function without considering any discount factor, which is just ψ g(x) (xt , z, 0, 0M , g0 , g, t, T ), where 0M represents a M × 1 vector containing exclusively zeros. Applying Theorem 2.2.3 to our hypothetical claim with a solution according to equation (2.30), we take advantage of the Feynman-Kac representation to derive the partial diﬀerential equation. Simplifying and suppressing unnecessary notation, we write henceforth ψ(xt , z, w0 , w, g0 , g, τ ) ≡ ψ g(x) (xt , z, w0 , w, g0 , g, t, T ) and then get the partial diﬀerential equation ∂ψ(xt , z, w0 , w, g0 , g, τ ) ∂ψ(xt , z, w0 , w, g0 , g, τ ) + µQ (xt ) ∂t ∂xt

2 ∂ ψ(x , z, w , w, g , g, τ ) 1 t 0 0 + tr Σ(xt )Σ(xt ) 2 ∂xt ∂xt

(2.31)

+ EJ [ψ(xt , z, w0 , w, g0 , g, J, τ ) − ψ(xt , z, w0 , w, g0 , g, τ )] λQ = ψ(xt , z, w0 , w, g0 , g, τ )r (xt ) , where the complex-valued vector ψ(xt , z, w0 , w, g0 , g, J, τ ) contains all jump components with particular elements (ψ(xt , z, w0 , w, g0 , g, J, τ ))n = ψ(xt + jn , z, w0 , w, g0 , g, τ ). The vector jn ∈ RM contains as mth element the random variable Jmn of the amplitude matrix J. Every contingent claim or function dependent on xt , an arbitrage-free environment presupposed, has to satisfy the same Partial diﬀerential equation structure as given in equation (2.31). For example, the corresponding risk-neutral transition density for the characteristic function ψ(xt , z, w0 , w, w0 , w, τ ), with g (xT ) = r (xT ), which is actually p(r(xt ), r(xT ), w0 , w, t, T ) need to satisfy the same partial diﬀerential equation as the characteristic function itself45 . The only diﬀerence between them would be the particular terminal payoﬀ condition. Hence, solving the above partial diﬀerential equation for p(r(xt ), r(xT ), w0 , w, t, T ), we would impose the Dirac delta function as the relevant terminal condition, having its density mass exclusively concentrated in an inﬁnite spike for r(xT ) at time T . Solving equation (2.31) together with this type of boundary condition can be quite challenging and is in many cases just impossible46 . Thus, it is feasible to ﬁrst solve equation (2.31) for the general characteristic function, with its smooth and continuous boundary function at T , and afterwards do some sort 45 46

See Heston (1993), p. 331. A prominent example is given with the stochastic volatility model of Heston (1993), for which no closed-form representation of the transition density of the underlying equity log-price variable exists.

28

2 A Multi-Factor Model and Characteristic Functions

of normalized integration, the inverse Fourier Transformation, probably in a numerical manner, to get the desired result. Proceeding like this is a very elegant way to ﬁnd some semi-analytic solution. In contrast, if we want to interpret the terminal payoﬀ function in equation (2.29) as a hypothetical futures-style contract, with solution ψ(xt , z, 0, 0M , g0 , g, τ ) = EQ eızg(xT ) , (2.32) we have a slightly diﬀerent partial diﬀerential equation. In this case the dynamic behavior of ψ(xt , z, 0, 0M , g0 , g, τ ) is described by the slightly altered PDE ∂ψ(xt , z, 0, 0M , g0 , g, τ ) ∂ψ(xt , z, 0, 0M , g0 , g, τ ) + µQ (xt ) ∂t ∂xt

2 1 ∂ ψ(xt , z, 0, 0M , g0 , g, τ ) + tr Σ(xt )Σ(xt ) 2 ∂xt ∂xt

+ EJ [ψ(xt , z, 0, 0M , g0 , g, J, τ ) − ψ(xt , z, 0, 0M , g0 , g, τ )] λ

(2.33) Q

= 0, Hence, the only diﬀerence to PDE (2.31) is that the right hand side is now equal to zero to contribute the missing discount rate. Moreover, we can use this futures-style characteristic function ψ(xt , z, 0, 0M , g0 , g, τ ) to obtain the particular values of the undiscounted transition density function. Thus, to compute the probability density function of the short rate r (xt ), we use this futures-style solution of the characteristic function together with the identity g (xt ) = r (xt ). Consequently, using a separation of variables approach, the partial diﬀerential equations in (2.31) and (2.33) can be decoupled into a system of ordinary diﬀerential equations. Therefore, we assume for ψ(xt , z, w0 , w, g0 , g, τ ) the exponential-aﬃne structure

ψ(xt , z, w0 , w, g0 , g, τ ) = ea(z,τ )+b(z,τ ) xt +ızg0 , with the scalar and complex-valued coeﬃcient function a(z, τ ) and (1) (1) ˜b (z, τ ) g (2) ˜(2) g b (z, τ ) ˜ τ ) + ızg, + ız . = b(z, b(z, τ ) = .. . . . ˜b(M) (z, τ ) g (M)

(2.34)

2.4 The Characteristic Function

29

denotes some complex-valued coeﬃcient vector. In the next step we plug the required expressions of the candidate function (2.34) into equation (2.31). Starting with the time derivative, we get ∂ψ(xt , z, w0 , w, g0 , g, τ ) ∂t = − (a(z, τ )τ + b(z, τ )τ xt ) ψ(xt , z, w0 , w, g0 , g, τ ),

(2.35)

where a(z, τ )τ and b(z, τ )τ are the ﬁrst derivatives with respect to the time to maturity variable τ . The gradient vector with respect to the state variables xt is given by ∂ψ(xt , z, w0 , w, g0 , g, τ ) = b(z, τ )ψ(xt , z, w0 , w, g0 , g, τ ), ∂xt

(2.36)

the Hesse matrix is ∂ 2 ψ(xt , z, w0 , w, g0 , g, τ ) = b(z, τ )b(z, τ ) ψ(xt , z, w0 , w, g0 , g, τ ), ∂xt ∂xt

(2.37)

and the jump component in equation (2.31) can be derived as EJ [ψ(xt , z, w0 , w, g0 , g, J, τ ) − ψ(xt , z, w0 , w, g0 , g, τ )] = EJ [ψ ∗ (z, w0 , w, g0 , g, J, τ ) − 1] ψ(xt , z, w0 , w, g0 , g, τ ),

(2.38)

with the normalized vector ψ(xt , z, w0 , w, g0 , g, J, τ ) ψ(xt , z, w0 , w, g0 , g, τ ) b(z,τ ) J 1 e b(z,τ ) J2 e . = .. .

ψ ∗ (z, w0 , w, g0 , g, J, τ ) =

(2.39)

eb(z,τ ) JN

In this aﬃne framework, it can be easily checked that the normalized amplitude vector ψ ∗ (z, w0 , w, g0 , g, J, τ ) is independent of the actual state of xt , which results in the special form given by equation (2.39). Therefore, we are able to express the system of ODEs resulting from equations (2.31) and (2.33), respectively, and the aﬃne form proposed in (2.34) in terms of the risk-neutral coeﬃcients derived in Section 2.3.2. According to Theorem 2.2.3, the ODE which has to be solved for the scalar coeﬃcient a(z, τ ) is then

30

2 A Multi-Factor Model and Characteristic Functions

1 a(z, τ )τ = −w0 + µQ 0 b(z, τ ) + b(z, τ ) Σ0 b(z, τ ) 2 + EJ [ψ ∗ (z, w0 , w, g0 , g, J, τ ) − 1] λQ ,

(2.40)

whereas for the vector coeﬃcient b(z, τ ) we have to solve 1 b(z, τ )τ = −w + µQ 1 b(z, τ ) + b(z, τ ) Σ1 b(z, τ ), 2

(2.41)

with boundary conditions a(z, 0) = 0 and b(z, 0) = ızg, respectively. The parameters w0 and w, determine whether we consider a discount rate or not for the characteristic function. The mth element of b(z, τ ) Σ1 b(z, τ ) can be computed as i,j b(z, τ )i (Σ1 )ijm b(z, τ )j 47 . Moreover, we want to emphasize that the trace operator is circular, meaning the equality tr [Σ(xt )Σ(xt ) b(z, τ )b(z, τ ) ] = tr [b(z, τ ) Σ(xt )Σ(xt ) b(z, τ )]

(2.42)

holds. Obviously, the right hand side of this last equation represents a scalar and therefore we are able to neglect the trace operator in equation (2.40) and equation (2.41), respectively. In order to calculate derivatives prices, the coeﬃcients a(z, τ ) and b(z, τ ) need not exhibit closed-form solutions in any case. There are several scenarios conceivable, e.g. the time integrated expectations of the jump amplitudes have no closed-form representations, or the processes themselves have such complicated structures that there simply does not exist a closed-form solution of the coeﬃcients a(z, τ ) or b(z, τ ) of the characteristic function. However, if we are able to represent a(z, τ ) and b(z, τ ) in terms of their ordinary diﬀerential equations (2.40) and (2.41), solutions can be eﬃciently obtained via a Runge-Kutta solver and appropriately integrated within our numerical pricing procedure, such that time consuming Monte-Carlo studies for the pricing of European interest-rate derivatives can be avoided.

47

See Duﬃe, Pan and Singleton (2000), p. 1351.

3 Theoretical Prices of European Interest-Rate Derivatives

3.1 Overview In this section, we want to give a representative selection of diﬀerent interestrate contracts for which the pricing framework used in this thesis is able to produce semi closed-form solutions48 . In doing this we distinguish, for didactical purposes, between contracts based on the short rate r(xt ) and contracts based on a simple yield Y (xt , t, T ) over a speciﬁed time period τ . These yields to maturity are often referred to as simple compound rates, e.g. LIBOR rates, and denote the constant compounding of wealth over a ﬁxed period of time τ , which is related to a zero bond with corresponding time to maturity. Deﬁnition 3.1.1 (Simply-Compounded Yield to Maturity). The simple yield to maturity Y (xt , t, T ) of a zero bond P (xt , t, T ), maturing after the time period τ , is deﬁned through the equality 1 = P (xt , t, T ) . 1 + τ Y (xt , t, T )

(3.1)

Therefore the simple yield to maturity can be derived as −1

Y (xt , t, T ) =

P (xt , t, T ) τ

−1

=

1 − P (xt , t, T ) . τ P (xt , t, T )

(3.2)

In the following sections, we generally distinguish in the derivation of theoretical prices of contingent claims between contracts based on the instantaneous interest rate r(xt ) and contracts depending on the simple yield 48

A comprehensive summary of diﬀerent valuation formulae of ﬁxed-income securities is given, e.g. Brigo and Mercurio (2001) and Musiela and Rutkowski (2005).

32

3 Theoretical Prices of European Interest-Rate Derivatives

Y (xt , t, T ). Moreover, we diﬀerentiate between contracts with unconditional and conditional exercise rights. This distinction is introduced because of the diﬀerent mathematical derivation of the particular model prices. For contracts with unconditional exercise, we obtain pricing formulae, which bear strong resemblance to moment-generating functions of the particular underlying state process whereas contracts with conditional exercise rights, i.e. option contracts, need an explicit integration due to the natural exercise boundary. All derivative prices for which we derive the corresponding pricing formulae are European-style derivatives, meaning that the exercise can only be performed at maturity T .

3.2 Derivatives with Unconditional Payoﬀ Functions This derivatives class is characterized by the trivial exercise of the contract at maturity. This means that the contract is always exercised, no matter if the holder suﬀers a loss or make a proﬁt as consequence of the exercise. Although trivially exercised, a zero-coupon bond is a special case of this class since it pays at maturity a predeﬁned riskless quantity of monetary units. Deﬁnition 3.2.1 (Zero-Coupon Bond). A zero-coupon bond maturing at time T guarantees its holder the payment of one monetary unit at maturity. The value of this contract at t < T is then denoted as P (xt , t, T ), which is the expected value of the discounted terminal condition G(xT ) = 1. This can be formally expressed as, T P (xt , t, T ) = EQ e

−

r(xs ) ds t

(3.3)

It is easily seen that the payoﬀ function G (xT ) used in equation (3.3) is independent both of the time variable and the state variables in xT . Using the formal deﬁnition in equation (3.3), a zero-coupon bond, or as shorthand a zero bond, is just the present value of one monetary unit paid at time T . Hence, we are able to interpret P (xt , t, T ) as the expected discount factor relevant for the time period t up to T . Due to this intuitive interpretation, these contracts are often used in calibrating interest-rate models to empirical data sets.

3.2 Derivatives with Unconditional Payoﬀ Functions

33

A slightly more elaborated contract is given by the combination of certain payments at diﬀerent times. We denote this contract then as a coupon-bearing bond.

Deﬁnition 3.2.2 (Coupon-Bearing Bond). A coupon-bearing bond guarantees its holder a number of A deterministic payments ca ∈ c at speciﬁc coupon dates Ta ∈ T for a = 1, . . . , A. Typically, at maturity TA , a nominal face value C is included in cA in addition to the ordinary coupon. The present value of a coupon bond CB(xt , c, t, T) is then given as T a A A − r(xs ) ds EQ e t ca = P (xt , t, Ta ) ca . CB(xt , c, t, T) = a=1

(3.4)

a=1

Obviously, a coupon-bearing bond, or as shorthand a coupon bond, is just the cumulation of payments ca discounted with the particular zero-bond prices P (xt , t, Ta ). If a ﬁrm is requiring a hedge position for a risk exposure in the form of a future payment of interest, due to an uncertain ﬂoating interest rate, we are able to conclude a forward-rate agreement.

Deﬁnition 3.2.3 (Forward-Rate Agreement). A forward-rate agreement concluded in time t guarantees its holder the right to exchange his variable interest payments to a ﬁxed rate K, scaled upon a notional principal N om. The contract is sold in t. The interest payments exchanged relate then to the time period, say [T, Tˆ] with t < T < Tˆ . We distinguish the cases, where the forward-rate agreement refers to the short rate r (xt ) and to the yield Y (xt , t, T ). Hence, for a contract based on the short rate, the relevant time interval is then [T, Tˆ ] = [T, T + dT ]. The price of this contract is given as F RAr (xt , K, N om, t, T ) T = EQ e

−

rs ds t

(K − r (xT )) N om

= K P (xt , t, T ) − EQ e

−

T

rs ds t

r (xT ) N om.

(3.5)

34

3 Theoretical Prices of European Interest-Rate Derivatives

The price for a forward-rate agreement over a discrete time period of length τˆ = Tˆ − T , written on a yield Y xT , T, Tˆ and paid in arrears, can be represented as49 F RAY (xt ,K, N om, t, T, Tˆ ) Tˆ = τˆEQ e

−

= EQ e

−

= E e

t

−

−1 + 1 N om τˆK − P xT , T, Tˆ

r(xs ) ds

τ K + 1) − 1 N om P xT , T, Tˆ (ˆ

T t

K − Y xT , T, Tˆ N om

r(xs ) ds

ˆ T t

Q

r(xs ) ds

(3.6)

N om ˜ = EQ e t P xT , T, Tˆ − K ˜ K ˜ (xt , t, T ) N om , = P xt , t, Tˆ − KP ˜ K −

˜ = with K

T

r(xs ) ds

1 τˆK+1 .

To give a more illustrative example, we consider a ﬁrm, which has to make a future payment subject to an uncertain, ﬂoating rate of interest. Reducing the immanent interest-rate risk exposure, this ﬁrm wants to transform this payment into a certain cash-ﬂow, locked at a ﬁxed rate K. This can be achieved by contracting a forward-rate agreement, therefore exchanging the ﬂoating interest rate to the ﬁxed rate K. Thus, the ﬁrm is, in its future calculation, independent of the evolution of the term structure. 49

Here we use the fact that the exponential-aﬃne model exhibits the Markov ÊTˆ

ability. Thus, the expectation E

Q

e− t r(xs ) ds

resented as the iterated expectation E EQ e

−

ÊT t

r(xs ) ds

to time T .

P xT , T, Tˆ

Q

e

−

ÊT t

= P xt , t, Tˆ

r(xs ) ds

E

QT

e

−

ÊTˆ T

can be repr(xs ) ds

=

, where the inner expectation is made with respect

3.2 Derivatives with Unconditional Payoﬀ Functions

35

Another point, we want to mention is the special strike value K = KF RA for which the yield-based forward-rate agreement becomes a fair zero value at time t. This value is commonly referred to as the forward rate and corresponds then to the simply-compounded rate Tˆ EQ e

−

r(xs ) ds t

−1 −1 P xT , T, Tˆ

τˆP xt , t, Tˆ P (xt , t, T ) − P xt , t, Tˆ = τˆP xt , t, Tˆ 1 P (xt , t, T ) − 1 . = τˆ P xt , t, Tˆ

KF RA =

(3.7)

Most of the time a ﬁrm does not want to insure itself against a ﬂoating interest payment for only one time period. For example, the ﬁrm has to serve a debt contract, which is linked to a LIBOR interest rate. In this case, the ﬁrm possibly wants to reduce its risk exposure due to the ﬂoating interest accrues over time and it is desired to make an exchange of interest payments for several successive time periods, where in each period the payment for the relevant ﬂoating rate is exchanged with a ﬁxed rate K. This task can be achieved buying a receiver swap contract.

Deﬁnition 3.2.4 (Swap). A forward-starting interest-rate receiver swap is deﬁned as a portfolio of forward-rate agreements for diﬀerent time periods Ta+1 − Ta with Ta ∈ T and t < Ta for a = 1, . . . , A on the same strike rate K. The payments of the contract are made at dates T2 , . . . , TA , whereas the contract is said to reset the ﬂoating rate at dates T1 , . . . , TA−1 . Due to the instantaneous character of the ﬂoating rate based swap contract, the payment and reset dates coincide. Hence, the swap contract in this case, with nominal principal N om and A payment dates contained in the vector T, can be represented as

36

3 Theoretical Prices of European Interest-Rate Derivatives

SW Ar (xt ,K, N om, t, T) Ta A − r(xs ) ds = EQ e t (K − r (xTa )) N om a=1

= N om

A

EQ e

−

Ta

r(xs ) ds

t

(K − r (xTa ))

a=1

= N om

K

A

P (xt , t, Ta )

a=1

−

A

(3.8)

EQ e

−

Ta t

r(xs ) ds

r (xTa ) .

a=1

The equivalent representation for a swap contract, exchanging a yield-based ﬂoating rate at A − 1 payment dates paid in-arrears is then SW AY (xt , K, N om, t, T) Ta+1 A−1 − r(xs ) ds = EQ e t (K − Y (xTa , Ta , Ta+1 )) τˆa+1 N om a=1

= N om× A−1

EQ e

−

Ta

r(xs ) ds

t

((K τˆa+1 + 1)P (xTa , Ta , Ta+1 ) − 1) (3.9)

a=1

= N om

A−1

(K τˆa+1 + 1) P (xt , t, Ta+1 ) − P (xt , t, Ta )

a=1

= N om

P (xt , t, TA ) − P (xt , t, T1 ) +K

A−1

τˆa+1 P (xt , t, Ta+1 ) ,

a=1

with τˆa+1 = Ta+1 − Ta . In contrast to the total number of A swap payments in equation (3.8), where these payments refer merely to speciﬁc time dates, for the yield-based swap contracts we have to consider A − 1 time periods, which explains the resulting summation term in equation (3.9). Subsequently, a swap contract can be interpreted as the sum of successive forward-rate agreements.

3.2 Derivatives with Unconditional Payoﬀ Functions

37

Similar to forward-rate agreements we are able to introduce the terminology of a special strike KS , which makes the yield-based swap contract a fair zero valued contract. This special strike is then denoted as the swap rate and can be represented in the case of a yield-based swap as A−1 (P (xt , t, Ta ) − P (xt , t, Ta+1 )) KS = a=1 A−1 ˆa+1 P (xt , t, Ta+1 ) a=1 τ P (xt , t, T1 ) − P (xt , t, TA ) = A−1 . ˆa+1 P (xt , t, Ta+1 ) a=1 τ

(3.10)

The last contract with unconditional exercise right which we include in the pricing methodology used is an Asian-type average-rate contract based on the ﬂoating rate r (xt ). These contracts do not belong to the class of traded derivatives in any exchange. However, this type of interest-rate derivative seems to be quite popular in over-the-counter markets50 . Asian contracts belong to the ﬁeld of path-dependent derivatives. Thus, the payoﬀ consists not only of the terminal value of the underlying rate at maturity but of the complete sample path over the averaging period. Deﬁnition 3.2.5 (Unconditional Average-Rate Contract). An unconditional average-rate agreement concluded in time t guarantees its holder the right at maturity T to exchange the continuously measured average of the ﬂoating rate r (xt ) over the period T − t against a ﬁxed strike rate K. The value of this diﬀerence is then scaled by a nominal principal N om. Hence, the price of this contract is given as U ARCr (xt , K, N om, t, T ) T =EQ e

−

r(xs ) ds t

K − 1 T −t

T

r(xs ) ds N om

t

T

− 1 =N om P (xt , t, T ) K − EQ e t τ

r(xs ) ds

T

(3.11)

r(xs ) ds .

t

Consequently, in contrast to the forward-rate agreement according to equation (3.5), where the sole expectation of r(xT ) played the major part, we are 50

See Ju (1997).

38

3 Theoretical Prices of European Interest-Rate Derivatives

interested in the discounted expectation of the integral of r(xt ) over the time to maturity at this point.

3.3 Derivatives with Conditional Payoﬀ Functions In the last subsection, we considered the pricing formulae for contracts with unconditional exercise at maturity under the risk-neutral measure Q. Obviously, these contracts can be expressed e.g. in terms of zero bonds and some constants. In this section we want to derive general pricing formulae for contracts with conditional or optional exercise rights at maturity. These derivatives contracts are therefore often referred to as option contracts. Basically, we are interested in calculating the particular option prices with underlying contracts of the form (3.5), (3.6), and (3.9) with optional exercise rights. Basically, the particular pricing formulae can be separated into zero bond and coupon-bond options, respectively, can be seen as a portfolio of several zerobond options in case of a yield-based swap contract. Hence, we begin the introduction with option contracts written on a zero bond.

Deﬁnition 3.3.1 (Zero Bond Option). We deﬁne a zero-bond call (put) option as a contract giving its holder the right, not the obligation, to buy (sell) a zero bond P xt , t, Tˆ for a strike price K at time T . The remaining time to maturity of this zero bond at the exercise date of the option is then given as τˆ. Formally, the price of a zero-bond call can be obtained as ZBC xt , K, t, T, Tˆ T − r(xs ) ds max P xT , T, Tˆ − K, 0 = EQ e t (3.12) T + − r(xs ) ds P xT , T, Tˆ − K , = EQ e t whereas a zero-bond put option can be calculated as T + − r(xs ) ds . ZBP xt , K, t, T, Tˆ = EQ e t K − P xT , T, Tˆ

(3.13)

3.3 Derivatives with Conditional Payoﬀ Functions

39

Zero bond options can be used to price two contracts commonly used to hedge interest-rate risk. Namely, we want to introduce cap and ﬂoor contracts. In this terminology, a cap contract is meant to hedge upside interest-rate risk exposure. This is often required for a ﬁrm which holds some debt position with interest payments on a ﬂoating rate base and fears that future interest rates are rising. So it wants the interest rate capped at some ﬁxed level, in order to limit its risk position due to this ﬁxed rate. In contrast to the above introduced forward-rate agreement or swap, a ﬁrm can now both participate on advantageously low interest rates and simultaneously cap its interest payments against high rates. The opposite eﬀect can be observed, if an institution or ﬁrm has outstanding loans based on a ﬂoating rate. In this case the ﬁrm is interested in limiting the downside risk, since low ﬂoating rates correspond to low interest payments. The contract with the desired properties is then a ﬂoor, where interest payments are exchanged under an agreed ﬁxed rate.

Deﬁnition 3.3.2 (Cap and Floor Contract). A cap (ﬂoor) contract is deﬁned as a portfolio of caplets (ﬂoorlets) for diﬀerent time periods Ta+1 − Ta with Ta ∈ T and t < Ta for a = 1, . . . , A on the same strike rate K. The payments of the contract are made at dates T2 , . . . , TA , whereas the contract is said to reset the ﬂoating rate at dates T1 , . . . , TA−1 . Due to the short rate, the character of the ﬂoating rate based swap contract, the payment and reset dates coincide. Hence, the model price of a caplet with nominal principal N om and A payment dates contained within the vector T, is then given by CP Lr (xt , K, N om, t, Ta ) = EQ e

−

Ta

r(xs ) ds

t

+ (r (xTa ) − K) N om. (3.14)

The price of a cap contract, as a simple summation of caplets for diﬀerent times Ta ∈ T, can then be represented as CAPr (xt , K, N om, t, T) =

A

CP Lr (xt , K, N om, t, Ta )

a=1

= N om

A a=1

EQ e

−

Ta t

r(xs ) ds

(r (xTa ) − K)+ .

(3.15)

40

3 Theoretical Prices of European Interest-Rate Derivatives

Subsequently, we have for a ﬂoor the pricing formula F LRr (xt , K, N om, t, T) T a A − r(xs ) ds + = N om EQ e t (K − r (xTa )) .

(3.16)

a=1

The particular yield-based cap and ﬂoor options, exchanging, if exercised, arbitrary yields with a ﬁxed rate K at A − 1 payment dates, are given by CAPY (xt , K, N om, t, T) T a A−1 + N om − r(xs ) ds ˜ a − P (xTa , Ta , Ta+1 ) K = EQ e t ˜a K a=1 =

A−1 a=1

(3.17)

˜ a , t, Ta , Ta+1 N om , ZBP xt , K ˜a K

and F LRY (xt , K, N om, t, T) T a A−1 + N om − r(xs ) ds ˜a P (xTa , Ta , Ta+1 ) − K = EQ e t ˜a K a=1

= ˜a = with K

(3.18)

˜ a , t, Ta , Ta+1 N om , ZBC xt , K ˜a K a=1

A−1

1 τˆa+1 K+1 .

Deﬁnition 3.3.2 shows that a cap or ﬂoor contract is just the summation of their legs, the caplets and ﬂoorlets, respectively. Especially for the more realistic case of yield-based contracts, we can identify the similarity to zerobond options, since contract prices can be obtained as the summation of these options. The yield-based options are said to be at the money if the modiﬁed strike ¯ a is equal to equation (3.10). A cap is therefore in the money if the rate K ¯ a > KS it is out of the money. modiﬁed strike rate is less than KS , and for K The opposite results hold for a ﬂoor contract. Furthermore, we can conclude that holding a cap contract long and a ﬂoor contract short, both with the

3.3 Derivatives with Conditional Payoﬀ Functions

41

same contract speciﬁcations, we are able to replicate a swap contract. This can be easily justiﬁed comparing the payoﬀ of such a portfolio given for a yield Y (ˆ τa+1 ), which is then (Y (xTa , Ta , Ta+1 ) − K)+ − (K − Y (xTa , Ta , Ta+1 ))+

(3.19)

= Y (xTa , Ta , Ta+1 ) − K,

and the corresponding swap payment. Taking the discounted expectation of the sum of terms in equation (3.19) for all periods, we have the equivalent swap contract. A more challenging contract in calculating model prices is a coupon-bond option. This option is only exercised if the coupon-bond price at maturity exceeds the strike K. Hence, we have to apply the maximum operator to the discounted sum of all outstanding coupon payments and the strike price. This is in contrast to the other option contracts mentioned above, where we applied the maximum operator to each term of the sum separately. Deﬁnition 3.3.3 (Coupon-Bond Option). A coupon-bond call (put) option is deﬁned as the right but not the obligation to buy (sell) a coupon bond CB(xT , c, t, T) with payment dates Ta ∈ T, with Ta > T for a = 1, . . . , A and strike price K. The price of a coupon-bond call option is given by T CBC (xt , c, K, t, T, T) = EQ e = EQ e

−

T

r(xs ) ds

A

t

−

r(xs ) ds t

(CB (xT , c, T, T) − K)+ +

P (xT , T, Ta ) ca − K

(3.20)

,

a=1

and the corresponding coupon-bond put option is given by T CBP (xt , c, K, t, T, T) = EQ e = EQ e

−

T

r(xs ) ds t

K−

−

r(xs ) ds t

A

(K − CB(xT , c, T, T))+

+ . P (xT , T, Ta )ca

(3.21)

a=1

Since the maximum operator is not distributive with respect to sums, the term inside the maximum operator in equation (3.20) and (3.21) cannot be

42

3 Theoretical Prices of European Interest-Rate Derivatives

decomposed easily without making further assumptions. Another popular option we want to discuss is an option on a swap contract or as shorthand often referred to as a swaption. With a swaption one can choose at the maturity of the option if it is advantageous to enter the underlying swap contract or otherwise leave the option unexercised. Deﬁnition 3.3.4 (Swaption). We deﬁne a forward-starting swaption as a contract conferring the right, but not the obligation to enter a forward starting receiver swap at maturity T . The particular underlying receiver swap contract is deﬁned according to deﬁnition 3.2.4, with T1 ≥ T . Formally, the yield-based forward-starting receiver swaption for an underlying swap with A − 1 payment periods is given as SW PY (xt , K, N om, t, T, T) T = EQ e Q

=E

e

−

r(xs ) ds t

−

+ (SW AY (xT , K, N om, T, T))

T

r(xs ) ds t

K

A−1

(3.22)

P (xT , T, Ta+1 )ˆ τa+1

a=1

+

+ P (xT , T, TA ) − P (xT , T, T1 )

N om.

Typically, the swaption maturity coincides with the ﬁrst reset date of the underlying swap contract. Thus, a yield-based receiver swaption with T1 = T , can be equivalently represented as a coupon-bond call option SW PY (xt , K, N om, t, T1 , T∗ ) = CBC (xt , cSW P , 1, t, T1 , T∗ ) , with

cSW P

and new time dates

K τˆ2 K τˆ3 × N om, = .. . 1 + K τˆA

T2

T3 T∗ = .. . . TA

(3.23)

3.3 Derivatives with Conditional Payoﬀ Functions

43

Subsequently, we reduce the valuation problem of a swaption to the calculation of an equivalent coupon-bond option with strike one, a coupon vector cSW P and a vector with payment dates T∗ . According to the unconditional contract deﬁned in equation (3.11), we are also able to price an average-rate option contract. The deﬁnition of the model price of an average-rate option is given below.

Deﬁnition 3.3.5 (Average-Rate Option). An average-rate cap option gives its holder the right, but not the obligation to exchange at expiration a ﬁxed strike rate K, over the period T − t, against the continuously measured average of the short rate r (xt ). Formally, the price of an average-rate cap option can be obtained as ARCr (xt , K, N om, t, T ) + T T − r(xs ) ds 1 r (xs ) ds − K N om. =EQ e t τ

(3.24)

t

Consequently, we have for an average-rate ﬂoor the pricing formula ARFr (xt , K, N om, t, T ) + T T 1 − r(xs ) ds K− r (xs ) ds N om. =EQ e t τ

(3.25)

t

Asian options show the advantageous ability to exhibit reduced risk positions in comparison to ordinary options because of the time-averaging of the underlying price process. Moreover, asian option contracts are more robust against price manipulations since the option payoﬀ includes the sample path over a ﬁnite time period. These options are not standard instruments traded on exchanges. However, they are popular over-the-counter contracts used by banks and corporations to hedge their interest-rate risk over a time period51 . For all theoretical option prices presented in this section, we give in Section 5.3 the corresponding pricing formulae which have to be used in a numerical 51

See, for example, Ju (1997).

44

3 Theoretical Prices of European Interest-Rate Derivatives

scheme. Thus, we distinguish between the calculation of a portfolio of options, e.g. used for the pricing of cap and ﬂoor contracts and as a special case for zero-bond options, respectively, and the computation of options on a portfolio which is the case for coupon-bond options and swaption contracts. This is done because only in case of a one-factor interest-rate process semi closedform solutions for swaptions and coupon bonds can be calculated.

4 Three Fourier Transform-Based Pricing Approaches

4.1 Overview Interest-rate derivatives are widely used instruments to cover possible interestrate risk exposures. However, to model the term structure more realistically, sophisticated models are required. One way to enhance the capability of the term-structure model is to incorporate more stochastic factors, by, for instance, incorporating a stochastic mean and/or a stochastic volatility, or modeling the term structure with help of an additive interest-rate process. Another way, which would especially enrich the model with the ability to reﬂect price shocks, lies in implementing jump components in the shape of diﬀerent Poisson processes with arbitrary stochastic jump amplitudes. Unfortunately, in most cases the pricing of derivatives securities, while incorporating for the underlying interest-rate process both features mentioned above, can only be accomplished with ineﬃcient Monte-Carlo simulations. Hence, more eﬃcient methods are needed to circumvent these time-consuming calculations. As shown in the prominent work of Heston (1993), a way out of this dilemma is achieved by using Fourier Transformation techniques. Doing this, we only need to solve one standardized inversion integral to evaluate the distribution function and then compute the desired derivative prices. The astonishing fact of the approach applied by Heston (1993) is that this Fourier-based valuation technique is independent of the underlying stochastic dynamics of the shortrate process and can be applied as long as the particular characteristic function

46

4 Three Fourier Transform-Based Pricing Approaches

exists52 . Bakshi and Madan (2000) generalized this method to interpret the characteristic function itself as a derivative contract with a trigonometric payoﬀ53 . Zhu (2000) derived various pricing formulae for options with underlying stock prices, where stochastic interest rates, volatilities and jumps were included in a modularized manner. There, the stochastic factors are integrated by parts and the author ends up with a system of ordinary diﬀerential equations, which then has to be solved. In this thesis, we go a step further and, by using the transform methods of Lewis (2001), are able to generalize the modular aspect of Fourier-based derivatives pricing into parts of the underlying stochastic behavior and the contract type. This enables us to present valuation techniques, which can be adapted to every desired European-style contract without greater eﬀort, assuming that the generalized Fourier Transformation of the payoﬀ function exists in closed form. We consider the general exponential-aﬃne model introduced in Section 2.1 for the short rate r (xt ) and derive a ﬂexible valuation procedure according to the approach given in Lewis (2001). Although we focus in our thesis on the exponential-aﬃne setup, we are also able to extend the framework to incorporate non-aﬃne term-structure models54 , such as the Longstaﬀ (1989) model or the class of quadratic Gaussian models as discussed in Beaglehole and Tenney (1992)55 and Filipovic (2001), respectively. All we need in the underlying model speciﬁcation is the exponential separability of the coeﬃcients in the general characteristic function. However, in applying these non-aﬃne model speciﬁcations, we have to ignore the possibility of jumps for non-aﬃne factors in order to avoid mixture terms in the fundamental partial diﬀerential equation, which would subsequently render the pricing procedure unattainable56 . 52

53 54 55

56

Due to our pricing framework we can relax this restriction to the existence of a system of ordinary diﬀerential equations. This methodology is covered in Section 4.2. See Chapter 10. In fact, the model of Longstaﬀ (1989) can be represented as a quadratic Gaussian model as shown in Beaglehole and Tenney (1992). The same holds for the term-structure model in Cheng and Scaillet (2004) where the terminology of a linear-quadratic jump-diﬀusion model is introduced. Despite the name, jump parts are only valid for linear factors, whereas the quadratic part is not allowed to bear jump parts. This issue is discussed in Section 9.3.

4.1 Overview

47

The outline of this chapter is as follows. We start with the comparison of three state of the art Fourier Transformation methodologies used in derivatives research. The Fourier-transformed Arrow-Debreu securities pricing approach is based on the work of Heston (1993)57 . Afterwards, we present the transform methodology as proposed by Carr and Madan (1999) and then discuss the generalized derivatives pricing setup of Lewis (2001), which display similarities in the derivation of the model price of a contingent claim. Both approaches focus on the Fourier Transformation of the payoﬀ function, whereas Carr and Madan (1999) apply the transform for the strike value, Lewis (2001) does a Fourier Transformation with respect to the state variable. Nevertheless, we provide an extension of the work in Lewis (2001), since we consider a multi-factor environment. One important diﬀerence between the pricing approach utilizing Fourier-transformed Arrow-Debreu securities, according to Heston (1993), the Carr and Madan (1999) methodology, and the method of Lewis (2001) is that the latter two approaches do not need to invoke Fourier Transformations for every single term in the pricing formula. Therefore, the transformation is applied on the entire contingent claim, which in a numerical sense is more efﬁcient. Additionally, the these two approaches provide a more stable solution due to the freedom of choosing a contour path for the integration parallel to the real axis in the inversion formulae58 . Generally, the derivatives we want to price are written on some functional of the underlying stochastic vector process xt , say g(xt ). Contingent claims on the short rate and on the yield are European-style derivatives and therefore pay only at maturity T a payoﬀ G (xT ). The solution of the pricing problems we seek then takes the following form.

Deﬁnition 4.1.1 (General Valuation Problem for European-Style Derivatives). We deﬁne the general valuation problem of a contract V (xt , t, T ) as the time T expectation of some (discounted) payoﬀ function G (xT ) under the risk-neutral probability measure Q, formally deﬁned as 57

Recent work with further development and uniﬁcation was made in Duﬃe, Pan and Singleton (2000), Bakshi and Madan (2000) and especially on the ﬁeld of

58

interest-rate derivatives in Chacko and Das (2002). See Carr and Madan (1999) and Lewis (2001).

48

4 Three Fourier Transform-Based Pricing Approaches

V (xt , t, T ) = E e Q

−

T

r(xs ) ds t

G (xT )

=

(4.1) G (xT ) p(xt , xT , w0 , w, t, T ) dxT .

RM

The contract can only be exercised at maturity T .

Apart from the underlying stochastic dynamics, the solution to equation (4.1) depends on how xT is incorporated within the payoﬀ function G (xT ). Thus, we follow Chacko and Das (2002) and distinguish for didactical purposes between payoﬀ functions which can be either linear, exponential-linear or integro-linear in xt . These idealized payoﬀ types are illustrated in Table 4.1 below59 .

Table 4.1. Idealized call option payoﬀ functions Payoﬀ type

G (xT )

Linear

G (xT ) = (g (xT ) − K)+

Exponential-linear

Integro-linear

G (xT ) = eg(xT ) − K

G (xT ) =

T t

+

g (xs ) ds − K

+

In contrast to option-pricing models written on equities, where constant interest rates are often assumed, in calculating equation (4.1), we are confronted with a more diﬃcult situation. Since both the discount factor and the payoﬀ function G (xT ) depend on the same stochastic process, we are not able to evaluate these expectations separately and multiply them afterwards60. We have 59 60

In case of unconditional payoﬀ functions, we use the same classiﬁcation. This is a direct consequence of the choice of numeraire made in Section 2.3.

4.2 Heston Approach

49

to consider that both expressions are obviously not independent and therefore have to derive the solution of equation (4.1) under their joint stochastic dynamics. However, thanks to the fact that the discount factor itself has an exponential-aﬃne representation61, we are still able to use the general characteristic function ψ(xt , z, w0 , w, g0 , g, τ ) in derivatives pricing. Consequently, equation (4.1) is the starting point for all of the following derivatives pricing approaches.

4.2 Heston Approach Pricing derivatives, using Fourier-transformed Arrow-Debreu securities and state prices, respectively, was introduced in Heston (1993). Since then, several articles utilizing Fourier Transformations in derivatives pricing have been published. Among others we want to mention, because of their relevance, Duﬃe, Pan and Singleton (2000) and Bakshi and Madan (2000). In the article of Duﬃe, Pan and Singleton (2000), a comprehensive survey is provided as to how this Fourier inversion methodology can be used to solve derivative prices for general stochastic dynamics. On the other hand, Bakshi and Madan (2000) oﬀer a rigorous survey, of how Fourier-transformed Arrow-Debreu securities can be used to span the underlying market and to price derivative prices. In principle, both articles use the same pricing mechanism, shown below62 . The basic principle behind the pricing approach with transformed ArrowDebreu securities is that all derivatives based on the interest rate r(xt ) described by equation (4.1) have to solve the same partial diﬀerential equations (2.31) and (2.33) for futures-style contracts, respectively. The only diﬀerence between them is that they need to satisfy diﬀerent terminal conditions. This statement holds also for the discounted probability density and the characteristic function of the interest-rate process. Therefore, they can be interpreted as hypothetical contingent claims solving the above-mentioned partial diﬀerential equations. Whereas derivative prices and probability densities are often 61

62

One can easily validate this statement by solving equation (2.30) and (2.34) and setting z equal to zero. In the context of interest-rate derivatives, Chacko and Das (2002) used this methodology to price the diﬀerent payoﬀ structures as given in Table 4.1.

50

4 Three Fourier Transform-Based Pricing Approaches

hard to obtain, due to their discontinuous terminal conditions63 , the solution for the particular general characteristic function can be recovered, even if jump components are encountered in the stochastic vector process xt . This is due to a special ability of characteristic functions; their terminal condition is inﬁnitely diﬀerentiable and smooth, which make them, from a mathematical point of view, more tractable.

Deﬁnition 4.2.1 (Arrow-Debreu Security). We deﬁne an Arrow-Debreu security as a contingent claim paying one unit of money at maturity T if and only if a speciﬁed state A occurs. The value AD(xt , t, T ) of an Arrow-Debreu security under probability measure Q∗ at time t is then given by AD(xt , t, T ) = EQ∗ [1A ] .

(4.2)

The expression 1A denotes the indicator function for the event A in time T , which is unity if the state A occurs and zero otherwise.

To demonstrate the pricing methodology, we consider the following example of a European call option with a linear payoﬀ function G (xT ) = (g (xT ) − K)+ and g (xT ) is given in equation (2.28)64 . The solution for this option can then be represented as V (xt , t, T ) = EQ e

−

r(xs ) ds t

= EQ e

T

−

T

r(xs ) ds t

− KEQ e 63

64

+ (g (xT ) − K)

−

g (xT ) 1g(xT )≥K

T

r(xs ) ds t

(4.3)

1g(xT )≥K ,

For many underlying stochastic dynamics, the solutions cannot be calculated in closed form. The derivation of option-pricing formulae for exponential-linear and integro-linear payoﬀ structures diﬀers slightly from the derivation of the theoretical option price formula of a linear payoﬀ function as given in this section. The derivation of the particular solutions for these payoﬀ functions can be looked up in Chacko and Das (2002), Sections 2 and 3.

4.2 Heston Approach

51

where the expectation is separated into parts. However, the expectations in equation (4.3) are not yet Arrow-Debreu securities in the sense of deﬁnition 4.2.1. These expressions still lack some sort of standardization to guarantee the outcome of one monetary unit. Thus, we need to apply the unconditional expectations65 T T EQ e

−

r(xs ) ds t

g (xT )

EQ e

and

−

r(xs ) ds t

= P (xt , t, T ).

Expanding the terms in equation (4.3) with their particular unconditional counterparts, we get

V (xt , t, T ) = EQ e

−

T

r(xs ) ds t

g (xT ) ×

− r(x ) ds s t g (x ) 1 e T g(xT )≥K Q T E EQ e− t r(xs ) ds g (xT ) T

e − KP (xt , t, T )EQ

−

(4.4)

T

r(xs ) ds t

1g(xT )≥K P (xt , t, T )

= Π0 (xt , t, T )Π1 (xt , t, T ) − KP (xt , t, T )Π2 (xt , t, T ). Obviously, the normalized functions Π1 (xt , t, T ) and Π2 (xt , t, T ) are two contingent claims and can be interpreted as Arrow-Debreu securities66 . On the other hand, Π1 (xt , t, T ) can be interpreted as the discounted forward price of the underlying contract. Introducing two artiﬁcial changes of measure deﬁned through the Radon-Nikodym derivatives, we get −

T

−

r(xs ) ds

dQ1 e t g(xT ) = dQ Π0 (xt , t, T )

and

T

r(xs ) ds

dQ2 e t = . dQ P (xt , t, T )

Consequently, we express the above call option price in terms of the particular Arrow-Debreu prices, which is 65 66

See Chacko and Das (2002), p. 205. In the last equation of (4.4), we adopted the notation given in Chacko and Das (2002).

52

4 Three Fourier Transform-Based Pricing Approaches

V (xt , t, T ) =Π0 (xt , t, T )EQ1 1g(xT )≥K − KP (xt , t, T )EQ2 1g(xT )≥K .

(4.5)

Obviously, in calculating the option price in equation (4.5), we need only the general characteristic function with terminal condition eızg(xT ) and its derivative with respect to z, respectively. However, calculations within this pricing framework for the particular functions Πi (xt , t, T ) are quite diﬀerent for linear, exponential-linear and integro-linear payoﬀ versions of G (xT )67 . Thus, only P (xt , t, T ) remains unchanged, since this quantity is completely independent of the characteristic payoﬀ part g (xT ). Recalling the formal structure of the general characteristic function in (2.30) and the connection between the moment-generating and characteristic function68 , we are able to express Π0 (xt , t, T ) with the help of the derivative of the general characteristic function with respect to the frequency parameter z, evaluated at z = 0, which is given by69 T Π0 (xt , t, T ) = EQ e =

−

r(xs ) ds t

1 d Q E e ı dz

−

g (xT ) T

r(xs ) ds t

ızg(xT ) e

(4.6) z=0

ψz (xt , 0, w0 , w, g0 , g, τ ) . = ı Here, the subscript denotes partial diﬀerentiation with respect to z 70 . Taking into account the exponential-aﬃne structure of the general characteristic function in (2.34), we are able to write equation (4.6) alternatively as 67 68 69 70

See Chacko and Das (2002). See Proposition 2.4.3. Compare with Theorem 1 (c) in Bakshi and Madan (2000). The result in equation (4.6) is always real, see e.g. Bakshi and Madan (2000). Therefore, the operator Re [. . .] in this calculation is not necessary at all, which can be justiﬁed by checking that all imaginary parts in this equation cancel out except in the term ıg (xT ).

4.2 Heston Approach

53

ψz (xt , 0, w0 , w, g0 , g, τ ) ψ(xt , 0, w0 , w, g0 , g, τ ) = × ı ı d ln [ψ(xt , z, w0 , w, g0 , g, τ )] dz z=0 ψ(xt , 0, w0 , w, g0 , g, τ ) φz (xt , 0, w0 , w, g0 , g, τ ). = ı In the last equation, we used the function φ(xt , z, w0 , w, g0 , g, τ ), which is just the natural logarithm of ψ(xt , z, w0 , w, g0 , g, τ ) in our exponential-aﬃne model setup. Thus, the derivative with respect to z of the exponent of the characteristic function is then ˜ z (z, τ ) xt + ıg (xt ) . φz (xt , z, w0 , w, g0 , g, τ ) = az (z, τ ) + b Using the same technique as before, we obtain the value of an ordinary zero bond as T T − r(xs ) ds − r(xs ) ds = E e t eızg(xT ) P (xt , t, T ) =EQ e t (4.7) z=0

=ψ(xt , 0, w0 , w, g0 , g, τ ). Finally, we are left with the calculation of the Arrow-Debreu prices. As mentioned before, these functions Π1 (xt , t, T ) and Π2 (xt , t, T ) can also be interpreted as probabilities. Hence, we apply a tool to determine probabilities from characteristic functions. This can be done with a Fourier inverse transform as proposed in Gil-Pelaez (1951). Theorem 4.2.2 (Inversion Theorem of Gil-Pelaez). If ψ xT (xt , z, t, T ) is the characteristic function of a one-dimensional stochastic variable xt then the probability Pr(xT ≥ K), given some state xt and some constant K, can be calculated as 1 1 Pr(xT ≥ K) = + 2 π

∞ 0+

ψ xT (xt , z, t, T )e−ızK Re ız

dz,

(4.8)

with z ∈ R. The expression 0+ in equation (4.8) denotes the right-sided limit to the origin. Obviously, the integrand is not deﬁned for a zero-valued transformation

54

4 Three Fourier Transform-Based Pricing Approaches

variable z 71 . Note that the inversion theorem in 4.2.2 is not limited to recover only probabilities for the case of symmetric probability density functions, which might be implicated due to the term 12 . Equation (4.9) holds for general probability distributions. The only condition to be satisﬁed is the existence of the characteristic function or its system of ODEs. Moreover, we are also able to use Theorem 4.2.2 for the linear combination g(xt ), as long as the outcome is a scalar random variable. As long as we are able to obtain the general characteristic functions ψ1 (xt , z, w0 , w, g0 , g, τ ) and ψ2 (xt , z, w0 , w, g0 , g, τ ) corresponding to the particular measures Q1 and Q2 , we are able to compute the values of Π1 (xt , t, T ) and Π2 (xt , t, T ). In analogy to equations (4.6) and (4.7), and keeping the normalization made in (4.4) in mind, we therefore have ψ1 (xt , z, w0 , w, g0 , g, τ ) =

ψz (xt , z, w0 , w, g0 , g, τ ) , ıΠ0 (xt , t, T )

ψ2 (xt , z, w0 , w, g0 , g, τ ) =

ψ(xt , z, w0 , w, g0 , g, τ ) . P (xt , t, T )

and

Subsequently, the values of the required Arrow-Debreu securities can be calculated as72 1 1 Π1,2 (xt , t, T ) = + 2 π

∞ Re 0+

ψ1,2 (xt , z, w0 , w, g0 , g, τ )e−ızK ız

dz.

(4.9)

Although the derivation of option prices within this methodology is comprehensible, this technique does entail some drawbacks. Firstly, a general advantage which holds for all pricing methodologies based on Fourier Transformation techniques is that we are not restricted to simple stochastic dynamics of the underlying short-rate process, where the probability density function p(xt , xT , w0 , w, t, T ) is explicitly known in closed form73 . With the continuum of characteristic functions at hand, we are able to calculate option prices for a much broader class of stochastic dynamics. Despite the apparent elegance of this approach, there are also some issues to discuss. Since we expressed the 71 72 73

More on this topic and residue calculus is discussed in Section 4.3. Compare with the general result in Bakshi and Madan (2000), Theorem 1. However, there exist density functions for which no characteristic function exists, e.g. a log-normal distributed random variable.

4.3 Carr-Madan Approach

55

option price as a decomposition of probabilities multiplied with their normalization factors, we have to calculate for a sum of N terms in G (xT ) the same number of separate Fourier inversions and therefore to perform N numerical integrations. Especially in one-factor interest-rate models, this fact can be avoided using a Fourier transform with respect to rT 74 . From a computational point of view, this can be very time consuming and therefore ineﬃcient compared to the pricing approaches of Carr and Madan (1999) and Lewis (2001). Additionally, the denominator in the integrand of equation (4.9) decays only linearly for the idealized payoﬀ functions, compared to the payoﬀ-transform approaches discussed in the subsequent sections75 . Another matter we want to address is the integration procedure itself. In equation (4.8), we need to consider carefully the pole at the origin. Sometimes, this can lead to rather unstable results. Another point to mention is that the structure of the option contract dictates the calculation procedure of the particular function Πj (xt , t, T ). Hence, it ﬁrst has to be determined whether the payoﬀ function G(xT ) exhibits linear, exponential-linear or integro-linear terms of g (xT )76 , which result in diﬀerent valuation formulae for the option price. This can complicate unnecessarily the computation of option prices in contrast to the approaches discussed in the following sections, where Fourier Transformations of the payoﬀ function are used.77 .

4.3 Carr-Madan Approach Carr and Madan (1999) develop a diﬀerent method for retrieving option prices using characteristic functions. Instead of applying general characteristic functions to obtain the exercise probabilities and the Arrow-Debreu security prices 74 75

See, for example, the pricing of coupon bonds in Section 5.3.3. The denominator in the payoﬀ transforms of the interest-rate option contracts in table 4.1 are quadratic and therefore have a higher rate of convergence. Compare

76

with the particular transformations given in Section 5.3. See Chacko and Das (2002) for a comprehensive discussion and classiﬁcation of payoﬀ functions and derivation of the particular option prices in this transformed

77

Arrow-Debreu security framework. See Bakshi and Madan (2000), pp. 218-220, cases 1-3, on how to derive the particular ψj (xt , z, w0 , w, g0 , g, τ ) for general payment structures. Chacko and Das (2002) also derive the respective valuation algorithms for these payoﬀ structures.

56

4 Three Fourier Transform-Based Pricing Approaches

under the particular probability measures Q1,2 as done in the last section, they propose an alternative approach. The intention behind this framework is to formulate a valuation procedure, that can incorporate the FFT, a very eﬃcient tool in deriving Fourier Transformations for diﬀerent values of the underlying random variable. However, they ﬁrst perform a Fourier Transformation on the payoﬀ function with respect to the strike variable K. Afterwards, interchanging the order of integration, they are able to compute the desired fair price of the option as an inverse Fourier Transformation, thus applying the relevant characteristic function, an example is given below. Obviously, a ﬁrst advantage of this strategy is that, since we deal with only one transform operation on the option price, in order to compute model price we need only one inverse transformation. As the authors mention, a closed-form solution of the option price in Fourier space is presupposed78 . Since option prices commonly have at least two terms in the payoﬀ function G (xT ), numerical calculations with this method are approximately twice as fast. A problem in this approach mostly arises if a Fourier transform on the payoﬀ function with a real-valued frequency variable z ∈ R is applied. As mentioned in Bakshi and Madan (2000), the transformed payoﬀ function would not exist at all, due to the unbounded option payoﬀ functions79 . To circumvent this issue, Carr and Madan (1999) introduce an artiﬁcial dampening parameter α and derive a modiﬁed transformed option price, upon which they apply the inverse transformation procedure. In the following presentation of this methodology we do not refer to an artiﬁcial dampening parameter α; rather we want to introduce a general Fourier Transformation as deﬁned in deﬁnition 2.4.1 with z ∈ C. Moreover, we show that the dampening parameter coincides with the negative ﬁxed imaginary part zi of the frequency variable z = zr + ızi . Following this trail, we get a more intuitive concept of the nature of the dampening factor α used by Carr and Madan (1999). Demonstrating the pricing technique, we rely on the same contract type as in (4.3) with G (xT ) = (g (xT ) − K)+ to maintain the comparability to 78

See Carr and Madan (1999), p. 61. We extend this methodology to allow for characteristic functions with no closed-form representations. This topic is discussed

79

in Chapter 6. See Bakshi and Madan (2000), p. 215. An exception would be a contract which is bounded on two sides, e.g. a butterﬂy contract.

4.3 Carr-Madan Approach

57

previously obtained solutions of our example in equation (4.5). Starting with a Fourier Transformation on the payoﬀ function with respect to K, we have ∞

∞

F K [G (xT )] =

+

eızK (g (xT ) − K) dK

eızK G (xT ) dK = −∞

−∞

g(x T)

eızK (g (xT ) − K) dK

= −∞

= −e =−

e

ızK 1

+ (g (xT ) − K)ız z2

(4.10)

g(xT ) −∞

ızg(xT )

z2

with Im(z) < 0.

The restriction in equation (4.10) upon the imaginary part of z guarantees the ﬁniteness of the transformed payoﬀ function. Thus, we are able to interpret (4.10) as a line integral, which is evaluated parallel to the real axis going through ızi . Apart from considerations about the regularity of the payoﬀ transform, the value of zi can also be used to optimize numerical accuracy of the valuation algorithm80 . Exploiting the symmetry of real-valued Fourier transforms, the payoﬀ function G (xT ) for our speciﬁc example, can be expressed by the following inverse transformation problem 1 G (xT ) = − π

∞

e−ızK

eızg(xT ) dz. z2

(4.11)

0

Carrying out this inverse operation, we need zi to be ﬁxed on the same strip used for the transformation. Otherwise, the original function and its image function in dual space would not correspond to each other81 . The essential part, in expressing the valuation formula as an inverse Fourier-style problem, is the interchanging of the integration order. Furthermore, we have in equation (4.11) an exponential term for both the underlying 80

See Lee (2004), for a comprehensive analysis of the eﬀect of zi on the accuracy of the computational result. Note, the derived error bounds in this article are only valid for one particular strike. These results have to be treated carefully for algorithms, where option prices for diﬀerent strike rates, such as ITM, ATM, and

81

OTM options, are computed simultaneously. This fact is discussed in Section 2.4.

58

4 Three Fourier Transform-Based Pricing Approaches

stochastic variable and the strike rate enabling the application of the characteristic function methodology and afterwards to calculate prices with the FFT. Denoting our exemplary valuation problem of (4.3) in terms of (4.11), we get the following integral representation + V (xt , t, T ) = (g (xT ) − K) p(xt , xT , w0 , w, t, T ) dxT 1 =− π

RM

RM

∞

ızg(xT ) e e−ızK dz p(xt , xT , w0 , w, t, T ) dxT . z2

(4.12)

0

Due to Fubini’s theorem, the order of integration can be interchanged82 . Therefore, we are able to use the alternative representation 1 V (xt , t, T ) = − π

∞

e−ızK z2

eızg(xT ) p(xt , xT , w0 , w, t, T ) dxT dz RM

0

"

#$

%

eqn. (2.30) 1 =− π

∞

e−ızK

(4.13)

ψ(xt , z, w0 , w, g0 , g, τ ) dz. z2

0

Eventually, we get the Fourier-style valuation formula for the price at time t of a European call option, based on the payoﬀ function G (xT ) = (g(xT ) − K)+ . The relationship between the artiﬁcial dampening factor α in Carr and Madan (1999) and zi becomes apparent if we substitute z = zr +ızi in equation (4.13), which gives 1 V (xt , t, T ) = − π

∞ 0

ezi K =− π

e−ı(zr +ızi )K ∞ 0

ψ(xt , zr + ızi , w0 , w, g0 , g, τ ) dzr (zr + ızi )2 (4.14)

ψ(xt , zr + ızi , w0 , w, g0 , g, τ ) e−ızr K dzr . zr2 + 2ızr zi − zi2

Obviously, compared to the corresponding option price formula in Lee (2004), it can easily be veriﬁed that the identity zi ≡ −α holds83 . 82

83

Since all parts of the integral are real-valued, we are able to change the order of integration without any problems. The modiﬁed transformed option price for our example is also given in Lee (2004) Theorem 4.2 as cˆα,G2 (u), where u matches zr . Also compare this result with the general Fourier-style valuation formula in Carr and Wu (2004), p. 136.

4.3 Carr-Madan Approach

59

In contrast to the Heston pricing approach, the Carr-Madan methodology provides an additional degree of freedom, since we are no longer limited to the case of a real-valued transformation variable z. This is of major importance in a numerical scheme for computing derivative prices84 . Furthermore, we are able to shift the integration contour around any existing pole. However, in these cases the residue of the particular pole must be taken into account85 . Proceeding like this, the accuracy of the valuation algorithm can be drastically increased86 . Nevertheless, we are also free to choose the imaginary part in (4.14), such that the contour integrals have to be performed right through a pole. Doing this we ﬁrst consider the residuals of the poles and then evaluate the integral due to Cauchy’s theorem87 . Generally, the advantage in this approach lies in the availability of a fast numerical integration routine, the FFT algorithm. A properly set procedure, based e.g. on our example in (4.14), can calculate a vast number of derivative prices for alternative strike rates in fractions of a second. On the other hand, Fourier-style solutions in this framework cannot be properly decomposed into parts of the general characteristic function and the transformed payoﬀ function88 . Thus, we needed a speciﬁc payoﬀ function in the derivation of the transformed option price. It would be more convenient and from a numerical perspective more desirable if the integral in (4.14) could be clearly separated into a part of the general characteristic function, which depends on the underlying stochastic dynamics, and a part determined by the contract we want to price. Moreover, there seems to exist a problem for particular models with speciﬁc parameter constellations89 . Finally, we do not prefer this methodology in the ﬁrst place because it cannot be properly applied for coupon-bond 84 85 86

The choice of the optimal value of zi is discussed in Section 6.3.3. See Lee (2004) equations (6) and (7). This can be validated by Tables 2 and 3 in Lee (2004). The error bounds presented there are up to a thousand times lower, if the integrals are evaluated on contours

87 88

89

with no existing poles. In the next section, we derive valuation formulae using diﬀerent values of zi . For example, the transformed option price in equation (4.13)

is

0 ,g,τ ) . − ψ(xt ,z,w0z,w,g 2 Itkin (2005) analyzed the FFT method of Carr and Madan (1999) for the case

of an underlying Variance-Gamma process and reports some numerical issues for diﬀerent lengths of time to maturity τ .

60

4 Three Fourier Transform-Based Pricing Approaches

options and swaptions, respectively, with an underlying one-factor interestrate process. The reason for this is that we need the exercise boundary to be explicit in rT in order to present the valuation formula in terms of the characteristic function. If this is not the case, we lose characteristics of the stochastic process, which are relevant in the valuation formula and therefore have to be considered within the integration. For example, in the case of coupon-bond options, we encounter the problem of determining numerically a critical value rT∗ 90 , thus making it impossible to compute the particular option prices. These problems can be circumvented with the approach discussed in the following section.

4.4 Lewis Approach Lewis (2001) presented in his work an alternative way to retrieve not only option prices, but general derivatives prices91 . The approach is similar to the previously discussed methodology of Carr and Madan (1999), but can be applied to a wider area of pricing problems. Thus, we are able to calculate all derivatives prices presented in Chapter 3 with a single general valuation formula. Fortunately, within this framework, it is also possible to use an eﬃcient numerical tool to compute derivative prices with comparable speed to the FFT algorithm, namely the IFFT algorithm. In contrast to the approach in Carr and Madan (1999), Lewis (2001) introduced a derivatives pricing framework starting with a Fourier Transformation of the payoﬀ function, but this time with respect to the underlying stochastic variable, where the frequency parameter z ∈ C is also supposed to be complex-valued. Thus, the advantages discussed in the last section still hold. As before, our starting point is the payoﬀ function G (xT ) of a derivatives contract. As in the previous section, the Fourier Transformation is performed on the payoﬀ function, in this case with respect to the scalar g (xT ). Accordingly, the transformed payoﬀ function is 90 91

See Jamshidian (1989). As mentioned before, the methodology was ﬁrstly used in Lewis (2000). However, we refer to Lewis (2001) because of the more detailed derivation and comprehensive discussion of this pricing framework.

4.4 Lewis Approach

61

∞ F

g(xT )

eızg(xT ) G(xT ) dg (xT ) .

[G(xT )] =

(4.15)

−∞

To guarantee the ﬁniteness of the integral in equation (4.15) and the existence of F g(xT ) [G (xT )], respectively, the imaginary part of z has to be restricted, where its domain depends on the speciﬁc contract. Continuing with our example in pricing an interest-rate cap of the form G (xT ) = (g (xT ) − K)+ , we ﬁrst calculate the transformed payoﬀ function with respect to g (xT ) as ∞ F

g(xT )

eızg(xT ) (g (xT ) − K)+ dg (xT )

[G(xT )] =

(4.16)

−∞

=−

e

ızK

z2

with Im(z) > 0. Although this formula bears a strong resemblance to equation (4.10), one remarkable diﬀerence between them is the interval of zi , for which the Fourier transform of the particular payoﬀ function exists92 . Another point we would like to mention is that the transformed payoﬀ function displays the strike rate K in the exponential function instead of g (xT ), according to the methodology of Carr and Madan (1999). Representing the time t option price with the help of the transformed payoﬀ function, we have at the general valuation formula93 ∞ 1 e−ızg(xT ) F g(xT ) [G (xT )] dz × V (xt , t, T ) = π RM

0

(4.17)

p(xt , xT , w0 , w, t, T ) dxT , which is for our speciﬁc example of an interest-rate cap 92 93

In comparison to equation (4.10), zi has to be negative. Again, we take advantage of the symmetry of Fourier Transformations for realvalued functions.

62

4 Three Fourier Transform-Based Pricing Approaches

=−

1 π

RM

∞

e−ızg(xT )

e

ızK

z2

dz p(xt , xT , w0 , w, t, T ) dxT .

(4.18)

0

Again, we apply Fubini’s theorem, implicating the possibility of interchanging the order of integration in (4.17). Thus, for general payoﬀ functions we obtain 1 V (xt , t, T ) = π

∞ F g(xT ) [G (xT )] × 0

(4.19)

e−ızg(xT ) p(xt , xT , w0 , w, t, T ) dxT dz.

RM

Firstly, we focus on the inner integral. In line with the formal deﬁnition of the characteristic function, according to equation (2.27), we are able to establish the relation

eı(−z)g(xT ) p(xt , xT , w0 ,w, t, T ) dxT (4.20)

RM

=ψ(xt , −z, w0 , w, g0 , g, τ ). Inserting this result into equation (4.19), we eventually get the general version of the Fourier-style valuation formula 1 V (xt , t, T ) = π

∞ F g(xT ) [G (xT )] ψ(xt , −z, w0 , w, g0 , g, τ ) dz,

(4.21)

0

which is for our example of a call contract with underlying variable g (xT ), 1 − π

∞

eızK ψ(xt , −z, w0 , w, g0 , g, τ ) dz z2

0

with Im(z) > 0.

In contrast to the pricing procedure introduced by Carr and Madan (1999), we have a strict separation of functionals, which depend either on the contract type or on the underlying stochastic dynamics. The respective part for the contract type is therefore represented by the transformed payoﬀ function, whereas

4.4 Lewis Approach

63

the stochastic dynamics of the underlying process is implemented in terms of the characteristic function. Hence, we have a real modular pricing framework, in which each part in (4.21) can be exchanged without greater eﬀort. Moreover, we can apply this methodology consistently to contracts, whether they are unconditionally exercised or bear an optional exercise right94 . In particular, for one-factor models with multiple jump components, we are able to take advantage of the fact that for most contracts the domains of zi are overlapping. This means that zi can be chosen arbitrarily, subject to compliance with numerical accuracy95. Thus, we usually have to evaluate ψ(xt , −z, w0 , w, g0 , g, τ ) only once for diﬀerent values of zr . Afterwards, these precomputed values can be used for all relevant contract types needed. This drastically improves the eﬃciency of the numerical valuation scheme. The payoﬀ-transform approach according to Lewis (2001) is extremely versatile. For example, with this pricing technique, we can also derive the quantities Π1 (xt , t, T ) and Π2 (xt , t, T ), without need of any derivative function ψz (xt , z, w0 , w, g0 , g, τ ), as done in formula (4.9). Although the numerical integration on a line integral (partly) including a pole exhibits the undesirable numerical properties discussed earlier, we want to show the derivation of the Gil-Pelaez style valuation formulae for Π2 (xt , t, T ), as given in Theorem 4.2.2 within the Lewis methodology96 , for demonstration purposes. Recalling that the payoﬀ of an Arrow-Debreu security can be formally represented by the indicator function, we apply a Fourier Transformation on this special function in order to calculate Π2 (xt , t, T ). Under the probability measure Q2 , the simple payoﬀ representation is then given by the incomplete Fourier Trans94 95

This is demonstrated in the next chapter. In addition to the restrictions for zi , due to the validity for the transformed payoﬀ function, in some cases we need to restrict the domain for the imaginary part of the transformation variable further to ensure the regularity of the characteristic function. One example, where zi has an additional constraint due to this issue is the characteristic function for the variance gamma process which is discussed in

96

Itkin (2005). In contrast to equation (4.9), we would get an alternative representation for Π1 (xt , t, T ), without needing any derivative of ψ(xt , z, w0 , w, g0 , g, τ ) and φ(xt , z, w0 , w, g0 , g, τ ), respectively.

64

4 Three Fourier Transform-Based Pricing Approaches

formation97

eızK F g(xT ) 1g(xT )>K = − ız

(4.22)

with Im(z) > 0. Using this formula, together with zi in the appropriate domain, we are almost ready to calculate Π2 (xt , t, T ). In fact, we consider the residue theorem and apply a suitable closed-contour integral to recover the exact formula according to equation (4.9). Hence, evaluating the integral including the pole at zi = 0 gives the desired result, which is demonstrated below. ˜ 2 (t, T ) to compensate for the We start with a slightly modiﬁed function Π inﬂuence of the probability law Q2 98 , which is deﬁned as ˜ 2 (xt , t, T ) = Π2 (xt , t, T )P (xt , t, T ). Π

(4.23)

Inserting the transformed payoﬀ function (4.22) into our general valuation formula (4.21) gives ˜ 2 (xt , t, T ) = − 1 Π π

∞

eızK ψ(xt , −z, w0 , w, g0 , g, τ ) dz, ız

(4.24)

0

with Im(z) > 0. Equation (4.24) can already be used for valuation purposes. Since we want to show the similarity of this formula to the transformed Arrow-Debreu security pricing approach, we encounter the problem of integrating through a pole, and therefore must apply Cauchy’s residue theorem for analytic functions99 .

Theorem 4.4.1 (Cauchy’s Residue Theorem). Assume the function f (z) is analytic within a closed, counter-clockwise performed integration contour C, 97

98

99

One-sided Fourier Transformations are commonly referred to as incomplete Fourier Transformations. This has to be done, since we use the general characteristic function ψ(xt , z, w0 , w, g0 , g, τ ). This means, the function has to satisfy the Cauchy-Riemann equations. See Duﬀy (2004), p. 16.

4.4 Lewis Approach

65

except at points zd ∈ C, where f (zd ) encounters singularities. Then the value of the closed contour integral for this function can be calculated as & f (z) dz = 2ıπ Res [f (z)|z = zd ] .

(4.25)

d

C

The residues at the singularities corresponding to points zd can be derived as Res [f (z)|z = zd ] = lim

z→zd

dn−1 1 [(z − zd )n f (z)] . (n − 1)! dz n−1

(4.26)

The parameter n represents the order of the pole. Hence, if we want to evaluate the integral in (4.24) for Im(z) = 0, we have to deal with a simple pole of order n = 1. To facilitate the calculations, we ˜ 2 (xt , t, T ) ﬁrst introduce the original, two-sided integral representation for Π in the manner of equation (2.25), which is simply ˜ 2 (xt , t, T ) = − 1 Π 2π

∞ −∞

eızK ψ(xt , −z, w0 , w, g0 , g, τ ) dz. ız

(4.27)

Proceeding like this, we add to the former line integral, which has to be evaluated parallel to the real axis with distance Im(z), several additional integral paths to build a rectangular shape on the upper imaginary half-plane100 . This gives us a contour C, which is performed, as illustrated in Figure 4.1. Setting

∞ ˜ 2 (xt , t, T ) = Π

f (z) dz,

(4.28)

−∞

with

eızK ψ(xt , −z, w0 , w, g0 , g, τ ) , 2πız we are able to express the contour integral as f (z) = − & f (z) dz = C 100

6 j=1 C

j

f (z) dz =

6

Ij .

(4.29)

j=1

In manipulating equation (4.27), we could also have chosen the lower half-plane. Subsequently, we would then have to be careful about the direction, how the pole is encircled, making its contribution to the integration either in a positive or negative sense.

66

4 Three Fourier Transform-Based Pricing Approaches

˜ 2 (xt , t, T ) in Fig. 4.1. Clockwise performed integral path for the derivation of Π equation (4.27) on the real line. The cross represents the pole.

Referring to Figure 4.1, the integral part I4 forms a half arc around the pole of the meromorphic function101 with radius . Thus, excluding the pole, we can state, due to Cauchy’s integral theorem, & f (z) dz = 0. (4.30) C

In the next step, we need to determine the values of the speciﬁc integrals ˜ 2 (xt , t, T ) given in equation Ij . Starting with I1 , we have just the value of Π (4.27). Recognizing that for 0 ≤ Im(z) < ∞, we have lim f (R + ızi ) = 0, R→±∞

we immediately obtain I2 + I6 = 0.

(4.31)

Subsequently, we are left with the computation of the remaining integral parts I3 , I4 and I5 . According to Theorem 4.4.1, if we consider an arc performed in a counter-clockwise fashion around a pole, we would have to take into account the entire contribution of the pole. Therefore, by assuming the radius of the half arc I4 to be inﬁnitesimally small, we eventually obtain half of the particular contribution. Thus, we have to consider the residue 101

A function f (z) is said to be meromorphic, if it only has some isolated singularities. This means that such a function is analytic everywhere, except at these poles. See Duﬀy (2004), p. 16.

4.4 Lewis Approach

67

I4 = ıπRes [f (z)|z = 0] = ıπ lim zf (z) z→0

P (xt , t, T ) ψ(xt , 0, w0 , w, g0 , g, τ ) =− . =− 2 2

(4.32)

Likewise, assuming the distance to the origin for integrals I3 and I5 to be inﬁnitesimally small, we are able to represent them in the limit as102 0+ I3 + I5 =

−∞ f (z) dz + f (z) dz.

∞

(4.33)

0−

Having derived the required expressions for all integral parts Ij in equations (4.31), (4.32) and (4.33), additionally using equation (4.30), we eventually end ˜ 2 (xt , t, T ), which is given by up with an alternative representation for Π ˜ 2 (xt , t, T ) = − (I3 + I4 + I5 ) Π + −∞ 0 P (xt , t, T ) = − f (z) dz + f (z) dz 2 ∞

P (xt , t, T ) −2 = 2

0−

(4.34)

−∞ f (z) dz 0−

In equation (4.34), the symmetry of characteristic functions for real-valued functions is exploited, due to Proposition 2.4.3. Therefore, the two integrals in the above equation can be aggregated. In a last step, we reinsert the detailed expression of f (z) and substitute z ∗ = −z. This results in the relation ˜ 2 (xt , t, T ) = P (xt , t, T ) Π 2

∞ −ız∗ K e 1 ψ(xt , z ∗ , w0 , w, g0 , g, τ ) Re + dz ∗ , π ız ∗

(4.35)

0+

with Im(z ∗ ) = 0. Dividing equation (4.35) by P (xt , t, T ) and considering only the relevant real part of the solution, we obtain the Heston-style solution of equation (4.9), 102

Here, we use again the convention 0± denoting the right- and left-hand sided limit towards zero.

68

4 Three Fourier Transform-Based Pricing Approaches

which we intentionally wanted to reproduce with the payoﬀ-transformation approach of Lewis (2001). In contrast to the Fourier-transform approach introduced in Carr and Madan (1999), the methodology discussed above is not that popular. One reason might be that the FFT algorithm cannot be applied to the valuation formula. Albeit, simply using an IFFT algorithm provides equivalent functionality and eﬃciency in solving derivatives prices. On the other hand, we prefer the method of Lewis (2001) because of the clear separation of diﬀerent valuation components in the pricing formula. Additionally, this framework enables us to consistently use the valuation formula presented in equation (4.21) for both unconditional and conditional derivatives contracts by using residue calculus. Moreover, with this methodology even swaptions and options on coupon bonds can be priced in case of one-factor interest-rate models.

5 Payoﬀ Transformations and the Pricing of European Interest-Rate Derivatives

5.1 Overview In this chapter we derive semi closed-form solutions of European interest-rate derivatives in terms of their transformed payoﬀ functions, for all contracts given in Chapter 3. Equipped with this frequency representation of the payoﬀ function, the contract can be priced with the general valuation formula according to equation (4.21). This procedure, combined with a standardized numerical integration routine, can then be used to compute the desired quantities. Apart from the generality of this method, we observe that all call and put option contracts exhibit identical payoﬀ representations in Fourier space. The diﬀerence between them are the diﬀerent strips in the imaginary plane, parallel to the real axis, on which the transform operation is valid for the particular contract. As before, we distinguish between contracts with unconditional and conditional exercise rights. The reason for this separation of the payoﬀ-transformed formulae is that contracts with unconditional exercise rights can be calculated as simple unconditional expectations. Using the residue theorem, solutions for the underlying contracts can be computed in terms of the general characteristic function, without evaluating numerically any integral at all. However, if the characteristic function is not known in closed form but can be represented as a system of ODEs, theoretical prices have to be numerically obtained via a Runge-Kutta algorithm. On the other hand, contracts with optional exercise rights are computed by numerical integration in every case.

70

5 Payoﬀ Transformations and European Interest-Rate Derivatives

5.2 Unconditional Payoﬀ Functions This section is organized as follows. First we compute some fundamental Fourier Transformations for functionals containing g (xT )103 , henceforth referred to as building blocks. These blocks, combined with the particular characteristic function, can then be used to compute the contract prices of Section 3.2 in the form of Fourier-style valuation formulae via equation (4.21). In calculating the payoﬀ transform, we do not have to pay attention to the question of whether the derivative to be priced is a normal or futures-style contract. This is captured by the choice of the relevant characteristic function, which can be either ψ(xt , z, w0 , w, g0 , g, τ ) or ψ(xt , z, 0, 0M , g0 , g, τ ). At ﬁrst sight, a problem arises in pricing unconditional interest-rate derivatives, due to the unbounded integration range of the expectation. As shown in the option-pricing example in Sections 4.3 and 4.4, the imaginary part of z can be used to ensure the existence of the payoﬀ transform by suﬃciently dampening the integral on one side, which could be either the upper or lower. Unfortunately, the dampening eﬀect cannot be accomplished simultaneously on both integration boundaries. Thus, we need additional considerations in order to derive an appropriate representation of the valuation formula in frequency space. Nevertheless, after some manipulation of the transformed payoﬀ, we derive in the upcoming section the particular valuation formulae. 5.2.1 General Results We begin with two basic interest-rate derivatives, the zero-bond contract as deﬁned in equation (4.7) and the expectation of g (xT ) as given by equation (4.6). According to Section 4.2, the value of a zero bond equals ψ(xt , 0, w0 , w, g0 , g, τ ) whereas the latter quantity can be obtained via the calculation of its ﬁrst derivative. These general results hold for arbitrary linear combinations g (xt ). In contrast, the payoﬀ-transformation technique as presented in Section 4.4 seems at ﬁrst sight to have diﬃculties in recovering these particular expectations, due to the unbounded integration domain. Hence, the ﬁrst step in this subsection is to prove the former results obtained 103

Although not explicitly displaying the variable g (xT ) in the payoﬀ function, we also interpret in the following the Fourier Transformation of a constant as encountered in zero-bond contracts as a building block.

5.2 Unconditional Payoﬀ Functions

71

in equations (4.6) and (4.7) and therefore show that the payoﬀ-transform methodology can be applied without exceptions. If we set G(xT ) = 1, which represents the riskless return of one unit money at maturity, it seems at ﬁrst that the ordinary payoﬀ transform is no longer ﬁnite. Unfortunately, with help of the imaginary part of the transformation variable z, we are only able to dampen the integrand on one side, which can be either in the direction of the positive or the negative real half-plane. Thus, we cannot dampen the underlying payoﬀ function for both sides simultaneously, and consequently cannot perform the inverse Fourier Transformation on the same strip in the imaginary plane. However, performing the integration on diﬀerent strips in the imaginary plane, we are again able to use the payoﬀtransform methodology. Dividing the integration domain (−∞, ∞) into two separate subdomains (−∞, ε) and (ε, ∞) with arbitrary ε ∈ R, we end up with two frequency functions deﬁned on diﬀerent strips in the imaginary plane. At ﬁrst glance, this seems to complicate the situation. In fact, with the help of Cauchy’s residue theorem, the calculations are rather simpliﬁed. The payoﬀ transform of an ordinary zero bond can be calculated as104 , ε ∞ eızε eızε g(xT ) ızg(xT ) F − , (5.1) [1] = e dg (xT ) + eızg(xT ) dg (xT ) = ız ız −∞

ε

with Im(z) < 0, and z representing the complex conjugate of z 105 . Working with this transformed payoﬀ function, we are already able to recover the zero-bond price due to the integral representation ∞ ızε e 1 P (xt , t, T ) = ψ(xt , −z, w0 , w, g0 , g, τ ) dz 2π ız −∞ (5.2) ∞ ızε e 1 ψ(xt , −z, w0 , w, g0 , g, τ ) dz. − 2π ız −∞

104

105

Obviously, in pricing a zero bond, the choice of g (xT ) is irrelevant. In fact, g (xT ) can be set to any value, since the payoﬀ function itself is independent of g (xT ). We make this assumption for convenience. Generally, the imaginary part of the transform variable used in the latter integral can be independently chosen on the positive half-axis.

72

5 Payoﬀ Transformations and European Interest-Rate Derivatives

Interchanging the integration boundaries of the latter integral in equation (5.2) and closing the contour with two additional paths from points (R, ızi ) to (R, −ızi ) for R → ±∞, thus forming a closed contour integral with the resulting four integrals, we are able to use Cauchy’s residue theorem again. The rectangular contour including the singularity is shown in Figure 5.1. Due to the direction of the path, we have to consider a counter-clockwise encircled simple pole at z = 0, which is completely inside the contour. Consequently, the contour integral equals 2πıRes [f (z)|z = 0] with f (z) =

eızε ψ(xt , −z, w0 , w, g0 , g, τ ), 2πız

and the value of a zero bond is106 ∞ P (xt , t, T ) =

−∞ f (z) dz + f (z) dz = 2πıRes [ f (z)| z = 0]

−∞

(5.3)

∞

=ψ(xt , 0, w0 , w, g0 , g, τ ). Here, the calculations for the residue are analogous to the ones made in equation (4.32), but this time considering the entire residue. The same result would have been obtained using the Dirac Delta function δ(z) in the transformed payoﬀ function. It is a well-known result that ∞ F

g(xT )

eızg(xT ) dg (xT ) = 2πδ(z),

[1] =

(5.4)

−∞

with Im(z) = 0. Hence, the fair value of a zero bond can be alternatively calculated as107 1 P (xt , t, T ) = 2π

∞ 2πδ(z)ψ(xt , −z, w0 , w, g0 , g, τ ) dz −∞

(5.5)

=ψ(xt , 0, w0 , w, g0 , g, τ ), 106

Starting from here, all zero-valued integrals are ignored.

107

Obviously, for arbitrary real-valued w, the relation

∞

−∞

holds.

δ(z − w)f (z) dz = f (w)

5.2 Unconditional Payoﬀ Functions

73

Fig. 5.1. Closed contour integral path for the derivation of P (xt , t, T ) in equation (5.2). The pole is completely encircled in a counter-clockwise manner.

which justiﬁes the above statement. So far, we have shown the result of one important building block, the model price of a zero bond, with the help of the payoﬀ-transform methodology. In order to price interest-rate contracts bearing unconditional exercise rights, we also need the expected value of the payoﬀ function G (xT ) = g (xT ) as given by equation (4.6). In the following, we want to prove this general result within the payoﬀ-transform methodology. Starting our calculations, we assume a linear payoﬀ function based on g (xT ) and then apply two incomplete Fourier Transformations, this time with an artiﬁcial integration boundary ε for the particular integrals. Hence, the transformed payoﬀ function of G (xT ) = g (xT ) can be calculated as

74

5 Payoﬀ Transformations and European Interest-Rate Derivatives

ε F

g(xT )

eızg(xT ) g (xT ) dg (xT )

[g (xT )] = −∞

∞ eızg(xT ) g (xT ) dg (xT )

+

(5.6)

ε

eızε (1 − ızε) eızε (1 − ızε) − , = z2 z2 with Im(z) < 0. This time, we build a rectangular integration path, performed in a clockwise manner which is depicted in Figure 5.2. Hence, we get for the discounted expectation108 T EQ e

−

r(xs ) ds t

g (xT )

1 =− 2π

−∞

∞ ∞

eızε (1 − ızε) ψ(xt , −z, w0 , w, g0 , g, τ ) dz z2

eızε (1 − ızε) ψ(xt , −z, w0 , w, g0 , g, τ ) dz z2 −∞

ızε e (1 − ızε) = −2πıRes − ψ(xt , −z, w0 , w, g0 , g, τ ) z = 0 . 2πz 2 (5.7) −

1 2π

Using again Cauchy’s residue theorem, the contribution of the pole at the origin109 can be derived as 108

109

According to the clockwise performed integration path, the contribution of the pole in this case is −2πı times the residue. According to a removable singularity, we have in fact at z = 0 two diﬀerent poles, a simple and a second order pole.

5.2 Unconditional Payoﬀ Functions

75

Fig. 5.2. Closed contour integral path for the discounted expectation of g (xT ). The pole is completely encircled in a clockwise manner.

eızε (1 − ızε)ψ (xt , −z, w0 , w, g0 , g, τ ) Res − z = 0 2πz 2 ızε

e ψ (xt , −z, w0 , w, g0 , g, τ ) =Res − z = 0 2πz 2 ızε

e ψ (xt , −z, w0 , w, g0 , g, τ ) ε + Res − z = 0 2πız ızε d e ψ (xt , −z, w0 , w, g0 , g, τ ) = lim − z→0 dz 2π ızε e ψ (xt , −z, w0 , w, g0 , g, τ ) ε + lim − z→0 2πı ψz (xt , 0, w0 , w, g0 , g, τ ) − ıψ (xt , 0, w0 , w, g0 , g, τ ) ε = 2π ψ (xt , 0, w0 , w, g0 , g, τ ) ε − 2πı ψz (xt , 0, w0 , w, g0 , g, τ ) = . 2π

(5.8)

76

5 Payoﬀ Transformations and European Interest-Rate Derivatives

Inserting this result in equation (5.7), we eventually obtain the general expression for the expected value of g (xT ), which is T EQ e

−

r(xs ) ds t

g (xT ) = −ıψz (xt , 0, w0 , w, g0 , g, τ ) (5.9) ψz (xt , 0, w0 , w, g0 , g, τ ) . = ı

Thus, we have also derived the result in equation (4.6) within the payoﬀtransform methodology. The remaining building block represents the unconditional expectation under the risk-neutral measure of an integro-linear variable where the payoﬀ T function satisﬁes G(xT ) = g(xs ) ds. Because of the integrated expression t

in the payoﬀ function, this quantity has to be treated diﬀerently. Pricing an unconditional contract, including such an integrated term, we are interested in the expected value EQ e

−

T

r(xs ) ds

T

t

g (xs ) ds .

(5.10)

t

In the following, we ﬁrst want to show how equation (5.10) can be recovered manipulating the expectation itself, as done in equations (4.6) and (4.7). Obviously, the calculations are very similar compared to equation (4.6). Afterwards, we show that the payoﬀ-transform methodology replicates the same result without any problems. Making the same considerations as for the derivation of the expected value of g (xT ), we compute (5.10) as the derivative with respect to the transform variable, evaluated at z = 0. Note that the characteristic function itself consists only of one sole exponential discounting term, since we have T T T EQ e

−

r(xs ) ds ız t

e

g(xs ) ds t

= EQ e

− (r(xs )−ızg(xs )) ds t

.

(5.11)

Obviously, this particular characteristic function is equivalent to the value of a zero-bond contract, but with a hypothetical complex-valued short rate of rA (xt , z) = r (xt ) − ızg (xt ) = (w0 − ızg0 ) + (w − ızg )xt .

(5.12)

5.2 Unconditional Payoﬀ Functions

77

In the last equation we considered that both the instantaneous interest rate r (xt ) and the payoﬀ-characterizing function g (xt ) are linear combinations of xt . Since we deal with a zero bond like contract, the solution for this model price also exhibits an exponential-aﬃne form. Thus, in analogy to the considerations made for zero bonds, we are able to represent the solution as an exponential-aﬃne function. Introducing new parameters characterizing the modiﬁed short rate, we have w0A (z) = w0 − ızg0

and wA (z) = w − ızg.

The resulting characteristic function for pricing average-rate derivatives is then ψ xt , z, w0A (z), wA (z), 0, 0M , τ , and the relevant payoﬀ function for this modiﬁed characteristic function is G(xT ) = 1. As mentioned above, this characteristic function exhibits a strong resemblance compared to the Fourier-style zero-bond representation in equation (5.3), where the original characteristic function was evaluated at some point z = 0. This can be traced back to the fact that both payoﬀ functions are independent of the Fourier Transformation variable. The diﬀerence between them is that the function ψ xt , z, w0A (z), wA (z), 0, 0M , τ generates zero-bond prices with respect to the modiﬁed short rate rA (xt , z), independently of the value of the transformation variable z. Thus, the coeﬃcient functions in this particular case, a(z, τ ) and b(z, τ ) solve again the system of ordinary differential equations (2.40) and (2.41), with terminal conditions a(z, 0) = 0, b(z, 0) = 0M . The hypothetical discount rate is deﬁned by w0A (z) and wA (z), respectively, whereas the terminal value is given by ψ xt , z, w0A (z), wA (z), 0, 0M , 0 = 1.

Having found the characteristic function for this special case, the same considerations can be applied as for the expected value of g (xT ). Using the technique of Fourier-transformed prices, we eventually express equation (5.11)

78

5 Payoﬀ Transformations and European Interest-Rate Derivatives

as110 EQ e

−

T

r(xs ) ds

T

t

t

T − (r(xs )−ızg(xs )) ds d Q t E g (xs ) ds = e dz z=0 ψz xt , 0, w0A (0), wA (0), 0, 0M , τ . = ı

(5.13)

Alternatively, we are also able to obtain this result using the payoﬀtransform methodology together with the contour integration technique. For convenience, we ﬁrst set up the substitution T γ(T ) =

g (xs ) ds, t

and afterwards perform the Fourier Transformation with respect to this new variable γ(T ). Thus, the transformation of the particular payoﬀ function is the same as the one used in deriving equation (5.6). Therefore, we can immediately adopt the result of equation (5.9) by exchanging the general characteristic function ψ (xt , z, w0 , w, g0 , g, τ ) with its modiﬁed pendant ψ xt , z, w0A (z), wA (z), 0, 0M , τ 111 . Afterwards, we get the desired result according to equation (5.13). In this section we proved the general results of unconditional expectations for zero bonds, and linear and integro-linear payoﬀ functions, respectively, obtained within the payoﬀ-transform framework112. Moreover, apart from the traditional formulae, where the desired value is derived by manipulation of the 110

Obviously, ψ

the

values

of

xt , z, w0A (z), wA (z), 0, 0M , τ

the

functions

ψz (xt , z, w0 , w, g0 , g, τ )

and

are equal for z = 0. However, the deriva-

tives with respect to z evaluated at this point, do not share this similarity. This is the reason why we make the dependence of z in the modiﬁed short rate 111

112

explicit, although w0A (0) = w0 and wA (0) = w. The path of the contour integral and the location of the pole is given in Figure 5.2. The particular derivation for the exponential-linear case was not derived in this section since it is not needed in this work. However, the calculations are straightforward using the integration-by-parts methodology, where the relevant pole is at z = ı.

5.2 Unconditional Payoﬀ Functions

79

expectation itself, as shown in Section 4.2, we have with the payoﬀ-transform approach the freedom to choose among a set of inﬁnite solution formulae due to the contour integration in the complex plane. This fact becomes especially important in computing the expectation EQ1 [1] and the expectation for the unconditional average-rate contract where the derivative of the characteristic function with respect to the transformation variable z has to be used. In these cases we are provided with the alternative to use the simple payoﬀ transform and apply equation (4.21) on the appropriate strip in the imaginary plane. Hence, using the building blocks above, we are able to price all interest-rate derivatives introduced in Section 3.2 with Fourier-style formulae. According to the results in equations (5.5), (5.9) and (5.13) we arrive at completely closed-form pricing formulae, which are illustrated in the next subsection113 .

5.2.2 Pricing Unconditional Interest-Rate Contracts So far, the three building blocks for general unconditional payoﬀ functions have been derived. In this section, these blocks are translated into the valuation formulae for the particular yield-based and level-based interest-rate contracts discussed in Section 3.2. Starting with yield-based contracts, we need ﬁrst a translation of yields into Fourier-style solutions. This is easily done as follows Y (xt , t, T ) =

ψ(xt , 0, w0 , w, g0 , g, τ )−1 − 1 . τ

(5.14)

The model price for zero bonds can then be obtained by using equation (5.5), whereas prices of coupon bonds can be calculated as CB(xt , c, t, T) =

A

ψ(xt , 0, w0 , w, g0 , g, τa )ca .

(5.15)

a=1

The price of a forward-rate agreement is given as 113

This statement is valid if the characteristic function or its derivative with respect to z can be displayed in closed form. In cases where the characteristic function cannot be explicitly expressed, but its coeﬃcient functions a(z, τ ) and b(z, τ ) are solutions to the system of ordinary diﬀerential equations according to (2.40) and (2.41), a Runge-Kutta algorithm can be used to obtain the relevant values.

80

5 Payoﬀ Transformations and European Interest-Rate Derivatives

F RAY (xt , K, N om, t, T, Tˆ ) ψ (xt , 0, w0 , w, g0 , g, τˆ) − ψ(xt , 0, w0 , w, g0 , g, τ ) , =N om ˜ K

(5.16)

and a yield-based swap can be similarly computed in terms of the general characteristic functions as SW AY (xt , K, N om, t, T) A−1 ψ (xt , 0, w0 , w, g0 , g, τa+1 ) =N om ˜a K a=1

−

A−1

(5.17)

ψ (xt , 0, w0 , w, g0 , g, τa ) .

a=1

On the other hand, pricing contracts linearly based on the function g (xT ), we foremost need the derivative of the general characteristic function ψ (xt , z, w0 , w, g0 , g, τ ) with respect to z. Hence, a level-based forward-rate agreement deﬁned in equation (3.5) is represented by F RAr (xt , K, N om, t, T ) ψz (xt , 0, w0 , w, g0 , g, τ ) =N om K ψ(xt , 0, w0 , w, g0 , g, τ ) − (5.18) ı φz (xt , 0, w0 , w, g0 , g, τ ) =N om K − ψ(xt , 0, w0 , w, g0 , g, τ ). ı Accordingly, the corresponding swap contract in this framework can be obtained as SW Ar (xt , K, N om, t, T) A =N om K ψ (xt , 0, w0 , w, g0 , g, τa ) a=1

−

A ψz (xt , 0, w0 , w, g0 , g, τa ) a=1

ı

A φz (xt , 0, w0 , w, g0 , g, τa ) =N om K− × ı a=1 ψ (xt , 0, w0 , w, g0 , g, τa ) .

(5.19)

5.3 Conditional Payoﬀ Functions

81

The last unconditional contract to be priced is the average-rate contract. Here, the integro-linear payoﬀ function can be interpreted as an interest-rate contract based on the short rate itself. According to equation (3.11) and (5.13), the price of this contract can be calculated as U ARCr (xt , K, N om, t, T ) =N om K ψ(xt , 0, w0 , w, g0 , g, τ ) −

ψz (xt , 0, w0A (0), wA (0), 0, 0M , τ ) ı

(5.20)

.

For the special case g (xT ) = r (xT ), we use the simpliﬁed versions w0A (z) = (1 − ız)w0 wA (z) = (1 − ız)w, respectively.

5.3 Conditional Payoﬀ Functions So far, we derived closed-form solutions for contracts with unconditional exercise rights. In contrast to the calculations in the last section, where contracts merely depended on the simple evaluation of the terms ψ(xt , 0, w0 , w, g0 , g, τ ), ψz (xt , 0, w0 , w, g0 , g, τ ) and ψz (xt , 0, w0A (0), wA (0), 0, 0M , τ ), respectively, the option-pricing problem confronts us with a diﬀerent situation. The integration by parts method is not of use anymore due to a natural integration boundary, characterized by some strike value K. Including this optional exercise right within the payoﬀ-transform methodology, we end up with some semi closedform solutions, which means we have to solve a standardized Fourier integral in order to compute the desired model prices of interest-rate options. Although the payoﬀ-transform methodology enables us to price consistently the option prices with payoﬀ functions according to Table 4.1, without adapting the valuation formula (4.21) to the diﬀerent cases, we distinguish for convenience between linear, exponential-linear and integro-linear payoﬀ functions. As before, we ﬁrst derive some basic payoﬀ transforms for general g (xT ) and afterwards take into account the interest-rate options discussed in Chapter 3. Eventually, we develop as a special case the Fourier-transformed payoﬀ function of a coupon-bond option for the case of a one-factor interest-rate model114 with xt = rt . 114

The term one-factor model refers to the fact that only one Brownian motion is incorporated in the model.

82

5 Payoﬀ Transformations and European Interest-Rate Derivatives

5.3.1 General Results Besides the elementary payoﬀ functions, we also diﬀerentiate between call and put options, because of the conditional exercise property of the contracts. The transformed payoﬀ functions for call and put contracts display a strong resemblance, which is demonstrated in this section, allowing a more general implementation of the valuation algorithms. Due to the exercise boundary and the diﬀerent ways of incorporating g (xT ) and its integro-linear counterpart, respectively, in the payoﬀ function G(xT ), we introduce the critical value ' ln[K] Exponential-linear Case. α(K) = (5.21) K Linear and Integro-linear Case, for which the option payoﬀ is exactly at the money. The Fourier Transformation for diﬀerent call payoﬀ structures can be generally represented as ∞ F

g(xT )

eızg(xT ) G(xT )1g(xT )≥α(K) dg (xT )

[G(xT )] = −∞ ∞

(5.22)

=

e

ızg(xT )

G(xT ) dg (xT ) ,

α(K)

whereas the particular put payoﬀ transform in its general form is given by ∞ F

g(xT )

eızg(xT ) G(xT )1g(xT )≤α(K) dg (xT )

[G(xT )] = −∞

(5.23)

α(K)

eızg(xT ) G(xT ) dg (xT ) .

= −∞

In deriving the solution for the exponential-linear case, we have to use the transform F

g(xT )

e

g(xT )

−K

+

∞ =

eızg(xT ) eg(xT ) − K dg (xT )

α(K)

Keızg(xT ) e(1+ız)g(xT ) − = 1 + ız ız (1+ız)α(K)

=

1+ız

K e = . ız(1 + ız) ız(1 + ız)

∞ α(K)

(5.24)

5.3 Conditional Payoﬀ Functions

83

Due to the exponential-linear dependence of the payoﬀ-characterizing variable we set α(K) = ln[K] and obtain the equivalent transformation as given in equation (2.26). Since the frequency representation of a call option payoﬀ only exists on a strip with Im(z) > 1, a general Fourier Transformation is needed. Although exhibiting diﬀerent payoﬀ structures the corresponding payoﬀ transform of a put option has the identical formal structure as given in equation (5.24). This can be easily proved by F

g(xT )

K −e

g(xT )

+

α(K)

eızg(xT ) K − eg(xT ) dg (xT )

=

(5.25)

−∞ 1+ız

=

K , ız(1 + ız)

but with Im(z) < 0. Based on this result, both call and put option prices can be recovered using the same payoﬀ transform and as a direct consequence, only one single program code is needed for evaluating values for both interest-rate option contracts. The only diﬀerence are the diﬀerent sets and strips on which Im(z) is valid for the inverse operation. Whereas the condition for the call contract assured the dampening of the integrand on the positive half-axis, we need for the put option the condition to guarantee the same on the negative equivalent. An interesting feature of the payoﬀ-transform methodology is, due to the equivalent transformed payoﬀ functions of calls and puts, the applicability of a closed contour integral to obtain in a very elegant way the particular put-call parity115 . Without loss of generality, we set f (z) =

K 1+ız ψ(xt , −z, w0 , w, g0 , g, τ ) . 2πız(1 + ız)

Thus, we have 115

The relevant integration path is depicted in Figure 5.3.

(5.26)

84

5 Payoﬀ Transformations and European Interest-Rate Derivatives

Fig. 5.3. Closed contour integral path for the derivation of the put-call parity in equation (5.27). The poles at z = 0 and z = ı are completely encircled in a clockwise manner.

EQ e

−

T

r(xs ) ds t

T + + − r(xs ) ds eg(xT ) − K − EQ e t K − eg(xT ) ∞

=

f (z) dz + −∞

−∞ f (z) dz

(5.27)

∞

= −2πı (Res [f (z)|z = 0] + Res [f (z)|z = ı]) , with Im(z) > 1. The imaginary part of the Fourier variable z in equation (5.27) can be chosen arbitrarily as long as the existence of the payoﬀ transformations is guaran-

5.3 Conditional Payoﬀ Functions

85

teed116 . Obviously, this clockwise performed contour integral now encircles two simple poles of the function f (z), one at the origin and the other one located at z = ı. Due to the closed contour, we only have to calculate the residues of all included poles in order to obtain the desired put-call parity. Comparing equation (5.26) with (5.5), the residue of f (z) at the origin is just Res [f (z)|z = 0] = K

ψ(xt , 0, w0 , w, g0 , g, τ ) , 2πı

whereas the residue at z = ı is Res [f (z)|z = ı] = −

ψ(xt , −ı, w0 , w, g0 , g, τ ) . 2πı

Hence, equation (5.27) equals T T + + − r(xs ) ds − r(xs ) ds EQ e t eg(xT ) − K − EQ e t K − eg(xT )

(5.28)

= ψ(xt , −ı, w0 , w, g0 , g, τ ) − K ψ(xt , 0, w0 , w, g0 , g, τ ). According to the result in equation (4.7), the term ψ(xt , 0, w0 , w, g0 , g, τ ) simply represents the price of a zero bond with maturity τ . The other term, the quantity ψ(xt , −ı, w0 , w, g0 , g, τ ) equals the discounted forward price of the exponential of the variable g (xt )117 . Therefore, setting z = −ı, we get T EQ e

−

r(xs ) ds t

eg(xT ) .

For a call option, linearly based on g (xT ), we get

∞ 1 + ız(K − g (xT )) + F g(xT ) (g (xT ) − K) = eızg(xT ) z2 α(K) eızα(K) eızK =− =− 2 , 2 z z

(5.29)

with 116

For convenience, we work with the complex conjugate for the latter integral. In fact, due to the exponential-linear payoﬀ function the restriction for the put

117

option transform can be independently chosen according to equation (5.25). See, for example, Bakshi and Madan (2000), p. 212. There, this quantity is alternatively denoted as the scaled-forward price.

86

5 Payoﬀ Transformations and European Interest-Rate Derivatives

Im(z) > 0. Similar to the call representation in Fourier space, the put option transform is eızK F g(xT ) (K − g (xT ))+ = − 2 . (5.30) z The only diﬀerence between the call and put option transform is that equation (5.30) is deﬁned on the opposite imaginary half-plane. Consequently, we use the complex conjugate of the Fourier variable in equation (5.29). The put-call parity for the linear case can be derived as118 T T EQ e

−

r(xs ) ds t

+ (g (xT ) − K) − EQ e

−

r(xs ) ds t

+ (K − g (xT ))

eızK ψ(xt , −z, w0 , w, g0 , g, τ ) = ıRes z = 0 z2 ψz (xt , 0, w0 , w, g0 , g, τ ) − K ψ(xt , 0, w0 , w, g0 , g, τ ). = ı

(5.31)

Due to the payoﬀ similarities of the linear and integro-linear case, the payoﬀ transformations are equivalent for both cases in Fourier space. Hence, to compute the average-rate option prices (3.24) and (3.25), equations (5.29), (5.30) and (5.31) can be used together with the modiﬁed characteristic function. Although not directly applicable for tradable option contracts, but nevertheless important for theoretical issues is the Fourier-transformed payoﬀ function of a hypothetical contingent claim according to the Dirac delta function, which is δ(g (xT ) − α(K)). As mentioned before, the Dirac delta function has an inﬁnite spike for g (xT ) = α(K). The Fourier Transformation of the Dirac delta function can be simply expressed as F g(xT ) [δ(g (xT ) − α(K))] = eızα(K) ,

(5.32)

with no need to set up any restriction on the imaginary part of the transform variable z. Since the Dirac delta function states the terminal condition of a probability density function, equation (5.32) may be used to recover the relevant transition density function. Especially for illustrating the behavior of a particular stochastic process g(xt ), the transition density function is useful to explain its characteristics. The other special function we want to derive, is 118

The relevant integration path is depicted in Figure 5.2.

5.3 Conditional Payoﬀ Functions

87

the Fourier Transformation of the cumulative probability function Pr(g(xT ) < α(K)). The payoﬀ corresponding to this terminal condition is given by the indicator function of the event g(xT ) < α(K). Thus, the transformed payoﬀ can be expressed as119 eızα(K) , F g(xT ) 1g(xT ) 1. In contrast, a zero-bond put option price can be derived via equation (5.35) but with the restriction Im(z) < 0.

According to equation (3.17) and (3.18), a yield-based cap and ﬂoor contract can be immediately expressed as the summation over the particular zero-bond options, scaled with some quantity NKom . Hence, the model price of a

a yield-based cap contract is 121

However, there exist some articles which derive approximated values for these contracts in a multi-factor framework, see e.g. Singleton and Umantsev (2002) or Collin-Dufresne and Goldstein (2002).

5.3 Conditional Payoﬀ Functions

CAPY (xt , K, N om, t, T) ız A−1 ∞ Ka N om × = π a=1 ız(1 + ız)

89

(5.36)

0

ψ(xt , −z, w0 , w, a(0, τˆa ), b(0, τˆa ), τa ) dz, with Im(z) > 0, τˆa = Ta+1 − Ta , and τa = Ta − t. Subsequently, a yield-based ﬂoor contract can be priced using equation (5.36) with Im(z) < 0.

Next, we derive the particular pricing formulae of level-based interestrate contracts and interest-rate options written on the short rate r (xt ) itself. Starting with a cap contract according to equation (3.15), we use the payoﬀ transform (5.29) with g0 = w0

and

g = w,

and therefore apply the characteristic function ψ(xt , z, w0 , w, w0 , w, τa ). Thus, the cap contract can be priced as CAPr (xt , K, N om, t, T) A ∞ N om eızK =− ψ(xt , −z, w0 , w, w0 , w, τa ) dz, π a=1 z2

(5.37)

0

with Im(z) > 0. Hence, the model price of a ﬂoor contract with equivalent input parameters can be recovered using equation (5.37) again but evaluating the integrals on the negative imaginary half-plane with Im(z) < 0.

90

5 Payoﬀ Transformations and European Interest-Rate Derivatives

The last option contracts for which we want to give a payoﬀ-transformed solution are the average-rate options due to equation (3.24) and (3.25). Thus, the payoﬀ of the average-rate cap option contract at expiration can be expressed as + T N om ∗ K − r(xs ) ds , τ t

with K ∗ = τ K. Taking the same considerations into account as done for the unconditional average-rate contract, the relevant characteristic function for T r(xs ) ds,

γ(T ) = t

is given by ψ(xt , z, w0A (z), wA (z), 0, 0M , τ ). Together with the payoﬀ transform in equation (5.29), we are able to postulate the model price of an average-rate cap as ARCr (xt , K, N om, t, T ) ∞ ızK ∗ (5.38) N om e A A =− ψ(x , −z, w (−z), w (−z), 0, 0 , τ ) dz, t M 0 τπ z2 0

with Im(z) > 0. The respective average-rate ﬂoor contract can be priced, using equation (5.38) with Im(z) < 0. 5.3.3 Pricing of Coupon-Bond Options and Yield-Based Swaptions So far, we have excluded the valuation formulae for coupon-bond options and yield-based swaptions, respectively. In contrast to the option contracts discussed in the last section, where we computed only a single option price and a portfolio of diﬀerent option prices, respectively, we deal here with a option on a portfolio of future cash ﬂows. Consequently, the determination of

5.3 Conditional Payoﬀ Functions

91

a unique critical exercise value α(K) in a multi-factor setting is not possible anymore122 . However, dealing with a one-factor interest-rate model setup with r(xt ) = rt 123 , we are able to circumvent this issue. Hence, we follow the technique proposed in Jamshidian (1989) to derive the theoretical price of a coupon-bond option using the payoﬀ-transform methodology in pricing this derivative contract, which is shown below. Setting xt = rt , we are able to exploit the coeﬃcient structure of the aﬃne term-structure model. The special form of the characteristic function is of the form T ψ(rt , z, 0, 1, 0, 1, τ ) = EQ e

−

rs ds+ızrT t

= ea(z,τ )+b(z,τ )rt .

Because a yield-based swaption can be interpreted as an option on a coupon bond124 , we focus on the valuation of the particular coupon-bond option. In a one-factor setup the coupon-bond call option payoﬀ is given by +

(CB(rT , c, T, T) − K) = =

A

+ P (rT , T, Ta )ca − K

a=1 A

+ e

a(0,τa )+b(0,τa )rT

ca − K

.

a=1

In the last equation, we inserted the particular Fourier-style zero-bond prices generated by the exponential-aﬃne model. The above presented payoﬀ function is then a continuous and strictly decreasing function in rT 125 . In these models we have126 ∂P (rt , t, T ) = b(0, τ ) < 0 ∂rt

∀ T > t.

Consequently, the payoﬀ function exhibits a unique zero value for the critical short rate rT∗ for which the coupon-bond call is exercised. However, dealing 122 123 124 125 126

See, for example, Singleton and Umantsev (2002). Without loss of generality, we set in the following w0 = 0 and w1 = 1. See the alternative presentation of a swaption payoﬀ in Section 3.3. The particular characteristic functions are derived in Chapter 8. See e.g. Duﬃe and Kan (1996) for the properties of b(0, τ ) in common one-factor interest-rate models.

92

5 Payoﬀ Transformations and European Interest-Rate Derivatives

with a single-factor environment, we cannot explicitly express this critical value rT∗ in closed form, which is due to the sum of exponentials in the payoﬀ function. Thus, the critical exercise value has to be computed numerically. Having determined the value of rT∗ , the Fourier Transformation of a couponbond call payoﬀ can be calculated as127 F rT (CB(rT , c, t, T, T) − K)+

∗

rT =

e −∞

=e

ızrT

∗ ızrT

A

e

a(0,τa )+b(0,τa )rT

ca − K

a=1 ∗ A ea(0,τa )+b(0,τa )rT

a=1

b(0, τa ) + ız

K ca − ız

drT

(5.39)

,

with Im(z) < min [b(0, τa )] . a

Note that in contrast to the valuation formula a zero-bond call option, where the Fourier Transformation of the payoﬀ function was made with respect to g(xT ), we now perform the transform operation with respect to rT . Therefore, we need a diﬀerent restriction for the imaginary part of the transform variable z. Because the coeﬃcient b(0, τa ) is generally negative, we take the smallest value of b(0, τa ) as an upper bound for the domain of valid values for Im(z), which is due to the monotonicity simply b(0, τA ). Eventually, using the general valuation formula (4.21), we are able to compute the price of a coupon-bond call option as CBC (rt , c, K, t, T, T) A ∞ ea(0,τa )+b(0,τa )rT∗ ∗ K 1 ızrT ca − × = e π b(0, τa ) + ız ız a=1

(5.40)

0

ψ(rt , −z, 0, 1, 0, 1, τ ) dz. As before, the payoﬀ transform of the particular put option is also given by equation (5.40), but with the slightly modiﬁed restriction Im(z) > 0. 127

Since the integration variable is no longer g (xT ), we have to switch the integration boundaries, due to the negativeness of b(0, τ ).

5.3 Conditional Payoﬀ Functions

93

Having derived the proper Fourier Transformation of a coupon-bond option payoﬀ, the equivalent expression for a yield-based swaption contract is given by the alternative representation of a swaption contract according to equation (3.23), with coupon payment vector cSW P and payment dates contained in T∗ . On the other hand, the particular forward-start payer swaption can be interpreted as a coupon-bond put option with strike one and the same coupon payment vector and the same payment dates as used before. Hence, for the transformed payoﬀ function to be existent, we have to ensure that the inequality Im(z) > 0 holds.

6 Numerical Computation of Model Prices

6.1 Overview In this chapter we develop a new pricing algorithm to compute model prices for the derivatives contracts previously discussed. Here, we distinguish, as before, between contracts with unconditional and conditional exercise rights. The distinction is made because of the separate fundamental calculation procedure for these prices. Whereas derivatives with unconditional exercise rights can be calculated in terms of the general characteristic function ψ(xt , z, w0 , w, g0 , g, τ ) and in terms of the relevant moment-generating function128 , respectively, without evaluating any integral at all if the characteristic function is known in closed form, we need for option-type contracts to apply a numerical integration scheme in order to calculate their model prices. Carr and Madan (1999) showed in their prominent article a very convenient method to compute option prices for a given strike range, using the FFT. The advantage in applying the FFT to option-pricing problems, is its considerable computational speed improvement compared to other numerical integration schemes. Due to the payoﬀ transform methodology, we use another pricing algorithm, which shares the same desirable, numerical properties of the FFT. Unfortunately, implementing the pricing approach according to Lewis (2001), it is necessary to impose the transform with respect to the strike. Therefore, one cannot use the FFT any longer to obtain option prices in one pass for a strike range129 . 128 129

See Section 5.2. See Lee (2004), p. 61. However, comparing the structure in equation (4.21) it is possible to obtain model prices with the help of a FFT procedure for diﬀerent levels of g (xt ).

96

6 Numerical Computation of Model Prices

In order to circumvent this problem within the payoﬀ-transform pricing approach, we need an another numerical algorithm. Therefore, we incorporate in our pricing algorithm the IFFT, to compute model prices for diﬀerent strike values130 . Furthermore, to enhance the quality of results131 , the fractional Fourier Transform of Bailey and Swarztrauber (1994) is used. This reﬁnement was introduced by Chourdakis (2005) in pricing equity option prices with the transformed option price methodology of Carr and Madan (1999). However, we sometimes encounter the problem that ψ(xt , z, w0 , w, g0 , g, τ ) cannot be calculated in closed form132 . For these cases, we implement a RungeKutta solver in our IFFT pricing algorithm. This algorithm is then used to compute the relevant values for diﬀerent z in ψ(xt , z, w0 , w, g0 , g, τ ) by solving the ODEs (2.40) and (2.41) numerically and providing the procedure with the needed values.

6.2 Contracts with Unconditional Exercise Rights As explained in Section 5.2.2 all contracts with unconditional exercise rights can be calculated as mere function evaluations of the general characteristic function ψ(xt , z, w0 , w, g0 , g, τ ), its ﬁrst order derivative with respect to z, and for integro-linear payoﬀ functions with the help of the ﬁrst order derivative ψz xt , z, w0A (z), wA (z), 0, 0M , τ . As shown, these unconditional expectations can be obtained by contour integration in closed form. Thus, we do not need to develop a numerical integration routine at all in order to calculate the relevant model prices. The calculations reduce in these cases to T EQ e 130

−

r(xs ) ds t

= ψ(xt , 0, w0 , w, g0 , g, τ ),

We ﬁnd it natural to use the FFT and the IFFT algorithm to obtain the desired Fourier Transformation. Other numerical integration schemes are also possible, like for example the numerical integration via Laguerre polynomials as used in

131

Tahani (2004). The ordinary IFFT pricing algorithm suﬀers, like the particular FFT algorithm, from the ﬁxed scale of increments of strike values and transformation variable,

132

which is discussed in Section 6.3.1. This could be the case e.g. for some subordinated processes rt or for jump components where EJ [ψ ∗ (z, w0 , w, g0 , g, J, τ )] cannot be solved explicitly.

6.3 Contracts with Conditional Exercise Rights

E e Q

and

EQ e

−

T

−

97

r(xs ) ds

t

g (xT ) =

ψz (xt , 0, w0 , w, g0 , g, τ ) , ı

T

r(xs ) ds t

ψz (xt , 0, w0A (0), wA (0), 0, 0M , τ ) , γ(T ) = ı

for arbitrary times to maturity τ . For normal contracts, the discount rate used in the characteristic function is based on the short rate r (xt ) and is zero for futures-style contracts. In case of an average-rate contract where the underlying is the geometric average of the short rate, we have to use the characteristic function with a modiﬁed discount rate rA (xt ). If the general characteristic function cannot be expressed in closed form although deﬁned by a system of ODEs, we apply a numerical algorithm to evaluate the needed values. In this case we implement a Runge-Kutta solver for the system of ODEs (2.40) and (2.41).

6.3 Contracts with Conditional Exercise Rights 6.3.1 Calculating Option Prices with the IFFT We start with the integral representation of the general option valuation formula (4.21). Since we are interested in calculating option prices in one pass for a given strike range simultaneously with the IFFT, we have to reduce the presence of K in the integral to the expression eızα(K) for both exponential-linear, linear, and integro-linear type payoﬀ functions. In the case of coupon-bond options and swaptions we have to divide the payoﬀ function up into A separate parts. The alternative representation of the valuation formula is eα(K)d V (xt , t, T ) = π

∞ eızα(K) gˆ(z)ψ(xt , −z, w0 , w, g0 , g, τ ) dz,

(6.1)

0

with F g(xT ) [G (xT )] = e(d+ız)α(K) gˆ(z), and α(K) = K for the case of a ﬂoating-rate based contract and an asian-type contract, respectively, and α(K) = ln[K] for a yield-based contract133 . The 133

See equation (5.21).

98

6 Numerical Computation of Model Prices

parameter d is chosen in a way to eliminate all dependency of α(K) in gˆ(z), which is crucial for the IFFT algorithm to work properly134 . A ﬁrst problem might arise using multi-valued functions, e.g. the complex-valued logarithm, square-root, and the conﬂuent hypergeometric function KU(a; b; y). Thus, we have to carefully keep track of the integration path to avoid any discontinuities135 . However, using a numerical algorithm to compute the particular values of the characteristic function such as a Runge-Kutta algorithm we do not encounter these problems136 . The ﬁrst step in deriving the IFFT pricing algorithm is to truncate the integration domain as ω eızα(K) gˆ(z)ψ(xt , −z, w0 , w, g0 , g, τ ) dz.

f (α(K)) ≈

(6.2)

0

Applying an U -point approximation with increment ∆ =

ω U,

we discretize the

domain of the transform variable into 1 zu = u − ∆ + ızi 2 with u = 1, . . . , U and zi corresponding to a ﬁxed value for which the Fouriertransformed payoﬀ function exists. The integration interval [0, ∞] is then replaced with a discrete, truncated region such that the integrand of f (α(K)) is negligible for zU . Hence, the discrete approximation to equation (6.2) is f (α(K)) ≈

U

eızu α(K) gˆ(zu )ψ(xt , −zu , w0 , w, g0 , g, τ ) ∆

u=1

= ∆e

−zi α(K)

U

(6.3) e

ı(u−1) ∆α(K)

e

ı∆ 2

α(K)

gˆu ψu ,

u=1 134

Otherwise, the IFFT algorithm is not applicable to the valuation problem at hand. Fortunately, we are able to reduce the dependency of K in the particular

135 136

integrals to the speciﬁc term eızα(K) , for all contracts discussed in Chapter 3. This topic is covered comprehensively in Nagel (2001), Appendix 4. In case of the Fong and Vasicek (1991a) model, we made the same experience as mentioned in Tahani (2004), Footnote 4, and compute values of the characteristic function with help of an explicit Runge-Kutta algorithm in the ﬁrst place. Thus, besides the prevention of discontinuities, the Runge-Kutta algorithm can be more eﬃcient than the explicit computation of the conﬂuent hypergeometric function.

6.3 Contracts with Conditional Exercise Rights

99

with gˆu = gˆ(zu )

ψu = ψ(xt , −zu , w0 , w, g0 , g, τ ).

and

The sum above is commonly referred to as a discrete inverse Fourier Transı∆

formation137 of the function e 2 α(K) gˆu ψu . We also want to mention that in computing this sum we eventually obtain the option price for only one particular strike value K. Since we are interested in calculating option prices for a strike range we also have to discretize α(K), which yields αv = α(K1 ) + (v − 1)η, with step size η and v = 1, . . . , U 138 . Thus, inserting the explicit expression for αv inside the brackets of equation (6.3) gives f (αv ) = ∆ e−zi αv

U

eı(u−1) ∆(α1 +(v−1)η) e

ı∆ 2 (α1 +(v−1)η)

gˆu ψu

u=1

= ∆e

−zi αv

e

ı ∆η 2 (v−1)

U

(6.4) e

ı(u−1)(v−1) ∆η ı ∆α1 (u− 12 )

e

gˆu ψu .

u=1

The form of f (αv ) is almost ready to be inserted into the IFFT algorithm. The IFFT algorithm is developed to calculate simultaneously the discrete inverse Fourier Transformation for a range of values αv . The main advantage is that it reduces the number of calculations from an order of U 2 to the order of U log2 [U ], which makes a signiﬁcant diﬀerence in computational speed139 . It eﬃciently computes the sum U 1 ı(u−1)(v−1) 2π U h e f (v, h) = u U u=1 137

for v = 1, . . . , U.

(6.5)

Although we deﬁned the transform operations in Section 2.4 vice versa, in this chapter we rely on the term discrete inverse transform, which belongs to engineering disciplines and is in line with the expression used afterwards for the

138

IFFT. We use the same discretization scheme for α(K) as used in Lee (2004). The advantage, in contrast to the discretization schemes applied in Carr and Madan (1999) and Raible (2000), is the possibility to adjust the numerical scheme for the lower bound of the strike rates. Thus, one does not necessarily have to compute

139

option prices for negligible strike rates, which is a more eﬃcient procedure. See Cooley and Tukey (1965).

100

6 Numerical Computation of Model Prices

Introducing the vectors

1 2 u=v= .. , . U

equation (6.5) can be displayed in a more compact form, which is f (h) = IFFT[h],

(6.6)

with h ∈ CU . By comparing equation (6.5) with (6.4), we obviously need the relation ∆η =

2π , U

in order to apply the IFFT algorithm properly to equation (6.4). Because

2π U

remains constant for a ﬁxed number of points U , we have only the freedom to choose either ∆ or η independently. Thus, there is a tradeoﬀ between the accuracy of the calculated results and the coarseness of the strike-value grid. According to these considerations, more accurate results of option prices corresponding to speciﬁc strike rates have to be paid with more points in the integration scheme due to the rule U × 2n . This rule ensures that the algorithm computes option prices for speciﬁc strike values and illustrates the exponential cost for more accurate results. Calculating the same number of option prices, most of them outside a desired strike range, entails a substantial waste of computational time140 . To give a more compact writing, we use henceforth the vectors α = U (αv )U g = (ˆ gu )U v=1 , ˆ u=1 and ψ = (ψu )u=1 . Eventually, the vector V(xt , t, T ) containing the option values for diﬀerent strikes, can be computed as

V(xt , t, T ) =

U ∆ e(d−zi )α π Re e

πı U (v−1)

IFFT[e

ı ∆α1 (u− 12 )

ˆ g ψ] ,

(6.7)

where the operator denotes the vector-dot product of two arbitrary vectors of the same length. This pricing algorithm is already capable of calculating 140

This particular problem is addressed in the next section.

6.3 Contracts with Conditional Exercise Rights

101

option prices. However, as stated before, equation (6.7) displays the problem of computing option prices for many irrelevant strike rates, given a desired level of accuracy. 6.3.2 Reﬁnement of the IFFT Pricing Algorithm The purpose of this subsection is to solve the problem of the inverse relationship of ∆ and η mentioned in the last section. The numerical eﬃciency can be enhanced by using a modiﬁed version of the ordinary IFFT algorithm to ensure that all calculated option prices are at least within an interval of relevant strike values. Bailey and Swarztrauber (1994) developed a method based on the FFT to choose ∆ and η independently. Their method, called the fractional Fourier Transformation, henceforth denoted as the FRFT, incorporates a new auxiliary parameter ζ 141 , which successfully dissects the otherwise ﬁxed relation ∆η ≡ 2π U . Chourdakis (2005) used this reﬁned algorithm in pricing European options on equities based on the Carr and Madan (1999) pricing framework. The FRFT was developed to eﬃciently compute the sum f (v, h, ζ) =

U

e−2πı(u−1)(v−1)ζ hu

for v = 1, . . . , U.

(6.8)

u=1

Thus, introducing the FRFT operator, we deﬁne the compact expression f (h, ζ) = FRFT [h; ζ] . Although, the parameter ζ is usually real-valued, it is not restricted to the set of R. Obviously, the FRFT is strongly connected to the FFT and the IFFT. For example, by comparing equation (6.5) with (6.8), we have the equivalence

1 1 IFFT [h] ≡ FRFT h; − . U U

The key insight to compute the FRFT in terms of the FFT and the IFFT algorithm, respectively, is to recognize that the product 2(u − 1)(v − 1) can be expressed as 141

The fractional Fourier Transformation parameter ζ in this thesis corresponds to α in the original article of Bailey and Swarztrauber (1994).

102

6 Numerical Computation of Model Prices

(u − 1)2 + (v − 1)2 − (v − u)2 . Inserting this relation into equation (6.8), subsequently doing some algebraic transformations and using the discrete version of the convolution theorem of Fourier Transformations142 , we are able to eﬃciently compute equation (6.8) with the help of both the FFT and the IFFT algorithm as follows143 . Deﬁning the vectors p and q with elements ' pu = and

' qu =

hu au

for

1≤u≤U

0

for

U < u ≤ 2U,

for au a(2U+2−u) for

1≤u≤U U < u ≤ 2U,

with 2

au = eıπζ(u−1) , we compute ﬁrst the raw transformation as ˆ f (h, ζ) = IFFT [FFT [p] FFT [q]] . The last U elements in ˆ f (h, ζ) can be discarded due to the zero padding made in the vector p. Thus, we store the ﬁrst half of the vector ˆ f (h, ζ) in a new − ˆ vector f (h, ζ). The FRFT is then f (h, ζ) = ˆ f − (h, ζ) a−u .

(6.9)

Obviously, by comparing the term inside the sum operator in equation (6.4) with the corresponding term inside the sum in equation (6.8) we have to establish the relation

∆η , 2π where both ∆ and η can be chosen arbitrarily144. Thus, our general optionpricing formula (6.7), can be rewritten in terms of the FRFT as ζ=−

142 143

144

See Proposition 2.4.3. The detailed derivation of the FRFT algorithm is given in Bailey and Swarztrauber (1994). Note that the factor

1 U

used in equation (6.4) is already included in ∆.

6.3 Contracts with Conditional Exercise Rights

103

V(xt , t, T ) =

∆ e(d−zi )α π

1

Re e−πı(v−1)ζ FRFT eı ∆α1 (u− 2 ) ˆ g ψ; −

∆η 2π

(6.10)

.

Although we have to compute two FFTs and one IFFT in order to obtain one FRFT, there is a substantial improvement due to the now independent choice of strike interval and integration domain, which saves in the end computer time. This fact becomes more important for the computation of characteristic functions for which no closed-form solutions exist and therefore the system of ODEs (2.40) and (2.41) must be solved numerically for each sampling point zu . 6.3.3 Determination of the Optimal Parameters for the Numerical Scheme As discussed in Lee (2004) and Lord and Kahl (2007), the choice of zi , determining the speciﬁc contour in the complex plane used for the numerical integration routine is crucial in computing option prices. Lee (2004) ﬁnds that for diﬀerent option payoﬀ functions, for diﬀerent strike values and driving processes, respectively, the optimal value of zi , thus minimizing the numerical error, varies substantially145 . Furthermore, the parameter ω concerning the truncation error is also of the utmost importance in a numerical option-pricing scheme. Thus, both parameters inﬂuence the accuracy of numerical solutions. This is illustrated in Figure 6.1 for zero-bond call options and the jumpenhanced models of Vasicek (1977) and Cox, Ingersoll and Ross (1985b)146 . Obviously, setting ω too small results in a highly oscillating solution vector. On the other hand choosing ω too high, the absolute error of the numerical so145

146

See Lee (2004) Table 2 and 3. The same observation is made in Lord and Kahl (2007), Figure 1. Both interest-rate models are enhanced with an exponentially distributed jump component. The coeﬃcients for the characteristic function of the jump-enhanced Vasicek model are given in equations (8.6), (8.7), and (8.8). The particular coefﬁcients in case of the jump-enhanced CIR model are given in equations (8.11), (8.12), and (8.13). A discussion of these models is given in Chapter 8.

104

6 Numerical Computation of Model Prices

lutions increase exponentially. The opposite statement holds for zi . Therefore, these parameters should be chosen to avoid minimize both eﬀects.

−12

x 10

x 10

−8

2 1

1

absolute error

absolute error

2

0

−1

0 −1 −2

−2 1400

−3 1600 1200 1000 800 600

ω

x 10

60

65

75

70

80

85

90

1400 1200 1000

ω

K

−12

60

65

75

70

80

85

90

K

−8

x 10

2

5

absolute error

absolute error

1 0 −1 −2

0

−5

−3 −4 25

−10 16 20 15

zi

10

60

65

75

70

K

80

85

90

14 12

zi

10

60

65

75

70

80

85

90

K

Fig. 6.1. Graphs in the ﬁrst row depict absolute errors of 512 zero-bond call prices for alternating values of ω. In the second row, the particular errors are depicted for varying values of zi . An exponential-jump version of the Vasicek (CIR) model is used in the left (right) column. The parameters are: rt = 0.05(0.03), κ = 0.4(0.3), θ = 0.05(0.03), σ = 0.01(0.1), η = 0.005(0.005), λ = 2(2), τ = 0.5(0.5), τˆ = 2(2).

Since we want to price a vector of option prices with the computation of one FRFT operation, thus considering one speciﬁc parameter setting for the entire strike range, we are interested in ﬁnding the optimal parameter setting for the pricing algorithm, (ω ∗ , zi∗ ), which minimizes the overall numerical error in equation (6.10). Hence, we need a criterion which measures the cumulative error of both positive and negative deviations from the theoretical solutions. Consequently, we apply in the following analysis the root mean-squared error (RMSE), which is

6.3 Contracts with Conditional Exercise Rights

( RMSE =

(VN um − VT rue ) (VN um − VT rue ) , U

105

(6.11)

where VN um denotes some numerical solution vector and VT rue represents the corresponding vector of closed-form solutions. To give an idea of the error behavior of the FRFT pricing algorithm, we ﬁrst compare quasi closed-form solutions computed with the QUADL integration routine in MATLAB147 according to equation (6.1) and the corresponding values due to the FRFT algorithm as deﬁned in equation (6.10) for a ﬁxed number of 512 diﬀerent strike rates. The particular natural logarithms of the RMSE for zero-bond call option prices are depicted in Figure 6.2. We make two remarkable observations. Firstly, for diﬀering values of ω and zi both models have a global minimum of the RMSE of computed option prices. Secondly, the logarithmic presentation of the RMSE implies a rapid and monotonic descent towards this minimum, starting with small values of ω and zi 148 . In case of the jumpenhanced CIR model, the speciﬁc error-minimizing parameter couple is clearly evident according to the contour plot of the logarithmic RMSE given in the lower right graph of Figure 6.2. On the other hand, the particular contour plot of the logarithmic RMSE for zero-bond call options under the jump-enhanced Vasicek model also clearly indicates a region of parameter couples exhibiting approximately the same RMSE magnitude. Consequently, we exploit this monotonic decrease of the RMSE to develop an algorithm, which is capable of ﬁnding an optimal parameter setting (ω ∗ , zi∗ ) and simultaneously giving an estimate of the magnitude of errors of numerical solutions even when the closed-form solutions are not known. The technique we use for the approximation of the numerical error is based on the exponential decreasing of the mean-squared error between two successive parameter values in the numerical scheme. Thus, we deﬁne the approximate RMSE as ) (VN um − VN um(+) ) (VN um − VN um(+) ) RMSEa = , (6.12) U where VN um and VN um

(+)

denote numerical solutions of two successive pa-

rameter values, whether in ω or in zi direction. 147

148

This integration routine uses an adaptive Lobatto quadrature scheme. In the calculation of quasi closed-form solutions, we set its error tolerance to 10−15 . This phenomenon shows up for all interest-rate model/payoﬀ combinations mentioned in this thesis.

106

6 Numerical Computation of Model Prices 80 20

70 30 60

10

10

50 i

0

z

ln(RMSE)

20

0 40

−10 −20

30

−30 80

20 60

−10

−20

10000 8000

40

10

6000 4000

20 0

z

−30

2000 0

1000

ω

i

2000

3000

4000

5000

ω

6000

7000

8000

9000 10000

80 25 70 20

40 60

30

15 10

50 10

5 i

0

z

ln(RMSE)

20

40

0

−10 −5

30

−20

−10

−30 80

20 60

−15

10000 8000

40

6000

−20

4000

20

zi

10

0

−25

2000 0

ω

1000

2000

3000

4000

5000

ω

6000

7000

8000

9000 10000

Fig. 6.2. Logarithmic RMSEs of 512 zero-bond call option prices. In the upper (lower) row the underlying interest rate is modeled by a jump-enhanced Vasicek (CIR) model. The parameters are: rt = 0.05(0.03), κ = 0.4(0.3), θ = 0.05(0.03), σ = 0.01(0.1), η = 0.005(0.005), λ = 2(2), τ = 0.5(0.5), τˆ = 2(2) and a strike range of K ∈ [60, 90].

In Figure 6.3, diﬀerences of the logarithmic RMSEa , for two successive parameter values of zi , and the logarithmic RMSE according to equation (6.11) are depicted for zero-bond call prices for varying zi values. Obviously, the approximate and exact RMSEs show nearly the same magnitude until the minimum RMSE is reached. Afterwards, the diﬀerence, still very small, becomes oscillating in case of the Vasicek model and experiences a decrease of its level in case of the CIR model, respectively. This characteristic behavior of the RMSEa is used in our algorithm to ﬁnd the optimal parameter couple (ω ∗ , zi∗ ) and enables the formulation of an approximate error bound for the numerical solution vector. As mentioned above, our algorithm to ﬁnd the optimal parameter couple (ω , zi∗ ) utilizes a steepest descent technique on the logarithm of the RMSEa . ∗

0

−20

−40

−1

0

2

4

6

8

10

12

14

16

18

12

1

0

0

−2 20

z

i

−12

−24

−1

0

2

4

6

8

10

12

14

16

18

ln(RMSEa)−ln(RMSE)

0

107

ln(RMSE)

1

ln(RMSE)

20

ln(RMSEa)−ln(RMSE)

6.3 Contracts with Conditional Exercise Rights

−2 20

z

i

Fig. 6.3. The dashed line represents the diﬀerence of the logarithmic RMSEa and the exact RMSE of 512 zero-bond call option prices and increasing values of zi . Both graphs are drawn for ω = 1400. The straight line depicts the logarithmic RMSE in dependence of zi . The underlying model in the left (right) graph is a jumpenhanced Vasicek (CIR) model with parameters: rt = 0.05(0.03), κ = 0.4(0.3), θ = 0.05(0.03), σ = 0.01(0.1), η = 0.005(0.005), λ = 2(2), τ = 0.5(0.5), τˆ = 2(2) and a strike range of K ∈ [60, 90].

Thus, initializing the algorithm, we ﬁrst evaluate the numerical solution VN um for some parameter values (ω o , zi0 )149 . Subsequently, we compute two additional solution vectors for ascending parameter values in the direction of both ω and zi which are then used to derive the particular ﬁrst order ﬁnite differences. Afterwards, if the slope in ω direction is smaller than the one in zi direction, thus more negative, the next numerical solution is computed with an exalted ω and vice versa. The next step in the numerical scheme is then again to obtain the necessary numerical solution vectors in order to derive the particular ﬁnite diﬀerences and so on. The algorithm aborts if the smallest value of ln(RMSEa ) is reached over some interval where the curve experienced its reversal point. In Figure 6.4, the paths with an initial value of zi0 = 2 and ω 0 = 10 for the jump-enhanced Vasicek and CIR model are shown. Obviously, the algorithm ﬁnds for both interest-rate models the optimal parameter setting, which can be justiﬁed by the graphs in the left column of Figure 6.4. In case of the optimal parameter couple using the jump-enhanced Vasicek (CIR) model, we get a diﬀerence of exact and approximate RMSEs of 9.02924×10−14 149

Since we observe the steepest descent starting at the origin the initial value for zi0 and ω 0 has to be near the origin subject to the particular regularity conditions of the Fourier-transformed payoﬀ function.

108

6 Numerical Computation of Model Prices 0

0

−5 −5

ln(RMSE)

ln(RMSE)

−10

−15

−20

−10

−15

−25 −20 −30

−35

0

100

200

300

400

500

600

700

800

−25

900

0

200

400

600

number of iterations

800

1000

1200

1400

1600

number of iterations

0

0

−5 −5

ln(RMSE)

ln(RMSE)

−10

−15

−20

−10

−15

−25 −20 −30

−35

0

2

4

6

8

10

12

14

16

18

−25

20

0

2

4

6

8

number of iterations

10

12

14

16

18

20

number of iterations

30

20

7

4

−11

−20

−8

14

−1

−1

9

−2

−23

−11

−5 20

16

1 −2 −5

−17

18 25

−17

i

z

10

−14

−20

−1

7

15

1

zi

12

−11

−11

−5

1

200

−17

ω

4

−5 1

1

400

600

−11 −14

−11

−23 5

6

−8

−5

−17

−29

1 −2 −5

8 10

800

2 200

−5 −2 1 400 600

−17 −14 −11 −8

−8

−5 −2 1

800

ω

1000 1200 1400 1600

Fig. 6.4. Search for the optimal parameter couple (ω ∗ , zi∗ ). In the ﬁrst (second) column particular graphs are shown for the Vasicek (CIR) model with the data used in Figure 6.3. In the ﬁrst row, the particular ln(RMSE) is depicted for the search algorithm with increments (∆ω, ∆zi ) = (1, 1). In the second row the same search is made with increments (100, 5). In the third row the dashed (dash-dotted) line denotes the particular search path for small (high) increment values, where the optimal choice is marked by a circle and cross, respectively.

6.3 Contracts with Conditional Exercise Rights

109

(8.27740×10−10), whereas the exact RMSE is 1.11766×10−13 (1.30453×10−9). Thus, we have in both models a diﬀerence which is of smaller order than the eﬀective error according to the RMSE. Consequently, the RMSEa gives a good prediction for the corresponding exact value, which justiﬁes the application of the approximate RMSE. In the ﬁrst row of Figure 6.4, we used very small increments for the search of the optimal parameter couple to give a detailed impression of the search path and the particular logarithmic RMSE. According to the graphs in the second row of Figure 6.4, a comparable result is achieved by running the algorithm with higher increments150 . However, due to the reduced number of iterations, the search algorithm with high increments is in case of the jump-enhanced Vasicek (CIR) model up to 71 (86) times faster. Dealing with a characteristic function known in closed form together with a FRFT-based pricing algorithm, the search takes only a second at all even for small increments. Thus, if the general characteristic function is known in closed form, the step-size does not matter. However, if values of the general characteristic need to be determined numerically via a Runge-Kutta algorithm, we usually set the increments high enough to keep the overall number of iterations small. Finally, we use the RMSEa to derive an upper error bound for the numerical solutions contained in VN um . The ﬁrst step in deriving this particular error bound is to consider a hypothetical solution vector VN um , where all elements equal their true solutions except the result given in the ﬁrst position of the solution vector, namely V1N um . Without loss of generality, we assume the numerical error of this particular option price to be of magnitude |a|. Therefore, solving equation (6.11) in this special case gives √ a = RMSE U . Additionally, we are also able to state the inequality * (VN um − VT rue ) (VN um − VT rue ) ≥ |VvN um − VvT rue |,

(6.13)

(6.14)

to hold for every element of the numerical solution vector VN um . According √ to equation (6.13), the RMSE scaled by some constant U states the value 150

The second run of the algorithm, with higher increments, gives an absolute error for the optimal parameter couple (ω ∗ , zi∗ ) for zero-bond calls under the jumpenhanced Vasicek (CIR) model of 1.13911 × 10−13 (1.46601 × 10−9 ).

110

6 Numerical Computation of Model Prices

of the maximum attainable error. Furthermore, this result together with the inequality in (6.14) generally implies that the absolute error of one particular numerical solution VvN um cannot exceed the absolute value |a|. Therefore, the RMSE can be used in formulating a boundary for the highest possible error. √ Consequently, we use the quantity RMSEa , scaled by some constant U , as a conservative upper error bound for the results generated by the pricing algorithm.

7 Jump Speciﬁcations for Aﬃne Term-Structure Models

7.1 Overview In this thesis, we discuss jump-diﬀusion interest-rate models. Thus, both diffusion and jump components are included in order to model more realistic term-structure models. The jump sizes considered are governed either by exponential, normal or gamma distributions. The exponential jump distribution is a very popular approach in modeling term structure and equity models151 , since it yields closed-form formulae for most derivatives contracts. Das and Foresi (1996) and Chacko and Das (2002) have conducted recent studies with a double-sided version of this jump type using a Vasicek model for the diffusion part152 . Our second candidate, the normal jump-size distribution is used in Baz and Das (1996) and Das (2002)153 . The last jump-size distribution candidate for the interest-rate process is a gamma distribution, which is used in Kispert (2005) to support the stochastic dynamics of the volatility in electricity derivative contracts. This jump type is used for the ﬁrst time in a jump-diﬀusion interest-rate model. As a special case, the gamma distribution covers the exponential distribution. Hence, we can build a more ﬂexible 151

152

Das and Foresi (1996) used this jump speciﬁcation in modeling short rates whereas Kou (2002) uses this type of jump-size distribution modeling equities. Jumps in the instantaneous interest rate are governed by an exponential jumpsize distribution. The direction of the jump itself is modeled either by a Bernoulli

153

distribution or via two diﬀerent Poisson processes. See Section 8.2. The articles consider a discrete version of the Vasicek interest-rate model with normally distributed jump shocks.

112

7 Jump Speciﬁcations for Aﬃne Term-Structure Models

jump shock component in contrast to the exponential case, by extending the repertoire of jump-size distributions to the gamma distribution case. In the following, we do not restrict ourselves solely to one jump component for each factor. Due to the independence of the jump distributions from the state of xt 154 , we are able to add an unlimited amount of diﬀerent jump components. However, we need to consider possible nonnegativity constraints of the particular diﬀusion process. Thus, we do not combine normally distributed jump parts, or negatively directed exponentially and gamma distributed jump parts, with a Square-Root diﬀusion model. This is in fact no real drawback, that is to say we can think of a bad news eﬀect rather as a discontinuous increase in interest rates than the opposite eﬀect155 . All possible combinations for diﬀusion and jump components are illustrated in Figure 7.1. According to equation (2.40), the coeﬃcient function a(z, τ ) can be split into a part containing the characteristics of the diﬀusion process and a part containing the additional jump characteristics156. This results in a modular representation of the ODE for the coeﬃcient function a(z, τ ), which is a(z, τ )τ = a0 (z, τ )τ + a1 (z, τ )τ , with a0 (z, τ )τ = −w0 + µQ 0 b(z, τ ) +

(7.1)

1 b(z, τ ) Σ0 b(z, τ ), 2

and a1 (z, τ )τ = EJ [ψ ∗ (z, w0 , w, g0 , g, J, τ ) − 1] λQ . Unless otherwise stated, the coeﬃcient function a0 (z, τ ) denotes the diﬀusion part, whereas a1 (z, τ ) represents the solution for the jump part, which is frequently called the jump transform157 . As mentioned above, each diﬀusion process can be augmented with an inﬁnite number of jump processes. Thus, taking the expectation in (2.39) 154

This statement also holds for diﬀerent jumps triggered by the same poisson pro-

155

cess. See equation (7.2). See, for example, Sch¨ obel and Zhu (2000), p. 5. Note that the jump part aﬀects the coeﬃcient a(z, τ ), whereas the coeﬃcient

156

157

vector b(z, τ ) is independent from the jump amplitude and intensity. See Duﬃe, Pan and Singleton (2000).

7.1 Overview

113

Fig. 7.1. Possible combinations of the Ornstein-Uhlenbeck (OU) process and the Square-Root (SR) process with the exponential (Ex), normal (No), and gamma (Ga) jump distributions.

with respect to the jump amplitudes J, we obviously are able to check that every element of the resulting vector is expressible as the product of diﬀerent expectations. Formally, we have Ej1 eb(z,τ ) j1 eb(z,τ ) j1 b(z,τ ) j2 E eb(z,τ ) j2 e j2 = . EJ .. .. . . b(z,τ ) jN b(z,τ ) j N e EjN e

Selecting one element of this vector as an example, say Ejn eb(z,τ ) jn , and manipulating the expectation operator, we get

114

7 Jump Speciﬁcations for Aﬃne Term-Structure Models

M + b(z,τ ) jn b(z,τ ) jn = Ejn e e ν(Jmn ) djn m=1

RM

=

M +

(m)

eb

(z,τ )Jmn

ν(Jmn ) dJmn

(7.2)

m=1 R

=

(m) EJmn eb (z,τ )Jmn .

M + m=1

The function ν(Jmn ) represents the probability density of the particular jump amplitude Jmn . As demonstrated in equation (7.2) the joint density function can be expressed as the product of diﬀerent density ν(Jmn ) on account of the independence of the jump amplitudes. Consequently, issues are simpliﬁed in equation (7.2) by successively evaluating all integrals one by one, which yields the cumulative product of diﬀerent expectations. Thus, we express the solution of the jump part as a1 (z, τ ) = −τ ιN λQ +

N n=1

(n)

λQ

τ + M

(m) EJmn eb (z,l)Jmn dl ,

(7.3)

0 m=1

with n = 1, . . . , N and each element of the vector ιN ∈ RN equals one. Since we need to calculate the integral over the time variable there is the possibility of ending up without any closed-form solution158 . In contrast, the Bates (1996) model, which is a jump-diﬀusion model in an equity context, in which a normally distributed jump component is used for the log-asset price process, the coeﬃcient for the jump size yields a nice closed-form expression in Fourier space159 . In term-structure models, normal and gamma size distributions allow only the formulation of the coeﬃcient a1 (z, τ ) in terms of its underlying diﬀerential equation. Thus, a possible reason why normal and gamma jump distributions are not as popular in interest-rate option pricing might be tracked back to the unavailability of appropriate valuation formulae for interest-rate contracts. Nevertheless, these jump size candidates provide a 158

Both gamma and normally distributed jump amplitudes have no closed-form jump

159

transform for all models discussed in this thesis.

In one-factor equity models the computation can be simpliﬁed to EJn eızJn , which is obviously easier to handle, since the exponential function inside the expectation operator is independent of the time to maturity variable τ . See Cont and Tankov (2004), p. 477.

7.2 Exponentially Distributed Jumps

115

valuable contribution in generating a realistic overall probability distribution of short rates. However, the algorithm presented in Chapter 6 can compute derivative prices under these interest-rate dynamics. The only condition that needs to be met is the availability of separable ODEs of the coeﬃcient functions a(z, τ ) and b(z, τ ), which lets us apply a Runge-Kutta algorithm160 .

7.2 Exponentially Distributed Jumps The exponential distribution is a widely used shock speciﬁcation in jumpdiﬀusion models. Thus, it can be found in both equity and interest-rate models161 . The probability density function pEx (J, η) of an exponentially distributed variable J ∼ Ex(η) is deﬁned as if 0 pEx (J, η) = 1 − J if e η η

J 0.

Hence, the expected value for J and its variance is EJ [J] = η, and VARJ [J] = η 2 . The shape of the density function pEx (J, η) for diﬀerent values of η is shown in Figure 7.2. For a positively directed jump, with distribution parameter η+ , we get 160

An interest-rate model which clearly opposes this separability ability of the coeﬃcient functions of the general characteristic function is given in Ahn and Gao (1999) and A¨ıt-Sahalia (1999), Example 3. However, closed-form solutions of zerobond prices under these short-rate dynamics can be derived. See Ahn and Gao

161

(1999), Proposition 1. Kou (2002), Kou and Wang (2004) implemented this jump speciﬁcation for equity models, whereas Das and Foresi (1996) integrated this jump type in an OrnsteinUhlenbeck short-rate model.

116

7 Jump Speciﬁcations for Aﬃne Term-Structure Models

200 0.005 0.01 0.015 0.02

180 160

probability density

140 120 100 80 60 40 20 0

0

0.01

0.02

0.03

0.04 J

0.05

0.06

0.07

0.08

Fig. 7.2. The density function pEx (J, η) for varying η of an exponentially distributed random variable. ∞ (m) b (z,τ )− η1 J (m) + e 1 −1 EJ eb (z,τ )J − 1 = η+ b(m) (z, τ ) − η1+ 0

1 = −1 1 − b(m) (z, τ )η+ =

b(m) (z, τ )η+ . 1 − b(m) (z, τ )η+

Accordingly, a negatively directed jump with parameter η− , has an expected value of ∞ (m) − b (z,τ )+ η1 J (m) − e 1 −1 EJ e−b (z,τ )J − 1 = − η− b(m) (z, τ ) + η1− 0

1 −1 = 1 + b(m) (z, τ )η− =−

b(m) (z, τ )η− . 1 + b(m) (z, τ )η−

7.3 Normally Distributed Jumps

117

In order to guarantee the existence of the jump transform, we need the real part Re b(m) (z, τ ) ≤ η1+ for the positively sized jump and Re b(m) (z, τ ) ≥ − η1− for the negatively directed jump162 , respectively. Thus, multiplying the recently derived expectations with the jump intensity of the particular Poisson jump, the jump part of the coeﬃcient function a(z, τ ) can be generally represented as, τ a1Ex± (z, τ )

=± 0

(n)

λQ b(m) (z, l)η± dl. 1 ∓ b(m) (z, l)η±

(7.4)

The transform for this jump candidate is the only one that can be expressed in closed form for the interest-rate models discussed in the next chapter.

7.3 Normally Distributed Jumps The second candidate we consider for the jump-size distribution is the normal distribution. As mentioned before, this speciﬁcation is not as popular in interest-rate pricing frameworks compared to the exponentially distributed case. One reason might be that the jump transform in an interest-rate jumpdiﬀusion framework cannot be expressed in closed form. In this setup, the jump amplitude J ∼ N µJ , σJ2 is distributed according to a probability density function: −

(J−µJ )2 2σ2 J

e pN o (J, µJ , σJ ) = √ 2πσJ

∀

J ∈ R,

with mean EJ [J] = µJ , and variance VARJ [J] = σJ2 . The shape of the density function pN o (J, µJ , σJ ) for diﬀerent values of σJ is shown in Figure 7.3. The few articles which mention this particular jump type can be quickly summarized. Baz and Das (1996), Durham (2005) and Durham (2006) implemented the Gaussian jump within a Vasicek base model. Since this type 162

Since b(m) (z, τ ) < 0 and be fulﬁlled.

1 η−

is usually very large, we assume both conditions to

118

7 Jump Speciﬁcations for Aﬃne Term-Structure Models

of jump might violate a non-negativity constraint of the underlying diﬀusion process, it is only meaningful in a context of a real-valued process. Therefore, we do not consider the normally distributed jump candidate in case of a Square-Root diﬀusion process.

80 0.005 0.01 0.015 0.02

70

probability density

60 50 40 30 20 10 0 −0.08

−0.06

−0.04

−0.02

0 J

0.02

0.04

0.06

0.08

Fig. 7.3. The density function pNo (J, µJ , σJ ) for ﬁxed µJ = 0 and varying σJ of a normally distributed random variable.

Baz and Das (1996) approximate the expectation in equation (7.3) via a Taylor series approximation163 . The series-approximation approach mentioned there considers two terms. Consequently, they ﬁrst approximate the expression inside the expectation operator, and then take the expectation of the resulting terms. Hence, the approximation is given as 163

The Taylor series approximation of the exponential function f (x) = ex is given by

∞

xi i=0 i! .

7.3 Normally Distributed Jumps

(m) b(m) (z, τ )2 2 J EJ eb (z,τ )J − 1 ≈ EJ b(m) (z, τ )J + 2 = b(m) (z, τ )µJ +

119

b(m) (z, τ )2 2 (µJ + σJ2 ). 2

In the last equation, the particular parameters of the normal distribution µJ and σJ2 are used. Of course, it is possible to use a Taylor series considering more terms in order to enhance the accuracy of the calculations. Another slightly diﬀerent approximation technique is presented in Durham (2005) and Durham (2006), respectively. Here, the author applies a Taylor series approximation after taking the expectation in equation (7.3). This incorporates the distributional parameters in a more explicit fashion. Applying a two-term Taylor expansion gives164 (m) b(m) (z, τ )2 2 µJ + σJ2 EJ eb (z,τ )J − 1 ≈ b(m) (z, τ )µJ + 2 b(m) (z, τ )3 b(m) (z, τ )4 4 2 + µJ σJ + σJ . 2 8

Obviously, there is an advantage in applying either one of these analytic approximations for the jump transform. Using these simpliﬁcations, we are able to solve the ODE (7.1) in a consistent manner, meaning that no numerical integration of the jump transform is needed anymore, since only terms of b(m) (z, τ ) are left, yielding an approximate closed-form solution for the characteristic function165 . As an additional beneﬁt of the analytical approximations, Baz and Das (1996) mention the computational speed enhancement, facilitating the calibration to empirical data. However, the major drawback of both approximation techniques results from the application of the Taylor series. In order to produce accurate results, the term b(m) (z, τ ) and the diﬀerence inside the expectation operator, respectively, must be very small. Hence, with an increasing mean of the jump component and increasing variance, the results 164

165

In Durham (2006) the negative sign of the coeﬃcient b(m) (z, τ ) is extracted which explains the slightly diﬀerent representation. For example, taking the Vasicek one-factor base model of equation (8.5) and the linear approximation due to Baz and Das (1996), we encounter the simple Q Q problem of solving equation (7.1) with extended parameters µ ˆQ 0 = µ0 + λ µJ and

σ ˆ0 =

σ02 + σJ2 + µ2J . Subsequently, the coeﬃcient function aτ (z, τ ) with these

modiﬁed parameters has to be solved.

120

7 Jump Speciﬁcations for Aﬃne Term-Structure Models

get more and more inaccurate. Consequently, the analytical approximation procedures should be applied only to scenarios where the jump component exhibits a small mean and variance. Since our numerical procedure is designed to handle implicitly the ODE (m) part of the jump transform, we need only the expected value of eb (z,τ )J to be explicit. Under a normally distributed jump size regime, this is (m) (b(m) (z,τ )σJ )2 b (z,τ )J b(m) (z,τ )µJ + 2 EJ e −1 =e − 1, which leads to the particular coeﬃcient function 2 τ b(m) (z,τ )σJ ) ( (n) (m) 2 dl . a1N o (z, τ ) = λQ −τ + eb (z,l)µJ +

(7.5)

(7.6)

0

The value of the integral can then be numerically approximated via a Runge-Kutta algorithm. Despite the numerical integration, the computational eﬀort is very small due to our implemented FRFT procedure. But in contrast to the Taylor-series approach mentioned above, our results do not suﬀer from inaccuracies due to high mean and volatility parameters of the jump component. Hence, with our valuation procedure we gain superior accuracy. Furthermore, we are able to compute model prices for this jump speciﬁcation for the ﬁrst time, not only for ordinary zero bonds, but for all derivatives contracts, presented in Sections 3.2 and 3.3.

7.4 Gamma Distributed Jumps The last jump-size distribution we want to implement in an interest-rate model is the gamma distribution. The probability density function pGa (J, η, p) of the random variable J ∼ Ga(η, p) is given as 0 if J 0. p η Γ (p) Thus, the expected value and variance for J is

7.4 Gamma Distributed Jumps

121

EJ [J] = ηp, and VARJ [J] = η 2 p. The function Γ (p) denotes the gamma function. Setting the parameter p = 1, the gamma distribution replicates the exponential distribution, since we have then the relation pGa (J, η, 1) = pEx (J, η). Additionally, we can use the gamma distribution to generate a chi-squared distribution. In this case we set η = 2 and p = 2q , where q is a positively valued integer. The resulting chi-squared distribution has then 2p and q degrees of freedom166 . Another special case of the gamma distribution is the Erlang distribution. Here, we only need p to be a positive integer value167 . Thus, the Erlang distribution can be interpreted as the sum of p independent exponentially distributed random variables with equal parameter η. The graph in Figure 7.4 shows the probability density function for diﬀerent values of p. Comparing the diﬀerent curves in Figure 7.4, it is obvious that this jump-size distribution is able to substantially enhance the short-rate model. In Heston (1995) a pure jump interest-rate model is proposed. Accordingly, instead of a diﬀusion component, the innovations of the process in this model are governed solely by gamma distributed jumps. To our knowledge, Kispert (2005) was the ﬁrst to use gamma distributed jump sizes within a jump-diﬀusion model. However, in pricing European options he needs inefﬁcient Monte-Carlo routines, using the gamma and normal jump amplitude speciﬁcation. These numerical problems can be circumvented by applying the FRFT-based algorithm together with a Runge-Kutta algorithm for the ODEs. Next, we want to derive the particular jump transform. The expectation for a positively directed jump can be computed as (m) EJ eb (z,τ )J − 1 = 166

167

1 p η+ Γ (p)

∞ e

−J

1 η+

−b(m) (z,τ )

J p−1 dJ − 1.

0

This is easily checked by comparing the particular moment-generating functions. See Stuart and Ord (1994), p. 541. See, for example, Balakrishnan, Johnson and Kotz (1994), p. 337.

122

7 Jump Speciﬁcations for Aﬃne Term-Structure Models

200 1 2 3 4

180 160

probability density

140 120 100 80 60 40 20 0

0

0.01

0.02

0.03 J

0.04

0.05

0.06

Fig. 7.4. The density function pGa (J, η, p) for ﬁxed η = 0.005 and varying p of a gamma distributed random variable.

Introducing the substitution m = J pler representation168 (m) EJ eb (z,τ )J − 1 =

Γ (p)

1 η+

− b(m) (z, τ ) , we arrive at the sim-

1

1 η+

p p − b(m) (z, τ ) η+

∞ e−m mp−1 dm − 1 0

"

#$

%

Γ (p) 1 p − 1. = 1 − b(m) (z, τ )η+ Hence, the corresponding expression for a negatively sized jump is (m) 1 p − 1. EJ e−b (z,τ )J − 1 = (m) 1 + b (z, τ )η− 168

To ensure theexistence of the jump transform, we have the same inequalities for Re b(m) (z, τ ) to be satisﬁed as in the case of exponentially distributed jumps.

7.4 Gamma Distributed Jumps

123

Having derived the relevant expectations, we immediately are able to formulate the respective jump transforms a1Ga± (z, τ ). Thus, integrating over the (n) time axis and multiplying the result by the relevant jump intensity λQ we obtain,

(n)

a1Ga± (z, τ ) = λQ

−τ +

τ 0

1 dl . (1 ∓ b(m) (z, l)η± )p

(7.7)

Since it is not possible to derive closed-form solutions for general values of p, we apply a Runge-Kutta algorithm implemented in our FRFT procedure in order to eﬃciently calculate option prices. Again, for a strictly positive interest-rate model, we use only a positively directed version of the jump candidate.

8 Jump-Enhanced One-Factor Interest-Rate Models

8.1 Overview In order to implement the previously proposed pricing procedure, we need in addition to the payoﬀ transformations derived in Chapter 5, the particular characteristic functions. The goal of this chapter is to provide these necessary functions for the case of an underlying one-factor interest-rate model and to examine the behavior of the particular density functions and prices of selected contingent claims, according to Table 4.1, inﬂuenced by jump components. Thus, we focus our eﬀorts exclusively on the exponential-aﬃne termstructure models generated by the one-factor version of equation (2.23). Since a one-factor model implicates the incorporation of one Brownian motion, this statement does not entail the restriction of including one sole jump component. Therefore, we apply diﬀerent jump components in our examples. The general version of the one-factor instantaneous interest rate is then given by rt = w0 + w1 xt , and the factor xt is deﬁned by the one-dimensional stochastic diﬀerential equation

dxt = µQ (xt ) dt + σ(xt ) dWtQ + j dN λQ t ,

(8.1)

where j ∈ RN , µQ (xt ) and σ(xt ) are the one-factor counterparts of the original parameters J, µQ (xt ) and Σ(xt ) used in equation (2.23). All parameters are postulated under the risk-neutral probability measure Q. Therefore, the solution of the general characteristic function ψ(xt , z, w0 , w1 , g0 , g1 , τ ) for these models is given by the simpliﬁed versions of the ODEs (2.40) and (2.41), which are

126

8 Jump-Enhanced One-Factor Interest-Rate Models

σ02 b(z, τ )2 − w0 , a0 (z, τ )τ = µQ b(z, τ ) + 0 2 a1 (z, τ )τ = Ej eb(z,τ )J1 , eb(z,τ )J2 , . . . , eb(z,τ )JN − 1 λQ ,

(8.2) (8.3)

and

σ12 b(z, τ )2 − w1 , 2 with terminal conditions a(z, 0) = 0 and b(z, 0) = ızg1. b(z, τ )τ = µQ 1 b(z, τ ) +

(8.4)

In the upcoming sections, we discuss jump-enhanced short-rate models where the diﬀusion part is either modeled as a Ornstein-Uhlenbeck or a Square-Root process and the jump components are governed by the distributions presented in the previous chapter.

8.2 The Ornstein-Uhlenbeck Model 8.2.1 Derivation of the Characteristic Function Modeling the factor xt as a stochastic process according to Ornstein and Uhlenbeck (1930) exhibits a strong resemblance to the well-known model given in Vasicek (1977)169 . The so-called Vasicek model has become very popular in interest-rate modeling. The instantaneous interest rate is modeled as an Ornstein-Uhlenbeck process, with a mean-reverting component and a Brownian motion, which yields a time-homogenous Markov process. The approach used in Vasicek (1977) to derive prices for contingent claims under the riskneutral probability measure, is similar to the methodology used in the article of Black and Scholes (1973), based on a hedging argument170 . Due to its popularity, many authors have made attempts to extend this diﬀusion model with jumps. Das and Foresi (1996) introduced a jump-enhanced Vasicek model, where the jump size is governed by an exponential distribution and the jump direction is modeled as a Bernoulli random variable which results in a double-sided jump component. However, given a jump intensity λ for 169

The process used in Vasicek (1977) and the process discussed in this section coincide for the case of rt = xt , thus setting the discount parameters to w0 = 0

170

and w1 = 1. In contrast to the Black-Scholes model, an appropriate market price of risk has to be additionally considered, since the short rate rt is no traded quantity. See Section 2.3.

8.2 The Ornstein-Uhlenbeck Model

127

the Poisson process, this model can be easily subsumed by applying a single exponentially distributed jump size and setting the modiﬁed intensity for the upward jump trigger to ψλ and the downward jump intensity to (1 − ψ)λ171 , respectively. Another model, where the Vasicek model is extended with a normally distributed jump size, is given in Baz and Das (1996), Das (2002), Durham (2005). In Baz and Das (1996) and Durham (2005), approximation techniques are presented for pricing option contracts under these interest-rate dynamics, as explained in Section 7.3. In Das (2002), the author utilizes this jump-diﬀusion model for the estimation of the term structure and subsequent calibration of the particular parameters according to Fed Funds data. The the risk-neutral coeﬃcients are µQ 0 = κθ,

µQ 1 = −κ,

σ0 = σ,

σ1 = 0,

where the mean-reverting feature of the instantaneous interest rate rt is guaranteed for κ > 0. Thus, the diﬀusion part of the SDE (8.1) is dxt = κ(θ − xt ) dt + σ dWtQ .

(8.5)

Under these dynamics the stochastic process starting with xt the OrnsteinUhlenbeck process reﬂects a normal distribution with expectation EQ [xT ] = xt e−κτ + θ(1 − e−κτ ), and variance VARQ [xT ] =

σ2 1 − e−2κτ . 2κ

Modeling the term structure with this Ornstein-Uhlenbeck type process, has the attractive feature that solutions for many important contingent claims can be derived within closed-form formulae. Moreover, the model is likely to be used for its high tractability. Finally, one major drawback of the model is the ability to produce negative short rates with a positive probability. According to equations (8.2) and (8.4), straightforward calculations show that the diﬀusion-related coeﬃcients of the general characteristic function can 171

In Das and Foresi (1996), the parameter ψ denotes the probability that the sign of the jump is positive.

128

8 Jump-Enhanced One-Factor Interest-Rate Models

be derived as172

˜b(z, τ ) = w1 + ızg1 e−κτ − 1 , κ

(8.6)

ızg1 σ 2 ˜ b(z, τ ) a0 (z, τ ) = − w0 τ − 2κ σ2 w1 σ 2 ˜ ˜b(z, τ )2 . − θ− b(z, τ ) + w1 τ − 2 2κ 4κ

(8.7)

and

Equipped with these time-dependent coeﬃcient functions corresponding to the diﬀusion parts of the short-rate model, we must determine in the next step the particular jump part a1 (z, τ ). Since this function is independent of a0 (z, τ ), we are able to derive it separately. Unfortunately, a closed-form solution for the coeﬃcient a1 (z, τ ) exists only in case of an exponentially distributed jump size. Thus, according to equation (7.4) we obtain for the Ornstein-Uhlenbeck model, where the nth jump in xt is governed by an exponentially distributed jump size, the relevant coeﬃcient function as (n)

a1Ex± = −λQ

τ+

(n) 1 ∓ b(z, τ )η± λQ ln . κ ± w1 η± (1 ∓ ızg1η± ) e−κτ

(8.8)

In equation (8.8), the signs in the index of a1Ex± denotes an upward and a downward jump, respectively. Considering normally and/or gamma distributed jumps, we have to apply a Runge-Kutta algorithm to solve equations (7.6) and (7.7). 8.2.2 Numerical Results Next, we want to examine and demonstrate the impact of the particular jump speciﬁcations for the case of a jump-enhanced Ornstein-Uhlenbeck process. Thus, we ﬁrst compare the probability density for diﬀerent jump amplitude speciﬁcations, and afterwards look brieﬂy at values of option prices for interest-rate derivatives corresponding to the payoﬀ structures given in Table 4.1. 172

˜ τ ) and therefore Here, the coeﬃcient ˜b(z, τ ) denotes the scalar version of b(z, complies with the relation b(z, τ ) = ˜b(z, τ ) + ızg1 .

8.2 The Ornstein-Uhlenbeck Model

129

Figures 8.1 - 8.3 depict probability density functions of short rates under diﬀerent jump regimes with diﬀusion parameters rt = 0.05, κ = 0.4, θ = 0.05, σ = 0.01 and T = 1. In each ﬁgure, we focus exclusively on one particular jump candidate, while ignoring other jump speciﬁcations. The probability density functions are then examined for diﬀerent arrival rates and jump amplitudes173 , respectively. Additionally, in case of a normally distributed jump component, we also examine the inﬂuence of the jump amplitude volatility, whereas in case of a gamma distributed jump distribution, the impact for diﬀerent values of p is displayed.

50

50 0 2 4 6

45 40

40 35

probability density

probability density

35 30 25 20

30 25 20

15

15

10

10

5 0

0.005 0.01 0.015 0.02

45

5 0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0

0

0.02

r

T

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

T

Fig. 8.1. Probability densities for a short rate governed by a Vasicek diﬀusion model enhanced with an exponentially distributed jump component. In the left (right) graph the density function for varying jump intensities (means) are depicted. The base parameters are: rt = 0.05, κ = 0.4, θ = 0.05, σ = 0.01, λ = 2, η = 0.005, T = 1.

The ﬁrst impression from Figures 8.1 - 8.3, is that increased jump intensity results in all three cases in a positively skewed density function with a slightly right-shifted mode174 . The asymmetric shape is in line with empirical ﬁndings175 . Increasing the mean of the jump amplitude, the density functions 173

In case of exponentially and gamma distributed jumps, the arrival rates belong only to positively directed jumps, thus leaving downward jumps with zero jump

174

175

intensities. This eﬀect becomes more apparent for higher values of jump amplitudes η and µJ , respectively. See Arapis and Gao (2006), Figure 3. The authors apply alternatively a nonparametric estimator for the short-rate probability density of three-month Treasury bill rates and seven-day Eurodollar deposit rates.

130

8 Jump-Enhanced One-Factor Interest-Rate Models 50

50 0 2 4 6

45 40

40 35

probability density

probability density

35 30 25 20

30 25 20

15

15

10

10

5 0

0.005 0.01 0.015 0.02

45

5 0

0.05

0.1

0.15

0.2

0.25

0.3

r

0

0

0.05

0.1

0.15

0.2

0.25

0.3

r

T

T

50 1 2 3 4

45 40

probability density

35 30 25 20 15 10 5 0

0

0.05

0.1

0.15

0.2

0.25

0.3

r

T

Fig. 8.2. Probability densities for a short rate governed by a Vasicek diﬀusion model enhanced with a gamma distributed jump component. In the upper left (right) graph the density functions for varying jump intensities (means) are depicted. The lower graph shows the density behavior for alternating values of p. The base parameters are: rt = 0.05, κ = 0.4, θ = 0.05, σ = 0.01, λ = 2, η = 0.005, p = 2, T = 1.

of all jump candidates show positive skewness while concurrently maintaining the mode of the density function. Comparing the particular density functions of an exponentially and gamma distributed jump-enhanced short-rate model, we encounter, in case of a gamma and normal distribution, a bi-modal density function. The impact of the volatility parameter σJ for a normally distributed jump is more complex. For high values of the jump volatility, the density function displays a leptokurtic behavior compared. In addition, we observe, due to the possibility of negative jump sizes, raised tails on both sides of the particular density function as well. This eﬀect is rather visible to the right tail, since we have a positive mean of the jump-size distribution. Due to the possibility to produce negative short rates in the Ornstein-Uhlenbeck case, we have the undesirable ability to obtain a density function with non-negligible

8.2 The Ornstein-Uhlenbeck Model 40

40 1 2 3 4

35

30

probability density

probability density

0 0.01 0.02 0.03

35

30 25 20 15

25 20 15

10

10

5

5

0

131

0

0.05

0.1

0.15

0.2

0.25

0.3

r

0

0

0.05

0.1

0.15

0.2

0.25

0.3

r

T

T

40 0.005 0.01 0.015 0.02

35

probability density

30 25 20 15 10 5 0

0

0.05

0.1

0.15

0.2

0.25

0.3

r

T

Fig. 8.3. Probability densities for a short rate governed by a Vasicek diﬀusion model enhanced with a normally distributed jump component. In the upper left (right) graph the density functions for varying jump intensities (means) are depicted. The lower graph shows the density behavior for alternating values of σJ . The base parameters are: rt = 0.05, κ = 0.4, θ = 0.05, σ = 0.01, λ = 1, µJ = 0.02, σJ = 0.01, T = 1.

probabilities for negative rates, which becomes more severe depending on the absolute height of the volatility. Obviously, besides the asymmetric shape, all density functions are skewed, in contrast to the plain Ornstein-Uhlenbeck model, which is another advantage in including jump components. Theoretical prices of interest-rate derivatives are computed with the following base parameters: rt = 0.05, κ = 0.4, θ = 0.05, σ = 0.01, λΓ = 2, η = 0.005, p = 2, λN = 2, µJ = 0.015, σJ = 0.01 and τ = 0.5. Table 8.1 reports values of zero-bond calls, according to equation (3.12), for a strike range from 60 to 90 units computed with the FRFT pricing algorithm. We choose this particular strike range to cover either ITM, ATM and OTM option

132

8 Jump-Enhanced One-Factor Interest-Rate Models

prices176 . Here, we only considered normally and gamma distributed jumps, thus excluding exponentially distributed jumps, because of the similarity to the gamma jump speciﬁcation. Solutions are given for diﬀerent jump intensities, amplitudes, and volatilities in case of normally distributed jumps and for diﬀerent values of p in case of the gamma jump size speciﬁcation, respectively. Examining values of zero-bond calls for diﬀerent values of the parameters η, λΓ and p one-by-one, we observe that the jump intensity has the greatest inﬂuence upon call values followed by the parameter p. The jump mean has the smallest eﬀect upon option prices although it signiﬁcantly alters the particular density function of the short rate. However, price diﬀerences for ITM options are relatively diminutive, whereas for OTM options the above mentioned impact is quite considerable. Applying a normally distributed jump component in the short-rate model, we observe for increased jump volatilities higher option prices, which can be explained based on the above mentioned two-sided enlargement of the probability density function compared to the cases where the jump mean or jump intensity is increased. Theoretical prices for cap contracts and average-rate caps on the short rate are presented in Tables 8.2 and 8.3, respectively, for a strike range from 2 to 8 units. The cap contracts have both only one payment date, which is paid at the maturity of the contract. Here, we observe the opposite eﬀect due to the direct inﬂuence of the short rate on the payoﬀ function. Since the contract is based on rT , positively directed jumps increase the contract value. Remarkably, the eﬀect of the jump volatility σJ is twofold. Firstly, it lowers the value of the option contract for ITM options. On the other hand cap values are being raised, if the option contract is OTM. Obviously, the geometric average is less sensitive to discontinuous jumps. Since the interest rate is inﬂuenced by positively sized jumps, we compute higher values for the ordinary cap contract compared to the corresponding average-rate contract due to the averaging process itself. Thus, we are able to validate the statement that average-rate options are more robust to price manipulations, thus reducing risk exposures177 .

176

177

The value of a zero bond with remaining time to maturity of two years priced with the base parameters is 83.768 units. Compare with the comments made on p. 41.

8.2 The Ornstein-Uhlenbeck Model

133

Table 8.1. Values of zero-bond call options for the jump-enhanced OU model, where the underlying zero-bond contract has a nominal value of 100 units. K

60

65

70

75

80

85

90

20.595

15.747

10.899

6.067

1.666

0.004

0

0.01

17.277

12.442

7.631

3.117

0.273

0

0

0.015

14.168

9.383

4.827

1.246

0.003

0

0

0.02

11.296

6.711

2.745

0.314

0

0

0

0

24.134

19.274

14.415

9.556

4.727

0.686

0

2

20.595

15.747

10.899

6.067

1.666

0.004

0

4

17.220

12.383

7.552

2.928

0.170

0

0

6

14.001

9.177

4.436

0.774

0

0

0

1

22.338

17.485

12.631

7.779

3.054

0.094

0

2

20.595

15.747

10.899

6.067

1.666

0.004

0

3

18.902

14.060

9.219

4.459

0.738

0

0

4

17.258

12.422

7.601

3.051

0.253

0

0

0.005

24.123

19.264

14.404

9.545

4.703

0.627

0

0.01

22.333

17.480

12.626

7.773

3.032

0.096

0

0.015

20.595

15.747

10.899

6.067

1.666

0.004

0

0.02

18.907

14.065

9.225

4.472

0.757

0

0

0

25.945

21.080

16.215

11.350

6.486

1.772

0

2

20.595

15.747

10.899

6.067

1.666

0.004

0

4

15.610

10.780

5.982

1.747

0.027

0

0

6

10.967

6.194

1.990

0.093

0

0

0

0.005

20.580

15.732

10.884

6.044

1.593

0.002

0

0.01

20.595

15.747

10.899

6.067

1.666

0.004

0

0.015

20.620

15.772

10.925

6.107

1.774

0.014

0

0.02

20.656

15.807

10.962

6.167

1.907

0.043

0

η 0.005

λΓ

p

µJ

λN

σJ

134

8 Jump-Enhanced One-Factor Interest-Rate Models

Table 8.2. Values of short-rate caps for the jump-enhanced OU model, with a nominal value of 100 units. K

2

3

4

5

6

7

8

0.005

5.096

4.126

3.160

2.241

1.476

0.909

0.525

0.01

5.952

4.984

4.020

3.097

2.296

1.642

1.136

0.015

6.798

5.833

4.871

3.948

3.135

2.446

1.878

0.02

7.635

6.672

5.712

4.790

3.974

3.267

2.665

0

4.230

3.258

2.295

1.430

0.826

0.445

0.223

2

5.096

4.126

3.160

2.241

1.476

0.909

0.525

4

5.957

4.990

4.024

3.080

2.223

1.512

0.972

6

6.815

5.850

4.886

3.931

3.024

2.215

1.543

1

4.664

3.693

2.727

1.822

1.115

0.636

0.338

2

5.096

4.126

3.160

2.241

1.476

0.909

0.525

3

5.526

4.557

3.592

2.668

1.872

1.243

0.782

4

5.954

4.986

4.022

3.096

2.284

1.613

1.092

0.005

4.231

3.261

2.306

1.435

0.786

0.388

0.175

0.01

4.664

3.694

2.730

1.825

1.104

0.613

0.314

0.015

5.096

4.126

3.160

2.241

1.476

0.909

0.525

0.02

5.525

4.557

3.591

2.666

1.877

1.257

0.800

0

3.797

2.824

1.858

0.993

0.441

0.177

0.065

2

5.096

4.126

3.160

2.241

1.476

0.909

0.525

4

6.385

5.419

4.454

3.512

2.645

1.901

1.302

6

7.665

6.702

5.741

4.789

3.875

3.031

2.288

0.005

5.097

4.127

3.160

2.233

1.444

0.856

0.467

0.01

5.096

4.126

3.160

2.241

1.476

0.909

0.525

0.015

5.093

4.126

3.166

2.266

1.531

0.988

0.610

0.02

5.093

4.132

3.188

2.312

1.606

1.084

0.710

η

λΓ

p

µJ

λN

σJ

8.2 The Ornstein-Uhlenbeck Model

135

Table 8.3. Values of average-rate caps for the jump-enhanced OU model, with a nominal value of 100 units. K

2

3

4

5

6

7

8

0.005

4.037

3.067

2.098

1.168

0.533

0.214

0.077

0.01

4.474

3.507

2.540

1.607

0.915

0.488

0.248

0.015

4.906

3.940

2.975

2.043

1.327

0.840

0.522

0.02

5.331

4.368

3.405

2.474

1.745

1.222

0.851

0

3.593

2.621

1.650

0.754

0.285

0.096

0.029

2

4.037

3.067

2.098

1.168

0.533

0.214

0.077

4

4.478

3.511

2.544

1.597

0.844

0.390

0.162

6

4.918

3.953

2.987

2.033

1.202

0.625

0.291

1

3.816

2.845

1.874

0.953

0.386

0.139

0.045

2

4.037

3.067

2.098

1.168

0.533

0.214

0.077

3

4.257

3.288

2.320

1.387

0.709

0.323

0.134

4

4.476

3.508

2.541

1.607

0.902

0.462

0.219

0.005

3.594

2.622

1.654

0.760

0.251

0.070

0.017

0.01

3.816

2.845

1.875

0.956

0.375

0.126

0.037

0.015

4.037

3.067

2.098

1.168

0.533

0.214

0.077

0.02

4.256

3.288

2.320

1.385

0.715

0.334

0.142

0

3.372

2.399

1.426

0.530

0.126

0.028

0.006

2

4.037

3.067

2.098

1.168

0.533

0.214

0.077

4

4.697

3.731

2.765

1.819

1.046

0.537

0.250

6

5.352

4.389

3.427

2.475

1.623

0.968

0.529

0.005

4.038

3.068

2.099

1.163

0.505

0.182

0.056

0.01

4.037

3.067

2.098

1.168

0.533

0.214

0.077

0.015

4.035

3.066

2.099

1.184

0.574

0.259

0.109

0.02

4.033

3.066

2.106

1.212

0.624

0.311

0.149

η

λΓ

p

µJ

λN

σJ

136

8 Jump-Enhanced One-Factor Interest-Rate Models

8.3 The Square-Root Model 8.3.1 Derivation of the Characteristic Function Modeling the short rate as a Square-Root process was introduced in Cox, Ingersoll and Ross (1985b) to demonstrate the equilibrium approach described in Cox, Ingersoll and Ross (1985a). In contrast to the arbitrage-based approach used in Vasicek (1977), the relevant interest-rate dynamics of the CIR model was derived within an equilibrium-based approach. The main advantage in modeling the short rate as a Square-Root process lies in its nonnegativity property. Thus, interest rates governed by a Square-Root process always stay positive178 . This ability, together with the maintained tractability, oﬀers a very useful tool in modeling the term structure of interest rates. Ahn and Thompson (1988) extend the diﬀusion model with a constant jump size, which is triggered by a Poisson process179 . Zhou (2001) uses a CIR model augmented with a uniformly distributed jump size for estimation purposes. Similar to the Vasicek model, this short-rate process has a mean-reverting component, which is crucial in depicting the term structure faithfully. However, the coeﬃcient governing the diﬀusion part has now a stochastic component governed by the factor xt itself. The the risk-neutral coeﬃcients are µQ 0 = κθ,

µQ 1 = −κ,

σ0 = 0,

√ σ1 = σ xt .

Thus, the diﬀusion part of the SDE (8.1) is √ dxt = κ(θ − xt ) dt + σ xt dWtQ .

(8.9)

Modeling the short-rate process this way bears several advantages. Firstly, as mentioned above, the interest-rate model displays a stochastic volatility without incorporating an additional factor. Secondly, as long as the initial value suﬃces xt ≥ 0 together with the condition 2κθ ≥ σ 2 , the model guarantees that the short rate never reaches the origin and therefore stays strictly positive180 . In contrast to the normally distributed short-rate process in Vasicek (1977), the mean-reverting Square-Root process exhibits a non-central Chi-Square distribution with expectation 178

Setting the discount parameters to w0 = 0 and w1 = 1, the general Square-Root

179

model as used in this thesis and the CIR model coincides. See Ahn and Thompson (1988), p. 168. See Feller (1951), p. 173.

180

8.3 The Square-Root Model

137

EQ [xT ] = xt e−κτ + θ(1 − e−κτ ), and variance 2 σ 2 −κτ σ2 e 1 − e−2κτ . − e−2κτ + θ κ 2κ √ Due to the stochastic volatility term σ xt , the derivation of the general characteristic function is more tedious, but also straightforward. In this case, the ordinary diﬀerential equation for the coeﬃcient function ˜b(z, τ ) has the form VARQ [xT ] = xt

of the well-known Riccati equation, for which several solution methods exist. In order to solve for ˜b(z, τ ), we prepare our diﬀerential equation by substituting the coeﬃcient b(z, τ ) in equation (8.4) with ˜b(z, τ ). This leads to the alternative representation 2 2 2 2 ˜b(z, τ )τ = − w1 + ızκg1 + σ z g1 + ızσ 2 g1 − κ ˜b(z, τ ) + σ ˜b(z, τ )2 . 2 2 Thus, introducing the parameters σ 2 z 2 g12 c0 (z) = − w1 + ızκg1 + , 2 c1 (z) = ızσ 2 g1 − κ, c2 (z) =

σ2 , 2

we are able to express this ODE simply as ˜b(z, τ )τ = c0 (z) + c1 (z)˜b(z, τ ) + c2 (z)˜b(z, τ )2 ,

(8.10)

for which standardized solution techniques exist. Eventually, we obtain the functional form of the coeﬃcient function ˜b(z, τ ) as181 ˜b(z, τ ) = with

(z, τ )

ϑ(z)τ 2

,

(8.11)

ϑ(z)τ ϑ(z)τ (z, τ ) = ϑ(z) cosh − c1 (z) sinh , 2 2

and ϑ(z) = 181

2c0 (z) sinh

0 c1 (z)2 − 4c0 (z)c2 (z).

The detailed derivation of the coeﬃcient functions ˜b(z, τ ) and a0 (z, τ ) is shown in Appendix A.

138

8 Jump-Enhanced One-Factor Interest-Rate Models

Given the coeﬃcient function ˜b(z, τ ), we can proceed onward with the calculation of a0 (z, τ ), which represents the antiderivative of b(z, τ ) = ˜b(z, τ ) + ızg1 , scaled by some constant factor κθ. Applying a logarithmic integration approach, the solution is formally given by

κθ (z, τ ) 0 a (z, τ ) = (ızκθg1 − w0 )τ − τ c1 (z) + 2 ln . 2c2 (z) ϑ(z)

(8.12)

Equipped with these two coeﬃcient functions, we are already able to price interest-rate derivatives for ordinary diﬀusion speciﬁcations of the short rate without considering any jump components. Implementing a jump component in the Square-Root model, one must be careful about the jump speciﬁcations. Due to the strict positiveness of the model, we have to limit ourselves to cases of positively sized exponentially and gamma distributed jump sizes, thus excluding the normal distribution for the jump size speciﬁcations182 . Similar to the Ornstein-Uhlenbeck model, a closed-form formula of the general characteristic function exists only in case of an exponentially distributed jump component. Thus, calculating the jump transform for this speciﬁcation, we obtain for a jump-enhanced SquareRoot model, where the nth (positively directed) jump in xt is governed by an exponential distribution with mean η, the coeﬃcient function (n)

(n)

a1Ex = − λQ τ + λQ × ) 1− c2 (z)c3 (z) + ηc12(z) τ − η ln (z,τ ϑ(z)

η c3 (z)

˜b(z, τ )

c2 (z)c23 (z) + η (c0 (z)η + c1 (z)c3 (z))

(8.13) ,

with c3 (z) = 1 − ızηg1 . For a gamma distributed jump size, we again use a Runge-Kutta solver to recover the relevant values for the coeﬃcient function a1 (z, τ ). 8.3.2 Numerical Results Given the two diﬀerent jump candidates, we want to demonstrate the impact on the density function as well as interest-rate derivative prices. Figures 8.4 182

However, Ahn and Thompson (1988) implemented a constant, negatively sized jump component in a CIR short-rate model. Accordingly, they have to choose carefully the ﬁxed jump amplitude to ensure that interest rates remain positive over the trading interval τ .

8.3 The Square-Root Model

139

and 8.5 depict density functions for short-rate models with diﬀusion parameters rt = 0.03, κ = 0.3, θ = 0.03, σ = 0.1 and T = 1. In each ﬁgure, we focus exclusively on one particular jump candidate, while ignoring other jump speciﬁcations.The probability density functions are examined for diﬀerent jump intensities, means, and in case of a gamma jump size speciﬁcation we also investigate the behavior of the density function for varying p. Thus, we have in each ﬁgure the diﬀusion base model exclusively combined with one jump speciﬁcation. Subsequently, model prices of idealized interest-rate contracts are derived similar to the payoﬀ functions in Table 4.1. Here, we only compute derivative prices for the gamma jump-enhanced diﬀusion model because the gamma distribution is able to generate the exponential distribution as a special case. In contrast to the Vasicek model, the density function of the pure diﬀusion model innately shows an asymmetric shape, since the instantaneous interest rate rt features a non-central chi-square probability density function. The effect of jump components can be seen by comparing the density function of the ordinary CIR diﬀusion model, which is depicted in the particular (upper) left graphs of Figure 8.4 and 8.5 for λ = 0, with the behavior of the jump-enhanced density function. Particularly, empirical ﬁndings of right-skewed density functions183 can be assembled within the jump-enhanced model. As mentioned earlier, we consider only positively sized, exponentially and gamma distributed jumps due to the positivity constraint of the Square-Root process, thus neglecting the normal distribution speciﬁcation for jump candidates in the CIR model. For both jump speciﬁcations we notice a higher skewness of the density function compared to the pure diﬀusion case. However, the jump intensity and jump size mean parameters inﬂuence the density function diﬀerently. According to the (upper) left graphs in Figures 8.4 and 8.5, increased arrival times show the eﬀect of distributing the probability mass over a broader range and shifting the mode of the density to the right, which is characteristic for the intensity parameter. Compared to the Vasicek model, this eﬀect is not that pronounced, which might be due to the non-central chi-squared distribution of the short rate. On the other hand, increasing the parameter of the jump size mean results in fat tails to the right. Accordingly, the density functions display a lower kurtosis. Comparing the particular graphs for the exponential 183

See, for example, Arapis and Gao (2006).

140

8 Jump-Enhanced One-Factor Interest-Rate Models 30

30 0 2 4 6

20

15

10

5

0

0.005 0.01 0.015 0.02

25

probability density

probability density

25

20

15

10

5

0

0.02

0.04

0.06

0.08

0.1

r

T

0.12

0.14

0.16

0.18

0.2

0

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

T

Fig. 8.4. Probability densities for a short rate governed by a CIR diﬀusion model enhanced with an exponentially distributed jump component. In the left (right) graph the density functions for varying jump intensities (means) are depicted. The base parameters are: rt = 0.03, κ = 0.3, θ = 0.03, σ = 0.1, λ = 2, η = 0.005, T = 1.

and gamma jump case and interpret Figure 8.4 as a special case of a gamma distributed jump size variable with p = 1, we clearly identify the multiplying eﬀect of p on the jump intensity. Especially in the upper right graph of Figure 8.5, we notice the extremely ﬂat tail of the density function for η = 0.02 compared to the behavior of the particular graph in Figure 8.4. Examining the eﬀect of jump parameters on derivative prices, we assume for all contingent claims the following base parameters: rt = 0.03, κ = 0.3, θ = 0.03, σ = 0.1, λ = 2, η = 0.005, p = 2 and τ = 0.5. Firstly, we have a look at Table 8.4, where values of zero-bond calls are computed according to a strike range of 60 to 90 units. The strike range is chosen in a way to include either ITM, ATM and OTM option prices184 . As in the Vasicek framework, we observe in Table 8.4 the jump intensity λ to have the greatest inﬂuence on zero-bond call values, followed by the jump mean parameter η and the parameter p. Since varying the jump mean η and the parameter p keeps the mode of the density nearly unchanged, a smaller amount of the probability mass is moved out of the exercise region of the zero-bond call. Comparing zero-bond call prices of the particular parameter settings and strike prices, thus keeping the overall expected jump size, based on η, λΓ and p, equal, we observe relatively low spreads between ITM option prices, while spreads for OTM option prices are high. Turning our attention to the cap contracts, 184

The value of a zero bond with remaining maturity of Tˆ − T = 3 is 85.525 units.

8.3 The Square-Root Model 30

30 0 2 4 6

0.005 0.01 0.015 0.02

25

20

probability density

probability density

25

15

10

5

0

141

20

15

10

5

0

0.05

0.1

0.15

0.2

0.25

0.3

r

0

0

0.05

0.1

0.15

0.2

0.25

0.3

r

T

T

30 1 2 3 4

probability density

25

20

15

10

5

0

0

0.05

0.1

0.15

0.2

0.25

0.3

r

T

Fig. 8.5. Probability densities for a short rate governed by a CIR diﬀusion model enhanced with a gamma distributed jump component. In the upper left (right) graph the density functions for varying jump intensities (means) are depicted. The lower graph shows the density behavior for alternating values of p. The base parameters are: rt = 0.03, κ = 0.3, θ = 0.03, σ = 0.1, λ = 2, η = 0.005, p = 2, T = 1.

we only consider one payment date at the maturity of both the ordinary and the average-rate cap. In Table 8.5, it is ﬁrst of all evident that the inﬂuence of jump parameters is reversed. Thus, the jump mean involves the greatest increase in cap prices, whereas the arrival rate results in a smaller increase in cap prices. Comparing contract values for alternating jump intensities, we have, in absence of any jump, relatively close values for ITM options of the ordinary and the average-rate cap. However, for ATM options we observe relatively large diﬀerences between both contracts. By neglecting positively sized jumps, the opposite eﬀect could be observed for rt < θ, because of the averaging process.

142

8 Jump-Enhanced One-Factor Interest-Rate Models

Table 8.4. Values of zero-bond call options for the jump-enhanced SR model, where the underlying zero-bond contract has a nominal value of 100 units. K

60

65

70

75

80

85

90

23.625

18.711

13.797

8.890

4.117

0.595

0

0.01

17.013

12.128

7.345

3.044

0.339

0

0

0.015

11.185

6.610

2.704

0.297

0

0

0

0.02

6.454

2.762

0.374

0

0

0

0

0

31.018

26.093

21.167

16.242

11.316

6.394

1.770

2

23.625

18.711

13.797

8.890

4.117

0.595

0

4

16.861

11.960

7.096

2.648

0.198

0

0

6

10.678

5.879

1.811

0.079

0

0

0

1

27.225

22.305

17.385

12.466

7.551

2.816

0.121

2

23.625

18.711

13.797

8.890

4.117

0.595

0

3

20.207

15.300

10.400

5.603

1.567

0.023

0

4

16.963

12.068

7.252

2.921

0.310

0

0

η 0.005

λ

p

8.3 The Square-Root Model

143

Table 8.5. Values of short-rate caps (Panel A) and average-rate caps (Panel B), for the jump-enhanced SR model, with a nominal value of 100 units. Panel A: Caps K

2

3

4

5

6

7

8

0.005

1.931

1.146

0.609

0.296

0.135

0.058

0.024

0.01

2.821

2.003

1.375

0.924

0.610

0.395

0.252

0.015

3.705

2.877

2.212

1.692

1.284

0.965

0.719

0.02

4.580

3.748

3.065

2.507

2.043

1.655

1.332

0

1.070

0.443

0.140

0.035

0.007

0.001

0

2

1.931

1.146

0.609

0.296

0.135

0.058

0.024

4

2.811

1.938

1.236

0.735

0.411

0.218

0.110

6

3.699

2.778

1.968

1.316

0.834

0.503

0.290

1

1.490

0.760

0.323

0.118

0.039

0.012

0.003

2

1.931

1.146

0.609

0.296

0.135

0.058

0.024

3

2.376

1.563

0.959

0.557

0.308

0.164

0.085

4

2.821

1.994

1.347

0.876

0.550

0.335

0.199

K

2

3

4

5

6

7

8

0.005

1.451

0.618

0.189

0.050

0.012

0.003

0.001

0.01

1.907

1.052

0.528

0.262

0.129

0.063

0.030

0.015

2.357

1.495

0.923

0.578

0.361

0.224

0.137

0.02

2.801

1.935

1.338

0.942

0.662

0.463

0.322

0

0.994

0.261

0.028

0.001

0

0

0

2

1.451

0.618

0.189

0.050

0.012

0.003

0.001

4

1.909

1.016

0.428

0.154

0.050

0.015

0.004

6

2.367

1.439

0.730

0.319

0.125

0.045

0.015

1

1.222

0.423

0.083

0.012

0.002

0

0

2

1.451

0.618

0.189

0.050

0.012

0.003

0.001

3

1.680

0.830

0.334

0.123

0.043

0.014

0.005

4

1.908

1.048

0.506

0.230

0.100

0.041

0.016

η

λ

p

Panel B: Average-Rate Caps η

λ

p

9 Jump-Enhanced Two-Factor Interest-Rate Models

9.1 Overview In this chapter, we derive the characteristic functions for one speciﬁc additive interest-rate model and one subordinated stochastic volatility interest-rate model. As in the one-factor case, we extend these pure diﬀusion models with additional jump components. The diﬀusion part of the additive model, which is discussed consists of both a factor governed by an Ornstein-Uhlenbeck process, and a factor modeled as a Square-root process. Other popular additive models are given by pure multi-factor versions of Ornstein-Uhlenbeck and Square-Root processes185 . The additive interest-rate model was ﬁrst introduced in Sch¨ obel and Zhu (2000). Here the authors apply a Heston-like transformation methodology to price interest-rate derivatives, as demonstrated in Section 4.2. The subordinated model we choose for our analysis was presented in Fong and Vasicek (1991a). Here, both the short rate and its stochastic volatility are modeled as Square-Root processes with additional jump components. In this thesis, we extend both models to incorporate various jump components. In each model, the behavior of the particular density function and numerical values of idealized interest-rate options are examined. 185

The interest rate is then modeled as the sum either of some Ornstein-Uhlenbeck processes deﬁned by the SDE (8.5) or of some mean-reverting Square-Root processes according to (8.9). Modeling the short rate as an additive Square-Root model, all Brownian motions have to be uncorrelated in order to derive closedform solutions for the general characteristic function.

146

9 Jump-Enhanced Two-Factor Interest-Rate Models

9.2 The Additive OU-SR Model 9.2.1 Derivation of the Characteristic Function Basically, additive multi-factor short-rate models consist only of either additive mean-reverting Ornstein-Uhlenbeck, or Square-Root processes. For example, an additive model for the short rate is used in Chen and Scott (1992), Longstaﬀ and Schwartz (1992), and Chen and Scott (1995). There, the short rate is modeled as the sum of two independent Square-Root processes. A multivariate, additive Gaussian interest-rate model with correlated factors is given in e.g. Langetieg (1980). Collin-Dufresne and Goldstein (2002) also consider both additive multivariate Ornstein-Uhlenbeck and Square Root processes in pricing swaptions. In the case of Ornstein-Uhlenbeck processes, the diﬀerent Brownian motions driving the particular factors can be correlated. On the other hand, if taking an additive Square-Root model, we have to impose the restriction that all Brownian motions be mutually uncorrelated. Otherwise, the separation approach is no longer valid and no closed-form solution for the general characteristic function would exist186 . Exemplary for the set of additive model candidates we select a term-structure model where the short-rate process consists of two factors. The ﬁrst factor xOU is modeled as an Ornsteint Uhlenbeck process, according to equation (8.5), whereas the second factor xSR t is governed by a Square-Root process, as given in equation (8.9)187 . Since we want to extend the model setup, we allow for both factors to include jump components subject to possible non-negativity constraints. Subsequently, the short rate is built as the weighted sum of those factors with a scaling factor w ∈ [0, 1], which gives r(xt ) = wxOU + (1 − w)xSR t t .

(9.1)

Therefore, the coeﬃcients characterizing the short rate are w = (w, 1 − w) and w0 = 0. Accordingly, we use a slightly modiﬁed version of the model setup introduced in Sch¨ obel and Zhu (2000). 186

187

In this case, even a Runge-Kutta solver cannot be applied to the valuation problem, due to the missing system of ODEs. We assume the parameters for the particular processes to be κi , θi , σ i with i ∈ {OU, SR}. Furthermore, we use the payoﬀ-characterizing coeﬃcients g0i and g1i .

9.2 The Additive OU-SR Model

147

Setting w = 1 we obtain the Vasicek model and for w = 0 we obtain the CIR model according to Section 8.3. Although the factors are linearly combined within the short rate, all derivative functions, e.g. the probability density function, are not just simple linear combinations of their particular one-factor counterparts, which is illustrated in Figure 9.1. Thus, the additive process allows more ﬂexibility in modeling the term structure of interest rates compared to the one-factor models discussed in Chapter 8, while maintaining the simple structure of coeﬃcients used in the general characteristic function.

35 0.25 0.5 0.75

30

probability density difference

25 20 15 10 5 0 −5 −10 −15

0

0.01

0.02

0.03

0.04

0.05 r

0.06

0.07

0.08

0.09

0.1

T

Fig. 9.1. Diﬀerences of the pure diﬀusion OU-SR model density function and the sum of the particular one-factor pendants for diﬀerent weighting factors. The parameters used are: xt = (0.05, 0.03) , κ = (0.4, 0.3) , θ = (0.05, 0.03) , σ = (0.01, 0.1) , T = 1. In case of the one-factor models the ﬁrst (last) elements correspond to the Vasicek (CIR) model.

Due to the independence of the two Brownian motions, the particular timedependent coeﬃcients exhibit the same formal structure as the ones derived in the one-factor Vasicek and CIR interest-rate models. Thus, the general char-

148

9 Jump-Enhanced Two-Factor Interest-Rate Models

acteristic function of this additive interest-rate model has the time-dependent vector function ˜ τ) = b(z,

˜bOU (z, τ ) , ˜bSR (z, τ )

with ˜bOU (z, τ ) and ˜bSR (z, τ ) given by their one-factor representations in equation (8.6) and (8.11) with adapted parameters. Consequently, we obtain for the coeﬃcient function a(z, τ ) the relation a0 (z, τ ) = a0OU (z, τ ) + a0SR (z, τ ), where a0OU (z, τ ) and a0SR (z, τ ) correspond to equation (8.7) and (8.12). The jumps contained in the vectors jxOU and jxSR are both triggered by the same Poisson vector process N(λQ )188 .Due to the independence of the two factors governing the short rate, we are also able to adapt the jump transforms of the particular one-factor models without altering their formal structure. 9.2.2 Numerical Results In this section, we show the behavior of the density function and compute values for some common interest-rate options under the additive jump-diﬀusion model. As base parameters for both the density function and the interestrate contracts, we use the particular parameters according to their one-factor counterparts. The default value of the scaling parameter w is set to 12 . The impact of jumps on the short-rate density is demonstrated in Figures 9.2 and 9.3. As before, we focus exclusively on one particular jump candidate, while ignoring other jump speciﬁcations. Thus, the graphs in the ﬁrst row and the left graph in the second row in Figure 9.2, respectively, display only the impact of gamma jump component of the Ornstein-Uhlenbeck process. The other three graphs in this ﬁgure depict the inﬂuence of the normal jump component on the short-rate density. Consequently, in Figure 9.3, we only consider the gamma jump component of the Square-Root process. Option prices for varying parameters of normally and gamma distributed jump amplitudes are given in Tables 9.1 - 9.6. Again, we focus on the sensitivity of option prices to jump parameters and neglect the exponentially distributed jump size since the exponential distribution is a special case of the gamma distribution. 188

However, setting elements in the jump vectors jxOU and jxSR to zero, it is possible to assign jump components to particular processes.

9.2 The Additive OU-SR Model

149

Comparing the ﬁgures of the densities functions depicted in 8.2, 8.3 and 8.5 with the corresponding graphs in Figures 9.2 and 9.3, we obviously notice more skewness in the densities of the two-factor model compared to the densities in the Vasicek model and a more leptokurtic behavior of the densities in the additive model compared to the ones in a CIR model. Firstly, taking a look at the inﬂuence of the gamma distributed jump size speciﬁcation, we determine a similar inﬂuence of the gamma jump component belonging to the Vasicek and CIR model. However, the Vasicek part has a signiﬁcantly weaker eﬀect on the probability density function compared to the relevant one-factor model. On the other hand, the gamma distributed jump size component of the Square-Root process has a considerable impact on the probability density, which can be justiﬁed by the similar shape of the probability density functions in Figures 8.5 and 9.3. Examining the impact of a normally distributed jump component in this model, we observe only small changes in contrast to the one-factor equivalent Vasicek model. Thus, in the multi-factor setup we no longer encounter the characteristic strong curvature in the probability density function and the bimodal distribution displayed in the upper right graph of Figure 8.3. However, the eﬀect of an increased volatility of the normally distributed jump size, which raises both tails of the probability density function, remains immanent. Computing numerical values of interest-rate derivatives in this model, we assume for all contingent claims the following base diﬀusion parameters: 0.05 0.4 0.05 0.01 xt = , κ= , θ= and σ = . 0.03 0.3 0.03 0.1 In each vector, the ﬁrst element corresponds to the Ornstein-Uhlenbeck part, whereas the second element states the parameter value for the Square-Root component in this particular additive interest-rate model. The default jump parameters used for the valuation are in case of a gamma distributed jump size candidate λiΓ = 2, η i = 0.005 and pi = 2 with i ∈ {OU, SR}. The normally distributed jump component of the Ornstein-Uhlenbeck process is governed by the parameters λOU = 2, µOU = 0.015 and σJOU = 0.01. All N J contracts have a remaining time to maturity of a half year. Let us discuss ﬁrst Tables 9.1 and 9.2, where numerical values of zero-bond calls are reported according to equation (3.12) for a strike range from 60 to 90 units. The strike

150

9 Jump-Enhanced Two-Factor Interest-Rate Models 50

50 0 2 4 6

45 40

40 35

probability density

probability density

35 30 25 20

30 25 20

15

15

10

10

5 0

0.005 0.01 0.015 0.02

45

5 0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0

0.2

0

0.02

0.04

0.06

0.08

r

40

0.18

0.2

1 2 3 4

40 35

probability density

probability density

0.16

45

35 30 25 20

30 25 20

15

15

10

10

5

5 0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0

0.2

0

0.02

0.04

0.06

0.08

r

0.1

0.12

0.14

0.16

0.18

0.2

r

T

T

50

50 0 0.01 0.02 0.03

45 40

0.005 0.01 0.015 0.02

45 40 35

probability density

35

probability density

0.14

50 1 2 3 4

45

30 25 20

30 25 20

15

15

10

10

5 0

0.12

T

50

0

0.1

r

T

5 0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0

0

0.02

0.04

rT

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

rT

Fig. 9.2. Probability densities for a short rate governed by an OU-SR diﬀusion model enhanced with either a gamma or normally distributed jump component for the OU process. In the upper left (right) graph density functions for varying jump intensities (means) of the gamma distributed jump component are depicted. The graphs in the second row show the density behavior for alternating values of pOU and the jump intensity λOU N of the normally distributed jump component. In the last row, the left (right) graph shows density functions for diﬀerent values of jump mean (volatility) of the normally distributed jump component. The base parameters are: = 2, η OU = xt = (0.05, 0.03) , κ = (0.4, 0.3) , θ = (0.05, 0.03) , σ = (0.01, 0.1) , λOU Γ OU 0.005, pOU = 2, λOU = 0.02, σJOU = 0.01, T = 1. N = 1, µJ

9.2 The Additive OU-SR Model 50

50 0 2 4 6

45 40

40 35

probability density

probability density

0.005 0.01 0.015 0.02

45

35 30 25 20

30 25 20

15

15

10

10

5 0

151

5 0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

0

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

T

T

50 1 2 3 4

45 40

probability density

35 30 25 20 15 10 5 0

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

T

Fig. 9.3. Probability densities for a short rate governed by an OU-SR diﬀusion model enhanced with a gamma distributed jump component for the SR process. In the upper left (right) graph the density functions for varying jump intensities (means) are depicted. The lower graph shows the density behavior for alternating values of p. The base parameters are: xt = (0.05, 0.03) , κ = (0.4, 0.3) , θ = = 2, η SR = 0.005, pSR = 2, T = 1. (0.05, 0.03) , σ = (0.01, 0.1) , λSR Γ

range is chosen to include either ITM, ATM and OTM option prices189 . As encountered in the one-factor framework, we observe in case of the gamma jump-size distribution that the intensity λiΓ has the greatest inﬂuence on zero-bond call values, followed by the parameter pi and the jump mean η i . This can be explained by the shifting of the density function to the right for increasing arrival rates. Contrary, varying jump mean η i and parameter pi keeps the mode of the density nearly unchanged, so that a smaller amount of the probability mass is moved out of the zero-bond call exercise region. Varying the parameters of the normally distributed jump component, we also 189

The value of a zero bond with remaining maturity of two years, priced with base parameters, is 87.359 units.

152

9 Jump-Enhanced Two-Factor Interest-Rate Models

observe a very similar behavior of option prices in comparison to the onefactor model. Comparing relative zero-bond call price diﬀerences of particular parameter settings and strike rates, thus keeping the overall expected jump size equal, we observe low spreads between ITM option prices, whereas spreads for OTM option prices are relative high. For the cap contracts, we only allow one payment date, which is at maturity. At ﬁrst, we observe the same positive eﬀect of the jump components as encountered in the one-factor pendants. Thus, the jump mean involves the greatest increase of cap prices, whereas the arrival rate results in a smaller increase of cap prices. However, due to the scaling factor w, the eﬀect of diﬀerent jump components is not as strong as we encountered in the particular one-factor models. Comparing the eﬀect of a gamma distributed jump on the average-rate cap, we compute nearly the same contract values, whether we have the jumps in the Ornstein-Uhlenbeck or in the Square-Root part of the model.

9.2 The Additive OU-SR Model

153

Table 9.1. Values of zero-bond call options for the jump-enhanced OU-SR model, where the underlying zero-bond contract has a nominal value of 100 units. K

60

65

70

75

80

85

90

24.825

19.944

15.063

10.183

5.304

0.875

0

OU

η

0.005 0.01

23.022

18.147

13.271

8.396

3.573

0.201

0

0.015

21.275

16.406

11.537

6.680

2.137

0.021

0

0.02

19.584

14.720

9.862

5.092

1.125

0

0

0

26.688

21.801

16.915

12.028

7.142

2.326

0.003

2

24.825

19.944

15.063

10.183

5.304

0.875

0

4

23.006

18.131

13.256

8.380

3.531

0.152

0

6

21.230

16.360

11.491

6.622

1.926

0.006

0

1

25.750

20.866

15.982

11.098

6.215

1.521

0

2

24.825

19.944

15.063

10.183

5.304

0.875

0

3

23.914

19.036

14.158

9.280

4.412

0.438

0

4

23.017

18.141

13.266

8.391

3.556

0.190

0

0.005

26.685

21.799

16.912

12.025

7.139

2.312

0.005

0.01

25.748

20.865

15.981

11.097

6.214

1.511

0

0.015

24.825

19.944

15.063

10.183

5.304

0.875

0

0.02

23.916

19.037

14.159

9.281

4.415

0.445

0

0

27.631

22.741

17.852

12.962

8.073

3.197

0.029

2

24.825

19.944

15.063

10.183

5.304

0.875

0

4

22.118

17.245

12.373

7.501

2.711

0.048

0

6

19.505

14.641

9.777

4.924

0.792

0

0

0.005

24.821

19.940

15.059

10.178

5.299

0.847

0

0.01

24.825

19.944

15.063

10.183

5.304

0.875

0

0.015

24.832

19.951

15.070

10.189

5.312

0.918

0

0.02

24.841

19.960

15.080

10.199

5.325

0.973

0.001

λOU Γ

OU

p

µOU J

λOU N

σJOU

154

9 Jump-Enhanced Two-Factor Interest-Rate Models

Table 9.2. Values of zero-bond call options for the jump-enhanced OU-SR model, where the underlying zero-bond contract has a nominal value of 100 units. K

60

65

70

75

80

85

90

24.825

19.944

15.063

10.183

5.304

0.875

0

SR

η

0.005 0.01

22.891

18.016

13.141

8.266

3.466

0.183

0

0.015

21.022

16.153

11.284

6.442

1.997

0.014

0

0.02

19.217

14.354

9.504

4.794

1.001

0

0

0

26.828

21.941

17.055

12.168

7.281

2.444

0.004

2

24.825

19.944

15.063

10.183

5.304

0.875

0

4

22.873

17.997

13.122

8.247

3.407

0.136

0

6

20.969

16.100

11.230

6.362

1.737

0.004

0

1

25.819

20.935

16.051

11.167

6.284

1.566

0

2

24.825

19.944

15.063

10.183

5.304

0.875

0

3

23.847

18.969

14.091

9.213

4.349

0.422

0

4

22.885

18.010

13.135

8.260

3.443

0.173

0

λSR Γ

SR

p

9.2 The Additive OU-SR Model

155

Table 9.3. Values of short-rate caps for the jump-enhanced OU-SR model, with a nominal value of 100 units. K

2

3

4

5

6

7

8

0.005

3.507

2.532

1.593

0.832

0.358

0.129

0.040

0.01

3.943

2.969

2.024

1.220

0.651

0.312

0.136

0.015

4.376

3.403

2.457

1.635

1.012

0.592

0.330

0.02

4.806

3.835

2.889

2.058

1.404

0.928

0.597

0

3.069

2.094

1.186

0.534

0.193

0.058

0.015

2

3.507

2.532

1.593

0.832

0.358

0.129

0.040

4

3.944

2.970

2.014

1.176

0.579

0.242

0.087

6

4.380

3.407

2.443

1.554

0.854

0.403

0.164

1

3.288

2.313

1.384

0.668

0.261

0.085

0.023

2

3.507

2.532

1.593

0.832

0.358

0.129

0.040

3

3.726

2.751

1.807

1.015

0.482

0.196

0.069

4

3.943

2.969

2.023

1.210

0.631

0.288

0.117

0.005

3.070

2.096

1.184

0.518

0.176

0.048

0.011

0.01

3.289

2.314

1.384

0.662

0.252

0.079

0.021

0.015

3.507

2.532

1.593

0.832

0.358

0.129

0.040

0.02

3.725

2.751

1.807

1.019

0.489

0.202

0.073

0

2.850

1.875

0.972

0.370

0.107

0.025

0.005

2

3.507

2.532

1.593

0.832

0.358

0.129

0.040

4

4.162

3.188

2.232

1.378

0.736

0.340

0.137

6

4.814

3.842

2.878

1.971

1.210

0.660

0.320

0.005

3.508

2.532

1.589

0.815

0.335

0.112

0.031

0.01

3.507

2.532

1.593

0.832

0.358

0.129

0.040

0.015

3.507

2.533

1.603

0.859

0.392

0.155

0.054

0.02

3.506

2.537

1.621

0.896

0.435

0.189

0.075

OU

η

λOU Γ

OU

p

µOU J

λOU N

σJOU

156

9 Jump-Enhanced Two-Factor Interest-Rate Models

Table 9.4. Values of short-rate caps for the jump-enhanced OU-SR model, with a nominal value of 100 units. K

2

3

4

5

6

7

8

0.005

3.507

2.532

1.593

0.832

0.358

0.129

0.040

0.01

3.953

2.979

2.035

1.232

0.664

0.325

0.147

0.015

4.397

3.424

2.478

1.657

1.037

0.618

0.353

0.02

4.838

3.866

2.920

2.089

1.439

0.965

0.633

0

3.059

2.083

1.173

0.518

0.181

0.052

0.013

2

3.507

2.532

1.593

0.832

0.358

0.129

0.040

4

3.955

2.980

2.026

1.190

0.593

0.252

0.093

6

4.402

3.428

2.465

1.579

0.881

0.427

0.181

1

3.283

2.308

1.378

0.661

0.254

0.080

0.021

2

3.507

2.532

1.593

0.832

0.358

0.129

0.040

3

3.731

2.756

1.813

1.021

0.490

0.202

0.073

4

3.954

2.980

2.034

1.222

0.645

0.301

0.127

SR

η

λSR Γ

SR

p

9.2 The Additive OU-SR Model

157

Table 9.5. Values of average-rate caps for the jump-enhanced OU-SR model, with a nominal value of 100 units. K

2

3

4

5

6

7

8

0.005

2.753

1.777

0.832

0.229

0.041

0.005

0.001

0.01

2.977

2.002

1.053

0.389

0.112

0.027

0.006

0.015

3.199

2.225

1.275

0.579

0.232

0.087

0.032

0.02

3.419

2.446

1.496

0.781

0.385

0.184

0.087

0

2.528

1.551

0.625

0.135

0.019

0.002

0

2

2.753

1.777

0.832

0.229

0.041

0.005

0.001

4

2.978

2.003

1.046

0.349

0.077

0.012

0.002

6

3.202

2.228

1.264

0.494

0.129

0.025

0.004

1

2.641

1.664

0.725

0.174

0.027

0.003

0

2

2.753

1.777

0.832

0.229

0.041

0.005

0.001

3

2.865

1.890

0.941

0.298

0.064

0.010

0.001

4

2.977

2.002

1.052

0.379

0.099

0.020

0.003

0.005

2.528

1.551

0.625

0.123

0.014

0.001

0

0.01

2.641

1.664

0.725

0.168

0.024

0.002

0

0.015

2.753

1.777

0.832

0.229

0.041

0.005

0.001

0.02

2.865

1.889

0.941

0.302

0.067

0.011

0.002

0

2.416

1.438

0.516

0.079

0.007

0.001

0

2

2.753

1.777

0.832

0.229

0.041

0.005

0.001

4

3.089

2.115

1.158

0.434

0.112

0.022

0.003

6

3.424

2.452

1.488

0.680

0.227

0.057

0.012

0.005

2.753

1.777

0.829

0.215

0.033

0.004

0

0.01

2.753

1.777

0.832

0.229

0.041

0.005

0.001

0.015

2.753

1.777

0.838

0.248

0.053

0.009

0.001

0.02

2.752

1.777

0.849

0.272

0.069

0.015

0.003

OU

η

λOU Γ

OU

p

µOU J

λOU N

σJOU

158

9 Jump-Enhanced Two-Factor Interest-Rate Models

Table 9.6. Values of average-rate caps for the jump-enhanced OU-SR model, with a nominal value of 100 units. K

2

3

4

5

6

7

8

0.005

2.753

1.777

0.832

0.229

0.041

0.005

0.001

0.01

2.980

2.005

1.056

0.393

0.116

0.030

0.007

0.015

3.206

2.232

1.282

0.587

0.240

0.094

0.035

0.02

3.429

2.457

1.507

0.792

0.396

0.195

0.094

0

2.524

1.547

0.621

0.130

0.017

0.002

0

2

2.753

1.777

0.832

0.229

0.041

0.005

0.001

4

2.981

2.006

1.050

0.354

0.079

0.013

0.002

6

3.209

2.235

1.272

0.502

0.136

0.027

0.004

1

2.639

1.662

0.723

0.171

0.025

0.003

0

2

2.753

1.777

0.832

0.229

0.041

0.005

0.001

3

2.867

1.891

0.943

0.300

0.066

0.011

0.002

4

2.980

2.005

1.055

0.383

0.103

0.022

0.004

SR

η

λSR Γ

SR

p

9.3 The Fong-Vasicek Model

159

9.3 The Fong-Vasicek Model 9.3.1 Derivation of the Characteristic Function Apart from the additive modeling approach, the term structure within our exponential-aﬃne framework can also be modeled with the help of subordinated factors. Hence, it is possible to explicitly incorporate the long term mean and/or the volatility as a stochastic factor itself190 . For example, the assumption of a constant volatility in one-factor short-rate models is frequently criticized. In Fong and Vasicek (1991b), the authors argue that a model with a deterministic volatility parameter cannot produce a meaningful volatility exposure. Therefore, they propose to model the variance of a CIR-like shortrate model as a Square-Root process itself. In this section, we use a slightly modiﬁed version of the Fong and Vasicek (1991a) model. Thus, the SDEs for the base diﬀusion model are191 √ drt = κ(θ − rt ) dt + σ vt dW1t , √ dvt = α(¯ v − vt ) dt + β vt dW2t .

(9.2) (9.3)

Equivalently to the target rate θ, often also referred to as the long term mean, of the short rate, v¯ expresses the parameter for a long-term mean of the variance factor vt . In addition to Fong and Vasicek (1991a), we extend this base diﬀusion model with additional jump components jr and jv for the short rate and its volatility factor192 , both triggered by the same vector of Poisson processes N(λQ ). In contrast to the additive OU-SR model discussed in the last section, the Brownian motions W1t and W2t can be correlated as follows: 190

Beaglehole and Tenney (1991) model the long term mean θ of a mean-reverting, normally distributed short rate as a subordinated factor governed by an OrnsteinUhlenbeck process. In Balduzzi, Das, Foresi and Sundaram (1996), the authors model a CIR like short rate with subordinated stochastic mean and volatility factor. However, in both cases, the authors do not give any option prices and

191

derive only zero-bond prices and yields of zero bonds, respectively. For this interest-rate model, we assume w0 = 0 and w = (1, 0) . However, the derivation of the time-dependent coeﬃcients of the characteristic function

192

is shown in Appendix B for general discounting parameter values w0 and w. We only allow strictly positively sized jumps, thus restricting ourselves to positively directed exponentially and gamma distributed jump amplitudes.

160

9 Jump-Enhanced Two-Factor Interest-Rate Models

dW1t dW2t = ρ dt.

(9.4)

However, according to Section 2.1, all Brownian motions have to be uncorrelated within our modeling approach. Fortunately, the feature of correlated Brownian motions can be easily incorporated into our framework193 by using the following diﬀusion-speciﬁc matrix √ Σ(vt ) = β vt

σ β

0

0 1 − ρ2

ρ

.

For convenience, we introduce for this model the following representation of time-dependent coeﬃcient functions194 a0 (z, τ ) = A0 (z, τ ),

and b(z, τ ) =

B(z, τ )

+ ız

C(z, τ )

¯ B C¯

.

Thus, in order to derive the general characteristic function of the state vector xt = (rt , vt ) , we explicitly have to solve the following system of ODEs ¯ + α¯ ¯ A0 (z, τ )τ =κθ(B(z, τ ) + ız B) v (C(z, τ ) + ız C) =A01 (z, τ )τ + A02 (z, τ )τ , ¯ − w, B(z, τ )τ = − κ(B(z, τ ) + ız B)

(9.5) (9.6)

¯ C(z, τ )τ = − α(C(z, τ ) + ız C) 2 σ2 ¯ 2 + β (C(z, τ ) + ız C) ¯ 2 (B(z, τ ) + ız B) 2 2 ¯ ¯ + σβρ(B(z, τ ) + ız B)(C(z, τ ) + ız C).

+

(9.7)

Hence, we are dealing with a system of coupled ODEs. Fortunately, there are no two-sided interdependencies, enabling us to successively solve the diﬀerential equations one-by-one. Starting with equation (9.6), the solution to this diﬀerential equation is easy to obtain and coincides with equation (8.6) for ¯ Also straightforward, but more tedious, is the derivation w1 = w and g1 = B. 193

A standard decomposition of two correlated Brownian motions is applied. Thus, two correlated Brownian motions as given in equation (9.4) allow for the alternative representation dW2t = ρ dW1t +

194

∗ ∗ 1 − ρ2 dW2t , where the processes W2t

and W1t are neither correlated. ¯ In this model setup, the constant parameter g0 is represented by the term A.

9.3 The Fong-Vasicek Model

161

of the coeﬃcient function solving the ODE (9.7). Performing some appropriate transformations on the particular ODE, the solution of the coeﬃcient function corresponding to vt can be obtained as195 C(z, τ ) = − M (z, τ ) + J(z, τ ) (1 + Q(z) − S(z))KU[Q(z) + 1; S(z); Y (z, τ )] + Υ(z) KM[Q(z) + 1; S(z); Y (z, τ )] , with κ M (z, τ ) = 2 β

f1 (z) + 2Q(z) − S(z) + 1+ κ

ρ

1+ 0 ρ2 − 1

(9.8)

Y (z, τ ) ,

2κQ(z) , β 2 (KU[Q(z); S(z); Y (z, τ )] + Υ(z) KM[Q(z); S(z); Y (z, τ )]) 0 σβ ρ2 − 1 w ¯ e−κτ , Y (z, τ ) = + ız B κ κ J(z, τ ) =

S(z) (f3 (z) + κ)ρ − βf2 (z) 0 , + 2 2κ ρ2 − 1 ( f3 (z)2 − 2β 2 f1 (z) S(z) = 1 + , κ2 M (z, 0)KU[Q(z); S(z); Y (z, 0)] Υ(z) = Ξ(z) (1 + Q(z) − S(z)) KU[Q(z) + 1; S(z); Y (z, 0)] − , β2 2κQ(z) Ξ(z) κ Ξ(z) = 2 2 Q(z)KM[Q(z) + 1; S(z); Y (z, 0)] β

Q(z) =

− M (z, 0)KM[Q(z); S(z); Y (z, 0)], and

2¯ 2 2 ¯ α − ızβ C + σβρw + σ w , f1 (z) = −ız C 2 κ 2κ2 σw , f2 (z) = ızβρC¯ − κ σβρw . f3 (z) = ızβ 2 C¯ − α − κ

In the equations above the function KM[a; b; y] is commonly referred to as the Kummer function (of the ﬁrst kind) and KU[a; b; y] represents a conﬂuent 195

The detailed derivation of the coeﬃcient function C(z, τ ) is given in Appendix B.

162

9 Jump-Enhanced Two-Factor Interest-Rate Models

hypergeometric function196 . Both functions represent the two independent solutions of the Kummer equation197 . In contrast to the last derivation, both parts of the coeﬃcient function A (z, τ ) are relatively easy to obtain. Due to the simple structure of the 0

solution given in equation (9.6), we are able to state immediately A01 (z, τ ) = −θ (B(z, τ ) + wτ ) .

(9.9)

For the second part of the diﬀusion component in A0 (z, τ ), we exploit the formal structure and attempt a logarithmic integration, which is shown in Appendix B. Thus, the solution of the second part of the coeﬃcient function A0 (z, τ ) can be written as198 A02 (z, τ ) = −

L(z, τ ) J(z, 0) 2α¯ v ¯ ln + ızα¯ v Cτ, β2 L(z, 0) J(z, τ )

with ln [L(z, τ )] =

S(z) − 1 +

f3 (z) κ

2

ρ

ln[Y (z, τ )]

Y (z, τ ) 1+ 0 2 ρ2 − 1 0 2 1 f3 (z) β ρ −1 + + . ln 2 2κ κ

−

(9.10)

(9.11)

Having calculated the diﬀusion-related coeﬃcients of the general characteristic function, we are now ready for the corresponding jump parts. In case 196

The conﬂuent hypergeometric function is sometimes also denoted as the Kummer function of the second kind and is – like the complex-valued square-root and logarithm – a multi-valued function. Thus, one has to track carefully the path of integration by using this type of function to avoid discontinuities according to the

197

principal branch used by standard mathematical programming environments. The functions KM[a; b; y] and KU[a; b; y] are two independent solutions of the diﬀerential equation y

d2 w(y) dy 2

+ (b − y) dw(y) = aw(y). More information on condy

ﬂuent hypergeometric functions, especially about the computation of KM[a; b; y] 198

and KU[a; b; y] can be found in Abramowitz and Stegun (1972), p. 504. The detailed derivation of A02 (z, τ ) is given in Appendix B.

9.3 The Fong-Vasicek Model

163

of a exponential jump size speciﬁcation on the short rate we are able to use the same jump transformation derived for the Vasicek model. In any other case – meaning gamma distributed jump sizes in the short rate, and for any jump components incorporated in the volatility factor199 – we have to apply a Runge-Kutta algorithm. 9.3.2 Numerical Results In this subsection, we want to demonstrate the impact of diﬀerent jump components and the correlation parameter ρ on the probability density function as well as on option prices with payoﬀ functions similar to Table 4.1, respectively. To the best of our knowledge, option prices for this model, whether of the exponential-aﬃne, linear, or integro-linear type, are presented for the ﬁrst time in this thesis. Articles do exist, which cover the computation of numerical values under the base model. However, only prices of unconditional contracts, such as zero bonds and likewise yields of zero-bond prices, are computed200 . These model prices are easy to obtain due to their similarity of the moment-generating function of the short rate201 . The base diﬀusion parameters we use in computing the probability density function are rt = 0.08, vt = 0.04, κ = 0.2, θ = 0.08, σ = 0.1, α = 0.4, v¯ = 0.04, β = 0.1, ρ = −0.5 and T = 1. Firstly, we examine the behavior of the probability density for alternating correlation speciﬁcations. To avoid the results being biased by the inﬂuence of jump components, we examine ﬁrst the density function of the pure diﬀusion model according to equations (9.2) and (9.3). Obviously, Figure 9.4 shows that the correlation between the short rate and its volatility has an eﬀect on the probability density function, and therefore on the price of any contingent claim. Thus, for low interest rates 199

200

Due to the complicated structure of the coeﬃcient function C(z, τ ), even for exponentially distributed jump amplitudes there exist no closed-form jump transforms. See Fong and Vasicek (1991b). Selby and Strickland (1995) also compute numerical values of zero-bond prices for the base model, but present a technique avoiding the application of hypergeometric functions. In Balduzzi, Das, Foresi and Sundaram (1996), zero-bond prices are computed for an extended version of the Fong-Vasicek model, where the mean of the short rate is also modeled as a

201

stochastic factor. See Proposition 2.4.3.

164

9 Jump-Enhanced Two-Factor Interest-Rate Models

18 −1 0 1

16

probability density

14 12 10 8 6 4 2 0

0

0.02

0.04

0.06

0.08

0.1 rT

0.12

0.14

0.16

0.18

0.2

Fig. 9.4. Probability density functions of the Fong-Vasicek pure diﬀusion model for diﬀerent values of the correlation parameter ρ. The parameters used are: rt = 0.08, vt = 0.04, κ = 0.2, θ = 0.08, σ = 0.1, α = 0.4, v¯ = 0.04, β = 0.1, T = 1.

and negatively correlated factors we observe substantially increased values of the probability density function, in contrast to the probability density function with positively correlated random variables. Since in this scenario we encounter a tendency toward higher volatilities in case of low interest rates, we are dealing with a more volatile process, which explains the behavior of the density. Finally, the correlation parameter can be used to adjust the skewness of the probability probability density function, which is advantageous for calibrating empirical term structures. Next, we examine the inﬂuence of jump speciﬁcations on the density function of the short rate. Thus, we ﬁx the correlation to ρ = −0.5. Comparing the graphs in Figure 9.5 with the particular graphs in Figure 8.5, we observe the same behavior of the density functions considering a gamma distributed jump component on the short rate. This fact is not surprising since the short rate in the Fong-Vasicek interest-rate model is governed by a Square-Root

9.3 The Fong-Vasicek Model 25

25 0 2 4 6

0.005 0.01 0.015 0.02

20

probability density

probability density

20

15

10

5

0

165

15

10

5

0

0.05

0.1

0.15

0.2

0.25

0.3

0

0

0.05

r

0.1

0.15

0.2

0.25

0.3

r

T

T

25 1 2 3 4

probability density

20

15

10

5

0

0

0.05

0.1

0.15

0.2

0.25

0.3

r

T

Fig. 9.5. Probability densities for a short rate governed by the Fong-Vasicek diffusion model enhanced with a gamma distributed jump component for the shortrate process. In the upper left (right) graph the density functions for varying jump intensities (means) are depicted. The lower graph shows the density behavior for alternating values of p. The base parameters are: rt = 0.08, vt = 0.04, κ = 0.2, θ = 0.08, σ = 0.1, α = 0.4, v¯ = 0.04, β = 0.1, ρ = −0.5, λr = 2, ηr = 0.005, pr = 2, T = 1.

process. However, allowing for a gamma distributed jump in the volatility process, the density functions in Figure 9.6 show rather increased values on both tails while maintaining their mode. Distinguishing between the impact of jump parameters on the density, we clearly identify the jump intensity to have the strongest inﬂuence on the short-rate process. The base diﬀusion parameters used for the computation of theoretical prices given in Tables 9.7 and 9.8 are the same as above. The default parameters for the diﬀerent gamma distributed jump components are λ = 2, η = 0.005 and p = 2 for both the short-rate and volatility process. All contracts have a remaining time to maturity of half a year. At ﬁrst, we turn our attention to Table 9.7, where zero-bond call option prices are given for

166

9 Jump-Enhanced Two-Factor Interest-Rate Models 25

25 0 2 4 6

20

probability density

probability density

20

15

10

5

0

0.005 0.0075 0.01 0.0125

15

10

5

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

0

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

T

T

25 1 2 3 4

probability density

20

15

10

5

0

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

T

Fig. 9.6. Probability densities for a short rate governed by the Fong-Vasicek diffusion model enhanced with a gamma distributed jump component for the volatility process. In the upper left (right) graph the density functions for varying jump intensities (means) are depicted. The lower graph shows the density behavior for alternating values of p. The base parameters are: rt = 0.08, vt = 0.04, κ = 0.2, θ = 0.08, σ = 0.1, α = 0.4, v¯ = 0.04, β = 0.1, ρ = −0.5, λv = 2, ηv = 0.005, pv = 2, T = 1.

a strike range from 60 to 90 units. The range is chosen to cover either ITM, ATM and OTM option prices202 . Similar to ﬁndings for the probability density function, the jump behavior of gamma distributed jump component on the short rate under the Fong-Vasicek model and also in the Square-Root model shows a strong resemblance on account of the Square-Root process. Looking at the gamma distributed jump component in the volatility process, we observe higher values of the particular derivative price for increasing jump parameters. This eﬀect is due to the more increased tails of the density in comparison to the ordinary diﬀusion case. Accordingly, the eﬀect of volatility 202

The value of a zero bond with remaining maturity of two years, priced with base parameters, is 82.335 units.

9.3 The Fong-Vasicek Model

167

jumps on derivative contracts is small, which explains the nearly linear relation between jump parameters and contract values. Turning our attention to Tables 9.8 and 9.9, we computed prices for caps with one payment date made at the maturity of the particular contract. Thus, we examine cap prices similar to the idealized payoﬀ function in Table 4.1. Since we restrict this model to positively sized jump components due to the Square-Root limitations, we have the ordinary cap dominating its average-rate counterpart. This is a direct consequence of the geometric average of an increasing function. Comparing the corresponding values of the one-factor CIR model with the cap values in this section, we indicate the same impact of jumps on the short rate generally. Thus, this model also has a less sensitive average-rate option in contrast to the ordinary cap contract. However, jumps on the volatility component have roughly speaking no eﬀect at all on average-rate options, whereas ordinary caps encounter small changes for either ITM, ATM and OTM option values.

168

9 Jump-Enhanced Two-Factor Interest-Rate Models

Table 9.7. Values of zero-bond call options for the jump-enhanced Fong-Vasicek model, where the underlying zero-bond contract has a nominal value of 100 units. K

60

65

70

75

80

85

90

20.232

15.439

10.647

5.867

1.562

0.052

0

ηr 0.005 0.01

16.456

11.68

6.965

2.689

0.255

0.002

0

0.015

12.970

8.316

4.050

0.921

0.021

0

0

0.02

9.868

5.579

2.088

0.181

0.001

0

0

0

24.302

19.498

14.694

9.890

5.087

0.910

0.015

2

20.232

15.439

10.647

5.867

1.562

0.052

0

4

16.378

11.597

6.828

2.393

0.169

0.001

0

6

12.728

7.967

3.409

0.469

0.006

0

0

1

22.232

17.434

12.636

7.838

3.103

0.237

0.002

2

20.232

15.439

10.647

5.867

1.562

0.052

0

3

18.299

13.512

8.730

4.067

0.660

0.010

0

4

16.430

11.651

6.910

2.596

0.238

0.002

0

0.005

20.232

15.439

10.647

5.867

1.562

0.052

0

0.01

20.247

15.455

10.663

5.885

1.597

0.064

0

0.015

20.263

15.471

10.678

5.902

1.631

0.077

0.001

0.02

20.279

15.486

10.694

5.920

1.664

0.090

0.002

0

20.216

15.424

10.631

5.850

1.525

0.041

0

2

20.232

15.439

10.647

5.867

1.562

0.052

0

4

20.247

15.455

10.663

5.884

1.598

0.063

0

6

20.263

15.471

10.678

5.901

1.634

0.075

0.001

1

20.224

15.432

10.639

5.859

1.543

0.046

0

2

20.232

15.439

10.647

5.867

1.562

0.052

0

3

20.240

15.447

10.655

5.876

1.580

0.057

0

4

20.247

15.455

10.663

5.884

1.597

0.063

0

λr

pr

ηv

λv

pv

9.3 The Fong-Vasicek Model

169

Table 9.8. Values of short-rate caps for the jump-enhanced Fong-Vasicek model, with a nominal value of 100 units. K

6

7

8

9

10

11

12

0.005

2.854

1.976

1.222

0.663

0.317

0.137

0.055

0.01

3.742

2.852

2.061

1.422

0.952

0.624

0.400

0.015

4.623

3.730

2.925

2.253

1.721

1.306

0.981

0.02

5.494

4.601

3.790

3.101

2.536

2.069

1.678

0

1.979

1.162

0.541

0.181

0.04

0.006

0.001

2

2.854

1.976

1.222

0.663

0.317

0.137

0.055

4

3.738

2.827

1.991

1.291

0.769

0.423

0.217

6

4.625

3.696

2.812

2.018

1.358

0.858

0.510

1

2.412

1.554

0.850

0.375

0.131

0.038

0.009

2

2.854

1.976

1.222

0.663

0.317

0.137

0.055

3

3.298

2.410

1.627

1.011

0.583

0.317

0.163

4

3.743

2.849

2.049

1.394

0.906

0.565

0.339

0.005

2.854

1.976

1.222

0.663

0.317

0.137

0.055

0.01

2.861

1.990

1.242

0.684

0.334

0.148

0.061

0.015

2.868

2.003

1.260

0.704

0.351

0.159

0.067

0.02

2.876

2.016

1.278

0.723

0.366

0.170

0.073

0

2.848

1.962

1.200

0.639

0.300

0.127

0.050

2

2.854

1.976

1.222

0.663

0.317

0.137

0.055

4

2.860

1.990

1.243

0.685

0.335

0.148

0.060

6

2.867

2.004

1.263

0.707

0.352

0.158

0.065

1

2.851

1.969

1.211

0.651

0.308

0.132

0.052

2

2.854

1.976

1.222

0.663

0.317

0.137

0.055

3

2.857

1.983

1.232

0.674

0.326

0.142

0.058

4

2.861

1.990

1.242

0.684

0.334

0.148

0.060

ηr

λr

pr

ηv

λv

pv

170

9 Jump-Enhanced Two-Factor Interest-Rate Models

Table 9.9. Values of average-rate caps for the jump-enhanced Fong-Vasicek model, with a nominal value of 100 units. K

6

7

8

9

10

11

12

0.005

2.377

1.443

0.655

0.205

0.051

0.012

0.003

0.01

2.827

1.892

1.077

0.543

0.268

0.131

0.063

0.015

3.271

2.336

1.512

0.936

0.586

0.366

0.226

0.02

3.708

2.775

1.946

1.346

0.948

0.668

0.468

0

1.922

1.004

0.308

0.036

0.001

0

0

2

2.377

1.443

0.655

0.205

0.051

0.012

0.003

4

2.831

1.889

1.040

0.449

0.159

0.050

0.014

6

3.283

2.337

1.450

0.751

0.329

0.127

0.044

1

2.150

1.221

0.466

0.096

0.012

0.001

0

2

2.377

1.443

0.655

0.205

0.051

0.012

0.003

3

2.603

1.668

0.859

0.352

0.127

0.043

0.013

4

2.828

1.892

1.072

0.523

0.237

0.101

0.041

0.005

2.377

1.443

0.655

0.205

0.051

0.012

0.003

0.01

2.377

1.446

0.661

0.211

0.053

0.012

0.003

0.015

2.378

1.449

0.667

0.216

0.056

0.013

0.003

0.02

2.378

1.452

0.672

0.221

0.058

0.014

0.003

0

2.377

1.441

0.648

0.200

0.049

0.011

0.002

2

2.377

1.443

0.655

0.205

0.051

0.012

0.003

4

2.377

1.446

0.661

0.211

0.053

0.012

0.003

6

2.377

1.449

0.668

0.216

0.055

0.013

0.003

1

2.377

1.442

0.651

0.202

0.05

0.012

0.002

2

2.377

1.443

0.655

0.205

0.051

0.012

0.003

3

2.377

1.445

0.658

0.208

0.052

0.012

0.003

4

2.377

1.446

0.661

0.211

0.053

0.012

0.003

ηr

λr

pr

ηv

λv

pv

10 Non-Aﬃne Term-Structure Models and Short-Rate Models with Stochastic Jump Intensity

10.1 Overview Although the model setup proposed in this thesis is of the exponential-aﬃne class, we can also extend the framework to allow for certain non-aﬃne models and models with state-dependent jump intensities λQ (xt ). Moreover, option prices under these more sophisticated model dynamics can be priced in our numerical scheme without greater eﬀort, due to an exponential separable structure of the governing characteristic function. However, working with a non-aﬃne model, we have to abandon jump components for those particular non-aﬃne factors. A stochastic jump intensity in the general exponentialaﬃne model framework is introduced in Duﬃe, Pan and Singleton (2000). Consequently, the jump transform is no longer independent of the coeﬃcient function a(z, τ ), and therefore a complicated system of ODEs has to be determined numerically anyway. Since both approaches need to establish further restrictions, they are only discussed as possibilities for extending and modifying the base model, respectively.

10.2 Quadratic Gaussian Models Non-Aﬃne exponential separable models are characterized by a non-aﬃne structure of the factors in the relevant moment-generating function, as well as the general characteristic function, while preserving the separability of coeﬃcient functions for diﬀerent powers of the particular factors included in the model. Thus, the essential system of ODEs can be derived. Prominent representatives of this model class are in an equity context the stochastic volatility

172

10 Non-Aﬃne and Stochastic Jump Intensity Term-Structure Models

model of Sch¨obel and Zhu (1999), which is a generalized version of the Stein and Stein (1991) model. In case of interest rates, we have, e.g. the Double Square-Root model of Longstaﬀ (1989), the quadratic Gaussian model approach of Beaglehole and Tenney (1991)203 , and the general linear-quadratic jump-diﬀusion model of Cheng and Scaillet (2004)204 . Although the quadratic Gaussian and the Double Square-Root model seem quite attractive to implement, it is impossible to compute theoretical model prices within the Fourier-based pricing framework if jumps are incorporated, while Monte-Carlo pricing approaches might still work. This stems from the fact that in equation (2.39), for the nth jump Jmn in the non-aﬃne fac(m) (m) tor xt , there would be a corresponding term (xt + Jmn )2 resulting in a mixed expression. Hence, the exponential separation approach will no longer be available in deriving the general characteristic function. Since none of the non-aﬃne interest-rate models are capable of exhibiting any jump component we completely ignored these models in our base setup according to Section 2.1. The one-factor quadratic Gaussian approach models the short rate under the risk-neutral measure, as the square of some factor xt governed by an Ornstein-Uhlenbeck process according to equation (8.5). In order to price interest-rate derivatives for this particular process, we need to have the general characteristic function to consider both the state variable xt and its square x2t . Thus, for the squared Gaussian interest-rate model we use the following form of the general characteristic function

ψ (yt , z, 0, w, g0 , g, τ ) = ea(z,τ )+b(z,τ ) yt +ızg0 ,

with yt =

xt x2t

and

0 w= . 1

For convenience, we use again the time-dependent coeﬃcient functions205 203

204

205

Ahn, Dittmar and Gallant (2002) give a good overview of general multidimensional linear-quadratic Gaussian interest-rate models. Linear-quadratic in this context means all factors contained in the state vector xt are allowed to enter the interest rate both in a linear and quadratic fashion. ¯ The constant parameter g0 is represented by the term A.

10.2 Quadratic Gaussian Models

173

a(z, τ ) = A(z, τ ),

and b(z, τ ) =

B(z, τ )

C(z, τ )

¯ B . + ız C¯

Inserting the above characteristic function in equation (2.33) and applying the separation approach result again in a system of coupled ODEs206 ¯ + σ 2 (C(z, τ ) + ız C) ¯ A(z, τ )τ =κθ(B(z, τ ) + ız B) ¯ 2, + 2σ 2 (B(z, τ ) + ız B) ¯ 2 (C(z, τ ) + ız C) ¯ − κ) B(z, τ )τ =(B(z, τ ) + ız B)(σ ¯ + 2κθ(C(z, τ ) + ız C), ¯ + 2σ 2 (C(z, τ ) + ız C) ¯ 2 − 1, C(z, τ )τ = − 2κ(C(z, τ ) + ız C) which can be solved successively. The advantage of this modeling approach lies in its tractability while describing a more elaborated interest-rate behavior. Additionally, the short rate in this approach is always positive, compared to possible negative short rates using the Vasicek model. In the Double SquareRoot model according to Longstaﬀ (1989), we encounter a very similar situation, since we are able to transform the model into a quadratic Gaussian model and vice versa but with additional restrictions on the parameter set207 . Cheng and Scaillet (2004) introduce a linear-quadratic jump-diﬀusion model. Here, the diﬀusion part of some random variable, for example the short rate r(xt ) or the payoﬀ characteristic function g(xt ), is built similarly to the multivariate quadratic Gaussian model in Beaglehole and Tenney (1991), as the sum of linear and quadratic terms of the state vector xt containing correlated Ornstein-Uhlenbeck processes. To gain a closed-form solution for the general characteristic function, additional jump parts only occur in the aﬃne terms of xt . Therefore, we can think of this interest-rate model as a simple combination of an additive multivariate Ornstein-Uhlenbeck model augmented with jump components and an additive multivariate quadratic Gaussian model. 206

Although the vector yt occurs in the characteristic function, derivatives remain still to be taken with respect to the unique state variables which is in this one-

207

dimensional model just the factor xt . See Beaglehole and Tenney (1992), pp. 346-347.

174

10 Non-Aﬃne and Stochastic Jump Intensity Term-Structure Models

10.3 Stochastic Jump Intensity Another possibility for extending the base model setup stated in Section 2.1 is to implement stochastic jump intensities. Duﬃe, Pan and Singleton (2000) introduced, with their aﬃne jump-diﬀusion model, a vector of stochastic jump intensities where the stochastic component is aﬃne in the state variable xt . Thus, they implement stochastic intensities without overly aggravating their solution technique. Deﬁning the vector of jump intensities as208 Q λQ (xt ) = λQ 0 + λ1 xt , Q M with (λQ × RM×M , we therefore get a slightly modiﬁed system of 0 , λ1 ) ∈ R

ODEs for the vector coeﬃcient b(z, τ ) compared to equation (2.41), which is 1 b(z, τ ) Σ1 b(z, τ ) 2 ∗ + λQ 1 EJ [ψ (z, w0 , w, g0 , g, J, τ ) − 1] .

b(z, τ )τ = − w + µQ 1 b(z, τ ) +

Obviously, in implementing this type of jump intensity, values of the coeﬃcient vector b(z, τ ) must be determined numerically due to the complicated structure of the relevant ODE. Subsequently, the same statement holds also for the coeﬃcient function a(z, τ ), which depends on b(z, τ ). Although this type of jump speciﬁcation enriches the modeling capabilities of the shortrate dynamics, it is infrequently implemented in interest-rate models because of the numerical diﬃculties mentioned above. However, our FRFT algorithm presented in Chapter 6 can be easily modiﬁed to handle this type of stochastic jump intensity.

208

To stay conform with our base model setup in equation (2.1), we suggest to include N Poisson processes with stochastic intensities.

11 Conclusion

In this thesis, we have introduced a general jump-diﬀusion short-rate model. The model approach we proposed extends the interest-rate model of Duﬃe and Kan (1996) by considering N diﬀerent possible Poison processes in the underlying factors. Using the ﬁndings in Carr and Madan (1999) and Lewis (2001) to interchange the order of integration of an integral-transformed option price in an equity context, we derived a general pricing formula valid for various popular interest-rate contracts. However, we eventually preferred the approach of Lewis (2001) over the technique presented in Carr and Madan (1999). The pricing scheme used in this thesis exhibits a rigorous modular structure. Thus, we took one step towards successfully extending the spirit of a modular pricing framework as proposed in Zhu (2000) by modularizing not only the stochastic parts, but also modularizing the derivative price in terms of its payoﬀ structure. Hence, all pricing formulae developed in this thesis can be split into parts of the Fourier-transformed payoﬀ function and of the underlying process, characterized by its characteristic function, respectively. Hence, we were able to state one single valuation formula, equation (4.21), to price derivatives of the linear, exponential-linear, and integro-linear types. Especially for the integro-linear case, the payoﬀ-transform approach oﬀers an elegant alternative to the methods proposed, e.g. in Bakshi and Madan (2000), Chacko and Das (2002), and Ju (1997). In addition, we presented within the pricing framework of Lewis (2001) for the ﬁrst time the consistent inclusion of both unconditional and conditional interest-rate derivatives. Hence, facilitating a integration by parts method, we successfully applied the residue theorem to recover prices of contracts with unconditional exercise rights and

176

11 Conclusion

the particular put-call parities, respectively. As a special case in Section 5.3.3, we also derived a Fourier-style solution of coupon-bond options, where the interest rate is governed by a one-factor process, which can be compared to the two-sided Laplace-style solution presented in Kluge (2005), Section 2.4. However, we want to emphasize that the term one-factor does not correspond to the number of jump components incorporated in the short-rate process. Thus, theoretically speaking, within our pricing scheme we are able to price coupon-bond options and swaptions, respectively, as long as the underlying process exhibits only one single Brownian motion. This is in fact a powerful result, since the additional inclusion of jump processes can result in more realistic models. In Chapter 6, we employed the IFFT algorithm for the ﬁrst time to compute option prices within the pricing framework of Lewis (2001). The obtained pricing algorithm is then reﬁned by translating the IFFT procedure into the FRFT algorithm. Subsequently, we dealt with the issue of ﬁnding the optimal and therefore error-minimizing parameter setting for the FRFT algorithm by utilizing a steepest descent technique. Doing this, we focused our eﬀorts to minimize the overall error of the solution vector generated by the FRFT pricing algorithm, rather than the error of one single option price209 . In our opinion, this procedure is a more powerful procedure, since the advantage of the FFT- and IFFT-based pricing algorithms is the simultaneous computation of option prices for a given strike range. Therefore, we used for the error measuring the RMSE of the numerical solution vector. Fortunately, it became apparent that the logarithmic RMSE of the numerical solution is a nearly linear descending function for increasing values of zi and ω, starting with the smallest possible values not violating any regularity conditions. Thus, we used a steepest-descent technique to identify the optimal parameters for the numerical algorithm. Furthermore, exploiting this linearity we were also able to formulate an approximate error bound for the numerical solution vector. After discussing the numerical algorithm, we analyzed a selection of both one-factor and two-factor jump-diﬀusion short-rate models. We ﬁrst speciﬁed our jump size candidates, which were the exponential, gamma, and normal 209

Lee (2004) and Lord and Kahl (2007) study the error behavior of Fourier transform-based algorithms for only a single strike value.

11 Conclusion

177

distribution, and derived their particular jump transforms. Subsequently, we derived the relevant general characteristic function of the jump-diﬀusion process, and then computed numerical values of the particular density functions and contract values. Widely used, the exponentially distributed jump size assumption presents no diﬃculties in derivatives pricing because of the closedform jump transforms for both Ornstein-Uhlenbeck and Square-Root diﬀusion processes. However, we also applied normal and gamma distributions for the jump size. The normal distribution for the jump component within a Vasicek model is also used in the articles of Baz and Das (1996) and Durham (2005), where an approximation technique for zero-bond prices is described. Unfortunately, under some circumstances both approaches deliver inaccurate values for the respective derivatives contracts210. Our pricing algorithm is able to circumvent these issues and compute accurate numerical values of interest-rate derivatives in any case. Moreover, we introduced gamma distributed jumps within a jump-diﬀusion short-rate model framework for the ﬁrst time. We then combined these jump candidates with the one-factor Ornstein-Uhlenbeck, the Square-Root processes, the two-factor version of a combined OU-SR, and the stochastic volatility model of Fong and Vasicek (1991a), and computed densities and option values. In particular, our contribution besides the implementation of the normal and gamma jump-size distribution in interest-rate option pricing, is to present an algorithm capable of computing option prices for the (jump-extended) Fong and Vasicek (1991a) interest-rate model. Up to now, only zero-bond prices have been computed for the jump-enhanced Vasicek and CIR model211 and the (pure diﬀusion) Fong and Vasicek (1991a) model212 , but no option prices have been presented so far. Due to the general applicability of the solution formula (4.21), we were able to compute numerical solutions for all important interest-rate derivatives. Comparing the diﬀerent results from the jump-diﬀusion term-structure models, it is obvious that jump components can enhance the stochastic dynamics. Accordingly, we were able to model probability density functions, which show bimodality and the important feature of fat tails. 210 211 212

See the concluding remarks in Durham (2006) and the comments in Section 7.3. Compare with the comments in Sections 8.2 and 8.3. See, for example, Selby and Strickland (1995).

178

11 Conclusion

Although the model setup used in this thesis is of the exponential-aﬃne type, the pricing technique can be extended to special non-aﬃne processes, namely to the family of quadratic Gaussian processes due to their exponentialaﬃne structure of the particular characteristic function. As discussed in Chapter 10, this model class cannot be enhanced with jump components since the resulting PDE would then no longer be separable. Another possible way of extending the base model speciﬁcation, which we brieﬂy discussed, is given by the inclusion of stochastic jump intensities, where the intensity is an aﬃne function of the state vector xt . However, we have then to numerically determine the coeﬃcient functions a(z, τ ) and b(z, τ ), which can be a challenging task due to the elaborated jump transforms. We presented a sophisticated alternative to time-consuming Monte-Carlo simulations, which have to be applied otherwise due to the complicated jumpdiﬀusion dynamics. Combined with the highly eﬃcient FRFT algorithm, this numerical pricing approach oﬀers an accuracy and eﬃciency, which can be hardly achieved by other methods. However, the methodology is restricted, in this form, to price only European-type derivatives. Thus, possible research can focus on developing a pricing procedure based on the algorithm in this thesis, which is also capable of valuing American-type derivatives. The early-exercise feature of these American-type derivatives might then be implemented by using some sort of time-stepping scheme of the Fourier-transformed derivative value or by using backward induction as proposed in Lord, Fang, Bervoets and Oosterlee (2007). Although we discussed one- and two-factor interestrate models, we can easily extend the pricing framework to include also jumpenhanced versions of higher factor models, such as e.g. the multi-factor models presented in Balduzzi, Das, Foresi and Sundaram (1996) and Collin-Dufresne and Goldstein (2002). Another possibility for further research might be an empirical validation of the family of gamma jump-enhanced diﬀusion models, as for example done in the studies by Lin and Yeh (1999) and Das (2002).

A Derivation of the Complex-Valued Coeﬃcients for the Characteristic Function in the Square-Root Model

Our starting point for deriving the time-dependent coeﬃcient function ˜b(z, τ ) is equation (8.10). Thus, making the standard transformation for this type of diﬀerential equation, we assume ˜b(z, τ ) = − 1 E(z, τ )τ . c2 (z) E(z, τ )

(A.1)

Consequently, substituting the particular expressions in equation (8.10), function E(z, τ ) satisﬁes the following homogeneous ODE E(z, τ )τ τ = c1 (z)E(z, τ )τ − c0 (z)c2 (z)E(z, τ ),

(A.2)

with E(z, 0)τ = 0, due to the terminal condition ˜b(z, 0) = 0. Additionally, we assume for the moment an unspeciﬁed constant E(z, 0) = E0 and guess a solution of the form E(z, τ ) = eυ(z)τ . Hence, plugging this function together with its particular derivatives into equation (A.2), we arrive at the so-called characteristic equation for this second order type ODE, which after some simpliﬁcations is υ 2 (z) − c1 (z)υ(z) + c0 (z)c2 (z) = 0. The solution of this quadratic form is given by υ± (z) =

c1 (z) ± ϑ(z) , 2

with ϑ(z) deﬁned according to Section 8.3. Since the discriminant of the square-root function ϑ(z) is

180

A Complex-Valued Coeﬃcients in the Square-Root Model

κ2 + 2σ 2 w1 > 0, the characteristic equation has two diﬀerent real-valued solutions and therefore the general solution can be represented by the linear combination E(z, τ ) = Ψ1 (z)eυ(z)+ τ + Ψ2 (z)eυ(z)− τ . Consequently, we get for τ = 0 the following terminal conditions E(z, 0) = Ψ1 (z) + Ψ2 (z), E(z, 0)τ = Ψ1 (z)υ+ (z) + Ψ2 (z)υ− (z). Keeping in mind that E(z, 0)τ ≡ 0, we use the two equations above to determine the coeﬃcient functions Ψ1 (z) and Ψ2 (z). Eventually, the solution of E(z, τ ) can be obtained as c1 (z) ϑ(z) ϑ(z) ϑ(z) ϑ(z) E0 e 2 τ e 2 τ + e− 2 τ e 2 τ − e− 2 τ E(z, τ ) = ϑ(z) − c1 (z) ϑ(z) 2 2 (A.3) c1 (z)

ϑ(z)τ E0 e 2 τ ϑ(z)τ = ϑ(z) cosh − c1 (z) sinh , ϑ(z) 2 2 and the particular derivative with respect to the time-to-maturity variable τ can be calculated as c1 (z)

E0 e 2 E(z, τ )τ = −2 ϑ(z)

τ

ϑ(z)τ c0 (z)c2 (z) sinh . 2

Finally, inserting the functions E(z, τ ), now up to a constant E0 determined, and E(z, τ )τ into equation (A.1), we end up with 2c0 (z) sinh ϑ(z)τ 2 ˜b(z, τ ) = , ϑ(z)τ − c ϑ(z) cosh ϑ(z)τ (z) sinh 1 2 2 which coincides with the solution given in equation (8.11)1 . Having obtained ˜b(z, τ ), it is a very simple task to derive the coeﬃcient function a0 (z, τ ) because of the approach taken in (A.1). Thus, using a logarithmic integration approach we immediately arrive at 1

The terminal condition ˜b(z, 0) = 0 is satisﬁed, which can be easily justiﬁed due to the relation sinh[0] = 0.

A Complex-Valued Coeﬃcients in the Square-Root Model

τ a0 (z, τ ) = −w0 τ + κθ

181

˜b(z, s) + ızg0 ds

0

κθ = (ızg0 κθ − w0 )τ − c2 (z)

τ

E(z, s)τ ds E(z, s)

(A.4)

0

κθ (ln [E(z, τ )] − ln [E(z, 0)]) . = (ızg0 κθ − w0 )τ − c2 (z) Because of the terminal condition a0 (z, 0) ≡ 0, we must set the constant E0 = 1 in equation (A.3). Eventually, after simplifying the resulting expression in equation (A.4) we are able to state the desired form given in (8.12).

B Derivation of the Complex-Valued Coeﬃcients for the Characteristic Function in the Fong-Vasicek Model

Starting with the time-dependent coeﬃcient function B(z, τ ), we adopt the solution according to equation (8.6). Thus, we exchange the parameter g1 with ¯ Subsequently, we show that the derivation of the time-dependent coeﬃcient B. A01 (z, τ ), the volatility-related part of A0 (z, τ ), states no problem on account of logarithmic integration. Thus, the next task is to recover the coeﬃcient function C(z, τ ). Therefore, plugging in the explicit solution of B(z, τ ) into ODE (9.7) results in 1 C(z, τ )τ =f1 (z) + f2 (z)X(z, τ ) + X(z, τ )2 + f3 (z)C(z, τ ) 2 β2 + βρX(z, τ )C(z, τ ) + C(z, τ )2 , 2

(B.1)

with time-independent coeﬃcients fi (z) according to Section 9.3 and w ¯ e−κτ . X(z, τ ) = σ + ız B κ Similar to the derivation of the time-dependent coeﬃcient function ˜b(z, τ ) in the SR model, we assume for C(z, τ ) a solution of the form C(z, τ ) = −

2 U (z, τ )τ . β 2 U (z, τ )

(B.2)

Inserting this alternative representation into equation (B.1) and simplifying the resulting ODE for the new function U (z, τ ) gives U (z, τ )τ τ = (f3 (z) + βρX(z, τ )) U (z, τ )τ β2 1 2 − f1 (z) + f2 (z)X(z, τ ) + X(z, τ ) U (z, τ ). 2 2

(B.3)

184

B Complex-Valued Coeﬃcients in the Fong-Vasicek Model

Subsequently, we apply another substitution V (X(z, τ )) = U (z, τ ),

(B.4)

and get a new ODE, with derivatives taken with respect to X(z, τ ). For convenience, we express the particular derivatives with V (X(z, τ ))X and V (X(z, τ ))XX , respectively. Thus, the resulting ODE has the formal structure X(z, τ )2 V (X(z, τ ))XX f3 (z) βρ X(z, τ )2 V (X(z, τ ))X + 1+ X(z, τ ) + κ κ 2 β 1 + 2 f1 (z) + f2 (z)X(z, τ ) + X(z, τ )2 V (X(z, τ )) = 0. 2κ 2

(B.5)

Finally, the solution of this particular ODE can be obtained by applying a last substitution of the form V (X(z, τ )) = L(z, τ ) W (Y (z, τ )), with L(z, τ ) and Y (z, τ ) as deﬁned in Section 9.3. Hence, inserting this substitution into equation (B.5) and simplifying the resulting ODE, we end up with − Q(z)W (Y (z, τ )) + (S(z) − Y (z, τ ))W (Y (z, τ ))Y

(B.6)

+ Y (z, τ )W (Y (z, τ ))Y Y = 0. Again, the explicit expressions of Q(z) and S(z) are given in Section 9.3. Equation (B.6) is better known as the prominent Kummer equation, which has the general solution2 W (Y (z, τ )) = Ψ1 (z)KM[Q(z), S(z), Y (z, τ )] + Ψ2 (z)KU[Q(z), S(z), Y (z, τ )]. Thus, in order to obtain the solution for the coeﬃcient C(z, τ ), we also need the ﬁrst derivative with respect to τ of the function U (z, τ ). Hence, according to the chain rule we have the relation U (z, τ )τ = −κX(z, τ )V (X(z, τ ))X . 2

See, for example, Abramowitz and Stegun (1972), p. 504. Our solution is customized to account for the parametric form due to the frequency representation.

B Complex-Valued Coeﬃcients in the Fong-Vasicek Model

185

The desired derivative of V (X(z, τ )) with respect to X(z, τ ) in the above equation can be represented as V (X(z, τ ))X = 1 − L(z, τ ) 2X(z, τ ) × −2Q(z) × Ψ1 (z)KM[Q(z) + 1, S(z), Y (z, τ )] + (1 + Q(z) − S(z))Ψ2 (z)KU[Q(z) + 1, S(z), Y (z, τ )] +

β2 M (z, τ ) κ

× Ψ1 (z)KM[Q(z), S(z), Y (z, τ )]

+ Ψ2 (z)KU[Q(z), S(z), Y (z, τ )] . Thus, according to the approach taken in equation (B.2), the coeﬃcient function C(z, τ ) can be recovered as (9.8), which is in terms of V (X(z, τ )) C(z, τ ) =

2κ V (X(z, τ ))X . X(z, τ ) β2 V (X(z, τ ))

Next, checking the validity of the terminal condition C(z, 0) = U (z, 0)τ = V (X(z, 0))X ≡ 0, we only need the explicit form of the time-independent function Υ (z), which is just the fraction Ψ1 (z) . Υ (z) = Ψ2 (z) Arranging terms for Ψ1 (z) and Ψ2 (z) in the ﬁrst derivative of V (X(z, τ )) ¯ evaluated at X(z, 0) = σ w κ + ız B , we get

186

B Complex-Valued Coeﬃcients in the Fong-Vasicek Model

2 β M (z, τ )KM[Q(z) + 1, S(z), Y (z, 0)] Ψ1 (z) κ − 2Q(z)KM[Q(z) + 1, S(z), Y (z, 0)] = Ψ2 (z) 2Q(z)(1 + Q(z) − S(z))KU[Q(z), S(z), Y (z, 0)] β2 − M (z, τ )KU[Q(z), S(z), Y (z, 0)] . κ Obviously, solving for the particular fraction, the speciﬁc form of Υ (z) can be validated by checking its deﬁnition given in Section 9.3. Thus, the coeﬃcient function C(z, τ ) with speciﬁed time-independent function Υ (z) coincides with the result given in equation (9.8). For the calculation of A02 (z, τ ), we exploit the functional form chosen in the derivation of the coeﬃcient function C(z, τ ). Thus, we apply a logarithmic integration approach and recover the antiderivative of A02 (z, τ )τ as A02 (z, τ ) = −

2α¯ v ¯ ln[U (z, τ )] + ızα¯ vCτ. 2 β

In order to guarantee the terminal condition of A02 (z, 0) = 0, at the maturity of the contract, we have to ensure that U (z, 0) = 1. Thus, rewriting U (z, τ ) as U (z, τ ) =L(z, τ )Ψ2 (z) × (Υ (z)KM[Q(z), S(z), Y (z, τ )] + KM[Q(z), S(z), Y (z, τ )]) , we immediately arrive at 1 = L(z, 0) (Υ (z)KM[Q(z), S(z), Y (z, 0)] + KM[Q(z), S(z), Y (z, 0)]) . Ψ2 (z) Therefore, the time-dependent function A02 (z, τ ) can be written in terms of J(z, τ ) =

2Q(z)κ β 2 (KU[Q(z); S(z); Y (z, τ )] + Υ(z) KM[Q(z); S(z); Y (z, τ )])

and L(z, τ ), given in (9.11), which concludes the derivation of the coeﬃcient functions in Section 9.3.

References

Abramowitz, M. and Stegun, I.: 1972, Handbook of Mathematical Functions with Formulas, Graphs and Mathematical Tables, Dover Publications. Ahn, C. and Thompson, H.: 1988, Jump-Diﬀusion Processes and the Term Structure of Interest Rates, Journal of Finance 43(1), 155–174. Ahn, D., Dittmar, R. and Gallant, A.: 2002, Quadratic Term Structure Models: Theory and Evidence, The Review of Financial Studies 15(1), 243–288. Ahn, D. and Gao, B.: 1999, A Parametric Nonlinear Model of Term Structure Dynamics, The Review of Financial Studies 12(4), 721–762. A¨ıt-Sahalia, Y.: 1996, Testing Continuous-Time Models of the Spot Interest, The Review of Financial Studies 9(2), 385–426. A¨ıt-Sahalia, Y.: 1999, Transition Densities for Interest Rate and Other Nonlinear Diﬀusions, Journal of Finance 54(4), 1361–1395. Arapis, M. and Gao, J.: 2006, Empirical Comparisons in Short-Term Interest Rate Models Using Nonparametric Methods, Journal of Financial Econometrics 4(2), 310–345. Bailey, D. and Swarztrauber, P.: 1994, A Fast Method for the Numerical Evaluation of Continuous Fourier and Laplace Transforms, Journal on Scientiﬁc Computing 15(5), 1105–1110. Bakshi, G. and Madan, D.: 2000, Spanning and Derivative-Security Valuation, Journal of Financial Economics 55(2), 205–238. Balakrishnan, N., Johnson, N. and Kotz, S.: 1994, Continuous Univariate Distributions, Vol. 1, 2nd edn, John Wiley and Sons. Balduzzi, P., Das, S., Foresi, S. and Sundaram, R.: 1996, A Simple Approach to Three-Factor Aﬃne Term Structure Models, Journal of Fixed Income

188

References

5(4), 43–53. Bates, D.: 1996, Jumps and Stochastic Volatility: Exchange Rate Processes Implicit in Deutsche Mark Options, The Review of Financial Studies 9(1), 69–107. Baz, J. and Das, S.: 1996, Analytical Approximation of the Term Structure for Jump-Diﬀusion Processes: A Numerical Analysis, Journal of Fixed Income 6(2), 78–86. Beaglehole, D. and Tenney, M.: 1991, General Solutions of some Interest RateContingent Claim Pricing Equations, Journal of Fixed Income 1(1), 69–83. Beaglehole, D. and Tenney, M.: 1992, Corrections and Additions to ’A Nonlinear Equilibrium Model of the Term Structure of Interest Rates’, Journal of Financial Economics 32(3), 345–353. Black, F. and Scholes, M.: 1973, The Pricing of Options and Corporate Liabilities, Journal of Political Economy 81(3), 637–659. Brennan, M. and Schwartz, E.: 1979, A Continuous Time Approach to the Pricing of Bonds, Journal of Banking and Finance 3(2), 133–155. Brennan, M. and Schwartz, E.: 1980, Conditional Predictions of Bond Prices and Returns, The Journal of Finance 35(2), 405–419. Brennan, M. and Schwartz, E.: 1982, An Equilibrium Model of Bond Pricing and a Test of Market Eﬃciency, Journal of Financial and Quantitative Analysis 17(3), 301–329. Brigo, D. and Mercurio, F.: 2001, Interest Rate Models, Springer. Briys, E., Crouhy, M. and Sch¨obel, R.: 1991, The Pricing of Default-Free Interest Rate Cap, Floor, and Collar Agreements, Journal of Finance 46(5), 1879–1892. Brown, S. and Dybvig, P.: 1986, The empirical Implications of the Cox, Ingersoll,Ross Theory of the Term Structure of Interest Rates, Journal of Finance 41(3), 617–630. Carr, P. and Madan, D.: 1999, Option Valuation Using the Fast Fourier Transform, Journal of Computational Finance 2(4), 61–73. Carr, P. and Wu, L.: 2004, Time-Changed Levy Processes and Option Pricing, Journal of Financial Economics 71(1), 113–141. Chacko, G. and Das, S.: 2002, Pricing Interest Rate Derivatives: A General Approach, The Review of Financial Studies 15(1), 195–241. Chan, K.: 2005, Modelling Conditional Heteroskedasticity and Jumps in Australian Short-Term Interest Rates, Accounting and Finance 45(4), 537–551.

References

189

Chen, R. and Scott, L.: 1992, Pricing Interest Rate Options in a Two-Factor Cox-Ingersoll-Ross Model of the Term Structure, The Review of Financial Studies 5(4), 613–636. Chen, R. and Scott, L.: 1995, Interest Rate Options in Multifactor CoxIngersoll-Ross Models of the Term Structure, Journal of Derivatives 3, 53– 72. Cheng, P. and Scaillet, O.: 2004, Linear-Quadratic Jump-Diﬀusion Modeling with Appliccation to Stochastic Volatility. FAME Research Paper. Chourdakis, K.: 2005, Option Pricing Using the Fractional FFT, Journal of Computational Finance 8(2), 7–24. Collin-Dufresne, P. and Goldstein, R.: 2002, Pricing Swaptions Within an Aﬃne Framework, Journal of Derivatives 10(1), 1–18. Cont, R. and Tankov, P.: 2004, Financial Modelling with Jump Processes, Chapman and Hall. Cooley, J. and Tukey, J.: 1965, An Algorithm for the Machine Computation of the Complex Fourier Series, Mathematics of Computation v19n91, 297– 301. Cox, J., Ingersoll, J. and Ross, S.: 1985a, An Intertemporal General Equilibrium model of Asset Prices, Econometrica 53(2), 363–384. Cox, J., Ingersoll, J. and Ross, S.: 1985b, A Theory of the Term Structure of Interest Rates, Econometrica 53(2), 385–408. Culot, M.: 2003, An Integrated Aﬃne Jump Diﬀusion Framework to Manage Power Portfolios in a Deregulated Market, PhD thesis, Universit´e Catholique de Louvain. Dai, Q. and Singleton, K.: 2003, Term Structure Dynamics in Theory and Reality, The Review of Financial Studies 16(3), 631–678. Das, S.: 2002, The Surprise Element: Jumps in Interest Rates, Journal of Econometrics 106(1), 27–65. Das, S. and Foresi, S.: 1996, Exact Solutions for Bond and Option Prices with Systematic Jump Risk, Review of Derivatives Research 1(1), 7–24. Delbaen, F. and Schachermayer, W.: 1995, The Existence of Absolutely Continuous Local Martingale Measures, Annals of Applied Probability 5(4), 926–945. Doetsch, G.: 1967, Anleitung zum praktischen Gebrauch der LaplaceTransformation und der Z-Transformation, 3rd edn, Oldenbourg.

190

References

Duﬃe, D.: 2001, Dynamic Asset Pricing Theory, 3rd edn, Princeton University Press. Duﬃe, D. and Kan, R.: 1996, A Yield-Factor Model of Interest Rates, Mathematical Finance 6(4), 379–406. Duﬃe, D., Pan, J. and Singleton, K.: 2000, Transform Analysis and Option Pricing for Aﬃne Jump-Diﬀusions, Econometrica 68(6), 1343–1376. Duﬀy, D.: 2004, Transform Methods for Solving Partial Diﬀerential Equations, 2nd edn, Springer. Durham, B.: 2005, Jump-Diﬀusion Processes and Aﬃne Term Structure Models: Additional Closed-Form Approximate Solutions, Distributional Assumptions for Jumps, and Parameter Estimates. FEDS Working Paper. Durham, B.: 2006, Additional Analytical Approximations of the Term Structure and Distributional Assumptions for Jump-Diﬀusion Processes, Journal of Fixed Income 15(1), 61–73. Feller, W.: 1951, Two Singular Diﬀusion Problems, Annals of Mathematics 54(1), 173–182. Filipovic, D.: 2001, A General Characterization of one Factor Aﬃne Term Structure Models, Finance and Stochastics 5(3), 389–412. Fong, H. and Vasicek, O.: 1991a, Interest-Rate Volatility as a Stochastic Factor. Giﬀord Fong Associates. Fong, H. and Vasicek, O.: 1991b, Fixed-Income Volatility Management, Journal of Portfolio Management 17(4), 41–46. Geman, H., Karoui, N. and Rochet, J.: 1995, Changes of Num´eraire, Changes of Probability Measure and Option Pricing, Journal of Applied Probability 32(2), 443–458. Gil-Pelaez, J.: 1951, A Note on the Inversion Theorem, Biometrika 38(4), 481–482. Hamilton, J.: 1996, The Daily Market for Federal Funds, Journal of Political Economy 104(1), 26–56. Harrison, J. and Kreps, D.: 1979, Martingales and Arbitrage in Multiperiod Security Markets, Journal of Economic Theory 20(3), 381–408. Harrison, J. and Pliska, S.: 1981, Martingales and Stochastic Integrals in the Theory of Continuous Trading, Stochastic Processes and their Applications 11(3), 215–260.

References

191

Heath, D., Jarrow, R. and Morton, A.: 1992, Bond Pricing and the Term Structure of Interest Rates: A New Methodology for Contingent Claim Valuation, Econometrica 60(1), 77–105. Heston, S.: 1993, A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options, The Review of Financial Studies 6(2), 327–343. Heston, S.: 1995, A Model of Discontinuous Interest Rate Behavior, Yield Curves and Volatility. Working Paper, University of Maryland. Itkin, A.: 2005, Pricing Options with VG model Using FFT. arXiv:physics/0503137.

eprint

Jamshidian, F.: 1989, An Exact Bond Option Formula, Journal of Finance 44(1), 205–209. Ju, N.: 1997, Fourier Transformation, Martingale, and the Pricing of AverageRate Derivatives. Working Paper, University of Maryland. Karatzas, I. and Shreve, E.: 1991, Brownian Motion and Stochastic Calculus, 2nd edn, Springer. Karlin, S. and Taylor, H.: 1981, A Second Course in Stochastic Processes, Academic Press. Kellerhals, B.: 2001, Financial Pricing Models in Continous Time and Kalman Filtering, Springer. Kispert, W.: 2005, Financial Contracts on Electricity in the Nordic Power Market, PhD thesis, University of T¨ ubingen. Kluge, W.: 2005, Time-Inhomogeneous L´evy Processes in Interest Rate and Credit Risk Models), PhD thesis, University of Freiburg. Kou, S.: 2002, A Jump-Diﬀusion Model for Option Pricing, Management Science 48(8), 1086–1101. Kou, S. and Wang, H.: 2004, Option Pricing Under a Double Exponential Jump Diﬀusion Model, Management Science 50(9), 1178–1192. Kushner, H.: 1967, Stochastic Stability and Control, Academic Press. Langetieg, T.: 1980, A Multivariate Model of the Term Structure, The Journal of Finance 35(1), 71–97. Lee, R.: 2004, Option Pricing by Transform Methods: Extensions, Uniﬁcation, and Error Control, Journal of Computational Finance 7(3), 51–86. Lewis, A.: 2000, Option Valuation under Stochastic Volatility, Finance Press.

192

References

Lewis, A.: 2001, A Simple Option Formula for General Jump-Diﬀusion and other Exponential Levy Processes. Manuscript, Envision Financial Systems and OptionCity.Net. Lin, B. and Yeh, S.: 1999, Jump-Diﬀusion Interest Rate Process: An Empirical Examination, Journal of Business Finance and Accounting 26(7+8), 967– 995. Longstaﬀ, F.: 1989, A Nonlinear General Equilibrium Model of the Term Structure of Interest Rates, Journal of Financial Economics 23(2), 195– 224. Longstaﬀ, F. and Schwartz, E.: 1992, Interest Rate Volatility and the Term Structure: A Two-Factor General Equilibrium Model, Journal of Finance 47(4), 1259–1282. Lord, R., Fang, F., Bervoets, F. and Oosterlee, C.: 2007, A Fast and Accurate FFT-Based Method for Pricing Early-Exercise Options under L´evy Processes. MPRA Paper. Lord, R. and Kahl, C.: 2007, Optimal Fourier Inversion in Semi-Analytical Option Pricing, Journal of Computational Finance 10(4), 1–30. Merton, R.: 1976, Option Pricing When Underlying Stock Returns are Discontinuous, Journal of Financial Economics 3(2), 125–144. Musiela, M. and Rutkowski, M.: 2005, Martingale Methods in Financial Modelling, 2nd edn, Springer. Nagel, H.: 2001, Optionsbewertung bei stochastischer Volatilit¨ at, Gabler. Øksendal, B.: 2003, Stochastic Diﬀerential Equations, 6th edn, Springer. Ornstein, L. and Uhlenbeck, G.: 1930, On the Theory of Brownian Motion, Phys. Rev. 36, 823–841. Pennacchi, G.: 1991, Identifying the Dynamics of Real Interest Rates and Inﬂation: Evidence Using Survey Data, The Review of Financial Studies 4(1), 53–86. Piazzesi, M.: 2003, Aﬃne Term Structure Models. unpublished manuscript. Raible, S.: 2000, L´evy Processes in Finance: Theory, Numerics, and Empirical Facts, PhD thesis, University of Freiburg. Sch¨ obel, R.: 1987, Zur Theorie der Rentenoption, Duncker & Humblot. Sch¨ obel, R.: 1995, Kapitalmarkt und zeitkontinuierliche Bewertung, PhysicaVerlag. Sch¨ obel, R. and Zhu, J.: 1999, Stochastic Volatility with an OrnsteinUhlenbeck Process: An Extension, European Finance Review 3(1), 23–46.

References

193

Sch¨ obel, R. and Zhu, J.: 2000, A Fourier Analysis Approach to the Valuation of Interest Rate Derivatives. Working Paper, University of T¨ ubingen. Selby, M. and Strickland, C.: 1995, Computing the Fong and Vasicek Pure Discount Bond Price Formula, Journal of Fixed Income 5(3), 78–84. Singleton, K. and Umantsev, L.: 2002, Pricing Coupon-Bond Options and Swaptions in Aﬃne Term Structure Models, Mathematical Finance 12(4), 427–446. Stein, E. and Stein, J.: 1991, Stock Price Distributions with Stochastic Volatiliy: An Analytic Approach, The Review of Pinancial Studies 4(4), 727–752. Stuart, A. and Ord, J.: 1994, Kendall’s Advanced Theory of Statistics Vol. I, 6th edn, A. Arnold. Tahani, N.: 2004, Valuing Credit Derivatives Using Gaussian Quadrature: A Stochastic Volatility Framework, Journal of Futures Markets 24(1), 3–35. Vasicek, O.: 1977, An Equilibrium Characterization of the Term Structure, Journal of Financial Economics 5(2), 177–188. Wilkens, C.: 2005, Are Jumps Inherent in European Interest Rate Data? Estimation using the Characteristic Function and Transform Analysis for Aﬃne Diﬀusions. Working Paper, Stockholm University. Zhou, H.: 2001, Jump-Diﬀusion Term Structure and Itˆ o Conditional Moment Generator. FEDS Working Paper. Zhu, J.: 2000, Modular Pricing of Options, Springer.

Lecture Notes in Economics and Mathematical Systems For information about Vols. 1–522 please contact your bookseller or Springer-Verlag Vol. 523: W.-B. Zhang, A Theory of Interregional Dynamics. XI, 231 pages. 2003.

Vol. 545: R. Hafner, Stochastic Implied Volatility. XI, 229 pages. 2004.

Vol. 524: M. Frölich, Programme Evaluation and Treatment Choise. VIII, 191 pages. 2003.

Vol. 546: D. Quadt, Lot-Sizing and Scheduling for Flexible Flow Lines. XVIII, 227 pages. 2004.

Vol. 525: S. Spinler, Capacity Reservation for CapitalIntensive Technologies. XVI, 139 pages. 2003.

Vol. 547: M. Wildi, Signal Extraction. XI, 279 pages. 2005.

Vol. 526: C. F. Daganzo, A Theory of Supply Chains. VIII, 123 pages. 2003. Vol. 527: C. E. Metz, Information Dissemination in Currency Crises. XI, 231 pages. 2003.

Vol. 548: D. Kuhn, Generalized Bounds for Convex Multistage Stochastic Programs. XI, 190 pages. 2005. Vol. 549: G. N. Krieg, Kanban-Controlled Manufacturing Systems. IX, 236 pages. 2005.

Vol. 528: R. Stolletz, Performance Analysis and Optimization of Inbound Call Centers. X, 219 pages. 2003.

Vol. 550: T. Lux, S. Reitz, E. Samanidou, Nonlinear Dynamics and Heterogeneous Interacting Agents. XIII, 327 pages. 2005.

Vol. 529: W. Krabs, S. W. Pickl, Analysis, Controllability and Optimization of Time-Discrete Systems and Dynamical Games. XII, 187 pages. 2003.

Vol. 551: J. Leskow, M. Puchet Anyul, L. F. Punzo, New Tools of Economic Dynamics. XIX, 392 pages. 2005.

Vol. 530: R. Wapler, Unemployment, Market Structure and Growth. XXVII, 207 pages. 2003.

Vol. 552: C. Suerie, Time Continuity in Discrete Time Models. XVIII, 229 pages. 2005.

Vol. 531: M. Gallegati, A. Kirman, M. Marsili (Eds.), The Complex Dynamics of Economic Interaction. XV, 402 pages, 2004.

Vol. 553: B. Mönch, Strategic Trading in Illiquid Markets. XIII, 116 pages. 2005.

Vol. 532: K. Marti, Y. Ermoliev, G. Pflug (Eds.), Dynamic Stochastic Optimization. VIII, 336 pages. 2004. Vol. 533: G. Dudek, Collaborative Planning in Supply Chains. X, 234 pages. 2004. Vol. 534: M. Runkel, Environmental and Resource Policy for Consumer Durables. X, 197 pages. 2004. Vol. 535: X. Gandibleux, M. Sevaux, K. Sörensen, V. T’kindt (Eds.), Metaheuristics for Multiobjective Optimisation. IX, 249 pages. 2004. Vol. 536: R. Brüggemann, Model Reduction Methods for Vector Autoregressive Processes. X, 218 pages. 2004. Vol. 537: A. Esser, Pricing in (In)Complete Markets. XI, 122 pages, 2004. Vol. 538: S. Kokot, The Econometrics of Sequential Trade Models. XI, 193 pages. 2004. Vol. 539: N. Hautsch, Modelling Irregularly Spaced Financial Data. XII, 291 pages. 2004. Vol. 540: H. Kraft, Optimal Portfolios with Stochastic Interest Rates and Defaultable Assets. X, 173 pages. 2004. Vol. 541: G.-y. Chen, X. Huang, X. Yang, Vector Optimization. X, 306 pages. 2005. Vol. 542: J. Lingens, Union Wage Bargaining and Economic Growth. XIII, 199 pages. 2004.

Vol. 554: R. Foellmi, Consumption Structure and Macroeconomics. IX, 152 pages. 2005. Vol. 555: J. Wenzelburger, Learning in Economic Systems with Expectations Feedback (planned) 2005. Vol. 556: R. Branzei, D. Dimitrov, S. Tijs, Models in Cooperative Game Theory. VIII, 135 pages. 2005. Vol. 557: S. Barbaro, Equity and Efficiency Considerations of Public Higer Education. XII, 128 pages. 2005. Vol. 558: M. Faliva, M. G. Zoia, Topics in Dynamic Model Analysis. X, 144 pages. 2005. Vol. 559: M. Schulmerich, Real Options Valuation. XVI, 357 pages. 2005. Vol. 560: A. von Schemde, Index and Stability in Bimatrix Games. X, 151 pages. 2005. Vol. 561: H. Bobzin, Principles of Network Economics. XX, 390 pages. 2006. Vol. 562: T. Langenberg, Standardization Expectations. IX, 132 pages. 2006.

and

Vol. 563: A. Seeger (Ed.), Recent Advances in Optimization. XI, 455 pages. 2006. Vol. 564: P. Mathieu, B. Beaufils, O. Brandouy (Eds.), Artificial Economics. XIII, 237 pages. 2005.

Vol. 543: C. Benkert, Default Risk in Bond and Credit Derivatives Markets. IX, 135 pages. 2004.

Vol. 565: W. Lemke, Term Structure Modeling and Estimation in a State Space Framework. IX, 224 pages. 2006.

Vol. 544: B. Fleischmann, A. Klose, Distribution Logistics. X, 284 pages. 2004.

Vol. 566: M. Genser, A Structural Framework for the Pricing of Corporate Securities. XIX, 176 pages. 2006.

Vol. 567: A. Namatame, T. Kaizouji, Y. Aruga (Eds.), The Complex Networks of Economic Interactions. XI, 343 pages. 2006.

Vol. 590: S. H. Oda, Developments on Experimental Economics: New Approaches to Solving Real-World Problems. XVI, 262 pages, 2007.

Vol. 568: M. Caliendo, Microeconometric Evaluation of Labour Market Policies. XVII, 258 pages. 2006.

Vol. 591: M. Lehmann-Waffenschmidt, Economic Evolution and Equilibrium: Bridging the Gap. VIII, 272 pages, 2007.

Vol. 569: L. Neubecker, Strategic Competition in Oligopolies with Fluctuating Demand. IX, 233 pages. 2006. Vol. 570: J. Woo, The Political Economy of Fiscal Policy. X, 169 pages. 2006. Vol. 571: T. Herwig, Market-Conform Valuation of Options. VIII, 104 pages. 2006.

Vol. 592: A. C.-L. Chian, Complex Systems Approach to Economic Dynamics. X, 95 pages, 2007. Vol. 593: J. Rubart, The Employment Effects of Technological Change: Heterogenous Labor, Wage Inequality and Unemployment. XII, 209 pages, 2007

Vol. 573: J. Emami Namini, International Trade and Multinational Activity. X, 159 pages, 2006.

Vol. 594: R. Hübner, Strategic Supply Chain Management in Process Industries: An Application to Specialty Chemicals Production Network Design. XII, 243 pages, 2007

Vol. 574: R. Kleber, Dynamic Inventory Management in Reverse Logistics. XII, 181 pages, 2006.

Vol. 595: H. Gimpel, Preferences in Negotiations: The Attachment Effect. XIV, 268 pages, 2007

Vol. 575: R. Hellermann, Capacity Options for Revenue Management. XV, 199 pages, 2006.

Vol. 596: M. Müller-Bungart, Revenue Management with Flexible Products: Models and Methods for the Broadcasting Industry. XXI, 297 pages, 2007

Vol. 572: M. F. Jäkel, Pensionomics. XII, 316 pages. 2006

Vol. 576: J. Zajac, Economics Dynamics, Information and Equilibnum. X, 284 pages, 2006. Vol. 577: K. Rudolph, Bargaining Power Effects in Financial Contracting. XVIII, 330 pages, 2006.

Vol. 597: C. Barz, Risk-Averse Capacity Control in Revenue Management. XIV, 163 pages, 2007

Vol. 578: J. Kühn, Optimal Risk-Return Trade-Offs of Commercial Banks. IX, 149 pages, 2006.

Vol. 598: A. Ule, Partner Choice and Cooperation in Networks: Theory and Experimental Evidence. Approx. 200 pages, 2007

Vol. 579: D. Sondermann, Introduction to Stochastic Calculus for Finance. X, 136 pages, 2006.

Vol. 599: A. Consiglio, Artificial Markets Modeling: Methods and Applications. XV, 277 pages, 2007

Vol. 580: S. Seifert, Posted Price Offers in Internet Auction Markets. IX, 186 pages, 2006.

Vol. 600: M. Hickman, P. Mirchandani, S. Voss (Eds.): Computer-Aided Scheduling of Public Transport. Approx. 424 pages, 2007

Vol. 581: K. Marti; Y. Ermoliev; M. Makowsk; G. Pflug (Eds.), Coping with Uncertainty. XIII, 330 pages, 2006. Vol. 582: J. Andritzky, Sovereign Default Risks Valuation: Implications of Debt Crises and Bond Restructurings. VIII, 251 pages, 2006. Vol. 583: I.V. Konnov, D.T. Luc, A.M. Rubinov† (Eds.), Generalized Convexity and Related Topics. IX, 469 pages, 2006. Vol. 584: C. Bruun, Adances in Artificial Economics: The Economy as a Complex Dynamic System. XVI, 296 pages, 2006. Vol. 585: R. Pope, J. Leitner, U. Leopold-Wildburger, The Knowledge Ahead Approach to Risk. XVI, 218 pages, 2007 (planned). Vol. 586: B.Lebreton, Strategic Closed-Loop Supply Chain Management. X, 150 pages, 2007 (planned). Vol. 587: P. N. Baecker, Real Options and Intellectual Property: Capital Budgeting Under Imperfect Patent Protection. X, 276 pages , 2007. Vol. 588: D. Grundel, R. Murphey, P. Panos , O. Prokopyev (Eds.), Cooperative Systems: Control and Optimization. IX, 401 pages , 2007. Vol. 589: M. Schwind, Dynamic Pricing and Automated Resource Allocation for Information Services: Reinforcement Learning and Combinatorial Auctions. XII, 293 pages , 2007.

Vol. 601: D. Radulescu, CGE Models and Capital Income Tax Reforms: The Case of a Dual Income Tax for Germany. XVI, 168 pages, 2007 Vol. 602: N. Ehrentreich, Agent-Based Modeling: The Santa Fe Institute Artificial Stock Market Model Revisited. XVI, 225 pages, 2007 Vol. 603: D. Briskorn, Sports Leagues Scheduling: Models, Combinatorial Properties, and Optimization Algorithms. XII, 164 pages, 2008 Vol. 604: D. Brown, F. Kubler, Computational Aspects of General Equilibrium Theory: Refutable Theories of Value. XII, 202 pages, 2008 Vol. 605: M. Puhle, Bond Portfolio Optimization. XIV, 137 pages, 2008 Vol. 606: S. von Widekind, Evolution of Non-Expected Utility Preferences. X, 130 pages, 2008 Vol. 607: M. Bouziane, Pricing Interest Rate Derivatives: A Fourier-Transform Based Approach. XII, 191 pages, 2008 Vol. 608: P. Nicola, Experimenting with Dynamic Macromodels: Growth and Cycles. XIII, 241 pages, 2008 Vol. 609: X. Fang, K.K. Lai, S. Wang, Fuzzy Portfolio Optimization: Theory and Models. IX, 173 pages, 2008 Vol. 610: M. Hillebrand, Pension Systems, Demographic Change, and the Stock Market. X, 176 pages, 2008

607

Markus Bouziane

Pricing Interest-Rate Derivatives A Fourier-Transform Based Approach

123

Dr. Markus Bouziane Landesbank Baden-Württemberg Am Hauptbahnhof 2 70173 Stuttgart Germany [email protected]

ISBN 978-3-540-77065-7

e-ISBN 978-3-540-77066-4

DOI 10.1007/978-3-540-77066-4 Lecture Notes in Economics and Mathematical Systems ISSN 0075-8442 Library of Congress Control Number: 2008920679 © 2008 Springer-Verlag Berlin Heidelberg This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Production: LE-TEX Jelonek, Schmidt & Vöckler GbR, Leipzig Cover design: WMX Design GmbH, Heidelberg Printed on acid-free paper 987654321 springer.com

To Sabine

Foreword

In a hypothetical conversation between a trader in interest-rate derivatives and a quantitative analyst, Brigo and Mercurio (2001) let the trader answer about the pros and cons of short rate models: ”... we should be careful in thinking market models are the ﬁnal and complete solution to all problems in interest rate models ... and who knows, maybe short rate models will come back one day...” In his dissertation Dr. Markus Bouziane contributes to this comeback of short rate models. Using Fourier Transform methods he develops a modular framework for the pricing of interest-rate derivatives within the class of exponential-aﬃne jump-diﬀusions. Based on a technique introduced by Lewis (2001) for equity options, the payoﬀs and the stochastic dynamics of interestrate derivatives are transformed separately. This not only simpliﬁes the application of the residue calculus but improves the eﬃciency of numerical evaluation schemes considerably. Dr. Bouziane introduces a reﬁned Fractional Inverse Fast Fourier Transformation algorithm which is able to calculate thousands of prices within seconds for a given strike range. The potential of this method is demonstrated for several one- and two-dimensional models. As a result the application of jump-enhanced short rate models for interestrate derivatives is on the agenda again. I hope, Dr. Bouziane’s monograph will stimulate further research in this direction.

T¨ ubingen, November 2007

Rainer Sch¨obel

Acknowledgements

This book is based on my Ph.D. thesis titled ”Pricing Interest-Rate Derivatives with Fourier Transform Techniques” accepted at the Eberhard Karls University of T¨ ubingen, Germany. Writing the dissertation, I am indebted to many people which contributed academic and personal development. Since any list would be insuﬃcient, I mention only those who bear in my opinion the closest relation to this work. First of all, I would like to thank my academic teacher and supervisor Prof. Dr.-Ing. Rainer Sch¨obel. He gave me valuable advice and support throughout the completion of my thesis. Furthermore, I would also express my gratitude to Prof. Dr. Joachim Grammig for being the co-referent of this thesis. Further thanks go to my colleagues from the faculty of Economics and Business Administration, especially Svenja Hager, Robert Frontczak, Wolfgang Kispert, Stefan Rostek and Martin Weiss for fruitful discussions and a pleasant working atmosphere. I very much enjoyed my time at the faculty. Financial support from the Stiftung Landesbank Baden-W¨ urttemberg is gratefully acknowledged. My deepest gratitude goes to my wife Sabine, my parents Ursula and Laredj Bouziane, and Norbert Gutbrod for their enduring support and encouragement.

T¨ ubingen, November 2007

Markus Bouziane

Contents

List of Abbreviations and Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . XV List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .XIX List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .XXI 1

2

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1

1.1 Motivation and Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Structure of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1 4

A General Multi-Factor Model of the Term Structure of Interest Rates and the Principles of Characteristic Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 An Extended Jump-Diﬀusion Term-Structure Model . . . . . . . . .

7 7

2.2 Technical Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.3 The Risk-Neutral Pricing Approach . . . . . . . . . . . . . . . . . . . . . . . . 13 2.3.1 Arbitrage and the Equivalent Martingale Measure . . . . . 15 2.3.2 Derivation of the Risk-Neutral Coeﬃcients . . . . . . . . . . . . 16 2.4 The Characteristic Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 3

Theoretical Prices of European Interest-Rate Derivatives . . 31 3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 3.2 Derivatives with Unconditional Payoﬀ Functions . . . . . . . . . . . . . 32 3.3 Derivatives with Conditional Payoﬀ Functions . . . . . . . . . . . . . . . 38

XII

4

Contents

Three Fourier Transform-Based Pricing Approaches . . . . . . . 45 4.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 4.2 Heston Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 4.3 Carr-Madan Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 4.4 Lewis Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

5

Payoﬀ Transformations and the Pricing of European Interest-Rate Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 5.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 5.2 Unconditional Payoﬀ Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 5.2.1 General Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 5.2.2 Pricing Unconditional Interest-Rate Contracts . . . . . . . . 79 5.3 Conditional Payoﬀ Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 5.3.1 General Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 5.3.2 Pricing of Zero-Bond Options and Interest-Rate Caps and Floors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 5.3.3 Pricing of Coupon-Bond Options and Yield-Based Swaptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

6

Numerical Computation of Model Prices . . . . . . . . . . . . . . . . . . 95 6.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 6.2 Contracts with Unconditional Exercise Rights . . . . . . . . . . . . . . . 96 6.3 Contracts with Conditional Exercise Rights . . . . . . . . . . . . . . . . . 97 6.3.1 Calculating Option Prices with the IFFT . . . . . . . . . . . . . 97 6.3.2 Reﬁnement of the IFFT Pricing Algorithm . . . . . . . . . . . 101 6.3.3 Determination of the Optimal Parameters for the Numerical Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

7

Jump Speciﬁcations for Aﬃne Term-Structure Models . . . . . 111 7.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 7.2 Exponentially Distributed Jumps . . . . . . . . . . . . . . . . . . . . . . . . . . 115 7.3 Normally Distributed Jumps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 7.4 Gamma Distributed Jumps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

8

Jump-Enhanced One-Factor Interest-Rate Models . . . . . . . . . 125 8.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 8.2 The Ornstein-Uhlenbeck Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

Contents

XIII

8.2.1 Derivation of the Characteristic Function . . . . . . . . . . . . . 126 8.2.2 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 8.3 The Square-Root Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 8.3.1 Derivation of the Characteristic Function . . . . . . . . . . . . . 136 8.3.2 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 9

Jump-Enhanced Two-Factor Interest-Rate Models . . . . . . . . . 145 9.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 9.2 The Additive OU-SR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 9.2.1 Derivation of the Characteristic Function . . . . . . . . . . . . . 146 9.2.2 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 9.3 The Fong-Vasicek Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 9.3.1 Derivation of the Characteristic Function . . . . . . . . . . . . . 159 9.3.2 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

10 Non-Aﬃne Term-Structure Models and Short-Rate Models with Stochastic Jump Intensity . . . . . . . . . . . . . . . . . . . . 171 10.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 10.2 Quadratic Gaussian Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 10.3 Stochastic Jump Intensity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 11 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175 A

Derivation of the Complex-Valued Coeﬃcients for the Characteristic Function in the Square-Root Model . . . . . . . . . 179

B

Derivation of the Complex-Valued Coeﬃcients for the Characteristic Function in the Fong-Vasicek Model . . . . . . . . 183

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

List of Abbreviations and Symbols

δ(x) Γ (x) ı ιM

Dirac delta function Gamma function √ imaginary unit, −1 diag[IM ]

F x [. . .], F −1 [. . .]

Fourier Transformation w.r.t. x and inverse Transformation operator

1A C

indicator function for the event A the set of complex-valued numbers

E[. . .], VAR[. . .] P, Q

expectation and variance operator real-world and equivalent martingale measure

R Ft

the set of real-valued numbers information set available up to time t

diag[. . .]

operator returning the diagonal elements of a quadratic matrix

FFT[. . .]

Fast Fourier Transformation operator

FRFT[. . . , ζ]

Fractional Fourier Transformation operator with parameter ζ

IFFT[. . .] Res[. . .]

inverse Fast Fourier Transformation operator residue operator

Re[z], Im[z]

real and imaginary part of the complex-valued variable z

RMSE RMSEa

root mean-squared error approximate root mean-squared error

tr[. . .]

trace operator

XVI

List of Abbreviations and Symbols

ψ(. . .), φ(. . .)

characteristic function and its logarithm

λ

vector of jump intensities governing the Poisson vector process

ΛΣ (xt ) , Λλ

vectors containing risk-compensating factors for the diﬀusion and jump parts, respectively

µ0 , µ1

constant coeﬃcients determining the drift component of xt

ν(J)

matrix of jump-size distributions for the ran-

Σ 0 , Σ1

dom variables J constant coeﬃcients determining the volatility

J

component of xt matrix of random jump sizes of xt

jn N(λt)

nth row of the matrix J vector of independent Poisson processes acting

Wt

with an intensity λ vector of independent Brownian motions

xt ξ(xt , t, T )

general stochastic vector process state-price kernel

a(z, τ ), b(z, τ )

complex-valued coeﬃcient functions of the general characteristic function

AD(. . .) ARCr (. . .)

Arrow-Debreu price level-based, average-rate contract

CAPr (. . .), CAPY (. . .)

level- and yield-based cap contract

CBC(. . .), CBP (. . .) F LRr (. . .), F LRY (. . .)

coupon-bond call and put option level- and yield-based ﬂoor contract

F RAr (. . .), F RAY (. . .) g0 , g 1

level- and yield-based forward rate agreement constant coeﬃcients determining the charac-

IM

teristic payoﬀ part of xt identity matrix of rank M

K N om

strike value of an option contract Nominal Value

P (. . .), CB(. . .) p(. . .)

zero bond, coupon bond probability density function

pEx (J, η)

density function of an exponentially distributed random variable J with mean and volatility η

List of Abbreviations and Symbols

pGa (J, η, p)

XVII

density function of a gamma distributed random variable J with mean ηp and volatility √ η p

pN o (J, µJ , σJ )

density function of a normally distributed random variable J with mean µJ and volatility σJ

SW Ar (. . .), SW AY (. . .) SW PY (. . .)

level- and yield-based receiver swap yield-based swaption contract

t, τ, T

calendar time, time to maturity and time of

U ARCr (. . .)

contract expiry level-based, unconditional average-rate contract

w0A (z), wA 1 (z)

complex-valued coeﬃcient functions determining the short rate for an average-rate contract

w0 , w1 (m) xt

constant coeﬃcients determining the short rate mth element of the vector xt

Y (. . .) z

simple yield to maturity Fourier-transform variable with real part zr

ZBC(. . .), ZBP (. . .)

and imaginary part zi , respectively zero-bond call and put option

ITM, ATM, OTM

in the money, at the money, and out of the money

ODE, PDE and SDE

ordinary, partial, and stochastic diﬀerential equation

OU

Ornstein-Uhlenbeck (process)

SR

Square-Root (process)

List of Tables

4.1

Idealized call option payoﬀ functions . . . . . . . . . . . . . . . . . . . . . . . . 48

8.1

Values of zero-bond call options for the jump-enhanced OU model, where the underlying zero-bond contract has a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

8.2

Values of short-rate caps for the jump-enhanced OU model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . . . . . . . . 134

8.3

Values of average-rate caps for the jump-enhanced OU model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . . . . . . . . 135

8.4

Values of zero-bond call options for the jump-enhanced SR model, where the underlying zero-bond contract has a nominal

8.5

value of 100 units. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 Values of short-rate (average-rate) caps for the jump-enhanced SR model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . 143

9.1

9.2

Values of zero-bond call options for the jump-enhanced OU-SR model, where the underlying zero-bond contract has a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 Values of zero-bond call options for the jump-enhanced OU-SR model, where the underlying zero-bond contract has a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

9.3

Values of short-rate caps for the jump-enhanced OU-SR model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . 155

9.4

Values of short-rate caps for the jump-enhanced OU-SR model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . 156

XX

List of Tables

9.5

Values of average-rate caps for the jump-enhanced OU-SR

9.6

model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . 157 Values of average-rate caps for the jump-enhanced OU-SR

9.7

model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . 158 Values of zero-bond call options for the jump-enhanced Fong-Vasicek model, where the underlying zero-bond contract has a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

9.8

Values of short-rate caps for the jump-enhanced Fong-Vasicek

9.9

model, with a nominal value of 100 units. . . . . . . . . . . . . . . . . . . . 169 Values of average-rate caps for the jump-enhanced Fong-Vasicek model, with a nominal value of 100 units. . . . . . . . 170

List of Figures

2.1

Diﬀerent contours of the Fourier transform in equation (2.26) for a strike of 90 units. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

4.1

Clockwise performed integral path for the derivation of ˜ 2 (xt , t, T ) in equation (4.27) on the real line. . . . . . . . . . . . . . . . 66 Π

5.1

Closed contour integral path for the derivation of P (xt , t, T ) in equation (5.2). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

5.2

Closed contour integral path for the discounted expectation of g (xT ). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

5.3

Closed contour integral path for the derivation of the put-call parity in equation (5.27). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

6.1

Absolute errors of zero-bond call prices for varying values of ω and zi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

6.2

Logarithmic RMSEs of zero-bond call options. . . . . . . . . . . . . . . . 106

6.3

Diﬀerences of the logarithmic RMSEa and the exact RMSE of zero-bond call options. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

6.4

Search for the optimal parameter couple (ω ∗ , zi∗ ). . . . . . . . . . . . . . 108

7.1

Possible combinations of basic diﬀusion processes and jump

7.2

parts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 The density function pEx (J, η) for varying η of an

7.3

exponentially distributed random variable. . . . . . . . . . . . . . . . . . . . 116 The density function pN o (J, µJ , σJ ) for ﬁxed µJ = 0 and varying σJ of a normally distributed random variable. . . . . . . . . . 118

XXII

7.4

List of Figures

The density function pGa (J, η, p) for ﬁxed η = 0.005 and varying p of a gamma distributed random variable. . . . . . . . . . . . 122

8.1

Probability densities for a short rate governed by a Vasicek diﬀusion model enhanced with an exponentially distributed jump component. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

8.2

Probability densities for a short rate governed by a Vasicek diﬀusion model enhanced with a gamma distributed jump component. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

8.3

8.4

Probability densities for a short rate governed by a Vasicek diﬀusion model enhanced with a normally distributed jump component. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 Probability densities for a short rate governed by a CIR diﬀusion model enhanced with an exponentially distributed jump component. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

8.5

Probability densities for a short rate governed by a CIR diﬀusion model enhanced with a gamma distributed jump component. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

9.1

Diﬀerences of the OU-SR model density function and the sum of the particular one-factor pendants for diﬀerent weighting factors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

9.2

Probability densities for a short rate governed by an OU-SR diﬀusion model enhanced with either a gamma or normally distributed jump component for the OU process. . . . . . . . . . . . . . 150

9.3

9.4

Probability densities for a short rate governed by an OU-SR diﬀusion model enhanced with a gamma distributed jump component for the SR process. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 Probability density functions of the Fong-Vasicek pure diﬀusion model for diﬀerent values of the correlation parameter ρ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164

9.5

9.6

Probability densities for a short rate governed by the Fong-Vasicek diﬀusion model enhanced with a gamma distributed jump component for the short-rate process. . . . . . . . . 165 Probability densities for a short rate governed by the Fong-Vasicek diﬀusion model enhanced with a gamma distributed jump component for the volatility process. . . . . . . . . 166

1 Introduction

1.1 Motivation and Objectives In the last few years the demand for sophisticated term-structure models, capable of reﬂecting the market behavior more realistically, e.g. models which can reproduce the feature of market shocks, has dramatically increased. For example, according to the results of their empirical study, Brown and Dybvig (1986) and A¨ıt-Sahalia (1996) question among others the use of pure diﬀusion models, such as the popular interest-rate models of Vasicek (1977) and Cox, Ingersoll and Ross (1985b), to describe the behavior of interest rates. Moreover, recent studies support the assumption of jump components in the term structure of interest rates. In the study of Hamilton (1996), Fed Funds rates on a daily base are analyzed. The author ﬁnds that settlement days and quarter-ends induce statistically signiﬁcant jumps in the term structure of interest rates. Das (2002) analyzed Fed Funds rates on daily bases over the period 1988-1997. As a result of this study, the proposed jump models show a substantially better ﬁt of the empirical data compared to the pure diﬀusion model. Durham (2005) also examined Fed Funds rates for the period 19882005. The model-generated yields of zero-bond prices are then calibrated to the Fed Funds Rate and one- and three-month U.S. Treasury bill rates. The author concludes that the so-called jump-diﬀusion models produce more accurate estimates of the interest-rate curves than the pure diﬀusion model1 . 1

Additional studies examining the empirical performance of jump-diﬀusion models are given in, e.g. Lin and Yeh (1999), Zhou (2001), Wilkens (2005), and Chan (2005).

2

1 Introduction

Thus, the ability of a term-structure model to reproduce these discount rate shocks, based e.g. on the adjustment of the discount rate by the European Central Bank, on an economic crisis, and quarter-end eﬀects, is highly appreciated. Accordingly, jump-diﬀusion interest-rate models were developed to cover this issue. Ahn and Thompson (1988) introduced one of the ﬁrst jumpdiﬀusion models for the term structure of interest rates. In their study, the interest-rate dynamics are derived within an equilibrium framework similar to the one used in Cox, Ingersoll and Ross (1985b) and particular approximate closed-form zero-bond prices are obtained. Das and Foresi (1996) also derived zero-bond prices for a jump-enhanced Vasicek (1977) model. The authors apply an exponentially distributed jump component, where the absolute value of the jump sign is drawn from a Bernoulli distribution. An alternative jump speciﬁcation for the mean-reverting normally distributed short rate is given in Baz and Das (1996)2 . In their approach, the jump-size distribution is given by a normal distribution and approximate zero-bond prices are derived. An empirical test of a Square-Root interest-rate model enhanced with uniformly distributed jumps is given in Zhou (2001). The author ﬁts the particular jump-diﬀusion model to weekly three-month Treasury bill yields. In Durham (2005), the author states an alternative approximation technique for zero-bond prices when the short rate follows the same dynamics as in Baz and Das (1996). Additionally, a bimodal normally distributed jump version of the Vasicek (1977) model together with a jump-enhanced two-factor model is presented3 . In addition, for derivatives research purposes, an important feature such interest-rate models should exhibit is the ability to generate analytical solutions for the derivatives contracts to be priced. If this can be accomplished, the interest-rate instrument can be examined in depth, e.g. doing some sensitivity analysis. However, dealing with jump components, we often have to rely on time-consuming Monte-Carlo methods in order to price interest-rate derivatives. Thus, more ambitious pricing approaches are needed. Recently, integral transformations have been found to be reliable in deriving semi closed-form 2 3

The same model speciﬁcation is used in Das (2002). The approximation technique is also discussed in depth in Durham (2006) for the bimodal normally and exponentially distributed jump extension of a Vasicek (1977) short-rate model.

1.1 Motivation and Objectives

3

solutions of derivatives contracts under more complicated stochastic dynamics. The term semi closed-form solutions in this case refers to closed-form solutions in the image space, according to the particular transformation rule. Especially the subclass of Fourier Transformations have been proven to be useful for pricing problems in ﬁnancial disciplines4 . Basically, the main advantage of this transform technique consists in providing distribution independent pricing formulae. However, even semi closed-form pricing formulae are hard to obtain, dealing with jump-size distributions such as the normal and the gamma. Accordingly, it is our objective to derive an eﬃcient and accurate pricing tool for interest-rate derivatives within a Fourier-transform pricing approach, which is generally applicable to exponential-aﬃne jump-diﬀusion models. This objective can be achieved within four steps. Firstly, we want a ﬂexible shortrate process, which is able to integrate both diﬀusion and jump components. Thus, we extend the exponential-aﬃne model presented in Duﬃe and Kan (1996) by introducing jump components. The second step is to reﬁne the concept of a modular option pricing as proposed in Zhu (2000) by applying the pricing methodology explained in Lewis (2000) and Lewis (2001)5 . Therefore, we want to formulate a distribution-independent pricing framework, where the particular interest-rate contract price can be clearly separated into stochastic and payoﬀ speciﬁc parts. Apart from the pricing theory, we also need a tool to obtain numerical values of the contracts to be priced. A very popular strategy to price derivatives is the Monte-Carlo approach. However, being generally applicable, this numerical pricing approach suﬀers from its time-consuming calculations and its poor convergence to true solutions. The third objective of this thesis is to develop an algorithm, which appropriately computes option prices in the Lewis (2001) pricing approach. In contrast to the Fast Fourier Transformation (FFT), as used in Carr and Madan (1999) for the pricing of 4

Heston (1993) is the seminal paper on this topic, where semi closed-form solutions for options on equities in a stochastic volatility model are derived for the ﬁrst time. Among others, we mention the inﬂuential work of Bakshi and Madan (2000) and Duﬃe, Pan and Singleton (2000) in deriving option prices using Fourier

5

Transformations. Even though this pricing method is mentioned for the ﬁrst time in Lewis (2000), we henceforth refer to Lewis (2001) as the source, because of the detailed discussion and derivation of the pricing methodology.

4

1 Introduction

equity options, we base our computations on the Inverse Fast Fourier Transform (IFFT). Consequently, we introduce in this thesis a new, IFFT-based pricing algorithm, which is able to calculate thousands of option prices within fractions of a second and is a straightforward application to option pricing in the Lewis (2001) framework. The last step is then to examine density functions and contract prices of some popular interest-rate diﬀusion models enhanced with three diﬀerent jump candidates.

1.2 Structure of the Thesis This thesis is organized as follows. We start in chapter two with the formulation of a general term-structure model, which is governed by a multivariate jump-diﬀusion process. After introducing some general concepts in stochastic calculus we demonstrate how the relevant risk-neutral coeﬃcients of the instantaneous interest-rate process can be obtained. Afterwards, we discuss the technique of performing a Fourier Transformation and its inverse and state the system of ordinary diﬀerential equations the general characteristic function has to solve. In chapter three we discuss a representative collection of some interest-rate derivative contracts which can be solved within the Fourierbased pricing mechanism. We distinguish between contracts with conditional and unconditional exercise rights, because of the diﬀerent pricing procedure. Subsequently, in chapter four we discuss three Fourier-based pricing approaches. We begin our summary with the pricing technique using Fouriertransformed Arrow-Debreu state prices. Since this type of valuation was ﬁrst applied by Heston (1993) and further discussed by Bakshi and Madan (2000), we henceforth refer to this approach as the Heston transform approach. Subsequently, we discuss the pricing procedure introduced by Carr and Madan (1999). In this thesis the authors exploit the Fourier Transformation applied not only to the state price densities but to the entire option price. They introduce a valuation approach where theoretical option prices can be subsequently recovered applying a highly eﬃcient algorithm, namely the Fast Fourier Transform, hereafter denoted as FFT. Finally, we discuss the valuation methodology applied by Lewis (2001). This approach features several advantages. Firstly, its composition is highly modularized. Secondly, employing Cauchy’s residue theorem, the approach can be consistently used both for

1.2 Structure of the Thesis

5

interest-rate derivatives with unconditional and conditional exercise rights. Fortunately, this methodology enables the application of an reﬁned IFFT algorithm which we implement in our pricing procedure. In chapter ﬁve, we derive the particular Fourier Transformations of payoﬀ functions needed in pricing the contract forms previously presented. Additionally, we derive in case of a one-factor term-structure model the Fourier representation of a swaption and a coupon-bond option, respectively. Chapter six gives an outline of the numerical algorithm used for pricing purposes. Again, we distinguish between the computation of derivatives with conditional and unconditional exercise rights. Subsequently, we present a further reﬁnement of the pricing algorithm for option contracts by the application of the Fractional Fourier Transformation according to the article of Bailey and Swarztrauber (1994). The last part of the chapter discusses the issue of ﬁnding the optimal parameter constellation of the numerical algorithm. In chapter seven we brieﬂy discuss three diﬀerent jump-size speciﬁcations and derive their general jump transforms. In chapters eight and nine we examine both jump-enhanced one-factor and two-factor interest-rate models and focus on the impact of diﬀerent jump speciﬁcations. The particular one-factor models we enhance with jump components are the prominent interest-rate models introduced in Vasicek (1977) and Cox, Ingersoll and Ross (1985b). For the class of two-factor models we exemplarily discuss an additive model used in Sch¨ obel and Zhu (2000) and a subordinated model according to Fong and Vasicek (1991a). To our knowledge, in case of the Fong and Vasicek (1991a) model, option prices are presented for the ﬁrst time. In chapter ten, we give a perspective of model extensions for which the pricing procedure is also capable in deriving numerical solutions. The ﬁrst extension is to consider a special model class of non-aﬃne interest-rate models. Another extension of our interest-rate model is to consider stochastic jump intensities. Since it ﬁts into the exponential-aﬃne model setup of Duﬃe, Pan and Singleton (2000), the implementation in our pricing procedure presents no greater diﬃculties. However, due to the non-existence of closed-form solutions in any case, we brieﬂy discuss these extensions. In the last chapter, we review the results of our study and give some concluding remarks.

2 A General Multi-Factor Model of the Term Structure of Interest Rates and the Principles of Characteristic Functions

2.1 An Extended Jump-Diﬀusion Term-Structure Model The evolution of the yield curve can be described in various ways. For instance, it is possible to use such quantities as zero-bond prices, instantaneous forward rates and short interest rates, respectively, to build the term structure of interest rates. If the transformation law from one quantity to the other is known, the choice of the independent variable is just a matter of convenience. In this thesis, we attempt to model the dynamics of the instantaneous interest rate, denoted hereafter by r(xt ), in order to construct our derivatives pricing framework. This instantaneous interest rate r(xt ) is also often referred to as the short-term interest rate or short rate, respectively, and characterizes the risk-free rate for borrowing or lending money over the inﬁnitesimal time period [t, t + dt]. Since we model the dynamics in a continuous trading environment, the relevant processes are described via stochastic diﬀerential equations. The economy we consider has the trading interval [0, T ]. The uncertainty under the physical probability measure is completely speciﬁed by the ﬁltered probability space (Ω, F, P). In this formulation Ω denotes the complete set of all possible outcome elements ω ∈ Ω. The information available in the economy is contained within the ﬁltration (F)t≥0 , such that the level of uncertainty is resolved over the trading interval with respect to the information ﬁltration. The last term, completing the probability space, is called the real-world probability measure P on (Ω, F), since it reﬂects the real-world probability law of the data.

8

2 A Multi-Factor Model and Characteristic Functions

We model the dynamic behavior of the term structure in the spirit of Duﬃe and Kan (1996) and Duﬃe, Pan and Singleton (2000), to preserve an exponential-aﬃne structure of the characteristic function. However, we extend the framework in Duﬃe and Kan (1996) to allow for N diﬀerent trigger processes6 , which oﬀers more ﬂexibility. The term structure is then modeled by a multi-factor structural Markov model of M factors, represented by a random vector xt , which solves the multivariate stochastic diﬀerential equation, (1) dxt dx(2) t .. P (2.1) dxt = = µP (xt ) dt + Σ(xt ) dWP t + J dN(λ t). . (M−1) dxt (M) dxt The coeﬃcient vector µP (xt ) has the aﬃne structure P µP (xt ) = µP 0 + µ1 xt

(2.2)

P M M×M and the variance-covariance matrix Σ(xt )Σ(xt ) with (µP 0 , µ1 ) ∈ R ×R suﬃces the relation

Σ(xt )Σ(xt ) = Σ0 + Σ1 xt ,

(2.3)

where Σ0 ∈ RM×M is a matrix and Σ1 ∈ RM×M×M is a third order tensor. The vector WP t in equation (2.1) represents M orthogonal Wiener processes. Thus, we have7

P EP ( dWP t dW t ) = IM dt

with IM as the M × M identity matrix. As mentioned above, we extend the ordinary diﬀusion model8 with N independent Poisson processes, condensed in the vector N(λP t). This vector process acts with constant and positive intensities9 λP . We allow for every 6

Chacko and Das (2002) model also the term structure with help of diﬀerent

7

Poisson processes. However, their approach consider a subordinated short rate. If not indicated otherwise, we subsequently use the shorthand notation E[ · ] for

8

the expression E[ · |Ft ]. This would be the original model approach presented in Duﬃe and Kan (1996). This exponential-aﬃne model can be easily extended to stochastic jump intensi-

9

P ties of the form λP (xt ) = λP 0 + λ 1 xt . See Chapter 10.

2.1 An Extended Jump-Diﬀusion Term-Structure Model

9

particular factor in xt an amount of N diﬀerent jumps drawn from a jump amplitude matrix J ∈ RM×N . Hence, the distribution functions of the particular jump amplitudes are given within the matrix ν(J). Finally, all jump amplitudes in J are independent of the state of the vector xt 10 . To preserve the exponential-aﬃne structure of any derivatives contract based on r(xt ) and xt , respectively, all random sources, the Brownian motions P WP t , intensities λ and jump amplitudes J are mutually independent. As a direct consequence of the independence of J and xt , there is no chance to generate an arbitrage opportunity according to available information before the particular jump occurs. Hence, given a jump time t∗ , we have formally J ∈ Ft∗− . Therefore, if a jump occurs at time t∗ , nobody is able to predict the exact jump amplitude and cannot gain an arbitrarily large proﬁt with certainty. In this thesis, the choice of jump amplitudes in J can draw on three different types of distribution. These are: •

Exponentially distributed jumps.

• •

Normally distributed jumps. Gamma distributed jumps.

These jump distributions and the resulting jump transforms, which are used in our pricing mechanism, are covered in Chapter 7. Basically, we prefer to model the term structure in terms of the instantaneous short interest rate r (xt )11 , because in this framework all fundamental quantities are properly deﬁned as the expectation of some functionals on the underlying process r (xt ). Accordingly, we are able to construct an arbitragefree economy and simultaneously guarantee a consistent pricing methodol10

From a technical point of view, it is either possible to introduce a dependence on xt for the jump intensity together with independent random jump amplitudes or a dependence on xt for the jump amplitude together with constant jump intensities.

11

See Zhou (2001), p. 4. Other approaches are possible, e.g. the direct approach as used in Sch¨ obel (1987) and Briys, Crouhy and Sch¨ obel (1991) or modeling the forward-rate process as done in Heath, Jarrow and Morton (1992).

10

2 A Multi-Factor Model and Characteristic Functions

ogy12 . The drawback of this approach is that we might not be able to explain perfectly the entire term structure extracted from observed bond market prices and therefore must content ourselves with a best ﬁt scenario. The literature distinguishes between two approaches in modeling the short interest rate in a multidimensional framework. Firstly, we can identify a strategy, which we call henceforth the subordinated modeling approach. Here, the short rate is modeled as (1)

(2)

(M)

r (xt ) = w0 + w1 xt (xt , . . . , xt

).

Consequently, the other M − 1 stochastic factors are subordinated loadings, containing e.g. a stochastic volatility and/or a stochastic mean13 . Apart from (1) the stochastic variable xt , we also consider the deterministic parameters w0 and w1 in modeling the short rate. Indeed, there are other factors, which can possibly have some other economic meaning worth to be included in the interest-rate model. The second method in modeling short rates, which we call the additive modeling approach, is to represent rt as a weighted sum over xt , formally given by r (xt ) = w0 + w xt , 12

13

This means that all derivative prices are based on the same price of risk. See Culot (2003), Section 2.1. In Brennan and Schwartz (1979), Brennan and Schwartz (1980), and Brennan and Schwartz (1982) the short-rate process is subordinated by a stochastic long-term rate. Beaglehole and Tenney (1991) discuss a two-factor interest-rate model with a stochastic long-term mean component and Fong and Vasicek (1991a) introduce a short-rate model with stochastic volatility. A model where the short rate depends on a stochastic inﬂation factor is modeled in Pennacchi (1991). Kellerhals (2001) analyzes an interest-rate model with a stochastic market price of risk component. In Balduzzi, Das, Foresi and Sundaram (1996), the authors present a short-rate model with a stochastic mean and volatility component.

2.2 Technical Preliminaries

11

where w is a M × 1 vector containing separate weights for the corresponding factor loadings in xt 14 . However, this model approach possibly entails diﬃculties in explaining the economic meaning of the variables xt 15 .

2.2 Technical Preliminaries Before we proceed any further, we have to discuss some general results and principles of stochastic analysis, which are commonly used in ﬁnancial engineering, namely the prominent Itˆ o’s Lemma and the equally famous FeynmanKac Theorem. These two principles play a major role in diﬀusion theory and are well connected. Since we consider discontinuous jumps in our model setup, we have to use extended versions of these two results. At ﬁrst we have to state some regularity conditions on the jump-diﬀusion process, in order to guarantee their application.

Deﬁnition 2.2.1 (Regularity Conditions for Jump-Diﬀusion Processes). If the vector process xt represents a multivariate jump-diﬀusion, the parameter coeﬃcients µ(xt ), Σ(xt ) have to satisfy the following technical conditions16 for all t ≥ 0 •

µ(xat ) − µ(xbt ) ≤ A1 xat − xbt

•

Σ(xat )) − Σ(xbt ) ≤ A2 xat − xbt

•

µ(xat ) ≤ A1 (1 + xat )

•

Σ(xat )) ≤ A2 (1 + xat )

where xat , xbt ∈ RM are two vectors containing diﬀerent realizations of xt and the constants A1 , A2 < ∞ denote some scalar barriers. Additionally, we need 14

Langetieg (1980) models the short rate as an additive process consisting of two correlated Ornstein-Uhlenbeck processes. In Beaglehole and Tenney (1991) an additive, multivariate quadratic Gaussian interest-rate model is given. Longstaﬀ and Schwartz (1992) and Chen and Scott (1992) model the interest-rate process

15 16

as the sum of two uncorrelated Square-Root processes. A comprehensive discussion on this topic is given in Piazzesi (2003). The ﬁrst two conditions are known as the Lipschitz conditions, the latter two represent the growth or polynomial growth conditions. See, for example, Karlin and Taylor (1981).

12

2 A Multi-Factor Model and Characteristic Functions

for the jump components the integral

R

ecJmn dν(Jmn ) to be well deﬁned for

every Jmn ∈ J and some constant c ∈ C.

If the conditions posed above are met, we are able to apply both Itˆo’s Lemma and the Feynman-Kac Theorem. We start with Itˆ o’s Lemma. This lemma enables us to determine the stochastic process driving some function f (xt , t, T ), depending on time t and a stochastic (vector) variable, e.g. the process xt given in equation (2.1). The variables t and xt , respectively, are hereafter denoted as the independent variables. The coeﬃcients µ(xt ) and λ used in this section have no superscripts, because the principles introduced here hold in general.

Theorem 2.2.2 (Itˆ o Formula for Jump-Diﬀusion Processes17 ). Assume the function f (xt , t, T ) is at least twice diﬀerentiable in xt and once diﬀerentiable in t. Then the canonical decomposition of the stochastic diﬀerential equation for f (xt , t, T ) is given by ∂f (xt , t, T ) ∂f (xt , t, T ) + µ(xt ) df (xt , t, T ) = ∂t ∂xt

2 1 ∂ f (xt , t, T ) + tr Σ(xt )Σ(xt ) dt 2 ∂xt ∂xt (2.4) ∂f (xt , t, T ) Σ(xt ) dWt + ∂xt + (f (xt , J, t, T ) − f (xt , t, T )) dN(λt), where the function f (xt , J, t, T ) contains all jump components with elements (f (xt , J, t, T ))n = f (xt + jn , t, T ) and jn ∈ RM contains as mth element Jmn of the amplitude matrix J.

Another key result which we use extensively is the Feynman-Kac theorem. This theorem provides us with a tool to determine the system of partial diﬀerential equations (PDEs), given an expectation. 17

See, Kushner (1967), p. 15, for the jump-extended version of Itˆ o’s lemma.

2.3 The Risk-Neutral Pricing Approach

13

Theorem 2.2.3 (Feynman-Kac). If the restrictions in deﬁnition 2.2.1 hold, we have the expectation f (xt , t, T ) = E e

−

T

h(xs ,s) ds t

f (xT , T, T ) ,

(2.5)

solving the partial diﬀerential equation

2 ∂f (xt , t, T ) 1 ∂f (xt , t) ∂ f (xt , t, T ) + µ(xt ) + tr Σ(xt )Σ(xt ) ∂t ∂xt 2 ∂xt ∂xt

(2.6)

+ EJ [f (xt , J, t, T ) − f (xt , t, T )] λ = h(xt , t)f (xt , t, T ), with boundary condition18 f (xT , T, T ) = G (xT )

(2.7)

and f (xt , J, t, T ) as deﬁned in theorem 2.2.2.

In diﬀusion theory, the function h(xt , t) is commonly addressed to as the killing rate of the expectation19 and can be interpreted as some short rate. Since we use equivalently as killing rate a short rate characterized by the time constant coeﬃcients w0 and w we set the relation h(xt , t) = r (xt ) . As we will see, these two principles are the fundamental tools in obtaining the solutions for our upcoming valuation problems, especially in calculating the general characteristic function of a stochastic process, which is discussed in the next sections.

2.3 The Risk-Neutral Pricing Approach So far, the stochastic behavior of the state vector xt was assumed to be modeled under the real-world probability measure P. This probability measure depends on the investor’s assessment of the market and therefore cannot be 18 19

The operator EJ [ · ] denotes the expectation with respect to the jump sizes J. See, for example, Øksendal (2003), p. 145.

14

2 A Multi-Factor Model and Characteristic Functions

used in calculating unique derivatives prices20 . However, for valuation purposes we need to derive contract prices under the condition of an arbitrage-free market21 , which will be shown in this section. According to the seminal papers of Harrison and Kreps (1979) and Harrison and Pliska (1981), it is a well known and rigorously proved fact, if one can ﬁnd at least one equivalent martingale measure with respect to P, then the observed market is arbitrage-free and therefore a derivatives pricing framework can be established. Thus, we establish the link between this equivalent martingale measure Q, also known as the risk-neutral probability measure22 , and the probability measure P in this section. Since we are dealing with M stochastic factors, primarily integrated in the short rate r (xt ), which are all non-tradable goods, we are confronted with an incomplete market. In contrast to other model frameworks in which factors represent prices of tradable goods, we encounter a somewhat more diﬃcult situation to end up in a consistent arbitrage-free pricing approach23 . Foremost, we need to introduce for every source of uncertainty a market price of risk reﬂecting the risk aversion of the market. The common procedure in this case is to choose a particular equivalent martingale measure, sometimes also called the pricing measure which determines the appropriate numeraire to be applied24 . Having chosen the numeraire, which has the function of a denominator of the expected contingent claim and determines the martingale condition for the expectation, we afterwards have to extract yields for diﬀerent maturities of zero-bond prices. In the next step the model prices of zero bonds 20 21

See, for example, Musiela and Rutkowski (2005), p. 10. The arbitrage-free approach is also known as the partial equilibrium approach. Including preferences of investors, i.e. working with utility functions would be a general equilibrium approach. Sch¨ obel (1995) gives a detailed overview of both

22

approaches. The terminology can be justiﬁed, since in a risk-neutral world, where all market participants act under a risk-neutral utility behavior, the probability measures P

23

and Q coincide. See, for example, Duﬃe (2001), p. 108. This statement holds only for tradable goods modeled by pure diﬀusion processes. Otherwise, due to the jump uncertainty one has again to implement some variable

24

compensating jump risk. See Merton (1976). This can be for example the money market account or zero-coupon bond prices. See Dai and Singleton (2003), pp. 635-637.

2.3 The Risk-Neutral Pricing Approach

15

are calibrated with respect to this empirical yield curve. In the calibration process for these parameters, two separate approaches can be utilized25 . In the ﬁrst approach one computes the particular model parameters under the P measure together with the diﬀerent market prices of risk. The other method would be to calibrate the model onto the parameters under the objective measure Q. A problem which is common to all model frameworks, where the instantaneous interest rate r(xt ) is used to describe the term structure of interest rates is that in general the given yield curve is not matched perfectly. Hence, we rather want an arbitrage-free model, which might not be able to explain perfectly all observed yields, but to state a model with an internally consistent stochastic environment. In the upcoming subsections, we will ﬁrst give an outline how the riskneutral measure is deﬁned and how the particular coeﬃcients under this probability measure Q can be derived for our aﬃne term-structure model. Due to the jump-diﬀusion framework, we also focus on the topic that our martingale measure should consider for discontinuous price shocks. 2.3.1 Arbitrage and the Equivalent Martingale Measure Before we start with the formulation of our option-pricing methodology, we need to ensure the existence of an arbitrage-free pricing system. A very useful insight for this delicate matter is given in the above mentioned work of Harrison and Kreps (1979) and Harrison and Pliska (1981). Using measure theory, they judge the market to be arbitrage free enabling the consistent calculation of derivative prices if at least one equivalent martingale measure can be found, corresponding to the physical measure P. Hence, using the money market account as numeraire in order to derive Q, the price of a derivative contract would be just the discounted expectation of its terminal payoﬀ G (xT )26 . So our ﬁrst step is to deﬁne the relevant conditions for an equivalent martingale measure. 25 26

See Duﬃe, Pan and Singleton (2000), p. 1354. See, for example, Geman, Karoui and Rochet (1995) and Dai and Singleton (2003), p. 635.

16

2 A Multi-Factor Model and Characteristic Functions

Deﬁnition 2.3.1 (Equivalent Probability Measure). Two probability measures P and Q are equivalent, if for any event A, P(A) > 0 if and only if Q(A) > 0. According to deﬁnition 2.3.1, the equivalent probability measure Q must only agree on the same null sets given by P. The next property we need, in order to obtain the probability measure Q, is the martingale property.

Deﬁnition 2.3.2 (Martingale Property). A stochastic process f (xt , t) is a martingale under the probability measure Q if and only if the equality f (xt , t, T ) = EQ [f (xT , T, T )]

(2.8)

holds for any t ≤ T .

This last deﬁnition ensures the fair game ability of our interest-rate market. Combining deﬁnitions 2.3.1 and 2.3.2 lead us to the equivalent martingale measure Q with respect to P. Thus, to be a fair game, respectively a martingale, the probability measure Q transforms the probability law for xt , leaving the null sets of P untouched. In the next subsection we show the transition of the probability law from the real-world measure P to the risk-neutral measure Q. 2.3.2 Derivation of the Risk-Neutral Coeﬃcients Having found the formal conditions of an equivalent martingale measure, we now want to derive the transformation rule from measure P to Q. This rule, also called the Radon-Nikodym derivative ξ(xt , t, T ), is represented by dQ ξ(xT , T, T ) . (2.9) = dP Ft ξ(xt , t, T ) In order to derive the risk-neutral coeﬃcients, we adopt the corresponding pricing-kernel methodology. Doing this, the pricing kernel or Radon-Nikodym derivative ξ(xt , t, T ), belongs itself to the class of exponential-aﬃne functions of xt 27 . The principle of risk-neutrality implies for the state-price kernel an 27

See, for example, Dai and Singleton (2003), p. 642.

2.3 The Risk-Neutral Pricing Approach

17

expected discount rate equal to the instantaneous risk-free rate r (xt ). Thus, we need the equation P

E

dξ(xt , t, T ) = −r (xt ) dt, ξ(xt , t, T )

(2.10)

to hold. Using this type of state-price kernel, we have the discounted expectation of an interest-rate derivatives price to fulﬁll the deﬁnition of a martingale as described in theorem 2.3.2. Consequently, ensuring the expectation made above holds and considering the systematic risk factors, we choose the speciﬁc form of ξ(xt , t, T ) to satisfy dξ(xt , t, T ) = −r (xt ) dt − ΛΣ (xt ) dWP − Λλ dN(λP t) − λP dt . (2.11) ξ(xt , t, T ) The vectors ΛΣ (xt ) and Λλ compensate the sources of risk under the riskneutral measure Q for the vector of Brownian motions and the vector of Poisson processes, respectively. The vector ΛΣ (xt ) is characterized by the two relations28

ΛΣ (xt ) ΛΣ (xt ) = l0 + l1 xt Σ (xt ) ΛΣ (xt ) = s0 + s1 xt with l0 ∈ R, l1 , s0 ∈ RM , and s1 ∈ RM×M . Deﬁning ΛΣ (xt ) like this, we ensure the exponential-aﬃne structure in the pricing kernel ξ(xt , t, T ). In contrast to the constant, N -dimensional vector Λλ , we need to establish in ΛΣ (xt ) a dependence on the state vector xt because of a possibly non(m) zero matrix Σ1 29 . Thus, if a particular factor xt has a constant volatility coeﬃcient, meaning its volatility does not depend on any element in xt , there is either no dependence on xt for the respective element in the the vector ΛΣ (xt ) and vice versa. Since λP is the vector of expected arrival rates, we have with

EP dN(λP t) − λP dt = 0N ,

a P-martingale, representing a vector of compensated Poisson processes30 . 28

29

30

Compare, for example, with Duﬃe, Pan and Singleton (2000), Culot (2003), and Dai and Singleton (2003). Dealing with a Square-Root process, we cannot set the particular market price of risk to a constant value, see Cox, Ingersoll and Ross (1985b), Section 5. A compensated Poisson process can be roughly seen as a discontinuous equivalent of a Brownian motion. See, for example, Karatzas and Shreve (1991), p. 12.

18

2 A Multi-Factor Model and Characteristic Functions

As a consequence of this incomplete market, the vectors ΛΣ (xt ) and Λλ are not uniquely deﬁned. Therefore, the pricing kernel itself is not uniquely deﬁned either and we have to determine these risk price vectors with a calibration of yields generated by the model to the empirical yield curve as mentioned earlier. We assume this calibration to depend on the yields of traded zero-coupon bonds P (xt , t, T ) with diﬀerent times to maturities31 . Suppressing unnecessary notations for convenience and applying Itˆo’s Lemma, we get the following SDE for the P-dynamics of a zero-coupon bond dP (xt , t, T ) = µP dt + σ P dWP + JP dN(λP t)

(2.12)

with drift, diﬀusion and jump components32 ∂P (xt , t, T ) ∂P (xt , t, T ) + µP (xt ) ∂t ∂xt

2 ∂ P (xt , t, T ) 1 + tr Σ(xt )Σ(xt ) , 2 ∂xt ∂xt ∂P (xt , t, T ) , σ P = Σ(xt ) ∂xt

µP =

JP = P(xt , J, t, T ) − P (xt , t, T ) .

(2.13)

(2.14) (2.15)

On the other hand, we impose the martingale condition for traded contracts, which is due to the chosen numeraire, T P (xt , t, T ) =EQ e =EP

−

r(xs ) ds t

P (xT , T, T )

ξ(xT , T, T ) P (xT , T, T ) . ξ(xt , t, T )

(2.16)

Multiplying this last equation with ξ(xt , t, T ), which is known at time t and therefore a certain quantity, we consequently have ξ(xt , t, T )P (xt , t, T ) to be a martingale and the inﬁnitesimal increment d (ξ(xt , t, T )P (xt , t, T )) to be a local martingale33 . According to Theorem 2.2.2 we have 31

32

33

Since coupon bonds are commonly traded, zero-bond values can be synthetically generated by coupon stripping. P(xt , J, t, T ) has the equivalent deﬁnition as f (xt , J, t, T ) with all calculations made with respect to P (xt , t, T ). See Theorem 2.2.2. The existence of a local martingale under the new measure Q is suﬃcient for the no-arbitrage condition. See Delbaen and Schachermayer (1995) and Øksendal (2003) Section 12.1., respectively.

2.3 The Risk-Neutral Pricing Approach

19

d(ξ(xt , t, T )P (xt , t, T )) = ξ(xt , t, T ) dP (xt , t, T ) + P (xt , t, T ) dξ(xt , t, T ) + dP (xt , t, T ) dξ(xt , t, T ) = ξ(xt , t, T )µP dt + ξ(xt , t, T )σP dWP + ξ(xt , t, T )JP N(λP )

(2.17)

− P (xt , t, T ) ξ(xt , t, T )r (xt ) dt

− P (xt , t, T ) ξ(xt , t, T )ΛΣ (xt ) dWP − P (xt , t, T ) ξ(xt , t, T )Λλ dN(λP t) − λP dt P

− ξ(xt , t, T )σP ΛΣ (xt ) dt − ξ(xt , t, T )JP Iλ N Λλ dt. In the last equation, we used for the inﬁnitesimal time increments the relation dt dt = 0, and for the vector of uncorrelated Brownian motions

dWP dWP = IM dt. Similarly, the corresponding expression for the vector of independent Poisson processes is P

dN(λP t) dN(λP t) = Iλ N dt, P

where Iλ N represents a matrix consisting of the diagonal elements P = λP , diag Iλ N and zeros otherwise. In the next step, we divide for notational ease all coefﬁcients of the zero-bond SDE (2.12) by P (xt , t, T ). Hence, we use hereafter the normalized coeﬃcients, µP , P (xt , t, T ) σP ˜P = σ , P (xt , t, T ) JP ˜P = J . P (xt , t, T ) µ ˜P =

Combining condition (2.16) and equation (2.17), and keeping in mind that under P-dynamics, the Brownian motions and the compensated Poisson processes in equation (2.11) are martingales, we get for the expectation

20

2 A Multi-Factor Model and Characteristic Functions

EP

d (ξ(xt , t, T )P (xt , t, T )) ˜ P λP dt =µ ˜P dt + EJ J ξ(xt , t, T )P (xt , t, T ) ˜ P ΛΣ (xt ) dt − r (xt ) dt − σ P ˜ P Iλ Λλ dt ≡ 0. − EJ J N

(2.18)

If we now solve equation (2.18) for the modiﬁed drift coeﬃcient µ ˜ P , subsequently eliminating all dt terms, we eventually end up with the relation P ˜ P Iλ Λλ − λP , ˜ P ΛΣ (xt ) + EJ J µ ˜P = r (xt ) + σ (2.19) N which means that the rate of return of a zero bond must be equal to the risk free short rate plus some terms reﬂecting the particular risk premiums of the diﬀerent sources of uncertainty. We are now ready to identify the corresponding formal expressions under Q-dynamics of the coeﬃcient parameters µP and λP . Comparing equation (2.13) with (2.19) lead us to the fundamental partial diﬀerential equation for zero-bond prices34 ∂P (xt , t, T ) ∂P (xt , t, T ) P + µ − Σ(xt )ΛΣ (xt ) ∂t ∂x t

1 ∂ 2 P (xt , t, T ) + tr Σ(xt )Σ(xt ) 2 ∂xt ∂xt P + EJ [JP ] λP − Iλ N Λλ = r (xt ) P (xt , t, T ) .

(2.20)

According to equation (2.20), together with Itˆ o’s Lemma, and the FeynmanKac representation, we are able to express the risk-neutral parameters as Q µQ = µP − Σ(xt )ΛΣ (xt ) = µQ 0 + µ1 x t , Q

P

λ =λ −

P Iλ N Λλ .

(2.21) (2.22)

Since the jump intensities λQ have to be positive, we need Λλ small enough to ensure the positiveness of the jump intensities under the risk-neutral measure Q given the intensity vector λP . The constant coeﬃcients in the variancecovariance matrix (2.3) remain unchanged under the new measure Q. This 34

Once the risk-neutral coeﬃcients for the interest-rate process are determined, equation (2.20) can be used to price any European contingent claim by exchanging the terminal condition and replacing P (xt , t, T ) with the particular function representing the price of the derivative security to be calculated.

2.4 The Characteristic Function

21

phenomenon is often referred to as the diﬀusion invariance principle, although this terminology is not completely correct. We want to emphasize that the variations of the Brownian motions only coincide under both measures P and Q, if the variance-covariance matrix exclusively exhibits constant coeﬃcients35 . Otherwise, we are implicitly dealing with a diﬀerent time-dependent variance-covariance matrix, since the vector xt experiences a drift correction and therefore aﬀects the relation given in equation (2.3). Consequently, the probability transformation law of the process xt from P to Q does not only contain a drift compensation. Moreover, besides the jump intensity correction, the very shape of the probability density itself can be changed, due to the implicitly altered variations of the diﬀusion terms. Hence, calibrating the theoretical term-structure model to zero-bond yields, whether estimating the parameters of the left or the right sides of equations 2.21 and 2.22, results in the following SDE governing the particular factors under risk-neutral dynamics Q dxt = µQ (xt ) dt + Σ(xt ) dWQ t + J dN λ t ,

(2.23)

which we use in the subsequent sections as starting point for our calculations.

2.4 The Characteristic Function In this section, we ﬁrst give a brief overview of the abilities of characteristic functions and show afterwards how the characteristic function of an exponential-aﬃne process, as given in equation (2.1), can be derived. We generalize the principle of building characteristic functions for some scalar process g(xt ), which is essential for our derivatives pricing technique. Since characteristic functions play a major part in our derivation of semi closed-form solutions for interest-rate derivatives, we discuss also some of their fundamental properties. Before we introduce the characteristic function itself, we ﬁrst need to state a deﬁnition of Fourier Transformations of some deterministic variable x36 . 35 36

In this case, we would deal with the matrix Σ(x)Σ(x) = Σ0 . In the literature, there seems to exist various deﬁnitions for this type of transformation. Thus, we want to clarify the issue by giving a straightforward deﬁnition

22

2 A Multi-Factor Model and Characteristic Functions

This concept belongs to the ﬁeld of integral transformations37 and is a widely used tool in engineering disciplines, especially in signal processing. Deﬁnition 2.4.1 (General one-dimensional Fourier Transformation and its Inversion). We deﬁne the Fourier Transformation F x [ · ] of some function f (x) with respect to the independent variable x as ∞ eızx f (x) dx = fˆ(z),

F [f (x)] = x

(2.24)

−∞

where z ∈ C denotes the transform variable in Fourier space, satisfying the restriction Im(z) ∈ (χ, χ) with χ and χ denoting some lower and upper bound√ aries guaranteeing the existence of the Fourier Transformation, ı = −1 as the standard imaginary unit, and fˆ(z) as the shorthand notation for the Fourier Transformation of f (x) with respect to its argument x. Accordingly, the inverse transformation operator F −1 [ · ] is then deﬁned by F

−1

1 [fˆ(z)] = 2π

∞

e−ızx fˆ(z) dz = f (x).

(2.25)

−∞

Due to the exponential character of the Fourier Transformation, we need to establish in equation (2.25) a normalization factor of 2π. The terminology general one-dimensional Fourier Transformation, in contrast to an ordinary one-dimensional Fourier Transformation, is used because we do not limit the transformation variable z to be on the real line38 . Thus, we allow z to be complex-valued, which makes equation (2.24) and (2.25) a line integral, performed parallel to the real line. Note that both the transform and its inverse in this section. In ﬁnancial studies our deﬁnition according to equation (2.24) of a Fourier Transformation seems to be commonly accepted. See, for example, Carr and Madan (1999), Bakshi and Madan (2000) and Raible (2000). On the other hand in engineering sciences, the opposite deﬁnition of a Fourier Transformation 37

and its inverse operation does exist. See, for example, Duﬀy (2004). Other popular integral transformations are e.g. the Laplace transformation or the z-transformation. A comprehensive discussion of the Laplace Transformation is

38

given in Doetsch (1967). Hence, the equivalent expression complex Fourier Transformation is sometimes used in the literature.

2.4 The Characteristic Function

23

operation have to take place on the same strip going through Im(z), in order to reconstruct the original function f (x). The advantage in performing this general Fourier Transformation is the possibility to derive image functions in cases where the ordinary transform approach would fail, e.g. for functions which are unbounded39 . However, in these cases, the general approach enables us to derive solutions for their Fourier Transformations. For example, if we want to compute the Fourier Transformation of a function40 G(x) = max(ex − K, 0), the ordinary transformation approach appears to be useless, since F x [G(x)] → ∞. Performing a general transformation, in this case within the strip Im(z) ∈ (1, ∞), we get41

K 1+ız , (2.26) ız(1 + ız) where Im(z) can be ﬁxed at every value within the above mentioned strip to derive the original function by applying the inverse Fourier Transformation. F x [G(x)] =

The diﬀerent contours in Fourier space of the transformed payoﬀ function given in equation (2.26) are depicted in Figure 2.1. Having derived the fundamental technique to compute Fourier Transformations, which is an essential part in this thesis, we go further and have a look at Fourier Transformations of density functions of stochastic variables, which are commonly known as characteristic functions. Deﬁnition 2.4.2 (Scalar Characteristic Functions). We deﬁne the scalar (m)

characteristic function ψ x (xt , z, w0 , w, t, T ) as the expected value of the ter(m) minal condition G (xT ) = eızxT , given the state xt at time t ≤ T . This can be expressed more formally as 39

40

This is the case for most payoﬀ structures of option contracts, e.g. plain vanilla call or put options. This function represents, for instance, the payoﬀ function of a plain vanilla call option in an asset pricing environment, where x is the natural logarithm of the

41

underlying asset price. In Section 5.3, Fourier Transformations are derived in detail for diﬀerent types of payoﬀ functions.

24

2 A Multi-Factor Model and Characteristic Functions

2 1.5

Re(f(z))

1 0.5 0 −0.5 −1 2 1.5

Im(z)

1

−4

−6

−2

2

0

4

6

Re(z)

Fig. 2.1. Diﬀerent contours of the Fourier transform in equation (2.26) for a strike of 90 units.

ψx

(m)

(xt , z, w0 , w, t, T ) =E e

−

=

T t

(m)

r(xs ) ds+ızxT

(2.27)

e

(m)

ızxT

p(xt , xT , w0 , w, t, T ) dxT ,

RM

for all m = 1, . . . , M . In the last equality of equation (2.27), the function p(xt , xT , w0 , w, t, T ) represents the (discounted) transition probability density, starting with an initial state xt and ending up in time T at xT . The continuous discounting is conducted with respect to r (xt∗ ) for t > t∗ ≥ T . Obviously, if the stochastic process consists only of one variable xt , the characteristic function ψ x (xt , z, 0, 0, t, T ) is then just the Fourier Transformation of the particular transition density function p(xt , xT , 0, 0, t, T ). Although the transform operation in equation (2.27) is performed with respect to the (m)

terminal state of one single random variable xT , we have to consider the state of the vector xt as an argument of the characteristic function. In fact,

2.4 The Characteristic Function

25

since we are looking at the overall expectation, equation (2.27) is generally built as the M -dimensional integral over the entire state vector xT 42 . Therefore, we are also able to apply the deﬁnition presented above of building a characteristic function for the more general case g (xT ) = g0 + g xT

(2.28)

with g0 ∈ R and g ∈ RM . This implies, as long as g (xT ) is a linear combination of the elements in xT that only one single transformation variable z necessary. Hence, if we are able to build the characteristic function for the scalar g (xT )43 , there is only a one-dimensional integral for the inverse operation to be performed, independent of the number of state variables included in g (xT ). Note, this powerful result will be used in our multi-factor framework. Equipped with these deﬁnitions we state next some general and important properties of Fourier Transformations on which we rely in our thesis.

Proposition 2.4.3 (Important Properties of Characteristic Functions and Fourier Transformations). Let α, β, x, y ∈ R, and f (x), g(y) some real-valued functions with Fourier transforms fˆ(z), gˆ(z) and Fourier Transformation variable z ∈ C. Then the following relations hold: 1. Linearity: F x [αf (x) + βg(x)] = αfˆ(z) + βˆ g (z). 2. Diﬀerentiation:

Fx

dα f (x) = (ız)α fˆ(z). dxα

3. Convolution: F x [f (x) ∗ g(x)] = fˆ(z)ˆ g (z). 4. Symmetry: ∞ πf (x) =

e

−ızx

0 42

If

(m) xt

0 fˆ(z) dz =

e−ızx fˆ(z) dz.

−∞

would be no subordinated process and independent from all other

state variables, equation (2.27) could still utilize the joint density function 43

p(xt , xT , w0 , w, t, T ) due to the possible discount factor including r(xt ). For example, calculating the general characteristic function for the short rate r (xt ) itself, we set g (xT ) = r (xT ).

26

2 A Multi-Factor Model and Characteristic Functions

5. Relation of the Moment-Generating and the Characteristic Function: dα ψ x (xt , z, 0, 0M , t, T ) . E [xα ] = (−ı)α dz α z=0 Taking a second glance at Figure 2.1, we are able to justify the symmetry of the Fourier Transformation (2.26) of a real-valued function, mentioned in Proposition 2.4.3. Furthermore, one can clearly identify the dampening property of the characteristic function which is essential in developing a numerical algorithm to compute derivative prices. In the following, we show how the characteristic function for a scalar function g (xT ) is derived within the exponential-aﬃne framework. Following Bakshi and Madan (2000), we interpret the characteristic function as a hypothetical contingent claim. Taking more elaborated payoﬀ structures into account, we have to extend the list of permissible arguments for the characteristic function. The more general representation of the characteristic function, which we use hereafter is ψ g(x) (xt , z, w0 , w, g0 , g, t, T ) with the complex-valued payoﬀ representation at maturity T , ψ g(x) (xt , z, w0 , w, g0 , g, T, T ) = eızg(xT ) .

(2.29)

As discussed in the last section, we have to consider that all contingent claims need to be priced under the risk-neutral probability measure Q. Hence, all prices are derived as discounted expectations. Consequently, the discounted expectation of the general form of the terminal condition can be represented as T ψ g(x) (xt , z, w0 , w, g0 , g, t, T ) = EQ e

−

r(xs ) ds+ızg(xT ) t

.

(2.30)

However, we need to compute discounted expectations, e.g. for vanilla zerobond calls, or undiscounted expectations, e.g. in the case of futures instruments. Hence, for futures-style contracts, w0 equals zero and w is a zero valued vector44 . In calculating European derivative prices, we rather need the general characteristic function ψ g(x) (xt , z, w0 , w, g0 , g, t, T ) than the special case of the 44

The characteristic marking to market for standardized futures-style contracts results in the non-existence of a discount factor in the pricing formula and the relevant PDE, respectively, of such a contract under the risk-neutral measure Q.

2.4 The Characteristic Function

27

characteristic function without considering any discount factor, which is just ψ g(x) (xt , z, 0, 0M , g0 , g, t, T ), where 0M represents a M × 1 vector containing exclusively zeros. Applying Theorem 2.2.3 to our hypothetical claim with a solution according to equation (2.30), we take advantage of the Feynman-Kac representation to derive the partial diﬀerential equation. Simplifying and suppressing unnecessary notation, we write henceforth ψ(xt , z, w0 , w, g0 , g, τ ) ≡ ψ g(x) (xt , z, w0 , w, g0 , g, t, T ) and then get the partial diﬀerential equation ∂ψ(xt , z, w0 , w, g0 , g, τ ) ∂ψ(xt , z, w0 , w, g0 , g, τ ) + µQ (xt ) ∂t ∂xt

2 ∂ ψ(x , z, w , w, g , g, τ ) 1 t 0 0 + tr Σ(xt )Σ(xt ) 2 ∂xt ∂xt

(2.31)

+ EJ [ψ(xt , z, w0 , w, g0 , g, J, τ ) − ψ(xt , z, w0 , w, g0 , g, τ )] λQ = ψ(xt , z, w0 , w, g0 , g, τ )r (xt ) , where the complex-valued vector ψ(xt , z, w0 , w, g0 , g, J, τ ) contains all jump components with particular elements (ψ(xt , z, w0 , w, g0 , g, J, τ ))n = ψ(xt + jn , z, w0 , w, g0 , g, τ ). The vector jn ∈ RM contains as mth element the random variable Jmn of the amplitude matrix J. Every contingent claim or function dependent on xt , an arbitrage-free environment presupposed, has to satisfy the same Partial diﬀerential equation structure as given in equation (2.31). For example, the corresponding risk-neutral transition density for the characteristic function ψ(xt , z, w0 , w, w0 , w, τ ), with g (xT ) = r (xT ), which is actually p(r(xt ), r(xT ), w0 , w, t, T ) need to satisfy the same partial diﬀerential equation as the characteristic function itself45 . The only diﬀerence between them would be the particular terminal payoﬀ condition. Hence, solving the above partial diﬀerential equation for p(r(xt ), r(xT ), w0 , w, t, T ), we would impose the Dirac delta function as the relevant terminal condition, having its density mass exclusively concentrated in an inﬁnite spike for r(xT ) at time T . Solving equation (2.31) together with this type of boundary condition can be quite challenging and is in many cases just impossible46 . Thus, it is feasible to ﬁrst solve equation (2.31) for the general characteristic function, with its smooth and continuous boundary function at T , and afterwards do some sort 45 46

See Heston (1993), p. 331. A prominent example is given with the stochastic volatility model of Heston (1993), for which no closed-form representation of the transition density of the underlying equity log-price variable exists.

28

2 A Multi-Factor Model and Characteristic Functions

of normalized integration, the inverse Fourier Transformation, probably in a numerical manner, to get the desired result. Proceeding like this is a very elegant way to ﬁnd some semi-analytic solution. In contrast, if we want to interpret the terminal payoﬀ function in equation (2.29) as a hypothetical futures-style contract, with solution ψ(xt , z, 0, 0M , g0 , g, τ ) = EQ eızg(xT ) , (2.32) we have a slightly diﬀerent partial diﬀerential equation. In this case the dynamic behavior of ψ(xt , z, 0, 0M , g0 , g, τ ) is described by the slightly altered PDE ∂ψ(xt , z, 0, 0M , g0 , g, τ ) ∂ψ(xt , z, 0, 0M , g0 , g, τ ) + µQ (xt ) ∂t ∂xt

2 1 ∂ ψ(xt , z, 0, 0M , g0 , g, τ ) + tr Σ(xt )Σ(xt ) 2 ∂xt ∂xt

+ EJ [ψ(xt , z, 0, 0M , g0 , g, J, τ ) − ψ(xt , z, 0, 0M , g0 , g, τ )] λ

(2.33) Q

= 0, Hence, the only diﬀerence to PDE (2.31) is that the right hand side is now equal to zero to contribute the missing discount rate. Moreover, we can use this futures-style characteristic function ψ(xt , z, 0, 0M , g0 , g, τ ) to obtain the particular values of the undiscounted transition density function. Thus, to compute the probability density function of the short rate r (xt ), we use this futures-style solution of the characteristic function together with the identity g (xt ) = r (xt ). Consequently, using a separation of variables approach, the partial diﬀerential equations in (2.31) and (2.33) can be decoupled into a system of ordinary diﬀerential equations. Therefore, we assume for ψ(xt , z, w0 , w, g0 , g, τ ) the exponential-aﬃne structure

ψ(xt , z, w0 , w, g0 , g, τ ) = ea(z,τ )+b(z,τ ) xt +ızg0 , with the scalar and complex-valued coeﬃcient function a(z, τ ) and (1) (1) ˜b (z, τ ) g (2) ˜(2) g b (z, τ ) ˜ τ ) + ızg, + ız . = b(z, b(z, τ ) = .. . . . ˜b(M) (z, τ ) g (M)

(2.34)

2.4 The Characteristic Function

29

denotes some complex-valued coeﬃcient vector. In the next step we plug the required expressions of the candidate function (2.34) into equation (2.31). Starting with the time derivative, we get ∂ψ(xt , z, w0 , w, g0 , g, τ ) ∂t = − (a(z, τ )τ + b(z, τ )τ xt ) ψ(xt , z, w0 , w, g0 , g, τ ),

(2.35)

where a(z, τ )τ and b(z, τ )τ are the ﬁrst derivatives with respect to the time to maturity variable τ . The gradient vector with respect to the state variables xt is given by ∂ψ(xt , z, w0 , w, g0 , g, τ ) = b(z, τ )ψ(xt , z, w0 , w, g0 , g, τ ), ∂xt

(2.36)

the Hesse matrix is ∂ 2 ψ(xt , z, w0 , w, g0 , g, τ ) = b(z, τ )b(z, τ ) ψ(xt , z, w0 , w, g0 , g, τ ), ∂xt ∂xt

(2.37)

and the jump component in equation (2.31) can be derived as EJ [ψ(xt , z, w0 , w, g0 , g, J, τ ) − ψ(xt , z, w0 , w, g0 , g, τ )] = EJ [ψ ∗ (z, w0 , w, g0 , g, J, τ ) − 1] ψ(xt , z, w0 , w, g0 , g, τ ),

(2.38)

with the normalized vector ψ(xt , z, w0 , w, g0 , g, J, τ ) ψ(xt , z, w0 , w, g0 , g, τ ) b(z,τ ) J 1 e b(z,τ ) J2 e . = .. .

ψ ∗ (z, w0 , w, g0 , g, J, τ ) =

(2.39)

eb(z,τ ) JN

In this aﬃne framework, it can be easily checked that the normalized amplitude vector ψ ∗ (z, w0 , w, g0 , g, J, τ ) is independent of the actual state of xt , which results in the special form given by equation (2.39). Therefore, we are able to express the system of ODEs resulting from equations (2.31) and (2.33), respectively, and the aﬃne form proposed in (2.34) in terms of the risk-neutral coeﬃcients derived in Section 2.3.2. According to Theorem 2.2.3, the ODE which has to be solved for the scalar coeﬃcient a(z, τ ) is then

30

2 A Multi-Factor Model and Characteristic Functions

1 a(z, τ )τ = −w0 + µQ 0 b(z, τ ) + b(z, τ ) Σ0 b(z, τ ) 2 + EJ [ψ ∗ (z, w0 , w, g0 , g, J, τ ) − 1] λQ ,

(2.40)

whereas for the vector coeﬃcient b(z, τ ) we have to solve 1 b(z, τ )τ = −w + µQ 1 b(z, τ ) + b(z, τ ) Σ1 b(z, τ ), 2

(2.41)

with boundary conditions a(z, 0) = 0 and b(z, 0) = ızg, respectively. The parameters w0 and w, determine whether we consider a discount rate or not for the characteristic function. The mth element of b(z, τ ) Σ1 b(z, τ ) can be computed as i,j b(z, τ )i (Σ1 )ijm b(z, τ )j 47 . Moreover, we want to emphasize that the trace operator is circular, meaning the equality tr [Σ(xt )Σ(xt ) b(z, τ )b(z, τ ) ] = tr [b(z, τ ) Σ(xt )Σ(xt ) b(z, τ )]

(2.42)

holds. Obviously, the right hand side of this last equation represents a scalar and therefore we are able to neglect the trace operator in equation (2.40) and equation (2.41), respectively. In order to calculate derivatives prices, the coeﬃcients a(z, τ ) and b(z, τ ) need not exhibit closed-form solutions in any case. There are several scenarios conceivable, e.g. the time integrated expectations of the jump amplitudes have no closed-form representations, or the processes themselves have such complicated structures that there simply does not exist a closed-form solution of the coeﬃcients a(z, τ ) or b(z, τ ) of the characteristic function. However, if we are able to represent a(z, τ ) and b(z, τ ) in terms of their ordinary diﬀerential equations (2.40) and (2.41), solutions can be eﬃciently obtained via a Runge-Kutta solver and appropriately integrated within our numerical pricing procedure, such that time consuming Monte-Carlo studies for the pricing of European interest-rate derivatives can be avoided.

47

See Duﬃe, Pan and Singleton (2000), p. 1351.

3 Theoretical Prices of European Interest-Rate Derivatives

3.1 Overview In this section, we want to give a representative selection of diﬀerent interestrate contracts for which the pricing framework used in this thesis is able to produce semi closed-form solutions48 . In doing this we distinguish, for didactical purposes, between contracts based on the short rate r(xt ) and contracts based on a simple yield Y (xt , t, T ) over a speciﬁed time period τ . These yields to maturity are often referred to as simple compound rates, e.g. LIBOR rates, and denote the constant compounding of wealth over a ﬁxed period of time τ , which is related to a zero bond with corresponding time to maturity. Deﬁnition 3.1.1 (Simply-Compounded Yield to Maturity). The simple yield to maturity Y (xt , t, T ) of a zero bond P (xt , t, T ), maturing after the time period τ , is deﬁned through the equality 1 = P (xt , t, T ) . 1 + τ Y (xt , t, T )

(3.1)

Therefore the simple yield to maturity can be derived as −1

Y (xt , t, T ) =

P (xt , t, T ) τ

−1

=

1 − P (xt , t, T ) . τ P (xt , t, T )

(3.2)

In the following sections, we generally distinguish in the derivation of theoretical prices of contingent claims between contracts based on the instantaneous interest rate r(xt ) and contracts depending on the simple yield 48

A comprehensive summary of diﬀerent valuation formulae of ﬁxed-income securities is given, e.g. Brigo and Mercurio (2001) and Musiela and Rutkowski (2005).

32

3 Theoretical Prices of European Interest-Rate Derivatives

Y (xt , t, T ). Moreover, we diﬀerentiate between contracts with unconditional and conditional exercise rights. This distinction is introduced because of the diﬀerent mathematical derivation of the particular model prices. For contracts with unconditional exercise, we obtain pricing formulae, which bear strong resemblance to moment-generating functions of the particular underlying state process whereas contracts with conditional exercise rights, i.e. option contracts, need an explicit integration due to the natural exercise boundary. All derivative prices for which we derive the corresponding pricing formulae are European-style derivatives, meaning that the exercise can only be performed at maturity T .

3.2 Derivatives with Unconditional Payoﬀ Functions This derivatives class is characterized by the trivial exercise of the contract at maturity. This means that the contract is always exercised, no matter if the holder suﬀers a loss or make a proﬁt as consequence of the exercise. Although trivially exercised, a zero-coupon bond is a special case of this class since it pays at maturity a predeﬁned riskless quantity of monetary units. Deﬁnition 3.2.1 (Zero-Coupon Bond). A zero-coupon bond maturing at time T guarantees its holder the payment of one monetary unit at maturity. The value of this contract at t < T is then denoted as P (xt , t, T ), which is the expected value of the discounted terminal condition G(xT ) = 1. This can be formally expressed as, T P (xt , t, T ) = EQ e

−

r(xs ) ds t

(3.3)

It is easily seen that the payoﬀ function G (xT ) used in equation (3.3) is independent both of the time variable and the state variables in xT . Using the formal deﬁnition in equation (3.3), a zero-coupon bond, or as shorthand a zero bond, is just the present value of one monetary unit paid at time T . Hence, we are able to interpret P (xt , t, T ) as the expected discount factor relevant for the time period t up to T . Due to this intuitive interpretation, these contracts are often used in calibrating interest-rate models to empirical data sets.

3.2 Derivatives with Unconditional Payoﬀ Functions

33

A slightly more elaborated contract is given by the combination of certain payments at diﬀerent times. We denote this contract then as a coupon-bearing bond.

Deﬁnition 3.2.2 (Coupon-Bearing Bond). A coupon-bearing bond guarantees its holder a number of A deterministic payments ca ∈ c at speciﬁc coupon dates Ta ∈ T for a = 1, . . . , A. Typically, at maturity TA , a nominal face value C is included in cA in addition to the ordinary coupon. The present value of a coupon bond CB(xt , c, t, T) is then given as T a A A − r(xs ) ds EQ e t ca = P (xt , t, Ta ) ca . CB(xt , c, t, T) = a=1

(3.4)

a=1

Obviously, a coupon-bearing bond, or as shorthand a coupon bond, is just the cumulation of payments ca discounted with the particular zero-bond prices P (xt , t, Ta ). If a ﬁrm is requiring a hedge position for a risk exposure in the form of a future payment of interest, due to an uncertain ﬂoating interest rate, we are able to conclude a forward-rate agreement.

Deﬁnition 3.2.3 (Forward-Rate Agreement). A forward-rate agreement concluded in time t guarantees its holder the right to exchange his variable interest payments to a ﬁxed rate K, scaled upon a notional principal N om. The contract is sold in t. The interest payments exchanged relate then to the time period, say [T, Tˆ] with t < T < Tˆ . We distinguish the cases, where the forward-rate agreement refers to the short rate r (xt ) and to the yield Y (xt , t, T ). Hence, for a contract based on the short rate, the relevant time interval is then [T, Tˆ ] = [T, T + dT ]. The price of this contract is given as F RAr (xt , K, N om, t, T ) T = EQ e

−

rs ds t

(K − r (xT )) N om

= K P (xt , t, T ) − EQ e

−

T

rs ds t

r (xT ) N om.

(3.5)

34

3 Theoretical Prices of European Interest-Rate Derivatives

The price for a forward-rate agreement over a discrete time period of length τˆ = Tˆ − T , written on a yield Y xT , T, Tˆ and paid in arrears, can be represented as49 F RAY (xt ,K, N om, t, T, Tˆ ) Tˆ = τˆEQ e

−

= EQ e

−

= E e

t

−

−1 + 1 N om τˆK − P xT , T, Tˆ

r(xs ) ds

τ K + 1) − 1 N om P xT , T, Tˆ (ˆ

T t

K − Y xT , T, Tˆ N om

r(xs ) ds

ˆ T t

Q

r(xs ) ds

(3.6)

N om ˜ = EQ e t P xT , T, Tˆ − K ˜ K ˜ (xt , t, T ) N om , = P xt , t, Tˆ − KP ˜ K −

˜ = with K

T

r(xs ) ds

1 τˆK+1 .

To give a more illustrative example, we consider a ﬁrm, which has to make a future payment subject to an uncertain, ﬂoating rate of interest. Reducing the immanent interest-rate risk exposure, this ﬁrm wants to transform this payment into a certain cash-ﬂow, locked at a ﬁxed rate K. This can be achieved by contracting a forward-rate agreement, therefore exchanging the ﬂoating interest rate to the ﬁxed rate K. Thus, the ﬁrm is, in its future calculation, independent of the evolution of the term structure. 49

Here we use the fact that the exponential-aﬃne model exhibits the Markov ÊTˆ

ability. Thus, the expectation E

Q

e− t r(xs ) ds

resented as the iterated expectation E EQ e

−

ÊT t

r(xs ) ds

to time T .

P xT , T, Tˆ

Q

e

−

ÊT t

= P xt , t, Tˆ

r(xs ) ds

E

QT

e

−

ÊTˆ T

can be repr(xs ) ds

=

, where the inner expectation is made with respect

3.2 Derivatives with Unconditional Payoﬀ Functions

35

Another point, we want to mention is the special strike value K = KF RA for which the yield-based forward-rate agreement becomes a fair zero value at time t. This value is commonly referred to as the forward rate and corresponds then to the simply-compounded rate Tˆ EQ e

−

r(xs ) ds t

−1 −1 P xT , T, Tˆ

τˆP xt , t, Tˆ P (xt , t, T ) − P xt , t, Tˆ = τˆP xt , t, Tˆ 1 P (xt , t, T ) − 1 . = τˆ P xt , t, Tˆ

KF RA =

(3.7)

Most of the time a ﬁrm does not want to insure itself against a ﬂoating interest payment for only one time period. For example, the ﬁrm has to serve a debt contract, which is linked to a LIBOR interest rate. In this case, the ﬁrm possibly wants to reduce its risk exposure due to the ﬂoating interest accrues over time and it is desired to make an exchange of interest payments for several successive time periods, where in each period the payment for the relevant ﬂoating rate is exchanged with a ﬁxed rate K. This task can be achieved buying a receiver swap contract.

Deﬁnition 3.2.4 (Swap). A forward-starting interest-rate receiver swap is deﬁned as a portfolio of forward-rate agreements for diﬀerent time periods Ta+1 − Ta with Ta ∈ T and t < Ta for a = 1, . . . , A on the same strike rate K. The payments of the contract are made at dates T2 , . . . , TA , whereas the contract is said to reset the ﬂoating rate at dates T1 , . . . , TA−1 . Due to the instantaneous character of the ﬂoating rate based swap contract, the payment and reset dates coincide. Hence, the swap contract in this case, with nominal principal N om and A payment dates contained in the vector T, can be represented as

36

3 Theoretical Prices of European Interest-Rate Derivatives

SW Ar (xt ,K, N om, t, T) Ta A − r(xs ) ds = EQ e t (K − r (xTa )) N om a=1

= N om

A

EQ e

−

Ta

r(xs ) ds

t

(K − r (xTa ))

a=1

= N om

K

A

P (xt , t, Ta )

a=1

−

A

(3.8)

EQ e

−

Ta t

r(xs ) ds

r (xTa ) .

a=1

The equivalent representation for a swap contract, exchanging a yield-based ﬂoating rate at A − 1 payment dates paid in-arrears is then SW AY (xt , K, N om, t, T) Ta+1 A−1 − r(xs ) ds = EQ e t (K − Y (xTa , Ta , Ta+1 )) τˆa+1 N om a=1

= N om× A−1

EQ e

−

Ta

r(xs ) ds

t

((K τˆa+1 + 1)P (xTa , Ta , Ta+1 ) − 1) (3.9)

a=1

= N om

A−1

(K τˆa+1 + 1) P (xt , t, Ta+1 ) − P (xt , t, Ta )

a=1

= N om

P (xt , t, TA ) − P (xt , t, T1 ) +K

A−1

τˆa+1 P (xt , t, Ta+1 ) ,

a=1

with τˆa+1 = Ta+1 − Ta . In contrast to the total number of A swap payments in equation (3.8), where these payments refer merely to speciﬁc time dates, for the yield-based swap contracts we have to consider A − 1 time periods, which explains the resulting summation term in equation (3.9). Subsequently, a swap contract can be interpreted as the sum of successive forward-rate agreements.

3.2 Derivatives with Unconditional Payoﬀ Functions

37

Similar to forward-rate agreements we are able to introduce the terminology of a special strike KS , which makes the yield-based swap contract a fair zero valued contract. This special strike is then denoted as the swap rate and can be represented in the case of a yield-based swap as A−1 (P (xt , t, Ta ) − P (xt , t, Ta+1 )) KS = a=1 A−1 ˆa+1 P (xt , t, Ta+1 ) a=1 τ P (xt , t, T1 ) − P (xt , t, TA ) = A−1 . ˆa+1 P (xt , t, Ta+1 ) a=1 τ

(3.10)

The last contract with unconditional exercise right which we include in the pricing methodology used is an Asian-type average-rate contract based on the ﬂoating rate r (xt ). These contracts do not belong to the class of traded derivatives in any exchange. However, this type of interest-rate derivative seems to be quite popular in over-the-counter markets50 . Asian contracts belong to the ﬁeld of path-dependent derivatives. Thus, the payoﬀ consists not only of the terminal value of the underlying rate at maturity but of the complete sample path over the averaging period. Deﬁnition 3.2.5 (Unconditional Average-Rate Contract). An unconditional average-rate agreement concluded in time t guarantees its holder the right at maturity T to exchange the continuously measured average of the ﬂoating rate r (xt ) over the period T − t against a ﬁxed strike rate K. The value of this diﬀerence is then scaled by a nominal principal N om. Hence, the price of this contract is given as U ARCr (xt , K, N om, t, T ) T =EQ e

−

r(xs ) ds t

K − 1 T −t

T

r(xs ) ds N om

t

T

− 1 =N om P (xt , t, T ) K − EQ e t τ

r(xs ) ds

T

(3.11)

r(xs ) ds .

t

Consequently, in contrast to the forward-rate agreement according to equation (3.5), where the sole expectation of r(xT ) played the major part, we are 50

See Ju (1997).

38

3 Theoretical Prices of European Interest-Rate Derivatives

interested in the discounted expectation of the integral of r(xt ) over the time to maturity at this point.

3.3 Derivatives with Conditional Payoﬀ Functions In the last subsection, we considered the pricing formulae for contracts with unconditional exercise at maturity under the risk-neutral measure Q. Obviously, these contracts can be expressed e.g. in terms of zero bonds and some constants. In this section we want to derive general pricing formulae for contracts with conditional or optional exercise rights at maturity. These derivatives contracts are therefore often referred to as option contracts. Basically, we are interested in calculating the particular option prices with underlying contracts of the form (3.5), (3.6), and (3.9) with optional exercise rights. Basically, the particular pricing formulae can be separated into zero bond and coupon-bond options, respectively, can be seen as a portfolio of several zerobond options in case of a yield-based swap contract. Hence, we begin the introduction with option contracts written on a zero bond.

Deﬁnition 3.3.1 (Zero Bond Option). We deﬁne a zero-bond call (put) option as a contract giving its holder the right, not the obligation, to buy (sell) a zero bond P xt , t, Tˆ for a strike price K at time T . The remaining time to maturity of this zero bond at the exercise date of the option is then given as τˆ. Formally, the price of a zero-bond call can be obtained as ZBC xt , K, t, T, Tˆ T − r(xs ) ds max P xT , T, Tˆ − K, 0 = EQ e t (3.12) T + − r(xs ) ds P xT , T, Tˆ − K , = EQ e t whereas a zero-bond put option can be calculated as T + − r(xs ) ds . ZBP xt , K, t, T, Tˆ = EQ e t K − P xT , T, Tˆ

(3.13)

3.3 Derivatives with Conditional Payoﬀ Functions

39

Zero bond options can be used to price two contracts commonly used to hedge interest-rate risk. Namely, we want to introduce cap and ﬂoor contracts. In this terminology, a cap contract is meant to hedge upside interest-rate risk exposure. This is often required for a ﬁrm which holds some debt position with interest payments on a ﬂoating rate base and fears that future interest rates are rising. So it wants the interest rate capped at some ﬁxed level, in order to limit its risk position due to this ﬁxed rate. In contrast to the above introduced forward-rate agreement or swap, a ﬁrm can now both participate on advantageously low interest rates and simultaneously cap its interest payments against high rates. The opposite eﬀect can be observed, if an institution or ﬁrm has outstanding loans based on a ﬂoating rate. In this case the ﬁrm is interested in limiting the downside risk, since low ﬂoating rates correspond to low interest payments. The contract with the desired properties is then a ﬂoor, where interest payments are exchanged under an agreed ﬁxed rate.

Deﬁnition 3.3.2 (Cap and Floor Contract). A cap (ﬂoor) contract is deﬁned as a portfolio of caplets (ﬂoorlets) for diﬀerent time periods Ta+1 − Ta with Ta ∈ T and t < Ta for a = 1, . . . , A on the same strike rate K. The payments of the contract are made at dates T2 , . . . , TA , whereas the contract is said to reset the ﬂoating rate at dates T1 , . . . , TA−1 . Due to the short rate, the character of the ﬂoating rate based swap contract, the payment and reset dates coincide. Hence, the model price of a caplet with nominal principal N om and A payment dates contained within the vector T, is then given by CP Lr (xt , K, N om, t, Ta ) = EQ e

−

Ta

r(xs ) ds

t

+ (r (xTa ) − K) N om. (3.14)

The price of a cap contract, as a simple summation of caplets for diﬀerent times Ta ∈ T, can then be represented as CAPr (xt , K, N om, t, T) =

A

CP Lr (xt , K, N om, t, Ta )

a=1

= N om

A a=1

EQ e

−

Ta t

r(xs ) ds

(r (xTa ) − K)+ .

(3.15)

40

3 Theoretical Prices of European Interest-Rate Derivatives

Subsequently, we have for a ﬂoor the pricing formula F LRr (xt , K, N om, t, T) T a A − r(xs ) ds + = N om EQ e t (K − r (xTa )) .

(3.16)

a=1

The particular yield-based cap and ﬂoor options, exchanging, if exercised, arbitrary yields with a ﬁxed rate K at A − 1 payment dates, are given by CAPY (xt , K, N om, t, T) T a A−1 + N om − r(xs ) ds ˜ a − P (xTa , Ta , Ta+1 ) K = EQ e t ˜a K a=1 =

A−1 a=1

(3.17)

˜ a , t, Ta , Ta+1 N om , ZBP xt , K ˜a K

and F LRY (xt , K, N om, t, T) T a A−1 + N om − r(xs ) ds ˜a P (xTa , Ta , Ta+1 ) − K = EQ e t ˜a K a=1

= ˜a = with K

(3.18)

˜ a , t, Ta , Ta+1 N om , ZBC xt , K ˜a K a=1

A−1

1 τˆa+1 K+1 .

Deﬁnition 3.3.2 shows that a cap or ﬂoor contract is just the summation of their legs, the caplets and ﬂoorlets, respectively. Especially for the more realistic case of yield-based contracts, we can identify the similarity to zerobond options, since contract prices can be obtained as the summation of these options. The yield-based options are said to be at the money if the modiﬁed strike ¯ a is equal to equation (3.10). A cap is therefore in the money if the rate K ¯ a > KS it is out of the money. modiﬁed strike rate is less than KS , and for K The opposite results hold for a ﬂoor contract. Furthermore, we can conclude that holding a cap contract long and a ﬂoor contract short, both with the

3.3 Derivatives with Conditional Payoﬀ Functions

41

same contract speciﬁcations, we are able to replicate a swap contract. This can be easily justiﬁed comparing the payoﬀ of such a portfolio given for a yield Y (ˆ τa+1 ), which is then (Y (xTa , Ta , Ta+1 ) − K)+ − (K − Y (xTa , Ta , Ta+1 ))+

(3.19)

= Y (xTa , Ta , Ta+1 ) − K,

and the corresponding swap payment. Taking the discounted expectation of the sum of terms in equation (3.19) for all periods, we have the equivalent swap contract. A more challenging contract in calculating model prices is a coupon-bond option. This option is only exercised if the coupon-bond price at maturity exceeds the strike K. Hence, we have to apply the maximum operator to the discounted sum of all outstanding coupon payments and the strike price. This is in contrast to the other option contracts mentioned above, where we applied the maximum operator to each term of the sum separately. Deﬁnition 3.3.3 (Coupon-Bond Option). A coupon-bond call (put) option is deﬁned as the right but not the obligation to buy (sell) a coupon bond CB(xT , c, t, T) with payment dates Ta ∈ T, with Ta > T for a = 1, . . . , A and strike price K. The price of a coupon-bond call option is given by T CBC (xt , c, K, t, T, T) = EQ e = EQ e

−

T

r(xs ) ds

A

t

−

r(xs ) ds t

(CB (xT , c, T, T) − K)+ +

P (xT , T, Ta ) ca − K

(3.20)

,

a=1

and the corresponding coupon-bond put option is given by T CBP (xt , c, K, t, T, T) = EQ e = EQ e

−

T

r(xs ) ds t

K−

−

r(xs ) ds t

A

(K − CB(xT , c, T, T))+

+ . P (xT , T, Ta )ca

(3.21)

a=1

Since the maximum operator is not distributive with respect to sums, the term inside the maximum operator in equation (3.20) and (3.21) cannot be

42

3 Theoretical Prices of European Interest-Rate Derivatives

decomposed easily without making further assumptions. Another popular option we want to discuss is an option on a swap contract or as shorthand often referred to as a swaption. With a swaption one can choose at the maturity of the option if it is advantageous to enter the underlying swap contract or otherwise leave the option unexercised. Deﬁnition 3.3.4 (Swaption). We deﬁne a forward-starting swaption as a contract conferring the right, but not the obligation to enter a forward starting receiver swap at maturity T . The particular underlying receiver swap contract is deﬁned according to deﬁnition 3.2.4, with T1 ≥ T . Formally, the yield-based forward-starting receiver swaption for an underlying swap with A − 1 payment periods is given as SW PY (xt , K, N om, t, T, T) T = EQ e Q

=E

e

−

r(xs ) ds t

−

+ (SW AY (xT , K, N om, T, T))

T

r(xs ) ds t

K

A−1

(3.22)

P (xT , T, Ta+1 )ˆ τa+1

a=1

+

+ P (xT , T, TA ) − P (xT , T, T1 )

N om.

Typically, the swaption maturity coincides with the ﬁrst reset date of the underlying swap contract. Thus, a yield-based receiver swaption with T1 = T , can be equivalently represented as a coupon-bond call option SW PY (xt , K, N om, t, T1 , T∗ ) = CBC (xt , cSW P , 1, t, T1 , T∗ ) , with

cSW P

and new time dates

K τˆ2 K τˆ3 × N om, = .. . 1 + K τˆA

T2

T3 T∗ = .. . . TA

(3.23)

3.3 Derivatives with Conditional Payoﬀ Functions

43

Subsequently, we reduce the valuation problem of a swaption to the calculation of an equivalent coupon-bond option with strike one, a coupon vector cSW P and a vector with payment dates T∗ . According to the unconditional contract deﬁned in equation (3.11), we are also able to price an average-rate option contract. The deﬁnition of the model price of an average-rate option is given below.

Deﬁnition 3.3.5 (Average-Rate Option). An average-rate cap option gives its holder the right, but not the obligation to exchange at expiration a ﬁxed strike rate K, over the period T − t, against the continuously measured average of the short rate r (xt ). Formally, the price of an average-rate cap option can be obtained as ARCr (xt , K, N om, t, T ) + T T − r(xs ) ds 1 r (xs ) ds − K N om. =EQ e t τ

(3.24)

t

Consequently, we have for an average-rate ﬂoor the pricing formula ARFr (xt , K, N om, t, T ) + T T 1 − r(xs ) ds K− r (xs ) ds N om. =EQ e t τ

(3.25)

t

Asian options show the advantageous ability to exhibit reduced risk positions in comparison to ordinary options because of the time-averaging of the underlying price process. Moreover, asian option contracts are more robust against price manipulations since the option payoﬀ includes the sample path over a ﬁnite time period. These options are not standard instruments traded on exchanges. However, they are popular over-the-counter contracts used by banks and corporations to hedge their interest-rate risk over a time period51 . For all theoretical option prices presented in this section, we give in Section 5.3 the corresponding pricing formulae which have to be used in a numerical 51

See, for example, Ju (1997).

44

3 Theoretical Prices of European Interest-Rate Derivatives

scheme. Thus, we distinguish between the calculation of a portfolio of options, e.g. used for the pricing of cap and ﬂoor contracts and as a special case for zero-bond options, respectively, and the computation of options on a portfolio which is the case for coupon-bond options and swaption contracts. This is done because only in case of a one-factor interest-rate process semi closedform solutions for swaptions and coupon bonds can be calculated.

4 Three Fourier Transform-Based Pricing Approaches

4.1 Overview Interest-rate derivatives are widely used instruments to cover possible interestrate risk exposures. However, to model the term structure more realistically, sophisticated models are required. One way to enhance the capability of the term-structure model is to incorporate more stochastic factors, by, for instance, incorporating a stochastic mean and/or a stochastic volatility, or modeling the term structure with help of an additive interest-rate process. Another way, which would especially enrich the model with the ability to reﬂect price shocks, lies in implementing jump components in the shape of diﬀerent Poisson processes with arbitrary stochastic jump amplitudes. Unfortunately, in most cases the pricing of derivatives securities, while incorporating for the underlying interest-rate process both features mentioned above, can only be accomplished with ineﬃcient Monte-Carlo simulations. Hence, more eﬃcient methods are needed to circumvent these time-consuming calculations. As shown in the prominent work of Heston (1993), a way out of this dilemma is achieved by using Fourier Transformation techniques. Doing this, we only need to solve one standardized inversion integral to evaluate the distribution function and then compute the desired derivative prices. The astonishing fact of the approach applied by Heston (1993) is that this Fourier-based valuation technique is independent of the underlying stochastic dynamics of the shortrate process and can be applied as long as the particular characteristic function

46

4 Three Fourier Transform-Based Pricing Approaches

exists52 . Bakshi and Madan (2000) generalized this method to interpret the characteristic function itself as a derivative contract with a trigonometric payoﬀ53 . Zhu (2000) derived various pricing formulae for options with underlying stock prices, where stochastic interest rates, volatilities and jumps were included in a modularized manner. There, the stochastic factors are integrated by parts and the author ends up with a system of ordinary diﬀerential equations, which then has to be solved. In this thesis, we go a step further and, by using the transform methods of Lewis (2001), are able to generalize the modular aspect of Fourier-based derivatives pricing into parts of the underlying stochastic behavior and the contract type. This enables us to present valuation techniques, which can be adapted to every desired European-style contract without greater eﬀort, assuming that the generalized Fourier Transformation of the payoﬀ function exists in closed form. We consider the general exponential-aﬃne model introduced in Section 2.1 for the short rate r (xt ) and derive a ﬂexible valuation procedure according to the approach given in Lewis (2001). Although we focus in our thesis on the exponential-aﬃne setup, we are also able to extend the framework to incorporate non-aﬃne term-structure models54 , such as the Longstaﬀ (1989) model or the class of quadratic Gaussian models as discussed in Beaglehole and Tenney (1992)55 and Filipovic (2001), respectively. All we need in the underlying model speciﬁcation is the exponential separability of the coeﬃcients in the general characteristic function. However, in applying these non-aﬃne model speciﬁcations, we have to ignore the possibility of jumps for non-aﬃne factors in order to avoid mixture terms in the fundamental partial diﬀerential equation, which would subsequently render the pricing procedure unattainable56 . 52

53 54 55

56

Due to our pricing framework we can relax this restriction to the existence of a system of ordinary diﬀerential equations. This methodology is covered in Section 4.2. See Chapter 10. In fact, the model of Longstaﬀ (1989) can be represented as a quadratic Gaussian model as shown in Beaglehole and Tenney (1992). The same holds for the term-structure model in Cheng and Scaillet (2004) where the terminology of a linear-quadratic jump-diﬀusion model is introduced. Despite the name, jump parts are only valid for linear factors, whereas the quadratic part is not allowed to bear jump parts. This issue is discussed in Section 9.3.

4.1 Overview

47

The outline of this chapter is as follows. We start with the comparison of three state of the art Fourier Transformation methodologies used in derivatives research. The Fourier-transformed Arrow-Debreu securities pricing approach is based on the work of Heston (1993)57 . Afterwards, we present the transform methodology as proposed by Carr and Madan (1999) and then discuss the generalized derivatives pricing setup of Lewis (2001), which display similarities in the derivation of the model price of a contingent claim. Both approaches focus on the Fourier Transformation of the payoﬀ function, whereas Carr and Madan (1999) apply the transform for the strike value, Lewis (2001) does a Fourier Transformation with respect to the state variable. Nevertheless, we provide an extension of the work in Lewis (2001), since we consider a multi-factor environment. One important diﬀerence between the pricing approach utilizing Fourier-transformed Arrow-Debreu securities, according to Heston (1993), the Carr and Madan (1999) methodology, and the method of Lewis (2001) is that the latter two approaches do not need to invoke Fourier Transformations for every single term in the pricing formula. Therefore, the transformation is applied on the entire contingent claim, which in a numerical sense is more efﬁcient. Additionally, the these two approaches provide a more stable solution due to the freedom of choosing a contour path for the integration parallel to the real axis in the inversion formulae58 . Generally, the derivatives we want to price are written on some functional of the underlying stochastic vector process xt , say g(xt ). Contingent claims on the short rate and on the yield are European-style derivatives and therefore pay only at maturity T a payoﬀ G (xT ). The solution of the pricing problems we seek then takes the following form.

Deﬁnition 4.1.1 (General Valuation Problem for European-Style Derivatives). We deﬁne the general valuation problem of a contract V (xt , t, T ) as the time T expectation of some (discounted) payoﬀ function G (xT ) under the risk-neutral probability measure Q, formally deﬁned as 57

Recent work with further development and uniﬁcation was made in Duﬃe, Pan and Singleton (2000), Bakshi and Madan (2000) and especially on the ﬁeld of

58

interest-rate derivatives in Chacko and Das (2002). See Carr and Madan (1999) and Lewis (2001).

48

4 Three Fourier Transform-Based Pricing Approaches

V (xt , t, T ) = E e Q

−

T

r(xs ) ds t

G (xT )

=

(4.1) G (xT ) p(xt , xT , w0 , w, t, T ) dxT .

RM

The contract can only be exercised at maturity T .

Apart from the underlying stochastic dynamics, the solution to equation (4.1) depends on how xT is incorporated within the payoﬀ function G (xT ). Thus, we follow Chacko and Das (2002) and distinguish for didactical purposes between payoﬀ functions which can be either linear, exponential-linear or integro-linear in xt . These idealized payoﬀ types are illustrated in Table 4.1 below59 .

Table 4.1. Idealized call option payoﬀ functions Payoﬀ type

G (xT )

Linear

G (xT ) = (g (xT ) − K)+

Exponential-linear

Integro-linear

G (xT ) = eg(xT ) − K

G (xT ) =

T t

+

g (xs ) ds − K

+

In contrast to option-pricing models written on equities, where constant interest rates are often assumed, in calculating equation (4.1), we are confronted with a more diﬃcult situation. Since both the discount factor and the payoﬀ function G (xT ) depend on the same stochastic process, we are not able to evaluate these expectations separately and multiply them afterwards60. We have 59 60

In case of unconditional payoﬀ functions, we use the same classiﬁcation. This is a direct consequence of the choice of numeraire made in Section 2.3.

4.2 Heston Approach

49

to consider that both expressions are obviously not independent and therefore have to derive the solution of equation (4.1) under their joint stochastic dynamics. However, thanks to the fact that the discount factor itself has an exponential-aﬃne representation61, we are still able to use the general characteristic function ψ(xt , z, w0 , w, g0 , g, τ ) in derivatives pricing. Consequently, equation (4.1) is the starting point for all of the following derivatives pricing approaches.

4.2 Heston Approach Pricing derivatives, using Fourier-transformed Arrow-Debreu securities and state prices, respectively, was introduced in Heston (1993). Since then, several articles utilizing Fourier Transformations in derivatives pricing have been published. Among others we want to mention, because of their relevance, Duﬃe, Pan and Singleton (2000) and Bakshi and Madan (2000). In the article of Duﬃe, Pan and Singleton (2000), a comprehensive survey is provided as to how this Fourier inversion methodology can be used to solve derivative prices for general stochastic dynamics. On the other hand, Bakshi and Madan (2000) oﬀer a rigorous survey, of how Fourier-transformed Arrow-Debreu securities can be used to span the underlying market and to price derivative prices. In principle, both articles use the same pricing mechanism, shown below62 . The basic principle behind the pricing approach with transformed ArrowDebreu securities is that all derivatives based on the interest rate r(xt ) described by equation (4.1) have to solve the same partial diﬀerential equations (2.31) and (2.33) for futures-style contracts, respectively. The only diﬀerence between them is that they need to satisfy diﬀerent terminal conditions. This statement holds also for the discounted probability density and the characteristic function of the interest-rate process. Therefore, they can be interpreted as hypothetical contingent claims solving the above-mentioned partial diﬀerential equations. Whereas derivative prices and probability densities are often 61

62

One can easily validate this statement by solving equation (2.30) and (2.34) and setting z equal to zero. In the context of interest-rate derivatives, Chacko and Das (2002) used this methodology to price the diﬀerent payoﬀ structures as given in Table 4.1.

50

4 Three Fourier Transform-Based Pricing Approaches

hard to obtain, due to their discontinuous terminal conditions63 , the solution for the particular general characteristic function can be recovered, even if jump components are encountered in the stochastic vector process xt . This is due to a special ability of characteristic functions; their terminal condition is inﬁnitely diﬀerentiable and smooth, which make them, from a mathematical point of view, more tractable.

Deﬁnition 4.2.1 (Arrow-Debreu Security). We deﬁne an Arrow-Debreu security as a contingent claim paying one unit of money at maturity T if and only if a speciﬁed state A occurs. The value AD(xt , t, T ) of an Arrow-Debreu security under probability measure Q∗ at time t is then given by AD(xt , t, T ) = EQ∗ [1A ] .

(4.2)

The expression 1A denotes the indicator function for the event A in time T , which is unity if the state A occurs and zero otherwise.

To demonstrate the pricing methodology, we consider the following example of a European call option with a linear payoﬀ function G (xT ) = (g (xT ) − K)+ and g (xT ) is given in equation (2.28)64 . The solution for this option can then be represented as V (xt , t, T ) = EQ e

−

r(xs ) ds t

= EQ e

T

−

T

r(xs ) ds t

− KEQ e 63

64

+ (g (xT ) − K)

−

g (xT ) 1g(xT )≥K

T

r(xs ) ds t

(4.3)

1g(xT )≥K ,

For many underlying stochastic dynamics, the solutions cannot be calculated in closed form. The derivation of option-pricing formulae for exponential-linear and integro-linear payoﬀ structures diﬀers slightly from the derivation of the theoretical option price formula of a linear payoﬀ function as given in this section. The derivation of the particular solutions for these payoﬀ functions can be looked up in Chacko and Das (2002), Sections 2 and 3.

4.2 Heston Approach

51

where the expectation is separated into parts. However, the expectations in equation (4.3) are not yet Arrow-Debreu securities in the sense of deﬁnition 4.2.1. These expressions still lack some sort of standardization to guarantee the outcome of one monetary unit. Thus, we need to apply the unconditional expectations65 T T EQ e

−

r(xs ) ds t

g (xT )

EQ e

and

−

r(xs ) ds t

= P (xt , t, T ).

Expanding the terms in equation (4.3) with their particular unconditional counterparts, we get

V (xt , t, T ) = EQ e

−

T

r(xs ) ds t

g (xT ) ×

− r(x ) ds s t g (x ) 1 e T g(xT )≥K Q T E EQ e− t r(xs ) ds g (xT ) T

e − KP (xt , t, T )EQ

−

(4.4)

T

r(xs ) ds t

1g(xT )≥K P (xt , t, T )

= Π0 (xt , t, T )Π1 (xt , t, T ) − KP (xt , t, T )Π2 (xt , t, T ). Obviously, the normalized functions Π1 (xt , t, T ) and Π2 (xt , t, T ) are two contingent claims and can be interpreted as Arrow-Debreu securities66 . On the other hand, Π1 (xt , t, T ) can be interpreted as the discounted forward price of the underlying contract. Introducing two artiﬁcial changes of measure deﬁned through the Radon-Nikodym derivatives, we get −

T

−

r(xs ) ds

dQ1 e t g(xT ) = dQ Π0 (xt , t, T )

and

T

r(xs ) ds

dQ2 e t = . dQ P (xt , t, T )

Consequently, we express the above call option price in terms of the particular Arrow-Debreu prices, which is 65 66

See Chacko and Das (2002), p. 205. In the last equation of (4.4), we adopted the notation given in Chacko and Das (2002).

52

4 Three Fourier Transform-Based Pricing Approaches

V (xt , t, T ) =Π0 (xt , t, T )EQ1 1g(xT )≥K − KP (xt , t, T )EQ2 1g(xT )≥K .

(4.5)

Obviously, in calculating the option price in equation (4.5), we need only the general characteristic function with terminal condition eızg(xT ) and its derivative with respect to z, respectively. However, calculations within this pricing framework for the particular functions Πi (xt , t, T ) are quite diﬀerent for linear, exponential-linear and integro-linear payoﬀ versions of G (xT )67 . Thus, only P (xt , t, T ) remains unchanged, since this quantity is completely independent of the characteristic payoﬀ part g (xT ). Recalling the formal structure of the general characteristic function in (2.30) and the connection between the moment-generating and characteristic function68 , we are able to express Π0 (xt , t, T ) with the help of the derivative of the general characteristic function with respect to the frequency parameter z, evaluated at z = 0, which is given by69 T Π0 (xt , t, T ) = EQ e =

−

r(xs ) ds t

1 d Q E e ı dz

−

g (xT ) T

r(xs ) ds t

ızg(xT ) e

(4.6) z=0

ψz (xt , 0, w0 , w, g0 , g, τ ) . = ı Here, the subscript denotes partial diﬀerentiation with respect to z 70 . Taking into account the exponential-aﬃne structure of the general characteristic function in (2.34), we are able to write equation (4.6) alternatively as 67 68 69 70

See Chacko and Das (2002). See Proposition 2.4.3. Compare with Theorem 1 (c) in Bakshi and Madan (2000). The result in equation (4.6) is always real, see e.g. Bakshi and Madan (2000). Therefore, the operator Re [. . .] in this calculation is not necessary at all, which can be justiﬁed by checking that all imaginary parts in this equation cancel out except in the term ıg (xT ).

4.2 Heston Approach

53

ψz (xt , 0, w0 , w, g0 , g, τ ) ψ(xt , 0, w0 , w, g0 , g, τ ) = × ı ı d ln [ψ(xt , z, w0 , w, g0 , g, τ )] dz z=0 ψ(xt , 0, w0 , w, g0 , g, τ ) φz (xt , 0, w0 , w, g0 , g, τ ). = ı In the last equation, we used the function φ(xt , z, w0 , w, g0 , g, τ ), which is just the natural logarithm of ψ(xt , z, w0 , w, g0 , g, τ ) in our exponential-aﬃne model setup. Thus, the derivative with respect to z of the exponent of the characteristic function is then ˜ z (z, τ ) xt + ıg (xt ) . φz (xt , z, w0 , w, g0 , g, τ ) = az (z, τ ) + b Using the same technique as before, we obtain the value of an ordinary zero bond as T T − r(xs ) ds − r(xs ) ds = E e t eızg(xT ) P (xt , t, T ) =EQ e t (4.7) z=0

=ψ(xt , 0, w0 , w, g0 , g, τ ). Finally, we are left with the calculation of the Arrow-Debreu prices. As mentioned before, these functions Π1 (xt , t, T ) and Π2 (xt , t, T ) can also be interpreted as probabilities. Hence, we apply a tool to determine probabilities from characteristic functions. This can be done with a Fourier inverse transform as proposed in Gil-Pelaez (1951). Theorem 4.2.2 (Inversion Theorem of Gil-Pelaez). If ψ xT (xt , z, t, T ) is the characteristic function of a one-dimensional stochastic variable xt then the probability Pr(xT ≥ K), given some state xt and some constant K, can be calculated as 1 1 Pr(xT ≥ K) = + 2 π

∞ 0+

ψ xT (xt , z, t, T )e−ızK Re ız

dz,

(4.8)

with z ∈ R. The expression 0+ in equation (4.8) denotes the right-sided limit to the origin. Obviously, the integrand is not deﬁned for a zero-valued transformation

54

4 Three Fourier Transform-Based Pricing Approaches

variable z 71 . Note that the inversion theorem in 4.2.2 is not limited to recover only probabilities for the case of symmetric probability density functions, which might be implicated due to the term 12 . Equation (4.9) holds for general probability distributions. The only condition to be satisﬁed is the existence of the characteristic function or its system of ODEs. Moreover, we are also able to use Theorem 4.2.2 for the linear combination g(xt ), as long as the outcome is a scalar random variable. As long as we are able to obtain the general characteristic functions ψ1 (xt , z, w0 , w, g0 , g, τ ) and ψ2 (xt , z, w0 , w, g0 , g, τ ) corresponding to the particular measures Q1 and Q2 , we are able to compute the values of Π1 (xt , t, T ) and Π2 (xt , t, T ). In analogy to equations (4.6) and (4.7), and keeping the normalization made in (4.4) in mind, we therefore have ψ1 (xt , z, w0 , w, g0 , g, τ ) =

ψz (xt , z, w0 , w, g0 , g, τ ) , ıΠ0 (xt , t, T )

ψ2 (xt , z, w0 , w, g0 , g, τ ) =

ψ(xt , z, w0 , w, g0 , g, τ ) . P (xt , t, T )

and

Subsequently, the values of the required Arrow-Debreu securities can be calculated as72 1 1 Π1,2 (xt , t, T ) = + 2 π

∞ Re 0+

ψ1,2 (xt , z, w0 , w, g0 , g, τ )e−ızK ız

dz.

(4.9)

Although the derivation of option prices within this methodology is comprehensible, this technique does entail some drawbacks. Firstly, a general advantage which holds for all pricing methodologies based on Fourier Transformation techniques is that we are not restricted to simple stochastic dynamics of the underlying short-rate process, where the probability density function p(xt , xT , w0 , w, t, T ) is explicitly known in closed form73 . With the continuum of characteristic functions at hand, we are able to calculate option prices for a much broader class of stochastic dynamics. Despite the apparent elegance of this approach, there are also some issues to discuss. Since we expressed the 71 72 73

More on this topic and residue calculus is discussed in Section 4.3. Compare with the general result in Bakshi and Madan (2000), Theorem 1. However, there exist density functions for which no characteristic function exists, e.g. a log-normal distributed random variable.

4.3 Carr-Madan Approach

55

option price as a decomposition of probabilities multiplied with their normalization factors, we have to calculate for a sum of N terms in G (xT ) the same number of separate Fourier inversions and therefore to perform N numerical integrations. Especially in one-factor interest-rate models, this fact can be avoided using a Fourier transform with respect to rT 74 . From a computational point of view, this can be very time consuming and therefore ineﬃcient compared to the pricing approaches of Carr and Madan (1999) and Lewis (2001). Additionally, the denominator in the integrand of equation (4.9) decays only linearly for the idealized payoﬀ functions, compared to the payoﬀ-transform approaches discussed in the subsequent sections75 . Another matter we want to address is the integration procedure itself. In equation (4.8), we need to consider carefully the pole at the origin. Sometimes, this can lead to rather unstable results. Another point to mention is that the structure of the option contract dictates the calculation procedure of the particular function Πj (xt , t, T ). Hence, it ﬁrst has to be determined whether the payoﬀ function G(xT ) exhibits linear, exponential-linear or integro-linear terms of g (xT )76 , which result in diﬀerent valuation formulae for the option price. This can complicate unnecessarily the computation of option prices in contrast to the approaches discussed in the following sections, where Fourier Transformations of the payoﬀ function are used.77 .

4.3 Carr-Madan Approach Carr and Madan (1999) develop a diﬀerent method for retrieving option prices using characteristic functions. Instead of applying general characteristic functions to obtain the exercise probabilities and the Arrow-Debreu security prices 74 75

See, for example, the pricing of coupon bonds in Section 5.3.3. The denominator in the payoﬀ transforms of the interest-rate option contracts in table 4.1 are quadratic and therefore have a higher rate of convergence. Compare

76

with the particular transformations given in Section 5.3. See Chacko and Das (2002) for a comprehensive discussion and classiﬁcation of payoﬀ functions and derivation of the particular option prices in this transformed

77

Arrow-Debreu security framework. See Bakshi and Madan (2000), pp. 218-220, cases 1-3, on how to derive the particular ψj (xt , z, w0 , w, g0 , g, τ ) for general payment structures. Chacko and Das (2002) also derive the respective valuation algorithms for these payoﬀ structures.

56

4 Three Fourier Transform-Based Pricing Approaches

under the particular probability measures Q1,2 as done in the last section, they propose an alternative approach. The intention behind this framework is to formulate a valuation procedure, that can incorporate the FFT, a very eﬃcient tool in deriving Fourier Transformations for diﬀerent values of the underlying random variable. However, they ﬁrst perform a Fourier Transformation on the payoﬀ function with respect to the strike variable K. Afterwards, interchanging the order of integration, they are able to compute the desired fair price of the option as an inverse Fourier Transformation, thus applying the relevant characteristic function, an example is given below. Obviously, a ﬁrst advantage of this strategy is that, since we deal with only one transform operation on the option price, in order to compute model price we need only one inverse transformation. As the authors mention, a closed-form solution of the option price in Fourier space is presupposed78 . Since option prices commonly have at least two terms in the payoﬀ function G (xT ), numerical calculations with this method are approximately twice as fast. A problem in this approach mostly arises if a Fourier transform on the payoﬀ function with a real-valued frequency variable z ∈ R is applied. As mentioned in Bakshi and Madan (2000), the transformed payoﬀ function would not exist at all, due to the unbounded option payoﬀ functions79 . To circumvent this issue, Carr and Madan (1999) introduce an artiﬁcial dampening parameter α and derive a modiﬁed transformed option price, upon which they apply the inverse transformation procedure. In the following presentation of this methodology we do not refer to an artiﬁcial dampening parameter α; rather we want to introduce a general Fourier Transformation as deﬁned in deﬁnition 2.4.1 with z ∈ C. Moreover, we show that the dampening parameter coincides with the negative ﬁxed imaginary part zi of the frequency variable z = zr + ızi . Following this trail, we get a more intuitive concept of the nature of the dampening factor α used by Carr and Madan (1999). Demonstrating the pricing technique, we rely on the same contract type as in (4.3) with G (xT ) = (g (xT ) − K)+ to maintain the comparability to 78

See Carr and Madan (1999), p. 61. We extend this methodology to allow for characteristic functions with no closed-form representations. This topic is discussed

79

in Chapter 6. See Bakshi and Madan (2000), p. 215. An exception would be a contract which is bounded on two sides, e.g. a butterﬂy contract.

4.3 Carr-Madan Approach

57

previously obtained solutions of our example in equation (4.5). Starting with a Fourier Transformation on the payoﬀ function with respect to K, we have ∞

∞

F K [G (xT )] =

+

eızK (g (xT ) − K) dK

eızK G (xT ) dK = −∞

−∞

g(x T)

eızK (g (xT ) − K) dK

= −∞

= −e =−

e

ızK 1

+ (g (xT ) − K)ız z2

(4.10)

g(xT ) −∞

ızg(xT )

z2

with Im(z) < 0.

The restriction in equation (4.10) upon the imaginary part of z guarantees the ﬁniteness of the transformed payoﬀ function. Thus, we are able to interpret (4.10) as a line integral, which is evaluated parallel to the real axis going through ızi . Apart from considerations about the regularity of the payoﬀ transform, the value of zi can also be used to optimize numerical accuracy of the valuation algorithm80 . Exploiting the symmetry of real-valued Fourier transforms, the payoﬀ function G (xT ) for our speciﬁc example, can be expressed by the following inverse transformation problem 1 G (xT ) = − π

∞

e−ızK

eızg(xT ) dz. z2

(4.11)

0

Carrying out this inverse operation, we need zi to be ﬁxed on the same strip used for the transformation. Otherwise, the original function and its image function in dual space would not correspond to each other81 . The essential part, in expressing the valuation formula as an inverse Fourier-style problem, is the interchanging of the integration order. Furthermore, we have in equation (4.11) an exponential term for both the underlying 80

See Lee (2004), for a comprehensive analysis of the eﬀect of zi on the accuracy of the computational result. Note, the derived error bounds in this article are only valid for one particular strike. These results have to be treated carefully for algorithms, where option prices for diﬀerent strike rates, such as ITM, ATM, and

81

OTM options, are computed simultaneously. This fact is discussed in Section 2.4.

58

4 Three Fourier Transform-Based Pricing Approaches

stochastic variable and the strike rate enabling the application of the characteristic function methodology and afterwards to calculate prices with the FFT. Denoting our exemplary valuation problem of (4.3) in terms of (4.11), we get the following integral representation + V (xt , t, T ) = (g (xT ) − K) p(xt , xT , w0 , w, t, T ) dxT 1 =− π

RM

RM

∞

ızg(xT ) e e−ızK dz p(xt , xT , w0 , w, t, T ) dxT . z2

(4.12)

0

Due to Fubini’s theorem, the order of integration can be interchanged82 . Therefore, we are able to use the alternative representation 1 V (xt , t, T ) = − π

∞

e−ızK z2

eızg(xT ) p(xt , xT , w0 , w, t, T ) dxT dz RM

0

"

#$

%

eqn. (2.30) 1 =− π

∞

e−ızK

(4.13)

ψ(xt , z, w0 , w, g0 , g, τ ) dz. z2

0

Eventually, we get the Fourier-style valuation formula for the price at time t of a European call option, based on the payoﬀ function G (xT ) = (g(xT ) − K)+ . The relationship between the artiﬁcial dampening factor α in Carr and Madan (1999) and zi becomes apparent if we substitute z = zr +ızi in equation (4.13), which gives 1 V (xt , t, T ) = − π

∞ 0

ezi K =− π

e−ı(zr +ızi )K ∞ 0

ψ(xt , zr + ızi , w0 , w, g0 , g, τ ) dzr (zr + ızi )2 (4.14)

ψ(xt , zr + ızi , w0 , w, g0 , g, τ ) e−ızr K dzr . zr2 + 2ızr zi − zi2

Obviously, compared to the corresponding option price formula in Lee (2004), it can easily be veriﬁed that the identity zi ≡ −α holds83 . 82

83

Since all parts of the integral are real-valued, we are able to change the order of integration without any problems. The modiﬁed transformed option price for our example is also given in Lee (2004) Theorem 4.2 as cˆα,G2 (u), where u matches zr . Also compare this result with the general Fourier-style valuation formula in Carr and Wu (2004), p. 136.

4.3 Carr-Madan Approach

59

In contrast to the Heston pricing approach, the Carr-Madan methodology provides an additional degree of freedom, since we are no longer limited to the case of a real-valued transformation variable z. This is of major importance in a numerical scheme for computing derivative prices84 . Furthermore, we are able to shift the integration contour around any existing pole. However, in these cases the residue of the particular pole must be taken into account85 . Proceeding like this, the accuracy of the valuation algorithm can be drastically increased86 . Nevertheless, we are also free to choose the imaginary part in (4.14), such that the contour integrals have to be performed right through a pole. Doing this we ﬁrst consider the residuals of the poles and then evaluate the integral due to Cauchy’s theorem87 . Generally, the advantage in this approach lies in the availability of a fast numerical integration routine, the FFT algorithm. A properly set procedure, based e.g. on our example in (4.14), can calculate a vast number of derivative prices for alternative strike rates in fractions of a second. On the other hand, Fourier-style solutions in this framework cannot be properly decomposed into parts of the general characteristic function and the transformed payoﬀ function88 . Thus, we needed a speciﬁc payoﬀ function in the derivation of the transformed option price. It would be more convenient and from a numerical perspective more desirable if the integral in (4.14) could be clearly separated into a part of the general characteristic function, which depends on the underlying stochastic dynamics, and a part determined by the contract we want to price. Moreover, there seems to exist a problem for particular models with speciﬁc parameter constellations89 . Finally, we do not prefer this methodology in the ﬁrst place because it cannot be properly applied for coupon-bond 84 85 86

The choice of the optimal value of zi is discussed in Section 6.3.3. See Lee (2004) equations (6) and (7). This can be validated by Tables 2 and 3 in Lee (2004). The error bounds presented there are up to a thousand times lower, if the integrals are evaluated on contours

87 88

89

with no existing poles. In the next section, we derive valuation formulae using diﬀerent values of zi . For example, the transformed option price in equation (4.13)

is

0 ,g,τ ) . − ψ(xt ,z,w0z,w,g 2 Itkin (2005) analyzed the FFT method of Carr and Madan (1999) for the case

of an underlying Variance-Gamma process and reports some numerical issues for diﬀerent lengths of time to maturity τ .

60

4 Three Fourier Transform-Based Pricing Approaches

options and swaptions, respectively, with an underlying one-factor interestrate process. The reason for this is that we need the exercise boundary to be explicit in rT in order to present the valuation formula in terms of the characteristic function. If this is not the case, we lose characteristics of the stochastic process, which are relevant in the valuation formula and therefore have to be considered within the integration. For example, in the case of coupon-bond options, we encounter the problem of determining numerically a critical value rT∗ 90 , thus making it impossible to compute the particular option prices. These problems can be circumvented with the approach discussed in the following section.

4.4 Lewis Approach Lewis (2001) presented in his work an alternative way to retrieve not only option prices, but general derivatives prices91 . The approach is similar to the previously discussed methodology of Carr and Madan (1999), but can be applied to a wider area of pricing problems. Thus, we are able to calculate all derivatives prices presented in Chapter 3 with a single general valuation formula. Fortunately, within this framework, it is also possible to use an eﬃcient numerical tool to compute derivative prices with comparable speed to the FFT algorithm, namely the IFFT algorithm. In contrast to the approach in Carr and Madan (1999), Lewis (2001) introduced a derivatives pricing framework starting with a Fourier Transformation of the payoﬀ function, but this time with respect to the underlying stochastic variable, where the frequency parameter z ∈ C is also supposed to be complex-valued. Thus, the advantages discussed in the last section still hold. As before, our starting point is the payoﬀ function G (xT ) of a derivatives contract. As in the previous section, the Fourier Transformation is performed on the payoﬀ function, in this case with respect to the scalar g (xT ). Accordingly, the transformed payoﬀ function is 90 91

See Jamshidian (1989). As mentioned before, the methodology was ﬁrstly used in Lewis (2000). However, we refer to Lewis (2001) because of the more detailed derivation and comprehensive discussion of this pricing framework.

4.4 Lewis Approach

61

∞ F

g(xT )

eızg(xT ) G(xT ) dg (xT ) .

[G(xT )] =

(4.15)

−∞

To guarantee the ﬁniteness of the integral in equation (4.15) and the existence of F g(xT ) [G (xT )], respectively, the imaginary part of z has to be restricted, where its domain depends on the speciﬁc contract. Continuing with our example in pricing an interest-rate cap of the form G (xT ) = (g (xT ) − K)+ , we ﬁrst calculate the transformed payoﬀ function with respect to g (xT ) as ∞ F

g(xT )

eızg(xT ) (g (xT ) − K)+ dg (xT )

[G(xT )] =

(4.16)

−∞

=−

e

ızK

z2

with Im(z) > 0. Although this formula bears a strong resemblance to equation (4.10), one remarkable diﬀerence between them is the interval of zi , for which the Fourier transform of the particular payoﬀ function exists92 . Another point we would like to mention is that the transformed payoﬀ function displays the strike rate K in the exponential function instead of g (xT ), according to the methodology of Carr and Madan (1999). Representing the time t option price with the help of the transformed payoﬀ function, we have at the general valuation formula93 ∞ 1 e−ızg(xT ) F g(xT ) [G (xT )] dz × V (xt , t, T ) = π RM

0

(4.17)

p(xt , xT , w0 , w, t, T ) dxT , which is for our speciﬁc example of an interest-rate cap 92 93

In comparison to equation (4.10), zi has to be negative. Again, we take advantage of the symmetry of Fourier Transformations for realvalued functions.

62

4 Three Fourier Transform-Based Pricing Approaches

=−

1 π

RM

∞

e−ızg(xT )

e

ızK

z2

dz p(xt , xT , w0 , w, t, T ) dxT .

(4.18)

0

Again, we apply Fubini’s theorem, implicating the possibility of interchanging the order of integration in (4.17). Thus, for general payoﬀ functions we obtain 1 V (xt , t, T ) = π

∞ F g(xT ) [G (xT )] × 0

(4.19)

e−ızg(xT ) p(xt , xT , w0 , w, t, T ) dxT dz.

RM

Firstly, we focus on the inner integral. In line with the formal deﬁnition of the characteristic function, according to equation (2.27), we are able to establish the relation

eı(−z)g(xT ) p(xt , xT , w0 ,w, t, T ) dxT (4.20)

RM

=ψ(xt , −z, w0 , w, g0 , g, τ ). Inserting this result into equation (4.19), we eventually get the general version of the Fourier-style valuation formula 1 V (xt , t, T ) = π

∞ F g(xT ) [G (xT )] ψ(xt , −z, w0 , w, g0 , g, τ ) dz,

(4.21)

0

which is for our example of a call contract with underlying variable g (xT ), 1 − π

∞

eızK ψ(xt , −z, w0 , w, g0 , g, τ ) dz z2

0

with Im(z) > 0.

In contrast to the pricing procedure introduced by Carr and Madan (1999), we have a strict separation of functionals, which depend either on the contract type or on the underlying stochastic dynamics. The respective part for the contract type is therefore represented by the transformed payoﬀ function, whereas

4.4 Lewis Approach

63

the stochastic dynamics of the underlying process is implemented in terms of the characteristic function. Hence, we have a real modular pricing framework, in which each part in (4.21) can be exchanged without greater eﬀort. Moreover, we can apply this methodology consistently to contracts, whether they are unconditionally exercised or bear an optional exercise right94 . In particular, for one-factor models with multiple jump components, we are able to take advantage of the fact that for most contracts the domains of zi are overlapping. This means that zi can be chosen arbitrarily, subject to compliance with numerical accuracy95. Thus, we usually have to evaluate ψ(xt , −z, w0 , w, g0 , g, τ ) only once for diﬀerent values of zr . Afterwards, these precomputed values can be used for all relevant contract types needed. This drastically improves the eﬃciency of the numerical valuation scheme. The payoﬀ-transform approach according to Lewis (2001) is extremely versatile. For example, with this pricing technique, we can also derive the quantities Π1 (xt , t, T ) and Π2 (xt , t, T ), without need of any derivative function ψz (xt , z, w0 , w, g0 , g, τ ), as done in formula (4.9). Although the numerical integration on a line integral (partly) including a pole exhibits the undesirable numerical properties discussed earlier, we want to show the derivation of the Gil-Pelaez style valuation formulae for Π2 (xt , t, T ), as given in Theorem 4.2.2 within the Lewis methodology96 , for demonstration purposes. Recalling that the payoﬀ of an Arrow-Debreu security can be formally represented by the indicator function, we apply a Fourier Transformation on this special function in order to calculate Π2 (xt , t, T ). Under the probability measure Q2 , the simple payoﬀ representation is then given by the incomplete Fourier Trans94 95

This is demonstrated in the next chapter. In addition to the restrictions for zi , due to the validity for the transformed payoﬀ function, in some cases we need to restrict the domain for the imaginary part of the transformation variable further to ensure the regularity of the characteristic function. One example, where zi has an additional constraint due to this issue is the characteristic function for the variance gamma process which is discussed in

96

Itkin (2005). In contrast to equation (4.9), we would get an alternative representation for Π1 (xt , t, T ), without needing any derivative of ψ(xt , z, w0 , w, g0 , g, τ ) and φ(xt , z, w0 , w, g0 , g, τ ), respectively.

64

4 Three Fourier Transform-Based Pricing Approaches

formation97

eızK F g(xT ) 1g(xT )>K = − ız

(4.22)

with Im(z) > 0. Using this formula, together with zi in the appropriate domain, we are almost ready to calculate Π2 (xt , t, T ). In fact, we consider the residue theorem and apply a suitable closed-contour integral to recover the exact formula according to equation (4.9). Hence, evaluating the integral including the pole at zi = 0 gives the desired result, which is demonstrated below. ˜ 2 (t, T ) to compensate for the We start with a slightly modiﬁed function Π inﬂuence of the probability law Q2 98 , which is deﬁned as ˜ 2 (xt , t, T ) = Π2 (xt , t, T )P (xt , t, T ). Π

(4.23)

Inserting the transformed payoﬀ function (4.22) into our general valuation formula (4.21) gives ˜ 2 (xt , t, T ) = − 1 Π π

∞

eızK ψ(xt , −z, w0 , w, g0 , g, τ ) dz, ız

(4.24)

0

with Im(z) > 0. Equation (4.24) can already be used for valuation purposes. Since we want to show the similarity of this formula to the transformed Arrow-Debreu security pricing approach, we encounter the problem of integrating through a pole, and therefore must apply Cauchy’s residue theorem for analytic functions99 .

Theorem 4.4.1 (Cauchy’s Residue Theorem). Assume the function f (z) is analytic within a closed, counter-clockwise performed integration contour C, 97

98

99

One-sided Fourier Transformations are commonly referred to as incomplete Fourier Transformations. This has to be done, since we use the general characteristic function ψ(xt , z, w0 , w, g0 , g, τ ). This means, the function has to satisfy the Cauchy-Riemann equations. See Duﬀy (2004), p. 16.

4.4 Lewis Approach

65

except at points zd ∈ C, where f (zd ) encounters singularities. Then the value of the closed contour integral for this function can be calculated as & f (z) dz = 2ıπ Res [f (z)|z = zd ] .

(4.25)

d

C

The residues at the singularities corresponding to points zd can be derived as Res [f (z)|z = zd ] = lim

z→zd

dn−1 1 [(z − zd )n f (z)] . (n − 1)! dz n−1

(4.26)

The parameter n represents the order of the pole. Hence, if we want to evaluate the integral in (4.24) for Im(z) = 0, we have to deal with a simple pole of order n = 1. To facilitate the calculations, we ˜ 2 (xt , t, T ) ﬁrst introduce the original, two-sided integral representation for Π in the manner of equation (2.25), which is simply ˜ 2 (xt , t, T ) = − 1 Π 2π

∞ −∞

eızK ψ(xt , −z, w0 , w, g0 , g, τ ) dz. ız

(4.27)

Proceeding like this, we add to the former line integral, which has to be evaluated parallel to the real axis with distance Im(z), several additional integral paths to build a rectangular shape on the upper imaginary half-plane100 . This gives us a contour C, which is performed, as illustrated in Figure 4.1. Setting

∞ ˜ 2 (xt , t, T ) = Π

f (z) dz,

(4.28)

−∞

with

eızK ψ(xt , −z, w0 , w, g0 , g, τ ) , 2πız we are able to express the contour integral as f (z) = − & f (z) dz = C 100

6 j=1 C

j

f (z) dz =

6

Ij .

(4.29)

j=1

In manipulating equation (4.27), we could also have chosen the lower half-plane. Subsequently, we would then have to be careful about the direction, how the pole is encircled, making its contribution to the integration either in a positive or negative sense.

66

4 Three Fourier Transform-Based Pricing Approaches

˜ 2 (xt , t, T ) in Fig. 4.1. Clockwise performed integral path for the derivation of Π equation (4.27) on the real line. The cross represents the pole.

Referring to Figure 4.1, the integral part I4 forms a half arc around the pole of the meromorphic function101 with radius . Thus, excluding the pole, we can state, due to Cauchy’s integral theorem, & f (z) dz = 0. (4.30) C

In the next step, we need to determine the values of the speciﬁc integrals ˜ 2 (xt , t, T ) given in equation Ij . Starting with I1 , we have just the value of Π (4.27). Recognizing that for 0 ≤ Im(z) < ∞, we have lim f (R + ızi ) = 0, R→±∞

we immediately obtain I2 + I6 = 0.

(4.31)

Subsequently, we are left with the computation of the remaining integral parts I3 , I4 and I5 . According to Theorem 4.4.1, if we consider an arc performed in a counter-clockwise fashion around a pole, we would have to take into account the entire contribution of the pole. Therefore, by assuming the radius of the half arc I4 to be inﬁnitesimally small, we eventually obtain half of the particular contribution. Thus, we have to consider the residue 101

A function f (z) is said to be meromorphic, if it only has some isolated singularities. This means that such a function is analytic everywhere, except at these poles. See Duﬀy (2004), p. 16.

4.4 Lewis Approach

67

I4 = ıπRes [f (z)|z = 0] = ıπ lim zf (z) z→0

P (xt , t, T ) ψ(xt , 0, w0 , w, g0 , g, τ ) =− . =− 2 2

(4.32)

Likewise, assuming the distance to the origin for integrals I3 and I5 to be inﬁnitesimally small, we are able to represent them in the limit as102 0+ I3 + I5 =

−∞ f (z) dz + f (z) dz.

∞

(4.33)

0−

Having derived the required expressions for all integral parts Ij in equations (4.31), (4.32) and (4.33), additionally using equation (4.30), we eventually end ˜ 2 (xt , t, T ), which is given by up with an alternative representation for Π ˜ 2 (xt , t, T ) = − (I3 + I4 + I5 ) Π + −∞ 0 P (xt , t, T ) = − f (z) dz + f (z) dz 2 ∞

P (xt , t, T ) −2 = 2

0−

(4.34)

−∞ f (z) dz 0−

In equation (4.34), the symmetry of characteristic functions for real-valued functions is exploited, due to Proposition 2.4.3. Therefore, the two integrals in the above equation can be aggregated. In a last step, we reinsert the detailed expression of f (z) and substitute z ∗ = −z. This results in the relation ˜ 2 (xt , t, T ) = P (xt , t, T ) Π 2

∞ −ız∗ K e 1 ψ(xt , z ∗ , w0 , w, g0 , g, τ ) Re + dz ∗ , π ız ∗

(4.35)

0+

with Im(z ∗ ) = 0. Dividing equation (4.35) by P (xt , t, T ) and considering only the relevant real part of the solution, we obtain the Heston-style solution of equation (4.9), 102

Here, we use again the convention 0± denoting the right- and left-hand sided limit towards zero.

68

4 Three Fourier Transform-Based Pricing Approaches

which we intentionally wanted to reproduce with the payoﬀ-transformation approach of Lewis (2001). In contrast to the Fourier-transform approach introduced in Carr and Madan (1999), the methodology discussed above is not that popular. One reason might be that the FFT algorithm cannot be applied to the valuation formula. Albeit, simply using an IFFT algorithm provides equivalent functionality and eﬃciency in solving derivatives prices. On the other hand, we prefer the method of Lewis (2001) because of the clear separation of diﬀerent valuation components in the pricing formula. Additionally, this framework enables us to consistently use the valuation formula presented in equation (4.21) for both unconditional and conditional derivatives contracts by using residue calculus. Moreover, with this methodology even swaptions and options on coupon bonds can be priced in case of one-factor interest-rate models.

5 Payoﬀ Transformations and the Pricing of European Interest-Rate Derivatives

5.1 Overview In this chapter we derive semi closed-form solutions of European interest-rate derivatives in terms of their transformed payoﬀ functions, for all contracts given in Chapter 3. Equipped with this frequency representation of the payoﬀ function, the contract can be priced with the general valuation formula according to equation (4.21). This procedure, combined with a standardized numerical integration routine, can then be used to compute the desired quantities. Apart from the generality of this method, we observe that all call and put option contracts exhibit identical payoﬀ representations in Fourier space. The diﬀerence between them are the diﬀerent strips in the imaginary plane, parallel to the real axis, on which the transform operation is valid for the particular contract. As before, we distinguish between contracts with unconditional and conditional exercise rights. The reason for this separation of the payoﬀ-transformed formulae is that contracts with unconditional exercise rights can be calculated as simple unconditional expectations. Using the residue theorem, solutions for the underlying contracts can be computed in terms of the general characteristic function, without evaluating numerically any integral at all. However, if the characteristic function is not known in closed form but can be represented as a system of ODEs, theoretical prices have to be numerically obtained via a Runge-Kutta algorithm. On the other hand, contracts with optional exercise rights are computed by numerical integration in every case.

70

5 Payoﬀ Transformations and European Interest-Rate Derivatives

5.2 Unconditional Payoﬀ Functions This section is organized as follows. First we compute some fundamental Fourier Transformations for functionals containing g (xT )103 , henceforth referred to as building blocks. These blocks, combined with the particular characteristic function, can then be used to compute the contract prices of Section 3.2 in the form of Fourier-style valuation formulae via equation (4.21). In calculating the payoﬀ transform, we do not have to pay attention to the question of whether the derivative to be priced is a normal or futures-style contract. This is captured by the choice of the relevant characteristic function, which can be either ψ(xt , z, w0 , w, g0 , g, τ ) or ψ(xt , z, 0, 0M , g0 , g, τ ). At ﬁrst sight, a problem arises in pricing unconditional interest-rate derivatives, due to the unbounded integration range of the expectation. As shown in the option-pricing example in Sections 4.3 and 4.4, the imaginary part of z can be used to ensure the existence of the payoﬀ transform by suﬃciently dampening the integral on one side, which could be either the upper or lower. Unfortunately, the dampening eﬀect cannot be accomplished simultaneously on both integration boundaries. Thus, we need additional considerations in order to derive an appropriate representation of the valuation formula in frequency space. Nevertheless, after some manipulation of the transformed payoﬀ, we derive in the upcoming section the particular valuation formulae. 5.2.1 General Results We begin with two basic interest-rate derivatives, the zero-bond contract as deﬁned in equation (4.7) and the expectation of g (xT ) as given by equation (4.6). According to Section 4.2, the value of a zero bond equals ψ(xt , 0, w0 , w, g0 , g, τ ) whereas the latter quantity can be obtained via the calculation of its ﬁrst derivative. These general results hold for arbitrary linear combinations g (xt ). In contrast, the payoﬀ-transformation technique as presented in Section 4.4 seems at ﬁrst sight to have diﬃculties in recovering these particular expectations, due to the unbounded integration domain. Hence, the ﬁrst step in this subsection is to prove the former results obtained 103

Although not explicitly displaying the variable g (xT ) in the payoﬀ function, we also interpret in the following the Fourier Transformation of a constant as encountered in zero-bond contracts as a building block.

5.2 Unconditional Payoﬀ Functions

71

in equations (4.6) and (4.7) and therefore show that the payoﬀ-transform methodology can be applied without exceptions. If we set G(xT ) = 1, which represents the riskless return of one unit money at maturity, it seems at ﬁrst that the ordinary payoﬀ transform is no longer ﬁnite. Unfortunately, with help of the imaginary part of the transformation variable z, we are only able to dampen the integrand on one side, which can be either in the direction of the positive or the negative real half-plane. Thus, we cannot dampen the underlying payoﬀ function for both sides simultaneously, and consequently cannot perform the inverse Fourier Transformation on the same strip in the imaginary plane. However, performing the integration on diﬀerent strips in the imaginary plane, we are again able to use the payoﬀtransform methodology. Dividing the integration domain (−∞, ∞) into two separate subdomains (−∞, ε) and (ε, ∞) with arbitrary ε ∈ R, we end up with two frequency functions deﬁned on diﬀerent strips in the imaginary plane. At ﬁrst glance, this seems to complicate the situation. In fact, with the help of Cauchy’s residue theorem, the calculations are rather simpliﬁed. The payoﬀ transform of an ordinary zero bond can be calculated as104 , ε ∞ eızε eızε g(xT ) ızg(xT ) F − , (5.1) [1] = e dg (xT ) + eızg(xT ) dg (xT ) = ız ız −∞

ε

with Im(z) < 0, and z representing the complex conjugate of z 105 . Working with this transformed payoﬀ function, we are already able to recover the zero-bond price due to the integral representation ∞ ızε e 1 P (xt , t, T ) = ψ(xt , −z, w0 , w, g0 , g, τ ) dz 2π ız −∞ (5.2) ∞ ızε e 1 ψ(xt , −z, w0 , w, g0 , g, τ ) dz. − 2π ız −∞

104

105

Obviously, in pricing a zero bond, the choice of g (xT ) is irrelevant. In fact, g (xT ) can be set to any value, since the payoﬀ function itself is independent of g (xT ). We make this assumption for convenience. Generally, the imaginary part of the transform variable used in the latter integral can be independently chosen on the positive half-axis.

72

5 Payoﬀ Transformations and European Interest-Rate Derivatives

Interchanging the integration boundaries of the latter integral in equation (5.2) and closing the contour with two additional paths from points (R, ızi ) to (R, −ızi ) for R → ±∞, thus forming a closed contour integral with the resulting four integrals, we are able to use Cauchy’s residue theorem again. The rectangular contour including the singularity is shown in Figure 5.1. Due to the direction of the path, we have to consider a counter-clockwise encircled simple pole at z = 0, which is completely inside the contour. Consequently, the contour integral equals 2πıRes [f (z)|z = 0] with f (z) =

eızε ψ(xt , −z, w0 , w, g0 , g, τ ), 2πız

and the value of a zero bond is106 ∞ P (xt , t, T ) =

−∞ f (z) dz + f (z) dz = 2πıRes [ f (z)| z = 0]

−∞

(5.3)

∞

=ψ(xt , 0, w0 , w, g0 , g, τ ). Here, the calculations for the residue are analogous to the ones made in equation (4.32), but this time considering the entire residue. The same result would have been obtained using the Dirac Delta function δ(z) in the transformed payoﬀ function. It is a well-known result that ∞ F

g(xT )

eızg(xT ) dg (xT ) = 2πδ(z),

[1] =

(5.4)

−∞

with Im(z) = 0. Hence, the fair value of a zero bond can be alternatively calculated as107 1 P (xt , t, T ) = 2π

∞ 2πδ(z)ψ(xt , −z, w0 , w, g0 , g, τ ) dz −∞

(5.5)

=ψ(xt , 0, w0 , w, g0 , g, τ ), 106

Starting from here, all zero-valued integrals are ignored.

107

Obviously, for arbitrary real-valued w, the relation

∞

−∞

holds.

δ(z − w)f (z) dz = f (w)

5.2 Unconditional Payoﬀ Functions

73

Fig. 5.1. Closed contour integral path for the derivation of P (xt , t, T ) in equation (5.2). The pole is completely encircled in a counter-clockwise manner.

which justiﬁes the above statement. So far, we have shown the result of one important building block, the model price of a zero bond, with the help of the payoﬀ-transform methodology. In order to price interest-rate contracts bearing unconditional exercise rights, we also need the expected value of the payoﬀ function G (xT ) = g (xT ) as given by equation (4.6). In the following, we want to prove this general result within the payoﬀ-transform methodology. Starting our calculations, we assume a linear payoﬀ function based on g (xT ) and then apply two incomplete Fourier Transformations, this time with an artiﬁcial integration boundary ε for the particular integrals. Hence, the transformed payoﬀ function of G (xT ) = g (xT ) can be calculated as

74

5 Payoﬀ Transformations and European Interest-Rate Derivatives

ε F

g(xT )

eızg(xT ) g (xT ) dg (xT )

[g (xT )] = −∞

∞ eızg(xT ) g (xT ) dg (xT )

+

(5.6)

ε

eızε (1 − ızε) eızε (1 − ızε) − , = z2 z2 with Im(z) < 0. This time, we build a rectangular integration path, performed in a clockwise manner which is depicted in Figure 5.2. Hence, we get for the discounted expectation108 T EQ e

−

r(xs ) ds t

g (xT )

1 =− 2π

−∞

∞ ∞

eızε (1 − ızε) ψ(xt , −z, w0 , w, g0 , g, τ ) dz z2

eızε (1 − ızε) ψ(xt , −z, w0 , w, g0 , g, τ ) dz z2 −∞

ızε e (1 − ızε) = −2πıRes − ψ(xt , −z, w0 , w, g0 , g, τ ) z = 0 . 2πz 2 (5.7) −

1 2π

Using again Cauchy’s residue theorem, the contribution of the pole at the origin109 can be derived as 108

109

According to the clockwise performed integration path, the contribution of the pole in this case is −2πı times the residue. According to a removable singularity, we have in fact at z = 0 two diﬀerent poles, a simple and a second order pole.

5.2 Unconditional Payoﬀ Functions

75

Fig. 5.2. Closed contour integral path for the discounted expectation of g (xT ). The pole is completely encircled in a clockwise manner.

eızε (1 − ızε)ψ (xt , −z, w0 , w, g0 , g, τ ) Res − z = 0 2πz 2 ızε

e ψ (xt , −z, w0 , w, g0 , g, τ ) =Res − z = 0 2πz 2 ızε

e ψ (xt , −z, w0 , w, g0 , g, τ ) ε + Res − z = 0 2πız ızε d e ψ (xt , −z, w0 , w, g0 , g, τ ) = lim − z→0 dz 2π ızε e ψ (xt , −z, w0 , w, g0 , g, τ ) ε + lim − z→0 2πı ψz (xt , 0, w0 , w, g0 , g, τ ) − ıψ (xt , 0, w0 , w, g0 , g, τ ) ε = 2π ψ (xt , 0, w0 , w, g0 , g, τ ) ε − 2πı ψz (xt , 0, w0 , w, g0 , g, τ ) = . 2π

(5.8)

76

5 Payoﬀ Transformations and European Interest-Rate Derivatives

Inserting this result in equation (5.7), we eventually obtain the general expression for the expected value of g (xT ), which is T EQ e

−

r(xs ) ds t

g (xT ) = −ıψz (xt , 0, w0 , w, g0 , g, τ ) (5.9) ψz (xt , 0, w0 , w, g0 , g, τ ) . = ı

Thus, we have also derived the result in equation (4.6) within the payoﬀtransform methodology. The remaining building block represents the unconditional expectation under the risk-neutral measure of an integro-linear variable where the payoﬀ T function satisﬁes G(xT ) = g(xs ) ds. Because of the integrated expression t

in the payoﬀ function, this quantity has to be treated diﬀerently. Pricing an unconditional contract, including such an integrated term, we are interested in the expected value EQ e

−

T

r(xs ) ds

T

t

g (xs ) ds .

(5.10)

t

In the following, we ﬁrst want to show how equation (5.10) can be recovered manipulating the expectation itself, as done in equations (4.6) and (4.7). Obviously, the calculations are very similar compared to equation (4.6). Afterwards, we show that the payoﬀ-transform methodology replicates the same result without any problems. Making the same considerations as for the derivation of the expected value of g (xT ), we compute (5.10) as the derivative with respect to the transform variable, evaluated at z = 0. Note that the characteristic function itself consists only of one sole exponential discounting term, since we have T T T EQ e

−

r(xs ) ds ız t

e

g(xs ) ds t

= EQ e

− (r(xs )−ızg(xs )) ds t

.

(5.11)

Obviously, this particular characteristic function is equivalent to the value of a zero-bond contract, but with a hypothetical complex-valued short rate of rA (xt , z) = r (xt ) − ızg (xt ) = (w0 − ızg0 ) + (w − ızg )xt .

(5.12)

5.2 Unconditional Payoﬀ Functions

77

In the last equation we considered that both the instantaneous interest rate r (xt ) and the payoﬀ-characterizing function g (xt ) are linear combinations of xt . Since we deal with a zero bond like contract, the solution for this model price also exhibits an exponential-aﬃne form. Thus, in analogy to the considerations made for zero bonds, we are able to represent the solution as an exponential-aﬃne function. Introducing new parameters characterizing the modiﬁed short rate, we have w0A (z) = w0 − ızg0

and wA (z) = w − ızg.

The resulting characteristic function for pricing average-rate derivatives is then ψ xt , z, w0A (z), wA (z), 0, 0M , τ , and the relevant payoﬀ function for this modiﬁed characteristic function is G(xT ) = 1. As mentioned above, this characteristic function exhibits a strong resemblance compared to the Fourier-style zero-bond representation in equation (5.3), where the original characteristic function was evaluated at some point z = 0. This can be traced back to the fact that both payoﬀ functions are independent of the Fourier Transformation variable. The diﬀerence between them is that the function ψ xt , z, w0A (z), wA (z), 0, 0M , τ generates zero-bond prices with respect to the modiﬁed short rate rA (xt , z), independently of the value of the transformation variable z. Thus, the coeﬃcient functions in this particular case, a(z, τ ) and b(z, τ ) solve again the system of ordinary differential equations (2.40) and (2.41), with terminal conditions a(z, 0) = 0, b(z, 0) = 0M . The hypothetical discount rate is deﬁned by w0A (z) and wA (z), respectively, whereas the terminal value is given by ψ xt , z, w0A (z), wA (z), 0, 0M , 0 = 1.

Having found the characteristic function for this special case, the same considerations can be applied as for the expected value of g (xT ). Using the technique of Fourier-transformed prices, we eventually express equation (5.11)

78

5 Payoﬀ Transformations and European Interest-Rate Derivatives

as110 EQ e

−

T

r(xs ) ds

T

t

t

T − (r(xs )−ızg(xs )) ds d Q t E g (xs ) ds = e dz z=0 ψz xt , 0, w0A (0), wA (0), 0, 0M , τ . = ı

(5.13)

Alternatively, we are also able to obtain this result using the payoﬀtransform methodology together with the contour integration technique. For convenience, we ﬁrst set up the substitution T γ(T ) =

g (xs ) ds, t

and afterwards perform the Fourier Transformation with respect to this new variable γ(T ). Thus, the transformation of the particular payoﬀ function is the same as the one used in deriving equation (5.6). Therefore, we can immediately adopt the result of equation (5.9) by exchanging the general characteristic function ψ (xt , z, w0 , w, g0 , g, τ ) with its modiﬁed pendant ψ xt , z, w0A (z), wA (z), 0, 0M , τ 111 . Afterwards, we get the desired result according to equation (5.13). In this section we proved the general results of unconditional expectations for zero bonds, and linear and integro-linear payoﬀ functions, respectively, obtained within the payoﬀ-transform framework112. Moreover, apart from the traditional formulae, where the desired value is derived by manipulation of the 110

Obviously, ψ

the

values

of

xt , z, w0A (z), wA (z), 0, 0M , τ

the

functions

ψz (xt , z, w0 , w, g0 , g, τ )

and

are equal for z = 0. However, the deriva-

tives with respect to z evaluated at this point, do not share this similarity. This is the reason why we make the dependence of z in the modiﬁed short rate 111

112

explicit, although w0A (0) = w0 and wA (0) = w. The path of the contour integral and the location of the pole is given in Figure 5.2. The particular derivation for the exponential-linear case was not derived in this section since it is not needed in this work. However, the calculations are straightforward using the integration-by-parts methodology, where the relevant pole is at z = ı.

5.2 Unconditional Payoﬀ Functions

79

expectation itself, as shown in Section 4.2, we have with the payoﬀ-transform approach the freedom to choose among a set of inﬁnite solution formulae due to the contour integration in the complex plane. This fact becomes especially important in computing the expectation EQ1 [1] and the expectation for the unconditional average-rate contract where the derivative of the characteristic function with respect to the transformation variable z has to be used. In these cases we are provided with the alternative to use the simple payoﬀ transform and apply equation (4.21) on the appropriate strip in the imaginary plane. Hence, using the building blocks above, we are able to price all interest-rate derivatives introduced in Section 3.2 with Fourier-style formulae. According to the results in equations (5.5), (5.9) and (5.13) we arrive at completely closed-form pricing formulae, which are illustrated in the next subsection113 .

5.2.2 Pricing Unconditional Interest-Rate Contracts So far, the three building blocks for general unconditional payoﬀ functions have been derived. In this section, these blocks are translated into the valuation formulae for the particular yield-based and level-based interest-rate contracts discussed in Section 3.2. Starting with yield-based contracts, we need ﬁrst a translation of yields into Fourier-style solutions. This is easily done as follows Y (xt , t, T ) =

ψ(xt , 0, w0 , w, g0 , g, τ )−1 − 1 . τ

(5.14)

The model price for zero bonds can then be obtained by using equation (5.5), whereas prices of coupon bonds can be calculated as CB(xt , c, t, T) =

A

ψ(xt , 0, w0 , w, g0 , g, τa )ca .

(5.15)

a=1

The price of a forward-rate agreement is given as 113

This statement is valid if the characteristic function or its derivative with respect to z can be displayed in closed form. In cases where the characteristic function cannot be explicitly expressed, but its coeﬃcient functions a(z, τ ) and b(z, τ ) are solutions to the system of ordinary diﬀerential equations according to (2.40) and (2.41), a Runge-Kutta algorithm can be used to obtain the relevant values.

80

5 Payoﬀ Transformations and European Interest-Rate Derivatives

F RAY (xt , K, N om, t, T, Tˆ ) ψ (xt , 0, w0 , w, g0 , g, τˆ) − ψ(xt , 0, w0 , w, g0 , g, τ ) , =N om ˜ K

(5.16)

and a yield-based swap can be similarly computed in terms of the general characteristic functions as SW AY (xt , K, N om, t, T) A−1 ψ (xt , 0, w0 , w, g0 , g, τa+1 ) =N om ˜a K a=1

−

A−1

(5.17)

ψ (xt , 0, w0 , w, g0 , g, τa ) .

a=1

On the other hand, pricing contracts linearly based on the function g (xT ), we foremost need the derivative of the general characteristic function ψ (xt , z, w0 , w, g0 , g, τ ) with respect to z. Hence, a level-based forward-rate agreement deﬁned in equation (3.5) is represented by F RAr (xt , K, N om, t, T ) ψz (xt , 0, w0 , w, g0 , g, τ ) =N om K ψ(xt , 0, w0 , w, g0 , g, τ ) − (5.18) ı φz (xt , 0, w0 , w, g0 , g, τ ) =N om K − ψ(xt , 0, w0 , w, g0 , g, τ ). ı Accordingly, the corresponding swap contract in this framework can be obtained as SW Ar (xt , K, N om, t, T) A =N om K ψ (xt , 0, w0 , w, g0 , g, τa ) a=1

−

A ψz (xt , 0, w0 , w, g0 , g, τa ) a=1

ı

A φz (xt , 0, w0 , w, g0 , g, τa ) =N om K− × ı a=1 ψ (xt , 0, w0 , w, g0 , g, τa ) .

(5.19)

5.3 Conditional Payoﬀ Functions

81

The last unconditional contract to be priced is the average-rate contract. Here, the integro-linear payoﬀ function can be interpreted as an interest-rate contract based on the short rate itself. According to equation (3.11) and (5.13), the price of this contract can be calculated as U ARCr (xt , K, N om, t, T ) =N om K ψ(xt , 0, w0 , w, g0 , g, τ ) −

ψz (xt , 0, w0A (0), wA (0), 0, 0M , τ ) ı

(5.20)

.

For the special case g (xT ) = r (xT ), we use the simpliﬁed versions w0A (z) = (1 − ız)w0 wA (z) = (1 − ız)w, respectively.

5.3 Conditional Payoﬀ Functions So far, we derived closed-form solutions for contracts with unconditional exercise rights. In contrast to the calculations in the last section, where contracts merely depended on the simple evaluation of the terms ψ(xt , 0, w0 , w, g0 , g, τ ), ψz (xt , 0, w0 , w, g0 , g, τ ) and ψz (xt , 0, w0A (0), wA (0), 0, 0M , τ ), respectively, the option-pricing problem confronts us with a diﬀerent situation. The integration by parts method is not of use anymore due to a natural integration boundary, characterized by some strike value K. Including this optional exercise right within the payoﬀ-transform methodology, we end up with some semi closedform solutions, which means we have to solve a standardized Fourier integral in order to compute the desired model prices of interest-rate options. Although the payoﬀ-transform methodology enables us to price consistently the option prices with payoﬀ functions according to Table 4.1, without adapting the valuation formula (4.21) to the diﬀerent cases, we distinguish for convenience between linear, exponential-linear and integro-linear payoﬀ functions. As before, we ﬁrst derive some basic payoﬀ transforms for general g (xT ) and afterwards take into account the interest-rate options discussed in Chapter 3. Eventually, we develop as a special case the Fourier-transformed payoﬀ function of a coupon-bond option for the case of a one-factor interest-rate model114 with xt = rt . 114

The term one-factor model refers to the fact that only one Brownian motion is incorporated in the model.

82

5 Payoﬀ Transformations and European Interest-Rate Derivatives

5.3.1 General Results Besides the elementary payoﬀ functions, we also diﬀerentiate between call and put options, because of the conditional exercise property of the contracts. The transformed payoﬀ functions for call and put contracts display a strong resemblance, which is demonstrated in this section, allowing a more general implementation of the valuation algorithms. Due to the exercise boundary and the diﬀerent ways of incorporating g (xT ) and its integro-linear counterpart, respectively, in the payoﬀ function G(xT ), we introduce the critical value ' ln[K] Exponential-linear Case. α(K) = (5.21) K Linear and Integro-linear Case, for which the option payoﬀ is exactly at the money. The Fourier Transformation for diﬀerent call payoﬀ structures can be generally represented as ∞ F

g(xT )

eızg(xT ) G(xT )1g(xT )≥α(K) dg (xT )

[G(xT )] = −∞ ∞

(5.22)

=

e

ızg(xT )

G(xT ) dg (xT ) ,

α(K)

whereas the particular put payoﬀ transform in its general form is given by ∞ F

g(xT )

eızg(xT ) G(xT )1g(xT )≤α(K) dg (xT )

[G(xT )] = −∞

(5.23)

α(K)

eızg(xT ) G(xT ) dg (xT ) .

= −∞

In deriving the solution for the exponential-linear case, we have to use the transform F

g(xT )

e

g(xT )

−K

+

∞ =

eızg(xT ) eg(xT ) − K dg (xT )

α(K)

Keızg(xT ) e(1+ız)g(xT ) − = 1 + ız ız (1+ız)α(K)

=

1+ız

K e = . ız(1 + ız) ız(1 + ız)

∞ α(K)

(5.24)

5.3 Conditional Payoﬀ Functions

83

Due to the exponential-linear dependence of the payoﬀ-characterizing variable we set α(K) = ln[K] and obtain the equivalent transformation as given in equation (2.26). Since the frequency representation of a call option payoﬀ only exists on a strip with Im(z) > 1, a general Fourier Transformation is needed. Although exhibiting diﬀerent payoﬀ structures the corresponding payoﬀ transform of a put option has the identical formal structure as given in equation (5.24). This can be easily proved by F

g(xT )

K −e

g(xT )

+

α(K)

eızg(xT ) K − eg(xT ) dg (xT )

=

(5.25)

−∞ 1+ız

=

K , ız(1 + ız)

but with Im(z) < 0. Based on this result, both call and put option prices can be recovered using the same payoﬀ transform and as a direct consequence, only one single program code is needed for evaluating values for both interest-rate option contracts. The only diﬀerence are the diﬀerent sets and strips on which Im(z) is valid for the inverse operation. Whereas the condition for the call contract assured the dampening of the integrand on the positive half-axis, we need for the put option the condition to guarantee the same on the negative equivalent. An interesting feature of the payoﬀ-transform methodology is, due to the equivalent transformed payoﬀ functions of calls and puts, the applicability of a closed contour integral to obtain in a very elegant way the particular put-call parity115 . Without loss of generality, we set f (z) =

K 1+ız ψ(xt , −z, w0 , w, g0 , g, τ ) . 2πız(1 + ız)

Thus, we have 115

The relevant integration path is depicted in Figure 5.3.

(5.26)

84

5 Payoﬀ Transformations and European Interest-Rate Derivatives

Fig. 5.3. Closed contour integral path for the derivation of the put-call parity in equation (5.27). The poles at z = 0 and z = ı are completely encircled in a clockwise manner.

EQ e

−

T

r(xs ) ds t

T + + − r(xs ) ds eg(xT ) − K − EQ e t K − eg(xT ) ∞

=

f (z) dz + −∞

−∞ f (z) dz

(5.27)

∞

= −2πı (Res [f (z)|z = 0] + Res [f (z)|z = ı]) , with Im(z) > 1. The imaginary part of the Fourier variable z in equation (5.27) can be chosen arbitrarily as long as the existence of the payoﬀ transformations is guaran-

5.3 Conditional Payoﬀ Functions

85

teed116 . Obviously, this clockwise performed contour integral now encircles two simple poles of the function f (z), one at the origin and the other one located at z = ı. Due to the closed contour, we only have to calculate the residues of all included poles in order to obtain the desired put-call parity. Comparing equation (5.26) with (5.5), the residue of f (z) at the origin is just Res [f (z)|z = 0] = K

ψ(xt , 0, w0 , w, g0 , g, τ ) , 2πı

whereas the residue at z = ı is Res [f (z)|z = ı] = −

ψ(xt , −ı, w0 , w, g0 , g, τ ) . 2πı

Hence, equation (5.27) equals T T + + − r(xs ) ds − r(xs ) ds EQ e t eg(xT ) − K − EQ e t K − eg(xT )

(5.28)

= ψ(xt , −ı, w0 , w, g0 , g, τ ) − K ψ(xt , 0, w0 , w, g0 , g, τ ). According to the result in equation (4.7), the term ψ(xt , 0, w0 , w, g0 , g, τ ) simply represents the price of a zero bond with maturity τ . The other term, the quantity ψ(xt , −ı, w0 , w, g0 , g, τ ) equals the discounted forward price of the exponential of the variable g (xt )117 . Therefore, setting z = −ı, we get T EQ e

−

r(xs ) ds t

eg(xT ) .

For a call option, linearly based on g (xT ), we get

∞ 1 + ız(K − g (xT )) + F g(xT ) (g (xT ) − K) = eızg(xT ) z2 α(K) eızα(K) eızK =− =− 2 , 2 z z

(5.29)

with 116

For convenience, we work with the complex conjugate for the latter integral. In fact, due to the exponential-linear payoﬀ function the restriction for the put

117

option transform can be independently chosen according to equation (5.25). See, for example, Bakshi and Madan (2000), p. 212. There, this quantity is alternatively denoted as the scaled-forward price.

86

5 Payoﬀ Transformations and European Interest-Rate Derivatives

Im(z) > 0. Similar to the call representation in Fourier space, the put option transform is eızK F g(xT ) (K − g (xT ))+ = − 2 . (5.30) z The only diﬀerence between the call and put option transform is that equation (5.30) is deﬁned on the opposite imaginary half-plane. Consequently, we use the complex conjugate of the Fourier variable in equation (5.29). The put-call parity for the linear case can be derived as118 T T EQ e

−

r(xs ) ds t

+ (g (xT ) − K) − EQ e

−

r(xs ) ds t

+ (K − g (xT ))

eızK ψ(xt , −z, w0 , w, g0 , g, τ ) = ıRes z = 0 z2 ψz (xt , 0, w0 , w, g0 , g, τ ) − K ψ(xt , 0, w0 , w, g0 , g, τ ). = ı

(5.31)

Due to the payoﬀ similarities of the linear and integro-linear case, the payoﬀ transformations are equivalent for both cases in Fourier space. Hence, to compute the average-rate option prices (3.24) and (3.25), equations (5.29), (5.30) and (5.31) can be used together with the modiﬁed characteristic function. Although not directly applicable for tradable option contracts, but nevertheless important for theoretical issues is the Fourier-transformed payoﬀ function of a hypothetical contingent claim according to the Dirac delta function, which is δ(g (xT ) − α(K)). As mentioned before, the Dirac delta function has an inﬁnite spike for g (xT ) = α(K). The Fourier Transformation of the Dirac delta function can be simply expressed as F g(xT ) [δ(g (xT ) − α(K))] = eızα(K) ,

(5.32)

with no need to set up any restriction on the imaginary part of the transform variable z. Since the Dirac delta function states the terminal condition of a probability density function, equation (5.32) may be used to recover the relevant transition density function. Especially for illustrating the behavior of a particular stochastic process g(xt ), the transition density function is useful to explain its characteristics. The other special function we want to derive, is 118

The relevant integration path is depicted in Figure 5.2.

5.3 Conditional Payoﬀ Functions

87

the Fourier Transformation of the cumulative probability function Pr(g(xT ) < α(K)). The payoﬀ corresponding to this terminal condition is given by the indicator function of the event g(xT ) < α(K). Thus, the transformed payoﬀ can be expressed as119 eızα(K) , F g(xT ) 1g(xT ) 1. In contrast, a zero-bond put option price can be derived via equation (5.35) but with the restriction Im(z) < 0.

According to equation (3.17) and (3.18), a yield-based cap and ﬂoor contract can be immediately expressed as the summation over the particular zero-bond options, scaled with some quantity NKom . Hence, the model price of a

a yield-based cap contract is 121

However, there exist some articles which derive approximated values for these contracts in a multi-factor framework, see e.g. Singleton and Umantsev (2002) or Collin-Dufresne and Goldstein (2002).

5.3 Conditional Payoﬀ Functions

CAPY (xt , K, N om, t, T) ız A−1 ∞ Ka N om × = π a=1 ız(1 + ız)

89

(5.36)

0

ψ(xt , −z, w0 , w, a(0, τˆa ), b(0, τˆa ), τa ) dz, with Im(z) > 0, τˆa = Ta+1 − Ta , and τa = Ta − t. Subsequently, a yield-based ﬂoor contract can be priced using equation (5.36) with Im(z) < 0.

Next, we derive the particular pricing formulae of level-based interestrate contracts and interest-rate options written on the short rate r (xt ) itself. Starting with a cap contract according to equation (3.15), we use the payoﬀ transform (5.29) with g0 = w0

and

g = w,

and therefore apply the characteristic function ψ(xt , z, w0 , w, w0 , w, τa ). Thus, the cap contract can be priced as CAPr (xt , K, N om, t, T) A ∞ N om eızK =− ψ(xt , −z, w0 , w, w0 , w, τa ) dz, π a=1 z2

(5.37)

0

with Im(z) > 0. Hence, the model price of a ﬂoor contract with equivalent input parameters can be recovered using equation (5.37) again but evaluating the integrals on the negative imaginary half-plane with Im(z) < 0.

90

5 Payoﬀ Transformations and European Interest-Rate Derivatives

The last option contracts for which we want to give a payoﬀ-transformed solution are the average-rate options due to equation (3.24) and (3.25). Thus, the payoﬀ of the average-rate cap option contract at expiration can be expressed as + T N om ∗ K − r(xs ) ds , τ t

with K ∗ = τ K. Taking the same considerations into account as done for the unconditional average-rate contract, the relevant characteristic function for T r(xs ) ds,

γ(T ) = t

is given by ψ(xt , z, w0A (z), wA (z), 0, 0M , τ ). Together with the payoﬀ transform in equation (5.29), we are able to postulate the model price of an average-rate cap as ARCr (xt , K, N om, t, T ) ∞ ızK ∗ (5.38) N om e A A =− ψ(x , −z, w (−z), w (−z), 0, 0 , τ ) dz, t M 0 τπ z2 0

with Im(z) > 0. The respective average-rate ﬂoor contract can be priced, using equation (5.38) with Im(z) < 0. 5.3.3 Pricing of Coupon-Bond Options and Yield-Based Swaptions So far, we have excluded the valuation formulae for coupon-bond options and yield-based swaptions, respectively. In contrast to the option contracts discussed in the last section, where we computed only a single option price and a portfolio of diﬀerent option prices, respectively, we deal here with a option on a portfolio of future cash ﬂows. Consequently, the determination of

5.3 Conditional Payoﬀ Functions

91

a unique critical exercise value α(K) in a multi-factor setting is not possible anymore122 . However, dealing with a one-factor interest-rate model setup with r(xt ) = rt 123 , we are able to circumvent this issue. Hence, we follow the technique proposed in Jamshidian (1989) to derive the theoretical price of a coupon-bond option using the payoﬀ-transform methodology in pricing this derivative contract, which is shown below. Setting xt = rt , we are able to exploit the coeﬃcient structure of the aﬃne term-structure model. The special form of the characteristic function is of the form T ψ(rt , z, 0, 1, 0, 1, τ ) = EQ e

−

rs ds+ızrT t

= ea(z,τ )+b(z,τ )rt .

Because a yield-based swaption can be interpreted as an option on a coupon bond124 , we focus on the valuation of the particular coupon-bond option. In a one-factor setup the coupon-bond call option payoﬀ is given by +

(CB(rT , c, T, T) − K) = =

A

+ P (rT , T, Ta )ca − K

a=1 A

+ e

a(0,τa )+b(0,τa )rT

ca − K

.

a=1

In the last equation, we inserted the particular Fourier-style zero-bond prices generated by the exponential-aﬃne model. The above presented payoﬀ function is then a continuous and strictly decreasing function in rT 125 . In these models we have126 ∂P (rt , t, T ) = b(0, τ ) < 0 ∂rt

∀ T > t.

Consequently, the payoﬀ function exhibits a unique zero value for the critical short rate rT∗ for which the coupon-bond call is exercised. However, dealing 122 123 124 125 126

See, for example, Singleton and Umantsev (2002). Without loss of generality, we set in the following w0 = 0 and w1 = 1. See the alternative presentation of a swaption payoﬀ in Section 3.3. The particular characteristic functions are derived in Chapter 8. See e.g. Duﬃe and Kan (1996) for the properties of b(0, τ ) in common one-factor interest-rate models.

92

5 Payoﬀ Transformations and European Interest-Rate Derivatives

with a single-factor environment, we cannot explicitly express this critical value rT∗ in closed form, which is due to the sum of exponentials in the payoﬀ function. Thus, the critical exercise value has to be computed numerically. Having determined the value of rT∗ , the Fourier Transformation of a couponbond call payoﬀ can be calculated as127 F rT (CB(rT , c, t, T, T) − K)+

∗

rT =

e −∞

=e

ızrT

∗ ızrT

A

e

a(0,τa )+b(0,τa )rT

ca − K

a=1 ∗ A ea(0,τa )+b(0,τa )rT

a=1

b(0, τa ) + ız

K ca − ız

drT

(5.39)

,

with Im(z) < min [b(0, τa )] . a

Note that in contrast to the valuation formula a zero-bond call option, where the Fourier Transformation of the payoﬀ function was made with respect to g(xT ), we now perform the transform operation with respect to rT . Therefore, we need a diﬀerent restriction for the imaginary part of the transform variable z. Because the coeﬃcient b(0, τa ) is generally negative, we take the smallest value of b(0, τa ) as an upper bound for the domain of valid values for Im(z), which is due to the monotonicity simply b(0, τA ). Eventually, using the general valuation formula (4.21), we are able to compute the price of a coupon-bond call option as CBC (rt , c, K, t, T, T) A ∞ ea(0,τa )+b(0,τa )rT∗ ∗ K 1 ızrT ca − × = e π b(0, τa ) + ız ız a=1

(5.40)

0

ψ(rt , −z, 0, 1, 0, 1, τ ) dz. As before, the payoﬀ transform of the particular put option is also given by equation (5.40), but with the slightly modiﬁed restriction Im(z) > 0. 127

Since the integration variable is no longer g (xT ), we have to switch the integration boundaries, due to the negativeness of b(0, τ ).

5.3 Conditional Payoﬀ Functions

93

Having derived the proper Fourier Transformation of a coupon-bond option payoﬀ, the equivalent expression for a yield-based swaption contract is given by the alternative representation of a swaption contract according to equation (3.23), with coupon payment vector cSW P and payment dates contained in T∗ . On the other hand, the particular forward-start payer swaption can be interpreted as a coupon-bond put option with strike one and the same coupon payment vector and the same payment dates as used before. Hence, for the transformed payoﬀ function to be existent, we have to ensure that the inequality Im(z) > 0 holds.

6 Numerical Computation of Model Prices

6.1 Overview In this chapter we develop a new pricing algorithm to compute model prices for the derivatives contracts previously discussed. Here, we distinguish, as before, between contracts with unconditional and conditional exercise rights. The distinction is made because of the separate fundamental calculation procedure for these prices. Whereas derivatives with unconditional exercise rights can be calculated in terms of the general characteristic function ψ(xt , z, w0 , w, g0 , g, τ ) and in terms of the relevant moment-generating function128 , respectively, without evaluating any integral at all if the characteristic function is known in closed form, we need for option-type contracts to apply a numerical integration scheme in order to calculate their model prices. Carr and Madan (1999) showed in their prominent article a very convenient method to compute option prices for a given strike range, using the FFT. The advantage in applying the FFT to option-pricing problems, is its considerable computational speed improvement compared to other numerical integration schemes. Due to the payoﬀ transform methodology, we use another pricing algorithm, which shares the same desirable, numerical properties of the FFT. Unfortunately, implementing the pricing approach according to Lewis (2001), it is necessary to impose the transform with respect to the strike. Therefore, one cannot use the FFT any longer to obtain option prices in one pass for a strike range129 . 128 129

See Section 5.2. See Lee (2004), p. 61. However, comparing the structure in equation (4.21) it is possible to obtain model prices with the help of a FFT procedure for diﬀerent levels of g (xt ).

96

6 Numerical Computation of Model Prices

In order to circumvent this problem within the payoﬀ-transform pricing approach, we need an another numerical algorithm. Therefore, we incorporate in our pricing algorithm the IFFT, to compute model prices for diﬀerent strike values130 . Furthermore, to enhance the quality of results131 , the fractional Fourier Transform of Bailey and Swarztrauber (1994) is used. This reﬁnement was introduced by Chourdakis (2005) in pricing equity option prices with the transformed option price methodology of Carr and Madan (1999). However, we sometimes encounter the problem that ψ(xt , z, w0 , w, g0 , g, τ ) cannot be calculated in closed form132 . For these cases, we implement a RungeKutta solver in our IFFT pricing algorithm. This algorithm is then used to compute the relevant values for diﬀerent z in ψ(xt , z, w0 , w, g0 , g, τ ) by solving the ODEs (2.40) and (2.41) numerically and providing the procedure with the needed values.

6.2 Contracts with Unconditional Exercise Rights As explained in Section 5.2.2 all contracts with unconditional exercise rights can be calculated as mere function evaluations of the general characteristic function ψ(xt , z, w0 , w, g0 , g, τ ), its ﬁrst order derivative with respect to z, and for integro-linear payoﬀ functions with the help of the ﬁrst order derivative ψz xt , z, w0A (z), wA (z), 0, 0M , τ . As shown, these unconditional expectations can be obtained by contour integration in closed form. Thus, we do not need to develop a numerical integration routine at all in order to calculate the relevant model prices. The calculations reduce in these cases to T EQ e 130

−

r(xs ) ds t

= ψ(xt , 0, w0 , w, g0 , g, τ ),

We ﬁnd it natural to use the FFT and the IFFT algorithm to obtain the desired Fourier Transformation. Other numerical integration schemes are also possible, like for example the numerical integration via Laguerre polynomials as used in

131

Tahani (2004). The ordinary IFFT pricing algorithm suﬀers, like the particular FFT algorithm, from the ﬁxed scale of increments of strike values and transformation variable,

132

which is discussed in Section 6.3.1. This could be the case e.g. for some subordinated processes rt or for jump components where EJ [ψ ∗ (z, w0 , w, g0 , g, J, τ )] cannot be solved explicitly.

6.3 Contracts with Conditional Exercise Rights

E e Q

and

EQ e

−

T

−

97

r(xs ) ds

t

g (xT ) =

ψz (xt , 0, w0 , w, g0 , g, τ ) , ı

T

r(xs ) ds t

ψz (xt , 0, w0A (0), wA (0), 0, 0M , τ ) , γ(T ) = ı

for arbitrary times to maturity τ . For normal contracts, the discount rate used in the characteristic function is based on the short rate r (xt ) and is zero for futures-style contracts. In case of an average-rate contract where the underlying is the geometric average of the short rate, we have to use the characteristic function with a modiﬁed discount rate rA (xt ). If the general characteristic function cannot be expressed in closed form although deﬁned by a system of ODEs, we apply a numerical algorithm to evaluate the needed values. In this case we implement a Runge-Kutta solver for the system of ODEs (2.40) and (2.41).

6.3 Contracts with Conditional Exercise Rights 6.3.1 Calculating Option Prices with the IFFT We start with the integral representation of the general option valuation formula (4.21). Since we are interested in calculating option prices in one pass for a given strike range simultaneously with the IFFT, we have to reduce the presence of K in the integral to the expression eızα(K) for both exponential-linear, linear, and integro-linear type payoﬀ functions. In the case of coupon-bond options and swaptions we have to divide the payoﬀ function up into A separate parts. The alternative representation of the valuation formula is eα(K)d V (xt , t, T ) = π

∞ eızα(K) gˆ(z)ψ(xt , −z, w0 , w, g0 , g, τ ) dz,

(6.1)

0

with F g(xT ) [G (xT )] = e(d+ız)α(K) gˆ(z), and α(K) = K for the case of a ﬂoating-rate based contract and an asian-type contract, respectively, and α(K) = ln[K] for a yield-based contract133 . The 133

See equation (5.21).

98

6 Numerical Computation of Model Prices

parameter d is chosen in a way to eliminate all dependency of α(K) in gˆ(z), which is crucial for the IFFT algorithm to work properly134 . A ﬁrst problem might arise using multi-valued functions, e.g. the complex-valued logarithm, square-root, and the conﬂuent hypergeometric function KU(a; b; y). Thus, we have to carefully keep track of the integration path to avoid any discontinuities135 . However, using a numerical algorithm to compute the particular values of the characteristic function such as a Runge-Kutta algorithm we do not encounter these problems136 . The ﬁrst step in deriving the IFFT pricing algorithm is to truncate the integration domain as ω eızα(K) gˆ(z)ψ(xt , −z, w0 , w, g0 , g, τ ) dz.

f (α(K)) ≈

(6.2)

0

Applying an U -point approximation with increment ∆ =

ω U,

we discretize the

domain of the transform variable into 1 zu = u − ∆ + ızi 2 with u = 1, . . . , U and zi corresponding to a ﬁxed value for which the Fouriertransformed payoﬀ function exists. The integration interval [0, ∞] is then replaced with a discrete, truncated region such that the integrand of f (α(K)) is negligible for zU . Hence, the discrete approximation to equation (6.2) is f (α(K)) ≈

U

eızu α(K) gˆ(zu )ψ(xt , −zu , w0 , w, g0 , g, τ ) ∆

u=1

= ∆e

−zi α(K)

U

(6.3) e

ı(u−1) ∆α(K)

e

ı∆ 2

α(K)

gˆu ψu ,

u=1 134

Otherwise, the IFFT algorithm is not applicable to the valuation problem at hand. Fortunately, we are able to reduce the dependency of K in the particular

135 136

integrals to the speciﬁc term eızα(K) , for all contracts discussed in Chapter 3. This topic is covered comprehensively in Nagel (2001), Appendix 4. In case of the Fong and Vasicek (1991a) model, we made the same experience as mentioned in Tahani (2004), Footnote 4, and compute values of the characteristic function with help of an explicit Runge-Kutta algorithm in the ﬁrst place. Thus, besides the prevention of discontinuities, the Runge-Kutta algorithm can be more eﬃcient than the explicit computation of the conﬂuent hypergeometric function.

6.3 Contracts with Conditional Exercise Rights

99

with gˆu = gˆ(zu )

ψu = ψ(xt , −zu , w0 , w, g0 , g, τ ).

and

The sum above is commonly referred to as a discrete inverse Fourier Transı∆

formation137 of the function e 2 α(K) gˆu ψu . We also want to mention that in computing this sum we eventually obtain the option price for only one particular strike value K. Since we are interested in calculating option prices for a strike range we also have to discretize α(K), which yields αv = α(K1 ) + (v − 1)η, with step size η and v = 1, . . . , U 138 . Thus, inserting the explicit expression for αv inside the brackets of equation (6.3) gives f (αv ) = ∆ e−zi αv

U

eı(u−1) ∆(α1 +(v−1)η) e

ı∆ 2 (α1 +(v−1)η)

gˆu ψu

u=1

= ∆e

−zi αv

e

ı ∆η 2 (v−1)

U

(6.4) e

ı(u−1)(v−1) ∆η ı ∆α1 (u− 12 )

e

gˆu ψu .

u=1

The form of f (αv ) is almost ready to be inserted into the IFFT algorithm. The IFFT algorithm is developed to calculate simultaneously the discrete inverse Fourier Transformation for a range of values αv . The main advantage is that it reduces the number of calculations from an order of U 2 to the order of U log2 [U ], which makes a signiﬁcant diﬀerence in computational speed139 . It eﬃciently computes the sum U 1 ı(u−1)(v−1) 2π U h e f (v, h) = u U u=1 137

for v = 1, . . . , U.

(6.5)

Although we deﬁned the transform operations in Section 2.4 vice versa, in this chapter we rely on the term discrete inverse transform, which belongs to engineering disciplines and is in line with the expression used afterwards for the

138

IFFT. We use the same discretization scheme for α(K) as used in Lee (2004). The advantage, in contrast to the discretization schemes applied in Carr and Madan (1999) and Raible (2000), is the possibility to adjust the numerical scheme for the lower bound of the strike rates. Thus, one does not necessarily have to compute

139

option prices for negligible strike rates, which is a more eﬃcient procedure. See Cooley and Tukey (1965).

100

6 Numerical Computation of Model Prices

Introducing the vectors

1 2 u=v= .. , . U

equation (6.5) can be displayed in a more compact form, which is f (h) = IFFT[h],

(6.6)

with h ∈ CU . By comparing equation (6.5) with (6.4), we obviously need the relation ∆η =

2π , U

in order to apply the IFFT algorithm properly to equation (6.4). Because

2π U

remains constant for a ﬁxed number of points U , we have only the freedom to choose either ∆ or η independently. Thus, there is a tradeoﬀ between the accuracy of the calculated results and the coarseness of the strike-value grid. According to these considerations, more accurate results of option prices corresponding to speciﬁc strike rates have to be paid with more points in the integration scheme due to the rule U × 2n . This rule ensures that the algorithm computes option prices for speciﬁc strike values and illustrates the exponential cost for more accurate results. Calculating the same number of option prices, most of them outside a desired strike range, entails a substantial waste of computational time140 . To give a more compact writing, we use henceforth the vectors α = U (αv )U g = (ˆ gu )U v=1 , ˆ u=1 and ψ = (ψu )u=1 . Eventually, the vector V(xt , t, T ) containing the option values for diﬀerent strikes, can be computed as

V(xt , t, T ) =

U ∆ e(d−zi )α π Re e

πı U (v−1)

IFFT[e

ı ∆α1 (u− 12 )

ˆ g ψ] ,

(6.7)

where the operator denotes the vector-dot product of two arbitrary vectors of the same length. This pricing algorithm is already capable of calculating 140

This particular problem is addressed in the next section.

6.3 Contracts with Conditional Exercise Rights

101

option prices. However, as stated before, equation (6.7) displays the problem of computing option prices for many irrelevant strike rates, given a desired level of accuracy. 6.3.2 Reﬁnement of the IFFT Pricing Algorithm The purpose of this subsection is to solve the problem of the inverse relationship of ∆ and η mentioned in the last section. The numerical eﬃciency can be enhanced by using a modiﬁed version of the ordinary IFFT algorithm to ensure that all calculated option prices are at least within an interval of relevant strike values. Bailey and Swarztrauber (1994) developed a method based on the FFT to choose ∆ and η independently. Their method, called the fractional Fourier Transformation, henceforth denoted as the FRFT, incorporates a new auxiliary parameter ζ 141 , which successfully dissects the otherwise ﬁxed relation ∆η ≡ 2π U . Chourdakis (2005) used this reﬁned algorithm in pricing European options on equities based on the Carr and Madan (1999) pricing framework. The FRFT was developed to eﬃciently compute the sum f (v, h, ζ) =

U

e−2πı(u−1)(v−1)ζ hu

for v = 1, . . . , U.

(6.8)

u=1

Thus, introducing the FRFT operator, we deﬁne the compact expression f (h, ζ) = FRFT [h; ζ] . Although, the parameter ζ is usually real-valued, it is not restricted to the set of R. Obviously, the FRFT is strongly connected to the FFT and the IFFT. For example, by comparing equation (6.5) with (6.8), we have the equivalence

1 1 IFFT [h] ≡ FRFT h; − . U U

The key insight to compute the FRFT in terms of the FFT and the IFFT algorithm, respectively, is to recognize that the product 2(u − 1)(v − 1) can be expressed as 141

The fractional Fourier Transformation parameter ζ in this thesis corresponds to α in the original article of Bailey and Swarztrauber (1994).

102

6 Numerical Computation of Model Prices

(u − 1)2 + (v − 1)2 − (v − u)2 . Inserting this relation into equation (6.8), subsequently doing some algebraic transformations and using the discrete version of the convolution theorem of Fourier Transformations142 , we are able to eﬃciently compute equation (6.8) with the help of both the FFT and the IFFT algorithm as follows143 . Deﬁning the vectors p and q with elements ' pu = and

' qu =

hu au

for

1≤u≤U

0

for

U < u ≤ 2U,

for au a(2U+2−u) for

1≤u≤U U < u ≤ 2U,

with 2

au = eıπζ(u−1) , we compute ﬁrst the raw transformation as ˆ f (h, ζ) = IFFT [FFT [p] FFT [q]] . The last U elements in ˆ f (h, ζ) can be discarded due to the zero padding made in the vector p. Thus, we store the ﬁrst half of the vector ˆ f (h, ζ) in a new − ˆ vector f (h, ζ). The FRFT is then f (h, ζ) = ˆ f − (h, ζ) a−u .

(6.9)

Obviously, by comparing the term inside the sum operator in equation (6.4) with the corresponding term inside the sum in equation (6.8) we have to establish the relation

∆η , 2π where both ∆ and η can be chosen arbitrarily144. Thus, our general optionpricing formula (6.7), can be rewritten in terms of the FRFT as ζ=−

142 143

144

See Proposition 2.4.3. The detailed derivation of the FRFT algorithm is given in Bailey and Swarztrauber (1994). Note that the factor

1 U

used in equation (6.4) is already included in ∆.

6.3 Contracts with Conditional Exercise Rights

103

V(xt , t, T ) =

∆ e(d−zi )α π

1

Re e−πı(v−1)ζ FRFT eı ∆α1 (u− 2 ) ˆ g ψ; −

∆η 2π

(6.10)

.

Although we have to compute two FFTs and one IFFT in order to obtain one FRFT, there is a substantial improvement due to the now independent choice of strike interval and integration domain, which saves in the end computer time. This fact becomes more important for the computation of characteristic functions for which no closed-form solutions exist and therefore the system of ODEs (2.40) and (2.41) must be solved numerically for each sampling point zu . 6.3.3 Determination of the Optimal Parameters for the Numerical Scheme As discussed in Lee (2004) and Lord and Kahl (2007), the choice of zi , determining the speciﬁc contour in the complex plane used for the numerical integration routine is crucial in computing option prices. Lee (2004) ﬁnds that for diﬀerent option payoﬀ functions, for diﬀerent strike values and driving processes, respectively, the optimal value of zi , thus minimizing the numerical error, varies substantially145 . Furthermore, the parameter ω concerning the truncation error is also of the utmost importance in a numerical option-pricing scheme. Thus, both parameters inﬂuence the accuracy of numerical solutions. This is illustrated in Figure 6.1 for zero-bond call options and the jumpenhanced models of Vasicek (1977) and Cox, Ingersoll and Ross (1985b)146 . Obviously, setting ω too small results in a highly oscillating solution vector. On the other hand choosing ω too high, the absolute error of the numerical so145

146

See Lee (2004) Table 2 and 3. The same observation is made in Lord and Kahl (2007), Figure 1. Both interest-rate models are enhanced with an exponentially distributed jump component. The coeﬃcients for the characteristic function of the jump-enhanced Vasicek model are given in equations (8.6), (8.7), and (8.8). The particular coefﬁcients in case of the jump-enhanced CIR model are given in equations (8.11), (8.12), and (8.13). A discussion of these models is given in Chapter 8.

104

6 Numerical Computation of Model Prices

lutions increase exponentially. The opposite statement holds for zi . Therefore, these parameters should be chosen to avoid minimize both eﬀects.

−12

x 10

x 10

−8

2 1

1

absolute error

absolute error

2

0

−1

0 −1 −2

−2 1400

−3 1600 1200 1000 800 600

ω

x 10

60

65

75

70

80

85

90

1400 1200 1000

ω

K

−12

60

65

75

70

80

85

90

K

−8

x 10

2

5

absolute error

absolute error

1 0 −1 −2

0

−5

−3 −4 25

−10 16 20 15

zi

10

60

65

75

70

K

80

85

90

14 12

zi

10

60

65

75

70

80

85

90

K

Fig. 6.1. Graphs in the ﬁrst row depict absolute errors of 512 zero-bond call prices for alternating values of ω. In the second row, the particular errors are depicted for varying values of zi . An exponential-jump version of the Vasicek (CIR) model is used in the left (right) column. The parameters are: rt = 0.05(0.03), κ = 0.4(0.3), θ = 0.05(0.03), σ = 0.01(0.1), η = 0.005(0.005), λ = 2(2), τ = 0.5(0.5), τˆ = 2(2).

Since we want to price a vector of option prices with the computation of one FRFT operation, thus considering one speciﬁc parameter setting for the entire strike range, we are interested in ﬁnding the optimal parameter setting for the pricing algorithm, (ω ∗ , zi∗ ), which minimizes the overall numerical error in equation (6.10). Hence, we need a criterion which measures the cumulative error of both positive and negative deviations from the theoretical solutions. Consequently, we apply in the following analysis the root mean-squared error (RMSE), which is

6.3 Contracts with Conditional Exercise Rights

( RMSE =

(VN um − VT rue ) (VN um − VT rue ) , U

105

(6.11)

where VN um denotes some numerical solution vector and VT rue represents the corresponding vector of closed-form solutions. To give an idea of the error behavior of the FRFT pricing algorithm, we ﬁrst compare quasi closed-form solutions computed with the QUADL integration routine in MATLAB147 according to equation (6.1) and the corresponding values due to the FRFT algorithm as deﬁned in equation (6.10) for a ﬁxed number of 512 diﬀerent strike rates. The particular natural logarithms of the RMSE for zero-bond call option prices are depicted in Figure 6.2. We make two remarkable observations. Firstly, for diﬀering values of ω and zi both models have a global minimum of the RMSE of computed option prices. Secondly, the logarithmic presentation of the RMSE implies a rapid and monotonic descent towards this minimum, starting with small values of ω and zi 148 . In case of the jumpenhanced CIR model, the speciﬁc error-minimizing parameter couple is clearly evident according to the contour plot of the logarithmic RMSE given in the lower right graph of Figure 6.2. On the other hand, the particular contour plot of the logarithmic RMSE for zero-bond call options under the jump-enhanced Vasicek model also clearly indicates a region of parameter couples exhibiting approximately the same RMSE magnitude. Consequently, we exploit this monotonic decrease of the RMSE to develop an algorithm, which is capable of ﬁnding an optimal parameter setting (ω ∗ , zi∗ ) and simultaneously giving an estimate of the magnitude of errors of numerical solutions even when the closed-form solutions are not known. The technique we use for the approximation of the numerical error is based on the exponential decreasing of the mean-squared error between two successive parameter values in the numerical scheme. Thus, we deﬁne the approximate RMSE as ) (VN um − VN um(+) ) (VN um − VN um(+) ) RMSEa = , (6.12) U where VN um and VN um

(+)

denote numerical solutions of two successive pa-

rameter values, whether in ω or in zi direction. 147

148

This integration routine uses an adaptive Lobatto quadrature scheme. In the calculation of quasi closed-form solutions, we set its error tolerance to 10−15 . This phenomenon shows up for all interest-rate model/payoﬀ combinations mentioned in this thesis.

106

6 Numerical Computation of Model Prices 80 20

70 30 60

10

10

50 i

0

z

ln(RMSE)

20

0 40

−10 −20

30

−30 80

20 60

−10

−20

10000 8000

40

10

6000 4000

20 0

z

−30

2000 0

1000

ω

i

2000

3000

4000

5000

ω

6000

7000

8000

9000 10000

80 25 70 20

40 60

30

15 10

50 10

5 i

0

z

ln(RMSE)

20

40

0

−10 −5

30

−20

−10

−30 80

20 60

−15

10000 8000

40

6000

−20

4000

20

zi

10

0

−25

2000 0

ω

1000

2000

3000

4000

5000

ω

6000

7000

8000

9000 10000

Fig. 6.2. Logarithmic RMSEs of 512 zero-bond call option prices. In the upper (lower) row the underlying interest rate is modeled by a jump-enhanced Vasicek (CIR) model. The parameters are: rt = 0.05(0.03), κ = 0.4(0.3), θ = 0.05(0.03), σ = 0.01(0.1), η = 0.005(0.005), λ = 2(2), τ = 0.5(0.5), τˆ = 2(2) and a strike range of K ∈ [60, 90].

In Figure 6.3, diﬀerences of the logarithmic RMSEa , for two successive parameter values of zi , and the logarithmic RMSE according to equation (6.11) are depicted for zero-bond call prices for varying zi values. Obviously, the approximate and exact RMSEs show nearly the same magnitude until the minimum RMSE is reached. Afterwards, the diﬀerence, still very small, becomes oscillating in case of the Vasicek model and experiences a decrease of its level in case of the CIR model, respectively. This characteristic behavior of the RMSEa is used in our algorithm to ﬁnd the optimal parameter couple (ω ∗ , zi∗ ) and enables the formulation of an approximate error bound for the numerical solution vector. As mentioned above, our algorithm to ﬁnd the optimal parameter couple (ω , zi∗ ) utilizes a steepest descent technique on the logarithm of the RMSEa . ∗

0

−20

−40

−1

0

2

4

6

8

10

12

14

16

18

12

1

0

0

−2 20

z

i

−12

−24

−1

0

2

4

6

8

10

12

14

16

18

ln(RMSEa)−ln(RMSE)

0

107

ln(RMSE)

1

ln(RMSE)

20

ln(RMSEa)−ln(RMSE)

6.3 Contracts with Conditional Exercise Rights

−2 20

z

i

Fig. 6.3. The dashed line represents the diﬀerence of the logarithmic RMSEa and the exact RMSE of 512 zero-bond call option prices and increasing values of zi . Both graphs are drawn for ω = 1400. The straight line depicts the logarithmic RMSE in dependence of zi . The underlying model in the left (right) graph is a jumpenhanced Vasicek (CIR) model with parameters: rt = 0.05(0.03), κ = 0.4(0.3), θ = 0.05(0.03), σ = 0.01(0.1), η = 0.005(0.005), λ = 2(2), τ = 0.5(0.5), τˆ = 2(2) and a strike range of K ∈ [60, 90].

Thus, initializing the algorithm, we ﬁrst evaluate the numerical solution VN um for some parameter values (ω o , zi0 )149 . Subsequently, we compute two additional solution vectors for ascending parameter values in the direction of both ω and zi which are then used to derive the particular ﬁrst order ﬁnite differences. Afterwards, if the slope in ω direction is smaller than the one in zi direction, thus more negative, the next numerical solution is computed with an exalted ω and vice versa. The next step in the numerical scheme is then again to obtain the necessary numerical solution vectors in order to derive the particular ﬁnite diﬀerences and so on. The algorithm aborts if the smallest value of ln(RMSEa ) is reached over some interval where the curve experienced its reversal point. In Figure 6.4, the paths with an initial value of zi0 = 2 and ω 0 = 10 for the jump-enhanced Vasicek and CIR model are shown. Obviously, the algorithm ﬁnds for both interest-rate models the optimal parameter setting, which can be justiﬁed by the graphs in the left column of Figure 6.4. In case of the optimal parameter couple using the jump-enhanced Vasicek (CIR) model, we get a diﬀerence of exact and approximate RMSEs of 9.02924×10−14 149

Since we observe the steepest descent starting at the origin the initial value for zi0 and ω 0 has to be near the origin subject to the particular regularity conditions of the Fourier-transformed payoﬀ function.

108

6 Numerical Computation of Model Prices 0

0

−5 −5

ln(RMSE)

ln(RMSE)

−10

−15

−20

−10

−15

−25 −20 −30

−35

0

100

200

300

400

500

600

700

800

−25

900

0

200

400

600

number of iterations

800

1000

1200

1400

1600

number of iterations

0

0

−5 −5

ln(RMSE)

ln(RMSE)

−10

−15

−20

−10

−15

−25 −20 −30

−35

0

2

4

6

8

10

12

14

16

18

−25

20

0

2

4

6

8

number of iterations

10

12

14

16

18

20

number of iterations

30

20

7

4

−11

−20

−8

14

−1

−1

9

−2

−23

−11

−5 20

16

1 −2 −5

−17

18 25

−17

i

z

10

−14

−20

−1

7

15

1

zi

12

−11

−11

−5

1

200

−17

ω

4

−5 1

1

400

600

−11 −14

−11

−23 5

6

−8

−5

−17

−29

1 −2 −5

8 10

800

2 200

−5 −2 1 400 600

−17 −14 −11 −8

−8

−5 −2 1

800

ω

1000 1200 1400 1600

Fig. 6.4. Search for the optimal parameter couple (ω ∗ , zi∗ ). In the ﬁrst (second) column particular graphs are shown for the Vasicek (CIR) model with the data used in Figure 6.3. In the ﬁrst row, the particular ln(RMSE) is depicted for the search algorithm with increments (∆ω, ∆zi ) = (1, 1). In the second row the same search is made with increments (100, 5). In the third row the dashed (dash-dotted) line denotes the particular search path for small (high) increment values, where the optimal choice is marked by a circle and cross, respectively.

6.3 Contracts with Conditional Exercise Rights

109

(8.27740×10−10), whereas the exact RMSE is 1.11766×10−13 (1.30453×10−9). Thus, we have in both models a diﬀerence which is of smaller order than the eﬀective error according to the RMSE. Consequently, the RMSEa gives a good prediction for the corresponding exact value, which justiﬁes the application of the approximate RMSE. In the ﬁrst row of Figure 6.4, we used very small increments for the search of the optimal parameter couple to give a detailed impression of the search path and the particular logarithmic RMSE. According to the graphs in the second row of Figure 6.4, a comparable result is achieved by running the algorithm with higher increments150 . However, due to the reduced number of iterations, the search algorithm with high increments is in case of the jump-enhanced Vasicek (CIR) model up to 71 (86) times faster. Dealing with a characteristic function known in closed form together with a FRFT-based pricing algorithm, the search takes only a second at all even for small increments. Thus, if the general characteristic function is known in closed form, the step-size does not matter. However, if values of the general characteristic need to be determined numerically via a Runge-Kutta algorithm, we usually set the increments high enough to keep the overall number of iterations small. Finally, we use the RMSEa to derive an upper error bound for the numerical solutions contained in VN um . The ﬁrst step in deriving this particular error bound is to consider a hypothetical solution vector VN um , where all elements equal their true solutions except the result given in the ﬁrst position of the solution vector, namely V1N um . Without loss of generality, we assume the numerical error of this particular option price to be of magnitude |a|. Therefore, solving equation (6.11) in this special case gives √ a = RMSE U . Additionally, we are also able to state the inequality * (VN um − VT rue ) (VN um − VT rue ) ≥ |VvN um − VvT rue |,

(6.13)

(6.14)

to hold for every element of the numerical solution vector VN um . According √ to equation (6.13), the RMSE scaled by some constant U states the value 150

The second run of the algorithm, with higher increments, gives an absolute error for the optimal parameter couple (ω ∗ , zi∗ ) for zero-bond calls under the jumpenhanced Vasicek (CIR) model of 1.13911 × 10−13 (1.46601 × 10−9 ).

110

6 Numerical Computation of Model Prices

of the maximum attainable error. Furthermore, this result together with the inequality in (6.14) generally implies that the absolute error of one particular numerical solution VvN um cannot exceed the absolute value |a|. Therefore, the RMSE can be used in formulating a boundary for the highest possible error. √ Consequently, we use the quantity RMSEa , scaled by some constant U , as a conservative upper error bound for the results generated by the pricing algorithm.

7 Jump Speciﬁcations for Aﬃne Term-Structure Models

7.1 Overview In this thesis, we discuss jump-diﬀusion interest-rate models. Thus, both diffusion and jump components are included in order to model more realistic term-structure models. The jump sizes considered are governed either by exponential, normal or gamma distributions. The exponential jump distribution is a very popular approach in modeling term structure and equity models151 , since it yields closed-form formulae for most derivatives contracts. Das and Foresi (1996) and Chacko and Das (2002) have conducted recent studies with a double-sided version of this jump type using a Vasicek model for the diffusion part152 . Our second candidate, the normal jump-size distribution is used in Baz and Das (1996) and Das (2002)153 . The last jump-size distribution candidate for the interest-rate process is a gamma distribution, which is used in Kispert (2005) to support the stochastic dynamics of the volatility in electricity derivative contracts. This jump type is used for the ﬁrst time in a jump-diﬀusion interest-rate model. As a special case, the gamma distribution covers the exponential distribution. Hence, we can build a more ﬂexible 151

152

Das and Foresi (1996) used this jump speciﬁcation in modeling short rates whereas Kou (2002) uses this type of jump-size distribution modeling equities. Jumps in the instantaneous interest rate are governed by an exponential jumpsize distribution. The direction of the jump itself is modeled either by a Bernoulli

153

distribution or via two diﬀerent Poisson processes. See Section 8.2. The articles consider a discrete version of the Vasicek interest-rate model with normally distributed jump shocks.

112

7 Jump Speciﬁcations for Aﬃne Term-Structure Models

jump shock component in contrast to the exponential case, by extending the repertoire of jump-size distributions to the gamma distribution case. In the following, we do not restrict ourselves solely to one jump component for each factor. Due to the independence of the jump distributions from the state of xt 154 , we are able to add an unlimited amount of diﬀerent jump components. However, we need to consider possible nonnegativity constraints of the particular diﬀusion process. Thus, we do not combine normally distributed jump parts, or negatively directed exponentially and gamma distributed jump parts, with a Square-Root diﬀusion model. This is in fact no real drawback, that is to say we can think of a bad news eﬀect rather as a discontinuous increase in interest rates than the opposite eﬀect155 . All possible combinations for diﬀusion and jump components are illustrated in Figure 7.1. According to equation (2.40), the coeﬃcient function a(z, τ ) can be split into a part containing the characteristics of the diﬀusion process and a part containing the additional jump characteristics156. This results in a modular representation of the ODE for the coeﬃcient function a(z, τ ), which is a(z, τ )τ = a0 (z, τ )τ + a1 (z, τ )τ , with a0 (z, τ )τ = −w0 + µQ 0 b(z, τ ) +

(7.1)

1 b(z, τ ) Σ0 b(z, τ ), 2

and a1 (z, τ )τ = EJ [ψ ∗ (z, w0 , w, g0 , g, J, τ ) − 1] λQ . Unless otherwise stated, the coeﬃcient function a0 (z, τ ) denotes the diﬀusion part, whereas a1 (z, τ ) represents the solution for the jump part, which is frequently called the jump transform157 . As mentioned above, each diﬀusion process can be augmented with an inﬁnite number of jump processes. Thus, taking the expectation in (2.39) 154

This statement also holds for diﬀerent jumps triggered by the same poisson pro-

155

cess. See equation (7.2). See, for example, Sch¨ obel and Zhu (2000), p. 5. Note that the jump part aﬀects the coeﬃcient a(z, τ ), whereas the coeﬃcient

156

157

vector b(z, τ ) is independent from the jump amplitude and intensity. See Duﬃe, Pan and Singleton (2000).

7.1 Overview

113

Fig. 7.1. Possible combinations of the Ornstein-Uhlenbeck (OU) process and the Square-Root (SR) process with the exponential (Ex), normal (No), and gamma (Ga) jump distributions.

with respect to the jump amplitudes J, we obviously are able to check that every element of the resulting vector is expressible as the product of diﬀerent expectations. Formally, we have Ej1 eb(z,τ ) j1 eb(z,τ ) j1 b(z,τ ) j2 E eb(z,τ ) j2 e j2 = . EJ .. .. . . b(z,τ ) jN b(z,τ ) j N e EjN e

Selecting one element of this vector as an example, say Ejn eb(z,τ ) jn , and manipulating the expectation operator, we get

114

7 Jump Speciﬁcations for Aﬃne Term-Structure Models

M + b(z,τ ) jn b(z,τ ) jn = Ejn e e ν(Jmn ) djn m=1

RM

=

M +

(m)

eb

(z,τ )Jmn

ν(Jmn ) dJmn

(7.2)

m=1 R

=

(m) EJmn eb (z,τ )Jmn .

M + m=1

The function ν(Jmn ) represents the probability density of the particular jump amplitude Jmn . As demonstrated in equation (7.2) the joint density function can be expressed as the product of diﬀerent density ν(Jmn ) on account of the independence of the jump amplitudes. Consequently, issues are simpliﬁed in equation (7.2) by successively evaluating all integrals one by one, which yields the cumulative product of diﬀerent expectations. Thus, we express the solution of the jump part as a1 (z, τ ) = −τ ιN λQ +

N n=1

(n)

λQ

τ + M

(m) EJmn eb (z,l)Jmn dl ,

(7.3)

0 m=1

with n = 1, . . . , N and each element of the vector ιN ∈ RN equals one. Since we need to calculate the integral over the time variable there is the possibility of ending up without any closed-form solution158 . In contrast, the Bates (1996) model, which is a jump-diﬀusion model in an equity context, in which a normally distributed jump component is used for the log-asset price process, the coeﬃcient for the jump size yields a nice closed-form expression in Fourier space159 . In term-structure models, normal and gamma size distributions allow only the formulation of the coeﬃcient a1 (z, τ ) in terms of its underlying diﬀerential equation. Thus, a possible reason why normal and gamma jump distributions are not as popular in interest-rate option pricing might be tracked back to the unavailability of appropriate valuation formulae for interest-rate contracts. Nevertheless, these jump size candidates provide a 158

Both gamma and normally distributed jump amplitudes have no closed-form jump

159

transform for all models discussed in this thesis.

In one-factor equity models the computation can be simpliﬁed to EJn eızJn , which is obviously easier to handle, since the exponential function inside the expectation operator is independent of the time to maturity variable τ . See Cont and Tankov (2004), p. 477.

7.2 Exponentially Distributed Jumps

115

valuable contribution in generating a realistic overall probability distribution of short rates. However, the algorithm presented in Chapter 6 can compute derivative prices under these interest-rate dynamics. The only condition that needs to be met is the availability of separable ODEs of the coeﬃcient functions a(z, τ ) and b(z, τ ), which lets us apply a Runge-Kutta algorithm160 .

7.2 Exponentially Distributed Jumps The exponential distribution is a widely used shock speciﬁcation in jumpdiﬀusion models. Thus, it can be found in both equity and interest-rate models161 . The probability density function pEx (J, η) of an exponentially distributed variable J ∼ Ex(η) is deﬁned as if 0 pEx (J, η) = 1 − J if e η η

J 0.

Hence, the expected value for J and its variance is EJ [J] = η, and VARJ [J] = η 2 . The shape of the density function pEx (J, η) for diﬀerent values of η is shown in Figure 7.2. For a positively directed jump, with distribution parameter η+ , we get 160

An interest-rate model which clearly opposes this separability ability of the coeﬃcient functions of the general characteristic function is given in Ahn and Gao (1999) and A¨ıt-Sahalia (1999), Example 3. However, closed-form solutions of zerobond prices under these short-rate dynamics can be derived. See Ahn and Gao

161

(1999), Proposition 1. Kou (2002), Kou and Wang (2004) implemented this jump speciﬁcation for equity models, whereas Das and Foresi (1996) integrated this jump type in an OrnsteinUhlenbeck short-rate model.

116

7 Jump Speciﬁcations for Aﬃne Term-Structure Models

200 0.005 0.01 0.015 0.02

180 160

probability density

140 120 100 80 60 40 20 0

0

0.01

0.02

0.03

0.04 J

0.05

0.06

0.07

0.08

Fig. 7.2. The density function pEx (J, η) for varying η of an exponentially distributed random variable. ∞ (m) b (z,τ )− η1 J (m) + e 1 −1 EJ eb (z,τ )J − 1 = η+ b(m) (z, τ ) − η1+ 0

1 = −1 1 − b(m) (z, τ )η+ =

b(m) (z, τ )η+ . 1 − b(m) (z, τ )η+

Accordingly, a negatively directed jump with parameter η− , has an expected value of ∞ (m) − b (z,τ )+ η1 J (m) − e 1 −1 EJ e−b (z,τ )J − 1 = − η− b(m) (z, τ ) + η1− 0

1 −1 = 1 + b(m) (z, τ )η− =−

b(m) (z, τ )η− . 1 + b(m) (z, τ )η−

7.3 Normally Distributed Jumps

117

In order to guarantee the existence of the jump transform, we need the real part Re b(m) (z, τ ) ≤ η1+ for the positively sized jump and Re b(m) (z, τ ) ≥ − η1− for the negatively directed jump162 , respectively. Thus, multiplying the recently derived expectations with the jump intensity of the particular Poisson jump, the jump part of the coeﬃcient function a(z, τ ) can be generally represented as, τ a1Ex± (z, τ )

=± 0

(n)

λQ b(m) (z, l)η± dl. 1 ∓ b(m) (z, l)η±

(7.4)

The transform for this jump candidate is the only one that can be expressed in closed form for the interest-rate models discussed in the next chapter.

7.3 Normally Distributed Jumps The second candidate we consider for the jump-size distribution is the normal distribution. As mentioned before, this speciﬁcation is not as popular in interest-rate pricing frameworks compared to the exponentially distributed case. One reason might be that the jump transform in an interest-rate jumpdiﬀusion framework cannot be expressed in closed form. In this setup, the jump amplitude J ∼ N µJ , σJ2 is distributed according to a probability density function: −

(J−µJ )2 2σ2 J

e pN o (J, µJ , σJ ) = √ 2πσJ

∀

J ∈ R,

with mean EJ [J] = µJ , and variance VARJ [J] = σJ2 . The shape of the density function pN o (J, µJ , σJ ) for diﬀerent values of σJ is shown in Figure 7.3. The few articles which mention this particular jump type can be quickly summarized. Baz and Das (1996), Durham (2005) and Durham (2006) implemented the Gaussian jump within a Vasicek base model. Since this type 162

Since b(m) (z, τ ) < 0 and be fulﬁlled.

1 η−

is usually very large, we assume both conditions to

118

7 Jump Speciﬁcations for Aﬃne Term-Structure Models

of jump might violate a non-negativity constraint of the underlying diﬀusion process, it is only meaningful in a context of a real-valued process. Therefore, we do not consider the normally distributed jump candidate in case of a Square-Root diﬀusion process.

80 0.005 0.01 0.015 0.02

70

probability density

60 50 40 30 20 10 0 −0.08

−0.06

−0.04

−0.02

0 J

0.02

0.04

0.06

0.08

Fig. 7.3. The density function pNo (J, µJ , σJ ) for ﬁxed µJ = 0 and varying σJ of a normally distributed random variable.

Baz and Das (1996) approximate the expectation in equation (7.3) via a Taylor series approximation163 . The series-approximation approach mentioned there considers two terms. Consequently, they ﬁrst approximate the expression inside the expectation operator, and then take the expectation of the resulting terms. Hence, the approximation is given as 163

The Taylor series approximation of the exponential function f (x) = ex is given by

∞

xi i=0 i! .

7.3 Normally Distributed Jumps

(m) b(m) (z, τ )2 2 J EJ eb (z,τ )J − 1 ≈ EJ b(m) (z, τ )J + 2 = b(m) (z, τ )µJ +

119

b(m) (z, τ )2 2 (µJ + σJ2 ). 2

In the last equation, the particular parameters of the normal distribution µJ and σJ2 are used. Of course, it is possible to use a Taylor series considering more terms in order to enhance the accuracy of the calculations. Another slightly diﬀerent approximation technique is presented in Durham (2005) and Durham (2006), respectively. Here, the author applies a Taylor series approximation after taking the expectation in equation (7.3). This incorporates the distributional parameters in a more explicit fashion. Applying a two-term Taylor expansion gives164 (m) b(m) (z, τ )2 2 µJ + σJ2 EJ eb (z,τ )J − 1 ≈ b(m) (z, τ )µJ + 2 b(m) (z, τ )3 b(m) (z, τ )4 4 2 + µJ σJ + σJ . 2 8

Obviously, there is an advantage in applying either one of these analytic approximations for the jump transform. Using these simpliﬁcations, we are able to solve the ODE (7.1) in a consistent manner, meaning that no numerical integration of the jump transform is needed anymore, since only terms of b(m) (z, τ ) are left, yielding an approximate closed-form solution for the characteristic function165 . As an additional beneﬁt of the analytical approximations, Baz and Das (1996) mention the computational speed enhancement, facilitating the calibration to empirical data. However, the major drawback of both approximation techniques results from the application of the Taylor series. In order to produce accurate results, the term b(m) (z, τ ) and the diﬀerence inside the expectation operator, respectively, must be very small. Hence, with an increasing mean of the jump component and increasing variance, the results 164

165

In Durham (2006) the negative sign of the coeﬃcient b(m) (z, τ ) is extracted which explains the slightly diﬀerent representation. For example, taking the Vasicek one-factor base model of equation (8.5) and the linear approximation due to Baz and Das (1996), we encounter the simple Q Q problem of solving equation (7.1) with extended parameters µ ˆQ 0 = µ0 + λ µJ and

σ ˆ0 =

σ02 + σJ2 + µ2J . Subsequently, the coeﬃcient function aτ (z, τ ) with these

modiﬁed parameters has to be solved.

120

7 Jump Speciﬁcations for Aﬃne Term-Structure Models

get more and more inaccurate. Consequently, the analytical approximation procedures should be applied only to scenarios where the jump component exhibits a small mean and variance. Since our numerical procedure is designed to handle implicitly the ODE (m) part of the jump transform, we need only the expected value of eb (z,τ )J to be explicit. Under a normally distributed jump size regime, this is (m) (b(m) (z,τ )σJ )2 b (z,τ )J b(m) (z,τ )µJ + 2 EJ e −1 =e − 1, which leads to the particular coeﬃcient function 2 τ b(m) (z,τ )σJ ) ( (n) (m) 2 dl . a1N o (z, τ ) = λQ −τ + eb (z,l)µJ +

(7.5)

(7.6)

0

The value of the integral can then be numerically approximated via a Runge-Kutta algorithm. Despite the numerical integration, the computational eﬀort is very small due to our implemented FRFT procedure. But in contrast to the Taylor-series approach mentioned above, our results do not suﬀer from inaccuracies due to high mean and volatility parameters of the jump component. Hence, with our valuation procedure we gain superior accuracy. Furthermore, we are able to compute model prices for this jump speciﬁcation for the ﬁrst time, not only for ordinary zero bonds, but for all derivatives contracts, presented in Sections 3.2 and 3.3.

7.4 Gamma Distributed Jumps The last jump-size distribution we want to implement in an interest-rate model is the gamma distribution. The probability density function pGa (J, η, p) of the random variable J ∼ Ga(η, p) is given as 0 if J 0. p η Γ (p) Thus, the expected value and variance for J is

7.4 Gamma Distributed Jumps

121

EJ [J] = ηp, and VARJ [J] = η 2 p. The function Γ (p) denotes the gamma function. Setting the parameter p = 1, the gamma distribution replicates the exponential distribution, since we have then the relation pGa (J, η, 1) = pEx (J, η). Additionally, we can use the gamma distribution to generate a chi-squared distribution. In this case we set η = 2 and p = 2q , where q is a positively valued integer. The resulting chi-squared distribution has then 2p and q degrees of freedom166 . Another special case of the gamma distribution is the Erlang distribution. Here, we only need p to be a positive integer value167 . Thus, the Erlang distribution can be interpreted as the sum of p independent exponentially distributed random variables with equal parameter η. The graph in Figure 7.4 shows the probability density function for diﬀerent values of p. Comparing the diﬀerent curves in Figure 7.4, it is obvious that this jump-size distribution is able to substantially enhance the short-rate model. In Heston (1995) a pure jump interest-rate model is proposed. Accordingly, instead of a diﬀusion component, the innovations of the process in this model are governed solely by gamma distributed jumps. To our knowledge, Kispert (2005) was the ﬁrst to use gamma distributed jump sizes within a jump-diﬀusion model. However, in pricing European options he needs inefﬁcient Monte-Carlo routines, using the gamma and normal jump amplitude speciﬁcation. These numerical problems can be circumvented by applying the FRFT-based algorithm together with a Runge-Kutta algorithm for the ODEs. Next, we want to derive the particular jump transform. The expectation for a positively directed jump can be computed as (m) EJ eb (z,τ )J − 1 = 166

167

1 p η+ Γ (p)

∞ e

−J

1 η+

−b(m) (z,τ )

J p−1 dJ − 1.

0

This is easily checked by comparing the particular moment-generating functions. See Stuart and Ord (1994), p. 541. See, for example, Balakrishnan, Johnson and Kotz (1994), p. 337.

122

7 Jump Speciﬁcations for Aﬃne Term-Structure Models

200 1 2 3 4

180 160

probability density

140 120 100 80 60 40 20 0

0

0.01

0.02

0.03 J

0.04

0.05

0.06

Fig. 7.4. The density function pGa (J, η, p) for ﬁxed η = 0.005 and varying p of a gamma distributed random variable.

Introducing the substitution m = J pler representation168 (m) EJ eb (z,τ )J − 1 =

Γ (p)

1 η+

− b(m) (z, τ ) , we arrive at the sim-

1

1 η+

p p − b(m) (z, τ ) η+

∞ e−m mp−1 dm − 1 0

"

#$

%

Γ (p) 1 p − 1. = 1 − b(m) (z, τ )η+ Hence, the corresponding expression for a negatively sized jump is (m) 1 p − 1. EJ e−b (z,τ )J − 1 = (m) 1 + b (z, τ )η− 168

To ensure theexistence of the jump transform, we have the same inequalities for Re b(m) (z, τ ) to be satisﬁed as in the case of exponentially distributed jumps.

7.4 Gamma Distributed Jumps

123

Having derived the relevant expectations, we immediately are able to formulate the respective jump transforms a1Ga± (z, τ ). Thus, integrating over the (n) time axis and multiplying the result by the relevant jump intensity λQ we obtain,

(n)

a1Ga± (z, τ ) = λQ

−τ +

τ 0

1 dl . (1 ∓ b(m) (z, l)η± )p

(7.7)

Since it is not possible to derive closed-form solutions for general values of p, we apply a Runge-Kutta algorithm implemented in our FRFT procedure in order to eﬃciently calculate option prices. Again, for a strictly positive interest-rate model, we use only a positively directed version of the jump candidate.

8 Jump-Enhanced One-Factor Interest-Rate Models

8.1 Overview In order to implement the previously proposed pricing procedure, we need in addition to the payoﬀ transformations derived in Chapter 5, the particular characteristic functions. The goal of this chapter is to provide these necessary functions for the case of an underlying one-factor interest-rate model and to examine the behavior of the particular density functions and prices of selected contingent claims, according to Table 4.1, inﬂuenced by jump components. Thus, we focus our eﬀorts exclusively on the exponential-aﬃne termstructure models generated by the one-factor version of equation (2.23). Since a one-factor model implicates the incorporation of one Brownian motion, this statement does not entail the restriction of including one sole jump component. Therefore, we apply diﬀerent jump components in our examples. The general version of the one-factor instantaneous interest rate is then given by rt = w0 + w1 xt , and the factor xt is deﬁned by the one-dimensional stochastic diﬀerential equation

dxt = µQ (xt ) dt + σ(xt ) dWtQ + j dN λQ t ,

(8.1)

where j ∈ RN , µQ (xt ) and σ(xt ) are the one-factor counterparts of the original parameters J, µQ (xt ) and Σ(xt ) used in equation (2.23). All parameters are postulated under the risk-neutral probability measure Q. Therefore, the solution of the general characteristic function ψ(xt , z, w0 , w1 , g0 , g1 , τ ) for these models is given by the simpliﬁed versions of the ODEs (2.40) and (2.41), which are

126

8 Jump-Enhanced One-Factor Interest-Rate Models

σ02 b(z, τ )2 − w0 , a0 (z, τ )τ = µQ b(z, τ ) + 0 2 a1 (z, τ )τ = Ej eb(z,τ )J1 , eb(z,τ )J2 , . . . , eb(z,τ )JN − 1 λQ ,

(8.2) (8.3)

and

σ12 b(z, τ )2 − w1 , 2 with terminal conditions a(z, 0) = 0 and b(z, 0) = ızg1. b(z, τ )τ = µQ 1 b(z, τ ) +

(8.4)

In the upcoming sections, we discuss jump-enhanced short-rate models where the diﬀusion part is either modeled as a Ornstein-Uhlenbeck or a Square-Root process and the jump components are governed by the distributions presented in the previous chapter.

8.2 The Ornstein-Uhlenbeck Model 8.2.1 Derivation of the Characteristic Function Modeling the factor xt as a stochastic process according to Ornstein and Uhlenbeck (1930) exhibits a strong resemblance to the well-known model given in Vasicek (1977)169 . The so-called Vasicek model has become very popular in interest-rate modeling. The instantaneous interest rate is modeled as an Ornstein-Uhlenbeck process, with a mean-reverting component and a Brownian motion, which yields a time-homogenous Markov process. The approach used in Vasicek (1977) to derive prices for contingent claims under the riskneutral probability measure, is similar to the methodology used in the article of Black and Scholes (1973), based on a hedging argument170 . Due to its popularity, many authors have made attempts to extend this diﬀusion model with jumps. Das and Foresi (1996) introduced a jump-enhanced Vasicek model, where the jump size is governed by an exponential distribution and the jump direction is modeled as a Bernoulli random variable which results in a double-sided jump component. However, given a jump intensity λ for 169

The process used in Vasicek (1977) and the process discussed in this section coincide for the case of rt = xt , thus setting the discount parameters to w0 = 0

170

and w1 = 1. In contrast to the Black-Scholes model, an appropriate market price of risk has to be additionally considered, since the short rate rt is no traded quantity. See Section 2.3.

8.2 The Ornstein-Uhlenbeck Model

127

the Poisson process, this model can be easily subsumed by applying a single exponentially distributed jump size and setting the modiﬁed intensity for the upward jump trigger to ψλ and the downward jump intensity to (1 − ψ)λ171 , respectively. Another model, where the Vasicek model is extended with a normally distributed jump size, is given in Baz and Das (1996), Das (2002), Durham (2005). In Baz and Das (1996) and Durham (2005), approximation techniques are presented for pricing option contracts under these interest-rate dynamics, as explained in Section 7.3. In Das (2002), the author utilizes this jump-diﬀusion model for the estimation of the term structure and subsequent calibration of the particular parameters according to Fed Funds data. The the risk-neutral coeﬃcients are µQ 0 = κθ,

µQ 1 = −κ,

σ0 = σ,

σ1 = 0,

where the mean-reverting feature of the instantaneous interest rate rt is guaranteed for κ > 0. Thus, the diﬀusion part of the SDE (8.1) is dxt = κ(θ − xt ) dt + σ dWtQ .

(8.5)

Under these dynamics the stochastic process starting with xt the OrnsteinUhlenbeck process reﬂects a normal distribution with expectation EQ [xT ] = xt e−κτ + θ(1 − e−κτ ), and variance VARQ [xT ] =

σ2 1 − e−2κτ . 2κ

Modeling the term structure with this Ornstein-Uhlenbeck type process, has the attractive feature that solutions for many important contingent claims can be derived within closed-form formulae. Moreover, the model is likely to be used for its high tractability. Finally, one major drawback of the model is the ability to produce negative short rates with a positive probability. According to equations (8.2) and (8.4), straightforward calculations show that the diﬀusion-related coeﬃcients of the general characteristic function can 171

In Das and Foresi (1996), the parameter ψ denotes the probability that the sign of the jump is positive.

128

8 Jump-Enhanced One-Factor Interest-Rate Models

be derived as172

˜b(z, τ ) = w1 + ızg1 e−κτ − 1 , κ

(8.6)

ızg1 σ 2 ˜ b(z, τ ) a0 (z, τ ) = − w0 τ − 2κ σ2 w1 σ 2 ˜ ˜b(z, τ )2 . − θ− b(z, τ ) + w1 τ − 2 2κ 4κ

(8.7)

and

Equipped with these time-dependent coeﬃcient functions corresponding to the diﬀusion parts of the short-rate model, we must determine in the next step the particular jump part a1 (z, τ ). Since this function is independent of a0 (z, τ ), we are able to derive it separately. Unfortunately, a closed-form solution for the coeﬃcient a1 (z, τ ) exists only in case of an exponentially distributed jump size. Thus, according to equation (7.4) we obtain for the Ornstein-Uhlenbeck model, where the nth jump in xt is governed by an exponentially distributed jump size, the relevant coeﬃcient function as (n)

a1Ex± = −λQ

τ+

(n) 1 ∓ b(z, τ )η± λQ ln . κ ± w1 η± (1 ∓ ızg1η± ) e−κτ

(8.8)

In equation (8.8), the signs in the index of a1Ex± denotes an upward and a downward jump, respectively. Considering normally and/or gamma distributed jumps, we have to apply a Runge-Kutta algorithm to solve equations (7.6) and (7.7). 8.2.2 Numerical Results Next, we want to examine and demonstrate the impact of the particular jump speciﬁcations for the case of a jump-enhanced Ornstein-Uhlenbeck process. Thus, we ﬁrst compare the probability density for diﬀerent jump amplitude speciﬁcations, and afterwards look brieﬂy at values of option prices for interest-rate derivatives corresponding to the payoﬀ structures given in Table 4.1. 172

˜ τ ) and therefore Here, the coeﬃcient ˜b(z, τ ) denotes the scalar version of b(z, complies with the relation b(z, τ ) = ˜b(z, τ ) + ızg1 .

8.2 The Ornstein-Uhlenbeck Model

129

Figures 8.1 - 8.3 depict probability density functions of short rates under diﬀerent jump regimes with diﬀusion parameters rt = 0.05, κ = 0.4, θ = 0.05, σ = 0.01 and T = 1. In each ﬁgure, we focus exclusively on one particular jump candidate, while ignoring other jump speciﬁcations. The probability density functions are then examined for diﬀerent arrival rates and jump amplitudes173 , respectively. Additionally, in case of a normally distributed jump component, we also examine the inﬂuence of the jump amplitude volatility, whereas in case of a gamma distributed jump distribution, the impact for diﬀerent values of p is displayed.

50

50 0 2 4 6

45 40

40 35

probability density

probability density

35 30 25 20

30 25 20

15

15

10

10

5 0

0.005 0.01 0.015 0.02

45

5 0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0

0

0.02

r

T

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

T

Fig. 8.1. Probability densities for a short rate governed by a Vasicek diﬀusion model enhanced with an exponentially distributed jump component. In the left (right) graph the density function for varying jump intensities (means) are depicted. The base parameters are: rt = 0.05, κ = 0.4, θ = 0.05, σ = 0.01, λ = 2, η = 0.005, T = 1.

The ﬁrst impression from Figures 8.1 - 8.3, is that increased jump intensity results in all three cases in a positively skewed density function with a slightly right-shifted mode174 . The asymmetric shape is in line with empirical ﬁndings175 . Increasing the mean of the jump amplitude, the density functions 173

In case of exponentially and gamma distributed jumps, the arrival rates belong only to positively directed jumps, thus leaving downward jumps with zero jump

174

175

intensities. This eﬀect becomes more apparent for higher values of jump amplitudes η and µJ , respectively. See Arapis and Gao (2006), Figure 3. The authors apply alternatively a nonparametric estimator for the short-rate probability density of three-month Treasury bill rates and seven-day Eurodollar deposit rates.

130

8 Jump-Enhanced One-Factor Interest-Rate Models 50

50 0 2 4 6

45 40

40 35

probability density

probability density

35 30 25 20

30 25 20

15

15

10

10

5 0

0.005 0.01 0.015 0.02

45

5 0

0.05

0.1

0.15

0.2

0.25

0.3

r

0

0

0.05

0.1

0.15

0.2

0.25

0.3

r

T

T

50 1 2 3 4

45 40

probability density

35 30 25 20 15 10 5 0

0

0.05

0.1

0.15

0.2

0.25

0.3

r

T

Fig. 8.2. Probability densities for a short rate governed by a Vasicek diﬀusion model enhanced with a gamma distributed jump component. In the upper left (right) graph the density functions for varying jump intensities (means) are depicted. The lower graph shows the density behavior for alternating values of p. The base parameters are: rt = 0.05, κ = 0.4, θ = 0.05, σ = 0.01, λ = 2, η = 0.005, p = 2, T = 1.

of all jump candidates show positive skewness while concurrently maintaining the mode of the density function. Comparing the particular density functions of an exponentially and gamma distributed jump-enhanced short-rate model, we encounter, in case of a gamma and normal distribution, a bi-modal density function. The impact of the volatility parameter σJ for a normally distributed jump is more complex. For high values of the jump volatility, the density function displays a leptokurtic behavior compared. In addition, we observe, due to the possibility of negative jump sizes, raised tails on both sides of the particular density function as well. This eﬀect is rather visible to the right tail, since we have a positive mean of the jump-size distribution. Due to the possibility to produce negative short rates in the Ornstein-Uhlenbeck case, we have the undesirable ability to obtain a density function with non-negligible

8.2 The Ornstein-Uhlenbeck Model 40

40 1 2 3 4

35

30

probability density

probability density

0 0.01 0.02 0.03

35

30 25 20 15

25 20 15

10

10

5

5

0

131

0

0.05

0.1

0.15

0.2

0.25

0.3

r

0

0

0.05

0.1

0.15

0.2

0.25

0.3

r

T

T

40 0.005 0.01 0.015 0.02

35

probability density

30 25 20 15 10 5 0

0

0.05

0.1

0.15

0.2

0.25

0.3

r

T

Fig. 8.3. Probability densities for a short rate governed by a Vasicek diﬀusion model enhanced with a normally distributed jump component. In the upper left (right) graph the density functions for varying jump intensities (means) are depicted. The lower graph shows the density behavior for alternating values of σJ . The base parameters are: rt = 0.05, κ = 0.4, θ = 0.05, σ = 0.01, λ = 1, µJ = 0.02, σJ = 0.01, T = 1.

probabilities for negative rates, which becomes more severe depending on the absolute height of the volatility. Obviously, besides the asymmetric shape, all density functions are skewed, in contrast to the plain Ornstein-Uhlenbeck model, which is another advantage in including jump components. Theoretical prices of interest-rate derivatives are computed with the following base parameters: rt = 0.05, κ = 0.4, θ = 0.05, σ = 0.01, λΓ = 2, η = 0.005, p = 2, λN = 2, µJ = 0.015, σJ = 0.01 and τ = 0.5. Table 8.1 reports values of zero-bond calls, according to equation (3.12), for a strike range from 60 to 90 units computed with the FRFT pricing algorithm. We choose this particular strike range to cover either ITM, ATM and OTM option

132

8 Jump-Enhanced One-Factor Interest-Rate Models

prices176 . Here, we only considered normally and gamma distributed jumps, thus excluding exponentially distributed jumps, because of the similarity to the gamma jump speciﬁcation. Solutions are given for diﬀerent jump intensities, amplitudes, and volatilities in case of normally distributed jumps and for diﬀerent values of p in case of the gamma jump size speciﬁcation, respectively. Examining values of zero-bond calls for diﬀerent values of the parameters η, λΓ and p one-by-one, we observe that the jump intensity has the greatest inﬂuence upon call values followed by the parameter p. The jump mean has the smallest eﬀect upon option prices although it signiﬁcantly alters the particular density function of the short rate. However, price diﬀerences for ITM options are relatively diminutive, whereas for OTM options the above mentioned impact is quite considerable. Applying a normally distributed jump component in the short-rate model, we observe for increased jump volatilities higher option prices, which can be explained based on the above mentioned two-sided enlargement of the probability density function compared to the cases where the jump mean or jump intensity is increased. Theoretical prices for cap contracts and average-rate caps on the short rate are presented in Tables 8.2 and 8.3, respectively, for a strike range from 2 to 8 units. The cap contracts have both only one payment date, which is paid at the maturity of the contract. Here, we observe the opposite eﬀect due to the direct inﬂuence of the short rate on the payoﬀ function. Since the contract is based on rT , positively directed jumps increase the contract value. Remarkably, the eﬀect of the jump volatility σJ is twofold. Firstly, it lowers the value of the option contract for ITM options. On the other hand cap values are being raised, if the option contract is OTM. Obviously, the geometric average is less sensitive to discontinuous jumps. Since the interest rate is inﬂuenced by positively sized jumps, we compute higher values for the ordinary cap contract compared to the corresponding average-rate contract due to the averaging process itself. Thus, we are able to validate the statement that average-rate options are more robust to price manipulations, thus reducing risk exposures177 .

176

177

The value of a zero bond with remaining time to maturity of two years priced with the base parameters is 83.768 units. Compare with the comments made on p. 41.

8.2 The Ornstein-Uhlenbeck Model

133

Table 8.1. Values of zero-bond call options for the jump-enhanced OU model, where the underlying zero-bond contract has a nominal value of 100 units. K

60

65

70

75

80

85

90

20.595

15.747

10.899

6.067

1.666

0.004

0

0.01

17.277

12.442

7.631

3.117

0.273

0

0

0.015

14.168

9.383

4.827

1.246

0.003

0

0

0.02

11.296

6.711

2.745

0.314

0

0

0

0

24.134

19.274

14.415

9.556

4.727

0.686

0

2

20.595

15.747

10.899

6.067

1.666

0.004

0

4

17.220

12.383

7.552

2.928

0.170

0

0

6

14.001

9.177

4.436

0.774

0

0

0

1

22.338

17.485

12.631

7.779

3.054

0.094

0

2

20.595

15.747

10.899

6.067

1.666

0.004

0

3

18.902

14.060

9.219

4.459

0.738

0

0

4

17.258

12.422

7.601

3.051

0.253

0

0

0.005

24.123

19.264

14.404

9.545

4.703

0.627

0

0.01

22.333

17.480

12.626

7.773

3.032

0.096

0

0.015

20.595

15.747

10.899

6.067

1.666

0.004

0

0.02

18.907

14.065

9.225

4.472

0.757

0

0

0

25.945

21.080

16.215

11.350

6.486

1.772

0

2

20.595

15.747

10.899

6.067

1.666

0.004

0

4

15.610

10.780

5.982

1.747

0.027

0

0

6

10.967

6.194

1.990

0.093

0

0

0

0.005

20.580

15.732

10.884

6.044

1.593

0.002

0

0.01

20.595

15.747

10.899

6.067

1.666

0.004

0

0.015

20.620

15.772

10.925

6.107

1.774

0.014

0

0.02

20.656

15.807

10.962

6.167

1.907

0.043

0

η 0.005

λΓ

p

µJ

λN

σJ

134

8 Jump-Enhanced One-Factor Interest-Rate Models

Table 8.2. Values of short-rate caps for the jump-enhanced OU model, with a nominal value of 100 units. K

2

3

4

5

6

7

8

0.005

5.096

4.126

3.160

2.241

1.476

0.909

0.525

0.01

5.952

4.984

4.020

3.097

2.296

1.642

1.136

0.015

6.798

5.833

4.871

3.948

3.135

2.446

1.878

0.02

7.635

6.672

5.712

4.790

3.974

3.267

2.665

0

4.230

3.258

2.295

1.430

0.826

0.445

0.223

2

5.096

4.126

3.160

2.241

1.476

0.909

0.525

4

5.957

4.990

4.024

3.080

2.223

1.512

0.972

6

6.815

5.850

4.886

3.931

3.024

2.215

1.543

1

4.664

3.693

2.727

1.822

1.115

0.636

0.338

2

5.096

4.126

3.160

2.241

1.476

0.909

0.525

3

5.526

4.557

3.592

2.668

1.872

1.243

0.782

4

5.954

4.986

4.022

3.096

2.284

1.613

1.092

0.005

4.231

3.261

2.306

1.435

0.786

0.388

0.175

0.01

4.664

3.694

2.730

1.825

1.104

0.613

0.314

0.015

5.096

4.126

3.160

2.241

1.476

0.909

0.525

0.02

5.525

4.557

3.591

2.666

1.877

1.257

0.800

0

3.797

2.824

1.858

0.993

0.441

0.177

0.065

2

5.096

4.126

3.160

2.241

1.476

0.909

0.525

4

6.385

5.419

4.454

3.512

2.645

1.901

1.302

6

7.665

6.702

5.741

4.789

3.875

3.031

2.288

0.005

5.097

4.127

3.160

2.233

1.444

0.856

0.467

0.01

5.096

4.126

3.160

2.241

1.476

0.909

0.525

0.015

5.093

4.126

3.166

2.266

1.531

0.988

0.610

0.02

5.093

4.132

3.188

2.312

1.606

1.084

0.710

η

λΓ

p

µJ

λN

σJ

8.2 The Ornstein-Uhlenbeck Model

135

Table 8.3. Values of average-rate caps for the jump-enhanced OU model, with a nominal value of 100 units. K

2

3

4

5

6

7

8

0.005

4.037

3.067

2.098

1.168

0.533

0.214

0.077

0.01

4.474

3.507

2.540

1.607

0.915

0.488

0.248

0.015

4.906

3.940

2.975

2.043

1.327

0.840

0.522

0.02

5.331

4.368

3.405

2.474

1.745

1.222

0.851

0

3.593

2.621

1.650

0.754

0.285

0.096

0.029

2

4.037

3.067

2.098

1.168

0.533

0.214

0.077

4

4.478

3.511

2.544

1.597

0.844

0.390

0.162

6

4.918

3.953

2.987

2.033

1.202

0.625

0.291

1

3.816

2.845

1.874

0.953

0.386

0.139

0.045

2

4.037

3.067

2.098

1.168

0.533

0.214

0.077

3

4.257

3.288

2.320

1.387

0.709

0.323

0.134

4

4.476

3.508

2.541

1.607

0.902

0.462

0.219

0.005

3.594

2.622

1.654

0.760

0.251

0.070

0.017

0.01

3.816

2.845

1.875

0.956

0.375

0.126

0.037

0.015

4.037

3.067

2.098

1.168

0.533

0.214

0.077

0.02

4.256

3.288

2.320

1.385

0.715

0.334

0.142

0

3.372

2.399

1.426

0.530

0.126

0.028

0.006

2

4.037

3.067

2.098

1.168

0.533

0.214

0.077

4

4.697

3.731

2.765

1.819

1.046

0.537

0.250

6

5.352

4.389

3.427

2.475

1.623

0.968

0.529

0.005

4.038

3.068

2.099

1.163

0.505

0.182

0.056

0.01

4.037

3.067

2.098

1.168

0.533

0.214

0.077

0.015

4.035

3.066

2.099

1.184

0.574

0.259

0.109

0.02

4.033

3.066

2.106

1.212

0.624

0.311

0.149

η

λΓ

p

µJ

λN

σJ

136

8 Jump-Enhanced One-Factor Interest-Rate Models

8.3 The Square-Root Model 8.3.1 Derivation of the Characteristic Function Modeling the short rate as a Square-Root process was introduced in Cox, Ingersoll and Ross (1985b) to demonstrate the equilibrium approach described in Cox, Ingersoll and Ross (1985a). In contrast to the arbitrage-based approach used in Vasicek (1977), the relevant interest-rate dynamics of the CIR model was derived within an equilibrium-based approach. The main advantage in modeling the short rate as a Square-Root process lies in its nonnegativity property. Thus, interest rates governed by a Square-Root process always stay positive178 . This ability, together with the maintained tractability, oﬀers a very useful tool in modeling the term structure of interest rates. Ahn and Thompson (1988) extend the diﬀusion model with a constant jump size, which is triggered by a Poisson process179 . Zhou (2001) uses a CIR model augmented with a uniformly distributed jump size for estimation purposes. Similar to the Vasicek model, this short-rate process has a mean-reverting component, which is crucial in depicting the term structure faithfully. However, the coeﬃcient governing the diﬀusion part has now a stochastic component governed by the factor xt itself. The the risk-neutral coeﬃcients are µQ 0 = κθ,

µQ 1 = −κ,

σ0 = 0,

√ σ1 = σ xt .

Thus, the diﬀusion part of the SDE (8.1) is √ dxt = κ(θ − xt ) dt + σ xt dWtQ .

(8.9)

Modeling the short-rate process this way bears several advantages. Firstly, as mentioned above, the interest-rate model displays a stochastic volatility without incorporating an additional factor. Secondly, as long as the initial value suﬃces xt ≥ 0 together with the condition 2κθ ≥ σ 2 , the model guarantees that the short rate never reaches the origin and therefore stays strictly positive180 . In contrast to the normally distributed short-rate process in Vasicek (1977), the mean-reverting Square-Root process exhibits a non-central Chi-Square distribution with expectation 178

Setting the discount parameters to w0 = 0 and w1 = 1, the general Square-Root

179

model as used in this thesis and the CIR model coincides. See Ahn and Thompson (1988), p. 168. See Feller (1951), p. 173.

180

8.3 The Square-Root Model

137

EQ [xT ] = xt e−κτ + θ(1 − e−κτ ), and variance 2 σ 2 −κτ σ2 e 1 − e−2κτ . − e−2κτ + θ κ 2κ √ Due to the stochastic volatility term σ xt , the derivation of the general characteristic function is more tedious, but also straightforward. In this case, the ordinary diﬀerential equation for the coeﬃcient function ˜b(z, τ ) has the form VARQ [xT ] = xt

of the well-known Riccati equation, for which several solution methods exist. In order to solve for ˜b(z, τ ), we prepare our diﬀerential equation by substituting the coeﬃcient b(z, τ ) in equation (8.4) with ˜b(z, τ ). This leads to the alternative representation 2 2 2 2 ˜b(z, τ )τ = − w1 + ızκg1 + σ z g1 + ızσ 2 g1 − κ ˜b(z, τ ) + σ ˜b(z, τ )2 . 2 2 Thus, introducing the parameters σ 2 z 2 g12 c0 (z) = − w1 + ızκg1 + , 2 c1 (z) = ızσ 2 g1 − κ, c2 (z) =

σ2 , 2

we are able to express this ODE simply as ˜b(z, τ )τ = c0 (z) + c1 (z)˜b(z, τ ) + c2 (z)˜b(z, τ )2 ,

(8.10)

for which standardized solution techniques exist. Eventually, we obtain the functional form of the coeﬃcient function ˜b(z, τ ) as181 ˜b(z, τ ) = with

(z, τ )

ϑ(z)τ 2

,

(8.11)

ϑ(z)τ ϑ(z)τ (z, τ ) = ϑ(z) cosh − c1 (z) sinh , 2 2

and ϑ(z) = 181

2c0 (z) sinh

0 c1 (z)2 − 4c0 (z)c2 (z).

The detailed derivation of the coeﬃcient functions ˜b(z, τ ) and a0 (z, τ ) is shown in Appendix A.

138

8 Jump-Enhanced One-Factor Interest-Rate Models

Given the coeﬃcient function ˜b(z, τ ), we can proceed onward with the calculation of a0 (z, τ ), which represents the antiderivative of b(z, τ ) = ˜b(z, τ ) + ızg1 , scaled by some constant factor κθ. Applying a logarithmic integration approach, the solution is formally given by

κθ (z, τ ) 0 a (z, τ ) = (ızκθg1 − w0 )τ − τ c1 (z) + 2 ln . 2c2 (z) ϑ(z)

(8.12)

Equipped with these two coeﬃcient functions, we are already able to price interest-rate derivatives for ordinary diﬀusion speciﬁcations of the short rate without considering any jump components. Implementing a jump component in the Square-Root model, one must be careful about the jump speciﬁcations. Due to the strict positiveness of the model, we have to limit ourselves to cases of positively sized exponentially and gamma distributed jump sizes, thus excluding the normal distribution for the jump size speciﬁcations182 . Similar to the Ornstein-Uhlenbeck model, a closed-form formula of the general characteristic function exists only in case of an exponentially distributed jump component. Thus, calculating the jump transform for this speciﬁcation, we obtain for a jump-enhanced SquareRoot model, where the nth (positively directed) jump in xt is governed by an exponential distribution with mean η, the coeﬃcient function (n)

(n)

a1Ex = − λQ τ + λQ × ) 1− c2 (z)c3 (z) + ηc12(z) τ − η ln (z,τ ϑ(z)

η c3 (z)

˜b(z, τ )

c2 (z)c23 (z) + η (c0 (z)η + c1 (z)c3 (z))

(8.13) ,

with c3 (z) = 1 − ızηg1 . For a gamma distributed jump size, we again use a Runge-Kutta solver to recover the relevant values for the coeﬃcient function a1 (z, τ ). 8.3.2 Numerical Results Given the two diﬀerent jump candidates, we want to demonstrate the impact on the density function as well as interest-rate derivative prices. Figures 8.4 182

However, Ahn and Thompson (1988) implemented a constant, negatively sized jump component in a CIR short-rate model. Accordingly, they have to choose carefully the ﬁxed jump amplitude to ensure that interest rates remain positive over the trading interval τ .

8.3 The Square-Root Model

139

and 8.5 depict density functions for short-rate models with diﬀusion parameters rt = 0.03, κ = 0.3, θ = 0.03, σ = 0.1 and T = 1. In each ﬁgure, we focus exclusively on one particular jump candidate, while ignoring other jump speciﬁcations.The probability density functions are examined for diﬀerent jump intensities, means, and in case of a gamma jump size speciﬁcation we also investigate the behavior of the density function for varying p. Thus, we have in each ﬁgure the diﬀusion base model exclusively combined with one jump speciﬁcation. Subsequently, model prices of idealized interest-rate contracts are derived similar to the payoﬀ functions in Table 4.1. Here, we only compute derivative prices for the gamma jump-enhanced diﬀusion model because the gamma distribution is able to generate the exponential distribution as a special case. In contrast to the Vasicek model, the density function of the pure diﬀusion model innately shows an asymmetric shape, since the instantaneous interest rate rt features a non-central chi-square probability density function. The effect of jump components can be seen by comparing the density function of the ordinary CIR diﬀusion model, which is depicted in the particular (upper) left graphs of Figure 8.4 and 8.5 for λ = 0, with the behavior of the jump-enhanced density function. Particularly, empirical ﬁndings of right-skewed density functions183 can be assembled within the jump-enhanced model. As mentioned earlier, we consider only positively sized, exponentially and gamma distributed jumps due to the positivity constraint of the Square-Root process, thus neglecting the normal distribution speciﬁcation for jump candidates in the CIR model. For both jump speciﬁcations we notice a higher skewness of the density function compared to the pure diﬀusion case. However, the jump intensity and jump size mean parameters inﬂuence the density function diﬀerently. According to the (upper) left graphs in Figures 8.4 and 8.5, increased arrival times show the eﬀect of distributing the probability mass over a broader range and shifting the mode of the density to the right, which is characteristic for the intensity parameter. Compared to the Vasicek model, this eﬀect is not that pronounced, which might be due to the non-central chi-squared distribution of the short rate. On the other hand, increasing the parameter of the jump size mean results in fat tails to the right. Accordingly, the density functions display a lower kurtosis. Comparing the particular graphs for the exponential 183

See, for example, Arapis and Gao (2006).

140

8 Jump-Enhanced One-Factor Interest-Rate Models 30

30 0 2 4 6

20

15

10

5

0

0.005 0.01 0.015 0.02

25

probability density

probability density

25

20

15

10

5

0

0.02

0.04

0.06

0.08

0.1

r

T

0.12

0.14

0.16

0.18

0.2

0

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

T

Fig. 8.4. Probability densities for a short rate governed by a CIR diﬀusion model enhanced with an exponentially distributed jump component. In the left (right) graph the density functions for varying jump intensities (means) are depicted. The base parameters are: rt = 0.03, κ = 0.3, θ = 0.03, σ = 0.1, λ = 2, η = 0.005, T = 1.

and gamma jump case and interpret Figure 8.4 as a special case of a gamma distributed jump size variable with p = 1, we clearly identify the multiplying eﬀect of p on the jump intensity. Especially in the upper right graph of Figure 8.5, we notice the extremely ﬂat tail of the density function for η = 0.02 compared to the behavior of the particular graph in Figure 8.4. Examining the eﬀect of jump parameters on derivative prices, we assume for all contingent claims the following base parameters: rt = 0.03, κ = 0.3, θ = 0.03, σ = 0.1, λ = 2, η = 0.005, p = 2 and τ = 0.5. Firstly, we have a look at Table 8.4, where values of zero-bond calls are computed according to a strike range of 60 to 90 units. The strike range is chosen in a way to include either ITM, ATM and OTM option prices184 . As in the Vasicek framework, we observe in Table 8.4 the jump intensity λ to have the greatest inﬂuence on zero-bond call values, followed by the jump mean parameter η and the parameter p. Since varying the jump mean η and the parameter p keeps the mode of the density nearly unchanged, a smaller amount of the probability mass is moved out of the exercise region of the zero-bond call. Comparing zero-bond call prices of the particular parameter settings and strike prices, thus keeping the overall expected jump size, based on η, λΓ and p, equal, we observe relatively low spreads between ITM option prices, while spreads for OTM option prices are high. Turning our attention to the cap contracts, 184

The value of a zero bond with remaining maturity of Tˆ − T = 3 is 85.525 units.

8.3 The Square-Root Model 30

30 0 2 4 6

0.005 0.01 0.015 0.02

25

20

probability density

probability density

25

15

10

5

0

141

20

15

10

5

0

0.05

0.1

0.15

0.2

0.25

0.3

r

0

0

0.05

0.1

0.15

0.2

0.25

0.3

r

T

T

30 1 2 3 4

probability density

25

20

15

10

5

0

0

0.05

0.1

0.15

0.2

0.25

0.3

r

T

Fig. 8.5. Probability densities for a short rate governed by a CIR diﬀusion model enhanced with a gamma distributed jump component. In the upper left (right) graph the density functions for varying jump intensities (means) are depicted. The lower graph shows the density behavior for alternating values of p. The base parameters are: rt = 0.03, κ = 0.3, θ = 0.03, σ = 0.1, λ = 2, η = 0.005, p = 2, T = 1.

we only consider one payment date at the maturity of both the ordinary and the average-rate cap. In Table 8.5, it is ﬁrst of all evident that the inﬂuence of jump parameters is reversed. Thus, the jump mean involves the greatest increase in cap prices, whereas the arrival rate results in a smaller increase in cap prices. Comparing contract values for alternating jump intensities, we have, in absence of any jump, relatively close values for ITM options of the ordinary and the average-rate cap. However, for ATM options we observe relatively large diﬀerences between both contracts. By neglecting positively sized jumps, the opposite eﬀect could be observed for rt < θ, because of the averaging process.

142

8 Jump-Enhanced One-Factor Interest-Rate Models

Table 8.4. Values of zero-bond call options for the jump-enhanced SR model, where the underlying zero-bond contract has a nominal value of 100 units. K

60

65

70

75

80

85

90

23.625

18.711

13.797

8.890

4.117

0.595

0

0.01

17.013

12.128

7.345

3.044

0.339

0

0

0.015

11.185

6.610

2.704

0.297

0

0

0

0.02

6.454

2.762

0.374

0

0

0

0

0

31.018

26.093

21.167

16.242

11.316

6.394

1.770

2

23.625

18.711

13.797

8.890

4.117

0.595

0

4

16.861

11.960

7.096

2.648

0.198

0

0

6

10.678

5.879

1.811

0.079

0

0

0

1

27.225

22.305

17.385

12.466

7.551

2.816

0.121

2

23.625

18.711

13.797

8.890

4.117

0.595

0

3

20.207

15.300

10.400

5.603

1.567

0.023

0

4

16.963

12.068

7.252

2.921

0.310

0

0

η 0.005

λ

p

8.3 The Square-Root Model

143

Table 8.5. Values of short-rate caps (Panel A) and average-rate caps (Panel B), for the jump-enhanced SR model, with a nominal value of 100 units. Panel A: Caps K

2

3

4

5

6

7

8

0.005

1.931

1.146

0.609

0.296

0.135

0.058

0.024

0.01

2.821

2.003

1.375

0.924

0.610

0.395

0.252

0.015

3.705

2.877

2.212

1.692

1.284

0.965

0.719

0.02

4.580

3.748

3.065

2.507

2.043

1.655

1.332

0

1.070

0.443

0.140

0.035

0.007

0.001

0

2

1.931

1.146

0.609

0.296

0.135

0.058

0.024

4

2.811

1.938

1.236

0.735

0.411

0.218

0.110

6

3.699

2.778

1.968

1.316

0.834

0.503

0.290

1

1.490

0.760

0.323

0.118

0.039

0.012

0.003

2

1.931

1.146

0.609

0.296

0.135

0.058

0.024

3

2.376

1.563

0.959

0.557

0.308

0.164

0.085

4

2.821

1.994

1.347

0.876

0.550

0.335

0.199

K

2

3

4

5

6

7

8

0.005

1.451

0.618

0.189

0.050

0.012

0.003

0.001

0.01

1.907

1.052

0.528

0.262

0.129

0.063

0.030

0.015

2.357

1.495

0.923

0.578

0.361

0.224

0.137

0.02

2.801

1.935

1.338

0.942

0.662

0.463

0.322

0

0.994

0.261

0.028

0.001

0

0

0

2

1.451

0.618

0.189

0.050

0.012

0.003

0.001

4

1.909

1.016

0.428

0.154

0.050

0.015

0.004

6

2.367

1.439

0.730

0.319

0.125

0.045

0.015

1

1.222

0.423

0.083

0.012

0.002

0

0

2

1.451

0.618

0.189

0.050

0.012

0.003

0.001

3

1.680

0.830

0.334

0.123

0.043

0.014

0.005

4

1.908

1.048

0.506

0.230

0.100

0.041

0.016

η

λ

p

Panel B: Average-Rate Caps η

λ

p

9 Jump-Enhanced Two-Factor Interest-Rate Models

9.1 Overview In this chapter, we derive the characteristic functions for one speciﬁc additive interest-rate model and one subordinated stochastic volatility interest-rate model. As in the one-factor case, we extend these pure diﬀusion models with additional jump components. The diﬀusion part of the additive model, which is discussed consists of both a factor governed by an Ornstein-Uhlenbeck process, and a factor modeled as a Square-root process. Other popular additive models are given by pure multi-factor versions of Ornstein-Uhlenbeck and Square-Root processes185 . The additive interest-rate model was ﬁrst introduced in Sch¨ obel and Zhu (2000). Here the authors apply a Heston-like transformation methodology to price interest-rate derivatives, as demonstrated in Section 4.2. The subordinated model we choose for our analysis was presented in Fong and Vasicek (1991a). Here, both the short rate and its stochastic volatility are modeled as Square-Root processes with additional jump components. In this thesis, we extend both models to incorporate various jump components. In each model, the behavior of the particular density function and numerical values of idealized interest-rate options are examined. 185

The interest rate is then modeled as the sum either of some Ornstein-Uhlenbeck processes deﬁned by the SDE (8.5) or of some mean-reverting Square-Root processes according to (8.9). Modeling the short rate as an additive Square-Root model, all Brownian motions have to be uncorrelated in order to derive closedform solutions for the general characteristic function.

146

9 Jump-Enhanced Two-Factor Interest-Rate Models

9.2 The Additive OU-SR Model 9.2.1 Derivation of the Characteristic Function Basically, additive multi-factor short-rate models consist only of either additive mean-reverting Ornstein-Uhlenbeck, or Square-Root processes. For example, an additive model for the short rate is used in Chen and Scott (1992), Longstaﬀ and Schwartz (1992), and Chen and Scott (1995). There, the short rate is modeled as the sum of two independent Square-Root processes. A multivariate, additive Gaussian interest-rate model with correlated factors is given in e.g. Langetieg (1980). Collin-Dufresne and Goldstein (2002) also consider both additive multivariate Ornstein-Uhlenbeck and Square Root processes in pricing swaptions. In the case of Ornstein-Uhlenbeck processes, the diﬀerent Brownian motions driving the particular factors can be correlated. On the other hand, if taking an additive Square-Root model, we have to impose the restriction that all Brownian motions be mutually uncorrelated. Otherwise, the separation approach is no longer valid and no closed-form solution for the general characteristic function would exist186 . Exemplary for the set of additive model candidates we select a term-structure model where the short-rate process consists of two factors. The ﬁrst factor xOU is modeled as an Ornsteint Uhlenbeck process, according to equation (8.5), whereas the second factor xSR t is governed by a Square-Root process, as given in equation (8.9)187 . Since we want to extend the model setup, we allow for both factors to include jump components subject to possible non-negativity constraints. Subsequently, the short rate is built as the weighted sum of those factors with a scaling factor w ∈ [0, 1], which gives r(xt ) = wxOU + (1 − w)xSR t t .

(9.1)

Therefore, the coeﬃcients characterizing the short rate are w = (w, 1 − w) and w0 = 0. Accordingly, we use a slightly modiﬁed version of the model setup introduced in Sch¨ obel and Zhu (2000). 186

187

In this case, even a Runge-Kutta solver cannot be applied to the valuation problem, due to the missing system of ODEs. We assume the parameters for the particular processes to be κi , θi , σ i with i ∈ {OU, SR}. Furthermore, we use the payoﬀ-characterizing coeﬃcients g0i and g1i .

9.2 The Additive OU-SR Model

147

Setting w = 1 we obtain the Vasicek model and for w = 0 we obtain the CIR model according to Section 8.3. Although the factors are linearly combined within the short rate, all derivative functions, e.g. the probability density function, are not just simple linear combinations of their particular one-factor counterparts, which is illustrated in Figure 9.1. Thus, the additive process allows more ﬂexibility in modeling the term structure of interest rates compared to the one-factor models discussed in Chapter 8, while maintaining the simple structure of coeﬃcients used in the general characteristic function.

35 0.25 0.5 0.75

30

probability density difference

25 20 15 10 5 0 −5 −10 −15

0

0.01

0.02

0.03

0.04

0.05 r

0.06

0.07

0.08

0.09

0.1

T

Fig. 9.1. Diﬀerences of the pure diﬀusion OU-SR model density function and the sum of the particular one-factor pendants for diﬀerent weighting factors. The parameters used are: xt = (0.05, 0.03) , κ = (0.4, 0.3) , θ = (0.05, 0.03) , σ = (0.01, 0.1) , T = 1. In case of the one-factor models the ﬁrst (last) elements correspond to the Vasicek (CIR) model.

Due to the independence of the two Brownian motions, the particular timedependent coeﬃcients exhibit the same formal structure as the ones derived in the one-factor Vasicek and CIR interest-rate models. Thus, the general char-

148

9 Jump-Enhanced Two-Factor Interest-Rate Models

acteristic function of this additive interest-rate model has the time-dependent vector function ˜ τ) = b(z,

˜bOU (z, τ ) , ˜bSR (z, τ )

with ˜bOU (z, τ ) and ˜bSR (z, τ ) given by their one-factor representations in equation (8.6) and (8.11) with adapted parameters. Consequently, we obtain for the coeﬃcient function a(z, τ ) the relation a0 (z, τ ) = a0OU (z, τ ) + a0SR (z, τ ), where a0OU (z, τ ) and a0SR (z, τ ) correspond to equation (8.7) and (8.12). The jumps contained in the vectors jxOU and jxSR are both triggered by the same Poisson vector process N(λQ )188 .Due to the independence of the two factors governing the short rate, we are also able to adapt the jump transforms of the particular one-factor models without altering their formal structure. 9.2.2 Numerical Results In this section, we show the behavior of the density function and compute values for some common interest-rate options under the additive jump-diﬀusion model. As base parameters for both the density function and the interestrate contracts, we use the particular parameters according to their one-factor counterparts. The default value of the scaling parameter w is set to 12 . The impact of jumps on the short-rate density is demonstrated in Figures 9.2 and 9.3. As before, we focus exclusively on one particular jump candidate, while ignoring other jump speciﬁcations. Thus, the graphs in the ﬁrst row and the left graph in the second row in Figure 9.2, respectively, display only the impact of gamma jump component of the Ornstein-Uhlenbeck process. The other three graphs in this ﬁgure depict the inﬂuence of the normal jump component on the short-rate density. Consequently, in Figure 9.3, we only consider the gamma jump component of the Square-Root process. Option prices for varying parameters of normally and gamma distributed jump amplitudes are given in Tables 9.1 - 9.6. Again, we focus on the sensitivity of option prices to jump parameters and neglect the exponentially distributed jump size since the exponential distribution is a special case of the gamma distribution. 188

However, setting elements in the jump vectors jxOU and jxSR to zero, it is possible to assign jump components to particular processes.

9.2 The Additive OU-SR Model

149

Comparing the ﬁgures of the densities functions depicted in 8.2, 8.3 and 8.5 with the corresponding graphs in Figures 9.2 and 9.3, we obviously notice more skewness in the densities of the two-factor model compared to the densities in the Vasicek model and a more leptokurtic behavior of the densities in the additive model compared to the ones in a CIR model. Firstly, taking a look at the inﬂuence of the gamma distributed jump size speciﬁcation, we determine a similar inﬂuence of the gamma jump component belonging to the Vasicek and CIR model. However, the Vasicek part has a signiﬁcantly weaker eﬀect on the probability density function compared to the relevant one-factor model. On the other hand, the gamma distributed jump size component of the Square-Root process has a considerable impact on the probability density, which can be justiﬁed by the similar shape of the probability density functions in Figures 8.5 and 9.3. Examining the impact of a normally distributed jump component in this model, we observe only small changes in contrast to the one-factor equivalent Vasicek model. Thus, in the multi-factor setup we no longer encounter the characteristic strong curvature in the probability density function and the bimodal distribution displayed in the upper right graph of Figure 8.3. However, the eﬀect of an increased volatility of the normally distributed jump size, which raises both tails of the probability density function, remains immanent. Computing numerical values of interest-rate derivatives in this model, we assume for all contingent claims the following base diﬀusion parameters: 0.05 0.4 0.05 0.01 xt = , κ= , θ= and σ = . 0.03 0.3 0.03 0.1 In each vector, the ﬁrst element corresponds to the Ornstein-Uhlenbeck part, whereas the second element states the parameter value for the Square-Root component in this particular additive interest-rate model. The default jump parameters used for the valuation are in case of a gamma distributed jump size candidate λiΓ = 2, η i = 0.005 and pi = 2 with i ∈ {OU, SR}. The normally distributed jump component of the Ornstein-Uhlenbeck process is governed by the parameters λOU = 2, µOU = 0.015 and σJOU = 0.01. All N J contracts have a remaining time to maturity of a half year. Let us discuss ﬁrst Tables 9.1 and 9.2, where numerical values of zero-bond calls are reported according to equation (3.12) for a strike range from 60 to 90 units. The strike

150

9 Jump-Enhanced Two-Factor Interest-Rate Models 50

50 0 2 4 6

45 40

40 35

probability density

probability density

35 30 25 20

30 25 20

15

15

10

10

5 0

0.005 0.01 0.015 0.02

45

5 0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0

0.2

0

0.02

0.04

0.06

0.08

r

40

0.18

0.2

1 2 3 4

40 35

probability density

probability density

0.16

45

35 30 25 20

30 25 20

15

15

10

10

5

5 0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0

0.2

0

0.02

0.04

0.06

0.08

r

0.1

0.12

0.14

0.16

0.18

0.2

r

T

T

50

50 0 0.01 0.02 0.03

45 40

0.005 0.01 0.015 0.02

45 40 35

probability density

35

probability density

0.14

50 1 2 3 4

45

30 25 20

30 25 20

15

15

10

10

5 0

0.12

T

50

0

0.1

r

T

5 0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0

0

0.02

0.04

rT

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

rT

Fig. 9.2. Probability densities for a short rate governed by an OU-SR diﬀusion model enhanced with either a gamma or normally distributed jump component for the OU process. In the upper left (right) graph density functions for varying jump intensities (means) of the gamma distributed jump component are depicted. The graphs in the second row show the density behavior for alternating values of pOU and the jump intensity λOU N of the normally distributed jump component. In the last row, the left (right) graph shows density functions for diﬀerent values of jump mean (volatility) of the normally distributed jump component. The base parameters are: = 2, η OU = xt = (0.05, 0.03) , κ = (0.4, 0.3) , θ = (0.05, 0.03) , σ = (0.01, 0.1) , λOU Γ OU 0.005, pOU = 2, λOU = 0.02, σJOU = 0.01, T = 1. N = 1, µJ

9.2 The Additive OU-SR Model 50

50 0 2 4 6

45 40

40 35

probability density

probability density

0.005 0.01 0.015 0.02

45

35 30 25 20

30 25 20

15

15

10

10

5 0

151

5 0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

0

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

T

T

50 1 2 3 4

45 40

probability density

35 30 25 20 15 10 5 0

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

T

Fig. 9.3. Probability densities for a short rate governed by an OU-SR diﬀusion model enhanced with a gamma distributed jump component for the SR process. In the upper left (right) graph the density functions for varying jump intensities (means) are depicted. The lower graph shows the density behavior for alternating values of p. The base parameters are: xt = (0.05, 0.03) , κ = (0.4, 0.3) , θ = = 2, η SR = 0.005, pSR = 2, T = 1. (0.05, 0.03) , σ = (0.01, 0.1) , λSR Γ

range is chosen to include either ITM, ATM and OTM option prices189 . As encountered in the one-factor framework, we observe in case of the gamma jump-size distribution that the intensity λiΓ has the greatest inﬂuence on zero-bond call values, followed by the parameter pi and the jump mean η i . This can be explained by the shifting of the density function to the right for increasing arrival rates. Contrary, varying jump mean η i and parameter pi keeps the mode of the density nearly unchanged, so that a smaller amount of the probability mass is moved out of the zero-bond call exercise region. Varying the parameters of the normally distributed jump component, we also 189

The value of a zero bond with remaining maturity of two years, priced with base parameters, is 87.359 units.

152

9 Jump-Enhanced Two-Factor Interest-Rate Models

observe a very similar behavior of option prices in comparison to the onefactor model. Comparing relative zero-bond call price diﬀerences of particular parameter settings and strike rates, thus keeping the overall expected jump size equal, we observe low spreads between ITM option prices, whereas spreads for OTM option prices are relative high. For the cap contracts, we only allow one payment date, which is at maturity. At ﬁrst, we observe the same positive eﬀect of the jump components as encountered in the one-factor pendants. Thus, the jump mean involves the greatest increase of cap prices, whereas the arrival rate results in a smaller increase of cap prices. However, due to the scaling factor w, the eﬀect of diﬀerent jump components is not as strong as we encountered in the particular one-factor models. Comparing the eﬀect of a gamma distributed jump on the average-rate cap, we compute nearly the same contract values, whether we have the jumps in the Ornstein-Uhlenbeck or in the Square-Root part of the model.

9.2 The Additive OU-SR Model

153

Table 9.1. Values of zero-bond call options for the jump-enhanced OU-SR model, where the underlying zero-bond contract has a nominal value of 100 units. K

60

65

70

75

80

85

90

24.825

19.944

15.063

10.183

5.304

0.875

0

OU

η

0.005 0.01

23.022

18.147

13.271

8.396

3.573

0.201

0

0.015

21.275

16.406

11.537

6.680

2.137

0.021

0

0.02

19.584

14.720

9.862

5.092

1.125

0

0

0

26.688

21.801

16.915

12.028

7.142

2.326

0.003

2

24.825

19.944

15.063

10.183

5.304

0.875

0

4

23.006

18.131

13.256

8.380

3.531

0.152

0

6

21.230

16.360

11.491

6.622

1.926

0.006

0

1

25.750

20.866

15.982

11.098

6.215

1.521

0

2

24.825

19.944

15.063

10.183

5.304

0.875

0

3

23.914

19.036

14.158

9.280

4.412

0.438

0

4

23.017

18.141

13.266

8.391

3.556

0.190

0

0.005

26.685

21.799

16.912

12.025

7.139

2.312

0.005

0.01

25.748

20.865

15.981

11.097

6.214

1.511

0

0.015

24.825

19.944

15.063

10.183

5.304

0.875

0

0.02

23.916

19.037

14.159

9.281

4.415

0.445

0

0

27.631

22.741

17.852

12.962

8.073

3.197

0.029

2

24.825

19.944

15.063

10.183

5.304

0.875

0

4

22.118

17.245

12.373

7.501

2.711

0.048

0

6

19.505

14.641

9.777

4.924

0.792

0

0

0.005

24.821

19.940

15.059

10.178

5.299

0.847

0

0.01

24.825

19.944

15.063

10.183

5.304

0.875

0

0.015

24.832

19.951

15.070

10.189

5.312

0.918

0

0.02

24.841

19.960

15.080

10.199

5.325

0.973

0.001

λOU Γ

OU

p

µOU J

λOU N

σJOU

154

9 Jump-Enhanced Two-Factor Interest-Rate Models

Table 9.2. Values of zero-bond call options for the jump-enhanced OU-SR model, where the underlying zero-bond contract has a nominal value of 100 units. K

60

65

70

75

80

85

90

24.825

19.944

15.063

10.183

5.304

0.875

0

SR

η

0.005 0.01

22.891

18.016

13.141

8.266

3.466

0.183

0

0.015

21.022

16.153

11.284

6.442

1.997

0.014

0

0.02

19.217

14.354

9.504

4.794

1.001

0

0

0

26.828

21.941

17.055

12.168

7.281

2.444

0.004

2

24.825

19.944

15.063

10.183

5.304

0.875

0

4

22.873

17.997

13.122

8.247

3.407

0.136

0

6

20.969

16.100

11.230

6.362

1.737

0.004

0

1

25.819

20.935

16.051

11.167

6.284

1.566

0

2

24.825

19.944

15.063

10.183

5.304

0.875

0

3

23.847

18.969

14.091

9.213

4.349

0.422

0

4

22.885

18.010

13.135

8.260

3.443

0.173

0

λSR Γ

SR

p

9.2 The Additive OU-SR Model

155

Table 9.3. Values of short-rate caps for the jump-enhanced OU-SR model, with a nominal value of 100 units. K

2

3

4

5

6

7

8

0.005

3.507

2.532

1.593

0.832

0.358

0.129

0.040

0.01

3.943

2.969

2.024

1.220

0.651

0.312

0.136

0.015

4.376

3.403

2.457

1.635

1.012

0.592

0.330

0.02

4.806

3.835

2.889

2.058

1.404

0.928

0.597

0

3.069

2.094

1.186

0.534

0.193

0.058

0.015

2

3.507

2.532

1.593

0.832

0.358

0.129

0.040

4

3.944

2.970

2.014

1.176

0.579

0.242

0.087

6

4.380

3.407

2.443

1.554

0.854

0.403

0.164

1

3.288

2.313

1.384

0.668

0.261

0.085

0.023

2

3.507

2.532

1.593

0.832

0.358

0.129

0.040

3

3.726

2.751

1.807

1.015

0.482

0.196

0.069

4

3.943

2.969

2.023

1.210

0.631

0.288

0.117

0.005

3.070

2.096

1.184

0.518

0.176

0.048

0.011

0.01

3.289

2.314

1.384

0.662

0.252

0.079

0.021

0.015

3.507

2.532

1.593

0.832

0.358

0.129

0.040

0.02

3.725

2.751

1.807

1.019

0.489

0.202

0.073

0

2.850

1.875

0.972

0.370

0.107

0.025

0.005

2

3.507

2.532

1.593

0.832

0.358

0.129

0.040

4

4.162

3.188

2.232

1.378

0.736

0.340

0.137

6

4.814

3.842

2.878

1.971

1.210

0.660

0.320

0.005

3.508

2.532

1.589

0.815

0.335

0.112

0.031

0.01

3.507

2.532

1.593

0.832

0.358

0.129

0.040

0.015

3.507

2.533

1.603

0.859

0.392

0.155

0.054

0.02

3.506

2.537

1.621

0.896

0.435

0.189

0.075

OU

η

λOU Γ

OU

p

µOU J

λOU N

σJOU

156

9 Jump-Enhanced Two-Factor Interest-Rate Models

Table 9.4. Values of short-rate caps for the jump-enhanced OU-SR model, with a nominal value of 100 units. K

2

3

4

5

6

7

8

0.005

3.507

2.532

1.593

0.832

0.358

0.129

0.040

0.01

3.953

2.979

2.035

1.232

0.664

0.325

0.147

0.015

4.397

3.424

2.478

1.657

1.037

0.618

0.353

0.02

4.838

3.866

2.920

2.089

1.439

0.965

0.633

0

3.059

2.083

1.173

0.518

0.181

0.052

0.013

2

3.507

2.532

1.593

0.832

0.358

0.129

0.040

4

3.955

2.980

2.026

1.190

0.593

0.252

0.093

6

4.402

3.428

2.465

1.579

0.881

0.427

0.181

1

3.283

2.308

1.378

0.661

0.254

0.080

0.021

2

3.507

2.532

1.593

0.832

0.358

0.129

0.040

3

3.731

2.756

1.813

1.021

0.490

0.202

0.073

4

3.954

2.980

2.034

1.222

0.645

0.301

0.127

SR

η

λSR Γ

SR

p

9.2 The Additive OU-SR Model

157

Table 9.5. Values of average-rate caps for the jump-enhanced OU-SR model, with a nominal value of 100 units. K

2

3

4

5

6

7

8

0.005

2.753

1.777

0.832

0.229

0.041

0.005

0.001

0.01

2.977

2.002

1.053

0.389

0.112

0.027

0.006

0.015

3.199

2.225

1.275

0.579

0.232

0.087

0.032

0.02

3.419

2.446

1.496

0.781

0.385

0.184

0.087

0

2.528

1.551

0.625

0.135

0.019

0.002

0

2

2.753

1.777

0.832

0.229

0.041

0.005

0.001

4

2.978

2.003

1.046

0.349

0.077

0.012

0.002

6

3.202

2.228

1.264

0.494

0.129

0.025

0.004

1

2.641

1.664

0.725

0.174

0.027

0.003

0

2

2.753

1.777

0.832

0.229

0.041

0.005

0.001

3

2.865

1.890

0.941

0.298

0.064

0.010

0.001

4

2.977

2.002

1.052

0.379

0.099

0.020

0.003

0.005

2.528

1.551

0.625

0.123

0.014

0.001

0

0.01

2.641

1.664

0.725

0.168

0.024

0.002

0

0.015

2.753

1.777

0.832

0.229

0.041

0.005

0.001

0.02

2.865

1.889

0.941

0.302

0.067

0.011

0.002

0

2.416

1.438

0.516

0.079

0.007

0.001

0

2

2.753

1.777

0.832

0.229

0.041

0.005

0.001

4

3.089

2.115

1.158

0.434

0.112

0.022

0.003

6

3.424

2.452

1.488

0.680

0.227

0.057

0.012

0.005

2.753

1.777

0.829

0.215

0.033

0.004

0

0.01

2.753

1.777

0.832

0.229

0.041

0.005

0.001

0.015

2.753

1.777

0.838

0.248

0.053

0.009

0.001

0.02

2.752

1.777

0.849

0.272

0.069

0.015

0.003

OU

η

λOU Γ

OU

p

µOU J

λOU N

σJOU

158

9 Jump-Enhanced Two-Factor Interest-Rate Models

Table 9.6. Values of average-rate caps for the jump-enhanced OU-SR model, with a nominal value of 100 units. K

2

3

4

5

6

7

8

0.005

2.753

1.777

0.832

0.229

0.041

0.005

0.001

0.01

2.980

2.005

1.056

0.393

0.116

0.030

0.007

0.015

3.206

2.232

1.282

0.587

0.240

0.094

0.035

0.02

3.429

2.457

1.507

0.792

0.396

0.195

0.094

0

2.524

1.547

0.621

0.130

0.017

0.002

0

2

2.753

1.777

0.832

0.229

0.041

0.005

0.001

4

2.981

2.006

1.050

0.354

0.079

0.013

0.002

6

3.209

2.235

1.272

0.502

0.136

0.027

0.004

1

2.639

1.662

0.723

0.171

0.025

0.003

0

2

2.753

1.777

0.832

0.229

0.041

0.005

0.001

3

2.867

1.891

0.943

0.300

0.066

0.011

0.002

4

2.980

2.005

1.055

0.383

0.103

0.022

0.004

SR

η

λSR Γ

SR

p

9.3 The Fong-Vasicek Model

159

9.3 The Fong-Vasicek Model 9.3.1 Derivation of the Characteristic Function Apart from the additive modeling approach, the term structure within our exponential-aﬃne framework can also be modeled with the help of subordinated factors. Hence, it is possible to explicitly incorporate the long term mean and/or the volatility as a stochastic factor itself190 . For example, the assumption of a constant volatility in one-factor short-rate models is frequently criticized. In Fong and Vasicek (1991b), the authors argue that a model with a deterministic volatility parameter cannot produce a meaningful volatility exposure. Therefore, they propose to model the variance of a CIR-like shortrate model as a Square-Root process itself. In this section, we use a slightly modiﬁed version of the Fong and Vasicek (1991a) model. Thus, the SDEs for the base diﬀusion model are191 √ drt = κ(θ − rt ) dt + σ vt dW1t , √ dvt = α(¯ v − vt ) dt + β vt dW2t .

(9.2) (9.3)

Equivalently to the target rate θ, often also referred to as the long term mean, of the short rate, v¯ expresses the parameter for a long-term mean of the variance factor vt . In addition to Fong and Vasicek (1991a), we extend this base diﬀusion model with additional jump components jr and jv for the short rate and its volatility factor192 , both triggered by the same vector of Poisson processes N(λQ ). In contrast to the additive OU-SR model discussed in the last section, the Brownian motions W1t and W2t can be correlated as follows: 190

Beaglehole and Tenney (1991) model the long term mean θ of a mean-reverting, normally distributed short rate as a subordinated factor governed by an OrnsteinUhlenbeck process. In Balduzzi, Das, Foresi and Sundaram (1996), the authors model a CIR like short rate with subordinated stochastic mean and volatility factor. However, in both cases, the authors do not give any option prices and

191

derive only zero-bond prices and yields of zero bonds, respectively. For this interest-rate model, we assume w0 = 0 and w = (1, 0) . However, the derivation of the time-dependent coeﬃcients of the characteristic function

192

is shown in Appendix B for general discounting parameter values w0 and w. We only allow strictly positively sized jumps, thus restricting ourselves to positively directed exponentially and gamma distributed jump amplitudes.

160

9 Jump-Enhanced Two-Factor Interest-Rate Models

dW1t dW2t = ρ dt.

(9.4)

However, according to Section 2.1, all Brownian motions have to be uncorrelated within our modeling approach. Fortunately, the feature of correlated Brownian motions can be easily incorporated into our framework193 by using the following diﬀusion-speciﬁc matrix √ Σ(vt ) = β vt

σ β

0

0 1 − ρ2

ρ

.

For convenience, we introduce for this model the following representation of time-dependent coeﬃcient functions194 a0 (z, τ ) = A0 (z, τ ),

and b(z, τ ) =

B(z, τ )

+ ız

C(z, τ )

¯ B C¯

.

Thus, in order to derive the general characteristic function of the state vector xt = (rt , vt ) , we explicitly have to solve the following system of ODEs ¯ + α¯ ¯ A0 (z, τ )τ =κθ(B(z, τ ) + ız B) v (C(z, τ ) + ız C) =A01 (z, τ )τ + A02 (z, τ )τ , ¯ − w, B(z, τ )τ = − κ(B(z, τ ) + ız B)

(9.5) (9.6)

¯ C(z, τ )τ = − α(C(z, τ ) + ız C) 2 σ2 ¯ 2 + β (C(z, τ ) + ız C) ¯ 2 (B(z, τ ) + ız B) 2 2 ¯ ¯ + σβρ(B(z, τ ) + ız B)(C(z, τ ) + ız C).

+

(9.7)

Hence, we are dealing with a system of coupled ODEs. Fortunately, there are no two-sided interdependencies, enabling us to successively solve the diﬀerential equations one-by-one. Starting with equation (9.6), the solution to this diﬀerential equation is easy to obtain and coincides with equation (8.6) for ¯ Also straightforward, but more tedious, is the derivation w1 = w and g1 = B. 193

A standard decomposition of two correlated Brownian motions is applied. Thus, two correlated Brownian motions as given in equation (9.4) allow for the alternative representation dW2t = ρ dW1t +

194

∗ ∗ 1 − ρ2 dW2t , where the processes W2t

and W1t are neither correlated. ¯ In this model setup, the constant parameter g0 is represented by the term A.

9.3 The Fong-Vasicek Model

161

of the coeﬃcient function solving the ODE (9.7). Performing some appropriate transformations on the particular ODE, the solution of the coeﬃcient function corresponding to vt can be obtained as195 C(z, τ ) = − M (z, τ ) + J(z, τ ) (1 + Q(z) − S(z))KU[Q(z) + 1; S(z); Y (z, τ )] + Υ(z) KM[Q(z) + 1; S(z); Y (z, τ )] , with κ M (z, τ ) = 2 β

f1 (z) + 2Q(z) − S(z) + 1+ κ

ρ

1+ 0 ρ2 − 1

(9.8)

Y (z, τ ) ,

2κQ(z) , β 2 (KU[Q(z); S(z); Y (z, τ )] + Υ(z) KM[Q(z); S(z); Y (z, τ )]) 0 σβ ρ2 − 1 w ¯ e−κτ , Y (z, τ ) = + ız B κ κ J(z, τ ) =

S(z) (f3 (z) + κ)ρ − βf2 (z) 0 , + 2 2κ ρ2 − 1 ( f3 (z)2 − 2β 2 f1 (z) S(z) = 1 + , κ2 M (z, 0)KU[Q(z); S(z); Y (z, 0)] Υ(z) = Ξ(z) (1 + Q(z) − S(z)) KU[Q(z) + 1; S(z); Y (z, 0)] − , β2 2κQ(z) Ξ(z) κ Ξ(z) = 2 2 Q(z)KM[Q(z) + 1; S(z); Y (z, 0)] β

Q(z) =

− M (z, 0)KM[Q(z); S(z); Y (z, 0)], and

2¯ 2 2 ¯ α − ızβ C + σβρw + σ w , f1 (z) = −ız C 2 κ 2κ2 σw , f2 (z) = ızβρC¯ − κ σβρw . f3 (z) = ızβ 2 C¯ − α − κ

In the equations above the function KM[a; b; y] is commonly referred to as the Kummer function (of the ﬁrst kind) and KU[a; b; y] represents a conﬂuent 195

The detailed derivation of the coeﬃcient function C(z, τ ) is given in Appendix B.

162

9 Jump-Enhanced Two-Factor Interest-Rate Models

hypergeometric function196 . Both functions represent the two independent solutions of the Kummer equation197 . In contrast to the last derivation, both parts of the coeﬃcient function A (z, τ ) are relatively easy to obtain. Due to the simple structure of the 0

solution given in equation (9.6), we are able to state immediately A01 (z, τ ) = −θ (B(z, τ ) + wτ ) .

(9.9)

For the second part of the diﬀusion component in A0 (z, τ ), we exploit the formal structure and attempt a logarithmic integration, which is shown in Appendix B. Thus, the solution of the second part of the coeﬃcient function A0 (z, τ ) can be written as198 A02 (z, τ ) = −

L(z, τ ) J(z, 0) 2α¯ v ¯ ln + ızα¯ v Cτ, β2 L(z, 0) J(z, τ )

with ln [L(z, τ )] =

S(z) − 1 +

f3 (z) κ

2

ρ

ln[Y (z, τ )]

Y (z, τ ) 1+ 0 2 ρ2 − 1 0 2 1 f3 (z) β ρ −1 + + . ln 2 2κ κ

−

(9.10)

(9.11)

Having calculated the diﬀusion-related coeﬃcients of the general characteristic function, we are now ready for the corresponding jump parts. In case 196

The conﬂuent hypergeometric function is sometimes also denoted as the Kummer function of the second kind and is – like the complex-valued square-root and logarithm – a multi-valued function. Thus, one has to track carefully the path of integration by using this type of function to avoid discontinuities according to the

197

principal branch used by standard mathematical programming environments. The functions KM[a; b; y] and KU[a; b; y] are two independent solutions of the diﬀerential equation y

d2 w(y) dy 2

+ (b − y) dw(y) = aw(y). More information on condy

ﬂuent hypergeometric functions, especially about the computation of KM[a; b; y] 198

and KU[a; b; y] can be found in Abramowitz and Stegun (1972), p. 504. The detailed derivation of A02 (z, τ ) is given in Appendix B.

9.3 The Fong-Vasicek Model

163

of a exponential jump size speciﬁcation on the short rate we are able to use the same jump transformation derived for the Vasicek model. In any other case – meaning gamma distributed jump sizes in the short rate, and for any jump components incorporated in the volatility factor199 – we have to apply a Runge-Kutta algorithm. 9.3.2 Numerical Results In this subsection, we want to demonstrate the impact of diﬀerent jump components and the correlation parameter ρ on the probability density function as well as on option prices with payoﬀ functions similar to Table 4.1, respectively. To the best of our knowledge, option prices for this model, whether of the exponential-aﬃne, linear, or integro-linear type, are presented for the ﬁrst time in this thesis. Articles do exist, which cover the computation of numerical values under the base model. However, only prices of unconditional contracts, such as zero bonds and likewise yields of zero-bond prices, are computed200 . These model prices are easy to obtain due to their similarity of the moment-generating function of the short rate201 . The base diﬀusion parameters we use in computing the probability density function are rt = 0.08, vt = 0.04, κ = 0.2, θ = 0.08, σ = 0.1, α = 0.4, v¯ = 0.04, β = 0.1, ρ = −0.5 and T = 1. Firstly, we examine the behavior of the probability density for alternating correlation speciﬁcations. To avoid the results being biased by the inﬂuence of jump components, we examine ﬁrst the density function of the pure diﬀusion model according to equations (9.2) and (9.3). Obviously, Figure 9.4 shows that the correlation between the short rate and its volatility has an eﬀect on the probability density function, and therefore on the price of any contingent claim. Thus, for low interest rates 199

200

Due to the complicated structure of the coeﬃcient function C(z, τ ), even for exponentially distributed jump amplitudes there exist no closed-form jump transforms. See Fong and Vasicek (1991b). Selby and Strickland (1995) also compute numerical values of zero-bond prices for the base model, but present a technique avoiding the application of hypergeometric functions. In Balduzzi, Das, Foresi and Sundaram (1996), zero-bond prices are computed for an extended version of the Fong-Vasicek model, where the mean of the short rate is also modeled as a

201

stochastic factor. See Proposition 2.4.3.

164

9 Jump-Enhanced Two-Factor Interest-Rate Models

18 −1 0 1

16

probability density

14 12 10 8 6 4 2 0

0

0.02

0.04

0.06

0.08

0.1 rT

0.12

0.14

0.16

0.18

0.2

Fig. 9.4. Probability density functions of the Fong-Vasicek pure diﬀusion model for diﬀerent values of the correlation parameter ρ. The parameters used are: rt = 0.08, vt = 0.04, κ = 0.2, θ = 0.08, σ = 0.1, α = 0.4, v¯ = 0.04, β = 0.1, T = 1.

and negatively correlated factors we observe substantially increased values of the probability density function, in contrast to the probability density function with positively correlated random variables. Since in this scenario we encounter a tendency toward higher volatilities in case of low interest rates, we are dealing with a more volatile process, which explains the behavior of the density. Finally, the correlation parameter can be used to adjust the skewness of the probability probability density function, which is advantageous for calibrating empirical term structures. Next, we examine the inﬂuence of jump speciﬁcations on the density function of the short rate. Thus, we ﬁx the correlation to ρ = −0.5. Comparing the graphs in Figure 9.5 with the particular graphs in Figure 8.5, we observe the same behavior of the density functions considering a gamma distributed jump component on the short rate. This fact is not surprising since the short rate in the Fong-Vasicek interest-rate model is governed by a Square-Root

9.3 The Fong-Vasicek Model 25

25 0 2 4 6

0.005 0.01 0.015 0.02

20

probability density

probability density

20

15

10

5

0

165

15

10

5

0

0.05

0.1

0.15

0.2

0.25

0.3

0

0

0.05

r

0.1

0.15

0.2

0.25

0.3

r

T

T

25 1 2 3 4

probability density

20

15

10

5

0

0

0.05

0.1

0.15

0.2

0.25

0.3

r

T

Fig. 9.5. Probability densities for a short rate governed by the Fong-Vasicek diffusion model enhanced with a gamma distributed jump component for the shortrate process. In the upper left (right) graph the density functions for varying jump intensities (means) are depicted. The lower graph shows the density behavior for alternating values of p. The base parameters are: rt = 0.08, vt = 0.04, κ = 0.2, θ = 0.08, σ = 0.1, α = 0.4, v¯ = 0.04, β = 0.1, ρ = −0.5, λr = 2, ηr = 0.005, pr = 2, T = 1.

process. However, allowing for a gamma distributed jump in the volatility process, the density functions in Figure 9.6 show rather increased values on both tails while maintaining their mode. Distinguishing between the impact of jump parameters on the density, we clearly identify the jump intensity to have the strongest inﬂuence on the short-rate process. The base diﬀusion parameters used for the computation of theoretical prices given in Tables 9.7 and 9.8 are the same as above. The default parameters for the diﬀerent gamma distributed jump components are λ = 2, η = 0.005 and p = 2 for both the short-rate and volatility process. All contracts have a remaining time to maturity of half a year. At ﬁrst, we turn our attention to Table 9.7, where zero-bond call option prices are given for

166

9 Jump-Enhanced Two-Factor Interest-Rate Models 25

25 0 2 4 6

20

probability density

probability density

20

15

10

5

0

0.005 0.0075 0.01 0.0125

15

10

5

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

0

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

T

T

25 1 2 3 4

probability density

20

15

10

5

0

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

r

T

Fig. 9.6. Probability densities for a short rate governed by the Fong-Vasicek diffusion model enhanced with a gamma distributed jump component for the volatility process. In the upper left (right) graph the density functions for varying jump intensities (means) are depicted. The lower graph shows the density behavior for alternating values of p. The base parameters are: rt = 0.08, vt = 0.04, κ = 0.2, θ = 0.08, σ = 0.1, α = 0.4, v¯ = 0.04, β = 0.1, ρ = −0.5, λv = 2, ηv = 0.005, pv = 2, T = 1.

a strike range from 60 to 90 units. The range is chosen to cover either ITM, ATM and OTM option prices202 . Similar to ﬁndings for the probability density function, the jump behavior of gamma distributed jump component on the short rate under the Fong-Vasicek model and also in the Square-Root model shows a strong resemblance on account of the Square-Root process. Looking at the gamma distributed jump component in the volatility process, we observe higher values of the particular derivative price for increasing jump parameters. This eﬀect is due to the more increased tails of the density in comparison to the ordinary diﬀusion case. Accordingly, the eﬀect of volatility 202

The value of a zero bond with remaining maturity of two years, priced with base parameters, is 82.335 units.

9.3 The Fong-Vasicek Model

167

jumps on derivative contracts is small, which explains the nearly linear relation between jump parameters and contract values. Turning our attention to Tables 9.8 and 9.9, we computed prices for caps with one payment date made at the maturity of the particular contract. Thus, we examine cap prices similar to the idealized payoﬀ function in Table 4.1. Since we restrict this model to positively sized jump components due to the Square-Root limitations, we have the ordinary cap dominating its average-rate counterpart. This is a direct consequence of the geometric average of an increasing function. Comparing the corresponding values of the one-factor CIR model with the cap values in this section, we indicate the same impact of jumps on the short rate generally. Thus, this model also has a less sensitive average-rate option in contrast to the ordinary cap contract. However, jumps on the volatility component have roughly speaking no eﬀect at all on average-rate options, whereas ordinary caps encounter small changes for either ITM, ATM and OTM option values.

168

9 Jump-Enhanced Two-Factor Interest-Rate Models

Table 9.7. Values of zero-bond call options for the jump-enhanced Fong-Vasicek model, where the underlying zero-bond contract has a nominal value of 100 units. K

60

65

70

75

80

85

90

20.232

15.439

10.647

5.867

1.562

0.052

0

ηr 0.005 0.01

16.456

11.68

6.965

2.689

0.255

0.002

0

0.015

12.970

8.316

4.050

0.921

0.021

0

0

0.02

9.868

5.579

2.088

0.181

0.001

0

0

0

24.302

19.498

14.694

9.890

5.087

0.910

0.015

2

20.232

15.439

10.647

5.867

1.562

0.052

0

4

16.378

11.597

6.828

2.393

0.169

0.001

0

6

12.728

7.967

3.409

0.469

0.006

0

0

1

22.232

17.434

12.636

7.838

3.103

0.237

0.002

2

20.232

15.439

10.647

5.867

1.562

0.052

0

3

18.299

13.512

8.730

4.067

0.660

0.010

0

4

16.430

11.651

6.910

2.596

0.238

0.002

0

0.005

20.232

15.439

10.647

5.867

1.562

0.052

0

0.01

20.247

15.455

10.663

5.885

1.597

0.064

0

0.015

20.263

15.471

10.678

5.902

1.631

0.077

0.001

0.02

20.279

15.486

10.694

5.920

1.664

0.090

0.002

0

20.216

15.424

10.631

5.850

1.525

0.041

0

2

20.232

15.439

10.647

5.867

1.562

0.052

0

4

20.247

15.455

10.663

5.884

1.598

0.063

0

6

20.263

15.471

10.678

5.901

1.634

0.075

0.001

1

20.224

15.432

10.639

5.859

1.543

0.046

0

2

20.232

15.439

10.647

5.867

1.562

0.052

0

3

20.240

15.447

10.655

5.876

1.580

0.057

0

4

20.247

15.455

10.663

5.884

1.597

0.063

0

λr

pr

ηv

λv

pv

9.3 The Fong-Vasicek Model

169

Table 9.8. Values of short-rate caps for the jump-enhanced Fong-Vasicek model, with a nominal value of 100 units. K

6

7

8

9

10

11

12

0.005

2.854

1.976

1.222

0.663

0.317

0.137

0.055

0.01

3.742

2.852

2.061

1.422

0.952

0.624

0.400

0.015

4.623

3.730

2.925

2.253

1.721

1.306

0.981

0.02

5.494

4.601

3.790

3.101

2.536

2.069

1.678

0

1.979

1.162

0.541

0.181

0.04

0.006

0.001

2

2.854

1.976

1.222

0.663

0.317

0.137

0.055

4

3.738

2.827

1.991

1.291

0.769

0.423

0.217

6

4.625

3.696

2.812

2.018

1.358

0.858

0.510

1

2.412

1.554

0.850

0.375

0.131

0.038

0.009

2

2.854

1.976

1.222

0.663

0.317

0.137

0.055

3

3.298

2.410

1.627

1.011

0.583

0.317

0.163

4

3.743

2.849

2.049

1.394

0.906

0.565

0.339

0.005

2.854

1.976

1.222

0.663

0.317

0.137

0.055

0.01

2.861

1.990

1.242

0.684

0.334

0.148

0.061

0.015

2.868

2.003

1.260

0.704

0.351

0.159

0.067

0.02

2.876

2.016

1.278

0.723

0.366

0.170

0.073

0

2.848

1.962

1.200

0.639

0.300

0.127

0.050

2

2.854

1.976

1.222

0.663

0.317

0.137

0.055

4

2.860

1.990

1.243

0.685

0.335

0.148

0.060

6

2.867

2.004

1.263

0.707

0.352

0.158

0.065

1

2.851

1.969

1.211

0.651

0.308

0.132

0.052

2

2.854

1.976

1.222

0.663

0.317

0.137

0.055

3

2.857

1.983

1.232

0.674

0.326

0.142

0.058

4

2.861

1.990

1.242

0.684

0.334

0.148

0.060

ηr

λr

pr

ηv

λv

pv

170

9 Jump-Enhanced Two-Factor Interest-Rate Models

Table 9.9. Values of average-rate caps for the jump-enhanced Fong-Vasicek model, with a nominal value of 100 units. K

6

7

8

9

10

11

12

0.005

2.377

1.443

0.655

0.205

0.051

0.012

0.003

0.01

2.827

1.892

1.077

0.543

0.268

0.131

0.063

0.015

3.271

2.336

1.512

0.936

0.586

0.366

0.226

0.02

3.708

2.775

1.946

1.346

0.948

0.668

0.468

0

1.922

1.004

0.308

0.036

0.001

0

0

2

2.377

1.443

0.655

0.205

0.051

0.012

0.003

4

2.831

1.889

1.040

0.449

0.159

0.050

0.014

6

3.283

2.337

1.450

0.751

0.329

0.127

0.044

1

2.150

1.221

0.466

0.096

0.012

0.001

0

2

2.377

1.443

0.655

0.205

0.051

0.012

0.003

3

2.603

1.668

0.859

0.352

0.127

0.043

0.013

4

2.828

1.892

1.072

0.523

0.237

0.101

0.041

0.005

2.377

1.443

0.655

0.205

0.051

0.012

0.003

0.01

2.377

1.446

0.661

0.211

0.053

0.012

0.003

0.015

2.378

1.449

0.667

0.216

0.056

0.013

0.003

0.02

2.378

1.452

0.672

0.221

0.058

0.014

0.003

0

2.377

1.441

0.648

0.200

0.049

0.011

0.002

2

2.377

1.443

0.655

0.205

0.051

0.012

0.003

4

2.377

1.446

0.661

0.211

0.053

0.012

0.003

6

2.377

1.449

0.668

0.216

0.055

0.013

0.003

1

2.377

1.442

0.651

0.202

0.05

0.012

0.002

2

2.377

1.443

0.655

0.205

0.051

0.012

0.003

3

2.377

1.445

0.658

0.208

0.052

0.012

0.003

4

2.377

1.446

0.661

0.211

0.053

0.012

0.003

ηr

λr

pr

ηv

λv

pv

10 Non-Aﬃne Term-Structure Models and Short-Rate Models with Stochastic Jump Intensity

10.1 Overview Although the model setup proposed in this thesis is of the exponential-aﬃne class, we can also extend the framework to allow for certain non-aﬃne models and models with state-dependent jump intensities λQ (xt ). Moreover, option prices under these more sophisticated model dynamics can be priced in our numerical scheme without greater eﬀort, due to an exponential separable structure of the governing characteristic function. However, working with a non-aﬃne model, we have to abandon jump components for those particular non-aﬃne factors. A stochastic jump intensity in the general exponentialaﬃne model framework is introduced in Duﬃe, Pan and Singleton (2000). Consequently, the jump transform is no longer independent of the coeﬃcient function a(z, τ ), and therefore a complicated system of ODEs has to be determined numerically anyway. Since both approaches need to establish further restrictions, they are only discussed as possibilities for extending and modifying the base model, respectively.

10.2 Quadratic Gaussian Models Non-Aﬃne exponential separable models are characterized by a non-aﬃne structure of the factors in the relevant moment-generating function, as well as the general characteristic function, while preserving the separability of coeﬃcient functions for diﬀerent powers of the particular factors included in the model. Thus, the essential system of ODEs can be derived. Prominent representatives of this model class are in an equity context the stochastic volatility

172

10 Non-Aﬃne and Stochastic Jump Intensity Term-Structure Models

model of Sch¨obel and Zhu (1999), which is a generalized version of the Stein and Stein (1991) model. In case of interest rates, we have, e.g. the Double Square-Root model of Longstaﬀ (1989), the quadratic Gaussian model approach of Beaglehole and Tenney (1991)203 , and the general linear-quadratic jump-diﬀusion model of Cheng and Scaillet (2004)204 . Although the quadratic Gaussian and the Double Square-Root model seem quite attractive to implement, it is impossible to compute theoretical model prices within the Fourier-based pricing framework if jumps are incorporated, while Monte-Carlo pricing approaches might still work. This stems from the fact that in equation (2.39), for the nth jump Jmn in the non-aﬃne fac(m) (m) tor xt , there would be a corresponding term (xt + Jmn )2 resulting in a mixed expression. Hence, the exponential separation approach will no longer be available in deriving the general characteristic function. Since none of the non-aﬃne interest-rate models are capable of exhibiting any jump component we completely ignored these models in our base setup according to Section 2.1. The one-factor quadratic Gaussian approach models the short rate under the risk-neutral measure, as the square of some factor xt governed by an Ornstein-Uhlenbeck process according to equation (8.5). In order to price interest-rate derivatives for this particular process, we need to have the general characteristic function to consider both the state variable xt and its square x2t . Thus, for the squared Gaussian interest-rate model we use the following form of the general characteristic function

ψ (yt , z, 0, w, g0 , g, τ ) = ea(z,τ )+b(z,τ ) yt +ızg0 ,

with yt =

xt x2t

and

0 w= . 1

For convenience, we use again the time-dependent coeﬃcient functions205 203

204

205

Ahn, Dittmar and Gallant (2002) give a good overview of general multidimensional linear-quadratic Gaussian interest-rate models. Linear-quadratic in this context means all factors contained in the state vector xt are allowed to enter the interest rate both in a linear and quadratic fashion. ¯ The constant parameter g0 is represented by the term A.

10.2 Quadratic Gaussian Models

173

a(z, τ ) = A(z, τ ),

and b(z, τ ) =

B(z, τ )

C(z, τ )

¯ B . + ız C¯

Inserting the above characteristic function in equation (2.33) and applying the separation approach result again in a system of coupled ODEs206 ¯ + σ 2 (C(z, τ ) + ız C) ¯ A(z, τ )τ =κθ(B(z, τ ) + ız B) ¯ 2, + 2σ 2 (B(z, τ ) + ız B) ¯ 2 (C(z, τ ) + ız C) ¯ − κ) B(z, τ )τ =(B(z, τ ) + ız B)(σ ¯ + 2κθ(C(z, τ ) + ız C), ¯ + 2σ 2 (C(z, τ ) + ız C) ¯ 2 − 1, C(z, τ )τ = − 2κ(C(z, τ ) + ız C) which can be solved successively. The advantage of this modeling approach lies in its tractability while describing a more elaborated interest-rate behavior. Additionally, the short rate in this approach is always positive, compared to possible negative short rates using the Vasicek model. In the Double SquareRoot model according to Longstaﬀ (1989), we encounter a very similar situation, since we are able to transform the model into a quadratic Gaussian model and vice versa but with additional restrictions on the parameter set207 . Cheng and Scaillet (2004) introduce a linear-quadratic jump-diﬀusion model. Here, the diﬀusion part of some random variable, for example the short rate r(xt ) or the payoﬀ characteristic function g(xt ), is built similarly to the multivariate quadratic Gaussian model in Beaglehole and Tenney (1991), as the sum of linear and quadratic terms of the state vector xt containing correlated Ornstein-Uhlenbeck processes. To gain a closed-form solution for the general characteristic function, additional jump parts only occur in the aﬃne terms of xt . Therefore, we can think of this interest-rate model as a simple combination of an additive multivariate Ornstein-Uhlenbeck model augmented with jump components and an additive multivariate quadratic Gaussian model. 206

Although the vector yt occurs in the characteristic function, derivatives remain still to be taken with respect to the unique state variables which is in this one-

207

dimensional model just the factor xt . See Beaglehole and Tenney (1992), pp. 346-347.

174

10 Non-Aﬃne and Stochastic Jump Intensity Term-Structure Models

10.3 Stochastic Jump Intensity Another possibility for extending the base model setup stated in Section 2.1 is to implement stochastic jump intensities. Duﬃe, Pan and Singleton (2000) introduced, with their aﬃne jump-diﬀusion model, a vector of stochastic jump intensities where the stochastic component is aﬃne in the state variable xt . Thus, they implement stochastic intensities without overly aggravating their solution technique. Deﬁning the vector of jump intensities as208 Q λQ (xt ) = λQ 0 + λ1 xt , Q M with (λQ × RM×M , we therefore get a slightly modiﬁed system of 0 , λ1 ) ∈ R

ODEs for the vector coeﬃcient b(z, τ ) compared to equation (2.41), which is 1 b(z, τ ) Σ1 b(z, τ ) 2 ∗ + λQ 1 EJ [ψ (z, w0 , w, g0 , g, J, τ ) − 1] .

b(z, τ )τ = − w + µQ 1 b(z, τ ) +

Obviously, in implementing this type of jump intensity, values of the coeﬃcient vector b(z, τ ) must be determined numerically due to the complicated structure of the relevant ODE. Subsequently, the same statement holds also for the coeﬃcient function a(z, τ ), which depends on b(z, τ ). Although this type of jump speciﬁcation enriches the modeling capabilities of the shortrate dynamics, it is infrequently implemented in interest-rate models because of the numerical diﬃculties mentioned above. However, our FRFT algorithm presented in Chapter 6 can be easily modiﬁed to handle this type of stochastic jump intensity.

208

To stay conform with our base model setup in equation (2.1), we suggest to include N Poisson processes with stochastic intensities.

11 Conclusion

In this thesis, we have introduced a general jump-diﬀusion short-rate model. The model approach we proposed extends the interest-rate model of Duﬃe and Kan (1996) by considering N diﬀerent possible Poison processes in the underlying factors. Using the ﬁndings in Carr and Madan (1999) and Lewis (2001) to interchange the order of integration of an integral-transformed option price in an equity context, we derived a general pricing formula valid for various popular interest-rate contracts. However, we eventually preferred the approach of Lewis (2001) over the technique presented in Carr and Madan (1999). The pricing scheme used in this thesis exhibits a rigorous modular structure. Thus, we took one step towards successfully extending the spirit of a modular pricing framework as proposed in Zhu (2000) by modularizing not only the stochastic parts, but also modularizing the derivative price in terms of its payoﬀ structure. Hence, all pricing formulae developed in this thesis can be split into parts of the Fourier-transformed payoﬀ function and of the underlying process, characterized by its characteristic function, respectively. Hence, we were able to state one single valuation formula, equation (4.21), to price derivatives of the linear, exponential-linear, and integro-linear types. Especially for the integro-linear case, the payoﬀ-transform approach oﬀers an elegant alternative to the methods proposed, e.g. in Bakshi and Madan (2000), Chacko and Das (2002), and Ju (1997). In addition, we presented within the pricing framework of Lewis (2001) for the ﬁrst time the consistent inclusion of both unconditional and conditional interest-rate derivatives. Hence, facilitating a integration by parts method, we successfully applied the residue theorem to recover prices of contracts with unconditional exercise rights and

176

11 Conclusion

the particular put-call parities, respectively. As a special case in Section 5.3.3, we also derived a Fourier-style solution of coupon-bond options, where the interest rate is governed by a one-factor process, which can be compared to the two-sided Laplace-style solution presented in Kluge (2005), Section 2.4. However, we want to emphasize that the term one-factor does not correspond to the number of jump components incorporated in the short-rate process. Thus, theoretically speaking, within our pricing scheme we are able to price coupon-bond options and swaptions, respectively, as long as the underlying process exhibits only one single Brownian motion. This is in fact a powerful result, since the additional inclusion of jump processes can result in more realistic models. In Chapter 6, we employed the IFFT algorithm for the ﬁrst time to compute option prices within the pricing framework of Lewis (2001). The obtained pricing algorithm is then reﬁned by translating the IFFT procedure into the FRFT algorithm. Subsequently, we dealt with the issue of ﬁnding the optimal and therefore error-minimizing parameter setting for the FRFT algorithm by utilizing a steepest descent technique. Doing this, we focused our eﬀorts to minimize the overall error of the solution vector generated by the FRFT pricing algorithm, rather than the error of one single option price209 . In our opinion, this procedure is a more powerful procedure, since the advantage of the FFT- and IFFT-based pricing algorithms is the simultaneous computation of option prices for a given strike range. Therefore, we used for the error measuring the RMSE of the numerical solution vector. Fortunately, it became apparent that the logarithmic RMSE of the numerical solution is a nearly linear descending function for increasing values of zi and ω, starting with the smallest possible values not violating any regularity conditions. Thus, we used a steepest-descent technique to identify the optimal parameters for the numerical algorithm. Furthermore, exploiting this linearity we were also able to formulate an approximate error bound for the numerical solution vector. After discussing the numerical algorithm, we analyzed a selection of both one-factor and two-factor jump-diﬀusion short-rate models. We ﬁrst speciﬁed our jump size candidates, which were the exponential, gamma, and normal 209

Lee (2004) and Lord and Kahl (2007) study the error behavior of Fourier transform-based algorithms for only a single strike value.

11 Conclusion

177

distribution, and derived their particular jump transforms. Subsequently, we derived the relevant general characteristic function of the jump-diﬀusion process, and then computed numerical values of the particular density functions and contract values. Widely used, the exponentially distributed jump size assumption presents no diﬃculties in derivatives pricing because of the closedform jump transforms for both Ornstein-Uhlenbeck and Square-Root diﬀusion processes. However, we also applied normal and gamma distributions for the jump size. The normal distribution for the jump component within a Vasicek model is also used in the articles of Baz and Das (1996) and Durham (2005), where an approximation technique for zero-bond prices is described. Unfortunately, under some circumstances both approaches deliver inaccurate values for the respective derivatives contracts210. Our pricing algorithm is able to circumvent these issues and compute accurate numerical values of interest-rate derivatives in any case. Moreover, we introduced gamma distributed jumps within a jump-diﬀusion short-rate model framework for the ﬁrst time. We then combined these jump candidates with the one-factor Ornstein-Uhlenbeck, the Square-Root processes, the two-factor version of a combined OU-SR, and the stochastic volatility model of Fong and Vasicek (1991a), and computed densities and option values. In particular, our contribution besides the implementation of the normal and gamma jump-size distribution in interest-rate option pricing, is to present an algorithm capable of computing option prices for the (jump-extended) Fong and Vasicek (1991a) interest-rate model. Up to now, only zero-bond prices have been computed for the jump-enhanced Vasicek and CIR model211 and the (pure diﬀusion) Fong and Vasicek (1991a) model212 , but no option prices have been presented so far. Due to the general applicability of the solution formula (4.21), we were able to compute numerical solutions for all important interest-rate derivatives. Comparing the diﬀerent results from the jump-diﬀusion term-structure models, it is obvious that jump components can enhance the stochastic dynamics. Accordingly, we were able to model probability density functions, which show bimodality and the important feature of fat tails. 210 211 212

See the concluding remarks in Durham (2006) and the comments in Section 7.3. Compare with the comments in Sections 8.2 and 8.3. See, for example, Selby and Strickland (1995).

178

11 Conclusion

Although the model setup used in this thesis is of the exponential-aﬃne type, the pricing technique can be extended to special non-aﬃne processes, namely to the family of quadratic Gaussian processes due to their exponentialaﬃne structure of the particular characteristic function. As discussed in Chapter 10, this model class cannot be enhanced with jump components since the resulting PDE would then no longer be separable. Another possible way of extending the base model speciﬁcation, which we brieﬂy discussed, is given by the inclusion of stochastic jump intensities, where the intensity is an aﬃne function of the state vector xt . However, we have then to numerically determine the coeﬃcient functions a(z, τ ) and b(z, τ ), which can be a challenging task due to the elaborated jump transforms. We presented a sophisticated alternative to time-consuming Monte-Carlo simulations, which have to be applied otherwise due to the complicated jumpdiﬀusion dynamics. Combined with the highly eﬃcient FRFT algorithm, this numerical pricing approach oﬀers an accuracy and eﬃciency, which can be hardly achieved by other methods. However, the methodology is restricted, in this form, to price only European-type derivatives. Thus, possible research can focus on developing a pricing procedure based on the algorithm in this thesis, which is also capable of valuing American-type derivatives. The early-exercise feature of these American-type derivatives might then be implemented by using some sort of time-stepping scheme of the Fourier-transformed derivative value or by using backward induction as proposed in Lord, Fang, Bervoets and Oosterlee (2007). Although we discussed one- and two-factor interestrate models, we can easily extend the pricing framework to include also jumpenhanced versions of higher factor models, such as e.g. the multi-factor models presented in Balduzzi, Das, Foresi and Sundaram (1996) and Collin-Dufresne and Goldstein (2002). Another possibility for further research might be an empirical validation of the family of gamma jump-enhanced diﬀusion models, as for example done in the studies by Lin and Yeh (1999) and Das (2002).

A Derivation of the Complex-Valued Coeﬃcients for the Characteristic Function in the Square-Root Model

Our starting point for deriving the time-dependent coeﬃcient function ˜b(z, τ ) is equation (8.10). Thus, making the standard transformation for this type of diﬀerential equation, we assume ˜b(z, τ ) = − 1 E(z, τ )τ . c2 (z) E(z, τ )

(A.1)

Consequently, substituting the particular expressions in equation (8.10), function E(z, τ ) satisﬁes the following homogeneous ODE E(z, τ )τ τ = c1 (z)E(z, τ )τ − c0 (z)c2 (z)E(z, τ ),

(A.2)

with E(z, 0)τ = 0, due to the terminal condition ˜b(z, 0) = 0. Additionally, we assume for the moment an unspeciﬁed constant E(z, 0) = E0 and guess a solution of the form E(z, τ ) = eυ(z)τ . Hence, plugging this function together with its particular derivatives into equation (A.2), we arrive at the so-called characteristic equation for this second order type ODE, which after some simpliﬁcations is υ 2 (z) − c1 (z)υ(z) + c0 (z)c2 (z) = 0. The solution of this quadratic form is given by υ± (z) =

c1 (z) ± ϑ(z) , 2

with ϑ(z) deﬁned according to Section 8.3. Since the discriminant of the square-root function ϑ(z) is

180

A Complex-Valued Coeﬃcients in the Square-Root Model

κ2 + 2σ 2 w1 > 0, the characteristic equation has two diﬀerent real-valued solutions and therefore the general solution can be represented by the linear combination E(z, τ ) = Ψ1 (z)eυ(z)+ τ + Ψ2 (z)eυ(z)− τ . Consequently, we get for τ = 0 the following terminal conditions E(z, 0) = Ψ1 (z) + Ψ2 (z), E(z, 0)τ = Ψ1 (z)υ+ (z) + Ψ2 (z)υ− (z). Keeping in mind that E(z, 0)τ ≡ 0, we use the two equations above to determine the coeﬃcient functions Ψ1 (z) and Ψ2 (z). Eventually, the solution of E(z, τ ) can be obtained as c1 (z) ϑ(z) ϑ(z) ϑ(z) ϑ(z) E0 e 2 τ e 2 τ + e− 2 τ e 2 τ − e− 2 τ E(z, τ ) = ϑ(z) − c1 (z) ϑ(z) 2 2 (A.3) c1 (z)

ϑ(z)τ E0 e 2 τ ϑ(z)τ = ϑ(z) cosh − c1 (z) sinh , ϑ(z) 2 2 and the particular derivative with respect to the time-to-maturity variable τ can be calculated as c1 (z)

E0 e 2 E(z, τ )τ = −2 ϑ(z)

τ

ϑ(z)τ c0 (z)c2 (z) sinh . 2

Finally, inserting the functions E(z, τ ), now up to a constant E0 determined, and E(z, τ )τ into equation (A.1), we end up with 2c0 (z) sinh ϑ(z)τ 2 ˜b(z, τ ) = , ϑ(z)τ − c ϑ(z) cosh ϑ(z)τ (z) sinh 1 2 2 which coincides with the solution given in equation (8.11)1 . Having obtained ˜b(z, τ ), it is a very simple task to derive the coeﬃcient function a0 (z, τ ) because of the approach taken in (A.1). Thus, using a logarithmic integration approach we immediately arrive at 1

The terminal condition ˜b(z, 0) = 0 is satisﬁed, which can be easily justiﬁed due to the relation sinh[0] = 0.

A Complex-Valued Coeﬃcients in the Square-Root Model

τ a0 (z, τ ) = −w0 τ + κθ

181

˜b(z, s) + ızg0 ds

0

κθ = (ızg0 κθ − w0 )τ − c2 (z)

τ

E(z, s)τ ds E(z, s)

(A.4)

0

κθ (ln [E(z, τ )] − ln [E(z, 0)]) . = (ızg0 κθ − w0 )τ − c2 (z) Because of the terminal condition a0 (z, 0) ≡ 0, we must set the constant E0 = 1 in equation (A.3). Eventually, after simplifying the resulting expression in equation (A.4) we are able to state the desired form given in (8.12).

B Derivation of the Complex-Valued Coeﬃcients for the Characteristic Function in the Fong-Vasicek Model

Starting with the time-dependent coeﬃcient function B(z, τ ), we adopt the solution according to equation (8.6). Thus, we exchange the parameter g1 with ¯ Subsequently, we show that the derivation of the time-dependent coeﬃcient B. A01 (z, τ ), the volatility-related part of A0 (z, τ ), states no problem on account of logarithmic integration. Thus, the next task is to recover the coeﬃcient function C(z, τ ). Therefore, plugging in the explicit solution of B(z, τ ) into ODE (9.7) results in 1 C(z, τ )τ =f1 (z) + f2 (z)X(z, τ ) + X(z, τ )2 + f3 (z)C(z, τ ) 2 β2 + βρX(z, τ )C(z, τ ) + C(z, τ )2 , 2

(B.1)

with time-independent coeﬃcients fi (z) according to Section 9.3 and w ¯ e−κτ . X(z, τ ) = σ + ız B κ Similar to the derivation of the time-dependent coeﬃcient function ˜b(z, τ ) in the SR model, we assume for C(z, τ ) a solution of the form C(z, τ ) = −

2 U (z, τ )τ . β 2 U (z, τ )

(B.2)

Inserting this alternative representation into equation (B.1) and simplifying the resulting ODE for the new function U (z, τ ) gives U (z, τ )τ τ = (f3 (z) + βρX(z, τ )) U (z, τ )τ β2 1 2 − f1 (z) + f2 (z)X(z, τ ) + X(z, τ ) U (z, τ ). 2 2

(B.3)

184

B Complex-Valued Coeﬃcients in the Fong-Vasicek Model

Subsequently, we apply another substitution V (X(z, τ )) = U (z, τ ),

(B.4)

and get a new ODE, with derivatives taken with respect to X(z, τ ). For convenience, we express the particular derivatives with V (X(z, τ ))X and V (X(z, τ ))XX , respectively. Thus, the resulting ODE has the formal structure X(z, τ )2 V (X(z, τ ))XX f3 (z) βρ X(z, τ )2 V (X(z, τ ))X + 1+ X(z, τ ) + κ κ 2 β 1 + 2 f1 (z) + f2 (z)X(z, τ ) + X(z, τ )2 V (X(z, τ )) = 0. 2κ 2

(B.5)

Finally, the solution of this particular ODE can be obtained by applying a last substitution of the form V (X(z, τ )) = L(z, τ ) W (Y (z, τ )), with L(z, τ ) and Y (z, τ ) as deﬁned in Section 9.3. Hence, inserting this substitution into equation (B.5) and simplifying the resulting ODE, we end up with − Q(z)W (Y (z, τ )) + (S(z) − Y (z, τ ))W (Y (z, τ ))Y

(B.6)

+ Y (z, τ )W (Y (z, τ ))Y Y = 0. Again, the explicit expressions of Q(z) and S(z) are given in Section 9.3. Equation (B.6) is better known as the prominent Kummer equation, which has the general solution2 W (Y (z, τ )) = Ψ1 (z)KM[Q(z), S(z), Y (z, τ )] + Ψ2 (z)KU[Q(z), S(z), Y (z, τ )]. Thus, in order to obtain the solution for the coeﬃcient C(z, τ ), we also need the ﬁrst derivative with respect to τ of the function U (z, τ ). Hence, according to the chain rule we have the relation U (z, τ )τ = −κX(z, τ )V (X(z, τ ))X . 2

See, for example, Abramowitz and Stegun (1972), p. 504. Our solution is customized to account for the parametric form due to the frequency representation.

B Complex-Valued Coeﬃcients in the Fong-Vasicek Model

185

The desired derivative of V (X(z, τ )) with respect to X(z, τ ) in the above equation can be represented as V (X(z, τ ))X = 1 − L(z, τ ) 2X(z, τ ) × −2Q(z) × Ψ1 (z)KM[Q(z) + 1, S(z), Y (z, τ )] + (1 + Q(z) − S(z))Ψ2 (z)KU[Q(z) + 1, S(z), Y (z, τ )] +

β2 M (z, τ ) κ

× Ψ1 (z)KM[Q(z), S(z), Y (z, τ )]

+ Ψ2 (z)KU[Q(z), S(z), Y (z, τ )] . Thus, according to the approach taken in equation (B.2), the coeﬃcient function C(z, τ ) can be recovered as (9.8), which is in terms of V (X(z, τ )) C(z, τ ) =

2κ V (X(z, τ ))X . X(z, τ ) β2 V (X(z, τ ))

Next, checking the validity of the terminal condition C(z, 0) = U (z, 0)τ = V (X(z, 0))X ≡ 0, we only need the explicit form of the time-independent function Υ (z), which is just the fraction Ψ1 (z) . Υ (z) = Ψ2 (z) Arranging terms for Ψ1 (z) and Ψ2 (z) in the ﬁrst derivative of V (X(z, τ )) ¯ evaluated at X(z, 0) = σ w κ + ız B , we get

186

B Complex-Valued Coeﬃcients in the Fong-Vasicek Model

2 β M (z, τ )KM[Q(z) + 1, S(z), Y (z, 0)] Ψ1 (z) κ − 2Q(z)KM[Q(z) + 1, S(z), Y (z, 0)] = Ψ2 (z) 2Q(z)(1 + Q(z) − S(z))KU[Q(z), S(z), Y (z, 0)] β2 − M (z, τ )KU[Q(z), S(z), Y (z, 0)] . κ Obviously, solving for the particular fraction, the speciﬁc form of Υ (z) can be validated by checking its deﬁnition given in Section 9.3. Thus, the coeﬃcient function C(z, τ ) with speciﬁed time-independent function Υ (z) coincides with the result given in equation (9.8). For the calculation of A02 (z, τ ), we exploit the functional form chosen in the derivation of the coeﬃcient function C(z, τ ). Thus, we apply a logarithmic integration approach and recover the antiderivative of A02 (z, τ )τ as A02 (z, τ ) = −

2α¯ v ¯ ln[U (z, τ )] + ızα¯ vCτ. 2 β

In order to guarantee the terminal condition of A02 (z, 0) = 0, at the maturity of the contract, we have to ensure that U (z, 0) = 1. Thus, rewriting U (z, τ ) as U (z, τ ) =L(z, τ )Ψ2 (z) × (Υ (z)KM[Q(z), S(z), Y (z, τ )] + KM[Q(z), S(z), Y (z, τ )]) , we immediately arrive at 1 = L(z, 0) (Υ (z)KM[Q(z), S(z), Y (z, 0)] + KM[Q(z), S(z), Y (z, 0)]) . Ψ2 (z) Therefore, the time-dependent function A02 (z, τ ) can be written in terms of J(z, τ ) =

2Q(z)κ β 2 (KU[Q(z); S(z); Y (z, τ )] + Υ(z) KM[Q(z); S(z); Y (z, τ )])

and L(z, τ ), given in (9.11), which concludes the derivation of the coeﬃcient functions in Section 9.3.

References

Abramowitz, M. and Stegun, I.: 1972, Handbook of Mathematical Functions with Formulas, Graphs and Mathematical Tables, Dover Publications. Ahn, C. and Thompson, H.: 1988, Jump-Diﬀusion Processes and the Term Structure of Interest Rates, Journal of Finance 43(1), 155–174. Ahn, D., Dittmar, R. and Gallant, A.: 2002, Quadratic Term Structure Models: Theory and Evidence, The Review of Financial Studies 15(1), 243–288. Ahn, D. and Gao, B.: 1999, A Parametric Nonlinear Model of Term Structure Dynamics, The Review of Financial Studies 12(4), 721–762. A¨ıt-Sahalia, Y.: 1996, Testing Continuous-Time Models of the Spot Interest, The Review of Financial Studies 9(2), 385–426. A¨ıt-Sahalia, Y.: 1999, Transition Densities for Interest Rate and Other Nonlinear Diﬀusions, Journal of Finance 54(4), 1361–1395. Arapis, M. and Gao, J.: 2006, Empirical Comparisons in Short-Term Interest Rate Models Using Nonparametric Methods, Journal of Financial Econometrics 4(2), 310–345. Bailey, D. and Swarztrauber, P.: 1994, A Fast Method for the Numerical Evaluation of Continuous Fourier and Laplace Transforms, Journal on Scientiﬁc Computing 15(5), 1105–1110. Bakshi, G. and Madan, D.: 2000, Spanning and Derivative-Security Valuation, Journal of Financial Economics 55(2), 205–238. Balakrishnan, N., Johnson, N. and Kotz, S.: 1994, Continuous Univariate Distributions, Vol. 1, 2nd edn, John Wiley and Sons. Balduzzi, P., Das, S., Foresi, S. and Sundaram, R.: 1996, A Simple Approach to Three-Factor Aﬃne Term Structure Models, Journal of Fixed Income

188

References

5(4), 43–53. Bates, D.: 1996, Jumps and Stochastic Volatility: Exchange Rate Processes Implicit in Deutsche Mark Options, The Review of Financial Studies 9(1), 69–107. Baz, J. and Das, S.: 1996, Analytical Approximation of the Term Structure for Jump-Diﬀusion Processes: A Numerical Analysis, Journal of Fixed Income 6(2), 78–86. Beaglehole, D. and Tenney, M.: 1991, General Solutions of some Interest RateContingent Claim Pricing Equations, Journal of Fixed Income 1(1), 69–83. Beaglehole, D. and Tenney, M.: 1992, Corrections and Additions to ’A Nonlinear Equilibrium Model of the Term Structure of Interest Rates’, Journal of Financial Economics 32(3), 345–353. Black, F. and Scholes, M.: 1973, The Pricing of Options and Corporate Liabilities, Journal of Political Economy 81(3), 637–659. Brennan, M. and Schwartz, E.: 1979, A Continuous Time Approach to the Pricing of Bonds, Journal of Banking and Finance 3(2), 133–155. Brennan, M. and Schwartz, E.: 1980, Conditional Predictions of Bond Prices and Returns, The Journal of Finance 35(2), 405–419. Brennan, M. and Schwartz, E.: 1982, An Equilibrium Model of Bond Pricing and a Test of Market Eﬃciency, Journal of Financial and Quantitative Analysis 17(3), 301–329. Brigo, D. and Mercurio, F.: 2001, Interest Rate Models, Springer. Briys, E., Crouhy, M. and Sch¨obel, R.: 1991, The Pricing of Default-Free Interest Rate Cap, Floor, and Collar Agreements, Journal of Finance 46(5), 1879–1892. Brown, S. and Dybvig, P.: 1986, The empirical Implications of the Cox, Ingersoll,Ross Theory of the Term Structure of Interest Rates, Journal of Finance 41(3), 617–630. Carr, P. and Madan, D.: 1999, Option Valuation Using the Fast Fourier Transform, Journal of Computational Finance 2(4), 61–73. Carr, P. and Wu, L.: 2004, Time-Changed Levy Processes and Option Pricing, Journal of Financial Economics 71(1), 113–141. Chacko, G. and Das, S.: 2002, Pricing Interest Rate Derivatives: A General Approach, The Review of Financial Studies 15(1), 195–241. Chan, K.: 2005, Modelling Conditional Heteroskedasticity and Jumps in Australian Short-Term Interest Rates, Accounting and Finance 45(4), 537–551.

References

189

Chen, R. and Scott, L.: 1992, Pricing Interest Rate Options in a Two-Factor Cox-Ingersoll-Ross Model of the Term Structure, The Review of Financial Studies 5(4), 613–636. Chen, R. and Scott, L.: 1995, Interest Rate Options in Multifactor CoxIngersoll-Ross Models of the Term Structure, Journal of Derivatives 3, 53– 72. Cheng, P. and Scaillet, O.: 2004, Linear-Quadratic Jump-Diﬀusion Modeling with Appliccation to Stochastic Volatility. FAME Research Paper. Chourdakis, K.: 2005, Option Pricing Using the Fractional FFT, Journal of Computational Finance 8(2), 7–24. Collin-Dufresne, P. and Goldstein, R.: 2002, Pricing Swaptions Within an Aﬃne Framework, Journal of Derivatives 10(1), 1–18. Cont, R. and Tankov, P.: 2004, Financial Modelling with Jump Processes, Chapman and Hall. Cooley, J. and Tukey, J.: 1965, An Algorithm for the Machine Computation of the Complex Fourier Series, Mathematics of Computation v19n91, 297– 301. Cox, J., Ingersoll, J. and Ross, S.: 1985a, An Intertemporal General Equilibrium model of Asset Prices, Econometrica 53(2), 363–384. Cox, J., Ingersoll, J. and Ross, S.: 1985b, A Theory of the Term Structure of Interest Rates, Econometrica 53(2), 385–408. Culot, M.: 2003, An Integrated Aﬃne Jump Diﬀusion Framework to Manage Power Portfolios in a Deregulated Market, PhD thesis, Universit´e Catholique de Louvain. Dai, Q. and Singleton, K.: 2003, Term Structure Dynamics in Theory and Reality, The Review of Financial Studies 16(3), 631–678. Das, S.: 2002, The Surprise Element: Jumps in Interest Rates, Journal of Econometrics 106(1), 27–65. Das, S. and Foresi, S.: 1996, Exact Solutions for Bond and Option Prices with Systematic Jump Risk, Review of Derivatives Research 1(1), 7–24. Delbaen, F. and Schachermayer, W.: 1995, The Existence of Absolutely Continuous Local Martingale Measures, Annals of Applied Probability 5(4), 926–945. Doetsch, G.: 1967, Anleitung zum praktischen Gebrauch der LaplaceTransformation und der Z-Transformation, 3rd edn, Oldenbourg.

190

References

Duﬃe, D.: 2001, Dynamic Asset Pricing Theory, 3rd edn, Princeton University Press. Duﬃe, D. and Kan, R.: 1996, A Yield-Factor Model of Interest Rates, Mathematical Finance 6(4), 379–406. Duﬃe, D., Pan, J. and Singleton, K.: 2000, Transform Analysis and Option Pricing for Aﬃne Jump-Diﬀusions, Econometrica 68(6), 1343–1376. Duﬀy, D.: 2004, Transform Methods for Solving Partial Diﬀerential Equations, 2nd edn, Springer. Durham, B.: 2005, Jump-Diﬀusion Processes and Aﬃne Term Structure Models: Additional Closed-Form Approximate Solutions, Distributional Assumptions for Jumps, and Parameter Estimates. FEDS Working Paper. Durham, B.: 2006, Additional Analytical Approximations of the Term Structure and Distributional Assumptions for Jump-Diﬀusion Processes, Journal of Fixed Income 15(1), 61–73. Feller, W.: 1951, Two Singular Diﬀusion Problems, Annals of Mathematics 54(1), 173–182. Filipovic, D.: 2001, A General Characterization of one Factor Aﬃne Term Structure Models, Finance and Stochastics 5(3), 389–412. Fong, H. and Vasicek, O.: 1991a, Interest-Rate Volatility as a Stochastic Factor. Giﬀord Fong Associates. Fong, H. and Vasicek, O.: 1991b, Fixed-Income Volatility Management, Journal of Portfolio Management 17(4), 41–46. Geman, H., Karoui, N. and Rochet, J.: 1995, Changes of Num´eraire, Changes of Probability Measure and Option Pricing, Journal of Applied Probability 32(2), 443–458. Gil-Pelaez, J.: 1951, A Note on the Inversion Theorem, Biometrika 38(4), 481–482. Hamilton, J.: 1996, The Daily Market for Federal Funds, Journal of Political Economy 104(1), 26–56. Harrison, J. and Kreps, D.: 1979, Martingales and Arbitrage in Multiperiod Security Markets, Journal of Economic Theory 20(3), 381–408. Harrison, J. and Pliska, S.: 1981, Martingales and Stochastic Integrals in the Theory of Continuous Trading, Stochastic Processes and their Applications 11(3), 215–260.

References

191

Heath, D., Jarrow, R. and Morton, A.: 1992, Bond Pricing and the Term Structure of Interest Rates: A New Methodology for Contingent Claim Valuation, Econometrica 60(1), 77–105. Heston, S.: 1993, A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options, The Review of Financial Studies 6(2), 327–343. Heston, S.: 1995, A Model of Discontinuous Interest Rate Behavior, Yield Curves and Volatility. Working Paper, University of Maryland. Itkin, A.: 2005, Pricing Options with VG model Using FFT. arXiv:physics/0503137.

eprint

Jamshidian, F.: 1989, An Exact Bond Option Formula, Journal of Finance 44(1), 205–209. Ju, N.: 1997, Fourier Transformation, Martingale, and the Pricing of AverageRate Derivatives. Working Paper, University of Maryland. Karatzas, I. and Shreve, E.: 1991, Brownian Motion and Stochastic Calculus, 2nd edn, Springer. Karlin, S. and Taylor, H.: 1981, A Second Course in Stochastic Processes, Academic Press. Kellerhals, B.: 2001, Financial Pricing Models in Continous Time and Kalman Filtering, Springer. Kispert, W.: 2005, Financial Contracts on Electricity in the Nordic Power Market, PhD thesis, University of T¨ ubingen. Kluge, W.: 2005, Time-Inhomogeneous L´evy Processes in Interest Rate and Credit Risk Models), PhD thesis, University of Freiburg. Kou, S.: 2002, A Jump-Diﬀusion Model for Option Pricing, Management Science 48(8), 1086–1101. Kou, S. and Wang, H.: 2004, Option Pricing Under a Double Exponential Jump Diﬀusion Model, Management Science 50(9), 1178–1192. Kushner, H.: 1967, Stochastic Stability and Control, Academic Press. Langetieg, T.: 1980, A Multivariate Model of the Term Structure, The Journal of Finance 35(1), 71–97. Lee, R.: 2004, Option Pricing by Transform Methods: Extensions, Uniﬁcation, and Error Control, Journal of Computational Finance 7(3), 51–86. Lewis, A.: 2000, Option Valuation under Stochastic Volatility, Finance Press.

192

References

Lewis, A.: 2001, A Simple Option Formula for General Jump-Diﬀusion and other Exponential Levy Processes. Manuscript, Envision Financial Systems and OptionCity.Net. Lin, B. and Yeh, S.: 1999, Jump-Diﬀusion Interest Rate Process: An Empirical Examination, Journal of Business Finance and Accounting 26(7+8), 967– 995. Longstaﬀ, F.: 1989, A Nonlinear General Equilibrium Model of the Term Structure of Interest Rates, Journal of Financial Economics 23(2), 195– 224. Longstaﬀ, F. and Schwartz, E.: 1992, Interest Rate Volatility and the Term Structure: A Two-Factor General Equilibrium Model, Journal of Finance 47(4), 1259–1282. Lord, R., Fang, F., Bervoets, F. and Oosterlee, C.: 2007, A Fast and Accurate FFT-Based Method for Pricing Early-Exercise Options under L´evy Processes. MPRA Paper. Lord, R. and Kahl, C.: 2007, Optimal Fourier Inversion in Semi-Analytical Option Pricing, Journal of Computational Finance 10(4), 1–30. Merton, R.: 1976, Option Pricing When Underlying Stock Returns are Discontinuous, Journal of Financial Economics 3(2), 125–144. Musiela, M. and Rutkowski, M.: 2005, Martingale Methods in Financial Modelling, 2nd edn, Springer. Nagel, H.: 2001, Optionsbewertung bei stochastischer Volatilit¨ at, Gabler. Øksendal, B.: 2003, Stochastic Diﬀerential Equations, 6th edn, Springer. Ornstein, L. and Uhlenbeck, G.: 1930, On the Theory of Brownian Motion, Phys. Rev. 36, 823–841. Pennacchi, G.: 1991, Identifying the Dynamics of Real Interest Rates and Inﬂation: Evidence Using Survey Data, The Review of Financial Studies 4(1), 53–86. Piazzesi, M.: 2003, Aﬃne Term Structure Models. unpublished manuscript. Raible, S.: 2000, L´evy Processes in Finance: Theory, Numerics, and Empirical Facts, PhD thesis, University of Freiburg. Sch¨ obel, R.: 1987, Zur Theorie der Rentenoption, Duncker & Humblot. Sch¨ obel, R.: 1995, Kapitalmarkt und zeitkontinuierliche Bewertung, PhysicaVerlag. Sch¨ obel, R. and Zhu, J.: 1999, Stochastic Volatility with an OrnsteinUhlenbeck Process: An Extension, European Finance Review 3(1), 23–46.

References

193

Sch¨ obel, R. and Zhu, J.: 2000, A Fourier Analysis Approach to the Valuation of Interest Rate Derivatives. Working Paper, University of T¨ ubingen. Selby, M. and Strickland, C.: 1995, Computing the Fong and Vasicek Pure Discount Bond Price Formula, Journal of Fixed Income 5(3), 78–84. Singleton, K. and Umantsev, L.: 2002, Pricing Coupon-Bond Options and Swaptions in Aﬃne Term Structure Models, Mathematical Finance 12(4), 427–446. Stein, E. and Stein, J.: 1991, Stock Price Distributions with Stochastic Volatiliy: An Analytic Approach, The Review of Pinancial Studies 4(4), 727–752. Stuart, A. and Ord, J.: 1994, Kendall’s Advanced Theory of Statistics Vol. I, 6th edn, A. Arnold. Tahani, N.: 2004, Valuing Credit Derivatives Using Gaussian Quadrature: A Stochastic Volatility Framework, Journal of Futures Markets 24(1), 3–35. Vasicek, O.: 1977, An Equilibrium Characterization of the Term Structure, Journal of Financial Economics 5(2), 177–188. Wilkens, C.: 2005, Are Jumps Inherent in European Interest Rate Data? Estimation using the Characteristic Function and Transform Analysis for Aﬃne Diﬀusions. Working Paper, Stockholm University. Zhou, H.: 2001, Jump-Diﬀusion Term Structure and Itˆ o Conditional Moment Generator. FEDS Working Paper. Zhu, J.: 2000, Modular Pricing of Options, Springer.

Lecture Notes in Economics and Mathematical Systems For information about Vols. 1–522 please contact your bookseller or Springer-Verlag Vol. 523: W.-B. Zhang, A Theory of Interregional Dynamics. XI, 231 pages. 2003.

Vol. 545: R. Hafner, Stochastic Implied Volatility. XI, 229 pages. 2004.

Vol. 524: M. Frölich, Programme Evaluation and Treatment Choise. VIII, 191 pages. 2003.

Vol. 546: D. Quadt, Lot-Sizing and Scheduling for Flexible Flow Lines. XVIII, 227 pages. 2004.

Vol. 525: S. Spinler, Capacity Reservation for CapitalIntensive Technologies. XVI, 139 pages. 2003.

Vol. 547: M. Wildi, Signal Extraction. XI, 279 pages. 2005.

Vol. 526: C. F. Daganzo, A Theory of Supply Chains. VIII, 123 pages. 2003. Vol. 527: C. E. Metz, Information Dissemination in Currency Crises. XI, 231 pages. 2003.

Vol. 548: D. Kuhn, Generalized Bounds for Convex Multistage Stochastic Programs. XI, 190 pages. 2005. Vol. 549: G. N. Krieg, Kanban-Controlled Manufacturing Systems. IX, 236 pages. 2005.

Vol. 528: R. Stolletz, Performance Analysis and Optimization of Inbound Call Centers. X, 219 pages. 2003.

Vol. 550: T. Lux, S. Reitz, E. Samanidou, Nonlinear Dynamics and Heterogeneous Interacting Agents. XIII, 327 pages. 2005.

Vol. 529: W. Krabs, S. W. Pickl, Analysis, Controllability and Optimization of Time-Discrete Systems and Dynamical Games. XII, 187 pages. 2003.

Vol. 551: J. Leskow, M. Puchet Anyul, L. F. Punzo, New Tools of Economic Dynamics. XIX, 392 pages. 2005.

Vol. 530: R. Wapler, Unemployment, Market Structure and Growth. XXVII, 207 pages. 2003.

Vol. 552: C. Suerie, Time Continuity in Discrete Time Models. XVIII, 229 pages. 2005.

Vol. 531: M. Gallegati, A. Kirman, M. Marsili (Eds.), The Complex Dynamics of Economic Interaction. XV, 402 pages, 2004.

Vol. 553: B. Mönch, Strategic Trading in Illiquid Markets. XIII, 116 pages. 2005.

Vol. 532: K. Marti, Y. Ermoliev, G. Pflug (Eds.), Dynamic Stochastic Optimization. VIII, 336 pages. 2004. Vol. 533: G. Dudek, Collaborative Planning in Supply Chains. X, 234 pages. 2004. Vol. 534: M. Runkel, Environmental and Resource Policy for Consumer Durables. X, 197 pages. 2004. Vol. 535: X. Gandibleux, M. Sevaux, K. Sörensen, V. T’kindt (Eds.), Metaheuristics for Multiobjective Optimisation. IX, 249 pages. 2004. Vol. 536: R. Brüggemann, Model Reduction Methods for Vector Autoregressive Processes. X, 218 pages. 2004. Vol. 537: A. Esser, Pricing in (In)Complete Markets. XI, 122 pages, 2004. Vol. 538: S. Kokot, The Econometrics of Sequential Trade Models. XI, 193 pages. 2004. Vol. 539: N. Hautsch, Modelling Irregularly Spaced Financial Data. XII, 291 pages. 2004. Vol. 540: H. Kraft, Optimal Portfolios with Stochastic Interest Rates and Defaultable Assets. X, 173 pages. 2004. Vol. 541: G.-y. Chen, X. Huang, X. Yang, Vector Optimization. X, 306 pages. 2005. Vol. 542: J. Lingens, Union Wage Bargaining and Economic Growth. XIII, 199 pages. 2004.

Vol. 554: R. Foellmi, Consumption Structure and Macroeconomics. IX, 152 pages. 2005. Vol. 555: J. Wenzelburger, Learning in Economic Systems with Expectations Feedback (planned) 2005. Vol. 556: R. Branzei, D. Dimitrov, S. Tijs, Models in Cooperative Game Theory. VIII, 135 pages. 2005. Vol. 557: S. Barbaro, Equity and Efficiency Considerations of Public Higer Education. XII, 128 pages. 2005. Vol. 558: M. Faliva, M. G. Zoia, Topics in Dynamic Model Analysis. X, 144 pages. 2005. Vol. 559: M. Schulmerich, Real Options Valuation. XVI, 357 pages. 2005. Vol. 560: A. von Schemde, Index and Stability in Bimatrix Games. X, 151 pages. 2005. Vol. 561: H. Bobzin, Principles of Network Economics. XX, 390 pages. 2006. Vol. 562: T. Langenberg, Standardization Expectations. IX, 132 pages. 2006.

and

Vol. 563: A. Seeger (Ed.), Recent Advances in Optimization. XI, 455 pages. 2006. Vol. 564: P. Mathieu, B. Beaufils, O. Brandouy (Eds.), Artificial Economics. XIII, 237 pages. 2005.

Vol. 543: C. Benkert, Default Risk in Bond and Credit Derivatives Markets. IX, 135 pages. 2004.

Vol. 565: W. Lemke, Term Structure Modeling and Estimation in a State Space Framework. IX, 224 pages. 2006.

Vol. 544: B. Fleischmann, A. Klose, Distribution Logistics. X, 284 pages. 2004.

Vol. 566: M. Genser, A Structural Framework for the Pricing of Corporate Securities. XIX, 176 pages. 2006.

Vol. 567: A. Namatame, T. Kaizouji, Y. Aruga (Eds.), The Complex Networks of Economic Interactions. XI, 343 pages. 2006.

Vol. 590: S. H. Oda, Developments on Experimental Economics: New Approaches to Solving Real-World Problems. XVI, 262 pages, 2007.

Vol. 568: M. Caliendo, Microeconometric Evaluation of Labour Market Policies. XVII, 258 pages. 2006.

Vol. 591: M. Lehmann-Waffenschmidt, Economic Evolution and Equilibrium: Bridging the Gap. VIII, 272 pages, 2007.

Vol. 569: L. Neubecker, Strategic Competition in Oligopolies with Fluctuating Demand. IX, 233 pages. 2006. Vol. 570: J. Woo, The Political Economy of Fiscal Policy. X, 169 pages. 2006. Vol. 571: T. Herwig, Market-Conform Valuation of Options. VIII, 104 pages. 2006.

Vol. 592: A. C.-L. Chian, Complex Systems Approach to Economic Dynamics. X, 95 pages, 2007. Vol. 593: J. Rubart, The Employment Effects of Technological Change: Heterogenous Labor, Wage Inequality and Unemployment. XII, 209 pages, 2007

Vol. 573: J. Emami Namini, International Trade and Multinational Activity. X, 159 pages, 2006.

Vol. 594: R. Hübner, Strategic Supply Chain Management in Process Industries: An Application to Specialty Chemicals Production Network Design. XII, 243 pages, 2007

Vol. 574: R. Kleber, Dynamic Inventory Management in Reverse Logistics. XII, 181 pages, 2006.

Vol. 595: H. Gimpel, Preferences in Negotiations: The Attachment Effect. XIV, 268 pages, 2007

Vol. 575: R. Hellermann, Capacity Options for Revenue Management. XV, 199 pages, 2006.

Vol. 596: M. Müller-Bungart, Revenue Management with Flexible Products: Models and Methods for the Broadcasting Industry. XXI, 297 pages, 2007

Vol. 572: M. F. Jäkel, Pensionomics. XII, 316 pages. 2006

Vol. 576: J. Zajac, Economics Dynamics, Information and Equilibnum. X, 284 pages, 2006. Vol. 577: K. Rudolph, Bargaining Power Effects in Financial Contracting. XVIII, 330 pages, 2006.

Vol. 597: C. Barz, Risk-Averse Capacity Control in Revenue Management. XIV, 163 pages, 2007

Vol. 578: J. Kühn, Optimal Risk-Return Trade-Offs of Commercial Banks. IX, 149 pages, 2006.

Vol. 598: A. Ule, Partner Choice and Cooperation in Networks: Theory and Experimental Evidence. Approx. 200 pages, 2007

Vol. 579: D. Sondermann, Introduction to Stochastic Calculus for Finance. X, 136 pages, 2006.

Vol. 599: A. Consiglio, Artificial Markets Modeling: Methods and Applications. XV, 277 pages, 2007

Vol. 580: S. Seifert, Posted Price Offers in Internet Auction Markets. IX, 186 pages, 2006.

Vol. 600: M. Hickman, P. Mirchandani, S. Voss (Eds.): Computer-Aided Scheduling of Public Transport. Approx. 424 pages, 2007

Vol. 581: K. Marti; Y. Ermoliev; M. Makowsk; G. Pflug (Eds.), Coping with Uncertainty. XIII, 330 pages, 2006. Vol. 582: J. Andritzky, Sovereign Default Risks Valuation: Implications of Debt Crises and Bond Restructurings. VIII, 251 pages, 2006. Vol. 583: I.V. Konnov, D.T. Luc, A.M. Rubinov† (Eds.), Generalized Convexity and Related Topics. IX, 469 pages, 2006. Vol. 584: C. Bruun, Adances in Artificial Economics: The Economy as a Complex Dynamic System. XVI, 296 pages, 2006. Vol. 585: R. Pope, J. Leitner, U. Leopold-Wildburger, The Knowledge Ahead Approach to Risk. XVI, 218 pages, 2007 (planned). Vol. 586: B.Lebreton, Strategic Closed-Loop Supply Chain Management. X, 150 pages, 2007 (planned). Vol. 587: P. N. Baecker, Real Options and Intellectual Property: Capital Budgeting Under Imperfect Patent Protection. X, 276 pages , 2007. Vol. 588: D. Grundel, R. Murphey, P. Panos , O. Prokopyev (Eds.), Cooperative Systems: Control and Optimization. IX, 401 pages , 2007. Vol. 589: M. Schwind, Dynamic Pricing and Automated Resource Allocation for Information Services: Reinforcement Learning and Combinatorial Auctions. XII, 293 pages , 2007.

Vol. 601: D. Radulescu, CGE Models and Capital Income Tax Reforms: The Case of a Dual Income Tax for Germany. XVI, 168 pages, 2007 Vol. 602: N. Ehrentreich, Agent-Based Modeling: The Santa Fe Institute Artificial Stock Market Model Revisited. XVI, 225 pages, 2007 Vol. 603: D. Briskorn, Sports Leagues Scheduling: Models, Combinatorial Properties, and Optimization Algorithms. XII, 164 pages, 2008 Vol. 604: D. Brown, F. Kubler, Computational Aspects of General Equilibrium Theory: Refutable Theories of Value. XII, 202 pages, 2008 Vol. 605: M. Puhle, Bond Portfolio Optimization. XIV, 137 pages, 2008 Vol. 606: S. von Widekind, Evolution of Non-Expected Utility Preferences. X, 130 pages, 2008 Vol. 607: M. Bouziane, Pricing Interest Rate Derivatives: A Fourier-Transform Based Approach. XII, 191 pages, 2008 Vol. 608: P. Nicola, Experimenting with Dynamic Macromodels: Growth and Cycles. XIII, 241 pages, 2008 Vol. 609: X. Fang, K.K. Lai, S. Wang, Fuzzy Portfolio Optimization: Theory and Models. IX, 173 pages, 2008 Vol. 610: M. Hillebrand, Pension Systems, Demographic Change, and the Stock Market. X, 176 pages, 2008

Our partners will collect data and use cookies for ad personalization and measurement. Learn how we and our ad partner Google, collect and use data. Agree & close