Florida State University Libraries
Electronic Theses, Treatises and Dissertations The Graduate School
2006 Option Pricing with Selfsimilar Additive Processes Mack L. (Mack Laws) Galloway
Follow this and additional works at the FSU Digital Library. For more information, please contact [email protected] THE FLORIDA STATE UNIVERSITY
COLLEGE OF ARTS AND SCIENCES
OPTION PRICING WITH SELFSIMILAR ADDITIVE PROCESSES
By
MACK L. GALLOWAY
A Dissertation submitted to the Department of Mathematics in partial fulfillment of the requirements for the degree of Doctor of Philosophy
Degree Awarded: Fall Semester, 2006 The members of the Committee approve the Dissertation of Mack L. Galloway defended on August 7, 2006.
Craig Nolder Professor Directing Dissertation
Fred Huffer Outside Committee Member
Paul Beaumont Committee Member
Bettye Anne Case Committee Member
John R. Quine Committee Member
The Office of Graduate Studies has verified and approved the above named committee members.
ii ACKNOWLEDGEMENTS
I thank the following friends and faculty for their help and support: Craig Nolder, Warren Beaves, Bettye Anne Case, Jennifer Mann, Fred Huffer, Paul Beaumont, Alec Kercheval, David Kopriva, John Quine, Max Gunzberger, James Doran, and Giles Levy. I also thank my parents, Harriett and Louie Galloway, III for their love and support throughout my life.
iii TABLE OF CONTENTS
List of Tables ...... vi
List of Figures ...... vii
Abstract ...... ix
1. Introduction ...... 1
2. Review of Stochastic Processes ...... 5 2.1 Additive Processes ...... 5 2.2 L´evy Stochastic Volatility Models ...... 9
3. Selfsimilar Additive Processes ...... 11 3.1 Selfdecomposability and Selfsimilarity ...... 11 3.2 Strictly Stable L´evy Processes ...... 12 3.3 Sato Processes ...... 13
4. Subordinated, Additive Processes ...... 17 4.1 Supporting Theorems and Definition ...... 17 4.2 The Generalized Subordination Theorem ...... 20
5. Construction of Additive Models ...... 23 5.1 Derivation of Characteristic Functions ...... 23 5.2 Expressions for the Central Moments ...... 28 5.3 Properties of the Increments: ...... 30
6. Option Pricing and Calibration ...... 32 6.1 Definition of the Price Process ...... 32 6.2 Numerical Pricing Method for European Options ...... 35 6.3 Calibration ...... 37
7. Empirical Study ...... 39 7.1 Models of Interest ...... 39 7.2 Data Specifications ...... 39 7.3 Algorithmic Specifications ...... 40 7.4 Results ...... 41
iv 8. Conclusion ...... 51 8.1 Future Research ...... 51
9. Plots ...... 53
A. Truncation and Sampling Error: Lee’s Theorems ...... 67 A.1 Bounds on the Discounted Characteristic Function ...... 68
B. Plots of Difference between Calculated Pricing Error and Upper Bound on Theoretical Pricing Error ...... 73
C. Additive Models: Discounted Characteristic Function Bounds for the European Call Option ...... 78 C.1 Variance Gamma Family ...... 78 C.2 Normal Inverse Gaussian Family ...... 80
D. VG-OU-Γ Discounted Characteristic Function Bounds for the European Call Option ...... 83
E. NIG-OU-Γ Discounted Characteristic Function Bounds for the European Call Option ...... 91
REFERENCES ...... 101
BIOGRAPHICAL SKETCH ...... 103
v LIST OF TABLES
7.1 Classification of Calibrated Models ...... 39
7.2 Upper Bound on the Time-averaged Theoretical Average Pricing Error (Jan. - Dec. 2005) ...... 45
7.3 Upper Bound on the Time-averaged Theoretical Root Mean Square Pricing Error (Jan. - Dec. 2005) ...... 45
7.4 Intra-family Comparison: Twelve Quote Date Time-averaged Ratio of the Upper Bounds on the Theoretical Average Pricing Errors (Jan. - Dec. 2005) 46
7.5 Intra-family Comparison: Twelve Quote Date Time-averaged Ratio of the Upper Bounds on the Theoretical Root Mean Square Pricing Errors (Jan. - Dec. 2005) ...... 46
7.6 Intra-class Comparison: Twelve Quote Date Time-averaged Ratio of the Upper Bounds on the Theoretical Average Pricing Errors (Jan. - Dec. 2005) 47
7.7 Intra-class Comparison: Twelve Quote Date Time-averaged Ratio of the Upper Bounds on the Theoretical Root Mean Square Pricing Errors (Jan. - Dec. 2005) ...... 47 7.8 Variance Gamma Models: Twelve Quote Date Time-average of the Risk- Neutral Parameters (Jan. - Dec. 2005): “( )” denotes the sample standard deviation of the mean...... 49
7.9 Normal Inverse Gaussian Models: Twelve Quote Date Time-average of the Risk-Neutral Parameters (Jan. - Dec. 2005): “( )” denotes the sample standard deviation of the mean...... 49
7.10 Normal H-Inverse Gaussian Model: Three Quote Date Time-average of the Risk-Neutral Parameters (Sep. - Nov. 2005): “( )” denotes the sample standard deviation of the mean...... 50
vi LIST OF FIGURES
9.1 Upper Bounds on Theoretical Average Pricing Error [in-sample, out-of-the- money options] ...... 54
9.2 Upper Bounds on Theoretical Average Pricing Error [out-of-sample, out-of- the-money options] ...... 55
9.3 Upper Bounds on Theoretical Root Mean Square Error [in-sample, out-of-the- money options] ...... 56
9.4 Upper Bounds on Theoretical Root Mean Square Error [out-of-sample, out- of-the-money options] ...... 57
9.5 Market and Model Option Prices vs. Moneyness (K/S) for Variance Gamma family : April 2005 quote date, out-of-the-money options ...... 58
9.6 Market and Model Option Prices vs. Moneyness (K/S) for Variance Gamma family : August 2005 quote date, out-of-the-money options ...... 59
9.7 Market and Model Option Prices vs. Moneyness (K/S) for Variance Gamma family : December 2005 quote date, out-of-the-money options ...... 60
9.8 Market and Model Option Prices vs. Moneyness (K/S) for Normal Inverse Gaussian family : April 2005 quote date, out-of-the-money options ...... 61
9.9 Market and Model Option Prices vs. Moneyness (K/S) for Normal Inverse Gaussian family : August 2005 quote date, out-of-the-money options .... 62
9.10 Market and Model Option Prices vs. Moneyness (K/S) for Normal Inverse Gaussian family : December 2005 quote date, out-of-the-money options ... 63
9.11 Risk-neutral Mean, Variance, Skewness, Kurtosis for April 2005 quote date [VG models - top row][NIG models - bottom row] ...... 64
9.12 Risk-neutral Mean, Variance, Skewness, Kurtosis for August 2005 quote date [VG models - top row][NIG models - bottom row] ...... 65
9.13 Risk-neutral Mean, Variance, Skewness, Kurtosis for December 2005 quote date [VG models - top row][NIG models - bottom row] ...... 66
vii B.1 In-sample Average Pricing Error Differences: (Upper Bound of Theoretical Value - Calculated Value using out-of-the-money options) ...... 74
B.2 Out-of-sample Average Pricing Error Differences: (Upper Bound of Theoret- ical Value - Calculated Value using out-of-the-money options) ...... 75
B.3 In-sample Root Mean Square Error Differences: (Upper Bound of Theoretical Value - Calculated Value using out-of-the-money options) ...... 76
B.4 Out-of-sample Root Mean Square Error Differences: (Upper Bound of Theo- retical Value - Calculated Value using out-of-the-money options) ...... 77
viii ABSTRACT
The use of time-inhomogeneous additive models in option pricing has gained attention in recent years due to their potential to adequately price options across both strike and maturity with relatively few parameters. In this thesis two such classes of models based on the selfsimilar additive processes of Sato are developed. One class of models consists of the risk-neutral exponentials of a selfsimilar additive process, while the other consists of the risk-neutral exponentials of a Brownian motion time-changed by an independent, increasing, selfsimilar additive process. Examples from each class are constructed in which the time one distributions are Variance Gamma or Normal Inverse Gaussian distributed. Pricing errors are assessed for the case of Standard and Poor’s 500 index options from the year 2005. Both sets of time-inhomogeneous additive models show dramatic improvement in pricing error over their associated L´evy processes. Furthermore, with regard to the average of the pricing errors over the quote dates studied, the selfsimilar Normal Inverse Gaussian model yields a mean pricing error significantly less than that implied by the bid-ask spreads of the options, and also significantly less than that given by its associated, less parsimonious, L´evy stochastic volatility model.
ix CHAPTER 1
Introduction
The purpose of this thesis is to study the ability of certain non-L´evy, additive processes to price options on the Standard and Poor’s 500 index (SPX) across both strike and maturity. As noted by Carr and Wu [10] and Eberlein and Kluge [15], L´evy models are incapable of fitting implied volatility surfaces of equity options across both strike and maturity. Although stochastic volatility models with jumps have success in pricing both across strike and maturity [12], the use of such models likely incurs the cost of computing with a significantly larger number of parameters than in the L´evy case. By using additive models which allow for nonstationarity of increments, one obtains greater flexibility in pricing across maturity than with L´evy processes. Furthermore, the additive processes developed in this thesis require only one more parameter than a L´evy model of like time one distribution. This additional parameter is known as the Hurst exponent. One class of models studied consists of price processes whose logarithm is given by a H-selfsimilar additive process of Sato. These models were previously studied by Nolder in 2003 [23]. Recently Carr, Geman, Madan, and Yor [7] demonstrated that exponential H-selfsimilar additive models could be calibrated with great success using options on the Standard and Poor’s 500 index, as well as for individual equity options. One reason for their success is due to the fact that these processes may be constructed from any of a multitude of selfdecomposable distributions already used in finance. In contrast, selfsimilar L´evy models consist solely of Brownian motion and Paretian stable processes [21]. The selfsimilar additive models studied in this thesis are those whose time one distributions coincide with those of the Variance Gamma and Normal Inverse Gaussian processes, denoted by “HssVG” and “HssNIG,” respectively. The other class of models studied consists of those processes whose logarithm is given by
1 a L´evy process time-changed by an independent, increasing H-selfsimilar additive process. Using a specific subset of the class of subordinated additive processes, Madan and Yor [20] showed that for a prespecified set of zero mean probability densities, q , one could construct { t} a Markov martingale process, M for which the one-dimensional distributions were those { t} corresponding to the set q . Furthermore, such a Markov martingale is represented as a { t} Wiener process time-changed by an independent, increasing, additive process. This idea is crucial in the construction of mean-corrected exponential L´evy stochastic volatility models. Here, subordination of a L´evy process is studied with examples of a two parameter Brownian motion time-changed by an independent, increasing, selfsimilar additive process. When the directing process used for the time-change is a selfsimilar Gamma process (resp. selfsimilar Inverse Gaussian process) the resulting process is denoted by “VHG” (resp. “NHIG”). In order that this research may be viewed in historical context, a brief history of pricing with additive processes now follows. The use of additive processes to price options dates prior to 1900 when Louis Bachelier, a student of Henri Poincar´e, began the development of the theory of Brownian motion [14]. In his dissertation, “Theorie de la Speculation,” changes in the price of a stock were modelled 1 with what is now known as a zero mean, arithmetic Brownian motion, a 2 -selfsimilar additive process [21]. Later, Paul L´evy studied and further developed the theory associated with a class of stochastic processes which included Brownian motion as a special case, the stable processes [18]. As in the case of Brownian motion, the increments of these stable processes were both independent and stationary. Unlike the associated one-dimensional distributions of Brownian motion, the non-Gaussian stable distributions permitted existence of, at most, the first moment [25]. The forthcoming financial application came in 1963 when Mandelbrot coined the phrase “stable Paretian” to denote the set of non-Gaussian stable distributions. By modelling the change in the logarithm of the closing spot price of cotton [22] as a Paretian stable distributed random variable, he was able to demonstrate that the tail behavior of his model was consistent with that corresponding to the observed change in the logarithm of the spot price [21]. A special case of this stable Paretian distribution would later be revived by Carr and Wu for the purpose of option pricing [10]. In January of 1973 Peter Clark used Bochner’s idea of subordination to introduce a new pricing model, the Lognormal-Normal process. Given σ1,σ2,µ > 0, the time one
2 2 distribution of the process being time–changed was Normal(0,σ2) distributed while the 2 directing process (subordinator) had a time one distribution which was Lognormal(µ, σ1) distributed. Clark attributed the nonuniformity of the rate of change in cotton futures prices to the nonuniformity of the rate at which information became available to traders. In his argument, fast trading days were caused by the influx of new information inconsistent with previously held expectations, while slow trading days were caused from a lack of new information [11]. This idea would prove useful in the subsequent development of the L´evy stochastic volatility models [6]. During that same year Fisher Black and Myron Scholes produced their closed-form expression for the price of a European call option, where the underlying price process was modelled by a geometric Brownian motion [3]. Another closed-form option pricing formula appeared in 1998 in which the logarithm of the underlier was distributed according to a non-stable L´evy process, known as the Variance Gamma process. In this model, a Brownian motion was time-changed by an independent Gamma process, yielding a model which could better accommodate the leptokurtic and negatively skewed risk-neutral distributions of the logarithm of the return data [19]. Al- though this process lacked selfsimilarity, its increments were independent and stationary, with notably, finite moments. Other well-known L´evy processes include: Normal Inverse Gaussian, Meixner, Generalized Hyperbolic [26], the five parameter CGMY process with drift term included [5], and the more general 6 parameter KoBoL process with drift term included [4]. As noted by Carr and Wu, L´evy models have associated model implied volatility surfaces as a function of maturity, T, and moneyness, ln(K/S) , which tend to a constant value √T as maturity gets large. This was contrary to the observed behavior associated with SPX options where, as maturity increased, the market implied volatility tended to a function which appeared to be independent of maturity alone and decreasing with moneyness [10]. One answer to the problem of fitting option prices across both strike and maturity came in 2003 when Carr, Madan, Geman, and Yor showed that one may subordinate a L´evy process by the time integral of a mean-reverting, positive process in order to obtain a process which was quite flexible across both strike and maturity [6]. Since closed-form characteristic functions exist where the rate of time change is given by the solution to the Cox-Ingersoll Ross or Ornstein-Uhlenbeck stochastic differential equation with one-dimensional Gamma or Inverse Gausssian distribution, one could use Fourier transform methods to quickly price options under such models.
3 Another answer to this problem came with the simpler Finite Moment Log Stable process, a stable Paretian process in which the skewness parameter, β, was set to an extreme value of 1. With β fixed and the drift term set so that the martingale condition was − satisfied, the FMLS model was able to fit SPX option data relatively well with only two free parameters [10]. In an effort to find models which had more flexibility than the Finite Moment Log Stable process and were more parsimonious than the stochastic volatility models, researchers pursued the development of time-inhomogeneous additive processes. In 2003 Nolder proposed the modelling of price processes with the H-selfsimilar additive processes of Sato [23]. Eberlein and Kluge in 2004 used Euro-caplets and swaptions to calibrate a model which incorporated a piecewise stationary additive process [15]. More recently in 2005 Carr, Madan, Geman, and Yor used spot options on the Standard and Poor’s 500 index to calibrate exponential H-selfsimilar additive processes of Sato [7]. In this thesis the set of additive models developed and studied consists of the L´evy models, the H-selfsimilar additive models, and the models formed by time-changing a L´evy process by an independent, increasing H-selfsimilar additive process. Examples from each of these classes are constructed in which the models share like time one distribution. These additive models, along with their related L´evy stochastic volatility models, are calibrated using SPX option data in order to assess their relative pricing performance. Chapter two consists of introductory definitions and theorems regarding additive processes along with a review of L´evy stochastic volatility models. Chapter three explains Sato’s construction of H- selfsimilar additive processes and gives the L´evy characteristics for these processes. Chapter four generalizes Sato’s subordination theorem to the case of additive directing processes. It also contains the needed generalized Laplace transform for the additive case, along with the L´evy spot characteristics of the subordinated additive processes. Chapter five contains the derivations of characteristic functions for both the Variance Gamma and Normal Inverse Gaussian families of additive processes. Also included are the first four central moments of the one-dimensional distributions, as well as those of the increments. The formulation of the risk-neutral price processes and discussions of the numerical option pricing method and calibration problem are found in chapter six. Chapter seven contains the data filters, market parameter values used for pricing, and the calibration results. Chapter eight concludes and suggests future research endeavors.
4 CHAPTER 2
Review of Stochastic Processes
2.1 Additive Processes
The simplest of all distributions is the Dirac measure, or trivial distribution, defined as follows.
Definition 1 [1] Dirac measure, δ : for x Rd, x ∈
1 if x A d δx (A)= ∈ A R 0 if x / A ∀ ∈B ∈ Extending this definition to the case of random variables yields the constant random variables.
d Definition 2 [25] A random variable, X on R , is trivial or constant if its distribution PX is trivial. Furthermore X is non-zero if P = δ . X 6 0
Further extending to the case of processes, one obtains the deterministic processes as defined below.
Definition 3 [25]A Rd-valued stochastic process X is trivial or deterministic if for every { t} t 0, X is trivial. X is non-zero if t 0 with P = δ . ≥ t { t} ∃ ≥ Xt 6 0
Of course, one might not choose to model financial assets with such processes. However, as will be shown, these definitions are needed for technical reasons in the later proofs. In order to define more complicated distributions than the trivial ones, a function is sought which may determine them uniquely, the characteristic function.
5 Definition 4 [25] Characteristic function: The characteristic function µ (z) of a probability measure µ on Rd is defined by b i z,x d µ (z)= e h iµ (dx) , z R . Rd ∈ Z Proposition 5 [25] If µ (z)=b µ (z) for z Rd then µ = µ . 1 2 ∈ 1 2
Furthermore, one mayb determineb convergence in distribution by pointwise convergence of a sequence of characteristic functions as shown below.
Proposition 6 [1] (Glivenko) If µ (z) µ (z) for every z, then µ µ. n → n →
Much of the theory which willb be presentedb is based on a special subset of distributions, the infinitely divisible distributions. In order to define these distributions, one must first define the convolution of two distributions.
Definition 7 [25] Convolution of two distributions: The convolution of two distributions µ and µ on Rd, denoted by µ = µ µ , is a distribution defined by 1 2 1∗ 2
d µ (B)= 1B (x + y) µ1 (dx) µ2 (dy) , B R . Rd Rd ∈B Z Z × Definition 8 [25] Infinite Divisibility: A probability measure µ on Rd is infinitely divisible d if, for any positive integer n, there is a probability measure µn on R such that
n µ =(µn) .
The characteristic function of an infinitelyb b divisible distribution has a unique representa- tion in terms of three key quantities: a symmetric nonnegative-definite matrix, a Rd-valued parameter, and a positive measure on Rd which satisfies a certain integrability condition. The following theorem shows how these quantities are related to an infinitely divisible distribution.
Proposition 9 [25] (i) ( L´evy-Khintchine) Let D = x : x 1 . If µ is an infinitely { | |≤ } divisible distribution on Rd, then
1 i z,x d µ (z) = exp z, Az + i γ,z + e h i 1 i z, x 1D (x) ν (dx) , z R −2 h i h i Rd − − h i ∈ Z b 6 where A is a symmetric nonnegative-definite d d matrix, γ Rd, and ν is a measure on × ∈ Rd satisfying
ν ( 0 ) = 0, { } x 2 1 ν (dx) < . Rd | | ∧ ∞ Z (ii) Conversely, if A is a symmetric nonnegative-definite d d matrix, ν is a measure on × Rd with ν ( 0 ) = 0 and satisfies the previously stated integrability condition, and γ Rd, { } ∈ then there exists an infinitely divisible distribution, µ, whose characteristic function is given by the previously defined, µ (z) in (i). (iii) The representation of µ (z) in (i) by (A, ν, γ) is unique. b A few infinitely divisible distributionsb include the following: the dirac measure, the expo- nential, the L´evy, the student-t, and the Cauchy distributions. Some familiar distributions which are not infinitely divisible are the uniform and binomial distributions [25]. A rather special subset of infinitely divisible distributions are the set of selfdecomposable distributions. Such distributions will play a key role in the formulation of the characteristic functions of the distributions for the processes introduced in later chapters.
Definition 10 [25] An additive process in law is a Rd-valued stochastic process X : t 0 { t ≥ } such that
For any choice of n 1 and 0 t < t < < t , random variables X , X • ≥ ≤ 0 1 ··· n t0 t1 − X , X X ,..., X X − are independent. t0 t2 − t1 tn − tn 1 X =0 a.s. • 0 X is stochastically continuous. •
With the extra technical requirement of c`adl`ag sample paths, the definition of an additive process is obtained below.
Definition 11 [25] An additive process on a measurable space (Ω, ) is an additive process F in law in which there is a Ω with P [Ω ]=1 such that, for every ω Ω , X (ω) is 0 ∈ F 0 ∈ 0 t right-continuous in t 0 and has left limits in t> 0. ≥
7 The following proposition by Sato makes explicit the relationship between infinitely divisible distributions and additive processes in law.
Proposition 12 [25] Let X be a Rd-valued additive process in law and, for 0 s t< { t} ≤ ≤ , let µ be the distribution of X X . Then µ is infinitely divisible and ∞ s,t t − s s,t µ µ = µ for 0 s t u< s,t ∗ t,u s,u ≤ ≤ ≤ ∞ µ = δ for 0 s< s,s 0 ≤ ∞ µ δ as t s s,t → 0 ↑ µ δ as t s. s,t → 0 ↓
Thus, the increments of an additive process in law have distributions which are infinitely divisible. Additive processes in law share other special properties. Suppose one wishes to determine if an additive process X is identical in law with another additive process Y . { t} { t} It is unnecessary to check all of the finite-dimensional distributions of both processes. One d only needs to check equality of the one-dimensional distributions, namely Xt = Yt for each t> 0. This property is made explicit in the following theorem.
′ Rd Proposition 13 [25] If Xt and Xt are -valued additive processes in law such that ′ { } ′ X =d X for any t 0, then X and X are identical in law. t t ≥ { t} t It is noteworthy to mention of a special subset of these additive processes: the L´evy processes.
Definition 14 [25] A L´evyprocess is an additive process where the distribution of X X s+t − s does not depend on s for any s, t 0. ≥
Alternatively, a L´evy process (resp. L´evy process in law) is an additive process (resp. additive process in law) with stationary increments. Another important property of an additive process in law is illustrated in the following proposition.
Proposition 15 [25] Let X be a Rd-valued L´evyor additive process in law. Then it has a modification which is, respectively, a L´evyor additive process.
8 2.2 L´evy Stochastic Volatility Models
L´evy stochastic volatility models may be viewed as an extension of the work of Peter Clark who used Bochner’s idea of subordination in order to model asset prices with a L´evy process time-changed by an independent, increasing L´evy process, namely the Lognormal-Normal process. Instead of subordinating a L´evy process by another L´evy process and thus preserving time-homogeneity and independent increments, Carr, Madan, Geman, and Yor proposed time-changing a L´evy process by a process which could be represented as the time-integral of mean reverting, positive process. The property of mean reversion associated with the the stochastic rate of time change caused the sample paths of the subordinated process to exhibit volatility persistence, otherwise known as ”volatility clustering.” [6]. A couple of popular models used to define the rate of time change of the market are the Cox-Ingersoll Ross (CIR) and the Orstein-Uhlenbeck processes (OU). The CIR process, y is given by { t}
1 dy = κ (η y ) dt + λy 2 dW t − t t t with W as a standard Brownian motion, η as the long-run rate of time change, κ as the rate of mean reversion, and λ as a constant of proportionality with respect to the standard deviation of the solution, y . The OU process y , is defined by letting z be an increasing t { t} { t} L´evy process with dy = λy dt + dz , t − t λt where λ is a mean reversion parameter and y > 0. Another name for z is the 0 { t} “Background Driving L´evy Process,” BDLP [26], [12]. Let X be a Rd-valued L´evy process and y be a solution to one of the previously { t} { t} defined stochastic differential equations. The corresponding L´evy stochastic volatility process, Z is defined by { t}
t Z =d X where Y = y ds { t} { Yt } t s Z0 [26]. The following theorem of Bardorff-Nielsen shows how to choose a BDLP in order to obtain an OU process with a prespecified stationary distribution.
9 Proposition 16 [2] Let c (ξ) be a differentiable and selfdecomposable characteristic function (λ) and let κ (ξ) = log c (ξ) . Suppose that ξκ′ (ξ) is continuous at zero and let φ (ξ)= λξκ′ (ξ) for a λ > 0. Then exp φ(λ) (ξ) is an infinitely divisible characteristic function. Further- (λ) (λ) more, letting z (t) be the homogeneous L´evyprocess for which z (1) has characteristic function exp φ(λ) (ξ) and defining the process y by dy = λy dt + dz(λ) we have that t t − t t a stationary version of yt exists, with one-dimensional marginal distribution given by the characteristic function c (ξ) .
Two popular choices for the stationary distributions are the Gamma and Inverse Gaussian distributions. The characteristic function for the Gamma case is given below [26].
1 λt ϕGamma OU (ξ; t, λ, a, b, y0) = exp iy0λ− 1 e− u − { − λa b + blog itu iu λb b iλ 1 (1 e λt) u − } − − − − − Let X be a Rd-valued L´evy process, independent of Y = t y ds, where y is a { t} t 0 s { s} stationary OU process with Gamma(a, b) distribution. Let −→θ Rn, be the n-parameter X1 ∈R vector which determines the distribution of X1. It follows that the characteristic function,
ϕXYt , of XYt is given by
ϕX ξ; −→θ X , λ, a, b, y0 = ϕY i log ϕX ξ; −→θ X ; λ, a, b, y0 . Yt 1 t − 1 1 If X is a Variance Gamma process with characteristic function given by { t} GM Ct ϕ (ξ; C, G, M)= VGt GM +(M G) iξ + ξ2 − and X is time-changed by the time integral of an OU-Gamma process, then the { t} characteristic function of the time t distribution of the subordinated process is given by:
ϕX (ξ; C, G, M, λ, a, b, y0) = ϕY ( i log ϕX (ξ; C, G, M); λ, a, b, y0) Yt t − 1 = ϕ ( iCy log ϕ (ξ;1, G, M); λ, a, by , 1) Yt − 0 X1 0
= ϕXYt (ξ; Cy0, G, M, λ, a, by0, 1) . This representation is possible due to the fact that the characteristic functions of the time t distributions of X and Y could be factored by C and y , respectively. This is helpful, { t} { t} 0 as it reduces the dimensionality of the domain over which one must minimize a prescribed distance between a set of model and market prices.
10 CHAPTER 3
Selfsimilar Additive Processes
3.1 Selfdecomposability and Selfsimilarity
In this section the definitions of selfdecomposability and selfsimilarity are presented, along with several theorems which will be useful in the construction of time-inhomogeneous selfsimilar processes.
Definition 17 Selfdecomposable [25]: A probability measure µ on Rd is selfdecomposable if, d for any b> 1, there is a probability measure ρb on R such that z µ (z)= µ ρ (z) . b b Not only is a selfdecomposableb distributionb b infinitely divisible, as mentioned in the previous chapter, but the measure, ρ, in the previous definition is also infinitely divisible.
Proposition 18 [25, p.93] If µ is selfdecomposable, then µ is infinitely divisible and, for any b> 1, ρb is uniquely determined and infinitely divisible.
Furthermore, when µ is a selfdecomposable distribution on R, its characteristic function has a very useful representation, as shown in the following proposition.
Proposition 19 [25] A probability measure µ on R is selfdecomposable if and only if
1 2 izx k (x) µ (z) = exp Az + iγz + e 1 izx1[ 1,1] (x) dx −2 R − − − x Z | | 2 k(x) b R ∞ where A 0, γ , k (x) 0, x 1 x dx < , and k (x) is increasing on ≥ ∈ ≥ −∞ | | ∧ | | ∞ ( , 0) and decreasing on (0, ) . −∞ ∞ R
11 Definition 20 [25] A stochastic process X is selfsimilar if, for any a > 0, there is a { t} b> 0 such that X t 0 =d bX t 0 . (3.1) { at | ≥ } { t | ≥ } Of particular importance to the construction of these processes is the representation of the number b in the previous definition.
Proposition 21 [25, p.73] Let X be a Rd-valued, selfsimilar, stochastically continuous, { t} and non-trivial process with X0 = constant a.s. Denote by Γ the set of all a > 0 such that there are b> 0 such that X t 0 =d bX t 0 . It follows that there is a H > 0 such { at | ≥ } { t | ≥ } that, for every a Γ, b = aH . ∈ This H is known as the Hurst exponent, the exponent of the non-trivial selfsimilar process. In the case of time-homogeneous additive processes, the reciprocal of H defines an important parameter for the class of selfsimilar L´evy processes, the strictly stable processes. The value 1 H is known as the index of the non-trivial strictly stable process [25]. 3.2 Strictly Stable L´evy Processes
Before studying the strictly stable processes, their corresponding distributions are first defined.
Definition 22 [25] Strictly Stable Distribution: A probability measure, µ on Rd is strictly stable if for any a> 0, there is a b> 0 such that
µ (z)a = µ (bz) .
Furthermore, a L´evyprocess, X, is strictlyb stableb if the distribution of X1 is strictly stable.
Interestingly, the L´evy processes whose time one distributions are strictly stable coincide with the selfsimilar L´evy processes, as stated in the following.
Proposition 23 [25, p.71] A L´evyprocess is strictly stable iff it is selfsimilar.
Remark 24 The more difficult forward direction of the proof of the previous proposition makes use of Proposition 13 in Chapter 2. In fact, selfsimilarity follows directly upon application of this proposition.
12 It is also true that strictly stable processes have selfdecomposable distributions, as stated in the next proposition. Proposition 25 [25, p.91] Any strictly stable distribution on Rd is selfdecomposable.
Remark 26 As shown by Sato, this theorem follows from the definition of an α-strictly stable distribution, µ, where
a 1 µ (z) = µ a α z , z R,α> 0. ∀ ∈ By fixing b > 1, defining a = bα, and making the appropriate change of variables, ξ 1 ξ b b 7→ b for ξ R the proof is obtained. If 0 0 for all ξ R. It follows that for all a> 0 one may satisfy the definition | | ∈ of selfdecomposability by defining a residual measure, ρb, by b a 1 1 − ρ (ξ)= µ ξ . b b As will be shown in the next section, this relationship between selfsimilarity and b b selfdecomposability extends beyond the L´evy case. 3.3 Sato Processes
The properties of selfsimilarity and selfdecomposability are shared not only with the strictly stable processes but also with a larger subset of additive processes which are not necessarily time-homogeneous. They were recently named after Ken-Iti Sato, who proved existence of these processes in the following theorem. Proposition 27 [25, p.99] If µ is a non-trivial selfdecomposable distribution on Rd, then, for any H > 0, there exists, uniquely in law, a non-trivial H selfsimilar additive process − X : t 0 such that P = µ. { t ≥ } X1 The following proposition is a key element in the proof of the previous proposition. Proposition 28 [25, p.51] If µ :0 s t< is a system of probability measures on { s,t ≤ ≤ ∞} Rd with
µ µ = µ for 0 s t u< , (3.2) s,t ∗ t,u s,u ≤ ≤ ≤ ∞ µ = δ for 0 s< , s,s 0 ≤ ∞ µ δ as s t, s,t → 0 ↑ µ δ as t s. s,t → 0 ↓ 13 then there is an additive process in law X : t 0 such that, for 0 s t< , X X { t ≥ } ≤ ≤ ∞ t − s has the distribution µs,t.
The proof of Proposition 27 may now be explained in the following remark.
H Remark 29 Given a selfdecomposable distribution µ on R, Sato defines µt (z) = µ t z P and X1 = µ. This is a reasonable choice if the corresponding one-dimensional distributions d b b of an additive process X are µ , since for each t> 0, X = aH X as shown below. { t} { t} at t H µat (z) = µ [at] z = µ tH aH z b b H = µt a z b d Since X is additive, equality in distribution follows: X = aH X . In order to show { t} b { at} t that such an additive process X with the chosen set of distributions exists, Sato uses { t} Proposition 28. Making use of selfdecomposability of µ, it follows that for any b> 1,t> 0
H µt (z) = µ t z 1 = µ tHz ρ tH z . b b b b b t H b Since the above relationship holds for b = s > 1, it follows that
H s H H µt (z) = µ t z ρ t H t z t ( s ) h i H = µs (z) ρ t H t z . b b ( s ) b
H By defining µs,t by µs,t = ρ t H t z ,bthe followingb convolution equation is obtained. ( s ) µ = µ µ (3.3) b b t s ∗ s,t As it is necessary to define µ for 0 s t, one must define the following quantities: { t} ≤ ≤ µ0, µ0,t, and µt,t. The choices of µ0 = δ0, µ0,t = µt, and µt,t = δ0 are consistent with the convolution equations 3.2 and imply, along with 3.3, the following
µt µu µs,tµt,u = = µs,u for 0 s t u. µs µt ≤ ≤ ≤
b b H Let 0 s t. Since bµtb(z) µ t z band infinite divisibility of µ implies that µ is ≤ ≤ ≡b b continuous in its argument [1, p. 30] and has no zero [25, p.32], then there is left and b b b 14 right continuity of µs,t by 3.3. Now that the required set of transition measures has been produced, Sato lets X be the additive process in law such that for any t > s 0, µ is { t} ≥ s,t the distribution of X X by Proposition 28. By construction, µ = µ = P , so that t − s 0,1 X1 µ (z) = µ aH z implies X =d aH X for all t,a > 0. As mentioned earlier, since X at t at t { at} H and a Xt are both additive processes in law, by Proposition 13 there is identity in law and b b thus selfsimilarity of X . { t} Interestingly, for a non-trivial selfdecomposable distribution, µ, there is an associated L´evy process, as well as a selfsimilar additive process for which the time one distributions of both processes are given by µ. As Sato points out [25, p.100], these processes are identical in law if and only if H 1 and µ is strictly stable with index, α = 1 . ≥ 2 H Since the increments of an additive process are infinitely divisible, one may derive the corresponding system of L´evy characteristics.
Theorem 30 L´evyCharacteristics of the Selfsimilar Additive Process: Given a non-trivial selfdecomposable distribution, µ, on R, let Y be a R-valued H selfsimilar additive process { t} − on (Ω, , P) with H > 0 and P = µ. If the L´evytriplet corresponding to µ is given F Y1 by (A, ν, γ) where ρ is the L´evydensity of ν, then the generating triplet of Yt is given by
2H AYt = t A H H H t γ + t− [ tH ,tH ] [ 1,1] uρ t− u du 0