Option pricing with time-changed Lévy processes - PDF

Description
Option pricing with time-changed Lévy processes Sven Klingler a, Young Shin Kim b, Svetlozar T. Rachev c and Frank J. Fabozzi d Abstract In this paper, we introduce two new six-parameter processes based

Please download to get full document.

View again

of 17
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Information
Category:

Spiritual/ Inspirational

Publish on:

Views: 17 | Pages: 17

Extension: PDF | Download: 0

Share
Transcript
Option pricing with time-changed Lévy processes Sven Klingler a, Young Shin Kim b, Svetlozar T. Rachev c and Frank J. Fabozzi d Abstract In this paper, we introduce two new six-parameter processes based on time-changing tempered stable distributions and develop an option pricing model based on these processes. This model provides a good fit to observed option prices. To demonstrate the advantages of the new processes, we conduct two empirical studies to compare their performance to other processes that have been used in the literature. Keywords: Option pricing, stochastic volatility, stochastic-time change, Lévy processes, tempered stable distributions a Department of Finance, Copenhagen Business School,Solbjerg Plads 3, 2000 Frederiksberg - Denmark b Chair of Econometrics and Statistics, Department of Economics, Karlsruhe Institute of Technology, Schlossbezirk 12, Geb Karlsruhe - Germany c Department of Applied Mathematics & Statistics, Stony Brook University,Heavy Engineering S250, Stony Brook, NY USA Corresponding Author. d EDHEC Business School, 393/400 Promenade des Anglais, Nice cedex 3 - France I. Introduction Since the ground-breaking work of Black and Scholes (1973) and Merton (1973) typically referred to as the Black-Scholes model option pricing has been based on the assumption that asset returns are normally distributed. However, not only has the normal distribution assumption been rejected by numerous empirical studies, it is a well-documented fact that asset returns exhibit asymmetry and heavy tails. There are two ways that have been proposed in the literature to deal with non-normality. The first is to include stochastic volatility and allow the variance of the normal distribution to change over time. 1 The second approach uses jumps in the return model. 2 In the model we propose in this paper, we combine these two approaches. We use two probability distributions from the family of tempered stable distributions which are capable of capturing both asymmetry and heavy tails and apply a stochastic-time change to them. The application of the first distribution the classical tempered stable (CTS) distribution 3 to option pricing was studied by Carr et al. (2002). Carr et al. (2003) further developed this model and used a stochastic-time change to include stochastic volatility. Although the work of Carr et al. (2003) is as a milestone in the literature on option pricing with tempered stable distributions, the use of a stochastic-time change to derive a more realistic price process is not conceptually new. Clark (1973) applies the concept of stochastic-time change to Brownian motions in order to obtain more realistic speculative prices. Later, Ikeda and Watanabe (1981) offer insights in stochastic-time changes from the perspective of stochastic differential equations. Two centuries later, Barndorff-Nielsen and Shephard (2003) studied non-gaussion stochastic volatility models. Carr and Wu (2004) extended the approach further by providing an efficient way to include the correlation between the stock price process and the stochastictime change. Huang and Wu (2004) conducted a specification analysis of different option pricing models and concluded that the best pricing model is one based on a process with a high-frequency jump component and diffusion component, with one time change applied to the jump component and one time change applied to the diffusion component. The second 1 The model by Heston (1993) is the most well-known model using this approach. 2 This approach was first introduced by Merton (1976). 3 This distribution is also known as the CGMY distribution. 1 distribution the rapidly decreasing tempered stable (RDTS) distribution was introduced and studied by Kim et al. (2010) and Rachev et al. (2011). The contribution of this paper is twofold. First, we apply the techniques of Carr et al. (2003) to the RDTS process and present a simple way to reduce the number of parameters from seven to six. Second, we conduct an empirical study to illustrate that the reduction of the number of parameters does not influence the performance of our models. We do not follow the approaches of Carr and Wu (2004) and Huang and Wu (2004) in this work because we want to provide an option pricing model with as few parameters and components as possible. Providing an option pricing model with as few parameters as possible is desirable for two reasons. First, from a practitioner s point of view, less parameters are desirable because model calibration can be done faster and more efficiently. Second, from a theoretical perspective, by introducing a stochastic-time change we are walking the thin line between capturing the observed information content correctly and over-fitting our model. A reduction of the number of parameters reduces the risk of over-fitting our model. The remainder of this paper is organized as follows. In Section II, we briefly review the tempered stable distributions of interest and some important formulas from option pricing. After this introductory section, in Section III, we introduce stochastic volatility using a continuous-time change. We empirically evaluate the performance of our proposed models in Section IV and offer some concluding remarks in Section V. II. Tempered Stable Processes and Option Pricing In this section, we first introduce the risk-neutral stock price process as a mean-corrected ordinary exponential of a Lévy process and discuss the merits of the RDTS process afterwards. Let r denote the risk-free interest rate and assume that the dividend paid by S t is zero. The risk-neutral stock price process is given by S t := S 0 e rt+xt E[e Xt ] = S 0 exp((r + ω)t + X t ), (1) where ω is given by the equation e ωt = φ Xt ( i) and X is the return process. We consider two types of Lévy processes as return process: the CTS process and the RDTS process. We 2 state without proof the characteristic functions of the two definitions for the Lévy processes: CTS process: φ Xt (u) = exp(tcγ( α)((λ iu) α λ α + (λ + + iu) α λ α +)), where Γ(x) := 0 t x 1 e t dt denotes the gamma function. RDTS process: φ Xt (u; α, C, λ +, λ, m) = exp(ium + C(G(iu; α, λ + ) + G( iu; α, λ ))), where G(x; α, λ) = 2 α 2 1 λ α Γ ( α 2 ) ( M ( ) ) ( ( ( ) )) α 2, 1 2 ; x α 2λ λ α 1 xγ 1 α 2 M 1 α 2, 3 2 ; x2 1, 2λ 2 Γ is the gamma function and M is the confluent hypergeometric function. 4 It is also possible to add a drift term to the process X, but based on the empirical results of Carr et al. (2002), we do not extend our model with an additional drift term. We next define the expcts and exprdts processes as follows. Let S t be defined as in Equation (1) with return process X. If X is a CTS process, S t is called an exponential CTS (expcts) process. If X is a RDTS process, S t is called exponential RDTS (exprdts) process. An efficient way for pricing European contingent claims is through the characteristic function. Lewis (2001) derived the following pricing formula for European call options. If the stock price S t is given as S t = S 0 e rt+x t E[e X t], where X is a stochastic process and φ X t denotes the characteristic function of X t, then the value of an European call option is: ( C t = K1+ρ e r(t t) πs ρ iu log K/St e Re e t 0 (T t) log φ X 1 (u+iρ) (ρ iu)(1 + ρ iu) du ), (2) where ρ 1 such that φ X1 (u + iρ) for all u R. A similar result holds for European put options. Now we demonstrate why under certain conditions the RDTS model is superior to the CTS model. Consider a European power call option, which is a European call option with terminal payoff max (ST n K, 0), for n N fixed. We recall that φ X t (u + iρ) holds for λ + ρ λ if X is a CTS process and for ρ R if X is a RDTS process. We start calculating the expected final payoff to illustrate the problem. 4 The work of Carr et al. (2002) and Bianchi et al. (2011) offer a formal introduction of these processes over the Lévy measure. 3 E[(S n T K) + ] = = = ( ) = (S n 0 e nx K) + f XT (x)dx log(k/s n 0 ) (e nx+n log(s 0) K)f XT (x)dx (e nx+n log(s0) K) 1 log(k/s0 n) 2π F ubini = 1 2π = S n 0 = S 0 [ e i(u+iρ)x φ XT (u + iρ)dudx (e nx+n log(s0) K)e i(u+iρ)x dx φ XT (u + iρ)du log(k/s0 } n) {{} =( ) log(k/s n 0 ) (e nx+n log(s 0) K)e izx dx, z = u + iρ e nx izx dx K log(k/s0 } n {{ ) } e (n iz)x 1 iz =( ) ] log(k/s 0 ) [ e izx K iz log(k/s n 0 ) e izx dx ] log(k/s 0 ) Next, we conclude that ρ n is necessary to price European power call options with power n : e (n iz)x 1 iz = enx i(u+iρ)x 1 i(u + iρ) = ρ iu e(n+ρ)x e iux and e (n+ρ)x n 0 if and only if ρ n. Since λ + ρ λ is required for the CTS model, it is not possible to price European power call options in a CTS model with n λ +. In this case, the RDTS is a good alternative. III. Including Stochastic Volatility Three alternative ways for modeling stochastic volatility have been proposed in the literature regime-switching models, time-series models for the volatility, and stochastic process to model the stochastic volatility. The principal advantage of the third model class is that it can be implemented efficiently, thereby enabling model calibration. 4 In this paper, we explore the introduction of stochastic volatility by allowing the variance of the distributions to vary over time by introducing a stochastic-time change. This concept uses a stochastic process to model stochastic volatility and, as mentioned earlier, has been subject of researchers since the early 1970s. Carr et al. (2003) apply this technique using the CTS distribution. Here, we first apply the technique of stochastic-time change as introduced in Carr et al. (2003) to the RDTS distribution and then, by explaining the intuition behind the stochastic-time change, we reduce the number of parameters from seven to six. In the next section, we show empirically that this reduction has no negative effect on the ability of the model to capture the market s information content. Although there is a large class of stochastic processes capable of serving as stochastic-time change, we restrict our considerations to the Cox-Ingersoll-Ross (CIR) process. 5 A process (y t ) t T is referred to as a CIR process or a square-root process if the dynamics of y are given by the following stochastic differential equation: dy t = κ(η y t )dt + λ y t dw t, y 0 = ỹ, with κ, η, λ, and ỹ R. We choose the CIR process as intrinsic time because of two financial insights this process provides. First, the CIR process is mean reverting, meaning that it fluctuates around a fixed mean determined by the parameter η. This is desirable because we want the stochastic-time change to fluctuate around a fixed average increasing in turbulent times and falling below average in quiet times. The second desirable property of the CIR process is that it is a strictly positive process. From a modelling perspective, a negative time change would not make sense. We will fix the parameter η and reduce the number of variables of the CIR process from four to three. We justify the evidence for this step later. We next define Y t := t 0 y(s)ds, (3) where y t is a CIR process. We will refer to this random variable as the economic clock. If we let Y t be the random variable given by Equation (3), the characteristic function of Y t is known in closed form: 6 φ(u, t, y(0); κ, η, λ) = A(t, u)e B(t,u)y(0), 5 This process was formally introduced by Cox et al. (1985). 6 See Cox et al. (1985) for further details. 5 with γ = κ 2 2λ 2 iu, A(t, u) = B(t, u) = cosh( γt 2 )+ κ γ exp( κ2 ηt λ 2 ) γt sinh( 2iu κ+γ coth( γt 2 ). 2 )2κη/λ2, and T ime changed processes of interest We next introduce the formal definition of a time-changed Lévy processes. Let X = (X t ) t T denote a Lévy process and Y = (Y t ) t T as defined in Equation (3) with X and Y being independent. Then the process Z t := X Yt (4) is called a time-changed Lévy process. If X is a CTS process, Z is a time-changed CTS (TCCTS) process. If X is a RDTS process, Z is a time-changed RDTS (TCRDTS) process. If the parameter η of the stochastic-time change is fixed to η = 1, we use the notation TCCTSη if X is a CTS process, and TCRDTSη if X is a RDTS process. As above, we introduce the exptccts, the exptcctsη, the exptcrdts, and the exptcrdtsη processes as the mean-corrected ordinary exponential of the TCCTS, the TCCTSη, the TCRDTS, and the TCRDTSη processes, respectively. T heoretical evidence f or these processes Our objective is to provide an option pricing model with as few parameters as possible, therefore, we assume X and Y to be independent. Another beneficial effect of this assumption is that the characteristic function of such a time-changed process is known in closed form: 7 φ Z (u, t) = φ Y ( iψ X (u), t), where ψ X is the characteristic exponent of X. One drawback of these processes is that by introducing stochastic volatility, we will have four more parameters and hence model calibration becomes more difficult and the common black box optimization algorithms provided by several numerical software packages no longer 7 We refer to Carr et al. (2003, p. 353) for a proof of this result. 6 provide reliable results. Therefore, we offer a way to reduce the number of free parameters in the time change from four to three, by fixing η = 1. As we will see in the next section, this parameter reduction has no major effect on the performance of our model. 8 IV. Empirical Investigation To analyze the performance of our models, we performed two empirical studies. The first compares the performance of the time-changed CTS model to the time-changed RDTS model for options with different maturities. The purpose of this study is twofold. First, beyond its theoretical superiority, we want to empirically test if the time-changed RDTS model is superior to the time-changed CTS model. Second, we want to measure the potential adverse influence of fixing the parameter η in the different models. Since Carr et al. (2003) already explained that a time change is necessary to obtain good calibration results for options with different maturities, we did not compare the performance of the time-changed models to the models without time change in the first study. We conduct the second study to compare the performance of the time-changed models to the models without time change on options with the same maturities. Calibration to options with dif f erent maturities Our first empirical study is similar to the investigation by Carr et al. (2003). They obtained market prices of out-of-the-money (OTM) options on the S&P 500 index with maturities between one month and one year for the second Wednesday of each month of the year With these data, they calibrated their proposed models and concluded that the time-changed CTS model is the best time-changed model for option pricing. To be more precise, they did not evaluate the same time-changed CTS process as we do. They considered a more general version of this process, with different values for C and α for the positive and the negative parts of the Lévy measure. Instead of OTM option prices, we collected both out-of-the-money and in-the-money 8 Beside the empirical power of this approach, this approach can also be justified by the intuitive argument that fixing the parameter η to 1 means that the economic time should fluctuate around the business time, but in normal market periods we assume that business time and economic time are equal. 7 (ITM) call option prices on the S&P 500 index with maturities between one month and one year and strikes between 80% and 120% of the current spot level from OptionMetrics s Ivy DB in the Wharton Research Data Services. Our reason for choosing these call options instead of OTM options is twofold. First, options with strike prices close to the current spot level are more liquid than options with strikes far from the current spot level. Therefore, in our opinion, the observed prices of these near-the-money options are a better representation of the information content of the market. Second, empirical studies based on OTM options typically do not consider deep OTM options. For instance, Huang and Wu (2004) in conducting an empirical study using OTM options choose OTM options with strikes between 67.11% and % of the current spot level. 9 Therefore, the moneyness level in our approach does not substantially differ from using OTM options. We use the closing prices of European call options traded on the Chicago Board Option Exchange (CBOE) between January 1 and September 30, We restrict our study sample to options with maturities between four weeks and one year, applying the following three filters to the dataset. First, we remove options with no positive difference between bid and ask prices and options where the bid price is not strictly positive. To apply the next two filters, we consider options with the same maturity for each date. For the second filter, if there are less than 10 options with the same maturity, we remove them from the sample. Third, we remove options where the price difference is less than 0.05% of the spot level of the S&P 500 index. In order to obtain comparable sample data, we do not consider dates where less than two or more than five maturities are remaining. We further restrict our study sample to the dates where between 150 and 200 options are remaining. This leaves 62 sample dates and a total number of 11,903 options. We performed the calibration by applying the fast Fourier pricing scheme proposed above using least squares estimation.we used the Matlab command lsqcurvefit to calibrate the model They stated that moneyness of options they considered, defined as k = log(k/s), ranges between and We used the Matlab command lsqcurvefit to calibrate the model. 8 To assess the goodness of fit, we use the average percentage error (APE), defined as: AP E := n k=1 ( MarketP rice ModelP rice ) n k=1 ( MarketP rice ). There are two principal findings of the empirical analysis. First, as illustrated in panel (a) of Figure 1, the exptcctsη process is better able to fit observed option prices than the exptcrdtsη process in most cases. As we can see in Figure 1 (a), there are some sample days where the exptcrdtsη process performs better than the exptcctsη process. However, the problem with the exptcrdtsη process is that there are sample days where it exhibits a very poor performance. For instance, between 0 and 10 days, the APE of the exptcrdtsη is almost twice that of the APE of the expctsη. The same result holds when comparing the exptccts and the exptcrdts models, as illustrated in Figure 1 (b). Second, comparing panel (a) and (b) of Figure 1 leads to the conjecture that fixing the parameter η has no major negative effect on the performance of the option pricing models. The two panels in Figure 2 provide further evidence to support this conjecture. First, Figure 2 (a) motivates fixing the parameter η from an empirical perspective. Although at a first glance the estimates of η for the two processes seem to differ significantly, the parameter is between 1 and 1.3 for all sample days. Figure 2 (b) shows the difference between the APE for the processes with fixed and variable η. The average difference between the time-changed model with fixed and variable η is 0.3% for the CTS models and 0.015% for the RDTS models. In fact, Figure 2 (b) indicates that the APE for the processes with fixed η is smaller than the APE of the other processes in some cases. This result seems odd upon initial examination because fixing η is only a special case of the time-changed tempered stable processes with variable η. But the optimization problem underlying the model calibration is rather complex and the black box algorithms used for the calibration cannot guarantee that the result is a global minimum. 11 This result emphasizes the need for an option pricing mod
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks