Pricing of Exotic Options With Monte Carlo Simulation in a Levy Market Under Stochastic Volatility...

8/8/2019 Pricing of Exotic Options With Monte Carlo Simulation in a Levy Market Under Stochastic Volatility (Schoutens and S

1/29

The Pricing of Exotic Options by Monte-Carlo

Simulations in a Levy Market with Stochastic

Volatility

Wim Schoutens

and Stijn Symens

October 17, 2002

K.U.Leuven, Celestijnenlaan 200 B, B-3001 Leuven, Belgium. E-mail:

[email protected] of Antwerp, Universiteitsplein 1, B-2610 Wilrijk, Belgium. E-mail:

[email protected]

1


2/29

Abstract

Recently, stock price models based on Levy processes with stochastic

volatility were introduced. The resulting vanilla option prices can be

calibrated almost perfectly to empirical prices. Under this model, we will

price exotic options, like the barrier, lookback and cliquet options, by

Monte-Carlo simulation. The sampling of paths is based on a compound

Poisson approximation of the Levy process involved. The precise choice

of the terms in the approximation is crucial and investigated in detail.

In order to reduce the standard error of the Monte-Carlo simulation, we

make use of the technique of control variates. It turns out that there are

significant differences with the classical Black-Scholes prices.

2


3/29

1 Introduction

The most famous continuous-time model for stock prices or indices is the cel-ebrated Black-Scholes model (BS-model) [11]. It uses the Normal distributionto fit the log-returns of the underlying: the price process of the underlying isgiven by the geometric Brownian Motion

St = S0 exp

2

2

t + Wt

,

where {Wt, t 0} is standard Brownian motion, i.e. Wt follows a Normaldistribution with mean 0 and variance t. Under this model pricing formulae fora variety of options are available. We are particularly interested in the pricingof so-called exotic options of European nature, i.e. the payoff function can bepath-dependent, however there is a fix maturity date and no-early exercise isallowed.

Path-dependent options have become popular in the OTC market in the lastdecades. Examples of these exotic path-dependent options are lookback optionsand barrier options. The lookback call option has the particular feature ofallowing its holder to buy the stock at the minimum it has achieved over the lifeof the option. The payoff of a barrier options depends on whether the price ofthe underlying asset crosses a given threshold (the barrier) before maturity. Thesimplest barrier options are knock in options which come into existence whenthe price of the underlying asset touches the barrier and knock-out optionswhich come out of existence in that case. For example, an up-and-out call hasthe same payoff as a regular plain vanilla call if the price of the underlyingassets remains below the barrier over the life of the option but becomes worthlessas soon as the price of the underlying asset crosses the barrier. Under the Black-Scholes framework closed-form option pricing formulae for the above types ofbarrier and lookback options are available (see for example [24]). We also focuson cliquet options, these options depend on the relative returns of the asset overa series of predetermined periods. These options are popular in mutual fundswith capital protection. However, even in the BS-model, no closed-form pricingformulae exits for the pricing of these types of derivatives.

It is well known however that the log-returns of most financial assets are

asymmetrically distributed and have an actual kurtosis that is higher than thatof the Normal distribution. The BS-model is thus a very poor model to de-scribe stock price dynamics. In real markets traders are well aware that thefuture probability distribution of the underlying asset may not be lognormaland they use a volatility smile adjustment. The smile-effect is decreasing withtime to maturity. Moreover, smiles are frequently asymmetric. To price a set ofEuropean vanilla options, one uses for every strike K and for every maturity Tanother volatility parameter . This is fundamentally wrong since this impliesthat only one underlying stock/index is modeled by a number of completelydifferent stochastic processes. Moreover, one has no guarantee that the chosenvolatility parameters can be used to price exotic options.

3


4/29


5/29

process: the CIR process. These models are called the Levy Stochastic Volatilitymodels (Levy-SV models). In [17] and [40] it was shown that by following thisprocedure, one can almost perfectly calibrate Levy-SV model option prices tomarket prices. Finding explicite formulae for exotic options is almost hopelessin these models. However, once you have calibrated the model to a basic set ofoptions, it is possible to price other (exotic) options using Monte-Carlo simula-tions. Moreover, the complexity of the simulations does not increase drastically;besides the Levy process, one only has to simulate the time-changing process,which is in our case the classical and easily simulated CIR process. The Levyprocess can be simulated based on a compound-Poisson approximation. Specialcare has to be taken for the very small jumps; as proposed in [2] these smalljumps can in some cases be approximated by a Brownian motion.

Throughout this paper we make use of a data set with the mid-prices of aset of European call options on the SP500-index at the close of the market onthe 18h of April 2002. At this day the SP500 closed at 1124.47. The short ratewas at that time equal to r = 0.019 per year and we had a dividend yield ofq = 0.012 per year.

First, we look at the exotic options we want to price, together with generalpricing techniques for vanilla options. In Section 3, we give an overview ofsome popular Levy processes. We focus on the VG-process, the NIG-processand the Meixner process. In the next section, we will use these Levy processesin the construction of Levy-SV models. Basically, a Levy-SV model exist of a

combination of a Levy processes with a stochastic time changing process. Asin the paper [17], our rate of time change is the CIR process. We explain aprocedure to simulate all these ingredients of the Levy-SV models. Next, wecalibrate the different Levy-SV models to our data set of market prices. Thecalibration procedure gives us the risk-neutral parameters of our model whichwe will use to produce a significant number of stock price paths. Finally, wewill perform a number of simulations and compute option prices for all thementioned models. In order to reduce the standard error of the Monte-Carlosimulation, we make use of the technique of control variates. This techniqueis particularly useful in this setting, since we can make use of the vanilla callprices available in the market as control variates.

2 Pricing of Derivatives

Throughout the text we will denote by r the daily interest rate and q the divi-dend yield per year. We assume a fixed planning horizon T. Our market consistof one riskless asset (the bond) with price process given by B = {Bt = ert, 0 t T} and one risky asset (the stock) with price process S = {St, 0 t T}.We focus on European-type derivatives, i.e. no early exercise is possible. Givenour market model, let G({Su, 0 u T}) denote the payoff of the derivativeat its time of expiry T.

According to the fundamental theorem of asset pricing (see [18]) the arbi-

5


6/29

trage free price Vt of the derivative at time t [0, T] is given byVt = EQ[e

r(Tt)G({Su, 0 u T})|Ft],where the expectation is taken with respect to an equivalent martingale measureQ and F= {Ft, 0 t T} is the natural filtration of S = {St, 0 t T}.An equivalent martingale measure is a probability measure which is equivalent(it has the same null-sets) to the given (historical) probability measure and un-der which the discounted process {e(rq)tSt} is a martingale. Unfortunatelyfor most models, in particular the more realistic ones, the class of equivalent

measures is rather large and often covers the full no-arbitrage interval. In thisperspective the BS-model, where there is an unique equivalent martingale mea-sure, is very exceptional. Models with more than one equivalent measures arecalled incomplete.

2.1 Vanilla options

A pricing method which can be applied in general when the characteristic func-tion of the risk-neutral stock price process is known, was developed by Carrand Madan [15] for the classical vanilla options. More precisely, let C(K, T) bethe price of a European call option with strike K and maturity T. Let be apositive constant such that the th moment of the stock price exists. Carr andMadan then showed that

C(K, T) =exp( log(K))

+0

exp(iv log(K))(v)dv (1)

where

(v) =erTE[exp(i(v ( + 1)i) log(ST))]

2 + v2 + i(2 + 1)v .

The Fast Fourier Transform can be used to invert the generalized Fourier trans-form of the call price. Put options can be priced using the put-call parity. ThisFourier-method was generalized to other types of options, like power optionsand self-quanto options in [31].

2.2 Exotic options2.2.1 Barrier and Lookback Options

Let us consider contracts of duration T, and denote the maximum and minimumprocess, resp., of a process Y = {Yt, 0 t T} as

MYt = sup{Yu; 0 u t} and mYt = inf{Yu; 0 u t}, 0 t T.Using risk-neutral valuation, we have that the time t = 0 price of a lookback

call option is given byLC = erTEQ[ST mST].

For single barrier options, we will focus on the following types:

6


7/29

The down-and-out barrier is worthless unless its minimum remains abovesome low barrier H, in which case it retains the structure of a Europeancall with strike K. Its initial price is given by:

DOB = erTEQ[(ST K)+1(mST > H)]

The down-and-in barrier is a normal European call with strike K, if itsminimum went below some low barrier H. If this barrier was neverreached during the life-time of the option, the option is worthless. Itsinitial price is given by:

DIB = erTEQ[(ST K)+1(mST H)]

The up-and-in barrier is worthless unless its maximum crossed some highbarrier H, in which case it retains the structure of a European call withstrike K. Its price is given by:

U IB = erTEQ[(ST K)+1(MST H)]

The up-and-out barrier is worthless unless its maximum remains belowsome high barrier H, in which case it retains the structure of a Europeancall with strike K. Its price is given by:

U OB = erTEQ[(ST K)+1(MST < H)]

We note that the value, DI B, of the down-and-in barrier call option withbarrier H and strike K plus the value, DOB, of the down-and-out barrier optionwith same barrier H and same strike K, is equal to the value C of the vanilla callwith strike K. The same is true for the up-and-out together with the up-and-in:

DIB + DOB = exp(rT)EQ[(ST K)+(1(mST H) + 1(mST < H))]= exp(rT)EQ[(ST K)+]= C;

U IB + U OB = exp(rT)EQ[(ST K)+(1(MST H) + 1(MST < H))]= exp(

rT)E

Q[(S

T K)+]

= C. (2)

We thus clearly see that the price of a vanilla call is correlated with the pricesof the corresponding barrier options.

An important issue for barrier and lookback options is the frequency thatthe stock price is observed for purposes of determining whether the barrier hasbeen reached. The above given formula assume a continuous observation. Often,the terms of the contract are modified and there are only a discrete number ofobservations, for example at the close of each trading day. [13] provide a way ofadjusting the formulas under the Black-Scholes setting for the situation of dis-crete observations in case of lookback options. For barrier options the adjusting

7


8/29

is described in [12]: The barrier H is replaced by Hexp(0.582

T /m) for an

up-and-in or up-and-out option and by Hexp(0.582T /m) for a down-and-in and down-and-out barrier, where m is the number of times the stock pricesis observed; T /m is the time interval between observations. In the numericalcalculations below, we have assumed a discrete number of observations, namelyat the close of each trading day. Moreover, we have assumed a year consists of250 trading days.

2.2.2 Cliquet Options

We also consider a more involved option, a cliquet option. These options arepopular in mutual funds: investors capital is protected and they benifit in alimited way of possible stock price rises. It has a payoff function which dependson the relative returns of the stock after a series of predetermined dates (in ourcase after 1, 2 and 3 years). These (yearly) returns are first floored with zero(capital is protected) and capped with a return cap (gains are limited). We willconsider caps ranging in cap [0.05, 0.15]. The final payoff is the sum of themodified relative returns:

CLIQ = exp(rT)EQ

3i=1

min

max

Si Si1

Si1, 0

,cap

3 Levy Processes

Suppose (z) is the characteristic function of a distribution. If for every positiveinteger n, (z) is also the nth power of a characteristic function, we say thatthe distribution is infinitely divisible. One can define for every such an infinitelydivisible distribution a stochastic process, X = {Xt, t 0}, called Levy process,which starts at zero, has independent and stationary increments and such thatthe distribution of an increment over [s, s + t], s, t 0, i.e. Xt+s Xs, has((z))t as characteristic function.

The function (z) = log (z) is called the characteristic exponent and itsatisfies the following Levy-Khintchine formula [10]:

(z) = iz 2

2 z2 ++

(exp(izx) 1 izx1{|x|


9/29

3.1 Examples of Levy Processes

3.2 The Variance Gamma Process

The characteristic function of the VG(,,) law is given by

V G(u; ,,) = (1 iu+ 2u2/2)1/.This distribution is infinitely divisible and one can define the VG-process X(V G) =

{X(V G)t , t 0} as the process which starts at zero, has independent and station-ary increments and where the increment X(

V G)s+t

X(

V G)s over the time interval

[s, t + s] follows a VG(,/t,t) law. Clearly (take s = 0 and note that V0 = 0),

E[exp(iuX(V G)t )] = V G(u;

t,/t,t)

= (V G(u; ,,))t

= (1 iu+ 2u2/2)t/ .In [25], it was shown that the VG-process may also be expressed as the

difference of two independent Gamma processes. This characterization allowsthe Levy measure to be determined:

V G(dx) =

Cexp(Gx)|x|1dx x < 0Cexp(M x)x1dx x > 0 ,

where

C = 1/ > 0

G =

22

4+

2

2

2

1> 0

M =

22

4+

2

2+

2

1> 0.

The Levy measure has infinite mass, and hence a VG-process has infinitelymany jumps in any finite time interval. Since

+

|x|V G(dx) < , a VG-

process has paths of finite variation. A VG-process has no Brownian componentand its Levy triplet is given by [, 0, V G(dx)], where

=C(G(exp(M) 1) M(exp(G) 1))

M G

With the parameterization in terms of C, G and M, the characteristic func-

tion of X(V G)1 reads as follows:

V G(u; C,G,M) =

GM

GM + (M G)iu + u2C

.

In this notation we will refer to the distribution by the notation VG(C,G,M).

9


10/29

The class of Variance Gamma distributions was introduced by Madan andSeneta [27] in the late 1980s as a model for stock returns. There (and in [28]and [26]) the symmetric case ( = 0) was considered. In [25], the general casewith skewness is treated.

3.3 The Normal Inverse Gaussian Process

The Normal Inverse Gaussian (NIG) distribution with parameters > 0, < < and > 0, NIG(,,), has a characteristic function [3] given by:

N IG (u; ,,) = exp 2 (+ iu)2 2 2 .

Again, one can clearly see that this is an infinitely divisible characteristic func-

tion. Hence we can define the NIG-process X(N IG) = {X(NI G)t , t 0}, withX(

N IG)0 = 0, stationary and independent NIG distributed increments: To be

precise X(NI G)t has a NIG(,,t) law.

The Levy measure for the NIG process is given by

N IG (dx) =

exp(x)K1(|x|)|x| dx,

where K(x) denotes the modified Bessel function of the third kind with index (see [1]).

A NIG-process has no Brownian component and its Levy triplet is given by[, 0, N IG(dx)], where

= (2/)

10

sinh(x)K1(x)dx.

The density of the NIG(,,) distribution is given by

fN IG (x; ,,) =

exp

2 2 + x K1(2 + x2)

2 + x2.

The NIG distribution was introduced by Barndorff-Nielsen [3]. See also [4][32] [33] and [34].

3.4 The Meixner Process

The density of the Meixner distribution (Meixner(,,)) is given by

fMeixner(x; ,,) =(2 cos(/2))2

2(2d)exp

bx

a

+ix

2

,

where > 0, < < , > 0.The characteristic function of the Meixner(,,) distribution is given by

Meixner(u; ,,) =

cos(/2)

cosh ui2

2

10


11/29

Clearly, the Meixner(,,) distribution is infinitely divisible and we canassociate with it a Levy process which we call the Meixner process. More

precisely, a Meixner process X(Meixner) = {X(Meixner)t , t 0} is a stochasticprocess which starts at zero, i.e. X

(Meixner)0 = 0, has independent and sta-

tionary increments, and where the distribution of X(Meixner)t is given by the

Meixner distribution Meixner(,,t).

It is easy to show that our Meixner process X(Meixner) = {X(Meixner)t , t 0}has no Brownian part and a pure jump part governed by the Levy measure

(dx) = exp(x/)x sinh(x/) dx;

The first parameter in the Levy triplet equals

= tan(/2) 21

sinh(x/)/ sinh(x/)dx

Because+

|x|(dx) = the process is of infinite variation.The Meixner process was introduced in [37] (see also [38]) and originates

from the theory of orthogonal polynomials. Later on it was suggested to servefor fitting stock returns in [22]. This application in finance was worked out in[39] and [40].

4 The Levy-Stochastic Volatility Model

4.1 The Levy-Stochastic Volatility Model: Theory

It has been observed that the volatilities estimated (or more general the para-meters of uncertainty) change stochastically over time and are clustered as canbe seen in Figure 1, where the absolute log-returns of the SP500-index over aperiod of 32 years is plotted. One clearly sees that there are periods with highabsolute log-returns and periods with lower absolute log-returns.

In order to incorporate such an effect Carr, Madan, Geman and Yor [17]proposed the following: One increase or decrease the level of uncertainty byspeeding up or slowing down the rate at which time passes. Moreover, in order

to build clustering and to keep time going forward one employs a mean-revertingpositive process as a measure of the local rate of time change. They use as therate of time change the classical example of a mean-reverting positive stochasticprocess: the CIR process yt that solves the SDE

dyt = ( yt)dt + y1/2t dWt,where W = {Wt, t 0} is standard Brownian motion.

The economic time elapsed in t units of calender time is then given by Ytwhere

Yt =

t0

ysds

11


12/29

0

0.05

0.1

0.15

0.2

absolute log returns of SP500 (19702001)

time

absolute

logr

eturn

1970

2001

Figure 1: Volatility clusters: absolute log-returns SP500-index between 1970and 2001

The characteristic function of Yt is explicitly known:

(u, t) =exp(2t/2) exp(2y(0)iu/( + coth(t/2)))

(cosh(t/2) + sinh(t/2)/)2/2,

where =

2 22iuNote, that for c > 0, y = cy = {cyt, t 0}, satisfies the SDE

dyt = (c yt)dt +

cy1/2t dWt, (3)

and the initial condition is y0 = cy0.The (risk-neutral) price process of the stock S = {St, 0 t T} is now

modeled as follows:

St = S0exp((r q)t)

E[exp(XY(t))]exp(XY(t)), (4)

where q is the dividend yield and X = {Xt, 0 t T} is a Levy process withE[exp(iuXt)] = exp(tX (u)).

The characteristic function for the log of our stock price is given by:

(u) = E[exp(iu log(St)] = exp(iu((r q)t + log S0)) (iX (u), t)(iX (i), t)iu

. (5)

The characteristic function is important for the pricing of vanilla options (seeformula (1)). Recall that in these methods we only needed the characteristicfunction of log(St). By the above formula, explicit formulae are at hand.

12


13/29

Note that if our Levy process X = {Xt, t 0} is a VG-process, we have forc > 0, that X = {Xct, t 0} is again a Levy process of the same class, withthe same parameters except the C-parameter, which is now multiplied with theconstant c. The same can be said for the NIG and the Meixner process. Theparameter which takes into account the same time-scaling property is now the-parameter. In combination with (3) this means that in these cases there isone redundant parameter. We therefore, can set y0 = 1, and scale the presentrate of time change to one. More precisely, we have that that the characteristicfunction (u) of (5) satisfies:

V GCI R(u; C,G,M,,,,y0) = V GCI R(u; Cy0,G,M,,/y0, /y0, 1)MeixnerCI R(u; ,,,,,,y0) = MeixnerCI R(u; ,,y0,,/y0, /

y0, 1)

N IGCI R(u; ,,,,,,y0) = NI GCI R(u; ,,y0,,/y0, /

y0, 1)

Also, instead of setting the y0 parameter equal to one, other involved parame-ters, e.g. or C, can be scaled to 1.

Actually, this time-scaling effect lies at the heart of the idea of incorporatingstochastic volatility through making time stochastic. Here, it comes down tothe fact that instead of making the volatility parameter (of the BS-model) sto-chastic, we are making the parameter C, in the VG case, or the parameter , inthe NIG and the Meixner case, stochastic (via the time). Note that this effect,does not only influences the standard deviation (or volatility) of the processes;

also the skewness and the kurtosis are now fluctuating stochastically.

4.2 Calibration to Market Data

Using formula (1) one can easily compute plain vanilla option prices under theabove Levy-SV models. By this one can calibrate model prices to marketsprices for example in the least-squared sense. In Figures 2-4, one can see thatthe Levy-SV models give a very good fit to the empirical option prices of ourSP-500 data set. The o-signs are market prices the +-signs are model prices. InTable 1 an overview is given of the risk-neutral parameters coming out of thecalibration procedure.

For comparative purposes, one computes the average absolute error as apercentage of the mean price. This statistic, which we will denote with ape, isan overall measure of the quality of fit:

ape =1

mean option price

options

|Market price Model price|number of options

Other measures which give also an estimate of the goodness of fit are the averageabsolute error (aae), the root mean square error (rmse) and the average relativepercentage error (arpe):

aae =

options

|Market price Model price|number of options

13


14/29

rmse =

options

(Market price Model price)2number of options

arpe =1

number of options

options

|Market price Model price|Market price

In Table 2 an overview of these measures of fit are given.

1000 1100 1200 1300 1400 15000

20

40

60

80

100

120

140

160

180

SP500 / 18042002 / MeixnerCIR / ape = 0.68 %

strike

option

price

Figure 2: Meixner-CIR calibration on SP500 options (os are market prices, +sare model prices)

5 Monte Carlo simulation of SV-Levy processes:

Theory

5.1 Introduction

Basically, the method goes as follows: we simulate, say m, paths of our stockprices process and calculate for each path the value of the payoff function Vi,i = 1, . . . , m. Then the Monte-Carlo estimate of the expected value of the payoffis

V =1

m

mi=1

Vi. (6)

14


15/29

1000 1100 1200 1300 1400 15000

20

40

60

80

100

120

140

160

180

SP500 / 18042002 / NIGCIR / ape=0.67%

Figure 3: NIG-CIR calibration on SP500 options (os are market prices, +s aremodel prices)

The final option price is then obtained by discounting this estimate: exp(rT)V.The standard error (SE) of the estimate is given by: 1

(m 1)2m

i=1

(V Vi)2.

The standard error decreases with the square root of the number of samplepaths: to reduce the standard error by half, it is necessary to generate fourtimes as many sample paths.

To simulate a stock price path, we first simulated our time change. For theCIR process, this is quite easy and classical: we follow the Euler Scheme.Basically, we discretize the SDE as:

yt = ( yt)t + y1/2t dWt.

Next, we simulate our Levy process upto time YT =T0

ysds. To simulate a Levyprocess, we exploit the well-known (compound-Poisson) approximation of theprocess, which we describe below in detail. It reduces the simulation of a Levyprocess to the simulation of a number of independent Poisson process. Simu-lating independent Poisson distributed random numbers, and as such Poissonprocesses, is easy. We refer to [19].

15


16/29

1000 1100 1200 1300 1400 15000

20

40

60

80

100

120

140

160

180

SP500 / 18042002 / VGCIR / ape=0.69%

Figure 4: VG-CIR calibration on SP500 options (os are market prices, +s aremodel prices)

Finally, we rescale our Levy path according to the path of our stochasticbusiness time and plug this into our formula (4) for the stock price behavior.

5.2 The Compound-Poisson Approximation of a Levy Process

The compound-Poisson approximation procedure is explained in detail for ex-ample in [35]. Further support can be given by [2] and [29]. The procedure goesas follows: Let X be a Levy process with characteristic triplet [, 2, (dx)].

First, we will discretize the Levy measure (dx). We choose some small0 < < 1. Then we make a partition ofR \ [, ] of the following form. Wechoose real numbers a0 < a1 < .. . < ak = , = ak+1 < ak+2 < .. . < ad+1.

The jumps larger than are approximated by a sum of independent Poisson

processes in the following way: We take an independent Poisson process N(i)t

for each interval, [ai1, ai), 1 i k and [ai, ai+1), k + 1 i d, with intensityi given by the Levy measure of the interval. Furthermore we choose a point ciin each interval such that the variance of the Poisson process matches the partof the variance of the Levy process corresponding to this interval.

16


17/29

VG-CIR

C G M y011.9896 25.8523 35.5344 0.6020 1.5560 1.9992 1NIG-CIR

y018.4815 -4.8412 0.4685 0.5391 1.5746 1.8772 1Meixner-CIR

y00.1231 -0.5875 3.3588 0.5705 1.5863 1.9592 1

Table 1: Parameter estimation for Levy SV models

Model ape aae rmse arpe

VG-CIR 0.69 % 0.4269 0.5003 1.33 %NIG-CIR 0.67 % 0.4123 0.4814 1.32 %Meixner-CIR 0.68 % 0.4204 0.4896 1.34 %

Table 2: ape, aae, rmse and arpe for Levy SV models

5.2.1 Approximation of the Small Jumps by their Expected Value

Next, we look at the very small jumps. The first method is by just replacing

them by their expected value. Putting all things together, we approximateX by a process X(d), consisting of a Brownian motion W = {Wt, t 0} and dindependent Poisson processes N(i) = {N(i)t , t 0}, i = 1, . . . , d with parameteri:

X(d)t = t + Wt +

di=1

ci(N(i)t it1|ci|


18/29

We let all (compensated) jumps smaller than contribute to Brownian partof X. To be precise, we again approximate X by a process X(d), consistingof a Brownian motion W = {Wt, t 0} and d independent Poisson processesN(i) = {N(i)t , t 0}, i = 1, . . . , d with parameter i. Only the Brownian partis different from above. We have now:

X(d)t = t + Wt +

di=1

ci(N(i)t i1|ci| 0

lim0

(() )()

= 1. (9)

This condition is implied by

lim0

()

= . (10)

Moreover, if the Levy measure of the original Levy process does not have atomsin some neighborhood of the origin the condition (10) and condition (9) areequivalent. Results on the speed of convergence of the above approximation canbe found in [2].

We conclude by noting that the Meixner and the NIG process satisfy thecondition (9), but the VG does not. In the simulations below we thus use aBrownian motion in the approximation for the Meixner and the NIG, but notfor the VG process.

5.3 On the Choice of the Approximating Poisson Processes

The choice of the intervals [ai1, ai), 1 i k and [ai, ai+1), k + 1 i dis crucial. We typically set k = 100 and d = 2k, so we have the same number

of Poisson processes reflecting a positive as a negative jump. Next, we look atthree different ways to choose the intervals. First we look at equally spaced,then at equally weighted and finally at intervals with inverse linear boundaries.We illustrate this for the VG, NIG and Meixner processes, with parameterstaken from Table 1.

5.3.1 Equally Spaced Intervals

One can choose the intervals equally spaced, i.e. |ai1 ai| is kept fix for all1 i d + 1, i = k + 1. This choice is illustrated in Figure 5, where we plotfor all Levy processes i versus ci. A width equal to 0.001 was chosen and wezoomed in on the range [0.05, 0.05]; k = 100. Note the explosion near zero.

18


19/29

0.05 0 0.050

5

10

15

20

25

30

35

40

Meixner

0.05 0 0.050

5

10

15

20

25

30

35

40

45

NIG

0.05 0 0.050

2

4

6

8

10

12

14

16

18

20

VG

Figure 5: Equally spaced intervals

5.3.2 Equally Weighted Intervals

Here we opt to keep the intensities for the up-jumps and down-jumps corre-sponding to an interval constant. Thus, for equally weighted intervals, the Levymeasures of intervals on the negative part of the real line ([ai1ai)) are keptfix for all 1 i k. Similarly also the measure of intervals corresponding toup-jumps ([ai ai+1)) is kept fix for all k + 1 i d. Note, that for thischoice the outer intervals can become quite large.

5.3.3 Interval with Inverse Linear Boundaries

Finally, we propose the case where the boundaries are given by ai1 = i1and a2k+2i = i

1, 1 i k + 1 and > 0. This leads to much graduallydecreasing intensity parameters i as can be seen from Figure 6, where = 0.2and k = 100. Moreover, there is no explosion to infinity near zero; the intensitiescome even down again. Note, that in Figure 6, we now show the whole rangewith ci [0.2, 0.2] for the same examples as above. Note also that in all casesthe intensities of down jumps are slightly higher than those of the correspondingup jumps; this reflects the fact that log stock returns are negatively skewed.

0.2 0 0.20

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Meixner

0.2 0 0.20

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

NIG

0.2 0 0.20

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

VG

Figure 6: ai1 = 0.2/i and a2k+2i = 0.2/i, 1 i k + 1

19


20/29

5.4 Variance Reduction by Control Variates

If we want to price exotic barrier and lookback options or other exotics (ofEuropean type), we often have information on vanilla options available. Notethat we have obtained our parameters from calibration on market vanilla prices.In this case, where we thus have exact pricing information on related objects, wecan use the variance reduction technique of control variates. The method is ahighly speed-up method, but the implementation depends on the characteristicsof the instruments being valued.

The idea is as follows. Let us assume that we wish to calculate some expected

value, E[G] = E[G({St, 0 t T})] of a (payoff) function G and that thereis a related function H whose expectation E[H] = E[H({St, 0 t T})] weknow exactly. One has to think of G as the payoff function of the exotic optionwe want to price via Monte-Carlo and ofH as the payoff function of the vanillaoption whose price (and thus the expectation E[H]) we observe in the market.

Suppose that for a sample path the value of the function G and H arepositively correlated, e.g. the value of a up-and-in call is positively correlatedwith the value of a vanilla call with same strike price and time to expiry. Thiscan be seen for example from Equation (2).

Define for some number b R a new payoff function

G({St, 0 t T}) = G({St, 0 t T}) + b (H({St, 0 t T} E[H]) .

Note that the expected value of the new function G is the same as the expecta-tion of the original function G. However there can be a significant difference inthe variance: We have

Var[G] = Var[G] 2bcov[G, H] + b2Var[H].

This variance is minimized if b = cov[G, H]/Var[H]. For this minimizing valueof b we find

Var[G] = Var[G]

1 cov

2[G, H]

Var[G]Var[H]

= Var[G](1 corr2(G, H))

Var[G].So if the absolute value of the correlation between G and H is close to 1, thevariance of G will be very small. Clearly, if we find such a highly correlatedfunction H, very large computational savings may be made. H is called thecontrol variate. Note that the method is flexible enough to include severalcontrol variates.

The precise optimal value for b is not known but can be estimated fromthe same simulation. Special care has to be taken however since estimatingparameters determining the result from the same simulation can introduce abias. In the limit of very large numbers of iterations, this bias vanishes. Aremedy for the problem of bias due to the estimation of b is to use an initial

20


21/29

simulation, possibly with fewer iterates than the main run, to estimate b inisolation. The control variate technique usually provides such a substantialspeed-up in convergence that this initial parameter estimation is affordable.

To summarize, we give an overview of the procedure (with an initial estima-tion of b). Recall we want to price an European exotic option expiring at timeT with payoff function G({St, 0 t T}) and that we have a correlated optionexpiring also at time T with payoff H({St, 0 t T}) whose option price isobservable in the market and given by

exp(rT)E[H({St, 0 t T})] = exp(rT)E[H].The expectation is under the markets risk-neutral pricing measure.

We proceed as follows:

1. Estimate the optimal b:

a) Sample a significant number n of paths for the stock price S ={St, 0 t T} (see procedure below) and calculate for each path i: gi = G({St, 0 t T}) and hi = H({St, 0 t T}).

b) An estimate for b is

b =1

n

n

i=1

gihi E[H]n

i=1

gi

.

2. Simulate a significant number m of paths for the stock price S ={

St, 0t T} (see procedure below) and calculate for each path i: gi = G({St, 0

t T}) and hi = H({St, 0 t T}).3. Calculate an estimation of the expected payoff by:

g =1

n

m

i=1

gi b(hi E[H])

.

4. Discount the estimated payoff g at the risk-free rate r to get an estimateof the value of the derivative: The option price is given by exp(rT)g.

The simulation of the stock price process is summarized as follows:

i) Simulate the rate of time-change process y =

{yt, 0

t

T

}.

ii) Calculate from i), the time change Y = {Yt, 0 t T}.iii) Simulate the Levy process X = {Xt, 0 t YT}. Note that we sample

over the period [0, YT].

iv) Calculate the time-changed Levy process XYt , for t [0, T].v) Calculate the stock price process S = {St, 0 t T}.In the Figure 7, one sees in the case of the Meixner-CIR combination a

sample of all ingredients: the rate of time change yt, the stochastic businesstime Yt, the Levy process Xt, the time changes Levy process XYt , and finallythe stock price St.

21


22/29

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

5

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

2

4

0 0.5 1 1.5 2 2.5 3

0.3

00.1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.3

0.2

0.1

0

0.1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

1000

1200

1400

Figure 7: Simulation of yt, Yt, Xt, XYt , and St

6 Monte Carlo Pricing of Exotics under a SV-

Levy Model

6.1 Barrier and Lookback Prices

We take for all barrier options the time to maturity T = 1, the strike K = S0

and the barrier H equal toHU IB = 1.1S0, HUO B = 1.3S0, HDI B = 0.95S0, HDOB = 0.8S0.

For all models, we make n = 10000 simulations of paths covering a one yearperiod. The time is discretised in 250 equally small time steps. We run 100simulations to find an estimate for the optimal b of the control variate. We con-sider both Equally Weighted Intervals (EWI) and Interval with Inverse LinearBoundaries (IILB).

In Tables 3, we compare the price along all model considered together withBlack-Scholes prices. The standard error of the simulation is given below theoption prices in brackets. The volatility parameter in the BS-model is taken

22


23/29


24/29

Model LC UIB UOB DIB DOB

VG-CIR-EWI 135.27 78.50 63.18 17.71 86.07(0.4942) (0.2254) (0.6833) (0.5306) (0.0811)

NIG-CIR-EWI 135.24 79.08 63.54 16.47 86.12(0.4764) (0.2161) (0.6665) (0.4924) (0.0819)

Meixner-CIR-EWI 135.72 78.57 64.34 17.28 86.06(0.4853) (0.2239) (0.7091) (0.5168) (0.0836)

VG-CIR-IILB 134.77 78.66 62.89 17.42 86.16(0.4894) (0.2224) (0.7250) (0.5259) (0.0958)

NIG-CIR-IILB 135.48 78.66 63.27 16.76 86.18(0.4817) (0.2203) (0.6841) (0.5409) (0.0609)Meixner-CIR-IILB 134.83 78.66 63.87 17.24 86.08

(0.4712) (0.2193) (0.7087) (0.5560) (0.0794)

BS min 128.64 65.66 46.26 21.78 69.17BS lse 155.12 81.66 39.98 32.88 83.50BS max 189.76 102.20 30.32 54.73 101.60

Table 3: Exotic Option prices

6.3 Conclusion

If we look at the prices of the exotic options in the Black-Scholes world, weobserve that the BS-prices depend heavily one the choice of the volatility pa-rameter and that it is not clear which value to take. For the Levy-SV modelsthe prices are very close to each other. We conclude that the BS-model it notat all appropriate to price exotics. Moreover their is evidence that the Levy-SVmodels are much more reliable; they give a much better indication than theBS-model.

100

101

102

103

104

105

0

5

10

15

20

25

30

35

40

45

50

number of simulations

SE

Standard error for Upin Barrier

with CVwithout CV

100

101

102

103

104

105

0

5

10

15

20

25

30

35

40

45


SE

Standard error for lookback call

with CVwithout CV

Figure 9: Standard error with and without control variates

24


25/29

100

101

102

103

104

20

40

60

80

100

120

140


option

prices

option prices with MeixnerCIR and CVs

LCUIBUOBDIBDOB

Figure 10: Convergence of options prices

Acknowledgments

The first author is a Postdoctoral Fellow of the Fund for Scientific Research -

Flanders (Belgium) (F.W.O. - Vlaanderen) and wishes to thank Dilip Madan.The second author is a Research assistant of the Fund for Scientific Research -Flanders (Belgium) (F.W.O. - Vlaanderen).

25


26/29

0.05 0.06 0.07 0.08 0.09 0.1 0.11 0.12 0.13 0.14 0.150.05

0.1

0.15

0.2

cap

option

price

BS (sigma=0.1479)BS (sigma=0.1812)BS (sigma=0.2259)MeixnerCIR

Figure 11: Cliquet prices for different caps

References

[1] Abramowitz, M. and Stegun, I.A. (1968) Handbook of Mathematical Func-tions. Dover Publ., New York.

[2] Asmussen, S. and Rosinski, J. (2001) Approximations of small jumps ofLevy processes with a view towards simulation. J. Appl. Probab. 38 (2),482493.

[3] Barndorff-Nielsen, O.E. (1995) Normal inverse Gaussian distributions andthe modeling of stock returns. Research Report No. 300, Department ofTheoretical Statistics, Aarhus University.

[4] Barndorff-Nielsen, O.E. (1996) Normal Inverse Gaussian distributions andstochastic volatility models. Scandinavian Journal of Statistics 24, 113.

[5] Barndorff-Nielsen, O.E., Mikosch, T. and Resnick, S. (Eds.) (2001) LevyProcesses - Theory and Applications. Boston: Birkhauser.

[6] Barndorff-Nielsen, O.E. and Shephard, N. (2000) Modelling by LevyProcesses for Financial Econometrics. In: O.E. Barndorff-Nielsen, T.Mikosch and S. Resnick (Eds.): Levy Processes - Theory and Applications,Birkhauser, Boston, 283318.

[7] Barndorff-Nielsen, O.E. and Shephard, N. (2001) Non-Gaussian OrnsteinUhlenbeck-based models and some of their uses in financial economics.Journal of the Royal Statistical Society, Series B 63, 167241.

[8] Barndorff-Nielsen, O.E. and Shephard, N. (2001) Integrated OU processesand non-Gaussian OU-based stochastic volatility models. , working paper.

26


27/29

[9] Barndorff-Nielsen, O.E., Nicolata, E. and Shephard, N. (2002) Some recentdevelopments in stochastic volatility modelling. Quantitative Finance 2,1123.

[10] Bertoin, J. (1996) Levy Processes. Cambridge Tracts in Mathematics 121,Cambridge University Press, Cambridge.

[11] Black, F. and Scholes, M. (1973) The pricing of options and corporateliabilities. Journal of Political Economy 81, 637654.

[12] Broadie, M., Glasserman, P. and and Kou, S.G. (1997) A continuity cor-rection for discrete barrier options. Math. Finance 7 (4), 325349.

[13] Broadie, M., Glasserman, P. and and Kou, S.G. (1999) Connecting discreteand continuous path-dependent options. Finance and Stochastics 3, 5582.

[14] Boyarchenko, S.I. and Levendorski, S.Z. (2002) Barrier options and touch-and-out options under regular Levy processes of exponential type. Ann.Appl. Prob., to appear.

[15] Carr, P. and Madan, D. (1998) Option Valuation using the Fast FourierTransform. Journal of Computational Finance 2, 6173.

[16] Carr, P., Geman, H., Madan, D.H. and Yor, M. (2000) The fine structure

of asset returns: an empirical investigations. Preprint.[17] Carr, P., Geman, H., Madan, D.H. and Yor, M. (2001) Stochastic Volatil-

ity for Levy Processes. Prepublications du Laboratoire de Probabilites etModeles Aleatoires 645, Universites de Paris 6& Paris 7, Paris.

[18] Delbaen, F. and Schachermayer, W. (1994) A general version of the funda-mental theorem of asset pricing. Math. Ann. 300, 463520.

[19] Devroye, L. (1986) Non-Uniform Random Variate Generation. Springer-Verlag, New York.

[20] Eberlein, E. and Keller, U. (1995) Hyperbolic distributions in finance.Bernoulli 1, 281299.

[21] Eberlein, E., Keller, U. and Prause, K. (1998) New insights into smile,mispricing and value at risk: The hyperbolic model. Journal of Business71 (3), 371406.

[22] Grigelionis, B. (1999), Processes of Meixner Type. Lith. Math. J. 39 (1),3341.

[23] Grigelionis, B. (2000), Generalized z-Distributions and related StochasticProcesses. Matematikos Ir Informatikos Institutas Preprintas Nr. 2000-22,Vilnius.

27


28/29

[24] Hull, J.C. (2000) Options, Futures and Other Derivatives (4th edition).Prentice-Hall.

[25] Madan, D.B., Carr, P. and Chang, E.C. (1998) The variance gamma processand option pricing. European Finance Review 2, 79105.

[26] Madan, D.B. and Milne, F. (1991) Option pricing with V.G. martingalecomponents. Mathematical Finance 1(4), 3955.

[27] Madan, D.B. and Seneta, E. (1987) Chebyshev polynomial approximationsand characteristic function estimation. Journal of the Royal Statistical So-ciety Series B 49(2), 163169.

[28] Madan, D.B. and Seneta, E. (1990) The v.g. model for share market re-turns. Journal of Business 63, 511524.

[29] Pollard, D. (1984) Convergence of stochastic processes. Springer Series inStatistics. Springer-Verlag, New York.

[30] Prause, K. (1999) The Generalized Hyperbolic Model: Estimation, Finan-cial Derivatives, and Risk Measures. Ph.D. thesis, Freiburg i. Br.

[31] Raible, S. (2000) Levy Processes in Finance: Theory, Numerics, and Em-pirical Facts. Ph.D. thesis, Freiburg i. Br.

[32] Rydberg, T. (1996) The Normal Inverse Gaussian Levy Process: Simu-lations and Approximation. Research Report 344, Dept. Theor. Statistics,Aarhus University.

[33] Rydberg, T. (1996) Generalized Hyperbolic Diffusions with ApplicationsTowards Finance. Research Report 342, Dept. Theor. Statistics, AarhusUniversity.

[34] Rydberg, T. (1997) A note on the existence of unique equivalent martingalemeasures in a Markovian setting. Finance and Stochastics 1, 251257.

[35] Rydberg, T. (1998) Some Modelling Results in the Area of Interplay betweenStatistics, Mathematical Finance, Insurance and Econometrics. PhD thesis,

University of Aarhus.

[36] Sato, K. (2000) Levy Processes and Infinitely Divisible Distributions. Cam-bridge Studies in Advanced Mathematics 68. Cambridge University Press,Cambridge.

[37] Schoutens, W. and Teugels, J.L. (1998), Levy processes, polynomials andmartingales. Commun. Statist.- Stochastic Models 14 (1 and 2), 335349.

[38] Schoutens, W. (2000), Stochastic Processes and Orthogonal Polynomials.Lecture Notes in Statistics 146. Springer-Verlag, New York.

28


29/29

[39] Schoutens, W. (2001), The Meixner Process in Finance. EURANDOM Re-port 2001-002, EURANDOM, Eindhoven.

[40] Schoutens, W. (2002), The Meixner Process: Theory and Applications inFinance. EURANDOM Report 2002-004, EURANDOM, Eindhoven.

[41] Yor, M. and Nguyen, L. (2001) Wiener-Hopf Factorization and the Pric-ing of Barrier and Lookback Options under General Levy Processes.Prepublications du Laboratoire de Probabilites et Modeles Aleatoires 640,Universites de Paris 6 & Paris 7, Paris.

29

Pricing of Exotic Options With Monte Carlo Simulation in a Levy Market Under Stochastic Volatility...

Documents

Transcript of Pricing of Exotic Options With Monte Carlo Simulation in a Levy Market Under Stochastic Volatility...