¯

CHALMERS ¯¯ GÖTEBORG UNIVERSITY

MASTER’S THESIS

Option Pricing using Lévy Processes

YONGQIANG BU

Department of Mathematical Statistics

CHALMERS UNIVERSITY OF TECHNOLOGY
GÖTEBORG UNIVERSITY
Göteborg, Sweden 2007
Thesis for the Degree of Master of Science (20 credits)

Option Pricing using Lévy Processes

Yongqiang Bu

¯
CHALMERS ¯¯ GÖTEBORG UNIVERSITY

Department of Mathematical Statistics

Chalmers University of Technology and Göteborg University
SE − 412 96 Göteborg, Sweden
Göteborg, June 2007
 Abstract
In this Master’s thesis we price exotic options using Monte Carlo simulations. The asset
price process is modeled as an exponential Lévy process. First we use Lévy processes to fit
the log-returns of S&P 500 historical data. By means of both graphical and quantitative tests
we find that the NIG process and the Meixner perform better than the Brownian motion.
Secondly, we calibrate NIG, Meixner and CGMY Lévy process models using S&P 500 index
vanilla options. The calibration results show that non-Gaussian Lévy processes describes the
market price better than Brownian motion. At last, we use the calibrated models to price
exotic options.

Keywords: Barrier Option; Calibration; Exotic Option; Fast Fourier Transformation; Lévy
Process; Monte-Carlo Simulation.

i
ii
 Acknowledgements
I would like to thank my supervisor Prof. Patrik Albin for his guidance and support
throughout process of this thesis, as well as his encouragement and help during the past two
years. I would also like to thank all members of staff at Chalmers, especially Ivar Gustafsson,
Holger Rootzen, Hans Westergren, Nanny Wermuth, Nils Svanstedt, Peter Kumlin, Michael
Patriksson, Torgny Lindvall, Torbjorn Lundh, Serik Sagitov, Catalin Starica, Christer Borell,
Erik Brodin, Viktor Olsbo and Mattias Sunden for their support to my study. I must thank to
all my friends met at Gothenburg during the past two years. Finally, particularly to express
my gratitude to my parents and my girlfriend, for their continued support and encouragement.

iii
iv
Contents

1 Introduction 1

2 The Lévy process framework 3

2.1 Lévy processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Examples of Lévy processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2.1 Brownian motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2.2 Normal inverse Gaussian process (NIG) . . . . . . . . . . . . . . . . . 4
2.2.3 Meixner process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2.4 CGMY process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Modelling S&P 500 index with Lévy processes . . . . . . . . . . . . . . . . . . 5
2.3.1 Stylized facts of financial time series . . . . . . . . . . . . . . . . . . . 6
2.3.2 Parameter estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3.3 Test of distributional assumptions . . . . . . . . . . . . . . . . . . . . 8

3 Inverse model calibration 11

3.1 Risk-neutral option pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1.1 Equivalent martingale measure . . . . . . . . . . . . . . . . . . . . . . 12
3.2 Pricing formulas for European vanilla options . . . . . . . . . . . . . . . . . . 12
3.2.1 Black-Scholes formula . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2.2 Option pricing using fast Fourier transformation . . . . . . . . . . . . 13
3.3 Model calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.4 Calibration results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4.1 Data selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4.2 Calibration results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

4 Exotic option pricing 21

4.1 Exotic options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.1.1 Barrier option . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.1.2 Lookback option . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2 Pricing methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2.1 Black-Scholes formula . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2.2 Monte Carlo pricing using Lévy processes . . . . . . . . . . . . . . . . 23
4.2.3 Simulation techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

5 Conclusion 29

v
A S&P 500 call option prices 33

vi
Chapter 1

Introduction

The beginning of modern mathematical finance can be attributed to Louis Bachelier who in
year 1900 proposed to model the price process {S(t)}t≥0 of an financial asset as
S(t) = S(0) + σ W (t),
where σ > 0 is a parameter and {W (t)}t≥0 is a standard Brownian motion.
The main drawback of the Bachelier model is that it is possible for prices of financial assets
to becomes negative. Therefore Samuelson suggested the so called Bachelier-Samuelson model
2 /2) t+σ W (t)
S(t) = S(0) e(µ−σ , (1.1)
where µ ∈ R is another parameter. In this model it is instead the log-price process log(S(t))
that is a (not necessarily standard) Brownian motion (with drift).
In their seminal paper [3] Black and Scholes give a theoretically consistent framework
for option pricing based on the model (1.1). This paper changed the world of mathematical
finance and initiated an strong growth of derivative markets. The Bachelier-Samuelson model
is therefore also called the Black-Scholes model (BS), depending on the context.
The Black-Scholes model assumes log-increments of the stock price are Gaussian. However,
there is much empirical evidence for that these log-increments are not Gaussian. This has
led researchers to consider a variety of asset price models with non-Gaussian log-increments
during the last decade. One of the most important and natural family of such model is that
of exponential Lévy processes. In turns out that such processes fit many empirically observed
properties of real world data much better than the Black-Scholes model.
In an exponential Lévy process model the price process is given by
S(t) = S(0) eX(t) for t ≥ 0,
where {X(t)}t≥0 is a Lévy process. Some of the most common Lévy processes X that feature in
such exponential Lévy process models are normal inverse Gaussian processes (NIG), Meixner
processes and CGMY processes. Note that the Black-Scholes model is also an exponential
Lévy process model as Brownian motion with drift is a Lévy process.
In this report we first show that NIG and Meixner Lévy process models perform better
than the Brownian motion when fitted to log-return of stock prices (Chapter 2). Then we
calibrate NIG, Meixner and CGMY Lévy process models by an inverse approach where we fit
their predicted theoretical option prices to observed real world S&P 500 index vanilla option
prices (Chapter 3). Finally we use the latter calibration results together with Monte Carlo
simulations to price European exotic options (Chapter 4).

1
2
Chapter 2

The Lévy process framework

In this chapter we give the definitions of the Lévy processes we will use in our work. We also
fit the corresponding exponential Lévy process models to S&P 500 historical data.

2.1 Lévy processes

We use the following definition of a Lévy process from the book by Cont and Tankov [8]:

Definition 2.1 A cádlág1 real valued stochastic process {X(t)}t≥0 such that X(0) = 0 is
called a Lévy process if it has stationary independent increments and is stochastically con-
tinuous.

An important feature of Lévy process is their intimate link to infinite divisible distributions
(e.g., Sato [13]): If {X(t)}t≥0 }t≥0 is a Lévy process, then every process value X(t) is infinitely
divisible. Conversely, to each infinitely divisible distribution there exist a unique in law Lévy
process {X(t)}t≥0 such that X(1) has that distribution.
Recall that a probability distribution on the real line is said to be infinitely divisible if for
any integer n ≥ 1 there exists independent identically distributed random variables Y1 , . . . , Yn
such that Y1 + ... + Yn has that distribution.
From the above it follows that a Lévy process {X(t)}t≥0 }t≥0 has a unique so called char-
acteristic exponent in form of a continuous function ψ : R → R such that the characteristic
function of X(t) is given by

E{eiuX(t) } = etψ(u) for u ∈ R and t > 0.

2.2 Examples of Lévy processes

2.2.1 Brownian motion
Brownian motion with drift is a Lévy process {X(t)}t≥0 }t≥0 that has Gaussian increments.
Specifically, X(t) is N(µt, σ 2 t)-distributed where σ > 0 and µ ∈ R are parameters.
1
Right continuous with left limits.

3
2.2.2 Normal inverse Gaussian process (NIG)
The normal inverse Gaussian process (NIG) is a Lévy process {X(t)}t≥0 that has normal
inverse Gaussian distributed increments. Specifically, X(t) has a NIG(α, β, δt, µt)-distribution
with parameters α > 0, |β| < α, δ > 0 and µ ∈ R.
The NIG(α, β, δ, µ)-distribution has probability density function
¡ p ¢
αδ K1 α δ 2 − (x − µ)2 © p ª
fNIG (x; α, β, δ, µ) = p exp δ α2 − β 2 + β(x − µ) ,
π 2
δ + (x − µ) 2

where Z n z
1 ∞
1 o
Kv (z) = uv−1 exp − (u + ) du
2 0 2 u
is the modified Bessel function of the third kind, while the characteristic function is given by
³ ¡p p ¢´
φNIG (u; α, β, δ, µ) = exp −δ α2 − (β + iu)2 − α2 − β 2 eiµu .

A NIG(α, β, δ, µ)-distributed random variable has the following stylized features:

βδ
Mean p +µ
α2 − β 2
α2 δ
Variance
(α2 − β 2 )3/2
3β
Skewness √
α δ(α − β 2 )1/4
2
µ ¶
α2 + 4β 2
Kurtosis 3 1+ √
δα2 δ(α2 − β 2 )
See Barndorff-Nielsen [2] on more information about NIG processes.

2.2.3 Meixner process

The Meixner process is a Lévy process {X(t)}t≥0 that has Meixner distributed increments.
Specifically, X(t) has a Meixner(x; a, b, dt, mt)-distribution with parameters a > 0, |b| < π,
d > 0 and m ∈ R.
The Meixner(x; a, b, dt, mt)-distribution has probability density function

(2 cos(b/2))2d n b(x − m) o ¯ ³ i(x − m)) ´¯¯2

¯
fMeixner (x; a, b, d, m) = exp ¯ Γ d + ¯ ,
2aπΓ(2d) a a

where Γ denotes the Gamma function, while the characteristic function is given by
µ ¶2d
cos(b/2)
φMeixner (u; a, b, d, m) = eimu .
cosh(au − ib)/2

A Meixner(x; a, b, dt, mt)-distributed random variable has the following stylized features:

4
 Mean ad tan(b/2) + m
1 a2 d
Variance
2 cos2 (b/2)
r
2
Skewness sin(b/2)
d
2 − cos(b)
Kurtosis 3+
d
See Schoutens [14] on more information about Meixner processes.

2.2.4 CGMY process

The CGMY process is a Lévy process {X(t)}t≥0 such that X(1) is CGMY(C, G, M, Y )-
distributed with parameters C, G, M > 0 and Y < 2.
The probability density function of a CGMY(C, G, M, Y )-distribution takes an analyti-
cally very complicated form, while the characteristic function is given by
© ª
φCGMY (u; C, G, M, Y ) = exp CΓ(−Y )((M − iu)Y − M Y + (G + iu)Y − GY ) .

A CGMY(C, G, M, Y )-distributed random variable has the following stylized features:

Mean C(M Y −1 − GY −1 )Γ(1 − Y )

Variance C(M Y −2 + GY −2 )Γ(2 − Y )
C(M Y −3 + GY −3 )Γ(3 − Y )
Skewness
(C(M Y −2 + GY −2 )Γ(2 − Y ))3/2
C(M Y −4 + GY −4 )Γ(4 − Y )
Kurtosis 3+
(C(M Y −2 + GY −2 )Γ(2 − Y ))2
See Carr, Geman, Madan and Yor [4] on more information about CGMY processes.

2.3 Modelling S&P 500 index with Lévy processes

Our dataset will be the S&P 500 index adjust closed price from 3rd, June, 2002 to 3rd, June,
2007, 1259 trading days, as listed by Yahoo Finance.

1400

1200

1000

800

200 400 600 800 1000 1200

Figure 2.1: S&P 500 index adjusted closed prices

5
 Log returns

0.04

0.02

200 400 600 800 1000 1200

-0.02

-0.04

Figure 2.2: Log-return of S&P 500 index adjusted close prices

2.3.1 Stylized facts of financial time series

Now we discuss some stylized facts of financial time series, see Cont [6] for more information.

Skewness and kurtosis

Skewness and kurtosis of Gaussian distributions are 0 and 3, respectively. However, empirical
financial time series usually display non-zero skewness and higher kurtosis than 3. In our case
the skewness of the daily log-return of S&P 500 data is 0.191731 while the kurtosis is 6.68909.
Hence it cannot be completely correct to model this data set with a Black-Scholes model.

Autocorrelations
Here we check the empirical autocorrelations (ACF) for log-return and squared log-returns of
our data set. For the definition of ACF, please check with any text book on time series:

Sample Autocorrelation Function (ACF)

0.8
Sample Autocorrelation

0.6

0.4

0.2

−0.2
0 50 100 150 200 250
Lag

Figure 2.3: Empirical autocorrelations for log-returns

6
 Sample Autocorrelation Function (ACF)

0.8
Sample Autocorrelation

0.6

0.4

0.2

−0.2
0 50 100 150 200 250
Lag

Figure 2.4: Empirical autocorrelations for squared log-returns

From the above two figures we see that the log-returns are uncorrelated, while the squared
log-returns instead are correlated. Hence it cannot even be completely correct to model the
data with an exponential Lévy process model. However, we will not consider more general
models than that anyway.

Volatility clustering
Large changes in financial data tend to be followed by large changes, of either sign, while small
changes tend to be followed by small changes, see Cont [7]. This experience is supported by
Figure 2.4 above.

2.3.2 Parameter estimation

We will fit the empirical log-return of S&P 500 index to NIG process and Meixner process,
as well as to Brownian motion by means of maximum likelihood estimation (MLE).
Due to the high numbers of parameters of the NIG and Meixner distribution and high
numbers of data points, it turned out to be close to impossibly time consuming to make a
direct MLE fit with the help of standard mathematical software packages, see Jonsson [9].
Therefore we used the methods of moments to get a first parameter estimate to be used as
starting point for the MLE fit in order to significantly speed up the fitting procedure. The
results were as follows:

Brownian motion µ σ
0.00031 0.0098
NIG a b d m
78.3512 -5.70771 0.00756726 0.000862369
Meixner a b d m
0.0279247 -0.178417 0.244316 0.000919888

7
 The lack of an analytically tractable expression for the density function of CGMY distri-
butions made us refrain from trying to fit CGMY processes.

2.3.3 Test of distributional assumptions

Here we consider two ways to evaluate the corectness of the fitted distributions.

Graphical test of distributional assumption

According to the Glivekno-Cantelli theorem, if the sample X1 , ..., Xn has cummulative distri-
bution function F (.; θ), then the ordered sample X(1) ≤ ... ≤ X(n) satisfies

lim max |(i − 0.5)/n − F (X(i) ; θ)| = 0,

n→∞ 1≤i≤n

so that a so called QQ-plot of

{(X(i) , F −1 ((i − 0.5)/n); θ)}ni=1

is an approximative 45o , and a systematic deviation therefrom indicates that the F (.; θ)
assumption is not true.
The following three figures depict QQ-plots of our data set fitted to normal distribution,
Meixner distribution and NIG distribution, respectively.

0.03
0.02
0.01

-0.02 -0.01 0.01 0.02

-0.01
-0.02
-0.03

Figure 2.5: Normal QQ-plot

0.02

0.01

-0.02 -0.01 0.01 0.02

-0.01

-0.02

Figure 2.6: Meixner QQ-plot

8
 0.02

0.01

-0.02 -0.01 0.01 0.02

-0.01

-0.02

Figure 2.7: NIG QQ-plot

The QQ-plots indicate that the empirical data fits much better to the Meixner and NIG
Lévy process models than to the Brownian motion.

Statistical test of distributional assumptions

There are many analytical statistical tests for checking distributional assumptions. Among
them the Kolmogorov-Smirnov distance (K-S) and Anderson & Darling statistic (A-D) [1] are
two common choices.
Writing Femp for the empirical distribution of a data set and Ffit for the fitted distribution,
the K-S distance is given by

KS = max |Femp (x) − Ffit (x)|,

x∈R

while the A-D statistic is given by

|Femp (x) − Ffit (x)|

AD = max p .
x∈R Ffit (x)(1 − Ffit (x))

Note that the A-D statistic pays attention to the fit in the tails by mean of amplifying tail
deviations as compared with the K-S statistic. This can be convenient, e.g., in applivations
to risk analysis etc.
We obtained the following values of the K-S and A-D statistics for our fitted distributions.

KS AD
Normal 0.117641 0.272002
NIG 0.0321954 0.0963174
Meixner 0.0190565 0.0487207

Table 2.1: K-S and A-D test statistics

The smaller value of K-S and A-D means closer of empirical distribution and fitted dis-
tribution. Obviously, the statistic for Lévy Process are smaller than the value for Brownian
Motion. We find that the NIG process and the Meixner perform better than the Brownian
motion.

9
10
Chapter 3

Inverse model calibration

Before calibrating our Lévy process models we have to introduce the risk-neutral option
pricing model.

3.1 Risk-neutral option pricing

We assume that the price B(t) of a risk-free asset satisfies the ordinary differential equation

dB(t) = r B(t) dt,

where r ≥ 0 is the interest rate. Further, we assume that there is a risky asset whose price
S(t) is given by
S(t) = S(0) eX(t) ,
where X(t) is a Lévy process. In our case this Lévy process will be of the type normal,
Meixner, NIG or CGMY. The market does not admit arbitrage.
Recall that an arbitrage is a portfolio strategy such that one starts with zero capital and
at some later time T is sure not to have lost money and has a positive probability to make
money.
By the first fundamental theorem of asset pricing, if there exists a risk-neutral probability
measure, then there is no arbitrage. This risk-neutral probability is a martingale measure Q
that is equivalent to the original probability measure P and such that the underlying asset
price is a Q local martingale.
See Shreve [17] on more information about the above matters.
An European call option is the right but not obligation to buy a contingent claim at the
time of maturity T to a fix strike price K. Thus the payoff function is given by

max(S(T ) − K, 0).

The arbitrage-free value of the option at time t < T can be defined as

Πt = e−r(T −t) EQ [max(S(T ) − K, 0)], (3.1)

where Q is a risk-neutral measure.

11
3.1.1 Equivalent martingale measure
We must find an equivalent martingale measure in order to price the derivatives. For this we
will use the so called mean-correcting martingale measure.
After we have estimated all the parameters of some specific asset price process S(t), then
we add a drift term ω ∈ R in a way appropriate to make the in this way discounted process a
martingale. Specifically, writing q ∈ R for the dividend rate, in our exponential Lévy process
setting, the condition
EQ [S(t)] = S(0) et(r−q)
gives that
ω = r − q − log(φ(−i)),
where φ is the characteristic function of S(1).
Here is a list of the mean-correcting risk neutral drift terms for the Lévy processes we
consider:

Model ω
Normal r−q−µ
CGMY r − q − CΓ(−Y )((M − 1)Y − M Y + (G + 1)Y − GY )
p p
NIG r − q + δ( (α2 − (β + 1)2 − (α2 − β 2 ))
Meixner r − q − 2δ(log(cos β/2)) − log(cos((α + β)/2))

3.2 Pricing formulas for European vanilla options

We consider the case of vanilla options for which the payoff function only depends on the
terminal stock price. We can find an analytical price formula for Brownian motion based
price models, but require numerical solutions for other Lévy process based models.

3.2.1 Black-Scholes formula

With the volatility σ, the interest rate r and the dividend rate q in the exponential Brownian
motion Black-Scholes asset price model, the asset price S(t) at time t is given by
n 1 o
S(t) = S(0) exp σW (t) + (r − q − σ 2 )t .
2
As n o
1
S(T ) = S(t) exp σ(W (T ) − W (t)) + (r − q − σ 2 )(T − t) ,
2
using (3.1), we get the option price
£ ¤
Πt = EQ e−r(T −t) (S(T ) − K)+
£ 1 2 ¤
= EQ e−r(T −t) (x eσ(W (T )−W (t))+(r−q− 2 )(T −t)σ − K)+
£ √ 1 2 ¤
= EQ e−rτ (x e−σ τ Y )+(r−q− 2 σ )τ − K)+ ,

where τ = T − t and
W (T ) − W (t)
Y =− √
T −t

12
is a standard normal random variable. Writing
√ 1 h ³x´ ³ 1 ´ i
d1 = d2 + σ τ = √ log + r − q + σ2 τ ,
σ τ K 2
we thus obtain
Z h n √ ³
1 d2
1 ´ o i 1 2
Πt = e−rτ x exp −σ τ y + r − q − σ 2 τ − K e− 2 y dy
2π −∞ 2
Z d2 n √ ³ ´ Z d2
1 1 1 o 1 1 2
= x exp −σ τ y − q + σ 2 τ − y 2 dy − e−rτ K e− 2 y dy
2π −∞ 2 2 2π −∞
Z d2 n 1 o
1 √
= x e−qτ exp − (y + σ τ )2 dy − e−rτ KΦ(d2 )
2π −∞ 2
= x e−qτ Φ(d1 ) − e−rτ KΦ(d2 ) (3.2)

If we insert S(t) instead of x in the (3.2), then we get the option pricing at time t.

3.2.2 Option pricing using fast Fourier transformation

For more general Lévy process models than those based on Brownian motion we typically
cannot find analytical solutions in the fashion of (3.2). We will therefore now introduce
pricing methods based on characteristics function. When using these methods in practice fast
Fourier transformation can be employed, see Carr and Madan [5] on more information.

The Fourier transform of an option price

Here we will evaluate an European call option price based on the price asset price process
S(t), maturity time T and strike price K. Write k = log(K) and s(T ) = log(S(T )). Let
CT (k) denote the option price and qT the risk-neutral probability density function of the log
price sT .
The characteristic function of the density qT is given by
Z ∞
φT (u) = eius qT (s) ds.
−∞

The option value which is related to the risk-neutral density qT is given by

Z ∞
¡ ¢
CT (k) = e−rT es − ek qT (s) ds.
k

Here CT (k) is not square integrable because when k → −∞ so that K → 0, we have CT →

S(0). To obtain a square integrable function, Carr and Madan [5] suggested consideration of
the modified price cT (k) given by

cT (k) = eαk CT (k),

for a suitable α > 0. Here Carr and Madan suggested to choose α ≈ 0.25, while Schoutens
[15] suggests α ≈ 0.75. The value of α affects the speed of convergence.
The Fourier transform of cT (k) is given by
Z ∞
ψT (υ) = eiυk cT (k) dk.
−∞

13
The inverse corresponding inverse Fourier transform takes the form
Z ∞
1
cT (k) = e−iυk ψT (υ) dυ.
2π −∞

We can use these formulas to get the following option price formula for CT (k):
Z Z
exp(−αk) ∞ −iυk exp(−αk) ∞ −iυk
CT (k) = e ψT (υ) dυ = e ψT (υ) dυ, (3.3)
2π −∞ π 0

where we made use of the fact that the function ψT is odd in its imaginary part and even in
its real part since CT (k) is real.
We may express ψT in terms of φT (k) as
Z ∞ Z ∞
iυk
ψT (υ) = e eαk e−rT (es − ek )qT (s) dsdk
−∞ k
Z ∞ Z s
−rT
= e qT (s) (es+αk − ek+αk )eiυk dkds
−∞ −∞
Z ∞ µ (α+1+iυ)s ¶
−rT e e(α+1+iυ)s
= e qT (s) − ds
−∞ α + iυ α + 1 + iυ
Z ∞
e(α+1+iυ)s
= e−rT qT (s) ds
−∞ (α + iυ)(α + 1 + iυ)
e−rT φT (υ − (α + 1)i)
= . (3.4)
α2 + α − υ 2 + i(2α + 1)υ
Using known expressions for the characteristic function of NIG, CGMY and Meixner in
(3.4), we can use (3.3) to get the option price.

Fast Fourier transformation

Fast Fourier transformation (FFT) is an efficient algorithm to compute the following sum
N
X
ω(k) = e−i2π(j−1)(k−1)/N x(j), (3.5)
j=1

where N is usually a power of 2. FFT is a commonly employed discrete approximation

technique of Fourier transform used to reduce computational labour.
Here we will reexpress the relation (3.3) approximately using the FFT (3.5) as
N
exp(−αk) X −iυj k
CT (k) ≈ e ψT (υj )η (3.6)
π
j=1

with the following conventions and parameter values (as suggested by Carr and Madan [5])
Nλ 2b 2π
υj = η(j − 1), N = 4096, a = N η = 600, b= , ku = −b + (u − 1), λη = .
2 N N
Here a is the upper limit for the integration, while ku is a vector with N values of k and b
sets a bound on the log strike to range between −b and b.

14
 Our formula (3.6) for CT can now be rewritten as
N
exp(−αku ) X −iλη(j−1)(u−1) ibυj
CT (k) ≈ e e ψT (υj )η.
π
j=1

Here we cannot combine a too fine integration grid with a wide enough region for strikes, as
if we choose a too small η we get a fine integration grid but few strikes lying in the region.
Carr and Madan suggest to use Simpson’s weighting rule to obtain an accurate integration
with large η. Then we rewrite our price formula as
N
exp(−αku ) X −i2π(j−1)(u−1)/N ibυj η¡ ¢
CT (k) ≈ e e ψT (υj ) 3 + (−i)j − δj−1 , (3.7)
π 3
j=1

where δn is the Kronecker delta function.

We will use Black-Scholes formula for the normal model together with (3.7) for our other
Lévy process based models to compute call option prices.

3.3 Model calibration

In this section we use historical option prices to calibrate model parameters. In this way we
avoid many problems that are associated with calibrations that are based on the underlying
asset prices.
While the pricing problem is concerned with computing option values given the model
parameters, the calibration problem is concerned with computing the model parameters given
the option prices. Thus the calibration problem is the inverse problem to the pricing problem.
One of the most popular calibration methods are to use least squares, they idea of which is
very simple: Observed market prices (Ci )Ni=1 at t = 0 with different strikes (Ki ) and maturities
(Ti ) should be the same as those proposed by the risk-neutral model prices C θ described in
last section with the model parameters θ. Thus we find the best parameter values θ by means
of minimizing the sum of quadratic deviations between these prices
N
X
∗
θ = arg min (C θ (Ti , Ki ) − Ci )2 . (3.8)
i=1

We will compute the following statistics suggested by Schoutens [15] to measure the quality
of fits:
N
X ÁX
N
|C θ (Ti , Ki ) − Ci )| Ci
APE = ,
N N
i=1 i=1
XN
|C θ (T i , Ki ) − Ci )|
AAE = ,
N
i=1
N
1 X |C θ (Ti , Ki ) − Ci )|
ARPE = ,
N Ci
i=1
v
uN
uX (C θ (Ti , Ki ) − Ci ))2
RMSE = t .
N
i=1

15
3.4 Calibration results
We use S&P 500 historical call option prices on 1st of June 2007 from Yahoo Finance that
are listed in Appendix A below. The market prices were chosen from June 2007 to December
2008. The strike is from 1300 to 2000 with the increment of 25 from 1300 to 1700 and the
increment 100 from 1700 to 2000. The index closed price is 1536.34.
Some of the options have two different prices with the same maturity and strike. In that
case, we choose the price with highest trading volume. We didn’t include the option prices
that were smaller than 1.

3.4.1 Data selection

There are four different versions of the option prices, namely the close price, the bid, the ask
and the mean of the bid and ask.

350
Close Price
Bid Price
300 Ask Price
Mean of Bid and Ask Price

250

200
Price

150

100

50

0
1300 1400 1500 1600 1700 1800 1900 2000
Strike

Figure 3.1: Option data set

Although very few paper discuss the selection of data set, it is crucial for the calibration
results. In particular, for some option with low volume, there are big difference between the
close prices and bid ask prices. This can be explained by that for frequently traded options
the bid and ask prices match, while for some little traded options the last trade date might
be long time ago, so that the last trade does not express the true value of option.
From the above figure we see that the close prices have a lot of outliers compare to bid
and ask. Thus we conclude that the bid and ask prices are better to use than the close prices.
The following table show the results of the calibrations of NIG using the bid, the ask as well
as the mean of bid and ask:

16
 NIG Bid Ask Mean of bid and ask
APE 0.0176 0.0137 0.0140
AAE 2.2634 1.7956 1.8120
ARPE 0.1089 0.0840 0.0894
RMSE 0.9162 0.3381 0.1501

Table 3.1: NIG Calibration Statistics

From the above table we conclude that calibration using mean of bid and ask is slightly
better than bid and ask. All of these three in turn are much better than using close prices.
Thus we will use the mean of bid and ask as our option market data to calibrate models.

3.4.2 Calibration results

the following tables show the calibrated model parameters together with the corresponding
values of APE, AAE, ARPE and RMSE.

Models Parameters
Normal σ
0.1531
CGMY C G M Y
0.0156 0.0767 7.5500 1.2996
NIG α β θ
Meixner 5.0364 -3.3199 0.0881
α β θ
0.3400 -1.4900 0.2900

Table 3.2: Calibration results

Normal CGMY Meixner NIG

APE 0.0575 0.0121 0.0120 0.0140
AAE 7.4543 1.5632 1.5553 1.8120
ARPE 0.3093 0.0793 0.0846 0.0894
RMSE 1.0244 0.0026 0.0426 0.1501

Table 3.3: APE, AAE, ARPE and RMSE for calibrations

The calibrations for CGMY, NIG and Meixner are quite similar in quality and all perform
much better than calibration for Normal. Hence we can get improvements if we employ more
general Lévy processes that Brownian motion in option pricing.
We remark here that the model parameters we got from our calibrations are different
from those obtained by the more conventional calibration method to fit the exponential Lévy
process asset price model to real world asset prices underlying the option.
The following four figures show the theoretical option prices from our calibrations together
with the corresponding market option prices.

17
 350
Market Price
Model Price
300

250

200
Price

150

100

50

0
1300 1400 1500 1600 1700 1800 1900 2000
Strike

Figure 3.2: Normal calibration to S&P 500 option prices

350
Market Price
Model Price
300

250

200
Price

150

100

50

0
1300 1400 1500 1600 1700 1800 1900 2000
Strike

Figure 3.3: NIG calibration to S&P 500 option prices

18
 350
Market Price
Model Price
300

250

200
Price

150

100

50

0
1300 1400 1500 1600 1700 1800 1900 2000
Strike

Figure 3.4: CGMY calibration to S&P 500 option prices

350
Market Price
Model Price
300

250

200
Price

150

100

50

0
1300 1400 1500 1600 1700 1800 1900 2000
Strike

Figure 3.5: Meixner calibration to S&P 500 option prices

19
20
Chapter 4

Exotic option pricing

The option we have discussed until now is the vanilla option which means the payoff function
only depend on terminal value. However, path-dependent options have become popular in the
OTC market in the last twenty years. Barrier option and lookback option are two important
examples of such so called exotic options.
In this chapter we will consider barrier options and discuss their pricing methods based
on our previous calibration results. Before discussing the pricing methods we describe the
barrier option in more detail.

4.1 Exotic options

4.1.1 Barrier option
The holder of a barrier option has the right to buy or sell an asset at a specific price at the
end of the contract. The payoff function of a barrier option depends on whether the price of
the underlying asset crossses a given threshold the barrier before maturity.
There are two types barrier option, namely knock-in options and knock-out options. A
knock-in options is activated only when the underlying asset touches the barrier, while a
knock-out option is instead deactivated when it touches the barrier. For each type of barrier
option, there is an up option and down option version opf it. Thus we have four types of
barrier call options: down-and-out barrier call, down-and-in barrier call, up-and-in barrier
call and up-and-out barrier call.
Let us assume that the duration of the contract is T . Define the maximum and minimum
asset price process up til time t ∈ [0, T ] as

M (t) = sup{S(u) : 0 ≤ u ≤ t} and m(t) = inf{S(u ) : 0 ≤ u ≤ t},

respectively. See Schoutens [15, 16] on more details.

Up-and-in barrier call

The up-and-in barrier call option is a standard European call option with strike K when its
maximum lies above the barrier H, while it is worthless otherwise. The initial price is give
by £ ¤
CUI = e−rT EQ (S(T ) − K)+ 1M (T )≥H .

21
Up-and-out barrier call
The up-and-out barrier call option is a standard European call option with strike K when its
maximum lies below the barrier H, while it is worthless otherwise. The initial price is give
by £ ¤
CUO = e−rT EQ (S(T ) − K)+ 1M (T )<H .

Down-and-in barrier call

The down-and-in barrier call option is a standard European call option with strike K when
its minimum lies below the barrier H, while it is worthless otherwise. The initial price is give
by £ ¤
CDI = e−rT EQ (S(T ) − K)+ 1m(T )≤H .

Down-and-out barrier call

The down-and-out barrier call option is a standard European call with strike K when its lies
above some barrier H, while it is worthless otherwise. The initial price is give by
£ ¤
CDO = e−rT EQ (S(T ) − K)+ 1m(T )>H .

4.1.2 Lookback option

The paper by Nguyen-Ngo [11] treats exotic options for underlying exponential Lévy process
asset price models. For exotic option base on exponential Brownian motion, see Shreve [17].
There are two types of lookback options, namely fixed and floating strike lookback option.
Here we only consider the fixed strike option.
The payoff function of fixed strike lookback option is the difference between the stock
terminal value and its lowest value during the option lifetime. The price is given by
£ ¤
CL = e−rT EQ S(T ) − m(T ) .

4.2 Pricing methods

The main problem with barrier option pricing is to find the distribution of minimum and max-
imum processes m and M . It is possible to obtain the explicit proce formluas in Black-Scholes
normal framework. However, the distribution of minima and maxima of more general Lévy
processes is usually not known explicitly. We will therefore use the Monte Carlo simulation
techniques to estimate the exotic option process. See Schoutens [15, 16] on more details.

4.2.1 Black-Scholes formula

The closed-form solution for the Brownian motion framework attributed to Merton, Reiner
and Rubinstein gives the formulas for the types of exotic we consider as
√ ³ H ´2λ
CDO = S(0)Φ(x1 ) e−qT − K e−rT Φ(x1 − σ T − S(0) e−qT Φ(y1 )
S(0)
³ H ´2λ−2 √
+K e−rT Φ(y1 − σ T )
S(0)

22
 √ ³ H ´2λ
CUI = S(0)Φ(x1 ) e−qT − K e−rT Φ(x1 − σ T ) − S(0) e−qT (Φ(−y) − Φ(−y1 ))
S(0)
³ H ´2λ−2 ¡ √ √ ¢
+K e−rT Φ(−y + σ T ) − Φ(y1 + σ T )
S(0)
−rT Q
£ ¤
CDI = e E (S(T ) − K)+ − CDO
£ ¤
CUO = e−rT EQ (S(T ) − K)+ − CUI
for H > K, while
³ H ´2λ ³ H ´2λ−2 √
CDI = S(0) e−qT Φ(y) − K e−rT Φ(y − σ T )
S(0) S(0)
CUO = 0
£ ¤
CDO = e−rT EQ (S(T ) − K)+ − CDI
£ ¤
CUI = e−rT EQ (S(T ) − K)+
for H ≤ K, where
1³ σ2 ´
λ = r − q + ,
σ2 2
1 ³ H2 ´ √
y = √ log + λσ T ,
σ T S(0)K
1 ³ S(0) ´ √
x1 = √ log + λσ T ,
σ T H
1 ³ H ´ √
y1 = √ log + λσ T .
σ T S(0)
Moreover, for the lookback option in the Brownian motion framework, we have
³ σ2 ´ ³ σ2 ´
CL = S(0) e−qT Φ(a1 ) − Φ(−a1 ) − S(0) e−rT Φ(a2 ) − Φ(−a2 ) ,
2(r − q) 2(r − q)
where
1³ σ 2 ´√ 1³ σ 2 ´√
a1 = r−q+ T and a2 = r−q− T.
σ 2 σ 2

4.2.2 Monte Carlo pricing using Lévy processes

It is not possible to get the closed-form exotic option prices for the more general Lévy process
than Brownian motion we consider. Hence we will use Monte Carlo methods to find these
option prices.
The Monte Carlo Pricing procedure goes as follows:
1. Calibrate the model on the vanilla option prices available in the market (S&P 500 call
option in our case) and find the risk-neutral parameters of the model. (This procedure
has already been carried out in a previous chapter.)
2. Simulate N trajectories of the calibrated Lévy process based models.
3. Calculate the payoff function pi for each trajectory, i = 1, . . . , N .
PN
4. Calculate the sample mean payoff to get the estimated payoff p̂ = i=1 pi /N .

5. Discount the estimated payoff at the risk-free rate and get the derivative price erT p̂.

23
4.2.3 Simulation techniques
Normal inverse Gaussian processes can be described as time changed Brownian motions, which
is they key to simulate them, see the books by Cont and Tankov [8] and Schoutens [15].

2000

1900

1800

1700

1600

1500

1400

1300

1200
0 0.2 0.4 0.6 0.8 1

Figure 4.1: Ten simulated NIG process paths.

As for CGMY processes, they can be simulated by methods developed by Madan and Yor
[10] and Poirot and Tankov [12]2 .

2400

2200

2000

1800

1600

1400

1200

1000

800
0 0.2 0.4 0.6 0.8 1

Figure 4.2: Ten simulated CGMY paths

2
We are grateful to Peter Tankov for providing us with a copy of this paper.

24
 We will not discuss simulation techniques for Meixner processes, and thus not consider
exotic option pricing based on Meixner processes.

4.3 Results
Before pricing the exotic options we checked the accuracy of our simulation approach by
means of pricing the European vanilla option using Monte Carlo simulations for the maturity
date June 20, 2008, and compare the simulated prices with the corresponding real world
market prices from Appendix A. The results were as follows:

Strike BS NIG CGMY Mean of bid and ask Bid Ask Close
1300 282.9 298.8 299.7 296.4 294.9 297.9 286.5
1325 262.2 277.9 278.5 275.6 274.1 277.1 145.0
1350 242.2 257.3 257.5 255.2 253.7 256.7 150.3
1375 222.9 237.0 237.0 235.2 233.7 236.7 182.0
1400 204.3 217.1 216.9 215.6 214.1 217.1 140.5
1425 186.6 197.6 197.3 196.6 195.1 198.1 190.4
1450 169.8 178.7 178.3 178.0 176.5 179.5 176.0
1475 153.9 160.3 159.9 160.1 158.6 161.6 154.0
1500 138.9 142.6 142.3 142.8 141.3 144.3 138.0
1525 124.9 125.6 125.6 126.3 124.8 127.8 124.0
1550 111.9 109.5 109.8 110.6 109.1 112.1 96.3
1575 99.8 94.4 95.2 95.7 94.2 97.2 92.2
1600 88.7 80.5 81.7 81.9 80.4 83.4 75.0
1650 69.3 56.6 58.6 57.4 55.9 58.9 56.0
1700 53.3 38.5 40.6 37.9 36.4 39.4 33.6
1800 30.2 17.5 18.1 13.2 12.2 14.2 9.0

Table 4.1: Monte Carlo price for European Vanilla and Market Prices

From the above table we see that the Monte Carlo simulations give very satisfactory
results (for the mean of bid and ask price).
Next we apply the Monte Carlo approach to exotic option pricing. We selected the ma-
turity time T = 1.0521 and the K = 1500, while the barrier levels ranged from 0.5 S(0) to
1.5 S(0). We used N = 100000 simulated trajectories. The results from NIG and CGMY sim-
ulations are preceeded by those from the closed-form formulas fot the Black-Scoles Brownian
motion framework. Note that results for NIG processes and CGMY processes above are very
similar, while results for normal Brownian motions are a little bit different.

25
 140
BS−CUO
BS−CUI
120

100
Option Price

80

60

40

20

0
0.5 1 1.5
Barrier as Percentage of Spot

Figure 4.3: Brownian motion barrier as percentage of spot

140
BS−CDO
BS−CDI
120

100
Option Price

80

60

40

20

0
0.5 1 1.5
Barrier as Percentage of Spot

Figure 4.4: Brownian motion barrier as percentage of spot

26
 150
NIG−CUO
NIG−CUI

100
Option Price

50

0
0.5 1 1.5
Barrier as Percentage of Spot

Figure 4.5: NIG barrier as percentage of spot

150
NIG−CDO
NIG−CDI

100
Option Price

50

0
0.5 1 1.5
Barrier as Percentage of Spot

Figure 4.6: NIG barrier as percentage of spot

27
 150
CGMY−CUO
CGMY−CUI

100
Option Price

50

0
0.5 1 1.5
Barrier as Percentage of Spot

Figure 4.7: CGMY barrier as percentage of spot

150
CGMY−CDO
CGMY−CDI

100
Option Price

50

0
0.5 1 1.5
Barrier as Percentage of Spot

Figure 4.8: CGMY barrier as percentage of spot

28
Chapter 5

Conclusion

In this master thesis, we focus on the Option Pricing using Lévy processes. We started with
definition of Lévy process and the examples of Lévy process.

We use Maximum-likelihood estimation to estimate the parameters and show that Lévy
process fit the log-returns of of S&P 500 historical data better than Brownian Motion by
means of both graphical and quantitative tests. However, for the distribution which do not
have closed-form, this method cannot be used. Secondly, we price the option price using fast
fourier transformation for Lévy process since it is easy to find the characteristic function for
most of the Lévy process.The calibration results show that non-Gaussian Lévy processes de-
scribes the market price better than Brownian motion. At last, we use the calibration result
to price exotic option using Monte Carlo simulation.

We also test the selection of dataset. The results show that using mean of bid and ask is
slightly better than bid and ask. All of these three in turn are much better than using close
prices.

In the future work, it would be interested to calibrate inverse problem with more option
data. We also need to consider the Lévy process with stochastic volatility.

29
30
Bibliography

[1] Anderson, T.W. and Darling, D.A. (1954). A test of goodness of fit. Journal of the Amer-
ican Statistical Association 49 765-769.
[2] Barndorff-Nielsen,O.E. (1995). Normal inverse Gaussian distributions and the modell-
ing of stock returns. Reasearch Report, Department of Theoretical Statistics, Aarhus
University.
[3] Black, F. and Scholes, M. (1973). The pricing of options and corporate liabilities. Journal
of Political Economy 81 637-654.
[4] Carr, P., Geman, H., Madan, D.H. and Yor, M. (2002). The fine structure of asset
returns: an empirical investigation. Journal of Business 75 305-332.
[5] Carr, P. and Madan, D.H. (1998). Option valuation using the fast Fourier transform.
Journal of Computational Finance 2 61-73.
[6] Cont, R. (2001). Empirical properties of asset returns: stylized facts and statistical issues.
Quantitative Finance 1 223-236.
[7] Cont, R. (2005). Volatility clustering in financial markets: Empirical facts and agent-
based models. In: Long-Memory in Economics, G. Teyssire and A. Kirman Ed’s. Spring-
er Verlag, Heidelberg.
[8] Cont, R. and Tankov, P. (2004). Financial Modelling with Jump Processes. Chapman &
Hall, New York.
[9] Jonsson, F. (2004). On the fitting of generalized hyperbolical distributions to financial
data. Master’s thesis, Department of Mathematical Statistics, Chalmers University of
Technology.
[10] Madan, D.B. and Yor, M. (2005). CGMY and Meixner subordinators are absolutely
continuous with respect to one sided stable subordinators. Prépublication, Laboratoire
de Probabilités et Modeles Aléatories.
[11] Nguyen-Ngoc, L. (2003). Exotic options in general exponential Lévy models. Prépubli-
cation, Universités Paris 6.
[12] Poirot, J. and Tankov, P. (2006). Monte Carlo option pricing for tempered stable
(CGMY) processes. Unpublished manuscript.
[13] Sato, K. (1999). Lévy Process and Infinitely Divisible Distributions. Cambridge Univer-
sity Press, Cambridge.

31
[14] Schoutens, W. (2002). Meixner Process: Theory and application in Finance. EURAN-
DOM Report, EURANDOM, Eindhoven.

[15] Schoutens, W. (2003). Lévy Process in Finance: Pricing Financial Derivatives. Wiley,
New York.

[16] Schoutens, W. (2006). Exotic options under Lévy models: An overview. Journal of Com-
puational and Applied Mathematics 189 526-538.

[17] Shreve, S.E. (2004). Stochastic Calculus for Finance II: Continous-Time Models. Spring-
er Verlag, New York.

32
Appendix A

S&P 500 call option prices

We collected 100 call option prices for the S&P 500 index at the close of market on Jun,1,2007
from Yahoo Finance. The closed index price is S0 = 1536.34. We selected the risk free interest
rate 0.05 and dividend yield 0.019. The depicted prices are the mean of bid and ask prices
that we used for our calibrations.

Strike Jun 15 Jul 20 Sep 21 Dec 21 Mar 21 Jun 20 Dec 19

2007 2007 2007 2007 2008 2008 2008
1300 239.1 244.5 254.0 268.5 296.4 322.9
1325 214.2 220.0 230.4 246.0 275.6 303.2
1350 189.3 195.6 207.1 223.9 239.6 255.2 283.9
1375 164.5 171.4 184.1 202.2 235.2 265.0
1400 139.7 147.4 161.5 181.0 198.5 215.6 246.5
1425 114.9 123.7 139.4 160.4 178.8 196.6 228.3
1450 90.4 100.5 118.1 140.4 159.6 178.0 210.7
1475 66.05 78.2 97.7 121.2 141.1 160.1 193.5
1500 42.85 56.9 78.4 103.0 123.4 142.8 176.8
1525 22.25 38.3 60.6 85.8 106.5 126.3 160.8
1550 6.95 22.25 44.4 69.9 90.7 110.6 145.3
1575 1.275 10.75 31.0 55.4 75.9 95.7
1600 1.15 4.5 20.1 42.6 62.4 81.9 116.3
1650 6.6 22.8 39.6 57.4
1700 10.3 22.9 37.9 67.8
1800 1.25 13.2 34.2
1900 14.4
2000 5.0

33
34