Option Pricing

with the
Extreme Value Distributions

by

Richard A. Bell

Department of Mathematics
King’s College London
The Strand, London WC2R 2LS
United Kingdom
Email: richard.bell@tinkle.me.uk
Tel: +44 (0)1892 661459
14 September 2006

Report submitted in partial fulfillment of the

requirements for the degree of MSc in Financial
Mathematics in the University of London
 Abstract

Recent events in financial markets have highlighted the frequency with

which large scale financial losses occur. Extreme value theory focusses on
the statistical properties of such extreme events. In this paper, extreme
value distributions are applied in the context of option pricing. Analytic so-
lutions to the standard European option problem based on the assumption
that returns are Gumbel distributed are derived and found to perform bet-
ter than the corresponding Black-Scholes solutions. The stochastic process
underlying the Gumbel distribution is found to be a Lévy process and the
Lévy characteristics of an approximating process are derived.
Contents
Comments on Notation 1

1 Introduction 2

2 Modeling Extreme Values 5

2.1 Extreme Value Theory . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Extreme Value Distributions . . . . . . . . . . . . . . . . . . . . . . 6

3 Infinite Divisibility and Lévy Processes 8

3.1 Lévy Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Infinite Divisibility . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.3 Lévy-Khintchine Formula . . . . . . . . . . . . . . . . . . . . . . . . 9
3.4 Subordinators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

4 Extreme Value Market Model 13

4.1 Model Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2 Martingale Condition . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.3 Gumbel Infinite Divisibility . . . . . . . . . . . . . . . . . . . . . . 14
4.4 Gumbel Lévy Process . . . . . . . . . . . . . . . . . . . . . . . . . . 17

5 Gumbel European Option Pricing 19

5.1 Model Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
5.2 Distribution of Returns . . . . . . . . . . . . . . . . . . . . . . . . . 19
5.3 Generic European Call Option Pricing Formula . . . . . . . . . . . 20
5.4 No Arbitrage Condition . . . . . . . . . . . . . . . . . . . . . . . . 22
5.5 European Call Option Pricing Equation . . . . . . . . . . . . . . . . 22
5.6 European Put Option Pricing Equation . . . . . . . . . . . . . . . . 23
6 Numerical Analysis 24
6.1 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
6.2 Data and Software . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
6.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
6.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
6.5 Discussion of Results . . . . . . . . . . . . . . . . . . . . . . . . . . 32

7 Conclusion 34

References 35

Appendices 37

List of Tables
1 Active Trading Days per Contract by Strike Type . . . . . . . . . . 25
2 S&P500 Futures Contracts . . . . . . . . . . . . . . . . . . . . . . . 25
3 Number of Strikes Sampled by Moneyness . . . . . . . . . . . . . . 26
4 Number of Strikes Sampled by Maturity . . . . . . . . . . . . . . . 27
5 Calculated and Historical Dividend Yield by Month . . . . . . . . . 28
6 Number of Root Mean Square Errors by Size . . . . . . . . . . . . . 29
7 Average Root Mean Square Errors for Call Prices ($) . . . . . . . . 30
8 Average Root Mean Square Errors For Put Prices ($) . . . . . . . . 31
9 Average Pricing Bias by Maturity ($) . . . . . . . . . . . . . . . . . 32
10 Average Pricing Bias by Moneyness ($) . . . . . . . . . . . . . . . . 32

List of Figures
1 Extreme Value Density Functions (µ = 0, σ = 1) . . . . . . . . . . . 7
Comments on Notation
Random variables are denoted by X, Y and Z and the notation X = {Xt , t ∈ [0, T ]}
is used to indicate a stochastic process of a random variable, X, indexed by a time
parameter, t. The characteristic function and moment generating function of a
random variable, X, are denoted by φX (c) and MX (c) respectively for some c ∈ R.
Throughout this paper the existence of a filtered probability space, (Ω, F, F, P),
is assumed where F = FT and the filtration, F, is given by F = {Ft , t ∈ [0, T ]}.
d
The symbol = means equivalence in distribution. The notations E[.] and EQ [.] are
used to refer to the expectation under the natural measure, P, and the risk neutral
measure, Q respectively, and 1{A} is used to denote an indicator function which
takes the value of unity if the event A ∈ Ω is true and zero otherwise. In addition,
numbered references within the text refer to equations unless otherwise indicated.

1
1 Introduction
Recent events in financial markets have further endorsed the prevalent view amongst
both market practitioners and academics that extreme events occur more often
than would be expected if asset returns were normally distributed. The NASDAQ
crash of 2000 and the defaults of Enron and WorldCom in 2001 and 2002 respec-
tively are just a few of the recent events which have resulted in large scale financial
losses. It is generally accepted that empirical distributions of returns exhibit fatter
tails than would be expected from a normal distribution. One implication is that
realised volatility in returns may well exceed that which would be expected under
the assumption of normality with obvious implications for both risk management
and instrument pricing methodologies.

The world of finance was transformed by the seminal 1973 paper by Black and
Scholes in which explicit formulae were derived for the prices of European options.
Despite the fact that these formulae are based on the assumption that asset price
returns are normally distributed, some 33 years later, these formulae remain the
linchpin of any option calculator toolbox. The inconsistencies related to the Black-
Scholes model are well known and include the so called volatility smile which gains
its name from the higher levels of implied volatility generated for at the money and
in the money strikes. Undoubtedly, part of the continuing success of the Black-
Scholes model relates to its elegant simplicity, which in some sense is related to the
underlying assumption of normality. Assuming that returns are Gaussian in nature
typically tends to simplify calculations. It is for this reason that the vast majority
of systems which calculate portfolio level risks such as, for example, Value at Risk,
tend to make the assumption of normality. To do otherwise is an undoubted recipe
for complexity.

Nonetheless, there have been concerted efforts on a number of fronts to tackle

the issue of fat tails in return distributions. Stochastic volatility models relax the
Black-Scholes assumption of constant variance and model volatility directly as a
stochastic process. This technique was implemented by Hull and White [12] where
the volatility process was again driven by Brownian motion. More recently Carr et
al. [5] have proposed an approach whereby uncertainty is increased or decreased
by speeding up or slowing down time. The intuition behind the latter technique
can be understood in the context of the scaling property of Brownian motion.

An alternative approach which can be used for instruments such as European con-
tingent claims, is simply to assume an appropriate distribution for asset returns
at the maturity of the contract. Baskshi and Madan [3] and Carr and Madan [7]
showed that the price of plain vanilla European options can be derived by direct

2
integration when the form of the characteristic function of the assumed distribution
of asset returns at maturity is known. In cases where the characteristic function
is not known, Carr and Madan [7] illustrated the circumstances under which the
technique of Fast Fourier Transform can be employed to numerically solve for the
price of the option.

At the heart of both of these techniques is the either implicit, or explicit, ad-
mission that the stochastic process driving the volatility of asset returns in some
way deviates from the standard assumption of Brownian motion. This is of course
intimately related to the assumption of normality since it is well know that Brow-
nian motion follows a Gaussian distribution. A broader class of stochastic models
which retain the properties of stationarity and independence, and which incorpo-
rates Brownian motion as a special case, are referred to as Lévy processes after
the French Mathematician, Paul Lévy, who was the pioneer of the theory. Lévy
processes provide a rich framework within which to specify the asset price process
allowing for the development of models which generate distributions of returns
which more accurately reflect those which have been empirically observed. One
such approach is the CGMY model proposed by Carr et al.[5].

In terms of the distributions which may be considered for asset price returns,
more attention has recently been devoted towards statistical extremes. An anal-
ysis of extreme values has been a constant feature of the literature related to the
insurance industry, where the problem of assessing the likelihood of the occurrence
of rare events is a critical activity. More recently, applications are turning up in
the finance literature, no doubt motivated by the frequency of large scale losses in
financial markets already mentioned. See for example Bacmann and Gawron [2]
and LeBaron and Samanta [13] for recent applications of extreme value distribu-
tions to financial market returns.

In a recent excellent paper, Markose and Alentorn [14], extended the use of extreme
value distributions to the topic of option pricing. By assuming that asset price re-
turns at maturity are distributed as an extreme value, Markose and Alentorn were
able to report simple analytical solutions to the standard European option pric-
ing problem which produced a significantly improved pricing performance over the
standard Black-Scholes approach. It was the apparent simplicity of the approach
taken by Markose and Alentorn, together with the notion that the distribution of
asset returns are directly related to the probability of extreme events, which was
the primary motivation for this paper.

In this paper the use of extreme value distributions in pricing options is exam-
ined further. In particular the relationships between extreme value distributions
and their underlying stochastic processes are considered. The outline of the paper

3
is as follows. In Section 2, some of the basic results of Extreme Value Theory are
reviewed. Section 3 collects some of the main results relating to Lévy processes
and examines the relationship between Lévy processes and their underlying distri-
butions. In Section 4, the suitability of extreme value distributions in the context
of a simple market model are considered, and, some of the properties of their cor-
responding stochastic processes explored. Analytic solutions for European option
prices are derived in Section 5 under the assumption that asset price returns are
Gumbel distributed, and in Section 6, the performance of the derived equations is
tested.

4
2 Modeling Extreme Values
In this section some of the basic results of extreme value theory are reviewed. The
book by Embrechts et al. [9] is regarded as the standard text on extreme value
theory while Coles [8] provides an excellent introduction to the subject. The paper
by Markose and Alentorn [14] also contains further background material.

2.1 Extreme Value Theory

Extreme value theory focusses directly on the statistical behaviour of extreme val-
ues. The extreme value model centers around an examination of the the statistical
properties of the maximum value of a sequence given by
Mn = max{X1 , . . . , Xn }
for n ∈ N, where X1 , . . . , Xn are independent random variables having a common
distribution function F . Typically, the Xi represent extreme values measured at
fixed time intervals. Deriving the distribution of Mn presents no difficulties in
theory since, for for all values of n
P({Mn ≤ z}) = P({X1 ≤ z, . . . , Xn ≤ z})
Yn
= P({Xi ≤ z)}
i=1
= F (z)n .
In practice, however, even small sampling errors in the estimate of F can lead to
very significant discrepancies in the estimate of F n as n becomes large. Instead
theory related to the modeling of extreme values focusses on the asymptotic be-
havior of F n as n → ∞. It is immediately obvious that the distribution of Mn in
2.1, degenerates to a point of mass at its right end-point, z+ , since for any z ≤ z+ ,
F n (z) → 0 as n → ∞. In order to avoid this difficulty, extreme value modeling
focusses on the asymptotic behavior of
Mn − bn
Mn∗ = (2.1)
an
where appropriate choices for sequences of parameters {an > 0} and {bn }, can sta-
bilise the scale and location of Mn∗ as n increases. The main result of extreme value
theory is encapsulated in the following theorem (see for example Coles [8],p48).

Theorem 2.1 (Extremal Types Theorem) If there exist sequences of constants

{an > 0} and {bn } such that
P({(Mn − bn )/an ≤ z}) → G(z) as n → ∞

5
for a non-degenerative distribution G, then G is a member of the Generalised Ex-
treme Value family of distributions
(   −1/ξ )
z−µ
G(z) = exp − 1 + ξ , (2.2)
σ

defined on {z : 1 + ξ(z − µ)/σ > 0} and where µ ∈ (−∞, ∞), σ > 0 and ξ ∈
(−∞, ∞).

Essentially, Theorem 2.1 states that appropriately scaled maximal values converge
to a distribution of the form in 2.2. There are three types of limit distributions,
detailed below, which are characterised by the tail parameter, ξ, and which form
the only possible limit distributions for the scaled maximal values as defined in
2.1. In this sense Theorem 2.1 may be viewed as an extreme value analogy to the
Central Limit Theorem.

2.2 Extreme Value Distributions

There are three distinct classes of distributions underlying the Generalised Ex-
treme Value distribution, 2.2, which correspond to the value of the tail parameter
ξ. Their corresponding cumulative density functions are presented below.

Type I - Gumbel (limξ→0 ):

   
z−µ
G(z) = exp −exp − , (2.3)
σ

for z ∈ (−∞, ∞) and where σ > 0.

Type II - Fréchet (ξ > 0):

(
0, n z≤µ
G(z) = (2.4)
o
z−µ −α


exp − σ
, z>µ

where σ, α > 0.

Type III - Weibull (ξ < 0):

z−µ α
  

exp − − σ
, z<µ
G(z) = (2.5)
1, z≥µ

where σ, α > 0.

6
 Figure 1: Extreme Value Density Functions (µ = 0, σ = 1)

0.4
0.3

0.3
0.2

0.2
p(x)

p(x)
0.1

0.1
0.0

0.0
−5 0 5 10 −5 0 5 10

x x

(a) Fréchet (ξ = 0.3) (b) Weibull (ξ = −0.3)

0.3
0.2
p(x)

0.1
0.0

−5 0 5 10

(c) Gumbel (ξ = 0.0)

The associated probability density functions of the three extreme value distribu-
tions are illustrated graphically in Figure 1. None of the extreme value distribu-
tions are symmetric. It is readily seen from the form of the cumulative distribution
function in 2.4, that the tail of the Fréchet distribution decays polynomially. The
Fréchet distribution is therefore typically characterised as a fat tailed distribution
which is visibly apparent in Figure 1(a). The tails of the Gumbel distribution
decay exponentially and are regarded as moderately fat. Of the three extreme
value distributions, the Weibull distribution is typically characterised as having
thin tails.

7
3 Infinite Divisibility and Lévy Processes
In this section some of the main concepts relating to Lévy processes and their
relationship to infinitely divisible distributions are reviewed. The reader is referred
to Sato [16], Bertoin [4] and AppleBaum [1] as standard texts on Lévy processes.
The results collated in this section draw heavily on these references. A review of
the application of Lévy process in finance may be found in Schoutens [18]. The
ideas presented in this section will be used in examining the properties of the
Lévy processes associated with extreme value distributions in Section 4 and in the
development of analytic European option pricing equations in Section 5.

3.1 Lévy Processes

Processes with the properties of independent and stationary increments are called
Lévy processes after the French mathematician Paul Lévy (1886-1971) who first
made the link between such processes and their infinitely divisible laws. The con-
cept of a Lévy process is made more precise by the following definition which can
be found in Applebaum [1](p39).

Definition 3.1 (Lévy Process) The stochastic process X = {Xt , t ≥ 0}, of a

random variable, X, is said to be a Lévy process if the following conditions hold.

1. E[X0 ] = 0
2. X has independent and stationary increments.
3. X is stochastically continuous i.e. for a > 0 and s ≥ 0
lim P({|Xt − Xs | > a}) = 0
t→s

The simplest example of a Lévy process is linear drift which is of course a deter-
ministic process. The only example of a stochastic Lévy process which has almost
sure continuous sample paths is Brownian motion. This will become apparent once
the concept of an infinitely divisible law and its relationship to a Lévy process has
been defined. The following definition of an infinitely divisible law is again due to
Applebaum [1](p23).

3.2 Infinite Divisibility

Definition 3.2 (Infinite Divisibility) A random variable, X, is said to have

infinitely divisible law if, for all n ∈ N, there exist independently and identically

8
 (n) (n)
distributed random variables Y1 ,. . .,Yn such that
d (n)
X = Y1 + . . . + Yn(n) . (3.1)

Equivalently, an infinitely divisible random variable may be defined using its char-
acteristic function. This alternative definition will be convenient in the context
of defining an equivalent martingale measure for an asset price process via a no
arbitrage argument in Section 4.

Definition 3.3 (Infinite Divisibility) The law of a random variable X is in-

finitely divisible, if for all n ∈ N, there exists a random variable X (n) , such that

φX (c) = φX (n) (c)n (3.2)

The example below demonstrates the infinite divisibility of the Poisson distribution
using definition 3.3.

Example 3.4 (Poisson Distribution) Let X be a Poisson distributed random

variable with mean λ, then

φX (c) = exp λ eic − 1 ]



λ ic
 
= exp n e − 1


n
λ ic
n
  
= exp e −1
n
= (φX (n) (c))n

where X (n) is Poisson distributed with mean nλ .

3.3 Lévy-Khintchine Formula

The notion of infinite divisibility and Lévy processes are closely related. In fact
each infinitely divisible distribution is uniquely associated with a Lévy process.
The following important theorem provides a complete characterisation of an in-
finitely divisible random variable and its associated Lévy process in terms of its
characteristic function, and is called the Lévy Khintchine formula. For simplicity
the formula is stated for a random variable, although it can be easily restated for
a random vector.

9
Theorem 3.5 (Lévy-Khintchine) The law of a random variable X is infinitely
divisible if, and only if, there exists a triplet (α, σR2 , ν (dx)) with α ∈ R, σ 2 ∈
[0, +∞) and a measure satisfying ν ({0}) = 0 and R 1 ∧ |x| ν (dx) < ∞, such
2

that
 icX  i
 Z 
1 2 2
E e = exp icα − σ c + e − 1 − icx1{|x|<1} ν (dx)
cx


(3.3)
2 R
The triplet (α, σ 2 , ν (dx)) is referred to as the Lévy or characteristic triplet and
the exponent in 3.3

eicx − 1 − icx1{|x|<1} ν (dx)

Z
1 2 2
Ψ (c) = icα − σ c +


(3.4)
2 R−{0}
is called the Lévy or characteristic exponent. It is apparent that if in 3.3, both
σ = 0 and the measure ν (dx) is null everywhere, that the characteristic exponent
is reduced to iαc. For this reason, α is typically referred to as the drift component.
If σ > 0 and the Lévy measure is null, the formula reduces to the characteristic
function of a normal random variable with mean α and variance σ 2 . As a result,
the second term is typically referred to as the Gaussian component. The third
characteristic is referred to as the Lévy measure.

Example 3.6 (Normal Distribution) It is well known that the characteristic

function of a normally distributed random variable, X, with mean µ and variance
σ 2 is given by
φ (c) = eµic− 2 σ c .
1 2 2

With reference to definition 3.3, it is easy to see that the normal distribution is
infinitely divisible. The Lévy Khintchine formula reveals that for the process of a
normally distributed random variable, the Lévy measure is null and that the Lévy
triplet is given by (µ, σ 2 , 0).

In the context of a stochastic process, {Xt , t ∈ [0, T ]}, where X1 is normally dis-
tributed with mean α and variance σ 2 , and in view of the infinite divisibility of the
normal distribution, it can easily be seen that the Lévy triplet of Xt is given by
(αt, σ 2 t, 0), in which case the characteristic exponent is immediately recognisable
as that of Brownian motion with drift α.

The third term in the Lévy Khintchine formula is sometimes called the jump com-
ponent. An intuition behind the nature of this term may be obtained by assuming
(see Applebaum [1]) that the Lévy measure is finite, in which case, 3.4 may be
rewritten as
eicx − 1 ν (dx)
Z
1 2 2
Ψ (c) = icα − σ c +


(3.5)
2 R−{0}
10
where α = α − R−{0} x1{|x<1|} ν (dx). Now take ν (dx) = λδh (dx) where λ > 0
R

and δh is a Dirac measure concentrated at h ∈ R − {0}. Consider a Poisson

process, N = {Nt , t ∈ [0, T ]}. By by the infinite divisibility property of the Poisson
distribution and the form of the Poisson characteristic function it follows that

φN1 (c) = exp λ eich − 1

eicx − 1 λδh (dx) .

Z


=
R−{0}
The latter equality is immediately recognisable as the jump term in 3.5 with mea-
sure λδh (dx). Since the random variable, Nt , counts the number of arrival times
before t, it can be seen that the path of the process N will beP randomly inter-
spersed with jump discontinuities of size |h|. Taking ν (dx) = N k=1 λk δhk (dx),
where hk ∈ R − {0} and N ∈ N, the path of the process would be randomly
interspersed with jump discontinuities of size |hk |.

It becomes clear that the Lévy measure is intimately related to the path properties
of the Lévy process. If the Lévy measure is finite, then almost surely all paths
of the associated Lévy process have a finite number of jumps on every compact
interval. In this case the Lévy process is said to have finite activity. Alternatively,
if the Lévy measure is infinite, then almost surely all sample paths have an infinite
number of jumps on every compact interval, and the Lévy process is said to have
infinite activity (see Sato [16]).

3.4 Subordinators

A Lévy process which takes values on the real positive half line is called a subor-
dinator. This implies that the sample paths of a subordinator process are almost
surely increasing, haveR bounded variation, and in particular, the Lévy measure
satisfies the condition 0 (1 ∧ |x|) ν (dx) < ∞. This affords the important simpli-
∞

fication of being able to work with the Laplace transform as opposed to the Fourier
transform which follows from the bounded variation of the Lévy process and the
fact that Brownian motion has infinite variation. The latter point implies that the
gaussian part of a subordinator process must be zero, which allows the character-
istic exponent, 3.4, to be written in the following simplified form (see Bertoin [4],
p15 or Applebaum [1], p47)
Z ∞
Ψ (c) = icα + eicx − 1 ν (dx)


(3.6)
0

where α = α − R x1{|x<1|} ν (dx).

R

11
The infinite divisibility law of a random variable, Xt , implies that its Laplace
transform can be written in the following form
E e−uXt = e−tΦ(u)
 
(3.7)
for u ≥ 0 where Φ is called the Laplace exponent. It is now evident that by taking,
for example, c = iu, that 3.6 and the associated characteristic function can be
extended analytically on the complex upper half plane so that
φXt (c) = E e−uXt
 
(3.8)
whenever = (c) ≥ 0. Combining 3.7 and 3.8 yields that
Φ (c) = −Ψ (ui) (3.9)
which in turn provides an expression for the Laplace exponent.
Z ∞
1 − e−ux ν (dx)


Φ (c) = αc + (3.10)
0

for c ≥ 0.

The latter equality is often referred to in the literature as the Lévy-Khintchine

formula for subordinators. An important example of a subordinator, which will
be useful in the discussion in Section 4 of the Lévy process associated with the
Gumbel distribution, is the Gamma process. In the example below, the relation in
3.10 is used to ascertain the Lévy characteristics of the Gamma process.

Example 3.7 (Gamma Process)

The probability density function of Gamma distributed random variable may be

written as
θα α−1 −xθ
Gamma (x; α, θ) = x e ,x ≥ 0 (3.11)
Γ(α)
and the associated characteristic generating function is
φX (c) = (1 − ic/θ)−α . (3.12)
In reference to 3.9 it can be seen that the Laplace exponent is given by
ΦX (c; α, θ) = −Ψ (ic; α, θ)
= α log (1 + c/θ)
Z ∞
1 − e−cx αx−1 e−θx dx


=
0

where the latter equality follows directly from the Frullani Integral. It is now
apparent that the Gamma process is characterised by zero drift and a Lévy measure
ν (dx) = αx−1 e−θx dx. Note also the absence of a Gaussian term as expected.

12
4 Extreme Value Market Model
In this section the results collected so far are used to formulate a market model
in which asset returns are assumed to follow a Lévy process. The suitability of
the extreme value distributions as the marginal distribution of returns within the
context of this model are assessed. The Gumbel distribution is identified as a
potentially appropriate distribution for asset returns and some properties of the
associated Lévy process are explored.

4.1 Model Framework

It would be intuitively desirable for a model of the evolution of asset price returns
over time to restrict values of the asset price to the positive half line, but to
allow asset prices to rise or fall at any given point in time. These basic model
requirements suggest that the asset price process, S = {St , t ∈ [0, T ]}, may be
appropriately written as
St = S0 eXt . (4.1)
Here, the process X = {Xt , t ∈ [0, T ]} represents the evolution of asset price returns
and it is clear that X should have support in R. Standard no arbitrage results
require that the process S is a martingale which would imply that the following
condition must hold
St = EQ e−r(T −t) ST |FT
 
(4.2)
where r is the risk free rate of interest. It will now become apparent that, in terms
of financial modeling, it would be desirable if the distribution of Xt were infinitely
divisible which would in turn imply that the process X would be a Lévy process.
In this case, the martingale condition in 4.2 may be formulated in terms of moment
generating functions.

4.2 Martingale Condition

By taking t = 0 along with the usual convention that F0 = 1, 4.2 may be written
as

S0 = EQ e−rT ST
 

= S0 e−rT E eXT
 

13
where it has been assumed that the risk free rate of interest between times 0 and
T is known at time 0. The latter equality implies that

erT = E eXT
 

= E eX1
 T

= MX1 (1)T

where the second line follows from the infinite divisibility of Xt and the final line
from the definition of a moment generating function. It can now be seen that the
martingale condition, 4.2, can be expressed as

r = log [MX1 (1)] . (4.3)

This form of the martingale condition will be useful in Section 5 in deriving an

equivalent martingale measure for the asset price process. It is noted that that the
moment generating function of a random variable need not necessarily exist.

4.3 Gumbel Infinite Divisibility

Of the three types of extreme value distributions outlined in Section 2, only the
Gumbel (Type I), 2.3, and Weibull (Type III), 2.5, distributions are known to
be infinitely divisible, and, in the latter case infinitely divisibility is restricted to
0 < α ≤ 1. It appears that it is not known whether the Fréchet (Type II) distri-
bution is infinitely divisible (see Sato [16],p46 and Steutel [19]). In addition it is
noted that the support of the Fréchet the Weibull distributions are truncated on
the negative and positive real axis respectively, and it is therefore not immediately
apparent that they represent appropriate distributions for asset returns in the con-
text of 4.1. Therefore, the focus for the rest of this section will be the infinite
divisibility and Lévy process associated with the Gumbel distribution.

In order to explore the Lévy process associated with the Gumbel distribution,
it will be instructive to derive the associated characteristic function. It is readily
seen from the form of the cumulative distribution function in 2.3 that the Gumbel
probability density function with parameters α and σ is given by
1 − (x−α)  
F (x; α, σ) = e exp −e
(x−α)
− σ
σ (4.4)
σ

14
for x ∈ R. The characteristic function is then given by

φX (c) = E[eciX ]
1 ∞ cix − (x−α)
Z  
e e σ exp −e− σ dx
(x−α)
=
σ −∞
1 0 cix −u σ
Z
= − e ue ( )du
σ u
Z ∞ ∞
= ecix e−u du
0
Z ∞
−αci
= e u−σci e−u du
0

(x−α)
where in the third line the substitution u := e− σ was used. The characteristic
function of a Gumbel distributed random variable, X, is then given by

φX (c) = eαci Γ(1 − σci) (4.5)

for some c ∈ R and where Γ (α) is the Euler Gamma function.

In reference to 3.3, it is not immediately apparent from the form of the char-
acteristic function in 4.5 that the Gumbel distribution is indeed infinitely divisible.
However, a Gumbel distributed random variable may be represented in the form of
a limit sum of Exponentially distributed random variables, a representation which
facilitates the validation of the infinite divisibility property. This fact is recorded
in the following proposition (see Steutel [19]).

Proposition 4.1 A random variable, X, which is Gumbel distributed with param-

eters α and β may be represented as
( n )
d
X Yk
X = α + lim − log (n) (4.6)
n→∞
k=1
k
where the Yk are independent and identically Exponentially distributed random vari-
ables with mean β.

A proof of proposition 4.1 is instructive in terms of exploring the Lévy process as-
sociated with the Gumbel distribution and proceeds by demonstrating the equiv-
alence using characteristic functions. The characteristic function of the random

15
variable on the right hand side of equation 4.6 may be established as follows
" ( n )!#
E[eciX ] = E exp ciα + ci lim
X Yk
− log (n)
n→∞
k=1
k
( " n
!#)
= lim eci(α−log(n)) E exp ci
X Yk
n→∞
k=1
k
( n )
c iα −ci log(n)
 
Y1
= e lim e E exp ci
Y
n→∞
k=1
k
( n
)
c iα −ci log(n) 1
= e lim e
Y
n→∞
k=1
1 − cikβ
∞ ci
e− k
= eciα e−ciγ
Y

k=1
1 − cikβ
= eciα Γ (1 − ciβ)
where γ is Euler’s constant. In the second line the strict monotonicity of the expo-
nent function was used. In the third line it was used that the Yi are independent
and identically distributed and in the fourth line that the characteristic function
of an Exponential random variable, Y , with mean β is given by
1
φY (c) = .
1 − c iβ
The final line follows directly from the Weierstraß identity for the Gamma function
(∞ z
)
1 −zγ Y e k
Γ (z) = e z (4.7)
z k=1
1 + k

and that fact that Γ (1 + z) = zΓ (z). In view of 4.5 the equivalence has been
established.

With reference to a theorem by Feller [10](p585), it can be shown that if the

distribution of 4.6 tends to a limit distribution, and the Yk tend in probability to
zero uniformly, then the limit distribution of 4.6 is infinitely divisibility. This latter
condition will be satisfied if, for all k sufficiently large,
|1 − φYk (ζ)| < 
given  > 0 and ζ0 > 0 and for |ζ| < ζ0 . Given the form of the characteristic
function for Yk which is
φYk (c) = (1 − cβ i/k)−1 (4.8)
this is clearly the case, which establishes that the Gumbel distribution is indeed
infinitely divisible.

16
4.4 Gumbel Lévy Process

An interesting corollary to proposition 4.1 is that a process, {Xt , t ∈ [0, T ]}, which
is Gumbel distributed may be represented by by a limit process of the form
( n )
d
X Zk
Xt = αt + lim − log (n) t . (4.9)
n→∞
k=1
k

where the Zi are now independently and identically distributed Gamma distributed
random variables with with parameters t and β. This can be easily verified in
an analogous manner to proposition 4.1 by noting that the form of the Gamma
characteristic function in 3.12 is that of the Exponential characteristic function in
4.3 raised to a power, say t. It directly follows that the characteristic function of
the process in 4.9 is given by

φXt (c) = eciαt Γ (1 − ciβ)t (4.10)

which, together with the infinite divisibility property of the Gumbel distribution,
which has already been demonstrated, and the form of the Gumbel characteristic
function in 4.5, by equivalence of characteristic functions, establishes the claim.

Given the form of the Gumbel characteristic function in 4.5, the pathwise properties
of the Lévy process associated with the Gumbel distribution are not immediately
analytically tractable, as would typically be the case for distributions with expo-
nentiated moments. (Two such cases of distributions with exponentiated moments
are the Normal and Poisson distributions as shown in examples 3.4 and 3.6 respec-
tively.) The limit process in 4.9 is therefore interesting in the sense that, for large
enough n, a process, {Xt , t ∈ [0, T ]}, of the form

Xt = (α − log (n)) t + Yt (4.11)

Pn Zk
where Yt = k=1 k and the Zk are defined as in relation to 4.9, will approxi-
mate the behavior of the Gumbel Lévy process and would therefore be potentially
useful in terms of pathwise simulations of the Gumbel distributed random variable.

Since the Gamma distribution is defined on the positive half line, it follows from
the discussion in Section 3.4 that the process {Yt , t ∈ [0, T ]} in equation 4.11 is a
subordinator, and that therefore, the form of the associated characteristic exponent
will be given by 3.10. In addition, the exact form of the characteristic exponent of
the Gamma process has already been seen in example 3.7, from which the charac-
teristic exponent of the approximating Gumbel process in 4.11, for t = 1, can be

17
written directly as
n
X
Φx (c) = (α − log (n)) + log (1 + cβ/k)
k=1
n Z ∞ 
1−e x e dx
X
−cx −1 −βx/k


= (α − log (n)) +
k=1 0
Z ∞ n
1−e e−βx/k dx
X
−cx −1

 
= (α − log (n)) + x
0 k=1

where line 2 again follows from the Frullani Integral and the interchange of summa-
tion and integration is justified in the last line by the monotonicity of the exponent
function.

The form of this last equality can be recognised as the Laplace exponent in 3.7
and it follows that the Lévy
Ptriplet of the
 approximating Gumbel process in 4.11
is given by α − log (n) , 0, k=1 e
n
dx . As expected, there is no Gaussian
 −βx/k

component and as n increases the drift term becomes increasingly negative and
tends to negative infinity at the limit. Since the process is a subordinator, there
are no negative jumps and the process has finite variation. Plainly, as n increases,
the jump term tends to positive infinity. However, due to the equivalence in dis-
tribution of the approximating Gumbel process in 4.11 to the Gumbel process at
the limit, and in view of 4.5, it is evident that the approximating process in 4.11
converges almost surely to a finite limit.

18
5 Gumbel European Option Pricing
In this section analytic solutions for European put and call options are developed
for an equivalent martingale measure under the assumption that asset price losses,
defined as negative returns, are Gumbel distributed.

5.1 Model Outline

The framework for the model developed in this section was outlined in Section 4.1
with the asset price process satisfying 4.1. Asset returns are assumed to follow a
Lévy process and suitable return distributions should be well defined on the entire
real line. This in turn ensures that the asset price process can rise or fall but will
never fall below zero. It is noted that the Gumbel distribution has support in R and
that it has been shown to be infinitely divisible and by implication associated with
a unique Lévy process. The Gumbel distribution is therefore a suitable candidate
for the distribution of returns in 4.1. Throughout this section it is assumed that the
risk free interest, r, and dividend yield, q, are constant and that option contracts
mature at time T with a strike equal to K. Only European plain vanilla put and
call option contracts will be considered.

5.2 Distribution of Returns

In line with the literature, and following Markose and Alentorn [14], it is assumed
that negative asset returns or losses are Gumbel distributed. Intuitively this means
that the fattest tail associated with the Gumbel distribution occurs on the negative
axis. In terms of the asset return process defined in 4.1, this means that Yt = −Xt
is Gumbel distributed.

Given the form of the Gumbel probability density function in 4.4, the distribu-

19
tion of returns, Xt , can be easily derived as follows.

P(Xt ≤ x) = P(−Yt ≤ x)
= P(Yt > −x)
Z ∞
1 α−y  α−y 
= e σ exp −e σ dy
−x σ
Z −x
1 α−y  α−y 
= − e σ exp −e σ dy
σ
Z∞x
1 α+u  α+u 
= e σ exp −e σ du
−∞ σ

where in the penultimate line the substation u := −y was used. It follows directly
that the density function of returns, Xt , is given by
1 α+x  α+x 
fXt (x) = e σ exp −e σ (5.1)
σ
for x ∈ R. A calculation analogous to that which was used to derive the Gumbel
characteristic function, 4.5, implies that the moment generating function associated
with the probability density function in 5.1 is given by

MXt (c) = e−αc Γ(1 + σc). (5.2)

It is also apparent from definition 3.3 that the distribution of returns in 5.1 is also
infinitely divisible since

φXt (c) = φYt (−c)

= φY1 (−c)t
= φX1 (c)t

and therefore uniquely associated with a Lévy process.

5.3 Generic European Call Option Pricing Formula

Standard no arbitrage arguments allow the price of a European call option under
the stated assumptions to be written as follows

C0 = e−(r−q)T EQ [(ST − K)+ ]

= e−(r−q)T EQ [(S0 eXT − K)+ ]

20
where the form of the asset price process, 4.1, was used in the second equality, and
the expectation is subject to the risk neutral measure, Q. It will be convenient to
write the latter equation in the following form
Z ∞
C0 = e −(r−q)T
(S0 ex − K)+ fXQT (x)dx
Z−∞
∞
= e−(r−q)T (S0 ex − K)fXQT (x)dx
κ

where κ := ln(K/S0 ) and fXQT is the risk neutral probability density function
associated with XT . The price of a call option can then be written as
Z ∞ Z ∞
C0 = e −(r−q)T
(S0 e fXT (x)dx − K
x Q
fXQT (x)dx). (5.3)
κ κ

The integrals in 5.3 can now be evaluated under the assumption that XT is dis-
tributed according to 5.1. It should be noted that in this context, the parameters
of the probability density functions in subsequent calculations relate to the dis-
tribution of losses at maturity and not for all t. The first integral in 5.3 may be
simplified as follows.
Z ∞
1 ∞ x α+x
Z
e fXT (x)dx = e e σ exp(−e σ )dx
α+x
x
κ σ κ
1 ∞ x −u σ
Z
= e ue ( )du
σ e α+κ
σ u
Z ∞
= α+κ ex e−u du
e σZ
∞
=e −α
uσ e−u du
e
α+κ
σ

where in the second line the substitution u := e σ was used. The second integral
α+x

may be simplified in an analogous manner.

Z ∞
1 ∞ α+x −e α+x
Z
fXT (x)dx = e σ e σ dx
κ σ
Zκ
1 ∞ σ
= ue−u ( )du
Z ∞e σ
σ α+κ
u
= α+κ u0 e−u du
e σ

21
Where in the second line an identical substitution as above was used. Combining
results, the price of a European call option under the assumption that losses are
Gumbel distributed at maturity may be written as
 Z ∞ Z ∞ 
C0 = e −(r−q)T
S0 e−α
u e du + K α+κ u e du .
σ −u 0 −u
(5.4)
eα+κ
σ e σ

5.4 No Arbitrage Condition

The final step is to ensure that the martingale condition, 4.2, is satisfied. In
reference to Section 4.2, it can be seen that, after including the dividend yield, in
order for the martingale condition 4.2 to be satisfied, it is sufficient for the following
relationship to hold.
e(r−q)T = MXT (1)
In reference to the moment generating function of the return distribution in 5.2, it
can be seen that MXT (1) = e−α Γ(1 + σ) which implies that the martingale condi-
tion, 4.2, will be satisfied at maturity, if the parameters of the Gumbel distribution
at maturity satisfy the following constraint, which is effectively a no arbitrage con-
dition,
e(r−q)T
e−α = (5.5)
Γ(1 + σ)
where it is noted that Γ (z) is well defined for < (z) > 0. The validity of this
statement can be easily checked by calculating 4.2 under the constraint of 5.5 as
follows
EQ e−(r−q)T ST |F0 = −e(r−q)T S0 EQ eXT
   
Z ∞
=e −(r−q)T
S0 e −α
uσ e−u du
0
Z ∞
S0
= uσ e−u du
Γ (1 + σ) 0
= S0

where again the substitution u := e σ was used in the second line. The last
α+x

equality follows by recognising the kernel of a Gamma distribution with parameters

unity and σ on the right hand side of the penultimate line.

5.5 European Call Option Pricing Equation

Collecting results, an equation for a European call option which satisfies the mar-
tingale condition 4.2, under the assumption that asset price losses are Gumbel

22
distributed, may now be derived. The first step is to substitute the no arbitrage
condition 5.5 into equation 5.4.

e(r−q)T
Z ∞
C0 = e −(r−q)T
(S0 u−σ e−u du
Γ(1 + σ) e log(Γ(1−σ))+κ−(r−q)T
σ
Z ∞
− K log(Γ(1−σ))+κ−(r−q)T u0 e−u du)
e Z ∞σ
Z ∞
S0 −(r−q)T K
= u e du − e
−σ −u
u0 e−u du
Γ(1 + σ) ξ Γ(1) ξ

where ξ := Γ(1 + σ) σ e σ . Finally, the integrands on the right hand side of the
1 κ−(r−q)T

last equality can again be recognised as kernels of the Gamma distribution which
allows the price of a European call option to be written in the following simplified
form.

C0 = S0 [1 − Gamma (ξ; 1 + σ, 1)] − K e−(r−q)T [1 − Gamma (ξ; 1, 1)] (5.6)

5.6 European Put Option Pricing Equation

The price of a European put option may be derived in an entirely analogous man-
ner. However, since the calculations relating to the put option are essentially the
same as those for the call option, they are omitted here but included in Appendix
A for reference. The price of a put option under the same stated assumptions as
those for the call option above is

P0 = K e−(r−q)T Gamma (ξ; 1, 1) − S0 Gamma (ξ; 1 + σ, 1) (5.7)

where ξ := Γ(1 + σ) σ e
1 κ−(r−q)T
σ .

23
6 Numerical Analysis

6.1 Outline

This section presents results of applying the analytical equations for the prices of
European style call and put options derived in Section 5. Analysis was conducted
using call and put prices for options on the S&P500 US Stock index for a period
covering the first half of 2006. The analysis carried out was similar in nature
to that which was carried out by Markose and Alenthorn [14], with the notable
exception that the forms of the analytical solutions in 5.6 and 5.7 permitted the
use of estimated dividend yields in the analysis of the option prices.

6.2 Data and Software

The sample period for the study was January 1st 2006 until June 30th 2006. Daily
closing option prices for both calls and puts on the S&P 500 Index quoted on
the Chicago Board Options Exchange (CBOE) were sampled for a wide range of
strikes. The strikes for each day were obtained within a range of 0.94 to 1.06 times
the spot price for that day in order to capture options covering a broad spectrum of
moneyness, defined as the spot price divided by the strike (see Section 6.3 below).
In order to ensure that the sampled prices had traded, only strikes which recorded
non-zero volume were included in the sampled data.

Options on the S&P500 index are available on the CBOE for expiry on a fixed
date each month. Strikes are available at intervals of $5. Prices were sampled
for each contract in order to provide a wide range of option maturities within the
sample. The option contracts, their associated expiry dates and the number of
active days for each contract is detailed in Table 1 below. In total 34,488 option
contracts of differing strikes and maturities were sampled over the entire period.
There were no options traded on the October and November contracts during the
sample period.

Daily closing spot prices for the S&P500 index were obtained as the spot ref-
erence for the sampled option contracts. US dollar LIBOR quotes were used as a
proxy for the risk free rate in calculations. In order to be able to match maturity
funding rates exactly, US dollar LIBOR curves covering a sample of maturities
from 1 week to 12 months were obtained for each day in the sample period.

Daily historical data for dividends on the S&P500 index was not available and
was therefore imputed from the prices of futures contracts on the S&P500 index.

24
 Table 1: Active Trading Days per Contract by Strike Type

Expiry Date Puts Calls Total

January 12 12 24
February 32 32 2,64
March 51 51 4,102
April 62 62 124
May 62 62 124
June 113 113 226
July 49 49 98
August 58 58 116
September 120 120 240
December 123 123 250

Futures on the S&P500 index are quoted for quarterly expiry. Daily prices were
obtained for the futures contracts detailed in Table 2. The expiry dates for futures
contracts covered a wider period that that of the sampled option price data in or-
der to ensure that estimates for dividends for all options trading within the sample
period would be available.

Table 2: S&P500 Futures Contracts

Contract Expiry Date

December 2005 12 December 2005
March 2006 17 March 2006
June 2006 16 June 2006
September 2006 15 September 2006
December 2006 15 December 2006

The majority of the calculations were conducted using Mathematica Students Edi-
tion Version 5.1 [20]. Data was stored and manipulated using MySQL Version 5.0
[15], a freeware database package. The database was accessed from Mathematica
using the MySQL ODBC Driver v3.51.

25
6.3 Methodology

Option contracts were categorised in two ways essentially following the framework
used by Markose & Alenthorn [14]. In terms of moneyness, spot price divided by
strike, call options were categorised as in the money (ITM) if moneyness exceed
1.03, out of the money (OTM) if moneyness was less than 0.97 and at the money
(ATM) if moneyness fell between 0.97 and 1.03. Puts were correspondingly cate-
gorised as OTM, ITM and ATM respectively. Options were also categorised into
broad maturity ranges based on the days left until expiry. Tables 3 and 4 below
summarise the data in the context of these categorisations.

Table 3: Number of Strikes Sampled by Moneyness

Moneyness ATM ITM OTM Total

Calls
January 1,672 436 652 2,760
February 1,392 320 714 2,426
March 1,802 344 844 2,990
April 1,516 246 724 2,486
May 1,694 280 964 2,938
June 1,952 254 988 3,194
Total Calls 10,028 1,880 4,886 16,794
Puts
January 1,590 172 838 2,600
February 1,444 224 792 2,460
March 1,984 262 904 3,150
April 1,610 184 790 2,584
May 1,880 392 942 3,214
June 2,206 526 954 3,686
Total Puts 10,714 1,760 5,220 17,694
Total 20,742 3,640 10,106 34,488

For each option contract maturity on each day, the risk free rate was calculated
using the LIBOR curve for that day. The rate for the exact maturity of each op-
tion contract was calculated by fitting a curve to the underlying LIBOR data using
Mathematica’s Interpolation function. The fitted curve was then used to calculate
the LIBOR rate at a given maturity which was then converted to a continuously
compounded rate.

26
 Table 4: Number of Strikes Sampled by Maturity

Maturity (Days) > 30 30-60 60-90 90-180 180+ Total

Calls
January 1,040 938 286 214 282 2,760
February 1,392 320 714 2,426
March 960 1,012 430 288 300 2,990
April 962 894 318 148 164 2,486
May 1,108 838 496 252 244 2,938
June 882 1,230 602 334 146 3,194
Total Calls 5,912 5,630 2,426 1,428 1,398 16,794
Puts
January 962 802 310 198 328 2600
February 982 680 268 186 344 2,460
March 964 1,032 438 314 402 3,150
April 980 882 304 198 220 2584
May 1,224 846 532 284 328 3,214
June 970 1,424 686 396 210 3,686
Total Puts 6,082 5,666 2,538 1,576 1,832 17,694
Total 11,994 11,296 4,964 3,004 3,230 34,488

Dividends were calculated using a two-step process. The essential methodologi-

cal approach was to assume that the dividend yield, q, was known and that, in
the absence of arbitrage, the condition in 6.1, at time t, between the price of a
future expiring at time T, Ft , and the price of the underlying asset, St , would hold.
Naturally, r is taken to be the risk free rate of interest. For a discussion of as to
why this condition must hold, the reader is referred to Hull [11].

Ft = e(r−q)(T −t) St (6.1)

The first step was then to calculate the continuously compounded dividend yield
to the expiry date of each futures contract for each day in the sample, using the
closing price of the index, which was implied by equation 6.1. The risk free rate
to the expiry date of the futures contract was calculated from the LIBOR curve as
outlined above. This calculation effectively provided a continuously compounded
dividend yield curve for each day in the sample. The second step was to calculate
the implied dividend yield on each day for each option contract maturity. This was
achieved in an analogous manner to the risk free rate at maturity by interpolating
a dividend curve for each day in the sample and calculating the implied continu-
ously compounded dividend yield for each option contract maturity.

27
The obvious drawback from using this approach to calculating dividends is that
it is assumed that the both the risk free rate and the dividend yield are known.
While neither is known at the time at which each option contract is traded, approx-
imating the risk free rate with the LIBOR curve is a commonly accepted practice.
However, the dividend yield would clearly not have been known and should there-
fore be subject to a degree of uncertainty. In this case, the relation in 6.1 would
be expressed as an expectation as opposed to a deterministic relation. Any bias in
the calculation should, however, be limited by the relatively short time period over
which dividend yields were being approximated and by the broad nature of the
index which incorporates 500 securities. Moreover, the results of the calculations,
which are summarised in Table 5, appear consistent with monthly realised market
data. This approach has the advantage over using actual historical dividends by,
in some sense, capturing the market’s view of what dividends would be at the time
at which each option was traded.

Table 5: Calculated and Historical Dividend Yield by Month

Month Calculated Historical

January 1.88 1.26
February 2.06 2.88
March 1.65 1.77
April 1.85 1.60
May 2.27 2.65
June 1.89 1.58
Average 1.93 1.87

The essential analysis undertaken in the study again loosely follows that of Markose
and Alentorn [14] which was to calculate the pricing performance of the Gumbel
pricing model and to compare it to that of the Black-Scholes model. For each
day in the sample, and for a fixed maturity, t, the volatility parameter was cal-
culated which minimised the sum of squared differences, 6.2, over all strikes, N,
between the sampled quoted option prices, C, and the analytic solutions, C, of the
Gumbel pricing equations given in 5.6 and 5.7, and separately the corresponding
Black-Scholes equations.
" N #
X
2
SSEt = min C (t, Ki ) − C (t, Ki ) (6.2)
σ
i=1
It should be noted that the Gumbel pricing equations of Section 5 have been
derived subject to a risk neutral density ensuring that the martingale condition

28
4.2 is satisfied. The optimisation was conducted primarily using Mathematica’s
NonlinearRegress function using the gradient model on maturities for which there
were at least two traded strikes. However, in some cases it was more convenient
to locate the root using a simple bisection method (see Section 6.5). The Black-
Scholes option price calculations were facilitated by using the functions in the
package provided in Shaw [18]. The Gumbel option pricing equations were encoded
in an analogous manner.

6.4 Results

The results of the non-linear regression analysis are summarised in Table √ 6 which
shows the number of root mean squared errors (RMSE), defined as N1 SSE, mea-
sured in US dollars by broad ranges for each model type. As can be seen, a large
number of RMSEs exceeded $5. On examination of the data it was apparent that
the very large errors occurred on days when there appeared to be significant dis-
crepancies in the strikes in the sense that, for example, lower strikes were associated
with substantially lower premiums for call options. The cause of these discrepan-
cies is unknown but may potentially result from either simple errors in the quote,
or, from the fact that, even though the closing price was sampled, in actuality, the
quoted premiums relate to the last traded price. On days where there is signifi-
cant market movement, the latter circumstance could easily result in discrepancies.
Independent calculations which used real-time quotations generated substantially
reduced errors.

Table 6: Number of Root Mean Square Errors by Size

RMSE < 0.5 0.5-1.0 1.0-2.5 2.5-5.0 5.0+ Total

Black-Scholes
Calls 8 42 298 300 34 682
Puts 4 21 265 351 46 687
Total Black-Scholes 12 63 563 651 80 1,369
Gumbel
Calls 57 214 323 83 5 682
Puts 63 178 348 92 6 687
Total Gumbel 120 392 671 175 11 1,369
Total 132 455 1,234 826 91 2,738

29
The results relating to the relative performance of the Gumbel model are nonethe-
less encouraging. Table 7 reports average RMSE by maturity range for call options.
It can be seen that the average RMSEs for the Gumbel call model are lower than
those for the Black-Scholes call model for each month and for every maturity range.
Overall, the average RMSE of the Gumbel call model was $1.47 compared to $2.66
for the Black-Scholes call model. Interestingly, the relative performance of the
Gumbel call model improved with the maturity of the option. For options ex-
ceeding 180 days in maturity, the average RMSE for the Gumbel call model was
$1.47, less than half of that of the Black-Scholes call model, which was $3.43 for
the same maturities. It is notable that the performance of the Gumbel model did
not deteriorate for longer maturities.

Table 7: Average Root Mean Square Errors for Call Prices ($)

Maturity(Days) < 30 30-60 60-90 90-180 180+ Total

Black-Scholes
January 1.53 1.77 2.21 3.52 3.47 2.64
February 1.58 1.74 2.22 2.66 3.44 2.49
March 1.44 2.04 2.12 2.81 3.19 2.42
April 1.71 1.73 2.15 2.93 3.49 2.39
May 1.98 2.30 2.78 3.17 3.57 2.77
June 2.65 2.78 3.43 3.78 3.57 3.21
Total 1.80 2.13 2.62 3.15 3.43 2.66
Gumbel
January 1.33 1.22 1.15 1.18 1.30 1.25
February 1.42 1.11 1.26 1.19 1.16 1.22
March 1.31 1.44 1.27 1.30 1.50 1.38
April 1.48 1.29 1.20 1.65 1.94 1.52
May 1.54 1.39 1.36 1.67 1.49 1.48
June 1.98 1.53 2.20 2.12 2.02 1.95
Total 1.50 1.36 1.50 1.53 1.47 1.47

Table 8 reports RMSEs by maturity range for put options. The pattern of the
results is similar to that which was observed for call options. The Gumbel put
model exhibited lower RMSE in all time periods and for all maturities with an
overall RMSE of $1.55 compared to that of $2.86 for the Black-Scholes put model.
Reflecting the results for call options, it is interesting to note that the performance
of the Gumbel put model was similar for all maturities, whereas the performance
of the Black-Scholes put model deteriorated markedly with contract maturity. For

30
 Table 8: Average Root Mean Square Errors For Put Prices ($)

Maturity(Days) < 30 30-60 60-90 90-180 180+ Total

Black-Scholes
January 1.25 1.85 2.41 3.28 3.18 2.49
February 1.66 2.10 2.73 3.02 3.19 2.65
March 1.41 2.17 2.23 2.71 3.22 2.46
April 1.30 1.93 2.23 3.26 3.11 2.35
May 2.35 2.79 3.39 3.79 3.53 3.19
June 3.24 3.64 3.87 4.33 5.07 3.92
Total 1.84 2.52 2.98 3.41 3.37 2.86
Gumbel
January 0.96 1.05 1.01 1.28 1.43 1.19
February 1.42 1.27 1.29 1.30 1.33 1.33
March 1.19 1.30 1.14 1.21 1.17 1.20
April 0.99 1.06 1.11 1.51 1.09 1.15
May 1.99 1.73 1.94 1.80 1.61 1.83
June 2.80 2.19 2.47 2.58 3.20 2.53
Total 1.53 1.50 1.65 1.63 1.46 1.55

example, the average RMSE for all time periods for the Black-Scholes put model
for options with less than 30 days to expiry was $1.84 compared to an RMSE of
$3.37 for options with 180 days or more to expiry. The corresponding RMSEs for
the Gumbel model were $1.53 and $1.46 respectively.

The lower RMSEs of the Gumbel models naturally translate into a lower pric-
ing bias, which is defined simply as the difference between the quoted option price
and the price predicted by each model based on the fitted volatility parameters.
Table 9 shows the pricing bias for call options by maturity of the contracts. It can
be seen that the pricing bias of the Gumbel model for both call and put options
was less than that of the Black-Scholes model for all time periods and all maturi-
ties. Overall, the average pricing bias for the Black-Scholes call model was $1.66
compared to $0.99 for that of the Gumbel call model. Similarly, the pricing bias of
the Gumbel put model was also superior to that of the Black-Scholes put model.
The overall average pricing bias for the Gumbel put model was $1.11, almost half
that of the corresponding pricing bias for the Black-Scholes put model which was
$2.02. Again, it is interesting to note that the pricing bias of the Gumbel model
was more stable than that of the Black-Scholes model as maturity increased.

31
 Table 9: Average Pricing Bias by Maturity ($)

Maturity(Days) < 30 30-60 60-90 90-180 180+ Total

Calls
Black-Scholes 1.13 1.63 1.98 2.48 2.65 1.66
Gumbel 0.87 0.93 1.12 1.24 1.19 0.99
Puts
Black-Scholes 1.36 2.10 2.39 2.77 2.80 2.02
Gumbel 0.91 1.11 1.33 1.36 1.28 1.11

Finally, Table 10 shows pricing bias by moneyness. It can be seen than both the
put and call Gumbel models exhibited lower pricing bias than the Black-Scholes
model for each moneyness type. The pricing bias of both models was better for
both ATM and OTM strikes than was the case for ITM strikes. For example,
the pricing bias of Gumbel calls for ATM and OTM strikes was $0.95 and $0.56
respectively whereas the pricing bias for ITM Gumbel calls was $2.28.

Table 10: Average Pricing Bias by Moneyness ($)

Moneyness ATM ITM OTM Total

Calls
Black-Scholes 1.39 3.52 1.50 1.66
Gumbel 0.95 2.28 0.56 0.99
Puts
Black-Scholes 1.55 4.07 2.28 2.02
Gumbel 0.99 2.88 0.77 1.11

6.5 Discussion of Results

The results are in general in agreement with those found by Markose and Alentorn
[14]. The Gumbel model exhibits lower RMSE and improved pricing performance
relative to the Black-Scholes model for both puts and calls over a wide range of
different strikes and for different levels of moneyness.

32
The size of the RMSEs are, however, significantly higher that those which were
reported by Markose and Alentorn. Part of the reason may be related to the fact
that Markose and Alentorn assumed that the distribution of returns at maturity
was drawn from the Generalised Extreme Value distribution, allowing the implied
tail index to be estimated from the data. In particular, this would allow for fatter
tails. The Gumbel model effectively forces the tail index parameter to be zero.
However, as the discussion of Section 4 indicated, from the perspective of model-
ing log returns as a stochastic process, the Gumbel distribution is a more natural
choice.

While the form of the distribution may have been a relevant factor in relation
to the higher level of RMSE, data issues appear likely to have played a key role. In
addition to the comments in Section 6.4 above, it is worth noting that convergence
of the Gumbel model, in terms of locating the minimum SSE, was noticeably slower
in some cases than was the case for the Black-Scholes model. On visual examina-
tion it was apparent these cases typically related to days on which there appeared
to be some inconsistencies in the data. It is also notable that that the SSE was far
more sensitive to changes in the target volatility in the case of the Gumbel model
than was the case for the Black-Scholes model. These facts combined are the likely
cause of the slow convergence in the case of the Gumbel model, and may, at the
margin, have contributed to larger SSEs than would have otherwise been the case.

33
7 Conclusion
Given the well documented evidence of the existence of fat tails in empirical studies
of asset return distributions, it is perhaps not surprising that an option model
based on an assumed distribution of asset returns with even moderately fat tails,
such as the Gumbel distribution, should outperform the pricing performance of
the Black-Scholes model. Nevertheless, the numerical results presented in Section
6 of applying the analytic solutions for European put and call options developed
in Section 5, were nonetheless encouraging. It was also shown in Section 4 that
the Gumbel distribution is infinitely divisible, which, together with the fact that
the support of the Gumbel distribution is the entire real axis, allows the Gumbel
distribution to be used as basis for a model of log of asset returns where the
underlying stochastic dynamics are a well defined Lévy process. It was also shown
in Section 4 that, while the pathwise properties of the unique Levy Process which
is associated with the Gumbel distribution remain analytically elusive, the process
can nevertheless be approximated by an appropriate form of a large enough sum
of independent and identically distributed Gamma distributed random variables.
Simulating this latter process could be potentially useful in analysing problems
which require pathwise properties, such as models related to default.

34
References
[1] Applebaum, D., (1999), ”Lévy Processes and Stochastic Calculus”, CUP,
ISBN:0521553024.

[2] Bacmann, J. and G. Gawron, (2004), ”Fat Tail Risk in Portfolios of Hedge
Funds and Traditional Investments”, RMF Research Paper.

[3] Bakshi, G. and D. Madan, (2000), ”Spanning and derivative-security valua-

tion”, Journal of Financial Economics, 55, 205-238.

[4] Bertoin, J., (1996), ”Lev́y Processes”, CUP, ISBN:0521646324.

[5] Carr, P., H. Geman, D. Madan, and M. Yor, (2003), ”Stochastic volatility for
Lévy processes”, Mathematical Finance, 13, 345-382.

[6] Carr, P., H. Geman, D. Madan, and M. Yor, (2003), ”The fine structure of
asset returns: an emprical investigation”, Journal of Business, 75, 305-332,
13, 345-382.

[7] Carr, P. and D. Madan, (1998), ”Option valuation using the fast fourier trans-
form”, Journal of Computational Finance, 2, 61-73.

[8] Coles, C., (2001), ”An Introduction to Statistical Modeling of Extreme Val-
ues”, Springer, ISBN:1852334592.

[9] Embrechts, P., C. Klüppelberg, C. T. Mikosch, (1997) ”Modelling extremal

events for insurance and finance”, Springer-Verlag, ISBN:3540609318.

[10] Feller, W., (1971), ”An Introduction to Probability Theory and Its Applica-
tions”, Vol II, 2nd Ed., Wiley, New York, ISBN:0471257095.

[11] Hull, J.C., (2006), ”Options, Futures and Other Derivatives”, Prentice Hall,
ISBN:0131499084.

[12] Hull, J.C and A. White, (1988), ”The pricing of options on assets with stochas-
tic volatility”, Jounal of Finance, 42, 281-300.

[13] LeBaron, B. and R. Samanta, (2004), ”Extreme value theory and fat tails in
equity markets”, SSRN Working Paper Series.

[14] Markose, S. and A. Alentorn, (2005), ”The generalized extreme value (GEV)
distribution, implied tail index and option pricing”, University of Essex, Eco-
nomics Discussion Papers, 954.

[15] MySQL 5.0 - http://www.MySQL.org.

35
[16] Sato, K., (2004), ”Lévy Processes and Infinitely Divisible Distributions”. CUP,
ISBN:0521832632.

[17] Schoutens, W., (2003), ”Lévy Processes in Finance: Pricing Financial Deriva-
tives”, Wiley Series in Probability and Statistics, ISBN:0470851562.

[18] Shaw, W., (2002), ”Modelling Financial Derivatives with Mathematica”, CUP,
ISBN:052159233X.

[19] Steutel, F.W., (1973), ”Some recent results in infinite divisibility”, Stochastic
Processes and their Applications, 1, 125-143 47.

[20] Wolfram Research, Inc. (2004), ”Mathematica”, Version 5.1, Champaign, IL.

36
Appendix A
In this appendix an analytic solution for a standard European put option is de-
rived under the assumption that asset price losses, defined as negative returns, are
Gumbel distributed. The underlying assumptions are the same as those of section
5. The price of a European put option may be written as follows.

P0 = EQ [e−(r−q)T (K − ST )+ ]
= e−(r−q)T EQ [(K − ST )+ ]
= e−(r−q)T EQ [(K − S0 eXT )+ ]
Z ∞
=e −(r−q)T
(K − S0 ex )+ fXQT (x)dx
Z−∞
κ
=e −(r−q)T
(K − S0 ex )fXQT (x)dx
−∞
Z κ Z κ
=e −(r−q)T
(K fXT (x)dx − S0
Q
ex fXQT (x)dx)
−∞ ∞

where κ := ln(K/S0 ) and fXQT is the risk neutral probability density function of XT .
Under the assumption that the probability density function of XT is distributed as
5.1 and that the no arbitrage condition, 5.5, holds, the integrals in the last equality
may be evaluated as follows
Z e α+κ
σ Z e α+κ
σ

P0 = e−(r−q)T (K u0 e−u du − S0 e−α u−σ e−u du)

∞ ∞
Z e log(Γ(1−σ))+κ−(r−q)T
σ

= e−(r−q)T (K u0 e−u du
−∞
Z e log(Γ(1−σ))+κ−(r−q)T
e(r−q)T σ

− S0 u−σ e−u du)

Γ(1 + σ) −∞
e−(r−q)T K ξ ξ
Z Z
S0
P0 = ue
0 −u
du − uσ e−u du
Γ(1) −∞ Γ(1 + σ) −∞

where ξ := Γ(1 + σ) σ e σ . The latter equality may be written more succinctly

1 κ−(r−q)T

as
P0 = K e−(r−q)T Gamma (ξ; 1, 1) − S0 Gamma (ξ; 1 + σ, 1) .

37