Pricing Asian Options by the

Method of Moments Matching

by

Pak Keung Chan

A thesis
presented to the University of Waterloo
in fulfillment of the
thesis requirement for the degree of
Master of Quantitative Finance
in
Mathematics

Waterloo, Ontario, Canada, 2015

c Pak Keung Chan 2015

I hereby declare that I am the sole author of this thesis. This is a true copy of the thesis,
including any required final revisions, as accepted by my examiners.
I understand that my thesis may be made electronically available to the public.

ii

Abstract
This Masters Thesis explores the method of moments matching for pricing Asian options. In this thesis, the underlying asset is assumed to be non-dividend paying and its
price process either follows the standard geometric Brownian motion (GBM) or the more
advanced Heston volatility model. The average price process is either discretely monitored
or continuously monitored.
The thesis is organized as follow. In Chapter 1, we give a brief introduction to Asian
options, including their history, their advantage over the more classical European option,
and various methods of pricing. In Chapter 2, we outline the mathematical framework
for the equity price process, average process, and the risk-neutral price of a general Asian
option. In Chapter 3, we introduce the standard Black-Scholes model and the Heston
model. Then, for each model, we compute the first few moments of the average price
at maturity. In Chapter 4, we explore various distributions which are used to fit the
distribution of the average process at the maturity. In Chapter 5, based on moments fitting,
we derive the pricing formulae for fixed-strike Asian options. In Chapter 6, we present the
numerical results obtained by Monte Carlo simulation and by methods of moments fitting.
The latter results are then compared against those obtained by Monte Carlo simulation.
In Chapter 7, we draw some conclusions about various methods of moments fitting.

iii

Acknowledgements
I am deeply indebted to my supervisor Professor Tony Wirjanto for his immeasurable
guidance and valuable advice in the program. Without his timely motivation and encouragement, this thesis would have been impossible.
I would like to thank Professors Adam Kolkiewicz and Andrew Heunis for their wonderful courses in financial mathematics and stochastic calculus which provide me a solid
foundation for future research.

iv

Table of Contents
List of Tables

vii

1 Introduction

2 The Mathematical Framework for Asian Options

3 Moments of the Average Price AT

3.1

3.2

The Geometric Brownian Motion . . . . . . . . . . . . . . . . . . . . . . .

3.1.1

The case of continuous averaging . . . . . . . . . . . . . . . . . . .

3.1.2

The case of discrete averaging . . . . . . . . . . . . . . . . . . . . .

10

The Heston Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

20

4 Distributions and Moments Fitting

27

4.1

Log-Normal Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . .

27

4.2

Shifted-Log-Normal Distributions . . . . . . . . . . . . . . . . . . . . . . .

29

4.3

Skew-Normal Distributions and Log-Skew-Normal Distributions . . . . . .

33

4.3.1

Standard Skew-Normal Distributions . . . . . . . . . . . . . . . . .

33

4.3.2

Affine Transformation of Skew-Normal Distributions

. . . . . . . .

38

4.3.3

Log-Skew-Normal Distributions . . . . . . . . . . . . . . . . . . . .

41

5 Pricing Asian Options by Moments Fitting

44

5.1

Approximation by log-normal distributions . . . . . . . . . . . . . . . . . .

44

5.2

Approximation by shifted-log-normal distributions . . . . . . . . . . . . . .

46

5.3

Approximation by skew-normal fitting . . . . . . . . . . . . . . . . . . . .

46

5.4

Approximation by log-skew-normal fitting . . . . . . . . . . . . . . . . . .

47

6 Numerical Results

49

6.1

Pricing of Asian Call Options, for the GBM Model . . . . . . . . . . . . .

50

6.2

Pricing of Asian Call Options, for the Heston Model . . . . . . . . . . . . .

53

7 Conclusions

56

Bibliography

58

vi

List of Tables
6.1

Pricing of Asian call options, GBM model . . . . . . . . . . . . . . . . . .

51

6.2

Pricing of Asian call options, Heston model . . . . . . . . . . . . . . . . . .

54

vii

Chapter 1
Introduction
Asian options are derivatives whose payoff is determined by the average value of the underlying asset over some pre-set period of time. They were introduced in the oil market
in the late 1970s and are nowadays traded in over-the-counter (OTC) markets all over the
world. Asian options are path-dependent due to the averaging and they are often classified
as exotic options.
Compared to standard European options, Asian options have several advantages. First
of all, due to averaging, the volatility of the average value is generally lower than that of the
equitys price. Asian options, in general, are cheaper than the European counterpart and
hence are better suited for hedging purposes. Secondly, since the average price is computed
by sampling the underlying equitys price over a long period of time, Asian options reduce
the risk of price manipulation near the maturity date, especially when the underlying is a
thinly traded asset or commodity.
However, one major drawback of Asian options is that, except some rare cases (for
example, when the strike K = 0), in general, there is no known analytical formula for
computing its risk-neutral price. The difficulty comes from the fact that even under the
standard Black-Scholes model (in which the stock price follows the GBM model with
constant drift and constant volatility, and the short rate process is assumed to be a deterministic constant) the average price at the maturity AT is not log-normally distributed.
In order to obtain the risk-neutral price, we have to employ some approximation schemes.
Currently, we have the following three different approaches of approximation.
The first approach is to approximate the distribution of the random variable AT by a
known distribution. The methodology of this approach is to choose suitable parameters
for a known distribution such that the first few moments of that distribution match with
1

those of the average price AT . Then we approximate the expectation

of the payoff, say, in
R
Q
the case of fixed-strike Asian call, E [max(AT K, 0)] by max(x K, 0)d(x), where
is the distribution measure of a known distribution. We hope that the known distribution is simple enough, so the integral can be computed relatively quickly by a numerical
quadrature method, and even calculated analytically.
The second approach is to price Asian options by a PDE method. For the case of
continuous averaging, the pricing problem reduces to solving a three dimensional partial
differential equations, that is, there are two spatial variables and a time variable. For
both fixed-strike and floating-strike options, the spatial dimension can be reduced by one
by utilizing some special techniques. For the case of discretely monitored averaging, we
notice that one spatial variable is missing between any two successive observation times.
Therefore, the problem reduces to a pseudo one-dimensional problem.
The last approach is to price Asian options by Monte Carlo simulation. The advantage
of this approach is that it is intuitive and simple. However, due to the nature of Monte
Carlo simulation, this approach is rather slow.
The thesis is organized as follow. In Chapter 2, we lay down the mathematical setting,
review relevant concepts such as filtration, conditional expectation, martingale, and define
the average process, the payoff of an Asian option and its risk-neutral price. In Chapter
3, we compute the first few moments of the average price AT , where the equity price
process either follows the GBM model or the Heston model. Both discrete averaging and
continuous averaging are considered. In Chapter 4, we explore various distributions which
will be used to fit moments of the average price AT . For each distribution, its probability
density function will be given, its first few moments will be derived and expressed in terms
of the parameters of that distribution. Lastly, the parameters are solved and expressed in
terms of moments, either in closed-form or by Newtons method. In Chapter 5, we derive
the pricing formulae for fixed-strike Asian options, either in closed-form or as an integral
which is then calculated by Simpson rule. In Chapter 6, we show the numerical results of
pricing fixed-strike Asian call options. The numerical results obtained by various moments
fitting methods are then compared against those obtained by Monte Carlo simulation,
which serves as a benchmark. In Chapter 7, we draw some conclusions about various
methods of moments fitting.

Chapter 2
The Mathematical Framework for
Asian Options
In this chapter, we outline the mathematical setting and define the payoff, which is the
risk-neutral price of an Asian option. Throughout the thesis, the triplet (, F, Q) denotes
a complete probability space and (Ft )t[0,) denotes a filtration satisfying the usual conditions. That is, for each t [0, ), Ft is a -algebra on such that for any s, t [0, )
with s < t, Fs Ft F, and the following conditions are satisfied:
completeness of Q: For any A , A F whenever there exists B F such that
A B and Q(B) = 0.
richness of F0 : For any A F, if Q(A) = 0, then A F0 .
right-continuity of filtration: For any t [0, ), Ft = {Fu | u (t, )}.
Given a -algebra G with G F and an integrable random variable X, there always exists
a unique (up to a Q-measure Rzero set, i.e.
R Q-a.e.) G-measurable integrable random variable
Y such that for any A G, A Y dQ = A XdQ. Y is called the conditional expectation of
X given G, denoted by Y = E[X | G]. By a random process (Xt )tI , we mean a family of
random variables indexed by a sub-interval I of [0, ), i.e., I is a sub-interval of [0, )
such that for each t I, Xt : R is a F/B(R)-measurable mapping. We say that the
random process (Xt )tI is (Ft )tI -adapted if for each t I, Xt is Ft /B(R)-measurable. In
this case, we say that (Xt , Ft )tI is an adapted process. If the index set I is clear from
the context, we sometime simplify the notation and simply denote the adapted process by
3

(Xt , Ft )t or even (Xt )t if the filtration is understood. An adapted process (Xt , Ft )tI is
martingale if for each t I, E[|Xt |] < and Xs = E[Xt | Fs ] whenever s, t I with
s < t.
Throughout the thesis, we fix a time horizon T > 0. Let (rt , Ft )t[0,T ] be the short
rate process, which is assumed to be sufficiently integrable. Let (Bt )t[0,T ] be a risk-free
Rt
bank account, defined by Bt = exp( 0 ru du). Here, we suppose that Q is a martingale
measure in the following sense: If (t , Ft )t[0,T ] is the price process of a traded asset in the
t
, Ft )t[0,T ] is a martingale with respect
market, the discounted price process, namely ( B
t
to the Q-measure. Let (St , Ft )t[0,T ] be the price process of a non-dividend paying stock.
In this thesis, (St )t is assumed to follow either a standard geometric Brownian motion
(GBM) or Heston dynamics, which will be discussed in more detail at a later time. Let
(At , Ft )t[0,T ] be the average process associated with (St , Ft )t , defined as follow. For the
case of continuous averaging,
(
S0 ,
if t = 0,
At = 1 R t
S du, if t (0, T ].
t 0 u
For the case of discrete averaging, we fix a sequence (ti )N
i=0 , with 0 = t0 t1 < . . . < tN
T. Then we define

if t [0, t1 ),
0,P
k
At = k1 i=1 Sti ,
if t [tk , tk+1 ), for k = 1, . . . , N 1,

1 PN
if t [tN , T ].
i=1 Sti ,
N
Now, we are ready to introduce Asian options mathematically. An Asian option with
strike K and maturity T is a contingent T -cliam whose payoff at the maturity T is:
Fixed-strike Call: = max(AT K, 0),
Fixed-strike Put: = max(K AT , 0),
Floating-strike Call: = max(ST KAT , 0),
Floating-strike Put: = max(KAT ST , 0).
By the general pricing theory, for each t [0, T ], we have




t

Q T
Q
=E
| Ft = E
| Ft ,
Bt
BT
BT
h
i
t
and hence t = Bt E Q B
|
F
t , (Q-a.e.). In this thesis, we will only consider the case
t
of fixed-strike Asian options.
4

Chapter 3
Moments of the Average Price AT
In this chapter, we give an introduction to the GBM model and the Heston model. Under
each of the models, we compute the first few moments of the average price AT . These
results will be used in subsequent chapters for moments fitting to various distributions.

3.1

The Geometric Brownian Motion

The Geometric Brownian Motion (GBM) model, also known as the standard Black-Scholes
model, is one the oldest mathematical model used for pricing derivatives. It was first
published by Fischer Black and Myron Scholes in [3]. The main advantage of this model is
its simplicity. For example, under the Black-Scholes model, we have closed-form formulae
for the risk-neutral price of European options and continuously monitored barrier options.
However, it is well-known that Black-Scholes model is over simplified and cannot capture
the reality of the market. For example, the implied volatility of European options observed
from the market is not constant among different strikes and maturities. Therefore, more
advanced models like a local volatility model, a Heston model, etc. have been proposed
and studied in the literature. In this section, we define the GBM model and derive first
few moments of the average price AT .
In the GBM model, the short rate process (rt )t is assumed to be a deterministic constant, i.e., there is a constant r such that rt () = r for any (t, ) [0, T ] . Let
(Wt , Ft )t[0,) be a standard 1-dimensional Wiener process. The equity price process
(St )t[0,T ] is assumed to follow the dynamics:
dSt = St (rdt + dWt ) ,
5

where the initial stock price S0 > 0 and the volatility 6= 0 are deterministic constants.
In the following, we consider two modes of averaging, namely, continuous averaging and
discrete averaging.

3.1.1

The case of continuous averaging

In this case, we can compute recursively the moments of AT to an arbitrary order. However,
for our moments fitting purpose, only Rthe first three moments are needed. Define an
t
auxiliary integral process (It )t by It = 0 Su du. In order to compute the moments of IT ,
we need the following lemma:
Lemma 3.1.1. Using the above notations, let 0 t1 t2 . . . tn T. Then we have


E [St1 St2 Stn ] = S0n exp r(t1 + t2 + . . . + tn ) + 2 [(n 1)t1 + (n 2)t2 + . . . + tn1 ] .
Proof. By Itos lemma, we have
1 1
1
dSt + 2 dSt dSt
St
2 St
1
= (rdt + dWt ) 2 dt
2


1 2
=
r dt + dWt ,
2

d ln St =

so for 0 s t, k N,
ln St ln Ss
Stk


Z t
Z t
1 2
=
r du +
dWu ,
2
s
s
 


1 2
k
= Ss exp k r (t s) + k (Wt Ws ) .
2

Therefore, we have
 


 k

1 2
1 2 2
k
E St | Fs = Ss exp k r (t s) + k (t s)
2
2


k(k 1) 2
k
= Ss exp kr(t s) +
(t s) .
2

Now, by the tower property of conditional expectation, we obtain

 

E [St1 St2 Stn ] = E E St1 St2 Stn | Ftn1


= E St1 St2 Stn1 Stn1 exp {r(tn tn1 )}
 

= E E St1 St2 Stn1 Stn1 exp {r(tn tn1 )} | Ftn2



= E St1 St2 Stn2 exp 2r(tn1 tn2 ) + 2 (tn1 tn2 ) + r(tn tn1 )
=
(
)
n1
X
(n k)tk .
= S0n exp r(t1 + t2 + + tn ) + 2
k=1

Next we compute moments of IT .

Proposition 3.1.1. Using the above notations, the first three moments of IT are
S0
{exp(rT ) 1} ,
r




2S02
2
2
2
2
E[IT ] =
r
exp
(2r
+

)T

(2r
+

)
exp(rT
)
+
r
+

,
r(r + 2 )(2r + 2 )
(
 3





6S03
1
1
1
1
2
2
E IT =

exp
3(r
+

)T

exp
2r
+

T
r + 2 2 2r + 3 2 3(r + 2 )
r + 2 2r + 2


1
1
1
1

+
+
exp(rT ) +
2
2
2
2
r(r + ) r(2r + 3 )
r(2r + 3 ) (r + )(2r + 2 )
)
1
1

.
2
2
3(r + )(2r + 3 ) r(r + 2 )
E [IT ] =

Proof. Fix n N. Observe that the cube [0, T ]n can be represented as a disjoint union of
sets of the form {(t1, . . . , tn ) | 0 < t1 < . . . < tn < T } and a set of Lebesgue measure zero,
namely
[0, T ]n = {(t1 , . . . , tn ) | 0 < t1 < . . . < tn < T } A,
where A is a subset of Rn with Lebesgue measure zero and the union is taken over all
permuations on the set {1, 2, . . . , n}. For each permutation on the set {1, 2, . . . , n}, let
A = {(t1 , t2 , . . . , tn ) | 0 < t1 < t2 < . . . < tn < T }
7

and let
A0 = {(t1 , t2 , , tn ) | 0 < t1 < t2 < . . . < tn < T }.
By symmetry,
Q

E [(IT ) ] =

Z

Z
= (n!)
A0

St1 St2 Stn d

E Q [St1 St2 . . . Stn ] dn .

For small n, the integral in (3.1) can be computed directly. For example, we have
Q

E [IT ] =

T
Q

E [Su ] du =
0

S0 exp(ru)du =
0

S0
{exp(rT ) 1} .
r

By Lemma 3.1.1 , we also have

Z


Q
2
E (IT )
= 2!
E Q [St1 St2 ] d2
0
ZA


S02 exp r(t1 + t2 ) + 2 t1 d2
= 2
A0
Z T Z t2


= 2
S02 exp r(t1 + t2 ) + 2 t1 dt1 dt2
0
0
Z T




2S02
2
exp(rt
)
exp
(r
+

)t

1
dt2
=
2
2
r + 2 0



 1


2S02
1
(2r+ 2 )T
rT
e
1
=
e 1
r + 2 2r + 2
r
n
o
2S02
(2r+ 2 )T
2 rT
2
=
re
(2r + )e + r + .
r(r + 2 )(2r + 2 )

(3.1)

Finally, we have
E

(IT )

E Q [St1 St2 St3 ] d3

A
Z T Z t3 Z t2


3
exp r(t1 + t2 + t3 ) + 2 (2t1 + t2 ) dt1 dt2 dt3
6S0
0
0
0
Z T Z t3
3





6S0
exp r(t2 + t3 ) + 2 t2 exp (r + 2 2 )t2 1 dt2 dt3
2
r + 2 0 0

Z T



1
6S03
1  t3 (r+2 )
rt3
t3 (2r+3 2 )
e
e
1
e
1
dt3
r + 2 2 0
2r + 3 2
r + 2
Z T



6S03
1
1  (2r+2 )t3
3(r+ 2 )t3
rt3
rt3
e
e
e
e

dt3
r + 2 2 0
2r + 3 2
r + 2
(



 1


1
1
6S03
3(r+ 2 )T
rT
e
1
e 1
r + 2 2 2r + 3 2 3(r + 2 )
r

)

 1


1
1
2
e(2r+ )T 1
erT 1
r + 2 2r + 2
r
(
1
1
1
6S03
1
2
2

e3(r+ )T

e(2r+ )T +
2
2
2
2
2
r + 2 2r + 3 3(r + )
r + 2r +


1
1
1
1

erT +
2
2
2
2
r(r + ) r(2r + 3 )
r(2r + 3 ) (r + )(2r + 2 )
)
1
1

.
3(r + 2 )(2r + 3 2 ) r(r + 2 )

= 3!

=
=
=
=
=

Remark 3.1.1. Since AT =

1
I ,
T T

the first three moments of AT can be written down

immediately as
S0
{exp(rT ) 1} ,
rT




2S02
2
2
2
r
exp
(2r
+

)T

(2r
+

)
exp(rT
)
+
r
+

,
E[A2T ] =
r(r + 2 )(2r + 2 )T 2
(
3




1
1
6S
1
1
0
2
2
exp
3(r
+

)T

E[A3T ] =

exp
(2r
+

)T
(r + 2 2 )T 3 2r + 3 2 3(r + 2 )
r + 2 2r + 2


1
1
1
1
+

exp(rT ) +
+
2
2
2
2
r(r + ) r(2r + 3 )
r(2r + 3 ) (r + )(2r + 2 )
)
1
1

2
2
3(r + )(2r + 3 ) r(r + 2 )
E[AT ] =

3.1.2

The case of discrete averaging

Let T be the maturity time and let t1 , t2 , P

. . . , tn be discrete observing
Pn times, satisfying
n
1
0 t1 < t2 < . . . < tn T. Define IT = k=1 Stk and AT = n k=1 Stk . We have the
following proposition.
Proposition 3.1.2. Using the above notations, we have:
E[IT ] = S0

n
X

exp(rtk ),

k=1

E[IT2 ]

2S02

exp(r(ti + tj ) + ti ) +

i<j

E[IT3 ] = 6S03

S02

n
X

exp(2rti + 2 ti ),

i=1

exp r(ti + tj + tk ) + 2 (2ti + tj ) +

i<j<k

3S03

3S03

exp r(2ti + tj ) + 2 (3ti ) +

i<j

S03

exp r(ti + 2tj ) + 2 (2ti + tj ) +

i<j
n
X

exp 3rti + 3 2 ti .

i=1

10

Proof. By Lemma 3.1.1,

n
X

E[IT ] =

E[Sk ] = S0

k=1

E[IT2 ]

n
X

n
X

exp (rtk ) .

k=1

E[Sti Stj ]

i,j=1

= 2

E[Sti Stj ] +

i<j

2S02

n
X

E[St2i ]

i=1

exp r(ti + tj ) + ti +

S02

i<j

E[IT3 ] =

n
X

exp 2rti + 2 ti .

i=1

n
X

E[Sti Stj Stk ]

i,j,k=1

= 3!

E[Sti Stj Stk ] +3 C1

E[Sti St2j ] +

i<j

= 6S03

E[St2i Stj ]

i<j

i<j<k

+3 C1

n
X

E[St3i ]

i=1

exp r(ti + tj + tk ) + 2 (2ti + tj ) +

i<j<k

3S03

3S03

exp r(2ti + tj ) + 2 (3ti ) +

i<j

S03

exp r(ti + 2tj ) + 2 (2ti + tj ) +

i<j
n
X

exp 3rti + 3 2 ti .

i=1

Notice that if the observation times ti are evenly spaced and tn = T, the moments in
Proposition 3.1.2 can be simplified as follows.
11

Proposition 3.1.3. Using the above notations and further assuming that ti = ih, where
h = Tn , i = 1, . . . , n, we have:
E[IT ] = S0 erh
(
E[IT2 ] = S02

1 erhn
,
1 erh

e2rh erh(n+1)
2e(r+ )h

1 erh
1 e(r+2 )h
2

e2(2r+ )h e(2r+ )h(n+1)

2
(1 e(r+2 )h ) (1 e(2r+2 )h )
)
2
1 e(2r+ )hn (2r+2 )h
+
e
,
1 e(2r+2 )h
(
E[IT3 ]

S03

1 erh(n2)
6e(6r+4 )h

1 erh
(1 e(r+22 )h ) (1 e(r+2 )h )
2

1 e(2r+ )h(n2)
6e(7r+5 )h

(1 e(r+22 )h ) (1 e(r+2 )h )
1 e(2r+2 )h
2

6e(7r+6 )h
1 erh(n2)

2
2
1 erh
(1 e(r+2 )h ) (1 e(2r+3 )h )
2

6e9(r+ )h
1 e3(r+ )h(n2)
+

(1 e(r+22 )h ) (1 e(2r+32 )h )
1 e3(r+2 )h




3(r+ 2 )h(n1)
6(r+ 2 )h
(4r+3 2 )h
rh(n1)
1

e
3e
3e
1e
+

2
(1 e(2r+3 )h ) (1 erh ) (1 e(2r+32 )h ) (1 e3(r+2 )h )




2
2
2
2
3e(5r+4 )h 1 e(2r+ )h(n1)
3e6(r+ )h 1 e3(r+ )h(n1)
+

(1 e(r+22 )h ) (1 e(2r+2 )h )
(1 e(r+22 )h ) (1 e3(r+2 )h )

)
2
2
e3(r+ )h 1 e3(r+ )hn
+
.
1 e3(r+2 )h
Proof. By Proposition 3.1.2,
E[IT ] = S0

n
X
i=1

exp(rti ) = S0

n
X
i=1

12

exp(irh) = S0 erh

1 erhn
.
1 erh

Next, we compute
X

i<j

exp (r(ti + tj ) + 2 ti ) and

exp r(ti + tj ) + 2 ti

Pn

i=1

exp(2rti + 2 ti ).

i<j

j1
n X
X

exp rh(i + j) + 2 hi

j=2 i=1
n
X

1 e(r+2 )h(j1)
=
exp(rhj) exp (r + )h
1 e(r+2 )h
j=2
(
)
2
rh(n1)
(2r+ 2 )h(n1)
1

e
1

e
e(r+ )h
2
2
e2rh
e(r+ )h e2h(2r+ )
=
1 erh
1 e(r+2 )h
1 e(2r+2 )h
2

(2r+ 2 )h

e2rh erh(n+1)
e2(2r+ )h e(2r+ )h(n+1)
e(r+ )h

.
=
1 erh
1 e(r+2 )h
(1 e(r+2 )h ) (1 e(2r+2 )h )
n
X

exp(2rti + ti ) =

i=1

n
X

exp(2rhi + hi) = e

i=1

1 e(2r+ )hn
.

1 e(2r+2 )h

Hence, we have
(
E[IT2 ]

S02

2e(r+ )h
e2rh erh(n+1)

1 erh
1 e(r+2 )h
2

e2(2r+ )h e(2r+ )h(n+1)

2
(1 e(r+2 )h ) (1 e(2r+2 )h )
)
2
1 e(2r+ )hn (2r+2 )h
+
e
.
1 e(2r+2 )h
P
To compute the third moment E[IT3 ], we need i<j<k exp (r(ti + tj + tk ) + 2 (2ti + tj )) ,
P
P
Pn
2
2
2
i<j exp (r(2ti + tj ) + (3ti )) ,
i<j exp (r(ti + 2tj ) + (2ti + tj )) , and
i=1 exp (3rti + 3 ti ) .

13

exp r(ti + tj + tk ) + 2 (2ti + tj )

i<j<k

j1
n X
k1 X
X

exp rh(j + k) + 2 hj exp (r + 2 2 )hi

k=3 j=2 i=1

n X
k1
X

(r+2 2 )h

exp (r + )hj + rhk e

k=3 j=2

1 e(r+2 )h(j1)

1 e(r+22 )h

2
k1
n

e(r+2 )h X rhk X  (r+2 )hj
(r+2 2 )h (2r+3 2 )hj

e
e
e
e
=
1 e(r+22 )h k=3
j=2
(
n
(r+2 2 )h
(r+ 2 )h(k2)
X
e
rhk
2(r+ 2 )h 1 e
=
e
e

1 e(r+22 )h k=3
1 e(r+2 )h
)
(2r+3 2 )h(k2)
(r+2 2 )h
(2r+3 2 )2h 1 e
e
e

1 e(2r+32 )h
(
(3r+4 2 )h
rh(n2)
e
3rh 1 e
=
e
1 erh
(1 e(r+22 )h ) (1 e(r+2 )h )
)
(2r+ 2 )h(n2)
1e
2
2
e2(r+ )h e(2r+ )3h
1 e(2r+2 )h
(
2
rh(n2)
e(4r+6 )h
3rh 1 e

e
1 erh
(1 e(r+22 )h ) (1 e(2r+32 )h )
)
2
1 e3(r+ )h(n2)
2
2
e(4r+6 )h e9(r+ )h
1 e3(r+2 )h
2

e(6r+4 )h
1 erh(n2)
=

1 erh
(1 e(r+22 )h ) (1 e(r+2 )h )
2

e(7r+5 )h
1 e(2r+ )h(n2)

(1 e(r+22 )h ) (1 e(r+2 )h )
1 e(2r+2 )h
2

e(7r+6 )h
1 erh(n2)

1 erh
(1 e(r+22 )h ) (1 e(2r+32 )h )
2

1 e3(r+ )(n2)h
e9(r+ )h
+

.
2
2
(1 e(r+2 )h ) (1 e(2r+3 )h )
1 e3(r+2 )h
14

Next, we compute the quantity

X

i<j

exp (r(2ti + tj ) + 2 (3ti )) as follow.

exp r(2ti + tj ) + 2 (3ti )

i<j

j1
n X
X

exp (2r + 3 2 )hi + rhj

j=2 i=1

n
X

rhj

j1
X

j=2

n
X

e(2r+3

2 )hi

i=1
2

erhj e(2r+3

j=2

2 )h

1 e(2r+3 )h(j1)
1 e(2r+32 )h

" n
#
2
n
X
X
e(2r+3 )h
rhj
3(r+ 2 )hj
(2r+3 2 )h
=
e
e
e
1 e(2r+32 )h j=2
j=2
2

rh(n1)
e(2r+3 )h
2rh 1 e
e
=
1 erh
1 e(2r+32 )h
3(r+ 2 )h(n1)
1
6(r+ 2 )h 1 e

2 )h
1 e(2r+32 )h
1 e3(r+




6(r+ 2 )h
3(r+ 2 )h(n1)
rh(n1)
(4r+3 2 )h
e
1

e
1e
e
=

.
(1 e(2r+32 )h ) (1 erh ) (1 e(2r+32 )h ) (1 e3(r+2 )h )

15

For

i<j

exp (r(ti + 2tj ) + 2 (2ti + tj )) , we have

X

exp r(ti + 2tj ) + 2 (2ti + tj )

i<j

=
=

j1
n X
X

exp (2r + 2 )hj + (r + 2 2 )hi

j=2 i=1
n
X

(r+2 2 )h

exp (2r + )hj e

j=2
2

1 e(r+2 )h(j1)

1 e(r+22 )h

1 e(2r+ )h(n1)
e

1 e(2r+2 )h
)
3(r+ 2 )h(n1)
1

e
2
2
e(r+2 )h e6(r+ )h
1 e3(r+2 )h

e(r+2 )h

=
1 e(r+22 )h

2(2r+ 2 )h

1 e(2r+ )h(n1)
e6(r+ )h
1 e3(r+ )h(n1)
e(5r+4 )h

.
=
1 e(r+22 )h
1 e(2r+2 )h
1 e(r+22 )h
1 e3(r+2 )h
Lastly,
n
X
i=1

exp 3rti + 3 ti =

n
X

3(r+ 2 )hi

i=1

16

3(r+ 2 )h

=e

1 e3(r+ )hn

.
1 e3(r+2 )h

Therefore, we have
(
E[IT3 ] = S03

6e(6r+4 )h
1 erh(n2)

1 erh
(1 e(r+22 )h ) (1 e(r+2 )h )
2

1 e(2r+ )h(n2)
6e(7r+5 )h

(1 e(r+22 )h ) (1 e(r+2 )h )
1 e(2r+2 )h
2

6e(7r+6 )h
1 erh(n2)

1 erh
(1 e(r+22 )h ) (1 e(2r+32 )h )
2

6e9(r+ )h
1 e3(r+ )h(n2)
+

(1 e(r+22 )h ) (1 e(2r+32 )h )
1 e3(r+2 )h




6(r+ 2 )h
3(r+ 2 )h(n1)
rh(n1)
(4r+3 2 )h
3e
1

e
1e
3e
+

2
(1 e(2r+3 )h ) (1 erh ) (1 e(2r+32 )h ) (1 e3(r+2 )h )




(5r+4 2 )h
(2r+ 2 )h(n1)
6(r+ 2 )h
3(r+ 2 )h(n1)
3e
1e
3e
1e
+

2
2
2
(1 e(r+2 )h ) (1 e(2r+ )h )
(1 e(r+2 )h ) (1 e3(r+2 )h )
)

2
2
e3(r+ )h 1 e3(r+ )hn
+
.
1 e3(r+2 )h

Remark 3.1.2. Consider the case when n , or equivalently h 0. As expected, we

will see that moments
of AT approach those where AT is defined by
RT
Pcontinuous averaging,
(n)
i.e., AT = T1 0 Su du. Recall that in the discrete case, AT = n1 ni=1 Sti . Therefore, the
limits of moments are given as follow:

(n)

1 erhn
h0 T
1 erh
h
h
S0
=
(1 erT ) lim
h0 1 erh
T
S0 rT
=
(e 1).
rT

lim E[AT ] = lim

h0

S0

erh

In the above, we evaluate the limit by LHospital rule: limx0+

17

x
1erx

= limx0+

1
rerx

= 1r .

Similarly, we have:
(n)2

=
=
=
=

lim E[AT ]
h0


h
h(1 erT )
h
h
S02
(2r+ 2 )T
2 lim
2(1 e
) lim

2
2
2
h0 1 e(r+ )h
h0 1 e(r+ )h 1 e(2r+ )h
T2
1 erh


2S02
1 erT
1
1
1
(2r+ 2 )T

(1 e
)

T 2 (r + 2 )
r
(r + 2 ) (2r + 2 )
n
o
2S02
rT
2
(2r+ 2 )T
(1

e
)(2r
+

(1

e
)r
T 2 r(r + 2 )(2r + 2 )
n
o
2S02
2
2 rT
(2r+ 2 )T
r
+

(2r
+

)e
+
re
,
T 2 r(r + 2 )(2r + 2 )

18

and
(n)3

lim E[AT ]
h0
(


h
h
h
S03
rT

lim
6
1

2
2
h0 1 e(r+2 )h 1 e(r+ )h 1 erh
T3


h
h
h
2
6 1 e(2r+ )T lim
2 )h
2 )h
2 )h
(r+2
(r+
(2r+
h0 1 e
1e
1e


h
h
h
6 1 erT lim
2 )h
2 )h
(r+2
(2r+3
h0 1 e
1 erh
1e
)


h
h
h
2
+6 1 e3(r+ )T lim

2
2
2
h0 1 e(r+2 )h 1 e(2r+3 )h 1 e3(r+ )h
(


2
1 erT
1 e(2r+ )T
6S03
+
T 3 (r + 2 2 ) (r + 2 ) r (r + 2 2 ) (r + 2 ) (2r + 2 )
)
2
1 erT
1 e3(r+ )T
+

(r + 2 2 ) (2r + 3 2 ) r (r + 2 2 ) (2r + 3 2 ) 3 (r + 2 )
(

2


6S03
1
1
1 e(2r+ )T
rT
+
1e
+
T 3 (r + 2 2 )
r (r + 2 ) r (2r + 3 2 )
(r + 2 ) (2r + 2 )
)
2
1 e3(r+ )T

3 (2r + 3 2 ) (r + 2 )
(
2
2
6S03
e3(r+ )T
e(2r+ )T

(r + 2 2 ) T 3 3 (r + 2 ) (2r + 3 2 ) (r + 2 ) (2r + 2 )


1
1
1
1
+

erT +
+
2
2
2
2
r (r + ) r (2r + 3 )
r (2r + 3 ) (r + ) (2r + 2 )
)
1
1

.
2
2
3 (r + ) (2r + 3 ) r (r + 2 )

19

3.2

The Heston Model

The Heston Model, proposed by Steven Heston [5], is a more advanced mathematical model
for modeling the evolution of the price of an asset. Under the Heston model assumption,
the volatility of the price process of an asset is no longer deterministic. Rather, it is
stochastic and follows the Cox-Ingersoll-Ross model (CIR-model), first studied in [4]. The
main advantage of the Heston model over the GBM model is that the Heston model is more
flexible and can capture the reality of the market better. However, due to its complexity,
in general, it is much more difficult to obtain closed form formulae for pricing derivatives.
In the following, we define the Heston model, derive the first moment and approximate the
second moment of the average price AT , where the average price is computed in continuous
basis.
In the Heston model, the short rate process (rt )t is assumed to be a deterministic
constant, i.e., there is a constant r R such that rt () = r, for any (t, ) [0, ) . Let
(Wt , Ft )t[0,) be a correlated 2-dimensional Wiener process (under the Q-measure), i.e.,
Wt = (Wt1 , Wt2 ), with a quadratic covariance process [W 1 , W 2 ]t =t, where [1, 1],
known as the correlation (sometime denoted by dWt1 dWt2 = dt.). The variance process
(Vt )t[0,) is a (Ft )t[0,) -adapted process, satisfying the following conditions:
V0 is deterministic and V0 > 0, and

dVt = a(b Vt )dt + c Vt dWt2 , where a, b, c R are constants, with a > 0, b > 0,
c 6= 0, 2ab > c2 .
The constants a, b, c are known as the mean reversion rate, the anchor, and the vol-vol
respectively. Moreover, the condition 2ab > c2 is known as the Fellers condition which
guarantees the existence of a strictly positive solution of the SDE. The
stock price

process
1
(St )t[0,) is (Ft )t[0,) -adapted, satisfying the SDE: dSt = St rdt + Vt dWt , and the
initial condition
that S0 is deterministic and S0 > 0. We define the integral process (It )t
Rt
by It = 0 S(u)du.
We follow [2] and consider the transformation Xt = ln(St ) rt. By Itos lemma, we
have that
 
1
1 1
dXt =
dSt +
dSt dSt rdt
St
2 St2

 1
p
= rdt + Vt dWt1 Vt dt rdt
2
p
1
1
= Vt dt + Vt dWt .
2
20

Hence, the system of SDEs becomes

p
1
dXt = Vt dt + Vt dWt1 ,
2
p
dVt = a(b Vt )dt + c Vt dWt2 ,
with initial conditions X0 = x0 = ln(S0 ) and V0 = v0 . Fix t > 0. By page 4 of [2], there
exists an open neighborhood U of 0 such that for each u U, we have that


E Q euXt = eux0 A(t)B(t),
where

! 2ab2

e(acu) 2


sinh
cosh 2t P (u) + acu
P (u)

A(t) =

uu2
P (u)

(
B(t) = exp v0
and

cosh

t
P (u)
2

t
P (u)
2

t
P (u)
2
acu
sinh
P (u)

sinh
+

)
t
P (u)
2

q
P (u) = (a cu)2 + c2 (u u2 ).

Using this formula, we are able to compute moments of St , the first moment of It and
an approximation of
the second moment of It . Notice that Stu = erut euXt , and hence
 uX
Q
u
rut
Q
E [St ] = e E e t .
We follow the idea of [1] and have the following.
Proposition 3.2.1. Using the notations as above, for t [0, ), we have that


S0 rt
e 1 ,
r (
 
t2 3r + v0 3
1
E Q It2 2S02
+
t +
2
6
24
E Q [It ] =

)



2

r(7r + 5v0 ) + ab + v0 (2c a) + v0 t4 ,

where the right hand side of the second expression is the Taylor expansion of E Q [It2 ] about
0 up to order 4.
21

Proof. Observe that

u [0, ),

St
Bt

is a martingale under the measure Q. In particular, for any

t

 ##


Q
Q
E E Su  F 0

 # #
"
"

Q
Q Su 
E E
 F0 Bu
Bu 


Q S0
E
Bu
B0
eru S0 .
"

E Q [Su ] =
=
=
=

"

Therefore, we have
Q

E [Su ] du =

E [It ] =

S0 eru du =

S0 rt
e 1 .
r

Next, we compute the second moment.


Z t
Z t
 2
Q
Q
E It = E
Su du
Sv dv
0
0
Z tZ t
E Q [Su Sv ] dudv
=
0Z Z
0
= 2
E Q [Su Sv ] d2 (u, v)
{(u,v) | 0uvt}
v
E Q [Su Sv ] dudv
2
0 0
 ##
"
"
Z tZ v


Q
Q
2
E E Su Sv  Fu dudv

0 0

Z tZ

=
=

Z tZ



E Q er(vu) Su2 dudv
Z0 t 0 Z v
 
= 2
erv
eru E Q Su2 dudv.

= 2

For simplicity, we denote (t) = E Q [It2 ] and compute the derivatives of at t = 0. Clearly,

22

(0) = 0. 0 (t) = 2ert

00

Rt
0

eru E Q [Su2 ] du, so 0 (0) = 0.

Z

 
 
eru E Q Su2 du + 2ert ert E Q St2
0
 
= r0 (t) + 2E Q St2 .

(t) = 2re

rt

Therefore, we have 00 (0) = 2S02 .

d  2rt 2x0
e e A(t)B(t)
dt  
00
= r (t) + 4rE Q St2 + 2S02 e2rt {A0 (t)B(t) + A(t)B 0 (t)} ,

000 (t) = r00 (t) + 2

where A(t), B(t) are the functions defined in above with u = 2. Now,

! 2ab2 1
c
(a2c) 2t
2ab
e


A0 (t) =

c2
cosh( 2t P (2)) + a2c
sinh 2t P (2)
P (2)
(



t
a 2c a2c t
a 2c
t
cosh( P (2)) +
e 2
sinh
P (2)
2
2
P (2)
2



)

a2c
t
a

2c
t
P
(2)
sinh
P (2) +
cosh
P (2)
/
e 2 t
2
2
2
2




2
t
a 2c
t
cosh
P (2) +
sinh
P (2)
2
P (2)
2


sinh 2t P (2)
2ab  a2c t  2ab
c2
c2
=
e 2

+1



i 2ab
c2
P (2) h
c2
t
cosh 2t P (2) + a2c
sinh
P
(2)
P (2)
2


sinh 2t P (2)
2ab (a2c) ab2 t
c
=
e
h
,
+1

a2c

i 2ab
P (2)
c2
t
t
cosh 2 P (2) + P (2) sinh 2 P (2)

23

and
(






P (2)
t
t
a 2c
t
2v
0
0

cosh
P (2) cosh
P (2) +
sinh
P (2)
B (t) = B(t)
P (2)
2
2
2
P (2)
2





)
t
P (2)
t
a 2c
t
sinh
P (2)
sinh
P (2) +
cosh
P (2)
/
2
2
2
2
2




2
t
a 2c
t
cosh
P (2) +
sinh
P (2)
2
P (2)
2




P (2) 
cosh2 2t P (2) sinh2 2t P (2)
2v0
2
h
= B(t)



i2
P (2)
t
sinh
P
(2)
cosh 2t P (2) + a2c
P (2)
2
= 
cosh

v0 B(t)


t
P (2) + a2c
sinh
2
2

2

t
P (2)
2

Hence, we have
(
000 (t) = r00 (t) + 4rE


Q


2

St + 2S02 e2rt B(t)

sinh
h

cosh

t
P (2)
2

t
P (2)
2

a2c
P (2)

2ab (a2c) ab2 t

c
e

P (2)

+
+1

i 2ab
c2

t
P (2)
2

sinh

ab

v0 h

e(a2c) c2 t

+2

i 2ab
c2
t
+ a2c
sinh
P
(2)
P (2)
2


ab
= r00 (t) + 4rE Q St2 + 2S02 e(2r+(a2c) c2 )t B(t)

h

a2c

i

t
t
t
2ab
P (2) sinh 2 P (2) cosh 2 P (2) + P (2) sinh 2 P (2) + v0

.
h
+2

a2c

i 2ab

c2

t
t
cosh 2 P (2) + P (2) sinh 2 P (2)

cosh

t
P (2)
2

In particular, we have
000 (0) = 2rS02 + 4rS02 + 2S02 v0
= (6r + 2v0 ) S02 .
Lastly, we compute the quantity (4) (0). Denote

 
ab
C(t) = exp
2r + (a 2c) 2 t
c
24

and
D(t) =

2ab
P (2)

sinh

h
cosh

i
sinh 2t P (2) + v0
.
h
+2

a2c

i 2ab
c2
t
t
cosh 2 P (2) + P (2) sinh 2 P (2)
t
P (2)
2

t
P (2)
2

a2c
P (2)

We now compute the quantities C 0 (0) and D0 (0). We have that



ab
ab
0
C (t) = 2r + (a 2c) 2 e(2r+(a2c) c2 )t ,
c
so C 0 (0) = 2r + (a 2c)
2ab
f (t) =
sinh
P (2)
and

ab
.
c2

We further write D(t) =

f (t)
,
g(t)

where

 




t
t
a 2c
t
P (2) cosh
P (2) +
sinh
P (2) + v0 ,
2
2
P (2)
2



 2ab2 +2


c
t
t
a 2c
P (2) +
sinh
P (2)
g(t) = cosh
.
2
P (2)
2

Now
 




t
a 2c
t
t
P (2) cosh
P (2) +
sinh
P (2) +
f (t) = ab cosh
2
2
P (2)
2

 




t
t
a 2c
t
P (2) sinh
P (2) +
cosh
P (2)
ab sinh
2
2
P (2)
2


and hence f 0 (0) = ab.

 



 2ab2 +1
c
2ab
t
a

2c
t
+
2

cosh
P
(2)
+
sinh
P
(2)
g 0 (t) =
c2
2
P (2)
2





P (2)
t
a 2c
t

sinh
P (2) +
cosh
P (2) ,
2
2
2
2


so, we have

a 2c
2ab
g (0) =
+2 1
2
c
2


ab
=
+ 1 (a 2c) .
c2
0

25

Therefore, we obtain
f 0 (0)g(0) f (0)g 0 (0)
[g(0)]2


ab 1 v0 ab
+ 1 (a 2c)
c2
=
1

ab
= ab v0
+ 1 (a 2c).
c2

D0 (0) =

Finally,
(

ab
(4)
2
2
2
(0) = r (6r + 2v0 )S0 + 4r (2r + v0 )S0 + 2S0 2r + (a 2c) 2
c
)


ab
1 v0 + 1 v0 v0 + 1 1 ab v0 ( 2 + 1) (a 2c)
c
( 

ab
2
2
= rS0 (14r + 6v0 ) + 2S0 v0 2r + (a 2c) 2 + v02
c
)


ab
+ab v0
+ 1 (a 2c)
c2


= rS02 (14r + 6v0 ) + 2S02 ab + v0 (2r + 2c a) + v02


= rS02 (14r + 10v0 ) + 2S02 ab + v0 (2c a) + v02 ,
and hence we have obtained the following approximation
1
(t) (0) + 0 (0)t + 00 (0)t2 +
2!

1
1
= 2S02
t2 + (3r + v0 )t3 +
2
6

1 000
1
(0)t3 + (4) (0)t4
3!
4!


1 
2 4
r(7r + 5v0 ) + ab + v0 (2c a) + v0 t .
24

26

Chapter 4
Distributions and Moments Fitting
In this chapter, we explore several distributions, namely, log-normal distributions, shifted
log-normal distributions, skew-normal distributions, and log-skew-normal distributions.
We also compute their first few moments and express the parameters in terms of moments.
These results will be used in the next chapter for pricing Asian options.

4.1

Log-Normal Distributions

In this section, we firstly define a log-normal distribution, then we derive the formulae for
moments fitting of the average stock price.
Log-normal distribution is a very common distribution used in the financial industry. In
fact, if the stock price (St )t process follows the GBM model, the stock price ST at the time
horizon T is log-normally distributed. One of the advantages of the log-normal distribution
is its simplicity and its nice behavior. For example, for this distribution all moments exist
and have closed-form formulae. Another property of the log-normal distribution is that
a log-normally distributed random variable is strictly positive which is a nice feature to
model the price of a stock.
Definition 4.1.1. Let X be a random variable. We say that X follows a log-normal
distribution with parameters and if ln(X) is normally distributed with mean and
variance 2 .
Proposition 4.1.1. Let X be a log-normally distributed random variable with parameters
and . Define Z = ln(X) and let X , Z be the probability density functions of X
27

and Z respectively. Then X and Z are related as follow. X (x) = 0, if x 0 and

X (x) = x1 Z (ln x), if x > 0. Z (z) = X (ez )ez .
Proof. By definition, since ln(X) is well-defined, X > 0 everywhere. Therefore P [X 0] =
0. In particular, X (x) = 0 if x 0. For x > 0, P [X x] = P [Z ln(x)] . Differentiating
both sides with respect to x yields X (x) = Z (ln(x)) x1 . Let z R, then P [Z z] =
P [X ez ] . Differentiating both sides with respect to z, we obtain Z (z) = X (ez )ez .
Proposition 4.1.2. Let X be a log-normally distributed random variable with parameters
1 2
2
and . Then the first two moments of X are given by E [X] = e+ 2 and E [X 2 ] = e2+2 .
Proof. Let Z = ln(X) distributed normally with mean and variance 2 , then
 
E [X] = E eZ
Z
1
2
1
=
et
e 22 (t) dt
2
Z

1
2 2
1 2
1
e 22 [t(+ )] dt
=
e+ 2
2

= e+ 2 .
Note that 2Z N (2, 4 2 ), so by the previous result, we have
 
 
1
2
2
E X 2 = E e2Z = e2+ 2 4 = e2+2 .

We are now ready to state the general result for moments fitting by the log-normal
distribution.
Proposition
4.1.3. Let F be
such that its first two moments exist. Let
R
R a distribution
2
m1 = xdF (x) and m2 = x dF (x). Suppose further that m1 > 0 and m2 > 0. Then
there exists a log-normally distributed random variable X with parameters and such
that E [X] = m1 and E [X 2 ] = m2 . Moreover, and are given by
s  
 2 
m
m2
= ln 1 , and = ln
.
m2
m21

28

Proof. Let X be a log-normally distributed random variable with parameters and . By

1 2
2
the previous proposition, we have that E [X] = e+ 2 and E [X 2 ] = e2+2 . Suppose that
E [X] = m1 and E [X 2 ] = m2 , then
 + 1 2
e 2
= m1 ,
2+2 2
e
= m2 .
Solving the above simultaneous equation leads to the desired result. Conversely,
let X
 2 
m1
be a log-normally distributed random variable with parameters = ln m2 and =
r  
m2
2
ln m
2 , then by a direct calculation, we have E [X] = m1 and E [X ] = m2 .
1

4.2

Shifted-Log-Normal Distributions

Log-normal distribution has a drawback that it can only fit the first two moments because
there are only two parameters. We slightly modify it and define the notion of a shifted
log-normal distribution.
Definition 4.2.1. A random variable X is said to be shifted-log-normally distributed if
there exists R, and a normally distributed random variable Z N (, 2 ), where R
and > 0, such that X = + exp(Z).
A shifted-log-normal distribution admits a probability density function which can be
calculated as follows.
Proposition 4.2.1. Let R, R, > 0, and let Z be a normally distributed random
variable with Z N (, 2 ). Let X = + exp(Z), then X admits a probability density
function X . Moreover, X , and Z , the probability density function of Z, are related
through
(
0,
if x (, ]
X (x) =
1
, if x (, ),
Z (ln(x )) x
and Z (z) = X (ez + )ez .
Next, we consider the problem of matching the first three moments of a given distribution by a shifted-log-normal distribution. Before we state and prove the main theorem,
we need the following lemmas.
29

Lemma 4.2.1. Let a R be a constant, then the cubic equation x3 3x + a = 0 has at

most one root (counting multiplicity) in [2, ).
Proof. Define a function f : R R by f (x) = x3 3x + a. We prove by contradiction and
suppose to the contrary that the equation has at least two roots in [2, ). If x0 [2, )
is a repeated root, f 0 (x0 ) = 0. However, f 0 (x0 ) = 3x20 3 > 0 which is a contradiction.
Suppose that the equation has roots x1 , x2 [2, ) with with x1 6= x2 . By Rolles theorem,
there exists strictly between x1 and x2 such that f 0 () = 0. However, since > 2, we
have that f 0 () = 3 2 3 > 0 which is again a contradiction.
Lemma 4.2.2. Let a > 0, b > 0 such that ab = 1. Then a + b 2 and the equality holds
if and only if a = b = 1.

b 2.
If a = b = 1,
Proof. Observe that ( a b)2 0, so a + b 2 ab 0 and hence a +
2
we clearly have
a +b = 2. Conversely, suppose that a + b = 2, then ( a b) = 0. It
follows that a = b and hence a = b = 1.
Now, we are ready to state and prove the main result.
Theorem
4.2.1.
F be Ra distribution. Suppose that the first three moments of F , i.e.
R
R Let

2
xdF (x), x dF (x), x3 dF (x) exist and the centered second and third moments

R
R
are non-zero, i.e., (x x0 )2 dF (x) 6= 0 and (x x0 )3 dF (x) 6= 0, where x0 =
R
xdF (x). Then there exist unique , , R with > 0 such that if a random variable

R
Z is normally distributed with Z N (, 2 ) and X = + eZ , then E(X) = xdF (x),
R
R
E(X 2 ) = x2 dF (x) and E(X 3 ) = x3 dF (x). Moreover, , , are given explicitly
as follows.
centered secondR and the third moments by b and c respectively,
R Denote the
2

2
i.e., b = (x x0 ) dF (x) and c = (x x0 )3 dF (x) and let d = cb3 + 2. Let x0 be
p
the unique
x3 3x d = 0. Then = ln(x0 1),
 root in (2, ) of the cubic equation
R
1
1
1
= ln cb (x0 2)(x
,
and

=
xdF (x) cb (x0 2)(x
0 +1)
0 +1) .
0

x 1
Proof. Firstly, we prove the uniqueness part. Suppose that there exists
R , , R with
> 0 such that if Z N (, 2 ) and X = + eZ , then E(X) = xdF (x), E(X 2 ) =
R 2
R
x dF (x) and x3 dF (x). For convenience, we denote the first three moments of F

by m1 , m2 , and m3 respectively, and denote the mean, the second and the third central

30

moments by a, b, c respectively, i.e.

Z

xdF (x),

m1 =
Z

m2 =
Z

m3 =

x2 dF (x),
x3 dF (x),

a = m1 ,
Z
b =
(x m1 )2 dF (x),
Z

c =
(x m1 )3 dF (x).

We remark that a, b, and c can be expressed in terms of m1 , m2 and m3 , namely, a = m1 , b =

m2 m21 and c = m3 3m1 m2 +2m31 . Since 2Z N (2, 4 2 ) and 3Z N (3, 9 2 ), we have
that E(e2Z ) = exp(2 + 21 4 2 ) = exp(2 + 2 2 ) and E(e3Z ) = exp(3 + 21 9 2 ). It follows
1 2
1 2
9 2
2
2
that E(X 2 ) = 2 + 2e+ 2 + e2+2 and E(X 3 ) = 3 + 32 e+ 2 + 3e2+2 + e3+ 2 .
Now, we have
2
2
2
2
b = m2 m21 = e2+2 e2+ = e2+ (e 1),
and
c = m3 3m1 m2 + 2m31


9 2
1 2
2
= 3 + 32 e+ 2 + 3e2+2 + e3+ 2



1 2
1 2
2
2 + 2e+ 2 + e2+2
3 + e+ 2


3
2 + 12 2
2+ 2
3+ 32 2
+2 + 3 e
+ 3e
+e
9

= e3+ 2 3e3+ 2 + 2e3+ 2


3  2

+ 12 2
3
2
= e
e 3e + 2 .
1

Define = e+ 2 , = e , then we have

a = + ,
b = 2 ( 1),
c = 3 ( 1)2 ( + 2).
31

(4.1)
(4.2)
(4.3)

By assumption, b > 0, so cb3 is well-defined. By (4.2) and (4.3), cb3 = ( 1)( + 2)2 .
Therefore satisfies a cubic equation. We try to get rid of the square term by considering
the Tschirnhaus transformation x0 = + k, where k will be determined later. We have that
( 1)( + 2)2 = x03 + [(k + 1) 2(k 2)] x02 + , so we set (k + 1) 2(k 2) = 0,
i.e. k = 1. Formally, we define x0 = + 1, then
c2
= (x0 2)(x0 + 1)2 = x03 3x0 2.
b3
2

Define d = cb3 +2, then x03 3x0 d = 0. Observe that x0 = +1 > 2, so by Lemma 4.2.1, x0
p
p
is uniquely determined by d. Now = ln() = ln(x0 1). By considering (4.3)/(4.2),

1
. As = e , we have
we obtain cb = ( 1)( + 2) and hence = cb (1)(+2)





1
1
c
= ln

0
.
= ln

b (x0 2)(x0 + 1)
x 1
1
Lastly, = a = a cb (x0 2)(x
0 +1) . This shows that , , are uniquely determined by
m1 , m2 , m3 .

Secondly, we prove the existence part. By the previous discusssion, it prompts us to

consider the cubic equation x3 3x = d, where a = m1 , b = m2 m21 , c = m3 3m1 m2 +2m31 ,
2
and d = cb3 + 2. We show that the equation indeed has a solution in (2, ) and can be
solved explicitly by Cardanos method. Write x = u + v, where u and v will be determined
later, then we have:
x3 3x = u3 + v 3 + 3uv(u + v) 3(u + v)
= u3 + v 3 + 3(u + v)(uv 1).
Now we impose another condition on u and v, namely, uv = 1, and try to solve the following
simultaneous equations
(
u3 + v 3 = d,
uv = 1.

d 4
Clearly, u3 = d
. Notice that d > 2, so the roots are distinct and positive. If we
2
2
2
d+ d 4
3
choose u =
, then v 3 = u13 = d 2d 4 , and vice versa. Formally, we define
2
s

d2 4
3 d +
u =
,
2
s

d2 4
3 d
v =
2

32

and define x0 = u + v. By direct calculation, u3 v 3 = 1 and hence uv = 1. Clearly, u 6= v,

so by Lemma 4.2.2, x0 > 2. Lastly,
x30 3x0 =
=
=
=

(u + v)3 3(u + v)
u3 + v 3 3(u + v)(uv 1)
u3 + v 3
d.

This demonstrates that the cubic equation

x3 3x = d hasa root x0 (2, ). Lastly, we

p
1
1
define = ln(x0 1) > 0, = ln cb (x0 2)(x
x10 1 and = a cb (x0 2)(x
.
0 +1)
0 +1)
It is now a routine to check that if Z N (, 2 ) and X = + eZ , then E(X) = m1 ,
E(X 2 ) = m2 and E(X 3 ) = m3 .

4.3

4.3.1

Skew-Normal Distributions and Log-Skew-Normal

Distributions
Standard Skew-Normal Distributions

In this section, we firstly define a standard skew-normal distribution, then we compute the
first three moments of the standard skew-normal distribution. In the second half of this
section, we consider affine transformation of the standard skew-normal distribution.
Proposition 4.3.1. Let : R R, : R R be the probability density function and
the culmulative distributionRfunction of the standard normal distribution, that is, (x) =
1 2
1 2
x
1 e 2 x and (x) = 1
e 2 t dt. For each R, define f : R R by f (x) =
2
2
R
2(x)(x). Then f (x) > 0 for each x R and f (x)dx = 1. Hence, f is the
probability density function of some distribution.
R
Proof. Clearly f (x) > 0 for each x R. It remains to show that f (x)dx = 1.
R
Define g : R R by g() = (x)(x)dx. Note that the integrand is positive with
(x)(x) (x) and is integrable, so g() is well-defined. Observe that
Z
0
g () =
(x)(x)xdx = 0,

33

because the integrand is an odd function, |(x)(x)x| |(x)x|, and the right hand side
of the above first equation is integrable. Therefore, g is a constant function. In particular,
we have
Z
1
1
(x)dx = .
g() = g(0) =
2
2
R
It follows that f (x)dx = 1.
Definition 4.3.1. The standard skew-normal distribution with parameter is the distribution with probability density function f defined in Proposition 4.3.1.
Remark 4.3.1. If = 0, the skew-normal distribution defined in above reduces to the
standard normal distribution.
Proposition 4.3.2. Consider the standard skew-normal distribution with parameter and
probability density function f . Then:
(a) Moment of any order exists,
(b) The first three moments are given by
r
Z
xf (x)dx =

m1 ,
Z

m2 ,

1 + 2

x2 f (x)dx = 1,

r
Z
3

m3 ,

x f (x)dx =




2
2

3
.

1 + 2
1 + 2

Proof. Let and respectively be the probability density function and the culmulative
distribution function of the standard normal distribution. For any non-negative integer n,
we have that
Z
Z
n
|x |f (x)dx
|xn | 2(x)dx < ,

since the moment of any order exists for the standard normal distribution. Therefore
the moment of any order exists for the standard skew-normal distribution with parameter
. Let m1 , m2 , m3 be the first, the second, and the thrid moments of theR skew-normal

distribution with parameter respectively. Define g1 : R R by g1 () = xf (x)dx,

34

then
g10 ()

=
Z

d
[2(x)(x)] dx
d

x 2(x)(x)xdx

Z
1 2 1 (1+2 )x2
=
x e 2
dx




Z
x
1
1
12 (1+2 )x2
12 (1+2 )x2

e
e
dx
=
+

1 + 2
1 + 2

Z
1 2
1
1
1

e 2 y dy
=
2
1 + 1 + 2

 32

1
1
2
=

1+
r
1
2

=
.
(1 + 2 ) 23
Therefore, we have
g1 () =
=
=
=

r Z
2
d
3

(1 + 2 ) 2
r Z
2
sec2 d
( let =tan )

sec3
r
2
(sin + C)

r 

2

+C .

1 + 2

Observe that since the function x(x) is odd, we have

Z
g1 (0) =
x(x)dx = 0.

It follows that C = 0. Hence m1 = g1 () = 2 1+

2.
R 2
Next, we define g2 : R R, g2 () = x f (x)dx. We have that
Z
0
g2 () =
2x2 (x)(x)xdx = 0,

35

since the integrand is an odd function. Therefore g2 is a constant function. In particular,

Z
x2 (x)dx = 1.
g2 () = g2 (0) =

Hence m2 = 1.
R
Lastly, we define g3 : R R, g3 () = x3 f (x)dx, then
Z
0
x3 2(x)(x)xdx
g3 () =

Z
1 4 1 (1+2 )x2
=
dx
x e 2




Z
1
x3 1 (1+2 )x2
3x2
12 (1+2 )x2
2
e
dx
e
=
+

1 + 2
1 + 2

Z
1
1
3
2 2
=
x2 e 2 (1+ )x dx

2
1+


 32

1
3
1
=

2
1 + 2
1 + 2
r
1
2

.
= 3
(1 + 2 ) 25
Hence, we have
r
g3 () = 3

2
3

r
2
3

r
2
3

r
2
3

d
5

(1 + 2 ) 2
Z

sec2 d
(let =tan )
sec5

cos3 d


1 3
sin sin + C
=
3



2
=

1
+ C.
3(1 + 2 )
1 + 2
R
Note that x3 (x) is an odd fuction, so g3 (0) = x3 (x)dx = 0. It follows that C = 0
q
n
o
2
2
and hence m3 = 1+2 3 1+2 .
=

36

We can also compute the mean, the variance, and the skewness of the standard skewnormal distribution.
2
Proposition 4.3.3. The mean SN K , the variance SN
K , and the skewness of the standard
skew-normal distribution with parameter are respectively given by:
r
2

,
SN K =

1 + 2
2
2
2

SN
=
1

,
K
r 1 + 2

 
3
4

2
1

1
.
skewness =
3

1 + 2

Proof. The mean is just the first moment, and hence

r
2

1 + 2
For the variance, we have that
variance 2 = second moment mean2
2
2
.
= 1
1 + 2
Denote the probability density function of the standard skew-normal distribution with
parameter by f , and denote the first, the second, and the third moments by m1 , m2 ,
and m3 respectively. We have that
R
(x )3 f (x)dx

skewness =
3


1
2
3
=
m

3m
+
3
m

3
2
1
3
!3
r
r
r



2
1
2

3
=

1+2

3
1 + 2

1 + 2
1 + 2
1 + 2
r


1
2

2
4
2
=

3
3+
3

1 + 2
1 + 2
1 + 2
r

 
3
1
2
4

1
.
3

1 + 2

37

4.3.2

Affine Transformation of Skew-Normal Distributions

Since a standard skew-normal distribution is governed by one variable only, it is quite

restrictive. We increase the degree of freedom by considering an affine transformation of
it.
Definition 4.3.2. Let X be a random variable, distributed standard skew-normally with
parameter . Let R and > 0. Define Y = X +. We say that Y is a random variable,
distributed skew-normally with parameters , , and . is called the scale parameter while
is called the location parameter. We are ready to derive the mean, the variance, and the
skewness of Y.
Proposition 4.3.4. Let Y be a random variable distributed skew-normally with parameters
, , and . Let and be the probability density function and the culmulative distribution
function of the standard normal distribution. Then the mean, the variance, the skewness,
and the probability density function Y of Y are respectively given by:
r

+ ,
mean =

1 + 2


2
2
2
2
variance = 1
,
1 + 2
q

 3
4
2

1
1+2

,
skewness =

32
2
1 2 1+
2




y
1
y
Y (y) =
2

.

Proof. Define X = (Y )/, then X is distributed with the standard skew-normal distribution with parameter . That the mean, the variance, and the skewness of Y are given
as above follows immediately from the general theory of probability and the results in the
previous proposition. For example, the variance of Y is 2 times the variance of X, and the
skewness of Y is the same as the skewness of X. Denote the probability density function
of X by X . We have that


y
.
P [Y y] = P [X + y] = P X

Differentiating both sides with respect to y, we obtain







y
1
1
y
y
Y (y) = X
= 2

.

38

Remark 4.3.2. We can express the

, , and in terms of the moments
 parameters

of Y. More precisely, let mk = E Y k , for k = 1, 2, 3, be the first three moments of
Y , then we write the mean, the standard deviation , and the
p skewness in terms of the
moments. For example, mean = m1 , standard deviation = m2 m21 , and skewness =
(m3 3m1 m2 + 2m31 )/ 3 . Now denote the skewness by c, and solve for in the following
equation:
r 

 32
3

2 4

2
2

1
=c 1
.

1 + 2
1 + 2
 q

 13
1
2
4

, then the equation is simplified

If c 6= 0, we let = 1+2 and a = c 1
q
1
to a = 1 2 2 . Therefore 2 = a2 +
2 . Notice that a and are of the same sign and a

and c are of the same sign, so and c are of the same sign. Hence we have:

12 2 ,
if the given skewness c > 0,
a +
=
12 2 , if the given skewness c < 0.
a +

Note that and are of the same sign, so =

1 2

. If c = 0, we clearly have = 0.

Once we obtain , is given by

=q
1

2
1+2

Finally, is given by
r
= m1

1 + 2

Lastly, we derive the moment generating function of the standard skew-normal distribution with parameter . This result permits us to study the moments of log-skew-normal
distribution in the next section.
Proposition 4.3.5. Let f be the probability density function of the standard skew-normal
distribution with parameter . Denote the probability density function and the cumulative
distribution function of the standard normal distribution by and respectively. Then the
moment generating function of this distribution is given by


Z
u2
u
ux
(u) ,
e f (x)dx = 2e 2
.
1 + 2

39

Proof. Fix u and define g() =

eux f (x)dx, then we have

eux 2x(x)(x)dx

g () =

Z
1 ux 1 (1+2 )x2
xe e 2
=
dx

Z
u
u2
2 1
1 12 (1+2 )(x 1+
2 ) + 2 1+2
xe
dx
=

Z
u2
1 12 1+
1 (1+2 )(x u 2 )2
2
1+
e
dx.
xe 2
=

Now let y =

u
1 + 2 (x 1+
2 ), then the integral becomes:
Z
1 (1+2 )(x u 2 )2
1+
xe 2
dx


Z 
1 2
u
1
y

+
e 2 y
dy
=
2
2
1+
1+
1 + 2

u
2.
= 0+
3
(1 + 2 ) 2

Hence, we obtain
r

2
1
u
2
u
2 1+2

e
.
(1 + 2 ) 23
2
R
1
u
u
2 1+2
Next, we compute the indefinite integral

e
d. Let y =
3
2

g 0 () =

u ,
1+2

(1+ ) 2

dy = u
Also, y 2 =
have

2 u2
1+2

1 + 2
1+

1+2

d =

ud
3

(1 + 2 ) 2

then

u y
u
which implies 2 = u2yy2 . It follows that 12 1+
. Therefore, we
2 =
2
Z
Z
1
u2
u2 y 2
u2
u

2
2 1+
d = e 2 dy = e 2 2(y) + C 0
3 e
(1 + 2 ) 2

and hence, we have

u2
2

u2
2

g() = 2e (y) + C = 2e
40

1 + 2


+ C.

On the other hand,

Z

g(0) =
Z

=
Z

eux (x)dx
1 2
1
eux e 2 x dx
2
1 1 (xu)2 + u2
2 dx
e 2
2

u2
2

= e .
u2
2

It follows that C = 0 and hence (u) = g() = 2e

4.3.3

u
1+2

Log-Skew-Normal Distributions

One drawback of the skew-normal distributions is that the probability of the skew-normally
distributed random variable being negative is positive. This phenomenon makes them
inappropriate to model the average stock price since the latter one is always positive. One
way to overcome this difficulty is to consider the exponential of a skew-normal distribution.
Definition 4.3.3. We say that a random variable X is log-skew-normally distributed if
ln(X) is skew-normally distributed.
Proposition 4.3.6. Let X be a log-skew-normally distributed random variable with ln(X)
distributed skew-normally with parameters , , . Then the first three moments of X are
respectively given by
 



2

,
E [X] = 2 exp +
2
1 + 2


 2


2
2
,
E X
= 2 exp 2 + 2
1 + 2

 

 3
9 2
3
E X
= 2 exp 3 +
,
2
1 + 2
where is the culmulative distribution function of the standard normal distribution.
Proof. Let Y = ln(X)
which is a random
variable distributed standard skew-normally


with parameter . Let (u) = E euY be the moment generating function of Y. Observe
41

that X = exp(Y + ), and hence, by the previous proposition, we have



E [X] = e E eY
= e ()


= e 2e
1 + 2
 


2

,
= 2 exp +
2
1 + 2
 


E X 2 = e2 E e2Y
2
2

= e2 (2)

2
= e 2e
1 + 2




2
2
= 2 exp 2 + 2
,
1 + 2
 


E X 3 = e3 E e3Y
4 2
2

= e3 (3)

3
= e 2e
1 + 2
 


3
9 2

= 2 exp 3 +
.
2
1 + 2
9 2
2

Remark 4.3.3. Unlike the skew-normal case, it is very hard, if not impossible, to express
the parameters , , in terms of the moments of X analytically. Here, we solve the
problem numerically. Let mk = E[X k ], for k = 1, 2, 3, be the first three moments of X. Let

F1 (, , )
F (, , ) = F2 (, , ) ,
F3 (, , )
where
2
F1 (, , ) = 2 exp +
2


m1 ,
1 + 2




2
2
F2 (, , ) = 2 exp 2 + 2
m2 ,
1 + 2

 

9 2
3
F3 (, , ) = 2 exp 3 +

m3 .
2
1 + 2
42

Then the Jacobi matrix of F at (, , ) is

dF (, , ) =

F1

F
2
F3

F1

F2

F3

F1

F2
,
F3

where all partial derivatives are evaluated at the point (, , ). We solve the equation
F (, , ) = (0, 0, 0)T by multi-dimensional Newtons method. Let xn = (n , n , n )T , for
n = 0, 1, 2, . . . , where x0 is an initial guess, and xn+1 is defined implicitly in terms of xn by
the equation
dF (xn )(xn+1 xn ) = F (xn ).
The 3-by-3 equation in above is solved numerically. In the following, we work out the
partial derivatives.


2

F1
+ 2
= 2e

3 ,
2

1+
(1 + 2 ) 2


F2
2

2+2 2
= 4e

3 ,

1 + 2
(1 + 2 ) 2


F3
3

3+ 92 2
= 6e

3 ,

1 + 2
(1 + 2 ) 2


2
F1

+ 2
= 2e

,

1 + 2


F2
2
2+2 2
= 4e

,

1 + 2


F3
3
3+ 92 2
= 6e

,

1 + 2






2
F1

+ 2
+

,
= 2e

1 + 2
1 + 2
1 + 2






F2
2
2

2+2 2
= 4e
2
+

1 + 2
1 + 2
1 + 2






F3
3
3

3+ 92 2
3
= 6e
+

1 + 2
1 + 2
1 + 2

43

Chapter 5
Pricing Asian Options by Moments
Fitting
In this chapter, we derive the pricing formulae for Asian call and Asian put options, where
the distribution of the average price at the maturity T , i.e., AT , is approximated by the
respective distributions discussed in the previous chapter. Here, we consider the cases
where AT is approximated by log-normal distributions, shifted-log-normal distributions,
skew-normal distributions, and log-skew-normal distributions. Throughout the chapter,
we denote the maturity and the strike of Asian options by T and K respectively.

5.1

Approximation by log-normal distributions

Suppose that we approximate the distribution of the average price AT at the maturity
T by a log-normal distribution, i.e., AT eZ , where Z N (, 2 ), in the sense that
the first two moments of AT and eZ match. Then the pricing formulae for Asian call
and Asian put options reduce to the usual Black-Schole formulae. Let (x) and (x) be
the probability density function and the cumulative distribution function of the standard

44

normal distribution respectively. Then the price of Asian call at the time t = 0 is:


(AT K)+
Price of Asian Call = E
BT
rT
= e E[(AT K)+ ]
erT E[(eZ K)+ ]
Z
1
2
1
rT
= e
(ex K)
e 22 (x) dx
2
ln K
rT
= e (I1 I2 ),
R
R
1
1
2
2
1
1
where I1 and I2 are the integrals ln K ex 2
e 22 (x) dx and ln K K 2
e 22 (x) dx
respectively. For the first integral, we have

1
2 2
1
1
2

e 22 [x(+ )] e 2 ( +2) dx
ZlnK 2
1 2
1
1
2
e 2 y dy e 2 ( +2)
=
ln K(+ 2 )
2




1
ln K ( + 2 )
2 +2)
(
= e2
1

2
1
ln K +
2
).
= e 2 ( +2) (

I1 =

For the second integral, we have

Z
I2 = K

(y)dy
ln K

ln K
= K 1 (
)

ln K
= K(
).

Therefore, we obtain
Price of Asian call
1 2
ln K + 2
ln K
erT + 2 (
) KerT (
).

45

Lastly, by observing that

we obtain

(AT K)+
BT

T)
(KA
=
BT

AT K
BT

and taking expectation on both sides,

Price of Asian put

1 2
ln K 2
ln K
) erT + 2 (
).
KerT (

5.2

Approximation by shifted-log-normal distributions

The pricing formulae for Asian options, where AT is approximated by a shifted-log-normal

distribution are very similar to those in the previous section. Suppose that AT is approximated by a shifted-log-normal distribution AT + eZ , where R and Z N (, 2 ),
in the sense that the first three moments of AT and + eZ match. Let K1 = K , then
the price of Asian call option is
Price

 of Asian+ call
(AT K)
= E
BT
 Z

rT
e E (e K1 )+



erT e+ 12 2 K1 ,


=
erT + 21 2 ln K1 +2 K1 erT

if K1 0


ln K1

if K1 > 0

Similarly, the price of Asian put option is

Price of Asian put


(K AT )+
= E
BT


rT
e E (K1 eZ )+
(
0,




=
ln K1
ln K1 2
rT + 21 2
rT
K1 e
e

5.3

if K1 0
if K1 > 0

Approximation by skew-normal fitting

Suppose that we approximate AT by a skew-normal distribution: AT X + , where

> 0, R, and X is standard skew-normally distributed with parameter , in the sense
46

that the first three moments of AT and X + match. Recall that the random variable
)( x
),
X + admits a probability density function (x), given by (x) = 2 ( x

where and are the probability density fucntion and the cumulative distribution function
of the standard normal distribution respectively. Note that, due to the complexity of (x),
we are unable to obtain closed-form pricing formulae for Asian call nor Asian put options.
Rather, we compute the expectation by a numerical method. For the price of Asian call
we get
Price

 of Asian+ call
(AT K)
= E
BT


rT
e E (X + K)+
Z
rT
= e
(x K)(x)dx,
K

where the integral is calculated numerically by Simpsons rule. Similarly, the price of Asian
put is
Price of Asian put


(K AT )+
= E
BT


rT
e E (K (X + ))+
Z K
rT
(K x)(x)dx,
= e

where again the integral is computed by Simpsons rule.

5.4

Approximation by log-skew-normal fitting

Suppose that we approximate AT by a log-skew-normal distribution: AT eX+ , where

> 0, R, and X is standard skew-normally distributed with parameter , in the sense
that the first three moments of AT and eX+ match. Denote Y = X + and Z = eY .
Observe that for z > 0, P [Z z] = P [Y ln z]. Differentiating both sides of the equation
with respect to z, we obtain:
d
1
P [Z z] = Y (ln z)
dz
z
2 ln z
ln z 1
=
(
)(
) .

z
47

In another word, Z has a probability density function (x), given by

(
2
( ln x
)( ln x
) x1 , if x > 0

(x) =
.
0,
if x 0
Now the prices of Asian call and Asian put can be computed numerically as:
Z
rT
Price of Asian call = e
(x K)(x)dx,
K

and
rT

(K x)(x)dx.

Price of Asian put = e

48

Chapter 6
Numerical Results
In this chapter, we present the prices of various Asian call options computed by Monte
Carlo simulation and by moments matching approaches discussed in Chapter 5. The prices
obtained by the Monte Carlo simulation act as benchmark values and are used to compare
with those obtained by various moments matching methods.
For the GBM model, we sample the stock price 1000 times per year and compute the
average; while for the Heston model, we sample the stock price 100 times per year and
computer the average. More precisely, let T be the maturity of the option. For the GBM
k
k
T }, tk = 1000
, for k = 1, 2, . . . , n, and
model, let nP= max{k Z | k 0 and 1000
k
let
n
=
max{k

Z | k 0 and 100
T },
let AT = n1 nk=1 Stk . For the Heston model,
P
n
1
k
tk = 100 , for k = 1, 2 . . . , n, and let AT = n k=1 Stk . In this way, these discretely averaging
average prices can be considered good approximations of continuously averaged prices.
We employ Milsteins scheme for the Heston model. Recall that if = (t, x), =
(t, x) are C 1,2 functions and (Xt )t is an Itos process satisfying the SDE:
dXt = (t, Xt )dt + (t, Xt )dWt ,
Milsteins scheme states that for t 0 and small h > 0, we have:



1
Xt+h Xt + (t, Xt )h + (t, Xt )(Wt+h Wt ) + (t, Xt )x (t, Xt ) (Wt+h Wt )2 h .
2
We recall the Heston model:

49



p
dSt = St rdt + Vt dWt1 ,
p
dVt = a (b Vt ) dt + c Vt dWt2 ,
where (Wt1 , Wt2 )t is a two dimensional Wiener process with dWt1 dWt2 = dt. We apply
Milsteins scheme to the variance process (Vt )t . For t 0 and small h > 0, we have:
p

c2  2

2
Vt+h Vt + a(b Vt )h + c Vt Wt+h
Wt2 +
(Wt+h Wt2 )2 h .
4
To be more precise, let 0 = t0 < t1 < . . . < tN be time points. Let (z1 , z2 , . . . , z2N )
be a realization of a random vector (Z1 , Z2 , . . . , Z2N ), where Z1 , Z2 , . . . , Z2N are independent, identically distributed random variables with the standard normal N (0, 1) as

(1)
(2)
the common distribution. Define yi = ti ti1 z2i1 and yi = ( ti ti1 z2i1 ) +
p

1 2 ( ti ti1 z2i ) , for i = 1, 2, . . . , N. Suppose that we have already generated Sti

and Vti for i = 0, 1, . . . , k 1, and that Vti > 0 for i = 0, 1, . . . , k 1. We define


q
c2  (2) 2
(2)
Vtk := Vtk1 + a(b Vtk1 )(tk tk1 ) + c Vtk1 yk +
yk
(tk tk1 ) ,
4
and


q
1
(1)
.
Stk := Stk1 exp (r Vtk1 )(tk tk1 ) + Vtk1 yk
2


If Vtk < 0, we discard all the results, choose another random sample (z1 , z2 , . . . , z2N ), and
repeat the whole procedure.
When computing the prices of various Asian options by Monte Carlo simulation, we
use 100 million paths for both the GBM model and the Heston model.
In the following, the percentage error is defined by
Percentage Error =

6.1

Price by moments matching Price by Monte Carlo simulation

100%.
Price by Monte Carlo simulation

Pricing of Asian Call Options, for the GBM Model

In this section, we assume that the stock price follows the GBM model and show the prices
of various Asian call options computed by Monte Carlo simulation, matching moments by
50

log-normal distributions, matching moments by shifted-log-normal distributions, matching moments by skew-normal distributions, and matching moments by log-skew-normal
distributions. In the following, the initial stock price S0 for all cases is set to 100.00.
In Table 6.1, columns T , r, Sigma, Strike, MC, Std Err, LN, SLN, SN, LSN, and Err
(%) represent maturity T , constant short rate r, volatility , strike K, prices by Monte
Carlo simulation, standard error of the simulation, prices by log-normal fitting, prices by
shifted-log-normal fitting, prices by skew-normal fitting, prices by log-skew-normal fitting,
and percentage error respectively.
Table 6.1: Pricing of Asian call options, GBM model
T

Sigma

0.15

0.05

0.20

0.25
1.00
0.15

0.08

0.20

0.25

Strike

MC

Std Err

LN

Err (%)

SLN

Err (%)

50
75
100
125
150
50
75
100
125
150
50
75
100
125
150
50
75
100
125
150
50
75
100
125
150
50
75
100
125
150

49.9790
26.1984
4.6868
0.0429
0.0000
49.9787
26.2056
5.7638
0.2488
0.0030
49.9785
26.2510
6.8529
0.6669
0.0324
49.9480
26.8702
5.5127
0.0686
0.0001
49.9478
26.8747
6.5180
0.3327
0.0048
49.9476
26.9083
7.5545
0.8165
0.0440

8.52E-04
8.52E-04
6.06E-04
5.51E-05
8.74E-07
1.14E-03
1.14E-03
7.97E-04
1.66E-04
1.67E-05
1.43E-03
1.42E-03
9.96E-04
3.21E-04
6.71E-05
8.43E-04
8.43E-04
6.42E-04
7.03E-05
1.54E-06
1.13E-03
1.13E-03
8.33E-04
1.93E-04
2.10E-05
1.41E-03
1.40E-03
1.03E-03
3.55E-04
7.82E-05

49.9821
26.2017
4.7019
0.0377
0.0000
49.9821
26.2122
5.7875
0.2290
0.0020
49.9821
26.2703
6.8886
0.6285
0.0243
49.9526
26.8748
5.5323
0.0618
0.0000
49.9526
26.8818
6.5478
0.3109
0.0033
49.9526
26.9262
7.5977
0.7777
0.0341

0.01
0.01
0.32
-12.03
-50.79
0.01
0.03
0.41
-7.95
-34.39
0.01
0.07
0.52
-5.76
-24.99
0.01
0.02
0.36
-9.80
-46.40
0.01
0.03
0.46
-6.53
-31.18
0.01
0.07
0.57
-4.74
-22.55

49.9821
26.2016
4.6907
0.0429
0.0000
49.9821
26.2088
5.7690
0.2491
0.0030
49.9821
26.2535
6.8598
0.6682
0.0320
49.9526
26.8748
5.5176
0.0687
0.0001
49.9526
26.8793
6.5240
0.3333
0.0047
49.9526
26.9124
7.5620
0.8184
0.0436

0.01
0.01
0.08
0.04
-4.09
0.01
0.01
0.09
0.12
-1.96
0.01
0.01
0.10
0.20
-1.20
0.01
0.02
0.09
0.18
-3.16
0.01
0.02
0.09
0.19
-1.59
0.01
0.02
0.10
0.24
-0.91

51

LSN

Err (%)

49.9821
0.01
49.9821
26.2017
0.01
26.2016
4.6817
-0.11
4.6887
0.0421
-1.75
0.0435
0.0000
-66.04
0.0000
49.9821
0.01
49.9821
26.2109
0.02
26.2096
5.7592
-0.08
5.7649
0.2496
0.30
0.2506
0.0018
-40.87
0.0032
49.9821
0.01
49.9821
26.2543
0.01
26.2570
6.8582
0.08
6.8528
0.6782
1.69
0.6692
0.0240
-25.89
0.0333
49.9526
0.01
49.9526
26.8749
0.02
26.8748
5.5085
-0.08
5.5162
0.0684
-0.17
0.0695
0.0000
-58.68
0.0001
49.9526
0.01
49.9526
26.8811
0.02
26.8799
6.5109
-0.11
6.5207
0.3362
1.08
0.3346
0.0031
-34.87
0.0050
49.9526
0.01
49.9526
26.9138
0.02
26.9154
7.5528
-0.02
7.5561
0.8339
2.14
0.8186
0.0346
-21.45
0.0451
Continued on next page

SN

Err (%)

0.01
0.01
0.04
1.59
8.64
0.01
0.02
0.02
0.72
4.85
0.01
0.02
0.00
0.34
2.65
0.01
0.02
0.06
1.31
8.14
0.01
0.02
0.04
0.58
4.34
0.01
0.03
0.02
0.26
2.39

Table 6.1 Continued from previous page

T

Sigma

0.15

0.05

0.20

0.25
3.00
0.15

0.08

0.20

0.25

Strike

MC

Std Err

LN

Err (%)

SLN

Err (%)

SN

Err (%)

LSN

Err (%)

50
75
100
125
150
50
75
100
125
150
50
75
100
125
150
50
75
100
125
150
50
75
100
125
150
50
75
100
125
150

49.8250
28.3322
9.4629
1.4130
0.1132
49.8247
28.4953
11.0936
2.8727
0.5691
49.8264
28.8801
12.7795
4.5346
1.4202
49.5726
29.9160
11.7674
2.2739
0.2387
49.5724
30.0050
13.0935
3.9207
0.8981
49.5731
30.2625
14.5467
5.6658
1.9415

1.43E-03
1.43E-03
1.13E-03
4.76E-04
1.31E-04
1.92E-03
1.89E-03
1.49E-03
8.24E-04
3.70E-04
2.42E-03
2.34E-03
1.87E-03
1.21E-03
6.95E-04
1.39E-03
1.38E-03
1.19E-03
5.94E-04
1.89E-04
1.86E-03
1.84E-03
1.55E-03
9.40E-04
4.58E-04
2.35E-03
2.29E-03
1.92E-03
1.31E-03
7.97E-04

49.8283
28.3456
9.5180
1.3736
0.0943
49.8285
28.5503
11.1881
2.8255
0.5085
49.8345
29.0093
12.9293
4.4997
1.3180
49.5772
29.9256
11.8282
2.2480
0.2115
49.5773
30.0451
13.1997
3.9011
0.8331
49.5807
30.3665
14.7126
5.6705
1.8487

0.01
0.05
0.58
-2.79
-16.67
0.01
0.19
0.85
-1.64
-10.65
0.02
0.45
1.17
-0.77
-7.19
0.01
0.03
0.52
-1.14
-11.39
0.01
0.13
0.81
-0.50
-7.24
0.02
0.34
1.14
0.08
-4.78

49.8283
28.3346
9.4671
1.4151
0.1125
49.8283
28.4938
11.0999
2.8798
0.5688
49.8296
28.8705
12.7883
4.5500
1.4244
49.5772
29.9201
11.7713
2.2779
0.2382
49.5772
30.0059
13.0980
3.9298
0.8993
49.5778
30.2555
14.5518
5.6824
1.9488

0.01
0.01
0.04
0.15
-0.63
0.01
-0.01
0.06
0.25
-0.06
0.01
-0.03
0.07
0.34
0.30
0.01
0.01
0.03
0.18
-0.21
0.01
0.00
0.03
0.23
0.14
0.01
-0.02
0.04
0.29
0.38

49.8283
28.3356
9.4435
1.4486
0.1001
49.8283
28.4399
11.1040
2.9962
0.5519
49.8283
28.5178
12.9393
4.8357
1.4590
49.5772
29.9214
11.7376
2.3219
0.2292
49.5772
29.9719
13.0453
4.0658
0.9097
49.5772
29.9636
14.5906
5.9996
2.0478

0.01
0.01
-0.21
2.52
-11.54
0.01
-0.19
0.09
4.30
-3.03
0.00
-1.25
1.25
6.64
2.74
0.01
0.02
-0.25
2.11
-3.97
0.01
-0.11
-0.37
3.70
1.29
0.01
-0.99
0.30
5.89
5.48

49.8283
28.3371
9.4637
1.4126
0.1147
49.8284
28.5052
11.0918
2.8690
0.5717
49.8310
28.8952
12.7744
4.5270
1.4219
49.5772
29.9215
11.7721
2.2723
0.2405
49.5773
30.0148
13.0965
3.9157
0.8999
49.5786
30.2782
14.5470
5.6570
1.9412

0.01
0.02
0.01
-0.03
1.31
0.01
0.03
-0.02
-0.13
0.45
0.01
0.05
-0.04
-0.17
0.13
0.01
0.02
0.04
-0.07
0.75
0.01
0.03
0.02
-0.13
0.20
0.01
0.05
0.00
-0.16
-0.01

In Table 6.1, we notice that if the option is deeply in-the-money, the price computed
by any method of moments fitting is highly accurate. This phenomenon will be explained
in the next chapter. The accuracy of all methods of moments fitting decreases as we move
from deeply in-the-money towards deeply out-of-money. Since the method of moments
fitting by the log-normal distribution only matches the first moments of AT , it gives, as
expected, the least accurate results. For the other three methods, all of them match the
first three moments of AT , so they yield more accurate results. For the method of moments
fitting by the skew-normal distribution, we observe that the skew-normal distribution has
a defect that it has a positive probability of being negative, so it gives less accurate results
than the methods of fitting by the shifted-log-normal distribution or by the log-skew-normal
distribution. This problem is especially serious when the option is deeply out-of-money.
Regarding the method of moments fitting by the shifted-log-normal distribution and the
method of moments fitting by the log-skew-normal distribution, for options with short
maturity, the former one gives slightly better results. For options with long maturity, both
methods give results with similar degree of accuracy.

52

6.2

Pricing of Asian Call Options, for the Heston

Model

In this section, we assume that the stock price follows the Heston model and show the
prices of various Asian call options computed by Monte Carlo, and matching moments by
log-normal distribution. In the following, the constant short rate r, the initial stock price
S0 , and the mean reversion rate a are set to 0.05, 100.00, and 1.0 respectively for all cases.
In Table 6.2, columns T , V0 , b, c, rho, strike, MC, Std Err, LN, Err (%) represent
maturity T , initial variance V0 , anchor value b, vol-vol c, correlation of Wiener processes ,
strike K, prices by Monte Carlo method, standard error of simulation, prices by log-normal
fitting, and percentage error respectively.

53

Table 6.2: Pricing of Asian call options, Heston model

T

V0

rho

0.50
0.0625

0.0625

0.0225
-0.50

1.00
0.50
0.16

0.16

0.1
-0.50

0.50
0.0625

0.0625

0.0225
-0.50

3.00
0.50
0.16

0.16

0.1
-0.50

strike

MC

Std Err

LN

Err (%)

50
75
100
125
150
50
75
100
125
150
50
75
100
125
150
50
75
100
125
150
50
75
100
125
150
50
75
100
125
150
50
75
100
125
150
50
75
100
125
150

49.9770
26.2444
6.8477
0.6903
0.0377
49.9771
26.2549
6.8536
0.6421
0.0277
49.9757
26.8304
10.1334
2.9743
0.7883
49.9784
26.9515
10.1225
2.7223
0.5857
49.8262
28.8547
12.7694
4.5851
1.4809
49.8274
28.9062
12.7868
4.4803
1.3603
49.9547
31.0331
17.9288
10.1811
5.8516
50.0368
31.2507
17.8744
9.7521
5.2215

1.43E-03
1.42E-03
1.00E-03
3.33E-04
7.44E-05
1.42E-03
1.41E-03
9.88E-04
3.11E-04
6.03E-05
2.34E-03
2.25E-03
1.69E-03
9.89E-04
5.23E-04
2.29E-03
2.18E-03
1.60E-03
8.74E-04
4.03E-04
2.43E-03
2.36E-03
1.89E-03
1.23E-03
7.22E-04
2.41E-03
2.32E-03
1.85E-03
1.18E-03
6.61E-04
4.13E-03
3.88E-03
3.35E-03
2.74E-03
2.20E-03
3.90E-03
3.62E-03
3.05E-03
2.41E-03
1.83E-03

49.9797
26.2677
6.8846
0.6271
0.0242
49.9797
26.2652
6.8562
0.6144
0.0231
49.9849
27.0406
10.2997
2.8310
0.6232
49.9836
26.9714
10.1043
2.6778
0.5596
49.8314
28.9723
12.8215
4.3937
1.2586
49.8306
28.9358
12.7073
4.2812
1.1964
50.1537
31.6623
18.3896
10.1693
5.4922
50.0655
31.2259
17.6493
9.3808
4.8290

0.01
0.09
0.54
-9.16
-35.81
0.01
0.04
0.04
-4.32
-16.82
0.02
0.78
1.64
-4.82
-20.94
0.01
0.07
-0.18
-1.63
-4.47
0.01
0.41
0.41
-4.17
-15.01
0.01
0.10
-0.62
-4.45
-12.05
0.40
2.03
2.57
-0.12
-6.14
0.06
-0.08
-1.26
-3.81
-7.52

In Table 6.2, we notice that the method of moments fitting by the log-normal distribution gives accurate results if the option is deeply in-the-money. However, the accuracy of
the results decreases as we move from deeply in-the-money towards deeply out-of-money.
Also, we observe that initial volatility V0 , the anchor value b and the correlation have
an impact on the accuracy. In general, the method of moments fitting by the log-normal
54

distribution yields better result if the initial volatility V0 and the anchor value b are small.
Moreover, in general, for the Heston model, if the correlation is negative, the average price
A is more accurately approximated by the log-normal distribution, and hence it results in
a more accurate pricing.
Remark 6.2.1. In the above, we have tested only the pricing of Asian call options but not
Asian put options. This is because the prices of Asian put options can be computed by putcall parity. MoreR precisely, let T be the maturity, and let AT be the average price at time
T
T , i.e., AT = T1 0 Su du. Let r be a constant deterministic short rate and let Bt = exp(rt)
be the risk-free banking account. Let K be the strike. Let C and P be the payoffs
of an Asian call and an Asian put respectively, with maturity T and strike K. That is,
C = max(AT K, 0) and P = max(KAT , 0). Denote the prices of the Asian call and the
Asian put
i time 0h by Ci(0) and P (0) respectively. Observeh that
i C P = AT K,
h at the
Q AT K
Q AT
Q C P
so E
=E
, and hence C (0) P (0) = E BT KerT . The term
BT
h i BT
 
Q AT
E BT is computed by observing that BStt is a martingale, namely
t

AT
BT

erT
=
T

E
0




St rt
S0 erT rT
e dt =
e 1 .
Bt
rT

In the above, E Q [ ] denotes the expectation under the Q-measure.

55

Chapter 7
Conclusions
Firstly, we consider the pricing results where the stock price process follows the standard
GBM model. By comparing the numerical results obtained from Monte Carlo simulation,
matching moments by log-normal distributions, matching moments by shifted-log-normal
distributions, matching moments by skew-normal distributions, and matching moments
by log-skew-normal distributions, we conclude that, in general, the methods of matching
moments by shifted-log-normal distributions and by log-skew-normal distributions give the
most accurate results; while the method of matching moments by log-normal distributions
gives the least accurate results. In general, if the option is deeply in-the-money, all the
methods give very accurate results. This phenomenon can be explained as follow. Let T be
the maturity, K the strike, AT the average price at time T and r the constant deterministic
short rate. If K is small compared to AT , the payoff max(AT K, 0) of the Asian call
option can be approximated by AT well, and hence




AT
max(AT K, 0)
E rT ,
Price of the Asian call = E
erT
e
where the expectation in the last line is model independent. Regarding the methods of
matching moments by shifted-log-normal distributions and by log-skew-normal distributions, although both give fairly accurate results, fitting by shifted-log-normal distributions
results in a closed-form pricing formula; while fitting by log-skew-normal distributions requires solving a system of 3 3 non-linear equations, and computing the prices requires
a numerical integration. Due to the simplicity of the former one, that method is widely
adopted in the industry.
Secondly, we consider the pricing results where the stock price process follows the
Heston model. Due to the complexity of the model, we are only able to compute the first
56

moment of AT exactly and a good approximation of the second moment of AT . Therefore,

we consider the method of matching moments by the log-normal distributions only. Again,
this method gives accurate results for in-the-money options but inaccurate results for outof-money options.

57

Bibliography
[1] Martin Forde and Antoine Jacquier, (2010). Robust Approximations for Pricing Asian
Options and Volatility Swaps Under Stochastic Volatility, Applied Mathematical Finance, Vol.17 No.3 241-259, July 2010.
[2] Sebastian del Bano Rollin, Albert Ferreiro-Castilla and Frederic Utzet (2009). A New
Look at the Heston Characteristic Function, arXiv, arXiv:0902.2154.
[3] Fischer Black and Myron Scholes (1973). The Pricing of Options and Corporate Liabilities, The Journal of Political Economy, Vol. 81, No. 3 (May - Jun., 1973), P.637-654.
[4] John C. Cox, Jonathan E. Ingersoll and Stephen A. Ross. A Theory of the Term
Structure of Interest Rates, Econometrica, Volume 53, Issue 2 (Mar., 1985), 385-408.
[5] Steven L. Heston. A Closed-Form Solution for Options with Stochastic Volatility with
Applications to Bond and Currency Options, The Review of Financial Studies 6(2):
327-343.

58