Professional Documents
Culture Documents
Physica A: D. Lemmens, L.Z.J. Liang, J. Tempere, A. de Schepper
Physica A: D. Lemmens, L.Z.J. Liang, J. Tempere, A. de Schepper
Physica A
journal homepage: www.elsevier.com/locate/physa
Pricing bounds for discrete arithmetic Asian options under Lvy models
D. Lemmens a, , L.Z.J. Liang a , J. Tempere a,b , A. De Schepper c
a
Faculty of Applied Economics & StatUa Statistics Center, Universiteit Antwerpen, Prinsstraat 13, 2000 Antwerpen, Belgium
article
info
Article history:
Received 3 March 2010
Received in revised form 12 May 2010
Available online 27 July 2010
Keywords:
Asian options
Analytical bounds
Lvy models
abstract
Analytical bounds for Asian options are almost exclusively available in the BlackScholes
framework. In this paper we derive bounds for the price of a discretely monitored
arithmetic Asian option when the underlying asset follows an arbitrary Lvy process.
Explicit formulas are given for Kous model, Mertons model, the normal inverse Gaussian
model, the CGMY model and the variance gamma model. The results are compared with
the comonotonic upper bound, existing numerical results, Monte carlo simulations and in
the case of the variance gamma model with an existing lower bound. The method outlined
here provides lower and upper bounds that are quick to evaluate, and more accurate than
existing bounds.
2010 Elsevier B.V. All rights reserved.
1. Introduction
In this paper we investigate discretely monitored arithmetic Asian options, and we present a method to derive lower
and upper bounds for this type of derivatives under Lvy models. Afterwards, the method is applied to a selection of Lvy
models, including Kous model [1], Mertons model [2], the normal inverse Gaussian (NIG) model [3], the variance gamma
(VG) model [4] and the CGMY model [5].
The lower and upper bounds constructed in this paper are based on existing approaches in a BlackScholes framework, a
setting where next to other derivatives Asian options have been studied extensively. For the geometric Asian option an easy
to use closed form solution exists, see Refs. [6,7], but for arithmetic Asian options, an analytic closed form solution of the
same simplicity does not exist. However several trails for pricing arithmetic Asian options have been investigated. Among
the numerical methods the most efficient ones are probably the pde methods, see for example Refs. [812]. There also exists a
variety of cutting edge analytical techniques to find exact or approximate solutions, for more information see Refs. [1320]
for the continuous case and Refs. [9,2125] for the discrete case. These analytical techniques have their drawbacks, for
example they can be slow to evaluate, and when it concerns an approximation, the error might not be known. Fruitful
alternatives are methods allowing the calculation of preferably analytical, quick-to-evaluate bounds. For the lower bound
in the BlackScholes model we refer to Refs. [26,8]. Since these lower bounds are very accurate most of the work thereafter
concentrates on the upper bound, see for example Refs. [2729,25,20,19] and references therein. Dhaene et al. [28], discuss
a nearly analytical lower bound by conditioning on a variable , which is an appropriately chosen linear combination of
normally distributed random variables; these bounds are proven to be more accurate than the upper bound derived in the
same paper. Two improved comonotonic upper bounds for arithmetic Asian option are derived in Vanmaele et al. [25] and
one of them outperforms the bound obtained in Ref. [29].
Corresponding author.
E-mail addresses: Damiaan.Lemmens@ua.ac.be (D. Lemmens), Lingzhi.Liang@ua.ac.be (L.Z.J. Liang), Jacques.Tempere@ua.ac.be (J. Tempere),
Ann.Deschepper@ua.ac.be (A. De Schepper).
0378-4371/$ see front matter 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.physa.2010.07.026
5194
All of these methods concentrate on the BlackScholes model. However it is well known that this model does not meet
all the needs of financial markets [3036]. An important class of models constructed to improve the BlackScholes model
are the Lvy models [3741]. These models do not contain all the stylized facts of the dynamics of financial markets, as
e.g. stochastic volatility addressed in Refs. [35,36,31,36] is absent in the Lvy models. Nevertheless, the class of Lvy models
is valuable since they are more realistic than the BlackScholes model and there exist plenty of methods on how to price
(simple) financial products under these models.
As it comes to Asian option pricing, less results are available for Lvy models compared to the BlackScholes model.
In Ref. [42] closed form solutions for geometrically averaged Asian options are derived and a numerical scheme to price
arithmetically averaged Asian options is presented. In Ref. [43] analytic approximations and bounds for the Esscher price of
Asian options are derived within the NIG model. In Ref. [44] lower bounds for arithmetic Asian options are calculated for the
VG model, and in Ref. [45] comonotonic upper bounds are derived for the VG model, the NIG model and the Meixner model.
We contribute to the existing work by deriving new lower and upper bounds for discrete Asian options under a general
Lvy model. The derivation of the lower bound is a generalization of the method used in Ref. [8] in the BlackScholes
framework to the setting of a general Lvy process. As was also the case in the BlackScholes setting in Ref. [8], our lower
bound turns out to be very accurate. Furthermore this bound has the advantage of being very quick to evaluate, since for
this purpose we only need to find the root of a monotonic function and evaluate a numerical integral of a smooth function
composed out of elementary functions. The construction of the upper bound is based on the upper bound presented in
Ref. [29] which is an improvement of the upper bound presented in Ref. [8], with both papers being restricted to the
BlackScholes framework. This upper bound is associated with the lower bound as will be explained later. For at and in
the money options our upper bound, together with the lower bound, results in rather accurate prices for the Asian option,
for out of the money options the results are slightly inferior. Moreover, when the strike converges to zero both the lower
and this upper bound converge to the exact price. For Kous model, Mertons model, the NIG model and the CGMY model
we compare our bounds with the numerical results of Ref. [42], including a Monte Carlo simulation. For the VG model we
compare with the (model independent) lower bound of Ref. [44] and we also provide a Monte Carlo simulation. In addition
all bounds are compared to the comonotonic upper bound, which we calculated following the approach as explained in
general in Refs. [27,28] and in particular for Lvy processes in Ref. [45].
The paper is organized as follows. In Section 2 we discuss the model for the underlying process. In Section 3 our lower
bound for the arithmetic Asian option for a general Lvy model is derived. In Section 4 we derive the corresponding upper
bound and discuss the comonotonic upper bound. In Section 5 we apply these results to Kous model, Mertons model, the
NIG model, the VG model and the CGMY model. Section 6 contains a conclusion. Some of the derivations are discussed in
the appendix, as well as the construction of the comonotonic upper bound.
2. Setting the stage
The asset price at time t will be denoted by St . The arithmetic average A is defined by
A=
N
X
N + 1 k=0
Stk ,
(1)
PN
with tk = k. It is straightforward to adapt the results to other definitions of the average such as A = N1
k=1 Stk . The
definition as in (1) was also adopted in Ref. [42], while the alternative with N instead of N + 1 was used in Ref. [45]. The
demeaned logreturn Xt is related to St by the following formula
St = S0 exp [(r m) t + Xt ] ,
(2)
where r is the risk free interest rate and the drift parameter m will be specified later. The average of X will be denoted
by:
X T =
N
X
N + 1 k=0
X tk =
N
X
N + 1 k=1
X tk .
(3)
(4)
For simplicity of notation Stk will sometimes be written as Sk and Xtk as Xk . The basic assumption for our method is that
the logarithm of the asset price follows a Lvy process. This implies that non-overlapping increments are independent and
increments over equal time spans are identically distributed. The characteristic function X associated with the distribution
of the increments Xtk is defined by:
Xt (w) =
k
tk
(5)
5195
with P (x) the density function of Xtk . For later use in the calculations, it turns out to be convenient to introduce the function
f (w) defined by the following relation:
(6)
Note that the characteristic function X does not depend on k since all the increments are identically distributed. We didnt
tk
determine the drift parameter m yet. Because the market is incomplete several choices for a risk neutral martingale measure
rt
are possible. One
h way
i to determine m is to require that the asset price discounted by the money market account Bt = e is
ST
BT
a martingale: E
S0
B0
m = f (i) .
(7)
More information on equivalent martingale measures for Lvy processes can be found in Refs. [46,37,38,47].
3. Lower bound
In this section we derive a lower bound for the call price C of a discretely monitored arithmetic Asian option. Recall that
C is given by:
C = exp (rT ) E (A K )+ ,
= exp (rT ) E E (A K )+ |X T ,
(8)
where (x)+ = max(x, 0), E [X ] is the expectation value of X and E [X |Z ] is the conditional expectation value of X with
respect to Z . Here we chose the logarithm of the geometric average as a conditioning variable similar as in Refs. [8,29]. The
inequality of Jensen [48] tells us expression (8) is bigger than:
C exp (rT ) E
i
E A|X T K + .
(9)
This inequality has important implications in path integral analysis and quantum mechanics [49]. Feynman and Kleinert [50]
showed in this context how quantum-mechanical partition functions can be approximated by an effective classical partition
function, a technique which has been successfully applied to finance (see Ref. [49] and references therein, and Ref. [51] for
a recent application). If in the expression above A is written in full with the aid of Eq. (1) and then Sk is substituted by
expression (2), if the conditional expectation changes places with the sum and if the expectation over X T is written out
explicitly we get:
C e
rT
"
S0
1+
N +1
N
X
(r m)k
Xk
E e |X T = x
k=1
!
#
PX T (x) dx
(10)
exp (i1 x ) X T (1 ) d1 ,
(11)
with
!
(N k + 1)
X T (1 ) = exp
.
f 1
N +1
k=1
The conditional expectation E exp (Xk ) |X T present in Eq. (10) is equal to:
Z +
P (xk , x )
E exp (Xk ) |X T = x =
exp (xk )
dxk .
PX T (x)
N
X
(12)
(13)
The joint probability density distribution P (xk , x ) can be obtained in a similar way as the P (x), see Appendix A for the
derivation, the result is given by:
P (x, xk ) =
1
4 2
exp (i1 x i2 xk ) Xk ,X (1 , 2 ) d1 d2 ,
(14)
with
"
Xk ,X (1 , 2 ) = exp
k
X
f
l =1
X
! #
N
(N l + 1)
(N l + 1)
1
+ 2 +
f 1
.
N +1
N +1
l=k+1
(15)
5196
E eXk |X T = x =
1
4
2P
X T
(x)
ex k
+
2 PX T (x)
ei1 x i2 xk X
ei1 x X
k ,XT
k ,XT
(2 + i) ei1 x X
2 PX T (x)
1
4 2 PX T (x)
1
k ,XT
(1 , 2 ) d1 d2 dxk
k , XT
(1 , 2 ) d1 d2
(1 , 2 ) d1 d2
(1 , i) d1 .
(16)
An efficient way to evaluate expression (10) is to first determine the value for x for which the monotonically increasing
function
S0
1+
(N + 1)
N
X
2 PX T (x) k=1
(r m)k
i1 x
k , XT
(1 , i) d1 K
(17)
rT
"
S0
1+
N +1
N
X
(r m)k
Xk
E e |X T = x
!
#
K PX T (x) dx .
(18)
k=1
g (0)
2
i
2
d1 ,
(19)
with
g (1 ) =
S0
X (1 ) +
(N + 1)
N
X
!
e
(r m)k
Xk ,X T (1 , i) K X T (1 ) .
(20)
k=1
Formula (19) is our central result concerning the lower bound, as already mentioned this formula is quick to evaluate and
is very accurate, which will become clearer Section 5.
4. Upper bound
In this section we derive an upper bound which supplements the lower bound in the sense that it is constructed based
on the difference between the real price and the lower bound, which makes it possible to reliably qualify the price C . For
the calculation of the comonotonic upper bound in the setting of Lvy models, see Appendix C.
The upper bound is found by determining the error made when calculating the lower bound, i.e.:
n
h
io
= exp (rT ) E (A K )+ E E A|X T K + .
(21)
exp (rT )
2
v
"
u
u
t
E
Var
N
X
N + 1 k=0
#
Sk K X T 1{X T <b}
(22)
with
b = ln
K
S0
+
mr
2
(23)
and 1{X <b} the indicator function. Denote the right hand side of inequality (22) by 0 . It is easily seen that the derivation
given in Refs. [8,29] to go from formula (21) to formula (22) is model independent. In other words their derivation applies
5197
to our situation without any essential changes. The upper bound is then given by: upper bound = lower bound + 0 . A
straightforward calculation shows that the variance in formula (22) is equal to:
"
#
S02
Var
Sk K X T =
N + 1 k=0
(N + 1)2
N
X
N
X
e
2
k=2
X X
l
k
|X E e |X E e |X
e
E e
l =1
.
N
X
2
+
e2(r m)k E e2Xk |X E eXk |X
(r m)k
k1
X
(r m)l
Xk +Xl
(24)
k=1
In formula (16) an expression for E exp (Xk ) |X T = x is given. In a similar way E exp (2Xk ) |X T = x equals:
E exp (2Xk ) |X T = x =
2 PX T (x)
(25)
Also E exp (Xk + Xl ) |X T can be calculated in a similar way, though first P (xk , xl , x ) needs to be derived. This can be done
in the same way as P (xk , x ) was derived (see Appendix A), the result for P (xk , xl , x ) with l < k is given by:
P (xk , xl , x ) =
(2 )3
ei1 x i2 xk i3 xl Xk ,Xl ,X T (1 , 2 , 3 ) d1 d2 d3 ,
(26)
with
Xk ,Xl ,X T (1 , 2 , 3 ) = exp []
(27)
and
l
X
f 1
k2 =1
k
X
(N k2 + 1)
f 1
+ 2
(N k2 + 1) + 2 + 3 +
N +1
N +1
k =l +1
1
k2 =k+1
(N k2 + 1)
.
N +1
(28)
E eXk +Xl |X T = x =
ei1 x Xk ,Xl ,X T (1 , i, i) d1 .
2 PX T (x)
(29)
Combining formulas (16), (24), (25) and (29) expression (22) becomes:
0 =
S0 erT
2 (N + 1)
h (x)dx
(30)
with
PX T (x)
h (x) = 2
e(r m)k
k=2
k1
X
l=1
e(r m)l
N
X
k=1
PX T (x)
1
4 2
i1 x
ei1 x Xk ,Xl ,X T (1 , i, i) d1
ei1 x Xk ,X T (1 , i) d1
e
Xk ,X T (1 , 2i) d1
2
2
Z +
1
i1 x
e
Xk ,X T (1 , i) d1
2
e2(r m)k
ei1 x Xl ,X T (1 , i) d1
(31)
Substituting expression (11) and changing places of the summations and the integrations the above expression becomes:
=
0
S0
2 (N + 1)
s Z
dx
ei(1 +2 )x g (1 , 2 ) d1 d2
(32)
5198
with
g (1 , 2 ) =
N
X
e(r m)k
k=2
k1
X
e(r m)l
l =1
N
1X
X T (2 ) Xk ,Xl ,X T (1 , i, i)
Xk ,X T (1 , i) Xl ,X T (2 , i)
p1
(1 p) 2
f (w) =
+
+
1 .
(34)
2
2 + i
1 i
Here is the diffusion volatility, is the jump intensity, p, 1 p is the probability of an up, down jump respectively and
1 , 2 control the exponentially distributed up and down jump sizes. The function g (1 ) necessary for the lower bound and
2
!
N
X
(2N + 1)
g (1 ) = exp
T +
(1 , l, 0)
2
6 (N + 1)
l=1
(2N k + 1) k
rk + i 2 1
k [ (0, 0, i) 1]
X
S0
2 (N + 1)
1+
exp
X
(N + 1)
k=1
[ (1 , l, i) (1 , l, 0)]
+
2T
2
1
l=1
K , (35)
5199
Table 1
In this table our bounds for prices of arithmetic Asian call options for Kous model are shown. Our lower bound (LB) and upper bound (UB) are compared
with the numerical quadrature (NQ) and Monte Carlo results (MC) of Ref. [42] and with the comonotonic upper bound (UBC). The parameter values used
are: S0 = 100, T = 1, r = 0.0367, = 0.120381, = 0.330966, p = 0.2071, 1 = 9.65997, 2 = 3.13868. MR stands for the maximal relative error
which is made when the average of the lower and the upper bound is used as an approximation. RLB stands for the relative error compared with the Monte
Carlo result which is made when the lower bound is used as an approximation.
N
12
50
250
LB
NQ
MC
UB
UBC
MR (%)
RLB (%)
90
100
110
90
100
110
90
100
110
12.7082
5.01609
1.0409
12.7398
5.05717
1.06829
12.7484
5.06851
1.07594
12.71236
5.01712
1.04142
12.74369
5.05809
1.06878
12.75241
5.06949
1.07646
12.71298
5.01729
1.04141
12.74420
5.05849
1.06886
12.75141
5.06910
1.07632
12.8215
5.18289
1.25014
12.8546
5.22485
1.27793
12.8636
5.23644
1.28572
13.057
5.5087
1.4029
13.103
5.5729
1.4512
13.114
5.5880
1.4624
0.4
1.6
9.1
0.4
1.6
8.9
0.4
1.6
8.8
0.04
0.02
0.05
0.03
0.03
0.05
0.02
0.01
0.04
with
(, l, a) =
p1
(1 p) 2
+
.
(N l+1)
2 + i N +1 + a
1 i (NN+l+11) + a
(36)
After substituting the characteristic functions in formula (33) by their explicit expressions (i.e. formulas (12), (15) and (27))
and rearranging terms g (1 , 2 ) becomes:
2T
2N + 1
T
2T + f (1 , 2 , 0, 0, 1, N ) i 2 1
2
6 (N + 1)
2
!
N
k
1
X
X
1
exp h(k) + f (1 , 1 , 1, 0, 1, k)
q (l) + q(k) .
g (1 , 2 ) = exp
w22 + 12
k=1
l =1
(37)
With
exp 2 k + f (1 , 1 , 2, 1, 1, k)
q(k) = eh(k)
exp f (2 , 2 , 1, 0, 1, k)
p1
( 1 p) 2
h(k) = r
+
1 k,
2 + 1
1 1
,
(38)
(39)
and
(2N k + 1) k
2 (N + 1)
(1 p) 2
(1 p) 2
+c
N
k
+
1
k
2
X 2 + a + i1
2 + b + i2 N Nk+21+1
N +1
+
p1
p1
+
k2 =1
+c
N k2 +1
1 a i1 N +1
1 b i2 N Nk+21+1
f (1 , 2 , a, b, c , k) = i 2 1
(40)
The results for this model discussed in the beginning of this section are given in Table 1.
5.2. Mertons model
For this model f (w) can be derived from results in Ref. [2]:
f (w) =
2
2
+ exp i
2
2 2
2
1 .
(41)
The parameters and are again the diffusion volatility and the jump intensity, and are the mean and the standard
deviation of the jump size. The results for Kous model are easily adapted to Mertons model. The function g (1 ) becomes:
)
N
X
(2N + 1)
g (1 ) = exp
T +
(1 , l, 0)
2
6 (N + 1)
l =1
2T
2
1
5200
Table 2
In this table our bounds for prices of arithmetic Asian call options for Mertons model are shown. Our lower bound (LB) and upper bound (UB) are compared
with the numerical quadrature (NQ) and Monte Carlo results (MC) of Ref. [42] and with the comonotonic upper bound (UBC). The parameter values used
are: S0 = 100, T = 1, r = 0.0367, = 0.126349, = 0.390078, = 0.174814, = 0.338796. MR stands for the maximal relative error which is
made when the average of the lower and the upper bound is used as an approximation. RLB stands for the relative error compared with the Monte Carlo
result which is made when the lower bound is used as an approximation.
N
12
50
250
LB
NQ
MC
UB
UBC
MR (%)
RLB (%)
90
100
110
90
100
110
90
100
110
12.7061
5.00959
1.05101
12.7364
5.05080
1.07898
12.74465
5.06217
1.08679
12.71066
5.01127
1.05162
12.74093
5.05246
1.07959
12.74917
5.06381
1.08740
12.71035
5.01117
1.05141
12.74051
5.05226
1.07959
12.74964
5.06417
1.08772
12.7992
5.15688
1.24086
12.8309
5.19901
1.26910
12.8395
5.21062
1.27698
13.052
5.5057
1.4180
13.097
5.5704
1.4678
13.107
5.5856
1.4795
0.4
1.4
8.2
0.4
1.4
8.1
0.4
1.4
8.0
0.03
0.03
0.04
0.03
0.03
0.06
0.04
0.04
0.09
N
S
X
0
1+
exp
(N + 1)
k=1
(2N k + 1) k
2 (N + 1)
, (42)
k
X
[ (1 , l, i) (1 , l, 0)]
+
rk k [ (0, 0, i) 1] + i 2 1
l=1
(N l + 1)
(, l, b) = exp i
+b
N +1
(NN+l+11) + b
2
(43)
It follows that:
2T
2N + 1
T
2T + f (1 , 2 , 0, 0, 1, N ) i 2 1
2
6 (N + 1)
2
!
k
1
N
X
X
1
exp (h(k) + f (1 , 1 , i, 0, 1, k))
q (l) + q(k) ,
g (1 , 2 ) = exp
w22 + 12
k=1
l =1
(44)
where
h(k) = r exp +
q(k) = eh(k)
2
2
1
k,
exp 2 k + f (1 , 1 , 2i, i, 1, k)
exp [f (2 , 2 , i, 0, 1, k)]
(45)
(46)
and
(2N k + 1) k
2 (N + 1)
2
1 N Nk+21+1 + a 2
N
k
+
1
2
exp
+a
i 1
N
+
1
2
k
X
2
.
k2 =1
2 N Nk+21+1 + b 2
k
+
1
2
+c exp
+b
i 2
N +1
2
f (1 , 2 , a, b, c , k) = i 2 1
(47)
The numerical performance of our bounds for the Merton model are illustrated in Table 2.
5.3. The NIG model
For this model f (w) can be derived from results in Ref. [3]:
f (w) =
q
p
2 ( + i)2 2 2 .
(48)
5201
Table 3
In this table our bounds for prices of arithmetic Asian call options for the NIG model are shown. Our lower bound (LB) and upper bound (UB) are compared
with the numerical quadrature (NQ) and Monte Carlo results (MC) of Ref. [42] and with the comonotonic upper bound (UBC). The parameter values used
are: S0 = 100, T = 1, r = 0.0367, = 6.1882, = 3.8941, = 0.1622. MR stands for the maximal relative error which is made when the average of
the lower and the upper bound is used as an approximation. RLB stands for the relative error compared with the Monte Carlo result which is made when
the lower bound is used as an approximation.
N
12
50
250
LB
NQ
MC
UB
UBC
MR (%)
RLB (%)
90
100
110
90
100
110
90
100
110
12.6192
5.05928
1.01298
12.6581
5.10241
1.03714
12.6686
5.11427
1.04392
12.62243
5.06060
1.01355
12.66118
5.10367
1.03770
n.a.
n.a.
n.a.
12.62293
5.06077
1.01374
12.66112
5.10359
1.03770
12.67186
5.11558
1.04446
12.7091
5.1929
1.1865
12.7495
5.23697
1.21121
12.7604
5.24908
1.21812
13.012
5.5400
1.3351
13.067
5.5978
1.3723
13.080
5.6115
1.3811
0.4
1.3
7.9
0.4
1.3
7.7
0.4
1.3
7.7
0.03
0.03
0.08
0.02
0.02
0.05
0.03
0.03
0.05
Here is the scale parameter governs the tail heaviness and determines the asymmetry of the distribution. The function
g (1 ) becomes:
(
g (1 ) = exp T
N
X
)
(1 , l, 0)
l=1
i
p
2 2
N
X
S0
K ,
k
exp
X
(N + 1) 1 +
[ (1 , l, i) (1 , l, 0)]
k=1
rk + k (0, 0, i)
(49)
l =1
with now
2
(N l + 1)
+i
+b
.
N +1
(, l, b) =
(50)
Furthermore, we get:
g (1 , 2 ) = exp 2 T
N
X
i
p
2 2 + f (1 , 2 , 0, 0, 1, N )
k=1
k1
X
q(l) +
l=1
1
2
!
q(k) ,
(51)
with
h(k) = r +
q
( + 1)
(52)
(53)
k,
and
N k2 + 1
2
2 + i 1
+a
k
X
N +1
f (1 , 2 , a, b, c , k) =
s
2
k2 =1
N k2 + 1
+c 2 + i 2
+b
N +1
(54)
The accuracy of our bounds for the NIG model and the parameter values discussed in the beginning of this section are
revealed in Table 3.
5.4. The CGMY model
For this model f (w) can be derived from results in Ref. [5]:
f (w) = C (Y ) (M i)Y + (G + i)Y GY + M Y
(55)
5202
Table 4
In this table our bounds for prices of arithmetic Asian call options for the CGMY model are shown. Our lower bound (LB) and upper bound (UB) are compared
with the numerical quadrature (NQ) and Monte Carlo results (MC) of Ref. [42] and with the comonotonic upper bound (UBC). The parameter values used
are: S0 = 100, T = 1, r = 0.0367, C = 0.0244, G = 0.0765, M = 7.5515 and Y = 1.2945. MR stands for the maximal relative error which is made when
the average of the lower and the upper bound is used as an approximation. RLB stands for the relative error compared with the numerical quadrature result
which is made when the lower bound is used as an approximation.
N
12
50
250
LB
NQ
MC
UB
UBC
MR (%)
RLB (%)
90
100
110
90
100
110
90
100
110
12.7002
5.03303
1.02051
12.7337
5.07406
1.04612
12.7429
5.08543
1.05330
12.70625
5.03492
1.02115
12.73854
5.07570
1.04674
12.74737
5.08694
1.05389
12.69114
5.02787
1.01895
12.73548
5.07403
1.04594
12.74864
5.08744
1.05377
12.8686
5.2513
1.2796
12.7040
5.2443
1.3089
12.9136
5.3065
1.3162
13.066
5.5157
1.3543
13.114
5.5687
1.4027
13.125
5.5839
1.4118
0.7
2.1
11.3
0.7
1.7
11.2
0.7
2.1
11.1
0.05
0.04
0.06
0.04
0.03
0.06
0.04
0.03
0.06
where C lays down the overall level of activity, the parameters G and M determine the exponential decay on the right and
the left of the Lvy density and Y determines the fine structure of the stochastic process. The function g (1 ) becomes:
g (1 ) = exp C (Y ) G + M
Y
T + C (Y )
N
X
!
(1 , l, 0)
l =1
rk kC (Y ) (0, 0, i) GY + M Y
X
S
0
k
X
1+
exp
K ,
[ (1 , l, i) (1 , l, 0)]
+
C (Y )
(N + 1)
k=1
(56)
l =1
(, l, b) =
Y
Y
(N l + 1)
(N l + 1)
M i
+b
+ G+i
+b
.
N +1
N +1
(57)
Here, g (1 , 2 ) becomes:
g (1 , 2 ) = exp 2C (Y ) T GY + M Y + f (1 , 2 , 0, 0, 1, N )
N
X
k=1
k1
X
q(l) +
l =1
1
2
!
q(k) ,
(58)
where
h(k) = r C (Y ) (M 1)Y + (G + 1)Y GY + M Y
k,
(59)
(60)
f (1 , 2 , a, b, c , k) = C (Y )
Y
Y
N k2 + 1
N k2 + 1
M i 1
+a
+ G + i 1
+a
k
X
N +1
N +1
Y
Y ! .
k
+
1
N
k
+
1
2
2
k2 =1
+c
M i 2
+b
+ G + i 2
+b
N +1
N +1
(61)
Also for the CGMY model Ref. [42] presents numerical quadrature and Monte Carlo results for arithmetic Asian option prices.
For this model however Fusai and Meucci [42] claim that their Monte Carlo result is less reliable. Indeed we find that their
Monte Carlo result is often below our lower bound. Therefore we compared with the numerical quadrature result to calculate
the relative error made when the lower bound is chosen as an approximation. The comparison of our bounds with their
results is shown in Table 4. The accuracy and the computational cost of the lower bound are practically the same as with the
other models. For the upper bound though both the accuracy and the computational cost are inferior to the other models.
5203
N
1 X
Sk
N k=1
as definition for the discrete average. For this model f (w) can be derived from results in Ref. [4]:
f (w) =
ln 1 i vw +
2
2
vw 2 ,
(62)
where controls the volatility, the skewness and v the kurtosis. The function g (1 ) becomes:
)
N
X
g (1 ) = exp
(1 , l, 0)
v l =1
k
2
v
rk + ln 1 v
N
S
X
v
2
0
1+
exp
k
(N + 1)
1 X
k=1
[ (1 , l, i) (1 , l, 0)]
v l =1
K ,
(63)
2 !
2
(N l + 1)
(N l + 1)
(, l, b) = ln 1 i v
+b +
v
+b
.
N +1
2
N +1
(64)
N
X
k=1
k1
X
l =1
q(l) +
1
2
!
q(k) ,
(65)
with
h(k) = r +
ln 1 v
2
2
k,
(66)
(67)
and
f (1 , 2 , a, b, c , k)
2 !
2
(N k2 + 1)
(N k2 + 1)
+a +
v
+a
ln 1 i v
k
N +1
2
N +1
!!
=
.
2
2
v k =1
(N k2 + 1)
(N k2 + 1)
2
+b +
v
+b
+c ln 1 i v
N +1
2
N +1
(68)
Table 5 demonstrates that also for this model our bounds perform very well. The computational cost for this model is
comparable with the other models (with exception of the CGMY model for which the upper bound takes a longer time
to evaluate).
6. Conclusion
We derived analytical upper and lower bounds for the price of a discretely monitored arithmetic Asian option when the
underlying asset is assumed to follow a general Lvy process. We tested these bounds for Kous model, Mertons model, the
normal inverse Gaussian model, the CGMY model and the variance gamma model, and compared them with the comonotonic
upper bound.
As far as we know a lower bound has hitherto only been computed for the BlackScholes model, while an upper bound
was computed for the BlackScholes model and for the variance gamma, the normal inverse Gaussian and the Meixner
model, for the last three models see Ref. [45]. Our result extends previous work in the Black Scholes setting to general Lvy
processes, providing and improving analytical bounds for a variety of models.
Note that next to our bounds, other methods could be considered, leaning on previous results for other settings in the
literature. We refer to Ref. [25] where an improved upper bound for Asian options was derived in a BlackScholes framework,
5204
Table 5
In this table our bounds for prices of arithmetic Asian call options for the VG model are shown. Our lower bound (LB) and upper bound (UB) are compared
with the lower bound of Ref. [44] (LBA), a Monte Carlo simulation of 4 108 iterations and with the comonotonic upper bound (UBC). Next to the Monte
Carlo simulation the error needed to find the 99% certainty interval is shown. The parameter values used are: S0 = 100, T = 10, N = 120, r = 0.03, =
0.2684, v = 1.1737, = 0.1280. MR stands for the maximal relative error which is made when the average of the lower and the upper bound is
used as an approximation. RLB stands for the relative error compared with the Monte Carlo result which is made when the lower bound is used as an
approximation.
K
LBA
LB
MC
Error
UB
UBC
MR (%)
RLB (%)
60
70
80
90
100
110
120
130
140
150
42.6644
36.3172
30.6444
25.6772
21.4012
17.7698
14.7195
12.1783
10.0757
8.3415
43.44140
37.40021
31.95129
27.12407
22.91283
19.28586
16.19468
13.58198
11.38774
9.55342
43.6456
37.6212
32.1819
27.3582
23.1459
19.5142
16.4156
13.7938
11.5894
9.7442
0.0061
0.0051
0.0049
0.0047
0.0044
0.0042
0.0039
0.0037
0.0035
0.0033
43.7861
37.8623
32.5306
27.8170
23.7144
20.1914
17.2003
14.6843
12.5834
10.8385
44.3869
38.7776
33.7057
29.1785
25.1828
21.6796
18.6637
16.0577
13.8248
11.9173
0.4
0.6
0.9
1.3
1.7
2.3
3.0
3.9
5.0
6.3
0.5
0.6
0.7
0.9
1.0
1.2
1.4
1.6
1.8
2.0
outperforming the bound obtained in Ref. [29]. An extension of this approach to general Lvy models could lead to a more
accurate upper bound than the comonotonic upper bound of Ref. [45]. We also refer to Ref. [52] where upper and lower
bounds are derived for transition densities for rather general models. It is conceivable that an application of their approach
to the pricing of Asian options could lead to similar bounds as presented in this paper. An investigation of both alternatives,
however, is beyond the scope of the present contribution.
The numerical analysis leads us to the following observations. Our lower bound is quick to evaluate, very accurate and
performs better than the existing lower bound for the VG model derived in Ref. [44]. For all the models except for the CGMY
model the upper bound is still acceptably quick to evaluate although much slower than the lower bound; for the CGMY model
the evaluation of the upper bound becomes quite slow. For all the examined models our upper bound is significantly more
accurate than the comonotonic upper bound. Although it depends on the parameter values, for most relevant parameter
values the upper bound is less accurate than the lower bound. Therefore we recommend using the lower bound as an
approximation instead of the average of the lower and the upper bound.
Acknowledgements
The authors would like to thank S. Foulon for the fruitful discussions. This work is supported financially by the Fund for
Scientific Research Flanders, FWO project G.0125.08, and by the Special Research Fund of the University of Antwerp BOF
NOI UA 2007.
Appendix A. Derivation of the joint probability distribution P (xk , x )
Here we derive the joint probability distribution P (xk , x ) in terms of its characteristic function X
characteristic function is given by:
Xk ,X T (1 , 2 ) = E exp i1 X T + i2 Xk
!#
"
N
1 X
Xt + i2 Xk
= E exp i1
N + 1 l =1 l
"
!#
N
l
k
X
1 XX
= E exp i1
Xt + i2
X tl
N + 1 l =1 j =1 j
l=1
"
!#
N
k
X
1 X
= E exp i1
X tl
(N l + 1) Xtl + i2
N + 1 l =1
l =1
N
X
1
i1
(N l + 1) Xtl
N + 1 l=k+1
= E exp
k
X
1
+
i1
(N l + 1) + i2 Xtl
N
+
1
l =1
"
Y
#
k
N
Y
(N l + 1)
(N l + 1)
=E
exp i 1
+ 2 Xtl
exp i1
X tl
.
N +1
N +1
l =1
l=k+1
k ,X T
(1 , 2 ). This
(A.1)
(A.2)
5205
Because of the independency of the increments the expectation value and the multiplications can be changed places. We
then recognize the characteristic function of the increments. The characteristic function Xk ,X T (1 , 2 ) then becomes:
Y
N
(N l + 1)
(N l + 1)
Xt 1
Xt 1
+ 2
k
k
N +1
N +1
l =1
l=k+1
"
! #
N
k
X
X
(N l + 1)
(N l + 1)
f 1
+ 2 +
.
= exp
f 1
N +1
N +1
l=k+1
l=1
k
Y
Xk ,X T (1 , 2 ) =
(A.3)
(A.4)
The probability distribution can then be recovered from the characteristic function:
+
P (x, xk ) =
4 2
exp (i1 x i2 xk ) Xk ,X T (1 , 2 ) d1 d2 .
(A.5)
rT
"
S0
1+
N +1
N
X
(r m)k
Xk
E e |X T = x
!
#
K PX T (x) dx .
(B.1)
k=1
C e
rT
dx PX T (x)
"
S0
(N + 1)
1+
N
X
2 PX T
k=1
(r m)k
i1 x
Xk ,X T (1 , i) d1 K .
(B.2)
Bringing P (x) inside the brackets and filling in Eq. (11) for P (x) one obtains
C e
rT
"
erT
2
(N + 1)
ei1 x X T (1 ) d1 +
ei1 x Xk ,X T (1 , i) d1
S0
dx (x a)
d1 e
i1 x
S0
(N + 1)
"
N
1 X (r m)k
e
2 k=1
ei1 x X T (1 ) d1 dx ,
X T (1 ) +
N
X
!
e
(r m)k
Xk ,X T (1 , i) K X T (1 ) (B.3)
k=1
S0
X T (1 ) +
(N + 1)
N
X
!
e
(r m)k
Xk ,X T (1 , i) K X T (1 )
(B.4)
k=1
and the Fourier representation of the indicator function, expression (B.3) becomes:
erT
=
=
rT Z +
2
e
rT Z +
2
erT
i
q + i
Z +
i
q + i
i
q + i
dq
2
exp (i1 x ) g (1 ) d1 dx
exp(iqa) exp (i (1 + q) x ) dx g (1 )
exp (rT )
d1
exp(iqa) (1 + q) g (1 ) dqd1
i
exp (i1 a) g (1 ) d1 .
2 1 + i
For the numerical integration it is convenient to write this as:
dq
2
g (0)
2
which is expression (19).
i
2
(B.5)
d1 ,
(B.6)
5206
as in Ref. [45] the Asian option is defined with an averaging over N values whereas we use N + 1 values, we calculated
the comonotonic upper bounds again for all five models under consideration. Therefore, we recall here the most important
formulae necessary to calculate the bound within Lvy models. The recipe to calculate this upper bound goes as follows. To
start with we need to numerically determine for which value u the following equality holds:
N
X
FSk (u) = (N + 1) K
1
k=1
S0
N +1
(C.1)
With this value for u the upper bound for the Asian call price C is given by:
C
N
i
erT X h
E Sk FSk 1 (u)
.
+
N + 1 k=1
(C.2)
To deal with these two equations one has to be able to evaluate the inverse cumulative distribution function of the random
variable Sk , denoted here by FSk 1 (u). As a first step this function can be expressed as a function of the inverse cumulative
distribution function of the random variable Xk :
FSk 1 (u) = FS1exp[(r m)k+X ] (u),
k
i
h
= S0 exp (r m) k + FXk1 (u) .
(C.3)
Analogously to the derivation of formula (19), presented in Appendix B the cumulative distribution function of Xk can be
shown to equal:
FXk (x) =
ef (0)k
2
i
2
dw.
(C.4)
The inversion of this formula needs to be done numerically. Although Eqs. (C.1) and (C.2) form an elegant presentation of
the comonotonic upper bound, it is hard to implement this for a general model. For the BlackScholes model it is quite
straightforward to determine which integration grid should be used in formula (C.4) for every k. For Lvy models on the
contrary this can be a tedious task. For some models it might be preferable to calculate the cumulative distribution function
of Xk by numerically integrating its probability distribution function.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
S.G. Kou, A jump-diffusion model for option pricing, Manage. Sci. 48 (8) (2002) 10861101.
R.C. Merton, Option pricing when underlying stock returns are discontinuous, J. Finance Econ. 3 (12) (1976) 125144.
O.E. Barndorff-Nielsen, Normal inverse Gaussian distributions and stochastic volatility modelling, Scand. J. Statist. 24 (1) (1997) 113.
D.B. Madan, P.P. Carr, E.C. Chang, The variance gamma process and option pricing, Eur. Finance Rev. 2 (1) (1998) 79105.
P. Carr, H. Geman, D. Madan, M. Yor, The fine structure of asset returns: an empirical investigation, J. Bus. 75 (2) (2002) 305333.
A.G.Z. Kemna, A.C.F. Vorst, A pricing method for options based on average asset values, J. Bank Finance 14 (1) (1990) 113129.
V. Linetsky, The path integral approach to financial modeling and options pricing, Comput. Econ. 11 (1998) 129163.
L.C.G. Rogers, Z. Shi, The value of an Asian option, J. Appl. Probab. 32 (4) (1995) 10771088.
J. Andreasen, The pricing of discretely sampled Asian and lookback options: a change of numeraire approach, J. Comput. Finance 2 (1) (1998) 523.
J. Hoogland, D. Neumann, Asians and cash dividends: exploiting symmetries in pricing theory, SSRN eLibrary. URL: http://ssrn.com.ezpprod1.hul.harvard.edu/paper=234830.
J. Hoogland, D. Neumann, Tradable schemes, SSRN eLibrary. URL: http://ssrn.com.ezp-prod1.hul.harvard.edu/paper=241128.
J. Vecer, A new pde approach for pricing arithmetic average Asian options, J. Comput. Finance 4 (4) (2001) 105113.
H. Geman, M. Yor, Bessel processes, Asian options and perpetuities, Math. Finance 3 (4) (1993) 349375.
V. Linetsky, Spectral expansion for Asian (average price) options, Oper. Res. 52 (2004) 856857.
E. Lvy, Pricing European average rate currency options, J. Int. Money Financ. 11 (5) (1992) 474491.
G. Fusai, A. Tagliani, An accurate valuation of Asian options using moments, Int. J. Theor. Appl. Finance 5 (2) (2002) 147.
S.M. Turnbull, L.M. Wakeman, A quick algorithm for pricing European average options, J. Finance Quant. Anal. 26 (3) (1991) 377389.
N. Ju, Pricing Asian and basket options via Taylor expansion, J. Comput. Finance 5 (3) (2002) 79103.
R. Lord, Partially exact and bounded approximations for arithmetic Asian options, J. Comput. Finance 10 (2) (2006) 152.
S. Vanduffel, Z. Shang, L. Henrard, J. Dhaene, E.A. Valdez, Analytic bounds and approximations for annuities and Asian options, Insurance Math. Econom.
42 (3) (2008) 11091117.
A.P. Carverhill, L.J. Clewlow, Flexible convolution, Risk 3 (1990) 2529.
R. Zvan, P.A. Forsyth, K.R. Vetzal, Discrete Asian barrier options, J. Comput. Finance 4 (4) (1999) 4167.
E. Benhamou, Fast Fourier transform for discrete Asian options, SSRN eLibrary. URL: http://ssrn.com.ezp-prod1.hul.harvard.edu/paper=269491.
P. Den Iseger, E. Oldenkamp, Pricing guaranteed return rate products and discretely sampled Asian options, J. Comput. Finance 9 (3) (2006) 139.
M. Vanmaele, G. Deelstra, J. Liinev, J. Dhaene, M. Goovaerts, Bounds for the price of discretely sampled arithmetic Asian options, J. Comput. Appl. Math.
185 (1) (2006) 5190.
M. Curran, Valuing Asian and portfolio options by conditioning on the geometric mean price, Manage. Sci. 40 (12) (1994) 17051711.
J. Dhaene, M. Denuit, M.J. Goovaerts, R. Kaas, D. Vyncke, The concept of comonotonicity in actuarial science and finance: theory, Insurance Math.
Econom. 31 (1) (2002) 333.
J. Dhaene, M. Denuit, M.J. Goovaerts, R. Kaas, D. Vyncke, The concept of comonotonicity in actuarial science and finance: applications, Insurance Math.
Econom. 31 (2) (2002) 133161.
J.A. Nielsen, K. Sandmann, Pricing bounds on Asian options, J. Finan. Quant. Anal. 38 (2) (2003) 449473.
P. Gopikrishnan, V. Plerou, L.A. Nunes Amaral, M. Meyer, H.E. Stanley, Scaling of the distribution of fluctuations of financial market indices, Phys. Rev.
E 60 (5) (1999) 53055316.
5207
[31] Y. Liu, P. Gopikrishnan, Cizeau, Meyer, Peng, H.E. Stanley, Statistical properties of the volatility of price fluctuations, Phys. Rev. E 60 (2) (1999)
13901400.
[32] R.N. Mantegna, Z. Palgyi, H.E. Stanley, Applications of statistical mechanics to finance, Physica A 274 (12) (1999) 216221.
[33] E. Scalas, R. Gorenflo, F. Mainardi, Fractional calculus and continuous-time finance, Physica A 284 (14) (2000) 376384.
[34] E. Van der Straeten, C. Beck, Superstatistical fluctuations in time series: applications to share-price dynamics and turbulence, Phys. Rev. E 80 (3) (2009)
036108.
[35] M. Raberto, E. Scalas, G. Cuniberti, M. Riani, Volatility in the Italian stock market: an empirical study, Physica A 269 (1) (1999) 148155.
[36] S. Miccich, G. Bonanno, F. Lillo, R.N. Mantegna, Volatility in financial markets: stochastic models and empirical results, Physica A 314 (14) (2002)
756761.
[37] W. Schoutens, Lvy Processes in Finance: Pricing Financial Derivatives, Wiley, New York, 2003.
[38] R. Cont, P. Tankov, Financial Modelling with Jump Processes, Chapman and Hall/CRC, 2004.
[39] R.N. Mantegna, Lvy walks and enhanced diffusion in Milan stock exchange, Physica A 179 (2) (1991) 232242.
[40] M. Mariani, Y. Liu, Normalized truncated Lvy walks applied to the study of financial indices, Physica A 377 (2) (2007) 590598.
[41] R. Matsushita, P. Rathie, S.D. Silva, Exponentially damped Lvy flights, Physica A 326 (34) (2003) 544555.
[42] G. Fusai, A. Meucci, Pricing discretely monitored Asian options under Lvy processes, J. Bank Finance 32 (10) (2008) 20762088.
[43] H. Albrecher, M. Predota, On Asian option pricing for nig Lvy processes, J. Comput. Appl. Math. 172 (1) (2004) 153168.
[44] H. Albrecher, P.A. Mayer, W. Schoutens, General lower bounds for arithmetic Asian option prices, Appl. Math. Finance 15 (2) (2008) 123149.
[45] H. Albrecher, J. Dhaene, M. Goovaerts, W. Schoutens, Static hedging of Asian options under Lvy models, J. Derivatives 12 (3) (2005) 6372.
[46] T. Chan, Pricing contingent claims on stocks driven by Lvy processes, Ann. Appl. Probab. 9 (2) (1999) 504528.
[47] F. Hubalek, C. Sgarra, Esscher transforms and the minimal entropy martingale measure for exponential Lvy models, Quant. Finance 6 (2) (2006)
125145.
[48] J.L.W.V. Jensen, Sur les fonctions convexes et les ingalits entre les valeurs moyennes, Acta Math. 30 (1) (1906) 175193.
[49] H. Kleinert, Path Integrals in Quantum Mechanics, Statistics, Polymer Physics, and Financial Markets, 5th ed., World Scientific Publishing Co., Singapore,
2009.
[50] R.P. Feynman, H. Kleinert, Effective classical partition functions, Phys. Rev. A 34 (6) (1986) 50805084.
[51] D. Lemmens, M. Wouters, J. Tempere, S. Foulon, Path integral approach to closed-form option pricing formulas with applications to stochastic volatility
and interest rate models, Phys. Rev. E 78 (1) (2008) 016101.
[52] M. Goovaerts, A.D. Schepper, M. Decamps, Closed-form approximations for diffusion densities: a path integral approach, J. Comput. Appl. Math.
164165 (2004) 337364.