Fast narrow bounds on the value of Asian options

G. W. P. Thompson
Centre for Financial Research,
Judge Institute of Management,
University of Cambridge
Abstract
We consider the problem of nding bounds on the value of xed-strike and oating-
strike Asian options. A good lower bound for both types was derived in Rogers & Shi
(1995). We provide an alternative derivation, which leads to a simpler expression for
the bound, and also to the bound given by Curran (1992) for xed-strike options; we
derive an analogous bound for oating-strike options. Combining these results with a
new upper bound allows the accurate valuation of xed-strike and oating-strike Asian
options for typical parameter values.
Keywords: Brownian motion, Asian option, xed-strike, oating-strike.
1 Introduction
A xed-strike Asian (call) option with strike K > 0 and maturity T on an asset with price
process {S
t
} is a contract with value (
1
T
_
T
0
S
t
dtK)
+
at time T. We assume that T = 1 year
and that the asset price follows a geometric Brownian motion S
t
= S exp(B
t
+ct) where c
is a constant, S and are positive constants and B
t
is a Brownian motion. The arbitrage-
free time-0 value of this option is then equal to exp()E[(
_
1
0
S exp(B
t
+ t) dt K)
+
],
where is the risk-free interest rate (assumed constant) and =
1
2

2
(see Baxter &
Rennie (1996) for an introduction to arbitrage-free valuation). Thus the problem of valuing
the option boils down to calculating E[(S
_
1
0
exp(B
t
+t) dtK)
+
]. Another type of Asian
option, the oating-strike option, pays out (
_
1
0
S
t
dt S
1
)
+
, and leads to the consideration
of E[(
_
1
0
exp(B
t
+t) dt exp(B
1
+))
+
].
Several approaches to the problem of valuing Asian options have been tried. Carverhill
& Clewlow (1990) use a convolution method to compute the distribution of
_
1
0
S
t
dt, and
work by Yor (1992) and Geman & Yor (1993) has lead to a formula for the price as a triple
integral. Approximate formulae and bounds in the form of single and double integrals
1
by Levy (1992), Levy & Turnbull (1992), Curran (1992) and Rogers & Shi (1995) seem
to be faster to evaluate however, and the methods of this paper fall into this category.
Using intuition and simple optimization we derive bounds on the value of xed-strike and
oating-strike Asian options which can be computed quickly, and which are accurate for
typical parameter values. The method generalizes to other options on sums of lognormal
assets: discretely monitored Asians, currency basket options, and swaptions in Gaussian
HJM models, for example.
The plan of this paper is as follows: in Section 2 we state some useful facts about covari-
ances related to Brownian motion, and in Section 3 present the impressive lower bound of
Rogers & Shi (1995) for xed-strike and oating-strike options, and the approximation to
the xed-strike bound given by Curran (1992). This approximation is notable since it is very
close to the bound of Rogers & Shi (1995) and is much easier to compute, involving only
one-dimensional integrals rather than a troublesome two-dimensional integral. We give an
alternative derivation of the bounds of Rogers & Shi (1995), leading to an expression involv-
ing only one-dimensional integrals; we also provide a generalization Currans approximation
to oating-strike options.
In Section 4 we derive upper bounds to complement the lower bounds, and in Section 5
we present a numerical comparison of the various bounds discussed here, using the parameter
values from Curran (1992).
2 Useful covariances
Like most of the approximation formulae in the literature, we will exploit the high correlation
between
_
1
0
exp(B
t
+ t) dt and
_
1
0
B
t
dt for the values of and met in practice, and
the fact that the second integral is Gaussian. We will need the covariance matrix of the
bivariate Gaussian random variable (B
t
,
_
1
0
B
s
ds):
E
_
_
B
t
,
_
1
0
B
s
ds
_
T
_
B
t
,
_
1
0
B
s
ds
_
_
=
_
t t(1 t/2)
t(1 t/2)
1
3
_
.
Thus the conditional distribution of B
t
given
_
1
0
B
s
ds = z is normal with mean 3t(1t/2)z
and variance t 3t
2
(1 t/2)
2
, and the conditional distribution of
_
1
0
B
s
ds given B
t
= x is
normal with mean (1 t/2)x and variance
1
3
t(1 t/2)
2
. We will also need the covariance
matrix of (B
t
,
_
1
0
B
s
ds B
1
):
E
_
_
B
t
,
_
1
0
B
s
ds B
1
_
T
_
B
t
,
_
1
0
B
s
ds B
1
_
_
=
_
t t
2
/2
t
2
/2
1
3
_
,
2
so, given
_
1
0
B
s
dsB
1
= x, the conditional distribution of B
t
is normal with mean 3t
2
x/2
and variance t 3t
4
/4.
3 Lower bounds
In this section we present two derivations of the bound of Rogers & Shi (1995): their
own, and an alternative, which yields a simpler expression for the bound. It also leads
to the bound of Curran (1992) for xed-strike options. We show how a similar bound for
oating-strike options may be derived.
The derivation of Rogers & Shi (1995) exploits the inequality:
E
_
A
+

= E
_
E
_
A
+
C
_
E
_
(E(A C))
+

which holds for any random variables A and C.

For xed-strike Asian option they choose A =
_
1
0
S
t
dt K and C =
_
1
0
B
t
dt. Since the
inner expectation is
_
1
0
E
_
S exp (B
t
+t)
_
1
0
B
s
ds
_
dtK and, conditional
_
1
0
B
t
dt = z,
B
t
is normal with mean 3t(1 t/2)z and variance t 3t
2
(1 t/2)
2
, we have the bound
V
xed
e

3
_

3z
_
__
1
0
Se
3t(1t/2)z+t+
1
2

2
(t3t
2
(1t/2)
2
)
dt K
_
+
dz. (3.1)
To bound a oating-strike option they use A =
_
1
0
S
t
dt S
1
, C =
_
1
0
B
t
dt B
1
, and get
V
oating
e

3
_

3z
_
__
1
0
S exp
_
3t
2
z/2 +t +
1
2

2
_
t 3t
4
/4
__
dt
S exp
_
3z/2 + +
2
/8
__
_
+
dz.
Both of these formulae are slightly tricky to evaluate since the outer integration has a
non-smooth integrand.
An alternative approach is to approximate the event that the option eventually ends
in-the-money with something more tractable. Let A = { :
_
1
0
S
t
dt > K}, and note that
E
_
__
1
0
S exp (B
t
+t) dt K
_
+
_
=
_
1
0
E[(S exp (B
t
+t) K) I(A)] dt. (3.2)
If we replace A by some other event A

, we no longer have equality in (3.2); the right hand

side is now a lower bound. We will use A

= {
_
1
0
B
t
dt > }. This Gaussian form allows
the expectation to be written as a Black-Scholes type formula once has been determined,
and just leaves us with a one-dimensional integral of a smooth integrand, which should be
very fast.
3
To determine the optimal value of , let N
t
= B
t
+ t + log S and note that for any
random variable X with density f
X
(x)

_
1
0
E(exp(N
t
) K; X > ) dt =
_
1
0
E(exp (N
t
) K X = ) (f
X
()) dt.
Thus the optimal value of ,

satises
_
1
0
E(exp(N
t
) | X =

) dt = K. (3.3)
With our choice of X =
_
1
0
B
t
dt, we conclude that
_
1
0
S exp
_
3

t (1 t/2) +t +
1
2

2
_
t 3t
2
(1 t/2)
2
__
dt = K, (3.4)
which determines

uniquely. We now have the bound

V
xed
e

_
1
0
E
_
_
Se
B
t
+t
K
_
I
__
1
0
B
s
ds >

__
dt,
and it remains to calculate the expectation. Fix t (0, 1) and let N
1
= B
t
+ t + log S
and N
2
=
_
1
0
B
s
ds

. Write
i
= E(N
i
),
2
i
= Var(N
i
) and c = Cov(N
1
, N
2
), then using
E
__
e
N
1
K
_
I (N
2
> 0)

= e

1
+
1
2

2
1

2
+c

2
_
K
_

2
_
,
where is the normal distribution function, and substituting
1
= t + log S,
2
=

2
1
=
2
t,
2
2
=
1
3
, c = t(1 t/2), we have
V
xed
e

__
1
0
Se
t+
1
2

2
t

+t (1 t/2)
1/

3
_
dt K
_

1/

3
__
.
Integrating this numerically is signicantly easier than integrating (3.1).
To see that this bound gives the same answers as that of Rogers & Shi (1995), let
Y =
_
1
0
S
t
dt, Z =
_
1
0
B
t
dt and note that E[(E(Y K|Z))
+
] = E[(E(Y K|Z))I(E(Y
K|Z) > 0)], and since from (3.3) and (3.4), E(Y K|Z) is strictly increasing in Z, we
have: E(Y K|Z) > 0 if and only if Z >

. Thus

satises E[(E(Y K|Z))

+
] =
E[(E(Y K|Z))I(Z >

)] which is just E[(Y K)I(Z >

)].
The bound of Curran (1992) arises from solving (3.4) approximately, using the following
method: let f() = E(
_
1
0
S exp(B
t
+ t) dt|
_
1
0
B
s
ds = ), and note that a reasonable
approximation to f is

f() := S exp( + /2), obtained by interchanging the orders of
integration and exponentiation. Recall that we seek

= f
1
(K) and observe that if f

f
4
then f
1
(x)

f
1
(2x f

f
1
(x)), the approximation being exact if f

f is constant.
Thus

f
1
(2K f

f
1
(K)), giving

1
_
log
_
2K
S

_
1
0
e
3(log(K/S)/2)t(1t/2)+t+
1
2

2
(t3t
2
(1t/2)
2
)
dt
_
/2
_
,
which is the continuous limit of the bound given by Curran (1992).
For the oating-strike option, let A = { :
_
1
0
S
t
dt > S
1
} and use an approximation
to A of the form A

= {
_
1
0
B
t
dt B
1
> }. With this choice,

, the optimal value of ,

satises
E
__
1
0
(S exp (B
t
+t) S exp (B
1
+)) dt
_
1
0
B
s
ds B
1
=

_
= 0
giving
exp
_
3

/2
2
/8
_
_
1
0
exp
_
3

t
2
/2 +t +
1
2

2
_
t 3t
4
/4
__
dt = 1, (3.5)
which has a unique solution.
Our lower bound for the oating-strike case is thus
V
oating
e

_
1
0
E
_
_
S exp (B
t
+t) S exp (B
1
+)
_
I
__
1
0
B
s
ds B
1
>

__
dt,
which reduces to
V
oating
e

__
1
0
Se
t+
1
2

2
t

t
2
/2
1/

3
_
dt Se
+
1
2

1/2
1/

3
__
.
Again this bound gives the same answers as that of Rogers & Shi (1995). To see this, let
Y =
_
1
0
S
t
dt, Z =
_
1
0
B
t
dt, and note rst that

E(Y S
1
Z B
1
= ) =
_
1
0
(3t
2
/2) exp(A(t, )) dt + (3/2) exp(A(1, )),
where A(t, ) = log S 3t
2
/2 +t +
1
2

2
_
t 3t
4
/4
_
. Since

solves E(Y S
1
|Z B
1
=

) = 0, we have
_
1
0
exp(A(t,

)) dt = exp(A(1,

)), so

E(Y S
1
Z B
1
= ) =
_
1
0
(3(1 t
2
)/2) exp(A(t,

)) dt > 0.
Thus E(Y S
1
Z B
1
= ) > 0 if and only if >

. A similar argument to the xed-

strike case completes the proof.
We can now generalize Currans formula to the case of oating-strike options by solving
(3.5) approximately. Let f() = e
3/2
2
/8
E(
_
1
0
e
B
t
+t
|
_
1
0
B
t
dt B
1
= ) and let
5

f() = exp( /2) be the approximation to f obtained by interchanging the orders

of integration and exponentiation. Then the solution to (3.5) is given approximately by

f
1
(2 f

f
1
(1)), giving

1
_
/2 + log
_
2 e
/4
2
/8
_
1
0
exp
_
t 3t
2
/4 +
1
2

2
_
t 3t
4
/4
__
dt
__
.
4 Upper bounds
In this section we derive a new upper bound on the value of xed-strike and oating-strike
Asian options in the form of a double integral. Rogers & Shi (1995) obtained an upper
bound by considering the error made by their lower bound (see Section 3). As an indication
of their relative accuracy, if = 0.3, = 0.09, S = 100, and T = 1 year and k = 100, the
lower bounds of Section 3 are both 8.8275 and the upper bound of this section is 8.8333.
By comparison, the upper bound given by Rogers & Shi (1995) is 9.039.
The inequality underlying the new bound is the following: let X be a random variable,
and let f
t
() be a random function with
_
1
0
f
t
() dt = 1 for all . Then
E
_
__
1
0
S exp(B
t
+t) dt X
_
+
_
= E
_
__
1
0
(S exp(B
t
+t) Xf
t
) dt
_
+
_
E
__
1
0
(S exp(B
t
+t) Xf
t
)
+
dt
_
(4.1)
=
_
1
0
E
_
(S exp (B
t
+t) Xf
t
)
+

dt.
For both the xed-strike and oating-strike cases we will use f
t
=
t
+ (B
t

_
1
0
B
s
ds),
where
t
is a deterministic function satisfying
_
1
0

t
dt = 1; and derive an expression for
t
which is approximately optimal in each case. As the bounds have a very similar derivation
we will concentrate on the xed-strike option and give the appropriate modications for the
oating-strike case at the end of the section.
Take X = K in (4.1) and rst consider the choice f
t
=
t
. To choose
t
we will minimize
the right hand side of (4.1) over the set of deterministic functions f
t
such that
_
1
0
f
t
dt = 1.
Let L(, {f
t
}) = E[
_
1
0
(S exp(B
t
+t) Kf
t
)
+
dt (
_
1
0
f
t
dt 1)] be the Lagrangian, and
consider stationarity with respect to {f
t
} for the unconstrained problem. This gives the
condition
_
1
0
(KP(S exp (B
t
+t) Kf
t
) )
t
dt = 0,
where {
t
} is some small deterministic perturbation. Thus we see that P(S exp(B
t
+t)
Kf
t
) must be independent of t. Equivalently we have log(Kf
t
/S) t =

t for some
6
constant . Thus the optimal choice for f
t
is
f
t
= (S/K) exp
_

t +t
_
,
where the constant is chosen so that
_
1
0
f
t
dt = 1. Since
_
1
0
f
t
dt is monotone increasing
in , the correct value for is easy to estimate numerically.
If instead f
t
=
t
+(B
t

_
1
0
B
s
ds), the condition for stationarity with respect to small
deterministic perturbations is
P
_
S exp(B
t
+t) K
_

t
+
_
B
t

_
1
0
B
s
ds
___
= , for all t, (4.2)
but this cannot easily be re-arranged to give the dependence of
t
on . Instead we will
use the approximation exp(B
t
) 1 + B
t
which should be reasonable for small . This
leads to the condition P(S exp(t)(1 +B
t
) Kf
t
) = for all t. Letting N
t
= S exp(t) +
(S exp(t) K)B
t
+ K
_
1
0
B
s
ds, we conclude that P(N
t
K
t
) must be independent
of t. Using the facts about the joint distribution of (B
t
,
_
1
0
B
s
ds) given in Section 2, we
deduce that

t
=
1
K
(S exp (t) +

v
t
) , (4.3)
where
v
t
= Var(N
t
) = c
2
t
t + 2(K)c
t
t(1 t/2) + (K)
2
/3, (4.4)
c
t
= S exp(t) K. (4.5)
Imposing
_
1
0

t
dt = 1 gives
=
_
K S(e

1)/
_ _ _
1
0

v
t
dt. (4.6)
We estimate the integral
_
1
0

v
t
dt numerically. We now know the constant and hence
the function
t
for our upper bound:
V
xed
e

_
1
0
E
_
_
S exp (B
t
+t) K
_

t
+
_
B
t

_
1
0
B
s
ds
___
+
_
dt.
Conditioning on B
t
= x, this becomes
V
xed
e

_
1
0
_

_
x

t
_
E
_
(a(t, x) +b(t, x)N)
+

dxdt, (4.7)
7
where N has a N(0, 1) distribution, and the functions a and b are given by
a(t, x) = S exp (x +t) K(
t
+x) +K(1 t/2)x, (4.8)
b(t, x) = K
_
1
3
t (1 t/2)
2
. (4.9)
The calculation of E[(a+bN)
+
] is straightforward and gives a(a/b) +b(a/b). In the form
of (4.7) the integrand is badly behaved near (0, 0) so we perform the change of variables
v =

t, w = x/

t, giving
V
xed
e

_
1
0
_

2v(w)
_
a(t, x)
_
a(t, x)
b(t, x)
_
+b(t, x)
_
a(t, x)
b(t, x)
__
dwdv.
This expression, combined with (4.3), (4.4), (4.5), (4.6), (4.8) and (4.9), constitutes the
upper bound in the xed-strike case.
For the case of a oating-strike option, we now take X = S
1
. Setting Z =
_
1
0
B
t
dt,
the condition for stationarity with respect to small deterministic perturbations analogous
to (4.2) is
E[S exp(B
1
+); S exp(B
t
+t) (
t
+ (B
t
Z))S exp(B
1
+))] = , t.
We approximate this by
P[S exp(B
t
+t) (
t
+(B
t
Z))S exp(B
1
+)] =

, t,
and further, using the approximation exp((B
t
B
1
)) 1 +(B
t
B
1
), by
P[exp((t 1))(1 +(B
t
B
1
)
t
+(B
t
Z))] =

, t.
This implies that for some ,

t
= exp ((t 1)) +

v
t
where v
t
= Var[exp((t 1))(1+(B
t
B
1
)) (B
t
Z)]. Since
_
1
0

t
dt = 1 we must have
= (1 (1 exp())/)/
_
1
0

v
t
dt. Our bound is thus
V
oating
e

_
1
0
E
_
_
Se
B
t
+t
(
t
+ (B
t
Z)) Se
B
1
+
_
+
_
dt
= e

_
1
0
E
_
_
e
N
1
(t)
N
2
(t)e
N
3
(t)
_
+
_
dt,
where N
1
(t) = B
t
+t +log S, N
2
(t) =
t
+(B
t

_
1
0
B
s
ds) and N
3
(t) = B
1
++log S.
If we condition on N
2
(t) = x we can perform the remaining expectation analytically. Let
8

i
(t) = E(N
i
(t)),
ij
(t) = Cov(N
i
(t), N
j
(t)) and denote by tildes the conditional distribu-
tions given N
2
(t) = x:
i
(t, x) =
i
+ (x
2
)
i2
/
22
and
ij
=
ij

i2

j2
/
22
. Finally
let v
2
= Var(N
1
(t) N
3
(t)|N
2
(t) = x) =
11
2
13
+
33
.
Our upper bound on the price of a oating-strike Asian option is then
V
oating
e

_
1
0
_

22

_
x
2

22
__
e

1
+
1
2

11

_

1

3
log(x) +
11

13
v
_
xe

3
+
1
2

33

_

1

3
log(x) +
13

33
)
v
__
dxdt,
where we take log(x) = for x 0.
5 Numerical Results
In Table 1 we consider xed-strike options and show the upper bound of Rogers & Shi
(1995), the upper and lower bounds derived in Sections 3 and 4, the approximation of
Curran (1992) and the Monte-Carlo results of Levy & Turnbull (1992). All calculations
assume = 0.09, an initial stock price of S = 100 and an expiry time of 1 year. For the
lower bound and the new upper bound, the approximate time taken (on an HP 9000/730)
is parenthesized; for the Monte-Carlo studies, the estimated standard error is bracketed
beneath.
In Table 2 we show how the upper and lower bounds of Rogers & Shi (1995) compare to
the upper bound of Section 4 and the generalization to oating-strike options of Currans
lower bound, described in Section 3. The approximate time taken is parenthesized.
References
Baxter, M. W. & Rennie, A. (1996), Financial Calculus, Cambridge University Press.
Carverhill, A. & Clewlow, L. (1990), Flexible convolution, Risk 3(4), 2529.
Curran, M. (1992), Beyond average intelligence, Risk 5(10), 60.
Geman, H. & Yor, M. (1993), Bessel processes, Asian options and Perpetuities, Mathe-
matical Finance 3(4), 349375.
Levy, E. (1992), Pricing European average rate currency options, J. Intnl. Money and
Finance 11, 474491.
Levy, E. & Turnbull, S. (1992), Average Intelligence, Risk 5(2), 59.
9
Rogers, L. C. G. & Shi, Z. (1995), The value of an Asian option, J. Appl. Prob. 32, 1077
1088.
Yor, M. (1992), Some aspects of Brownian motion, Part 1: Some Special Functionals,
Birkauser.
10
Volatility Strike K Curran lower RS lower MC result Upper bound RS upper
95 8.8088 8.8088 8.81 8.8089 8.821
(0.00016) (0.0019) [0.00] (0.013)
0.05 100 4.3082 4.3082 4.31 4.3084 4.318
(0.00012) (0.0011) [0.00] (0.019)
105 0.9583 0.9583 0.95 0.9585 0.968
(0.00012) (0.0011) [0.00] (0.019)
95 8.9118 8.9118 8.91 8.9130 8.95
(0.00016) (0.0018) [0.00] (0.019)
0.10 100 4.9150 4.9150 4.91 4.9155 5.10
(0.00023) (0.0017) [0.00] (0.020)
105 2.0699 2.0699 2.06 2.0704 2.34
(0.00023) (0.0018) [0.00] (0.021)
90 14.9827 14.9827 14.96 14.9929 15.194
(0.00023) (0.0019) [0.01] (0.024)
0.30 100 8.8275 8.8275 8.81 8.8333 9.039
(0.00023) (0.0019) [0.01] (0.024)
110 4.6949 4.6949 4.68 4.7027 4.906
(0.00023) (0.0018) [0.01] (0.028)
90 18.1829 18.1829 18.14 18.2208 18.57
(0.00023) (0.0019) [0.03] (0.028)
0.50 100 13.0225 13.0225 12.98 13.0569 13.69
(0.00023) (0.0018) [0.03] (0.063)
110 9.1179 9.1179 9.10 9.1561 9.97
(0.00023) (0.0018) [0.03] (0.064)
Table 1 Comparison of various bounds on xed-strike Asian option prices for S = 100,
= 0.09, and an expiry time of 1 year. Parenthesized numbers are computation times
in seconds, bracketed numbers are estimates of standard errors (from Curran (1992)).
11
Volatility Interest Rate Generalized RS lower Upper bound RS upper
0.05 1.2454 1.2454 1.2457 1.355
(0.00018) (0.0036) (0.027)
0.1 0.09 0.6992 0.6992 0.6997 0.825
(0.00020) (0.0054) (0.025)
0.15 0.2516 0.2516 0.2525 0.415
(0.00020) (0.0078) (0.024)
0.05 3.4044 3.4044 3.4064 3.831
(0.00017) (0.0024) (0.031)
0.2 0.09 2.6216 2.6216 2.6237 3.062
(0.00020) (0.0032) (0.033)
0.15 1.7098 1.7098 1.7124 2.187
(0.00020) (0.0044) (0.028)
0.05 5.6246 5.6246 5.6318 6.584
(0.00017) (0.0018) (0.028)
0.3 0.09 4.7382 4.7382 4.7456 5.706
(0.00020) (0.0022) (0.033)
0.15 3.6085 3.6085 3.6166 4.604
(0.00020) (0.0032) (0.033)
Table 2 Comparison of bounds on oating-strike Asian option prices for S = 100 with
an expiry time of 1 year. Parenthesized numbers are computation times in seconds.
The parameter values are those used in Rogers & Shi (1995).
12