Mathematical Finance, Vol. 7, No.

3 (July 1997), 307–324

AN ASYMPTOTIC ANALYSIS OF AN OPTIMAL HEDGING MODEL FOR

OPTION PRICING WITH TRANSACTION COSTS

A. E. WHALLEY
Mathematical Institute, Oxford University
P. WILMOTT∗
Mathematical Institute, Oxford University
and
Department of Mathematics, Imperial College, London

Davis, Panas, and Zariphopoulou (1993) and Hodges and Neuberger (1989) have presented a very
appealing model for pricing European options in the presence of rehedging transaction costs. In their
papers the ‘maximization of utility’ leads to a hedging strategy and an option value. The latter is
different from the Black–Scholes fair value and is given by the solution of a three-dimensional free
boundary problem. This problem is computationally very time-consuming. In this paper we analyze
this problem in the realistic case of small transaction costs, applying simple ideas of asymptotic analysis.
The problem is then reduced to an inhomogeneous diffusion equation in only two independent variables,
the asset price and time. The advantages of this approach are to increase the speed at which the optimal
hedging strategy is calculated and to add insight generally. Indeed, we find a very simple analytical
expression for the hedging strategy involving the option’s gamma.

KEY WORDS: option pricing, transaction costs, asymptotic analysis, nonlinear diffusion

1. INTRODUCTION
Option pricing in the presence of transaction costs has recently become a very popular
subject for research. There are two main approaches to this work in the literature: local in
time and global in time. The former was started by Leland (1985) and extended by Boyle
and Vorst (1992), Hoggard, Whalley, and Wilmott (1994) and Whalley and Wilmott (1993).
The first three of these assume hedging takes place at given discrete intervals (Boyle and
Vorst is actually a binomial model) and the last assumes flexible trading periods. In all cases
the decision whether or not to rehedge is based on minimizing the current level of risk as
measured by the variance of the hedged portfolio. Such models are often used in practice
and are invariably quick to compute. They typically result in two-dimensional nonlinear or
inhomogeneous diffusion equations for the value of an option. The global-in-time models
can be illustrated by the model of Hodges and Neuberger (1989) and Davis, Panas, and
Zariphopoulou (1993). Such models achieve an element of ‘optimality,’ since they are
based on the approach of utility maximization. The appeal of optimality is obvious, but, on
the other hand, such models do have a number of disadvantages. Two of these disadvantages
are speed of computation and the necessity of prescribing the investor’s utility function.

∗ This author thanks the Royal Society for its support.

Manuscript received January 1994; final revision received July 1995.
Address correspondence to Paul Wilmott, Centre for Industrial and Applied Mathematics, Oxford University,
24-29 St. Giles, Oxford, OX1 3LB, UK.

°
c 1997 Blackwell Publishers, 350 Main St., Malden, MA 02148, USA, and 108 Cowley Road, Oxford,
OX4 1JF, UK.

307
308 A. E. WHALLEY AND P. WILMOTT

The models are slow to compute since they usually result in three- or four-dimensional free
boundary problems. There is great practitioner resistance to the idea of utility theory.
In this paper we perform a simple asymptotic analysis of the Davis, Panas, and Za-
riphopoulou (1993) model. We show how, in the limit of small transaction costs, their
three-dimensional free boundary problem reduces to a much simpler two-dimensional in-
homogeneous diffusion equation of the form found in the local-in-time models. We thus
bring together the competing philosophies behind modeling transaction costs. The asymp-
totic formulas for the hedging strategy we present here have been tested empirically by
Mohamed (1994), and found to be the best strategy he tested.
Perturbation analysis is a very powerful tool of applied mathematics. It is used to great
effect in areas such as fluid mechanics (Hinch 1991), because it reveals the salient features of
the problem while remaining a good approximation to the full but more complicated model.
As yet, the technique has, to our knowledge, rarely been used in finance. For this reason,
we shall at times walk the reader very slowly through the calculations. For comparison,
for an asymptotic analysis of the Morton and Pliska (1995) portfolio management problem
with transaction costs, see Atkinson and Wilmott (1993).
In Section 2 we very briefly describe the model of Davis, Panas, and Zariphopoulou
(1993), the interested reader should read that paper carefully in conjunction with this. In
Section 3 we consider the asymptotic limit of small transaction costs. This results in an
inhomogeneous diffusion equation for the price of an option. In Section 4 we compare the
model with others and draw conclusions.
Recall that in the absence of transaction costs the Black–Scholes equation for the value
of an option is

σ 2 S2
(1.1) Wt + r SW S + W SS − r W = 0.
2
Here S is the underlying asset price, t is time, r the interest rate, assumed deterministic, σ
the volatility of the underlying, and W (S, t) is the value of an option. This equation must be
solved for t < T and 0 ≤ S < ∞. On t = T we must impose a final condition, amounting
to the payoff function for the option in question. For example, for a call option with strike
price E we have

W (S, T ) = max(S − E, 0).

This is the problem to be solved in the absence of costs. In the presence of costs, we shall
find an equation similar to the Black–Scholes equation but with additional small terms that
allow for the cost of rehedging and that are nonlinear in the option’s gamma. In common
with the Davis, Panas, and Zariphopoulou paper, we are initially considering the valuation
of a short European call option. We shall continue to use W to denote the Black–Scholes
value of a European option.

2. THE MODEL OF DAVIS, PANAS, AND ZARIPHOPOULOU

In the model of Davis, Panas, and Zariphopoulou (1993) the writing price of a European
option is defined in terms of a utility maximization problem. Simply put, the option price
(to the writer) is obtained by a comparison of the maximum utilities of trading with and
without the obligation of fulfilling the option contract at expiry. When there are no costs
this results in the Black–Scholes value for the option (Black and Scholes 1973).
 ASYMPTOTIC ANALYSIS OF AN OPTIMAL HEDGING MODEL 309

The asset price S is assumed to follow the random walk

d S = µS dt + σ S d X,

where µ and σ are constant and X is a Brownian motion.

When the utility function takes the special form U(x) = 1 − exp(−γ x) (so that γ is the
index of risk aversion) Davis et al. find that the option price V (S, t) is given by

µ ¶
δ(T, t) Q w (S, 0, t)
(2.1) V (S, t) = log ,
γ Q 1 (S, 0, t)

where T is the expiry date, δ(T, t) = e−r (T −t) , and Q 1 (S, y, t) and Q w (S, y, t) both satisfy
the following equation

½ ¾
∂Q γ (1 + ²)S Q ∂Q γ (1 − ²)S Q ∂ Q ∂Q σ 2 S2 ∂ 2 Q
min + , − − , + µS + = 0.
∂y δ ∂y δ ∂t ∂S 2 ∂ S2

Here ² measures the transaction costs: A trade of N shares will result in a loss of ² N S.
This cost structure represents bid–offer spread,1 or more generally commissions and costs
that are proportional to the value of the assets traded. The independent variable y measures
the number of shares held in the optimally hedged portfolio. The two functions Q 1 and
Q w must satisfy certain final conditions, analogous to the payoff profile of the option; for
example, for a call option

(2.2) Q 1 (S, y, T ) = exp(−γ c(S, y))

and
½
exp(−γ c(S, y)) S≤E
(2.3) Q w (S, y, T ) =
exp(−γ (c(S, y) + E − S)) S>E
where
(
(1 + ²)y S y<0
c(S, y) =
(1 − ²)y S y ≥ 0.

So the final condition for the second problem (with subscript w) is equal to that of the first
problem (with subscript 1) modified by the effects of the potential liability at expiry of the
European call (after transaction costs). Note we are assuming here that the option is settled
in cash. For options with delivery of the asset on exercise the analysis below remains the
same; the final conditions merely alter.
Finally, to fully pose the problem we must specify that for t < T , Q, ∂ Q/∂ S, and
∂ 2 Q/∂ S 2 must all be continuous.

1 Davis et al. consider the slightly more general case in which there are different levels of cost for buying and
selling.
310 A. E. WHALLEY AND P. WILMOTT

FIGURE 2.1. A schematic diagram of (S, y) space showing the buy, sell, and no-transaction
regions.

This is a free boundary problem. It is explained by Davis et al. how the (S, y) space
divides into three regions, shown schematically in Figure 2.1. The writer of the option
must always maintain his portfolio in the region of the (S, y) space bounded by the two
outer curves, while inside this region he does not transact. Should a movement of the asset
price take the writer to the edge of this no-transaction region he must trade so as to just stay
inside. If he hits the top boundary he must sell shares, if he hits the bottom boundary he
must buy shares. The middle line in Figure 2.1 is the curve along which the investor must
move in the absence of transaction costs; this curve is denoted by

y = y ∗ (S, t).

Both y ∗ and the position of the upper and lower boundaries are to be found. We shall find
simple analytical expressions for all three of these curves.
In the buy region we have

∂Q γ (1 + ²)S Q
(2.4) + = 0.
∂y δ

In the sell region we have

∂Q γ (1 − ²)S Q
(2.5) + = 0.
∂y δ
 ASYMPTOTIC ANALYSIS OF AN OPTIMAL HEDGING MODEL 311

In the no-transaction region we have

∂Q ∂Q σ 2 S2 ∂ 2 Q
(2.6) + µS + = 0.
∂t ∂S 2 ∂ S2

This is the free boundary problem we shall shortly solve asymptotically. Because the two
problems for Q 1 and Q w are identical except for the final data, we need only perform the
analysis for one of them. When we come to apply the final data we will distinguish between
Q 1 and Q w as necessary. As mentioned above, across the two free boundaries (the outer
curves in Figure 2.1) Q, ∂ Q/∂ S, and ∂ 2 Q/∂ S 2 must all be continuous.

3. ASYMPTOTIC ANALYSIS FOR SMALL LEVELS OF TRANSACTION COSTS

Equation (2.4) is very easy to solve explicitly. The solution for Q(S, y, t) in the buy region
is found to be
µ ¶
γ Sy γ S²y −
(3.1) Q = exp − − + H (S, t; ²)
δ δ

where H − is, as yet, an arbitrary function of S and t that comes from solving the ordinary
differential equation (2.4): in this equation S and t are effectively parameters.
In the sell region we can similarly solve (2.5) to get

µ ¶
γ Sy γ S²y +
(3.2) Q = exp − + + H (S, t; ²) .
δ δ

This contains another arbitrary function H + . The two expressions (3.1) and (3.2) are the
exact, general solutions of (2.4) and (2.6).
The solution in the no-transaction region is much harder to find. Indeed, we shall not find
the general solution, rather we shall find the asymptotic solution valid for small ². The first
stage in determining this solution is to expand Q in an asymptotic series in powers of ².
We write the solution in the no-transaction region as

µ
γ Sy ∗ γ SY
(3.3) Q = exp − + H0 (S, t) − ² 1/3 + ² 1/3 H1 (S, t) + ² 2/3 H2 (S, t)
δ δ
¶
+ ² H3 (S, t) + ² 4/3 H4 (S, Y, t) + ² 5/3 H5 (S, Y, t) + · · · .

There are two very important things to note about this expression. First, we have chosen
to expand in powers of ² 1/3 . This is not an arbitrary choice. As we perform our analysis, we
shall see how such a choice is the natural one. (Shreve and Soner, 1994, have results that
suggest a similar asymptotic scale for the width of the no-transaction interval for an optimal
investment and consumption model with transaction costs under a different utility function,
and notes that Fleming, Grossman, Vila, and Zariphopoulou (1990) have also obtained this
312 A. E. WHALLEY AND P. WILMOTT

scale.) Second, we have translated the y coordinate according to

(3.4) y = y ∗ (S, t) + ² 1/3 Y.

Thus Y is a rescaled variable (see Figure 2.1). It is a measure of the difference between the
number of shares actually held in the portfolio and the ideal number we would hold in the
absence of transction costs, y ∗ . We shall find an explicit expression for y ∗ as a function of
S and t. Y turns out to be a more natural variable to use than y. The factor of ² 1/3 represents
the scale of the asymptotic width of the no-transaction region for this type of transaction
costs (proportional to value traded).
Observe how, in (3.3), there is Y dependence at O(² 1/3 ) and O(² 4/3 ). The former is
forced by the leading terms in (3.1) and (3.2) and continuity of slope at the boundary of
the no-transaction region. The reason for the latter is similar and the details will become
apparent. It is such continuity requirements that actually force on us the special choice
of ² 1/3 .
As yet (3.3) does not satisfy the equation in the no-transaction region. We must now
find the functions Hi such that this equation and all relevant boundary and smoothness
conditions are satisfied. We shall see, in performing this analysis, that the choice of a series
expansion in powers of ² 1/3 is inevitable.
Since the derivatives in (2.6) are with respect to t and S keeping y fixed, then

∂ ∂
→ ² −1/3 ,
∂y ∂Y

∂ ∂ ∂
→ − ² −1/3 yS∗ ,
∂S ∂S ∂Y

∂ ∂ ∂
→ − ² −1/3 yt∗ .
∂t ∂t ∂Y
Thus we readily find from (3.3) and (3.4) that

µ
∂Q γ Syt∗ r γ Sy ∗ r γ SY
≡ Qt = − + H0t (S, t) + + ² 1/3
∂t δ δ δ
+ ² 1/3 H1t (S, t) + ² 2/3 H2t (S, t) + ² H3t (S, t)
µ ¶ ¶
γS
+ ² 4/3 H4t (S, Y, t) + yt∗ − ² H4Y + · · · Q,
δ
µ
∂Q γ SyS∗ γ y∗ γY
≡ QS = − + H0S (S, t) − − ² 1/3
∂S δ δ δ
+ ² 1/3 H1S (S, t) + ² 2/3 H2S (S, t) + ² H3S (S, t)
µ ¶ ¶
γS
+ ² 4/3 H4S (S, Y, t) + yS∗ − ² H4Y + · · · Q,
δ
 ASYMPTOTIC ANALYSIS OF AN OPTIMAL HEDGING MODEL 313

and

µ
∂2 Q γ SyS∗ γ y∗ γY
≡ Q SS = − + H0S (S, t) − − ² 1/3 + ² 1/3 H1S (S, t)
∂S 2 δ δ δ
µ ¶ ¶2
∗ γS
+ ² H2S (S, t) + ² H3S (S, t) + yS
2/3
− ² H4Y + · · · Q
δ
¡
+ H0SS (S, t) + ² 1/3 H1SS (S, t) + ² 2/3 H2SS (S, t)
´
∗
H4Y + ² 2/3 yS∗ H4Y Y − 2²yS∗ H4Y S + ²ys∗ H5Y Y + · · · Q.
2 2
+ ² H3SS (S, t) − ²ySS

It will be observed that each of the above can be slightly simplified. We have retained
them in this form to help the reader perform his own calculations.
The advantage of asymptotic analysis will now become clear when we perform the next
step, to substitute these expressions into (2.6) and equate powers of ² 1/3 .

3.1. The O(1) Equation

To leading order (O(1)) we find that

µ ¶
r γ Sy ∗
Qt = H0t + Q,
δ

µ ¶
γ y∗
QS = H0S − Q,
δ

à µ ¶2 !
γ y∗
Q SS = H0SS + H0S − Q.
δ

Thus, to leading order equation (2.6) becomes

µ ¶ µ ¶
r γ Sy ∗ γ y∗ σ 2 S2 γ y∗ 2 σ 2 S2
(3.5) H0t + + µS H0S − + H0S − + H0SS = 0.
δ δ 2 δ 2

3.2. The O(² 1/3 ) Equation

We can take this procedure to the next order, equating powers of ² 1/3 . We find that

µ ¶ µ ¶µ ¶
r γ SY γY γY γ y∗ σ 2 S2
+ H1t +µS − + H1S +σ 2 S 2 − + H1S H0S − + H1SS = 0.
δ δ δ δ 2
314 A. E. WHALLEY AND P. WILMOTT

This equation contains a term proportional to Y and one independent of Y . Since all the
other terms in the equation are independent of Y , these terms must separately be zero. From
the first of these we find that

δ δ(µ − r )
(3.6) y ∗ (S, t) = H0 + .
γ S γ Sσ 2

Thus, if we can find H0 then we have found the leading order expression for y ∗ .
Equation (3.6) determines the hedging strategy in the absence of transaction costs y ∗ , in
terms of the leading order ‘option value’ H0 . If we substitute this back into (3.5) we find
that H0 satisfies
σ 2 S2 (µ − r )2
(3.7) H0t + H0SS + r S H0S = .
2 2σ 2
If we write
γ
H0 (S, t) = V0 (S, t),
δ
we have
δ(µ − r )
(3.8) y ∗ (S, t) = V0S +
γ Sσ 2
as given by Davis et al., and equation (3.7) becomes

σ 2 S2 δ(µ − r )2
(A) V0t + V0SS + r SV0S − r V0 = .
2 2γ σ 2

The particular solution of this with zero final data is

δ(µ − r )2 (T − t)
− .
2γ σ 2

The general solution is thus any solution satisfying the Black–Scholes equation plus this
particular solution.
We then retrace our steps to get from V0 to V , the option price, using (3.3) and (2.1)
(for both Q w and Q 1 ). We find that the leading order final data in the portfolio without the
option liability, (Q 1 ), is V0 (S, T ) = 0, whereas in the portfolio with the call option liability,
(Q w ), it has the usual payoff functional form V0 (S, T ) = −max(S − E, 0). So from the
linearity of (3.2(A)) we see that, to leading order, (or in the absence of any costs) the option
value is simply the Black–Scholes value. Similarly the extra number of shares required
in the portfolio with the additional option liability is, to leading order, the Black–Scholes
delta value.
We now consider the terms independent of Y , which give an equation for H1

µ ¶
γ y∗ σ 2 S2
H1t + µS H1S + σ 2 S 2 H1S H0S − + H1SS = 0.
δ 2
 ASYMPTOTIC ANALYSIS OF AN OPTIMAL HEDGING MODEL 315

∗
If we substitute for H0S − γ δy using (3.8), and set V1 = δ H1 /γ as above, we find that V1
satisfies the Black–Scholes equation. The final condition for this equation for both Q w and
Q 1 is V1 (S, t) = 0. (This is found by expanding the final conditions in powers of ² 1/3 and
considering the terms of O(² 1/3 ).)
Thus, V1 is identically zero for all S and t < T , and so the leading order correction to
the Black–Scholes value occurs at the O(² 2/3 ) level.

3.3. The O(² 2/3 ) Equation

We now take the analysis to higher order to find the correction to the Black–Scholes value
due to transaction costs. If we examine the O(² 2/3 ) terms in (2.6), we find that

σ 2 S2 σ 2 S 2 ∗2 γ 2σ 2 S2Y 2
H2t + r S H2S + H2SS + yS H4Y Y + = 0.
2 2 2δ 2

This is an ordinary differential equation2 for H4 which is easily integrated to give

µ ¶
Y2 σ 2 S2 γ 2Y 4
H4 (S, Y, t) = − H2t + r S H2S + H2SS − + aY + b.
σ 2 S 2 yS∗ 12δ 2 yS∗
2 2
2

We now have to join this solution in the no-transaction cost region with the solutions
(3.1) and (3.2) in the buy and sell regions respectively.
Let us use the notation Y + (S, t) and −Y − (S, t) to denote the Y -coordinates of the
boundaries of the no-transaction region. These are, of course, unknown and must be
determined as part of the solution by imposing suitable smoothness conditions. As stated
previously, we require Q and its first two derivatives with respect to Y to be continuous at
Y = Y + and Y = −Y − . From (3.1) and (3.2) we can see that continuity of the gradient of
Q at Y = Y + and Y = −Y − is ensured by

γS
H4Y = on Y = Y +
δ

and
γS
H4Y = − on Y = −Y − .
δ
Thus,
µ ¶
2Y + γ 2Y +
3
σ 2 S2 γS
− H2t + r S H2S + H2SS − +a =
σ 2 S 2 yS∗
2
2 3δ 2 yS∗
2
δ
and
µ ¶
2Y − γ 2Y −
3
σ 2 S2 γS
H2t + r S H2S + H2SS + +a =− .
σ 2 S 2 yS∗
2
2 3δ 2 yS∗
2
δ

2 The inhomogeneous term, proportional to Y 2 , at this order has forced Y dependence in H4 .

316 A. E. WHALLEY AND P. WILMOTT

The second derivative of Q with respect to Y must also be continuous, that is, zero, at
Y = Y + and Y = −Y − . If this were not the case then there could be no finite value for the
option price.3 Thus,

µ ¶
γ 2 Y +,−
2
2 σ 2 S2
H2t + r S H2S + H2SS =−
σ 2 S 2 yS∗ δ 2 yS∗
2 2
2

and a = 0. We conclude from this that the no-transaction region is to leading order
symmetric about the Black–Scholes hedging strategy, i.e., Y − = Y + . Eliminating Y + and
Y − from these equations we arrive at

à !2/3
3γ 2 S 4 σ 3 yS∗
2
σ 2 S2 1
(3.9) H2t + r S H2S + H2SS = − .
2 2 2δ 2

We also find that the edges of the ‘hedging bandwidth,’ Y = Y + and Y = −Y − , are
given by
à !1/3
3SδyS∗
2
+ −
(3.10) Y =Y = ,
2γ

to leading order.
We cannot stress the importance of this last result enough. As far as implementation of
the optimal hedging is concerned, we need to know the boundaries of the no-transaction
region. These are given by very simple analytic expressions in terms of yS∗ , via equation
(3.10), which in turn is simply related to the option’s gamma by equation (3.6). We shall
see this more clearly in the final section of this paper.
Equation (3.9) is to be solved subject to the final condition

H2 (S, T ) = 0.

By letting
γ
H2 = V2 (S, t)
δ
we can write (3.9) as

µ ¶2/3 µ¯ ¯¶4/3
σ 2 S2 δ 3γ 2 S 4 σ 3 ¯ ¯
(3.11) V2t +r SV2S + V2SS −r V2 = − ¯V0 − δ(µ − r ) ¯
2 2γ 2δ 2 ¯ SS γS σ ¯
2 2

It is now important to distinguish between the two problems for Q w , the problem including
the option liability, and Q 1 , the problem without the option. The V2 component of Q 1

3 Recall that the number of the underlying asset held contains a term V , as in Black–Scholes, to leading order.
0S
The infinite number of trades in a finite time required at the boundary of the no-transaction region would lead to
an infinite cost unless the gamma of the option is zero at the boundary.
 ASYMPTOTIC ANALYSIS OF AN OPTIMAL HEDGING MODEL 317

satisfies (3.11) with V0SS = 0, i.e.,

µ ¶2/3
σ 2 S2 1 3 δ(µ − r )4/3
V2t + r SV2S + V2SS − r V2 = − .
2 2 2σ γ

This has solution with zero final data

µ ¶2/3
1 3 δ(µ − r )4/3 (T − t)
V2 = .
2 2σ γ

Using W (S, t) to denote the Black–Scholes option value we see that the V2 component
of Q w satisfies (3.11) with V0SS being the Black–Scholes value for the gamma, i.e., W SS .
Thus we see that the option value correct to O(² 2/3 ) is simply

à µ ¶2/3 !
1 3 δ(µ − r )4/3 (T − t)
V (S, t) = W (S, t) + ² 2/3
V2 (S, t) − + ···,
2 2σ γ

where V2 satisfies (3.11) with V0SS = W SS .

3.4. The O(²) Equation

It is remarkable that the algebraic complexity of the problen is still manageable at the
O(²) level. We can thus take the asymptotic analysis even further.
The O(²) terms in expression (2.6) give

¡ ¢ σ 2 S2 ¡
(3.12) H3t − yt∗ H4Y + µS H3S − yS∗ H4Y + ∗
H3SS − ySS H4Y − 2yS∗ H4Y S
2
µ ¶ ¶
γ y∗ ¡ ¢ 2γ Y
+ yS∗ H5Y Y + 2 H0S − H3S − yS∗ H4Y −
2
H2S = 0.
δ δ

This may be interpreted first as an ordinary differential equation for H5 and then, given
sufficient boundary conditions, as a partial differential equation for H3 (just as in the H2 ,
H4 problem of Section 3.3). To determine the correct boundary conditions, recall that we
must have continuity of first and second derivatives with respect to Y at all orders of ².
Thus

γS
(3.13) H4Y + ² 1/3 H5Y = ±
δ

on the top and bottom free boundaries. By going to higher order we must also expand the
position of the free boundaries as power series in ² 1/3 . Transferring the boundary condition
318 A. E. WHALLEY AND P. WILMOTT

(3.13) onto the known leading order boundaries y = Y + and y = −Y − , we find that

H5Y = 0 on y = Y+ and y = −Y − ,

since H4Y Y = 0 on y = Y + and y = −Y − .

Now integrate (3.12) from y = −Y − to y = Y + . We find that

σ 2 S2
(3.14) H3t + r S H3S + H3SS = 0
2

R Y+
(since −Y − H4Y dY = 0). With H3 = γ V3 /δ we can now see that V3 satisfies the Black–
Scholes equation.
The final data for this equation is, for both the 1 and the w problems,

(µ − r )
H3 (S, T ) = .
σ2

This is found by expanding (2.2) and (2.3) in powers of ² 1/3 . The solution of (3.14) with
this final data is simply
(µ − r )
H3 (S, t) = .
σ2
The only remaining step in calculating the option value to O(²) is to apply continuity
between the no-transaction region and the buy region. We have

µ ¶
γS ∗
H − = H0 + ² 2/3 H2 + ² H3 + y .
δ

Finally, since the option value depends on Q(S, 0, t), we need the result

Q(S, 0, t) = exp(H − ).

From (2.1) we now have

à µ ¶2/3 !
1 3 δ(µ − r )4/3 (T − t)
V (S, t) = W (S, t)+² 2/3
V2 (S, t)− +² SW S + O(² 4/3 ),
2 2σ γ

where V2 satisfies (3.11) with V0SS = W SS . Observe that the O(²) correction to our earlier
result is simply the cost of changing the number of shares in the portfolio in order to set
up the initial hedge! Recall that it is assumed that the option obligation will be held until
maturity, and that the final condition incorporates any transaction costs payable at maturity
in order to unwind the hedge.
 ASYMPTOTIC ANALYSIS OF AN OPTIMAL HEDGING MODEL 319

FIGURE 4.1. The hedge ratio and no-transaction band as functions of S without the option
liability. See the text for details of parameters.

4. RESULTS AND CONCLUSIONS

In this section we give the results of our asymptotic limit of the Davis et al. model and make
comparisons with their numerical results. To be specific we have concentrated on examples
given in Davis et al.
First, we consider a European call option with exercise price E = 0.5 and time to expiry
0.3. Other parameters are r = 0.07, σ = 0.2, µ = 0.1, and γ = 1.0. The level of
transaction costs is such that ² = 0.002.
In Figure 4.1 we plot the solution for y ∗ and the hedging boundaries against S for the
first problem (denoted by subscript 1), which does not have the option liability at expiry.
The solution in the absence of costs, y ∗ , is the middle curve. The outer, bold curves are the
boundaries of the no-transaction region when there are nonzero transaction costs as detailed
above. Recalling our expressions for y ∗ , equation (3.6), and Y + and Y − , equation (3.10),
these three curves are given by

δ(µ − r )
y = y ∗ (S, t) =
γ Sσ 2

and

à !1/3
3SδyS∗
2
δ(µ − r )
y = ± ² 1/3
γ Sσ 2 2γ
µ ¶1/3
δ(µ − r ) 3δ 3 (µ − r )2
= ± ² 1/3 .
γ Sσ 2 2γ 3 σ 4 S 3
320 A. E. WHALLEY AND P. WILMOTT

FIGURE 4.2. The hedge ratio and no-transaction band as functions of S for the problem with
the option liability. See the text for details of parameters.

In Figure 4.2 we plot the equivalent solutions for the second problem (denoted by the
subscript w), which includes the option liability at expiry. Again y ∗ , the solution in the
absence of transaction costs, is the middle curve and two bold curves are the boundaries of
the no-transaction region. These three curves are given by

δ(µ − r )
y = y ∗ (S, t) = W S +
γ Sσ 2

and

à !1/3
3SδyS∗
2
δ(µ − r )
y = WS + ± ² 1/3
γ Sσ 2 2γ
à ¯ ¯ !1/3
δ(µ − r ) 3Sδ ¯¯ δ(µ − r ) ¯¯2
= WS + ± ² 1/3 W SS −
γ Sσ 2 2γ ¯ γ S2σ 2 ¯

where W is the Black–Scholes call value.

This plot is of particular interest. Because y ∗ has turning points, the width of the no-
transaction region (which is proportional to (yS∗ )2/3 ) goes to zero. This gives the ‘string of
sausages’ shape shown in Figure 4.2. This result has an obvious financial interpretation.
At the two turning points of y ∗ , a relatively large change in the share price can be tolerated
before rehedging is necessary. In stochastic terms, to leading order, we have

dy ∗ = yS∗ d S + · · · .
 ASYMPTOTIC ANALYSIS OF AN OPTIMAL HEDGING MODEL 321

FIGURE 4.3. The difference between the asymptotic limit of the Davis et al. model and the
Black–Scholes value for a European call. The bold curve is the sum of the other two curves.
See the text for details of parameters.

Away from turning points dy ∗ is of the same order as d S. However, at the two turning
points dy ∗ becomes deterministic and of higher order. Thus it is possible to impose tighter
bounds on the no-transaction region and this is exactly what is seen.
In deriving these plots we have not had to solve any differential equation since the
functions y ∗ and Y + depend only on W , the Black–Scholes call value.
We now move on to another example. The parameters in this case are ² = 0.002, γ = 1.0,
σ = 0.05, r = 0.085, and µ = 0.1. We consider a European call with exercise price 20
and with up to three years until expiry.
The plot in Figure 4.3 shows the difference between the asymptotic limit of the Davis
et al. model and the Black–Scholes call option value. This is the bold curve. It has two
components, the O(² 2/3 ) part and the O(²) part, and these two curves are also shown in
the figure. The bold curve is the sum of the other two curves. Note that the O(² 2/3 ) and
the O(²) curves are similar in magnitude. This is because they differ by a factor of order
O(² 1/3 ) which for ² = 0.002 is 0.13 and not very small.
This plot (and Figure 4.4) has required the solution of (3.11). The solution shown
in Figure 4.3 was computed by a simple explicit finite-difference scheme and thus took
approximately the same time to run as the binomial solution of an American option.
In Figure 4.4 we plot the time dependence of the difference between the asymptotic
limit of the Davis et al. model and the Black–Scholes value, for the same parameters as in
Figure 4.3 with S = 19. This is the bold curve and is the sum of the lower two curves.
Again these O(² 2/3 ) and O(²) curves are similar in magnitude. Nevertheless this asymptotic
solution shows very good agreement with the numerical results of Davis et al., also plotted.
To finish this paper, let us recall the model of Leland (1985) and of Hoggard et al. (1994).
In that model it is assumed that a delta-hedged portfolio is rehedged every fixed time period
δt. The option is then valued so as to give the hedged portfolio the same expected return as
that from a bank. With the same cost structure as above it is readily found that for a short
322 A. E. WHALLEY AND P. WILMOTT

FIGURE 4.4. The difference between the asymptotic limit of the Davis et al. model and the
Black–Scholes value for a European call as a function of time to expiry. The unlabeled
bold curve is the sum of the lower two curves. Numerical results taken from Davis et al.
are also shown.

position
r
σ 2 S2 2
Vt + r SVS + VSS − r V = − ²σ S 2 |VSS | .
2 π δt
By writing V (S, t) = W (S, t) + ²V2 (S, t) + · · · we have

r
σ 2 S2 2
(3.15) V2t + r SV2S + V2SS − r V2 = − σ S 2 |W SS | ,
2 π δt

with V2 (S, T ) = 0.
Now recall the model of Whalley and Wilmott (1993). In that model the investor delta
hedges with rehedging determined by ‘market movements.’ If the difference between
the delta and the number of assets actually held becomes greater than d(S, t)/S then the
portfolio is rehedged to the delta value giving the portfolio the minimum variance. The
function d(S, t) which specifies the hedging bandwidth must be prescribed by the investor.
The option value is again determined by assuming that the expected return is equal to the
risk-free rate. With V (S, t) = W (S, t) + ²V2 (S, t) + · · · it is found that this time the
correction term for a short position satisfies

σ 2 S2 σ 2 S4 2
(3.16) V2t + r SV2S + V2SS − r V2 = − 1/2 W SS ,
2 d

with V2 (S, T ) = 0. This models a strategy commonly used in practice.

Now we can see the similarities between the three different models. All of them give
the Black–Scholes value to leading order with a smaller order correction. This correction
differs between models, but in all cases satisfies an ‘inhomogeneous Black–Scholes-type
 ASYMPTOTIC ANALYSIS OF AN OPTIMAL HEDGING MODEL 323

equation,’ where the extra term resulting from the transaction costs depends on some power
of the Black–Scholes option gamma (W SS ).
In Whalley and Wilmott (1993) many issues arising from such equations are discussed.
Briefly, these include the following.

1. Nonlinearity. Since the right-hand side of the V2 equation is in each case a non-
linear function of the Black–Scholes value of gamma, W SS , there will inevitably
be different values for short and long positions. Also portfolios of options must be
treated as a whole and not as the sum of individually valued components.
2. Negative option prices. With the more general costs structure discussed in Hoggard
et al. and Whalley and Wilmott (not simply bid–offer spread) it is possible to arrive
at negative option prices. (To see this, consider the commission component of costs.
If a fixed amount is paid at each rehedge, then for small asset values the call option
can have a negative value.) This suggests modifying hedging strategies to allow
the possibility of not rehedging if to rehedge would make the option value negative.
This introduces a free boundary below which (for a call) the option should not be
rehedged. However, it is unlikely that the simple bid–offer spread considered here
would lead to negative option prices.
3. American options. As also mentioned in Davis et al. it is the owner of the American
option who controls its exercise. It is difficult to optimally value an American option
unless the owner’s hedging and exercise strategy is known. This entails at least
knowing all of his estimates of the parameters.

From the point of view of the numerical solution of these equations we can say that the
inhomogeneous equations will not take significantly longer to solve by finite-difference
methods than the basic inhomogeneous Black–Scholes equation. Thus, by performing this
simple asymptotic analysis of the Davis et al. model, we have made its use a practical
possibility.

REFERENCES

ATKINSON, C. AND P. WILMOTT (1995): “Portfolio management with transaction costs: An asymptotic
analysis.” Math. Finance, 5, 357–367.
BLACK, F. AND M. SCHOLES (1973): “The pricing of options and corporate liabilities.” J. Political
Economy, 81, 637–54.
BOYLE, P. P. AND T. VORST (1992): “Option replication in discrete time with transaction costs.”
J. Finance, 47, 271.
DAVIS, M. H. A., V. G. PANAS, AND T. ZARIPHOPOULOU (1993): “European option pricing with
transaction costs.” SIAM J. Control Optim., 31, 470–493.
FLEMING, W. H., S. D. GROSSMAN, J.-L. VILA, AND T. ZARIPHOPOULOU (1990): “Optimal portfolio
rebalancing with transaction costs.” Working paper, Division of Applied Mathematics, Brown
University.
HINCH, E. J. (1991): Perturbation Methods. New York: Cambridge University Press.
HODGES, S. D. AND A. NEUBERGER (1989): “Optimal replication of contingent claims under trans-
action costs.” Rev. Futures Markets, 8, 222–239.
HOGGARD, T., A. E. WHALLEY, AND P. WILMOTT (1994): “Hedging option portfolios in the presence
of transaction costs.” Adv. Futures Opt. Res., 7, 21.
LELAND, H. E. (1985): “Option pricing and replication with transaction costs.” J. Finance, 40, 1283–
301.
324 A. E. WHALLEY AND P. WILMOTT

MOHAMED, B. (1994): “Simulations of transaction costs and optimal rehedging.” Appl. Math. Fin.,
1, 49–63.
MORTON, A. AND S. PLISKA (1995): “Optimal portfolio management with fixed transaction costs.”
Math. Finance, 5, 337–356.
SHREVE, S. E. AND H. M. SONER (1994): “Optimal nvestment and consumption with transaction
costs.” Ann. Applied Prob., 4, 680–692.
WHALLEY, A. E. AND P. WILMOTT (1993). “A hedging strategy and option valuation model with
transaction costs.” OCIAM Working paper, Mathematical Institute, Oxford.