PRICING DISCRETELY SAMPLED

PATH-DEPENDENT EXOTIC OPTIONS USING

REPLICATION METHODS
M. S. JOSHI

Abstract. A semi-static replication method is introduced for pricing discretely sampled path-dependent options. It depends upon
buying and selling options at the reset times of the option but does
not involve trading at intervening times. The method is model independent in that it only depends upon the existence of a pricing
function for vanilla call options which depends purely on current
time, time to expiry, spot and strike. For the special case of a
discrete barrier, an alternative method is developed which involves
trading only at the initial time and the knockout time or expiry of
the option.

1. Introduction
In this paper, we examine the problem of pricing path-dependent
exotic options via the use of a new replication method. Our fundamental assumption is that the only state variable is spot and that it in
combination with the current time determines the price of any specified
call or put option. One advantage of this approach is that it naturally
allows the inclusion of smile information in the pricing.
Thus we assume the existence of a deterministic function,
Price(S, K, t, T ),
which gives the price of a call option with expiry T and strike K,
at time t when spot is S. We shall say that a model which implies
this property is a deterministic-smile model as it implies that smiles
at future times are determined by the value of spot. Several popular
models can be fit into this framework. In particular, the Black-Scholes
model, the Dupire model, [10], Mertons jump-diffusion model, [18],
and the variance gamma model, [15, 16, 17], all have this property.
We remark that stochastic volatility models, eg [12], do not have this
property as the volatility is a second state variable.
1

M. S. JOSHI

In Section 2, we prove that under certain hypotheses knowledge of

the pricing function determines the price of discretely sampled exotic
options. These results extend to options with Bermudan features.
In the rest of the paper, we then examine how these results can be
applied in practical terms. Whilst our method is quite general and
applies to a large class of options, we develop it in detail in Sections 3,
4 for the discretely sampled arithmetic Asian option, and then indicate
in Section 5 how to adapt it to many other options. Recall that a
discretely sampled Asian option is an option on the average of the spot
at certain specified times. Thus if the specified times are t1 < t2 <
< tn , and the strike of the option is K, then the Asian call and
Asian put will pay at time tn
!
!
n
n
1X
1X
(1.1)
St K
and K
St
,
n j=1 j
n j=1 j
+

respectively.
Our replication method is semi-static in that we only rehedge at the
times tj but dynamic in that at those times, we buy and sell portfolios
of options depending on the value of spot and the realized path up to
that point. We emphasize that we regard our trading strategy as a
computational device rather than as a strategy which would actually
be carried out by a trader.
The interesting fact about (1.1) is that whilst the payoff depends
on the value of spot at the times t1 , . . . , tn it does so in a particularly
simple way which is essentially one-dimensional. In particular, if we let
i

(1.2)

1X
St ,
Ai =
i j=1 j

then we have that

i
1
Ai +
St .
i+1
i + 1 i+1
Thus the value of an Asian option at time t = ti , is a function of Sti
and Ai the precise values of St1 , . . . , Sti1 are irrelevant as all relevant
information has been encoded in Ai .

(1.3)

Ai+1 =

The upshot of this is that we can apply backwards methods to the

pricing of Asian options, provided we introduce the auxiliary variables,
Ai , the realized average of the path up to time ti . In fact, this is a
standard approach to implementing PDE methods for the pricing of
Asian options. See [2, 7, 13, 22].

REPLICATION METHODS

The fundamental difference between our approach and standard numerical PDE methods, is that we only solve for option values at the
times ti without looking at the intervening time steps. Indeed, if the
reader is particularly wedded to PDE approaches, and the model being used has a PDE interpretation then one can view the replicating
options as being a basis for the solutions of the PDE, [14]. It is also
important to realize that this approach does not require the existence
of a PDE describing the price evolution. The similarity here between
PDE methods and replication methods suggests a vague general rule
that if an option can be priced using PDEs in the Black-Scholes world
then it can be priced using replication methods for any deterministic
smile model.
One advantage of our approach is that the initial portfolio for the
replication is independent of the initial value of spot. This means
that the price, delta and gamma can be immediately read off from
the portfolio for any value of spot. As the replicating portfolio does
not change until the first averaging time, we can in fact get value,
delta, gamma and theta for any value of spot and time before the
first averaging time just by evaluating the relevant quantity for the
replicating portfolio.
Another advantage from a more conceptual viewpoint is that we require no strong assumptions on the process of the underlying except
the existence of the pricing function, and so we have proven our replication result simultaneously for all processes with deterministic pricing
functions. This means that we do not need a process; if we wish to
examine the effect of an arbitrarily specified pricing function then we
can do so provided the function is arbitrage-free. Note that this means
that we have a natural way of including smile information. From a
computational point of view this means that the model can be implemented simultaneously for many different models simply by specifying
different pricing functions.
In the special case of a discrete barrier option, a simpler trading
strategy can be employed which involves buying a portfolio of vanilla
options at time zero and selling it when the option knockouts or expires. We develop this approach in Section 7. That a discrete barrier
option can be hedged in this manner for the Black-Scholes model and
Mertons jump-diffusion model has already been observed by Andersen, Andreasen and Eliezer, [1]. However, they do not appear to have
developed the method as a pricing tool and their arguments are model
dependent.

M. S. JOSHI

We place our results in historical context. Replication is one of the

main ideas in modern derivatives pricing. We can think in terms of a
hierachy of assumptions. We list some possible assumptions and then
discuss which are needed for each trading strategy. We also make the
general assumptions of no trading costs and heteroskedasticity throughout.
Assumptions on the underlying:
A1 there exists a liquid market in the underlying (or forwards) at
all times
A2 the underlying follows a Markovian process
A3 the underlying follows a continuous process
A4 the underlying follows a diffusive process
A5 the underlying follows a log-normal process
Assumptions on the vanilla options markets:
B1 there exists a liquid market in calls and puts of all strikes and
maturities today
B2 there exists a liquid market in calls and puts of all strikes and
maturities at all times
B3 the prices of calls and puts satisfy call-put symmetry conditions at all times
B4 the price of calls and puts are a known deterministic function
of calendar time, spot, strike and maturity.
We give a classification of replication methods also.
C1 strong static : the option pay-off can be perfectly replicated by
a finite portfolio of calls, puts and the underlying set-up today
with no further trading
C2 mezzo static: the option pay-off can be perfectly replicated by
a finite portfolio of calls, and puts set-up today together with
a finite number of trades in the underlying
C3 weak static: the option pay-off can perfectly replicated by setting up a finite portfolio of calls and puts today which may be
sold before expiry
C4 feeble static: the option pay-off can perfectly replicated by trading a finite number of calls and puts at a finite number of times
C5 dynamic: the option pay-off can be perfectly replicated by continuous trading in the underlying

REPLICATION METHODS

We shall also use the term almost to indicate that the pay-off can be
replicated arbitrarily well with a finite portfolio rather than perfectly.
If the underyling satisfies A1-5 then C5 holds; this is the fundamental
result of Black and Scholes, [3]. This still holds under A1-A4, Dupire
[10].
Under assumption B1 then C1 holds for a straddle with no assumptions on the underlying. We also have that under the assumption B1
that digital European options can be almost strong statically replicated, by approximating using a call-spread. Unfortunately, strong
static replication holds for very few options.
If we make the assumptions A1, A3 as well as B1 then more options
can be mezzo statically replicated. For example, an up-and-in put
option struck at K with up-barrier at K can be replicated. Purchase
a call option struck at K. At the first time that the spot reaches K,
go short a forward contract struck at K (which is of zero cost.) The
forward turns the call into a put and the pay-off is replicated. If the
spot never crosses K then the call option and the original both pay
zero at expiry. See [5, 11].
Under B1, B2, B3, A3 it was shown in [6] that a class of barrier
options can be weak statically replicated. For example, a down-andout call can be replicated by holding a call option with the same expiry
and going short a put option with strike below the barrier and the same
expiry in such a way as to guarantee that the resultant portfolio has
zero value on the boundary.
If we assume B1, B2 and B4 then we are in the situation of this
paper. The method we present for hedging discrete barrier options
under these assumptions is then an almost weak static replication. Note
that we can also hedge continuous barrier options arbitrarily well by
approximating with a discrete-barrier option with an arbitrarily large
number of sampling dates.
Our method of replication for a general path-dependent exotic option
makes the same assumptions, but it is almost feeble static in that it
requires trading in options at multiple times.
Note that whilst these methods make no assumptions on the underlying, it is difficult to imagine a situation where B1, B2 and B4 hold
but A2 does not.

M. S. JOSHI

If we make the additional assumption of continuity of sample paths,

A3, then a simpler method can be used to almost weak replicate continuous barrier options. This method relies on dissolution of the portfolio at the instant the barrier is crossed, and therefore only requires
the replicating portfolio to be of zero value on the barrier rather than
behind it. See [21]. Under these assumptions, it is also possible to
replicate American options, [14]. In fact, one does not really need
continuity of sample paths, the crucial property is that the spot cannot jump across the barrier for knock-out options or into the exercise
domain for American options. These techniques could therefore be applied in markets where only down jumps occur, for example equities,
to the pricing of up-and-out barrier options and American call options.
This research was carried out while working in the Quantitative Research Centre of the Royal Bank of Scotland Group and I am grateful
to my colleagues, Christopher Hunter, Peter Jackel, and Riccardo Rebonato, for many helpful conversations.

2. The uniqueness of prices and measures

In this section, we prove that given a pricing function, F (S, K, t, T ),
for call options which is continuous and a pricing function, G(S, K, t, T ),
for digitals which is continuous for t < T, that the finite-dimensional
distributions of a martingale measure for a given numeraire are uniquely
determined. We work within a market in which spot moves according
to some unknown measure, and derivative prices are given by their
expectation under some unknown martingale measure. An immediate
consequence is therefore that the prices of derivatives depending on
spot at a finite number of dates are determined uniquely. Given the
existence of the pricing function F for calls, the existence of a pricing
function for digitals is weak in that under general conditions such as
the no free lunch with vanishing risk condition, the value of a digital
can be proven to be the derivative of the call prices with respect to
strike, via approximation by call spreads, and this will follow from our
assumption that prices are given by expectations. It is possible to construct a market in which the digital price is not continuous for t < T.
For example, if the stock price has no volatility then digital prices will
jump at S = K, however such an example is highly contrived. We
make the additional assumption that the probability of spot (in either
measure) taking any given value at a fixed time is zero. We also assume

REPLICATION METHODS

that interest rates are known and deterministic; in any case it is difficult to imagine circumstances in which future prices are a deterministic
function of spot without deterministic interest rates.
Our approach is to show that the prices of certain digital options
based on the value of spot at multiple times are determined by noarbitrage conditions. By risk-neutral valuation, it then follows that the
risk-neutral joint probability of spot lying in these intervals is determined. Our approach relies purely on using finite portfolios of options
involving trading at a finite number of times. We therefore do not use
the result of Breeden and Litzenberger, [4], that any single time horizon
derivative can be written as an integral of call options.
We construct a self-financing portfolio involving trading at a finite
number of times which approximates the pay-off arbitrarily accurately.
This shows that there is a unique arbitrage-free price for the option
and fixed the probabilities.
Definition 2.1. Given intervals [aj , bj ] and an increasing sequence of
times tj , j = 1, . . . , n, the associated multi-digital options pays 1 at time
tn if and only if the underlying asset price lies in the interval [aj , bj ] at
times tj for all j.
Let Z(T ) denote the value of a zero-coupon bond expiring at time T.
Let P(E) denote the probability of an event in the martingale measure
with Z(T ) as numeraire. We immediately have that the value of a
multi-digital option, D, is equal to
Z(tn )P(Stj [aj , bj ], j),
for all j. Thus the knowing the price of the option is equivalent to
knowing probability of the event
{Stj [aj , bj ], j}.
In other words, since we have assumed that the probability of spot
taking a single value is zero, the finite-dimensional distributions are
determined by the price of the option.
We now proceed to the proof. First, we prove a lemma on the replicability of a compactly supported single-time horizon derivative.
Lemma 2.1. Let the derivative D pay a continuous function f (S) at
time T provided S [a, b] with < a < b < and zero otherwise.
Given any  > 0 there then exists a finite portfolio, P, of calls and
digital-calls such that the value of P D is less than  for all values of
spot.

M. S. JOSHI

Proof. By taking a digital call struck at a, a digital call struck at b and

a call options struck at a and b we can clearly replicate a derivative, E,
paying zero below a and above b with pay-off increasing linearly from
f (a) to f (b) on the interval [a, b].
By taking D E, we are now reduced to the case where the payoff is continuous everywhere and zero at a and b. As the pay-off is
continuous on a compact interval, it is uniformly continuous. (See for
example [19].) Therefore, given  > 0, there exists > 0, such that if
|x y| then |f (x) f (y)| < .
Take n such n > b a. We take a sequence of points xj such that
x0 = a, xn = b and
|xj+1 xj | .
For example, we could take
ba
xj = a + j
.
n
Define a function, g, by g(x) = 0 for x 6 [a, b], g(xj ) = f (xj ) for all
j, and g is linear on the intervals [xj , xj+1 ] for all j. We then have using
the uniform continuity that |f (x) g(x)| is less than  for all x.
We show that a derivative with pay-off g at time T can be perfectly
replicated by call options and then we are done. To replicate g we
purchase call options struck at xj of notional equal to j j1 where
j is the slope of g on the interval [xj , xj+1 ]. We take 1 = n = 0.
As these call options give precisely the change in notional required to
match the new slope, g is replicated.

With this lemma, we can now prove
Theorem 2.1. Let D be a multi-look digital option associated to intervals [aj , bj ] and times tj then given  > 0 there exists a self-financing
portfolio involving trading a finite number of calls, digital calls and
bonds at the times tj which replicates the pay-off of D within .
Proof. We proceed by induction on the number of times. If n = 1, the
replication is trivial as it involves going long one digital struck at a1
and short one struck at b1 .
Suppose the result is true for n 1. Let D0 be the multi-look digital
associated to the times and intervals for j equal to 2 through n. Given
 > 0, by the inductive hypothesis, there exists a portfolio P involving
trading at the times tj for j > 1 which replicates the pay-off of D0
within /2.

REPLICATION METHODS

The portfolio P is static up to time t2 . Using our assumptions on

the pricing functions for calls and digital calls, its value at time t1 is a
continuous function of spot. If spot is in the interval [a1 , b1 ] at time t1
then the value of D will be equal to the value of D0 and otherwise the
value of D will be zero.
Thus at time t1 , the value of D agrees with the value of P up to /2
within the interval [a1 , b1 ]. By the lemma, we can construct a portfolio,
Q, which replicates the value of P within Z(t1 , tn )/2 on [a1 , b1 ] at time
t1 and has zero value otherwise, where Z(t1 , tn ) is the discount factor
from time t1 to time tn .
Our new portfolio R is therefore equal to Q initially. At time t1 , if
spot is outside [a1 , b1 ] then both D and R have zero value and we have
replicated precisely. Otherwise, we use the pay-off of the options in Q
to purchase the portfolio P, any excess (possibly negative) is used to
purchase zero-coupon bonds expiring at time tn . The magnitude of the
notional of the zero-coupon bonds will be at most /2 by construction.
At time tn , P will replicate the pay-off of D with /2 by hypothesis
so R will replicate D with  and we are done.

Corollary 2.1. Within the assumptions of this section, there is a
unique arbitrage-free price for a multi-look digital option.
Proof. For each n construct a portfolio Pn which replicates D within
1/n. Let Pn (0) denote the initial value of Pn . The value of Pn (0)Pm (0)
must be less than n1 + m1 and therefore tends to zero as m, n tend to
infinity.
It follows that Pn (0) is a Cauchy sequence and therefore converges.
(See [19].) The limit must be the value of D as D is replicated within
1/n by Pn : let P (0) denote the limit; we then have
(2.1)

|D(0) P (0)| |D(0) Pn (0)| + |Pn (0) P (0)| ,

the first term is less than 1/n and the second term tends to zero as
n .

Corollary 2.2. If an exotic option, O, pays H(St1 , . . . , Stn ) at time tn
then its price is uniquely determined by F and G.
Proof. The value of O is equal to
Z(T )E(H(St1 , . . . , Stn )),
where the expectation is taken with respect to the numeraire Z(T ). The
expectation only involves the probabilities for the finite-dimensional

10

M. S. JOSHI

distributions over (St1 , . . . , Stn ) which have been determined by noarbitrage.


An easy corollary is that the prices of options with Bermudan features are also determined.
Corollary 2.3. Suppose the option, D, carries the right to receive
Hl (St1 , . . . , Stl ) at time tl for l = 1, . . . , n, provided the right to receive
has not been exercised at a previous time then its price is uniquely
determined by F and G.
Proof. We prove the result by induction. If n = 1 then the option is
equivalent to a single time horizon option and the result is immediate.
If the result is true for n 1 then at time 1 we have the right to
receive H1 (S1 ) or hold an early exercisable option, D0 , associated to
n 1 times. The value of D0 is uniquely determined at time t1 for
any value of spot by our inductive hypothesis. The value of D at time
zero is therefore equal to the value of an derivative, E, which pays
the maximum of the value of D0 and H1 (S1 ) at time 1. The value of
E is uniquely determined by our previous results and the corollary
follows.

3. The method for Asian options
In Section 2, we proved some theoretical results about prices of exotic
options, in this section, we show how this results can be turned into
practical tools for option pricing. We observed that the value of an
Asian option at a time ti depends only on the current value of spot
and the value of the auxiliary variable Ai . Our approach is therefore
to construct approximating portfolios of options with the composition
of the portfolio depending upon spot and Ai . We do this by backwards
iteration.
At the final time, T = tn , the value of the option is (An K)+ . We
can rewrite this as
(3.1)


1
1
n1
An1 + Stn K
= (Stn (nK (n 1)An1 ))+ .
n
n
n
+
This means that at time tn1 a call option struck (nK (n 1)An1 )
of notional 1/n precisely replicates the final pay-off. Note that the
replicating portfolio here depends upon An1 but not Stn1 .

REPLICATION METHODS

11

We have assumed the existence of a pricing function so we immediately have that the value of this replicating call option is determined
as a function of Stn1 for each value of An1 . By no-arbitrage the value
of the Asian call option must be equal to the value of this replicating
option and we therefore know its value as a function of Stn1 and An1 .
At time tn2 , we now wish to construct portfolios of options expiring
at time tn1 whose payoffs precisely replicate the value of the Asian
call at time tn1 .
Recall that the value of An1 is equal to
n2
1
An2 +
St .
n1
n 1 n1
This means that the set of points in the (Stn1 , An1 ) plane reachable
from a point (Stn2 , An2 ) is a line, and which line depends purely
upon the value of An2 and not Stn2 . In particular, the line of points
reachable is


n2
1
Stn1 ,
An2 +
St
.
n1
n 1 n1
Thus given a value of An2 , the value of Stn1 determines a point in
the (Stn1 , An1 ) plane and a price for the Asian option at time tn1 .
Call this price fAn2 (Stn1 ). This means that for each value of An2 ,
we can replicate the value of the Asian call at time tn1 by a European
option paying fAn2 (Stn1 ) at time tn1 .
This European option can be approximated arbitrarily well by using
a portfolio of vanilla call and put options. The given pricing function
can then be used to assess the value of this portfolio for any value of
Stn2 , at time tn2 . Thus by valuing the portfolio associated to each
value of An2 for every value of Stn2 , we develop the price of the Asian
call option as a function of Stn2 and An2 at time tn2 .
We can repeat this method to get the value of the call option across
each plane (Stj , Aj ) for j = 1, . . . , n 1.
At time t1 , we have, of course, that A1 = S1 and only the values
along that line are relevant. Thus in setting up the initial replicating
portfolio at time zero, we replicate the values along this line in the
(S1 , A1 ) at time t1 .
The value of this initial replicating portfolio is now the value of the
Asian call option at time zero. Note that the initial value of spot was
not used in the construction so we can get the value of the Asian call
as a function of spot immediately just by revaluing the initial portfolio.

12

M. S. JOSHI

This also means that the delta and gamma are equal to the delta and
gamma of the initial portfolio. As the replicating portfolio does not
change up to the first reset time, we can also read off the theta as
the theta of the replicating portfolio. Note that this argument does
not extend to Greeks with respect to the other model parameters as
changing the other parameters will in general affect the composition of
the initial portfolio.
Having used a replicating argument to price the derivative, what is
the actual trading strategy? We set up the initial replicating portfolio
and hold it until the first reset time. At the first reset time, we dissolve
the portfolio by exercising all the options which are in the money,
as all the options are at their expiry. The sum of money received is
by construction precisely the cost of setting up a new portfolio which
depends upon the value of A1 and which replicates out to the second
reset time. We then exercise again and use the money to buy a new
portfolio out to the third reset time and so on. At all stages, the new
portfolio set up will depend on the value of the auxiliary variable, and
will be equal in value to exercised value of the previous portfolio.
The existence of a deterministic pricing function is crucial as any
indeterminacy in the set-up cost of the later replicating portfolios will
destroy our argument.
4. The implementation
The argument in the previous section implicitly assumed a perfect
replication of the value of the Asian call option across all values of spot
and the auxiliary variable. Clearly, if we are to implement the pricing
method in a computer we need to use a discrete approximation. The
approach tested was to use the same square two-dimensional grid at
each time step. The grid was taken to have a uniform size of squares
in log-space.
In the implementation of the method it was sometimes necessary to
assess the value of a replicating portfolio at a point for which the auxiliary value was not on the grid. This was done by linearly interpolating
the prices between the neighbouring auxiliary values.
Prices obtained were compared against quasi-Monte Carlo prices for
Asian call options in Black-Scholes and jump-diffusion models and
found to agree to high levels of accuracy. We give computed prices
for a five-year Asian call option with yearly resets. Spot and strike
were taken to be 100 and interest rates were taken to be zero.

REPLICATION METHODS

13

We present numbers for the Black-Scholes model with volatility 10

percent. We present simulation results for various different step sizes.
With step-size log(1.005), we get
Steps
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80

Price
5.100
5.520
5.535
5.715
5.721
5.839
5.838
5.890
5.887
5.909
5.906
5.914
5.912
5.914
5.913

Delta
0.2815
0.3368
0.3735
0.4111
0.4391
0.4642
0.4815
0.4957
0.5053
0.5127
0.5174
0.5208
0.5229
0.5243
0.5251

Gamma
0.01662
0.01896
0.02041
0.02232
0.02320
0.02462
0.02506
0.02590
0.02606
0.02650
0.02653
0.02673
0.02673
0.02681
0.02680

Price
5.267
5.733
5.771
5.894
5.896
5.916
5.915
5.917
5.917
5.917
5.917
5.917
5.917
5.917
5.917

Delta
0.3570
0.4310
0.4743
0.5026
0.5156
0.5223
0.5249
0.5259
0.5262
0.5263
0.5264
0.5264
0.5264
0.5264
0.5264

Gamma
0.01897
0.02294
0.02428
0.02603
0.02627
0.02672
0.02673
0.02680
0.02680
0.02680
0.02680
0.02680
0.02680
0.02680
0.02680

With step-size log(1.01),

Steps
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80

14

M. S. JOSHI

With step-size log(1.02), we obtain

Steps
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80

Price
3.101
3.814
4.155
4.651
4.927
5.260
5.429
5.617
5.706
5.798
5.839
5.880
5.897
5.913
5.920

Delta Gamma
0.172 0.006
0.253 0.007
0.297 0.008
0.350 0.010
0.381 0.011
0.422 0.014
0.445 0.016
0.472 0.019
0.486 0.021
0.501 0.023
0.508 0.024
0.516 0.025
0.519 0.025
0.522 0.026
0.524 0.026

We compare with a Quasi-monte-carlo method in which Greeks are

computed by finite differencing.
Steps
2097152
1048576
524288
262144
131072
65536
32768
16384
8192
4096

Price
5.910
5.910
5.910
5.910
5.910
5.909
5.908
5.910
5.910
5.899

Delta Gamma
0.5263 0.0166
0.5264 0.0184
0.5263 0.0141
0.5262 0.0091
0.5262 0.0060
0.5262 0.0000
0.5265 0.0000
0.5277 0.0000
0.5272 0.0000
0.5270 0.0000

The Gamma here has clearly not converged even after two million
paths.

5. Other options
So far we have concentrated on the study of the discrete Asian call
option, however the method can be used to price other path-dependent

REPLICATION METHODS

15

exotic options. We also remark that the addition of Bermudan-style

features causes no new difficulties.
What properties are necessary for implementing this method? The
pay-off should be dependent on the value of spot on a finite set of times
t1 , . . . , tn . There should exist a sequence of functions Aj such Aj is a
function of (St1 , . . . , Stj ), and a second sequence fj of updating function
such that
(5.1)

Aj+1 = fj+1 (Stj+1 , Aj ).

The crucial property is that the final pay-off should depend only on An
and Stn .
The geometric-average Asian option is easily fit into the framework.
It pays on the geometric average of the spot at the sampling dates
instead of the arithmetic average. We take
! 1j
j
Y
.
(5.2)
Aj =
Sti
i=1

The updating function is then

(5.3)

fj+1 (Stj+1 , Aj ) = Ajj Stj+1

1

j+1

A look-back option pays the positive part of the maximum value of

Stj , j = 1, . . . , n, minus the strike. In this case, Aj is the maximum of
Sti for i j. The updating function fj is just the maximum of Aj1
and Stj .
A forward-starting option pays the sum
(St2 St1 )+
for some specified . In this case,
A1 = A2 = St1
and the pay-off is
(St2 At2 )+ .
This option is sometimes called a cliquet. A related option is the ratio
cliquet which pays


St2
1 .
St1
+
In this case, we take A1 = A2 = St1 , and the pay-off is trivially purelydependent on A2 and St2 .

16

M. S. JOSHI

We can also price a growth certificate which pays a multiple of the

stock price at time t3 if the price at time t2 is greater than the price
at time t1 , and zero otherwise. Here we take A1 = St1 . We set A2
equal to one or zero according to whether St2 is greater than A1 . We
set A3 = A2 . The pay-off is then St3 A3 . Note that there are only two
possible auxiliary values for this option which massively speeds up implementations.
Discrete barrier options can similarly be fit into this framework. Suppose the option pays f (Stn ) at time tn , provided the spot lies in a given
interval Ij at each of the times tj . Let gj (x) equal one if x Ij and zero
otherwise. We let A1 equal g1 (St1 ) and Aj equal gj (Stj )Aj1 , for each j
greater than 1. The final pay-off is then f (Stn )An . As with the growth
certificate, we require only two auxiliary values for this implementation.
In conclusion, the method presented here allows the rapid pricing
of many path-dependent options using alternative models which allow
the incorporation of smile information.

6. Additional Simplifications
If our pricing function comes from a model that admits certain homogeneity properties, and the option is also homogeneous then we can
greatly reduce the number of computations necessary. In particular,
suppose our pay-off is that of an Asian call option. The class of payoffs of call options is homogeneous in that we have
(6.1)

(x K)+ = (x K/)+ .

This means to price for all values of K and , we only need to price
for one value of and all values of K. We can adapt this to speed up
the pricing of Asian options for certain pricing functions.
Our assumption on the pricing function is that it is the discounted
expectation of a Markovian process in which for some a, the distribution
of log(St + a) log(Ss + a) for t > s is independent of the value
of Ss . With a = 0, this assumption is satisfied by the Black-Scholes
model with determinsistic time-dependent volatility, Mertons jumpdiffusion model and the Variance Gamma model. For a non-zero, we
can obtained the displaced diffusion model.

REPLICATION METHODS

is

17

Given this assumption, at time tj we have that the value of the option

!
!
n
X
1
St + iAi K P (tj , T )
E
n j=i+1 j
+

where Ai is as in Section 3 and P (tj , T ) is the value of a zero coupon

bond expiring at time T at time tj . Our assumption implies that
(6.2)

(Stj + a) = (Sti + a)Xj ,

where {Xj } is a collection of random variables (which will not be independent.) We can therefore rewrite the value of the Asian option
as

!
!
n
1 X
(6.3) E
(St + a)Xi a + iAi K P (tj , T )
n j=i+1 i
+

!
n
X
iAi nK (n i)a
1
Xi +
P (tj , T ).
= (Sti + a)E
n j=i+1
n(Sti + a)
+

The interesting thing about the right-hand-side of (6.3) is that putting

(6.4)

(Ai , Si ) =

iAi nK (n i)a
,
n(Sti + a)

we see that the dependence of the expectation upon Ai and Sti is now
one-dimensional.
Here we have adapted a method of Rogers and Shi, [20], developed for
pricing arithmetic continuous average options in a Black-Scholes world
using PDE methods. This has previously been applied to PDE pricing
in a Black-Scholes setting for discrete arithmetic average options by
Benhamou and Duguet, [8]. Note that the Black-Scholes case would
have a = 0 and the variables Xj all being log-normal.
7. Replicating Discrete Barriers
Suppose we wish to price a discrete-barrier knock-out option. The
method we have already presented works, however it requires buying
and selling options at each knock-out date. In this section, we present
a method that requires buying and selling only at the initial time and
at the time of knock-out or expiry. Recall that a discrete barrier option
pays off at some time T, unless the price of the underlying is outside a
specified range on a predetermined finite set of dates, t1 < t2 < <

18

M. S. JOSHI

tn < T, in which case the option pays zero (or possibly a fixed rebate
at the time of knockout.)
The technique we present here is related to that in [1] and will produce the same replicating portfolio. The essential difference is that our
approach is algorithmic and relies on the existence of a deterministic
pricing function for options rather on a specific process. A consequence
of this is that our approach can be used for pricing whereas the arguments presented in [1] relied on the price of the knock-out option
already being known by different means.
The essential idea is that we construct a portfolio of plain vanilla
options which has the same value as the option being modelled at payoff time, and has zero value at the points in spot-time where the original
option knocks-out. So the option knocks-out if the spot passes below
the value B at any of the times t1 through tn . Our portfolio should
therefore have zero value on the set
[0, B] {t1 } [0, B] {t2 } [0, B] {tn }.
The idea here is that if the option knocks-out then the replicating
portfolio would be immediately liquidated at zero cost. Unlike our
construction for a general path-dependent exotic, the strategy here
involves selling options before their expiry which is the crucial point
where we use the existence of a deterministic pricing function. In what
follows, we concentrate on the case of a down-and-out option for concreteness. However, the same techniques apply with little change to
pricing a double-barrier option or an up-and-out option.
To construct this portfolio, we induct backwards. First, we choose a
portfolio of vanilla options with expiry T which approximates the final
pay-off as accurately as we desire. Of course, if we were modelling a
knock-out call or put, this would just be the call or put without the
knock-out condition. Call this initial portfolio, P0 . For a portfolio P,
we denote its value at the point (S, t) by P (S, t).
As we know the price of any unexpired vanilla option for any value of
spot and time. We can value P0 along the last barrier [0, B] {tn }. We
can kill the value of P0 at the point (B, tn ) by shorting a digital option
with value equal to P0 (B, tn ), below B and zero above. In order to
retain the property that the portfolio consist solely of vanilla options,
we approximate the digital by a tight put-spread. Call our new portfolio
P10 . This portfolio then has correct final pay-off profile and has zero
value at (B, tn ) but may have non-zero value along [0, B) {tn }. We
partition [0, B) into [0, x1 ], [x1 , x2 ], . . . [xk1 , B). We then approximate

REPLICATION METHODS

19

the value of P10 along [0, B) {tn } by assuming it is affine on each of

these subintervals. We now remove this value by moving successively
inward. The portfolio P10 has zero value at (B, tn ) so we if short put
options struck at B with expiry at tn and notional P11 (xk1 , tn ), we obtain
a portfolio P11 which has zero value at both (B, tn ) and (xn1 , tn ). If
our partition is suitably small, the linear approximation will be close
in value to the original and the value will be small on the interval
[xn1 , B]{tn }. We now iterate along the barrier at each stage shorting
put options struck at xnj with expiry tn and notional P (xnj1 , tn )
to obtain a sequence of portfolios, P1j . Note that the put options only
affect value for spot below their strike so will not affect the value of the
portfolio at the points already fixed. The portfolio P1k1 will then have
close to zero value along barrier as desired. Let P1 be the portfolio
P1k1 .
We can repeat this procedure along the barrier [0, B] {tn1 } using
the portfolio P1 instead of P0 to obtain a portfolio P2 . Repeating, we
obtain a sequence of portfolios, Pl . If we regard an option as having
zero value after expiry then the portfolio Pn then has the property
of having close to zero value along all the barriers [0, B] {tj } and
approximates the final payoff profile. As we immediately liquidate the
replicating portfolio when the original options knocks out, options will
only expire in the money at a time of liquidation. This means that we
do not need to keep track of any pay-offs from options maturing. The
value of Pn at time zero and todays spot will therefore be the value of
the knock-out option.
A similar procedure would be effective for up barriers, simply replacing puts by calls and inducting upwards instead of downwards. To do
double barriers, we simply do each independently at each barrier time
as there is no interaction between the two pieces at a given knock-out
time, although the portfolio is of course affected by both barriers at
previous times. We note that our procedure does not require the barrier
level to be constant.
One disadvantage of this approach over that presented in Section 3 is
that for each option added the entire portfolio of already added options
has to be valued for a new value of spot and time. This means that
the time required to evaluate using replication will be proportional to
the square of the total number of options. The total number of options
will be the number of steps per barrier times the number of barriers.
On the other hand, the fact that there is only one rehedging makes the
approach conceptually nicer.

20

M. S. JOSHI

We present the results of some simulations for a jump-diffusion model.

We first present the pricing of a down-and-out call option with the
following contract,
Barrier Level 90
Strike
100
Notional
1
Expiry
1
and with knock-out dates 0.1, 0.2, 0.3, . . . , 0.9. We use a log-normal
model for stock evolution with log-normal jumps as follows,
Interest Rate
0.05
Dividend Rate
0
Initial Spot
100
Diffusive Volatility 0.1
Jump Intensity
0.2
Jump Mean
0.8
Jump Sigma
0.2
The prices produced by the replication method given above were,
Steps per barrier
8
16
32
64
128
256

Price
8.685
8.686
8.687
8.687
8.687
8.687

These compare with prices produced by Monte Carlo as follows,

Number of Paths
1000000
500000
250000
125000
62500
31250
15625

Price
8.687
8.689
8.691
8.687
8.693
8.691
8.649

It is also interesting to compare the implied price with that obtained

from pricing with a Black-Scholes model. Pricing the same option using

REPLICATION METHODS

21

Monte Carlo but in a Black-Scholes world with the at-the-money implied volatility, the price was 8.42 and using the at-the-barrier implied
volatility was 9.25.
We now present prices with the more extreme parameters,
Diffusive Volatility
Jump Intensity
Jump Mean
Jump Sigma

0.2
0.5
0.6
0.3

For the replication method, we get

Steps per barrier
8
16
32
64
128
256

Price
17.377
17.389
17.393
17.391
17.391
17.391

and for Monte Carlo,

Paths
1000000
500000
250000
125000
62500
31250
15625

Price
17.394
17.400
17.406
17.392
17.397
17.436
17.374

Pricing the same option using Monte Carlo but in a Black-Scholes world
with the at-the-money implied volatility, the price was 13.89 and using
the at-the-barrier implied volatility was 14.47.
References
[1] L. Andersen, J. Andreasen, D. Eliezer, Static Replication of Barrier Options:
Some General Results, preprint 2000
[2] Y.Z. Bergman, Pricing Path Contingent Claims, Research in Finance, 5, 229241
[3] F. Black, M. Scholes, The Pricing of Options and Corporate Liabilities, Journal
of Political Economy, 81, 637-654
[4] D. Breeden, R. Litzenberger, Prices of state-contingent claims implicit in option prices, J. Business 51, 621-651.

22

M. S. JOSHI

[5] H. Brown, D. Hobson, L.C.G. Rogers, Robust Hedging of Barrier Options,

preprint 2000
[6] P. Carr, K. Ellis, V.Gupta, Static Hedging of Exotic Options, Journal of Finance, LIII, 1165-1191, 1998
[7] J.N. Dewynne, P. Wilmott, Partial to the Exotic, 6 (3) 38-46
[8] E. Benhamou, A. Duguet, Small Dimension PDE for Discrete Asian Options,
preprint 2000
[9] E. Dinenis, D. Flamouris, J. Hatgioannides, Implied Valuation of Asian Options, preprint 2000
[10] B. Dupire, Pricing with a smile, Risk, 1994
[11] S. Hodges, A. Neuberger, Rational Bounds for Exotic Options, preprint
[12] J. Hull, A. White, The pricing of options on assets with stochastic volatilities,
The Journal of Finance, Vol XLII no 2, 1987, 281-300
[13] J.E. Ingersoll, Theory of Financial Decison Making, Rowman and Littlefield
1987
[14] P. J
ackel, R. Rebonato, An efficient and general method to value Americanstyle equity and exotic options in the presence of user-defined smiles and timedependent volatility, preprint 2000
[15] D. Madan, P. Carr, E. Chang, The Variance Gamma Process and Option
Pricing, European Finance Review Vol. 2, No. 1 1998
[16] D. Madan, F. Milne, Option Pricing with V.G. Martingale Components, Mathematical Finance, 1, 4, 39-55
[17] D. Madan, E. Seneta, The Variance Gamma Model for Share Market Returns,
Journal of Business, 63, 4, 511-524
[18] R. Merton, Option pricing when underlying stock returns are discontinuous,
Journal of Financial Economics 3, 1973, 141-83
[19] W. Rudin, Principles of Mathematical Analysis, McGraw-Hill Publishing Company
[20] L. Rogers, Z. Shi, The Value of an Asian Option, J. Appl. Probability 32,
1995, 1077-1088
[21] R. Rebonato, Volatility and Correlation in the Pricing of Equity, FX and
Interest-rate Options, Wiley Chichester New York Weinheim Brisbane Singapore Toronto 1999
[22] P. Wilmott, Derivatives, The Theory and Practice of Financial Engineering,
Wiley 1998
Royal Bank of Scotland Group Risk, 135 Bishopsgate, London EC2M
3UR