Basic Monte Carlo Techniques

Earlier we derived the Black-Scholes problem to price a European option

2
S
2

2
V
S
2
+r
V
S
rV = 0,
V (S, T) = P (S) .
The BSE is a linear parabolic PDE and as such the solution can be expressed
as an integral of the form
V (S, t) = e
r(Tt)
0

p (S, t; S

, T) V (S

, T) dS

where p (S, t; S

, T) represents the transition density and is the solution of the

2
S
2

2
p
S
2
+rS
p
S
= 0,
p (S, t; S

, T) = (S

S) .
We have discussed that the function p (S, t; S

, T) can be considered as one of

2
S
2

2
U
S
2
+rS
U
S
= 0,
U (S, T) = f (S) .
Secondly, and more importantly for this section in probabilistic terms it is the
probability density function for the risk-neutral random walk mentioned earlier.
This is also called the risk-neutral measure.
We can write the value of the option in the form
V (S, t) = e
r(Tt)
E
Q
[P (S)]
which is the present value of the expected payo wrt the risk-neutral probability
density Q and recall
E
Q
[P (S)] =

0
p (S, t; S

, T) P (S

) dS

.
1
More often than not we must solve option-pricing problems by numerical
means. It is highly unlikely that analytical solutions can be obtained to a
pricing problem unless it is simple and ideal.
The most useful numerical techniques are The Monte Carlo Scheme and
nite-dierence methods.
The former is a simulation based approach hence uses probability as the
underlying method.
2
Relationship between derivative values and simulations
Simulations are at the very heart of nance and used to forecast future
scenarios.
More importantly they can also be used to price options.
The fair value of an option is the present value of the expected payo at
expiry under the risk-neutral random walk for the underlying. (Phelim Boyle
1977)
Recall an amount of cash M = M (t) in the bank grows according to
dM
dt
= r (t) M
where r (t) is the variable (risk-free) interest rate. This dierential equation
has solution
M (t) = M (T) exp
_

T
t
r () d
_
i.e. the present value (time t) of a future cash ow (time T). The exponential
term is the discount factor. In the simple case of a xed rate of interest it
becomes e
(r(Tt))
.
This gives the fair price of an option V to be
V = E
Q
_
e

T
t
r()d
Payo(S)
_
where
S =asset price, r = stochastic domestic interest rate, T = expiry, t =
current time, Q= risk neutral density.
3
Scheme: The Monte Carlo method consists of the following steps
1. Simulate sample paths/realizations for the underlying asset price (e.g.
equities or interest rates) over the relevant time horizon, according to the
risk-neutral measure. Here we use the discretized SDE
S
i+1
= S
i
_
1 +rt +

t
_
.
2. Evaluate the discounted cashows (using domestic rate of interest) of a
derivative on each sample path, as determined by the structure of the
security being priced.
3. Average the discounted cashows over sample paths.
So the option price becomes
e
(r(Tt))
.
1
N
N

n=1
Payo(S)
The payo for a European Call Option with strike E is C = max (S (T) E, 0) , where
S (T) is obtained from (5) in discrete form, for each value 1 n N.
Although we are not concerned with the path followed by the process S (t) in
getting to S (T) we will nevertheless simulate this as we can price other options
which are path dependent.
Based upon the N realizations an estimate for the price of an option becomes
C (S, t)
C (S, t) =
1
N
N

n=1
C
(n)
(S, T)
which is equivalent to
e
(r(Tt))
.
1
N
N

n=1
max
_
S
(n)
(T) E, 0
_
.
4
Variance Reduction
For the simulation of an asset price, samples are drawn from a probability
(normal) distribution.
If these samples are generated in a fashion, which is not entirely random,
but in a manner that reduces the uctuations (i.e. volatility) of the resulting
samples, computational time can be reduced considerably to obtain the desired
degree of accuracy. A similar eect can be obtained by performing suitable
transformations on the function, which forms the basis of the simulation, so
that dependency upon the uctuations arising in the samples is reduced. The
disadvantage is that correlations are introduced. It then becomes a choice,
whether to compromise computational time over the risk of correlations being
introduced. Recall that the standard error associated with the Monte Carlo
method is
=

N
.
Increasing the number of sample paths generated,by increasing N, leads to a
reduction in . In addition we are able to manipulate the variance, i.e. reduce
the value of . For this reason a whole area of Monte Carlo, namely variance
reduction techniques has been developed.
5
Antithetic Variable Technique
A very simple technique is by use of antithetic variates, which can reduce
computational time and be implemented at no additional eort, was introduced
to option pricing by Boyle (1977).
The method, which was initially used in the pricing of a European call option
on a dividend paying stock, is outlined below. It is based upon the observation
that if
(n)
N (0, 1) , then
(n)
also has a standard Normal distribution.
In this technique, by using the one set of random numbers generated, two
estimates for an option are calculated. If
i
is used to obtain C, then
i
gives

S (t) and hence a second approximation for the option price

C where

C = e
(r
d
(Tt))
.
1
N
N
n=1
max
_

S
(n)
(T) E, 0
_
6
The estimate for the option C

is now the average of the two values, C &

C, so
C

=
C +

C
2
The technique converges because of the symmetry of the Normal Distribu-
tion.
Justication for obtaining C

is based upon the distribution of the antithetic

(n)
,
(n)
__
are distributed more regularly than a collection of
2n independent samples with the sample mean over the antithetic pairs always
equal to the population mean of 0. The data set has a lower variance.
7
Transformation Methods
Earlier we discussed how to create random variables following N (0, 1) in
excel using NORMSINV(RAND()).
This represents the inverse cumulative density function with a uniformly
distributed RV Unif [0, 1] as the parameter.
Suppose y is a RV which is Unif [0, 1] and is to be converted into another
variable x with pdf p (x) . If its CDF F (x) is dened as
F (x) =
x

p (s) ds
then we wish to invert this to obtain
x = F
1
(y) : y Unif [0, 1]
in a computationally ecient manner, assuming of course that F
1
exists. Re-
call that the NORMSINV() function is an accurate but slow mode of numer-
ically inverting the integral.
Consider the following simple case of generating x from a unit Cauchy dis-
tribution p (x) =

1
1 +x
2
.
Firstly write
F (x) =
1

1
1 +s
2
ds
=
1

_
arctan x +

2
_
upon rearranging we have
F (x)
1
2
=
1

arctan x
x = tan
_

_
F (x)
1
2
__
i.e.
x = F
1
(y)
= tan
_

_
y
1
2
__
.
8
We need a numerical technique for the conversion. Most programming lan-
guages generate uniformly distributed random variables over 0 and 1. How can
these be transformed to standard normals N (0, 1)?
The Box M uller Method
The CDF for the standardized normal distribution is dened as
N (x) =
1

2
x

e
s
2
/2
ds.
In this technique the inecient inversion process of N (x) is not required as
the random number x N (0, 1) can be directly captured using the following
algorithm:
1. generate two independent uniformly distributed random variables y
1
, y
2

Unif [0, 1]
2. compute x
1
, x
2
N (0, 1) by
x
1
=
_
2 log y
1
cos (2y
2
)
x
2
=
_
2 log y
1
sin (2y
2
) .
9
The disadvantage with this method lies in the computation of the trigono-
metric and transcendental functions sin, cos and log . This leads on to a more
ecient scheme which employs an acceptance-rejection method.
10