Introduction

We first consider some basic models for returns on a single investment.

There are the so-called deterministic models in which the returns are as-
sumed to be governed by known functions. There are also stochastic models
in which the returns are governed by a random process of known type. Some
of the deterministic returns we consider can be calculated by very elemen-
tary means, others require the knowledge of ordinary differential equations
and/or the numerical solution of ordinary differential equations. We will
not treat these topics here.
Next, we consider the simulation of the most basic stochastic models. This
is the beginning of the study of Monte Carlo Simulation. We relate the basic
problem to one numerical integration and briefly describe a deterministic
approach to the problem that can speed up calculations a great deal.
Finally, we briefly indicate the advantages and disadvantages of the Monte
Carlo approach.

Deterministic Interest Rate Modeling

The simplest investments are those whose return is a fixed number—the
interest rate. We consider here mathematical models for this most basic
class of investments. The first two sections below involve constant rates and
should be familiar. The third is a natural extension to a class of non-constant
rates. While the concept presented in the third section is important, the
application to real-world interest rates is somewhat non-realistic.

Simple Interest. The formula

I = P rt
expresses the interest earned I on a principal amount of P invested over the
time period t if the annual rate of interest is r. If the interest is compounded
annually, then at then end of t years the original principal will have grown
to the value of P + I = P (1 + rt). In particular, after one year the principal
is P (1 + r)

Continuous Compounding. If the interest is compounded over N time

intervals all of duration ∆t, then the principal grows to P (1 + r∆t)N . If
we subdivide one year into N time intervals all of length ∆t = N1 , then
the principal grows to P (1 + Nr )N . If we let the compounding interval get
arbitrarily small and are good at math, we will discover the number e:
r
lim P (1 + )N = P er .
N →∞ N
Therefore, under the assumption of continuous compounding at a constant
known rate, we can say that the principal at time t > 0 is given by the
formula
P (t) = P (0)ert .
1
Monte Carlo Simulation c 2009-11, Stephen Demko

One can check that P (t) satisfies the ordinary differential equation
dP
= rP
dt
or
dP = rP dt.
This last formulation says that in a tiny time interval of length dt the
change in the principal is equal to the amount of principal times the annual
rate times the length of the time interval. As the length of the interval gets
smaller and smaller this assertion becomes more and more accurate.
Continuous, Deterministic Interest Rates. There are several ways to
extend the idea of continuously compounded interest with a constant rate.
The first way is to allow non-constant rates. A very general deterministic
mathematical model that would allow this would be
dP = f (P, t)dt.
or
dP
= f (P, t)
dt
where f is allowed to be a function of P and t. More simply one would have
f (P (t), t) = r(t)P (t) where r(t) is some known, reasonably nice1 function.
To describe the solution completely, one needs to add an initial condition,
say P (t0 ) = c, where t0 is some initial time and c is a constant.
This model is tractable in the sense that depending on the nature of f one
can either invoke results from the theory of ordinary differential equations
to state a solution formula or use numerical methods to find very good
approximate solutions. For example, if we require that f (P, t) = r(t)P (t),
then we have to solve the differential equation
dP
= r(t)dt
P (t)
which, after integrating both sides and taking into account the constant that
arises, has the solution
Rt
P (t) = P (0)e 0 r(s)ds
With this sort of model of interest rates one can explore scenarios based
on different assumptions about how rates will change. For example, there
have been times recently when the short term direction of the Federal Re-
serve’s discount rate has been easy to predict–based on announcements by
the Chairman. Currently, one might make some guesses about when rate
will start to move up and with what speed and then explore the effects of
this on a portfolio. However, interest rates are notoriously difficult to pre-
dict and are subject to shocks—we have seen these in the last year in the
sudden lowering of interest rates by the central banks.
1For example, we could allow it to have a finite number of discontinuities.
2
Monte Carlo Simulation c 2009-11, Stephen Demko

Change in Principal Expressed as an Integral. This little excursion is op-

tional; it is meant to be a reminder of basic calculus as well as a preparation
for a major concept in stochastic calculus.
Let’s break the long time interval [0, T ] into small pieces all of length ∆t.
Say 0 = t0 < t1 < · · · < tN = T where ti − ti−1 = ∆t. Then, following the
model dP = f (P (t), t)dt with continuous P , for each small interval [ti−1 , ti ]
there is a point2 ξi ∈ [ti−1 , ti ] such that the change in P over this interval
satisfies
P (ti ) − P (ti−1 ) = f (P (ξi ), ξi )∆t
If we add up all of these changes, we see that
N
X
P (T ) − P (0) = f (P (ξi ), ξi )∆t
i=1
Since we never know the special points ξi , a much more realistic approach
is to select an arbitrary point ζi ∈ [ti−1 , ti ] and replace the equality by
approximate equality:
N
X
P (T ) − P (0) ≈ f (P (ζi ), ζi )∆t
i=1
Now, as ∆t → 0, the arbitrariness of the ζi becomes severely constrained,
and we can say (or hope?)
N
X
P (T ) − P (0) = lim f (P (ζi ), ζi )∆t
∆t→0
i=1
One recognizes the limit above as the familiar Riemann Integral
XN Z T
lim f (P (ζi ), ζi )∆t = f (P (t), t)dt
∆t→0 0
i=1

The case of random interest rates. There are a number models for interest
rates that allow for randomness. We will mention some very soon. One
major complication
Rt
that arises beyond the model itself is how to interpret
r(s)ds
something like e 0 when r is random.

2We only know that such a point exists; we do not have a method of finding it.
3
Monte Carlo Simulation c 2009-11, Stephen Demko

A Model for Random Returns

It is clear that daily returns on stocks cannot be modeled with the ideas
discussed above. In this section we describe first the most basic model of
a random process—Brownian Motion—and then a way to simulate a single
path, or trajectory, of Brownian Motion. Finally we consider the case that
the returns of an investment behave like Brownian Motion. This leads to
the idea of Geometric Brownian Motion. In the section that follows this, we
will discuss the numerical simulation of Geometric Brownian Motion.
A General Stochastic Model. We extend the scope of the model consid-
ered above to include any asset that is subject to random fluctuations. Or
more broadly to any quantity whose changes can be thought of as being in-
fluenced by both deterministic forces and random forces. Examples of such
quantities include equities, interest rates, and volatility.
Let S = S(t) denote the quantity being modeled. It is convenient to think
of S as being the price of a stock at time t. A general model of changes in
S could be written as
dS = f (S, t)dt + g(S, t)dR
where g(S, t) is some deterministic function and dR captures the randomness
of change in the small time interval dt.
The Wiener Process. One standard model of randomness is based on the
mathematical concept of Brownian Motion or Wiener Process which is a
continuous time stochastic process. This has the consequence that the ana-
logue of dS for Brownian motion is not a number, but a random variable.
We will denote this infinitesimal increment of Brownian motion by dW .
Brownian motion is characterized by the four properties:
(1) W (0) = 0. This is a convenient normalization
(2) If 0 < s < u < t, then the increments W (t)−W (u) and W (u)−W (s)
are independent random variables
(3) If s < t, then W (t) − W (s) is Normally distributed with mean 0 and
variance t − s. Random variables having a normal distribution are
also called Gaussian.
(4) W (t) is continuous (almost surely).

4
Monte Carlo Simulation c 2009-11, Stephen Demko

Simulation of Brownian Motion. In certain cases one can derive formulas

for the solution of dS = f (S, t)dt + g(S, t)dW . The formulas will necessar-
ily involve random variables, so that one usually makes only probabilistic
statements about the solution. However, from the definition of Brownian
motion one can see how to simulate a realization of Brownian motion:
(1) INITIALIZE: Select a positive number step to be “step size” and set
x0 = 0 and y0 = 0.
(2) For k = 1, 2, . . .
• select a random z from a Gaussian distribution (that is, a Nor-
mal distribution) with mean 0 and variance 1. The computer
command to do this varies with the language and/or the par-
ticular code library being used. √
• define xi = xi−1 + step and yi = yi−1 + z step
• draw a straight line between (xi−1 , yi−1 ) and (xi , yi ).
For example, to make a Brownian motion simulation over an interval
[0, T ] consisting of N steps one would take step = T /N and let k vary
from 1 to N . The figure labeled Brownian Trajectory Demo illustrates eight
(approximate) realizations of Brownian Motion in which each successive
realization reduces the step size by one-half. Notice how the curves progress
from a piecewise linear form to a very jaggy form in which much of the
change appears to be nearly vertical.

5
Monte Carlo Simulation c 2009-11, Stephen Demko

The Ito Process. At the level of intuition, an Ito Process is a stochastic

process, X(t), whose infinitesimal increments, dX, satisfy the equation
dX = a(X, t)dt + b(X, t)dW
where a and b are continuous functions and dW is a Wiener Process. When
we say X is a stochastic process, we mean that at each instant of time t
X(t) is a random variable. It may be possible to relax the conditions on
the functions a and b, but we will not concern ourselves with these sort
of technical matters here. Notice that an Ito process is a specialization of
the general stochastic process: the random increment must be Brownian
motion.
Ito processes are used to model a variety of financial phenomena
• Interest rates: Models proposed by Vasicek; Cox, Ingersoll, and Ross;
and Hull and White assume that the rate satisfies dr = u(r, t)dt +
w(r, t)dW for appropriate functions u and w.
• Equity Prices: The classical Black-Scholes framework starts with the
equation dS = µSdt + σSdW .
• Volatility:
√ Heston’s
√ model of stochastic volatility assumes dσ =
κ(θ − σ)dt + ζ σdW for appropriate parameters.
The full meaning of the notation used above is well beyond the scope of
this course. However, it is good to know a few “rules of thumb” for dealing
with Ito Processes:
(1) dW 2 = dt
(2) Classical calculus-like manipulations with partial derivatives can gen-
erally be done using this fact that dW 2 = dt.
One can generate approximate realizations of the stochastic process de-
scribed by an Ito Process by following the idea behind the simulation of
approximate realizations of Brownian motion. Suppose we have the Ito
Process dX = a(X, t)dt + b(X, t)dW with initial condition X(0) = x0 . The
the Euler-Maruyama method proceeds as follows:
(1) INITIALIZE: Select a positive number step to be “step size” and set
t0 = 0 and ti = i ∗ step.
(2) For i = 1, 2, . . .
• select a random z from a Gaussian distribution. √
• define xi = a(xi−1 , ti−1 )step + b(xi−1 , ti−1 )z step.
• draw a straight line between (ti−1 , xi−1 ) and (ti , xi ).
Geometric Brownian Motionis a special type of Ito process in which
the functions a and b are positive constant multiples of the random variable
X. When used to model the returns of equities, the coefficient of the dt term
is called the drift. The coefficient of the dW term is called the volatility.
With this we typically see the model for stock returns:
dS = µSdt + σSdW
where S(0) = S0 is a given initial spot price.
6
Monte Carlo Simulation c 2009-11, Stephen Demko

One can create approximate realizations of Geometric Brownian Motion

by simulating it as an Ito Process. While one can implement the Euler-
Maruyama method and take small steps of size ∆t and select from the
distribution at each step to get current value, there is formula that expresses
the distribution of the random variable S(t) in terms of S(0). With this
formula, one needs only to take one step of size T . This is discussed in the
next section.

Option Pricing by Monte Carlo Simulation

The Problem. The assumptions here are (1) the risk-free rate r and the
volatility σ are constant; (2) the option is European and does not depend
on the trajectory of the underlying. If P (S) denotes the price at t = T
and V (S, t) denotes the price of the option at time 0 ≤ t < T and under-
lying value S, the Black-Scholes Theorem says that V satisfies the partial
differential equation
1
(1) Vt + σ 2 S 2 VSS + rSVS − rV = 0
2
with terminal condition
V (S, T ) = P (S) for all S
It can be shown that the solution of (1) at time t0 and underlying spot
value S0 , V (S0 , t0 ), is the discounted expected value of the payoff applied
to the solution of the SDE
(2) dS = rSdt + σSdW
subject to the initial condition S(t0 ) = S0 .
Therefore, if we have only the present value of the option, we can solve
(2) starting with S(0) = S0 .

The Solution of the SDE. An application of Ito’s Lemma3 to the SDE

(3) dS = µSdt + σSdW
where volatility σ and drift µ are assumed constant, yields the following ex-
pression for the random variable S(t + ∆t) for arbitrary t ≥ 0 and increment
∆t > 0:
σ2
(4) S(t + ∆t) = S(t) exp{(µ − )∆t + σ[W (t + ∆t) − W (t)]}
2
where W is a standard, normalized Brownian Motion. While one might
2
guess at part of the formula, the − σ2 t is a “surprise” that arises from the
stochastic nature of the process.
3Ito’s Lemma if a fundamental result in stochastic calculus. It is similar to the chain
rule of regular Calculus. In this example, it is applied to calculate dY where Y = ln(S)
and S satisfies a Geometric Brownian Motion
7
Monte Carlo Simulation c 2009-11, Stephen Demko

Taking t = 0 and ∆t = T , since W (0) = 0, one can make one big jump
and write:
σ2
(5) S(T ) = S(0) exp{(µ − )T + σW (T )}
2
Realizations of ST can be calculated as
σ2 √
(6) ST (ω) = S0 exp{(µ − )T + σ T ω}
2
where ω is a sample from a Gaussian distribution4.
In principle one can use this to estimate probabilities related to ST (e.g
Probability {m ≤ ST ≤ M } where m and M are prescribed values) by
taking ω to be independent samples from a normal distribution.
The computer calculation of (6) requires the ability to generate samples
of a Gaussian distribution. Either a computing environment will contain a
functionality to manufacture independent Gaussian (pseudo-)random5 num-
bers, or the programmer will write such a tool. For example, in Matlab the
function randn does the job and in Java the method Gaussian() of the
Random class does it. Methods for generating a sequence of independent
Gaussian samples typically involve a transformation of independent uni-
formly distributed samples. In the last ten years the “Mersenne Twister”
algorithm has become very widely used because of both its speed and the
quality of the uniformly distributed random numbers it produces. Code for
this is readily available on the internet6. See the Wikipedia article on the
Mersenne Twister for appropriate links. A good reference on the generation
of pseudo-random numbers is Donald Knuth’s classic The Art of Computer
Programming, Volume 2: Seminumerical Algorithms (3rd Edition).
The current discussion of option pricing has as its focus the case that
the underlying satisfies a GBM. As noted above this allows us to “jump”
from the present spot value to spot values at expiration and, thus, avoid
the discrete time approximation of the Euler-Maruyama method. Certain
options that depend on the history (or path, or trajectory) of the underlying
can also be easily priced with the method discussed here, for example, a call
option for which the settlement price is not the spot at expiration but an
average of the closing prices of the underlying at a finite sets of trading
days. Such an option is called an Asian-tail option 7. Some other types of
options, like American Puts, can require complex modifications of the basic
4A Gaussian has a normal distribution with mean 0 and variance 1. We also call this
the Standard Normal distribution.
5We call the output of algorithms that attempt to produce random numbers of any
sort “pseudo-random numbers”. This is in contrast to typos which should be deemed as
random, and hopefully rare, events.
6As a rule, if you have both legal access and user rights to well written, useful code, it
makes lots of sense to use it if you have the need.
7A good exercise is to plan the coding for an Asian tail option, assuming that strike,
rate, and volatility are given
8
Monte Carlo Simulation c 2009-11, Stephen Demko

Monte Carlo method because the option contract can cease to exist before
expiration.
Recalling that the present value of a non-path dependent European option
with payoff function P (S) is the discounted expectation of the payoff give
the present spot value, we write this in symbols as:

(7) V0 = e−rT E[P (S(T ))|S(0) = S0 ]

The expectation is with respect to the weighting of trajectories given by

the SDE (2).
In the Monte-Carlo method of pricing, one approximates the discounted
expectation in (7) by generating a large number of pseudo-random numbers
ω1 , ω2 , . . . ,ωN and the associated values of the underlying ST (ωi ) and use
the standard unbiased linear estimator of expectation to get V̂N
N
1 X
(8) V̂N = e −rT
P [ST (ωi )]
N
i=1

A standard computer implementation of the above formula would go as

follows: set sum = 0.0, then for k = 1, 2, . . . do the following:
Get random z from N (0, 1): Get a new pseudo-random number (Gauss-
ian) at each step. √
2
Set ST = S0 exp{(r − σ2 )T + σ T z}: Generate a sample of the ter-
minal distribution.
Set Pi = P (ST ): Apply the payoff.
Set sum = sum + Pi : . Accumulate the sum of payoff values to date
Set Avek = sumk : Compute this only if you want the approximate op-
tion price through k steps. In this case, V̂k = e−rT Avek .
It is sometimes useful to save the values in the last (optional) step to
create a chart that shows the convergence of the method. Such charts are
illustrated near the end of this document.

Error Theory. The basic error theory of the Monte Carlo method is phrased
in terms of concepts from probability and statistics, in particular in terms
of confidence intervals. This is no surprise since the idea is essentially to
estimate the expectation of a random variable by averaging a largeR number of
x
samples of the random variable. We use the notation Φ(x) = √12π −∞ e−0.5s ds
If we let Vi denote the estimate computed at step i, Vi = e−rT Pi , then
with VˆN = N1 (V1 + V2 + · · · + VN ), we can compute the sample standard
deviation of V1 , V2 , . . . , VN by
r
1
sN = (VˆN − Vi )2
N −1
9
Monte Carlo Simulation c 2009-11, Stephen Demko

With V denoting the true value of the option, the Central Limit Theorem
asserts that the probability of
VˆN − V
−x ≤ √ ≤x
SN / N
Rx 2
converges, as N → ∞, to 2Φ(x) − 1 = √12π −x e−0.5s ds.
If we want this probability to be large, say ≥ 1 − ǫ, then we should take x
such that 2Φ(x) − 1 = 1 − ǫ, or Φ(x) = 1 − 2ǫ . Let’s denote this value of x by
xǫ : Φ(xǫ ) = 1 − 2ǫ .. So with N simulations and given ǫ reasonable small, we
would take some asymptotic probabilistic comfort that the value of V lay in
the interval (VˆN − √ SN
x , Vˆ + √
N ǫ N
SN
x ). We can make the interval smaller by
N ǫ
taking more simulations, but since the length of the interval is controlled by
1 th
N which appears under a square root sign, to make the interval about 10
smaller we would need to go from N simulations to 100N .
This points out the main limitation of the method. Accuracy depends on
many simulations. Normally, tens of thousands or even millions are required.
Monte Carlo As a Numerical Integration Method. Since the stan-
dard normal distribution (i.e., N (0, 1)) has distribution function d(x) =
2
√1 exp(− x ), we can write the expectation of the payoff function as an
2π 2
integral
1
Z ∞
σ2 √ ω2
(9) E[P (S)] = √ P [S0 exp{(r − )T + σ T ω}]e− 2 dω
2π −∞ 2
It is useful to normalize the interval of integration to [0, 1]. So, with
Z ω
1 t2
N (ω) = √ exp{− } dt
2π −∞ 2
we make the change of variables z = N (ω) and conclude from standard
Calculus that
Z 1
σ2 √
(10) E[P (S)] = P [S0 exp (r − )T + σ T N −1 (z)] dz
0 2
In this second integral we see that as z moves throughout the interval [0, 1],
the values of N −1 (z) vary over the entire set of real numbers. Moreover,
if the values of z are distributed uniformly over [0.1], then the values of
w = N −1 (z) are distributed like a Gaussian over (−∞, ∞). It is clear that
if f is continuous, then approximations like

N √
1 X σ2
(11) E[P (S)] ≈ P [S0 exp (r − )T + σ T N −1 (zi )]
N 2
i=1
should converge to the value of the integral as the number of points N
grows, provided that the points z1 , z2 , . . . , zN fill up the interval [0, 1] as
N → ∞.
10
Monte Carlo Simulation c 2009-11, Stephen Demko

With this view the method is seen to be an attempt at approximating

an integral by evaluating the integrand at a number of points ωi and taking
the average value. Since we know that the distribution function is a positive
function whose integral is 1, this makes some sense. However, it is clear
there are good choices as well as bad choices for the placements of the ωi .
For example, zi that tend to cluster in the “wrong places” are probably not
good. For this reason the calculation in (8) may be repeated several times in
order to diminish the effect of a “bad” sequence of pseudo-random numbers.
Because a poor distribution of the pseudo-random numbers can lead to
a less than optimal approximation, the “seed” that sets the initial state of
the pseudo-random number generator can be an issue in convergence of the
method. Some seeds result in better distributions than others. Often these
seeds are published for the benefit of the user community. Another advan-
tage of setting and knowing the seed is that, provided that the computations
proceed in exactly the same way in each run of the program, results of simu-
lations can be reproduced exactly if the seed is retained8. If the user does not
prescribe a seed, then the pseudo-random number generator creates a seed
from the computing environment, often from the clock. Standard practice
is to perform several MonteCarlo evaluations of the option–each evaluation
using a large number of sample and average these separate evaluations.
Given the potential for poor distribution, it is reasonable to question the
need for pseudo-random numbers in the approximation (11). One might rea-
sonably suggest that the zi should be taken to be uniformly spaced, or even
that a standard numerical integration method, like the Trapezoid Rule or
Simpson’s Rule, be used. These ideas appeal to both the underlying geom-
etry of the interval and to the unstated assumption that the payoff function
is often continuous, maybe piecewise smooth. The idea of using one of the
standard numerical integration methods will surely work in the simplest
cases. However, if one’s model does not have constant rate or volatility, this
won’t work. Also, in higher dimensional problems, the standard methods
are not feasible because they depend in too many points. However, soon
we will address the issue of generating points that might be better than
random.

Error Estimates. We investigate the convergence of the approximations

(11) to the integral (expectation) in (10). It costs little to consider the more
general case of estimating an integral over a d-dimensional set Ω of “measure
1” in Rd . Measure here is the natural sense of size for the dimension: area
for 2-dimension, volume for 3-dimensions, and so on. The main case of
interest is when Ω is the set {(x1 , x2 , . . . , xd ) : 0 ≤ xi ≤ 1} because we
are thinking of the problem of pricing an option that depends on several
underlying securities. In this case the payoff function is integrated against

8However, if you modify your code, then a given seed need not replicate the results of
an earlier experiment
11
Monte Carlo Simulation c 2009-11, Stephen Demko

the density of a multivariate normal distribution. This would be transformed

to an integral over this special Ω.
continuous function9 on Ω and denote Ω f (x) dx by E(f ).
R
Let f be any
Let σ 2 (f ) = Ω (f − E(f ))2 dx We are approximating the integral of f by an
R
average of values of f at uniformly distributed points in Ω:
N
1 X
E(f ) ≈ f (xi )
N
i=1
The expected square error is obtained by averaging over all choices of
(x1 , x2 , . . . , xN ):
N N
1 X 1 X
Z Z Z
2
E[E(f ) − f (xi )] = . . . [E(f ) − f (xi )]2 dx1 dx2 . . . dxN
N Ω Ω Ω N
i=1 i=1
A careful calculation reveals that
N
X
(12) E[E(f ) − N1
f (xi )]2 = 1 2
N σ (f )
i=1

Bottom line: as N → ∞ error is only O( √1N )10. So, to achieve a ten-fold

decrease in error one should expect to take N one hundred times large!
At first this appears discouraging. If we price a single option by solving
the Black-Scholes PDE with appropriate side conditions, we can expect our
error to be at worst O( N1 ) and often O( N12 ) where N is the number of input
parameters to our scheme.
On the other hand, the Monte Carlo convergence rate does not depend on
the dimension of the underlying space. This makes it possible to apply the
Monte Carlo method in situations11 where the standard methods for solving
PDEs choke on data management.12
In the sections that follow we introduce some methods for accelerating
convergence of the Monte Carlo method. This is meant to be only an in-
troduction to some basic ideas. We will consider taking advantage to the
geometry of Ω soon. But first, we introduce a technique for variance re-
duction and a technique for comparing one Monte Carlo simulation with
another in hopes of generating obtaining an improved approximation.
9We could just require that f 2 be integrable.
10The meaning of the “‘big Oh” notation: to say that an expression, say “E is O(f (n))
as n → ∞” means (1) that the expression depends on the parameter n and (2) that as
n → ∞ the expression is bounded by a constant times f (n).
11Typically in problems in which the interactions of several SDEs are used to model
dynamics of a financial derivative.
12To achieve error of the order O(N −2 ), finite difference methods typically seek a
discrete solution depending on N parameters in each dimension of the problem. When
the dimension gets above 3 or 4, the number of variables that need to be managed becomes
a problem. On the other hand, when the data can be managed, these methods provide a
fairly complete picture of the whole solution V (S, t), whereas basic Monte Carlo methods
yield only a single number at a time.
12
Monte Carlo Simulation c 2009-11, Stephen Demko

Variance Reduction by Antithetic Variates. The idea here exploits

symmetry in the governing probability to generate two realizations from
each random number. Under certain conditions Rthe variance isR reduced. For
1 1
example, if f is integrable on [0, 1], then E[f ] = 0 f (x) dx = 0 f (1 − x) dx.
So if we define
f (x) + f (1 − x)
g(x) =
2
we have that E[f ] = E[g] and

N
1 X σ 2 (g)
E[ g(xi ) − E[g]]2 =
N N
i=1

One can show that if f is continuous and monotone, then σ 2 (g) ≤ 21 σ 2 (f ).

Therefore, in this case

N
1 X σ 2 (f ) σ 2 (f )
E[ g(xi ) − E[f ]]2 = <
N 2N N
i=1

With the Normal distribution, we use the symmetry to generate 2 real-

izations from each random number and take the average of the payoff values
at them as the approximation. The coding is simple; we add two lines to
the pseudo-code following (8) to get:
• Get random z from N (0, 1). √
2
• Set r+ = P [S0 exp{(r − σ2 )T + σ T z}]
2 √
• Set r− = P [S0 exp{(r − σ2 )T − σ T z}]
• Set ave = 0.5 ∗ (r+ + r− )
• Set sum = sum + ave.
• If we want the approximate option price through k steps, set Avek =
ave
k and multiply by the discount factor.

This method is widely used because it is easy to implement and promises

a reduction of variance of the error in many cases—even though its effect
may be hardly noticed.

Control Variates. The idea here is to simultaneously simulate both the

integral whose value you need and the integral of a related quantity that
you know exactly, or to high precision and that is highly correlated with the
integral you need. For example, suppose you know the exact value of a par-
ticular option to be V ∗ and wish to find an approximation to another option
whose (unknown) value is V . We calculate Monte Carlo approximations to
both options with the same sequence of random values x1 , x2 , . . . , xN and
13
Monte Carlo Simulation c 2009-11, Stephen Demko

at step N have the values

N
1 X
VN∗ = PV ∗ (xi )
N
i=0
N
1 X
VN = PV (xi )
N
i=0

The computable quantity V − ∗ VN∗

is viewed as a correction term to VN
and a new approximation to V is defined by
VCV = VN + V ∗ − VN∗
The quantity V ∗ does not have to be the value of an option. It just has
to be a quantity that (1) is highly correlated to the option you want to
price and (2) is known to a high precision. One can quantify, in terms of
the correlation between the two options being priced, the improvement in
efficiency that the Control Variate method yields .13
Sometimes the idea here is used to approximate American Options with
the Binomial Method. One prices the European version of the option with
both the same Binomial Tree and the appropriate Black-Scholes Formula.
Then, one uses the approximation
AmerP rice = AmerP riceBinomial +EuroP riceBlackScholes−EuroP riceBinomial

Low Discrepancy Sequences. For routine option calculations, the Monte

Carlo method does not provide the accuracy of the binomial method or
PDE methods because its accuracy is limited by probabilistic arguments.
However, taking the view of integration, one can see that randomness is not
needed. One can get better estimates of integrals in [0, 1] by using uniformly
space points. However, if one is not satisfied with the accuracy of a given
method, one may have to either greatly increase the grid or start all over
with a new grid–i.e., restart the process.
A better idea is to generate a sequence of numbers xi ∈ [0, 1] with the
property that for each value of N the numbers x1 , x2 , . . . , xN are nearly
uniformly distributed in [0, 1]. Now if N is prescribed ahead of time, we could
simply take uniformly spaced points. But, in Monte Carlo applications N
is typically not known in advance and is likely to be changed (raised) in the
process of the calculation. Various authors have presented algorithms for
generating such “low discrepancy” sequences.
Perhaps the simplest way of doing this is to generate a Halton sequence.
The simple idea is to place down an point and then place the next point in
[0, 1] to be as far away from the first one as possible and continue this way.
For example, if we start with 0, 1, then the next point is 12 . Then the next
two points are 14 and 43 , then we move to the eights.

13Tavella, Quantitative Methods in Derivative Pricing , page 137.

14
Monte Carlo Simulation c 2009-11, Stephen Demko

More generally, if p is a prime, then any integer k can be represented as

a sum of powers of the prime (i.e., in “base-p”):
k = a0 p0 + a1 p1 + a2 p3 + . . . ai pi
where 0 ≤ ai ≤ p − 1 and i ≤ logp (k).
With this define the kth -Halton number of base p to be
X
φp (k) = ai p−(i+1)
i
If we have an integral over the unit cube in d-dimensions Ω = {(x1 , x2 , . . . , xd ) :
0 ≤ xi ≤ 1}, then it is known that (φ2 (k), φ3 (k), . . . , φp (k)) is a higher di-
mensional analogue where p is the dth prime.
Several sequences of quasi-random or uniform discrepancy numbers are
known: Sobel, Faure, Niederreiter, ...
d
Convergence rate in d−dimensions is of order O( (lnNN ) ).
Graphical Representation of Convergence of the Method. Below
are graphical representations of the convergence paths of five independent
Monte Carlo approximations to the price of a particular option. In each case
the quantity being charted is the current estimate of the expected value of
the payoff. The first picture shows these values over the range of 1 to 5000
realizations of the payoff.
5
"R5000-1.out"
"R5000-2.out"
"R5000-3.out"
"R5000-4.out"
"R5000-5.out"
4.8

4.6

4.4

4.2

4
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000

15
Monte Carlo Simulation c 2009-11, Stephen Demko

The second picture is a close up of the first over the range of 2500 to 5000
realizations.
4.7
"R5000-1.out"
"R5000-2.out"
"R5000-3.out"
"R5000-4.out"
4.6 "R5000-5.out"

4.5

4.4

4.3

4.2

4.1

4
2500 3000 3500 4000 4500 5000

The third picture shows the result of applying the method using the Hal-
ton Sequence to generate uniformly distributed points and then generating
the associated normally distributed points. Note that there is one answer
here since the process is deterministic. Also, note the rate of convergence.
5
"QR5000.out"

4.8

4.6

4.4

4.2

4
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000

16
Monte Carlo Simulation c 2009-11, Stephen Demko

Summary of Advantages and Disadvantages. Below are some of the

advantages and disadvantages of the Monte Carlo method.
Advantages: •
• Higher dimensional problems14 can be treated with Monte Carlo
Simulation in a fairly straight-forward manner; whereas, such
problems become intractable with Binomial and PDE methods
after a small number of dimensions. The applicability of the
approach to multi-asset problems is a major advantage. For
this reason all (well-rounded) students of the subject must un-
derstand the approach and gain some experience with it.
• One can consider stochastic σ and r in a straight forward way as
long as they are generated as solutions to another SDE. In this
case however one cannot use the formula (1) but must discretize
time and deal with the inaccuracies of approximating the SDE.
In this case one has both N → ∞ and ∆t → 0 where ∆t is the
time discretization parameter used in the SDE for σ or r (or
both) as the case may be.
• Certain types of options whose payoff depends on the average
daily closing value of the underlying over a given period of time
can be evaluated via this technique in a straight forward way.
Disadvantages: • The method is really, really slow. Especially
when compared with Binomial Trees or PDE methods. This all
but makes Monte Carlo useless for vanilla options.
• All you get for your efforts is the option price. No Greeks, no
graphs.
• One can treat American options, but the simplicity of the overall
approach is lost.
References.
(1) Monte Carlo Methods in Financial Engineering by Paul Glasserman,
Springer, 2004. The basic reference.
(2) http://www.espenhaug.com/quasi_random.html
contains some examples.

14These arise for example when the value of an option depends on more than one
underlying.
17