Suresh Chandra

S. Dharmaraja
Aparna Mehra
Reshma Khemchandani

α
Alpha Science International Ltd.
Oxford, U.K.
Financial Mathematics
An Introduction
516 pgs. | 48 figs. | 23 tbls.

Suresh Chandra
S. Dharmaraja
Aparna Mehra
Department of Mathematics
Indian Institute of Technology
Hauz Khas, New Delhi
Reshma Khemchandani
RBS Global Banking and Markets
DLF Building, 7, Cybercity
DLF Phase 3, Sector 25 A
Gurgaon, Haryana

Copyright © 2013
ALPHA SCIENCE INTERNATIONAL LTD.
7200 The Quorum, Oxford Business Park North
Garsington Road, Oxford OX4 2JZ, U.K.

www.alphasci.com

All rights reserved. No part of this publication may be reproduced, stored in a retrieval
system, or transmitted in any form or by any means, electronic, mechanical, photocopying,
recording or otherwise, without prior written permission of the publisher.

Printed from the camera-ready copy provided by the Authors.

ISBN 978-1-84265-654-9
E-ISBN 978-1-78332-002-8

Printed in India
viii Preface
Preface ix
xii Contents
Contents xiii
xiv Contents
Contents xv
1
Financial Mathematics: An Overview

1.1 Introduction
The word financial mathematics refers to the application of mathematics to study
the problems arising in the area of finance. Here the word mathematics has been
used in a wider sense so as to include subjects like probability and statistics,
stochastic processes, optimization, econometrics, numerical analysis and partial
differential equations. Also the scope of the word finance could be very broad but
here it should be understood in terms of investment of money for the purpose of
receiving more money (hopefully!!) at sometime in future. Thus financial mathe-
matics could also be termed as investment science or investment theory.
Though investment as an art form has always been there, it developed as sci-
ence/theory due to the pioneering work of Fisher Black, Myron Scholes, Robert
Merton, Harry Markowitz, William Sharpe and John Lintner amongst many oth-
ers. In fact, Merton and Scholes got Nobel Prize in Economics (in 1997) for their
work on pricing of option carried out during late 60’s and early 70’s. Black could
not share this prize as he died of illness before 1997. Earlier Markowitz along with
Sharpe (Markowitz’s Ph.D. student) were awarded Nobel Prize (in 1990) for their
work on portfolio optimization carried out during late 50’s and early 60’s. They
shared Nobel Prize with another famous economist Merton Miller.
The above extra ordinary theoretical developments had great impetus on re-
search in finance. This together with globalization of investment activities and
tremendous power of computing technology has made the financial mathematics
area intellectually very fertile. There is a significant growth in theory as well as
in applications in the areas of derivative pricing, portfolio management, interest
rate modeling, credit risk methodologies and algorithmic trading.
The aim of this chapter is to introduce basic notions and assumptions of a
simple financial market model so as to facilitate the presentation of some simple
2 Financial Mathematics: An Introduction

investment problems for illustration purposes. Our discussion here is motivated

by some of the excellent texts. We shall particularly like to mention Capinski and
Zastawniak [25], Hull [65] and Luenberger [85].

1.2 Some Basic Terminology

Any financial market model has to deal with the trading of financial assets. By a
financial asset we mean a financial instrument that can be bought or sold. Here an
asset will always mean a financial asset. It is common knowledge that some assets
are risk free and some are risky. For example a bank deposit or a bond issued by a
government are risk free where as stocks are risky. To introduce certain notations
and terminology for a financial market we first assume a very simple scenario.
We assume that only two assets are being traded. These are, one risk free asset
(bond) and one risky asset (stock). The other parameter in modeling a financial
market is the time horizon. The simplest of these is the single period model where
we restrict the time scale to be two instants only, namely t = 0 and t = 1 and any
transaction takes place at these two time instants only. Realistically it has to be
a multi period or continuous time scenario. But here we restrict our discussion to
the case of single period only, the multi period and continuous time models will
be discussed later.
Let t = 0 and t = 1 be two time instants. Let B(0) denote the price of one bond
at t = 0 and B(1) be the corresponding price at t = 1. Similarly let S(0) denote
the price of one share of stock at t = 0 and S(1) be the corresponding price at
t = 1. As bond is a risk free asset, both B(0) and B(1) are known at t = 0. But
for the case of stock, which is a risky asset, though S(0) is known, S(1) remains
uncertain. The stock price at t = 1, namely S(1), may go up or down with certain
probability, i.e. S(1) is a random variable. Thus for a risk free asset both B(0)
and B(1) are deterministic, but for a risky asset S(0) is deterministic while S(1) is
a random variable. We next define the rate of return of an asset. For the case of
bond it is defined as
B(1) − B(0)
rB = ,
B(0)
and for the case of stock it is defined as
S(1) − S(0)
rS = .
S(0)
Here we note that rB is known exactly but rS is a random variable because S(1) is
a random variable. In the finance literature, rate of return is called return only.
 Some Basic Terminology 3

Thus when we refer to the return of an asset we really mean the rate of return of
that asset.
Let an investor hold x shares of stock and y units of bond at t = 0. The pair
(x, y) is called a portfolio and it is denoted by P : (x, y). The value of the portfolio
P : (x, y) at time t = 0 is given by
VP (0) = x S(0) + y B(0).
Similarly the value of the portfolio at time instant t = 1 is given by
VP (1) = x S(1) + y B(1).
The quantity
Vp (1) − Vp (0)
rP = ,
Vp (0)
is called the rate of return of the portfolio P : (x, y). We note that rP is a random
variable because VP (1) is so.
The units of B(0), B(1), S(0) and S(1) are units of money and here they are
always taken in Rupees. Thus B(0) = 100 etc. will always mean B(0) = Rs 100 etc.
We shall follow this convention throughout the book.
Since the portfolio return rP is a random variable we can talk of its expected
value and variance. The quantity E(rP ) is called the expected return of the portfolio
P : (x, y) and the standard deviation σP of rP is called its risk. Obviously if we are
given a choice between two portfolios with the same expected return, we should
choose the one for which the risk is least. In a similar manner, if the risk levels of
two portfolios are same we should choose the one for which return is maximum.
This is a typical problem of portfolio optimization. We shall have opportunity
of discussing a general n-asset problem of portfolio optimization in context of
mean-variance theory of Markowitz in later chapters of the book. The concept of
risk itself has been a topic of research in the recent past. Is standard deviation
or variance the correct measure of risk? Are there some other better definitions
of risk? All these questions need to be answered and we plan to discuss them in
greater detail in a later chapter.
Example 1.2.1 Let B(0) = 100, B(1) = 110, S(0) = 80 and


100, with probability 0 · 8
S(1) =
60, with probability 0 · 2.
What is the expected return on the stock? Consider the portfolio P : (x = 50, y =
60). Determine the expected return and the risk for the given portfolio.
4 Financial Mathematics: An Introduction

Solution The return on bond is obviously 10%. But for the stock
⎧
⎪
⎪ 100 − 80
⎪
⎪ , with probability 0 · 8
⎪
⎪ 80
⎨
rS = ⎪
⎪
⎪
⎪ 60 − 80
⎪
⎪
⎩ , with probability 0 · 2,
80
i.e.

0 · 25, with probability 0 · 8
rS =
−0 · 25, with probability 0 · 2.
Therefore E(rS ) = (0 · 25)(0 · 8) + (−0 · 25)(0 · 2) = 0 · 15.

Further as x = 50 and y = 60, we have

VP (0) = (50 × 80) + (60 × 100) = 10, 000.

Also


(50 × 100) + (60 × 110), with probability 0 · 8
VP (1) =
(50 × 60) + (60 × 110), with probability 0 · 2


11, 600, with probability 0 · 8
=
9, 600, with probability 0 · 2.

Therefore
VP (1) − VP (0)
rP =
VP (0)


0 · 16, with probability 0 · 8
=
−0 · 04, with probability 0 · 2.

This gives the expected return on the portfolio as

E(rP ) = (0 · 16)(0 · 8) + (−0 · 04)(0 · 2) = 0 · 12.

We have defined the standard deviation of the portfolio as its risk. Therefore the
risk of given portfolio is

 2
σP = E rp − E(rp ) = (0 · 16 − 0 · 12)2 × (0 · 8) + (−0 · 04 − 0 · 12)2 × (0 · 2)
= 0 · 08. 

 Certain Assumptions 5

1.3 Certain Assumptions

We now state certain assumptions which help in building a simple mathematical
model of a typical financial market. The real world financial market is extremely
complex and any mathematical model has to trade off between reality and sim-
plifications which make the model amenable to analytical or numerical solution.
Apart from the time horizon, it is the dynamics of the assets which plays a very
crucial role in the modeling aspect of a financial market. Thus, for a general case,
we need to describe the dynamics of the stock price S(t) as t varies over a contin-
uum. Since dynamics generally refers to a differential equation, we expect that the
stock price movement will be represented by a differential equation. But things are
not that simple as for each t, S(t) is a random variable i.e. it is a random process.
So there has to be an appropriate stochastic process component in this differential
equation leading to something called a stochastic differential equation (SDE). So
for a proper understanding of this dynamics we need to study stochastic calculus
and stochastic differential equations more deeply. Since more than one stock may
be in the portfolio, one has to deal with more than one SDE and also to take into
account their interaction as stocks may be correlated. But as agreed in Section
1.2, in this chapter we restrict our discussion to the case when only two assets
(one bond and one stock) are being traded and the time horizon is only one period
(from t = 0 to t = 1).
For the present we do not take the general case and state certain assumptions
for our single period case only. We should take all these assumptions only as a
first step towards modeling and try to modify them later so as to make the model
more realistic.
Assumption 1 (Randomness)
The stock price S(1) at t = 1 is a random variable with at least two different values.

Assumption 2 (Positivity of Prices)

All bond and stock prices are strictly positive, i.e. B(0) > 0, B(1) > 0, S(0) > 0
and S(1) > 0.

Assumption 3 (Divisibility, Liquidity and Short Selling)

Let the investor hold x shares of stock and y units of bond. Then x ∈ R and y ∈ R.
Thus the investor is permitted to hold any units of bond and any number
of shares of stock. These numbers may be zero, positive or negative reals. This
property is referred to as divisibility property. In a real market this property is
expected to hold when the volume of transactions is large compared to unit prices.
6 Financial Mathematics: An Introduction

As there are no bounds on x and y they can be as large or as small as we wish.

In other words, an investor can buy or sell arbitrary quantities at the market price
on demand. This property is called liquidity. In reality there is always a restriction
on the volume of trading so this property may not hold in general.
The next thing to understand is the interpretation of sign restriction on x and
y. If x > 0 then it obviously means that at t = 0, the investor is buying x shares
of stock and this we state as “The investor has taken a long position on stock”.
Similarly we interpret y > 0. However, if x < 0 (respectively y < 0), we say that
“the investor has taken a short position on stock (respectively bond)”. Thus if
x > 0 and y < 0, we say that investor has taken a long position on stock and a
short position on bond. If the investor has taken a short position on an asset we
say that the asset is shorted.
But what is the physical interpretation of a short position ? A short position in a
risk free asset (bond) means borrowing cash, where the interest rate is determined
by the bond price. Thus if y = −10, B(0) = 100 and B(1) = 110, then it means
that the investor is borrowing Rs 1000 from the bank where the interest rate is
10%. We say that the short position in the bond is closed when the borrowed cash
along with interest is repaid to the bank or lender.
A short position in the stock is slightly more complicated. This is best under-
stood in terms of short selling. Short selling is a way to make profit if price goes
down. Suppose that the investor feels that the price of a stock is overvalued, so it
makes sense in selling the stock and purchasing it when the price goes down. In
this situation the investor can resort to short selling. Here the investor borrows
the stock from someone (say broker), sells it and uses the amount received to
make some other investment. But the original owner of the stock has all the rights
vested to him/her. Thus he/she receives all dividends due and has the full right
to sell the stock any time. In this situation, the investor has to return the stock
to the original owner. If he/she does not have stock with him/her, he/she may
have to buy from the market at the prevailing market price. Therefore it becomes
imperative for the investor to keep sufficient resources (cash) with him/her to fulfil
this obligation. We say that the short position on the stock is closed when the
stock is returned to the original owner by the investor. The investor may return
the stock of his/her own or when the owner asks for the same. In practice the
investor is also asked to give a security deposit with the broker from which the
stock is borrowed. But in theory we do not consider this.
The above discussion suggests that to meet his/her obligations as per Assump-
tion 3, the wealth of the investor should never be negative. This is precisely the
Assumption 4 given below.
 Certain Assumptions 7

Assumption 4 (Solvency)
The wealth of the investor should never be negative, i.e. VP (t) ≥ 0 at all times t.
For the single period case it simply means that VP (0) ≥ 0 and VP (1) ≥ 0.
A portfolio (x, y) satisfying the solvency property is termed as an admissible
portfolio.
Short selling is generally considered to be extremely risky because in short
selling there are possibilities of unlimited losses. Let the borrowed asset be sold
for an amount X0 at t = 0 and later at t = 1 it be purchased for an amount X1
so as to close the short position. If X1 is less than X0 then a profit of (X0 − X1 ) is
made. In other words short selling is profitable if the asset price goes down. But
if the asset value increases then the loss is (X1 − X0 ), which can be arbitrary large
since X1 can increase arbitrary. Because of this not all financial institutions allow
short selling. In some institutions it may be allowed only on some assets and that
too with certain restrictions.
Let us consider an example in this regard. Let S(0) = Rs 10 and let an investor
decide to short 100 shares of this stock at t = 0. For this he/she borrows 100
shares from a broker and sells these in the stock market to receive an amount
of Rs 1000. Let at t = 1, the stock price S(1) become Rs 9, so the investor buys
back 100 shares from the market at Rs 900 and gives these shares to the broker
to close the short position. This has been a favorable position to the investor as
stock price has gone down. In fact the investor has made a profit of Rs 100 due
to short selling. If the price had gone up, the investor would obviously had made
losses.
Now imagine another investor who has not gone for short selling but rather has
purchased 100 shares in the beginning at t = 0. Imagine that this investor has to
sell these shares at t = 1 (the problem being a single period only). This will entail
a loss of Rs 100 to this investor. The rate of return on the asset for this investor
will be
900 − 1000
r= = −0 · 10
1000
i.e. −10%.
This example clearly illustrates that short selling allows an investor to convert
a negative rate of return into a profit. This is because the original investment
of the investor is also negative. It gives a profit of (original investment)×(rate of
return), i.e. (−1000)(−0 · 10) = 100.
8 Financial Mathematics: An Introduction

Assumption 5 (Discrete Unit Prices)

By Assumption 1, the stock price S(1) at time t = 1, is a random variable taking
at least two distinct values. What we assume here is that S(1) is a random variable
taking only finitely many values.
This assumption is again not realistic as S(t) at any time t can take values
in a continuum. But again it is taken only for the sake of simplicity and will be
generalized later.
We next present no arbitrage principle and the law of one price. In fact the
law of one price is a consequence of no arbitrage principle. We say that there is
an arbitrage opportunity in the market when someone, with no invested capital,
has a positive probability of achieving a positive return with no risk of loss. So
in our common usage of English language, arbitrage opportunity really means
free lunch. It is certainly not desirable in a fair market, and therefore we make the
assumption given below.

Assumption 6 (No Arbitrage Principle)

There is no portfolio with initial value V(0) = 0 such that V(1) ≥ 0 with probability
1 and V(1) > 0 with non-zero probability. Thus if the initial value of a portfolio
is zero, i.e. V(0) = 0 (no initial investment), and V(1) ≥ 0 (no risk of loss), then
V(1) = 0 (no profit) with probability 1.
As a consequence of no arbitrage principle, no investor can lock in a profit
without risk of loss and no initial endowment. In other words, there are no zero
cost investment strategies that allow an investor to make a profit without taking
some risk of a loss. In this context, we may note that investing in a Government
bond (or fixed deposit) is not an arbitrage though it guarantees a positive net
return because capital is invested initially. Some readers may like to analyze if
short selling is an arbitrage opportunity? The answer is obviously in negative.
Assumption 6 can also be stated mathematically as follows
There is no portfolio such that for some t
(i) V(0) = 0,
(ii) V(t) ≥ 0 with probability 1,
(iii) P[V(t) > 0] > 0 ,
where V(t) is the wealth of the portfolio at time t. For the single period case it
means that there is no portfolio such that
(i) V(0) = 0,
(ii) V(1) ≥ 0 with probability 1,
(iii) P[V(1) > 0] > 0.
 Certain Assumptions 9

The following is rather an artificial example of arbitrage opportunity. Consider

two banks (say Bank 1 and Bank 2) that offer to loan money or accept deposits at
the same rate of interest but these are different in the two banks. Let the rate at
Bank 1 for loan and deposits be 10% and at Bank 2 be 12%. So an investor can go
to Bank 1 and borrow say Rs 10, 000 at 10% and then deposit that Rs 10, 000 at
12% interest at Bank 2. At t = 1, the investor will get Rs 11, 200 and then he/she
can return Rs 11, 100 (borrowed money and 10% interest on it) to Bank 1. In this
process he/she has earned Rs 100 (guaranteed) without making any investment
of his/her own. Obviously such opportunities are rare and instantaneously both
banks will adjust the same rate of interest and there will be no arbitrage.
The above discussion suggests that in a financial market, arbitrage opportunity
cannot prevail for extended periods of time and therefore it justifies Assumption
6. This assumption is extremely important for any meaningful study in financial
mathematics and we shall have opportunity to see its application throughout the
book.
As a consequence of no arbitrage principle we have the law of one price or also
called the comparison principle. It essentially tells that if two financial instruments
have exactly the same pay-offs, then they have the same price.
To have a better idea of the law of one price, let us imagine two portfolios, say
Portfolio P and Portfolio Q. Let us consider the case of single period only and refer
to the time instants as t = 0 and t = 1. The two portfolios P and Q have known
(determinate) value at t = 0 but their values at t = 1 (also called pay-offs) are
random variables. Thus VP (0) and VQ (0) are deterministic, but VP (1) and VQ (1)
are random variables, whose values will depend upon the state of economy ω ∈ Ω.
Therefore,
VP (1) ∈ {VP,1 (ω) : ω ∈ Ω} ,
where VP,1 (ω) refers to the value taken by the random variable VP (1) when the
state of the economy is ω. In particular, if there are m possible states of economy,
i.e. Ω = {ω1 , ω2 , . . . , ωm } then

VP (1) ∈ {VP,1 (ω1 ), . . . , VP,1 (ωm )} ,

where VP,1 (ωi ) refers to the value of the random variable VP (1) when the state of
economy is ωi (i = 1, 2, . . . , m). In a similar manner

VQ (1) ∈ {VQ,1 (ω) : ω ∈ Ω}.

The law of one principle states that if VP,1 (ω) = VQ,1 (ω), for all ω ∈ Ω, then
VP (0) = VQ (0).
10 Financial Mathematics: An Introduction

To see that it should be true, suppose that VP (0) > VQ (0). Then at t = 0
an investor can sell the costlier portfolio (Portfolio P) and purchase the cheaper
portfolio (Portfolio Q), pocketing the difference. At t = 1, no matter what the state
of economy is in, the investor receives the common final value of the portfolios.
Thus the investor looses nothing at the end and made profit. But this violates the
principle of no arbitrage. Hence VP (0) cannot be more than VQ (0). In a similar
manner we can show that VQ (0) cannot be more than VP (0). Hence VP (0) = VQ (0).
The law of one price has proved to be a very effective tool in pricing derivative
securities, which we shall discuss in detail in later chapters.
Apart from the above main assumptions it also assumed that there are no
commissions or transaction costs and the lending rate is equal to the borrowing
rate.

1.4 Derivative Securities

Some financial securities have the property that their values depend upon the
value of another security. In this case, the former security is called a derivative
security of the later security which is called the underlying security for the given
derivative security. We often call a derivative security simply as derivative and
the corresponding underlying security simply as underlying. Derivative securities
are also called contingent claims.
Our aim here is to introduce some simple derivatives and illustrate them
through examples. In particular we shall describe a forward contract, a call op-
tion and a put option. These derivative securities in their pure form have stock
as underlying security. But sometimes it is possible to have a derivative whose
underlying is another derivative. So we can have options on futures contracts.
Thus a given financial entity can be derivative under some circumstances and
an underlying in other circumstances. In fact sometimes underlying may even be
interest rate and currency exchange rate which do not look like typical financial
securities. Our discussion here remains valid even when some other asset, other
than stock, is the underlying security. We shall also illustrate the application of
no arbitrage principle (Assumption 6) for their pricing. All these topics will be
further continued in the later chapters of the book.

Forward Contracts
A forward contract is a derivative security whose underlying is a stock and it has
the below given basic characteristics.
 Derivative Securities 11

(i) It is an agreement to buy or sell the stock at a specified future time, called
delivery date, for a fixed price F, called the forward price which is agreed at
t = 0. Obviously for the single period model, delivery date has to be t = 1.
(ii) An investor who agrees to buy the stock is said to enter into a long forward
contract or take a long forward position.
(iii) An investor who agrees to sell the stock is said to enter into a short forward
contract or take a short forward position.
(iv) No money is paid at t = 0 when a forward contract is exchanged.
(v) A forward contract guarantees that the stock will be bought (for long position)
or sold (for short position) for forward price F at the delivery date.
In general, the party holding the long forward contract will benefit if S(1) > F
and suffer a loss if S(1) < F. Therefore the pay-off of a long forward contract is
S(1) − F, which could be zero, positive or negative. For a short forward position,
the pay-off is obviously F − S(1).
The question which we have to answer is as follows: What should be the forward
price F? Though a formal theoretical formula will be derived in Chapter 2, here
we explain the logic through an example.
Example 1.4.1 Let B(0) = 100, B(1) = 110, S(0) = 50. Determine the forward
price F of a forward contract on the given stock.
Solution Here delivery is one period (from t = 0 to t = 1) which let us take as
one year. As B(0) = 100 and B(1) = 110, the nominal rate of interest is 10%. Since
S(0) = 50, the forward price F should at least be the amount which the holder of
the forward contract would have got by depositing Rs 50 in the bank for one year
with rate of interest 10%. Thus F should be at least Rs 55.
We shall now argue that F should be exactly Rs 55. If possible let F > 55. Then
at t = 0, we construct the portfolio P : (x = 1, y = − 12 , z = −1), where z denotes
the units of forward contracts in the portfolio. The interpretation of the portfolio
P so constructed is that at time t = 0 we borrow Rs 50 as (y = − 12 ), buy the stock
for S(0) = 50 as (x = 1) and enter into a short forward contract as (z = −1) with
forward price F and delivery date t = 1. This is the interpretation of x = 1, y = − 12
and z = −1.This gives
1
VP (0) = x S(0) + y B(0) = (1 × 50) + (− × 100) = 0.
2
Here it may be noted that in VP (0) there is no term corresponding to the short
forward contract (z = −1) because in forward contract there is no exchange of
money at the time when the contract is initialized, i.e. t = 0.
12 Financial Mathematics: An Introduction

Now at t = 1, we close the short forward position by selling the asset for Rs F
and close the risk free position by paying 21 B(1) = 12 × 110 = Rs 55. Therefore at
t = 1, VP (1) = F − 55, which is strictly positive as we have assumed that F > 55.
This clearly violates the no arbitrage principle. Hence F = 55.


Call and Put Options
Similar to forward contracts, the call and put options in their basic form are also
derivative securities whose underlying is stock. An option has the below given
characteristics.
(i) The owner of a call option gets the right to buy stock at a future date (called
maturity or expiration date) at a predetermined price (called strike or exercise
price). In a similar manner, the owner of a put option gets the right to sell
stock at maturity at a predetermined strike price.
(ii) The owner of the option (whether call or put) has the right without any
obligation. Thus at maturity the owner may not exercise the option if he/she
so wishes.
(iii) The person who has bought the option is called its holder and the person
who has sold the option is called its writer.
(iv) As explained in (ii) above, the holder of the option has the right but no
obligation whereas the writer of the option does have a potential obligation.
Thus in case of a call option, the writer is duty bound to sell the stock should
the holder choose to exercise the call option. Similarly in case of a put option,
the writer is duty bound to purchase the stock should the option is exercised
by the holder.
What we have described above are basic characteristics of a European call/European
put option. Here the option can be exercised at maturity only and not any time
before that. Thus for a single period model, maturity has to be t = 1. In contrast,
if the option can be exercised at any time before or at maturity, then it is called
an American option. Thus for a single period model, an American option can be
exercised at t = 1 or at t = 0. Unless we mention it specifically, for us an option
shall always mean a European option.
Since an option confers on its holder a right without any obligation, the holder
needs to pay some amount at t = 0 to have this right. This amount which is paid
at t = 0 when option is bought is called the premium or price of the option. Thus
we need to determine the price of a European call, American call, European put
and American put etc.
 Derivative Securities 13

Remark 1.4.1 At first, a call option resembles very much with a long forward
contract because both involve buying stock at a future date for a fixed price in
advance. But the holder of long forward contract is committed for buying the asset
for the fixed price whereas the holder of a call option has the right but no obligation
to do so. For this reason there is no premium in case of a long forward contract
but there is a premium in the case of an option.
Remark 1.4.2 In the case of derivatives e.g. forward contracts, options etc., there
are two parties which get together and set a rule by which one of two parties will
receive a payment from the other party depending upon the value of some financial
variables. Also profit of one party is the loss of other party. This is a typical
situation of a two person zero sum game. Therefore one can visualize many of the
derivative pricing and other finance related problems as game theoretic problems
and use the rich theory available there.
Example 1.4.2 Let B(0) = 100, B(1) = 110, S(0) = 100 and


120, with probability 0 · 8
S(1) =
80, with probability 0 · 2.

Consider a European call option C on the given stock for which the strike price
K = 100 and expiration date is 1 year. Determine C(0), i.e. the call price of this
option.
Solution Here S(1), the stock price at t = 1, takes two values. Denoting these by
u S(0) and d S(0), we obtain u = 1.2 and d = 0.8. Such a stock dynamics is called
a single period binomial model. Here


u S(0), with probability p
S(1) =
d S(0), with probability (1 − p) ,

with 0 < p < 1.

Also as B(1) = 110, B(0) = 100 we have nominal rate of interest r = 10% = 0 · 1.
It is important to note here that d < (1 + r) < u which in the context of binomial
model is a requirement for no arbitrage principle to hold. This result is proved
analytically in Chapter 3.
Though this dynamics looks extremely simple and probably unrealistic, it forms
the basis of other multi period and continuous pricing models.
To determine the call price C(0) we shall make use of the law of one price. For
this we shall first determine the pay-off of the given call, i.e its value C(1) at t = 1;
14 Financial Mathematics: An Introduction

and then construct a portfolio P : (x, y) of stock and bond such that VP (1) equals
C(1). Then the law of one price tells that the required price of the call has to be
the value of the portfolio at t = 0 ; i.e. C(0) = VP (0).
To compute the pay-off C(1) we argue that the holder will exercise the call
option only when the stock price S(1) is more than K and receive the amount
S(1) − K. If S(1) ≤ K, then it is not beneficial for the holder and therefore he/she
will use his/her right not to exercise the option. Thus

S(1) − K , if S(1) > K
C(1) =
0, if S(1) ≤ K ,

i.e.
C(1) = max ((S(1) − K), 0) = (S(1) − K)+ .
Here for x ∈ R, x+ denotes max(x, 0).
For our problem


20, with probability 0 · 8
C(1) =
0, with probability 0 · 2 .

Next we need to determine x and y such that for the portfolio P : (x, y), VP (1) =
C(1). This gives
x S(1) + y B(1) = C(1) ,
i.e.

120x + 110y = 20
80x + 110y = 0.

1 4
Solving above two equations we get x = and y = − . Hence by the law of one
2 11
price we obtain

C(0) = VP (0) = x S(0) + y B(0)

1 4
= (100) − (100)
2 11
= 13 · 6364.

Thus the buyer (holder) of the call option has to pay a premium of Rs 13.64
to the seller (writer) of the option. 

 Derivative Securities 15

Here it must be noted that the above arguments are not valid if the inequality
d < (1 + r) < u is not true. This is because if this inequality is not true, then no
arbitrage principle does not hold, and hence the law of one price is not valid.
Remark 1.4.3 At the time of writing the option S(1) is unknown. Therefore the
first problem of option pricing is to find how much the holder should pay the writer
for an asset (stock) worth (S(1)−K)+ at time t = 1? From the writers point of view
it is imperative to use the premium C(0) in constructing a portfolio (x, y) such that
it generates an amount (S(1) − K)+ at t = 1. This is the problem of hedging the
option and the portfolio (x, y) is called the replicating portfolio because in terms of
pay-off at t = 1 it replicates the given option. Thus while pricing a derivative, two
problems are being solved simultaneously namely the pricing problem for the holder
and hedging problem for the writer. Here again a game theoretic interpretation
could be given so as to interpret these two problems as primal-dual pair.
Remark 1.4.4 On first reading, option may look superfluous in a market because
we are replicating in terms of two basic securities of stock and bond. This is cer-
tainly true for a single period model. But for multi period and continuous time
models, this is not that simple because it involves re-balancing of the replicating
portfolio at every time instant. Therefore options and other derivatives are treated
as separate and independent financial instruments and not something which can
be expressed in terms of stock and bond.
Example 1.4.3 Let B(0), S(0), B(1), S(1), K and T be same as in Example 1.4.2.
Let the option P be a put option. Determine P(0).
Solution It is simple to argue that for the case of put option the pay-off at time
t = 1 is Max ((K − S(1)), 0) = (K − S(1))+ .
Thus


0, with probability 0 · 80
P(1) =
20, with probability 0 · 20.

This gives

120x + 110y = 0
80x + 110y = 20 ,
1 6
i.e. x = − and y = . Therefore,
2 11
1 6 

P(0) = x S(0) + y B(0) = − (100) + (100) = 4.54.
2 11
16 Financial Mathematics: An Introduction

Remark 1.4.5 In above examples we have seen that while determining C(0) and
P(0) we have never used the actual probabilities p = 0 · 8 and (1 − p) = 0 · 2. It is
probably not possible to explain its reason at this stage because it has to be explained
in terms of risk neutral probability measure (RNPM) which will be discussed in
Chapter 3 and also in other chapters on option pricing.

1.5 Some More Terminology

We have already introduced terms like bonds, stocks, portfolios, securities and
derivatives. In this section we first introduce some more terminology with respect
to bond and stock and then proceed to discuss time value of money and money
market account.
Security is a general term used for bonds, stocks and derivatives. Stocks are
also referred as equities. We can treat bonds as fixed income securities, because
at maturity they pay fixed amounts of money to their owners. Regular savings
accounts, certificates of deposits are other examples of fixed income securities.
Thus a bond is a security that gives its owner the right to a fixed predetermined
payment at a future predetermined date, called maturity. The amount of money
that a bond will pay in future is called nominal value, face value, par value or
principal . Thus bonds represent the general paradigm of risk free securities. In
fact we can treat money (cash) also as bond with zero interest rate and zero
(immediate) maturity.
There are bonds that involve only an initial payment (the initial price) and a
final payment (the final value or principal). Such bonds are called pure discount
bonds or zero coupon bonds. In contrast there are coupon bonds where the holder
of the bond gets periodic payments from the issuing institution during the life of
the bond. These periodic payments are specified percentage of the principal and
are called coupons. If the price at which the bond is sold is same as its nominal
value we say that the bond sells at par . The terms above par and below par are
defined in an obvious manner.
Sometimes the institution issuing the bond does not meet the promise and it
defaults in payment. This type of risk is called credit risk or default risk . The
bonds issued by the Government are considered to be free of default risk. Even
when there is no default risk, there is always inflation risk because future prices
of goods are uncertain. If the rate of interest does not take into consideration
the inflation risk then it is called nominal interest rate. If a bond guarantees a
payment that takes care of inflation level, it is called inflation indexed bond.
 Some More Terminology 17

In context of stock, we call the owner of the stock as stockholder . The profit
that the company distributes to the stockholders is called dividend. Dividend in
general is not known in advance because it depends upon the company’s profit and
its policy. The difference between the selling price and the initial price is called
capital gain or loss.

Time Value of Money

The way in which money changes its value in time is of utmost importance in
finance. Though a general scenario may be very complex, the below given questions
are simple to answer.
(i) What is the future value of an amount invested or borrowed today?
(ii) What is the present value of an amount to be paid or received at a certain
time in future?
We all have studied simple and compound interest in our schools. Let an initial
amount (called principal ) P be deposited into a bank account where it earns
interest. The future value of this investment consists of the principal P together
with all the interest earned for the period the money remained deposited. Let the
interest rate be r% per year and t be the time (expressed in years) for which the
money was deposited. Let V(t) denote the value of this investment at time t. Then
under the assumption that the principal P attracts simple interest rate only, we
have
V(t) = P(1 + rt).
We next consider the case of periodic compounding. Here again an amount P
is deposited in a bank account attracting a constant interest rate r% per year.
But in contrast to the simple interest case, here the interest earned is added
to the principal periodically. This is called periodic compounding which could
be annually, semi-annually, quarterly, monthly, weekly or daily basis. Therefore
interest will be attracted not only on the principal P but also by all the interest
earned so far. In this case we have


r tm
V(t) = P 1 + ,
m
where m refers to the period used in the compounding. Thus if it is compounded
semi-annually then m = 2 because there are two periods in an year. For quarterly
compounding m = 4 and for monthly compounding m = 12, etc. Here the number


r tm
1+ is called the growth factor.
m
18 Financial Mathematics: An Introduction

Example 1.5.1 How long will it take for a sum of Rs 800 attracting simple in-
terest to became Rs 830 if the rate of interest is 9% per year? Also compute the
return on this investment.
Solution From the formula for the simple interest we have
830 = 800(1 + 0 · 09 t) ,
which gives t = 0 · 417 year which is approximately 152 days. Further the return
on this investment will be
V(t) − V(0) 830 − 800
= = 0 · 0375 ,
V(0) 800
i.e. 3 · 75%.


Example 1.5.2 Which will deliver a higher future value after one year, a deposit
of Rs 1, 000 attracting interest at 15% compounded daily, or at 15·5% compounded
semi-annually?
Solution At 15% compounded daily the deposit will grow to


0 · 15 1×365
V(t) = 1000 1 +
Rs 1161 · 80 ,
365
after one year. If the interest is compounded semi-annually at 15 · 5%, the value
after one year will be


0 · 155 1×2
V(t) = 1000 1 +
Rs 1161 · 01 .
2
So interest at 15% compounded daily will give higher future value.


We next introduce the concept of continuous compounding. Let us recall the
formula

r tm
V(t) = P 1 + ,
m
which can be written as ⎡ ⎤rt
⎢⎢
 m ⎥⎥
⎢ r r ⎥⎥
V(t) = P ⎢⎢⎢⎢ 1 + ⎥⎥ .
⎣ m ⎦⎥

Therefore if we take the limit as m → ∞ we get V(t) = Pert . This is known as

continuous compounding. Here the growth factor is ert .
 Some More Terminology 19

Example 1.5.3 It is given that the future value of Rs 950 subject to continuous
compounding will be Rs 1, 000 after half year. Find the interest rate.
Solution The rate r satisfies

950 er×0.5 = 1000.

This gives r = 0.1026, i.e. 10.26%.



Let us recall the simple interest formula V(t) = P(1 + rt). Denoting the initial
deposit P as V(0) (value of this investment at t = 0) we have V(t) = V(0)(1 + rt),
which gives V(0) = V(t)(1 + rt)−1 . This last formula gives the present or discounted
value of V(t) and (1 + rt)−1 is the discount factor.
For the case of periodic compounding the present or discounted value V(0) of



r −tm r −tm
V(t) is given by V(t) 1 + and 1 + is the discount factor. In case
m m
of continuous compounding we have the present or discounted value as V(0) =
V(t)e−rt and the discount factor as e−rt . If the principal P is invested at time s
rather than at time 0, then the modifications are obvious.
Example 1.5.4 Find the present value of Rs 1, 00, 000 to be received after 100
years if the interest rate is assumed to be 5% throughout the whole period and (a)
daily or (b) annual compounding applies.
Solution Under daily compounding the present value is given by


0.05 −100×365
V(0) = 100000 1 +
Rs 674 · 03.
365
Under annual compounding the present value is given by
V(0) = 100000(1 + 0.05)−100
Rs 760 · 45. 

It is obvious that frequent compounding will produce a higher future value
than less frequent compounding if the initial deposit P and the interest rate r are
same. But for the general case, to compare two compounding methods we need to
compare the corresponding growth factor over the same base, say one year. This
motivates the below given definition.
Definition 1.5.1 (Comparison of compounding methods) Two compound-
ing methods are said to be equivalent if their corresponding growth factors over a
period of one year are same. If one of the growth factor exceeds the other, then the
compounding method having the higher growth factor is preferred over the other.
20 Financial Mathematics: An Introduction

Example 1.5.5 Is semi-annual compounding at 10% is preferable to monthly

compounding at 9%?.
Solution For semi-annual compounding at 10%, the growth factor over a period
of one year is


0·1 2
1+ = 1 · 1025.
2
But for monthly compounding at 9%, the growth factor is

12
0 · 09
1+
1 · 0938.
12
Therefore 10% semi annual compounding is preferable to 9% monthly compound-
ing.


There is another equivalent convention to compare the two compounding meth-
ods. Here rather than comparing the corresponding growth factors, we compare
the corresponding effective rate re as defined below.
Definition 1.5.2 (Effective Rate) For a given compounding method with in-
terest rate r, the effective rate re is the one which gives the same growth factor
over a one year period under annual compounding.
In view of Definition 1.5.2, for the case of periodic compounding, the effective
rate re satisfies the equation


r m
1+ = (1 + re ) .
m
If it is the case of continuous compounding then re satisfies the equation

e r = 1 + re .

In terms of effective rates, two compounding methods are equivalent if the

 
corresponding effective rates re and re are equal. if re > re , then the compounding
method with effective rate re is preferred over the compounding method with

effective rate re . This follows directly from Definitions 1.5.1 and 1.5.2.
Example 1.5.6 Find the effective rate re if it is semi-annual compounding at
10%.
 Some More Terminology 21

Solution We have
2
0.01
1+ = (1 + re ) ,
2
which gives re = 10 · 25%.


Example 1.5.7 Is daily compounding at 15% preferable to semi-annual com-
pounding at 15 · 5%?

Solution Let the corresponding effective rates be re and re . Then


0.15 365
1 + re = 1 +
1 · 1618 ,
365
and
2
 0.155
1 + re = 1 +
1 · 1610.
2

Thus re
16 · 18% and re
16 · 10%. Therefore daily compounding at 15% is
preferable.


We can also express future value in terms of effective rate re because of its
relationship with the growth rate. In fact we can show that

V(t) = P (1 + re )t

for all t ≥ 0. This formula holds both for the periodic compounding and con-
tinuous compounding, but it cannot hold for the case of simple interest. This is
because for the simple interest case, V(t) is a linear function of t, whereas for the
periodic/continuous compounding it is an exponential function of t.

Streams of Payments
An annuity is a sequence of finitely many payments of a fixed amount due at
equal time intervals. Let payment of an amount C are to be made once a year for
n years, the first one is due a year hence. If we assume that annual compounding
applies, then the present value of such a stream of payments is given by
C C C C
+ + + . . . + .
(1 + r) (1 + r)2 (1 + r)3 (1 + r)n
If we denote
22 Financial Mathematics: An Introduction

1 1 1
PA(r, n) = + + ... + ,
(1 + r) (1 + r) 2 (1 + r)n
then the present value of the given stream of payments is (C × PA(r, n)). Here
the number PA(r, n) is called the present value factor for an annuity. Obviously

1 − (1 + r)−n
PA(r, n) = .
r
We define perpetuity as an infinite stream of payments of a fixed amount C
occurring at the end of each year. Therefore the perpetuity can be computed by
C
limn→∞ PA(r, n) × C which equals .
r

Example 1.5.8 Consider a loan of Rs 1, 000 to be paid back in 5 equal instalments

due at yearly intervals. The instalments include both the interest payable each year
calculated at 15% of the current outstanding balance and the repayment of the
fraction of the loan. Determine the amount of each instalment.
Solution We can look the loan as equivalent to annuity from the point of view
of the lender. Therefore if C is the amount of each instalment then

1000 = C × PA(0 · 15, 5).

But
1 − (1 + 0 · 15)−5
PA(0 · 15, 5) = .
0 · 15
This gives C
Rs 298 · 32.


Example 1.5.9 Let the interest rate be 18% and a person can afford to pay
Rs 10, 000 at the end of each year for next 10 years. Determine the amount which
the person can borrow?
Solution The person can borrow the amount PA(0 · 18, 10) × 10000 = Rs 44, 941.


Money Market
By the money market we mean default free securities. Thus money market
consists of risk free bonds like treasury bills and notes and other financial securities
which promise the holder a sequence of guaranteed future payments. Risk free here
means default free i.e. these payments will be delivered with certainty. However in
 Some More Terminology 23

practice the risk, even for risk free securities, cannot be completely avoided since
the market prices of such securities may fluctuate unpredictably and there may
be default in payments. But in our discussion here we shall assume that it is a
default free, and hence risk free situation.
As mentioned earlier, the simplest case of a bond is zero coupon bond. Here the
issuing institution (a government, a bank, or a company) promises to exchange
the bond for a certain amount of money F, called the face value, on a given day T,
called the maturity date. A typical zero coupon bond has the face value of Rs 100
(or some other round figure) and maturity date as one year (or multiple of a year).
If an investor buys a bond, he/she becomes its holder and its effect is that he/she
is lending money to the bond writer (a government, a bank, or a company). Here
it must be noted that investor need not always be an individual, it could also be
a financial institution.
Given the interest rate, the present value of such a bond can be computed
easily. Thus if the face value of a bond is F = Rs 100 and maturity date T = 1
year, then for 12% compounding annually we have V(0) = F(1 + r)−1
Rs 89 · 29.
For simplicity, we shall assume that the face value F of the given bond is
Rs 1. Such bonds are called unit bonds. Typically a bond can be sold any time
prior to maturity at the prevailing market price. Let B(t, T) denote the price of
the given unit bond at time t when the maturity is T. Thus B(T, T) = 1 and
B(t, T) = e−r(T−t) . Here we have assumed that r remains constant throughout the
period up to maturity.
But in reality the bonds are freely traded in the market and their prices are
determined by market forces. Therefore the market interest rate (depending upon
the market price of the given bond) may be different from the implied interest
F
rate r = − 1 given by the formula V(0) = F(1 + r)−1 . In fact the implied
V(0)
interest rate should not be constant as it should depend on the trading time t as
well as on the maturity time T. Not only this, but also the bond prices B(t, T) are
determined by (random) market forces. Therefore the implied interest rate should
be taken as a random process r(t, T). We shall have occasion to discuss this aspect
in a later chapter.
As our bond is a unit bond we can treat B(0, T) as the discount factor and
B(0, T)−1 as the growth factor. This observation allows to compute time value of
money without restoring to the corresponding interest rates.

Coupon Bonds
In zero coupon bonds, there is just a single payment at maturity. But in coupon
24 Financial Mathematics: An Introduction

bonds, there is a sequence of payments. These payments consist of the face value
due at maturity, and coupons paid regularly, say annually, semi-annually or quar-
terly. The last coupon becomes due at maturity only. If the interest rate r is
assumed to be constant throughout, then it is simple to compute the price of a
coupon bond by discounting all the future payments. We explain this process with
the help of the next example.

Example 1.5.10 Find the price of a bond with face value Rs 100 and Rs 5 annual
coupons that matures in 4 years, if the continuous compounding rate is 8%.
Solution The price of the given bond is obtained as

5e−0.08 + 5e−2×0.08 + 5e−3×0.08 + 105e−4×0.08
Rs 89 · 06.

Here we should note that as maturity is 4 years, the payments have to be

Rs 5, 5, 5 and105.


Money Market Account
An investment in the money market can be realized by means of an investment
bank which buys and sells bonds on behalf of its customers. The risk free position
of an investor is given by the level of his/her account with the bank. We can think
of this account as a tradable asset, since the bond themselves are tradable. A long
position in the money market involves investing the money and a short position
means borrowing money.
Thus an investment of amount A(0) in the money market amounts to purchas-
A(0)
ing bonds at t = 0. At time t, the value of each bond will be
B(0, T)

B(t, T) = e−r(T−t) = ert e−rT = ert B(0, T).

As a result, at time t, t ≤ T, the investment A(0) will become

A(0)
A(t) = B(t, T) = A(0)ert .
B(0, T)
The investment in bond has a finite time horizon, say T. To extend the position
in the money market beyond T, the investor can invest the amount received at
 
maturity into a newly issued bond at time T, but maturing at T (T > T). We
explain this by the below given example.
 Summary and Additional Notes 25

Example 1.5.11 Suppose Rs 1 is invested in zero coupon bonds maturing after

one year. At the end of each year the proceeds are re-invested in new bonds of the
same kind. How many bonds will be purchased at the end of year n (n ≥ 1)? It may
be assumed that the interest rate is constant at r % and continuous compounding
is followed.
1
Solution At t = 0, the investor buys = er bonds. At time t = 1, increases
B(0, 1)
er
his/her holdings to = e2r bonds. Therefore at the end of year n, the investor
B(1, 2)
will purchase e(n+1)r , one year bonds.



1.6 Summary and Additional Notes

• The date of March 29, 1900 should be considered as the origin of Financial
Mathematics. On that day, a French postgraduate student, Louis Bachelier
successfully defended his thesis Théorie de la spéculation. Bachelier presented
a model of Brownian motion (see Chapter7) while deriving the evolution of
Paris asset prices. This work of Bachelier was neglected by the economic com-
munity for many years but probabilists took great interest in this development.
Kolmogorov’s analytical theory of Markov process has its origin in the theory
proposed by Bachelier. The economic community also later recognized Bache-
lier’s contribution and named International Finance Society after him.
Bachelier’s thesis topic at that time was out of ordinary. The appropriate topic
of research in that era were on the theory of functions. Therefore this topic was
not initially accepted as a research topic. For these and other historical details
on Bachelier’s work we may refer to Vassiliou [144].
• In this chapter we have presented an overview of the topic of financial math-
ematics. Here our main aim has been to introduce the basic terminology of
finance and motivate the readers to the art of mathematical modeling to deal
with the problems arising in this area. In the process, we have also introduced
the readers to the three core topics of the book, namely option pricing, portfolio
optimization and interest rate theory through elementary illustrative examples.
• Though the illustrative examples presented here require only an elementary
knowledge of algebra, calculus and probability theory, it has been emphasized
that real life problems in finance are extremely complex. Therefore we need
to get ourselves equipped with certain advanced topics in mathematics, e.g.
26 Financial Mathematics: An Introduction

stochastic process and stochastic calculus, numerical optimization, and partial

differential equations etc.
• The importance of no arbitrage principle and the law of one price is illustrated
repeatedly in various illustrative examples.
• It has been remarked that the standard deviation is not always a good measure
of risk for studying problems in the finance area and therefore the definition of
risk becomes significant.
• The relevant texts for this chapter are Capinski and Zastawniak [25], Hull [65]
and Luenberger [85].

1.7 Exercises
Exercise 1.1 Let B(0) = Rs 100, B(1) = Rs 110 and S(0) = Rs 80. Also let


100, with probability p = 0.80
S(1) =
60, with probability p = 0.20.

Design a portfolio with initial wealth of Rs 100, 000, split fifty-fifty between stock
and bond. Compute the expected return and the risk of the portfolio so constructed.
Exercise 1.2 Let B(0), B(1) S(0) and S(1) be as specified in Exercise 1.1. Also let
C and P respectively be a European call and European put with K = Rs 100 and
T = 1 year.
(i) Determine C(0) and P(0).
(ii) Find the final wealth of an investment with initial capital of Rs 900 being
invested equally in the given stock, the given call and the given put.
Exercise 1.3 In the data of Exercise 1.2, let there be a minor change. Instead of
B(1) = Rs 110, let B(1) = Rs 120. Find the replicating portfolio (x, y) for the call C.
Are we justified in taking C(0) = x S(0) + y B(0)? Give appropriate mathematical
reason for your answer. Will your answer change if instead of call C we take the
put P ?
Exercise 1.4 Spot an arbitrage opportunity (if it exists) in the following situation.
Suppose that a dealer A offers to buy British pounds in an year from now at a
rate of Rs 79 a pound, while dealer B would sell British pounds immediately at a
rate of Rs 80 a pound. Assume that a rupee can be borrowed at an annual rate of
4% and a British pound can be invested in a bank at 6% annual interest.
 Exercises 27

Exercise 1.5 Let B(0) = Rs 100, B(1) = Rs 112, S(0) = Rs 34 and T = 1. Find
the forward price F. Also find an arbitrage opportunity if F is taken to be Rs 38.60.
Exercise 1.6 Assume that a person can afford to pay Rs 10, 000 at the end of
each year. How much the person can borrow if the interest rate is 12% and he/she
wishes to clear the loan in 5 years?
Exercise 1.7 What will be the difference between the value after 1 year of
Rs 10, 000 deposited at 10% compounded monthly and compounded continuously.
How frequently should the periodic compounding be done for the difference to be
less than Rs 1?
Exercise 1.8 Consider a loan of Rs 10, 000 to be paid back in 10 equal instalments
at yearly intervals. The instalments include both the interest payable each year
calculated at 15% of the current outstanding balance and repayment of a fraction
of the loan. Determine the amount of each instalment. What is the amount of
interest included in each instalment and what is the outstanding balance of the
loan after each instalment is paid?
Exercise 1.9 Find the return on a 75-day investment in zero-coupon bonds if
B(0, 1) = Rs 0.89.
Exercise 1.10 After how many days will a bond purchased for B(0, 1) = Rs 0.92
produces a 5% return?
2
Forward and Futures Contracts

2.1 Introduction
A forward contract is probably the simplest of all derivative securities. It has a
simple pricing mechanism and has wide applications, particularly in commodity
and foreign exchange markets. We had already initiated some discussion on for-
ward contracts in Chapter 1 where we had also explained the pricing methodology
through a simple example.
Though futures contracts are very much in the spirit of forward contracts, they
have specially been designed to standardize these contracts so as to eliminate the
risk of default by the party suffering the loss. For this, a process called marking to
market is conceptualized which requires an individual to open a margin account
which is managed by an organized clearing house/exchange.
This chapter continues our earlier discussion on forward contracts and explains
various concepts related with the working of futures contracts. We also describe
very briefly another derivative security called swap which is again very popular in
commodity and foreign exchange markets.

2.2 Forward Contract

Let us recall the definition of a forward contract from Chapter 1 and note that it is
an agreement between two investors (also called parties) to buy or sell a risky asset
at a specified future time (called the delivery date) for a price F (called the forward
price) fixed at the present time (say t = 0). The party which agrees to buy the asset
is said to enter into a long forward contract or to take a long forward position. The
party which agrees to sell the asset is said to enter into a short forward contract
or to take a short forward position. There is no payment of money by either party
30 Financial Mathematics: An Introduction

when the agreement for the forward contract is made. Some typical examples of
forward contract could be a farmer wishing to fix the sale price of his/her crops in
advance, an importer arranging to buy foreign currency at a fixed rate in future, a
fund manager wishing to sell stock for a fixed price in future or a country wishing
to import commodities (like wheat, sugar, oil etc.) from another country at a fixed
rate at some specified future date. In these situations, forward contracts become
very handy because they provide an opportunity to hedge against the unknown
future price of the underlying risky asset.
Let t = 0 denote the time when the two parties enter into the specific forward
contract agreement and t = T be the delivery date. Let the agreed forward price be
F(0, T). Here the two arguments in F denote the time t = 0 and t = T respectively.
In case the context is clear and there is no ambiguity, we write F(0, T) simply as
F.
Let us now analyze the two scenarios, namely F(0, T) < S(T) and F(0, T) > S(T).
Let us first consider the case F(0, T) < S(T). In this case the party having the long
forward contract will benefit because it can buy the asset for F(0, T) and sell the
same for the market price S(T), making a profit of S(T) − F(0, T). But the other
party which has taken a short forward position will suffer loss of S(T) − F(0, T)
because it will have to sell below the market price. Fig. 2.1 clearly exhibits these
two positions.

Pay−off0
1
Pay−off 1
0
0
1
0
1
111111111111111
000000000000000
0
1 0
1
000000000000000
111111111111111
0
1 0
1
000000000000000
111111111111111
0
1 0
1
0
1
000000000000000
111111111111111
0
1 0
1
000000000000000
111111111111111
0
1
000000000000000
111111111111111
0
1
000000000000000
111111111111111
0
1
000000000000000
111111111111111
0
1
0
1
0
1
0
1
0
1
01
000000000000000
111111111111111
00
11 111111111
000000000
0
1 0
10
1 0
1
000000000000000
111111111111111
00
11
S(T)−F(0,T) 0
1
000000000000000
111111111111111 01
10 0
1
0
1
000000000000000
111111111111111 0
1
0
1 0
1
0
1
000000000000000
111111111111111
0
1 0
1 0
1
000000000000000
111111111111111
0
1 0
1 0
1
0
1
000000000000000
111111111111111
0
1
000000000000000
111111111111111 0
1 0
1
0
1
000000000000000
111111111111111 0
1 0
1
0
1
000000000000000
111111111111111 0
1
0
1 0
1
0
1
000000000000000
111111111111111
0
1 0
1 0
1
0
1
000000000000000
111111111111111
11111111111111111
00000000000000000
0
1
0
1 1
0
000000000000000
111111111111111
000000000000000
111111111111111
0
1
000000000000000
111111111111111
0
1 F(0,T)
0
1
11
00
S(T) Price
0
1
0
1
0
1
0
1
11
00
1111111111111111
0000000000000000
F(0,T) S(T) 1
0
Price
000000000000000
111111111111111
0
1 0
1
000000000000000
111111111111111
0
1 0
1
000000000000000
111111111111111
0
1
000000000000000
111111111111111
0
1
000000000000000
111111111111111
0
1
000000000000000
111111111111111
0
1
000000000000000
111111111111111
11
00
00
11
0
1
0
1
F(0.T)−S(T) 1
0
0
1
0
1
000000000000000
111111111111111 0
1
0
1
000000000000000
111111111111111
0
1 0
1
000000000000000
111111111111111
0
1 0
1
000000000000000
111111111111111
0
1 0
1
0
1 0
1

Fig. 2.1. Pay-off of a forward contract

If F(0, T) > S(T) then these arguments will simply get reversed. In this case the
 Forward Price Formula 31

party taking a short forward position will benefit and make a profit of F(0, T)−S(T)
whereas the other party (which has taken a long forward position) will make a
loss of F(0, T) − S(T) and analogue of Fig. 2.1 can be drawn for this case as well.
Therefore in either case the pay-off of a long forward contract at delivery is
S(T) − F(0, T) and that of a short forward contract is F(0, T) − S(T). Sometimes
the contract may be initiated at time t < T rather than t = 0. In this scenario the
pay-offs of a long forward position and a short forward position are S(T) − F(t, T)
and F(t, T) − S(T) respectively where F(t, T) denotes the forward price.

2.3 Forward Price Formula

We now look into the problem of determining the price of the forward contract,
i.e. the forward price F(0, T). We shall make use of the no arbitrage principle and
follow the methodology indicated in Chapter 1 to get the next theorem.
Theorem 2.3.1 (Forward Price Formula) Let the price of an asset at t = 0
be S(0). Then the forward price F(0, T) is given by
S(0)
F(0, T) = , (2.1)
d(0, T)
where d(0, T) is the discount factor between t = 0 and t = T.
Proof. If possible let F(0, T) > (S(0)/d(0, T)). Then at t = 0 we construct the
portfolio P as follows
(i) borrow S(0) amount of cash until t = T,
(ii) buy one unit of the underlying asset on the spot market at price S(0) (if the
underlying is the stock then it will mean buying one share of stock for amount
S(0)),
(iii) take a short forward position with delivery at t = T and forward price as
F(0, T).
This gives the value of the portfolio P at t = 0 as

VP (0) = −S(0) + S(0) = 0.

Now at t = T, we sell the asset for F(0, T) and pay the amount (S(0)/d(0, T)) to
clear the loan with interest and thereby close the short position. This gives

VP (1) = F(0, T) − (S(0)/d(0, T)) ,

32 Financial Mathematics: An Introduction

which is strictly positive under the stated assumption. Thus there is a strictly
positive amount of risk free profit with zero net investment. This is contrary to
the no arbitrage principle.
Next suppose that F(0, T) < (S(0)/d(0, T)). In this case, at t = 0, we construct
the portfolio Q as follows
(i) short sell one unit of the underlying asset (thus if the underlying asset is stock,
then we borrow one share of stock and sell the same in the market with all
conditions and procedure as specified in short selling),
(ii) invest the proceeds (i.e. the amount S(0) received by selling one unit of the
asset) at the risk free rate up to t = T,
(iii) enter into a long forward contract with delivery as t = T and forward price
F(0, T).
This gives
VQ (0) = S(0) − S(0) = 0.
Now at t = 1, we
(i) cash the risk free investment with interest, i.e. collect cash (S(0)/d(0, T)) ,
(ii) buy the underlying asset for F(0, T) using the long forward position,
(iii) close the short position on the underlying asset by returning the borrowed
unit of asset to the owner.
This gives
VQ (T) = (S(0)/d(0, T)) − F(0, T) ,
which is strictly positive. This again allows a strictly positive risk free profit
with zero net investment contradicting the no arbitrage principle. Thus F(0, T) =
S(0)/d(0, T).


Remark 2.3.1 For the case of constant interest rate r being compounded contin-
uously, we have d(0, T) = e−rT and hence

F(0, T) = S(0) erT . (2.2)

Remark 2.3.2 If the contract is initiated at some intermediate time t, 0 < t < T,
then d(t, T) = e−r(T−t) and
F(t, T) = S(t) er(T−t) , (2.3)
where r is same as in Remark 2.3.1.
 Forward Price Formula 33

Remark 2.3.3 In terms of zero coupon bond prices, the forward price formula
2.1 becomes
S(0)
F(0, T) = , (2.4)
B(0, T)
which is more convenient to use as it does not require the assumption that the
interest rate r is constant.
It is a matter of common knowledge that holding of physical assets like gold,
sugar, oil, wheat etc. entail inventory carrying costs, such as rental for storage
and insurance fees etc. These costs affect the theoretical forward price formula
as given at (2.1). The standard forward price formula (2.1) can be generalized
in several ways to take into consideration the costs of carry and also to include
dividends. We have the below given result in this regard.
Theorem 2.3.2 (Forward Price Formula with Carrying Costs) Let an as-
set carry a holding cost of c(i) per unit in period i (i = 0, 1, 2, . . . (n − 1)). Also let
at t = 0, the price of this asset be S(0) and short selling be allowed. Then

S(0)  c(i)
n−1
F(0, T) = + , (2.5)
d(0, T) i=0 d(i, n)
where delivery date is t = T and between t = 0 and t = T there are n periods which
have appropriately been identified as per the given context.
Next we consider the case when the underlying asset (e.g. stock) is dividend
paying.
Theorem 2.3.3 (Forward Price Formula with Dividend) Let an asset be
stored at zero cost and also sold short. Let the price of this asset at t = 0 be S(0)
and a dividend of Rs div be paid at time τ, 0 < τ < T. Then
[S(0) − (div) (d(0, τ))]
F(0, T) = . (2.6)
d(0, T)
For the case of constant interest rate r being compounded continuously, the above
formula becomes  
F(0, T) = S(0) − (div) e−rτ erT . (2.7)
Sometimes the asset (stock) may pay dividends continuously at a rate of rdiv >
0. This is called continuous dividend yield. If the dividends are reinvested in the
stock, then an investment in one share held at t = 0 will become e(rdiv )T shares at
t = T. This is very similar to continuous compounding and therefore e−(rdiv )T shares
at t = 0 will give one share at t = T. This observation leads to Theorem 2.3.4.
34 Financial Mathematics: An Introduction

Theorem 2.3.4 (Forward Price Formula with Continuous Dividend ) Let

an asset be stored at zero cost and also sold short. Let the price of the asset at
t = 0 be S(0) and the asset pay dividends continuously at a rate of rdiv . Then

F(0, T) = S(0) e(r−rdiv )T . (2.8)

Theorems 2.3.2, 2.3.3 and 2.3.4 are not proved here as their proofs can be
constructed analogous the proof of Theorem 2.3.1. Interested readers may refer to
Luenberger [85] and Capinski and Zastawniak [25] in this regard.
Forward contracts in foreign currency market are very common and we wish
to derive a formula for determining the price of the same. To be specific, let us
consider the two currencies as British pound and US dollars with the latter as the
underlying. The readers may imagine a British importer of US goods requiring
US dollars after t = T. So this British importer may think of taking a forward
contract on US dollars with delivery as t = T to hedge against the fluctuating
exchange rate of British pound versus US dollars.
Let at t = τ, the buying and selling exchange rate be: 1 British pound = P(t)
US dollars. Also let the risk free interest rates for investments in British pounds
and US dollars be rGBP and rUSD respectively. Then the forward price is given by

F(0, T) = P(0) e(rUSD −rGBP )T . (2.9)

The above formula gives the agreed exchange rate at t = T, i.e. at t = T, the
British importer will be able to buy F(0, T) US dollars for one British pound.
To prove formula (2.9), let us consider two strategies as described below.

Strategy A
Invest P(0) US dollars at the rate of rUSD until t = T.

Strategy B
Buy one British pound for P(0) US dollars, invest it until t = T at the rate of
rGBP , and take a short position in exp(rGBP · T) British pound forward contract
with delivery as t = T and forward price as F(0, T). This gives

P(0) e(rUSD )T = F(0, T) e(rGBP )T ,

and hence the result.

Example 2.3.1 The current price of gold is Rs 18, 000 per 10gm. Assuming a
constant interest rate of 8% per year compounded continuously, find the theoretical
price of gold for delivery after 9 months.
 Forward Price Formula 35

Solution Here S(0) = 18, 000, r = 8% per year and T = 9 months = 3/4 year.
Hence by the forward price formula

F = S(0) erT
= (18000) e(0.08×3/4)
= (18000) e0.06
= Rs 19113.06.


Example 2.3.2 Find the forward price of a non-dividend paying stock traded to-
day at Rs 100, with the continuously compounded interest rate of 8% per year, for
a contract expiring seven months from today.
Solution We have S(0) = 100, r = 8% per year and T = 7/12 year. This gives

F = 100 e(0.08)(7/12) = Rs 104.78.



Example 2.3.3 The current price of sugar is Rs 60 per Kg and its carrying cost
is 10 paisa per Kg per month to be paid at the beginning of each month. Let the
constant interest rate r be 9% per annum. Find the forward price of the sugar
(Rupees per Kg) to be delivered in 5 months.
Solution The interest rate is (.09)/12 = .0075 per month. Therefore the reciprocal
of one month discount rate for any month is (1.0075). Therefore we have


5 
F(0, 5) = (1.0075)5 (60) + (1.0075)i (.1)
i=1
= Rs 62.79.


Example 2.3.4 Let the price of a stock on 1st April 2010 be 10% lower than it
was on 1st January 2010. Let the risk free rate be constant at r = 6%. Find the
percentage drop of the forward price on 1st April 2010 as compared to the one on
1st January 2010 for a forward contract with delivery on 1st October 2010.
Solution It is convenient to take 1st January 2010 as t = 0. Then 1st October
2010 is 9 months, i.e. 3/4 year. Thus t = 3/4. Also t = τ is 1st April 2010, i.e.
3/12 = 1/4. Therefore using Theorem 2.3.1 we get
36 Financial Mathematics: An Introduction

F(0, 3/4) = S(0) e(0.06×3/4)

F(1/4, 3/4) = S(0)(0.9) e(0.06×2/4) .

Here it may be noted that on 1st April 2010 the price of the stock is 10% lower
than that on 1st January 2010, i.e. (0.9)S(0). Now

F(0, 3/4) − F(1/4, 3/4) e(0.06×3/4) − (0.9) e(0.06×1/2)

=
F(0, 3/4) e(0.06×3/4)
= 0.1134.

Therefore the percentage drop in the forward price is 11.34%.



Example 2.3.5 An Indian importer wants to arrange a forward contract to buy
US dollars in half a year. The interest rates for investment in Indian rupees and
US dollars are rINR = 6% and rUSD = 4% respectively. Also the current rate of
1
exchange is Rs 46 to a dollar, i.e. dollars to a rupee. Find the forward price
46
of dollar-rupee exchange rate.
Solution Here US dollars play the role of underlying. Therefore from (2.9) we get

F(0, 1/2) = (1/46) e(0.5×(0.06−0.04))

= (1/46) e(0.5×0.02)
= 0.022,
1
dollars to a rupee, i.e. = 45 · 54 rupees to a dollar.
(0 · 022)



2.4 The Value of a Forward Contract

Let a forward contract be initiated at t = 0 with delivery t = T. Also let F(0, T)
be the forward price of this contract. Consider an intermediate time 0 < τ < T
and let F(τ, T) be the forward price of the contract initiated at t = τ with delivery
t = T.
Thus we have two forward contracts; one initiated at t = 0 and other initiated
at t = τ, both having the same delivery date, namely t = T. Then F(0, T) and
F(τ, T) are the respectively the forward prices of these two contracts. Now as time
 The Value of a Forward Contract 37

progress, the value of the forward contract initiated at t = 0 will be changing. Let
at t = τ, its value be f (τ). Is there a relationship connecting f (τ), F(0, T) and
F(τ, T)? The below given theorem is precisely the answer to this question.
Theorem 2.4.1 (Value of a Forward Contract) Let f (τ), F(0, T) and F(τ, T)
be as explained above. Then

f (τ) = (F(τ, T) − F(0, T)) d(τ, T) , (2.10)

where d(τ, T) is the risk free discount factor over the period t = τ to t = T.
Proof. If possible let

f (τ) < (F(τ, T) − F(0, T)) d(τ, T).

Now at t = τ we construct a portfolio P as follows

(i) borrow the amount f (τ) to enter into a long forward contract with forward
price F(0, T) and delivery t = T,
(ii) enter into a short forward position with forward price F(τ, T) for which there
is no cost as per the definition of forward contract.
Then the value of the portfolio P at t = τ is given by

VP (τ) = 0.

Next at time t = T, we
(i) close the forward contracts by collecting (or paying, depending upon the sign of
pay-offs) the amounts S(T) − F(0, T) for the long forward position and −S(T) +
F(τ, T) for the short forward position,  
(ii) pay back the loan amount with interest, i.e. amount f (τ)/d(τ, T) .
Therefore the value of the portfolio at t = T is
 
VP (T) = F(τ, T) − F(0, T) − f (τ)/d(τ, T) ,

which is strictly positive and risk free. This violates no arbitrage principle.
Let us now consider the second case, namely f (τ) > [F(τ, T) − F(0, T)] d(τ, T).
In this case our strategy is to construct a portfolio Q at t = τ as follows
(i) sell the forward contract which was initiated at t = 0 for the amount f (τ),
(ii) invest this amount f (τ) risk free from t = τ to t = T,
38 Financial Mathematics: An Introduction

(iii) enter into a long forward contract with delivery time t = T and forward price
as F(τ, T).
Then VQ (τ) = 0 and VQ (T) is given by
f (τ)
VQ (T) = + (S(T) − F(τ, T)) + (F(0, T) − S(T))
d(τ, T)
f (τ)
= + F(0, T) − F(τ, T) ,
d(τ, T)
which is strictly positive and risk free. This is not possible due to no arbitrage
principle.


Example 2.4.1 Let at the beginning of the year, a stock be sold for Rs 45 and risk
free interest rate be 6%. Consider a forward contract on this stock with delivery
date as one year. Find its forward price. Also find its value after 9 months if it is
given that the stock price at that time turns out to be Rs 49.
Solution From (2.2), the initial forward price F(0, 1) is given by

F(0, 1) = S(0) erT = 45 e0.06 = Rs 47.78.

Also it is given that S(9/12) = 49 and hence

F(9/12, 1) = 49 exp ((·06)(3/12)) = Rs 49 · 74.

Therefore by Theorem 2.4.1, the value of the forward contract after 9 months is

f (9/12) = [F(9/12, 1) − F(0, 1)] exp((−0 · 06)(1 − (9/12)))

= 1 · 93.


Example 2.4.2 Consider the data of Example 2.4.1 and assume that a dividend
of Rs 2 is being paid after 6 months. Find the forward price of the contract and
also its value after 9 months.
Solution By (2.7), the initial forward price F(0, 1) is given by

F(0, 1) = [S(0) − (div)e−rt ] erT

 
= 45 − 2e−0.06(1/2) e·06
= Rs 45 · 72.
 Futures Contract 39

Also by (2.3)
F(9/12, 1) = S(9/12) e(0·06)(1−9/12) = Rs 49 · 74.
Hence by Theorem 2.4.1,
f (9/12) = Rs 3 · 96.



2.5 Futures Contract

It is obvious that when two parties enter into a forward contract, one of them is
certainly going to lose money. Therefore a forward contract is always exposed to
a risk of default by the party suffering a loss. To eliminate this risk of default,
futures contracts have been introduced which are managed by an organized ex-
change. Thus even though futures contracts are very much in the spirit of forward
contracts, their working is entirely different. Here the role of an organized ex-
change becomes very important because individual contracts are made with the
exchange, which itself becomes counter party to both long and short trades. There-
fore individuals themselves do not need to search for an appropriate counterparty
because this job is taken care off by the exchange itself. The exchange also takes
appropriate measures so as to eliminate the risk of counterparty default. This is
done by a process called marking to market. Before we explain the working of the
process, we list the basic features of a futures contract.
(i) Similar to a forward contact, a futures contract also has an underlying asset
(e.g. stock) and a delivery date, say time t = T.
(ii) In addition to the (underlying) asset prices, the market also dictates the fu-
tures price f (t, T) at discrete time steps, say f (n, T) for n = 0, 1, 2, . . . with
nt ≤ T. In practice these discrete time steps may refer to day1, day2 etc.
(iii) f (0, T) is known but f (n, T) for n = 1, 2 . . . with nt ≤ T are unknown and
are treated as random variables. Also the (underlying) asset prices S(n) for
n = 1, 2 . . . with nt ≤ T are random variables.
(iv) As in the case of a forward contract, there is no cost involved in initiating
a futures contract, but there is a major difference in the cash flow. A long
forward contract involves just a single payment S(T) − F(0, T) at the delivery
time t = T, whereas a futures contract involves a random cash flow at each
time step n = 0, 1, 2 . . . with nt ≤ T. Thus the holder of a long futures position
will receive the amount f (n, T) − f (n − 1, T) if positive, or will have to pay if
it is negative. For a short futures position, exactly opposite payments apply.
These payments are managed by a clearing house for the futures market.
40 Financial Mathematics: An Introduction

(v) The futures price satisfy the condition that f (t, T) = S(T). This condition has
to hold because the futures cost of immediate delivery of goods has to be the
market price or spot price.
(vi) It costs nothing to close, open or alter a futures position at any time step
between t = 0 and t = T. This condition can be met if at each time step
n = 0, 1, 2 . . . with nt ≤ T, the value of futures position is zero. For n ≥ 1, this
value is computed after marking to market.
But what is the physical meaning of futures price f (n, T)? Suppose a forward
contract initiated at time t = 0 has forward price F(0, T), where time t = T is the
delivery date. Let the forward price for a new contract initiated at time t = 1 with
delivery date time t = T be F(1, T). Now the clearing house comes into picture.
At the second day it revises all earlier contracts to the new delivery price F(1, T)
and accordingly an investor which holds a long forward contract initiated at t = 0
receives or pays the difference of two prices depending upon if the change in price
reflects a loss or gain. So if F(1, T) > F(0, T), then an investor holding a long
forward contract receives F(1, T) − F(0, T) from the clearing house because at the
delivery he/she has to pay F(1, T) rather than F(0, T). Continuing in this manner
and assuming that the investor stays until maturity, the profit/loss pay-off of a
long futures position will be [F(1, T) − F(0, T)] + [F(2, T) − F(1, T)] + . . . + [F(T, T) −
F(T − 1, T)], which equals S(T) − F(0, T) because F(T, T) = S(T).
The above discussion demonstrates that the pay-off of a futures contract is
the pay-off of the corresponding forward contract, except that the pay-off is paid
throughout the life of the contract rather than at maturity. However there is
no requirement that the investor of a futures contract has to stay till maturity.
He/she can come out of the contract any time by taking the opposite position in
the contract with the same maturity. Thus f (n, t), n = 0, 1, 2, . . . with nt ≤ T are
essentially the forward prices as perceived by the market.
Here it may be noted that f (n, T) are not known and are taken as random
variables dictated by the market. This is because interest rate r is rarely constant
and, in general, it is stochastic in nature.
This process of adjusting the contract is called marking to market. Here an
individual is required to open a margin account with the clearing house. This
account must have a specified amount of cash for each futures contract. In practice
it is about 5 to 10% of the value of futures contract. The margin account is
compulsory for all contract holders whether they have long or short position.
If the price of futures contract increased that day, then the parties having long
position receive an amount which equals (change in price)× (the contract quantity)
which is deposited in their account. The short parties loose the same amount and
 Futures Contract 41

therefore this amount is deducted from their account. This process is termed as
marking of accounts to the market which is carried out at the end of each trading
day. This guarantees that both parties of the contract cover their obligations.
Thus each margin account value fluctuates from day to day according to change
in futures price. At the delivery date, delivery is made at the futures contract price
at that time which may be different from the futures price when the contract was
first initiated.
Margin accounts serve a dual role. They serve as accounts to collect or pay
out daily profits and also guarantee that contract holders do not default on their
obligations. If the value of a margin account drops below a pre-defined margin level
(in practice about 75% of the initial margin requirement), a margin call is issued to
the contract holder demanding additional margin. Otherwise futures position will
be closed by taking an equal and opposite position. Also any excess amount above
the initial margin can be withdrawn by the investor. This margin account is totally
managed by the futures clearing house. There are other rules/procedures/practices
for managing the market account but these details are not presented here.
Example 2.5.1 Suppose that the initial margin is set at 10% and the mainte-
nance margin at 5% of the futures price. Suppose for n = 0, 1, 2, 3, 4, the futures
prices are 140, 138, 130, 140 and 150 respectively. Show the working of marking to
market and margin account in a tabular form.
Solution In the below given table, the two columns termed as Margin 1 and
Margin 2, respectively refer to the deposits at the beginning and end of the day.

n f(n,T) Cash Flow Margin 1 Payment Margin 2

0 140 opening 0 -14 14
1 138 -2 12 0 12
2 130 -8 4 -9 13
3 140 10 23 9 14
4 150 10 24 9 15
closing 15 15 0
Total 10
Here on day 0, a futures position is opened and 10% deposit (i.e. Rs 14) paid
to open the market account. On day 1, the futures price drops by Rs 2 which is
subtracted from the deposit. On day 2, there is further drop of Rs 8 in the futures
price which makes the deposit below 5%. Therefore there is a margin call and the
investor has to pay Rs 9 to restore the deposit to 10% level. On day 3, the futures
42 Financial Mathematics: An Introduction

price increases and Rs 9 is withdrawn leaving a 10% margin. On day 4, the futures
price again goes up, which allows the investor to further withdraw Rs 9. At the
end of day, the investor decides to close the position collecting the balance of the
deposit. The total of all payments is Rs 10 which is the same as the increase in
future price from day 0 to day 4.



2.6 Futures Pricing

We now wish to establish that if the interest rate r is constant and it is being
compounded continuously then f (0, T) = F(0, T). Obviously there is nothing very
special here about t = 0. We can take any t, 0 < t < T and obtain

f (t, T) = F(t, T) = S(t) er(T−t) , (2.11)

for an asset which is non dividend paying. For a dividend paying stock, the above
formula (2.11) could be modified in an obvious manner.
Earlier we have noted that the futures prices are random and dictated by the
market perception. What is important to note here is that if the market futures
prices depart significantly from the values given by formula (2.11), it indicates
that the market does not believe that the interest rate is constant. The difference
between the two values signifies the market’s perception of future interest rate
changes.
Theorem 2.6.1 Let the interest rate r be constant and compounded continuously.
Then
f (0, T) = F(0, T).
Proof. For the sake of simplicity, let us assume that in the entire life of futures
contract (i.e. from t = 0 to t = T), the marking to market is carried out at two
time steps only. Let these be denoted by t = t1 and t = t2 with 0 < t1 < t2 < T. The
general case of n time steps can be dealt with by making suitable modifications
in the arguments.
We now consider two strategies, namely strategy A and strategy B, as follows

Strategy A
At t = 0: we take a long forward position with delivery as time T and forward
price as F(0, T). Also we invest the amount e−rT F(0, T) risk free until time T.
 Futures Pricing 43

At t = T: we close the risk free investment and collect the amount F(0, T). We then
use this amount to purchase one share of the asset, and sell the same at market
price S(T).
Therefore at time T our final wealth will be F(0, T) + S(T) − F(0, T) i.e. S(T).

Strategy B
At t = 0: we initiate e−r(T−t1 ) units of long futures position. This does not involve
any cost. Also we invest the amount e−rT f (0, T) risk free until time T.
At t = t1 : we receive (or pay) the amount e−r(T−t1 ) [ f (t1 , T) − f (0, T)] as a re-
sult of marking to market. We invest (or borrow, depending upon to sign)
e−r(T−t1 ) [ f (t1 , T) − f (0, T)] risk free. We further increase our long futures position
to e−r(T−t2 ) units; which does not involve any cost.
At t = t2 : we cash (or pay) e−r(T−t2 ) [ f (t2 , T) − f (t1 , T)] as a result of mark-
ing to market. We invest (or borrow, depending upon the sign) the amount
e−r(T−t2 ) [ f (t2 , T) − f (t1 , T)] risk free until time T. We increase the long futures
position to 1, which does not involve any cost.
At t = T: we collect the risk free investments which are f (0, T)+[ f (t1 , T)− f (0, T)]+
[ f (t2 , T) − f (t1 , T)], i.e. f (t2 , T). We close the futures position, receiving (or paying)
the amount S(T) − f (t2 , T).
Therefore at t = T, the final wealth of strategy B is f (t2 , T)+S(T)− f (t2 , T) = S(T).
Since at t = T, the final wealth of both strategies are same, under the principle
of no arbitrage, the initial wealth needed to initiate Strategy A and Strategy B
should be same. Thus
e−rT f (0, T) = e−rT F(0, T),
i.e.
f (0, T) = F(0, T).


Example 2.6.1 Let the interest rate r be 6% which is compounded continuously.
Let S(0) = Rs 80, and S(1) be the asset price after one day. If the marking to
market of a futures contract initiated on day 0 (i.e. first day of the year) with
delivery in 3 months is zero, then find S(1).
1
Solution We are given that t = T = 3 months= 3
4
year, t = 0 = day 1 = year,
365
r = 6% and S(0) = Rs 80. Hence by formula (2.11)




1 3 3 1
f , − f 0, =S e0.6((3/4)−(1/365)) − S(0) e(0.6×3/4) .
365 4 4 365
44 Financial Mathematics: An Introduction

From the given condition, on day 1 the marking to market is zero, i.e.



1 3 3
f , − f 0, = 0.
365 4 4
This gives

1
S = 80 e0.6/365 ,
365

which is the same as investing Rs 80 for one day with risk free rate of 6%.



2.7 Swaps
There are many investment situations where we desire to transform one cash flow
stream into another by appropriate market activity. For example, a company which
has sold fixed-coupon bonds and is paying fixed interest may wish to switch into
paying the floating rate instead. This can be realized by writing a floating coupon
bond and paying a fixed coupon bond with the same present value. This may be
achieved by a financial instrument (contract/derivate) called swap. Here one party
swaps a series of fixed-level payments for a series of variable level payments. We
can visualize a swap as a series of forward contracts, and hence price the same by
using the concept of forward pricing.
We shall have occasion to discuss interest rate swaps in a later chapter and
therefore, here, we shall introduce a commodity swap only. For this consider the
below given example.
Let us consider an electric power company that has to purchase oil every month
for its power generation facility. If it purchases the oil from the spot market, the
company will experience randomly fluctuating cash flows caused by fluctuating
spot prices. Therefore the company may desire to swap this payment scheme for
the one that is constant. But for this, the company needs to find a counter party
willing to swap. Here the swap counter party agrees to pay to the power company
the amount ((spot price of oil) × (a fixed number of barrels)), and in return the
power company pays to the counter party a fixed price per barrel for the same
number of barrels over the life of the swap. In this way, the variable cash flow
stream is transformed to a fixed cash flow stream which is depicted in Fig 2.2.
 Value of Commodity Swap 45

Fixed payments
Swap buyer
Swap seller
(Power company)
(Counter party)
Variable payments

Fig. 2.2. Swap

2.8 Value of Commodity Swap

Let us consider an agreement where party A receives spot price for N units of a
commodity each period while paying a fixed amount X per unit for N units. If the
agreement is made for M periods (i.e. the life of the swap in M periods), the net
cash flow stream received by A is (S1 − X, S2 − X, . . . , SM − X) multiplied by the
number of units N, where Si denotes the spot price of the commodity at time i.
We can value this stream using the concepts of forward markets. At time zero
forward price of the unit of commodity to be received at time i is F(i). By dis-
counting back at time zero, we note that the current value of receiving Si at time
i is d(0, i)F(i), where d(0, i) is the discount factor at time zero for cash received at
time i. By applying this argument each period, we find that the total value of the
stream under consideration is


M
V= d(0, i) (F(i) − X) N.
i=1

Hence the value of the swap can be determined by the series of forward prices.
We choose X to make the value V as zero, so that the swap represents an equal
exchange.
46 Financial Mathematics: An Introduction

2.9 Summary and Additional Notes

• In this chapter we have discussed two simple, but very popular derivative se-
curities. These are forward contracts and futures contracts. A pricing formula
for forward contract has been derived for both, a non-dividend paying stock as
well as a dividend paying stock.
• The process of marking to market is explained in the context of futures con-
tracts, and a small numerical example is presented for illustration.
• A very brief discussion is included for swap and that too in the context of
commodity swaps only.
• We have not discussed these (forward contracts, futures and swaps) derivative
securities in the context of interest rates or foreign exchange as we plan to
discuss them in a later chapter.
• Forward and futures contracts are used to hedge risk in commercial transac-
tions. The simplest type of hedge is the perfect hedge, where the risk associated
with a future commitment to deliver or receive an asset is completely eliminated
by taking an equal and opposite position in the futures market.
• The perfect hedge may not always be available. This is because there may be
no hedging instrument available that matches the commodity of the obligation
exactly. In that situation a minimum-variance hedge can be construed using in-
struments that are correlated with the obligation. We shall refer to Luenberger
[85] for details on minimum variance hedging.

2.10 Exercises
Exercise 2.1 A share of INFOSYS stock can be purchased at Rs 2, 500 today or
at Rs 2, 850 six months from now. Which of these prices is the spot price, and
which is the forward price?
Exercise 2.2 An investor enters a futures contract on SBI stock at Rs 534 today.
Tomorrow the futures price is Rs 535. How much goes into or out of investor’s
margin account? What will be the answer if tomorrow’s price in equal to Rs 532?
Exercise 2.3 Let B(0) = 100, B(1) = 112, S(0) = 34 and T = 1. Can the forward
price F of the stock can be Rs 38.60? Justify your answer mathematically.
Exercise 2.4 Explain the difference between entering into a long forward contract
with the forward price of Rs 50 and buying a call option with strike price of Rs 50.
You may assume that the other parameters, namely S(0), S(1), r and T remain
same.
 Exercises 47

Exercise 2.5 At the beginning of April 1 year, the silver forward price (in Rs/Kg)
were as follows
April 60, 650
July 61, 664
Sept 62, 348
Dec 63, 384
Assume that contracts settle at the end of the given month. The carrying cost of
silver is Rs 2, 000 per kg per year, paid at the beginning of each month. Estimate
the interest rate at that time.
Exercise 2.6 Suppose that interest rate r is constant. Given S(0), find the price
S(1) of the stock after one day such that the marking to market of futures with
delivery in 3 months is zero on that day.
Exercise 2.7 The current price of gold is Rs 25, 000 per 10 gm. The storage cost
is Rs 200 per gm per year, payable quarterly in advance. Assuming a constant
interest rate of 9% compounded quarterly, find the theoretical forward price of
gold for delivery in 9 months.
Exercise 2.8 The April 14, 2012 edition of the Economic Times gives the fol-
lowing listing for the price of a USD
• today: Rs 46.50
• 90 days forward: Rs 47.60
In other words, one can purchase 1 USD today at the price of Rs 46.50. In addition,
one can sign a contract to purchase 1 USD in 90 days at a price to be paid on
delivery of Rs 47.60. Let rIR and rUSD be the nominal yearly interest rates being
compounded continuously. Obtain the value of (rINR − rUSD ).
Exercise 2.9 Suppose that the value of a stock exchange index (say SENSEX) is
16,500, the futures price for delivery in 9 months is 17,100 index points. If the
interest rate r is 8%, find the dividend yield rdiv .
Exercise 2.10 The difference between the spot and future prices is called the basis
b(t, T). Show that as t → T, b(t, T) → 0. Derive an explicit formula for b(t, T) in
a market with constant interest rate r.
3
Basic Theory of Option Pricing-I

3.1 Introduction
We have already seen some examples of European options in Chapter 1. Though
these examples were specific to single step discrete time scenario, they were general
enough to guide the basic principle of pricing methodology. There the strategy had
been to replicate the given option in terms of stock and bond so that at the end of
expiry, the pay-off of the option matches with the value of the replicating portfolio.
Then, by no arbitrage principle, the initial value of the portfolio became the price
of the given option.
The aim of this chapter is to formalize the above pricing methodology and
introduce single and multi-period binomial lattice models for pricing of European
and American call/put options. This discussion is continued in Chapter 4 as well,
where first the Cox-Ross-Rubinstein (C.R.R.) model is presented, and then the cel-
ebrated Black-Scholes formula of option pricing is introduced. The Black-Scholes
formula will be re-visited in Chapter 10 after readers have acquired the necessary
background in stochastic calculus.

3.2 Basic Definitions and Preliminaries

In this section we define European and American call/put options and discuss
some of their simple properties.
Definition 3.2.1 (European Call Option) A European call option is a con-
tract giving the holder the right to buy an asset, called the underlying (e.g. stock),
for a price K fixed in advance, called the strike price or exercise price, at a specified
future time T, called the exercise time or the expiry time.
 50 Financial Mathematics: An Introduction

Here the term ‘underlying’ has a general meaning. It could be stock, com-
modity, foreign currency, stock index or even interest rate. However, unless it is
otherwise stated, we shall always mean stock option, i.e. the underlying asset is
being taken as stock. Thus a European (stock) call option is a derivative security
whose underlying security is stock. It gives the holder the right, to buy stock under
specified terms as prescribed in Definition 3.2.1. There is absolutely no obligation
for the holder to buy stock, but he/she has got the right to buy if he/she so wishes.
A European put option is exactly similar to a European call option, except that
the word ‘buy’ changes to ‘sell’. Therefore a European put option is a contract
giving the holder the right to sell the underlying asset for the strike price K at the
exercise time T.
Along with the European call/put options another term, namely American
call/put options, is also very commonly used. The basic distinction between a
European option and an American option is that an American option allows exer-
cise at any time before and including the expiry; whereas a European option can
be exercised only at the expiry. Here we may note that the word European and
American refer to two different conventions of exercise rather than having any
geographical significance. Thus the words European and American have become
two standards in option market, referring to two different structures, no matter
where they are issued.
The above discussion suggests that an option can be described by describing
its four basic features. These are as listed below.
(i) The description of the underlying asset.
(ii) The nature of the option - whether a call or a put; a European or an American.
(iii) The strike or exercise price.
(iv) The exercise or expiry date.
Next we try to understand the meaning of the term option pricing. By definition,
the holder of an option gets the right to buy or sell the asset (depending upon
the nature of the option) but has no obligation, so some amount has to be paid
at the time of contract to get this right. This amount is termed as the premium
or price of the option, and the problem of option pricing revolves around the
methodologies to find this price in a fair way.
An option has two sides or parties. The party that grants the option is said
to write the option and the party that obtains the option is said to purchase it.
The party that purchases the option becomes its holder, and it has no risk of
loss other than the original premium paid because it has the right to exercise the
option and has no obligation attached to it. Whereas the party that writes the
 Basic Definitions and Preliminaries 51

option has major risk. This is because if the option, say a call, is exercised then
the writer has to arrange for the asset. If the writer does not already own the
asset then it might have to be acquired from the market at a price higher than
the agreed strike price. Similarly in the case of a put option, the writer may have
to accept the asset at a much lower price (namely the strike price) than what
is prevailing in the market. The problem of option pricing is to be fair to both
the parties and determine the fair price of the option under consideration. There
is another terminology used for the writer and the purchaser of the option. The
buyer (purchaser) of the option is said to have taken the long position and the
seller (writer) is said to have taken the short position. Thus we have terms like
long call , short call , long put and short put etc.
The next important concept to understand is the pay-off or the value of the
option at expiry. To fix our ideas, let us consider the case of a European call option
with strike price K and expiry as T. Let S(T) denote the price of the underlying
asset (stock) at the expiry. Then, the holder will not exercise the option at the
expiry if S(T) ≤ K but will certainly exercise the option if S(T) > K. Therefore the
pay-off to the holder is zero for S(T) ≤ K and S(T) − K for S(T) > K. If we now
introduce the notation x+ as


+ x, x>0
x =
0, otherwise,
then the pay-off of the given call option can be written as (S(T) − K)+ .
Definition 3.2.2 (Pay-off of a European Call Option) Let C be a European
call option with specifications as prescribed in Definition 3.2.1. Then
C(T, K, S) = (S(T) − K)+ = Max (S(T) − K, 0) ,
is called the value or the pay-off of the call option C. Here S(T) denotes the price
of the underlying at the exercise time T.
It is simple to define the pay-off or the value of a European put option P.
Obviously the holder of a European put option will exercise the option only when
K > S(T) and get the profit as K − S(T). Therefore the pay-off or the value of the
given European put option is defined as
P(T, K, S) = (K − S(T))+ = Max(K − S(T), 0) .
The left hand side diagram of Fig. 3.1 depicts the pay-off of a European call option
whereas the pay-off of a European put option is depicted in the right hand side
diagram of Fig. 3.1.
52 Financial Mathematics: An Introduction

C(T) P(T)

K S(T) K S(T)

Fig. 3.1. Pay-off curves of call and put options

We say that a European call option is in the money, at the money or out of
money depending upon S(T) > K, S(T) = K or S(T) < K. A European put option is
in the money, at the money or out of money depending upon S(T) < K, S(T) = K
or S(T) > K.
We also define the gain of an option. The gain of an option buyer is the pay-
off modified by the premium C(0) or P(0) paid for the option. Thus for the call
option C, the gain is (S(T) − K)+ − C(0)erT . Similarly the gain of the buyer of the
put option P is (K − S(T))+ − P(0)erT .
We shall discuss the pay-off corresponding to an American option at a later
place. Also sometimes to make the context specific, we shall write CE or CA (re-
spectively PE or PA ) to identify whether the option is European or American.
However, if only C or P is used, it will always mean a European option.
Though there are many results connecting CE , PE and CA , PA , the following
lemma, called put-call parity is interesting and useful.

Lemma 3.2.1. (Put-Call Parity) Let CE and PE be the prices of a European

call and a European put defined over the same stock with price S. Let the given
call and the given put have the same strike price K and the same expiry date T.
Further, let the underlying stock pay no dividend. Then

CE (0) − PE (0) = S(0) − Ke−rT

where r is the constant risk-free interest rate under continuous compounding.

 Basic Definitions and Preliminaries 53

Though a formal proof could be given to the above lemma, we present here only
an intuitive argument. Suppose we construct a portfolio by writing and selling one
put and buying one call option, both with the same strike price K and expiry date
T. Now if S(T) ≥ K, then the call will pay S(T) − K and the put will be worthless.
If S(T) < K then the call will be worthless and the writer of the put will need to
pay K − S(T). In either case, the value of the portfolio will be S(T) − K at expiry.
But S(T) − K is also the pay-off of a long forward contract with forward price
K and delivery time T. Therefore by no arbitrage principle, the current value of
the constructed portfolio of options should be that of the forward contract, i.e.
CE (0) − PE (0) should be S(0) − Ke−rT . Thus CE (0) − PE (0) = S(0) − Ke−rT . Fig. 3.2.
depicts this argument.

Call pay-off
Put pay-off

− =
K S(T) K S(T) K S(T)

Fig. 3.2. Put-call parity

It is also quite obvious that CE (0) ≤ CA (0) and PE (0) ≤ PA (0). This is because
an American option gives us more freedom to exercise the right and therefore we
have to pay more premium in comparison to a European option. Here of course,
it is being assumed that both American and European options have the same
parameters, T, K and S.
The below given lemma states that for a non-dividend paying stock, the Amer-
ican call option will never be exercised prior to expiry. This being so, it should be
equivalent to the European call option, i.e. CA (0) = CE (0).

Lemma 3.2.2. Let CE and CA be respectively the prices of a European call and
an American call defined over the same stock with price S. Let both these calls
have the same strike price and the same expiry date. Further, let the stock be
non-dividend paying. Then CA (0) = CE (0).
54 Financial Mathematics: An Introduction

Again rather than giving a formal proof, let us give an intuitive argument for
the above lemma. We have already seen that CA (0) ≥ CE (0) because in comparison
to a European call option, an American option gives more right with regard to
the time of exercise of the given option. Also, by using the put-call parity, CE (0) ≥
S(0) − Ke−rT , as PE (0) ≥ 0. Therefore CA (0) ≥ CE (0) ≥ S(0) − Ke−rT . But r > 0 and
hence this last inequality gives CA (0) > S(0) − K.
But CA (0) > (S(0)−K) implies that the price of the given American call option is
greater than the pay-off. Therefore the option should be sold sooner than exercised
at t = 0. Here the choice of t = 0 is for the purpose of reference time and is totally
arbitrary. Hence taking any t < T instead of t = 0, the above arguments show
that the American call option will not be exercised at time t, i.e. the American
call option will not be exercised prior to expiry, and then it is equivalent to the
European call option. It is important to note here that the assumption of stock
being non-dividend paying is crucial for the above assertion. The situation would
be different for dividend paying stock. Moreover, this result does not hold for
American put, i.e. PA (0) is different from PE (0) even when the stock is non-dividend
paying (see Example 3.8.1).
Though the above lemma convinces that for a non dividend paying stock,
CA (0) = CE (0), we attempt to give an independent formal proof of this equality.
This proof does not make use of the result that a non-dividend paying stock, an
American call option will always be exercised at the expiry.
Lemma 3.2.3. Let the stock be non-dividend paying. Then for the same K, T and r
we have CA (0) = CE (0).
Proof. We know that CA (0) ≥ CE (0). Therefore we assume that CA (0) > CE (0)
and then arrive at a contradiction.
At time t = 0, let the investor write and sell an American call so as to get the
amount CA (0). Further let a European call be bought for the amount CE (0) and
the balance CA (0) − CE (0) be invested risk free at the interest rate r.
If the American call is exercised at time t < T, then the investor borrows a share
of stock and sells the same for K to the buyer so as to settle his/her obligation
as writer of the American call option. Further the investor invests this amount K
risk free up to time T. Now at time t = T, the investor exercises his/her European
call option to buy a share of stock for K and closes the short position on stock.
This will result in an arbitrage profit of CA (0) − CE (0) erT + Ker(T−t) − K > 0. If
the American option is not exercised at all, then  the investor will end up with
the European call and even an arbitrage profit CA (0) − CE (0) erT > 0. Therefore
CA (0) = CE (0). 

 Basic Definitions and Preliminaries 55

An obvious question at this stage is as follows : Do we have put call parity

for American option as well? Unfortunately there is no exact put call parity for
American option. But we have certain estimates for CA (0) − PA (0) as detailed in
the below given lemma.

Lemma 3.2.4. (Put-Call parity Estimate for American Option) Let the
stock be non dividend paying. Then for the same K, T and S, we have

S(0) − Ke−rT ≥ CA (0) − PA (0) ≥ S(0) − K.

Proof. We shall first prove that S(0) − Ke−rT ≥ CA (0) − PA (0). For this, we assume
that CA (0) − PA (0) − S(0) + Ke−rT > 0, and arrive at some contradiction. Now at
t = 0, we write and sell an American call, buy an American put, buy a share of
stock, and finance the transactions in the money market.
If the holder of the American call chooses to exercise it at time t ≤ T, then we
shall receive K for the share of stock and settle the money market position, ending
up with the put and a positive amount
   
K + CA (0) − PA (0) − S(0) ert = Ke−rt + CA (0) − PA (0) − S(0) ert
 
≥ Ke−rT + CA (0) − PA (0) − S(0) ert
> 0.

This argument presumes that CA (0) − PA (0) − S(0) > 0. In case CA (0) − PA (0) −
S(0) < 0, then this much money is borrowed from the bank at t = 0. Then
at t, we get K because of the sold American call. But under our assumption
K > CA (0) − PA (0) − S(0) ert and therefore we close the position in the money
market and still having a positive amount. This violates no arbitrage principle.
Next suppose that CA (0) − PA (0) − S(0) + K < 0. In this case at t = 0, we
write and sell a put, buy a call, sell short one share of stock and finance the
transactions in the money market. If the American put is exercised at t ≤ T,
then we can withdraw K from the money market to buy a share of stock and
close
 the short sale. We
 shall be left with the call option and a positive amount
−C (0) + P (0) + S(0) ert − K > Kert − K ≥ 0. If the put is not exercised at all,
A A

then we can buy a share of stock for K by exercising the call at time T and close
the short position on stock. On closing the money market position, we shall also
end up with a positive amount. This again contradicts the no arbitrage principle.


 56 Financial Mathematics: An Introduction

3.3 Behavior of Option Prices With Respect to Variables

The price of an option depends on a number of variables. These are
(i) the strike price K
(ii) the expiry time T
(iii) the current stock price S(0)
(iv) the dividend rate rdiv , and
(v) the risk free rate r.
Both from theoretical as well as from applications point of view, it should be useful
to analyze the behavior of option price as function of one of the variables, keeping
the remaining variables constant. We now present a sample of some results and
refer to the text by Capinski and Zastawniak [25] for further details. The notation
CE (K) denotes the price of European call options over the same underlying and
with the same exercise time, but with different values of the strike price K. Similar
interpretation is given for other notations, namely CE (S) and CE (T) etc.
Lemma 3.3.1. Let K1 < K2 . Then
(i) CE (K1 ) ≥ CE (K2 ),
(ii) PE (K1 ) ≤ PE (K2 ).
Thus CE (K) is a non-increasing and PE (K) is a non-decreasing function of K. These
assertions are obvious because to have the right to buy at a lower price, we have
to pay more premium than to have the right to buy at a higher price.
Lemma 3.3.2. Let K1 < K2 . Then
(i) CE (K1 ) − CE (K2 ) ≤ e−rT (K2 − K1 ),
(ii) PE (K2 ) − PE (K1 ) ≤ e−rT (K2 − K1 ).
Proof. From put-call parity, we have

CE (K1 ) − PE (K1 ) = S(0) − K1 e−rt ,

and
CE (K2 ) − PE (K2 ) = S(0) − K2 e−rt .
If we now subtract the above two equations, we get

CE (K1 ) − CE (K2 ) + PE (K2 ) − PE (K1 ) = (K2 − K1 )e−rt .

But then an application of Lemma 3.3.1 gives the result. 


 Behavior of Option Prices With Respect to Variables 57

Lemma 3.3.3. Let K1 < K2 and 0 ≤ α ≤ 1. Then CE (K) and PE (K) satisfy
(i) CE (αK1 + (1 − α)K2 ) ≤ αCE (K1 ) + (1 − α)CE (K2 )
(ii) PE (αK1 + (1 − α)K2 ) ≤ αPE (K1 ) + (1 − α)PE (K2 ).
This means that CE (K)and PE (K) are convex functions of K.

Proof. Let us denote K  = α K1 + (1 − α) K2 . Suppose that CE (K)  > α CE (K1 ) +

(1 − α) CE (K2 ).
Now at t = 0, let the investor write and sell the option with strike K  and get
the amount (premium) CE (K).  Then let the investor purchase α option with strike
price K1 and (1 − α) options with strike price K2 . This will involve a payment
(premium) of α CE (K1 ) + (1 − α) CE (K2 ). But then under the assumed inequality
 − α CE (K1 ) + (1 − α) CE (K2 ) > 0
the investor is still left with the amount CE (K)
which he/she invests risk free.

At the expiry, if the option with + price Kis exercised, then the investor
 the strike
shall have to pay the amount S(T) − K  = Max S(T) − K,  0 . But he/she can
+ +
raise the amount α (S(T) − K1 ) + (1 − α) (S(T) − K2 ) by exercising α calls with
strike price K1 and (1 − α) calls with strike price K2 . But it is known that
 +
S(T) − K ≤ α (S(T) − K1 )+ + (1 − α) (S(T) − K2 )+ ,

because the function (S(T) − K)+ is

 a convex function of K. Therefore the investor
E 
  rT
will realize an arbitrage profit of C (K) − α C (K1 ) + (1 − α) C (K2 ) e at ex-
E E

piry. Therefore our supposition is wrong and we get the result.



The below given Lemmas can also be proved on the similar lines.

Lemma 3.3.4. Let T1 < T2 . Then

(i) CE (T1 ) ≤ CE (T2 )
(ii) PE (T1 ) ≤ PE (T2 ).

Proof. If possible let CE (T1 ) > CE (T2 ). We write and sell one option expiring at
time T1 and buy one with the same strike price but expiry as T2 , investing the
balance without risk. If the written option is exercised at T1 , we can exercised
the option immediately to cover our liability. The balance (CE (T1 ) − CE (T2 )) > 0
invested without risk will be our arbitrage profit.
The inequality (ii) can be proved on similar lines. 

58 Financial Mathematics: An Introduction

Lemma 3.3.5. Let 0 < x1 < x2 . Further let at time t = 0, S = x1 S(0) and
Ŝ = x2 S(0). Then
(i) CE (S) ≤ CE (Ŝ)
(ii) PE (S) ≥ PE (Ŝ).

Proof. We shall prove the first inequality. If possible let CE (S) > CE (Ŝ). We can
write and sell a call on a portfolio with x1 shares and buy a call on a portfolio with
x2 shares having the same strike price K and exercise time T. Also we invest the
amount CE (S) − CE (Ŝ) risk-free. As x1 < x2 , we have (x1 S(T) − K)+ ≤ (x2 S(T) − K)+ .
If the sold option is exercised at time T, we can exercised other option to cover
our liability. The balance CE (S) − CE (Ŝ) > 0 invested risk-free will be our arbitrage
profit.



Lemma 3.3.6. Let S and Ŝ be as defined in Lemma 3.3.5. Then

(i) CE (Ŝ) − CE (S) ≤ (Ŝ − S)
(ii) PE (S) − PE (Ŝ) ≤ (Ŝ − S).

Proof. Using put-call parity we have

CE (Ŝ) − PE (Ŝ) = Ŝ − Ke−rT ,

and

CE (S) − PE (S) = S − Ke−rT .

On subtraction, we get
   
CE (Ŝ) − CE (S) + PE (S) − PE (Ŝ) = (Ŝ − S),

which gives inequalities (i) and (ii) because both terms on the right hand side are
non-negative.


Results similar to above Lemmas also hold for American options. For these
details we may refer to Capinski and Zastawniak [25].
 Pay-off Curves of Options Combinations 59

3.4 Pay-off Curves of Options Combinations

The graph between the pay-off and the stock price at the expiry is called the pay-
off curve of the option. Often we need to invest in combinations of options to take
care of specific hedging or speculative strategies. Knowing the pay-off curves of
individual options we can obtain the pay-off curve of the combinations of options.
This curve will be a combination of connected straight line segments depending
upon the number and nature of securities in the specific combination. This pay-off
curve of the combination is called its spread . We now give examples of some of
the most common spreads.

Bull Spread
Consider a scenario in which the investor expects the stock price to rise and wants
to speculate on that. Obviously the investor should buy a call option say CK1 , with
strike price K1 which is close to the current stock price. Here we have used the
symbol CK1 rather than C(K1 ) for the sake of simplicity. The premium may be
reduced by selling a call option, say CK2 , with strike price K2 > K1 . This strategy
should bring good returns provided the stock price increases are moderate. The
spread of the combination CK1 − CK2 is called the bull spread , which is depicted in
Fig 3.3.

Pay-off of Ck1 Pay-off of -CK2 Pay-off of

(long position) (short position) CK1 − CK2

+ =
K1 S(T) K1 K2 S(T) K1 K2 S(T)

Fig. 3.3. Bull spread

Bear Spread
The pay-off curve of the combination CK1 − CK2 with (K1 > K2 ) gives rise to a
bear spread . This is usually employed by an investor who is expecting a moderate
decline in the stock price. The bear spread is depicted in Fig. 3.4.
60 Financial Mathematics: An Introduction

Pay-off of Ck1 Pay-off of -CK2 Pay-off of

(long position) (short position) C K1 − C K2

K1
+ = K1
S(T) K2 K1 S(T) K2
S(T)

Fig. 3.4. Bear spread

Butterfly Spread
Let K1 < K2 < K3 . Let CK1 , CK2 and CK3 be call options with strike prices K1 ,

Pay−off of
Pay-off of Ck1 −2CK2
(long position)

K2
+ +
K1 K2 K3 S(T) K1 K3 S(T)

Pay−off of
CK3 Pay−off of
CK1 − 2CK2 + CK3

=
K1 K2 K3 S(T) K1 K2 S(T)
K3

Fig. 3.5. Butterfly spread

 Single Period Binomial Lattice Model for Option Pricing 61

K2 and K3 respectively. Let these calls be defined over the same stock and have
common expiry T. Then the pay-off curve of the combination CK1 − 2CK2 + CK3
gives rise to a butterfly spread. This option combination is used by an investor who
feels that the stock price will generally remain unaltered, i.e. it will not change
significantly. The butterfly spread is depicted in Fig 3.5.
The above discussion suggests that by forming the combinations of options,
any pay-off function can be approximated by a sequence of straight line segments.
In other words, any continuous pay-off function can be made close to pay-off of
an appropriate option combination.
Now similar to pay-off curve, we can also sketch gain curve of a given option
combination. The readers can sketch the gain curves for the above examples in an
obvious manner by utilizing the definition of gain of an option.
Example 3.4.1 Let the sale and purchase of options over the same stock and
having the same expiry be given by the expression

−P100 + P120 + 2C150 − C180 .

Sketch the spread of the given option combination.

Solution To draw the spread, we draw the individual pay-off curves and then
combine them. The details in Fig. 3.6 are self explanatory.



3.5 Single Period Binomial Lattice Model for Option

Pricing
The simple example of pricing a European option, presented in Chapter 1, is in
fact an example of a single period binomial lattice model. Here we formally develop
the theory for the single period case and extend the same for multi-period case in
the next section. For this model we assume the following
(i) The initial value of the stock is S(0), i.e. the stock price at t = 0 is S(0).
(ii) At the end of the period, the price is either u S(0) with probability p, or d S(0)
with probability (1 − p), 0 < p < 1.
(iii) u > d > 0.
(iv) At every period, it is possible to borrow or lend at a common risk free interest
rate r. Let R = (1 + r), and u > R > d. This assumption is needed to avoid
arbitrage.
62 Financial Mathematics: An Introduction

P120 C150
−P100

+ +
100 120 100 120 150
100

C150

−C180 −P100 + P120 − C180 + 2C150

+ =
100 120 150 180 100 120 150 180

Fig. 3.6. Spread of Example 3.4.1

Let C be the given European call option on the stock with strike price K and
expiry at the end of period (i.e. t = 1). Our aim is to find C(0). Since at t = 1, the
stock price could be either u S(0) (with probability p) or d S(0) (with probability
(1 − p)). We have


Cu = Max(u S(0) − K, 0), with probability p
C(1) =
Cd = Max(u S(1) − K, 0), with probability (1 − p).

Though the value of the risk free asset (say bond) is deterministic, we can take
it as a (degenerate) derivative of the stock-degenerate in the sense that the same
value R is assigned at the end of each arc. Thus we have the following figures

p j4 u S(0) p mm6 R
m u C = Max(u S(0) − K, 0)
jjjj
j mmm eee2
eeeeee
p
S(0) TTTT 1 QQQQQ C(0) YYYYYYYY
TT* Q( Y,
1−p R 1−p
1−p d S(0) Cd = Max(d S(0) − K, 0)

Before we actually derive the pricing formula for the given scenario, let us try
to justify the assumption u > R > d. For this we have below given lemma.
 Single Period Binomial Lattice Model for Option Pricing 63

Lemma 3.5.1. If u > R > d does not hold then “no arbitrage principle” is vio-
lated.
Proof.
 We first consider
 the case R ≥ u > d. Now construct the portfolio P :
1 1
x=− , y= , where B(0) denotes the price of the bond at t = 0. Let
S(0) B(0)
VP (0) and VP (1) respectively denote the value of the portfolio P at t = 0 and
t = 1. Then
   
1 1
VP (0) = x S(0) + y B(0) = − S(0) + B(0) = 0,
S(0) B(0)
and
VP (1) = x S(1) + y B(1)
⎧
⎪
⎪ 1 1
⎪
⎪ − u S(0) + R B(0), with probability p
⎨ S(0) B(0)
=⎪⎪
⎪
⎪ 1 1
⎩− d S(0) + R B(0), with probability (1 − p)
S(0) B(0)


R − u, with probability p
=
R − d, with probability (1 − p).
But under the assumption R ≥ u > d, (R − u) ≥ 0 and (R − d) > 0. As 0 < p < 1,
VP (1) ≥ 0 with probability 1 and it can take positive value with positive probability
even though VP (0) = 0. This clearly violates the “no arbitrage principle”.
We next consider the case u > d ≥ R, and construct the portfolio Q :
 1 1 
x= ,y=− and note that VQ (0) = 0. Further VQ (1) ≥ 0 with probability
S(0) B(0)
1 and it can take positive value (u − R) with positive probability, which again
violates the “no arbitrage principle”. Therefore to avoid the arbitrage opportunity,
we need to assume that u > R > d.


Now to find the price C(0) of the call, we need to replicate it in terms of stock
and bond so that at t = 1, the value of this replicating portfolio equals the pay-off
of the call at t = 1. Let the replicating portfolio be RP : (x = a, y = b), where a is
the number of shares of stock and b is the units of bond. Then we need to find a
and b such that VRP (1) = C(1). But
VRP (1) = a S(1) + b B(1)


a u S(0) + b R B(0), with probability p
=
a d S(0) + b R B(0), with probability (1 − p),
64 Financial Mathematics: An Introduction

and


Cu , with probability p
C(1) =
Cd , with probability (1 − p).

Hence VRP (1) = C(1) gives

a u S(0) + b R B(0) = Cu (3.1)

a d S(0) + b R B(0) = Cd . (3.2)

Solving the above system we obtain

Cu − Cd
a= , (3.3)
S(0)(u − d)
and
u C d − d Cu
b= . (3.4)
R (u − d)B(0)
Also
VRP (0) = a S(0) + b B(0). (3.5)
Therefore using (3.3), (3.4) and (3.5) we have

VRP (0) = a S(0) + b B(0)

   
Cu − Cd u Cd − d Cu
= S(0) + B(0)
S(0)(u − d) R (u − d)A(0)
   
Cu − Cd u Cd − d Cu
= +
(u − d) R (u − d)
 
 
1 R−d u−R
= Cu + Cd
R u−d u−d
1 
= p̂ Cu + (1 − p̂) Cd , (3.6)
R
 

R−d u−R
where p̂ = and (1 − p̂) = .
u−d u−d
Since at t = 1, the value VRP (1) of the replicating portfolio and the pay-off C(1)
of the call are equal, by no arbitrage principle, we should have C(0) = VRP (0).
Hence from (3.6)
1 
C(0) = p̂ Cu + (1 − p̂) Cd . (3.7)
R
 Single Period Binomial Lattice Model for Option Pricing 65

In (3.7) we should note that 0 < p̂ < 1 because u > R > d. Hence p̂ can be
considered a probability. This probability p̂ is called the risk neutral probability.
Before we discuss its interpretation and nomenclature, we observe from (3.7) that
1
C(0) = Ep̂ (C(1)), where Ep̂ denotes the expectation under risk neutral probability
R
p̂. The formula
1
C(0) = Ep̂ (C(1)) (3.8)
R
is very general and is valid for any derivative security with “appropriate” modifi-
cations.
In the above derivation, there is no role of B(0). In fact we may take B(0) = 1,
giving B(1) = R. Therefore we can think of bond as cash, as Rs 1 gives Rs (1 + r) =
Rs R after one time period. In this light, formula (3.8) tells that to find the price
C(0) of the call, we should first take the expectation of the pay-off at expiry with
respect to the risk neutral probability, and then discount the same according to
the risk free rate.
Here we should note that risk neutral probability p̂ is different from p. The
actual probability p describes the (stochastic) price movement but the risk neutral
probability p̂ has a totally different interpretation. In fact the actual probability p
enters nowhere in the derivation of C(0). However it gives a motivation to introduce
risk neutral probability. This is because one would invest in stock only if the
expected growth rate of stock is higher than that of the rate at money market
(bank/bond/cash), i.e.

p (u S(0)) + (1 − p) (d S(0)) > R S(0),

i.e.
1 
S(0) < p (u S(0)) + (1 − p) (d S(0)) . (3.9)
R
But  

R−d u−R
p̂ (u S(0)) + (1 − p̂) (d S(0)) = (u S(0)) + (d S(0)) ,
u−d u−d
which on simplification gives
1 
p̂ (u S(0)) + (1 − p̂) (d S(0)) = S(0). (3.10)
R
Equation (3.10) tells that under risk neutral probability p̂, the mean rate of
growth of stock equals the rate of growth in the money market account which is
risk free. Now if this is the case then the investor must be neutral about the risk
66 Financial Mathematics: An Introduction

and hence the name risk neutral probability. The formula (3.7) is therefore referred
as the risk neutral pricing formula for European call option.
The role of risk neutral probability is very important because it provides fair
opportunity to the writer of the option to hedge his/her risk and thereby provides
a fair price of the call. The hedging problem in the present context is to find an
equivalent replicating portfolio, i.e. a portfolio of stock and bond at t = 0 such
that the value of this portfolio at t = 1 matches with the pay-off of the call. This
process of replicating the option is called the hedging problem, and the replicating
portfolio is called the hedge of the given option.
Remark 3.5.1 Though the actual probabilities p and (1 − p) have not entered
in the pricing formula (3.7), they have played an important role indirectly. Thus
the investor will like to buy a call option provided he/she feels that there is high
probability of the stock price going up at t = 1. Also the agreed strike price K
depends to some extent on these probabilities. This is because if there is a feeling
that at t = 1, the stock price will go up with high value of p then he/she may agree
for a higher value of K.
The above procedure of finding the call price is valid provided we guarantee
that the risk neutral probability measure (RNPM) always exists and is unique.
The below given result tells that under no arbitrage principle RNPM always exists.
Further if in addition, the market is complete, then the RNPM is unique.

Lemma 3.5.2. A risk neutral probability measure exists if and only if no arbitrage
principle holds. Further if the market is complete then, the RNPM is unique.

We shall partly prove this Lemma in Section 3.7 by making use of the concept
of duality in linear programming. Specifically we shall prove that RNPM exists
if and only if, no arbitrage principle holds. The discussion about the uniqueness
of RNPM will be postponed till the concept of completeness of the market is
introduced.
Example 3.5.1 Find the price of a European call option with the given data as
B(0) = 100, B(1) = 110, S(0) = 100, K = 100, T = 1 and


120, with probability 0 · 6
S(1) =
80, with probability 0 · 4.

Will the price change if the probabilities p and (1 − p) are taken as 0.3 and 0.7
respectively?
 Single Period Binomial Lattice Model for Option Pricing 67

Solution As B(1) = 110, B(0) = 100, we get r = 10%, i.e. R = 1 + r = 1.1. Further
S(0) = 100, u S(0) = 120 and d S(0) = 80 yield u = 1.2 and d = 0.8. Here the risk
3
neutral probability p̂ is as shown below.
4
R − d 1.1 − 0.8 3
p̂ = = =
u − d 1.2 − 0.8 4
3 1
i.e. 1 − p̂ = 1 − = .
4 4
Also
Cu = Max(u S(0) − K, 0) = Max(120 − 100, 0) = 20
Cd = Max(d S(0) − K, 0) = Max(80 − 100, 0) = 0.
Therefore
1 
C(0) = p̂ Cu + (1 − p̂) Cd
R  
1 3 1
= (20) + (0)
1.1 4 4
= 13.63.
We exhibit the above details in the form of following tables

t 0 1
t 0 1 C(t) 3
p̂= 4 hh
C (= 20)
h3 u
S(t) hhhh
p
ii4 120 h h
iiii C(0) hhh
100 UUUUU V
U (?) VVVVVVVVV
(1−p) * 80 VV+
(1−p̂)= 1 4 Cd (= 0)
This gives C(0) = Rs 13.64. As call price does not depend on p, it will not change
if 0.6 and 0.4 are changed to 0.7 and 0.3 respectively.


To find the price of a European put option we can follow exactly the same
derivation as for the European call. We need to use pay-off of the put at expiry
instead of the pay-off of the call to get the following pricing formula
1 
p̂ Pu + (1 − p̂) Pd ,
P(0) = (3.11)
R
where Pu = Max (K − u S(0), 0) and Pd = Max (K − d S(0), 0).
68 Financial Mathematics: An Introduction

Example 3.5.2 For the data given in Example 3.5.1, find the price of the corre-
sponding European put option.
3 1
Solution We have already obtained p̂ = , (1 − p̂) = and R = 1.1. Further
4 4
Pu = Max (K − u S(0), 0) = 0 and Pd = Max (K − d S(0), 0) = 20. Therefore

1 
P(0) = p̂ Pu + (1 − p̂) Pd
R 


1 3 1
= ×0 + × 20
1.1 4 4
= 4.54,

and the corresponding table is

t 0 1
P(t)
h
p̂ h4 Pu (= 0)
hhhh
P(0) VVVV
V*
(1−p̂) Pd (= 20)

At this stage we can also verify the put-call parity. For the given data we
have already obtained C(0) =Rs 13.63 and P(0) =Rs 4.54. Hence the value of the
expression C(0) − P(0) + d(0, 1)K comes out to be 13.64 − 4.54 + (1.1)−1 × 100 = 100,
1
which is same as the value S(0). Here d(0, 1) = is the discount factor between
R
t = 0 and t = 1.



3.6 Multi Period Binomial Lattice Model for Option

Pricing
We shall now extend the pricing formula (3.7) to multi period binomial lattice
model. For this let us first take the two period case. Our notations are same as
those in Section 3.5. Thus S(0) is the price of the stock at t = 0 (initial time), u
and d are up and down factors of the stock price movement, K is the strike price,
t = 2 is expiry and R = (1 + r). The following table illustrates the price movement
of the stock along the binomial lattice
 Multi Period Binomial Lattice Model for Option Pricing 69

t 0 1 2
S(t) 2
hh4 u S(0)
hhhh
ii4
u S(0) VVV
iiii VV+
S(0) UUUU h3 u d S(0)
UU* hhhhh
d S(0) VVVV
VVV*
d2 S(0)

Along the line of the single period case, the table for the call price should be

t 0 1 2
C(t)
kk5 Cuu
kkkk
C
oo7 u NNNNN
oooo N'
C(0)
7 Cud
=? OOOOOO ppppp
' p
Cd SSSS
SS)
Cdd

Here Cuu , Cud and Cdd are pay-offs of the call at the expiry (i.e. at t = 2). As
per our definition, it is obvious that

Cuu = Max(u2 S(0) − K, 0)

Cud = Max(u d S(0) − K, 0)
Cdd = Max(d2 S(0) − K, 0).

But what are the values of Cu and Cd ? For this let us re-look the binomial lattice
for the price of the stock. We can visualize this bigger lattice as a combination of
three binomial lattices of the following type

u S(0)
 2
hh4 u S(0)

 h3 u d S(0)
ii4 h hhhh
iiii  hhh 
S(0) UUUU  u S(0) VVV  d S(0) VVVV
UU* VV+ VVV*
 
d S(0) u d S(0) d2 S(0)

Let above lattices be referred as L(1), L(2) and L(3) respectively. Treating lattice
L(3) as the single period binomial lattice and using formula (3.7), we obtain

1 
Cd = p̂ Cud + (1 − p̂) Cdd , (3.12)
R
70 Financial Mathematics: An Introduction

R−d
where as before p̂ is the risk neutral probability given by p̂ = . Similarly
u−d
using lattice L(2) we obtain
1 
Cu = p̂ Cuu + (1 − p̂) Cud . (3.13)
R
Finally using lattice L(1) we obtain
1 
C0 = p̂ Cu + (1 − p̂) Cd . (3.14)
R
Here the same risk neutral probability p̂ has been used for all three lattices L(1),
L(2) and L(3). The reason being that for these single period lattices the up-tick
probability p, the down-tick probability 1 − p and the factors u and d do not
change - they remain same for t = 0 as well as for t = 1. Now substituting for Cd
and Cu from (3.12) and (3.13), equation (3.14) gives

1  2 
C0 = p̂ Cuu + 2 p̂ (1 − p̂) Cud + (1 − p̂) 2
Cdd . (3.15)
R2
Formula (3.15) can also be written as
⎡ ⎤
⎢
⎢
1 ⎢⎢ 
2   ⎥⎥
2! + ⎥
C0 = 2 ⎢⎢ (p̂) (1 − p̂)
j 2− j
u d S(0) − K ⎥⎥⎥ ,
j 2− j
(3.16)
⎣
R j=0 j!(2 − j)! ⎦

 +  
where u j d2−j S(0) − K = Max u j d2− j S(0) − K, 0 .
Remark 3.6.1 Here, unlike the single period case, Cu and Cd are NOT the pay-
offs - Max(u S(0) − K, 0) and Max(d S(0) − K, 0) at t = 1, because the call is being
exercised at t = 2 only.
In deriving formula (3.15), we have not made use of the concept of replicating
portfolio directly. This has been possible because we have visualized the given 2-
period binomial lattice as a combination of three 1-period binomial lattices which
have already been studied. But it will be interesting to derive the same formula
by employing the replicating portfolio strategy as well.
Though we are not giving the complete details here, we are giving enough hint
to get the complete solution. In order to have replication at maturity t = 2 we
start replicating backwards, from the end of lattice. For this at t = 1, we consider
the upper node of the lattice L2 and determine the scalars a1 and b1 such that
 Multi Period Binomial Lattice Model for Option Pricing 71

 
a1 u2 S(0) + b1 R = Cuu ,

and
a1 (udS(0)) + b1 R = Cud .
This system can be solved to get
Cuu − Cud
a1 = ,
uS(0)(u − d)
and
Cuu − a1 (u2 S(0))
b1 = .
R
Then Cu = a1 (uS(0)) + b1 becomes the value of the replicating portfolio in the
upper node at t = 1. We can similarly find the value of Cd . Then we need to
replicate these two values in the first period. Thus we need to find a0 and b0 such
that
a0 (uS(0)) + b0 = Cu ,
and
a0 (dS(0)) + b0 = Cd .
The above system can be solved as before to get the values of a0 and b0 . Now
a little manipulation and the expression C(0) = a0 S(0) + b0 R will give the formula
(3.15).
Remark 3.6.2 It is to be noted that in this model, the number of shares in the
replicating portfolio is always equal to the ratio of ∆C and ∆S, where ∆C is the
change in the future values of the call and ∆S is the change in the future values of
the stock. This ratio is called the delta of the call, and it is different at different
nodes of the lattice. Delta is one of the Greeks to be studied later for a derivative
security defined over a given undertaking. Greeks have been used extensively in
derivative pricing for hedging purposes.
Example 3.6.1 Find the price of a European call option with the given data as
S(0) = 100, K = 100, u = 1.2, d = 0.8, r = 10% per year and time to expiry T = 2
years.
Solution We have
72 Financial Mathematics: An Introduction

t 0 1 2
t 0 1 2 C(t)
kk5 Cuu
S(t) i4 144 kkkk
iiiii C
oo7 u NNNNN
ii4 120 UUUUUU ooo N'
iiii * C(0) o
100 UUUUU ii4 96 7 Cud
U* iiii i =? OOOOOO ppp
80 UUUUUU ' ppp
U* Cd SSSS
64 SS)
Cdd

R − d 1.1 − 0.8 3
From the given data R = 1 + r = 1.1 and p̂ = = = . Also
u − d 1.2 − 0.8 4
Cuu = Max(u2 S(0) − K, 0) = Max(144 − 100, 0) = 44
Cud = Max(u d S(0) − K, 0) = Max(96 − 100, 0) = 0
Cdd = Max(d2 S(0) − K, 0) = Max(64 − 100, 0) = 0.
Therefore



1  1 3 1
Cu = p̂ Cuu + (1 − p̂) Cud = × 44 + × 0 = 30
R 1.1 4 4



1 3 1
Cd = ×0 + × 0 = 0,
1.1 4 4
which gives
1 
C(0) = p̂ Cu + (1 − p̂) Cd
R 


1 3 1
= × 30 + ×0
1.1 4 4
= 20.45.

Alternatively, we can use the formula (3.15) to get

1  2 
C(0) = p̂ C uu + 2 p̂ (1 − p̂) C ud + (1 − p̂) 2
Cdd
R2 

1 9
= 2
× 44
1.1 64
= 20.45.


Example 3.6.2 For the data given in Example 3.6.1, find the price of the corre-
sponding European put option.
 Multi Period Binomial Lattice Model for Option Pricing 73

3 1
Solution We have already obtained p̂ = , (1 − p̂) = and R = 1.1. Further
4 4
Puu = Max(K − u2 S(0), 0) = Max(−44, 0) = 0
Pud = Max(K − u d S(0), 0) = Max(4, 0) = 4
Pdd = Max(K − d2 S(0), 0) = Max(36, 0) = 36.

Therefore



1 3 1 1
Pu = ×0 + ×4 =
1.1 4 4 1.1



1 3 1 12
Pd = ×4 + × 36 = .
1.1 4 4 1.1
This gives
1 
P(0) = p̂ Pu + (1 − p̂) Pd
R 


1 3 1 1 12
= × + ×
1.1 4 1.1 4 1.1
= 3.10.

Alternatively, we can get

1  2 
P(0) = p̂ Puu + 2p̂ (1 − p̂) Pud + (1 − p̂) 2
Pdd
R2  

2  
2 
1 3 3 1 1
= ×0 + 2× × ×4 + × 36
1.12 4 4 4 4
= 3.10.

We can also use put-call parity to find P(0), once C(0) is known. Specifically
C(0) − P(0) + d(0, 2)K = S(0),
gives

P(0) = C(0) + d(0, 2)K − S(0) = 20.45 + d(0, 2)100 − 100.

We can take d(0, 2) = (1 + 2r)−1 or e−2r depending upon the nature of the com-
pounding of the interest rate r; but it has to be same for the entire calculations.


74 Financial Mathematics: An Introduction

The development of the pricing methodology for the case of multi period bino-
mial lattice model is similar to the one discussed for the two period lattice model.
The single period risk free discounting is carried out at every node of the lattice
as has been done for the two period case. Obviously we need to start from the
final period (t = N) and then proceed backward till we reach the initial time
period (t = 0). For the case of European call option, this process will result in the
following formula
⎡ ⎤
⎢
1 ⎢⎢N
N!  +⎥⎥⎥
C(0) = N ⎢⎢⎢ (p̂) j (1 − p̂)N− j u j dN− j S(0) − K ⎥⎥⎥ , (3.17)
R ⎣ j!(N − j)!
j=0
⎦

which reduces to (3.16) for N = 2.

3.7 Existence of Risk Neutral Probability Measure

We have already seen in Section 3.5 that for the single period binomial pricing
model, risk neutral probability measure (RNPM) plays the most fundamental
role. We have also exhibited the RNPM p̂ and used the same for deriving the
option pricing formula (3.7). Here it may be noted that in this scenario the state
space Ω consists of only two points say Ω = {ω1 , ω2 } where ω1 is the up-tick
movement of the stock price and ω2 is the down-tick movement of the stock price.
But realistically Ω may consist of m points Ω = {ω1 , ω2 , . . . , ωm } or Ω may be
an interval [a, b]. Can we still guarantee the existence of RNPM? To answer this
question is important if we wish to discuss derivative pricing for more general
scenarios.
Let us consider a more general model where Ω = {ω1 , ω2 , . . . , ωm }, ω j being
the jth state of economy. Thus Ω represents the finite set of possible values; e.g.
these could be m possible values of the stock price S(1) (at t = 1), namely S1 (ω1 ),
S1 (ω2 ), . . ., S1 (ωm ). For the single period binomial lattice, Ω = {ω1 , ω2 } with
S1 (ω1 ) = uS(0) and S1 (ω2 ) = dS(0).
Also we could have more than two securities, say S(k) (k = 0, 1, 2, . . . , n). Let
(k)
S1 (ω j ), for (k = 0, 1, 2, . . . , n) and ( j = 1, 2, . . . , m) denote the price of the kth
security at t = 1 when the state of the economy is ω j . We may think of k securities
as bond, stock, option, forward contracts, etc having different defining variables
but over the same underlying.
Let S(k)
0
(k = 0, 1, 2, . . . , n) denote the current (t = 0) price of the kth security.
Here we may note that S(k) 0
is deterministic,
! but S(k)
1
is a random variable tak-
(k)
ing values S1 (ω j ), (j = 1, 2, . . . , m) . Since all derivative securities are defined on
 Existence of Risk Neutral Probability Measure 75

the same underlying (stock), they all will be random variables taking m possible
values.
As a convention, we take k = 0 for bond, k = 1 for the underlying (stock)
and (k = 2, . . . , n) for other (derivative) securities. Let r be the risk free interest
rate for the period t = 0 to t = 1. It is convenient to assume that S(0) 0
= 1 and
(0)
S1 (ω j ) = R = (1 + r), (j = 1, 2, . . . , m).
Definition 3.7.1 (Risk Neutral Probability Measure) A risk neutral prob-
ability measure (RNPM) is a vector p̂ = (p̂1 , p̂2 , . . . , p̂m )T such that
(i) p̂ j > 0 ( j = 1, 2, . . . , m),
 m
(ii) p̂ j = 1,
j=1

and for every security k (k = 0, 1, 2, . . . , n), we have

⎛ ⎞
⎜⎜ m ⎟⎟
1 ⎜⎜ ⎟⎟
S(k) = ⎜ p̂ S(k)
(ω j ⎟
) ⎟⎠ . (3.18)
R ⎜⎝
0 j 1
j=1

 
If we denote by Ep̂ S1(k) the expected value of S(k)
1
with respect to RNPM p̂, then
(3.18) can be written as
1  (k) 
S(k)
0
= Ep̂ S1 . (3.19)
R
Here S(k)
1
denotes the value of the kth security at t = 1 and S0(k) is the price of the
same security at t = 0.
Regarding the existence and uniqueness of RNPM we have the following two
main theorems
Theorem 3.7.1 (First Fundamental Theorem of Asset Pricing) A risk
neutral probability measure p̂ exists if and only if no arbitrage principle holds.
Theorem 3.7.2 (Second Fundamental Theorem of Asset Pricing) The
RNPM is unique if and only if the market is complete.
Thus for an arbitrage free market, there is unique RNPM p̂ if and only if the
market is complete. We have not yet discussed the meaning of market completeness
but that we postpone for the time being.
We shall now give a linear programming based proof of Theorem 3.7.1. Let us
recall the following primal-dual pair of linear programming problem
76 Financial Mathematics: An Introduction

(LP) Min cT x
subject to
Ax ≥ b
x ≥ 0,
and
(LD) Max bT y
subject to
AT y ≤ c
y ≥ 0.
It is well known in duality theory that if (LP) and (LD) both are feasible then both
have optimal solutions. The below given theorem, called strict complementarity
theorem, gives some additional information as well.
Theorem 3.7.3 (Goldman-Tucker Theorem) Let (LP) and (LD) both be fea-
sible. Then they both have optimal solutions x∗ (for(LP)) and y∗ (for(LD)) satis-
fying
x∗ + (c − AT y∗ ) > 0. (3.20)
The condition (3.20) is called the strict complementarity condition. Here it may
be noted that Theorem 3.7.3 does not tell that (3.20) holds for every pair (x, y)
of optimal solution of (LP) and (LD). But rather it guarantees the existence of a
pair (x∗ , y∗ ) of optimal solution of (LP) and (LD) for which (3.20) holds. We can
refer to Goldman and Tucker [52] for the proof of Theorem 3.7.3.

We now proceed to prove Theorem 3.7.1. For this let us consider a portfolio
P : (x0 , x1 , . . . , xn ). Then the value of this portfolio at t = 0 is

n
VP (0) = xk S(k)
0
.
k=0

Further its value at t = 1 will be one of the m values, namely VP,1 (ω j ) =

n
xk S(k)
1
(ω j ), ( j = 1, 2, . . . , m), depending upon the state of economy ω j where
k=0
Ω = {ω1 , ω2 , . . . , ωm }. We now consider the following linear programming problem
(LPP)
Min VP (0)
subject to
VP,1 (ω j ) ≥ 0 ( j = 1, 2, . . . , m),
 Existence of Risk Neutral Probability Measure 77

i.e.

n
Min xk S(k)
0
k=0
subject to

n
xk S1(k) (ω j ) ≥ 0 ( j = 1, 2, . . . , m). (3.21)
k=0

The dual of problem (3.21) is


m
Max 0.p j
j=1

subject to

m
S(k)
1
(ω j ) p j = S(k)
0
(k = 0, 1, 2, . . . , n)
j=1
(3.22)
pj ≥ 0 ( j = 1, 2, . . . , m).

Next we note that LPP (3.21) is feasible because {xk = 0, (k = 0, 1, 2, . . . , n)}

satisfies all its constraints. Also this is optimal because by definition to meet
the assumption of solvency, only admissible portfolios are to be considered, i.e.

n
VP (0) ≥ 0, i.e. xk S(k)
0
≥ 0. But (3.21) is a minimization problem and therefore
k=0
{xk = 0, (k = 0, 1, 2, . . . , n)} is optimal to (3.21) and the optimal value is zero.
Therefore its dual LPP (3.22) also has an optimal solution with optimal value as
zero.

We now write the primal-dual pair (3.21) - (3.22) in the form of the pair (LP)-
(LD) and then apply the Goldman-Tucker theorem (Theorem 3.7.3) to this pair.
This implies that there exists x optimal to (3.21) and p optimal to (3.22) such
that
78 Financial Mathematics: An Introduction


n
− S(k)
1
(ω j )xk ≤ 0 ( j = 1, 2, . . . , m), (3.23)
k=0

m
S(k)
1
(ω j ) p j = S(k)
0
(k = 0, 1, 2, . . . , n), (3.24)
j=1

pj ≥ 0 ( j = 1, 2, . . . , m), (3.25)

m n
0.p j = S(k)
0
xk = 0, (3.26)
j=1 k=0

and
⎛ ⎛ n ⎞⎞
⎜⎜ ⎜⎜  (k) ⎟⎟⎟⎟
p j + ⎜⎜⎜⎝0 − ⎜⎜⎜⎝− S1 (ω j ) xk ⎟⎟⎟⎠⎟⎟⎟⎠ > 0 ( j = 1, 2, . . . , m). (3.27)
k=0

But from (3.26), Vp (0) = 0 for the portfolio (xk , (k = 0, 1, 2, . . . , n)). Hence by no
n
arbitrage principle, Vp (1) should be zero with probability 1, i.e. S(k)
1
(ω j ) xk =
k=0
0 ( j = 1, 2, . . . , m). Therefore (3.27) gives p j > 0 ( j = 1, 2, . . . , m).
Now we recall that k = 0 refers to the bond. Taking S(0)
0
= 1 we get S(0)
1
(ω j ) =
1 + r = R ( j = 1, 2, . . . , m). Then (3.24) gives

m
R p j = 1,
j=1

i.e.

m
1
pj = . (3.28)
j=1
R

If we now define p̂ such that p̂ j = R p j ( j = 1, 2, . . . , m), then (3.24), (3.25) and


m
(3.28) give p̂ j > 0 ( j = 1, 2, . . . , m), p̂ j = 1 and
j=1
⎛ ⎞
⎜
1 ⎜⎜⎜ m ⎟⎟
⎟⎟
⎜⎜ S(k) (ω ) p̂ (k)
⎟⎟ = S0 (k = 0, 1, 2, . . . , n). (3.29)
R⎝ ⎠
1 j j
j=1

But then this shows that p̂ = (p̂1 , p̂2 , . . . , p̂ j , . . . , p̂m ) is a risk neutral probability
measure.
 Existence of Risk Neutral Probability Measure 79

We can also write (3.29) in a more familiar notation as

1  (k) 
S(k)
0
= Ep̂ S1 (k = 0, 1, 2, . . . , n), (3.30)
R
where Ep̂ denotes the expectation under RNPM p̂. This proves that if no arbitrage
principle holds then RNPM exists. The converse can be proved on similar lines.


Remark 3.7.1 The formula (3.30) is very general. Apart from the fact that it
holds for the stock and bond, it should hold for other securities (k = 2, 3, . . . , n) as
well, be it European call, put, forward contract etc. What we simply need to do is
to find the expectation of the pay-off of the given security at expiry under RNPM
and discount the same for t = 0.
Remark 3.7.2 For the case of single period binomial lattice model the linear pro-
gramming problem to find RNPM is

Max 0 p̂1 + 0 p̂2

subject to
(u S(0)) p̂1 + (d S(0)) p̂2 = RS(0)
p̂1 + p̂2 = 1
p̂1 ≥ 0, p̂2 ≥ 0,

such that at optimality

 p̂1 > 0, p̂2 > 0. It is not difficult to see that the unique
R−d R−u
optimal solution is p̂1 = , p̂2 = , which is the same as obtained earlier
u−d u−d
by replicating portfolio arguments.
Also if k = 2 and k = 3 respectively refer to European call and European put
options, then formula (3.30) gives

1 
C(0) = p̂1 Cu + p̂2 Cd ,
R
and
1 
P(0) = p̂1 Pu + p̂2 Pd .
R
Example 3.7.1 Consider a forward contract with the given data as S(0) =
100, u = 1.2, d = 0.8, T = 1 year and r = 10% per year. Determine the for-
ward price F.
80 Financial Mathematics: An Introduction

Solution We know that F = RS(0) = (1.1) × 100 = Rs 110. But here we have to
determine F by utilizing the formula (3.30). For this we note that the pay-off of
the forward contract is (S − F) if S > F and −(F − S) if S ≤ F. Also from the given
3 1
data (p̂1 = , p̂2 = ). Therefore formula (3.30) gives
4 4
 
1 3 1
F(0) = (120 − F) + (80 − F) . (3.31)
1.1 4 4
But in the case of forward contract F(0) = 0, and then (3.31) gives F = 110.


Let us again look at formula (3.29) and concentrate for the case k = 1. We recall
that k = 1 refers to the stock and therefore if we write S10 = S(0) and S11 = S(1)
then
1
S(0) = Ep̂ (S(1))
R 
S(1)
= Ep̂
R
 
S(1)
= Ep̂
B(1)
= Ep̂ ((
S(1)). (3.32)

Here we have taken B(0) = 1, (

S(1) = (S(1)/B(1)) and written Ep̂ to emphasize that
expectation has been taken with respect to RNPM p̂. Therefore (3.32) gives

S(0) = Ep̂ ((
( S(1)) (3.33)

The expectation in (3.32) is in fact conditional because it is computed once the

stock price S(0) becomes known at t = 0. Therefore (3.33) is actually
 
S(0) = Ep̂ (
( S(1))/(
S(0) ,

which, in general, is expressed as

 
S(l) = Ep̂ ((
( S(l + 1))/(
S(l) . (3.34)

The relationship (3.34) is expressed as: the discounted stock prices (

S(0), (
S(1), (
S(2), . . .
form a martingale with respect to RNPM p̂. This holds for other securities as well,
i.e.
 Existence of Risk Neutral Probability Measure 81

 
S(k) (l) = Ep̂ (
( S(k) (l + 1)/(
S(k) (l) (k = 0, 1, 2, . . . , n).
Therefore the problem of asset pricing gets translated into the problem of find-
ing a unique RNPM p̂ or to be precise p̂-martingale on the set of scenarios Ω.
Moreover Ep̂ is called the risk neutral or martingle expectation with respect to p̂.
These concepts will probably get much clearer once we are familiar with basics of
stochastic process and stochastic calculus.
Since there are no goods in the model other than money, it is convenient to
pick one of the security as reference and normalize others with respect to it. The
security so chosen for normalization purpose is called numeraire. In our context,
the usual choice of numeraire is the bond price and that is what exactly we have
done in defining (S(0) = S(0)/B(0), (S(1) = S(1)/B(1) etc.
Example 3.7.2 Consider the data: S(0) = 100, u = 1.2, d = 0.9, r = 10% per
year and T = 2 years. Determine RNPM p̂ and show that discounted stock prices
form a p̂-martingale.
Solution We have p̂ = 2
3
and (1 − p̂) = 13 . We have the dynamics of the stock price
as
t t=0 t=1 t=2
S(t) i4 144
iiiii
ii4 120 UUUUUU
iiii *
100 UUUUU iii 4 108
U* iii
90 UUUUUU
U*
81
We can check the following


1 2 1
(144) + (108) = 120
1.1 3 3


1 2 1
(108) + (81) = 90
1.1 3 3


1 2 1
(120) + (90) = 100.
1.1 3 3

These give
 )  
S(2) S(1) 120 120
Ep̂ = =
(1.1)2 1.1 1.1 1.1
 )  
S(2) S(1) 90 90
Ep̂ = =
(1.1)2 1.1 1.1 1.1
82 Financial Mathematics: An Introduction

etc. Thus  
Ep̂ (S(2)/(
S(1) = (
S(1).


The below given theorem is very general and is valid for any European derivative
security.
Theorem 3.7.4 Let D be a European derivative security whose pay-off in the
N-period binomial model is f (S(N)). Then

1  
D(0) = Ep̂ f (S(N)) . (3.35)
RN
The above theorem essentially tells that the price of a European derivative
security D with pay-off f (S(N)) in the N-period binomial model is the expectation
of the discounted pay-off under the risk neutral probably measure.
We know that for a European call, f (S(N)) = (S(N) − K)+ ; for a European put
f (S(N)) = (K − S(N))+ and for a forward contract f (S(N)) = (S(N) − F). Therefore
the same formula (3.35) can be used to price each of these derivative securities.
Example 3.7.3 Use Theorem 3.7.1 to justify the pricing formula (3.15) for the
European call option.
Solution Here N = 2 and the three possible value for f (S(2)) are
 
Cuu = Max u2 S(0) − K, 0
Cud = Max (u d S(0) − K, 0)
 
Cdd = Max d2 S(0) − K, 0 .

(R − d)
Also p̂ = . Therefore if we define p̂1 = (p̂)2 , p̂2 = 2p̂(1 − p̂) and p̂3 = (1 − p̂)2
(u − d)
then
(i) p̂i > 0 (i = 1, 2, 3)
 3
(ii) p̂i = 1 and
i=1
(iii) Ep̂ (S(2)) = R2 S(0).
The third assertion can be verified as follows
 Existence of Risk Neutral Probability Measure 83

   
Ep̂ (S(2)) = p̂1 u2 S(0) + p̂2 (u d S(0)) + p̂3 d2 S(0)
 
= S(0) u2 (p̂)2 + 2p̂ (1 − p̂) u d S(0) + (1 − p̂)2 d2 S(0)
 2
= S(0) u p̂ + (1 − p̂) d
 2
u(R − d) + d(u − R)
= S(0)
(u − d)
 2
R(u − d)
= S(0) = R2 S(0),
u−d
i.e.
1  
S(0) = Ep̂ (S(2)) .
R2
Therefore p̂i (i = 1, 2, 3) as defined above is in fact RNPM. Hence
1   
C(0) = Ep̂ f (S(2))
R2
1  
= 2 (p̂)2 Cuu + 2p̂ (1 − p̂) Cud + (1 − p̂)2 Cdd ,
R
as obtained by formula (3.15).


Remark 3.7.3 Apparently we have employed two distinct approaches to price a
given derivative. These are the replicating portfolio approach and the RNPM ap-
proach. Various illustrative examples presented above and also the derivation of
single period binomial lattice model suggest that these two approaches are related.
In fact for a complete market, the two approaches are equivalent. This is because
the replicating portfolio approach uses the law of one price which is essentially a
consequence of no arbitrage principle. This is because then the unique RNPM ex-
ists which can be used to price any contingent claim. For the existence of RNPM,
the market has to be arbitrage free. Therefore if no arbitrage principle does not
hold then we cannot price the given derivative even if the corresponding unique
replicating portfolio exists. The below given example illustrates the point. We shall
further discuss this aspect in Section 3.10.
Example 3.7.4 Let B(0) = 100, B(1) = 120, S(0) = 100 and S(1) take values 120
and 80 with probabilities 0.8 and 0.2 respectively. Let C be a European call with
K = 100 and T = 1 year. Find the replicating portfolio (x, y) for the call C. Are
we justifying in taking C(0) = x S(0) + y B(0)? Determine RNPM if it exists.
84 Financial Mathematics: An Introduction

Solution The pay-off of the call C is

20, with probability 0.8
C(1) =
0, with probability 0.2.

Therefore if (x, y) is the replicating portfolio, then we have x S(1) + y B(1) = C(1).
This gives the following two equations

120x + 120y = 20

80x + 120y = 0,
having solution as (x = 1/2, y = −1/3). Here we cannot take C(0) = 1/2S(0) −
1/3B(0) because this utilizes the law of one price which is valid only under no
arbitrage principle. But as u = 1.2 and R = 1 + r = 1.2 we do not have the
required condition u > R > d for no arbitrage principle to hold.
To determine the RNPM we need to solve the system

p̂1 + p̂2 = 1
1  
120p̂1 + 80p̂2 = 100
1.2
p̂1 > 0, p̂2 > 0.
The above system does not have a solution because the first two equations give
pˆ1 = 1, pˆ2 = 0. Therefore the RNPM does not exist. This is again because u =
1.2 = R which violates the condition u > R > d and therefore no arbitrage principle
does not hold.



3.8 Pricing American Options: A Binomial Lattice

Model
Following earlier notations, let D denote the derivative security under consider-
ation and stock be its underlying. Let DA (n) denote the price of an American
option at the time period n, 0 ≤ n ≤ N. To be specific, let N = 2 and D denote
either an American call or an American put. We now wish to analyze the given
American option expiring after two time periods.
If the option has not already been exercised then at the expiry it will have
the value DA (2) = f (S(2)), where f is the pay-off function of the derivative D.
 Pricing American Options: A Binomial Lattice Model 85

Here we may note that for the two period binomial lattice model, f (S(2)) will
have three possible values, depending upon the three possible values of S(2). If D
denote an American put then these values are denoted by PAuu , PAud and PAdd . For
the case of an American call these values are denoted by CAuu , CAud and CAdd . Since
D is a American derivative and the binomial lattice is of two periods, the holder
can exercise his/her option at t = 0, t = 1 or t = 2.
At t = 1, the option holder has a choice. He/She can exercise immediately or
wait until t = 2. If he/she exercises his/her option at t = 1, the pay-off is f (S(1)),
otherwise it is f (S(2)). But is it worth waiting until t = 2 to exercise the option?
This question can be answered by computing the value of waiting. For this we
may treat f (S(2)) as a one step European option to be priced at t = 1. Let as
(R − d)
before p̂ = denote one step RNPM. Then the value of the given derivative
(u − d)
1
at t = 1 is given by Ep̂ ( f (S(2))), where
R
1 1 
Ep̂ ( f (S(2))) = p̂ f (u S(1)) + (1 − p̂) f (d S(1)) . (3.36)
R R
Therefore the option holder has to choose the higher of two values f (S(1)) and
1  
Ep̂ f (S(2)) as given at (3.36). Thus at t = 1, the given American option is worth
R
the price DA (1), where
 
1
DA (1) = Max f (S(1)), (p̂ f (u S(1)) + (1 − p̂) f (d S(1))) .
R
In a similar manner we can argue to get
 
1
D (0) = Max f (S(0)), (p̂ f1 (u S(0)) + (1 − p̂) f1 (d S(0)) ,
A
(3.37)
R
where
 
1
f1 (x) = Max f (x), (p̂ f (u x) + (1 − p̂) f (d x)) . (3.38)
R
Remark 3.8.1 As we have assumed that the stock is non-dividend paying, an
American call is the same as a European call and CA (0) = CE (0). Therefore the
formula (3.37) is essentially for the case of an American put, though it can be
applied for an American call as well if one so wishes. For a dividend paying stock,
an American call will behave differently from a European call.
86 Financial Mathematics: An Introduction

Remark 3.8.2 Using backward induction, formula (3.37) can be extended to an

N-period binomial lattice model in an obvious manner. Unlike the closed form
formula (3.35) for the case of European scenario, we do not have any closed form
formula for pricing of an American derivative. We do not have any choice except
to resort to the backward induction as explained above.
Example 3.8.1 Find the price of an American put option with the given data as
S(0) = 80, K = 80, u = 1.8, d = 0.95, r = 5% per year and time to expiry T = 2
years.
Solution We detail the solution in below given steps
Step 1. The table for the stock price at different periods is
n 0 1 2
S(n)
i4 96.80
iiiii
U
kk5 88 UUUUU*
kkkk
S
80 SSSS i4 83.60
S) iiiii
76 UUUUUU
*
72.20
R − d 1.05 − 0.95 2 1
Also p̂ = = = , (1 − p̂) = .
u−d 1.8 − 0.95 3 3
Step 2. Let PA (n), (n = 0, 1, 2), denote the price of given American put. To
determine PA (n), we first need to compute
f
 
Puu = Max K − u2 S(0), 0 = Max(80 − 96.80, 0) = 0
f
Pud = Max (K − u d S(0), 0) = Max (80 − 83.60, 0) = 0
f
 
Pdd = Max K − d2 S(0), 0 = Max(80 − 72.2, 0) = 7.80.
f
Here Puu etc are same as for the European put option. We have used the suffix
‘ f  to denote the forward value. The notation Pbu etc will be used to denote the
backward value to be computed in step 3. Thus f (S(2)) is a random variable tak-
f f f
ing values as Puu , Pud and Pdd .

Step 3. Compute
 
1 f f
 1 2 1
Pbu = p̂ Puu + (1 − p̂) Pud = ( × 0) + ( × 0) = 0
R 1.05 3 3
   
1 f f 1 2 1
Pbd = p̂ Pud + (1 − p̂) Pdd = ( × 0) + ( × 7.80) = 2.476.
R 1.05 3 3
 Pricing American Options: A Binomial Lattice Model 87

Step 4. Evaluate
f
Pu = Max (K − u S(0), 0) = Max(80 − 88, 0) = 0
f
Pd = Max (K − d S(0), 0) = Max(80 − 76, 0) = 4.

Step 5. Evaluate

f
Max(Pu , Pbu ) = Max(0, 0) = 0 = PMax
P (1) =
A
f
u
Max(Pd , Pbd ) = Max(4, 2.476) = 4 = PMax
d
.

Step 6. Compute


1  Max
PA (0) = Max K − S(0), p̂ Pu + (1 − p̂) PMax
d
R

 
1 2 1
= Max (80 − 80), ( × 0) + ( × 4)
1.05 3 3
= 1.27.

The above calculation are depicted in the following tables which are now self ex-
planatory

n 0 1 2
A
P (n) f
(Puu = 0)
iii4
iiii
7 (?) UUUUU
ooooo UU*
o f
(?) OOO
i4 (Pud = 0)
OOO i i
' iiii
(?) UUUUU
U*
f
(Pdd = 7.80)

n 0 1 2
A
P (n) ff3 0.0
ffffff
PMax = [0, 0] W
gg3 WWWWW
u
gWgggg WW+
(?) WWWWW g3 0.0
+ ggggg
PMax = [4, 2.47]
d XXXXX
+
7.80
88 Financial Mathematics: An Introduction

n 0 1 2
A
P (n) ff3 0.0
ffffff
Max
e2 Pu = [0, 0] W
eeeeee WWWWW
WW+
[80 − 80, 1.27] YY ggggg3 0.0
YYY, gg
PMax = [4, 2.47]
d XXXXX
+
7.80
Here the notation [a, b] is used to denote that b is the maximum of a and b,
i.e. b = Max(a, b). Therefore PA (0) = 1.27, and the given American put should be
exercised early in the down state at t = 1. This will give a pay-off of Rs 4.00 which
is more than the value of holding it to the expiry for the value of Rs 2.47. We may
also compute PE (0) and get PE (0) = 0.79. Clearly PE (0) < PA (0). Hence these are
different even for a non-dividend paying stock.


Example 3.8.2 Find the price of an American call and a European call for the
data as S(0) = 120, K = 120, u = 1.2, d = 0.9, r = 10% per year and time to
expiry T = 2 years.
Solution As the stock is non-dividend paying, CA (0) = CE (0). Computing CE (0)
by formula (3.15) or utilizing the below given tables we get CE (0) = 22.92. Here
2
p̂ = .
3
n 0 1 2
S(n) 3 172.80 h
hhhh
jj5 144 VVVVV+
jjjj
12 TTTTT h3 129.60
T) hhhhh
V
96 VVVVVV
+
97.20
n 0 1 2
CE (n) h3 52.80
hhhh
h3 34.91 VVVVVV
hhhh +
22.92 VVVVV
V+ hhhhhh3 9.60
h
5.82 VVVVV
VV+
0.0
 Options on Dividend Paying Stock 89

It may also be interesting to compute CA (0) as per the procedure outlined in

this section and verify that CA (0) also comes out to be Rs 22.92. This calculation
is shown in the below given tables

n 0 1 2
A
C (n)
jj4 52.80
jjjj
ll6 (?) SSSS
llll SS)
(?) RRRR kk5 9.60
RR( kkkk
(?) TTTTT
TT*
0.0
n 0 1 2
A
C (n) g3 52.80
gggg
4
[24, 34.9] V
i VVVV
iiiii V*
(?) UUUUU hhh 4 9.60
UU* hhhh
[0, 5.8] WWW
WWWW+
0.00



1 2 1
C (0) = Max (120 − 120),
A
(34.9) + (5.80)
1.1 3 3
= Max(0, 22.92)
= 22.92.

At time T = 2, both CE and CA have the same pay-off. At t = 1, The given

American option will not be exercised in the up state because it will bring only
Rs 24 which is less than the value of holding it up to expiry. In the down state,
the American call will be out of money and will not be exercised either.



3.9 Options on Dividend Paying Stock

It has been remarked in the last section that for a non-dividend paying stock,
CA (0) equals CE (0). But for a dividend paying stock, in general, CA (0) and CE (0)
will have different value.
Let the stock pay a dividend of amount Q at time τ, 0 ≤ τ ≤ T. We wish to
determine the price of a European call option on this stock, using the multi period
90 Financial Mathematics: An Introduction

binomial model. For this we divide the time interval [0, T] into N time periods.
Assuming that the dividend date τ occurs in the kth period, we form a lattice
in the usual way, but subtract the amount Q from the nodes at the kth period.
But then the nodes at time period k do not recombine and therefore we get a
binomial tree model rather than a binomial lattice model. Except for this change
and some obvious modifications, the procedure for finding PE (0), CE (0), PA (0) and
CA (0) remains same as discussed in Section 3.7 and Section 3.8. We illustrate the
procedure in the below given exercise.
Example 3.9.1 Compute the price of the European put with the given data as
S(0) = 12, K = 14, u = 1.1, d = 0.95, r = 2% per year and time to expiry in two
years. Assume that a dividend of Rs 2.00 is paid at time t = 1.
Solution The ex-dividend price are as given in the below given table

n 0 1 2
S(n)
(ex-div) h3 12.32
hhhh
11.20
8 VVVV
ppppp V+
pp 10.64
12.00 VVVVV 3
h 10.34
V+ hhhhh
9.40 WWWWWW
W+
8.93
7
Then following the methodology discussed in Section 3.8 we get p̂ = and
15
n 0 1 2
PE i4 1.68
iiiii
2.53 UUUUUU
u: *
uuu 3.36
u
uu
(?) TTTTT jj4 3.66
T) jjjj
4.33 UUUUUU
*
5.07
Here
 
1 7 8
2.53 = (1.68) + (3.36)
1.02 15 15
 
1 7 8
4.33 = (3.66) + (5.07) .
1.02 15 15
 Options on Dividend Paying Stock 91

Therefore  
1 7 8
P (0) =
E
(2.53) + (4.33) = 3.42.
1.02 15 15


Example 3.9.2 For the data given in Example 3.9.1, evaluate PA (0).
Solution Continuing with the data of Example 3.9.1, we have

n 0 1 2
PA 3 1.68
ggggg
[2.80, 2.53] W
p8 WWWW+
ppppp 3.36
ppp
(?) VVVVV 3 3.66
* hhhhh
[4.60, 4.33] W
WWWW+
5.07
Therefore



1 7 8
P (0) = Max K − S(0),
A
(2.80) + (4.60)
1.02 15 15
 
1 7 8
= (2.80) + (4.60)
1.02 15 15
= 3.69.


Example 3.9.3 Find the price CE (0) and CA (0) for the data, S(0) = 120, K = 120,
u = 1.2, d = 0.90, r = 10% per year and time to expiry is T = 2 periods. What are
the value of CE (0) and CA (0) if it is additionally given that a dividend of Rs 14
is paid at T = 2 ?
Solution If there is no dividend on the stock then CA (0) = CE (0). The value of
2
CE (0) can be easily determined as Rs 22.92. Here p̂ = .
3
We next consider the case when there is a dividend of Rs 14.00 at time t = 2.
This gives the following table for the stock price at different periods.
92 Financial Mathematics: An Introduction

n 0 1 2
S(n)
(ex-div) (1.2)2 (120) − 14 = 158.80
ddddddd2
j5 144 ZddZZZZZZZ
j Z-
jjjj
120 TTTTT (1.2)(0.9)(120)
1
− 14 = 115.60
T) d dddddddd
108 ZZZZZZZZZ,
(0.9)2 (120) − 14 = 83.20

n 0 1 2 n 0 1 2
CA CA g3 38.80
jj4 38.80 gggg
jjjj
ll6 (?) SSSSSS [24, 21.5] VV
llll i4 VVVVV
S) iiiii *
(?) RRRR k 5 0.0
kk (?) UUUUU hhhh3 0.0
UUU* h
RR( kkkk hhh h
(?) TTTTT [0, 0] WWWWW
TT* WWWW+
0.0 0.0

Therefore
 
1 2 1
C (0) =
A
(24) + (0)
1.1 3 3
= 14.54.

To compute CE (0) we follow the usual methodology to get the following table

n 0 1 2
CE
h4 38.80
hhhh
jj5 23.5 UUUUUUU
jjjj *
(?) TTTT
TT) iiiiii4 0.0
i
0.0 VVVVVV
V*
0.0
This gives  
1 2 1
C (0) =
E
(23.5) + (0) = 14.24.
1.1 3 3
As expected, here CE (0) and CA (0) are not equal because the stock is dividend
paying.


 Notion of Complete Markets 93

3.10 Notion of Complete Markets

The concept of completeness property for a financial market is very important.
This is because, if the market is complete then the unique RNPM exists and
therefore every financial contract has unique fair price. In view of Theorems 3.7.1
and 3.7.2 a complete market has certainly to be arbitrage free.
In order to introduce the notion of completeness, we first introduce the notion
of contingent claim. A contingent claim is a financial contract with random pay-off
that can be positive or negative. The term ‘contingent’ reflects the situation that
because the pay-off is random, it is contingent (i.e. dependent) on which state of
nature is realized at the time of pay-off. All the securities that we have described
earlier are contingent claims. Thus stock, bond, call, put, forward contract, futures
etc are all contingent claims.
We say that a portfolio replicates a contingent claim when the pay-off of the
portfolio matches the pay-off the claim in all possible states. Here ‘matching’ has
to be understood in the sense that ‘matches with probability one’. Such a portfolio
is called a replicating portfolio. Here it must be added that we require replicating
portfolio strategy to be self financing, i.e. the investor can not use funds other than
the initial wealth to finance the position in the market and he/she is not allowed
to spend money outside market. We say that the market is complete when it is
arbitrage free and we are able to uniquely replicate any contingent claim with the
existing securities.
Some readers must have realized that single and multi-period binomial lattice
models are complete market models as they are capable of pricing any contingent
claim. But what about a trinomial model? Here we have three possible outcomes
for the price of stock in the next period which are: it goes up, it does not change
or it goes down. Thus if S(0) is the price of the stock at t = 0 then S(1) takes three
values uS(0), S(0) and dS(0) with probabilities p1 , p2 and p3 respectively. The
trinomial model is certainly arbitrage free because we have again assumed that
u > R > d. But what about the existence of replicating portfolio? Let C denote
European call with strike price K and expiry t = 1. Then the portfolio (x, y) will
replicate the pay-off of the call if

xuS(0) + yB(1) = Cu
xS(0) + yB(1) = Cn
xdS(0) + yB(1) = Cd ,

where Cu = Max(uS(0) − K, 0), Cn = Max(S(0) − K, 0) and Cd = Max(dS(0) − K, 0).

94 Financial Mathematics: An Introduction

We see that the above linear system only permits a solution for a select subset of
contingent claims (e.g. European calls here) with pay-offs within a 2-dimensional
subspace. Some readers may like to verify that the linear system under consider-
ation is consistent only when

(1 − d)Cu + (u − 1)Cd − (u − d)Cn = 0.

Thus not all contigent claims can be priced by the trinomial model. This means
that the trinomial model is not complete.
In fact if we try to determine the unique RNPM for the trinomial model we
shall again face the problem. Here we need to find p̂1 > 0, p̂2 > 0, and p̂3 > 0 such
that

p̂1 + p̂2 + p̂3 = 1,

and

p̂1 (uS(0)) + p̂2 S(0) + p̂3 (dS(0)) = RS(0).

Obviously this system is consistent but it cannot have unique solution. This again
shows that the trinomial model is not complete.
To tackle such situation we enlarge the set of basic securities. For example, we
may take bond, stock and (say) an option whose price C(0) is known. Now we
wish to obtain p̂1 , p̂2 , p̂3 (p̂i > 0, i = 1, 2, 3) such that

p̂1 + p̂2 + p̂3 = 1

p̂1 (uS(0)) + p̂2 (S(0)) + p̂3 (dS(0)) = RS(0)
p̂1 Cu + p̂2 Cn + p̂3 Cd = RC(0).

The above system has unique solution giving the unique RNPM. This unique
RNPM can now be used to price other contingent claims. In this way we have
artificially made the trinomial model as a complete market model.
Example 3.10.1 Let S(0) = 100 and S(1) takes three values 120, 100 and 90 with
probabilities 0.4, 0.2 and 0.4 respectively. The price of the European call on this
stock with the strike price 105 is 5 and the price of the European call with strike
price 95 is 10, with both option expiring at t = 1. Replicate the risk-free security
that pays Rs 1 at t = 1 regardless of what happens.
Solution Let x denote the number of shares of stock, z1 the number of calls with
strike price 100 and z2 the number of calls with strike price 95. Therefore in order
to replicate the risk-free security we need to solve
 Summary and Additional Notes 95

120x + 15z1 + 25z2 = 1

100x + 0z1 + 5z2 = 1
90x + 0z1 + 0z2 = 1.

The solution of the above system is x = 1/90, z1 = 2/135 and z3 = −1/45. Further
the cost of this portfolio is (1/90)100 + (2/135)5 − (1/45)10, i.e. 0.963.
Therefore the price of risk-free security is 0.963. Since its relative return is
(1/0.963) − 1 = 3.85%, the risk-free rate in this model is 3.85%.



3.11 Summary and Additional Notes

• Certain basic definitions and preliminaries with regard to European call/put
options are presented in Section 3.2. An interesting and useful result of this sec-
tion is the put- call parity. Here it is also mentioned that there is no such exact
put-call parity for American options, but only certain estimates are provided
for CA (0) − PA (0).
• Section 3.3 is devoted to the analysis of the behavior of option prices with
respect to defining variables K, S and T. Pay-off curves/gain curves of option
combinations is the topic of discussion in Section 3.4. In particular, we have
introduced bull spread, bear spread, and butterfly spread; and also a general
procedure to draw spread of a given option combination.
• The core of this chapter has been discussion on single and multi period bi-
nomial lattice models for option pricing. The single period case is discussed
in Section 3.5, while multi period case is presented in Section 3.6. These are
very important topics on option pricing because they form the basis of the
elaborated Black-Scholes formula to be discussed in the next chapter.
• An important question about the existence of RNPM is answered in Section
3.7. Here the readers are able to see the application of the linear programming
duality in proving that under no arbitrage principle RNPM certainly exists.
Another important aspect of this section is to motivate the readers to study
martingales so as to be ready for detailed study in a later chapter.
• Section 3.8 discusses the applicability of binomial lattice approach for pricing
American options. Here, unlike the case of European option, we cannot have
any closed form pricing formula and backward induction is the only recourse.
• The case of dividend paying stock is considered in Section 3.9 and results of
Section 3.5 and 3.6 are restated in this new scenario.
96 Financial Mathematics: An Introduction

• Most of the material presented in this chapter is motivated by Capinski and

Zastawrick [25], and Luenberger [85]. The readers are encouraged to read ‘Sum-
mary and additional notes’ section of the next chapter (Chapter 4) to be fa-
miliar with certain recent topics on binomial lattice models for option pricing.
• The globalization (business anywhere in world) and democratization (informa-
tion advantage) of finance has led to designing more new products which offer
a wide array of investment choices to investors. The European and the Ameri-
can calls and puts are the simplest type of investment options, therefore they
are sometimes referred to as “plain vanilla options”. There are however several
other sophisticated options available in the market. They are referred as ex-
otic options. In the earlier years, exotic options were traded over the counter
(OTC) and not in exchanges; but more recently some of these options can be
found on exchange markets too. The prominent among many exotic options are
Asian options, barrier options, basket options, digital (or binary) options, ‘as
you like it’ (or chooser) options, look back options, rainbow options, to name
a few. These options have different characteristic and execution rules. Pricing
these options is rather hard and requires advance tools from Econometric and
Mathematics. This concept is therefore not touched here in this book. But we
sincerely encourage readers to explore the sea of exotic options, pricing, and
how they can be used in hedging risk in different situations; one can refer to
[65, 87, 145].

3.12 Exercises
Exercise 3.1 Let S(T) be the price of a given security (stock) at time T. All of the
following options have exercise time T and (unless stated otherwise) strike price
K. Give the pay-off at time T that is earned by an investor who
(i) owns one call and one put option.
(ii) owns two calls and has sold short one share of stock.
(iii) owns one share of stock and has sold one call.
(iv) owns one call having strike price K1 and has sold one put having strike price
K2 .
Exercise 3.2 Consider a family of call options on a non-dividend paying stock,
each option being identical except for its strike price. The value of the call with
strike price K is denoted by C(K). Let K1 < K2 < K3 . Show that



K3 − K2 K2 − K1
C(K2 ) ≤ C(K1 ) + C(K3 ).
K3 − K1 K3 − K1
 Exercises 97

Exercise 3.3 In a binomial lattice model, let possible values of S(2) be 32, 28 and
x. Find the value of x.
Exercise 3.4 Draw the pay-off curves of the following option portfolios
(i) −P80 + P100 + 2C130 − C150 .
(ii) −P100 + P120 + C150 .
Exercise 3.5 Let B(0) = 100, B(1) = 110, S(0) = 100 and S(1) take values
120 and 80 with probabilities 0.8 and 0.2 respectively. Also let CE and PE re-
spectively be a European call and a European put with K = 100 and T = 2.
(i) Verify put call pairty.
(ii) What can you say if the options are American options CA and PA ?
(iii) What difficulty (if any) will be faced if instead of B(1) = 110 we have B(1) =
120? Give mathematical justification to your answer.
Exercise 3.6 Let r = 0.2. Find the risk neutral conditional expectation of S(3)
given that S(2) = Rs 110.
Exercise 3.7 A certain stock is selling for Rs 50. The feeling is that for each
month, for the next two months, the stock price will rise by 10% or fall by 10%.
Assuming that the risk free rate is 1%, calculate the price of the European call
with the strike price of Rs 48.
Exercise 3.8 Prove the following
(i) S(0) − Ke−rt − D(0) ≤ CE (0)
(ii) Ke−rt + D(0) − S(0) ≤ PE (0),
where D(0) is the present value of the dividend paid by the stock. The rest of
terminologies stand for usual meanings.
Exercise 3.9 Consider the data S(0) = 60, K = 62, u = 1.1, d = 0.95, r =
0.03 and T = 3. Find CE (0), PE (0), CA (0) and PA (0). Identify the time instants
when CA and PA will be exercised.
Exercise 3.10 Let S(0) = 120, u = 1.2, d = 0.9 and r = 1%. Consider a call option
with strike price K = 120 and T = 2. Find the option price and the replicating
strategy.
Exercise 3.11 For the single period trinomial model, let S(0) = 100 and S(1)
take three values 120, 100 and 90 with probabilities 0.4, 0.2 and 0.4 respectively.
Let r = 3.85%. Further let the price of the European call on this stock with the
strike price 105 be 5. Find the price of the European call with the strike price 95.
4
Basic Theory of Option Pricing-II

4.1 Introduction
The world of options underwent a revolutionary change in 1973 when Fisher Black
and Maryon Scholes [15] and Robert C. Merton [94] published their seminal pa-
pers on theory of option pricing. The basic idea behind their studies is that the
price of an option is determined implicitly by the price of the underlying stock
and certain other parameters whose values, except one of them, are easily ob-
servable. Black-Scholes option pricing model constructed a continuously hedging
strategy to protect the writer’s short position in option. Also Merton designed a
self-financing and dynamically replicating hedged position containing the options,
the underlying risky stock and a riskless asset. Both approaches led to a partial
differential equation, now famously known as the Black-Scholes partial differential
equation or simply Black-Scholes equation, with appropriate boundary conditions
defined according to the contractual specifications of the option. These models will
be presented and discussed later in Chapter 10. The price of a particular deriva-
tive security or option is obtained by solving the Black-Scholes equation subject
to boundary conditions. In recognition to their pioneering contributions, Scholes
and Merton were awarded 1997 Nobel prize in Economics. Black’s contribution
was also recorded by the Swedish Academy though he was found ineligible for
prize as he died in 1995.
Despite its richness and simplicity, the Black-Scholes model has one major
limitation. It cannot be used to accurately price options with an American-style
exercise options as it only calculates the option price at expiration. It does not
consider the steps along the way where there could be the possibility of early
exercise of an option. To this end, came the alternative proposal of using binomial
option pricing model. Both the Black-Scholes model and the binomial model are
based on the same theoretical foundations and assumptions but there are also some
100 Financial Mathematics: An Introduction

important differences between the two. The mathematical tools Black, Scholes and
Merton employed are quite advanced, like solving a partial differential equation,
and this fact does not make the underlying economics more clear. In contrast to
the Black-Scholes model the binomial option-pricing model is simple and can easily
be understood with undergraduate knowledge. Partly the same has already been
experienced in previous chapter wherein the readers have been provided with the
glimpse of multi period binomial lattice models for option pricing through simple
two period examples.
As already exhibited in the previous chapter, the guiding principle of the bi-
nomial model is to break down the time to expiration into a large number of time
intervals or steps. Thereafter, at each step, it is assumed that the stock price will
move up or down by certain amounts. This produces a binomial lattice of un-
derlying stock prices. Subsequently the option prices at each node of the lattice
are calculated, working backward from expiration to the present. For almost all
practical purposes, in practice the discrete model of option pricing is preferred.
In this chapter, we aim to present the binomial lattice model proposed by John
C. Cox, Stephen A. Ross and Mark Rubinstein [33] in 1979 for option pricing.
This model is now famously called the CRR model for option pricing. Quoting
Cox, Ross and Rubinstein, “Our formulation, by its very construction, leads to
an alternative numerical procedure which is both simpler, and for many purposes,
computationally more efficient”. In fact the Black-Scholes formula turns out to be
a particular limiting case of the discrete binomial CRR model. In Section 4.2, we
describe the construction of the CRR model. In Section 4.3, we present the case
when the CRR model can be matched with the multi-period binomial model. This
matching finally leads to the Black-Scholes formula in the limiting case. Section
4.4 presents an extension of the Black-Scholes formula for options on dividend
paying underlying stocks. Section 4.5 briefly describes the parameters, called the
Greeks, which are vital tools in risk management.

4.2 The CRR Model

We illustrate how the knowledge gained in previous chapter on a single period
binomial model and the strategy of replication can be used to extend the option
pricing theory to multi-period scenario. The assumptions on the financial market
made for single period binomial model are carried forward here too. For the sake
of completeness, let us recall them as follows.
(i) The underlying stock on which option is written is perfectly divisible.
 The CRR Model 101

(ii) The underlying stock pays no dividend.

(iii) There is no transaction costs in buying or selling the option, and no taxes.
(iv) Short selling is allowed.
(v) The riskless interest rate is known and constant till time to expiration of
option.
(vi) No arbitrage principle holds.
Let us now divide the time horizon interval [0, T] into n subintervals each of
length ∆t = T/n, and assume that in each subinterval the stock price changes
like a one period binomial case. Thus in each period the underlying stock price
either moves up by a constant factor u or moves down by a constant factor d with
positive probability. For k = 1, . . . , n, we also define


u, with probability p
Ek =
d, with probability 1 − p,

with 0 < p < 1. Here we note that Ek is a Bernoulli random variable.

At time to expiration T, the stock price S(T) is given by

S(T) = S(0)E1 E2 · · · En
= S(0) eH , (say)

where eH = E1 E2 · · · En . Thus,

n
ln(S(T)) = ln(S(0)) + H; H= ln(Ek ).
k=1

Here, H represents the logarithmic growth of the stock price.

The above expression helps us to visualize it as a discrete approximation to
the continuous price process. The same is shown in Figure 4.1.
We recollect that if a random variable X, defined as


a, with probability p
X=
b, with probability 1 − p,

the expectation of X is E(X) = ap + b(1 − p) and the variance of X is Var(X) =

E(X2 ) − (E(X))2 = p(1 − p)(a − b)2 . Using this we can easily derive that

E(ln(Ek )) = pln(u) + (1 − p)ln(d) , (4.1)

Var(ln(Ek )) = p(1 − p)(ln(u) − ln(d))2 . (4.2)
102 Financial Mathematics: An Introduction

uuuS(0)
p
uuS(0)
p
uS(0) 1−p
uudS(0)
p
1−p
S(0) udS(0)
1−p p
uddS(0)
dS(0) p
1−p
ddS(0)
1−p
dddS(0)

Fig. 4.1. The underline binomial lattice model

We now introduce two new parameters µ and σ2 as follows

µ∆t = E(ln(Ek )) and σ2

t = Var(ln(Ek )). (4.3)

The parameter µ is called the drift and the parameter σ is called the volatility.
It may appear a little strange as to how and why we suddenly bring in two
new parameters µ and σ2 . As of now we urge to take them on face value as we
shall be providing their physical interpretation later in this chapter. Now for each
k = 1, . . . , n, let us introduce another random variable

ln(Ek ) − E(ln(Ek ))
Xk = .
Var(ln(Ek ))

Then, by using (4.1), (4.2) and definition of Ek , we can easily work out that
⎧ 1−p
⎪
⎪
⎪
⎪ , with probability p
⎪
⎪
⎨ p(1 − p)
Xk = ⎪
⎪
⎪
⎪ −p
⎪ , with probability 1 − p.
⎩
⎪
p(1 − p)

Clearly, for each (k = 1, . . . , n), E(Xk ) = 0 and Var(Xk ) = 1. Now,

 The CRR Model 103


n
H= ln(Ek )
k=1
n
√
= (µ
t + σ
tXk )
k=1 √
= µT + σ
tY,

where

n
Y= Xk , (4.4)
k=1

is a simple random walk. Though the notion of simple random walk will be ex-
plained in details in Chapter 7, here we just wish to get familiar with it for clarity.
Definition 4.2.1 (Random Walk) A random walk is a stochastic process {Sn , n =
0, 1, . . .}, with S0 = 0, defined by

n
Sn = Xk ,
k=1

where {Xk } are independent and identically distributed random variables. The ran-
dom walk is simple if for each k = 1, . . . , n, Xk takes value from {a, b} only, a and
b are real constants, with P(Xk = a) = p and P(Xk = b) = 1 − p, p ∈ [0, 1].
It is important to observe that the constants a, b, p are independent of k. Generally
a and b are taken as 1 and −1 respectively. A simple random walk can be visualized
as a game involving two persons, which consists of a sequence of independent
identically distributed moves, and the sum Sn represents the score of the first
person, say, after n moves, with the assumption that the score of the second
person is −Sn .
In view of the above discussion we have the following lemma.
Lemma 4.2.1 For a CRR model with probability of up tick u equals p and prob-
ability of down tick d equals 1 − p, life time T and time increment
t = T/n, the
stock price is given by
 √ 
S(T) = S(0) exp µT + σ ∆T Y ,

where µ is the drift and σ is the volatility described by (4.3) and Y is a simple
random walk given by (4.4).
104 Financial Mathematics: An Introduction

We calculate expectation and variance of the random variable H.

  S(T) 
E ln = E(H)
S(0) √
= E(µT + σ
tY)
√  n
= µT + σ
t E(Xk )
k=1
= µT (as E(Xk ) = 0, ∀ k) ,
and
  S(T) 
Var ln = Var(H)
S(0) √
= Var(µT + σ
tY)

n
= σ
t
2
Var(Xk )
k=1
= σ2 T (as Var(Xk ) = 1, ∀ k) .
The above two relations made us ready to give interpretations of the two
parameters - the drift and the volatility. Suppose we have n quotes of stock
prices per year. A typical example is the daily closing stock prices for T = 1
year (or 252 trading days per year), so in this case n = 252 and
t = 1 day.
We consider the random stock prices, initially S(0) = S0 and on subsequent
days S1 , S2 , . . . , S251 and finally at the end of one year S252 = S(T) = S(1).
S(1)
As an investor we are interested in the yearly return , or its logarithm
S(0)
 S(1) 
ln . Note that the logarithmic returns of the daily observed stock prices
S(0)
are {ln(S1 ) − ln(S0 ), ln(S2 ) − ln(S1 ), . . . , ln(S252 ) − ln(S251 )}. The sum of these daily
returns yield (ln(S252 ) − ln(S0 )), which is the logarithmic yearly return. This simple
S1 − S0 S2 − S1 Sn − Sn−1
characteristic is not available in decimal returns { , ,..., }.
S0 S1 Sn−1
By using the logarithmic prices we can convert an exponential problem to a linear
problem; the logarithmic returns can simply be added. For this simple reason the
logarithmic yearly returns are often used. What we observe is that the drift µ
is the expectation of the logarithmic return and the volatility σ is the standard
deviation of the logarithmic return.
The volatility is the most critical of all option trading concepts. It provides in-
vestors with an estimate of how much movement a stock can be expected to make
over a given time frame. Two types of volatilities are generally talked about - his-
torical and implied. At present we shall not be talking about the implied volatility
 Matching of CRR Model with a Multi-Period Binomial Model 105

as it requires other conceptual knowledge which we have not yet acquired. (Please
refer to the third point in the notes, Section 4.6).
Also, there are many sophisticated models in financial econometric for comput-
ing historical volatility, like ARCH (autoregressive conditional heteroskedasticity),
GARCH (generalized ARCH), risk metrics EWMA (equally weighted moving av-
erage), threshold autoregressive model, and so on so forth. Again we are not in a
position to elaborate on them. However, for the sake of understanding, we briefly
explain a very basic model for computing the historical volatility. First calculate
the natural log of the ratio of a stock price from the current day t to the previous
day t − 1, that is, Rt = ln(St /St−1 ). Then find the *naverage of all these for the given
period n (for example, n = 20 days), say, R = ( t=1 Rt )/n. The historical volatility
is the standard deviation of Rt , (t = 1, . . . , n), from the mean R. If a stock price
follows a Brownian motion or a geometric Brownian motion then the above result
is multiplied by the square root of the average number of trading days in a year
to quote the volatility on an annual basis. The reason for the last step lies in the
fact that the volatility increases with the square root of the unit of time. Though
we agree that one may find this discussion a bit incomplete, but the details have
been avoided just to skip the complications involved with volatility. However, we
encourage the readers to search the web and find sufficient material/examples on
this concept or refer to an excellent text [138].
We shall next attempt to approximate the CRR model by a multi-period bi-
nomial model. It is important to point out here that not every CRR model is a
multi-period binomial model.

4.3 Matching of CRR Model with a Multi-Period

Binomial Model
The primary question is to find the values of u, d, and p for given values of µ and
σ2 from (4.1), (4.2), (4.3). It is worth to mention that, in the conclusive analysis,
the value of µ will interestingly not be required. However, it is the actual value of
σ which is rather hard to capture. Nevertheless, we assume that σ is known from
some past data of the stock (i.e. say historical volatility).
Now, in lieu of (4.3) we have two relations (4.1) and (4.2) to determine three
parameters u, d, p. In order to describe these parameters uniquely we need a
third condition consistent with the above two relations. To meet this need, let us
denote by U = ln(u) , D = ln(d), and assume that D = −U. In other words, we
shall be matching the CRR model with a particular multi-period binomial model
106 Financial Mathematics: An Introduction

1
in which d = . Now, (4.1) and (4.2), on using (4.3), can be re-expressed as follows
u
(2p − 1)U = µ
t
4p(1 − p)U2 = σ2
t. (4.5)

Simplifying the above equations, we get

(2p − 1)2 U2 + 4p(1 − p)U2 = µ2 (

t)2 + σ2
t,

that is, +
U= µ2 (
t)2 + σ2
t, and D = −U.
Furthermore, (4.5) yields

1 µ
t
p= + .
2 2 µ2 (
t)2 + σ2
t

Let us take n sufficiently large (equivalently, the time steps between successive
trading instances approach zero) so that (
t)2 terms can be neglected. We get
√ √
U = ln(u) = σ
t and D = −U = −σ
t.

Consequently, we have
√ √ 1 µ √
u = eσ
t
, d = e−σ
t
and p = +
t. (4.6)
2 2σ
We take a small break here for some examples to illustrate the choice of the
parameters.
Example 4.3.1 A non-dividend paying stock is currently selling at Rs 100 with
annual volatility 20%. Assume that the continuously compounded risk-free interest
rate is 5%. Using a two-period CRR binomial option pricing model, find the price
of one European call option on this stock with a strike price of Rs 80 and time to
expiration 4 years.
Solution We are given S(0) = Rs 100, K = Rs 80, T = 4,
t = 2, r = 0.05, σ =
0.2. Applying the CRR model, the up-factor and the down-factor are respectively
given by
√ 1
u = eσ
t = 1.3269, d = = 0.7536.
u
 Matching of CRR Model with a Multi-Period Binomial Model 107

Now,

Suu = u2 S(0) = 176.0664, Sud = udS(0) = 100.00, Sdd = 56.7913.

Thus,
Cuu = Max{Suu − K, 0} = 96.0664,
Cud = Max{Sud − K, 0} = 20.00,
Cdd = Max{Sdd − K, 0} = 0.
Since the annual rate of risk-free

interest is continuously compounded (so instead
of 1 + r or R we shall use er t ), the risk neutral probability is
er
t − d
p̂ = = 0.6132.
u−d
Hence for a two period binomial model,
C(0) = e−r
t ((p̂)2 Cuu + 2p̂(1 − p̂)Cud + (1 − p̂)2 Cdd )
= e−0.1 ((0.3760)(96.0664) + (0.4744)(20.00) + 0)
= Rs 41.27.


For computational purpose, we take the annualized risk-free interest rate and
volatility. Although we take 365 days for counting a year but the exchanges across
the world remain closed and no trading take place on Saturday and Sunday. More-
over, there are some other holidays observed by exchanges. Consequently the num-
ber of trading days are less. In practice options are priced taking into account the
number of trading days. The following example throw light on this aspect.
Example 4.3.2 A non-dividend paying stock is selling at Rs 1, 500 on March 1,
2010, with annual volatility of 22%. Assume that the continuously compounded
risk-free interest rate is 3%. Compute the price of a European call option written
on this stock struck at Rs 1470 expiring in April 2010 using a single-period CRR
binomial model.
Solution It is given that S(0) = Rs 1, 500, K = Rs 1, 470, r = 0.03, σ = 0.22.
The option is written on March 1 and it expires in the business day immediately
preceding the last business day of the contract month which is April 29, 2010. The
total number of trading days are 44 days. The number of trading days in 2010
is 252. Thus,√T = 44/252 = 0.1746. For a single-period model,
t = T = 0.1746.
Then, u = eσ
t = 1.0963, d = u−1 = 0.9124. The risk neutral probability measure
er
t − d
is p̂ = = 0.505. Thus,
u−d
108 Financial Mathematics: An Introduction

C(0) = e−rT (p̂Cu + (1 − p̂)Cd ) = (0.9948)((0.505)174.45 + (0.495)0) = Rs 87.64.



However, for text purpose we continue to take 365 days or 12 months in a
year. This may lead to calculation differences between the theoretical value, when
applied on a real data, and the actual financial quotation (available in newspapers
or www) for the option. The readers are cautioned.
Continuing with our previous discussion, we next define a counter on up tick
and down tick movements on stock price at time k
t, (k = 1, . . . , n), as a Bernoulli
random variable


1, with probability p if stock goes up
Yk =
0, with probability 1 − p if stock goes down.

Then,
S(n
t) = S(T)
*n *n
= S(0)u k=1 Yk
d(n− k=1 Yk )

 u *nk=1 Yk
= d S(0)
n
,
d
giving
S(T) T
 u *T/
t Yk
= d
t .
k=1

S(0) d
*T/
t
Observe that Y = k=1 Yk is a simple random walk with E(Y) = p(T/
t) and
Var(Y) = p(1 − p)(T/
t).
Using the value of ln(d) from (4.6), we have

 S(T) 
−Tσ √  T/
t
ln = √ + 2σ
t Yk .
S(0)
t k=1

Hence,
  S(T)  −Tσ √  T/
t
E ln = √ + 2σ
t E(Yk )
S(0)
t k=1
−Tσ √ T
= √ + 2σ
t p

t
t
= µT,
where the last equality follows on using the value of p from (4.6). Also,
 Matching of CRR Model with a Multi-Period Binomial Model 109

  S(T)  *
Var ln = 4σ2
t T/
t
k=1
Var(Yk )
S(0)
= 4σ2 Tp(1 − p)
1
→ σ2 T, as p → when n → ∞.
2
Furthermore, by application of the central limit theorem, we can assume that
Yk follows a normal distribution when time steps approach zero. Summarizing the
above discussion we can conclude that
 S(T) 
ln ∼ N(µT, σ2 T).
S(0)
Moving ahead, observe that

√ T σ2 T
u = eσ T/n

1+σ + (4.7)
n 2n

√
−σ T/n T σ2 T
d=e +

1−σ . (4.8)
n 2n
So, the risk neutral probability measure (RNPM) is given by

R−d
p̂ =
u−d
rT
1+ −d
= n
u−d

rT T σ2 T
+σ −
n n 2n

 .
T
2σ
n
Thus, 
1 2r − σ2 T
p̂ = + . (4.9)
2 4σ n
We have seen that the European call or put options are simple to price. We
therefore concentrate on pricing the European call option using the above de-
scribed scheme. As established in previous chapter, the European call option price
for an n-period binomial lattice model is described as follows.
110 Financial Mathematics: An Introduction

 rT −n
C(0) = 1 + Ep̂ ((S(T) − K)+ )
n
 rT −n
= 1+ Ep̂ ((S(0)uY dn−Y − K)+ )
n
 rT −n  u Y
= 1+ Ep̂ ((S(0) dn − K)+ ).
n d
It follows from (4.7) and (4.8), that

 rT −n T √
C(0) = 1 + Ep̂ ((S(0)ew − K)+ ), where w = 2σ Y − σ nT. (4.10)
n n

It is to be noted here that


T √
E(w) = 2σ E(Y) − σ nT
n

T √
= 2σ np − σ nT
n
√
= (2p − 1)σ nT
√ √
= µ
t nT
= µT,

where the second

√ last equality follows by virtue of first relation in (4.5) and the
value of U = σ
t. Also, using that Var(Y) = p(1 − p)(T/
t) and n
t = T, we
have

T
Var(w) = 4σ2 Var(Y)
n
= 4p(1 − p)σ2 T
→ σ2 T,
where the last relation follows in the limiting case when n → ∞ (or
t → 0), and
p → 21 from (4.6). Moreover it is important to note that
 Matching of CRR Model with a Multi-Period Binomial Model 111


T √
Ep̂ (w) = 2σ Ep̂ (Y) − σ nT
n

T ∗ √
= 2σ np − σ nT
n

√  1 2r − σ2 T  √
= 2σ nT + − σ nT (using (4.9))
2 4σ n
 σ 2
= r− T
2
and
Var(w)p̂ = 4σ2 p̂(1 − p̂)T
→ σ2 T,
(using the limiting case when n → ∞ and p → 1
2
in (4.9)).

As a consequence of above relations, (4.10) gives

1  w − (r − 2 )T 2
σ2
, ∞ − √
1
C(0) = e −rT
(S(0)ew − K)+ √ e 2 σ T dw.
−∞ σ 2πT
 K 
Now S(0)ew > K, implies w > ln = w1 . Therefore,
S(0)

1  w − (r − 2 )T 2
σ2
, − √
1 2 σ
C(0) = e−rT (S(0)ew − K) √ e T dw.
w>w1 σ 2πT
To evaluate this integral, substitute
 σ2 
w− r− T
y= 2 .
√
σ T
Then,
√  σ2  √
w = yσ T + r − T, so dw = σ T dy.
2
Moreover, w > w1 gives y > y1 , where
112 Financial Mathematics: An Introduction

1   K   σ2  
y1 = √ ln − r− T.
σ T S(0) 2
Subsequently,
,
e−rT √ 2
T+(r− σ2 )T
y2
C(0) = √ (S(0)e yσ − K)e− 2 dy
2π y>y1

, , ∞
e−rT √ 2
yσ T+(r− σ2 )T −
y2 1 y2
= √ S(0)e e 2 dy − √ Ke−rT e− 2 dy
2π y>y1 2π y1

≡ I − Ke−rT Φ(−y1 ), (4.11)

, x
1 z2
where, Φ(x) = √ e− 2 dz is the distribution function of a standard normal
2π −∞
random variable and
,
e−rT √ σ2 y2
I= √ S(0)e yσ T+(r− 2 )T e− 2 dy
2π y>y1
√
, (y − σ T)2
S(0) −
= √ e 2 dy.
2π y>y1
√
To evaluate I we√ define a new variable s = y − σ T. Then for y > y1 , we have
s > s1 = y1 − σ T. Thus,
,
S(0) s2
I= √ e− 2 ds
2π s>s1
= S(0)Φ(−s1 )
 √ 1   K   σ2  
= S(0)Φ σ T − √ ln − r− T
σ T S(0) 2
= S(0)Φ(d1 ),
where
 S(0)   σ2 
ln + r+ T
d1 = K 2 .
√ (4.12)
σ T
Summarizing the discussion, from (4.11), we have
 Matching of CRR Model with a Multi-Period Binomial Model 113

C(0) = S(0)Φ(d1 ) − Ke−rT Φ(d2 ), (4.13)

√
where d2 = −y1 = d1 − σ T and d1 is as described in (4.12). The expression in
(4.13) is known as the Black-Scholes formula for computing the European call
option price under the set of assumptions described at the beginning of Section
4.2.
Observe that the expected rate of return µ on the stock does not appear in
the Black-Scholes formula. Moreover all parameters except for σ, the volatility of
the underlying stock, can be observed in the market. Regarding σ, we assume to
estimate it from historical market data (historical volatility) even though it may
not present the true picture. We shall briefly be touching the concept of‘implied
volatility at the end of this chapter.
The next example underlines the strength of the Black-Scholes formula.
Example 4.3.3 Suppose a non-dividend paying stock is currently selling at Rs
100 and the stock’s volatility is 24%. Assume that the continuously compounded
risk-free interest rate is 5%. A European call option is offered on this stock with
time to maturity 3 months and strike price Rs 125. Calculate the price of the block
of 100 options in the Black-Scholes framework.
Solution Here, S(0) = Rs 100, K = Rs 125, T = 3/12 = 0.25, r = 0.05, σ = 0.24.
ln(100/125) + (0.05 + 21 (0.24)2 ) 0.25
d1 = √ = −1.695363
0.24 0.25
and √
d2 = d1 − σ T = −1.815363.
Because d1 and d2 are negative, we use, Φ(d1 ) = 1−Φ(−d1 ) and Φ(d2 ) = 1−Φ(−d2 ).
We round off d1 to 1.70 and d2 to 1.82 before looking up the standard normal
distribution table. Thus,
Φ(d1 ) = 1 − 0.9554 = 0.0446, Φ(d2 ) = 1 − 0.9656 = 0.0344.
Invoking (4.13), we get

C(0) = 100(0.0446) − 125e(−0.05)(0.25) (0.0344) = 0.2134.

Therefore, the cost of the block of 100 options is Rs 21.34.



In case we wish to compute the European call option price at any time t, 0 ≤
t < T, then we simply need to make a small change in the Black-Scholes formula
as follows.
114 Financial Mathematics: An Introduction

C(t) = S(t)Φ(d1 ) − Ke−r(T−t) Φ(d2 )

 S(t) 
 σ2 
ln + r+ (T − t)
d1 = K 2
√
√σ T −t
d2 = d1 − σ T − t,
where S(t) is the spot price of the stock at time t and T − t is the remaining time
to maturity.
The price of the European put option with the strike price K and time to
maturity T − t, 0 ≤ t < T, can either be calculated using the put-call parity or
directly by the following Black-Scholes formula.

P(t) = Ke−r(T−t) Φ(−d2 ) − S(t)Φ(−d1 )

 S(t) 
 σ2 
ln + r+ (T − t)
d1 = K 2
√
√σ T −t
d2 = d1 − σ T − t.

Obviously if we enter into the European put option on the day it is offered then
the put price P(0) can be obtained by taking t = 0 in the above formula. Here, we
encouraged the readers to derive the Black-Scholes formula for pricing European
put option using the analogous arguments as done above in the derivation of C(0).
So far, we have discussed a European call or a European put options cases. It
is well known that, for a non-dividend paying stock, the value of an American call
option is same as that of a European call option with the same strike price and
time to expiration. Hence (4.13) can be used to compute price of an American
call option on a non-dividend paying underlying stock. Notably, the Black-Scholes
formula for an American put option is not the same as for European put option as
it may pay to exercise them early. An American put price has to be approximated
using the binomial method explained in previous chapter wherein we can simulate
sufficiently large binomial lattice by taking sufficiently small time steps for more
accurate approximation.

4.4 Black-Scholes Formula for Dividend Paying Stock

One of the assumptions in the Black-Scholes basic model is that the underlying
stock pays no dividend. We can easily relax this by adjusting the initial stock
 Black-Scholes Formula for Dividend Paying Stock 115

price by the present value of the dividend. To make us understand this, let us first
explore how the option gets affected by dividends on the underlying stock.
Dividends are a way by which the companies distribute part of their profits
to their investors. The question of who should be paid dividends becomes com-
plex as the composition of shareholders changes each day. To settle this question,
companies designate a date known as the record date. Dividends are paid to the
list of shareholders who hold stock on the record date. In order to allow time for
settling stock purchases, stock exchanges set a date generally two business days
prior to the record date known as the ex-dividend date. Someone who purchases
the stock on or after the ex-dividend date is not eligible to receive dividends.
Now suppose we own a call option on a dividend paying stock, but we don’t own
any of the actual underlying stock. Then the bad news is that we are not entitled
to get any dividends because dividends are paid only to the actual shareholders.
We can only receive the dividend if we own the stock before the ex-dividend date.
All what we own is the right to become a shareholder and do so only when, for
instance, we exercise the call option. But then all is not lost. The good news is
that not getting the dividends while holding only the option is compensated on the
price of option. This is due to the fact that on the ex-dividend date the underlying
stock price falls by roughly the amount of the dividend. This drop causes call
option price becoming cheaper relative to the amount of the dividend and put
option price becoming slightly more expensive. In other words, with dividends,
the premium we have to pay for call option gets reduced while the put option
premium increases.
The Black-Scholes option pricing formula has to be adjusted when the underly-
ing stock pays dividends. For adjusting the price of European options, for known
discrete dividends, we merely have to subtract the present value of the dividend
from the current price of the underlying asset in calculating the Black-Scholes
value. That is, to take into account the drop in stock price on ex-dividend date,
instead of S(0) work with S(0) − div e−r tdiv in the Black-Scholes formula, where div
denotes the dividend amount in proportions to stock, and tdiv denotes the dividend
time. In case if dividends div1 , . . . , divp are paid at discrete times tdiv1 , . . . , tdivp then
*p −r t
take S(0) − j=1 div j e div j instead of S(0) in the Black-Scholes formula (4.13) and
*p −r t
(4.12). A word of caution: be careful to see that S(0) > j=1 div j e div j for we have
to take its logarithm while computing d1 in the Black-Scholes formula.
The following examples help us to appreciate the discussed scenario.
Example 4.4.1 A stock currently trades for Rs 100 per share. The annual con-
tinuously compounded risk-free interest rate is 5% and the annual price volatility
116 Financial Mathematics: An Introduction

relevant for the Black-Scholes formula is 30%. Call options are written on this
stock with a strike price of Rs 80 and time to expiration of 5 years. The stock will
pay a dividend of Rs 20 in 2 years and another dividend of Rs 30 in 3 years. Use
the Black-Scholes formula to find the price of one such call option.
Solution Here, S(0) = Rs 100, K = Rs 80, T = 5, r = 0.05, σ = 0.3, div = Rs 20
for first 2 years and Rs 30 for next 3 years. We first compute the adjusted stock
price by decreasing the current stock price by the present value of the dividend.
The adjusted stock price is denoted by say Sa .
Sa = S(0) − (div)e−r tdiv = 100 − 20e−(0.05)2 − 30e−(0.05)3 = 56.082.
Now we use the formula
ln(Sa /K) + (r + 0.5σ2 )T
d1 = √
σ T
ln(56.08/80) + (0.05 + 0.5(0.3)2 )5
= √ = 0.1786.
0.3 5
√
d2 = d1 − σ T = −0.4922.
Φ(d1 ) = Φ(0.18) = 0.5714, Φ(d2 ) = 1 − Φ(−d2 ) = 1 − 0.6879 = 0.3121.
C(0) = Sa Φ(d1 ) − Ke−rT Φ(d2 ).
Thus,
C(0) = (56.082)(0.5714) − (62.304)(0.3121) = Rs 12.60.


Remark 4.4.1 Suppose in the above example the underlying stock pays no div-
idend in the first two years but thereafter pays a dividend of Rs 30 in 3 years.
Then,
Sa = 74.1787, d1 = 0.5955, d2 = −0.0753.
Φ(d1 ) = 0.7123, Φ(d2 ) = 1 − Φ(−d2 ) = 0.4721 and C(0) = Rs 23.42.
Now assume the stock pays no dividend in its entire lifespan of 5 years, that is,
div = 0. Then
ln(100/80) + (0.05 + 0.5(0.3)2 )5
d1 = √ = 1.0407, d2 = 0.3699.
0.3 5
Φ(d1 ) = 0.8508, Φ(d2 ) = 0.6443,
and hence C(0) = Rs 44.94.
Thus, the dividend paid by the underlying stock gets reflected in price reduction
of the European call option.
 Black-Scholes Formula for Dividend Paying Stock 117

Remark 4.4.2 Working with the same data as in Example 4.4.1, we assume that
a European put has been offered on the stock. Although we can directly compute
the put price using the Black-Scholes formula for European put option, we use the
put-call parity with dividends to do so.
(i) In case the stock pays dividends exactly in the same manner as described in
the Example 4.4.1 then
P(0) = C(0) − (S(0) − (div)e−r tdiv − Ke−rT ) = Rs 18.82.
(ii) Consider the other case when the stock pays no dividends in the first 2 years
and pays a dividend of Rs 30 in the next 3 years. Then
P(0) = C(0) − (S(0) − (div)e−r tdiv − Ke−rT ) = Rs 11.55.
(iii) If stock pays no dividend in 5 years then P(0) = C(0)−(S(0)−Ke−rT ) = Rs 7.24.
This clearly illustrates that the European put option become more costly on a
dividend paying underlying stock.
It is worth to note that if the underlying stock pays continuous dividend at a
rate δdiv , then the price of the European call option on such a stock is computed
as follows.
C(0) = S(0)e−δdiv T Φ(d1 ) − Ke−rT Φ(d2 ),
 S(0)e−δdiv T 
 σ2 
ln + r+ T
d1 = K 2 ,
√
√ σ T
d2 = d1 − σ T.
Similarly the price of a European put option is given by

P(0) = Ke−rT Φ(−d2 ) − S(0)e−δdiv T Φ(−d1 ),

where d1 and d2 are described immediately above.

Example 4.4.2 A stock currently trades for Rs 100 per share. The annual con-
tinuously compounded risk-free interest rate is 5% and the annual price volatility
relevant for the Black-Scholes formula is 20%. European call option is written on
this stock with a strike price of Rs 80 and time to expiration of 3 months. The
stock pays dividend continuously at the rate of 2%. Use the Black-Scholes formula
to find the price of one such call option.
Solution The data is S(0) = Rs 100, K = Rs 80, T = 3/12 = 0.25, r = 0.05, σ =
0.2, and δdiv = 0.02.
118 Financial Mathematics: An Introduction

ln(S(0)e−δdiv T /K) + (r + 0.5σ2 )T

d1 = √ = 2.3564, d2 = 2.2564.
σ T
Φ(d1 ) = 0.9909, Φ(d2 ) = 0.9881.
C(0) = 100e−(0.02)0.25 (0.9909) − 80e−(0.05)0.25 (0.9881) = Rs 20.53.



4.5 The Greeks

Let us shift our attention to another noteworthy aspect of the Black-Scholes for-
mula that we had already worked out for the equilibrium price of the European
call option on a non-dividend paying stock. Recall that

C(t) = S(t)Φ(d1 ) − Ke−r(T−t) Φ(d2 ), (4.14)

√  
ln(S(t)/K) + r + (σ2 /2) (T − t)
d2 = d1 − σ T − t, d1 = √ . (4.15)
σ T−t
Note that, besides the two constants - time to maturity (T) and the strike price
(K), quite a few parameters are involved in the Black-Scholes formula, viz., the
price of the underlying asset S(t); time to price t; the volatility of the underlying
asset σ; and the interest rate r. Even a minor change in one or more of them
will affect the price of the call. It is therefore essential to understand the effect of
changes in these parameters on the call price.
In calculus, the notion of “derivative” is a pivotal tool used to measure the rate
of change of a dependent variable as an independent variable is changed. A similar
situation is prevalent in the aforementioned context. Here we will calculate the
partial derivatives of option value formula (4.14), and they in turn will allow us to
determine how sensitive the value of option is to the changes in the parameters. In
this section we shall be learning about the delta, the gamma, the theta, the vega,
and the rho. These famous fives are collectively known as the Greeks.
Let us first begin with some simple derivatives. The cumulative distribution
function is , x
1
e− y /2 dy.
2
Φ(x) = √
2π −∞
By the fundamental theorem of calculus, the derivative of Φ with respect to x is
1
Φ (x) = √ e− x /2 .
2
(4.16)
2π
 The Greeks 119

Also, a basic calculus working in (4.15) yield the following.

∂d1 1  ln(S/K) σ2 
= √ −r− . (4.17)
∂t 2σ T − t T − t 2
∂d1 1
= √ . (4.18)
∂S Sσ T − t
√
∂d1 T−t
= . (4.19)
∂r σ
∂d1 √ d1
= T−t− . (4.20)
∂σ σ
We encourage the readers to compute the above expressions.

Since time is the only deterministic independent variable we are certain, we

first introduce the Greek related to time; also called the Theta denoted by Θ. The
∂C
theta of the European call option is defined as . From (4.14), we have
∂t

∂C ∂d1  ∂d1 σ 
Θ= = SΦ (d1 ) − rKe−r(T−t) Φ(d2 ) − Ke−r(T−t) Φ (d2 ) + √ .
∂t ∂t ∂t 2 T−t

We can plug in the values of the partial derivatives from (4.16) and (4.17) to
get the final expression for Θ. Moreover, using the put-call parity, we can easily
compute the Θ of the non-dividend paying European put option.
Perhaps the most significant Greek is the delta denoted by ∆. For the European
call option, it is defined to be the partial derivative of call price C with respect
to the price of the stock (underlying derivative security). From (4.14), and noting
that ∂d2 /∂S = ∂d1 /∂S, we have

∂C ∂d1
∆= = Φ(d1 ) + (SΦ (d1 ) − Ke−r(T−t) Φ (d2 )) . (4.21)
∂S ∂S

Now concentrate on the second term in the expression (4.21), we will try to
simplify it. Using (4.16) and (4.18), we obtain
120 Financial Mathematics: An Introduction

∂d1 1
(Se−d1 /2 − Ke−r(T−t)−d2 /2 )
2 2
(SΦ (d1 ) − Ke−r(T−t) Φ (d2 )) =
∂S Sσ 2π(T − t) √
1 −d21 /2
(d −σ T−t)2
−r(T−t)− 1 2
= (Se − Ke )
Sσ 2π(T − t)
e−d1 /2
2
√
(S − Ke−(r+σ /2)(T−t)+σd1 T−t )
2
=
Sσ 2π(T − t)
e−d1 /2
2

= (S − Keln(S/K) )
Sσ 2π(T − t)
= 0.

Therefore, from (4.21) we get ∆ = Φ(d1 ) for the non-dividend paying European
call. Notice that ∆ > 0, and it is less than 1. Using the put-call parity we can easily
work out that for a non-dividend paying European put option ∆ = Φ(d1 ) − 1.
Example 4.5.1 The current price of a stock is Rs 100 and its volatility is 30%
per year. The risk-free interest rate is 4% per year. A portfolio is constructed
consisting of one 6-month European call option with a strike price of Rs 80 and
the cash obtained from shorting ∆ shares of the stock. The portfolio value is non-
random. What is ∆?
Solution By assumption that the portfolio value is non-random, we have, C−∆S =
0. Thus
∂C
= ∆ = Φ(d1 ),
∂S
where
ln(100/80) + (0.04 + (0.3)2 /2)(0.5 − 0)
d1 = √ = 1.2522.
0.3 0.5 − 0
Consequently, ∆ = 0.894752.


Before we proceed in describing the other Greeks, we would like to get more
comfortable as to what roles these Greeks play in option theory. We present a
situation below in this context.
A company has sold a 3-month European call option on 1, 000 shares of a non-
dividend paying stock at a strike price of Rs 100. The share is currently trading at
Rs 95 per unit, the interest free rate of return is 5% per annum, and the market
volatility is estimated at 20% per annum. Using (4.14) and (4.15) we can easily
compute the value of 1, 000 call options C(0) = Rs 2269.30. The company is a
writer of call, so it is exposed to a risk on the call options which is the difference
 The Greeks 121

between the value of the call premium (for 1000 calls) invested at a risk free
rate of interest for 3 months and the value of the call at the exercise time. Thus
the risk profile of the company is C(0)erT − NMax{S(T) − K, 0} = Rs (2297.84 −
1000Max{S( 14 ) − 100, 0}). A natural question is that how does the company hedge
this risk? If it decides not to take any action (called the naked position) then
the company is exposed to lose if the stock price rises from present Rs 95. See,
if S( 14 ) = Rs 103, then the company loses Rs 2, 702.16 (the amount may appears
to be meagre but the numeric data in the example can be changed to make this
figure large). On the other hand, if the company decided to buy say 1, 000 shares
of the underlying stock at t = 0 (called the covered position), then it is exposed
to lose if stock decreases. See, if S( 14 ) = Rs 90, the company will lose Rs 2, 702.16.
Both positions expose the company to risk! One can easily think of designing
some other strategy, like wait and buy 1, 000 shares as soon as price of the stock
reaches Rs 100 and sell 1, 000 shares as soon as the price of the stock declines
below Rs 100. But simple hedging strategy sometime does not work well.
What we are looking for is a more concrete and workable strategy to hedge the
risk positions in options created due to changes in the stock price. It is here when
the delta, ∆, becomes a handy tool. In the immediate aforementioned situation,
the company has S(0) = Rs 95, C(0) = Rs 2, 269.30, and ∆ = Φ(d1 ) = Φ(−0.338) =
0.3685. By linear approximation from the classical calculus, we have, dC = ∆ dS.
So, if stock price goes up by Rs 1 (i.e. it becomes Rs 96) then the call price
increases approximately by Rs 0.3685 or Rs368.50 for 1, 000 call options. Now,
since the company is the writer of 1, 000 call, it must own 368.50 shares of stock
so that a change of one unit in the stock price is offset by the change in the short
call.
Generally speaking, a writer (short position) of any derivative security can take
a hedge position against risk in the security whose price depends on the underlying
stock. An important point is that the writer should frequently re-balance the
hedge position by constantly monitoring the underlying stock price movements.
A portfolio embedding the derivative security can be constructed such that it
becomes delta-neutral where by the delta-neutral portfolio we mean the portfolio
having delta equal to zero. Traders usually ensure that their portfolios are delta-
neutral at least once a day. A trading strategy that dynamically maintains a
delta-neutral portfolio is called the delta hedging which is based on the simple
trading rule of “buy high, sell low”.
Continuing with our discussion on Greeks, another important Greek is the
gamma denoted by Γ. The gamma for the non-dividend paying European call
122 Financial Mathematics: An Introduction

∂2 C
option is the second order partial derivative of C with respect to S, thus, Γ = .
∂S2
Obviously,
e−d1 /2
2
∂2 C  ∂d1
Γ= = Φ (d1 ) = .
∂S2 ∂S Sσ 2π(T − t)
Note that since Γ > 0 for the European call option, thus C is a convex function
of S, keeping other parameters constants. Again, using the put-call parity one can
work out the gamma of the European put option. We left this for the readers to
complete.
Just like delta hedging, there is another hedging strategy called the delta-
gamma hedging. A delta-gamma hedge is a delta hedge that maintains zero port-
folio gamma. The additional gamma neutrality is maintained by constructing a
portfolio comprising of a short position in the derivative security, a long position
in the underlying stock, and a long position in another hedging call such that the
delta and the gamma of the portfolio become zero. Note that the delta hedge is
based on the first-order approximation; and when change in S is not small enough,
the second-order approximation (or gamma) helps. For this reason, a delta-gamma
neutral portfolio generally offers better protection against the stock price changes
than a simple delta neutral portfolio.
Also, we shall see later in Chapter 10 that the Black-Scholes PDE is

∂f 1 2 ∂2 f ∂f
+ σ (S(t))2 2 + rS(t) = r f.
∂t 2 ∂S ∂S
∂f ∂f ∂2 f
Since = Θ, = ∆, and = Γ, the Black-Scholes PDE can be thought
∂t ∂S ∂S2
of as
1
Θ + rS(t)∆ + σ2 S2 Γ = r f. (4.22)
2
Here, f in (4.22) can be price of any derivative security which is governed by
the Black-Scholes formulas (4.14) and (4.15). Also, (4.22) provides a relationship
between the three Greeks Θ, ∆ and Γ. The readers can skip this discussion for
the time being and can come back to it after going through Chapter 10. It will
provide better understanding then. But we have included this text here for the
sake of completeness.
The fourth Greek that we would like to get familiar with is called the vega
which is the partial derivative with respect to volatility σ. It is denoted by V. For
∂C
the European call option it is . From (4.14), we have
∂σ
 The Greeks 123

∂C ∂d1 ∂d1 √
V= = SΦ (d1 ) − Ke−r(T−t) Φ (d2 )( − T − t).
∂σ ∂σ ∂σ
Using (4.20), we get
√
T − t −d2 /2 d1
V= √ Se 1 + (Ke−r(T−t) Φ (d2 ) − SΦ (d1 ))
2π σ
√
T − t − d2 /2
= √ Se 1 .
2π

The latter equation follows on account of Ke−r(T−t) Φ (d2 ) − SΦ (d1 ) = 0. For detail
working of this point, we urge the readers to take a look back at the analogous
instance while computing ∆ = Φ(d1 ) for the European call option. It is immediate
from the put-call parity that the vega of the European put option is identical to
that of the European call option.
Example 4.5.2 Consider a 3-month European put option on a stock whose cur-
rent value is Rs 100 and whose volatility is 30% per annum. The option has a
strike price of Rs 95 and the risk-free interest rate is 3.25% per annum. Find the
vega of the option. If the volatility of the stock increases to 31%, approximate the
change in the value of the put.
Solution We provide the hints to solve; the gaps can √be filled by the readers.
T − t −d2 /2
First compute d1 using (4.15), thereafter compute V = √ Se 1 for the put
2π
option. Use the linear approximation dP = Vdσ, to find approximate change in
the value of the put.


If volatility changes then the delta-gamma hedge may not work well. An en-
hancement is the delta-gamma-vega hedge, which also maintains vega zero portfo-
lio. This is accomplished by bringing in one more security in the portfolio.
The final Greek to be introduced herein is the rho. It is denoted by ρ and is the
partial derivative with respect to the risk-free interest rate r. For the European
call option it is described as follows.

∂C ∂d1 ∂d2
ρ= = SΦ (d1 ) + K(T − t)e−r(T−t) Φ(d2 ) − Ke−r(T−t) Φ (d2 )
∂r ∂r ∂r
= K(T − t)e−r(T−t) Φ(d2 ).
124 Financial Mathematics: An Introduction

Again the last equation follows on similar argument as the one given above along
∂d2 ∂d1
with (4.19) and the fact that = . We encourage you to find the ρ of the
∂r ∂r
European put.
Example 4.5.3 Consider a 3-month European put option on a stock whose cur-
rent value is Rs 100 and whose volatility is 30% per annum. The option has a
strike price of Rs 90 and the risk-free interest rate is 3.25% per annum. Find the
rho of the option. If the interest rate increases to 4%, then approximate the change
in the value of the put option.
Solution Given that K = Rs 90, S(0) = Rs 100, r = 0.0325, σ = 0.3, T = 3
12
=
0.25, t = 0. We first compute d1 as follows

ln(S/K) + (r + σ2 /2)(T − t)
d1 = √ = 0.83157.
σ T−t
The rho of the European put is given by
√
ρ = −K(T − t)e−r(T−t) Φ(σ T − t − d1 )
= (−22.13733)Φ(−0.68157) = −5.48463.

Using the linear approximation from the classical calculus, the change in value of
put option is given by

dP = ρdr = (−5.48463)(0.0075) = 0.041134.




4.6 Summary and Additional Notes

• American options are much harder to deal with than European ones. Except
for American call option on non-dividend paying underlying stock, there is no
explicit formula as Black-Scholes for American put options (on dividend or
non-dividend paying stock) or American call options on dividend paying stock.
There is only one known analytical formula for American call options due to
Roll-Geske-Whaley [65] on a stock that pays a known dividend. Also refer to
http://www.bus.lsu.edu/academics/finance/
faculty/dchance/Instructional/TN98-01.pdf. Otherwise, the CRR binomial model
can be applied to compute prices of such options. This has already been ex-
plained in previous chapter. For more accuracy in pricing these options, one
 Summary and Additional Notes 125

can take sufficiently small time steps in the CRR binomial model. Some nu-
merical methods for solving partial differential equations can also be applied
for computing American option prices. For more details on these tools, we refer
to [57, 70].
• Another interesting area where the Black-Scholes theory has been extended
is the foreign currency options. The potential of market of foreign currency
options has been fully trapped in the United States of America, Europe and
some other parts of the world. These markets are used to hedge foreign currency
risks. For example, with currency option, one can insure against the adverse
affects of changes in foreign exchange rates. A modified Black-Scholes formula
have been developed by Garman and Kohlhagen [48]. The formula involves
both the foreign and domestic interest rates. The shortcomings of the proposed
model has also been rectified in some subsequent studies. One can take a look
at research articles [48] for further insight into this concept.
• The volatility σ is the only parameter in the Black-Scholes formula which is
not explicitly observable in the market. Although we stated that we can use
historical volatility yet it has been observed that the theoretical value of the
option computed from the Black-Scholes formula fails to match the actual
option price quoted in financial circuits for that option. There could be more
than one cause for this difference. However, non availability of the actual market
value of σ is a key reason for it. This forces us to have a deeper look at this
issue. We then ask a reverse question: knowing the option price, say C(0), along
with other parameters K, T, r in the Black-Scholes formula, can we compute
σ for that stock? The σ so obtained is called implied volatility, and should be
distinguished from the historical volatility. Notably, there is no direct formula to
obtain σ from the Black-Scholes formula as we do not have an explicit formula
for inverting the Φ function. In other words, the Black-Scholes formula can not
be inverted to get the explicit expression for σ in terms of C(0), K, r, T. Again
numerical techniques such as the Newton Raphson method for computing root
of an equation can come to our rescue.
• One of the critical limitation of the CRR binomial model is that the volatility is
assumed to be constant on all the nodes of the binomial lattice. To acknowledge
the critical role of implied volatility and to model option prices consistent
with the market, many new approaches have been proposed. Rubeinstein [115]
proposed the concept of implied binomial tree (called IBT in short) which has
been extended by Derman, Kani and Chriss [38]. The principal idea in these
studies is to compute local volatility at each node of the tree. This is done by
using the Arrow-Debreu prices. Consequently, unlike in the CRR model, the
126 Financial Mathematics: An Introduction

risk neutral probability also see a change at each node. For further details on
IBT, please refer to [31].
• In recent years, innumerable contributions to lattice approaches have been
published. We would like to share some of them with our readers. Jarrow and
Rudd [68] constructed a binomial model where the first two moments, mean
and variance, of the discrete and continuous models coincide. Boyle [20] con-
structed a trinomial lattice. Tian [136] proposed binomial and trinomial models
where the model parameters are derived as unique solutions to some equation
systems derived from the first three moments. Leisen and Reimer [82] change
the formulas to determine the constant up and down factors, and proposed a
new binomial model which converges to the Black-Scholes formula with a high
order of convergence than the previously quoted methods. The main limitation
of almost all lattice models are their relatively slow speed. They do not offer
a practical solution for the calculation of thousands of prices in a few seconds.
Rapid calculations are the need of today’s market. In fact the issue of rate of
convergence of different lattice methods is crucial and has attracted attention
off late.
• The Black-Scholes no-arbitrage argument fails to capture the correct picture
of financial market involving proportional transaction costs. In such a market,
there is no portfolio that can replicate the European call option. Thus the
argument of replication of portfolio, as explained in previous chapter, can no
longer be applicable. This makes it harder to relax the assumption of “no
transaction cost” in binomial models, and consequently in the Black-Scholes
theory. Some alternative approaches have been suggested in literature for option
pricing with transactions costs. These studies are beyond the scope of present
discussion. Interested readers can refer to the works of Boyle and Vorst [20],
Leland [81], Sonar et al. [125] and Perrakis and Lefoll [108].
• While discussing the matching of CRR model with a multi period bino-
mial model in Section 4.3, it has been assumed that ud = 1, i.e. S(0) =

(uS(0))(dS(0)). In other words S(0) is the geometric mean of uS(0) and dS(0).
This opens the possibility of using arithmetic mean and harmonic mean of
uS(0) and dS(0) as well. We may refer to Chawala [26] for further details in
this regard.

4.7 Exercises
Exercise 4.1 Consider the following data: S(0) = Rs 51, K = Rs 50, σ =
30%, r = 8%. Assuming the Black-Scholes framework, and that the stock pays
 Exercises 127

no dividend, compute 3-months European call price and 3-months European put
price using the Black-Scholes formula. Also compute the put price using the put-
call parity. Are the two values same?
Exercise 4.2 The price of a stock is Rs 260. A 6-month European call option on
the stock with strike price Rs 256 is priced using Black-Scholes formula. It is given
that the continuously compounded risk-free rate is 4%; the stock pays no dividend;
the volatility of the stock is 25%. Determine the price of the call option.
Exercise 4.3 You own 100 shares of a stock whose current price is Rs 42. You
would like to hedge your downside exposure by buying 6-month European put option
with a strike price of Rs 40. It is given that the continuously compounded risk-free
rate is 5%; the stock pays no dividends; the stock volatility is 22%. Assuming the
Black-Scholes framework determine the cost of the put option.
Exercise 4.4 Consider purchase of 100 units of 3-month Rs 25-strike European
call option. It s given that the stock is currently selling for Rs 20; the continuous
compounding risk free interest is 5%; the stocks volatility is 24% per annum. If
the stock pays dividends continuously at the rate of 3% per annum, determine the
price of block of 100 call options, assuming the Black-Scholes framework.
Exercise 4.5 For European call and put options on a stock having the same expiry
and strike price, it is given that the stock price is Rs 85; the strike price is Rs 90;
the continuously compounded risk free rate is 4%; the continuously compounded
dividend rate on the stock is 2%. If the call option has premium Rs 9.91 and a put
option has premium Rs 12.63, then determine the time to expiry for the options.
Exercise 4.6 Consider a 1-year European call option on a non-dividend paying
stock which is currently priced at Rs 40. The call strike price is Rs 45 and the
continuously compounded risk-free rate is 5%. It has been observed that if the stock
price increases Rs 0.50, the price of the option increases by Rs 0.25. Assuming
the price of the stock follows the Black-Scholes framework, determine the implied
volatility of the stock.
Exercise 4.7 Consider a stock which is currently trading at Rs 100.
(i) Assuming the stock pays no dividend, compute the Black-Scholes European call
price for T years to maturity with a strike price Rs 120, stock price volatility
30%, continuous compounding return 8%.
(ii) What happens to the option price as T → ∞?
(iii) Suppose the same stock pays a dividend of 0.1%. Repeat (i).
128 Financial Mathematics: An Introduction

(iv) Now what happens to the option price in (iii) when T → ∞?

(v) What accounts for the difference in (ii) and (iv)?
5
Portfolio Optimization-I

5.1 Introduction
The readers have already come across certain terms like portfolio, portfolio return,
and portfolio risk in Chapter 1. There, while giving example of a simple portfolio
optimization problem, it was remarked that we need to minimize the portfolio risk
for given aspiration level of portfolio returns and thereby do a trade-off between
the two. This has precisely been the approach of the celebrated mean-variance
theory of Markowitz [90] for the single period portfolio optimization problem. The
aim of this chapter is to continue this discussion in greater detail for a general
n-asset problem and discuss other related results, in particular the capital asset
pricing model (CAPM).

5.2 Risk and Return of an Asset

Let us recollect the usual definition of asset return and asset risk from Chapter 1.
Definition 5.2.1 (Return of an Asset) The return of an asset is an indicator
of gain/loss in the investment of an asset in the financial market. It is determined
by the following formula
amount received − amount invested
return = .
amount invested
In the context of a single period model, let at the initial time t = 0 the price of
an asset be B(0), and at the final time t = T its price be A(T). Then its return r
is given by
B(T) − B(0)
r= ,
B(0)
130 Financial Mathematics: An Introduction

and is normally expressed in percentage.

The positive value of return on an asset signifies gain while the negative value
of return signifies loss. The zero value of return means neither gain nor loss from
the investment.

Remark 5.2.1 It is important to mention here that Definition 5.2.1 is actually

the definition of the rate of return on an asset. However, we shall continue to call
it as return on an asset for consistency with the financial market glossary.
Definition 5.2.2 (Risk of an Asset) The risk is often defined as the degree of
uncertainty of return on an asset. It signifies the possibility of loss in the invest-
ment.

The risk can either be zero, implying that the asset is risk-free, or positive,
implying the asset is risky. If the asset is risk-free then the future value of the
asset is known with certainty otherwise the future value of the risky asset is
uncertain. Any financial asset can thus be classified either as risk-free asset (like a
fixed deposit) or as a risky asset (like share of a stock or share of a mutual fund).
Unfortunately there is no unique measure of risk of an asset. In practice we
most often measure risk of an asset in terms of variance of returns on that asset,
but that may not be the best thing to do. However, unless otherwise stated, the
variance will always be used as a risk measure. But we shall have a much deeper
discussion on various risk measures in the next chapter.
In order to obtain the mathematical model of a typical portfolio optimization
problem we make certain assumptions. These are as follows

(i) The prices of all assets at any time are strictly positive.
(ii) The return r of an asset is a random variable.
(iii) An investor can own a fraction of an asset. This assumption is known as
divisibility. Also, we assume that this fraction can be positive or negative, i.e.
short selling is allowed. We shall mention specifically if short selling is not
allowed.
(iv) An asset can be bought or sold on demand in any quantity at the market
price. This assumption is known as liquidity.
(v) No arbitrage principle holds.
(vi) Unless stated otherwise, there are no commissions/transaction costs.
 Portfolio Optimization Problem 131

5.3 Portfolio Optimization Problem

We are now in a position to move towards our main goal, i.e. to study the portfolio
optimization problem. We first define a portfolio and other associated terms.
Definition 5.3.1 (Portfolio) A portfolio is a collection of two or more assets,
say, a1 , . . . , an , represented by an ordered n-tuple Θ = (x1 , . . . , xn ), where xi ∈ R is
the number of units of the asset ai (i = 1, . . . , n) owned by the investor.
We consider only a single period model, that is, in between the initial time
taken as t = 0 and the final transaction time taken as t = T, no transaction ever
takes place.
Let Vi (0) and Vi (T) be the values of the i-th asset at t = 0 and t = T, respec-
tively. Let VΘ (0) and VΘ (T) denote the values of the portfolio Θ = (x1 , . . . , xn ) at
t = 0 and t = T, respectively. Then, we have

n 
n
VΘ (0) = xi Vi (0), and VΘ (T) = xi Vi (T).
i=1 i=1

VΘ (T) − VΘ (0)
The quantity rΘ (T) = is referred as the return of the portfolio Θ.
VΘ (0)
Definition 5.3.2 (Asset Weights) The weight wi of the asset ai is the propor-
tion of the value of the asset in the portfolio at t = 0, i.e.
xi Vi (0)
wi = (i = 1, . . . , n).

n
x j V j (0)
j=1

It can be observed that w1 + · · · + wn = 1.

In view of Definition 5.3.2, a portfolio can now be represented by the weights
wi , (i = 1, 2, . . . , n) such that w1 + · · · + wn = 1. Thus we can imagine that we
are having one unit of money (say Rs 1) and if its allocation in the ith asset is
wi , (i = 1, 2, . . . , n) then the resulting portfolio is (w1 , w2 , . . . , wn ).
Remark 5.3.1 In a portfolio, if wi < 0, for some i, it indicates that the investor
has taken a short position on the i-th asset ai .
Let ri be the return on the i-th asset. Then
Vi (T) − Vi (0)
ri = (i = 1, . . . , n).
Vi (0)
132 Financial Mathematics: An Introduction

Definition 5.3.3 (Mean of the Portfolio Return) Let (w1 , w2 , . . . , wn ) be a

portfolio of n assets a1 , a2 , . . . , an . Let ri (i = 1, 2, . . . , n), be the return on the ith
asset ai and E(ri ) = µi (i = 1, 2, . . . , n), be its expected value. Then the mean of the
portfolio return is defined as
n 
n 
n
µ = E( w i ri ) = wi E(ri ) = wi µi .
i=1 i=1 i=1

Definition 5.3.4 (Variance of the Portfolio) Let (w1 , w2 , . . . , wn ) be a portfolio

of n assets a1 , a2 , . . . , an . Let ri be the return on the ith asset ai and E(ri ) = µi , (i =
1, 2, . . . , n). Let Cov (ri , r j ) = σij , for i, j = 1, 2, . . . n. Then the variance of the
portfolio return is defined as

n 
n 
n
σ2 = Var( w i ri ) = wi w j σij .
i=1 i=1 j=1

Here the covariance term σij equals ρij σi σ j , where Var(ri ) = σ2i , Var(r j ) = σ2j and
ρij is the correlation coefficient between ri and r j .
In practice, the mean of the portfolio return is simply referred as the return of
the portfolio. Also the variance or rather the standard deviation of the portfolio
return is referred as the risk of the portfolio. Therefore given a portfolio A :
(w1 , w2 , . . . , wn ), we can compute its mean µA and standard deviation σA and
therefore get the point A : (σA , µA ) in (σ, µ)-plane. Thus irrespective of the number
of assets, a portfolio can always be identified as a point in the (σ, µ)-plane. This
representation is called (σ, µ)-diagram or (σ, µ)-graph (see Fig. 5.1) and is very
convenient for further discussion.
The portfolio optimization problem refers to the problem of determining weights
wi (i = 1, 2, . . . , n), such that the return of the portfolio is maximum and the risk of
the portfolio is minimum. Thus we aim to solve the following optimization problem
*n
Min i,j=1 wi w j σij
and *
Max ni=1 wi µi
subject to
w1 + w2 + . . . + wn = 1.
The above problem, is a bi-criteria optimization problem which has been studied
extensively in the mathematical programming literature. As in general, the two
 Two Assets Portfolio Optimization 133

11
00
00
11 B (σB , µB )

01A (σA , µA )

Fig. 5.1. (σ, µ)-diagram

objective functions may not attain their optimum at the same point, it is not pos-
sible to define the meaning of optimization in the usual way. Therefore we consider
either (i) to minimize risk for a given level of return or (ii) to maximize return for
a given level of risk. From the algorithmic point of view the first approach is more
convenient because it results into a linearly constrained quadratic programming
problem. We shall first discuss the two asset case and then move to the multi-asset
case.

5.4 Two Assets Portfolio Optimization

Consider a portfolio with two assets, say, a1 , a2 with weights w1 , w2 , returns r1 , r2
and standard deviations σ1 , σ2 , respectively. Then the portfolio expected return
µ and portfolio variance σ2 are respectively given by

µ = E(w1 r1 + w2 r2 ) = w1 µ1 + w2 µ2 , (5.1)
σ2 = Var(w1 r1 + w2 r2 ) = w21 σ21 + w22 σ22 + 2ρw1 w2 σ1 σ2 . (5.2)

Here ρ is the coefficient of correlation between r1 and r2 , and the value of ρ

lies in [−1, 1]. We pause here to analyze the effect of ρ on the risk involved in a
portfolio. What we shall be observing is that the value of ρ provides a measure of
the extent of diversification of portfolio so as to reduce risk. The more negative
the value of ρ, the greater are the benefits of the portfolio diversification.
As w1 and w2 are weights representing the proportions of total investment in
two assets a1 and a2 , respectively, we have w1 + w2 = 1. Moreover, in case of short
134 Financial Mathematics: An Introduction

selling, the weights can be negative. Subsequently, we write w1 = 1 − s, and so,

w2 = s, s ∈ R. Now, it follows from relations (5.1) and (5.2) that
µ = (1 − s)µ1 + sµ2 , (5.3)
σ2 = (1 − s)2 σ21 + s2 σ22 + 2ρ(1 − s)sσ1 σ2 . (5.4)
Relation (5.4) can be simplified as
σ2 = (σ21 + σ22 − 2ρσ1 σ2 )s2 − 2σ1 (σ1 − ρσ2 )s + σ21 . (5.5)
Without loss of generality we assume that 0 < σ1 ≤ σ2 . We discuss the following
two independent cases
(i) ρ = ±1 (ii) − 1 < ρ < 1.
Case (i) ρ = ±1. From relations (5.3) and (5.5), the portfolio expected return
and standard deviation are respectively given by
µ = (1 − s)µ1 + sµ2
σ = |(1 − s)σ1 ± sσ2 |.
For s ∈ [0, 1] both weights are non-negative and thus the portfolio has no short
positions. If s > 1 then w1 < 0, and therefore that asset a1 is held short. If s < 0
then w2 < 0 and therefore it indicates that asset a2 is held short. It may be noted
that an investor can not take short position on both the assets simultaneously.
For µ = (1−s)µ1 +sµ2 , σ̃ = (1−s)σ1 +sσ2 , we plot the points (σ̃, µ) in (σ̃, µ)-plane,
and get the first graph of Fig. 5.2. The second graph in Fig. 5.2 corresponds to
the case when µ = (1 − s)µ1 + sµ2 , σ̃ = (1 − s)σ1 − sσ2 . These graphs are essentially
straight lines with the bold parts corresponding to s ∈ [0, 1].
Subsequently, we plot the standard deviation-mean diagram of the portfolio,
i.e. the (σ, µ)-graph, for ρ = ±1. Observe that σ = | σ̃ | and therefore to get the
required graphics we simply need to flip the portion of the line that lies in the left
half-plane over the µ axis. The two graphs so obtained are depicted in Fig. 5.3.
From Fig. 5.3 we have the following observations to share.
Remark 5.4.1 When the returns of the two assets are perfectly positively or neg-
atively correlated, i.e. ρ = ±1, the expected portfolio return increases with the
gradual decrease in the overall risk of the portfolio (point A to point P Fig. 5.3)
till the risk is completely eliminated (point P). After that the higher return comes
with higher risk as weight of riskier asset increases. For the case ρ = +1, the point
P corresponds to a portfolio in which a short position is taken on asset a2 ; where
as for the case ρ = −1, the point P corresponds to a portfolio in which weights of
both assets are non-negative.
 Two Assets Portfolio Optimization 135

µ µ

(σ2 , µ2 )
(σ1 , µ1 ) (−σ2 , µ2 )
P
P (σ1 , µ1 )
σ̃ σ̃

ρ=1 ρ = −1
Fig. 5.2. ((
σ, µ)-graph forρ = ±1.

µ µ

(σ2 , µ2 )
(σ2 , µ2 )
P (σ1 , µ1 ) P
(σ1 , µ1 )
A
A
σ σ
ρ=1 ρ = −1
Fig. 5.3. (σ, µ)-graph forρ = ±1.

Below we provide a mathematical justification of the above observations. For

the case when ρ = 1 and σ1 = σ2 , then σmin = σ1 for all s ∈ R. Therefore we need
to consider following two cases only.
(a) ρ = 1 and σ1 < σ2 . In this case we have

σ2 = (1 − s)2 σ21 + s2 σ22 + 2(1 − s)sσ1 σ2 .

Our aim is to minimize σ, or equivalently minimize σ2 . Now

dσ2
= 2[s(σ1 − σ2 )2 − σ1 (σ1 − σ2 )] , (5.6)
ds
and
d 2 σ2
2
= 2(σ1 − σ2 )2 > 0. (5.7)
ds
136 Financial Mathematics: An Introduction

dσ2
Thus, to minimize the risk σ, we must choose the weight s such that = 0.
ds
Thereby (5.6) yields smin (the value of s for which σ2 is minimum) as
σ1
smin = < 0.
σ1 − σ2
σ2
Hence (1 − smin ) = > 0. Let µmin and σ2min respectively denote the
σ2 − σ1
expected return and variance of the portfolio with (w1 = 1 − smin , w2 = smin )
then
σ1 µ2 − σ2 µ1
µmin = and σ2min = 0.
σ1 − σ2
Since smin < 0 (i.e. w2 < 0), an investor can eliminate risk in the portfolio by
taking a short position with asset a2 .
(b) When ρ = −1 and σ1 ≤ σ2 . In this case we have

σ2 = (1 − s)2 σ21 + s2 σ22 − 2(1 − s)sσ1 σ2 .

Thus, we have
σ1
smin = >0
σ1 + σ2
σ2
1 − smin = >0
σ1 + σ2
σ1 µ2 + σ2 µ1
µmin =
σ1 + σ2
σ2min = 0.
Since smin > 0 and 1 − smin > 0 hence the investor can eliminate the risk in the
portfolio without restoring to short selling.
Case (ii) We now consider the second case when −1 < ρ < 1. Recalling relations
(5.3) and (5.5), we have

µ = (1 − s)µ1 + sµ2 ,

and

σ2 = (σ21 + σ22 − 2ρσ1 σ2 )s2 − 2σ1 (σ1 − ρσ2 )s + σ21 ,

which represents the parametric equation of a parabola in (σ2 , µ)-plane. Now

 Two Assets Portfolio Optimization 137

σ2

Fig. 5.4. (σ2 , µ)-graph for − 1 < ρ < 1.

dσ2 σ2 − ρσ1 σ2
=0 ⇒ s= 2 1 2 .
ds σ1 + σ2 − 2ρσ1 σ2

Also
d2 σ2
= 2((σ1 − ρσ2 )2 + σ22 (1 − ρ2 )) > 0.
ds2
σ2 − ρσ1 σ2
Consequently smin = 2 1 2 , and the minimum value of σ2 is given by
σ1 + σ2 − 2ρσ1 σ2

σ21 σ22 (1 − ρ2 )
σ2min = .
σ21 + σ22 − 2ρσ1 σ2

Further, the corresponding expected portfolio return equals

µmin = (µ2 − µ1 )smin + µ1 .

Remark 5.4.2 It is important to take note of the following points.

σ1
(i) The condition −1 ≤ ρ < is equivalent to 0 < smin < 1. Thus the minimum
σ2
risk can be achieved without short selling. Also, in this case

σ2min = 0 ⇔ ρ = −1.
σ1
(ii) ρ = ⇔ smin = 0 ⇔ σ2min = σ21 .
σ2
138 Financial Mathematics: An Introduction

σ1
(iii) The condition < ρ ≤ 1 is equivalent to smin < 0. In this case the investor
σ2
has taken a short position on asset a2 in order to minimize the portfolio risk.
Further, in this case

σ2min = 0 ⇔ ρ = 1.

The entire theory of this section is summarized in Fig. 5.5.

µ
B

ρ = −1 −1 < ρ < 1

P ρ=1

ρ = −1 11
00
A
σ
Fig. 5.5. Feasible region for two asset problem.

The risk-return relation of two assets for various values of ρ provides us with
a triangle APB. The points A and B signify undiversified portfolios. Since −1 ≤
ρ ≤ 1,
APB specifies the limit of diversification. The risk-return relation for
all values of ρ except ±1 lie within this triangle. Here the bold portions of the
graphs represent the case 0 ≤ s ≤ 1. We verify the two-assets portfolio theory by
considering the below given example.
Let A and B be two assets with expected returns 12% and 16% and standard
deviation 16% and 20%, respectively. Let xA and xB denote the number of units of
asset A and asset B, respectively, in a portfolio. Then we have the following table
 Multi Asset Portfolio Optimization 139

σ
xA xB µ
ρ=1 ρ = 0·5 ρ = 0 ρ = −0 · 5 ρ = −1
100 0 12·00 16·00 16·00 16·00 16·00 16·00
90 10 12·40 16·40 15·50 14·54 13·51 12·40
80 20 12·80 16·80 15·20 13·41 11·34 8·80
70 30 13·20 17·20 15·12 12·71 9·71 5·20
60 40 13·60 17·60 15·26 12·50 8·91 1·60
50 50 14·00 18·00 15·62 12·81 9·4 2·00
40 60 14·40 18·40 16·18 13·60 10·40 5·60
30 70 14·80 18·80 16·92 14·80 12·32 9·20
20 80 15·20 19·20 17·82 16·32 14·66 12·80
10 90 15·60 19·60 18·85 18·07 17·26 16·40
0 100 16·00 20·00 20·00 20·00 20·00 20·00

From the above table we observe the following.

(i) The two assets can be combined in such a way that the portfolio risk is less
than the individual risks. For instance, when the assets are taken in the ratio
80 : 20 and ρ = 0·5 then the risk in the portfolio is 15.20% whereas if the entire
investment is put in asset A only or in asset B only then the risks involved are
16% and 20%, respectively. While if ρ = −0 · 5 then 60 : 40 ratio can bring
down the risk to 8.91%. It signifies that in many circumstances diversification
of the portfolio is advisable for reducing the risk. The underline principle is
thus that ‘do not put all the eggs in one basket’.
(ii) For a given weight combination the risk reduces as ρ moves from 1 to -1.
(iii) When ρ < 1 then certain combinations of the assets are better than the others.
For example, for ρ = 0 the 60:40 combination with µ = 13.60% and σ = 12.50%
is better than the 70:30 combination with µ = 13.20% and σ = 12.71%. For
ρ = 0.5, the 70 : 30 combination yields return 13 · 20% and risk 15.12% which
is better than the 80 : 20 ratio that gives lower return 12.80% with higher risk
15.20%
From the above discussion we thus conclude that by choosing appropriate ratio
of investment between two assets, the risk can be reduced considerably.

5.5 Multi Asset Portfolio Optimization

The weights of the various assets a1 , . . . , an in the portfolio are written in the
vector form wT = (w1 , . . . , wn ). Let eT = (1, . . . , 1) ∈ Rn . Then w1 + · · · + wn = 1
140 Financial Mathematics: An Introduction

can be expressed as eT w = 1. Let mT = (µ1 , . . . , µn ) be the expected return vector

of the portfolio, where, µi = E(ri ) (i = 1, . . . , n), and C = [cij ] denotes the n × n
variance-covariance matrix with entries cij = Cov(ri , r j ) (i, j = 1, . . . , n). Note that
cii = σ2i (i = 1, . . . , n). Obviously C is a symmetric matrix. Now the expected
return µ of the portfolio is given by
⎛ n ⎞
⎜⎜ ⎟⎟ n
µ = E ⎜⎝⎜ ⎟
wi ri ⎟⎠ = wi µi = mT w ,
i=1 i=1

and the risk σ2 of the portfolio is

⎛ n ⎞
⎜  ⎟⎟ n
⎜
⎜ ⎟
σ = Var ⎜⎝
2
wi ri ⎟⎠ = cij wi w j = wT Cw. (5.8)
i=1 i, j=1

Here C is certainly positive semidefinite. In practice, it is also assumed to be

positive definite (and hence invertible) because the minimum risk of a general
n-asset portfolio is rarely zero.
During a portfolio selection, every investor is faced with a choice of either
minimizing a risk with respect to certain value of return or maximizing a return
with respect to certain value of risk. Now, from (5.8), we observe that the portfolio
risk σ2 depends on three factors, viz.,
(i) risk of each individual asset;
(ii) coefficient of correlation between assets returns;
(iii) weights of the assets.
Out of these contributing factors, the only factor that an investor can control is
the weights of the assets. Our main aim is to examine the optimal choice of these
weights.

The Feasible Region of a Portfolio Problem

Let W = {w ∈ Rn : eT w = 1} be the collection of all portfolios. We have earlier
observed that every portfolio w ∈ W corresponds to a point in the (σ, µ)-plane,
say (σ(w) , µ(w) ). The set {(σ(w) , µ(w) ) | w ∈ W} is called the feasible region or feasible
set of the given portfolio optimization problem. It should be of interest to know
the geometry of the feasible region and that is the point of our discussion now.
Consider the n-dimensional hyperplane eT w = 1 in which the weight vector
w resides. Let f be the mapping that takes each weight vector in the weight
hyperplane to the corresponding portfolio point in the (σ, µ)-plane. We try and
 Multi Asset Portfolio Optimization 141

find the image of any straight line in the weight hyperplane eT w = 1 under the
mapping f . For this we note that the parametric equation of any line in the weight
hyperplane is of the form

l(ξ) = (s1 ξ + b1 , . . . , sn ξ + bn )T
= ξs + b, −∞ < ξ < ∞,

where s = (s1 , . . . , sn )T and b = (b1 , . . . , bn )T . Let w be any point on this line. Then

µ = mT w
= mT (ξs + b)
= ξ(mT s) + (mT b).

−1 −1
Let α = (mT s) , β = − (mT b)(mT s) . Then, ξ = αµ + β. Moreover,

σ2 = wT Cw
= (ξs + b)T C(ξs + b)
= (sT Cs)ξ2 + (sT Cb + bT Cs)ξ + bT Cb
≡ γξ2 + δξ + η.

Substituting the value of ξ we get

σ2 = γ(αµ + β)2 + δ(αµ + β) + η. (5.9)

As ξ varies from −∞ to ∞, the ordered pair (σ2 , µ) traces out a parabola given
by (5.9) which lies in (σ, µ)-plane with axis parallel to σ-axis and sides open on
the right.
We are actually interested in (σ, µ)-graph. Taking the square root of σ2 , the re-
sulting curve is +
σ= γ(αµ + β)2 + δ(αµ + β) + η . (5.10)
This curve is called a Markowitz curve. Thus, each line in the weight hyperplane
is mapped onto a Markowitz curve. This phenomena is depicted in Fig. 5.6.
Remark 5.5.1 Here it is important to note that the Markowitz curve (5.10) is
not a parabola. In fact the main difference between the parabola (5.9) and the
Markowitz curve (5.10) in (σ, µ)-graph is that a tangent can be drawn to the
parabola (5.9) from any point on the µ-axis, whereas the Markowitz curve behaves
almost as a straight line as µ → ∞, thereby, it is not possible to draw a tangent to
142 Financial Mathematics: An Introduction

w2 µ
w1 + w2 + w3 = 1 Markowitz Curves

w1 σ

Weight Lines

w3
Fig. 5.6. Markowitz curves.

the Markowitz curve as µ → ∞. This difference may not sound significant right
now but it plays a vital role when the portfolio consists of one risk-free asset as
well. We shall be addressing to this type of portfolio in the next section. For the
current discussion we have assumed that all the assets in the portfolio are risky.
Remark 5.5.2 As we cover the weight hyperplane by taking weight lines, we trace
a family of Markowitz curves in the (σ, µ)-plane. It is not difficult to get convinced
that this region in the (σ, µ)-plane is going to be a solid region and its shape will
be like a bullet, which is appropriately called the Markowitz bullet.
There is another much simpler way to get convinced that the feasible set in the
(σ, µ) plane should be of the type as described above. Suppose we have three assets
a1 , a2 and a3 . Let these be represented by three points A : (σa1 , µa1 ), B : (σa2 , µa2 )
and C : (σa3 , µa3 ) in the (σ, µ)-plane. As already seen in the case of two asset
portfolio diagram, any two assets when combined to form portfolios give rise to a
curved line or a straight line between them. Whether the line will be straight or
curved depends upon the correlation coefficient between them. The three (curved)
lines between the possible three pairs are shown in the left hand side figure of Fig.
5.7. Now if we take any asset D formed by the combination of B and C, then
again combination of assets A and D will give us a (curved) line connecting A
and D and the process continues. Eventually we get the feasible region which is
a solid two dimensional region in the (σ, µ)-plane. This region is depicted in the
right hand side figure of Fig. 5.7. Further the feasible region is convex to the left,
i.e. given any two points in the region, the straight line connecting them does not
cross the left boundary.
 The Minimum Variance Set, The Minimum Variance Point and the Efficient Frontier 143

A
σ

Fig. 5.7. Feasible region for three asset problem.

In the feasible region as depicted in Fig. 5.7, the outer region corresponds to
the case when shorting is allowed and the inner region corresponds to the case
when shorting is not allowed.

5.6 The Minimum Variance Set, The Minimum Variance

Point and the Efficient Frontier
Consider the feasible set of a given portfolio optimization problem as depicted
in (σ, µ)-diagram (Fig. 5.8). Its left boundary is called the minimum variance set
because for any level of the mean rate return, the feasible point with the smallest
variance (or standard deviation) is the corresponding left boundary point. This
situation is illustrated in Fig. 5.8.
Here when the return is at level µ0 , the feasible point with the smallest variance
is P0 . Similarly for return levels µ1 and µ2 we get the respective points P1 and P2 .
Amongst all such points (σ, µ) lying on the left boundary of the feasible region,
there is a point Pmin which has the least variance. This point is called the minimum
variance point.
Mathematically, to find the minimum variance point we need to solve the fol-
lowing risk minimization problem
Min σ2 = wT Cw
subject to
eT w = 1.
The following theorem gives a closed form solution of the above problem
144 Financial Mathematics: An Introduction

P1
µ1

µ0 P0

Pmin

µ2
P2

σ
Fig. 5.8. Minimum variance set.

Theorem 5.6.1 A portfolio with minimum risk has weights given by

C−1 e
w= .
eT C−1 e
Proof. We desire to solve the following optimization problem

Min σ2 = wT Cw
subject to
eT w = 1. (5.11)

Using the Lagrange multiplier λ ∈ R, we minimize the Lagrangian

L(w, λ) = wT Cw + λ(1 − eT w). (5.12)

Note that λ is unrestricted in sign because the constraint in the risk minimiza-
tion problem is an equation eT w = 1. Now, differentiating (5.12) with respect to
w, we obtain
λ
2wT C − λeT = 0 =⇒ w = C−1 e.
2
Using (5.11), we get


λ λ 1
eT C−1 e = 1 =⇒ = T −1 .
2 2 e C e
Thus the requisite result follows. 

 The Minimum Variance Set, The Minimum Variance Point and the Efficient Frontier 145

Markowitz Efficient Frontier

Looking at the minimum variance set in the (σ, µ)-diagram, we observe that for
a given level of risk, (say σ0 ), there are two values of returns (µ(L)
0
and µ(U)
0
with
(U) (L)
µ0 > µ0 ). Since our aim in portfolio optimization is to maximize return for a
given level of risk, we shall obviously choose the larger of these two returns, i.e
µ(U)
0
. Therefore in the minimum variance set, it is only the upper half which is of
importance for investment. This upper half portion of the minimum variance set
is called the Markowitz efficient frontier. A typical efficient frontier is illustrated
in Fig. 5.9.

efficient frontier
µ(U)
0 •
• min variance point

µ(L) •
0
min variance set

σ0 σ

Fig. 5.9. Efficient frontier.

We now formalize the definition of efficient frontier and outline a procedure to

determine the same.
Definition 5.6.1 (Dominating Point and Efficient Frontier) Let P : (σ1 , µ1 )
and Q : (σ2 , µ2 ) be two points in the feasible region of a given portfolio optimization
problem. Then the point P is said to dominate Q if σ1 ≤ σ2 and µ1 ≥ µ2 .
Definition 5.6.2 (Non-dominated Point and Efficient Frontier) Let A be
a point in the feasible region. The point A is said to be a non-dominated point
if there is no point P in the feasible region which dominates A. The set of all
non-dominated points in the feasible region is called its efficient frontier.
Geometrically the efficient frontier is the upper portion of the minimum vari-
ance set as depicted in Fig. 5.9. Obviously for an investor it is only the efficient
frontier which will be of interest and therefore we should explore methodologies
to determine the same.
Before we proceed in that direction we notice that in many cases, it is more
likely that an investor provides a fixed value of the expected return, say µ, that
146 Financial Mathematics: An Introduction

he/she desires to achieve. Therefore the investor’s problem is to decide the right
investment strategy to obtain the return µ with the minimum risk. We look at
this scenario in the result to follow.
Theorem 5.6.2 For a given expected return µ, the portfolio with minimum risk
has weights given by
   T −1 
µ mT C−1 e −1 m C m µ
det C m + det C−1 e
1 eT C−1 e eT C−1 m 1
w=  T −1  . (5.13)
m C m mT C−1 e
det
eT C−1 m eT C−1 e
Proof. We wish to solve the following quadratic programming problem
1
Min σ2 = wT Cw
2
subject to
mT w = µ
(5.14)
eT w = 1.

This is a convex quadratic programming problem with unrestricted variable vector

w. Therefore we define the Lagrangian
1
L(w, α, β) = wT Cw + α(µ − mT w) + β(1 − eT w)
2
∂L
where α, β ∈ R are Lagrange multipliers. Now solving = 0 gives
∂w
wT C − αmT − β eT = 0

i.e.
w = C−1 (αm + β e). (5.15)
Substituting the value of w in (5.14), we get

(mT C−1 m)α + (mT C−1 e)β = µ

(eT C−1 m)α + (eT C−1 e)β = 1.
Solving the above two equations for α and β, we obtain
 The Minimum Variance Set, The Minimum Variance Point and the Efficient Frontier 147

   
µ mT C−1 e mT C−1 m µ
det det
1 eT C−1 m eT C−1 m 1.
α=  T −1 , β=  T −1 .
m C m mT C−1 e m C m mT C−1 e
det det
eT C−1 m eT C−1 e eT C−1 m eT C−1 e

Substituting these values in the expression (5.15) for w, we get the required ex-
pression (5.13).


Now to generate the entire efficient frontier we need to solve problems of type
(5.14) for all values of µ ∈ R. This is almost impossible. But then an extremely
interesting observation is made here. Recall from (5.14) and (5.15) that, for a
given value of return µ, the points of minimum variance must satisfy the following
system of (n + 2) linear equations in (n + 2) unknowns w ∈ Rn , α ∈ R, β ∈ R

wT C − αmT − β eT = 0
mT w = µ
eT w = 1. (5.16)

Suppose we solve the system (5.16) for two distinct values of expected return µ,
 T
say µ̄(1) and µ̄(2) . Let the two solutions be (w(1) )T = (w(1)1
, . . . , w (1)
n ), α , β
(1) (1)
and
 (2) (2)
T
(w(2) )T = (w1 , . . . , wn ), α(2) , β(2) , respectively. Then it is simple to verify that
the combination portfolio, λ(w(1) , α(1) , β(1) )T + (1 − λ)(w(2) , α(2) , β(2) )T , λ ∈ R, is
also a solution of the system (5.16) corresponding to the return λµ̄(1) + (1 − λ)µ̄(2) .
Therefore, in order to solve (5.16) for every value of µ, one is only required to solve
it for two distinct values of µ and then form the combination of the two solutions.
Thus, the knowledge of two distinct portfolios yielding the minimum variances is
sufficient to generate the entire minimum variance set. This result is significant
from investor’s point of view. Also, it demonstrates a very good application of
Karush-Kuhn-Tucker optimality conditions. The result is known as the two fund
theorem.
Theorem 5.6.3 (Two Fund Theorem) Two efficient portfolios can be estab-
lished so that any other efficient portfolio can be duplicated, in terms of mean and
variance, as a linear combination of these two assets. In other words, it says that,
an investor seeking an efficient portfolio need to invest only in the combination of
these two assets.
148 Financial Mathematics: An Introduction

The most convenient way to get two solutions of (5.16) is to assign two distinct
values to α and β, and then work out the solutions. The most convenient choices
are α = 1, β = 0 and α = 0, β = 1. The above discussion is illustrated through
the below given example.
Example 5.6.1 Consider three risky assets with the variance-covariance matrix
and expected returns as follows.

variance - covariance matrix(C) return(M)

2 1 0 0·4
1 2 1 0·8
0 1 2 0·8

Find two portfolios yielding the minimum variance. Also, determine the expected
returns from these two portfolios. Using the two fund theorem, construct the port-
folio giving the return of 33 · 4% with minimum risk.
*3 (1)
Solution Taking α = 0, β = 1 in (5.16), we need to solve: j=1 σij v j = 1 (i =
1, 2, 3), resulting in the following system of linear equations

2v(1)
1
+ v(1)
2
=1
v(1)
1
+ 2v (1)
2
+ v(1)
3
=1
(1) (1)
v2 + 2v3 = 1.

The solution is V (1) = (0 · 5, 0, 0 · 5)T . We next take α = 1, β = 0 in (5.16), to

*
solve 3j=1 σij v(2)
j
= µi (i = 1, 2, 3), i.e.

2v(2)
1
+ v(2)
2
=0·4
(2) (2) (2)
v1 + 2v2 + v3 = 0 · 8
v(2)
2
+ 2v(2)
3
= 0 · 4.

The solution of above system is V (2) = (0 · 1, 0 · 2, 0 · 3)T .

*3the (1)
Note that j=1 v j = 1, thus we take w(1) = V (1) = (0.5, 0, 0.5). Normalizing
*
V (2) , we get, w(2) = (1/6, 1/3, 1/2)T (so that, 3j=1 w(2) j
= 1). The corresponding
returns from the two portfolios with weights w and w are µ̄(1) = mT w(1) = 0 · 6
(1) (2)

and µ̄(2) = mT w(2) = 0 · 733, respectively.

Next, we consider the case when the investor desired a return of µ = 0 · 334 at
minimum risk. It is simple to check that for λ = 3, λµ̄(1) + (1 − λ)µ̄(2) = 0 · 334.
 Capital Asset Pricing Model (CAPM) 149

Thus, by the two fund theorem the requisite portfolio is given by w = λw(1) + (1 −
λ)w(2) = (7/6, −2/3, 1/2). Observe that the second asset has a short position in
this portfolio. The variance corresponding
⎛ to this
⎞ ⎛portfolio
⎞ is
 ⎜
 ⎜⎜ 2 1 0 ⎟ ⎜
⎟⎟ ⎜⎜ 7/6 ⎟⎟
⎟
w Cw = 7/6 −2/3 1/2 ⎜⎜ 1 2 1 ⎟⎟ ⎜⎜ −2/3 ⎟⎟⎟ = 2/9.
T ⎜ ⎟ ⎜
⎝ ⎠⎝ ⎠
0 1 2 1/2



5.7 Capital Asset Pricing Model (CAPM)

So far we have assumed that all assets in the portfolio are risky assets. So it is
natural to query as to what would be the scenario if one risk-free asset is included
in the portfolio? In this section we make an attempt to study this aspect of
portfolio selection.
Consider a portfolio with n risky assets, a1 , . . . , an with weights w1 , . . . , wn and
one risk-free asset arf with weight wrf . Then

n
wrisky + wrf = wi + wrf = 1. (5.17)
i=1

Also, the expected return and the variance associated with this portfolio are
respectively given by

n
µ= wi µi + wrf µrf = µrisky + wrf µrf ,
i=1

and
⎛ n ⎞ ⎛ n ⎞
⎜⎜ ⎟⎟ ⎜⎜ ⎟⎟
σ = Var ⎜⎝⎜
2
wi ri + wrf rrf ⎟⎠⎟ = Var ⎜⎝⎜ wi ri ⎟⎟⎠ = σ2risky .
i=1 i=1

If we remove the risk-free asset from the portfolio and readjust the weights of
the risky assets so that their sum remain 1, the resultant portfolio so obtained
is referred to as the derived risky portfolio. We use µder and σ2der to denote the
derived risky portfolio expected return and risk, respectively. Then,
150 Financial Mathematics: An Introduction


n
µ= wi µi + wrf µrf
i=1 ⎛ n ⎞
⎜⎜ wi ⎟⎟
= wrisky ⎜⎜⎝ µi ⎟⎟⎠ + wrf µrf (5.18)
i=1
wrisky
= wrisky µder + wrf µrf
= wrisky µder + (1 − wrisky )µrf
= wrisky (µder − µrf ) + µrf .

Also,
⎛ n ⎞
⎜ ⎟⎟
⎜
σ2 = Var ⎜⎜⎝ wi ri ⎟⎟⎠
i=1 ⎛ ⎞
⎜⎜n
w ⎟⎟ (5.19)
= wrisky Var ⎜⎜⎝ ri ⎟⎟⎠
2 i
w i=1
risky
= w2risky σ2der ,

σ
which gives wrisky = . From (5.18) and (5.19) we get
σder

µ − µ 
der rf
µ = µrf + σ, (5.20)
σder
which is an equation of the line joining (0, µrf ) and (σder , µder ) in the (σ, µ)-graph.
Now, for a given risk σ, if we choose various weight combinations of risk-free
asset and risky assets satisfying (5.17), we generate different lines represented by
(5.20) in (σ, µ)-graph. Obviously, among all such lines, the line that produces the
point with highest expected return for a given risk is tangent to the upper portion
of the Markowitz bullet. This is illustrated in Fig. 5.10.
Definition 5.7.1 (Capital Market Line) Among all the lines (5.20) for various
weight combinations of risk-free asset and risky assets, the line giving the highest
return for a given risk is called the capital market line.
Definition 5.7.2 (Market Portfolio) The point on the Markowitz bullet where
the capital market line is tangential is said to represent the market portfolio.
Theoretically, the market portfolio must contain all risky assets, for if some
asset is not in it then it will wither and die. Since the market portfolio contains
all risky assets, it is a completely diversified portfolio with no unsystematic risk.
 Capital Asset Pricing Model (CAPM) 151

µ
Capital Market Line

(σM , µM )

µr f (σder , µder )

σ
Market Portfolio

Fig. 5.10. Capital market line.

The basic idea of the capital asset pricing model (CAPM) is that an investor
can improve the risk-expected return balance by investing partially in a portfolio
of risky assets and partially in a risk-free asset. All investors will end up with
portfolios along the capital market line as all efficient portfolios lie along this line
while any other combination of risk-free asset and risky assets, except those which
are efficient, lies below the capital market line. It is thus important to observe that
all investors will hold combinations of only two assets, viz. the market portfolio M
and a risk-free asset. This fund scenario is summarized in the following theorem.
Theorem 5.7.1 (One Fund Theorem) There exists a single portfolio, namely
the market portfolio M, of risky assets such that any efficient portfolio can be
constructed as a linear combination of the market portfolio M and the risk-free
asset.
Unlike with the two fund theorem where any two efficient portfolios are suffi-
cient, in this case, the tangent portfolio is a specific portfolio.
Theorem 5.7.2 For any expected risk-free return µrf , the weight vector wM of the
market portfolio is given by

C−1 (m − µrf e)
wM = T −1 .
e C (m − µrf e)

Proof. From Fig. 5.10, we observe that for any point (σ, µ) in the Markowitz
bullet, the slope of the line joining (0, µrf ) and (σ, µ) is
152 Financial Mathematics: An Introduction

*n
µ − µrf i=1 µi wi − µrf
s= = *n .
σ 1
( i, j=1 cij wi w j ) 2

For the line joining (0, µrf ) to (σ, µ) to be a tangent line to the Markowitz bullet,
we need to solve the following optimization problem

mT w − µrf
Max
(wT Cw)1/2
subject to eT w = 1. (5.21)

The Lagrange function L : Rn × R → R for the problem (5.21) is described as

mT w − µrf
L(w, λ) = T 1/2
+ λ(1 − eT w).
(w Cw)

Now, solving (5.21) is same as maximizing L(w, λ). So, ∇w L(w, λ) = 0, giving

1  T −1/2

(w Cw)1/2
m − (m T
w − µrf )(w T
Cw) Cw = λe.
wT Cw
The above expression can be rewritten as
Cw
σm − (µ − µrf ) = λσ2 e.
σ
Multiplying by σ, we obtain

σ2 m − (µ − µrf )Cw = λσ3 e, (5.22)

which in turn yields

σ2 wT m − (µ − µrf )wT Cw = λσ3 wT e.

Since eT w = 1, µ = wT m, and σ = wT Cw, we get

µrf
λ= . (5.23)
σ
The requisite value of weight vector wM now follows from (5.22) and (5.23).


 Capital Asset Pricing Model (CAPM) 153

Remark 5.7.1 Suppose the market portfolio (σM , µM ) is known. Then, from
(5.20), the equation of the capital market line is given by

µ − µ 
M rf
µ = µrf + σ.
σM

µ − µis willing to take a positive risk σ, he/she can earn an additional
If the investor
M rf
return σ over and above the risk-free return µrf to compensate the risk
σM
µ − µ 
M rf
taken by him/her. Therefore sometimes the quantity is called the price
σM
of risk.
Example 5.7.1 Suppose a portfolio comprises of one risk-free asset with return
0.5, and three mutually independent risky assets with expected returns 1, 2, 3 and
variances 1, 1, 1, respectively. Determine the equation of the capital market line.
Solution The given information gives, mT = (µ1 , µ2 , µ3 ) = (1, 2, 3), µrf = 0.5, C =
[σij ] = I3×3 , eT = (1, 1, 1). Therefore, the weight vector of the market portfolio is
given by ⎛ ⎞
C−1 (m − µrf e) ⎜⎜ 1/9 ⎟⎟
⎜ ⎟
wM = T −1 = ⎜⎜⎜ 1/3 ⎟⎟⎟ .
e C (m − µrf e) ⎝ ⎠
5/9
Consequently, the expected return and variance of the market portfolio are
√
22 35
µM = m wM = , σM = ((wM ) CwM ) =
T T 1/2
.
9 9
Thus, the equation of the capital market line is

µ − µ 
M rf
µ = µrf + σ
√ σM
1 35
= + σ.
2 2


In practice there are certain assets which are listed in the stock called index
stocks. These limited assets are significant ones that can capture the pulse of the
whole market. The most regularly quoted market indices are broad-base indices
comprising of the stocks of large companies listed on a nation’s largest stock
exchanges, such as the American Dow Jones Industrial Average and S&P 500
Index, the British FTSE 100, the French CAC 40, the Japanese Nikkei 225. The
154 Financial Mathematics: An Introduction

Bombay Stock Exchange is the largest in India, with over 6000 stocks listed and it
accounts for over two thirds of the total trading volume in the country. The index
stocks finally help us to compute the market portfolio (σM , µM ). The knowledge
of the market portfolio yields the equation of capital market line, see Remark
5.7.1. Now suppose an investor P is willing to take risk σP . Then for this risk, the
expected return µP is maximum if the point (σP , µP ) lies on the capital market
line. Thus,
µ − µ 
M rf
µP = µrf + σP .
σM
σP
If we let wP = then
σM
µP = wP µM + (1 − wP )µrf .
Remark 5.7.2 The above relation suggests that if an investor is willing to take
σP
a risk σP , then he/she should invest wP = proportion of investment in index
σM
fund and (1 − wP ) proportion of investment in the risk-free investment schemes.
We now aim to examine how an individual asset behaves with respect to the
market portfolio. For this, we attempt to build a relationship between the expected
return along with the risk of an individual asset with the market portfolio. This
gives the CAPM formula (5.24).
Theorem 5.7.3 Suppose the market portfolio is (σM , µM ). The expected return of
an asset ai is given by
Cov (ri , rM )
µi = µrf + βi (µM − µrf ), where βi = . (5.24)
σ2M

Proof. Suppose an investor portfolio comprises of asset ai with weight w and the
market portfolio M with weight 1 − w. Then the expected return and risk of the
investor portfolio are respectively given by

µ = wµi + (1 − w)µM
σ2 = w2 σ2i + (1 − w)2 σ2M + 2ρw(1 − w)σi σM (5.25)

where ρ is the coefficient of correlation between the returns of asset ai and the
market portfolio M.
As w varies, these values trace out a curve in the (σ, µ)-graph. It can be observed
from Fig. 5.11 that as w passes through zero, the capital market line becomes
tangent to the curve at M. This tangency condition can be translated into the
 Capital Asset Pricing Model (CAPM) 155

µ
Capital Market Line

(σM , µM )

µr f (σak , µak )

σ
Fig. 5.11. Market portfolio.

condition that the slope of the curve is equal to the slope of the capital market
line at M (corresponding to w = 0).
Now the slope of the curve at M is given by
- -
dµ --- dµ dw ---
- = -
dσ -(w=0) dw dσ -(w=0)
--
dw --
= (µi − µM ) - .
dσ -(w=0)

Differentiating (5.25) with respect to w and computing its value at w = 0, we

get
-- -
dσ -- wσ2i − (1 − w)σ2M + ρσi σM (1 − 2w) --
- = --
dw -(w=0) σ -
(w=0)
σiM − σ2M
= , σiM = ρσi σM .
σM
Consequently, -
dµ --- (µi − µM )σM
-- = . (5.26)
dσ (w=0) σiM − σ2M
As discussed above, the slope of the curve needs to be equal to the slope of the
capital market line at M, thereby yielding that
µM − µrf dµ
= |(w=0) .
σM dσ
156 Financial Mathematics: An Introduction

The above relation along with (5.26), on simplification, yields

 
µM − µrf
µi = µrf + σiM
σ2M
= µrf + βi (µM − µrf ).


σiM
Remark 5.7.3 Here, βi = 2 is called the beta of an asset. Note that, for the
σM
market portfolio, βM = 1. Beta is generally calculated for individual assets using
regression analysis. As can be observed, beta measures an asset’s volatility or risk
in relation to the rest of the market. It is thus appropriately referred to as finan-
cial elasticity or correlated relative volatility, and it is all what is required to be
known about the asset’s risk characteristics in CAPM formula. In other words, an
investor ready to bear some systematic risk gets rewarded for it. For instance, if
βi = 2, it indicates that the ith asset return is expected to increase (decrease) by 2%
when the market increases (decreases) by 1%. Equivalently, if the market return
fluctuates over a specific range of values, the asset returns will fluctuate over a
larger range of values. Thus, the market risk is magnified in the asset risk.
Remark 5.7.4 If we take µrf as a base, then the above theorem  states that the
expected return of a particular asset
 over the risk-free
 asset namely (µi − µrf ) is
proportional to that of the market namely (µM − µrf ) and the constant of propor-
tionality is βi . The CAPM essentially tells that the expected excess rate of return
of an assets is directly proportional to its covariance with the market. So it is
important to understand this covariance or equivalently the beta of the asset.
If βi = 0, the given asset is completely uncorrelated with the market and there-
fore CAPM gives µi = µrf . At first this looks surprising because it tells that no
matter how risky the given asset is (even if σi is very large) the expected rate of
return of this asset cannot be improved over the base point µrf and therefore there
is no premium for risk. The main reason being that as the given asset is uncorre-
lated with the market its risk can be diversified away by having a small amount of
this asset in the portfolio.
If βi < 0 then CAPM gives µi < µrf . This again looks odd because it says that
even though asset risk σi may be large, its expected rate of return is less than
the base µrf . But then if, this asset is combined with the market it reduces the
overall portfolio risk. Therefore the investor may be willing to have this asset in
the portfolio as it has the capability of reducing its overall risk. Such an asset is
expected to do well even when other assets are not doing that well and therefore it
is called an insurance.
 Capital Asset Pricing Model (CAPM) 157

The above discussion suggests that though for a portfolio an appropriate mea-
sure of risk is σ but for an individual asset the proper measure of risk is its beta.
Thus there is a paradigm shift in understanding the risk of an asset.
Example 5.7.2 Let the risk-free rate µr f be 8% and the market has µM = 12%
and σM = 15%. Let an asset a be given which has covariance of 0.045 with the
market. Determine the expected rate of return of the given asset.
0 · 045
Solution From the given data we have βa = = 2. Then CAPM gives
(0 · 015)2

µa = µr f + β(µM − µr f ) = 0.08 + 2(0.12 − 0.08) = 0 · 16.

Therefore the expected rate of return of the given asset is 16%.



Definition 5.7.3 (Beta of the Portfolio) The overall β of the portfolio is the
weighted average of the betas of the individual assets in the portfolio, with the

n
weights being those that define the portfolio, i.e. β = wi βi .
i=1

The above formula is*immediate because for the

n *nportfolio w = (w1* , w2 , . . . , wn ),
n
the
*n rate of return is w r
i=1 i i and Cov(r, r M ) = w
i=1 i Cov(r ,
i Mr ) = i=1 wi σiM =
i=1 ρσi σM .

Definition 5.7.4 (Security Market Line) A linear equation

Cov(r, rM )
µ = µrf + β (µM − µrf ), where β=
σ2M

that describes the expected return for all assets in the market is called the security
market line.
The security market line highlights the essence of CAPM formula. It says that
under the equilibrium conditions assumed by CAPM, all portfolio investments lie
along the security market line in the beta-return space. It emphasizes that the
risk of an asset is a function of its covariance with the market, or equivalently a
function of its beta. The security market line is depicted in bold in Fig. 5.12
158 Financial Mathematics: An Introduction

Security Market Line

µM
Market Portfolio

µr f

β
1
Fig. 5.12. Security market line.

CAPM as a Pricing Formula

We now give another interpretation of CAPM as a pricing formula. Let an asset
be purchased at price P and later sold at price Q. Then the rate of return is
(Q − P)
r= . Here P is known but Q is random. If we write E(Q) = Q̄, then the
P
CAPM formula gives

Q̄ − P
= µr f + β(µM − µr f )
P
i.e.
Q̄
P= . (5.27)
1 + µr f + β(µM − µr f )

Here β is the beta of the given asset.

The formula (5.27) can be viewed as a pricing formula. Here for the random
pay-off Q, Q̄ is known and the aim is to determine the price P.
Q
For the deterministic scenario, it is well known that P = . Therefore,
1 + µr f
the quantity µr f + β(µM − µr f ) can be interpreted as risk adjusted interest rate.
This gives another interpretation of formula (5.27).
Example 5.7.3 Consider a mutual fund that invests 10% it funds at a risk free
rate of 7%. The remaining 90% is invested in a widely diversified portfolio (re-
sembling the market portfolio) which is expected to give a return of 15%. Further
it is known that the beta of the fund is 0.90, and one share of the mutual fund
costs Rs 100. Is this price of a share of the mutual fund fair? Justify your answer.
 Capital Asset Pricing Model (CAPM) 159

Solution: The value of a share after one year will be (10×1.07)+(90×1.15) = 114.20.
Thus Q̄ = 114.2. Therefore
114.20
P= = Rs 100.
(1.07) + (0.90)(0.15 − 0.07)
This shows that the price of the share, namely Rs 100, represents Rs 100 of assets
in the fund, and therefore CAPM tells that the price is right.


The CAPM as a Factor Model
The CAPM can be derived as a special case of a single factor model. Let us assume
that the asset return ri and market return rM (taken as a factor) are related as
follows
   
ri − µrf = αi + βi rM − µrf + i . (5.28)

Here µrf is the risk-free interest rate and E(i ) = 0. Also i is uncorrelated with
the market return rM and also with other  j ’s. Further αi and βi are the usual
coefficients appearing in a single factor model.
Taking expectation in (5.27) gives
   
µi − µrf = αi + βi µM − µrf . (5.29)

Here we note that (5.29) is identical with CAPM except that in CAPM, αi = 0.
If we further take the covariance of both sides in (5.29) we get

σiM = βi σ2M .

Hence,
σiM
βi = ,
σ2M

which is the same expression as used in CAPM. The equation (5.29) represents
a line between the quantities (µM − µr f ) and (µi − µr f ). This line is called the
characteristic equation or characteristic line.
The characteristic line in a sense is more general than CAPM because here
αi need not be zero. In fact αi can have a very nice economic interpretation. A
stock with non zero αi can be regarded as mispriced. If αi > 0 then in view of
CAPM, the asset is performing better than it should. Similarly if αi < 0, then it
is performing worse than it should.
160 Financial Mathematics: An Introduction

Though we have tried to explain CAPM as a single factor model, we must note
that the two are not equivalent. In CAPM we assume that the market is efficient,
but in a single factor model we have taken arbitrary covariance matrix σiM and
made no assumption on market efficiency.
The CAPM and βi can be understood from a different angle if we take the
following model
ri = rr f + βi (rM − rr f ) + i ,
where i is a random variable. Taking expectation in the above equation and
using CAPM we get E(i ) = 0. Further taking the correlation with rM in the above
equation we get Cov(i , rM ) = 0. Therefore we have

σ2i = β2i σ2M + Var(i ).

This equation tells that σ2i is the sum of two expressions. The first expression β2i σ2M
is called the systematic risk. This is the risk associated with the market as a whole.
There is no chance of reducing this risk by diversification because all assets with
nonzero beta have this risk. The second expression Var(i ) is uncorrelated with
the market and therefore can be reduced by diversification. The quantity Var(i )
is called unsystematic risk of the asset. Therefore the systematic risk measured by
β becomes more important because it directly combines with the systematic risk
of other assets.

5.8 Summary and Additional Notes

• In this chapter we analyzed the advantages of diversifying the total investment
among several assets optimally so as to get a ‘decent’ return with minimum
risk. The theory described here is mainly based on the work of Markowitz [90].
• The requisite terminologies were introduced in Section 5.2, followed by a simple
case of two assets portfolio optimization in Section 5.3, to get the feel of the
subject. The ideas of Section 5.3 lead us to extend the analysis to the multi
asset scenario in Section 5.6.
• Section 5.7 continues with the multi asset portfolio optimization with the differ-
ence that one risk-free asset is included in the portfolio. This inclusion results in
the Capital Asset Pricing model (CAPM) and a new concept of market portfo-
lio. It was shown that the market portfolio can guide the investor to determine
the advantage of taking more risk with his investment.
• The derivation of CAPM assumes the efficient market hypothesis. This hy-
pothesis tells that in an efficient market, asset prices fully reflect all available
 Summary and Additional Notes 161

information. Therefore no relevant information is ignored, and systematic er-

rors are not made. As a consequence, prices are always at levels consistent with
fundamentals.
• The contents of the chapter are kept very simple with the intention to familiar-
ize the readers with the basics behind portfolio optimization. However things
can be very complex in financial circuits. One needs to realize that any study
related to financial problems requires extra skills and high level of understand-
ing. Several interesting books covering various aspects of portfolio theory can
be referred to, like, Bartholomew-Biggs [7], Capinski and Zastwawniak [25],
Cornuejols and Tütüncü [32], Elton et al. [41], Luenberger [85], Roman [113],
Ross [114], to name a few.
• A prominent area where portfolio optimization has gained momentum in recent
years is the asset-liability management of the institutional investors, like, in-
surance companies, pension funds or mutual funds. The institutional investors
make huge investments in the markets and simultaneously repay the maturity
amounts to the other investors who had invested with them. For this reasons,
they need to constantly rebalance their portfolio after every time frame, which
is generally very small. An institutional investor will get some inflow of money
at t instance of time as the return from various investments that had been
made in the market earlier and which had subsequently matured at t time, and
also the institutional investor needs to pay the maturity amount to all those
investors who had invested with him and whose funds have matured at the end
of t − 1 time. The remaining amount is reinvested in the market. The time scale
involved in such asset-liability problems has been captured by using stochastic
linear programming models (see Kall [74] for stochastic LPP). Number of re-
search papers can be found on asset-liability management, like Sodhi [125], Yu
et al. [148] and Steuer et al. [130] and references therein.
• There are several commercial packages, for instance, CPLEX, LINGO, MAT-
LAB, SAS that provide lot of inbuilt functions for Portfolio Analysis. The major
disappointment with all the commercial packages is that they can best gener-
ate only the approximate piecewise linear representation of the efficient frontier
in portfolio optimization. With large number of assets involved, say 600-800,
the performance of these software in computing the efficient frontier deterio-
rates. The MPQ (multi-parametric quadratic programming), programmed in
Java and available in public domain, performs exceptionally well on large-scale
applications in a reasonable time and yields the exact efficient frontier. For
more on MPQ, we refer to Steuer et al. [130].
• The effect of introduction of transaction costs and/or different lending and
162 Financial Mathematics: An Introduction

borrowing rates in portfolio optimization theory has also been analyzed in

literature. Some of the books mentioned above contain subject matter on this
issue.
• Some researchers have re-visited Markowitz’s mean- variance model so as to
simplify the analysis and computations in the determination of efficient frontier.
The interested readers may refer to Chawla [27] and Steinbach [129].

5.9 Exercises
Exercise 5.1 Suppose there are three financial market scenarios Ω = {w1 , w2 , w3 }
with different probabilities of occurrence. Consider the following table showing the
returns on two different stocks in these three scenarios

scenario prob return k1 % return k2 %

w1 0·2 −10 −30
w2 0·5 0 20
w3 0·3 20 15

(a) What are the expected returns on the stocks?

(b) Suppose 60% of the available fund is invested in stock 1 and the remaining is
invested in stock 2, then what is the expected return of the portfolio?
(c) Compute the weights if the expected return on a portfolio is 20%.
Exercise 5.2 Consider the following data for two different stocks

scenario prob return k1 % return k2 %

w1 0·4 −10 20
w2 0·2 0 20
w3 0·4 20 10

Suppose a portfolio comprises of 40% of total investment in stock 1 and 60% in

stock 2. Compare the risk of the portfolio with the risks of its individual compo-
nents. What will be the risk situation if a portfolio is designed with investment of
80% in stock 1 and the remaining in stock 2.
Exercise 5.3 Prove that if short sales are not allowed then the risk of the portfolio
can not exceed the greater of the risks of the individual components of the portfolio.
 Exercises 163

Exercise 5.4 Let a portfolio be designed with investment of 50% in stock 1 and
the remaning 50% in stock 2. Further let short sale be allowed in stock 1 and all
the other data being the same as in Exercise 5.2. Does the conclusion of Exercise
5.3 hold.
Exercise 5.5 Suppose the portfolios are constructed using three securities a1 , a2 , a3
with expected returns, µ1 = 20%, µ2 = 13%, µ3 = 4%, standard deviations of re-
turns, σ1 = 25%, σ2 = 28%, σ3 = 20%, and the correlation between returns,
ρ12 = 0 · 3, ρ13 = 0 · 15 and ρ23 = 0 · 4. Among all the attainable portfolios, find
the one with minimum variance. What are the weights of the three securities in
this portfolio? Also compute the expected return and standard deviation of this
portfolio.
Exercise 5.6 Among all attainable portfolios with expected return 20% con-
structed using the data provided in Exercise 5.5, find the portfolio with minimum
variance. Compute the weights of individual assets in this portfolio.
Exercise 5.7 Consider the following data
µ σ
asset 1 10% 5%
asset 2 8% 2%
For each correlation coefficient ρ = −1, −0 · 5, 0, 0 · 5, 1, what is the combination
of the two assets that yields the minimum standard deviation and what is the
minimum value of the standard deviation?
Exercise 5.8 Compute the minimum risk portfolio for the following rate return
(%) data
Jan Feb Mar Apr May June
asset 1 12 10 5 7 15 12
asset 2 7 12 10 10 12 15
Also compute the expected return for the optimal portfolio.
Exercise 5.9 Consider three risky assets with the variance-covariance matrix and
expected returns (all data in %) as follows.
variance - covariance matrix(C) return(M)
10 4 0 5
4 12 6 6
0 6 10 1
164 Financial Mathematics: An Introduction

Find two efficient portfolios. Also construct the portfolio giving the return of 2·8%
with minimum risk. Will this portfolio be also efficient?
(Hint : use two fund theorem).
Exercise 5.10 Suppose an investor is interested in constructing a portfolio with
one risk-free asset a1 , and three risky assets a2 , a3 and a4 . Let the expected re-
turns of a1 , a2 , a3 and a4 be 6%, 10%, 12% and 18% respectively. Let the variance-
covariance matrix C of the three risky assets be
⎛ ⎞
⎜⎜ 4 20 40 ⎟⎟
⎜ ⎟
C = ⎜⎜⎜ 20 10 70 ⎟⎟⎟ .
⎝ ⎠
40 70 14
Determine all efficient portfolios for the investor.
Exercise 5.11 Consider the data of two risky assets a1 , a2 with µ1 = 12·5%, µ2 =
10 · 5%, σ1 = 14 · 9%, σ2 = 14%, ρ = 0 · 33.
(a) Is it advisable to diversify the investment? If so then what composition of the
assets will minimize the risk?
(b) What is the minimum value of the risk?
(c) If the risk-free rate of return is 5% then derive the equation of the capital
market line?
Exercise 5.12 Given the following information about the one risk-free asset and
three risky assets, find the expected return and standard deviation of the market
portfolio. Also determine the equation of the capital market line.

µr f = 5%, µ1 = 14%, µ2 = 8%, µ3 = 20% ;

σ1 = 6%, σ2 = 3%, σ3 = 15% ; σ12 = 0 · 5%, σ13 = 0 · 2%, σ23 = 0 · 4%.

Exercise 5.13 Assume that the following assets are correctly priced according to
the security market line. Derive the security market line.

µ1 = 6%, β1 = 0 · 5; µ2 = 12%, β2 = 1 · 5.

What is the expected return on an asset with β = 2?

Exercise 5.14 If the following two assets are correctly priced according to the
security market line, what is the return of the market portfolio? What is the risk-
free return?

µ1 = 9 · 5%, β1 = 0 · 8; µ2 = 13 · 5%, β2 = 1 · 3.
 Exercises 165

Exercise 5.15 Let the expected rate of return on the market portfolio be 23% and
that of the risk free asset be 7%. Also let the standard deviation of the market
portfolio be 32% and let us assume that the market is efficient.
(a) What is the equation of capital market line?
(b) If Rs 300 is invested in the risk free asset, and Rs 700 in the market portfolio
then what is the expected return at the end of the year?
(c) If an investor has Rs 1000 to invest and he/she desires a return of 39%, then
what should be his/her portfolio?
6
Portfolio Optimization-II

6.1 Introduction
We have presented Markowitz’s mean variance model for portfolio optimization
in the last chapter. But contrary to its theoretical reputation, this model in its
original form has not found much favor with the practitioners to construct large
scale portfolios. There are several theoretical and practical reasons for not using
this model extensively in practice - particularly when the number of assets in the
portfolio is large. This chapter aims to understand these reasons and then present
some other models which have been developed to improve Markowitz’s model both
theoretically and computationally.

6.2 Markowitz’s Model: Some Theoretical and

Computational Issues
Theoretically, Markowitz’s model is known to be valid if returns ri s are multivari-
ate normally distributed and the investor is risk averse in the sense that he/she
prefers less standard deviation of the portfolio to more. But one is not fully con-
vinced of the validity of the standard deviation as a measure of risk. An investor is
certainly unhappy to have small or negative profit, but feels happy to have larger
profit. In other words, this means that the investor’s perception about risk is not
symmetric about the mean. There are several empirical studies, which reveal that
most ri are not normally or even symmetrically distributed. In this scenario, one
possible approach seems to be to consider the skewness and kurtosis of the distri-
bution in addition to the mean and variance and extend the Markowitz’s model to
generate the efficient frontier in (mean, variance, skewness, kurtosis)-space. This
has been the approach of some of the models, e.g. Konno and Suzuki [79], and
168 Financial Mathematics: An Introduction

Joro and Na [71], but has not found much favor because the resulting optimiza-
tion problem is not easy to handle. An alternative and popular approach is to
introduce certain new measures of risk which carry information about the possi-
ble portfolio losses implied by the tail of the return distribution, even in the case
when the distribution is not symmetric. This takes care of those situations where
the return distribution is heavily tailed. These risk measures are called downside
or safety-first risk measures which aim to maximize the probability that the port-
folio loss is below a certain acceptable level, commonly referred as the benchmark
or the disaster level. Thus these risk measures are quantile based risk measures
and are different from standard deviation or other moment based risk measures.
Some of the most popular quantile based risk measures are value at risk (VaR)
and conditional value at risk (CVaR). Since downside risk measures of individual
securities cannot be easily aggregated into portfolio downside risk measures (we
need the entire joint distribution of security returns), their application in prac-
tice requires computationally intensive non parametric estimation, simulation and
optimization techniques.
There is another major problem associated with the classical Markowitz’s
model. This model gives us an optimal portfolio assuming that we have perfect
information about µi ’s and σij ’s for the assets that we are considering. Therefore
an important practical issue is the estimation of the µi ’s and σij ’s. A reasonable
approach for estimating these data is to use time series of past returns rit which
represents the return of ith asset from time (t − 1) to t, where t = 1, 2, . . . , T. How-
ever, it has been observed that small changes in the time series rit lead to changes
in the µi ’s and σi j ’s that often lead to significant changes in the optimal portfolio.
This is a fundamental weakness of the Markowitz model, no matter how cleverly
µi ’s and σij ’s are computed. This is because the optimal portfolio construction is
very sensitive to small changes in the data. Only one small change in one µi may
produce a totally different portfolio. In fact recent research (Chopra and Ziemba
[30]) has revealed that errors in the estimation of means µi can be more damaging
than errors in other parameters. This has motivated researchers to employ robust
optimization techniques in Markowitz’s model, e.g. Ben-Tal [9], and Tütüncü and
Koenig [139]. A much simpler approach is to consider portfolio optimization under
a minimax rule (Cai et al. [23]) and provide some flexibility by allowing µi to lie
in some interval ai ≤ µi ≤ bi (Deng et al. [36]).
The mean-variance model of Markowitz, in general, results in a dense quadratic
programming problem. If the number of assets in the portfolio is large then it
becomes very difficult to obtain an optimal solution of such large-scale dense
quadratic programming problem on a real time basis. This has motivated re-
 Mean Absolute Deviation Based Portfolio Optimization: A L1 -Risk Model 169

searchers to consider mean-absolute deviation of the portfolio as a measure of risk

(Konno and Yamazaki [80]) so that the resulting optimization problem reduces to
a linear programming problem. These models are called L1 -risk models because
these are based on L1 metric on Rn . In this terminology, Markowitz’s model can be
termed as a L2 -risk model since it uses variance as a risk measure which is based
on the notion of L2 metric. In the same spirit, the models based on the minimax
rule portfolio selection strategies can be termed as L∞ -risk models. It is simple to
note that L1 and L∞ risk models will result in a linear programming formulation.
This is in contrast to an L2 - risk model which results in a quadratic programming
formulation. Therefore computationally, L1 and L∞ risk models are easier to han-
dle in comparison to L2 risk model. The L1 , L2 and L∞ are three standard moment
based risk models which have been studied in the literature.
The above discussion suggests that the choice of a proper risk measure is very
crucial to study portfolio selection problems. But then what would be the guiding
principles for this choice? The concept of coherent risk measures (Artzner et al. [5])
is an important contribution in this regard. This stipulates an axiomatic approach
to the study of risk measure by presenting certain desirable properties.
Some other issues associated with Markowitz’s model are very natural and
these have been incorporated in the original model. These are with regard to its
extension for the multi-period scenario and also to incorporate transaction costs.
We shall discuss some of the models in the subsequent sections which addresses
the above issues.

6.3 Mean Absolute Deviation Based Portfolio

Optimization: A L1 -Risk Model
We first introduce the L1 -risk measure and formulate the corresponding portfolio
optimization problem with this risk measure. Let an investor has initial wealth
M0 which is to be invested in n assets ai (i=1,2,. . . ,n). Let ri be the return of the
asset ai which is a random variable. Also let xi be the amount of money to be
invested in the asset ai out of the total fund M0 .
We now define µi = E(ri ) and qi = E(|ri − µi |), i=1,2,. . . ,n. Then µi denotes the
expected return rate of the asset ai and qi denotes the expected absolute deviation
of ri from its mean. Then the expected return of a portfolio (x1 , x2 , . . . , xn ) is given
by ⎛ n ⎞
⎜⎜ ⎟⎟ n
µ = E ⎜⎝⎜ ⎟
ri xi ⎟⎠ = µi xi .
i=1 i=1
170 Financial Mathematics: An Introduction

We now define L1 -risk measure or the mean absolute deviation of the portfolio
(x1 , x2 , . . . , xn ).
Definition 6.3.1 (L1 -Risk Measure of a Portfolio) Let (x1 , x2 , . . . , xn ) be the
given portfolio. Then its L1 -risk measure or mean absolute deviation is defined as
⎡-- n ⎛ n ⎞--⎤
⎢⎢-- ⎜
⎜
 ⎟⎟--⎥⎥
wL1 (x1 , x2 , . . . , xn ) = E ⎢⎢⎢⎣-- ri xi − E ⎜⎜⎝ ri xi ⎟⎟⎠--⎥⎥⎥⎦ .
- i=1 i=1
-

In terms of the L1 -risk measure wL1 (x1 , x2 , . . . , xn ), the L1 -risk model of the portfolio
optimization problem is formulated as
⎡-- n --⎤
⎢⎢-- --⎥⎥
Min wL1 (x1 , x2 , . . . , xn ) = E ⎢⎢⎢⎣-- (ri − µi )xi --⎥⎥⎥⎦
- i=1 -
subject to

n
µi xi ≥ αM0
i=1

n
(6.1)
xi = M0
i=1
0 ≤ xi ≤ ui (i = 1, 2, . . . , n),

where (α > 1) is a parameter representing the minimum rate of return required by

the investor. Also ui is the maximum amount of the money which can be invested
in the asset ai .
In (6.1) it may be noted that by defining wi = xi /M0 , (i = 1, 2, . . . , n), the
portfolio (x1 , x2 , . . . , xn ) can also be represented in terms of its weight vector
(w1 , w2 , . . . , wn ). Traditionally in the portfolio optimization literature, a portfo-
lio has been represented by its weight vector (w1 , w2 , . . . , wn ) but it can also be
equivalently represented in terms of its amount allocation vector (x1 , x2 , . . . , xn ).
In practice, the historical data is used to estimate the parameters in the opti-
mization problem (6.1). Let rit be the realization of the random variable ri during
the period t (t = 1, 2, . . . , T). We assume that rit is available through the historical
data and the expected value of ri can be approximated by the average derived
from the data. This gives
1
T
µi = E(ri ) = rit . (6.2)
T t=1
Also wL (x1 , x2 , . . . , xn ) can be approximated by
 Mean Absolute Deviation Based Portfolio Optimization: A L1 -Risk Model 171

⎛-- n --⎞ -
T --
--
⎜⎜-- --⎟⎟ 1  --
n --
E ⎜⎜⎜⎝-- (ri − µi )xi --⎟⎟⎟⎠ = (r − µ )x -- . (6.3)
T t=1 -- i=1
it i i
- i=1 - -

In (6.3) it may be noted that, due to the absolute value function, the expres-
sion on the right hand side becomes a nonlinear and non smooth function of
(x1 , x2 , . . . , xn ).
Using (6.2) and (6.3), problem (6.2) can be reformulated as
- --
T - n
1  -- -
Min -- (rit − µi )xi ---
T t=1 - i=1 -
subject to

n
µi xi ≥ αM0
i=1

n
(6.4)
xi = M0
i=1
0 ≤ xi ≤ ui (i = 1, 2, . . . , n).
-- --
-- n --
If we now denote -- (rit − µi )xi -- by yt and employ the definition of the absolute
- i=1 -
function then we get
-- --
-- n --
yt = -- (rit − µi )xi --
- i=1 -
⎛ n ⎞
 n
⎜⎜⎜ ⎟⎟
= Max ⎜⎝ (rit − µi )xi , − (rit − µi )xi ⎟⎟⎠ . (6.5)
i=1 i=1

Therefore problem (6.4) gets transformed to

172 Financial Mathematics: An Introduction

1
T
Min yt
T t=1
subject to

n
(rit − µi )xi ≤ yt
i=1

n
− (rit − µi )xi ≤ yt
i=1

n
(6.6)
µi xi ≥ αM0
i=1

n
xi = M 0
i=1
0 ≤ xi ≤ ui (i = 1, 2, . . . , n),

where the first two constraints in (6.6) follow from (6.5) and the definition of
maximum.
Denoting (rit − µi ) by cit (i = 1, 2, . . . , n; t = 1, 2, . . . , T), we can rewrite (6.6) as

1
T
Min yt
T t=1
subject to

n
yt − cit xi ≥ 0 (t = 1, 2, . . . , T)
i=1
n
yt + cit xi ≥ 0 (t = 1, 2, . . . , T)
i=1

n
(6.7)
µi xi ≥ αM0
i=1

n
xi = M0
i=1
0 ≤ xi ≤ ui (i = 1, 2, . . . , n) .

Remark 6.3.1 The formulation (6.4) is essentially nonlinear and non smooth.
But since this nonlinearity and non smoothness occurs due to the presence of
absolute value function only, it can be handled in a reasonably simple manner
 Mean Absolute Deviation Based Portfolio Optimization: A L1 -Risk Model 173

as explained above. The resulting linear programming problem (6.7) can be solved
efficiently even when n is large.
One obvious question at this stage is to enquire if there is any relationship
between L1 and L2 -risk models. The below given theorem answers this question.
Theorem 6.3.1 Let (r1 , r2 , . . . , rn ) be multivariate normally distributed. Then for
a given portfolio x = (x1 , x2 , . . . , xn )

2
wL1 (x) = σ(x),
π
where the standard deviation σ(x) is given by
.
/
0 ⎡⎧ n
/ ⎛ n ⎞⎫ 2⎤
⎢⎢⎪
⎪  ⎜  ⎟ ⎪
⎪ ⎥
⎢⎨ ⎜⎜ ⎟⎟⎬ ⎥⎥⎥
σ(x1 , x2 , . . . , xn ) = E ⎢⎢⎢⎪ r x − E ⎜
⎝ r x ⎟
i i ⎠⎪ ⎥.
⎣⎪
⎩ i i
⎭ ⎥⎦
⎪
i=1 i=1

Proof. Let (σi j ) ∈ Rn×n be the variance-covariance matrix of (r1 , r2 , . . . , rn ). Then

n  n
under the given hypothesis, ri xi is normally distributed with mean ri µi and
i=1 i=1
standard deviation .
0 n n

σ(x) = σij xi x j .
i=1 j=1

Therefore
, ∞  
1 u2
wL1 (x) = √ | u | exp − 2 du
2πσ(x) −∞ 2σ (x)
, ∞  
2 u2
= √ u exp − 2 du ,
2πσ(x) 0 2σ (x)

which on substitution (u2 /2σ2 ) = s gives

 , 
2 σ2 (x) ∞ −s 2
wL1 (x) = e ds = σ(x).
π σ(x) 0 π


In view of Theorem 6.3.1, for the multivariate normal case, both L1 and L2 -
risk models are equivalent. In other words, Markowitz’s mean-variance portfolio
174 Financial Mathematics: An Introduction

selection strategy will be the same as the one given by mean-absolute deviation
selection strategy. Even in the case when normality assumption does not hold,
through certain case studies, it has been shown that minimizing the L1 -risk pro-
duces portfolios which are comparable to Markowitz’s mean-variance model which
minimizes L2 -risk, i.e. standard deviation. However as one expects the variance of
mean-absolute deviation portfolio is always at least as large as the corresponding
mean-variance portfolio. But in actual applications, this difference is small.
In fact Konno and Yamazaki [80] applied both L1 and L2 -risk models in Tokyo
Stock Market by using historical data of 224 stocks in NIKKEL 225 index. They
generated efficient frontiers and observed that the difference of the standard devi-
ation of the optimal portfolio generated by L2 and L1 -risk models is at most 10%
for what ever value of α. Of course two frontiers will coincide if ri ’s are multivari-
ate normally distributed. Thus this difference can be largely attributed to the non
normality of the data. Therefore irrespective of the distribution scenario, L1 -risk
model provides a good alternative to Markowitz’s L2 -risk model.

Advantages of the Formulation (6.7)

We now list some of the advantages of L1 -risk model (6.7) over the classical
Markowitz’s model which minimizes the L2 -risk.
(i) The formulation (6.7) is a linear programming problem and hence can be solved
much more efficiently in comparison to quadratic programming problem which
is obtained in Markowitz’s model. This is particulary significant for portfolios
having large number of assets.
(ii) We do not need to calculate the variance-covariance matrix to set up the
L1 -risk model.
(iii) In the formulation (6.7) there are always (2T+2) constraints regardless of
the number of assets included in the model. This allows to handle very large
portfolios on a real time basis.
(iv) An optimal solution x∗ = (x∗1 , x∗2 , . . . , x∗n ) of problem (6.7) contains at most
(2T+2) positive components if ui = +∞ for i = 1, 2, . . . , n. This means that
an optimal portfolio consists of at most (2T+2) assets regardless of the size of
n. Therefore we can use T as a control variable when we wish to restrict the
number of assets in the portfolio.
 Minimax Rule Based Portfolio Optimization: An L∞ -Risk Model 175

6.4 Minimax Rule Based Portfolio Optimization: An

L∞ -Risk Model
We discuss the minimax rule based portfolio optimization model due to Cai et
al. [23]. We first introduce the L∞ -risk measure which is based on L∞ -metric on
Rn . We then discuss the corresponding portfolio optimization model, namely the
L∞ -risk model, and compare the same with Markowitz’s mean-variance model.
We continue with the notations introduced in Section 6.3. Thus M0 denotes
the initial wealth to be invested in n assets ai (i = 1, 2, . . . , n). Further for (i =
1, 2, . . . , n), ri denotes the return of the asset ai and xi denotes the allocation from
M0 in this asset. Then µi = E(ri ) and qi = E(|ri − µi |) respectively denote the
expected return rate of the asset ai and the expected absolute deviation of ri from
its mean µi .
Let us now introduce the L∞ -risk measure. For this let

n
F = {x = (x1 , x2 , . . . , xn ) : xi = M0 , xi ≥ 0 (i = 1, 2, . . . , n)}. (6.8)
i=1

Then x ∈ F is a portfolio of the assets a1 , a2 , . . . , an and F is called the feasible

region of the portfolio optimization problem. As explained earlier, the expected

n
return of the portfolio x : (x1 , x2 , . . . , xn ) is given by µ = µi xi . We now have the
i=1
following definition.
Definition 6.4.1 (L∞ -Risk Measure) Let x : (x1 , x2 , . . . , xn ) be the given port-
folio. Then its L∞ -risk measure is defined as

wL∞ (x1 , x2 , . . . , xn ) = Max E( |ri xi − µi xi | ) = Max (qi xi ).

1≤i≤n 1≤i≤n

In terms of the L∞ -risk measure wL∞ (x1 , x2 , . . . , xn ), the L∞ -risk model of the
portfolio optimization problem, denoted by (POL∞ ) is formulated as
⎛ ⎞
⎜⎜  n
⎟⎟
Min ⎜⎜⎝Max (qi xi ), − µi xi ⎟⎟⎠ . (6.9)
x∈F 1≤i≤n
i=1

The problem (6.9) is a bi-criteria optimization problem as there are two ob-
jectives involved in the optimization. These two objectives are the portfolio risk
and the portfolio return. This scenario is again same as for the L1 and L2 -risk
176 Financial Mathematics: An Introduction

models studied earlier. There we had minimized the portfolio risk for a given level
of portfolio return. Following the same strategy, we get the following problem


Min Max (qi xi )
x∈F 1≤i≤n
subject to

n
µi xi ≥ αM0 . (6.10)
i=1

Writing y = Max (qi xi ), we rewrite (6.11) as

1≤i≤n

Min y
x∈F
subject to

n
µi xi ≥ αM0
(6.11)
i=1
qi xi ≤ y (i = 1, 2, . . . , n).
For each α, (6.11) is a linear programming problem and hence can be solved
efficiently. However to generate the entire efficient frontier we need to solve (6.11)
for every α, which is practically not possible. In the case of Markowitz’s L2 -risk
model, the famous two fund theorem comes to our rescue which allows to generate
the entire efficient frontier from the knowledge of only two efficient points. In the
case of L∞ -risk model, we do not have any analogue of two fund theorem, so we
look to some other suitable option which does not require simplex algorithm to
solve (6.11) for every α, but rather gives a solution in close form for every level
of return. This is the approach which we shall be following now and describe the
details below.
The solution of the bi-criteria optimization problem (6.9) is to be understood in
terms of an efficient point which in this context is termed as an efficient portfolio.
Definition 6.4.2 (Efficient Portfolio) A feasible portfolio x̄ = (x̄1 , x̄2 , . . . , x̄n ) ∈
F is said to be an efficient portfolio if there does not exist any other portfolio
x = (x1 , x2 , . . . , xn ) ∈ F such that
(i) Max (qi x̄i ) ≤ Max (qi xi ),
1≤i≤n 1≤i≤n

and

n 
n
(ii) µi x̄i ≥ µi xi .
i=1 i=1
 Minimax Rule Based Portfolio Optimization: An L∞ -Risk Model 177

The collection of all efficient portfolios is called the efficient frontier. As ex-
plained in the context of Markowitz’s model, an investor is always interested in
determining the efficient frontier so as to select that portfolio which gives maxi-
mum return for the chosen level of risk. This principle remains valid for L∞ -risk
model as well.
The problem (POL∞ ) for the L∞ -risk model can be rewritten as

n
Min (y , − µi xi )
i=1
subject to
qi xi ≤ y (i = 1, 2, . . . , n)
(6.12)
x ∈ F.

Now borrowing the standard technique from the multi-objective literature, we

introduce the following parametric optimization problem PO(λ) for 0 < λ < 1

n
Min λy + (1 − λ)(− µi xi )
i=1
subject to
qi xi ≤ y (i = 1, 2, . . . , n)
(6.13)
x ∈ F.

We now have the following lemma connecting problems (POL∞ ) and PO(λ).
Lemma 6.4.1. The pair (x̄, ȳ) is an efficient solution of (POL∞ ) if and only if
there exists 0 < λ < 1 such that (x̄, ȳ) is an optimal solution of PO(λ).
Here λ can be considered as an investor’s risk tolerance parameter which helps to
do appropriate trade-off between the risk and the return.
In view of Lemma 6.4.1, finding efficient frontier of (POL∞ ) is equivalent to
finding solution of PO(λ) for all 0 < λ < 1. The discussion given below illustrates
that given an arbitrary λ̄ ∈ (0, 1), the solution (x̄, ȳ) of PO(λ̄) can be obtained
explicitly in a close form and there is no need of applying any simplex like nu-
merical optimization algorithm. This greatly reduces the computational burden
in determining the efficient frontier.
Analytical Solution of PO(λ)
Without any loss of generality we can assume that (i) µ1 ≤ µ2 ≤ . . . ≤ µn and (ii)
there do not exist two assets ai and a j , i  j, such that µi = µ j and qi = q j . There
178 Financial Mathematics: An Introduction

is obviously no problem with the ordering of asset returns µi and therefore the
first assumption can always be met. The second assumption is also valid because
if µi = µ j and qi = q j for two assets i, j (i  j), then we may treat them as a single
aggregated asset. We first consider the case when all assets are risky assets.
Theorem 6.4.1 Let all asset a1 , a2 , . . . , an be risky assets. Then for any 0 < λ < 1,
an optimal solution (x∗ , y∗ ) of the problem PO(λ) is given by

y∗ /qi i ∈ T∗ (λ)
x∗i =
0 otherwise,

where ⎛ ⎞−1
⎜⎜  ⎟⎟
y∗ = M0 ⎜⎜⎜⎝ (1/ql )⎟⎟⎟⎠ ,
l∈T∗ (λ)

and the set T∗ (λ) is identified by the following rule.

(a)

(µn − µn−1 ) λ
<
qn 1−λ
(µn − µn−2 ) (µn−1 − µn−2 ) λ
+ <
qn qn−1 1−λ
.. .. ..
. . .
(µn − µn−k ) (µn−1 − µn−k ) (µn−k+1 − µn−k ) λ
+ + ... + < ,
qn qn−1 qn−k+1 1−λ

and
(µn − µn−k−1 ) (µn−1 − µn−k−1 ) (µn−k − µn−k−1 ) λ
+ + ... + ≥ .
qn qn−1 qn−k 1−λ

Then
T∗ (λ) = {n, n − 1, . . . , n − k}. (6.14)
(b) Otherwise
T∗ (λ) = {n, n − 1, . . . , 1}. (6.15)
Proof. We apply the Karush Kuhn-Tucker (KKT) conditions to PO(λ). For this,
we first introduce the Lagrangian of PO(λ) as
 Minimax Rule Based Portfolio Optimization: An L∞ -Risk Model 179

 
n  
n
L(x, y, β, δ, γ) = λy + (1 − λ) − µi xi + βi (qi xi − y)
i=1 i=1

n   n
+δ xi − M0 − γi xi ,
i=1 i=1

where x ∈ Rn , y ∈ R, β ∈ Rn , δ ∈ R and γ ∈ Rn . We now write the KKT conditions

that an optimal solution of PO(λ) must satisfy. These are

∂L  n
=λ− βi = 0 (6.16)
∂y i=1
∂L
= −(1 − λ)µi + βi qi + δ − γi = 0 (i = 1, . . . , n) (6.17)
∂xi

n
xi = M0 (6.18)
i=1
(qi xi − y)βi =0 (i = 1, . . . , n) (6.19)
γi xi =0 (i = 1, . . . , n) (6.20)
βi ≥0 (i = 1, . . . , n) (6.21)
γi ≥0 (i = 1, . . . , n). (6.22)

Our aim now is to obtain a solution of the above system. Define T∗ (λ) = {i :
βi > 0}. This means that for (6.19) to be satisfied we must have

qi xi − y = 0 for i ∈ T∗ (λ) ,

i.e.
y
xi = , for i ∈ T∗ (λ). (6.23)
qi
We let

xi = 0, for i  T∗ (λ). (6.24)


n   
But M0 = xi = xi + xi = xi , and hence (6.23) gives
i=1 i∈T∗ (λ) iT∗ (λ) i∈T∗ (λ)

M0 = y (1/qi ) (6.25)
i∈T∗ (λ)
180 Financial Mathematics: An Introduction

i.e.
⎛ ⎞−1
⎜⎜  1 ⎟⎟
y = M0 ⎜⎜⎜⎝ ⎟⎟⎟ .
⎠ (6.26)
∗
ql
l∈T (λ)

Therefore, we have
⎧ ⎛ ⎞−1
⎪
⎪ ⎜⎜  1 ⎟⎟
⎪
⎪ M0 ⎜⎜ ⎟⎟ ,
∗
⎪
⎨ ⎜⎝ ⎟⎠ i ∈ T∗ (λ)
xi = ⎪
⎪ qi q (6.27)
⎪
⎪ ∗
l∈T (λ)
l
⎪
⎩ 0, otherwise.

From (6.20), for xi > 0 we have γi = 0. Hence γi = 0 whenever i ∈ T∗ (λ). Thus

from (6.17), we have for i ∈ T∗ (λ)
1 
βi = (1 − λ)µi − δ . (6.28)
qi
From (6.16) and (6.28), we get
 1   
λ= (1 − λ)µl − δ = (1 − λ) (µl /ql ) − δ (1/ql ).
∗
ql ∗ ∗
l∈T (λ) l∈T (λ) l∈T (λ)

Therefore we get
⎛ ⎞−1 ⎛ ⎞
⎜⎜  1 ⎟⎟ ⎜⎜  µl ⎟⎟
δ = ⎜⎜⎜⎝ ⎟⎟ ⎜⎜(1 − λ)
⎟⎠ ⎜⎝ − λ⎟⎟⎟⎠ . (6.29)
∗
ql ∗
ql
l∈T (λ) l∈T (λ)

Thus, from (6.28) for i ∈ T∗ (λ)

⎡ ⎛ ⎞−1 ⎛ ⎞⎤
⎢
⎢
1 ⎢⎢ ⎜
⎜⎜  ⎟
1 ⎟⎟ ⎜⎜⎜  µ ⎟⎟⎥⎥⎥
− λ⎟⎟⎟⎠⎥⎥⎥ .
l
βi = ⎢⎢(1 − λ)µi − ⎜⎝⎜ ⎟ ⎜
⎟⎠ ⎜⎝(1 − λ) (6.30)
qi ⎣ ∗
ql ∗
ql ⎦
l∈T (λ) l∈T (λ)

Also from (6.17) and noting that for i  T∗ (λ), βi = 0, we have for i  T∗ (λ),

γi = −(1 − λ)µi + δ. (6.31)

Clearly, if it is possible to determine the set T∗ (λ) which ensures that βi and γi
as given by (6.30) and (6.31) are all non negative, then y and xi as given by (6.26)
and (6.27), respectively, will be a solution of KKT conditions (6.16) to (6.22).
 Minimax Rule Based Portfolio Optimization: An L∞ -Risk Model 181

Now to prove our theorem, it is required that if there exists an integer k ∈

[0, n − 2] such that inequalities specified in (a) hold, then T∗ (λ) as given by (6.14)
will ensure that µi ≥ 0 and γi ≥ 0. The following analysis proves the above
arguments.
By (6.30), it follows that, for any i ∈ T∗ (λ) = {n, n − 1, . . . , n − k},
⎛ ⎞−1 ⎡ ⎤
⎜⎜  qi ⎟⎟ ⎢⎢  µi − µl ⎥⎥
βi = ⎜⎜⎜⎝ ⎟⎟ ⎢⎢(1 − λ)
⎟⎠ ⎢⎣ + λ⎥⎥⎥⎦
ql ql
l∈T∗ (λ) l∈T∗ (λ)
⎛ ⎞−1 ⎡ ⎛ n ⎞ ⎤
⎜⎜  qi ⎟⎟ ⎢⎢ ⎜⎜  µl − µi λ ⎟⎟  i
µ − µ ⎥⎥
= ⎜⎜⎜⎝ ⎟⎟⎟ (1 − λ) ⎢⎢⎢ ⎜⎜⎜− ⎟⎟ + ⎥⎥ ,
i l
⎠ ⎣ ⎝ + ⎟
⎠ ⎥⎦
∗
ql ql (1 − λ) ql
l∈T (λ) l=i+1 l=n−k

which by inequalities in (a), is strictly positive. On the other hand, for i =

1, 2, . . . , (n − k − 1), from (6.31) and (6.29), we have
⎛ ⎞−1 ⎛ ⎞
⎜⎜  1 ⎟⎟ ⎜⎜  µl ⎟⎟
γi = −(1 − λ)µi + ⎜⎜⎜⎝ ⎟⎟ ⎜⎜(1 − λ)
⎟⎠ ⎜⎝ − λ⎟⎟⎟⎠
ql ql
l∈T∗ (λ) l∈T∗ (λ)
⎛ ⎞−1 ⎛ ⎞
⎜⎜  1 ⎟⎟ ⎜⎜  µl − µi ⎟⎟
= ⎜⎝ ⎜
⎜ ⎟
⎟ ⎜
⎜ − λ⎟⎟⎟⎠
⎟⎠ ⎜⎝(1 − λ)
ql ql
l∈T∗ (λ) l∈T∗ (λ)
⎛ ⎞−1 ⎛ ⎞
⎜⎜  1 ⎟⎟ ⎜⎜  µl − µn−k−1 ⎟⎟
≥ ⎜⎜⎜⎝ ⎟⎟⎟ ⎜⎜⎜(1 − λ)
⎠ ⎝ − λ ⎟⎟⎟ ,
⎠
∗
ql ∗
q l
l∈T (λ) l∈T (λ)

which by (6.14), is nonnegative. The above details show that the KKT conditions
(6.21) and (6.22) are satisfied. In the case when there does not exist any integer
k ∈ [0, n − 2] such that inequalities in (a) are satisfied, we need to show that the
solution given by (6.26) and (6.27) with T∗ (λ) = {n, n − 1, . . . , 2, 1}, will satisfy all
KKT conditions. To do this, we introduce a dummy asset a0 with µ0 = −L and
q0 = L, where L is a sufficiently large positive number. Now following in the same
manner as above, we can show that all KKT conditions are satisfied.
Thus above analysis shows that the KKT conditions (6.16) to (6.22) are satis-
fied, if we select T∗ (λ) by the selection procedure (6.14) to (6.15) and solutions y
and xi are given by (6.26) and (6.27). Now since PO(λ) is a convex programming
problem the KKT conditions become necessary and sufficient for optimality. So
solution given by (6.12) to (6.15) is an optimal solution. This proves the theorem.


182 Financial Mathematics: An Introduction

We next consider the case when one risk-free asset is included in the portfolio.
To be specific, let the asset a1 be risk-free. It is natural to assume that the risk-free
asset has the lowest return. This is because choosing assets having lower returns
than that of the risk-free asset in the portfolio can not make it optimal.
Now for the risk-free asset a1 , we have q1 = 0. Hence Theorem 6.4.1 can not
be directly applied in this scenario. However, we can take q1 =  where  > 0 is
a small number. Now we can apply Theorem6.4.1 and get the desired result by
taking  → 0+ .
Taking q1 = (> 0), when we apply Theorem 6.4.1 to determine the set
T∗ (λ),there are two possibilities. These are (i) 1  T∗ (λ) and (ii) 1 ∈ T∗ (λ).
When 1  T∗ (λ), the situation is exactly same as discussed earlier and so the
optimal solution for PO(λ) as given in Theorem 6.4.1 remains unchanged.
When 1 ∈ T∗ (λ). Then by Theorem 6.4.1, the optimal solution of PO(λ) is given
by
⎧ ∗
⎪
⎪ y
∗
⎪
⎨ , i ∈ T∗ (λ)
xi = ⎪⎪ q (6.32)
⎪
⎩0 ,
i
i  T∗ (λ),

where
⎛ ⎞−1
⎜⎜ 1 
n
1 ⎟⎟
y∗ = M0 ⎜⎜⎜⎝ + ⎟⎟ .
⎟ (6.33)
 ql ⎠
l1

In (6.32) and (6.33), if we now take the limit as  → 0+ , we obtain x∗i = 0 for all
i > 1, x∗1 = M0 and y∗ = 0.
An obvious question now is to state certain condition so that the two cases,
namely 1  T∗ (λ) and 1 ∈ T∗ (λ), can be verified easily. In this context, we state
the following condition
(µn − µ1 ) (µn−1 − µ1 ) (µ2 − µ1 ) λ
+ + ... + < . (6.34)
qn qn−1 q2 (1 − λ)

Here µ1 is the risk-free return. For the risk-free asset a1 , q1 = 0 but it does not
enter in (6.34).
Form the statement of Theorem 6.4.1 and related condition for the determina-
tion of T∗ (λ) for k = (n − 2) it is clear that if the condition (6.34) is not satisfied
then 1  T∗ (λ). Otherwise if condition (6.34) is satisfied then 1 ∈ T∗ (λ). Therefore
the above discussion leads to the following theorem
 Minimax Rule Based Portfolio Optimization: An L∞ -Risk Model 183

Theorem 6.4.2 Let λ ∈ (0, 1) be given. If the condition (6.34) is not satisfied then
T∗ (λ) and associated (x∗ , y∗ ) should be determined as per Theorem 6.4.1. Otherwise
if (6.34) is satisfied, then all wealth M0 should be invested in the risk free asset,
i.e. x∗1 = M, x∗i = 0, i  1 and y∗ = 0.
Certain Observations about the Solution of Problem PO(λ)
(i) Consider the solution where

(µn − µn−1 ) λ
< ,
qn (1 − λ)

and
(µn − µn−2 ) (µn−1 − µn−2 ) λ
+ ≥ .
qn qn−1 (1 − λ)
Then Theorem 6.4.1 tells that the optimal portfolio should consist of assets
an and an−1 only. Further, the actual amounts of investments for these assets
should be  −1  −1
∗ M0 1 1 qn
xn = + = M0 1 + ,
qn qn−1 qn qn−1
and  −1  −1
M0 1 1 qn−1
x∗n−1 = + = M0 1 + .
qn−1 qn−1 qn qn
Therefore, if qn is much larger than qn−1 , then it is possible that x∗n is nearly
zero while x∗n−1 is nearly equal to M0 .
The above discussion suggests that the process of constructing an optimal
portfolio is a two phase decision process. In the first phase, the assets are
selected according to their rates of return, thereby giving the set of investable
assets. In the second phase, the actual amounts allocated to the investable
assets (those assets which have been selected in Phase-1) are determined based
on their risk levels. Here if the risk of any investable asset is very high, then
its wealth allocation is very small and therefore it may be neglected in the
optimal portfolio. Thus in Phase-1, an asset may be eliminated if its return is
very low, while in Phase-2, it may be eliminated if its risk is very high.
(ii) For the optimal portfolio x∗ as given by Theorem 6.4.1, we have

∗
∗ y, i ∈ T∗ (λ)
xi qi = (6.35)
0, otherwise.
 184 Financial Mathematics: An Introduction

Since qi x∗i is the risk associated with the ith asset when the amount x∗i is in-
vestable in it, (6.35) tells that for the assets selected for investment, we invest
them with the amounts such that they have the same risk y∗ . Thus the optimal
solution of PO(λ) gives an investment plan in which we invest a small amount
for assets having high risk, and invest a large amount for assets having low
risk. This strategy will not increase the maximum risk but certainly increase
the overall expected return.
(iii) The amounts x∗i (i ∈ T∗ (λ)) as given in Theorem 6.4.1 does not depend on µi
as long as the set T∗ (λ) is selected. Thus the information about the expected
returns µi is used to determine the set of investable assets only. Later only risks
of the investable assets are used to determine the actual allocation of wealth
in the investable assets.
(iv) In Theorem 6.4.1, the inequalities are used to define the ranking rule for se-
lecting the investable assets. Because of the presence of these inequalities, this
model has some robustness and is not that much sensitive against errors in
the parameters µi (i = 1, 2, . . . , n). In fact Cai et al. [23] gave an explicit result
for allowable perturbation δi in µi so that the optimal portfolio as obtained by
Theorem 6.4.1 remains unchanged.
We now present some illustrative examples to verify the above points.
Example 6.4.1 Consider a six asset (a1 , a2 , a3 , a4 , a5 , a6 ) portfolio optimization
problem PO(λ) with the following data

µ = col(µ1 , µ2 , µ3 , µ4 , µ5 , µ6 ) = (1, 3, 5, 6, 7, 9)T

q = col(q1 , q2 , q3 , q4 , q5 , q6 ) = (5, 10, 6, 7, 400, 8)
M0 = Rs 60, 000
λ = 0 · 35.

Use Theorem 6.4.1 to determine the optimal portfolio.

Solution Here the asset returns µi are already in the ascending order. Next we
note that
λ
= 0.5384,
(1 − λ)
and
 Minimax Rule Based Portfolio Optimization: An L∞ -Risk Model 185

(µ6 − µ5 ) (9 − 7)
= = 0 · 25 < 0 · 5384
q6 8
(µ6 − µ4 ) (µ5 − µ4 ) (9 − 6) (7 − 6)
+ = + = 0 · 3775 < 0 · 5384
q6 q5 8 400
(µ6 − µ3 ) (µ5 − µ3 ) (µ4 − µ3 ) (9 − 5) (7 − 5) (6 − 5)
+ + = + + = 0 · 647 > 0 · 5384.
q6 q5 q4 8 400 7
= 0 · 647 > 0 · 5384.
Hence (n − k) = 4, which gives k = 6 − 4 = 2. This gives
T∗ (0.35) = {n, n − 1, . . . , n − k} = {6, 5, 4},
and therefore the investable assets are a4 , a5 and a6 .

Next we determine the allocation in these investable assets using the formulae
given in Theorem 6.4.1, we have


∗ 60, 000 1 1 1 −1
x4 = + + = 31704 · 09
7 7 400 8


60, 000 1 1 1 −1
x∗5 = + + = 554 · 82
400 7 400 8


60, 000 1 1 1 −1
x∗6 = + + = 27741 · 08,
8 7 400 8

and

1 1 1 −1
y∗ = 60, 000 + + = 221928 · 66.
7 400 8
Therefore the optimal portfolio consists of the allocation x∗1 = 0, x∗2 = 0, x∗3 =
0, x∗4 = 31704 · 09, x∗5 = 554 · 82 and x∗6 = 27741 · 08 from the given wealth
M0 = 60, 000. Further the minimum value of the L∞ -risk measure for the optimal
portfolio x∗ , i.e. wL∞ (x∗ ) is y∗ = 221928 · 66. In fact y∗ = q4 x∗4 = q5 x∗5 = q6 x∗6 =
221928 · 66 and therefore all investable assets with the chosen optimal allocation
have the same risk y∗ . Further the maximum value of the portfolio expected return
is
 n
µi x∗i = µ4 x∗4 + µ5 x∗5 + µ6 x∗6 = 443778 · 0.
i=1
Here we see that the returns of the asset a1 , a2 and a3 are small in comparison to
a4 , a5 and a6 and they are not selected for the investment. This is precisely the
186 Financial Mathematics: An Introduction

Phase-1 of the investable assets selection procedure. Further, as risk of the asset
a5 is very large, it is therefore allocated a very small amount of investment. This
is the Phase-2 of the investable selection procedure.


Some More Observations on L∞ -Risk Model

(i) Looking at the definition of the L∞ risk measure wL∞ (x) or the formula for the
optimal allocation x∗ , one gets the impression that none of these depend on
the covariances between the assets. Further it seems that only the risks of the
individual assets, rather than the risk of the entire portfolio, are taken care off
in this development. But this is not true as being explained ⎛-- nbelow. n --⎞
⎜⎜--  --⎟⎟
Let us recall that the total portfolio risk is given by E ⎜⎜⎜⎝-- ri xi − µi xi --⎟⎟⎟⎠.
- i=1 i=1
-
Let ⎛ --ξ nbe a given positive
-- number.
⎞ We are interested in making
⎜⎜ -- n -- ⎟⎟
P ⎜⎜⎜⎝ -- ri x i − µi xi -- ≥ ξ⎟⎟⎟⎠ as small as possible because this ensures that the
- i=1 i=1
-
deviation of the actual total return from the expected total return is as small
as possible. But by Markov inequality we have

n 
n
1  n  n
P(| ri xi − µi xi |≥ ξ) ≤ E(| ri xi − µi xi |)
i=1 i=1
ξ i=1 i=1

1
n
≤ E(| ri − µi | xi )
ξ i=1
n
≤ wL∞ (x). (6.36)
ξ
The inequality (6.36) illustrates that the total portfolio risk is small if wL∞ (x)
is kept small. But in the expression for the total portfolio risk, the covariances
among various assets are certainly involved. This explains that in the L∞ -
risk model, the covariances are not ignored and minimizing wL∞ (x) indirectly
attempts minimizing the total portfolio risk.
(ii) In analogy with Markowitz’s mean-variance model, it is natural to ask certain
questions with regard to the efficient frontier and the status of CAPM type
model for L∞ -risk model. Cai et al. [23] developed an explicit procedure for
tracing the efficient frontier for L∞ -risk model but did not present any CAPM
type result.
 Value-at-Risk of an Asset 187

(iii) For Markowitz’s model we can generate efficient frontier easily provided short
selling is allowed (we refer to the two fund theorem in this regard). For L∞ -risk
model discussed here, generation of efficient frontier becomes easier provided
short selling is not allowed (we refer to Theorem 6.4.1 in this regard).
(iv) Cai et al. [23] L∞ -risk model can be made more robust if more flexibility is
allowed with regard to the parameters µi (i = 1, 2, . . . , n). Deng et al. [36] pre-
sented a minimax type model for portfolio optimization where µi s are allowed
to be in the interval, ai ≤ µi ≤ bi (i=1,2,. . . ,n). They used the celebrated min-
imax theorem along with the results of Cai et al. [23] to solve the resulting
optimization problem.
So far we have considered certain variations of Markowitz’s mean-variance
model where the portfolio risk is taken different from the standard deviation (e.g.
L1 -risk or L∞ -risk) but it is still essentially based on moments of the portfolio re-
turn. In the coming sections, we shall discuss few variations of Markowitz’s model
where portfolio risk is quantile based. In particular, we discuss VaR and CVaR
based portfolio optimization problems.

6.5 Value-at-Risk of an Asset

Before defining value-at-risk of an asset (denoted by VaR), we consider the below
given example for the sake of motivation.
Consider two time instances t = 0 and t = T. Let S(0) = 100 and assume that
we buy a share of stock at t = 0 to sell it at t = T. As S(T) is a random variable,
we cannot predict the quantum of profit/loss. Obviously we shall suffer a loss if
S(T) < 100 erT , where r is the risk-free rate under continuous compounding.
A natural question at this stage could be to determine the probability that loss
is less than or equal to a specified amount, say Rs 20, i.e. P[(100 erT − S(T)) ≤ 20].
This probability can be easily be computed once the probability distribution of
S(T) is known. But it becomes more interesting if this question is reversed. Here
we fix the probability, say 95%, and seek the amount such that the probability
of loss not exceeding this amount is more than or equal to 95%. The amount so
obtained is essentially called the value-at-risk of the given asset. Thus

P[(100 erT − S(T)) ≤ VaR] ≥ 0 · 95 , (6.37)

and VaR represents the predicted maximum loss with specified probability (0.95
in our example) over a certain period of time which is T in our case.
188 Financial Mathematics: An Introduction

The expression (6.37) suggests that VaR is a measure related to percentiles of

loss distributions. In certain sense, it tries to answer the basic question which every
investor seems to ask at some point in time, namely, what is the most he/she can
lose on his/her investments? Value-at-Risk tries to answer this question within a
reasonable bound. VaR and also other downside risk measures are very useful in
assessing the risk for securities with asymmetric return distributions, such as call
and put options.
VaR has been developed by JP Morgan, and made available through Risk
Metrics software in October 1994. We now give the general definition of VaR for
a random variable X. This random variable may represent the loss distribution of
the asset return with −X being represented as gain.
Definition 6.5.1 (Value at Risk) Let X be the given random variable and α be
the given probability level. Then the VaR of X with confidence level (1 − α), 0 <
α < 1, denoted by VaR(1−α) (X) is defined as

VaR(1−α) (X) = Min{z : FX (z) ≥ (1 − α)} = F−1

X (1 − α) , (6.38)

where FX denotes the cumulative distribution function of the random variable X.

In view of the above definition VaR(1−α) (X) is a lower (1 − α) percentile of the
random variable X. If we substitute the expression for FX (z) in (6.38), we obtain

VaR(1−α) (X) = Min{z : P(X ≤ z) ≥ (1 − α)}. (6.39)

Thus VaR(1−α) (X) for an asset is the value z such that the probability that the
maximum loss X is at most z, is at least (1 − α).
The use of VaR involves two chosen parameters. These are confidence level
(1 − α) and the holding period T of the asset. The choice of α, and hence(1 − α),
depends on the purpose to which our risk measure is utilized. In practice α is
typically taken as 10%, 5% and 1%, so that the typical confidence levels are
90%, 95% and 99%. The usual holding periods are one day or one month, but it
can be even one quarter or more. Given the confidence level (1 − α) and horizon
T, VaR is a bound such that the loss over the horizon is less than this bound
with probability equal to the confidence coefficient. For example, if horizon is one
week, the confidence level is 99% (so α = 0 · 01) and VaR is Rs 50,000, then there
is only a 1% chance of loss exceeding Rs 50,000 over the next week.
From (6.39) we note that VaR(1−α) (X) = F−1
X
(1−α) and therefore for a continuous
loss distribution VaR(1−α) (X) is simply the loss such that

P(X ≤ VaR(1−α) (X)) = (1 − α).

 Value-at-Risk of an Asset 189

Let us recall that if X ∼ N(µ, σ2 ) then the q-percentile of X is µ + σΦ−1 (q),

where Φ is the standard normal density function. Therefore

VaR(1−α) (X) = µ + σΦ−1 (1 − α) ,

where (1 − α)-percentile of Φ, namely Φ−1 (1 − α), is that value of the standard

normal variate for which the area in the left is (1 − α).
The above relation shows that for normally distributed random variables, VaR
is proportional to the standard deviation.
Example 6.5.1 Suppose that the stock price is lognormal with mean 12% and
standard deviation 30%. Let the interest rate r for the period t = 0 to t = 1 be
8% and stock price S(0) at time t=0 be Rs 100. Determine the VaR for the given
stock at 95% confidence level.
Solution We are given that
 S(1) 
ln ∼ N(0.12, (0.30)2 ) ,
S(0)

and our aim is to find VaR for (1 − α) = 0.95, i.e. for α = 0.05. By definition, loss
equals (100e0.08 − S(1)) and

P[(100e0.08 − S(1)) ≤ VaR] = 0.95. (6.40)

Now from the given information

 S(1) 
ln − 0.12
S(0)
∼ N(0, 1). (6.41)
0.30
Therefore (6.40) gives
 
S(1) (100e0.08 − VaR)
P ≥ = 0.95
S(0) 100

i.e.  
 S(1) 
(100e0.08 − VaR)
P ln ≥ ln = 0.95
S(0) 100
i.e.
190 Financial Mathematics: An Introduction

⎡    ⎤
⎢⎢ S(1)  (100e0.08 − VaR) ⎥
⎢⎢ ln − 0.12 ⎥⎥⎥
⎢⎢ S(0) − 0.12
ln
100 ⎥⎥
P ⎢⎢⎢⎢ ≥ ⎥⎥ = 0.95.
⎥⎥ (6.42)
⎢⎢ 0.30 0.30 ⎥⎥
⎢⎣ ⎥⎦

But from (6.41) and (6.42), we get

 
100e0·08 − VaR
ln − 0.12
100
= −1.645,
0.30
which determines
VaR = 100(e0.08 − e−0.3750 ) = Rs 39.50.


Example 6.5.2 Let an investment A return a gain of Rs 100 with probability 0.96
and a loss of Rs 200 with probability 0.04. Obtain VaR of the given investment
with 95% confidence level.
Solution We note that our definition of VaR is based on the probability distribu-
tion of the random variable X, where X represents the loss of the given investment.
Therefore from the given information we have


−100, with probability 0 · 96
X=
200, with probability 0 · 04.
For determining the value of VaR of the given investment we need to use the
relation
VaR(1−α) (X) = Min{z : P(X ≤ z) ≥ (1 − α)}
with (1 − α) = 0 · 95. Since X is a discrete random variable we have
P(X ≤ −100) = P(X = −100) = 0 · 96 ≥ 0 · 95,
and P(X ≤ 200) = 1 ≥ 0 · 95. But -100=Min(-100, 200) and therefore the required
VaR is Rs − 100.


Example 6.5.3 Let there be two identical bonds A and B. Each of these defaults
with probability 0.04 giving a loss of Rs 100. Further there is no loss if default does
not occur. Let default occur independently and C be a portfolio consisting of these
two bonds A and B. Obtain VaR of bonds A and B; and also of the portfolio C at
95% confidence level.
 Value-at-Risk of an Asset 191

Solution Following on lines similar to Example 6.5.2, it is simple to get VaR(A)=0

and VaR(B)=0. Since C is the portfolio consisting of bonds A and B, its return
R becomes the sum of returns of bonds A and B. Writing returns in terms of the
loss function, we obtain the distribution of loss X of investment C as
⎧
⎪
⎪ 0, with probability (0 · 96)2 = 0 · 9216
⎪
⎨
X=⎪ ⎪ 200, with probability (0 · 04)2 = ·0016
⎪
⎩ 100, with probability (1 − (0 · 96)2 − (0 · 04)2 ) = 0 · 0768.
Therefore at 95% confidence level,
VaR(C) = Rs 100 ,
because P(loss of investment C ≤ 100) = 0.9216 + 0.0768 = .9984 > 0.95, and
P(loss of investment C ≤ 200) = 1 > 0 · 95, but 100 = Min(100, 200). Here we may
note that P(X ≤ 0) = P(X = 0) = 0 · 9216 < 0 · 95.



Some Theoretical and Computational Difficulties with VaR

(i) In Example 6.5.3, at 95% confidence level VaR(C)=100 but VaR(A)+VaR(B)=0.
Thus VaR(C) > VaR(A) + VaR(B). But this violates the principle that diversi-
fication reduces risk. We expect a good risk measure to respect this principle.
Unfortunately VaR does not do so. Mathematically it means that VaR is not
a subadditive risk measure. We call a risk measure f to be subadditive if for two
different investments A and B,
f (A + B) ≤ f (A) + f (B) ,
i.e. the total risk of two different investment portfolios does not exceed the
sum of individual risks.
In Example 6.5.3, diversification has actually increased the risk if VaR is used
as a risk measure.
(ii) VaR does not pay any attention to the magnitude of losses beyond the VaR
value. For example it is very unlikely that an investor will take a neutral
view for two portfolios with identical expected return and VaR, but return
distribution of one portfolio having short left tail and other having a long left
tail.
(iii) There is additional difficulty with VaR in its computation and optimization.
When VaR is computed by generating scenarios, it turns out to be a non
smooth and non convex optimization problem is required to be solved.
192 Financial Mathematics: An Introduction

The above shortcomings of VaR has motivated researchers to look for other quan-
tile based risk measures and CVaR (conditional value at risk) is one such risk
measure.
Before we proceed with the discussion of CVaR, we remark that inspite of the
difficulties outlined above, VaR is still very popular in the market. Therefore we
need to discuss some non-parametric and parametric methods for its estimation.

Nonparametric Estimation of VaR

Here we describe a procedure for the nonparametric estimation of VaR. Since
it is a nonparametric estimation, the loss distribution is not assumed to be in any
parametric family of distributions such as normal distributions.
To motivate the procedure, let us consider a simple example. Suppose that we
hold a Rs 20,000 position in an index fund, e.g. NIFTY, so that our returns are
those of this index. Suppose we require a 24-hour VaR by using 1000 daily returns
on NIFTY for the period ending on December 31,2010. These are approximately
the last four years of daily returns in the time series of NIFTY.
Let the confidence level (1 − α) be 95%, i.e. α = 5%. Since 5% of 1000 is 50 and
VaR is a loss which is minus of the revenue, we can estimate the required VaR
by taking the 50th smallest daily return. Let this be r(50) , the 50th order statistic
of the sample of historic returns. This means that a daily return of r(50) or less
occurred only 5% time in the historic data. Equivalently this means that a daily
loss of −r(50) or more occurred only 5% time in the historic data. A return of r(50)
on a Rs 20, 000 investment yields a revenue of 20, 000r(50) which can be taken as
an estimate of VaR at 95% confidence level. Specifically if r(50) = −0 · 0227, then
−r(50) = 0 · 0227 and the estimated VaR is 20, 000(0 · 0227), i.e. Rs 454.
The above procedure can be generalized easily. We first note that (1 − α)th
percentile of loss equals the αth percentile of the returns. Suppose that there are
n returns r1 , r2 , . . . , rn in the historic sample and let k be equal to (nα) rounded to
the nearest integer. Then α percentile of the sample of returns is the kth smallest
return, that is, the kth order statistic r(k) of the sample of returns. If S is the size
of the initial investment then the required estimate of VaR is −Sr(k) . Here minus
sign converts revenue (return times initial investment) to a loss.
Remark 6.5.1 Estimation using sample percentiles is only feasible if the sample
size is large. If T(holding period) is taken a quarter rather than a day, then in
a four years duration we will have only 16 observations. In such a situation,
increasing the number of years for historical data would also not increase sample
size substantially, and also because of volatility there may be bias in our estimate.
 Conditional Value-at-Risk 193

Parametric Estimation of VaR

For small sample size, VaR can be best estimated by parametric technique. We
discuss only the case of normal distribution because it is extremely simple.
Since, in general, the historical data or returns is given, we have

VaR(1−α) (loss) = −VaR(α) (return).

Therefore
VaR(1−α) (X) = −S(µ + Φ−1 (α)σ),
which can be estimated by −S(X + Φ−1 (α)σ̂). Here X and σ̂ are the mean and
standard deviation of sample of returns and S is the size of the initial investment.
Let us assume that from the given historical data, X = −3 · 107 × 10−4 and
σ̂ = 0 · 0151. Also Φ−1 (α) = −1 · 645 for α = 0 · 05. Then the estimate of the
required VaR is ((−3 · 107 × 10−4 ) + (−1 · 645)(0 · 0151)(20, 000)) = Rs 471.
In this example, the estimate of the expected return, is negative. This may be
because the particular years used include a prolonged bear market. However, we
certainly do not expect average future returns to be negative, otherwise we would
not have invested Rs 20,000 in the market.
Since in actual practice, normality assumption is not going to hold, we need to
apply historical simulation and Monte Carlo strategies to generate a large number
of possible scenarios. Also we need to understand VaR for a portfolio of assets
which could include options along with certain number of stocks. The estimation
of VaR in this situation also requires Monte Carlo simulation. We shall discuss
Monte carlo simulation in Chapter 14.

6.6 Conditional Value-at-Risk

One major criticism of VaR is that it pays no attention to the magnitude of
losses beyond the VaR value. Also VaR is not subadditive, which not only violates
the principle of diversification, but also creates computational difficulty in the
resulting portfolio optimization problem. These undesirable features of VaR led to
the development of conditional value-at-risk , denoted by CVaR, which is obtained
by computing the expected loss given that the loss exceeds VaR. Some other names
for CVaR are expected shortfall , expected tail loss and tail VaR. CVaR has been
introduced by Rockafellar and Uryasev [111] who studied many mathematical
properties of CVaR in detail.
194 Financial Mathematics: An Introduction

To define CVaR we first define the occurrence of a tail event. We say that a tail
event occurs if the loss exceeds the VaR. Then CVaR is the conditional expectation
of loss given that the tail event occurs. We now proceed to define CVaR and discuss
its minimization mathematically. To be specific, we define CVaR in context of
portfolio optimization.
We consider a portfolio of assets with random returns. Let f (w, r) denote the
loss function when we choose the investment w from a set of feasible portfolios and
r is the realization of random returns. In the context of our portfolio optimization
problem, f (w, r) = −rT w where r = (r1 , r2 , . . . , rn )T is the vector of random returns
and w = (w1 , w2 , . . . , wn )T with eT w = 1, e being the vector (1, 1, . . . , 1)T . Here in
the expression of f (w, r), minus sign has been taken to express return rT w as loss
function. We assume that the return vector r has a probability density function
p(r), e.g., the random returns may have a multivariate normal distribution.
For a fixed weight vector w, we define
,
ψ(w, q) = p(r)dr, (6.43)
f (w,r)≤q

which represents the cumulative distribution function of the loss associated with
the weight vector w. In fact ψ(w, q) = P(loss ≤ q). Therefore for a given confidence
level (1 − α), the VaR(1−α) (w) associated with the portfolio is given by

VaR(1−α) (w) = Min{q ∈ R : ψ(w, q) ≥ (1 − α)}.

Definition 6.6.1 (Conditional Value-at-Risk) Let w be a given portfolio and

(1−α) be the given confidence level. Then CVaR(1−α) (w) associated with the portfolio
w is defined as
,
1
CVaR(1−α) (w) = f (w, r)p(r)dr. (6.44)
α f (w,r)≥VaR(1−α) (w)

Remark 6.6.1 In special case of returns having discrete probability distribution,

we have ⎛ ⎞
1 ⎜⎜⎜⎜ ⎟⎟
⎟
CVaR(1−α) (w) = ⎜⎜ p j f (w, r(j) )⎟⎟⎟ , (6.45)
α⎝ j∈J1
⎠

where J = {1, 2, . . . , m}, J1 = {j ∈ J : f (w, r(j) ) ≥ VaR(w)} and the vector r(j) is the
jth realization of the return vector r with probability p j .
 Conditional Value-at-Risk 195

Remark 6.6.2 In certain sense, (6.44) and (6.45) tell that CVaR(1−α) (w) is the
average of the outcomes greater than VaR(1−α) (w). This is certainly true for contin-
uous distribution functions; but for general distributions this is not exactly true.
There are certain subtle aspects which need to be explained. We shall refer to
Sriboonchitta et el [128] in this regard.
Example 6.6.1 Let the loss function f (w, r) be given by f (w, r) = −r where r =
75 − j, (j = 0, 1, 2, . . . , 99) with probability 1%. Evaluate VaR(w) and CVaR(w) at
95% confidence level.
Solution The loss function f (w, r) = −r, where r = 75 − j, (j = 0, 1, 2, . . . , 99),
takes 100 values given by (−75, −74, −73, . . . , 0, 1, . . . , 19, 20, 21, 22, 23 and 24) with
equal probability p = 0 · 01. Therefore VaR(w) = 20 for 1 − α = 0 · 95.
We next evaluate CVaR(1−α) (w) for (1 − α) = 0.95 by using formula (6.45). This
gives
1
CVaR(w) = (20 + 21 + 22 + 23 + 24)(0 · 01) = 22.
0 · 05
Thus at 95% confidence level VaR is 20 and CVaR is 22.


In the above example for the same confidence level CVaR is more than VaR.
We have the following result in this regard.
Lemma 6.6.1. For the given confidence level (1−α), CVaR(1−α) (w) ≥ VaR(1−α) (w).
Proof. We have
,
1
CVaR(1−α) (w) = f (w, r)p(r)dr
α f (w,r)≥VaR(1−α) (w)
,
1
≥ VaR(1−α) (w)p(r)dr
α f (w,r)≥VaR(1−α) (w)
,
VaR(1−α) (w)
= p(r)dr
α f (w,r)≥VaR(1−α) (w)
= VaR(1−α) (w).


Remark 6.6.3 As CVaR of a portfolio is always more than or equal to VaR for
same (1 − α), portfolios with small CVaR will also have small VaR. But this does
not mean that the minimization of CVaR is equivalent to the minimization of VaR.
196 Financial Mathematics: An Introduction

Minimization of CVaR
We now discuss Rockafellar and Uryasev’s [111] procedure to minimize CVaR.
Since the definition of CVaR involves the VaR function explicitly, it is not very
convenient to optimize CVaR directly. Therefore we introduce the auxiliary func-
tion ,
1
F(1−α) (w, q) = q + ( f (w, r) − q)p(r)dr. (6.46)
α f (w,r)≥q
If we denote a+ = Max(a, 0), a ∈ R, then from (6.46) we get
,
1
F(1−α) (w, q) = q + ( f (w, r) − q)+ p(r)dr. (6.47)
α
We refer to Rockafellar and Uryasev [111] for below given results.
Lemma 6.6.2. The auxiliary function F(1−α) (w, q) is a convex function of q.
Lemma 6.6.3. VaR(1−α) (w) is a minimizer of F(1−α) (w, q) over q.
Lemma 6.6.4. The minimum value of F(1−α) (w, q) over q is CVaR(1−α) (w).
In view of the above Lemmas, we have for a given w,

CVaR(1−α) (w) = Minq (F(1−α) (w, q)) = F(1−α) (w, VaR(1−α) (w)). (6.48)

The left equality of (6.48) tells that we can minimize CVaR directly without
computing VaR first. Since for portfolios, the loss function f (w, r) = −rT w is a
linear and hence also convex function of w, the auxiliary function F(1−α) (w, q) is a
convex function of w. Therefore the problem

Min CVaR(1−α) (w)

subject to
eT w = 1, (6.49)

is equivalent to

Min F(1−α) (w, q)

w,q

subject to
eT w = 1, (6.50)

which is a smooth convex optimization problem. Though problem (6.50) can be

solved by using standard convex optimization techniques, the formulation (6.50)
 Conditional Value-at-Risk 197

still needs the computation/determination of joint density function p(r) of random

return vector r. In practice it is not simple or even possible, so we present another
approach which is based on generation of scenarios via computer simulation. An
added advantage of this approach is that it results in a linear programming for-
mulation.
Suppose we have scenarios r(s) , (s=1,2,. . . ,ns ), which may represent historical
values of the random vector of returns or obtained via computer simulation. We
shall assume that all scenarios have equal probability and define the following
empirical distribution of the random returns based on the available scenarios

1 
ns
(
F(1−α) (w, q) = a + ( f (w, r(s) ) − q)+ . (6.51)
α ns s=1

Using (6.51) to approximate F(1−α) (w, q), we get an approximation to the problem
(6.50) as

Min (
F(1−α) (w, q)
w,q

subject to
eT w = 1. (6.52)

Now writing ( f (w, r(s) ) − q)+ as zs and using the definition of a+ , we obtain

1 
ns
Min q+ zs
w,q,zs α.ns s=1
subject to
zs ≥ f (w, r(s) ) − q, s = 1, 2, . . . , ns
eT w = 1 (6.53)
zs ≥ 0, s = 1, 2, . . . , ns .

In the context of portfolio optimization, f (w, r(s) ) = −wT r(s) is linear and so the
problem (13.5) becomes a linear programming problem.
Most often we try to optimize a suitable performance measure (e.g, expected
return) while making sure that certain risk measures do not exceed a threshold
value. It could be variance or absolute deviation as has been discussed earlier.
When the risk measure is CVaR, the resulting optimization problem is
198 Financial Mathematics: An Introduction

Max µT w
subject to
CVaR(1−α) (w) ≤ uα
(6.54)
eT w = 1,

which can be approximated as

Max µT w
w,z,q

subject to
1 
ns
uα ≥ q + zs
α ns s=1
zs ≥ f (w, r(s) ) − q (s = 1, 2, . . . , ns )
e w=1
T

zs ≥ 0 (s = 1, 2, . . . , ns ).

In (6.54), we can have more than one CVaR constraint for different levels α. Also
similar to Markowitz’s model, we can have a trade off between return and CVaR.
We can refer to Mansini et al. [89] for more details in this regard.

6.7 Preference Relation, Utility theory and Decision

Making
The mean-variance criterion used in the Markowitz portfolio model can also be ex-
plained in terms of the expected utility maximization principle in decision making.
To understand this aspect we first wish to motivate the readers to the celebrated
Von Neumann- Morgenstern’s model for decision making. This model tries to ex-
plain how a rational person makes his/her optimal choice in a given situation.
This requires some discussion on preference relation and related results in utility
theory.
Choice under certainty
We first consider the situation when there is no uncertainty present, i.e. the alter-
natives are certain. Let this set of alternatives be denoted by Z. As the choice is
being made under certainty, we shall get the element of our choice for sure.
In order to perform our choice, i.e. the selection procedure, we prescribe our
own preferences, and use them to select the desired optimal element of Z. These
preferences are expressed as a preference relation, say  on Z. Thus we say that
 Preference Relation, Utility theory and Decision Making 199

for x, y ∈ Z, x  y states that we strictly prefer x to y. Here our preference relation

should reflect our valuation of different alternatives, i.e. different elements of Z.
These valuations give a quantitative representation of our preference relation.
The above discussion can now be formalized mathematically. A preference re-
lation  is a binary relation on Z i.e. a subset A of Z × Z where (x, y) ∈ A when
x  y. We demand the following properties with our preference relation 
(i) The preference relation  is asymmetric i.e. x  y ⇒ y  x. Thus if x is strictly
preferred to y then y is not strictly preferred to x.
(ii) The preference relation  is negative transitive i.e. x  y and y  z ⇒ x  z.
Thus if x is not strictly preferred to y and y is not strictly preferred to z then
x is not strictly preferred to z.
We write x  y if x  y and y  x. The relation  is called an indifference or
equivalence relation because it is reflexive, symmetric and transitive. Thus given
x, y ∈ Z, exactly one of x  y, y  x, or x  y holds. We further write
x  y when y  x, i.e. either x  y or x  y. The relation  is called the weak
preference relation. It is simple to verify that x  y when y  x. Also x  y when
x  y and y  x. For making an optimal choice we need to prescribe a numerical
representation of our strict preference relation  . This is done via the utility
function.
Definition 6.7.1 (Utility Function) A function u : Z → R is called a utility
function if
x  y ⇔ u(x) > u(y).
An obvious question at this stage is about the existence of such a function u.
We have the following theorem in this regard (Sriboonchitta et al. [128]).
Theorem 6.7.1 Let Z be finite or countably infinite. Then a utility function u as
defined above certainly exists.
Choice Under Uncertainty
We next discuss the more realistic situation of making an optimal choice under
uncertainty. A typical example of this scenario could be possible losses in invest-
ment portfolio which are real-valued random variables. Here the range space Z is
the real line R. Taking motivation from the earlier case of choice under certainty,
we attempt to give a numerical representation of a preference relation  on a
class of random variables X. We say that for X, Y ∈ X and a given preference
relation , there exists a utility function u : Z → R (unique up to a positive linear
transformation) such that
200 Financial Mathematics: An Introduction

X  Y ⇔ E(u(X)) > E(u(Y)),

provided that the expectations exist. Thus for making their optimal choices under
uncertainty, decision makers try to maximize their expected utilities. This princi-
ple is called Von Neumann- Mangenstern expected utility maximization principle.
Is this principle always applicable? Does there always exist a utility function as
described above for the random variables in the class X? Apart from various
mathematical conditions for its existence it was argued that the existence of util-
ity function is restricted to “rational” people. This latter part has been a topic of
debate among economists and psychologists. For this as well as other aspects of
Von Neumann - Mangenstern theory we shall refer to Fishburn [47].
Utilities and Risk Attitudes
We all know that different people react differently in risky situations. They
make decision based upon their own attitudes towards risk. Thus, making an
investment choice in risky and risk free assets, each person will act differently
depending upon his/her risk attitude. Roughly there are three main risk attitudes
of an investor. Let the random variable X be a risky prospect, and u be the utility
function of the decision maker. We say that the investor is
(i) risk neutral if, when facing two risky prospects with same expected value will
feel indifferent. In terms of the utility function u, this means that u(E(X)) =
E(u(Y)). This will happen for example if u is linear, say u(x) = x.
(ii) risk averse if, when facing two risky prospects with the same expected value,
will prefer the less risky one. In terms of the utility function u, this means
that u(E(X)) > E(u(X)). This will happen for example if u is concave, say
u(x) = log x or u(x) = −x2 .
(iii) risk seeking if, when facing two risky prospects with the same expected value,
will prefer the more risky one. In terms of the utility function u, this means that
u(E(X)) < E(u(X)). This will happen for example if u is convex, say u(x) = x2 .
It is obvious that given a utility function u(x), any function of the form v(x) =
au(x) + b with a > 0 is a utility function equivalent to u(x). This is because
equivalent utility functions give identical rankings in terms of the principle of
expected utility maximization.
To explain the meaning of risk averse scenario, let us consider an investor who
has two alternatives for future wealth.The first alternative is based on the outcome
of tossing of a coin. If the coin turns up ‘head’ the investor gets Rs
x, otherwise
x + y
he/she gets Rs y. The second alternative is a sure event of getting Rs with
2
certainty. Let the investor’s utility function be given by a strictly concave function
 Preference Relation, Utility theory and Decision Making 201

u(x) + u(y)
u(x). Then the expected utility of first alternative is , whereas the

x + y 2
expected utility of the second alternative is u . Since u is strictly concave
2
we have
x + y  u(x) + u(y)
u > .
2 2
Hence by the principle of expected utility maximization we shall prefer the sure
(x + y)
wealth of Rs to a 50-50 chance of x or y. Here we must note that both
2
(x + y)
alternatives have the same expected value of Rs , but the one without risk
2
(certainty of getting) is preferred. This explains the risk averse attitude of the
investor.
The below given result is useful in this regard.
Theorem 6.7.2 An investor is risk averse if and only if his/her utility function
is concave.
Remark 6.7.1 Although it has not been mentioned in the statement of above
theorem, u is non-decreasing by the definition of utility function. We refer to
Sriboonchitta et al. [128] for the proof of the above theorem. The proof essentially
uses Jensen’s inequality and the defining condition u(E(X)) > E(u(X)) for a risk
averse investor.
Let us take another example to understand the risk attitudes. For an investor
let there be three possible scenarios: a 5% chance of loosing Rs 20, 000, a 10%
chance of loosing Rs 10, 000, and a sure chance of loosing Rs 1, 000. A risk neutral
investor will not find any of these situations worse than the other. For a risk averse
investor, the first situation is worse than the second one, and the second one is
worse than the last one. This is because risk averse investors dislike uncertainty
about the size of losses, so they do not prefer even a small possibility of large
amount of loss.
Certainty Equivalent
The certainty equivalent of a random wealth X is defined to be the amount of a
certain (i.e. risk-free) wealth that has a utility level equals to the expected utility
of X. Thus if C is the certainty equivalent of a random wealth X then by definition
u(C) = E(u(X)).
The certainty equivalent of a random variable is same for all equivalent utilities
and is measured in units of wealth.
202 Financial Mathematics: An Introduction

6.8 Mean-Variance Model Revisited in Utility Frame

Work
We now discuss Markowitz’s mean-variance model in the frame work of expected
utility maximization principle. Here we need to solve the following optimization
problem
 
Max E u(rT w)
subject to
eT w = 1, (6.55)

where rT = (r1 , r2 , . . . ri , . . . , rn ) is the vector of asset returns, u is investor’s utility

function and wT = (w1 , w2 , . . . wi , . . . , wn ) is the decision vector of asset weights.
Earlier we had agreed to represent a portfolio by w : (w1 , w2 , . . . wi , . . . , wn )
such that eT w = 1, where eT = (1, 1, . . . , 1). Therefore X = rT w denotes the return
(random) of the portfolio w. The mean-variance rule of Markowitz states that the
portfolio w(1) is preferred to the portfolio w(2) if
(i) E(rT w(1) ) ≥ E(rT w(2) ),
and
(ii) σ2 (rT w(1) ) ≤ σ2 (rT w(2) ).
If we denote the random variable rT w(1) by X and the random variable rT w(2)
by Y then the mean-variance rule says that X is preferred to Y when E(X) ≥ E(Y)
and Var(X) ≤ Var(Y). The essence of Markowitz’s theory is that the above mean-
variance rule is constrained with the principle of expected utility maximization for
1
the utility function u(x) = ax− b x2 with a > 0 and b ≥ 0. This utility function is
2
really meaningful only in the range x ≤ a/b, where the function is increasing. Also
as b > 0, the function u is strictly concave every where and therefore corresponds
to the risk aversion scenario.
Let the given portfolio have a random wealth value of y. Then


  1 2
E u(y) = E ay − b y
2
1 1
= aE(y) − b(E(y))2 − bVar(y). (6.56)
2 2
Now as per the expected utility maximization principle the optimal portfolio y
is the one that maximizes (6.56) with respect to all feasible choices of the random
 Risk Modeling and Financial Risk Measures 203

   
wealth variable y. Thus E u( ŷ) ≥ E u(y) for all feasible y. Let E( ŷ) = M and
S = {y : y is feasible and E(y) = M}. We shall now show that Var( ŷ) ≤ Var(y) for
all y ∈ S. If possible let there exists y ∈ S such that Var(y) ≤ Var( ŷ). Then as
y ∈ S it is feasible and E(y) = M. Further
1 1 1 1
aM − bM2 − bVar( ŷ) ≤ aM − bM2 − bVar(y). (6.57)
2 2 2 2
But then (6.57) contradicts that ŷ is optimal. Therefore Var( ŷ) ≤ Var(y) for
all y ∈ S, i.e. ŷ has the minimum variance with respect to all feasible y s with
E(y) = M. Hence ŷ corresponds to a mean-variance efficient point. Different mean-
variance efficient points are obtained by providing different values for the param-
eters a and b. The readers may identify y = rT w, E(y) = µT w and Var(y) = wT Cw
with C as the variance-covariance matrix.
Now we have a very natural question. What happens if we take a general utility
function u? How different are the mean-variance efficient portfolios and portfolios
that maximize the expected utility i.e. those which are the solution of the problem
(6.55) for a general u? In this context we have the following theorem
Theorem 6.8.1 Let X and Y be two random variables, normally distributed with
means µ1 , µ2 and variances σ21 , σ22 respectively. Then following are equivalent
(i) E (u(X)) ≥ E (u(Y)) for any u : R → R which is non decreasing and concave.
(ii) µ1 ≥ µ2 and σ21 ≤ σ22 .
Proof. See details in Sriboorchitta et al. [128].


Remark 6.8.1 In view of the above theorem, if the returns are normally dis-
tributed and the utility function is concave, then the portfolio that maximizes the
expected utility is on Markowitz’s efficient frontier. Exactly which portfolio on the
efficient frontier will be obtained, will depend on the choice of utility function.

6.9 Risk Modeling and Financial Risk Measures

Financial risk management is a broad concept involving various perspectives. It
is about managing exposure to risk, such as market risk and credit risk. For these
purposes we need a quantitative concept of risk in its own right.
Mathematically, risk management is a procedure for shaping a loss function
or loss distribution. We have already studied several moment based risk measures
(e.g. L1 , L2 and L∞ -risk measures) and also some quantile based risk measures (e.g.
204 Financial Mathematics: An Introduction

VaR and CVaR). Here we attempt to understand risk measures in a broader math-
ematical framework and introduce the concept of coherent risk measures(Artzner
[5]). An excellent tutorial on this topic is due to Rockafellor [112].
Basically, we wish to answer the following question: How can we assess the risk
of a financial position mathematically? For this we note that a financial position
is captured by a random variable X, e.g. loss. Information about loss is in its
cumulative distribution function F. In the financial context where the risk is
about the risk of loosing money, we wish to define numerical risk for X as some
appropriate number ρ(X) and call it as a risk measure. It seems natural to choose
ρ(·) in such a manner that it captures our risk perception and it is also suitable
for financial management purposes.
Domain and Range of Risk Measures
As stated above, we wish to assign a numerical value ρ(X) to each random variable
X to describe its risk. Here X stands for the loss of an investment portfolio, the
loss of a financial position, or capital needed to hold for an insurance company to
avoid insolvency.
It is very clear that the range of ρ has to be R. But what about its domain?
Let U denote the real vector space of all possible real valued random variables.
Mathematically we do not need the entire vector space U as the domain of ρ. But
rather we need a particular subset of U namely a convex cone X.
Definition 6.9.1 (Cone) A subset X in U is called a cone if for X ∈ X and
λ > 0, we have λX ∈ X.
Definition 6.9.2 (Convex Cone) A subset X in U is called a convex cone if it
is a cone and also a convex set. Thus X is a convex cone if for X, Y ∈ X, λ > 0,
we have X + Y ∈ X and λX ∈ X.
It is economically meaningful to take the domain of risk measures as a convex
cone of the vector space £. This is because we know that diversification should
reduce risk. But for this
* we need to evaluate our * risk measure on an investment
portfolio of type X = ki=1 λi Yi , where λi > 0, ki=1 λi = 1 and Yi denotes the rate
th
of return of i* asset in the portfolio.
* Thus if Yi are in a domain X of £, we also
require that ni=1 λi Yi , (λi > 0, ki=1 λi = 1), should also be in X; i.e. X is a convex
cone.
Desirable properties of Risk Measures

Let X denote the convex cone of loss random variable X. Then, as per Artzner
[5], we consider the following desirable properties for a risk measure ρ : X → R.
 Risk Modeling and Financial Risk Measures 205

(A1) (Monotonicity) If X ≤ Y almost surely, i.e. P(X ≤ Y) = 1, then ρ(X) ≤ ρ(Y).

(A2) (Positive Homogeneity) ρ(λX) = λρ(X) for λ ≥ 0.
(A3) (Translation Invariance) ρ(X + a) = ρ(X) + a, a ∈ R.
(A4) (Subadditivity) ρ(X + Y) ≤ ρ(X) + ρ(Y).
Definition 6.9.3 (Coherent Risk Measure) A functional ρ : X → R is called
a coherent risk measure if it satisfies axioms (A1)-(A4) listed above.
Remark 6.9.1 Though we are calling above properties as axioms, our approach
is not axiomatic as these do not characterize a risk measure. These are simply
the desirable properties and therefore even if a functional satisfies these axioms, it
may not have any meaning in terms of risk. This can be verified for ρ(X) = E(X).
The above axioms are based on our perception about risk. In case of loss
random variables, it is natural to believe that smaller losses have smaller risks.
This is precisely the axiom (A1). Also the risk of a loss of a financial position
should be proportional to the size of the loss of the position. This leads to the
axioms (A2) and (A3). The principle that diversification should reduce risk leads
to the axiom (A4). Also axioms (A2) and (A4) imply that the risk measure is
convex, which is certainly a desirable property.
Before we analyze some of already studied risk measures, we need to introduce
the notion of stochastic dominance. Stochastic dominance rules refer to criteria
to order distributions of random variables modelling uncertain quantities such as
future returns or investments. Thus these are various partial orders on the space
of distribution functions of random variables and they are based on the principle
of maximization of expected utility (Sriboonchitta et al. [128]). While stochastic
dominance rules shed light on which future prospects are better than others, they
do not provide us with any numerical value which we need to manage further our
financial positions. This numerical value is provided by risk measures. Therefore
for further analysis of risk measures, apart from axioms (A1)-(A4), we also need
to check consistency of risk assessment with respect to stochastic dominance.
Definition 6.9.4 (First Order Stochastic Dominance) Let X, Y be two ran-
dom variables with distribution functions F, G respectively. Then X is said to
dominate Y in the first order stochastic dominance (FSD), denoted by X 1 Y, if
F(.) ≤ G(.).
Remark 6.9.2 If X 1 Y then E(X) ≥ E(Y) but the converse is not true. This is
because , ∞ , 0
E(X) = (1 − F(t)) dt − F(t)dt.
0 −∞
206 Financial Mathematics: An Introduction

As preference relations on the space of distribution functions, stochastic dom-

inance rules are essentially stochastic orders which have been introduced much
earlier in statistics. These were rediscovered and made use in Economics/Finance.
We may refer to the survey article by Levy [83] in this regard. An obvious ques-
tion at this stage is: How can we rank distributions when they intersect? In this
situation FSD cannot be applied and therefore we need to weaken the criteria
which is consistent with FSD.
Definition 6.9.5 (Second Order Stochastic Dominance) Let X, Y be two
random variables with distribution functions F, G respectively. Then X is said to
dominate Y in the second order stochastic dominance (SSD), denoted by X 2 Y,
if , t
(G(u) − F(u))du ≥ 0,
−∞
for all t ∈ R.
To analyze any proposed risk measure we have to take care of two main criteria.
These are being coherent and being consistent with the stochastic dominance rules.
In general we say that X precedes Y when X  Y, and a risk measure ρ is said to
be consistent with respect to the partial order  if

(X  Y) ⇒ (ρ(X) ≤ ρ(Y)).

Here  could be 1 or if need be 2 or some other higher order stochastic domi-

nance rule. Since risks are related to preferences, such a consistency with respect
to a suitable partial order  is important.
Remark 6.9.3 First order stochastic dominance (FSD) implies the Second order
stochastic dominance (SSD). Thus (X 1 Y) ⇒ (X 2 Y). Also if (X 2 Y), then
E(X) ≥ E(Y). Further if (X 2 Y) and E(X) = E(Y), then Var(X) ≤ Var(Y).
Though higher order stochastic dominance rules have also been introduced in
the literature, in context of portfolio optimization only FSD and SSD have been
popular.

Some Already Studied Popular Risk Measures

(i) Variance In the early theory of portfolio selection, Markowitz proposed the
use of variance as a risk measure. Thus the risk measure ρ is given by

ρ(X) = Var(X) = E(X − E(X))2 ,

 Risk Modeling and Financial Risk Measures 207

and portfolio with small variance is considered less risky.

Here we note that variance as a risk measure is not a coherent risk measure as
it is not subadditive. Further it is not consistent with the first order stochastic
dominance because (X 1 Y) does not imply Var(X) ≤ Var(Y). Due to these
reasons, and also to take care of non symmetry in the loss distributions, risk
measures based on downside tail distribution of returns have been introduced.

(ii) Value-at-Risk For a loss random variable X with the cumulative distribution
function F, the value-at-risk of X with confidence level (1 − α), 0 < α < 1 is
defined as

VaR(1−α) (X) = F−1 (1 − α) = Min{x ∈ R : F(x) ≥ (1 − α)}.

Here (1 − α)-quantile is the position such that P(X ≤ F−1 (1 − α)) ≥ (1 − α)).
Thus, P(X > F−1 (1 − α)) ≤ α.
It can now be shown that VaR is again not a coherent risk measure because
it is not subadditive. We have already given such an example but below given
another example which is probably more general.
Example 6.9.1 Let X,Y be two independent and identically distributed Bernoulli
random variables with parameter p, i.e. P(X = 0) = (1 − p) and P(X = 1) = p.
Let (1 − p)2 < α < 1. Show that VaR is not sub additive in this situation.
Solution As (1 − p)2 < α < 1, we have 0 < (1 − α) < 2p(1 − p). Therefore
F−1
X
(1 − α) = 0 = F−1
Y
(1 − α). But F−1
X+Y
(1 − α) > 0. Therefore for this choice of
α, the subadditive law does not hold.


Let us recall the definition of SSD. We say that X 2 Y if
, t
(FY − FX )(u)du ≥ 0, ∀ t ∈ R.
−∞

With respect to SSD, 2 we define 2 as follows: X 2 Y if

, ∞
(FY − FX )(u)du ≥ 0, ∀ t ∈ R.
t

The order 2 is called risk seeking stochastic dominance (RSSD), also called
stop-loss-order which is very popular in acturial science.
Definition 6.9.6 (Distortion Function) A function g : [0, 1] → [0, 1] is
called a distortion function if (i) it is nondecreasing and (ii) g(0)=0, g(1)=1.
208 Financial Mathematics: An Introduction

Definition 6.9.7 (Risk Measure ρg ) Let g be the given distortion function.

A risk measure ρg (risk measure with respect to the given distortion function
g) is defined as
, ∞ , 0
ρ g (X) = g (1 − F(t)) dt + [g (1 − F(t)) − 1]dt,
0 −∞

where F is the cumulative distribution function of the random variable X (say

loss random variable).
Remark 6.9.4 VaR is a risk measure of type ρg . Some readers may like to
obtain the distortion function g for VaR.
Lemma 6.9.1. If the distortion function g is concave then
X 2 Y ⇒ ρg (X) ≤ ρg (Y).
Thus any risk measure ρg defined with respect to a concave distortion function
g is 2 always consistent.
Using the above lemma it can be shown that VaR is consistent with re-
spect to the stochastic dominance rule induced by the partial order 2 . Thus
(X 2 Y) ⇒ VaR(1−α) (X) ≤ VaR(1−α) (Y) for all α ∈ (0, 1).

(iii) Conditional Value-at-Risk Let us recall that CVaR(1−α) is the conditional

expectation of loss given that the loss exceeds VaR(1−α) . It can be shown that
unlike other risk measures CVaR is a coherent risk measure. Further using
Lemma 6.9.1, it is coherent with respect to the stochastic dominance rules
induced by the partial order 2 . Thus (X 2 Y) implies that CVaR(1−α)(X) ≤
CVaR(1−α)(Y) for all α ∈ (0, 1).
Remark 6.9.5 The proofs of most of results in this section have been omitted.
This has been done deliberately because the proofs are involved and need a good
knowledge of measure theory, in particular the theory of Choquet integrals. Inter-
ested readers are encouraged to refer to Sriboonchitta et al. [128] and references
cited therein.

6.10 Stochastic Dominance and Portfolio Optimization

Let rT = (r1 , r2 , . . . , rn ) be the vector of random returns for n assets and w :
(w1 , w2 , . . . , wn ) be the given portfolio. Then the portfolio return R(w) is given by
 Stochastic Dominance and Portfolio Optimization 209

the random variable rT w where wT = (w1 , w2 , . . . , wn ). In the context of portfolio

optimization we consider the stochastic dominance relations between the random
returns rT w(1) and rT w(2) corresponding to the two portfolios w(1) and w(2) . Thus
we say that portfolio w(1) dominates portfolio w(2) under the first order stochastic
dominance, written as w(1) 1 w(1) if R(w(1) ) 1 R(w(2) ). Thus w(1) 1 w(2) if
   
F R(w(1) ); η ≤ F R(w(2) ); η ∀ η ∈ R. (6.58)
 
Here F R(w(1) ), · is the cumulative distribution function of the random variable
R(w).
We say that w(1) 1 w(2) if R(w(1) ) 1 R(w(2) ) and R(w(2) ) 1 R(w(1) ). This
amounts to saying that at least one inequality is strict at (6.58).
Similarly w(1) 2 w(2) if R(w(1) ) 2 R(w(2) ), i.e.
   
F2 R(w(1) ); η ≤ F2 R(w(2) ); η ∀ η ∈ R, (6.59)
where , η
F2 (V; η) = F(V, ξ)dξ, ∀ η ∈ R.
−∞
We say that w(1) 2 w(2) if R(w(1) ) 2 R(w(2) ) and R(w(2) )  R(w(1) ). This amounts
to saying that at least one inequality is strict at (6.59).
Since stochastic dominance rules also try to capture preference between two
random alternatives, it is natural that there should be some connection between
these rules and associated utility functions. We have following results in this regard
(Sriboonchitta et al. [128])
Theorem 6.10.1 Let w(1) and w(2) be two portfolios. Then R(w(1) ) 1 R(w(2) ) if
and only if    
E u(R(w(1) )) ≥ E u(R(w(2) )) ,
for any nondecreasing function u : R → R for which these expectations are finite.
Theorem 6.10.2 Let w(1) and w(1) be two portfolios. Then R(w(1) ) 2 R(w(2) ) if
and only if    
E u(R(w(1) )) ≥ E u(R(w(2) )) ,
for any nondecreasing function u : R → R for which these expectations are finite.
Definition 6.10.1 (FSD-efficient portfolio) A feasible portfolio w is called
FSD- efficient if there is no other feasible portfolio w such that R(w) 1 R(w).
Definition 6.10.2 (SSD-efficient portfolio) A feasible portfolio w is called
SSD- efficient if there is no other feasible portfolio w such that R(w) 2 R(w).
210 Financial Mathematics: An Introduction

6.11 Portfolio Optimization with Stochastic Dominance

Constraints
In view of the above discussion it makes sense to study a typical portfolio opti-
mization problem having stochastic dominance constraints. Let us assume that a
reference random return Y having a finite expected value is available. For example
this could be the (random) return of a reference portfolio like an index portfolio
or the current portfolio itself. Then we introduce the following problem

Max E (R(w))
subject to,
R(w) 2 Y
(6.60)
eT w = 1,

where R(w) = rT w.
The problem (6.60) aims to determine those portfolios whose return dominate
the reference return under second order stochastic dominance rule. This problem
has continum of constraints because of the presence of the constraints R(w) 2 Y,
and as such is difficult to handle. Never the less, the application of stochastic dom-
inance in portfolio selection problem has been an important recent contribution in
the literature. Some notable contributions in this direction are due to Dentcheva
and Ruszczynski [37], Ruszczynski and Vanderbei [116], and Fábián, Mitra and
Roman [44].

6.12 Summary and Additional Notes

• This chapter presents certain developments post Markowitz’s mean-variance
theory with regard to other risk measures.
• Section 6.2 discusses certain theoritical and computational issues pertaining to
Markowitz’s model and need to discuss other risk measures.
• Section 6.3 presents mean absolute deviation based portfolio optimization
whereas Section 6.4 discusses minimax rule based portfolio optimization.
• Sections 6.5 and 6.6 are devoted to the study of value at risk and conditional
value at risk respectively. These are quantile based risk measures, in contrast
to moment based risk measures (L1 , L2 and L∞ ) which are discussed in earlier
sections.
• The expected utility maximization principle of Von Neumann - Mongestern’s
is important for rational decision making. This requires some discussion on
 Summary and Additional Notes 211

preference relation. After presenting a very brief introduction of these topics

in Section 6.7, Markowitz’s model is revisited in Section 6.8 in utility theory
frame work.
• Sections 6.9 and 6.10 respectively introduce coherent risk measures and stochas-
tic dominance. The application of stochastic dominance in portfolio optimiza-
tion is relatively new. These two topics provide a very basic theoretical frame
work of portfolio optimization. Much of the current research on portfolio opti-
mization is devoted to the application of this theoretical frame work to study
the real life portfolio selection problems.
• We have not discussed certain other interesting portfolio optimization prob-
lems. For example we have not included transaction costs in our study. Also
we have not looked into multi period and continuous portfolio optimization
problems. We may refer to Yu et al. [148], Zhou and Li [150] and references
cited therein in this regard. An interesting application of dynamic program-
ming to study multiperiod portfolio optimization problem is presented in Sun
et al. [131].
• Is there any tool which can help us to describe the possible dependency of
several random variables? Let us first try to understand the question itself. We
explain it through a simple example. Suppose X1 and X2 are two real-valued
random variables that can take values from {1, 2, . . . , 6}. Assume that we know
the value of X1 and we wish to infer the value of X2 from it. An important
thing to know here is that how much information we gained from knowing
the value of X1 ! In other words, it is same if we ask what is the interrelation
or dependence of these two random variables? In certain simple situations we
can indeed find the answer. For instance, match the above situation with the
procedure of throwing a fair die twice and the outcome on the first throw is X1
and the one of the second is X2 . The variables are independent; the knowledge
of X1 gives us no information about X2 . The situation will be quite different
if it is given that X1 is always the smaller number of the two throws and X2
is the larger one. We then have a strict monotonic relation X1 ≤ X2 . Now, if
X1 = 6, then we also know X2 = 6. If X1 = 5, we would guess that X2 is either
5 or 6, both with a probability 1/2, and so on. We need some mathematical
tool to depict the latter dependency. It is precisely what we ask for in the
beginning. The answer lies in the notion called Copulas. Copulas are the tools
for modeling and describing dependence of several random variables. The name
first appeared, in context of finance, in the work of Sklar (1959) and is derived
from the Latin word ‘copulare’, meaning ’to connect or to join’.
The concept of copula has received growing attention in finance and economics
212 Financial Mathematics: An Introduction

in recent years. In fact one of the main issues of risk management is the aggre-
gation of individual risks. Copulas have been used to aggregate risk in [42] and
[18] and many more references cited therein. They are often used to identify
market risk, credit risk and operational risk. Though in this book we have used
correlation coefficient to compute interdependence of returns of two or more
assets in a portfolio, yet it is equally important to inform readers about a rel-
atively complex but highly useful notion of copula. Useful because, correlation
parameter works well with normal distributions, while distributions in finan-
cial markets are mostly skewed and copulas are very handy to deal with the
skewness. For this reason, they have been applied in option pricing, portfolio
value-at-risk, and interest rates derivatives, to name a few. There are various
types of copulas, like, Frank copula, normal (Gaussian) copula, student copula,
singular copula, and many more. For more details on how copulas are used in
risk management, we can refer to [17, 18, 29], and numerous excellent articles
on web. We encourage you to hunt for this concept and its applications in
finance.

6.13 Exercises
Exercise 6.1 Consider the data of Exercise 1.1 of Chapter 1. Let there be
Rs 50,000 to be invested on these three assets. Formulate and solve the result-
ing portfolio optimization problem if
(i) L1 risk measure is used.
(ii) L2 risk measure is used.
(iii) L∞ risk measure is used.
Exercise 6.2 Suppose that the portfolios are constructed using three securities
a1 , a2 , a3 with expected returns µ1 = 20%, µ2 = 15% and µ3 = 4%. Further, let
expected absolute deviations be q1 = 50, q2 = 400 and q3 = 10 respectively. Let
there be M0 = Rs 50, 000 to be invested. Then for the risk tolerance parameter
λ = 0.5, obtain the minimax rule based optimal portfolio.
Exercise 6.3 Solve the above problem if q3 = 0 and there is no change in the
other data.
Exercise 6.4 An investor has utility function u(x) = x1/4 for his/her salary.
He/she has a new job offer which pays Rs 80,000 with a bonus. The bonus will be
Rs 0, Rs 10,000, Rs 20,000, Rs 30,000, Rs 40,000, Rs 50,000, or Rs 60,000
each with equal probability. Find the certainty equivalent of this job offer.
 Exercises 213

Exercise 6.5 Let X ∼ N(µ, σ2 ). Show that

(i) for each α ∈ (0, 1), F−1 X
(α) is proportional to α.
(ii) if X ≤ Y, a.s. then FX (α) ≤ F−1
−1
Y
(α).
(iii) F−1
λX
(α) = λF −1
X
(α) for λ > 0.
(iv) F−1
X+a
(α) = F −1
X
(α) + a for any a ∈ R.

Exercise 6.6 Find Var(1−α) (X) when

(i) X ∼ uniform[a,b].
(ii) X ∼ N(µ, σ2 ).
(iii) log X ∼ N(µ, σ2 ).

Exercise 6.7 Compute CVar(1−α) (X) when

(i) X ∼ N(µ, σ2 ).
(ii) log X ∼ N(µ, σ2 ).

Exercise 6.8 Let X be the class of profit or loss random variables with finite
variances. Verify that

(i) X is a convex cone.

(ii) If ρ : X → R+ given by ρ(X) = Var(X), then ρ is not a coherent risk measure.

Exercise 6.9 Let X and Y be two random variables such that

P(X = 2) = 1 − P(X = 12) = 1/5,

and
P(Y = 8) = 1 − P(Y = 18) = 4/5.

(i) Verify that E(X) = E(Y) and V(X) = V(Y).

(ii) Consider the utility function u(x) = log x for x > 0. Does any risk averse
investor feel the indifference between the two risky prospects X and Y?
(Hint: Is E (u(X)) = E (u(Y))?
214 Financial Mathematics: An Introduction

Exercise 6.10 Let  be a preference relation on a given set Z. Show that

(i) z  x and x ≈ y ⇒ z  y.
(ii) z  x and z ≈ y ⇒ y  x.
Exercise 6.11 Prove that
(i) X 1 Y ⇒ X 2 Y.
(ii) X 2 Y, E(X) = E(Y) ⇒ V(X) ≤ V(Y).
(iii) X 1 Y ⇔ F−1X
(·) ≤ F−1
Y
(·).
Exercise 6.12 Let (t − X)+ = Max(t − X, 0). Show that X 2 Y if and only if
E ((t − X)+ ) ≤ E ((t − Y)+ ) for all t ∈ R.
7
Stochastic Processes

7.1 Introduction
In our earlier chapters, we have been making frequent use of terms like random
variables and random vectors. In fact, any first course on probability theory dis-
cusses these concepts in detail. The aim of this chapter is to go beyond the notion
of a random vector, and introduce a collection of random variables parameterized
by a parameter, say time. Such a collection of random variables is essentially a
stochastic process.
Stochastic processes occur naturally in real life applications. For instance con-
sider the (opening) exchange rate: Indian rupees (INR)/US dollar (USD) at every
day between August 18, 2011 and February 2, 2012 except weekends. The actual
exchange rate is shown in Fig. 7.1. As on any given day, the exchange rate is
random, we can interpret this figure as a realization xn of the random variable Xn ,
where n is the day 1, 2, . . .. In order to make a guess of the interest rate on a fu-
ture date, it is reasonable to look at the whole evolution of Xn between August 18,
2011 and February 2, 2012. Therefore, there is a need into develop theory which
provides almost continuous information about the process considered. A mathe-
matical representation for describing such a phenomenon leads to the notion of
stochastic process.
Stochastic processes play a vital role in financial mathematics, because the
proper understanding of asset dynamics is crucial for making any meaningful
financial decision. The derivation of Black-Scholes formula for option pricing il-
lustrates this point very well. The derivative pricing and interest rate modeling
can be studied in greater depth using the concept of stochastic process. In this
chapter, we aim to present a brief introduction of stochastic processes and some
of the related concepts keeping finance in view.
216 Financial Mathematics: An Introduction

54

53

52

Indian Rupees per USD

51

50

49

48

47

46

45
0 20 40 60 80 100 120
Days (August 18, 2011 − February 2, 2012)

Fig. 7.1. Exchange Rate: Indian Rupees (INR) / US Dollar (USD) from August 18, 2011 to February
2, 2012

7.2 Definitions and Simple Stochastic Processes

In this section, we first review certain definitions for random variables and then
proceed formally to define a stochastic process. We shall denote the random ex-
periment by E and the corresponding sample space by Ω.
Definition 7.2.1 (σ-field) A σ-field F (or σ-algebra) is a family of subsets of
Ω which satisfies the following properties
(i) ∅ ∈ F ,
(ii) if A ∈ F then its complement Ac ∈ F ,
(iii) if a countable sequence of sets A1 , A2 , . . . is in F , then ∪i Ai ∈ F .
From the above it is simple to conclude that Ω ∈ F , and if a countable sequence
of sets A1 , A2 , . . . is in F then ∩i Ai ∈ F .
Obviously, the smallest possible σ-field is {∅, Ω}. It is called the trivial σ-field.
When the set Ω is finite, the collection of all subsets of Ω is a σ-field, which is
same as the power set of Ω. It is called the total σ-field and is the largest σ-field
on Ω. The set {∅, A, Ac , Ω} where A is a proper subset of Ω (∅ ⊂ A ⊂ Ω), is
another σ-field on Ω; it is called the σ-field generated by the set A. The σ-field F
generated by two proper subsets A and B is
 Definitions and Simple Stochastic Processes 217

F = {∅, A, B, Ac , Bc , A ∩ B, A ∩ Bc , Ac ∩ B, Ac ∩ Bc , (A
B)c , (A
B),

A ∪ B, A ∪ Bc , Ac ∪ B, (A ∩ B)c , Ω} .

Here A
B is defined as (A ∩ Bc ) ∪ (B ∩ Ac ) and is called the symmetric difference

of sets A and B. Further it may be noted that given two σ-fields F1 and F2 , F1 ∩F2
is always a σ-field but F1 ∪ F2 may not be a σ-field.
Example 7.2.1 Let Ω = {1, 2, 3} and F = {∅, {1}, {2}, {3}, {1, 2}, {2, 3}, {1, 3}, Ω}.
Verify that F is a σ-field on Ω.
Solution We must show that properties (i), (ii) and (iii) of Definition 7.2.1 of σ-
field are satisfied by given F . We have ∅ ∈ F . Also ∅c = Ω, {1}c = {2, 3}, {1, 2}c = {3}
and so on. Hence (ii) is satisfied. Also, {1}∪{2, 3} = {1, 2, 3}, {1}∪{3} = {1, 3}. Further
this property can be checked for other sets as well. Hence F satisfies (i), (ii) and
(iii). Therefore, F is a σ-field on Ω.


Example 7.2.2 Let Ω = {a, b, c, d}. Let F1 = {∅, {a}, {b, c, d}, Ω} and F2 =
{∅, {a, b}, {c, d}, Ω}. Show that F1 ∪ F2 is not a σ-field. What is the largest σ-field
in Ω?
Solution By the definition of σ-field, F1 and F2 are σ-fields. Now, F1 ∪ F2 =
{∅, {a}, {a, b}, {c, d}, {b, c, d}, Ω}. Clearly, F1 ∪ F2 is not a σ-field because {a} ∪
{c, d} = {a, c, d}  F1 ∪ F2 . Further, F = {∅, {a}, {b}, {c}, {d}, {a, b}, {c, d}, {a, c}, {b, d},
{a, d}, {b, c}, {a, b, c}, {b, c, d}, {a, c, d}, {a, b, d}, Ω} is the largest σ-field.


Let F be the given σ-field. Then F consists of those subsets of Ω which satisfy
conditions (i), (ii) and (iii) of Definition 7.2.1. These subsets of Ω are called
F -events or F -measurable sets. The pair (Ω, F ) is called a measurable space.
Measurable spaces are the building blocks for defining probability spaces.
Definition 7.2.2 (Probability Measure) Let F be a σ-field over Ω. Let P be
a real-valued function defined on F such that
(i) P(A) ≥ 0 for all A ∈ F ,
(ii) P(Ω) = 1,
(iii) if A1 , A2 , A3 , . . . are mutually exclusive events, then
⎛∞ ⎞
⎜⎜4 ⎟⎟  ∞
P ⎜⎜⎝ Ai ⎟⎟⎠ = P(Ai ).
i=1 i=1
218 Financial Mathematics: An Introduction

Then P(.) is called a probability measure. The triplet (Ω, F , P) is called a proba-
bility space.
In most of the random experiments, either it is not feasible or it is difficult
to obtain the probability of an event using definition. However, it is possible
to calculate the probability of an event and related probability distributions by
mapping the set of all possible outcomes into real line. This leads to the definition
of a random variable.
Definition 7.2.3 (Random Variable) Let (Ω, F , P) be given probability space.
A function X : Ω → R is called a random variable if for any x in R, X−1 {(−∞, x]}
belongs to F . Here
X−1 {(−∞, x]} = {w ∈ Ω : X(w) ∈ (−∞, x]} .
Remark 7.2.1 For a given probability space (Ω, F , P), when F is the largest σ-
field on Ω, any real-valued function defined on Ω will be a random variable.
Remark 7.2.2 Let Ω = R and U (1) = {(a, b] : −∞ < a < b < ∞}. Then the σ-field
generated by U (1) is called the Borel σ-field, and its elements are called Borel sets.
Thus Borel σ-field is the σ-field generated by the family of semi-open intervals of
type (a, b], −∞ < a < b < ∞.
Remark 7.2.3 A function f from a measurable space (Ω, F ) into (R, B) is a said
to be measurable function if for each Borel set B ∈ B, the set {w ∈ Ω : f (w) ∈
B} ∈ F . Here B is the Borel σ-field of subsets of R. From the above definition, a
random variable X is a measurable function from (Ω, F ) to (R, B).
Remark 7.2.4 As X : Ω → R, a statement of type a ≤ X ≤ b is to be understood
as {w ∈ Ω : a ≤ X(w) ≤ b}. In that case P(a ≤ X ≤ b) = P{w ∈ Ω : a ≤ X(w) ≤ b}.
In general for a Borel set B, P(X ∈ B) = P{w ∈ Ω : X(w) ∈ B}. Here it is assumed
that (Ω, F , P) is the underlying probability space.
The function which records the probabilities associated with the random vari-
able is defined as below.
Definition 7.2.4 (Cumulative Distribution Function of a Random Vari-
able) The cumulative distribution function (CDF) of a random variable X is given
as
FX (x) = P(X ≤ x) = P(w ∈ Ω : X(w) ≤ x), − ∞ < x < ∞ .
For a < b, we have
P(a < X ≤ b) = P(w ∈ Ω : a < X(w) ≤ b) = FX (b) − FX (a) .
 Definitions and Simple Stochastic Processes 219

Example 7.2.3 Let Ω = {a, b, c, d} and F = {∅, {a}, {b, c, d}, Ω}. Let (Ω, F , P) be
the given probability space where P({a}) = 1/4 and P({b, c, d}) = 3/4. Define
⎧
⎪
⎪ 0, w=a
⎪
⎪ 1,
⎪
⎨ w=b
X(w) = ⎪⎪
⎪ 2,
⎪ w=c
⎪
⎩ 3, w = d,

and

0, w=a
Y(w) =
1, w ∈ {b, c, d}.
Check whether X and Y are random variables? If so, find the corresponding CDFs.
Solution As X−1 {(−∞, 1]} = {a, b}  F , the function X is not a random variable.
Hence CDF of X is not defined.
Next, for Y we have
⎧
⎪
⎪ ∅, −∞ < y < 0
−1
⎪
⎨
Y {(−∞, y]} = ⎪ ⎪ {a}, 0≤y<1
⎪
⎩ Ω, 1≤y<∞.
5 6
Since, for every y ∈ R, Y−1 (−∞, y] ∈ F , Y is a random variable. The CDF of Y
is given by ⎧
⎪
⎪ 0, −∞ < y < 0
⎪
⎨1
FY (y) = ⎪
⎪ , 0≤y<1
⎪ 4
⎩ 1, 1≤y<∞.


To understand uncertainty in a static system, we employ standard probability
theory. But to handle uncertainty in a dynamic system, e.g. the price of a share of
stock, we need to go beyond to capture evolution over time. This can be studied
via a collection of random variables namely a stochastic process with time as index
set. Now, we define the stochastic process formally.
Definition 7.2.5 (Stochastic Process) Let (Ω, F , P) be a given probability
space. A collection of random variables {X(t), t ∈ T} defined on the probability
space (Ω, F , P) is called a stochastic process.
A stochastic process is also called a random process or a chance process.
220 Financial Mathematics: An Introduction

Remark 7.2.5 Given a probability space (Ω, F , P), the stochastic process {X(t), t ∈
T} can be identified as a real-valued function X : T × Ω → R of two independent
variables t ∈ T, w ∈ Ω such that X−1 (t) {(−∞, x]} belongs to F for every x ∈ R and
for every t ∈ T. Here

X−1 (t) {(−∞, x]} = {w ∈ Ω : X(t, w) ∈ (−∞, x]} .

In this sense, the stochastic process {X(t) , t ∈ T} can be thought of as

{X(t, w) : t ∈ T, w ∈ Ω}.
Remark 7.2.6 The mapping X gives rise to two mappings, X(·, w) and X(t, ·) as
follows. For fixed w ∈ Ω, the mapping X(., w) : T → R takes t to X(t, w). The
sequence of real numbers {X(t, w), t ∈ T} for fixed w ∈ Ω is called the realization or
trajectory or sample path of the stochastic process. This is not a random variable.
Next, for each fixed t ∈ T, the mapping X(t, .) : Ω → R takes w to X(t, w) which
is a random variable.
Definition 7.2.6 (Parameter Space and State Space) Let {X(t), t ∈ T} be
a given stochastic process. The set {t ∈ T} is called the parameter space or index
set. The collection of all possible values of X(t) for t ∈ T is called the state space.
The state space is denoted by S.
A given stochastic process is classified based on the values of state space
and parameter space. For example, if both the state space and parameter space
are finite or countably infinite, then it is called discrete time discrete space
stochastic process. We shall be denoting a discrete time stochastic process by
{Xt , t = 0, ±1, ±2, . . .} while the continuous time stochastic process will be denoted
by {X(t), t ∈ T}. Note that, a time series {Xt , t = 0, ±1, ±2, . . . } is a discrete time
stochastic process with T = Z = {0, ±1, ±2, . . .}.
Based on parameter space T and state space S, we now present some of the
examples of different stochastic processes.
(i) Continuous time discrete space stochastic process:
T = R+ = [0, ∞), S = N, e.g. total number of shares {X(t), t ∈ [0, ∞)} held by
an investor at any time t,
(ii) Discrete time continuous space stochastic process:
T = N, S = R+ = [0, ∞), e.g. {Xn , n = 1, 2, 3, . . .} value of one US dollar (USD)
in rupees at the end of n-th day.
(iii) Continuous time continuous space stochastic process:
T = R+ = [0, ∞), S = R+ = [0, ∞), e.g. {X(t), t ∈ [0, ∞)} stock price of a
particular item at any time t.
 Definitions and Simple Stochastic Processes 221

(iv) Discrete time discrete space stochastic process:

T = N, S = N, e.g. total number of shares {Xn , n = 1, 2, 3 . . .} held by a
particular person at the end of n-th day.
We already know that a random variable X can be characterized by its CDF
FX (.). The main question now is to characterize a stochastic process. For this we
need to define finite dimensional distributions of the given process.
Definition 7.2.7 (Finite Dimensional Distributions of a Stochastic Pro-
cess) Let {X(t), t ∈ T} be a given stochastic process. Then finite dimensional distri-
butions of the given process are the distributions of the finite dimensional random
vectors (X(t1 ), X(t2 ), . . . , X(tn )), t1 , t2 , . . . , tn ∈ T, for all possible choices of times
t1 , t2 , . . . , tn ∈ T and for all n ≥ 1.
It can be shown that the stochastic process {X(t), t ∈ T}, is known completely if
all its finite dimensional distributions are known.
Now we introduce certain properties of stochastic process {X(t), t ∈ T}.
Definition 7.2.8 (Strict Sense Stationary Stochastic Process) The stochas-
tic process {X(t), t ≥ 0} is called strict sense stationary stochastic process if, for
arbitrarily chosen n, and for 0 < t1 < t2 < . . . < tn , the finite dimensional ran-
dom vectors (X(t1 ), X(t2 ), . . . , X(tn )) and (X(t1 + h), X(t2 + h), . . . , X(tn + h)) have
the same joint distributions for all h > 0 and all t1 , t2 , . . . , tn .
Note that a strict sense stationary stochastic process is invariant with respect to
time translation. It is also called a strong stationary stochastic process.
Definition 7.2.9 (Wide Sense Stationary Stochastic Process) The stochas-
tic process {X(t), t ≥ 0} is wide sense stationary if it satisfies the following
(i) E (X(t)) is independent of t,
(ii) Cov (X(t), X(s)) depends only on the time difference | t − s | for all t, s,
 
(iii) E (X(t))2 < ∞ (finite second order moment).
A wide sense stationary stochastic process is also called a covariance stationary
or weak stationary or second order stationary stochastic process.
In other words, a stochastic process with finite second order moments is said to
be weak stationary if its mean (or expectation) function is independent of t and
covariance function is invariant under time shift.
Example 7.2.4 Consider the process X(t) = A cos(θt) + B sin(θt) where A and B
are uncorrelated random variables with mean 0 and variance 1 and θ is a positive
constant. Is {X(t), t ≥ 0} covariance or wide sense stationary stochastic process?
222 Financial Mathematics: An Introduction

Solution Since E(A) = 0 and E(B) = 0, we have, for a fixed t,

E(X(t)) = cos(θt)E(A) + sin(θt)E(B) = 0.

Hence E(X(t)) is independent of t.

Now for s < t,

Cov(X(t), X(s)) = E((X(t) − 0)(X(s) − 0))

= E(X(t)X(s))
= E ((A cos(θt) + B sin(θt)) (A cos(st) + B sin(st))) .

Since E(A2 ) = 1 and E(AB) = E(A)E(B) = 0, we get

Cov(X(t), X(s)) = cos(θt) cos(θs) + sin(θt) sin(θs)

= cos(t − s)θ.

Hence Cov(X(t), X(s)) depends only on | t − s |.

Next for fixed t,

E(X2 (t)) = Var(X(t)) − (E(X(t)))2

= cos2 (θt)Var(A) + sin2 (θt)Var(B) + 2 cos(θt) sin(θt)Cov(A, B)
= cos2 (θt) + sin2 (θt)
= 1.

Hence, by Definition 7.2.9, {X(t), t ≥ 0} is a wide sense stationary process.



Definition 7.2.10 (Stationary Increments) The process {X(t), t ∈ T} is said to
have stationary increments if X(t + τ) − X(t) is a strict sense stationary process for
all τ > 0, i.e. whenever s < t, X(t)−X(s) has the same distribution of X(t−s)−X(0).
Definition 7.2.11 (Independent Increments) If for all n and for t1 < t2 <
. . . < tn , X(t2 ) − X(t1 ), X(t3 ) − X(t2 ),. . . , X(tn ) − X(tn−1 ) are independent random
variables, then the process is said to have independent increments.
Definition 7.2.12 (Independent 7 Process) If for all n and for t1 < t2 <
. . . < tn the CDF satisfies F(X, T) = ni=1 F(X(ti ), ti ) where T = (t1 , t2 , . . . , tn ) and
X = (X(t1 ), X(t2 ), . . . X(tn )), then the process is called an independent process. Here
F(X, T) is the joint distribution function of the random vector X.
 Definitions and Simple Stochastic Processes 223

Note that, a stochastic process having independent increments is different from a

stochastic process which is an independent process.
Definition 7.2.13 (Markov Property) A given process {X(t), t ∈ T} is said
to have Markov property if for all n and for all 0 < t1 < t2 < . . . < tn < t, the
conditional CDF satisfies

P (X(t) ≤ x/X(0) = x0 , X(t1 ) = x1 , . . . , X(tn ) = xn ) = P (X(t) ≤ x/X(tn ) = xn ) .

In other words, if the future prediction depends only on the current state of the
stochastic process and does not depend on the past information, then it has the
Markov property.
Definition 7.2.14 (Markov Process) A given stochastic process {X(t), t ∈ T}
is said to be a Markov process if it satisfies Markov property.
A Markov process is a stochastic process with property that, given the value
of X(s), the values of X(t), t > s, do not depend on the values of X(u), u < s, i.e.
the probability of any particular future behavior of the process, when it’s present
state is known exactly, is not altered by additional knowledge concerning it’s past
behavior. For instance, binomial process, Poisson process and Brownian motion
are examples of Markov process whereas Gaussian process is an example of non
Markov process. All these processes will be discussed later in this chapter.
Stochastic processes can be classified according to different criteria. One of
them is the distribution of finite dimensional random variables. Based on this,
we list few stochastic processes which are related to stochastic study of financial
mathematics.
Definition 7.2.15 (Bernoulli Process) A discrete time discrete space stochas-
tic process {Xn , n = 1, 2, . . .} is called a Bernoulli process if for each n, Xn is a
Bernoulli random variable with parameter p, 0 < p < 1.
Here we may note that the Bernoulli process is strict sense as well as wide sense
stationary process.
Definition 7.2.16 (Binomial Process) Let for n = 1, 2, . . ., Sn = X1 + X2 +
. . .+Xn where Xi ’s are mutually independent Bernoulli distributed random variables
with parameter p, 0 < p < 1. Then {Sn , n = 1, 2, . . .} is called a binomial process.
Here we may note that {Sn , n = 1, 2, . . .} is a discrete time discrete space stochastic
process. Also each Sn is a binomial distributed random variable.
224 Financial Mathematics: An Introduction

Lemma 7.2.1 The binomial process {Sn , n = 1, 2, . . .} is a Markov process.

Proof. For the binomial process {Sn , n = 1, 2, . . .}, we have, for n = 2, 3, . . .,

Sn = Sn−1 + Xn .

Hence,
P(Sn = x, Sn−1 = xn−1 , . . . , S1 = x1 )
P(Sn = x/Sn−1 = xn−1 , . . . , S1 = x1 ) =
P(Sn−1 = xn−1 , . . . , S1 = x1 )

P(Sn = x/Sn−1 = xn−1 )P(Sn−1 = xn−1 /Sn−2 = xn−2 ) . . . P(S2 = x2 /S1 = x1 )P(S1 = x1 )
=
P(Sn−1 = xn−1 /Sn−2 = xn−2 ) . . . P(S2 = x2 /S1 = x1 )P(S1 = x1 )
P(Xn = x − xn−1 )P(Xn−1 = xn−1 − xn−2 ) . . . P(X2 = x2 − x1 )P(S1 = x1 )
=
P(Xn−1 = xn−1 − xn−2 ) . . . P(X2 = x2 − x1 )P(S1 = x1 )
= P(Xn = x − xn−1 )
= P(Sn = x/Sn−1 = xn−1 ) .

Thus {Sn , n = 1, 2, . . .} satisfies the Markov property. Hence {Sn , n = 1, 2, . . .} is a

Markov process.


Definition 7.2.17 (Gaussian Process) A given stochastic process {X(t), t ∈ T}
is said to be a Gaussian process if, for t1 , t2 , . . . , tn and all n, (X(t1 ), X(t2 ), . . . , X(tn ))
is a multivariate normal distributed random vector.
We may observe the following for a Gaussian process.
(i) The joint pdf of (X(t1 ), X(t2 ), . . . , X(tn )) is given by
 
1 1 T −1
fX (x) = exp − (X − µ) Σ (X − µ) ,
(2π)n/2 (det(Σ))1/2 2

where

µ = E(X) = (E(X(t1 )), E(X(t2 )), . . . , E(X(tn ))) ,

Σ = (Cov(X(ti ), X(t j )); i, j = 1, 2, . . . , n),

and
Cov(X(ti ), X(t j )) = E[(X(ti ) − E(X(ti ))(X(t j ) − E(X(t j )))] .
 Definitions and Simple Stochastic Processes 225

(ii) A Gaussian process which is weak stationary is also strict sense stationary.
Further, in any Gaussian process, its finite dimensional random variables are
independent if and only if those random variables are uncorrelated, i.e. the
corresponding covariance matrix Σ is a diagonal matrix.
Definition 7.2.18 (Symmetric Random Walk) Consider a random exper-
iment of tossing a fair coin infinitely many times. Let the successive outcomes
be denoted as w = (w1 , w2 , w3 , . . .) e.g. w = (w1 , w2 , w3 , . . .) = (H, T, T, . . .) or
(T, H, T, . . .) etc. We now define, for j = 1, 2, . . .


1, if w j = H
Xj =
−1, if w j = T,
and
P(X j = 1) = P(w j = H) = 0.5 ; P(X j = −1) = P(w j = T) = 0.5 .
Set S0 = 0. Let

k
Sk = X j, (k = 1, 2, . . .) .
j=1

Then, {Sk , k = 0, 1, . . .} is known as a symmetric random walk.

Now, we present the sample path of a random walk {S(k), k = 0, 1, . . .} for fixed
w ∈ Ω. Suppose w = (w1 , w2 , w3 , . . .) = (H, H, H, T, H, H, T, H, H, H, H, . . .). This
realization is shown in Fig. 7.2. This symmetric random walk {Sk , k = 0, 1, . . .} is
a discrete time, discrete space stochastic process. By considering continuous time
and continuous space, we get Brownian motion which will be discussed in the next
section.
Next, we present the following results for a symmetric random walk.
Theorem 7.2.1 Let {Sk , k = 0, 1, 2, . . .} be a symmetric random walk. Then
(i) for each k, E (Sk ) = 0 and Var(Sk ) = k,
(ii) it has independent increments,
(iii) it has stationary increments,
(iv) it is a Markov process.
Proof.
(i) For j = 1, 2, . . ., we have E(X j ) = 0 and Var(X j ) = 1. Therefore, for k = 1, 2, . . .,
E(Sk ) = 0 and Var(Sk ) = k .
Here we are using the fact that Xi s are mutually independent.
226 Financial Mathematics: An Introduction

0
Sk

−2

−4

−6

−8
0 20 40 60 80 100
k

Fig. 7.2. Sample path of symmetric random walk

(ii) We choose an arbitrary positive integer n and then choose non-negative inte-
gers 0 = k0 < k1 < . . . < kn . Then


ki+1
Ski+1 − Ski = Xj .
j=ki +1

Since X j are independent and identically distributed (i.i.d) random variables

having Bernoulli distribution, Sk1 − Sk0 , Sk2 − Sk1 , . . . , Skn − Skn−1 are mutually
independent variables. Hence the stochastic process {Sk , k = 0, 1, . . .} has the
independent increment property.
(iii) Choose non-negative integers k1 < k2 . Then


k2
Sk2 − Sk1 = X( j) .
j=k1

Since X j are i.i.d random variables having Bernoulli distribution, Sk2 − Sk1 has
the same distribution of Sk2 −k1 −S0 . Hence the stochastic process {Sk , k = 0, 1, . . .}
has the stationary increment property.
(iv) We have for k = 1, 2, . . .
Sk = Sk−1 + Xk .
 Definitions and Simple Stochastic Processes 227

Now,
P(Sk ≤ x, Sk−1 = xk−1 , . . . , S1 = x1 )
P(Sk ≤ x/Sk−1 = xk−1 , . . . , S1 = x1 =
P(Sk−1 = xk−1 , . . . , S1 = x1 )

P(Sk ≤ x/Sk−1 = xk−1 ) . . . P(S2 = x2 /S1 = x1 )P(S1 = x1 )

=
P(Sk−1 = xk−1 /Sk−2 = xk−2 ) . . . P(S2 = x2 /S1 = x1 )P(S1 = x1 )
= P(Sk ≤ x/Sk−1 = xk−1 ) .

From the above discussion we can easily conclude that {Sk , k = 1, 2, . . .} is a

Markov process.


Definition 7.2.19 (Poisson Process) A stochastic process {N(t), t ≥ 0} is said
to be a Poisson process with parameter λ > 0, if it satisfies the following properties
(i) N(0) = 0,
(ii) For all n and for all 0 ≤ t0 < t1 < t2 < · · · < tn , increments N(ti ) − N(ti−1 ), i =
1, 2, . . . , n are independent and stationary,
(iii) For 0 ≤ s < t, N(t) − N(s) is a Poisson distributed random variable with
parameter λ(t − s), i.e.

(λ(t − s))n e−λ(t−s)

P(N(t) − N(s) = n) = , (n = 0, 1, 2, . . .) .
n!
The sample path of a Poisson process with parameter λ = 2 is shown in Fig.
7.3. It is observed that in any finite interval, the sample path has finite jumps of
size one and in the interval [0, ∞) it has countably infinite jumps.
Let {N(t), t ≥ 0} be a Poisson process. Then, for each t, E(N(t)) = λt and
Var(N(t)) = λt.
Remark 7.2.7 By property (ii) of Definition 7.2.19, for every s < t, the incre-
ment N(t) − N(s) is independent from the history of the process up to time s and
has the same law as N(t − s). Therefore {N(t), t ≥ 0} satisfies Markov property.
Hence, it is a Markov process.
Remark 7.2.8 Suppose N(t) denotes the number of events that occur in the in-
terval [0, t]. Let Ti , i = 1, 2, . . . denote the occurrence of events. The inter arrival
of successive events T1 , T2 − T1 , . . . , Ti+1 − Ti are independent and are exponentially
distributed with parameter λ.
228 Financial Mathematics: An Introduction

14

12

10

N(t) 8

0
0 2 4 6 8 10
t

Fig. 7.3. Sample path of Poisson process with λ = 2

Example 7.2.5 Consider a financial risk model that may occur in an insurance
company. Let N(t) denote the number of claims received by time t, t ≥ 0. Thus,
N(t) counts the number of claims received in (0, t]. Suppose the process {N(t), t ≥ 0}
is a Poisson process with claim arrival rate 9 per month of 30 days. In a randomly
chosen month of 30 days,
(i) what is the probability that there are exactly 4 claims were received in the first
15 days?
(ii) given that exactly 4 claims were received in the first 15 days, what is the
probability that all the four claims were received in the last 7 days out of these
15 days?
Solution We have
e−λt (λt)k
P(N(t) = k) = , (k = 0, 1, . . .),
k!
9
where λ = per day.
30
(i) The required probability is
9
e− 2 94
P(N(15) = 4) = .
384
 Brownian Motion and its Properties 229

(ii) The required probability is

4
P(N(8) = 0)P(N(7) = 4) e−36/15 e−63/30 (63/30)4 7
= = .
P(N(15) = 4) −9/2
e (9/2) 4 15



7.3 Brownian Motion and its Properties

The long studied model known as Brownian motion, is named after the English
botanist Robert Brown. In 1827, Brown described the unusual motion exhibited
by a small particle totally immersed in a liquid or a gas. Early nineteenth cen-
tury, Brownian motion is introduced to model the price movements of stocks
and commodities. A formal mathematical description of Brownian motion and its
properties was first given by the great mathematician Norbort Wiener beginning
1918.
Definition 7.3.1 (Brownian Motion) A stochastic process {W(t), t ≥ 0} is
said to be a Brownian motion (BM), if it satisfies the following properties
(i) W(0) = 0,
(ii) for t > 0, the sample path of W(t) is continuous,
(iii) for all n and for all 0 ≤ t0 < t1 < t2 < · · · < tn , increments W(ti ) − W(ti−1 ), i =
1, 2, . . . , n are independent and stationary,
(iv) for 0 ≤ s < t < ∞, W(t) − W(s) is normally distributed random variable, with
mean 0 and variance (t − s).
A Brownian motion is also called a Wiener process. The sample path of Wiener
process is shown in Fig. 7.4. Wiener process can also be visualized as a scaling
limit of a symmetric random walk.
Remark 7.3.1 We observe that, the sample paths are always continuous func-
tions. In other words, the sample paths do not have any jumps. On the other
hand, these paths are also essentially nowhere differentiable, that is, it is not pos-
sible to define a unique tangent line at any point on a curve. We can show this by
using the convergence in the second order moment. For this we shall like to show
that  
W(t0 + ∆t) − W(t0 )
lim Var
∆t→0 ∆t
does not exist for any arbitrary point t0 .
230 Financial Mathematics: An Introduction

3.5

2.5

W(t) 1.5

0.5

−0.5

−1

−1.5
0 2 4 6 8 10
t

Fig. 7.4. Sample path of Wiener process

We know that, W(t0 + ∆t) − W(t0 ) has normal distribution with mean zero and
variance ∆t, i.e. N(0, ∆t). Hence,
 
W(t0 + ∆t) − W(t0 ) 1 1
lim Var = lim × ∆t = lim .
∆t→0 ∆t ∆t→0 (∆t)2 ∆t→0 ∆t

Thus the above limit does not exist. Hence, we conclude that Brownian motion
has non-differentiability at a fixed point. We observe that nowhere differentiability
of Brownian motion requires a more careful argument than non-differentiability
at a fixed point. The precise mathematical proof is not presented in this book but
rather the above argument may convince to conclude nowhere differentiability of
Brownian motion (refer Karatzas and Shreve [75] for proof ).
Remark 7.3.2 The Wiener process is not wide sense stationary. This is because
for s < t, the covariance function Cov (W(t), W(s)) is not a function of (t − s). In
fact
Cov (W(t), W(s)) = E[(W(t) − E(W(t)))(W(s) − E(W(s)))]
= E[W(t)W(s)]
= E[(W(t) − W(s) + W(s))W(s)]
= E[W(t) − W(s)]E[W(s)] + E[(W(s))2 ]
= (0 × 0) + s = s .
 Brownian Motion and its Properties 231

Hence, Cov (W(t), W(s)) = Min(s, t).

Remark 7.3.3 Given W(t), the future W(t + h) for any h > 0 only depends on the
increment W(t + h) − W(t) and this is independent of the past. Hence by Definition
7.3.1, the Markov property is satisfied. Thus, {W(t), t ≥ 0} is a Markov process.
Lemma 7.3.1 The joint probability density function of (W(t1 ), W(t2 )) is given by

2 8
1 1 x1 (x2 − x1 )2
f (x1 , x2 ) = exp − + .
2π t1 (t2 − t1 ) 2 t1 (t2 − t1 )

Proof. From Definition 7.2.17, a Wiener process is also a Gaussian process. Since
{W(t), t ≥ 0} is a Markov process as well as a Gaussian process, we have

P (W(t) ≤ x/W(tn ) = xn ) = P (W(t) − W(tn ) ≤ x − xn )

, x−xn  
1 s2
= exp − ds .
−∞ 2π(t − tn ) 2(t − tn )

Let us now consider the joint distribution of (W(t1 ), W(t2 )). We know that, W(t1 )
and (W(t2 ) − W(t1 )) are independent. Also, W(t1 ) is N(0, t1 ) and W(t2 ) − W(t1 ) is
N(0, t2 − t1 ). Therefore, the joint probability density function of (W(t1 ), W(t2 )) is

f (x1 , x2 ) = p(x1 , t1 )p(x2 − x1 , t2 − t1 ),

where  2
1 x
p(x, t) = √ exp − , t > 0; −∞ < x < ∞ .
2πt 2t
Thus,

2 8
1 1 x1 (x2 − x1 )2
f (x1 , x2 ) = exp − + .
2π t1 (t2 − t1 ) 2 t1 (t2 − t1 )


Remark 7.3.4 The above Lemma can be generalized in the sense that for 0 <
t1 < t2 < · · · < tn , X = (W(t1 ), W(t2 ), . . . , W(tn )) is jointly normally distributed
with the joint probability density function given by
 
1 1 T −1
fX (x) = exp − (X − µ) Σ (X − µ) ,
(2π)n/2 (det(Σ))1/2 2
232 Financial Mathematics: An Introduction

where

µ = E(X) = (E(W(t1 )), E(W(t2 )), . . . , E(W(tn ))) ∈ Rn

and

Σ = [Cov(W(ti ), W(t j )); i, j = 1, 2, . . . , n] ∈ Rn×n .

We now present some important properties of Brownian motion and refer to

Shreve [122] for their proofs.
Result 7.3.1 Let {W(t), t ≥ 0} be a Wiener process. Then,
(i) {−W(t), t ≥ 0} is a Wiener process. (symmetric property)
1
(ii) { √ W(ct), t ≥ 0} is a Wiener process for each fixed c > 0. (scaling property)
c
(
(iii) Taking W(0) ( = tW( 1 ), t > 0, the process {W(t),
= 0, W(t) ( t ≥ 0} is a Wiener
t
process. (time inversion property)
Example 7.3.1 Let {W(t), t ≥ 0} be a Wiener process. Find the conditional dis-
tribution of W(s) for 0 < s < t, given W(t) = x .
Solution There are two possible approaches to solve this problem. In the first ap-
proach we find the joint distribution of (W(t), W(s)) and then find the conditional
distribution. In the second approach we use the time inversion property in Result
1 1
7.3.1. The required conditional distribution is sW given that tW = x. This
  1 s    t
1 1 x
is same as the distribution of sW 1s − sW + sW given W = . Hence,

 t  1  t sx t t
1
we need to find the distribution of s W −W + . This distribution is a
s t
t
sx 1 1
normal distribution with mean and variance s2 − .
t s t



7.4 Processes Derived from Brownian Motion

A general Brownian motion need not have W(0) = 0 and σ2 = 1. Therefore we
define a general Brownian motion with drift µ and variance σ2 as follows.
 Processes Derived from Brownian Motion 233

Definition 7.4.1 (Brownian motion with drift µ and volatility σ) A

stochastic process {X(t), t ≥ 0} is said to be a Brownian motion with drift µ and
volatility σ if X(t) = µt + σW(t) where (i) W(t) is a standard Brownian motion,
(ii) −∞ < µ < ∞ is a constant and (iii) σ > 0 is a constant.
This is a generalization of standard Brownian motion. In this process, the mean
function E(X(t)) = µt and covariance function Cov (W(s), W(t)) = σ2 Min(s, t), s, t ≥
0. Fig. 7.5 shows a sample path of Brownian motion with drift µ = 0.1 and
volatility σ2 = 1.5.

2
X(t)

−1

−2
0 2 4 6 8 10
t

Fig. 7.5. Sample path of Brownian motion with µ = 0.1 σ2 = 1.5

Definition 7.4.2 (Brownian Bridge) A standard Brownian bridge {X(t), 0 ≤

t ≤ 1} is defined as X(t) = W(t) − tW(1), where W(t) is a standard Brownian
motion. Clearly, X(0) = 0 = X(1).
Since for 0 < t < 1, W(t) ∼ N(0, t), X(t) ∼ N(0, t(1 − t)). The covariance
function is s(1 − t) for 0 ≤ s ≤ t ≤ 1. Therefore, the Brownian bridge is a Gaussian
process but is not a Brownian motion. For fixed T > 0, the general Brownian
bridge {X(t), 0 ≤ t ≤ T} can be defined as
t
X(t) = W(t) − W(T), 0 ≤ t ≤ T .
T
234 Financial Mathematics: An Introduction

The covariance function is given by

Cov(X(s), X(t)) = E[X(s)X(t)] − E(X(s))E(X(t))

 
st 2 t s
= E W(t)W(s) + 2 W (T) − W(s)W(T) − W(t)W(T)
T T T
st t s
= Min{s, t} + − Min{s, T} − Min{t, T}
T T T
st
= Min{s, t} − .
T
Brownian motion model for stock prices is open to the objection that prices
are by definition positive quantities while W(t), being normally distributed, is
negative with strictly positive probability. In 1965, Samuelson [117] introduced
what has now become the standard model, namely geometric Brownian motion.
Definition 7.4.3 (Geometric Brownian Motion) A stochastic process {X(t), t ≥
0} is said to be a geometric Brownian motion (GBM) if X(t) = X(0) eW(t) where
W(t) is a standard Brownian motion.
Result 7.4.1 For any h > 0, we have

X(t + h) = X(0) eW(t+h)

= X(0) eW(t)+W(t+h)−W(t)
= X(t) eW(t+h)−W(t) .

We note that, BM has independent increments. Hence given X(t), the future X(t +
h) only depends on the future increment of the BM. Thus future is independent
of the past and therefore the Markov property is satisfied. Hence, {X(t), t ≥ 0} is a
Markov process.
Because a geometric Brownian motion is nonnegative, it provides for a more
realistic model of stock prices. Also, the GBM model considers the ratio of stock
prices to have the same normal distribution. Therefore, the percentage change in
price as opposed to the absolute change in price is modeled by a GBM. Fig. 7.6
shows a sample path of geometric Brownian motion.
How does geometric Brownian motion relate to stock prices? One possibility is
to think of modeling the rate of return of the stock price as a Brownian motion.
Suppose that the stock price S(t) at time t is given by

S(t) = S(0) eH(t) ,

 Processes Derived from Brownian Motion 235

1.6

1.4

1.2

1
X(t)

0.8

0.6

0.4

0.2
0 0.2 0.4 0.6 0.8 1
t

Fig. 7.6. Sample path of geometric Brownian motion

where S(0) is the initial price and H(t) = µt + σW(t) is a Brownian motion with
drift. In this case, H(t) represents a continuously compounded rate of return of
the stock price over the period of time [0, t]. Here, H(t) refer to the logarithmic
growth of the stock price, satisfies
 
S(t)
H(t) = ln .
S(0)
This gives
ln(S(t)) = ln(S(0)) + H(t) .
Therefore, ln(S(t)) has a normal distribution with mean µt + ln(S(0)) and variance
σ2 t. As we have seen, if a random variable X has the property that lnX has normal
distribution, then the random variable X is said to have a lognormal distribution.
Accordingly, S(t)/S(0) is lognormal distributed random variable.
Example 7.4.1 Suppose that the stock price S(t) at time t is given by S(t) =
S(0) eH(t) where S(0) is the initial price and H(t) = µt + σW(t) is a Brownian
motion with drift µ and volatility σ. Prove that
  
(i) E(S(t)) = S(0) exp µ + σ2 t .
2

   2    
(ii) Var(S(t)) = S(0) exp µ + σ2 t
2
exp σ2 t − 1 .
236 Financial Mathematics: An Introduction

Solution We know that, if a random variable X is normally distributed with mean

µ and variance σ2 , then the moment generating function of X is given by
 

θX 1 2 2
MX (θ) = E e = exp µθ + σ θ . (7.1)
2
Since for every t, W(t) is normally distributed with mean zero and variance t,
H(t) = µt + σW(t) is normally distributed with mean µt and variance σ2 t. Hence,
 

θH(t) 1 2 2
MH(t) (θ) = E e = exp µtθ + σ tθ . (7.2)
2
(i) Now
 
E(S(t)) = E S(0)eH(t)
 
= S(0)E eH(t)

Using (7.2) and substituting θ = 1, we obtain

  
σ2
E(S(t)) = S(0) exp µ + t .
2

(ii)

Var(S(t)) = E(S2 (t)) − (E(S(t)))2

   
  σ2
= E S (0)e
2 2H(t)
− S (0) exp 2 µ +
2
t
2

Using (7.2) and substituting θ = 2, we obtain

   
    σ2
Var(S(t)) = S (0) exp 2 µ + σ t − S (0) exp 2 µ +
2 2 2
t
2
   2
σ2    
= S(0) exp µ + t exp σ2 t − 1 .
2


Remark 7.4.1 Letting r = µ + 12 σ2 , we get

E(S(t)) = S(0) ert

 Processes Derived from Brownian Motion 237

2
2  
(r+ σ2 )t
eσ t − 1 .
2
Var(S(t)) = S(0) e

Here we observe that, the expected stock price depends not only on the drift µ of
H(t) but also on the volatility σ. Further, it shows that, the expected price grows
like a fixed-income security with continuously compounded interest rate r. In real
scenario, r is much lower than r, the real fixed-income interest rate, that is why
one invests in stocks. But the stock has variability due to the randomness of the
underlying Brownian motion and hence a risk is involved here.
Example 7.4.2 Suppose that stock price {S(t), t ≥ 0} follows geometric Brownian
motion with drift µ = 0.12 per year and volatility σ = 0.24 per annum. Assume
that, the current price of the stock is S(0) = Rs 40. What is the probability that
a European call option having four years to exercise time and with a strike price
K = Rs 42, will be exercised?
Solution We have
 
S(4) 42
P (S(4) > 42) = P >
40 40
  

S(4) 42
= P ln > ln .
40 40
 
S(4)
Since ln follows normal distribution with mean 0.48 and variance (0.48)2 ,
40
we get
⎛  

 ⎞
⎜⎜ S(4) 42 ⎟⎟
  
 ⎜⎜ ln − 0.48 ln − 0.48 ⎟⎟⎟
S(4) 42 ⎜
⎜ 40 ⎟⎟
= P ⎜⎜⎜⎜
40 ⎟⎟
P ln > ln > ⎟⎟
40 40 ⎜⎜ 0.48 0.48 ⎟⎟
⎜⎝ ⎠
⎛
 ⎞
⎜⎜ ln 42 − 0.48 ⎟⎟
⎜⎜ ⎟⎟
⎜ 40 ⎟⎟
= 1 − Φ ⎜⎜⎜ ⎟⎟
⎜⎜ 0.48 ⎟⎟
⎝ ⎠
= 1 − Φ(−0.8983)
= Φ(0.8983) = 0.3133,

where Φ is the cumulative standard normal distribution function

238 Financial Mathematics: An Introduction

, x
y2
1 −
Φ(x) = √ e 2 dy.
2π −∞


Definition 7.4.4 (Ornstein-Uhlenbeck Process) Let {W(t), t ≥ 0} be a
Wiener process. Define

X(t) = X(0)e−at + be−at W(e2at − 1) ,

where a and b are strictly positive real numbers and X(0) is independent of W(t).
Then, we say {X(t), t ≥ 0} is an Ornstein-Uhlenbeck process.
We note that, {X(t), t ≥ 0} is a Markov process. It is also a Gaussian process if
X(0) = x(0) is fixed or X(0) is Gaussian.
In the “real” world, we observe that asset price processes have jumps or spikes
and risk-managers have to take them into account. We need processes that can
describe the observed reality of financial markets in a more accurate way than
models based on Brownian motion. Levy processes provide us with the appropriate
framework to model both in the “real” and in the “risk-neutral” world.
Definition 7.4.5 (Levy Process) A stochastic process {X(t), t ≥ 0} is said to
be a Levy process if it satisfies the following properties
(i) X(0) = 0,
(ii) for all n and for 0 ≤ t0 < t1 < t2 < . . . < tn , increments X(ti ) − X(ti−1 ), i =
1, 2, . . . , n, are independent and stationary,
(iii) for a > 0, P(| X(t) − X(s) |> a) → 0 when t → s.
Remark 7.4.2 Let b be a constant and X(t) = bt. Then {X(t), t ≥ 0} is a Levy
process.
Remark 7.4.3 A Wiener process {W(t), t ≥ 0} defined in Definition 7.3.1, is a
Levy process in R that has continuous paths and has the Gaussian distribution
with mean zero and variance ∆t for its increments W(t + ∆t) − W(t). The most
general continuous Levy process in R has the form X(t) = bt + cW(t), t ≥ 0, where
b and c are real constants.
Remark 7.4.4 A Poisson process {N(t), t ≥ 0} with parameter λ defined in Def-
inition 7.2.19 is a Levy process that is a counting process having the Poisson
distribution with mean λ∆t for its increments N(t + ∆t) − N(t).
 Exercises 239

The above three processes namely deterministic process X(t) = bt, Wiener
process and Poisson process are Levy processes. It turns out that all Levy processes
can be built up out of these building blocks. We may refer to Applebaum [4] for
further reading in this regard.

7.5 Summary and Additional Notes

• As early as 1900, Louis Bachelier proposed Brownian motion (also, named
as Wiener process) as a model of the fluctuations of stock prices. Since the
1970s, the Wiener process has been widely applied in financial mathematics
and economics to model the evolution in time of stock prices and bond interest
rates.
• The earlist attempt to model Brownian motion mathematically can be traced to
three different sources. The first was that of T.N. Thiele in 1880 who effectively
created a model of Brownian motion while studying time series; the second was
that of L. Bachelier (mentioned above) and the third was that of Einstien, who
in 1905 proposed a model of the motion of small particles suspended in a liquid
so as to convince other physicists of the molecular nature of matter. The readers
may refer to Vassiliou [144] for further historical details.
• Section 7.2 presents the basics of stochastic process while the Brownian motion
is introduced in Section 7.3.
• In Section 7.4, processes derived from Brownian motion are presented. It in-
cludes geometric Brownian motion which has been of tremendous use in modern
finance theory.

7.6 Exercises
Exercise 7.1 Consider the binomial model for trading in stock, t = 1, 2, where at
each time the stock can go up by the factor u or down by the factor d. The sample
space Ω = {(u, u), (u, d), (d, u), (d, d)}. Create one non-trivial σ-field and the largest
σ-field on Ω.
Exercise 7.2 Construct an example of σ-fields F (∞) and F (∈) such that F (∞)∪
F (∈) is not a σ-field.
Exercise 7.3 Prove that there does not exist a σ-field which contains exactly 6
elements.
240 Financial Mathematics: An Introduction

Exercise 7.4 Consider Ω = {1, 2, 3, 4}. Let σ-algebra F = {∅, {1}, {2, 3, 4}, Ω}.
Construct a random variable on the measurable space (Ω, F ).
Exercise 7.5 Consider a random experiment of tossing an unbiased coin three
times. Let Ω denote the set of all possible outcomes. Let the random variable X be
the number of heads observed. Find the cumulative distribution function F(X, x).
Exercise 7.6 Let Xn , for n√ even take values +1 and −1 each with probability 0.5,
and for n odd, take values a, −1 √ with probability 1 , a respectively (a > 0, a 
a a+1 a+1
1). Further, let Xn ’s be independent, show that the stochastic process {Xn , n ≥ 1}
is wide sense stationary but not strict stationary.
x
Exercise 7.7 Let {X(t), t ≥ 0} be a stochastic process with independent increments
and X(0) = 0. Show that Cov (X(s), X(t)) = Var (X(Min(s, t))), for any s, t (t, s > 0) .
Exercise 7.8 Let X and Y be i.i.d. random variables each having uniform distri-
bution on the interval (−π, π). Let Z(t) = cos(tX + Y), t ≥ 0. Is {Z(t), t ≥ 0} wide
sense stationary process?
Exercise 7.9 Consider an urn containing 100 red balls and 100 black balls. Balls
are drawn one by one without replacement. Let Xn be the number of red balls
remaining in the urn after the nth ball is drawn. Is {Xn , n = 0, 1, . . .} a Markov
process?
Exercise 7.10 Let Yn = a0 Xn + a1 Xn−1 ; n = 1, 2, . . . where a0 , a1 are constants and
Xn , (n = 0, 1, . . .) are i.i.d random variables with mean 0 and variance σ2 . Is
{Yn , n = 1, 2, . . .} a Markov process?
Exercise 7.11 Let {N(t), t ≥ 0} be a Poisson process with parameter λ. Suppose
N(t) denotes the number of events that occur in the interval [0, t], prove that the
inter arrival of successive events are independent and are exponentially distributed
with parameter λ.
Exercise 7.12 Let {W(t), t ≥ 0} be the Brownian motion. Prove that, E((W(t) −
W(s))4 ) = 3(t − s)2 .
Exercise 7.13 Let {W(t), t ≥ 0} be the Brownian motion. Prove that
(W(t1 ), W(t2 ), . . . , W(tn )) is jointly normal distributed with CDF for 0 < t1 < t2 <
· · · < tn is given by
 Exercises 241

1
P(W(t1 ) ≤ a1 , W(t2 ) ≤ a2 , . . . , W(tn ) ≤ an ) = ×
(2π)n t1 (t2 − t1 ) · · · (tn − tn−1 )
, a1 , an   2 
1 x1 (x2 − x1 )2 (xn − xn−1 )2
··· exp − + + ··· + dx1 · · · dxn .
−∞ −∞ 2 t1 t2 − t1 tn − tn−1

Exercise 7.14 Let X be a normally distributed random variable with mean µ and
variance σ2 . Let u be a fixed number in R and define the convex function φ(x) = eux
for all x ∈ R. Prove that
1 2 2
(i) E(φ(X)) = euµ+ 2 u σ .
(ii) Verify the Jensen’s inequality holds E(φ(X)) ≥ φ(E(X)).
Exercise 7.15 Let {W(t), t ≥ 0} be a Wiener process. Find the conditional distri-
bution of W(t) given that W(s) = c (where c is a constant) when s < t.
Exercise 7.16 Let {W(t), t ≥ 0} be a Wiener process. Find the conditional distri-
bution of W(s/2) given that W(s) = x.
Exercise 7.17 Show that for any T > 0, V(t) = W(t + T) − W(T) is a Wiener
process if W(t) is a Wiener process.
Exercise 7.18 Let {W(t), t ≥ 0} be a Brownian motion. Prove that {tW(1/t), t ≥ 0}
where tW(1/t) is taken to be zero when t = 0, is a Brownian motion.
Exercise 7.19 Let {W(t), t ≥ 0} be a Wiener process. Prove that distribution of
{W(t), 0 ≤ t ≤ 1} and {W(1) − W(1 − t), 0 ≤ t ≤ 1} are the same.
Exercise 7.20 Consider the process


W(t), t<1
Y(t) =
W(t) + Z, t≥1

where W(t) is a Wiener process and Z ∼ N(0, 1) and is independent of W(t). Show
that {Y(t), t ≥ 0} not a Levy process.
8
Filtration and Martingale

8.1 Introduction
Conditional expectation is an extremely important concept in probability theory.
Traditionally in a typical probability course, the well known gambler’s ruin prob-
lem is presented as a motivational example. But now the concept of conditional
expectation has found much favor in financial mathematics because it provides
the basis for two of the most important concepts namely, the filtration and the
martingale.
In gambler’s ruin problem, a gambler starts with an amount of Rs N. Then an
unbiased coin is flipped, landing head with probability 0.5 and landing tail with
probability 0.5. If the coin lands head, he/she gains Rs 1, otherwise he/she loses
Rs 1. The game continues until he/she loses all his/her amount of Rs N. Let Xn
be the fortune at the nth game, and Sn be his/her capital after n games. An inter-
esting question here is to know his/her fortune, on an average, on the next game
given his/her current fortune. To answer this question, we need the conditional
expectation of the random variables {Xn , n = 0, 1, . . .} given the information up to
r, r < n.
We next present another example which is more relevant to us. Let X(t) denote
the share price of a particular stock at time t. Assume that, we have the informa-
tion about the share price up to time s > 0. We may like to know the expected or
average share price of the stock at a future time, say, s + 5 given that X(s) = 10
(say). To make such statements precise and also to answer them, we need the
conditional expectation of random variables {X(t), t ≥ 0} given the information up
to time s, s < t.
The above two examples certainly make a case for the study of conditional
expectation. But for a better understanding and deeper study of various other
244 Financial Mathematics: An Introduction

aspects we need to develop the concepts of filtration and martingale property of

the underlying stochastic process.
A filtration is an increasing sequence of σ-fields on a measurable space (Ω, F ).
Since a σ-field defines the set of events that can be measured, a filtration is often
used to represent the change in the set of events that can be measured through
gain or loss of information. In financial mathematics, a filtration represents the
information available at each time t, and it becomes more and more precise as
information from the present becomes available.
The concept of martingale in probability theory is referred to a class of betting
strategies. It says that, the conditional expected value of the next observation,
given all the past observations is equal to the last observation. It is important
to note that the property of being a martingale involves both the filtration and
the probability measure. It is possible that one stochastic process could be a
martingale with respect to one probability measure but not with respect to another
one.
In this chapter, we first discuss the conditional expectation and then proceed
to study filtrations and martingales in detail.

8.2 Conditional Expectation E(X/Y) and Calculation

Rules
Let (Ω, F , P) be the given probability space and X be the given random variable.
Then its expectation E(X) is defined as
,
E(X) = X(w)dP(w) , (8.1)
Ω

where the integration is to be understood in the Lebesgue-Stieltjes sense. In case

X is a discrete random variable taking values α1 , α2 , . . . , αi , . . . we have

E(X) = αi P(X = αi ) ,
i

where P(X = αi ) = P(w ∈ Ω : X(w) = αi ). If X is a continuous random variable

having probability density function (p.d.f.) f (x) then
, ∞
E(X) = x f (x)dx . (8.2)
−∞
 Conditional Expectation E(X/Y) and Calculation Rules 245

We say that the expectation E(X) exists provided E(X) < ∞. Obviously not all
random variables have expectation. For example, if the discrete random variable
X takes value 2n , n = 1, 2, . . . with probability mass function (p.m.f.)

P(X = 2n ) = 2−n , (n = 1, 2, . . .) ,

then E(X) does not exist. In a similar manner, if the continuous random variable
X has the p.d.f.
1
f (x) = , −∞<x<∞ ,
π(1 + x2 )
then E(X) does not exist.
Now, we define the conditional expectation and present some of its properties.
Definition 8.2.1 (Conditional Expectation of a Random Variable) For
a discrete random variable X and a discrete random variable Y, the conditional
expectation of X given the event Y = y is defined as

E(X/Y = y) = x P(X = x/Y = y) .
x

Remark 8.2.1 Looking at the expression of E(X/Y = y), it is simple to note that
it is a function of y. If we write this function of y as f (y) = E(X/Y = y), then
we write E(X/Y) for f (Y). This is a discrete random variable. It is called the
conditional expectation of X given Y and is denoted by E(X/Y).
Remark 8.2.2 Definition 8.2.1 of conditional expectation can be extended in an
obvious manner even when X is continuous random variable. We note that Y is
still a discrete random variable. Therefore if we define

E(X/Y)(w) = E(X/Ai ) for w ∈ Ai (i = 1, 2, . . .) ,

and Ai = {w ∈ Ω : Y = yi }, then it takes care of both discrete as well as continuous

cases of the random variable X. Here again E(X/Y) is a discrete random variable.
Remark 8.2.3 Definition 8.2.1 of conditional expectation can further be extended
when the event Y = y is replaced by a general event B for which P(B) > 0. Thus
when P(B) > 0, the conditional expectation of the random variable X given B is
defined as
E(X1B )
E(X/B) = ,
P(B)
where
246 Financial Mathematics: An Introduction

1, w∈B
1B (w) =
0, wB,
denotes the indicator function of the set B.
Example 8.2.1 Let X denote the outcome of tossing of an unbiased die. Let Y
be a random variable which takes the value +1 if the outcome is an even number
and it takes the value −1 if the outcome is an odd number.
(i) Find E(X/Y = 1) and E(X/Y = −1).
(ii) Let B = {1, 2, 3}. Find E(X/B).
Solution Here Ω = {1, 2, 3, 4, 5, 6} and P(X = i) = 1
6
for i = 1, 2, 3, 4, 5, 6.
(i) We have A1 = {w ∈ Ω : Y(w) = 1} = {2, 4, 6} and A2 = {w ∈ Ω : Y(w) = −1} =
{1, 3, 5}. Hence
E(X1A1 )
E(X/Y = 1) = E(X/A1 ) = .
P(A1 )
But
1 1 1
E(X1A1 ) = 2 × +4× +6× ,
6 6 6
and
1 1 1
P(A1 ) = P({2, 4, 6}) = + + .
6 6 6
Hence,
E(X1A1 )
E(X/Y = 1) = =4.
P(A1 )
Similarly,
E(X1A2 )
E(X/Y = −1) = =3.
P(A2 )
(ii) We have
E(X1B )
E(X/B) = =2.
P(B)


Properties of Conditional Expectation
Using the above definition of conditional expectation we present the following
properties for E(X/Y). We do not prove these results here and shall refer to Mikosch
[97] for the same.
 Conditional Expectation E(X/Y) and Calculation Rules 247

(i) For random variables X1 , X2 and reals a1 , a2 , we have

E ((a1 X1 + a2 X2 )/Y) = a1 E(X1 /Y) + a2 E(X2 /Y) .

This property is called the linear property of E(X/Y).

(ii) E(E(X/Y)) = E(X). This property is called the average property of E(X/Y).
(iii) For Borel measurable function g, we have E(Xg(Y)/Y) = g(Y)E(X/Y). This
property is called the out property.
(iv) E(E(X/Y, Z)/Z) = E(X/Z). This property is called the tower property.
(v) If X and Y are independent random variables then E(X/Y) = E(X).
Remark 8.2.4 Conditional expectation tells us how to use the observation of the
random variable Y to estimate another random variable X. This conditional expec-
tation is the best guess about X given Y if we want to minimize the mean square
error.
Example 8.2.2 Consider two i.i.d random variables X and Y each having uni-
form distribution between the intervals 0 and 1. Define Z = X + Y. Find
(i) E(Z/X).
(ii) E(XZ/X).
Solution Since X and Y are uniformly distributed random variables, we have
E(X) = E(Y) = 12 .
(i) We have
1
E(Z/X) = E[(X + Y)/X] = E(X/X) + E(Y/X) = X + E(Y) = X + .
2
(ii) Using property (iii) of conditional expectation, we have


1
E(XZ/X) = XE(Z/X) = X X + .
2


Example 8.2.3 Consider a 2-period binomial model. Let S(n) be the stock price
at period n. Assume that, P(up) = p and P(down) = 1 − p such that 0 < p < 1.
Find E (S(2)/S(1)).
Solution The stock price S(1) is a discrete random variable taking two values
uS(0) and dS(0) with probabilities p and 1 − p respectively. This can be formed as
partition B = {B1 , B2 } with B1 = {S(1) = uS(0)} and B2 = {S(1) = dS(0)}. Then the
forecast of S(2) given the information B1 is given by
248 Financial Mathematics: An Introduction

E(S(2)/S(1) = uS(0)) = pS(0)u2 + (1 − p)S(0)ud .

In a similar manner the forecast of S(2) given the information B2 is given by

E (S(2)/S(1) = dS(0)) = pS(0)ud + (1 − p)S(0)d2 .

Note that, this forecast differs depending on whether B1 or B2 actually occurs in

period n = 1. 

Example 8.2.4 Consider Example 7.3.1. Find the conditional expectation of
W(s) given W(t) = x for 0 < s < t.
Solution We know that the conditional distribution of W(s) given
W(t) = x for
sx 1 1
0 < s < t is a normal distribution with mean and variance s2 − . Hence,
t s t
sx
E (W(s)/W(t) = x) = .
t



8.3 More on σ-Fields

In the previous section we have defined the conditional expectation E(X/Y) where
Y is a discrete random variable. We have observed that E(X/Y) is a function of Y
which is again a discrete random variable. The definition required the construction
of sets Ai = {w ∈ Ω : Y(w) = yi } and then define

E(X/Y)(w) = E(X/Ai ) for w ∈ Ai (i = 1, 2, . . .) .

Thus E(X/Y) can be understood as a discrete random variable constructed from

a collection of subsets Ai of Ω.
In Chapter 7, we have studied σ-fields and also learnt how to generate a σ-field
from a given collection of subsets of Ω. Let us recall the following definitions.
Definition 8.3.1 (σ-Field) A σ-field on Ω is a collection F of subsets of Ω
satisfying
(i) ∅ ∈ F ,
(ii) if A ∈ F then Ac ∈ F ,
(iii) if a countable collection of sets A1 , A2 , . . . is in F , then ∪∞
i=1 Ai ∈ F .
 More on σ-Fields 249

Definition 8.3.2 (σ-Field Generated by a Collection of Subsets of Ω) Let

U be a collection of subsets of Ω. Then the smallest σ-field containing U is called
the σ-field generated by the collection U of subsets of Ω and it is denoted by σ(U).
Based on Definition 8.3.2, in this section we wish to introduce σ-field generated
by a random variable, σ-field generated by a random vector and σ-field generated
by a stochastic process. This is important because we need the σ-fields generated
by stochastic processes to define filtration and martingale.
Definition 8.3.3 (σ-Field Generated by a Random Variable) Let Y be a
discrete random variable taking values yi , (i = 1, 2, . . .). Let Ai = {w ∈ Ω : Y(w) =
yi } (i = 1, 2, . . .), and U = {A1 , A2 , . . .} be the corresponding family of subsets of
Ω. Then the σ-field generated by the random variable Y is defined as the σ-field
generated by the family U. In that case we denote σ(Y) = σ(U).
Remark 8.3.1 The σ-field σ(Y) certainly consists of all sets of form A = ∪i∈I Ai ,
where I is any subsets of N = {1, 2, . . .} including I = ∅. For I = ∅, we get A = ∅
and for I = N we get A = Ω. Also σ(Y) consists of all subsets of the form
Aa,b = {Y ∈ (a, b]} = {w ∈ Ω : a < Y(w) ≤ b}, −∞ < a < b < ∞ ,
because
Aa,b = ∪i∈I {w : Y(w) = yi } ∈ σ(Y) ,
where I = {i : a < yi ≤ b}.
Remark 8.3.2 Recall, if Ω = R and U (1) = {(a, b] : −∞ < a < b < ∞}, then
the σ-field generated by U (1) is called the Borel σ-field, and its elements are called
Borel sets. Thus the Borel σ-field is the σ-field generated by the family of semi-open
intervals of type (a, b], −∞ < a < b < ∞.
In case Ω = Rn , we can consider the family of rectangles
U (n) = {((a1 , b1 ] × (a2 , b2 ] × . . . × (an , bn ]) , −∞ < ai < bi < ∞, (i = 1, 2, . . . , n)} ,
and the σ-field generated by U (n) , i.e. σ(U (n) ). This σ-field is called the n-
dimensional Borel field. Thus n-dimensional Borel field is the σ-field generated
by the collection of rectangles in Rn .
Example 8.3.1 Consider a binomial model with t = 1, 2. Let St be the stock price
at time t. Let Ω = {(u, u), (u, d), (d, u), (d, d)}. Define a discrete random variable X
as

0.5, w ∈ {(u, u), (u, d)}
X(w) =
1.5, w ∈ {(d, u), (d, d)} .
Find σ(X).
250 Financial Mathematics: An Introduction

Solution The random variable X is a discrete random variable on F1 . We have

A1 = {w ∈ Ω : X(w) = 0.5} = {(u, u), (u, d)} ,

and
A2 = {w ∈ Ω : X(w) = 1.5} = {(d, u), (d, d)} .
Hence to determine the required σ-field we need to consider the family U =
{A1 , A2 } of subsets of Ω. Thus

U = {{(u, u), (u, d)}, {(d, u), (d, d)}} .

Then the σ-field generated by the random variable X is defined as the σ-field
generated by the family U. In this case,

σ(X) = σ(U)
= {∅, {(u, u), (u, d)}, {(d, u), (d, d)}, Ω} .


Definition 8.3.4 (σ-Field Generated by a Random Vector)
Let Y = (Y1 , . . . , Yn ) be an n-dimensional random vector. Then the σ-field gener-
ated by the random vector Y = (Y1 , Y2 , . . . , Yn ) is the smallest σ-field containing
all n-dimensional Borel sets, i.e. sets of the form

{w ∈ Ω : ai < Yi (w) ≤ bi (i = 1, 2, . . . , n), −∞ < ai < bi < ∞ (i = 1, 2, . . . , n)} .

This σ-field is denoted by σ(Y) = σ(Y1 , Y2 , . . . , Yn ).

Example 8.3.2 Consider a 2-period binomial model. Let Sn be the stock price
at period n. Let Ω = {(u, u), (u, d), (d, u), (d, d)}. Define a discrete random vector
X = (X1 , X2 ) where


0.5, w ∈ {(u, u), (u, d)}
X1 (w) =
1.5, w ∈ {(d, u), (d, d)} ,

and ⎧
⎪
⎪ 0.25, w = (u, u)
⎪
⎪ 0.5,
⎪
⎨ w = (u, d)
X2 (w) = ⎪
⎪
⎪
⎪ 0.75, w = (d, u)
⎪
⎩ 1.00, w = (d, d) .
Find the σ-field generated by the random vector X = (X1 , X2 ).
 More on σ-Fields 251

Solution We have

A1 = {w ∈ Ω : X1 (w) = 0.5} = {(u, u), (u, d)} ,

and
A2 = {w ∈ Ω : X1 (w) = 1.5} = {(d, u), (d, d)} .
Similarly, we have

B1 = {w ∈ Ω : X2 (w) = 0.25} = {(u, u)}

B2 = {w ∈ Ω : X2 (w) = 0.5} = {(u, d)}
B3 = {w ∈ Ω : X(w) = 0.75} = {(d, u)} ,

and
B4 = {w ∈ Ω : X2 (w) = 1.00} = {(d, d)} .
Hence σ-field generated by X = (X1 , X2 ) is the σ-field generated by the family U
where
U = {A1 , A2 , B1 , B2 , B3 , B4 } .
Therefore the desired σ-field is

σ(X) = σ(X1 , X2 ) = {∅, A1 , A2 , B1 , B2 , B3 , B4 , B1 ∪ B3 , B1 ∪ B4 , B2 ∪ B3 , B2 ∪ B4 ,

B1 ∪ B2 ∪ B3 , B1 ∪ B2 ∪ B4 , B2 ∪ B3 ∪ B4 , B1 ∪ B3 ∪ B4 , Ω} .


Definition 8.3.5 (σ-Field Generated by a Stochastic Process) Let {Y(t), t ∈
T} be the given stochastic process. Then the σ-field generated by the stochastic
process {Y(t), t ∈ T} is the smallest σ-field containing all sets of the form

{w : the sample path (Y(t, w), t ∈ T) belongs to C}

for all suitable sets C of functions on T.

Remark 8.3.3 The set C in Definition 8.3.5 probably needs some more specifi-
cations. But at this stage and with our limited background, it is not possible to
answer this question for a general stochastic process. However we shall present
the σ-field generated by a Brownian motion, as this is the one which we shall be
using most often.
252 Financial Mathematics: An Introduction

Definition 8.3.6 (σ-Field Generated by a Brownian Motion) Let W =

{W(s), 0 < s ≤ t} be the given Brownian motion on [0, t]. Then the σ-field generated
by all sets of the form

At1 ,t2 ,...,tn = {w ∈ Ω : (W(t1 , w), W(t2 , w), . . . , W(tn , w)) ∈ C} ,

for any n-dimensional Borel set C, and for any choice of ti ∈ [0, t], i ≥ 1 is called
the σ-field generated by the Brownian motion W.
Remark 8.3.4 For a random variable, a random vector or a stochastic process Y
on Ω, the σ-field σ(Y) generated by Y contains all the essential information about
the structure of Y as a function of w ∈ Ω. It consists of all subsets {w : Y(w) ∈ C}
for all suitable sets C. In general, C has to be any n-dimensional Borel set (n ≥ 1),
n being equal to 1 for the case of a random variable. In this situation we agree to
the terminology that Y contains information represented by σ(Y) or Y carries the
information σ(Y).

8.4 Filtration
Filtration is important in financial mathematics, because it allows to model the
flow of information. Of course, the information increases as time goes by. Before
we give the formal definition of a filtration, we consider below the given situation.
Let us consider the random experiment of tossing a coin three times. Here the
set of all possible outcomes is

Ω = {(HHH, (HHT), (HTH), (HTT), (THH), (THT), (TTH), (TTT)} .

We choose Ω as the universal or global set as it contains all information on

the process. Let EH be consists of all possible outcomes starting with head in
1st toss and ET consists of all possible outcomes starting with tail in 1st toss.
Now, we consider the family F1 = {∅, {EH }, {ET }, Ω} of subsets of Ω. Then it
can be verified that F1 is a σ-field. Note that, F1 is the set of events that
have been decided by the end of 1st toss. Similarly, let EHH consists of all
possible outcomes starting with head in 1st toss and head in 2nd toss. We
define EHT , ETH and ETT analogously. Then, we consider the following fam-
ily F2 = {∅, {EH }, {ET }, {EHH }, {EHT }, {ETH }, {ETT }, {EcHH }, {EcHT }, {EcTH }, {EcTT }, {EHH ∪
ETH }, {EHH ∪ ETT }, {EHT ∪ ETH }, {EHT ∪ ETT }, Ω} of subsets of Ω. We can again verify
that F2 is a σ-field. Observe that, this σ-field F2 contains the information learned
by observing the first two tosses of a coin.
 Filtration 253

Now, after the coin is tossed three times, every subset of Ω is resolved. Hence
the family F3 of 256 subsets of Ω which is the set of all subsets of Ω. That is, F3
is the total σ-field. Since Fi is the set of events that have been decided by the end
of ith toss, we have more information at the end of (i + 1)th toss than at the end
of ith toss.
Taking motivation from the above example we give the definition of filtration
in discrete time.
Definition 8.4.1 (Filtration in a Discrete Time) Let Ω be the set of all
possible outcomes of a random experiment and F0 = {∅, Ω}. Then a filtration in
discrete time is an increasing sequence F0 ⊂ F1 ⊂ . . . of σ-fields, one per time
instant.
The σ-field Fn may be thought of as the events of which the occurrence is de-
termined at or before time n, the “known events” at time n. The interpretation
of F0 = {∅, Ω} is that, at the beginning, one has no information. Further the
interpretation of F0 ⊂ F1 ⊂ · · · is that the state of information increases over
time.
Example 8.4.1 Consider Ω = {a, b, c, d}. Construct σ-fields Fi , (i = 0, 1, 2), such
that F0 ⊂ F1 ⊂ F2 .
Solution Obviously F0 = {∅, Ω}. Let F1 = {∅, {a, b}, {c, d}, Ω} and F2 = {∅, {a},
{b}, {c}, {d}, {a, b}, {c, d}, {a, d}, {a, c}, {b, c}, {b, d}, {a, b, c}, {c, d, a}, {d, a, b}, {b, c, d}, Ω}.
Then, F0 ⊂ F1 ⊂ F2 .


Definition 8.4.2 (Filtration in a Continuous Time) Let Ω be the set of all
possible outcomes of a random experiment. Let T be a fixed positive number and
assume that for each t ∈ [0, T], there is a σ-field Ft . Assume further that, if s ≤ t,
then every set in Fs is also in Ft . Then, the collection of σ-fields {Ft , 0 ≤ t ≤ T}
is called a filtration in continuous time.
Thus a collection of σ-fields {Ft , t ≥ 0} is called a filtration in continuous time if
Fs ⊂ Ft for all 0 ≤ s ≤ t.
Remark 8.4.1 Filtration is used to model the flow of information over time. As
an example, we think of Xt as the price of some asset at time t and Ft as the
information obtained by watching all the prices in the market up to time t.
Definition 8.4.3 (Natural Filtration) The natural filtration of a discrete time
stochastic process {X0 , X1 , . . .} is defined by the filtration {Fn , n = 0, 1, . . .} where
Fn is the σ-field generated by the random vector (X0 , X1 , . . . , Xn ).
254 Financial Mathematics: An Introduction

Thus Fn contains all events that depend on the first (n + 1) elements of the
stochastic process. It gives the “history” of the process up till time n. A convenient
notation to describe a σ-field corresponding to observing a random vector X is
σ(X). Thus σ(X), the σ-field generated by X, consists of all events that can be
expressed in X: events of the type {X ∈ C}, C being (n + 1) dimensional Borel
sets. In this notation, the natural filtration of a discrete time stochastic process
{X0 , X1 , . . .} can be written as Fn = {(X0 , X1 , . . . , Xn )}.
The natural filtration of a continuous time stochastic process {X(t), t ≥ 0} is
defined by the filtration {Ft , t ≥ 0} where Ft = σ(X(s), s ≤ t).
Definition 8.4.4 (Adapted Process) We say that a discrete time stochastic
process {X0 , X1 , . . .} is adapted to a given filtration {Fn , n = 0, 1, . . .} if the σ-field
generated by Xn is a subset of Fn means σ(Xn ) ⊂ Fn , for every n. In a similar
manner, a continuous time stochastic process {X(t), t ≥ 0} is said to be adapted to
a given filtration Ft , t ≥ 0} if σ (X(t)) ⊂ Ft for all t ≥ 0.
Thus the events connected to an adapted process up to time n are known at time
n. For instance, suppose Sn is the price of a stock at the end of nth day then the
price process {Sn , n = 0, 1, 2, . . .} is adapted to natural filtration {Fn , n = 0, 1, 2, . . .}
where Fn is the history up to the end of nth day, and F0 = {∅, Ω}, Ω = [0, ∞). In
a similar manner for continuous time stochastic process {S(t), t ≥ 0} is adapted to
{Ft , t ≥ 0} where Ft is the history up to time t.
Remark 8.4.2 The natural filtration corresponding to a process is the smallest
filtration to which it is adapted. If the process {Y0 , Y1 , . . .} is adapted to the natural
filtration of a stochastic process {X0 , X1 , . . .} then for each n the variable Yn is a
function σ(X0 , X1 , . . . , Xn ) of the sample path of the process X up till time n.
Example 8.4.2 Consider Example 8.4.1. Define
X0 (w) = {1, w ∈ {a, b, c, d}} ,


1, w ∈ {a, b}
X1 (w) =
−1, w ∈ {c, d} ,
and
⎧
⎪
⎪ 1, w=a
⎪
⎪
⎪
⎨ 2, w=b
X2 (w) = ⎪
⎪
⎪
⎪ 3, w=c
⎪
⎩ 4, w=d.
Verify that {Xi , i = 0, 1, 2} is an adapted process to Fi , (i = 0, 1, 2).
 Conditional Expectation E(X/F ) and Calculation Rules 255

Solution From Example 8.4.1, we have Fi , (i = 0, 1, 2) are the natural filtration

of the stochastic process {Xi , i = 0, 1, 2}. The σ-field generated by the random
variable X0 is defined as the σ-field generated by the family {a, b, c, d}.
σ(X0 ) = {∅, Ω} ⊆ F0 .
Similarly, the σ-field generated by the random variable X1 is defined as the σ-field
generated by the family {{a, b}, {c, d}}.
σ(X1 ) = {∅, {a, b}, {c, d}, Ω} ⊆ F1 .
Finally, the σ-field generated by the random variable X2 is defined as the σ-field
generated by the family {{a}, {b}, {c}, {d}}.
σ(X2 ) = {∅, {a}, {b}, {c}, {d}, {a, b}, {c, d}, {a, d}, {a, c}, {b, c}, {b, d},
{a, b, c}, {c, d, a}, {d, a, b}, {b, c, d}, Ω}
⊆ F2 .
Hence, {Xi , (i = 0, 1, 2)} is an adapted process to Fi (i = 0, 1, 2).


Remark 8.4.3 If a stochastic process {X(t), t ≥ 0} is adapted to the natural filtra-
tion {Ft , t ≥ 0} of the Brownian motion {W(t), t ≥ 0}, then it means that X(t) is
a function of W(s), s ≤ t. For example, X(t) = W 2 (t) − t and Y(t) = sup W(s) are
s≤t
adapted processes whereas Z(t) = sup W(s) and U(t) = W(1)−W(t) are not adapted
s≤t+1
processes to the natural filtration {Ft , t ≥ 0} of the Brownian motion {W(t), t ≥ 0}.
Remark 8.4.4 Many stochastic calculus texts specify a probability space as (Ω,
{Ft , t ≥ 0}, P) thereby, specifying the filtration explicitly from the very beginning.
Remark 8.4.5 The natural filtration is the name given to the filtration for which
Ft consists of those sets that can be decided by observing trajectories of the specified
process up to time t.

8.5 Conditional Expectation E(X/F ) and Calculation

Rules
In Section 8.2, we discussed the conditional expectation E(X/Y) while in Section
8.3, we defined the σ-field generated by the random variable Y. In this section, we
study the conditional expectation E(X/σ(Y)).
256 Financial Mathematics: An Introduction

Definition 8.5.1 (Conditional Expectation of a Random Variable given

σ-field) Let X be a random variable defined on the probability space (Ω, F , P).
Let G be a σ-field of subsets of Ω. A random variable Z is called the conditional
expectation of X given the σ-field G if
(i) Z does not contain more information than that contained in G, i.e. σ(Z) ⊂ G.
(ii) For every A ∈ G, Z satisfies E(X1A ) = E(Z1A ) where 1A is the indicator
function of the event A.
We denote Z by E(X/G) and call the same as the conditional expectation of X
given G.
Remark 8.5.1 Thus we may have Z and Z such that Z = E(X/G) and also
Z = E(X/G), but then we have Z = Z almost surely. In that sense there is
a unique G-measurable random variable Z which satisfies E(X1A ) = E(Z1A ) for
every A ∈ G.
Remark 8.5.2 The interpretation of (i) in the above definition is that, the infor-
mation in G is sufficient to determine Z. Then, we say that Z is G-measurable.
Remark 8.5.3 Obviously, every random variable is F -measurable where F is
the largest or total σ-field on Ω. Also, if X is G-measurable, then h(X) is also
G-measurable for any Borel-measurable function h. The other extreme is when a
random variable is independent of a σ-field. In this case, the information contained
in the σ-field gives no clue about the random variable.
Remark 8.5.4 When G = σ(Y), the conditional expectation E(X/Y) is a special
case of the above definition of E(X/G). We know that, every element A of σ(Y) is
of the form
A = ∪i∈I Ai = ∪i∈I {w ∈ Ω : Y(w) = yi }, I ⊂ N .
Hence, ⎛ ⎞
⎜⎜  ⎟⎟ 
E(X1A ) = E ⎜⎜⎝X 1Ai ⎟⎟⎠ = E(X1Ai ) .
i∈I i∈I
For the case of discrete random variable Z1A , its expectation is given by
 
E(Z1A ) = E(X/Ai )P(Ai ) = E(X1Ai ) .
i∈I i∈I

From Definition 8.5.1, we conclude that when G is the σ-field generated by Y, then
we write E(X/G) for the random variable E(X/Y). Thus E(X/G) is the expected
value of X given the information G. This result is valid for the continuous random
variable also.
 Conditional Expectation E(X/F ) and Calculation Rules 257

Example 8.5.1 Consider a binomial model with t = 1, 2. Let St be the stock price
at time t. Let Ω = {(u, u), (u, d), (d, u), (d, d)}. Let σ-field F be the power set of Ω
and P(w) = 14 for all w ∈ Ω. Let G = {∅, {(u, u), (u, d)}, {(d, u), (d, d)}, Ω}. Define a
discrete random variable X as
⎧
⎪
⎪ 0.25, w = (u, u)
⎪
⎪
⎪
⎨ 0.5, w = (u, d)
X(w) = ⎪ ⎪
⎪
⎪ 0.75, w = (d, u)
⎪
⎩ 1.00, w = (d, d) .
Find E(X/G).
Solution Define
3
, w ∈ {(u, u), (u, d)}
Z(w) = 8
7
8
, w ∈ {(d, u), (d, d)} .
Then we have
⎧
⎪
⎪ 3
⎪
⎪ ∅, −∞ < z <
⎪
⎪ 8
⎪
⎪
⎪
⎨ {(u, u), (u, d)}, 3 7
Z−1 {(−∞, z]} = ⎪ ≤z<
⎪
⎪ 8 8
⎪
⎪
⎪
⎪ 7
⎪ Ω,
⎪ ≤z<∞.
⎩
8
Hence, Z is G-measurable.
For A1 = {(u, u), (u, d)}, we have
1 1 3
E(X1A1 ) = 0.25 × + 0.50 × = ,
4 4 16
and for A2 = {(d, u), (d, d)}, we have
1 1 7
E(X1A2 ) = 0.75 × + 1.00 × = .
4 4 16
Now, for A1 = {(u, u), (u, d)}, we have


3 1 1 3
E(Z1A1 ) = × + = ,
8 4 4 16
and for A2 = {(d, u), (d, d)}, we have


7 1 1 7
E(Z1A2 ) = × + = .
8 4 4 16
By Definition 8.5.1, the random variable Z is the conditional expectation of X
given the σ-field G.
258 Financial Mathematics: An Introduction



Result 8.5.1 Let X and Y be two integrable random variables and a and b be two
real numbers, and G be the sub-σ-field of F . Then,
(i) E((aX + bY)/G) = aE(X/G) + bE(Y/G),
(ii) If X and G are independent, then E(X/G) = E(X). It means that we do not
gain any information about X, if we know G and vice versa.
(iii) If σ-field σ(X) ⊂ G, then E(X/G) = X. This means that the information con-
tained in G provides us with the whole information about the random variable
X. Hence, X can be treated as a constant.
(iv) If X is G-measurable, then E(XY/G) = XE(Y/G). Given G, we can deal with
X as if it is a constant, hence we can pull X(w) out of the updated expectation
and write it in front of E(Y/G).
(v) If G1 and G2 are two sub σ-fields of F with G1 ⊆ G2 , then E(E(X/G1 )/G2 ) =
E(E(X/G2 )/G1 ) = E(X/G1 ).
For the proofs of above results we shall refer to Mikosch [97].
Example 8.5.2 Suppose G = {∅, Ω}. Find E(X/G).
Solution The given σ-field {∅, Ω} is the trivial σ-field containing no information.
The only random variables which are measurable with respect to the trivial σ-field
are constants. Hence, E(X/G) = E(X) = c, where c is a constant.


√
Example
√ 8.5.3 Consider a discrete random variable X which takes value − b
and b with probabilities 13 and 23 respectively. Find E(X/σ(X2 )). Here σ(X2 ) is
the σ-field generated by the random variable X2 .
Solution We have
√ 2 √ 1
P(X = b/X2 = b) = ; and P(X = − b/X2 = b) = .
3 3
Hence, √ √ √
2 b b b
E(X/σ(X )) =2
− = .
3 3 3


 Martingales 259

8.6 Martingales
The theory of martingale plays a very important and useful role in the study of
financial mathematics. A formal definition is given below.
Definition 8.6.1 (Discrete Time Martingale) Let (Ω, F , P) be a probability
space. Let {Xn , n = 0, 1, . . .} be a stochastic process and {Fn , n = 0, 1, . . .} be the
filtration. The stochastic process {Xn , n = 0, 1, . . .} is said to be a martingale cor-
responding to the filtration {Fn , n = 0, 1, . . .} if it satisfies the following conditions
(i) For every n, E(Xn ) exists.
(ii) Each Xn is Fn -measurable.
(iii) For every n, E(Xn+1 /Fn ) = Xn .
Remark 8.6.1 The definition of martingale depends on the collection of σ-fields
Fn . For clarity, one can say that (Xn , Fn ) is a martingale.
Remark 8.6.2 From the definition of martingale and using the properties of con-
ditional expectation, we observe that if {Xn } is a martingale then E(Xn+1 ) = E(Xn )
for every n. This implies that E(Xn ) = c, a constant. Therefore if, for some n > 0,
E(Xn ) < ∞ and the increments Xn+1 − Xn of the martingale {Xn } are bounded, then
E(Xn ) = E(X0 ).
Result 8.6.1 We can generate martingale sequences by the following procedure.
Given any increasing family of σ-fields {Fn }, and any integrable random variable
X on (Ω, F , P), we take Xn = E(X/Fn ) and it is easy to check that {(Xn , Fn )} is a
martingale sequence. Of course, every finite martingale sequence is generated this
way for we can always take X to be Xn , the last one.
Result 8.6.2 In equation (8.6.1), if ‘ = is replaced by ‘ ≥ then Xn : n = 0, 1, 2, . . .
is called a submartingale while if it is replaced by ‘ ≤ then Xn : n = 0, 1, 2, . . . is
called a supermartingale. Obviously, {Xn } is a supermartingale if and only if {−Xn }
is a submartingale.
Result 8.6.3 Let {Xn , n = 0, 1, . . .}, {Yn , n = 0, 1, . . .} be discrete time stochastic
processes. We say {Xn } is a martingale with respect to {Yn } if E(|Xn |) < ∞ and
E(Xn+1 /Y0 , Y1 , . . . , Yn ) = Xn . We may think of Y0 , Y1 , . . . , Yn as the information or
history upto stage (n+1). It may include more information than just X0 , X1 , . . . , Xn .
By calculating the conditional expectation value of Xn+1 given the information
about Xn (or Yn ) upto time n, we are making a forecast for the random variable.
The martingale relation implies that the “best” forecast for the next value of the
random variable is its current value.
260 Financial Mathematics: An Introduction

Result 8.6.4 Martingale theory provides a classification scheme for the time se-
ries. If a time series exhibits no discernible trend then it has a martingale like
behavior. On the other hand, if the trend is an increasing (decreasing) one, then
the time series behaves like submartingale (supermartingale).
Example 8.6.1 Consider the gambler’s ruin problem. A gambler starts with Rs
N. Then a coin is flipped, landing head with probability p and landing tail with
probability 1−p. If the coin lands heads, he/she gains one rupee, otherwise he /she
loses a rupee. The game continues until he/she loses all his/her N rupees. Let Xn
be fortune at the nth game, and Sn be his/her capital after n games. What will
be his/her fortune, on an average, on the next game given that his/her current
fortune?
Solution For this we suppose that Sn is adapted to the filtration Fn . Then we can
interpret a martingale as a fair game. Thinking (Sn − Sr ) as the net winnings of the
game per unit stake in time frame (r, n], the best prediction of the net winnings
given the information at the time r < n has value
E ((Sn − Sr )/Fr ) = E(Sn /Fr ) − Sr .
If Sn is a martingale then E(Sn /Fr )−Sr = 0. This means that the best prediction of
the future net winnings per unit stake in the interval (r, n] is zero. This is exactly
what we expect to be a fair game. It says that gambler expected capital after one
more game played with the knowledge of the entire past and present is exactly
equal to his/her current capital.


Example 8.6.2 Let X1 , X2 , . . . be a sequence of i.i.d random variables each taking
two values +1 and -1 with equal probabilities. Let us define S0 = 0 and Sn =
n
X j , (n = 1, 2, . . .). This discrete time stochastic process {Sn , n = 0, 1, . . .} is a
j=1
symmetric random walk. Prove that, {Sn , n = 0, 1, . . .} is a martingale with respect
to {Xn , n = 1, 2, . . .}.
Solution We have E(|Sn |) ≤ E(|X1 |) + E(|X2 |) + · · · + E(|Xn |) < ∞ . Also
E(Sn+1 /X1 , X2 , . . . , Xn ) = E ((Sn + Xn+1 )/X1 , X2 , . . . , Xn )
= E(Sn /X1 , X2 , . . . , Xn ) + E(Xn+1 /X1 , X2 , . . . , Xn )
= Sn + E(Xn+1 ) (using independent of X1 , X2 , . . . , Xn , Xn+1 )
= Sn + 0
= Sn .
 Martingales 261

Hence {Sn , n = 0, 1, . . .} is a martingale with respect to {Xn , n = 1, 2, . . .}. Suppose

Fk is the σ-field of information corresponding to the first k random variables Xk ,
we have for non-negative integers k < n, E(Sn /Fk ) = Sk .


Example 8.6.3 Consider a symmetric random walk {Sn , n = 0, 1, . . .} which is a
martingale with respect to the filtration {Fn , n = 0, 1, . . .} where F0 = {∅, Ω} and
Fn = σ(X1 , X2 , . . . , Xn ), (n ≥ 1), is the σ-field of information corresponding to the
first n random variables Xn . Verify if {S2n , n = 0, 1, . . .} is a martingale with respect
to the filtration {Fn , n = 0, 1, . . .}.
Solution For each n = 1, 2, . . ., S2n is Fn -measurable. Also

n
E(S2n ) = E(Xi2 ) < ∞ .
i=1

Now,
 
E(S2n+1 /Fn ) = E (Sn+1 − Sn + Sn )2 /Fn
 
= E (Sn+1 − Sn )2 /Fn − 2E [Sn (Sn+1 − Sn )/Fn ] + E(S2n /Fn )

= E(Xn+1
2
/Fn ) − 2E(Xn+1 Ṡn /Fn ) + E(S2n /Fn ).

Since Xn+1 is independent of Fn and since S2n is Fn -measurable, we have

E(S2n+1 /Fn ) = E(Xn+1

2
) − 2Sn E(Xn+1 ) + S2n
= 1 − 0 + S2n
= 1 + S2n .

Hence, E(S2n+1 /Fn )  S2n . Therefore, {S2n , n = 0, 1, . . .} is not a martingale. Since

E(S2n+1 /Fn ) ≥ S2n , (n = 1, 2, . . .), hence {S2n , n = 0, 1, . . .} is a submartingale.


Example 8.6.4 Let Y0 = 0 and let Y1 , Y2 , . . . be a sequence* of i.i.d random vari-
2
ables with E(|Yk |) = 0 and E(Yk ) = σ . Let X0 = 0 and Xn = nk=1 Yk − nσ2 . Show
2 2

that {Xn , n = 0, 1, . . .} is a martingale with respect to {Yn , n = 0, 1, . . .} .

Solution We have
262 Financial Mathematics: An Introduction

⎛ n+1 ⎞2
⎜⎜ ⎟⎟
Xn+1 = ⎜⎜⎜⎝ Yk ⎟⎟⎟⎠ − (n + 1)σ2
k=1
⎛⎛ n ⎞ ⎞2
⎜⎜⎜⎜ ⎟⎟ ⎟⎟
= ⎜⎜⎜⎝⎜⎜⎜⎝ Yk ⎟⎟⎟⎠ + Yn+1 ⎟⎟⎟⎠ − nσ2 − σ2
k=1
⎛ n ⎞
⎜⎜ ⎟⎟
= Xn + 2Yn+1 ⎜⎜⎜⎝ Yk ⎟⎟⎟⎠ + Yn+1
2
− σ2 .
k=1

⎡ ⎛ n ⎞ ⎤
⎢⎢ ⎜⎜ ⎟⎟ ⎥⎥
E [Xn+1 /Y0 , Y1 , . . . , Yn ] = E ⎢⎢⎢⎣(Xn + 2Yn+1 ⎜⎜⎜⎝ Yk ⎟⎟⎟⎠ + Yn+1
2
− σ2 )/(Y0 , Y1 , . . . , Yn )⎥⎥⎥⎦
k=1
= E [Xn /(Y0 , Y1 , . . . , Yn )]
⎡ ⎤
⎢⎢⎢ n
⎥⎥
+2E ⎢⎢⎣Yn+1 Yk /(Y0 , Y1 , . . . , Yn )⎥⎥⎥⎦
k=1
 
+E Yn+1 /(Y0 , Y1 , . . . , Yn ) − σ2
2


n
= Xn + 2 Yk E(Yn+1 ) + E(Yn+1 2
) − σ2
k=1
= Xn .

Thus {Xn , n = 0, 1, . . .} is a martingale with respect to {Yn , n = 0, 1, . . .} .



Example 8.6.5 Let a person start with Rs 1. A fair coin is tossed infinitely many
times. For nth toss, if it turns up ‘head’, the person gets Rs 2, but if turn up ‘tail’,
the person does not get any amount. Let Yn be his/her fortune at the end of nth
toss. Prove that Yn is a martingale.
Solution Let X1 , X2 , . . . be a sequence of i.i.d random variables each defined by


2, with probability 0.5
Xi =
0, with probability 0.5 .

Since the game is double or nothing, his/her fortune at the end of nth toss is given
by
Yn = X1 X2 · · · Xn (n = 1, 2, . . .) .
 Martingales 263

Let Fn be the σ-field generated by X1 , X2 , . . . , Xn . We note that 0 ≤ Yn ≤ 2n and

E[Xn+1 ] = 1. Now,

E [Yn+1 /Fn ] = E [Yn Xn+1 /Fn ]

= Yn E [Xn+1 /Fn ]
= Yn E[Xn+1 ]
= Yn .

Hence, {Yn , n = 1, 2, . . .} is a martingale.



Example 8.6.6 Consider a binomial lattice model. Let Sn be the stock price at
period n and


uSn , with probability p
Sn+1 =
dSn , with probability 1 − p .

Define a related process Rn as

 
Rn = ln(Sn ) − n p ln(u) + (1 − p) ln(d) .

Prove that {ln(Sn ), n = 1, 2, . . .} is not a martingale whereas {Rn , n = 1, 2, . . .} is a

martingale with respect to {Sn , n = 1, 2, . . .}. Also, prove that the discounted stock
process {S0 , e−r S1 , e−2r S2 , . . .} is a martingale only if
er − d
p= ,
u−d
where r is the nominal interest rate.
Solution In this binomial lattice model {S0 , S1 , . . .} with the natural filtration
{F0 , F1 , . . .}, we have

P(Sn+1 = uSn /Fn ) = 1 − P(Sn+1 = dSn /Fn ) = p .

Hence,
E(Sn+1 /Fn ) = p uSn + (1 − p) dSn = Sn [p u + (1 − p) d] .
We consider the variable ln(Sn ) and observe that


 
Sn
E ln /Sn−1 , Sn−2 , . . . , S0 = p ln(u) + (1 − p) ln(d) .
Sn−1
264 Financial Mathematics: An Introduction

Therefore,

E (ln(Sn )/Sn−1 , Sn−2 , . . . , S0 ) = ln(Sn−1 ) + p ln(u) + (1 − p) ln(d) . (8.3)

Here {ln(Sn ), n = 1, 2, . . .} is not a martingale and depending upon the values of p, u

and d it may be either a submartingale or a supermartingale. Next, we consider
the process Rn .
   
E(Rn /Rn−1 , Rn−2 , . . . , R0 ) = E ln(Sn ) − n p ln(u) + (1 − p) ln(d) /Rn−1 , Rn−2 , . . . , R0

Using equation (8.3), and noting that the history of Sn−1 , Sn−2 , . . . , S0 yields the
history of Rn−1 , Rn−2 , . . . , R0 and vice-versa, we get
 
E(Rn /Rn−1 , Rn−2 , . . . , R0 ) = ln(Sn−1 ) − (n − 1) p ln(u) + (1 − p) ln(d)
= Rn−1 .

Therefore {Rn , n = 1, 2, . . .} is martingale.

Now, consider the discounted process {S0 , e−r S1 , e−2r S2 , . . .} where r is the interest
rate. We have

E(e−(n+1)r Sn+1 /Fn ) = p ue−(n+1)r Sn + (1 − p) de−(n+1)r Sn .

The discounted process is a martingale only if the right hand side of the above
equation is equal to e−nr Sn . That is,

e−nr Sn = p ue−(n+1)r Sn + (1 − p) de−(n+1)r Sn

or
er = p u + (1 − p) d .
Thus, the discounted process is a martingale only if

er − d
p= .
u−d


Remark 8.6.3 It may be noted that in Chapter 3 the RNPM for the binomial
lattice model has been derived via replicating portfolio arguments. But the approach
presented here in Example 8.6.6 is based on the fact that discounted stock process
under RNPM form a martingale. This approach is more general and is applicable
in a large variety of derivative pricing problems.
 Martingales 265

Now, one of the important results of martingale theory is Doob’s decompo-

sition, which connects the martingale and submartingale. This result is stated
below. We shall refer to Roman [113] for its proof.
Lemma 8.6.1 (Doob’s Decomposition) Let {Xn , n = 0, 1, . . .} be a submartin-
gale with respect to the filtration {Fn , n = 0, 1, . . .}. Then there exists a martingale
M = {Mn , n = 0, 1, . . .} with respect to the filtration {Fn , n = 0, 1, . . .} and a stochas-
tic process A = {An , n = 0, 1, . . .} such that:
(i) Xn = Mn + An , (n = 0, 1, . . .) .
(ii) A is an increasing stochastic process, i.e. An ≤ An+1 , (n = 0, 1, . . .) .
(iii) An is Fn−1 - measurable, for all n .
Remark 8.6.4 In Example 8.6.6, ln(Sn ) is not a martingale whereas Rn is a mar-
tingale. This is simply an illustration (and not a consequence) of Lemma 8.6.1.
Example 8.6.7 Prove that, under risk neutral probability measure Q the dis-
counted stock price process {(1 + r)−k Sk , k = 1, 2, . . .} is a martingale.
Solution This is a simple interest version of the second part of Example 8.6.1. We
are proving it separately even at the cost of repetition. Let Ω be all sequences of
length n of heads H and tails T. Let S0 be a fixed number. Define Sk (w) = u j dk− j S0 if
the first k elements of a given w ∈ Ω has j occurrences of H and k − j occurrences
of T. It means that if H occurs then the stock price goes up by a factor u; if
T occurs then it goes down by a factor d. Let Fk be the σ-field generated by
(1 + r) − d u − (1 + r)
S0 , S1 , . . . , Sk . Let p = , q = and define Q(w) = (p) j (q)n− j
u−d u−d
if w has j appearances of H and n − j appearances of T. We note that under
probability measure Q, the random variables Sk+1 /Sk = Yk (say) are independent
and identically distributed with probability mass function

P(Y = u) = p, P(Y = d) = q .

Since Yk i.i.d random variables, they satisfy

P(Y1 = y1 , , . . . , Yn = yn ) = P(Y1 = y1 ) . . . P(Yn = yn ) = (p) j (q)n− j .

Also we may note that Yk is the factor by which the stock price goes up or down
at time k. Since the random variable Yk is independent of Fk , we have
 
EQ (1 + r)−(k+1) Sk+1 /Fk = (1 + r)−k Sk (1 + r)−1 EQ [(Sk+1 /Sk ) /Fk ] .
266 Financial Mathematics: An Introduction

But
EQ [(Sk+1 /Sk ) /Fk ] = EQ [(Sk+1 /Sk )] = p u + q d = (1 + r) .
Therefore on substitution, we conclude
 
EQ (1 + r)−(k+1) Sk+1 /Fk = (1 + r)−k Sk .
This proves that the process {(1 + r)−k Sk , k = 1, 2, . . .} is a martingale.


Wealth Process
Let ∆k be the number of shares of a stock held between time k and k + 1. We
assume that ∆k is Fk -measurable and X0 is the amount of money we have started
with time t = 0. If we have ∆k shares between time k and k + 1, then at time
k + 1 those shares will be worth ∆k Sk+1 , where Sk+1 is the share price at time k + 1.
The amount of cash we hold between time k and k + 1 is Xk minus the amount
held in stock, that is Xk − ∆k Sk . Hence, the worth of this amount at time k + 1 is
(1 + r)[Xk − ∆k Sk ]. Therefore, the amount of money we have at time k + 1 is
Xk+1 = ∆k Sk+1 + (1 + r) [Xk − ∆k Sk ] .
When r = 0, this reduces to
Xk+1 − Xk = ∆k (Sk+1 − Sk ) .
Thus,

k
Xk+1 = X0 + ∆i (Si+1 − Si ) .
i=0
The stochastic process {Xk , k = 0, 1, . . .} is called the wealth process.
We shall now show that under risk neutral probability measure Q the dis-
counted wealth process is a martingale.
EQ [Xk+1 − Xk /Fk ] = EQ [∆k (Sk+1 − Sk )/Fk ]
= ∆k EQ [(Sk+1 − Sk )/Fk ] (∆k is Fk − measurable)
= 0 (Sk is a martingale) .
Now writing Xk+1 as Xk + ∆k (Sk+1 − Sk ) and noting that r > 0, we have
     
EQ (1 + r)−(k+1) Xk+1 − (1 + r)−k Xk )/Fk = EQ ∆k (1 + r)−(k+1) Sk+1 − (1 + r)−k Sk /Fk
 
= ∆k EQ (1 + r)−(k+1) Sk+1 − (1 + r)−k Sk /Fk
= 0 (Sk is a martingale under Q) .
Hence, discounted wealth process {(1 + r)−k Xk , k = 1, 2, . . .} is a martingale.
 Martingales 267

Definition 8.6.2 (Continuous Time Martingale) Let (Ω, F , P) be a proba-

bility space. Let {X(t), t ≥ 0} be a stochastic process and {Ft , t ≥ 0} be a filtration.
The stochastic process {X(t), t ≥ 0} is said to be a martingale corresponding to the
filtration {Ft , t ≥ 0} if it satisfies the following conditions
(i) For every t, E(X(t)) exists.
(ii) Each X(t) is Ft -measurable.
(iii) For every 0 < s < t,
E(X(t)/Fs ) = X(s) . (8.4)
Example 8.6.8 Prove that {W(t), t ≥ 0} is a martingale.
Solution For 0 < s < t,

E(W(t)/Fs ) = E(W(t) − W(s) + W(s)/Fs )

= E(W(t) − W(s)/Fs ) + E(W(s)/Fs )
= 0 + W(s) (from the property of Brownian motion) .

Therefore, {W(t), t ≥ 0} is a martingale.



Example 8.6.9 Prove that the Poisson process {N(t), t ≥ 0} is not a martingale.
Solution First E(| N(t) |) ≤ λt < ∞ . For 0 < s < t,

E(N(t)/Fs ) = E ((N(t) − N(s) + N(s))/Fs )

= E ((N(t) − N(s))/Fs ) + E(N(s)/Fs )
= λ(t − s) + N(s) (from the property of Poisson process) .

Hence, E(N(t)/Fs )  N(s), {N(t), t ≥ 0} is not a martingale. Note that, {N(t), t ≥ 0}

is a submartingale since E(N(t)/Fs ) ≥ N(s), for 0 < s < t.




t
Example 8.6.10 Show that exp W(t) − is a martingale.
2
Solution Let 0 ≤ s < t. Since W(t) − W(s) is independent of Fs and W(s) is
Fs -measurable, we have
   
E eW(t) /Fs = E eW(t)−W(s) eW(s) /Fs
 
= eW(s) E eW(t)−W(s) /Fs
 
= eW(s) E eW(t)−W(s) .
268 Financial Mathematics: An Introduction

Since W(t) − W(s) has normal distribution with mean zero and variance (t − s), we
have
t−s
 
E e W(t)−W(s)
=e 2 .
Hence,
t−s
 
E eW(t) /Fs = eW(s) e 2 .
This gives, for 0 ≤ s < t,
⎛ ⎞ −t
⎜⎜ W(t)− t ⎟⎟  
s
⎜⎜ ⎟
2 /Fs ⎟⎟⎟ = e 2 E eW(t) /Fs = e
W(s)−
2.
E ⎜⎜e
⎝ ⎠
 
It follows that exp W(t) − 2t is a martingale.


The cornerstone of martingale theory is Doob’s optional sampling theorem.
This states, roughly, that “stopping” a martingale at a random time τ does not
alter the expected “payoff”, provided the decision about when to stop is based
solely on information available up to τ. Such random times are called stopping
times. Stopping times are also called as Markov times or optional times.
If τ is optional and c > 0 is a positive constant, then τ + c is a stopping time.
Before stating Doob’s optional sampling theorem we give below the definitions.
Definition 8.6.3 (Stopping Time) A stopping time relative to a filtration
{Fn , n ≥ 0} is a non-negative integer-valued random variable τ such that for each
n the event τ = n ∈ Fn . In continuous case, a random variable γ : Ω → [a, b] is
called a stopping time with respect to a filtration {Ft , a ≤ t ≤ b} if {w, γ(w) ≤ t} ∈ Ft
for all t ∈ [a, b].
From the definition, we can think of γ as the time to stop playing a game. The
decision to stop playing the game before or at time t should be determined by the
information provided by Ft which will give the condition for γ to a stopping time.
Definition 8.6.4 (Hitting Time) Let {X(t), t ≥ 0} be a stochastic process. Let
A be Borel set in Rn . Define

τ = in f {t > 0 : X(t) ∈ A} .

Then τ is called a hitting time of A for the stochastic process {X(t), t ≥ 0}.
 Martingales 269

Example 8.6.11 Consider a symmetric random walk {Sn , n = 0, 1, . . .}. Let T be

the first time that Sn = 2. Prove that T is a stopping time.
Solution Since T is the first time that Sn = 2, we have


Min{n ≥ 0 : Sn = 2}, if Sn = 2 for some n ∈ N
T=
∞, otherwise

In other words,
T = in f {n ≥ 0 : Sn = 2} .
Further
4
n
{T ≤ n} = {Si = 2} ∈ Fn
i=0

where {Si = 2} ∈ Fi ⊆ Fn for i ≤ n because {Fn , n = 0, 1, . . .} is a filtration. Hence

it follows that T is a stopping time.


Result 8.6.5 A stopping time T is said to be bounded if there exists a constant
K < ∞ such that P(T ≤ K) = 1, and is said to be finite a.s. if P(T < ∞) = 1. If T is
a bounded stopping time and {X(t), t ≥ 0} is a martingale, then E(X(T)) = E(X(0)).
For instance, in the above example , T is a finite a.s. but not a bounded stopping
time. Hence E(S(T)) = 2  0 = E(S(0)) .
Example 8.6.12 Let τ be a stopping time. Prove that W(t + τ) − W(τ) is a Brow-
nian motion.
Solution: By observing the Brownian motion, we can determine whether or not
τ ≤ t. For a fixed a, the hitting time of a is defined by

τ = in f {t > 0 : W(t) = a} .

Since the sample paths of Brownian motion are continuous, it is easy to see that
τ is a stopping time. That is,

{τ ≤ t} = {W(s) = a for some s, 0 ≤ s ≤ t} ,

which depends only on {W(s), 0 ≤ s ≤ t}. Also, note that, if τ < ∞, then W(τ) = a
by continuity of sample paths. Since τ is a stopping time, W(t + τ) − W(τ) is a
Brownian motion.


270 Financial Mathematics: An Introduction

The following theorem is an important result in martingale theory. We shall

refer to Roman [113] for the proof of the above theorem.
Theorem 8.6.1 (Doob’s Optional Sampling Theorem) Let {Xn , n = 0, 1, . . .}
be a martingale with respect to {Yn , n = 0, 1, . . .} and let τ be a stopping time. Then
for any n = 1, 2, . . .

E(Xτ 9 n /Y0 , Y1 , . . . Yn−1 ) = Xτ 9 n−1

9
where τ n = Min(τ, n). i.e. the stopped process is also a martingale.
We shall refer to Roman [113] for the proof of the above Theorem. This theorem
is also known as optional stopping theorem.
Remark 8.6.5 In case {Xn , n = 0, 1, . . .} is a submartingale with respect to
{Yn , n = 0, 1, . . .} then

E(Xτ 9 n /Y0 , Y1 , . . . Yn−1 ) ≥ Xτ 9 n−1 .

Remark 8.6.6 In case {Xn , n = 0, 1, . . .} is a supermartingale with respect to

{Yn , n = 0, 1, . . .} then

E(Xτ 9 n /Y0 , Y1 , . . . Yn−1 ) ≤ Xτ 9 n−1 .

Example 8.6.13 Suppose that two players A and B start respectively with Rs a
and Rs b, and they bet against each other one rupee at a time by tossing a fair
coin. What is the probability that player B will be broke and A ends up with all
the money?
Solution Let Xn be player A fortune at the nth game, and Sn be his/her capital
after n games. We have
n
Sn = a + Xn ,
i=1

where Xi are i.i.d random variables with

1
P(Xn = ±1) = .
2
 Summary and Additional Notes 271

Let
τ0 = in f {n : Sn = 0}
τa+b = in f {n : Sn = a + b} ,
9
and T = τ0 τa+b . Therefore, T is the first time that player A either makes b
extra rupees or goes ruin i.e.
T = in f {n : Sn = 0 or Sn = b + a} .
Hence, T is the stopping time with respect to the martingale {Sn , n = 0, 1, . . .}. By
using optional stopping theorem, we have
E(ST ) = S0 = a .
But
E(ST ) = 0 × P(T = τ0 ) + (a + b) × P(T = τa+b ) .
Hence,
a
P (T = τa+b ) = .
a+b
This is exactly the probability that player B goes broke and player A wins all.



8.7 Summary and Additional Notes

• Conditional expectation plays a key role in derivative pricing models. In Section
8.2, the definition of conditional expectation and its properties are presented.
• In Section 8.3, σ-fields generated by a random variable, a random vector and a
stochastic process are defined. Finally, σ-field generated by Brownian motion
is presented.
• Filtration is presented in Section 8.4 with simple examples. Both discrete and
continuous time filtration are explained with some examples.
• We provide a detailed evaluation procedure of conditional expectation with
respect to the filtration. Further, their properties are presented along with
some examples.
• We know that a conditional expectation is itself a random variable. The com-
plicated conditional expectations such as conditional expectation of a random
variable with respect to the filtration arise in martingale theory and the concept
of martingale theory is used in a betting strategy. Kolmogorov (1932) invented
the martingale theory whereas its first systematic treatment was provided by
Doob (1942).
272 Financial Mathematics: An Introduction

• Two important results in martingale theory namely Doob’s decomposition and

Doob’s optional sampling theorem are presented. Optional sampling concept
reveals that the expectation of a martingale remains constant. It has immediate
implications concerning the pathwise behaviour of martingales, submartingales
and supermartingales. We may refer to Roman [113] for some more details on
martingale theory.

8.8 Exercises
Exercise 8.1 Let {Xn , n ≥ 1} be a sequence of independent and identically dis-
tributed random variables each having uniform distribution with values -2, -1, 0,
1, 2, 3. Let {N(t),
* t ≥ 0} be a Poisson process with parameter 2. Find the mean
and variance of N(t)i=1
Xi .
Exercise 8.2 Consider two i.i.d random variables X and Y each having uniform
distribution between the intervals 0 and 1. Define Z = X + Y. Prove that E(X/
Z
Z) = .
2
Exercise 8.3 Consider the successive rolling of an unbiased die. Let X and Y
denote the number of rolls necessary to obtain a two and a three respectively.
Obtain (i) E(X/Y = 2), (ii) E(X/Y = 4).
Exercise 8.4 Let (Ω, F , P) be a probability space and let X be an integrable ran-
dom variable and let G ⊂ F be a σ-field. Prove that, the conditional expectation
E(X/G) exists. (Hint: Use Radon-Nikodym theorem)
Exercise 8.5 Consider Ω = {a, b, c, d}. Construct 4 distinct σ-fields F1 , F3 , F3 , F4
such that F1 ⊂ F2 ⊂ F3 ⊂ F4 .
Exercise 8.6 Construct an example of σ-fields F1 and F2 such that F1  F2 and
F2 ⊂ F 1 .
Exercise 8.7 Consider a binomial model with t = 1, 2, 3. Let St be the stock price
at time t. Let Ω = {(u, u, u), (u, u, d), (u, d, u), (u, d, d), (d, u, d), (d, u, u), (d, d, u), (d, d, d)}.
1
Let σ-field F be the power set of Ω and P(w) = for all w ∈ Ω. Let
8
G = {∅, {(u, u, u), (u, u, d), (u, d, u), (u, d, d)}, {(d, u, d), (d, u, u), (d, d, u), (d, d, d)}Ω}.
Define a discrete random variable X as
X(wi ) = i, i = 1, 2, . . . , 8 .
Find E(X/G).
 Exercises 273

Exercise 8.8 Let {Xn , n = 0, 1, . . .} and {Yn , n = 0, 1, . . .} be stochastic processes.

We say {Xn } is a martingale with respect to {Yn } if
(i) E(| Xn |) < ∞.
(ii) E(Xn+1 /Y0 , Y1 , . . . , Yn ) = Xn .
Prove that {Xn , n = 0, 1, . . .} is a martingale with respect to {Yn , n = 0, 1, . . .} where
Xn = Y1 + Y2 + . . . + Yn , n ≥ 1, Y0 = 0, {Yi , i = 1, 2, . . .} are independent random
variables with E(Yn ) = 0.
Exercise 8.9 Let {Sn , n = 0, 1, . . .} be a symmetric random walk and Fn be a
filtration. Show that Yn = (−1)n cos(πSn ) is a martingale with respect to Fn .
Exercise 8.10 Consider the tossing of an unbiased coin three times. Suppose that
the toss of a head wins Rs 1 while an outcome of a tail loses Rs 1. Let Xn denote
the sum of the winnings at time n. Prove that {Xn , n = 1, 2, 3} is an adapted
process with respect to {Fi , i = 1, 2, 3} defined in Section 7.4. Further prove that
{Xn , n = 1, 2, 3} is a martingale with respect to {Fi , i = 1, 2, 3}.
Exercise 8.11 {N(t), t ≥ 0} be a Poisson process with parameter λ. Prove that,
{N(t) − λt, t ≥ 0} is a martingale.
Exercise 8.12 Let X(t) = µt + σW(t), − ∞ < µ < ∞, 0 < σ < ∞. Prove
that {X(t), t ≥ 0} is a martingale for µ = 0. Also prove that {X(t), t ≥ 0} is a
submartingale for µ > 0, and a supermartingale for µ < 0.
Exercise 8.13 Show that the maximum of two submartingales (with respect to
the same filtration) is a submartingale.
Exercise 8.14 Let Y1 , Y2 , . . . be a sequence of positive independent
7n random vari-
ables with E(Yi ) = 1 for all i. Define X0 = 1, and Xn = i=1 Yi for n ≥ 1. Show
that {Xn , n = 0, 1, . . .} is a martingale with respect to its natural filtration.
Exercise 8.15 Let Y1 , Y2 , . . . be a sequence of i.i.d 7
random variables with P(Yi =
n Yi
1) = p and P(Yi = 0) = 1 − p. Define Xn = i=1 p . Prove that, E(Xn+1 /
Y1 , Y2 , . . . , Yn ) = Xn .
Exercise 8.16 Let X be a random variable and let φ be a convex function on
R. Suppose both X and φ(X) are integrable, and G ⊂ F is a σ-field. Prove that,
φ(E(X/G)) ≤ E(φ(X)/G).
274 Financial Mathematics: An Introduction

Exercise 8.17 Let X1 , X2 , . . . be a sequence of square integrable random variables.

Show that if {Xn , n = 1, 2, . . .} is a martingale with respect to a filtration Fn , then
Xn2 is a sub martingale with respect to the same filtration.
(Hint: Use Jensen’s inequality with convex function φ(x) = x2 )
Exercise 8.18 Prove that, if {Xn , n = 1, 2, . . .} is a martingale, then | Xn |, Xn2 ,
eXn and e−Xn are all submartingales.
√
Exercise 8.19 Prove that, if Xn > 0 and is a martingale, then Xn and ln(Xn )
are supermartingales.
Exercise 8.20 Prove that, if Xn is a submartingale and K a constant, then
Max {Xn , K} is a submartingale, while if Xn is a supermartingale, so is Min {Xn , K}.
Exercise 8.21 Let X1 , X2 , . . . be i.i.d random variables each having normal dis-
tribution with
*n zero mean and unit variance. Show that the sequence
Yn = exp i=1 Xi − 2 n forms a martingale.
1

Exercise 8.22 Prove that {W 2 (t) − t, t ≥ 0} is a martingale, where {W(t), t ≥ 0} is

a Brownian motion.
σ2
Exercise 8.23 Let {W(t), t ≥ 0} be a Wiener process. Is exp(σW(t) − t) a mar-
2
tingale where σ is a positive constant?
Exercise 8.24 Let {N(t), t ≥ 0} be a Poisson process with parameter 1. Which of
the following are martingales.
(i) {N(t) − t, t ≥ 0}.
(ii) {N(t)2 − t, t ≥ 0}.
(iii) {(N(t) − t)2 − t, t ≥ 0}.
Exercise 8.25 Let X be a integrable random variable and G1 and G2 be two sub-σ-
fields of F . Prove that, if G1 ⊆ G2 , then E(E(X/G1 )/G2 ) = E(E(X/G2 )/G1 ) = E(X/
G1 ).
Exercise 8.26 Let {Fn , n = 1, 2, . . .} be a filtration. Let X be any random variable
with E(| X |) < ∞. Define Xn = E(X/Fn ), n = 1, 2, . . . Prove that, {Xn , (n =
1, 2, . . .)} is a martingale with respect to the filtration {Fn , n = 1, 2, . . .}.
Exercise 8.27 Let {Yn , (n = 0, 1, . . .)} be an arbitrary sequence of random vari-
ables. Suppose X is a random variable with E(| X |) < ∞. Define

Xn = E(X/Y0 , Y1 , . . . , Yn ), (n = 0, 1, . . .) .
 Exercises 275

Show that {Xn , n = 0, 1, . . .} forms a martingale with respect to {Yn , n = 0, 1, . . .}.

We say that, {Xn , n = 0, 1, . . .} is a Doob’s martingale.
Exercise 8.28 Let {W(t), t ≥ 0} be a Weiner process and let τ be a stopping time.
Define

W(t), t≤τ
W̃(t) =
2W(τ) − W(t), t>τ.
Prove that {W̃(t), t ≥ 0} is a Weiner process.
Exercise 8.29 Let τ is a stopping time for the Weiner process. Using the optimal
sampling theorem, show that E(W(τ)) = 0. Use this to deduce that P(W( τ) = a) =
b a
and P(W(τ) = b) = .
a+b a+b
Exercise 8.30 For a ∈ R, let τa be the first time that {W(t), t ≥ 0}
9
hits a. Suppose
τa τb
that a > 0 > b. By considering the stopped martingale {(W(t)) , t ≥ 0}, prove
−b
that P(τa < τb ) = .
a−b
9
Stochastic Calculus

9.1 Introduction
This chapter attempts to present certain introductory topics of stochastic calculus,
also called Ito’s calculus. But do we really need the apparatus of Ito’s calculus
in finance? The answer is YES. This is because to make any meaningful financial
decision, we need to understand the dynamics of the asset under consideration.
For example, to price a stock option, we need to understand the dynamics of
the underlying stock. Traditionally calculus in general and differential equations
in particular have played a very vital role in the mathematical modeling of any
dynamical system. We expect that these topics will again come to our rescue to
model the asset dynamics. But since the asset dynamics is mostly stochastic, we
need a different type of calculus to take care of functions which are no where
differentiable and are not of bounded variation (do not worry! we shall be defining
it shortly). Ito’s calculus serves our goal very well in this scenario.
When in the 19th century the German mathematician Weierstrass constructed
a real-valued function which is continuous everywhere, but differentiable nowhere,
(see Section 9.10) this was considered as nothing else but a mathematical curios-
ity. Interestingly, this “curiosity” is at the core of mathematical finance. High
frequency data show that prices of exchange rates, interest rates, and liquid assets
are practically continuous, but are of unbounded variation in every given time in-
terval. In particular, they are nowhere differentiable. Therefore classical calculus is
required to be extended to functions of unbounded variation, a task overlooked by
mathematicians for long. This gap was bridged by the development of stochastic
calculus, which can be considered as the theory of differentiation and integration
of stochastic processes.
Although the subject of stochastic calculus is very broad and complex, we keep
our goal very modest. We present a very brief and introductory discussion on
278 Financial Mathematics: An Introduction

certain topics of stochastic calculus, stochastic differential equations, Ito’s integral

and Ito’s formula. Our discussion here is motivated by the writings of Mikosch
[97], Shreve [122], Karatzas and Shreve [75], and Klebaner [78].

9.2 Variations of Real-Valued Functions

Before we discuss the variations of Brownian motion, we first discuss the variations
of real-valued functions. Here for the sake of simplicity we consider the interval
[0, T] rather than a general interval [a, b].
Definition 9.2.1 (First Variation of a Real-Valued Function) Let g :
[0, T] → R and Π be the set of all partitions π = {0 = t0 < t1 < . . . < tn = T} of
[0, T]. Then the first variation of g over [0, T] is defined as
⎛ n−1 ⎞
⎜⎜ ⎟⎟
V g (T) = sup ⎜⎜⎝⎜ | g(ti+1 ) − g(ti ) |⎟⎟⎟⎠ .
π∈Π i=0

Here we are using the symbol V g (T) rather than V g [0, T] since left end point 0
is fixed throughout discussion while right end point T can vary.
Definition 9.2.2 (Function of Bounded Variation) Let 0 < t ≤ T. Then g
is said to be of finite variation if V g (t) < ∞, for all t. Further if for 0 < t ≤ T,
V g (t) < K, a constant independent of t, then g is said to be of bounded variation
on [0, T].
Definition 9.2.3 (p-Variation of a Real-Valued Function) Let 0 < t ≤ T.
Then the p-variation, p > 1, of the function g in the interval [0, t] is defined as
⎛ n−1 ⎞
⎜⎜ ⎟⎟
V g (t) = sup ⎜⎜⎜⎝ | g(ti+1 ) − g(ti ) |p ⎟⎟⎟⎠ .
p
(9.1)
π∈Π i=0

Remark 9.2.1 (Goffman [51]) In case g is a continuous function then V g (T) can
alternatively be expressed as
⎛ n−1 ⎞
⎜⎜ ⎟⎟
V g (T) = lim ⎜⎜⎜⎝ | g(ti+1 ) − g(ti ) |⎟⎟⎟⎠ ,
π→0
i=0

where π is an arbitrary partition {0 = t0 < t1 < . . . < tn = T} and  π =

Max (tk+1 − tk ).
0≤k≤n−1
 Variations of Real-Valued Functions 279

In a similar manner, the p-variation can be expressed as

⎛ n−1 ⎞
⎜
⎜ ⎟⎟
V g (t) = lim ⎜⎜⎜⎝ | g(ti+1 ) − g(ti ) |p ⎟⎟⎟⎠ .
p
(9.2)
π→0
i=0

In our discussion here we shall mostly be using these alternate expressions of

p
V g and V g because our functions of interest will be continuous.
Result 9.2.1 (Goffman [51]) Let the p-variation of the function g in the interval
p p
[0, T] be denoted by V g (T). Let V g (T) < ∞. Then for r < p, V rg (T) = ∞, and for
q
q > p, V g (T) = 0.
Theorem 9.2.1 Let T > 0 and g : [0, T] → R be continuously differentiable
function. Then
:T 
(i) V g (T) = 0 | g (t) | dt < ∞.
(ii) V 2g (T) = 0.
Proof.
(i) We have

n−1 - --
--
V g (T) = lim g(ti+1 ) − g(ti )- .
π→0
i=0

Since g is continuously differentiable function, by applying Mean Value Theo-

rem, we have for each subinterval [ti , ti+1 ]

g(ti+1 ) − g(ti ) = g (t∗i )(ti+1 − ti ) ,
for some point t∗i in (ti , ti+1 ). Hence,

n−1
V g (T) = lim | g(ti+1 ) − g(ti ) |
π→0
i=0

n−1

= lim | g (t∗i )(ti+1 − ti ) |
π→0
i=0
, T

= | g (t) | dt < ∞ .
0

The last integral is finite because g (t) is continuous on [0, T] and therefore

| g (t) | is Riemann integrable.
280 Financial Mathematics: An Introduction

(ii) Next, the 2-variation (or quadratic variation) of g is given by


n−1 - --2
--
V (2)
g (T) = lim g(ti+1 ) − g(ti )-
π→0
i=0
n−1 --
 -2
-- g(ti+1 ) − g(ti ) ---
= lim - ti+1 − ti - (ti+1 − ti ) .
2
π→0
i=0

By applying Mean Value Theorem for some t∗i ∈ (ti , ti+1 ), i = 0, 1, . . . , n − 1, we

have

n−1 -
--  ∗ ---2
V (2)
g (T) = lim g (ti ) (ti+1 − ti )2 ,
π→0
i=0

n−1 -
--  ∗ ---2
≤ lim  π  g (ti ) (ti+1 − ti )
π→0
i=0

n−1 -
--  ∗ ---2
= lim  π  lim  g (ti ) (ti+1 − ti )
π→0 π→0
i=0
, T --  --2
= lim  π  - g (t)- dt
π→0 0
=0 .

--  --2
Here we have used the fact that g (t) is continuous on [0, T] and therefore - g (t)-
is Riemann integrable.


p
Example 9.2.1 Consider g(t) = t2 . Determine V g (1) for p = 1, 2.
Solution We have , 1
 
g (t) = 2t, | g (t) | dt = 1 .
0
Hence , , -1
1

1
t3 -- 4
| g (t) | dt = 4
2
t dt = 4 -- = .
2

0 0 3 0 3
But , 1 --  --2
lim  π  - g (t)- dt = 0 .
π→0 0
 Variations of Brownian Motion 281

Therefore
V g (1) = 1; V 2g (1) = 0 .
Now by applying Result 9.2.1, we have


p 1, p=1
V g (1) =
0, p>1.



9.3 Variations of Brownian Motion

Now, we discuss the first and second variations for the Brownian motion over
[0, T]. We have already shown in Chapter 7 that sample paths of W(t) are nowhere
differentiable. Therefore the first variation does not make sense as it does not exist
for the Brownian motion. Next we evaluate the second variation, also called the
quadratic variation of the Brownian motion.
The quadratic variation for Brownian motion {W(t), t ≥ 0} over the interval
[0, T], denoted by [W, W](T), is given by

[W, W](T) = VW(t)

2
(T) = lim Qπ ,
π→0

where

n−1
Qπ = (W(ti+1 ) − W(ti ))2 . (9.3)
i=0

But now we face certain difficulties and these need to be addressed. Clearly, Qπ is
a function of the sample points w ∈ Ω. Hence, the quadratic variation calculated
for the Brownian motion for each partition is itself a random variable. Here we
note that, the limit is to be taken over all partitions of [0, T], with  π → 0 as
n → ∞. Since, for each partition π, Qπ is a random variable, we need to specify
the sense in which we are finding the limiting distribution of Qπ for large n. In
other words we need to specify the proper mode of convergence in these random
variables. We shall use the convergence in mean square (convergence in L2 ) sense
as defined below.
Definition 9.3.1 Let {Xn , n ≥ 1} and X be random variables defined on a common
probability space (Ω, F , P). We say that Xn converges to X in mean square sense
(in L2 sense) if
lim E(|Xn − X|2 ) = 0 .
n→∞
282 Financial Mathematics: An Introduction

In the case of Brownian motion, we will show that {Qπ } converges to T in mean
square sense, i.e.
lim E(|Qπ − T|2 ) = 0 . (9.4)
π→0

When the above result holds good, we say that the quadratic variation accumu-
lated by the Brownian motion over the interval [0, T] is T almost surely and is
denoted as [W, W](T) = T.
Theorem 9.3.1 Let Qπ be defined as in (9.3). Then
(i) E(Qπ ) = T,
(ii) Var(Qπ ) ≤ 2  π  T .
Proof.
(i) We have

n−1
E(Qπ ) = E (W(ti+1 ) − W(ti ))2 . (9.5)
i=0

Since, for fixed i, W(ti+1 ) − W(ti ) has normal distribution with mean zero and
variance (ti+1 − ti ), equation (9.5) gives


n−1
E(Qπ ) = (ti+1 − ti ) = T .
i=0

(ii) Next we have


n−1
Var(Qπ ) = Var(W(ti+1 ) − W(ti ))2 . (9.6)
i=0

But

Var(W(ti+1 )−W(ti ))2 = E(W(ti+1 )−W(ti ))4 −2E(W(ti+1 )−W(ti ))2 (ti+1 −ti )+(ti+1 −ti )2 .
(9.7)
Since the fourth order moment of normal distribution with mean zero and
variance (ti+1 − ti ) is 3(ti+1 − ii )2 (see Exercise 7.12 in Chapter 7), we get from
(9.7)

Var(W(ti+1 ) − W(ti ))2 = 3(ti+1 − ti )2 − 2(ti+1 − ti )2 + (ti+1 − ti )2

= 2(ti+1 − ti )2 . (9.8)

Substituting (9.8) in (9.6), we get

 Variations of Brownian Motion 283


n−1
Var(Qπ ) = 2(ti+1 − ti )2
i=0

n−1
≤ 2  π  (ti+1 − ti ) = 2  π  T . (9.9)
i=0


Remark 9.3.1 The above theorem tells that for the Brownian motion {W(t), t ≥
0}, [W, W](T) = T for all T ≥ 0 and almost surely. This is because Var(Qπ ) =
E (Qπ − E(Qπ ))2 = E (Qπ − T)2 , which from (9.9) gives

lim E (Qπ − T)2 = lim Var(Qπ ) = 0 .

π→0 π→0

Therefore
[W, W](T) = lim Qπ = T ,
π→0

where the limit is understood in the mean square sense because

lim E(|Qπ − T|2 ) = 0 .

π→0

Hence we write [W, W](T) = T almost surely. Here the terminology almost surely
means that there can be some paths of the Brownian motion for which the assertion
[W, W](T) = T is not true. But the ‘the set of all such paths’ has zero probability.
Though we write [W, W](T) = T, we must realise that it is to be understood in the
sense as described above.
Remark 9.3.2 In view of Theorem 9.2.1, we have
⎧
⎪
⎪ ∞, p=1
p ⎪
⎨
VW (T) = ⎪
⎪ T, p=2
⎪
⎩ 0, p>2.

This concludes that, the Brownian motion {W(t), t ≥ 0} is of unbounded variation

and finite quadratic variation for every t > 0.
Remark 9.3.3 We know that for the given Brownian motion {W(t), t ≥ 0},
[W, W](T) = T, i.e. ⎛ n−1 ⎞
⎜⎜ ⎟⎟
lim ⎜⎜⎜⎝ (W(ti+1 ) − W(ti ))2 ⎟⎟⎟⎠ = T . (9.10)
π→0
i=0
284 Financial Mathematics: An Introduction

Also for 0 < T1 < T2 , [W, W](T2 ) − [W, W](T1 ) = T2 − T1 , the Brownian mo-
tion accumulates (T2 − T1 ) units of quadratic variation over the interval [T1 , T2 ].
Since this is true for every interval, we infer that the Brownian motion accumu-
lates quadratic variation at rate one per unit time. This last statement we write
informally as
dW(t) dW(t) = dt , (9.11)
and remember that the dt in (9.11) is in fact 1.dt.
The mathematical justification of the formula (9.11) is essentially equation
(9.10). At this stage we shall also like to know some formula for dW(t) dt and
dt dt. To determine these, we need to compute the cross variation of W(t) and t
and also the quadratic variation of t itself. We have the below given theorem in
this regard.
Theorem 9.3.2 Let {W(t), t ≥ 0} be the given Brownian motion and π = {0 =
t0 < t1 < . . . < tn = T} be a partition of [0, T]. Then
(i) ⎛ n−1 ⎞
⎜⎜ ⎟⎟
lim ⎜⎜⎜⎝ (W(ti+1 ) − W(ti )) (ti+1 − ti )⎟⎟⎟⎠ = 0 ,
π→0
i=0

(ii) ⎛ n−1 ⎞
⎜⎜ ⎟⎟
lim ⎜⎜⎜⎝ (ti+1 − ti )2 ⎟⎟⎟⎠ = 0 .
π→0
i=0

Proof.
(i) We observe that

|(W(ti+1 ) − W(ti )) (ti+1 − ti )| ≤ Max |W(ti+1 ) − W(ti )| (ti+1 − ti ) .

0≤i<n

Therefore
-- --
-- n−1 -
-- (W(ti+1 ) − W(ti )) (ti+1 − ti )--- ≤ Max |W(tk+1 ) − W(tk )| .T . (9.12)
- i=0 - 0≤k≤n

Since W(t) is continuous, the R.H.S of (9.12) goes to zero as  π → 0.

 The Riemann and The Riemann-Stieltjes Integrals 285

(ii) We note that


n−1
 n−1
2
(ti+1 − ti ) ≤ Max (ti+1 − ti ) . (ti+1 − ti ) = π  T ,
0≤k≤n−1
i=0 i=0

which goes to zero as  π → 0.



Remark 9.3.4 In view of Theorem 9.3.2 and in analogy with (9.10) and (9.11),
we can informally write
dW(t) dt = 0 , dt dt = 0 .
These relations together with dW(t) dW(t) = dt become very handy in the working
of stochastic calculus. We shall like to emphasize the word ‘informally’ here because
actual derivation is rather involved and uses law of large numbers.

9.4 The Riemann and The Riemann-Stieltjes Integrals

In this section, we discuss the notion of ordinary integral or Riemann integral
and also consider an extension, the Riemann-Stieltjes integral. At the end of this
section, we explain why classical integrals may fail when the integrand or the
integrator are Brownian sample paths.
For the sake of consistency, we continue to work with [0, T]; however it is so
very natural to change it, if desired, to any arbitrary [a, b].

Riemann Integral

Let g be a real-valued function defined on [0, T]. We assume that g is continuous

in [0, T]. Consider the integral
, T
S= g(t) dt . (9.13)
0

Let π = {0 = t0 < t1 < . . . < tn = T} be an arbitrary partition of the interval [0, T].
For each subinterval [ti , ti+1 ], we set Mi = Max g(t) and mi = Min g(t). Then the
ti ≤t≤ti+1 ti ≤t≤ti+1
upper Riemann sum is defined as

n−1
S+π = Mi (ti+1 − ti ) ,
i=0
286 Financial Mathematics: An Introduction

and the lower Riemann sum is defined as


n−1
S−π = mi (ti+1 − ti ) .
i=0

Let the maximum step size of the partition π be defined as

 π = Max (tk+1 − tk ) .
0≤k≤n−1

When the number n of partition points goes to infinity and the length of the
longest subinterval tk+1 − tk goes to zero (i.e.  π → 0), the upper Riemann sum
S+π and the lower Riemann sum S−π converge to the same limit, which we call
:T
0
g(t)dt. Equivalently the Riemann integral can also be defined as
, T 
n−1
S= g(t) dt = lim g(t∗i )(ti+1 − ti ) , (9.14)
0 π→0
i=0

where t∗i is arbitrary point in [ti , ti+1 ].

In the above definition we must note that the function is Riemann integrable
provided the limit does not depend on which partition π of [0, T] has been taken
and which point t∗ in [ti , ti+1 ] has been chosen. Thus irrespective of the way the
interval [0, T] has been partitioned and the point t∗ in [ti , ti+1 ] has been chosen,
the limit of the sum (9.14), as  π → 0, remains the same. We can show that all
continuous functions and all piecewise continuous bounded functions are Riemann
integrable.

Riemann-Stieltjes Integral

Suppose that g and f are real-valued functions defined on [0, T]. We assume
that on [0, T], g is continuous and f is monotonically nondecreasing. We aim to
define the integral , T
S= g(t) d f (t) . (9.15)
0

As for the Riemann integral, we consider an arbitrary partition π = {0 = t0 < t1 <

. . . < tn = T} of [0, T]. For each subinterval [ti , ti+1 ], we set Mi = Max g(t) and
ti ≤t≤ti+1
mi = Min g(t). The corresponding upper and lower Riemann-Stieltjes sums are
ti ≤t≤ti+1
respectively
 The Riemann and The Riemann-Stieltjes Integrals 287


n−1
S+π = Mi ( f (ti+1 ) − f (ti )) ,
i=0

and

n−1
S−π = mi ( f (ti+1 ) − f (ti )) .
i=0

When  π → 0, the upper Riemann-Stieltjes sum S+π and the lower Riemann-
:T
Stieltjes sum S−π converge to the same limit, which we call 0 g(t)d f (t). Equiva-
lently the Riemann-Stieltjes integral can also be defined as

, T 
n−1
 
S= g(t) d f (t) = lim g(t∗i ) f (ti+1 ) − f (ti ) , (9.16)
0 π→0
i=0

where t∗i is arbitrary point in [ti , ti+1 ].

We note that if g is continuous on [0, T], then g is Riemann-Stieltjes integrable

with respect to every f which is monotonically nondecreasing on [0, T].
The Riemann-Stieltjes integral has application in probability theory to find the
expectation of a random variable X

, ∞
E(X) = t dF(t) ,
−∞

where F(t) is the cumulative distributive function of X. For instance, if g(t) = t

and F(t) is given by

⎧
⎪
⎪ 0, t<0
⎪
⎨
F(t) = ⎪
⎪ p + (1 − p)(1 − e−λt ), 0≤t<T
⎪
⎩ 1, t≥T,

where 0 < p < 1, λ > 0 and T > 0, then

288 Financial Mathematics: An Introduction

, ∞
E(X) = tdF(t)
−∞
, 0 , T , ∞
= tdF(t) + 0 × P(X = 0) + tdF(t) + T × P(X = T) + tdF(t)
−∞ 0 T
, T
=0+0+ tλ(1 − p)e−λt dt + T(1 − p)e−λT + 0
0
,
(1 − p) λT −y
= ye dy + T(1 − p)e−λT
λ 0
(1 − p)
= (1 − λTe−λT − e−λT ) + T(1 − p)e−λT
λ
1  
= (1 − p) 1 − e−λT .
λ

9.5 Stochastic Integral and its Properties

In the last section we have presented the classical Riemann and Riemann-Stieltjes
integrals. In this section we wish to extend the concept of ‘integration’ of a stochas-
tic process {X(t), t ≥ 0} with respect to a given Brownian motion {W(t), t ≥ 0}.
But then the meaning of the word ‘integration’ is to be understood appropriately.
Let {X(t), t ≥ 0} be a stochastic process which is adapted to the natural filtration
{F
t , t ≥ 0} of Wiener
 process {W(t), t ≥ 0}, i.e. X(t) is Ft -measurable. Let us assume
:T
E 0 X2 (t) dt < ∞.
We next consider a partition π of [0, T] where π = {0 = t0 < t1 < . . . < tn = T},

n−1
and form the sum X(ti ) (W(ti+1 ) − W(ti )). Now if we take the limit of this sum
i=0
as  π → 0, then in analogy with the procedure discussed in Section 9.4 we can
write , T 
n−1
I(T) = X(s) dW(s) = lim X(ti ) (W(ti+1 ) − W(ti )) . (9.17)
0 π→0
i=0

Here it must be noted that at present equation (9.17) is purely informal because
we have not yet attached any meaning to the ‘limiting process’ of the sum of
random variables involved in the R.H.S of this equation. In defining stochastic
integral we use the mean square convergence as in Definition 9.3.1. Once we agree
to this mode of convergence, then (9.17) is well defined. We therefore take (9.17)
 Stochastic Integral and its Properties 289

as the definition of Stochastic integral or Ito integral of the stochastic process

{X(t), t ≥ 0} with respect to the Brownian motion {W(t), t ≥ 0}.
Remark 9.5.1 Looking at the partial sums involved in (9.17) and (9.12) (or
(9.11)), we find one major difference. In (9.11) and (9.12) we were free to choose
any t∗i in the ith subinterval [ti , ti+1 ] and evaluate the integrand g at t∗i . But here we
are constrained to choose the left end point ti of the subinterval [ti , ti+1 ] and take
t∗i to be ti only. This will require evaluation of the integrand X(t) at ti and not at
any other point t∗i of our choice in [ti , ti+1 ]. This has been done so as to make Ito
:t
integral I(t) = 0 X(s) dW(s), 0 < t ≤ T, a martingale.
Properties of Ito integral

The stochastic integral I(t), 0 < t ≤ T, satisfies the following properties.

(i) E(I(t)) = 0.
: t 2 : t 
(ii) E 0 X(s) dW(s) = E 0 X2 (s) ds . This property is called Ito isometry.
(iii) Let {X(1) (t), t ≥ 0} and {X(2) (t), t ≥ 0} be stochastic processes having stochastic
integrals with respect to {W(t), t ≥ 0}. Let α and β be constants. Then
, t  , t , t
αX (s) + βX (s) dW(s) = α
(1) (2)
X (s) dW(s) + β
(1)
X(2) (s) dW(s) .
0 0 0

This is called the linearity property of the Ito integral.

(iv)
, t , t1 , t
X(s) dW(s) = X(s) dW(s) + X(s) dW(s)
0 0 t1

for 0 ≤ t1 ≤ t.
(v) The process I(t) has continuous sample path.
(vi) For each t, I(t) is Ft - measurable.
:t
(vii) [I, I](t) = 0 X2 (s) ds.
:t
(viii) The process I(t) = 0 X(s) dW(s), t ∈ [0, T], is a martingale with respect to
the natural Brownian filtration Ft (0 ≤ t ≤ T).
This last property is very important in applications. It essentially follows be-
cause in the definition of Ito integral, in the subinterval [ti , ti+1 ] we have taken t∗i
as the left end point. So this choice of t∗i very crucial to make I(t) a martingale.
We are not proving above results, and shall refer to Karatzas and Shreve [75]
for the same.
290 Financial Mathematics: An Introduction

Remark 9.5.2 The Ito integral does not : T have the monotonicity
:T property. Thus
X(t) ≤ Y(t) does not necessarily mean 0 X(t) dW(t) ≤ 0 Y(t) dW(t). We may
:T :T
take X(t) = 0 and Y(t) = 1. Then 0 1.dW(t) = W(T) and 0 0.dW(t) = 0. But
1
W(T) is smaller than 0 with probability .
2
We shall now have certain examples to illustrate some of the points discussed
above.
Example 9.5.1 Evaluate the Ito integral
, T
W(s) dW(s) .
0

Solution Let π = {0 = t0 < t1 < . . . < tn = T} be an arbitrary partition of [0, T].

We have
, T 
n−1
W(s) dW(s) = lim W(ti ) (W(ti+1 ) − W(ti )) . (9.18)
0 π→0
i=0

But, for each i, W(ti ) and W(ti+1 ) − W(ti ) are independent random variables and
are having normal distributions. Hence, the right hand side terms within the
summation are nothing but the sum of independent random variables. Hence, the
integral is nothing but the limit of sum of such random variables. Now, we have


n−1
Qπ = (W(ti+1 ) − W(ti ))2
i=0

n−1  
= W 2 (ti+1 ) − W(ti )2 − 2W(ti )(W(ti+1 ) − W(ti ))
i=0

n−1
= W (T) − W (0) − 2
2 2
W(ti )(W(ti+1 ) − W(ti )) ,
i=0

i.e.

n−1
1 2 
W(ti ) (W(ti+1 ) − W(ti )) = W (T) − W 2 (0) − Qπ . (9.19)
i=0
2
Now taking limit as  π → 0 and using E(Qπ ) = T, we get
 Stochastic Integral and its Properties 291

, T
W 2 (T) − T
W(s) dW(s) = .
0 2

We see that, unlike the Riemann-Stieltjes integral, we have an extra term T/2,
which arises on account of the finite quadratic variation of the Brownian motion.


*n−1
Remark 9.5.3 In Example 9.5.1, if we form the sum i=0 W(t∗i ) (W(ti+1 ) − W(ti ))
ti + ti+1
with t∗i = , and take its limit as  π → 0, then we obtain this limit as
2 :T
0.5W 2 (T). But can we say that the Ito integral 0 W(s) dW(s) equals 0.5W 2 (T)?
This is not correct because we desire Ito integral to be a martingale, and 0.5W 2 (T)
is not a martingale. However when we take t∗i = ti , the left end point of the interval
[ti , ti+1 ], then the limit is (W 2 (T) − T)/2 which is a martingale. The requirement
that I(t) should be a martingale tells us to choose t∗i as ti .
Remark 9.5.4 There is a more compelling practical reason to choose t∗i as ti (the
left point of [ti , ti+1 ]). Let t > 0 and π = {0 = t0 < t1 < . . . < tn = t}. We can think
of t0 , t1 , . . . , tn−1 as the trading dates in the asset and ∆(t0 ), ∆(t1 ), . . . , ∆(tn−1 ) as the
position (number of shares) taken in the asset at each trading date and held to the
next trading date. If I(t) denote the gain from trading at each limit t, then
⎧
⎪
⎪ ∆(t0 ) [W(t) − W(0)] = ∆(0)W(t), 0 ≤ t ≤ t1
⎪
⎪
⎪ ∆(t0 )W(t1 ) + ∆(t1 ) [W(t) − W(t1 )] ,
⎪ t 1 ≤ t ≤ t2
⎪
⎪
⎪ .
⎨ ..
I(t) = ⎪ ⎪
⎪
⎪ *k−1
⎪
⎪ i=0 ∆(ti ) [W(ti+1 ) − W(ti )] + ∆(tk ) [W(t) − W(tk )] , tk ≤ t ≤ tk+1
⎪
⎪
⎪ .
⎩ ..
:t
Obviously the process I(t) defined above becomes our Ito integral 0 ∆(s) dW(s) of
the simple process ∆(t).
:t
Remark 9.5.5 In the Ito integral 0 X(s) dW(s), we can have a financial inter-
pretation of integrand and integrator. The integrand represents a position in an
asset and integrator represents the price of that asset. Since we need to decide
the position at the beginning of each interval, we have to take t∗i as ti , rather an
arbitrary point in [ti , ti+1 ]. Thus there are theoretical as well as practical reasons
for choosing t∗i as the left end point of the interval [ti , ti+1 ].
292 Financial Mathematics: An Introduction

Example 9.5.2 Evaluate

, t
W(1) dW(s), 0 ≤ t ≤ 1 .
0

Solution Note that, W(1) is not adapted to the filtration σ{W(s), 0 < s ≤ t}, 0 ≤
t ≤ 1, because it depends on future events. Hence, this Ito integral does not exist.
This example shows that, assumption of the integrand adapted to the filtration
{Ft , t ≥ 0} is needed to have existence of the Ito integral.


:T
Example 9.5.3 Let X(t) and Y(t) be suitable processes so that I1 = 0 X(t) dW(t)
:T :T
and I2 = 0 Y(t) dW(t) exist. Show that E(I1 .I2 ) = 0 E (X(t)Y(t)) dt.
1 
Solution We know that I1 I2 = (I1 + I2 )2 − I12 − I22 . Now we use the isometry
2
property and get the result.



9.6 Ito-Doeblin Formula and its Variants

In many situations we are interested in ‘differentiating’ an expression of the form
f (W(t)). For example, if we assume that the stock price at any time t is of the form
S(t) = f (W(t)), where f is a twice continuously differentiable real-valued function,
then to study the dynamical behavior of S(t) we need to ‘differentiate’ f (W(t)).
For the real-valued differentiable functions f and g, we have the standard chain
rule
d  
f (g(t) = f  (g(t)) g (t) ,
dt
i.e.
 
d f (g(t) = f  (g(t)) dg(t).

But do we have similar result for f (W(t))? The answer is provided by the Ito-
Doeblin formula.

1. Ito-Doeblin Formula for Brownian Motion: First Version

 Ito-Doeblin Formula and its Variants 293

Let f be at least twice continuously differentiable function of x and {W(t), t ≥ 0}

be a Weiner process. Then
1 
d f (W(t)) = f  (W(t)) dW(t) + f (W(t)) dt , (9.20)
2
or equivalently
, t , t
 1
f (W(t)) = f (W(0)) + f (W(u)) dW(u) + f  (W(u)) du . (9.21)
0 2 0

Here the first integral in (9.21) is an Ito integral whereas the second integral in
(9.21) is a Reimann integral.
Remark 9.6.1 Borrowing ideas from the classical Taylor’s series we can infor-
mally write

d f (W(t)) = f (W(t) + dW(t)) − f (W(t))

1
= f  (W(t)) dW(t) + f  (W(t)) (dW(t))2 + . . . .
2
But then as per our understanding

(dW(t))2 = (dW(t)) (dW(t)) = dt

(dW(t))3 = (dW(t))2 (dW(t)) = dt (dW(t)) = 0 ,

and
(dW(t))4 = (dW(t))2 (dW(t))2 = dt (dt) = 0 .
This is true for other higher order terms as well. Though it is not a proof of
Ito-Doeblin formula (9.20), it certainly convinces about its validity.
:T
Example 9.6.1. Find 0 W(t) dW(t) using the Ito-Doeblin formula.
Solution Consider the first version of the Ito-Doeblin formula
1 
d f (W(t)) = f  (W(t)) dW(t) + f (W(t)) dt ,
2
or equivalently
, T , T
 1
f (W(T)) − f (W(0)) = f (W(t)) dW(t) + f  (W(t)) dt .
0 2 0
294 Financial Mathematics: An Introduction

:t
Taking motivation from the fact that 0 x dx = t2 /2, we choose f (x) = x2 /2. This
gives f  (x) = x and f  (x) = 1. Hence the above Ito-Doeblin formula gives
, T , T
W 2 (T) 1
−0= W(t) dW(t) + dt .
2 0 2 0

Therefore , T
W 2 (T) − T
W(t) dW(t) = . (9.22)
0 2


:T
Remark 9.6.2 Here we may note that, the integral 0 W(t) dW(t) cannot be de-
fined path by path in Riemann-Stieltjes sense, because sample path of Brownian
motion is of unbounded variation on each time interval. In the formula
, T
W 2 (T) − T
W(t) dW(t) = , t≥0,
0 2

T
there is an additional term of − . This is because the local increment of the
2
Wiener √ process over an increment of length ∆t is of the size of its standard de-
viation ∆t. However for a smooth continuously differentiable function f (t), the
second term in (9.22) is zero, as it should be.
Example 9.6.1 Show that
, t , t
1
eW(t)
=1+ e W(s)
dW(s) + eW(s) ds .
0 2 0

Solution We take f (x) = ex and then use the Ito-Doeblin formula.



2. Ito-Doeblin Formula for Brownian Motion: Second Version

Let f (t, x) have continuous partial derivatives of at least second order and
{W(t), t ≥ 0} be the given Weiner process. Then

1
d f (t, W(t)) = ft (t, W(t)) dt + fx (t, W(t)) dW(t) + fxx (t, W(t)) dt ,
2
or equivalently
 Ito-Doeblin Formula and its Variants 295

, t  , t
1
f (t, W(t))− f (0, W(0)) = ft (u, W(u)) + fxx (u, W(u)) du+ fx (u, W(u)) dW(u) .
0 2 0

This formula can again be justified by considering the classical Taylor’s expansion
for a function of two variables. In particular we may take
 
∂ f (t, W(t)) ∂ f (t, W(t))
f (t + ∆t, W(t + ∆t)) − f (t, W(t)) = dt + dW(t)
∂t ∂x

1 ∂2 f (t, W(t)) ∂2 f (t, W(t))
+ (dt) 2
+ 2 dt dW(t)
2 ∂t2 ∂t∂x

∂2 f (t, W(t))
+ dW(t) dW(t) + . . .
∂x2
But since we have dW(t) dW(t) = dt, dW(t) dt = 0 and dt dt = 0, therefore,
 
∂ f (t, W(t)) 1 ∂2 f (t, W(t))
f (t + ∆t, W(t + ∆t)) − f (t, W(t)) = + dt
∂t 2 ∂x2
∂ f (t, W(t))
+ dW(t),
∂x

where
- - -
∂ f (t, W(t)) ∂ f (t, x) -- ∂ f (t, W(t)) ∂ f (t, x) -- ∂2 f (t, W(t)) ∂2 f (t, x) ---
= - , = - , = - .
∂x ∂x -x=W(t) ∂t ∂t -x=W(t) ∂x2 ∂2 x -x=W(t)

, T
Example 9.6.2 Using second version of the Ito-Doeblin formula, find W(t) dW(t).
0

Solution For any T > 0, we have

, T  , T
1
f (T, W(T))− f (0, W(0)) = ft (u, W(u)) + fxx (u, W(u)) du+ fx (u, W(u)) dW(u) .
0 2 0

x2
Choose f (t, x) = . Then fx (t, x) = x, ft (t, x) = 0 and fxx (t, x) = 1. Substituting,
2
we get , T
, T

W 2 (T) 1
−0= 0+ du + W(u) dW(u)
2 0 2 0
296 Financial Mathematics: An Introduction

Hence,
, T
W 2 (T) T
W(t) dW(t) = − .
0 2 2


W(t)− 2t
Example 9.6.3 Find the stochastic differential dX(t) of X(t) = e .
t
Solution We take f (t, x) = ex− 2 and then use the second version of the Ito-Doeblin
formula. This gives

dX(t) = d f (t, W(t)) = f (t, W(t)) dW(t) = X(t) dW(t) .




9.7 Stochastic Differential Equation

In this section we introduce the concept of stochastic differential equations. For
this let us first consider the initial value problem (IVP)

dx(t)
= f (t, x(t)), t ∈ [0, T], x(0) = x0 , (9.23)
dt
where f : [0, T] × R → R is a continuous function. A continuously differentiable
function x : [0, T] → R is called a solution of the given IVP if x(0) = x0 and x(t)
satisfies the ordinary differential equation (ODE) (9.23) for all t ∈ [0, T]. We note
that the IVP (9.23) is equivalent to the integral equation given by
, t
x(t) = x0 + f (s, x(s)) ds . (9.24)
0

Let us recollect that f satisfies a Lipschitz condition if there exists a constant

k > 0 such that

| f (t, x) − f (t, y) | ≤ k | x − y | for all t ∈ [0, T], x, y ∈ R .

Then it is well known that for such a function f the IVP (9.23)(or the correspond-
ing integral equation (9.24)) has a unique solution. Further this unique solution
x(t) can be obtained by applying the standard Picard’s method
, t
xn+1 (t) = x0 + f (s, xn (s)) ds, (n = 0, 1, . . .) ,
0
 Stochastic Differential Equation 297

giving
x(t) = lim xn (t) .
n→∞

Suppose we now introduce randomness in the above IVP (9.23). This random-
ness could either be introduced in x(0) or in f . If x(0) = x0 not fixed, but rather a
random variable, then for each w ∈ Ω, the IVP (9.23) can be solved. The solution
in this case will depend on Ω. Thus the solution x(t) will be a stochastic process
{X(t, w), t ∈ [0, T], w ∈ Ω}. Similarly, if the function f depends on w ∈ Ω, then
again the solution {X(t, w), t ∈ [0, T], w ∈ Ω} is a stochastic process. Such differen-
tial equations are known as random differential equations. Here we observe that,
in both these cases of introducing randomness, we solve IVP’s for each w ∈ Ω.
We now discuss the stochastic differential equations. These are different from
random differential equations. Here we introduce uncertainties in the equation
(9.23) by introducing an additive term, which is the ‘derivative’ of Brownian
motion W(t). Symbolically we write

dX(t) dW(t)
= b(t, X(t)) + σ(t, X(t)) , (0 ≤ t ≤ T) ,
dt dt
where b : [0, T] × R → R and σ : [0, T] × R → R are two given functions. The
above equation can also be symbolically written

dX(t) = b(t, X(t)) dt + σ(t, X(t)) dW(t) . (9.25)

A naive interpretation of (9.25) tells us that the change dX(t) = X(t + dt) − X(t)
is caused by a change dt of time, with factor b(t, X(t)) in combination with a
change dW(t) = W(t + dt) − W(t) of Brownian motion with factor σ(t, X(t)). Now
borrowing analogy from classical ODE we also symbolically express (9.25) in terms
of integral equation
, t , t
X(t) = X(0) + b(s, X(s)) ds + σ(s, X(s)) dW(s) , (0 < t ≤ T) . (9.26)
0 0

Here the initial condition X(0) and the coefficient functions b(t, x) and σ(t, x) are
given. The differential equation (9.25) is termed as the stochastic differential equa-
tion and (9.26) is termed as stochastic integral equation. We shall now discuss the
types of possible solution of SDE (9.25) or its equivalent integral equation version
(9.26).
We must note that so far stochastic differential equation (9.25) and stochastic
integral equation (9.26) are only symbolic. They will make sense only after we
298 Financial Mathematics: An Introduction

define them in a proper mathematical way. In this context we shall like to empha-
size that mathematically only stochastic integral equation has a meaning, because
we have defined a stochastic integral in strict mathematical sense. The stochastic
differential equation shall always remain a notational convenience whose interpre-
tation shall always be in terms of the corresponding stochastic integral equation.
There are two types of solution concepts for stochastic differential equations.

1. Strong solution
A strong solution to the SDE (9.25) is a stochastic process {X(t); t ∈ [0, T]} which
satisfies the following
(i) {X(t), t ∈ [0, T]} is adapted to the Brownian motion, i.e. at time t it is a function
of W(s), s ≤ t.
(ii) The integrals in (9.26) are well defined and {X(t); t ∈ [0, T]} satisfies the same.
(iii) {X(t); t ∈ [0, T]} is a function of the underlying Brownian sample path and of
the coefficients b(t, x) and σ(t, x).
Thus a strong solution is an explicit function f such that X(t) = f (t, W(s) : s ≤ t).
A strong solution to (9.26) is based on the path of the underlying Brownian
motion. The solution {X(t); t ∈ [0, T]} is said to be unique strong solution if given
any other solution {Y(t), t ∈ [0, T]}, P(X(t) = Y(t)) = 1 for all t ∈ [0, T]

2. Weak solution
For a weak solution, the path behaviour is not essential. Hence we are only
interested in the distribution of X. Thus weak solutions are sufficient to determine
the expectation, variance and covariance functions of the process. In this case we
do not have to know the sample paths of X.
A strong or weak solution X of the given SDE is called a diffusion. We may
note that Brownian motion is also a diffusion process because in (9.26) we can
take b(t, x) = 0 and σ(t, x) = 1.
We now have the following existence theorem.
Theorem 9.7.1 (Existence Theorem) Let E(X2 (0)) < ∞ and X(0) be inde-
pendent of {W(t), t ≥ 0}. Let for all t ∈ [0, T] and x, y ∈ R, b(t, x) and σ(t, x) be
continuous and satisfy Lipschitz condition with respect to second variable, i.e.

| b(t, x) − b(t, y) | + | σ(t, x) − σ(t, y) | ≤ K | x − y | .

Then the SDE (9.26) has a unique strong solution {X(t), 0 ≤ t ≤ T}.
 Stochastic Differential Equation 299

Example 9.7.1 Consider the SDE

dX(t) = X(t) dW(t), with X(0) = 1 .

Find the strong solution using Ito-Doeblin formula.

Solution We have b(t, x) = 0 and σ(t, x) = x. Hence, Lipschitz condition is satis-
t
fied. Now we use the second version of Ito-Doeblin formula with f (t, x) = e(x− 2 ) .
This gives

∂ f (t, x) −1 x− t ∂ f (t, x) t ∂2 f (t, x) x− 2t

= e 2, = ex− 2 and = e .
∂t 2 ∂x ∂x2
On substituting these expressions in the second version of Ito-Doeblin formula,
we get

dX(t) = d f (t, W(t))

1
= ft (t, W(t)) dt + fx (t, W(t)) dW(t) + fxx (t, W(t)) dt

 2
t t
= − X(t) + X(t) dt + X(t) dW(t)
2 2
= X(t) dW(t) .
t
Hence, the required strong solution is X(t) = eW(t)− 2 .


Example 9.7.2 A stochastic process {S(t), t ≥ 0} is governed by

dS(t) = a S(t) dt + bS(t) dW(t)

where a and b are constants. Find the SDE of

(i) S(t).
(ii) ln(S(t)).
Solution
(i) Choose f (x) = x1/2 , then

1 1
ft = 0; fx = √ ; fxx = − 3/2 .
2 x 4x

now applying Ito-Doeblin formula, we get

300 Financial Mathematics: An Introduction

1
d( S(t)) = ft dt + fx dS(t) + fxx dS(t) dS(t)
2
1 1
=0+ dS(t) − √ dS(t) dS(t)
2 S(t) 8 x3/2
1 1
= [a S(t) dt + bS(t) dW(t)] − b2 (S(t))2 dt
2 S(t) 8 (S(t)) 3/2
 
a b 2 b
= − S(t) dt + S(t) dW(t) .
2 8 2

(ii) Choose f (x) = ln(x), then

1 1
ft = 0; fx = ; fxx = − 2 .
x x
Apply Ito-Doeblin formula, we get
1
d(ln(S(t))) = ft dt + fx dSt + fxx dSt dSt
2
1 1
=0+ dS(t) − dS(t) dS(t)
S(t) 2(S(t))2
1 1
= [a S(t) dt + bS(t) dW(t)] − b2 (S(t))2 dt
S(t) 2(S(t))2
 
b2
= a− dt + b dW(t) .
2


Definition 9.7.1 ( Ito process) Let {W(t), t ≥ 0} be a Brownian motion and let
{Ft , t ≥ 0} be the associated natural filtration. An Ito process is a stochastic process
{X(t), t ≥ 0} of the form
, t , t
X(t) = X(0) + ∆(u) dW(u) + Θ(u) du ,
0 0

where X(0) is non random, ∆(u) and Θ(u) are adapted processes.
A stochastic differential equation form of the Ito process {X(t), t ≥ 0} is

dX(t) = ∆t dW(t) + Θt dt .

All stochastic processes except those that have jumps are actually Ito processes.
 Stochastic Differential Equation 301

Another Extension of Ito-Doeblin Formula

Let X = {X(t), t ≥ 0} be an Ito process and f (t, x) be a function whose second
order partial derivatives are continuous. Then
, t , t
f (t, X(t)) − f (0, X(0)) = ft (s, X(s)) ds + fx (s, X(s)) dX(s)
0 0
,
1 t
+ fxx (s, X(s)) d[X, X](s) .
2 0

Equivalently
1
d f (t, X(t)) = ft (t, X(t)) dt + fx (t, X(t)) dX(t) + fxx (t, X(t)) dX(t) dX(t) .
2
We may further rewrite above expression as
, t , t
f (t, X(t)) = f (0, X(0)) + ft (s, X(s)) ds + fx (s, X(s)) ∆(s)dW(s)
0 0
, t ,
1 t
+ fx (s, X(s)) Θ(s)ds + fxx (s, X(s)) ∆2 (s)ds .
0 2 0

This is because dXt = ∆(t) dW(t) + Θ(t) dt and dX(t) dX(t) = ∆2 (t) dt.

Example 9.7.1. Find the stochastic differential of W 3 (t) and show that W 3 (t) is
an Ito process.

Solution Using the first version of Ito-Doeblin formula, we get

dW 3 (t) = 3W(t) dt + 3W 2 (t) dW(t) .

Now using the equivalent integral equation with the condition W(0) = 0, we have
, t , t
W (t) = 3
3
W(s) ds + 3 W 2 (s) dW(s) .
0 0
, t
Since 3W(s) is adapted process, the integral 3W(s) ds exist and is finite. Also,
:t 0
3W 2 (s) is mean square integrable, i.e. E 0 (3W 2 (s))2 ds < ∞. By using the Defini-
tion 9.7.1, W 3 (t) is an Ito process.


302 Financial Mathematics: An Introduction

Example 9.7.3 Show that tW(t) is an Ito process and find the stochastic differ-
ential of tW(t).
Solution We take f (t, x) = tx and then use the second version of Ito-Doeblin
formula to get , t , t
tW(t) = W(s) ds + s dW(s)
0 0
i.e. , ,
t t
s dW(s) = t W(t) − W(s) ds .
0 0
Since W(t) is adapted process and f (t) = t is a mean square integrable, for any
t > 0, tW(t) is an Ito process. The stochastic differential of tW(t) is
d(tW(t)) = W(t) dt + tdW(t) .


Recall that, we had proved earlier that Brownian motion {W(t), t ≥ 0}, W(0) =
0, a martingale and have the continuous sample path with [W, W](t) = t. But we
now have a natural converse question. If the above stated properties hold for a
process then will the process be a Brownian motion? The below given theorem
provides an answer to the question.
Theorem 9.7.2 (Levy’s Theorem) Let {M(t), t ≥ 0} be a martingale with re-
spect to a filtration {Ft , t ≥ 0}. Suppose M(0) = 0 and M(t) has continuous paths.
Further let [M, M](t) = t for all t ≥ 0. Then {M(t), t ≥ 0} is a Brownian motion
with associated filtration {Ft , t ≥ 0}.
Proof. We know that Brownian motion is a martingale and increments W(t)−W(s)
for s < t are normally distributed with mean 0 and variance t − s. Here we are
already given that M(t) is a martingale. Therefore if we prove that M(t) is normally
distributed with mean 0 and variance t, then it is a Brownian motion. For this
we apply Ito-Doeblin formula to M(t). Therefore for any function f (t, x) whose
partial derivatives exist and are continuous, we have
1
d f (t, M(t)) = ft (t, M(t)) dt + fx (t, M(t)) dM(t) + fxx (t, M(t0) dt . (9.27)
2
Here dM(t) dM(t) = dt since [M, M](t) = t. Equation (9.27) in integral form is
, t , t
1
f (t, M(t)) = f (0, M(0)) + [ ft (t, M(0)) + fxx (s, M(0))] ds + fx (s, M(s)) dM(s) .
0 2 0
(9.28)
 Some Important SDE’s and their Solutions 303

, t
Since M(t) is a martingale, I(t) = fx (s, M(s)) dM(s) is also a martingale. Also
0
I(0) = 0 = E(I(t)). Now taking expectation on both sides in (9.28) we get
, t
 
1
E f (t, M(t)) = f (0, M(0)) + E ft + fxx ds + 0 . (9.29)
0 2

If we now prescribe f (t, x) = exp(ux − 21 u2 t), for fixed u, then

1 1
ft = − u2 f (t, x), fx = u f (t, x), fxx = u2 f (t, x), and ft + fxx = 0 .
2 2
Hence from (9.29) we get E( f (t, M(t))) = f (0, M(0)), which gives
1
E(exp(u M(t) − u2 t)) = 1
2
i.e.
1 2
E(euM(t) ) = e 2 u t .
By using the uniqueness property of moment generating function, we conclude
that M(t) is normally distributed with mean 0 and variance t.



9.8 Some Important SDE’s and their Solutions

In this section we discuss some important SDE’s which are frequently encountered
in practice.

1. Geometric Brownian Motion

Let S(t) be the stock price at time t. Let −∞ < µ < ∞ be the constant growth
rate of the stock and σ > 0 be the volatility. Let us consider the SDE

dS(t) = µS(t) dt + σS(t) dW(t), S(0) is known.

We are interested in finding the strong solution of S(t), if it exists. For this we
first verify that conditions of Theorem 9.7.1 are satisfied. But this is true because
µ and σ are constants. Now we assume that S(t) = f (t, W(t)) and make use of the
second version of Ito-Doeblin formula. This gives
1
d f (t) = ft dt + fx dW(t) + fxx dt ,
2
304 Financial Mathematics: An Introduction

∂ f (t, x) ∂ f (t, x) ∂2 f (t, x)

where ft = , fx = and fxx = . Now, on comparing with the
∂t ∂x ∂x2
given SDE, we get

fx = σ f (9.30)
1
ft + fxx = µ f . (9.31)
2
Now solving equation (9.30), we get f (t, x) = eσx k(t) , for some function k(t). From

here we get ft = k (t) eσx and fxx = σ2 eσx k(t) which on substituting in equation
(9.31), we get  
 σ 2
k (t) eσx = µ − eσx k(t) .
2
Solving the above equation, we get
⎛ ⎞
⎜⎜
⎜⎜ σ2 ⎟⎟⎟⎟
⎜⎜µ− ⎟⎟ t
⎜⎝
k(t) = S(0) e 2 ⎟⎠ .

Therefore, the required solution is

⎛ ⎞
⎜⎜
⎜⎜ σ2 ⎟⎟⎟⎟
⎜⎜µ− ⎟⎟ t+σ W(t)
⎜⎝
S(t) = S(0) e 2 ⎟⎠ .

Here we may observe that for fixed t, S(t) follows lognormal distribution. Hence,
it can be verified that
(i) E(S(t)) = E(S(0)) eµt .
2
(ii) E((S(t))2 ) = E(S2 (0)) e(2µ+σ )t .
Further, we observe
σ2
(i) If µ > then S(t) → ∞ as t → ∞ almost surely.
2
σ2
(ii) If µ < then S(t) → 0 as t → ∞ almost surely.
2
σ 2
(iii) If µ = then S(t) will fluctuate between arbitrary large and arbitrary small
2
value as t → ∞.
 Summary and Additional Notes 305

2. Ornstein-Uhlenbeck Process

We consider another SDE

dX(t) = −µ X(t) dt + σ dW(t), X(0) = x0 , (9.32)

where −∞ < µ < ∞ and σ > 0 are constants. This equation is often referred to
as Langevin equation. This equation is useful to model the velocity at time t of a
free particle that performs a Brownian motion. If we set dt = 1 then

X(t + 1) − X(t) = −µ X(t) + σ (W(t + 1) − W(t))

i.e.

X(t + 1) = µ̂ X(t) + σ (W(t + 1) − W(t)) ,

where µ̂ = 1 − µ. This model can be considered as a discrete analogue of the

solution of the Langevin equation (9.32).
We derive the solution of (9.32) using Ito-Doeblin formula. Let f (t, x) = eµt x.
Then
∂f ∂f ∂2 f
= µeµt x, = eµt x, =0.
∂t ∂x ∂x2
Now using the second version of the Ito-Doeblin formula, we get
, t
−µt
X(t) = e x0 + σ e−µ(t−s) dW(s) . (9.33)
0

This stochastic process {X(t), t ≥ 0} is called a Ornstein-Uhlenbeck process. Equa-

tion (9.33) represents the strong solution of the SDE (9.32).
Several other important SDE’s, particularly those occurring in interest rate
modeling, will be presented in Chapter 11.

9.9 Summary and Additional Notes

• An example of a function which is continuous everywhere but nowhere differ-
entiable was given by Weierstrass in 1872. This famous example is

∞
cos(3n t)
f (t) = , (0 ≤ t ≤ 2π) .
n=1
2n
306 Financial Mathematics: An Introduction

• This chapter gives a brief description of certain basic concepts and important
results of stochastic calculus keeping finance in view. By ignoring the various
technical conditions that are required to make our definitions rigorous, these
concepts are discussed.
• This chapter starts with variation of real-valued functions and then variation
of Brownian motion in Sections 9.2 and 9.3 respectively.
• In Section 9.5, stochastic integral is introduced and its properties are discussed
with few examples. For results concerning the existence of the general Ito
stochastic integral we may refer to Mikosch [97].
• Kiyosi Ito developed Ito formula in 1951 and that is how these calculus rules
were referred in earlier texts and research papers. Independently the same was
studied by Wolfgang Doeblin before him, although Doeblin’s work remained
secret, hidden away in the safe of the French Academy of Science. In May
2000, the sealed envelope sent in February 1940 by Doeblin was finally opened.
In recognition of the Doeblin’s work, the Ito’s formula is now referred as Ito-
Doeblin formula. Interested readers may refer to Vassiliou [144] for these his-
torical details.
• Ito-Doeblin formula with two versions are presented and illustrated with some
examples in Section 9.6. The proofs of various versions of the Ito formula can
be found in standard textbooks on stochastic calculus, for example, Karatzas
and Shreve [75] or Shreve [122].
• The counter part of stochastic integral, stochastic differential equation is pre-
sented in Section 9.7. SDE’s which admit an explicit solution are the exception
from the rule. Therefore using numerical techniques, the approximation of the
solution to a SDE can be obtained. Such an approximation is called a numerical
solution. Numerical solutions allow us to simulate the sample paths which is
the basis for Monte Carlo techniques. In this text book, we have discussed only
exact and closed form solutions of stochastic differential equations and Monte
Carlo simulation will be discussed in a later chapter.
• Some important SDE’s which are frequently occurring in finance are presented
with their solutions in Section 9.8.
• We may want to allow for the possibility that a stock price can experience
sudden jumps. It may be useful to have models that incorporate jump processes
and can be studied using Ito’rule for jump processes. The topic of jump process
is not presented in this text book. Option pricing with jumps is studied in
Merton [93].
• Stochastic calculus for processes with jumps is very important in financial appli-
cations. But as this chapter is only introductory in nature, we have avoided any
 Exercises 307

discussion on this important topic because it requires a high degree of math-

ematical maturity. Interested readers may refer to Shreve [122] and Karatzas
and Shreve [75] in this regard.
• In many economic processes, volatility of the stock may itself be a stochastic
process changing randomly over time. This flexibility produces more realis-
tic models for pricing options. To study the models with stochastic volatility,
interested readers may refer to Hull and White [66] and Heston [62].

9.10 Exercises
Exercise 9.1 Let f : [−1, 1] → R be given by
⎧

⎪
⎪ 1
⎨ x sin , x0
f (x) = ⎪
⎪ x
⎩ 0, x=0.
Is f continuous? Is f of bounded variations? What will be your answer if for x  0,


2 1
f (x) is taken as x sin ?
x
Exercise 9.2 Let Y1 , Y2 , . . . be independent random variables each taking *n two val-
ues +1 and -1 with equal probabilities. Define X(0) = 0 and X(n) = j=1 Y j , (n =
1, 2, . . .). This stochastic process {X(n), n = 0, 1, 2 . . .} is a symmetric random walk.
Show that, the quadratic variation of X(n) up to k is k, i.e. [X(n), X(n)](k) = k.
Exercise 9.3 For a Poisson process {X(t), t ≥ 0} with rate 1, find
: t 
(i) E 0 X(s) dW(s)
: t 
(ii) Var 0 X(s) dW(s) .
Exercise 9.4 Find the stochastic differentials of sin(W(t)) and cos(W(t)).

Exercise 9.5 Show that the process {X(t), t ≥ 0} given by

,
1 t W(s)
X(t) = −1 + e W(t)
− e ds
2 0
is a martingale.
Exercise
:t 9.6 Let h(s) be a real-valued function which is differentiable and such
that 0 h (s)ds < ∞.
2
308 Financial Mathematics: An Introduction

(i) Show that h(s) is Ito-integrable.

(ii) Use Ito formula to prove the identity
, t , t

h(s) dW(s) = h(t) W(t) − h (s) W(s) ds .
0 0
:t
(iii) Find the distribution of 0
h(s) dW(s).
Exercise 9.7 Consider the SDE of the form

dX(t) = µ dt + σ dW(t), X(0) = x .

Find a deterministic function A(t) such that exp(X(t) + A(t)) is a martingale.

Exercise 9.8 Consider the SDE of the form

dX(t) = −µX(t) dt + σ dW(t) ,

where X(0), µ and σ > 0 are constants. Find the strong solution of the above SDE.
Also, find the distribution of X(t).
Exercise 9.9 Prove that , t
I(t) = X(s) dW(s)
0
is a martingale.
Exercise 9.10 Prove that , T
W(T) = dW(t)
0
is an Ito process.
Exercise 9.11 Consider the SDE of the form dX(t) = X(t) dW(t) with X(0) = 1.
t
Prove that its solution X(t) = eW(t)− 2 is an Ito process.
Exercise 9.12 Find the stochastic differential of W 2 (t) and show that W 2 (t) is an
Ito process.
Exercise
, T 9.13 Using the first version of Ito-Doeblin formula, to evaluate
W 2 (t)dW(t).
0
 Exercises 309

:T :T
Exercise 9.14 Are the random variables 0 t dW(t) and 0 W(t) dt independent?
Also, find the mean and variance of these random variables.
Exercise 9.15 An option is called digital option if the pay-off is 1 for S(T) > S(0)
at the time of exercise T, and zero otherwise. Find the arbitrage free price of a
digital option (European) with strike price K = S(0). You may assume that the
stock price follows the SDE.
dS(t) = r S(t) dt + σ S(t) dW(t) ,
where r is the interest rate and W(t) is the Brownian motion under risk neutral
probability measure.
Exercise 9.16 Consider the SDE
dX(t) = c(t) X(t) dt + σ(t) X(t) dW(t), t ∈ [0, T] .
Using the second version of Ito-Doeblin formula, prove that, the solution is

, t
 , t 8
1 2
X(t) = X(0) exp c(s) − σ (s) ds + σ(s) dW(s) , t ∈ [0, T] .
0 2 0

Exercise 9.17 Consider the SDE

dX(t) = A(t) dt + B(t) dW(t), X(0) = x ,
where A(t) and B(t) are two time-dependent functions. Find A(t) such that Z(t) =
exp(X(t)) is an exponential martingale?
Exercise 9.18 Let µn (t) be the n-th order moment about the origin for the Brow-
nian motion {W(t), t ≥ 0}. Using Ito-Doeblin formula, prove that
, t
1
µn (t) = n(n − 1) µn−2 (t), (n = 2, 3, . . .) .
2 0

Also, deduce that µ4 (t) = 3t2 .

Exercise 9.19 Consider the SDE
X(t)
dX(t) = − dt + dW(t), 0 ≤ t < 1 ,
1−t
with X(0) = 0. Prove that it’s solution
, t
1
X(t) = (1 − t) dW(s), 0 ≤ t < 1
0 1−s

is a Brownian bridge, between time 0 and time 1.

310 Financial Mathematics: An Introduction

Exercise 9.20 Consider the SDE

, t
X(t) = x0 + sgn(X(s)) dW(s) .
0

Prove that, it has a weak solution, but does not have a strong solution. Further,
prove that, when x0 = 0, the weak solution of X(t) is implicitly given by
, t
W(t) = sgn(X(s)) dX(s) .
0

Exercise 9.21 Consider the SDE

dX(t) = X(t)dt + dW(t) ,

with initial condition X(o) = c. Obtain the strong solution of X(t). Prove that,
, t
X(t) = c e + e
t t
e−s dW(t) .
0
 
Exercise 9.22 Let Q(t) = ln S(t) and dQ(t) = µ − 21 σ2 dt + σ dW(t). Find dS(t).
10
Black-Scholes Formula Revisited

10.1 Introduction
It is seldom a surprise to find in literature that the approach taken by mathe-
maticians to study concepts of finance is quite different from the one adopted
by financial economists. The two groups speak different languages and take their
own routes to develop theories for financial mathematics, but arrive at the same
conclusions. Nothing can illustrate our point better than the theory of derivative
pricing. We have already seen that one can resort to any of the several methods for
derivative pricing. For instance, for European options pricing, no matter whether
we follow the binomial lattice approach or the CRR approach, we arrive at the
same Black-Scholes (BS) formula. The two approaches have been fairly simple to
understand, and hence favorite with many.
In this chapter we aim to introduce the readers to a fascinating though com-
plex world of change of probability measure (one can recollect how the method
of substitution works so well in the classical Riemann integral calculus). We shall
also see how this new concept in stochastic integration help us to derive the BS
formula for option pricing. But wait! why should we derive the BS formula again
and that too with an altogether new theory? We would like to emphasize here
that our aim is not to derive the BS formula only, but more importantly, to get
familiarize with a more formal mathematical approach in financial instruments
pricing. On first reading the initial part of this chapter one may feel like asking,
what is really going on! We agree with you that the discussion is more mathemat-
ical than financial but we all agree that it is useful sometimes, even for financial
economists, to recognize and appreciate what the mathematicians are saying.
Our primary aim is to present two important theorems from stochastic calcu-
lus namely, the Girsanov’s theorem and the Feynman-Kac theorem. It is worth
mentioning that both the theorems provide glimpse of confluence of mathematical
312 Financial Mathematics: An Introduction

theory with financial theory. The theorems initially appear to be a somewhat com-
plex rules but enable us to perform change of probability measure in stochastic
processes. The same will subsequently be applied to obtain the BS formula. The
two theorems provide a rigorous analytical framework to build not only the BS
formula but many other financial theories. They also open the gates to look for
advance theories in probability measure theory and partial differential equations
from the perspective of financial concepts. We urge the readers to go through the
entire chapter to appreciate the presented approach.

10.2 Change of Probability Measure

We initiate our discussion by first describing how a random variable behaves when
an original probability measure is changed to an equivalent probability measure,
in fact we need to define what we meant by an equivalent probability measure.
In order to facilitate understanding of otherwise a difficult concept, we first
consider a very simple situation as follows.
Let Ω be a finite sample space, specifically say Ω = {ω1 , ω2 , ω3 }. A σ-field F of
Ω is simply the power set P(Ω). The cardinality of F is 8, and we write it as F =
{∅, A1 , A2 , A3 , B1 , B2 , B3 , Ω}, where Ai = {ωi }, i = 1, 2, 3, and B1 = {ω1 , ω2 }, B2 =
{ω1 , ω3 }, B3 = {ω2 , ω3 }. In order to speak of change of measure we need to consider
at least two probability measures defined on the same measure space (Ω, F ). Let
us denote these probability measures by P and Q. Thus, we have two probability
spaces (Ω, F , P) and (Ω, F , Q). We aim to spell out what we mean by changing
from one measure P to another measure Q. Suppose * we define P and Q respectively
as P(∅) = 0, P(Ai ) = pi , pi > 0, i = 1, 2, 3, 3i=1 pi = 1, and Q(∅) = 0, Q(Ai ) =
*
qi , qi > 0, i = 1, 2, 3, 3i=1 qi = 1. It is worth noting at the outset that both P and
Q agree for the null event ∅, and sure event Ω. Let us now explore a relationship
between P and Q on other events in F . We note that
qi
Q(Ai ) = qi = P(Ai ), i = 1, 2, 3. (10.1)
pi
q1 q2
Q(B1 ) = q1 + q2 = p1 + p2
p1 p2
 q1 p1 q 2 p2 
= + P(B1 ). (10.2)
p1 p1 + p2 p2 p1 + p2

Similar expressions can be obtained for Q(B2 ) and Q(B3 ).

 Change of Probability Measure 313

qi
We now define a non-negative random variable Z on Ω as, Z(ωi ) = , i = 1, 2, 3.
pi
Then Z can be written as

3
qi
Z(ω) = 1Ai (ω),
i=1
pi
where 1Ai is the indicator function of the set Ai (which takes value 1 only when
ω = ωi else value 0). With this Z, note that we can rewrite (10.1) and (10.2) as
follows

Q(Ai ) = EP (Z1Ai ) = EP (Z/Ai )P(Ai ), i = 1, 2, 3,

Q(Bi ) = EP (Z1Bi ) = EP (Z/Bi )P(Bi ), i = 1, 2, 3,

where EP denotes the expectation operator with respect to measure P, and EP (Z/
C) denotes the conditional expectation of Z given C ∈ F with respect to P.
Since EP (Z) = EP (Z/Ω) = 1, we have, in general,

Q(C) = EP (Z1C ) = EP (Z/C)P(C), for any C ∈ F .

Now, let us define a natural random variable Y on Ω as Y(ωi ) = yi , i = 1, 2, 3.

Obviously Y has two alternative probability distributions on Ω depending on
probability measure P or Q. We can now give an expression for the expectation
of Y under measure in terms of and P, Q, Z as follows.

3 
3 
3
qi
EQ (Y) = Y(ωi )Q(ωi ) = yi qi = yi pi = EP (ZY).
i=1 i=1 i=1
pi

Finally, we have, EQ (Y) = EP (ZY), for any random variable Y on Ω. The random
variable Z can be thought of as a constructed random variable such that Z(ω) =
Q(ω)
, ω ∈ Ω. Note that we have specifically assumed P(ω) > 0 in the above
P(ω)
discussion to avoid division by zero. However, it is not hard to overcome this
point as we can also write Z(ω)P(ω) = Q(ω). Note that if Z(ω) > 1 (or Z(ω) < 1)
for some ω ∈ Ω, then Q(ω) > P(ω) (or Q(ω) < P(ω)), that is, the probability
P(ω) is revised upwards
* (or downwards). Also, for Q to be a probability measure,
we must have, ω∈Ω Z(ω)P(ω) = 1. Thus the random variable Z is acting as a
revision factor such that EP (Z) = 1.
We now have a set-up to build on and extend the above idea to an arbitrary
set Ω, which may be uncountably infinite. But before that, we define equivalent
measures.
314 Financial Mathematics: An Introduction

Let (Ω, F , P) be :a measure space and Z a non-negative F -measurable stochastic

variable such that ω Z(ω)dP(ω) = 1. We define a new probability measure Q on
Ω by dQ(ω) = Z(ω)dP(ω), that is
, ,
X(ω)dQ(ω) = X(ω)Z(ω)dP(ω),
Ω Ω

for all F -measurable functions X such that the integrals exist. In this case,
(Ω, F , Q) is also a measure space, but in general Q(A)  P(A), A ∈ F . Note
that if P(A) = 0, then 1A (ω)Z(ω) = 0 almost surely with respect to P, implying,
Q(A) = 0. Opposite is not true. For instance, if Z(ω) = 0 for ω ∈ A where P(A) > 0
then Q(A) = 0 despite P(A)  0. But if Z(ω) > 0, ∀ ω ∈ Ω, then it works both
ways.
Definition 10.2.1 (Equivalent measure) Let (Ω, F ) be a measurable space,
that is, F is a σ-field on Ω. Consider two probability measure P and Q on F . We
say that Q is absolutely continuous with respect to P (denoted by Q  P) if

P(A) = 0 ⇒ Q(A) = 0, ∀ A ∈ F .

If P  Q and Q  P, then P and Q are called equivalent measures.

The objective of change of probability measure is typically to obtain a more
favorable probability measure to work with. The vital question is that when an
equivalent measure exists? The answer lies in the following theorem known as the
Radon-Nikodym theorem. We skip its proof but interested readers can find the
same in [67].
Theorem 10.2.1 Let P be a probability measure on measurable space (Ω, F ),
and let Q be a finite measure on (Ω, F ). If Q  P then there exists a nonnegative
random variable Z such that Q(A) = EP (Z1A ), ∀ A ∈ F .
The variable Z is P-unique almost surely and it is referred to as the Radon-
Nikodym derivative of Q with respect to P. If there is no confusion in symbol, we
dQ
denote it by Z = .
dP
We now make some important observations regarding the Radon-Nikodym the-
orem.
(i) Since Q is a finite measure, that is Q(ω) is finite, it can be normalized by
Q(A)
. So, Q is a probability measure on Ω. The normalization also implies
Q(Ω)
that EP (Z) = 1.
 Change of Probability Measure 315

(ii) Z is P-unique almost surely means that if there is another version of Z, say
Z∗ , then we have P({ω ∈ Ω : Z(ω) = Z∗ (ω)}) = 1. That is the set where Z and
Z∗ do not agree has a measure zero with respect to P.
(iii) The Radon-Nikodym states the existence of Z without providing an explicit
expression for Z. In a practical situation, Z is estimated by a series of obser-
vations or simulation of sufficiently large sample size.
(iv) The expectations of any random variable X : Ω → Rn under the two proba-
bility measures are related by
, ,
f (ω)dQ(ω) = f (ω)Z(ω)dP(ω), A ∈ F .
ω∈A ω∈A

The above theorem is significant for financial engineers as it provides a frame-

work to switch from real market probability measure P to a risk neutral probability
measure (RNPM) Q = P ( world. Though finding a RNPM P ( itself is challenging, yet
it is its very existence that ensures arbitrage free market. The latter assumption
helps to develop theories for general equilibrium pricing of products and services
in markets. The same had been witnessed in first few chapters of the book.
Let us illustrate the concept of the Radon-Nikoym derivative through some
simple examples.
Example 10.2.1 Let X be a random variable on a (R, B, P) where B is a Borel
field of R. Let the density function of X with respect to P be fX . We assume that
fX (x) is not zero for any x ∈ R. Consider a transformation of X to a random
variable Y according to Y = g(X), where g(·) is a strictly increasing differentiable
function on R. Find the Radon-Nikodym derivative.
Solution From basic probability theory, we know
FY (y) = P(Y ≤ y)
= P(g(X) ≤ y)
= P(X ≤ g−1 (y))
= FX (g−1 (y)).
Differentiate both sides with respect to y, we get
1
fY (y) = fX (g−1 (y))  ,
g (g−1 (y))
d
where fY is the density function of Y. Here we have used that fY (y) = (FY (y)).
dy
Define a random variable Z as follows
316 Financial Mathematics: An Introduction


fY (g(X))g (X)
Z= .
fX (X)

Then Z ≥ 0, Ep (Z) = 1, and

, ,
fY (y)dy = Z fX (x)dx.
A A

Thus Z is a Radon-Nikodym derivative transforming X into Y.



Example 10.2.2 Suppose a random variable X on (R, B, P), where B is a Borel
field of R, has an exponential distribution according to P measure with parameter
1
λ. The density function of X is fX (x) = λe−λx , with mean . Let a random variable
λ
Z(x) = ξekx be the Radon-Nikodym derivative of X such that the distribution of X
in the new probability Q remains exponential. Find Z.
Solution We wish Z to be non-negative and E(Z) = 1. For these two to happen,
ξ ≥ 0, and in fact
1 λ−k
ξ= = ,
MX (k) λ
where MX (k) is the moment generating function of X evaluated at k. Thus, λ ≥ k.
Also, note that

ekx
Z(x) fX (x) = λe−λx
MX (k)
λ − k
= ekx λe−λx
λ
= (λ − k)e−(λ−k)x .

Thus, X remains an exponentially distributed random variable under Q-measure

with parameter λ − k.


Example 10.2.3 Consider two measure spaces (R, B, P) and (R, B, Q), where B
is a Borel field of R. Let X be a N(0, 1) random variable on (R, B, P) with induced
probability measure given by
,
1 1 2
µX (A) =
P
√ e− 2 x dx, A ∈ B.
A 2π
 Change of Probability Measure 317

For an arbitrary constant θ, construct an induced measure µQ

X
, so that (X + θ) ∼
Q
N(0, 1) random variable with respect to µX .
Solution Note (X + θ) ∼ N(0, 1) means X ∼ N(−θ, 1). So, we wish X to be
N(−θ, 1) with respect to µQ X
. Let f P (x) be a probability density function of X
with respect to probability measure P. Then,
,
µX (A) =
P
f P (x)dx, A ∈ B.
A

Applying Theorem 10.2.1, we have

:
µQ (A) = Z(x)dµPX (x)
X :A
= Z(x)1A (x)dµPX (x)
:R
= Z(x)1A (x) f P (x)dx
:R
= A
Z(x) f P (x)dx.
: :
Now, µQX
(x) = A
dµQ
X
(x) = A
f Q (x)dx, where f Q (x) is a probability density function
of X with respect to measure Q. From above two relations, it is clear that
, ,
f (x)dx =
Q
Z(x) f P (x)dx, A ∈ B.
A A

The Radon-Nikodym derivative is given by

f Q (x) e−(x+θ) /2
2
1 2
Z(x) = P = −x2 /2 = e−θx− 2 θ .
f (x) e
The new measure is thus defined as follows
,
1 2
Q
µX (A) = e−θx− 2 θ dµPX (x), A ∈ B.
A


Example 10.2.3 plays motivating role in developing theory for change of mea-
sure, in general. What we have seen is that if a random variable X ∼ N(0, 1)
with respect to probability measure P then (X + θ) ∼ N(0, 1) with respect
: to an
equivalent probability measure Q, for any constant θ, where Q(A) = A Z(x)dP(x).
We now wish to see how the above description of change of measure on random
variable can be extended to a stochastic processes. In the section to follow we shall
throw more light on this aspect. We shall see how a stochastic process with a non-
zero drift can be transformed to another stochastic process with either drift zero
or decreasing.
318 Financial Mathematics: An Introduction

10.3 Girsanov Theorem

Definition 10.3.1 (Radon-Nikoym derivative process) Let Z be a random
variable on space
: (Ω, F , P) such that Z(ω) > 0 almost surely with respect to P,
and EP (Z) = Ω Z(ω)dP(ω) = 1. The Randon-Nikodym derivative process {Z(t), 0 ≤
t ≤ T} is defined by
Z(t) = EP (Z/Ft ), t ∈ [0, T].
Define a new probability measure Q on (Ω, F ) by dQ = ZdP, that is,
,
Q(A) = Z(ω)dP(ω) = EP (Z(ω)1A (ω)) = EP (EP (Z/Fω )1A (ω)), A ∈ F .
ω∈A

We simply write it as
Q(A) = EP (EP (Z/F )1A ), A ∈ F.
The much awaited Girsanov theorem is stated as follows. The result was first
proved by R. H. Cameron and W. T. Martin in the 1940’s and by I. V. Girsanov
in 1960.
Theorem 10.3.1 Let (Ω, F , P) be a probability space. Let {W(t), 0 ≤ t ≤ T} be a
Brownian motion with associated filtration {Ft , 0 ≤ t ≤ T}. Let {θ(t), 0 ≤ t ≤ T} be
an adapted measurable process adapted to the filtration {Ft , 0 ≤ t ≤ T}. Define
,
:t 1 t 2
− 0 θ(t)dW(t)− θ (t)dt
Z(t) = e 2 0 , t ∈ [0, T].
Let Z = Z(T). A sufficient condition, known as the Novikov’s condition, for Z to
be a martingale is that , T
 1 θ2 (t)dt 
E e2 0 < ∞.

Assume Z is a P-martingale. Define a new measure Q on (Ω, F ) as

Q(A) = EP (ZIA ).
(
Then the process {W(t), 0 ≤ t ≤ T}, defined by
, t
(
W(t) = W(t) + θ(t)dt,
0

is a Brownian motion on (Ω, F , Q).

 Girsanov Theorem 319

Proof. We first observe that Z(t) is a martingale under P. because Z(t) is a

particular case of the generalized Brownian motion defined by

: :t 
t
σ(t)dW(t)− 0 (α(t)− 12 σ2 (t))dt
Z(t) = Z(0)e 0
,

with α(t) = 0, Z(0) = 1, and the generalized Brownian motion is a martingale.

Consider 0 ≤ s ≤ t ≤ T. We have

EP (Z(t)/Fs ) = EP (EP (Z(T)/Ft )/Fs ) = EP (Z/Fs ) = Z(s),

and EP (Z) = EP (Z(T)) = 1. Thus {Z(t), 0 ≤ t ≤ T} is the Radon-Nikodym process.

Moreover, if {Y(t), 0 ≤ t ≤ T} is any Ft -measurable process, then

EQ (Y) = EP (YZ)
= EP (Y(t)Z(t)/Ft )
= EP (YEP (Z(t)/Ft ))
= EP (YZ(t)).

Hence, Q is equivalent measure to P. The probability measure Q is called an

equivalent martingale measure.
(
Next we show that W(t)Z(t) is a martingale under P.

(
d(W(t)Z(t)) ( + W(t)dZ(t)
= Z(t)dW(t) ( (
+ dZ(t)dW(t)
(
= Z(t)(dW(t) + θ(t)dt) + W(t)(−θ(t)Z(t)dW(t))
+(−θ(t)Z(t)dW(t))(dW(t) + θ(t)dt)
(
= Z(t)dW(t) − θ(t)W(t)Z(t)dW(t)
(
= (1 − θ(t)W(t))Z(t)dW(t),

where in the second last relation we have used that dW(t)dW(t) = dt and dW(t)dt =
0. Therefore,
, t , t
(
W(t)Z(t) (
= W(0)Z(0) + Z(t)dW(t) − (
θ(t)W(t)Z(t)dW(t).
0 0

The two Ito integrals on the right hand side of the above expression are
(
martingale, so the process W(t)Z(t) is a martingale under P. Therefore, for any
0 ≤ s ≤ t ≤ T, we have,
320 Financial Mathematics: An Introduction

(
EP (W(t)Z(t)/F (
s ) = W(s)Z(s).

Now, for 0 ≤ s ≤ t ≤ T,
(
EQ (W(t)/F (
(s)) = EQ (1A W(t)), A ∈ Fs
(
= EP (1A W(t)Z(t))
(
= EP (1A EP (W(t)Z/F s ))
1 (
= EP ((1A EP (W(t)Z(t)/F s ))Z(s))
Z(s)
1 (
= EQ (1A EP (W(t)Z(t)/Fs ))
Z(s)
1 (
= EP (W(t)Z(t)/F s)
Z(s)
1 (
= W(s)Z(s)
Z(s)
(
= W(s).
( is a martingale under probability measure Q.
Hence, W(t)
(
Invoking Levy’s theorem (Theorem 9.7.2), with notices that dW(t)d ( = dt,
W(t)
:t
and 0 θ(t), being a Riemann integration, is a continuous function of t, it follows
( is a Brownian motion. The process {W(t),
that W(t) ( 0 ≤ t ≤ T} is adapted to the
filtration {Ft , t ≥ 0}.



10.4 Discounted Portfolio Process

We attempt to build a general theory for derivative security which shall be used
in the subsequent section to derive BS formula for option pricing. Our discussion
in this section is centered around the Girsonov’s theorem.
Let the stock having a price S(t) per unit follows a generalized geometric Brow-
nian motion with a constant mean return µ and a constant volatility σ > 0. The
price process is governed by the linear SDE

dS(t) = µS(t)dt + σS(t)dW(t), t ∈ [0, T]. (10.3)

We also let β(t) to be the price of some risk-free asset which satisfies the fol-
lowing ordinary differential equation
 Discounted Portfolio Process 321

dβ(t) = rβ(t)dt, (10.4)

where r is the constant risk-free interest rate.

Suppose at time t we take a portfolio consisting of a(t) shares of stock and b(t)
shares of risk-free asset. Let V(t) be the value of this portfolio at t, that is,

V(t) = a(t)S(t) + b(t)β(t), t ∈ [0, T].

Then,
dV(t) = a(t)dS(t) + b(t)dβ(t).
The discounted price of one share of stock is (
S(t) = e−rt S(t), t ∈ [0, T]. Apply the
Ito’s Lemma given at (9.20) on it, we have,

d(
S(t) = −re−rt S(t)dt + e−rt dS(t)
= −re−rt S(t)dt + e−rt S(t)(µdt + σdW(t))
=(S(t)((µ − r)dt + σdW(t))
= σ( (
S(t)dW(t) (10.5)
µ−r
where we denote by W(t)( = W(t) + t, t ∈ [0, T]. A natural interpretation for
σ
µ−r
is that the numerator is the expected return minus the risk-free rate, or the
σ
risk premium while the denominator is risk. The ratio is the risk premium per
unit of risk and is called the market price of risk . In other words, the ratio reflects
the additional expected return necessary to induce investors to take risk. Refresh
that in the study of general market equilibrium such as the capital asset pricing
model (CAPM), the appropriate risk is the systematic risk beta. It is important
to note that in (10.5), the drift term is completely removed. When we remove the
drift, what we are doing is removing the risk premium and the risk-free rate.
From the Girsanov’s
: theorem (Theorem 10.2.1), there exists an equivalent mea-
sure Q, Q(A) = A Z(T)dP, A ∈ F , and Z is a Radon-Nikodym derivative of Q
( into the Brownian motion with respect to Q.
with respect to P, which turns W(t)
Note that, EP (W(t)) = 0, on account of W(t) being a Brownian motion under P,
µ−r
(
hence, EP (W(t)) = ( is not a Brownian motion under P but it is
t. Thus, W(t)
σ
a Brownian motion under Q. Furthermore, solution of equation (10.5), given by,

−σ2 (
t + σW(t)
(
S(t) = (
S(0)e 2 , t ∈ [0, T],
322 Financial Mathematics: An Introduction

is a martingale under Q. This measure Q is nothing else but the risk neutral
probability measure (RNPM) (the probability measure in which the discounted
( as a convention.
price process becomes a martingale). We shall be denoting Q by P,
To summarize, we have to adjust the drift of the stock price process by changing
the probability measures such that we obtain a martingale.
Theorem 10.4.1 There exist two probability measures namely the Wiener mea-
( such that the stock price S(t) at time t, has
sure P and the risk neutral measure P,
the following properties:
(i) In probability measure P, S(t) = S(0)eµt+σW(t) .
( S(t) = S(0)e(r− σ2 )t+σW(t) (
2
(
(ii) In probability measure P, , and {W(t), t ≥ 0} is a Wiener
process.
Consider the discounted portfolio value process
( = e−rt V(t) = e−rt a(t)S(t) + b(t)β(t) .
V(t)

Then,
( = −re−rt V(t)dt + e−rt dV(t)
dV(t)
 
= −re−rt a(t)S(t) + b(t)β(t) dt + e−rt (a(t)dS(t) + b(t)dβ(t))
 
= a(t)e−rt (−rS(t)dt + dS(t)) + b(t)e−rt −rβ(t)dt + dβ(t)
= a(t)d(
S(t),

where the last relation follows on account of (10.4) and definition of ( S(t). Thus,
the increment in the value of the discounted portfolio process {V(t), 0 ≤ t ≤ T} is
coming from the discounted stock price process {S(t), 0 ≤ t ≤ T}. Also, note that
( = V(0). We thus have
V(0)
, t
(
V(t) = V(0) + a(t)d(
S(t)
0
, t
= V(0) + σ a(t)( (
S(t)dW(t).
0

(
Since W(t) ( and a(t)(
is a Brownian motion under P S(t) is an adapted process to
( constitutes a martingale. Therefore,
Ft , t ∈ [0, T], hence V(t)
 
V(t) (
( = E( V(T)/F t , t ∈ [0, T],
P
 Risk Neutral Pricing Formula 323

implying
 
V(t) = ert EP( e−rT V(T)/Ft , t ∈ [0, T]. (10.6)
 
It is easy to note that V(0) = EP( e−rT V(T) , because F0 = {∅, Ω}.

10.5 Risk Neutral Pricing Formula

Recall, how, in earlier part of this book, we have used risk neutral probability
measure (RNPM) in replication and thereafter pricing a derivative security. The
basic philosophy behind replication is that there exists a portfolio which can hedge
a short position on the derivative security on the stock priced at S(t). The second
fundamental theorem of asset pricing guarantees the existence of such a replicating
(hedging) portfolio.
For the sake of completeness we state, with a partial proof, the aforementioned
theorem.
Theorem 10.5.1 Suppose the market model has a risk neutral probability measure
(RNPM) (in other words, market follows no arbitrage principle). The market is
complete (that is, every derivative security can be hedged by some portfolio) if and
only if the RNPM is unique.
Proof. Suppose the market is complete. Let there be two RNPMs in the market,
(1 and P
say, P (2 . Define the value of the derivative security as

1
X(T) = 1A , A ∈ FT = F ,
D(T)

where {D(t), 0 ≤ t ≤ T} is a discounted price process (a stochastic process analo-

gous to e−rt ) adapted to filtration {Ft , 0 ≤ t ≤ T}. Under assumption of complete
market, this portfolio can be hedged; there exists a portfolio V(t), t ∈ [0, T], such
that X(T) = V(T). Also, the discounted value of portfolio, V(t)( = D(t)V(t) is a
martingale under RNPM. We encourage , readers to verify the same for a gen-
t
eral discount process, D(t) = exp(− R(t)dt), where {R(t), 0 ≤ t ≤ T} is the
0
interest rate process adapted to filtration {Ft , 0 ≤ t ≤ T}, on the similar lines as
explained in previous section with a particular choice of constant rate of interest
R(t) = r, t ∈ [0, T].
Continuing our discussion, we have,
324 Financial Mathematics: An Introduction

(
EP(1 (V(T)) = V(0),
(
EP(2 (V(T)) = V(0).

Consequently, because of the portfolio hedging, we have

(
EP(1 (X(T)) = EP(1 (1A ) = V(0)
(
EP(2 (X(T)) = EP(2 (1A ) = V(0).

Finally yielding that P (2 (A), for allA ∈ F , thereby means, P

(1 (A) = P (1 = P
(2 . Thus
in a complete market RNPM is unique.


From now onwards, we shall assume that a short position in the derivative
security can be hedged by some portfolio.
Let X(T) be a payoff on maturity from a derivative security written on a stock
whose price process is {S(t), 0 ≤ t ≤ T}, and let V(t), t ∈ [0, T], be the value of
the associated hedging portfolio. That is, we can construct a portfolio compris-
ing shares of the underline stock and risk-free asset. By hedging mechanism (or
replicating), we have,
X(T) = V(T).
For fair price of the derivative security (no arbitrage), we must have,

X(t) = V(t), for all t ∈ [0, T). (10.7)

From (10.6), we have

e−rt V(t) = EP((e−rT V(T)/Ft ), t ∈ [0, T],

which, in view of (10.7), yields

X(t) = ert EP((e−rT X(T)/Ft ), t ∈ [0, T]. (10.8)

The above is called the risk neutral pricing formula. The applicability of the
risk neutral pricing formula is much wider than only the BS formula as many of the
assumptions made in say CRR model had been dropped here and we are working
in a more general scenario for any derivative security which can be hedged.
Furthermore, the price of derivative security at t = 0 is
X(0) = EP((e−rT X(T)).
We urge you to take a pause and carefully examine the above formula. Is it not
somewhat a familiar expression that we have came across earlier too, for instance,
 Risk Neutral Pricing Formula 325

in Chapter 3; except that there we have used notations V(0) for what is X(0) here
and VRP (t) for what is V(t) in the above discussion. The readers must appreciate
that all the hard mathematics concepts finally lead to a path we are so familiar
with, giving us a hope that we are close to deriving the BS formula. The discussion
to follow is the ultimate in this context.
Consider a European call option derivative security on the stock. The payoff,
to the holder of call, on maturity is X(T) = C(T) = (S(T) − K)+ , K is the strike
price of call.
C(t) = ert EP((e−rT C(T)/Ft ), t ∈ [0, T]
−r(T−t) +
= EP((e (S(T) − K) /Ft ). (10.9)
Since S(t) is a random process following the generalized Brownian motion under
(
P, we have,
(
S(t) = S(0)eσW(t)+(r−(σ /2))t ,
2
t ∈ [0, T].
In other words,
( (
S(T) = S(t)eσ(W(T)−W(t))+(r−(σ /2))(T−t) ,
2
t ∈ [0, T].
Let ς = T − t. Then,
( (
C(t) = EP((e−rς (S(t)eσ(W(T)−W(t))+(r−(σ /2))ς − K)+ /Ft ).
2

(
Now, W(T) (
− W(t) is the increment in the Brownian motion for remaining
time ς, so it is independent of Ft . Moreover S(t) is Ft -measurable. Thus, the
conditional expectation in immediate above formula is superfluous. Furthermore,
(
W(T) ( ∼ N(0, τ). Setting
− W(t)
(
W(T) (
− W(t)
−Y = √ ∼ N(0, 1),
T−t
we get,
 √ 
C(t) = EP( e−rς (S(t)eσ ς(−Y)+(r−(σ /2))τ − K)+
2

, ∞
1 √ 1 2
e−rς (S(t)eσ ς(−y)+(r−(σ /2))τ − K)+ e− 2 y dy.
2
= √
2π −∞
From now onwards, mimic the argument used in the proof presented in Section
4.3, Chapter 4 to get the Black-scholes formula for European Option. We skip its
redoing here.
326 Financial Mathematics: An Introduction

10.6 Feynman-Kac Theorem and BS Formula

Let us now turn our attention to present another famous and significant theorem
in stochastic of finance. The Feynman-Kac theorem (formula), named after R.
Feynman and M. Kac, establishes a link between parabolic partial differential
equation (PDE) and stochastic processes.
Theorem 10.6.1 Let the stochastic process {X(t), 0 ≤ t ≤ T} satisfy the following
SDE
dX(t) = µ(t, X(t))dt + σ(t, X(t))dW(t), (10.10)
where µ(t, X(t)) and σ(t, X(t)) are smooth functions on [0, T] × R, called the drift
and diffusion functions respectively. In addition, let the initial condition on X(t)
be X(0) = x, for some x ∈ R. Then the solution of the following PDE
1
gt (t, x) + µ(t, x)gx (t, x) + σ2 (t, x)gxx (t, X) − rg(t, x) = 0, (10.11)
2
subject to the boundary condition g(T, X(T) = x) = h(x), x ∈ R, is a function
g : [0, T] × R → R given by

g(t, x) = E(e−r(T−t) h(X(T))/X(t) = x). (10.12)

The Feynman-Kac theorem can be used in both directions. That is, if we know
that X(t) follows the SDE (10.10) and we are given the PDE with boundary
condition, then we can always obtain the solution g(t, x) as described in (10.12).
On the other hand if we know that the solution g(t, x) is given by (10.12) and that
X(t) follows the process in (10.10), then we are assured that g(t, x) satisfies the
PDE (10.11).
The generator of the process in (10.10) is defined as the operator

∂ 1 ∂2
A = µ(t, x(t)) + σ2 (t, x(t)) 2 .
∂x 2 ∂x
The PDE (10.11) can then be rewritten as follows

∂g
+ Ag − rg = 0.
∂t
Let us now examine how we can apply the above theorem to derive the BS
formula for a derivative security.
Let the stock price S(t) be driven by the process
 Feynman-Kac Theorem and BS Formula 327

dS(t) = µS(t)dt + σS(t)dW(t).

As explained in (10.3) and (10.5), the risk neutral process is given by

(
dS(t) = rS(t)dt + σS(t)dW(t),
µ−r
( = W(t) +
where W(t) t.
σ
Suppose a derivative is written on this stock. Let V(t, S(t)) be the price of this
security at any time t ∈ [0, T), and V(T, S(T)) be its payoff on maturity. Note
V : [0, T] × R+ → R+ , R+ is the set of non negative reals. By Ito’s Lemma, we
have
 
∂V ∂V 1 2 2 ∂2 V ∂V (
dV(t) = dV(t, S(t)) = + rS(t) + σ S (t) 2 dt + σS(t) dW(t). (10.13)
∂t ∂x 2 ∂x ∂x
Suppose the derivative security can be hedged. We replicate the derivative
security by taking a portfolio in which we take a(t) shares of underline stock and
b(t) shares of a risk-free asset whose price is governed by an ordinary differential
equation (10.4). Then, we will have,

dV(t) = a(t)dS(t) + b(t)rβ(t)dt

(
= a(t)(rS(t)dt + σS(t)dW(t)) + rb(t)β(t)dt. (10.14)

From (10.13) and (10.14), it follows that

∂V ∂V 1 2 2 ∂2 V
= a(t) and + σ S (t) 2 = rb(t)β(t).
∂x ∂t 2 ∂x
∂V
Now using b(t)β(t) = V(t) − a(t)S(t) = V(t) − S(t), we finally get that the
∂x
derivative price satisfies the following Black- Scholes PDE
∂V ∂V 1 2 ∂2 V
+ rS + σ − rV = 0. (10.15)
∂t ∂S 2 ∂S2
∂ ∂ 2
The generator of the process given by A = rS ∂S + 12 σ2 ∂S 2 . By the Feynman-Kac

theorem, the time t value of the derivative is the solution of

V(t, S(t)) = e−r(T−t) EP((h(S(T))/Ft ), (10.16)

where P( is a RNPM, and h(S(T)) is payoff of the derivative security on maturity.

For instance, h(T) = (S(T)−K)+ for European call option maturing at T with strike
328 Financial Mathematics: An Introduction

price K. Note that (10.16) is same as (10.8) except that in (10.8) the notation has
been X(t) for denoting derivative price on stock whose price is S(t). Now onwards,
we can follow the discussion of Section 10.5 after relation (10.8) to get the BS
formula for price of European call option. We leave it for readers to recognize that
all approaches for derivative pricing finally lead to a similar result.
It is also a good opportunity to have a closer look at the Black-Scholes (BS)
PDE (10.16). We would like to mention here that the PDE like (10.16) has moti-
vated a community of numerical analysts to play a serious role in financial math-
ematics. Though, this PDE has a closed form solution which is nothing but the
BS formula, there are many other PDE’s which naturally arise in financial theory
which require efficient numerical methods to come up with their analytic solutions.
We sketch below the method by which the closed form solution of (10.16) can
be obtained.
Introduce the variables τ = ν(T − t) and ξ = α(ln( Kx ) + β(T − t)), where the
constants α, β, and ν will be appropriately chosen later. Define
C(t, x) = e−r(T−t) y(τ, ξ).
Then,

∂C
Ct = = re−r(T−t) y + e−r(T−t) yt
∂t
= re−r(T−t) y + e−r(T−t) (−αβyξ − νyτ ).
∂C
xCx = x = xe−r(T−t) yx = e−r(T−t) αyξ .
∂c
∂Cx ∂ α
Cxx = = (e−r(T−t) yξ ).
∂x ∂x x
The latter on some simplification yields

x2 Cxx = e−r(T−t) (α2 yξξ − αyξ ).

We left the above for the readers to verify. Since C satisfies the BS PDE,
1
Ct + rxCx + σ2 x2 Cxx − rC = 0,
2
we must have,
1
−αβyξ − νyτ + ναyξ + σ2 (α2 yξξ − αyξ ) = 0. (10.17)
2
 Summary and Additional Notes 329

Choose β and ν such that ν − β = 21 σ2 . Then, the coefficient of yξ in (10.17) is

zero. Thus, (10.17) reduces to the following equation
1
−νyτ + α2 σ2 yξξ = 0.
2
Choose ν = α2 σ2 , (α  0). Then the above PDE becomes the following heat
equation.
1
yτ = yξξ .
2
Let us take a particular instance when C(t, x) is the payoff of a European call
option with strike K and maturity T.
Now, we started with, C(t, x) = e−r(T−t) y(τ, ξ). In order to satisfy C(T, x) =
(x − K)+ , we must have that, when t = T, τ = ν(T − t) = 0 and ξ = αln( Kx ).
Moreover, we can choose any α  0, so, we take it to be equal to 1. Then, x = Keξ ,
and hence we get the boundary condition y(0, ξ) = (Keξ − K)+ . Thus, pricing the
European call option with strike K and maturity T is equivalent to solving the
following heat equation with boundary condition.
1
yτ = yξξ , y(0, ξ) = (Keξ − K)+ .
2
We leave it here itself; as to solve the heat equation one needs an appropriate
background and that is beyond the scope of our present work. But we ensure the
readers that it will indeed be the BS formula obtained in Chapter 4.

10.7 Summary and Additional Notes

• In this chapter, we have introduced the concept of change of probability mea-
sure in probability measure spaces. The idea is extremely powerful and used
in many occasions in financial theories besides alone the classical derivative
pricing. In fact the Girsanov’s theorem studied in this chapter has been also
applied in literature to price exotic options and insurance premium risks. The
Girsanov’s theorem has been extended to more general classes of processes.
Further, several variants of this theorem have appeared in literature that in-
clude its multidimensional version. The theorem has also been proved in the
Hilbert space setting. To keep the theory simple, we have focussed our discus-
sion on a basic version of the Girsanov’s theorem. Of course there is no dearth
of literature. We encourage you to search on a web crawler and many research
articles of interest will appear.
330 Financial Mathematics: An Introduction

• We have discounted the stock price and then worked with the discounted stock
price to change its probability measure to RNPM. We then evaluate the option
price by applying the probability distribution of the discounted stock price to
the option payoff on maturity. In this way the option price is its expected payoff
at expiration without discounting. Mathematicians prefer to convert the stock
price to a martingale, requiring that the discounting be done beforehand.
On the other hand, without discounting the stock price, we could have worked
(
directly with the SDE dS(t) = rS(t)dt + σS(t)dW(t), instead of dS(t) = µS(t)dt +
σS(t)dW(t) (see, (10.3)). We would then have first evaluated the expected op-
tion payoff at expiration and do the discounting thereafter only. The financial
economists prefer to follow this path. That is because it is in line with the idea
that the price of any asset is its expected future value discounted to the present
at an appropriate rate.
Whichever approach one adopts, the appreciating point is that fundamental
process of taking the expectation would not have altered except the simple
linear adjustment by e−rt . The same is depicted in (10.9).
• The other important highlight of the chapter is application of the Feynman-Kac
theorem in derivative pricing. Again, this theorem also has many variants. It
has been proved for multidimensional case as well as in variety of settings. We
again urge the interested readers to explore the abstract world of this theorem
on web.
• The chapter also emphasize that be it economists, financial engineers, hard-core
mathematicians, optimization researchers, numerical analysts, or real practi-
tioners, all have significant role to play in developing the subject of finance.
The beautiful confluence of many streams of mathematics into one is indeed
remarkable. All this make ‘finance’ awesome.

10.8 Exercises
1. Let X be a random variable such that X ∼ N(0, 1) with respect to a
probability measure P. Suppose Q is another probability measure such that
1 2
EQ (X) = EP (XG), where G = e−γX− 2 γ . Prove that X ∼ N(−γ, 1) with respect
to Q.
(Note: Changes of probability measures can be used to shift the means of
random variables.)
2. (a)Let X ∼ N(µ, 1). Define another random variable Y = X + θ. What is the
new probability measure Q such that Y, under Q, has the same distribution
 Exercises 331

as X under the original probability measure?

(b)Do the same exercise in part (a) but with X ∼ N(µ, σ2 ).

 −θ(x − µ) − 0.5θ2 
(Hint for (b): Take Z = exp ).
σ2
(Note: Question 2 is an equivalent exercise of Question 1 above).
3. Let X ∼ N(µ1 , σ21 ) under probability measure P. Suppose the Radon-Nikodym
derivative of another probability measure Q is given by

dQ σ1  (X − µ1 )2 (X − µ2 )2 
(X) = exp − .
dP σ2 2σ21 2σ22

Show that X ∼ N(µ2 , σ22 ) under Q.

(Hint: Try to work out the moment generating function of X under Q. Also
see, Klebaner [78].)
4. Let {N(t), 0 ≤ t ≤ T} be a Poisson process with rate 1 under P. For a constant
λ > 0 define Q, by

dQ
= exp((1 − λ)T + N(T)ln(λ)).
dP
Prove that {N(t), 0 ≤ t ≤ T} is a poisson process under Q with rate λ. Also
see, Klebaner [78].)
5. Let {W(t), t > 0} be a Brownian motion on probability measure space (Ω, F , P).
Find probability measure(s) Q on (Ω, F ) that is mutually absolutely continu-
ous with respect to P, and under which the following Y(t) becomes martingale
(a) Y(t) = 2dt − 3dW(t) (b) Y(t) = 4dt + dW(t), (c) Y(t) = −2dW(t).
6. Let {W(t), t > 0} be a Brownian motion with respect to a probability measure
P and associated filtration {F(t), t > 0}. If dW(t) = θdt + dW(t), then prove
that there exists a probability measure Q such that W(t) is a Brownian motion
with respect to Q.
1 2
(Hint: Take Z(t) = e−θt− 2 θ t , in the Girsanov theorem.)
(Note: The above result is a particular case of the Girsanov theorem when the
drift term is a constant θ. However, we encourage the readers to rework the
proof independently.)
7. Let Q be a risk-neutral probability measure for the investor investing in rupees.
Suppose the dollar/rupee exchange rate Y(t) obeys a stochastic differential
equation
dY(t) = µ(t)Y(t)dt + σ(t)Y(t)dW(t).
332 Financial Mathematics: An Introduction

If the risk-less rates of return for dollar investors and rupee investors are rd (t)
and rrp (t), then describe the exchange rate Y(t) under Q? (Hint: Note the drift
is rd (t) − rrp (t), then use the Girsanov theorem.)
8. Let W(t) = (W1 (t), W2 (t)), t ≤ T} , be a 2-dimensional Brownian motion on
probability measure space (Ω, F , P). Find an equivalent probability measure
Q on (Ω, F ) under which the following Y(t) becomes a martingale
    
2 1 1 dW1 (t)
Y(t) = dt + , t ≤ T.
4 1 −1 dW2 (t)

(Hint: Explore the multidimensional version of the Girsanov’s theorem and use
it.)
11
Interest Rates and Interest Rate Derivatives

11.1 Introduction
Up till now we have assumed that the market interest rate is constant and ob-
servable. Be it a general concept building or a discussion on option pricing theory,
European or American, the market interest rate is taken to remain the same from
the time of purchase/investment till the maturity of the instrument. This has
greatly simplified the discussion on pricing and hedging of these instruments. In
practice, however, it cannot be the case when options and derivatives are written
either on interest rates or specifically on securities whose values depend on inter-
est rates (e.g., bond options, swaps, caps, floors etc.). Any fluctuation in interest
rates can bring risk to investment on these instruments, and it is exactly what an
investor is looking to hedge. It thus makes no sense to assume a constant interest
rate. In fact pricing and hedging interest rate derivatives require models that can
describe the evolution of the entire term structure of interest rates. This provides
a good reason to study and understand what and how the interest rate theory is
build.
One thing that we have to understand is that besides equity derivatives (like
the one we had studied in option pricing theory or portfolio optimization), there
are several other types of derivatives available in the market. Among them, the
interest rate derivatives are the popular ones. The interest rate derivatives market
constitutes the largest derivatives market in the world. An interest rate derivative
is a derivative where the underlying asset is the right to pay or receive a notional
amount of money at a given interest rate. Interest rate theory is fairly complex and
requires greater insight so much so that it is generally covered only in Appendix of
few introductory textbooks on financial mathematics till late. Except when either
the text is aimed for advance level learning/modelling or it primarily focus on
interest rate theory, there are not much references available to report at a very
334 Financial Mathematics: An Introduction

basic level. The difficulties are mainly due to the very nature of the fixed income
instruments, implementation and calibration of various models, and above all the
interest rate theory is not standardized yet. There is no well-accepted standard
general model that can capture the dynamic of interest rates like the Black-Scholes
model for equities.
This chapter is not intended to present the entire contents of otherwise a vast
theory on interest rates, but rather to briefly provide the fragments of it so that
the readers can appreciate the significance of the concept as a whole. Of course,
we cite here few good books and notes [21, 24, 87, 92, 127, 146, 149] on interest
rates that have been published in the last decade or so.

11.2 Bond Price

A good starting point is to understand the meaning of bond (or fixed securities)
and some notions related to bond interest.
A government bond is a bond issued by a national government denominated
in the country’s own currency. Bonds issued by national governments in foreign
currencies are normally referred to as sovereign bonds. The first ever government
bond was issued by the English government in 1693 to raise money to fund a war
against France. Bonds are also issued by public authorities, credit institutions,
companies and supranational institutions in the primary markets. The most com-
mon process of issuing bonds is through underwriting. In underwriting, one or
more securities firms or banks, forming a syndicate, buy an entire issue of bonds
from an issuer and re-sell them to investors.
Definition 11.2.1 (Bond) A bond is a contract/certificate in which the issuer
promises to pay the holder a sequence of interest payments for a specified period
of time, and to repay a definite amount at a specified terminal date. The promised
final amount is called the face value of the bond, and the terminal date of the
contract is called the maturity or redemption date. The sequence of payments at
regular interval until the redemption date are done through coupons. A zero coupon
bonds (ZCB) are the ones which makes no payment (in other words no coupons)
before the redemption date.
Thus a bond is like a loan; the issuer is the borrower (debtor), the holder is
the lender (creditor), and the coupon is the interest. Bonds provide the borrower
with external funds to finance long-term investments, or, in the case of govern-
ment bonds, to finance current expenditure. Bonds can be looked at as long term
 Bond Price 335

debt investment and because of a relative low risk of default they do not usually
come with a high interest rate and would largely be for investment rather than
speculation purpose.
Unless otherwise stated, by a bond we shall mean a coupon bond. Whenever
we wish to talk about a zero coupon bond we shall be stating it explicitly as ZCB.
We shall be making two assumptions in the sequel. One that the issuer of the
bond fulfills all commitments made at the issue date of bond, and second that the
ZCBs are traded for all maturities T in the market.
We shall be using the notation B(t, T) to denote the price of a T-maturity bond
at time t, 0 ≤ t ≤ T.
Let the redemption date of a ZCB be T, and assume that its face value is Rs
1, that is, B(T, T) = 1. Note that B(T, T) might be less than 1 if the issuer of the
T-bond defaults. If the market rate of interest (continuously compounding) r is
constant, then
B(t, T)er(T−t) = 1,
equivalently,
B(t, T) = e−r(T−t) .
If the face value of a ZCB is F then B(t, T) = Fe−r(T−t) .
Example 11.2.1 A ZCB is issued by the government on August 16, 2010 at an
interest rate 2.53% payable semiannually. The bond will mature on August 16,
2040, with a face value Rs 100. Find the price of the bond.
Solution This is a 30-year bond with F = Rs 100 and r = 0.0253/2 = 0.01265, T =
60 (because of semiannual payments). Hence,

B(0, 60) = 100e−60∗0.01265 = Rs 46.81.

Thus, one unit of the ZCB will be trading at Rs 46.81 on the issue date. So if an
investor purchases 500 units of this bond on August 16, 2010 then he has invested
Rs 23405 and will get Rs 50000 after 30 years.


Example 11.2.2 Consider the same data as in Example 11.2.1. (a) Suppose an
investor wishes to purchase the bond on August 16, 2015, then what price he has
to pay for the bond? (b) If the same bond is purchased on December 1, 2015, then
find the bond price?
336 Financial Mathematics: An Introduction

Solution (a) The main difference is in the purchase date which is after 5 years of
issue of bond. Because of semiannual payment, t = 10, and thus

B(10, 60) = 100e−50∗0.01265 = Rs 53.13.

(b) There are 106 days between August 16, 2015 and December 1, 2015 (assuming
that the bond can be traded on all 365 days of the year). Therefore, t = 2(5 +
(106/365)) = 10.581, where the factor of 2 is because of semiannual payment and
5 is the number of years from the issue of the bond to August 16, 2015. So,

B(t, 60) = 100e−(60−10.581)∗0.01265 = Rs 53.52.



Let us now talk about the bond which pays coupons at regular intervals.
Definition 11.2.2 (Coupon Rate) It is the rate at which the bond pays interest
on its face value to the holder at regular time intervals until the redemption date.
We shall denote the coupon rate by c. Let n coupons be paid at regular time
intervals from the date of purchase of bond, say t, to the redemption date T. Let
h = (T − t)/n. We assume that c and r are measured per coupon-payment period.
The following figure illustrates the cash flow on a coupon paying bond to its
holder.

Bond Price B(0,T)

Time
0 1 2 n−2 n−1 n

Fc Fc Fc Fc Fc + F
Fig. 11.1. Cash flow of a coupon paying bond

Example 11.2.3 Assume that a company issues a 2-year bond with a coupon rate
of 10% payable semiannually and face value of Rs 100. Find the cash flow of the
bondholder.
 Bond Price 337

Solution The bondholder receives the following cash flows.

Year 0.5 1.0 1.5 2.0

CashFlow(Rs) 5 5 5 105


Observe that, for a coupon paying bond with face value F at redemption time
T, the bond price is described by

B(t, T) = Fc(e−hr + e−2hr + · · · + e−nhr ) + Fe−r(T−t) , (11.1)

Simplifying the expression, we get

 1 − e−r(T−t)  T−t
B(t, T) = F c + Fe−r(T−t) , h = .
ehr −1 n
Although we have taken the same coupon rate c throughout the life of the
bond but it may differ with the time intervals. It is worth to note from (11.1) that
the price of a T-bond with face value F paying coupons at the rate ci at dates
ti , (i = 1, . . . n), 0 < t1 < · · · < tn−1 < tn , tn = T, can be expressed in terms of the
prices of ti -ZCBs as follows.

n
B(0, T) = ĉi B(0, ti ),
i=1

where ĉi = ci , (i = 1, . . . , n − 1), and ĉn = cn + F. The left side B(0, T) means price
of T-(coupon) bond and the right side B(0, ti ) means price of the ti -ZCB. In turn
it amounts to say that if we can develop and analyze the price dynamics of ZCBs
of all maturities T, then we can translate the same to analyze the dynamics of
(coupon) bonds.
Example 11.2.4 A bond is issued by the government on August 16, 2010 at an
interest rate 8.53% and coupon rate 7.5% payable semiannually. The bond will
mature on August 16, 2040, with a face value Rs 100. Find the price of the bond.
Solution The bond is a 30-years bond with c = 0.075/2 = 0.0375, r = 0.0853/2 =
0.04265, F = Rs 100, T = 60, n = 60. Hence, using the aforementioned expression,
the bond price on date of issue is given by
 1 − e−(60)(0.04265) 
B(0, 60) = 100(0.0375) + 100e−60∗0.04265 = Rs 87.11.
e0.04265 − 1


338 Financial Mathematics: An Introduction

Example 11.2.5 A Rs 100 face value 10-year bond and a coupon rate of 8%
payable semiannually is purchased by an investor at the price of Rs 98. Find the
annual market rate of interest on the bond.
Solution We need to find r given that T = 20 (because of semiannual data),
B(0, 20) = Rs 98, F = Rs 100, c = 0.04 (again because of semiannual coupon).
Thus,  
1 − e−20r
98 = (100)(0.04) + 100e−20r .
er − 1
Set er = ξ. Then r = lnξ, and ξ is a solution of the following equation.
49ξ21 − 51ξ20 − 50ξ + 52 = 0, ξ  1.
A numerical technique can be applied to solve the above equation for ξ. We can
verify that ξ is close to 1.0415 and so r is close to 4.066%. Thus, the interest rate
is 8.132% per annum convertible semiannually.


Sometime instead of coupon rate, coupon payments of the bond are specified.
For instance, suppose it is specified that a bond makes n coupon payments per
year for T years in the amount Cn , and pays F at maturity. Let r be the annualized
T
interest rate and h = . Then the price of the bond is given by
n
C  −ihr
n
B(0, T) = e + Fe−rT .
n i=1

It can be observed that bond prices have close knit relationship with the in-
terest rate r (that is B(0, T) can be treated as a function of r, although it is not
explicitly specified notational in B(0, T)), and that is precisely what we are aim-
ing to investigate in sections to follow. Some bonds have greater sensitivity to
changes in interest rates. This risk is measure by computing how its price is likely
to change when market interest rates go up or down. A bond’s modified duration,
a figure derived from several factors, measures this risk. The change in the bond
price for a unit change in the yield, r, is described by

dB C  n  T
= −h ie−ihr + nFe−rnh , h = .
dr n i=1 n

The percentage change in the bond price for a unit change in the interest
is called modified duration of the bond. Thus, the modified duration of bond is
 Term Structure of Interest 339

1 dB
− . Modified duration is stated in years. For example, a 3-year duration
B(0, T) dr
means the bond will decrease in value by 3% if interest rate rises 1% and increase
in value by 3% if interest rate falls by 1%. For more details on this aspect, we
refer to [87].
There are many other risks involved in investing in bonds, including default
risk (when the bond issuer is unable to make interest payments and/or redemption
repayment), market risk (the risk that the bond market as a whole declines),
currency risk for foreign investors, interest rate risk (if the prevailing interest rate
rises, the price of the bond will fall making the bond less attractive), inflation
risk, to name a few. Some examples do exist where a government has defaulted on
its domestic currency debt, such as the one in Russia in 1998, and very recently
(2012) in Greece, however such examples are rare. Therefore our already assumed
condition that there is no default payment on part of the issuer does not seem out
of place. In the section to follow we shall be talking more about interest rate risk
on bonds.

11.3 Term Structure of Interest

The real markets interest rates on bonds are not constants. Instead, real markets
have yield curve which is a function of not only the redemption time T but also
the time of purchase t of the bond.
Definition 11.3.1 (Bond Yield) A rate that gives face value 1 at maturity T
on investment of B(t, T) in a ZCB at time t is called bond yield from t to T. It is
denoted by Y(t, T). The continuously compounding bond yield Y(t, T) is given by

B(t, T) = e−Y(t,T)(T−t) ,

that is,
1
Y(t, T) = − ln(B(t, T)).
T−t
Note that Y(t, T) > 0 since B(t, T) < 1 for t < T.
1  B(t, T) 
In case if the face value of a ZCB is F then, Y(t, T) = − ln . Bond
T−t F
yield is an example of a long rate of interest.
Definition 11.3.2 (Yield to Maturity (YTM)) It is the yield earned on a
bond when the bond is held until maturity, assuming that all coupons and principal
payments shall be made on schedule.
340 Financial Mathematics: An Introduction

For a T-year annual coupon bond with face value F and purchase price B(t, T),
the YTM, denoted by Ym , is the solution of the following equation

nr
T−t
B(t, T) = F c e−ihYm + Fe−(T−t)Ym , h= . (11.2)
i=1
nr

where c is the coupon rate and nr is the remaining number of coupon payment
periods from date of purchase t to date of maturity T of the bond. Note that the
YTM, Ym , of a ZCB is same as the bond yield Y(t, T). However, for a (coupon)
bond, the YTM is not the same as Y(t, T).
Definition 11.3.3 ( Term Structure and Spot Rates) The function Y(t, T)
of two variables t and T, t < T, corresponding to ZCBs, is called the term structure
of interest rates. The yields Y(0, T), T > 0, described by the current bond prices of
T-ZCBs are called spot rates.
Let us present an example to clarify the difference between spot rates and
YTM.
Example 11.3.1 Consider a Rs 100 face value 2-year bond which pays annual
coupon at the rate of 5%. Suppose the spot rates are Y(0, 1) = 0.08 and Y(0, 2) =
0.1. Find the YTM of the bond.
Solution First we compute the bond price with the data F = Rs 100, T = 2, c =
0.05.

2
B(0, 2) = Fc e−iY(0,i) + Fe−2Y(0,2)
i=1
= (100)(0.05)e−0.08 + (100)(0.05)e−(2)(0.1) + 100e−(2)(0.1)
= Rs 90.58.

The YTM of the bond is the interest, say Ym , is computed using (11.2) as follows.

B(0, 2) = 5e−Ym + 5e−2Ym + 100e−2Ym

that is,
90.58 = 5e−Ym + 105e−2Ym .
After some simplification we get Ym = 0.099495 or the YTM is 9.95%.


 Term Structure of Interest 341

In practice, the yield rates or term structure on bonds, Y(t, T), 0 < t ≤ T ≤ T̂,
where T̂ is some hypothetical maximum time horizon such that all ZCBs trading in
markets will mature at or before T̂, are not observable while the current (purchase)
price of the bonds can be observed from the market. What we can do then is to
solve for the YTM that equates the discounted future cash flows (coupon payments
and redemption value) to the price of the bond.
Example 11.3.2 Consider a Rs 100 face value 10-year bond issued on August 16,
2010 which pays coupon semiannually at a rate 10% on August 16 and February
16 of each year. An investor purchase a bond at a price Rs 102.54 on December
31, 2017. Find the YTM of an investor.
Solution There are 183 days between two coupon dates and 46 days between
the date of purchase December 31, 2017, and the date of the next coupon on
February 16, 2018 (assuming 365 trading days in an year). An investor will also
receive the next 6 coupon payments starting on February 16, 2018. A little insight
in the problem will give the YTM Ym of an investor as a solution of the following
equation
46
−Ym −2Ym −3Ym −4Ym −5Ym −5Ym
− Ym
102.54 = ((100)(0.05)(1 + e +e +e +e +e ) + 100e )e 183 ,

that is,

(105ξ6 − 100ξ5 − 5)ξ0.25137 + 102.54(1 − ξ) = 0, ξ = e−Ym , ξ  1.

The above can be solved for Ym by some numerical technique. We can verify that
ξ is close to 0.95. The YTM per annum is thus close to 2Ym ∗ 100% = 10.26%.


The term structure of interest rates is also known as yield curve. It is a very
common bond valuation method. The yield curve is constructed by interpolat-
ing the market data of yield to maturities and the corresponding maturity dates
of finitely many benchmark fixed-income bonds. The yield curve measures the
market’s expectations of future interest rate. The short-term interest rate can be
thought of as shortest maturity yield or perhaps the overnight rate offered by the
market. The exact shape of the curve can be different at any point in time. Any
change in the shape of a normal yield curve is an indication that the investors
need to change their outlook on economy.
Under normal market conditions, wherein investors believe that that there will
be no significant changes in the economy, such as in inflation rates, and that the
342 Financial Mathematics: An Introduction

rve

interest rate %
cu
yield

maturity
Fig. 11.2. Normal yield curve

economy will continue to grow at a normal rate, the yield curve generally looks
like the one shown in Fig. 11.2.
A flat yield curve usually occurs when there are mixed indications in the mar-
ket. There are speculations that short-term interest rates will rise and simulta-
neously there are signals that long-term interest rates will fall. If the initial term
structure is flat, then the yields Y(0, T) may be independent of T as shown in Fig.
11.3(a).
The market expects long-term fixed income securities to offer higher yields than
short-term fixed income securities. This is a normal expectation because short-
term instruments generally hold less risk than long-term instruments. Sometimes,
however, abnormal conditions in the economic environment can result in short-
term interest rates rising above that offered by long-term fixed income investments.
The result is a negative yield curve depicted in Fig. 11.3(b).
interest rate %
interest rate %

yield curve
yie
ld c
urv
e

maturity maturity
(a) (b)

Fig. 11.3. (a) Flat yield curve; (b) Negative yield curve
 Term Structure of Interest 343

To keep the matter simple for understanding, in the example to follow, we

assume that the bond yield Y(t, T) is independent of bond maturity time T. So,
we denote it by Y(t).
Example 11.3.3 Consider a Rs 100 face value 2-year ZCB which is currently
trading at Rs 90. Analyze how the ZCB price gets affected with a change in the
bond yield.
Solution We first compute the present yield value Y(0) of the ZCB under consid-
eration.
B(0, 2) = 100e−2Y(0) with B(0, 2) = Rs 90,
yields
ln(0.9)
Y(0) = = 0.0527.
2
Thus the bond yield at time t = 0 is Y(0) = 5.27%.
Suppose after one year, the bond yield Y(1) changes to 6%. It amounts to
assuming that Y(1, T) = 0.06, for all T > 1. Then

B(1, 2) = 100e−0.06 = 94.18,

which means that one unit of the same bond would be trading at a cost of Rs
94.18 if bond yield becomes 6%. The logarithmic return on the bond is
 94.18 
ln = 0.0454 or 4.54%.
90
On the other hand if the bond yield reduces to 4% after 1 year. Then the price
of the bond becomes
B(1, 2) = 100e−0.04 = 96.08,
which means that one unit of the bond will be trading at Rs 96.08. The logarithmic
return, in this case, is
 96.08 
ln = 0.0654 or 6.54%.
90
In case if the yield Y(1) remains the same, that is, 5.27%, then it can easily be
seen that B(1, 2) = Rs 94.87 with logarithmic return 5.265%.


344 Financial Mathematics: An Introduction

From above discussion we can realize that an investor in a ZCB is better off if

the bond yield decreases.
Next suppose we wish to have a logarithmic return on investment to be say
7%. Then the bond price B(1, 2) and the corresponding bond yield Y(1) should
respectively be

B(1, 2) = B(0, 2)e0.07 = 0.9653 or Rs 96.53 ,

Y(1) = −ln(0.9653) = 0.0354 or 3.54% .

Let us ask another question. Can we get a yield of 11% on the bond after
one year? Suppose yes; then B(1, 2) = 90e−0.11 = 100.46, which is not possible as
B(1, 2) ≯ Rs 100, the face value of the bond.
When the term structure is not flat, the premium of a bond with respect to
its face value may vary with the time to maturity even for bonds with the same
coupon rate of interest. The same is illustrated in the following example.
Example 11.3.4 Consider a Rs 100 face value bond with coupon rate of 4.0%
per annum paid semiannually. The spot rates of interest in the market are given
as follows.

Year 0.5 1 1.5 2 2.5 3 3.5 4

Spot Rate (in %) 3 3 3.5 3.5 4 4 4.5 4.5

Find the price of the bond if it matures in (a) 6 months (b) 2 years, and (c) 4
years.
Solution (a) Note that T = 1 and Y(0, 1) = 0.03/2 = 0.015. Thus,

B(0, T) = FceY(0,1) + FeY(0,1) = 102e−0.015 = Rs 100.48.

(b) Here, T = 4 and Y(0, i) = (ith spot rate)/2.


T
B(0, T) = Fc e−iY(0,i) + FeY(0,T)T
i=1
−0.015
= 2(e + e−(2)(0.015) + e−(3)(0.0175) + e−(4)(0.0175) ) + 100e−(4)(0.0175)
= Rs 96.92.

(c) In this case T = 8. So,

 Forward Rates 345


T
B(0, T) = Fc e−iY(0,i) + FeY(0,T)T
i=1
−(3)(0.015)
= 2(e + e−(7)(0.0175) + e−(11)(0.02) + e−(15)(0.0225) ) + 100e−(8)(0.0225)
= Rs 90.24.

Note that we have a premium bond for the case of (a) with the price higher
than the face value while for the cases (b) and (c), we have a discount bond with
the price lower than the face value.



11.4 Forward Rates

We begin this section with an example to familiarize the readers with another very
important type of interest rate closely related to term structure of ZCB. This will
also highlight the relationship between spot rates of different time maturity.
We have already seen that the yield vary through time primarily because of
expected variation in the inflation rates in the markets. This in turn causes bond
price fluctuations even if the bonds have same maturity value.
Example 11.4.1 Consider two ZCBs. Bond A is a 1-year bond and bond B is a
2-year bond. They both have face values Rs 100. The one-year spot rate is 5% and
the two-year spot rate is 8%. Find the two bonds prices.
Solution We can easily calculate the present value for bond A and bond B as
follows.

BA (0, 1) = 100e−0.05 = Rs 95.12

BB (0, 2) = 100e−(2)(0.08) = Rs 85.21.


Through the above example we would like to note something important about
the relationship between one-year and two-year spot rates. If one invests in a 2-
year ZCB that yields 8%, his wealth at the end of two years is the same as if
he/she received yields of 5% over the first year and an 11% over the second year
because 100e(2)(0.08) = 100(e0.05 )(e0.11 ). This hypothetical yield of 11% for the second
year is called the forward rate. Thus, we can think of an investor with a 2-year
ZCB as getting the one-year spot rate of 5% and locking in 11% over the second
year.
346 Financial Mathematics: An Introduction

Definition 11.4.1 (Forward Rate) The rate at time t, denoted by f (0, t, T),
such that the present price of a ZCB with maturity T can be generated at time
T by locking a present price of a ZCB with maturity t at time t. That is, the
continuously compounded forward rate for [t, T] prevailing currently (at time zero)
satisfies
B(0, T) = B(0, t)e−(T−t) f (0,t,T) .
In other words,
1
f (0, t, T) = − (ln(B(0, T)) − ln(B(0, t)))
T−t

TY(0, T) − tY(0, t)
= . (11.3)
T−t
Since B(t, t) = 1, hence
1
f (t, t, T) = − ln(B(t, T)) = Y(t, T).
T−t
Let us see some more examples and interpretation to clarify what we mean by
forward rate.
Example 11.4.2 If the one-year spot rate is 8% and the two-year spot rate is
9%, what is f (0, 1, 2)? Interpret the result from an investor perspective.
Solution From (11.3), we have f (0, 1, 2) = 2(0.09) − 0.08 = 0.10 or 10%. Consider
an individual investing in a 2-year ZCB yielding 9%. Equivalently, it is same as if
an investor receives 8% over the first year and simultaneously locks in 10% over
the second year. 
Forward rate can also be viewed as an interest rate which is specified at a
current time t for a loan that will occur at a specified future date T. Forward
interest rates also include a term structure which shows the different forward
rates offered to loans of different maturities.
Example 11.4.3 Suppose you wish to take a loan of Rs 100000 for your child
admission one month from now, and you expect to have means to repay the loan
along with interest on loan after 6 months from now. Assume that the market spot
rates for 1 month and 6 months are respectively 0.35% and 0.55% . What interest
rate does the bank will offer you to construct this loan?
Solution You can arrange Rs 100000 by choosing the following strategy. Compute
the discounted value of Rs 100000 for 1 month from now, that is, (100000)e−0.0035 =
 Forward Rates 347

Rs 99650.61. So, you can take a loan of Rs 99650.61 today for the period of 6
months. Invest this loan in purchasing 996.5061 units of 1-month ZCB of Rs 100
face value. This will give (996.5061)(100e0.0035 ) = Rs 100000 after 1 month from
today, and this can be used for child’s admission purpose. Now think of repaying
the loan after 6 months from today. The principal along with interest comes out
to be (99650.61)e(6)(0.0055) = Rs 102993.94.
Instead suppose you decided to wait and take a loan of Rs 100000 after 1-month
from now. In order to maintain no arbitrage position, the interest rate that the
1  102993.94 
bank should charge on your loan to be paid after 5 months is ln =
5 100000
0.0059. It amounts to saying that an interest 7.08% (per annum) will be charged
by the bank on your loan. Observe that the same interest rate can be obtained
directly from (11.3).


To clarify how the forward rates are computed in real markets, we construct a
forward rate agreement (FRA). A prototypical FRA is an over the counter (OTC)
contract involving three time instants say t, T and T + τ, t < T < (T + τ), where t
is the current time, T is the expiry time, and (T + τ) is the maturity time. Suppose
that today is day t, and that at time T we want to lend Rs 1 for the period τ ,
earning the implied forward rate f (t, T, T + τ) over the interval from T to T + τ.
In other words, we want to accomplish on day t the position that allows cash
going out on day T and coming in on day T + τ. To meet this cash flow, we need
to borrow on day t with a T-day maturity (to generate a cash outflow on day
T) and lend with a T + τ-day maturity (to generate a cash inflow on day T + τ).
Moreover, we want that the borrowing and lending be equal on day t so that there
is no initial cash flow. In FRA, the same is achieved, by setting up the following
portfolio at time t.
(a) Take a short position of one unit of T-maturity ZCB having face value 1.
B(t, T)
(b) Take a long position by purchasing units of T + τ-maturity ZCB
B(t, T + τ)
having face value Rs 1.
B(t, T)
The value of the portfolio at t is B(t, T) − B(t, T + τ) = 0. At time T,
B(t, T + τ)
close the short position of Rs 1 in T-maturity bond. At later time T + τ, receive an
B(t, T)
amount from the long position in T + τ-maturity ZCB. In other words,
B(t, T + τ)
B(t, T)
under no arbitrage condition, the discounted value of at time T must
B(t, T + τ)
348 Financial Mathematics: An Introduction

be equal to Rs 1. The yield Y that can explain this payment is the forward rate
for [T, T + τ] prevailing at time t and is given by
B(t, T)
= 1 ∗ e((T+τ)−T) f (t,T,T+τ) , (11.4)
B(t, T + τ)
which means,
1
f (t, T, T + τ) = − (lnB(t, T + τ) − lnB(t, T)).
τ
We define instantaneous forward rate prevailing at time t for investing at time
T as follows.
lnB(t, T + τ) − lnB(t, T)
f (t, T) = − lim
τ↓0 τ
∂
=− lnB(t, T).
∂T
If we know f (t, T) for all 0 < t ≤ T, we can easily recover B(t, T) for all values
of 0 < t ≤ T as :T
B(t, T) = e− t f (t,s) ds .
Definition 11.4.2 The function f (0, T), T ≥ 0, of variable T, is called the ini-
tial forward rate curve, while the interest rate f (t, t) is called instantaneous
short rate that we can lock in at time t to borrow at time t.
Thus instantaneous short rate can be thought of a market rate at which the
money can be borrowed for a very short duration which could be overnight charges
to be repaid later. Suppose a bank in India requires an amount say S to clear their
liabilities due for next day, and presently it is not in a position to arrange for S on
its own. Then the bank can take an overnight loan S1 from say RBI (central bank
of India) at an instantaneous rate (viz., 1 day lending rate) offered by the RBI
which makes S1 grows to S overnight. The bank can repay the amount S1 along
with its interest to RBI at some later date on the borrowing rate offered by RBI.
Now, if instead of continuously compounding forward interest rate, we take
simple interest rate to explain the heretofore payment of FRA in (11.4). Let the
simple interest rate be L(t, T). Then, by the similar argument leading to (11.4),
we have
B(t, T)
= 1 + τL(t, T),
B(t, T + τ)
equivalently,
 Forward Rates 349

 
1 B(t, T)
L(t, T) = −1 .
τ B(t, T + τ)
For 0 ≤ t < T, L(t, T) is the forward LIBOR, the interest rate locked at time
t for investment over the time [T, T + τ]. Also, L(T, T) is called spot LIBOR (or
simply LIBOR), and τ is tenor of LIBOR and is usually taken as 0.25 or 0.5 year.
Many large financial institutions trade with each other deposits for maturities
ranging from just overnight to one year at a given currency. These are traded
on market interest rates. The most commonly used market interest is LIBOR
(London Interbank Offered Rate). The LIBOR is the rate at which financial insti-
tutions are willing to lend, on average (actually there are fifteen different LIBOR
rates for fifteen maturities: overnight, one week, 2 weeks, one month, and so on).
It is an average indicative quote of the interbank lending market. It is calculated
by Thomson Reuters for ten currencies (including USD, AUD, GBP, DKK, EUR,
CAD, JPY, to name a few), and published daily by the British Bankers Associ-
ation. On the other hand, the LIBID (London Interbank Bid Rate) is the rate
that these financial institutions are prepared to pay to borrow money, on average.
Normally, LIBID < LIBOR. The LIBOR is a fundamental point of reference to
financial institutions. Moreover, many fixed income instruments like, forward rate
agreements or mortgage rates, are indexed to the LIBOR. The following is a small
example of EUR LIBOR interest rates having maturity 1 day

Date Rate (in %)

9 Aug 2011 1.18875
10 Aug 2011 0.98750
11 Aug 2011 0.87250
12 Aug 2011 0.85250
15 Aug 2011 0.84500

When reference is made to the Indian interest rate this often refers to the MI-
BOR (Mumbai Inter-Bank Offered Rate) and MIBID (Mumbai Inter-Bank Bid
Rate). The MIBOR was launched on June 15, 1998 by the Committee for the
Development of the Debt Market, as an overnight rate. The National Stock Ex-
change of India launched the 14-day MIBOR on November 10, 1998, and the one
month and three month MIBORs on December 1, 1998. Since the launch, MIBOR
rates have been used as benchmark rates for the majority of money market deals
made in India.
Besides bonds, there are various other interest rate derivative (debt) instru-
ments in the market. Some are easy to understand and model while some others
350 Financial Mathematics: An Introduction

involves complex financial intricacies. The interest rate derivative market is enor-
mous and it is not possible for us to cover all aspects of it in one chapter. Still we
would like to briefly touch upon interest rate swaps, one of the largest and fastest
growing derivative instruments. A “plain vanilla” interest rate swap is a contract
between two parties, often called counter-parties, in which they agree to exchange
their interest payments of two different kind on a predefined principal amount,
on a periodic basis over the fixed time period. Typically payments made by one
counter-party is based on a fixed interest rate for the term of the contract, while
payments made by other counter-party is based on floating interest rate for the
same term. In doing so, the principal amount is not physically exchanged rather
interest payments are exchanged on notional principal. It is a convention to des-
ignate a fixed-rate counter-party as the buyer of the swap and the floating-rate
counter-party is the seller of the swap. This type of contract is based on the needs
and estimates of the level and changes in interest rates during the period of the
swap contract.
For example, when we say a 3-year 8% fixed for six-month LIBOR floating
Rs 10 lakh swap we mean a fixed-rate party is required to pay 8% fixed-rate
interest on a notional principal of 10 lakh to a floating-rate party in exchange for
a variable-rate interest that depends on a pre-specific six-month LIBOR rate on
10 lakh, and the transaction is to be settled every six months. Fig. 11.4 explains
this kind of transaction.

Fixed Rate

Counterparty A Counterparty B
buyer seller
Floating Rate

Fig. 11.4. Vanilla interest rate swap

Example 11.4.4 A swap buyer company A and the swap seller company B enter
into a 5-year swap on January 10, 2001, on terms that company A pays fixed
interest 3.25% annually on the notional principal of Rs 10 lakh and company B
 Forward Rates 351

pays floating interest equal 3-month MIBOR annually on a notional principal of

Rs 10 lakh. What is the cash flow between the parties.
Solution The company A has to pay 3.25% of Rs 1000000 each year to company
B. Thus, B will receive Rs 32500 periodically every year from A for 5 years. The
following table depicts the amount that will get exchanged between two counter-
parties A and B.

Reset Date Payment Date 3-month MIBOR Floating Net Flow

on Reset Date Payment
by B to A
Jan 10, 2001 Jan 10, 2002 4.564 45640 13140
Jan 10, 2002 Jan 10, 2003 3.475 34750 2250
Jan 10, 2003 Jan 10, 2004 3.485 34850 2350
Jan 10, 2004 Jan 10, 2005 2.22 22200 -10300
Jan 10, 2005 Jan 10, 2006 2.314 23140 -9360

The net cash flow figures shown in the last column above are expressed from
company A’s point of view and indicate that company A must pay company B
on each of the last two payment dates. On first three payment dates, since the
floating payment received by company A exceeds the fixed payment, company
A will receive a net cash inflow on these dates, while in last two payments it is
company B that is benefitted . All payments are in rupees.


Another very basic interest rate derivative is repo. A repo (Repurchase agree-
ment) is a way of borrowing against a collateral. Suppose a financial institution
A borrows money from another financial institution B (usually banks or RBI; in
context to follow we assume it to be RBI) to meet its short term needs by selling
certain securities (like bonds) with an agreement that it will buy back the security
at some fixed point in future time (the next day, after a week, etc.) at a predeter-
mined price. It is equivalent to saying that A gets a loan against a collateral (the
security) and pays an interest rate to B. The following figure captures this idea.
Repo rate is also called short term lending rate. When the repo rate increases
borrowing from Reserve Bank of India (RBI) becomes more expensive. Therefore,
we can say that in case, RBI wants to make it more expensive for the banks to
borrow money, it increases the repo rate; similarly, if it wants to make it cheaper
for banks to borrow money, it reduces the repo rate. If banks are short of funds
they can borrow rupees from the RBI at the repo rate, the interest rate with a 1
day maturity. On the other hand, the reverse repo rate is the interest rate that
352 Financial Mathematics: An Introduction

asset (start)

Rs 10000 (start)
Bank A Bank B

borrower lender

Rs 10000 + 5%

asset (end)

Fig. 11.5. Repo rate

banks receive if they deposit money with the RBI. This reverse repo rate is always
lower than the repo rate. The RBI uses the reverse repo rate tool when it feels
there is too much money floating in the banking system. Increases or decreases
in the repo and reverse repo rate have an effect on the interest rate on banking
products such as loans, mortgages and savings. An increase in the reverse repo
rate means that the RBI will borrow money from the banks at a higher rate of
interest. As a result, banks would prefer to keep their money with the RBI. To
conclude the discussion on short-term interest rate, we would like to bring to
the readers note a very common news line when the repo rate and/or reversed
repo rate are changed: “The RBI today hiked short-term lending and borrowing
rates sharply by 50 basis points”. What we mean by one basis point is 1/100 of
1%, so, 50 basis point means 0.5% change in the existing short-term lending and
borrowing interest rates.

11.5 Binomial Lattice Approach for Term Structure

We all agree that the lattice models (particularly the binomial lattice models
studied in previous chapters) may not be good enough to capture the prevalent
scenario exactly but they indeed provide a basic framework for understanding
and building more sophisticated continuous-time models. A very natural question
 Binomial Lattice Approach for Term Structure 353

therefore, “can our most reliable friend, the binomial lattice, be again called to
help us to model the short rate?” Luckily, the answer is YES.
We first partition the time horizon for which we intend to design short term
interest rate into finite number of periods, like per day or per week etc. The
lattice is drawn in a right angled triangle form; it is assumed that at time t, the
two branches from any node are either in “up” state or in “flat” state. An index
i is used to denote how many ups have been taken to reach the node. Thus, each
node of the lattice is indexed by a pair (t, i), where t is the time and i is the node
index at time t. Then, each node (t, i) is assigned a short rate rti ≥ 0. Fig. 11.6
shows a binomial lattice model say for 5-day short rate interest with a basic time
span of 1 day. The time t is shown at the bottom of the lattice while index i
denotes the ups (or height) of that node.

3 3

2 2 2
i
1 1 1 1

0 0 0 0 0
0 1 2 3 4
t
Fig. 11.6. Binomial lattice model for term structure

This lattice forms the basis for pricing interest rate securities by using the mar-
tingale pricing. For example, if Sti is the value of a non-dividend/coupon paying
security at time t and state i, then we insist that
1
Sti = (pu St+1,i+1 + pd St+1,i ), (11.5)
1 + rti
where pd = 1 − pu and pu is the risk neutral probability measure (RNPM). Since
the probability pu is assigned rather than computed, here, it is a convention to
take it 12 . Observe that such a model is arbitrage-free by construction.
If the security pays a coupon/dividend Cti at node (t, i), then formula (11.5)
should be taken as follows.
354 Financial Mathematics: An Introduction

1
Sti = (pu St+1,i+1 + pd St+1,i ) + Cti .
1 + rti
Example 11.5.1 Suppose the current short term rate is 7% per annum, and the
up factor is 1.2 while the flat factor is 0.8. Construct a short term interest lattice
for 6 years from now.
Solution The lattice is described in Table 11.1.

0.1742
0.1452 0.0116
0.1210 0.0968 0.0774
0.1008 0.0806 0.0645 0.0516
0.084 0.0672 0.0538 0.043 0.0344
0.07 0.056 0.0448 0.0358 0.0287 0.0229
t=0 t=1 t=2 t=3 t=4 t=5
Table 11.1. Short term interest rate lattice


Example 11.5.2 Assuming the data of Example 11.5.1, find the price of a ZCB
maturing 4-years from now.
Solution The term structure lattice is described in Table 11.1. To compute the
bond price of a 5-year ZCB, we assign a face value Rs 1 to this bond at t = 4.
Then we work backwards and use formula (11.5) with rti read from Table 11.1.

1.0000
0.8921 1.0000
0.8255 0.9254 1.0000
0.7858 0.8782 0.9490 1.0000
0.7642 0.8496 0.9161 0.9654 1.0000
t=0 t=1 t=2 t=3 t=4
Table 11.2. 4 year ZCB price

For instance, the top entry in second column , i.e. node (1, 1), is computed using
1 1
r11 = 0.084 from Table 11.1 and formula (11.5), to get, (0.8255) +
1 + 0.084 2
1 
(0.8782) = 0.7858.
2
 Binomial Lattice Approach for Term Structure 355



The option contracts in which the underlying asset is a bond are called bond
options. We have to realize that the price of a call bond option on bond increases
and the price of a put bond option on bond decreases as the short-term interest rate
rises (through the impact of short term interest rate on underlying bond price).
There is no significant difference between the stock options and the bond options
except that the underlying asset is the bond than the stock. Another characteristic
difference to appreciate is that bonds are long term investment and because of a
very low risk of default, they are more for an investment rather than speculation.
This is why the bond options are traded on OTC (over the counter) basis, that is,
traded between two private parties and are not listed on an exchange, unlike stock
options which are highly speculative and generally traded in exchange markets.
Despite this difference, the mathematics of bond option price is similar to the one
we are already familiar with stock option pricing.
For instance, let us look at the put-call parity for European bond options.
Consider a portfolio where we purchase one ZCB, take a short position (sell) on
one European call bond option, and take a long position (buy) on one European
put bond option, both bond options have same time to maturity T and same strike
price K.
At t = 0, the portfolio worth is V(0) = B(0, 0) − C(0, 0) + P(0, 0), where B(0, 0)
denotes the current price of a ZCB, and C(0, 0), P(0, 0) are prices of a call option
and a put option on the underlying ZCB at t = 0 respectively. At T, the value of
this portfolio is


B(0, T) + K − B(0, T) , if B(0, T) < K
V(T) =
B(0, T) − (B(0, T) − K) , if B(0, T) ≥ K.

Thus, no matter what is the state, the portfolio is worth K at time of expiration.
With no arbitrage in force, the payoff from the portfolio is risk-free, and we can
discount its value at the spot rate Y(0, T), YTM of a T-ZCB, to get B(0, 0)−C(0, 0)+
P(0, 0) = Ke−Y(0,T)T , which is similar to the put-call parity for stock options.
Let us investigate pricing of bond option through some simple examples.
Example 11.5.3 Compute the price of a European call option on the ZCB of
Example 11.5.2 that expires in 3 years and has strike price Rs 93.
Solution At time of expiry t = 3, K = 93, hence C(i, 3) = Max{100 B(i, 3) −
K, 0}, i = 0, 1, 2, 3. The same is depicted in the last column in Table 11.3. There-
after the call price is computed iterating backward and using (11.5).
356 Financial Mathematics: An Introduction

0
0 0
0.4102 0.8894 1.8983
0.9643 1.6534 2.6025 3.54
t=0 t=1 t=2 t=3
Table 11.3. 3-year European call bond option price

The European call price is Rs 0.9643 on a ZCB governed by short term interest
rates given in Table 11.1.


Example 11.5.4 Compute the price of a European put option on the ZCB of
Example 11.5.2 that expires in 3 years and has strike price Rs 93.
Solution We shall work out the put price using Table 11.2 and formula (11.5),
working backward. The Table 11.4 depicts the complete working of the same.

3.7908
1.9318 0.4622
0.9909 0.2166 0
0.511 0.1025 0 0
t=0 t=1 t=2 t=3
Table 11.4. 3-year European put bond option price


By now we have realized that the term structure dynamics is characterized by
the evolution of the short term interest rates. In the sections to follow we shall
be describing two basic models which capture the dynamics of short term interest
rates and consequently the bond pricing.

11.6 Vasicek Model

Vasiceks [143] pioneering work in 1977 is the first account of a bond pricing model
that incorporates stochastic interest rate. The short interest rate or instantaneous
rate (simply called short rate from now onwards) dynamics is modeled as a diffu-
sion process with constant parameters. Remember we have already discussed that
the short-term interest rate is the shortest maturity yield or perhaps the overnight
rate offered by the market, that is, f (t, t) which is same as Y(t).
 Vasicek Model 357

Vasiceks model is a special version of Ornstein-Uhlenbeck (O-U) process, with

constant volatility. In fact the O-U process (named after Leonard Ornstein and
George Eugene Uhlenbeck) is a stochastic process that describes the velocity of a
massive Brownian particle under the influence of friction. Over time, the process
tends to drift towards its long-term mean. Such a process is thus called mean-
reverting .
We shall be denoting the short rate stochastic process by {r(t), t ≥ 0} to keep
unanimity with the existing literature in this context. The short rate is modeled
through a stochastic differential equation (SDE) involving three parameters. The
model assumes that r(t) follows the following SDE.
dr(t) = (α − βr(t))dt + σdW(t), (11.6)
where α, β, σ are positive constants and W(t) is a Wiener process.
This model is very tractable, and there are explicit solutions for a number of
derivatives based on it. In order to find the distribution of r(t) in (11.6) we apply
the Ito Lemma. Consider the function g(t, r) = eβt r. Apply Ito Lemma, we get
∂g ∂g 1 ∂2 g
dg = dt + dr + (dr)2
∂t ∂r 2 ∂2 r
= βeβt r(t)dt + eβt dr(t).
Since the right hand side is independent of g, on stochastic integration equiv-
alent, we get
, t , t
βs
g(t, r) − g(r, 0) = βe r(s)ds + eβs ((α − βr(s))dt + σdW(s))
0, , t0
t
βs
=α e ds + σ eβs dW(s)
0 , 0t
α βt
= (e − 1) + σ eβs dW(s),
β 0
where the name of the integration variable has been changed from t to s to avoid
confusions.
Remembering that g(t, r) = eβt r, the above relation yields
α
E(g(t, r)) = E(eβt r) = r(0) + (eβt − 1), (11.7)
β
:t
where in computing (11.7) we have use the fact that 0 eβs dW(s), being an Ito
integral with non-random integrand, so a martingale, consequently its expected
value is zero.
358 Financial Mathematics: An Introduction

Moreover, as we have
, t
−βt α
r(t) = r(0)e + (1 − e−βt ) + σe−βt eβs dW(s), (11.8)
β 0

its variance is given by

, t
2 −2βt
Var(r(t)) = σ e e2βs ds.
0

The immediate above relation needs a bit of explanation. Recall that from
the definition of the Brownian motion : t the increments
Wi = W(ti+1 ) − W(ti )
*
have variance ti+1 − ti . So, for I = 0 f (s)dW(s)  i f (ti )
Wi , the variance is,
* * :t
Var(I) = i f 2 (ti )Var(
Wi ) = i f 2 (ti )(ti+1 − ti ), implying, Var(I) = 0 f 2 (s)ds. We
have used this fact in computing Var(r(t)) from (11.8). Next on simplifying the
Var(r(t)) relation, we get

σ2
Var(r(t)) = (1 − e−2βt ). (11.9)
2β
Moreover, as the increments of the Brownian motion are independent and normally
distributed, so from (11.9), r(t) is also normally distributed.
The above discussion can be summarized to explicitly state the distribution of
r(t) as follows.
 α σ2 
r(t) ∼ N r(0)e−βt + (1 − e−βt ) , (1 − e−2βt ) .
β 2β

α σ2
Observe that when t → ∞, the E(r(t)) → while Var(r(t)) → . Hence, in
β 2β
ασ2 
this case, r(t) ∼ N , . Observe that the variance of the short rate converges
β 2β
to a finite value in contrast to the case of Brownian motion.
α
We now provide an interpretation of the term . Recall the Vasicek model SDE
β

dr(t) = (α − βr(t))dt + σdW(t)

α 
= β − r(t) dt + σdW(t).
β
 Cox-Ingersoll-Ross Model 359

α
(i) If r(t) = , then the drift (term associated with dt) is zero.
β
α
(ii)If r(t) > , then the drift is negative. The process will try to pull up r(t).
β
α
(iii)If r(t) < , then the drift is positive. The process will try to pull down r(t).
β
α
Hence the drift is always directed to , that may thus be interpreted as a long
β
run mean of the short rate r(t). The parameter β represents then the strength of
this mean reversion.
Let us see what happens to the distribution of r(t) when the volatility σ = 0.
In that case the Vasicek model SDE (11.6) reduces to

dr(t) = (α − βr(t))dt,

that is,
dr
α = βdt.
−r
β
The solution is
α  α
r(t) = + r(0) − e−βt .
β β
α α
Again observe that when t → ∞, r(t) → from below when r(0) < and from
β β
α
above when r(0) > . Thereby depicting the mean reverting nature of the short
β
interest rate {r(t), t ≥ 0} process.
An obvious limit of this model is that r(t), having a normal distribution, can
always assume negative values with positive probability. Although practically, on
calibration of the Vasieck model, it is found that the probability of r(t) taking
on negative values is negligible, yet it could not be completely ruled out. The
Vasicek model is not realistic enough because the short interest rate r(t) < 0 with
a positive probability is ambiguous.

11.7 Cox-Ingersoll-Ross Model

The Cox-Ingersoll-Ross (CIR) model [34] is another well-known mathematical
model which describes interest rate movements driven by only one source of market
360 Financial Mathematics: An Introduction

risk. It was introduced in 1985 as an extension of the Vasicek model. The dynamic
of the short interest rate process, {r(t), t ≥ 0}, is described by the following SDE.

dr(t) = (α − βr(t))dt + σ r(t)dW(t), (11.10)
where α, β, σ are positive constants, W(t) is a Wiener process which models the
random market risk factor. The parameter σ determines the volatility of the in-
terest rate.
Example 11.7.1 Assume a particular short interest rate follows the following
CIR model
dr(t) = 0.22(0.06 − r(t))dt + 0.45 r(t)dW(t).
At some particular time t, r(t) = 0.05, and then r(t) suddenly becomes 0.02. What
is the resulting change in the volatility?

Solution The volatility in the CIR model is σ r(t). For σ = 0.45 and r(t) = 0.05,
√ √
√ for r = 0.02,
the volatility is 0.45 0.05, while √ the volatility is 0.45 0.02. The
change in volatility is thus 0.45 0.02 − 0.443 0.05 = −0.09426.


Unlike the Vasicek model, the CIR model is not Gaussian and is therefore
considerably more difficult to analyze. Furthermore, it does not have the closed
form solution like the one (11.8) we have for the Vasicek model. But still we can
find the distribution of r(t). For this text we restrict ourselves to find only the first
two moments of r(t) by applying the Ito Lemma.
Let g(t, r) = eβt r. From Ito Lemma, we have
∂g ∂g 1 ∂2 g
dg = dt + dr + (dr)2
∂t ∂r 2 ∂2 r
= βeβt r(t)dt + eβt dr

= βeβt dt + eβt ((α − βr(t))dt + σ r(t)dW(t))

= αeβt dt + σ r(t)dW(t).
An equivalent integration form is
, t , t

βt βs
e r(t) = r(0) + α e ds + σ eβs r(s)dW(s)
0 0
, t
α
= r(0) + (eβt − 1) + σ eβs r(s)dW(s). (11.11)
β 0
 Cox-Ingersoll-Ross Model 361

Though (11.11) looks very similar to the one we obtain in (11.8) for the Vasicek
model, yet there is a noticeable difference. Observe that the term r(s) appears
in the right hand side Ito integral in (11.11). Thus we are not able to obtain the
closed form expression for r(t). We need to restore to some numerical techniques
for computing r(t) value at given t.
Going back to our discussion, we compute the first moment, that is expectation
of r(t), from (11.11). Recall that the expectation of an Ito integral is zero, we get
α  βt 
E(eβt r(t)) = r(0) + e −1 , (11.12)
β

equivalently
α 
E(r(t)) = r(0)e−βt + 1 − e−βt .
β
Note that the E(r(t)) in the CIR model is the same as the E(r(t)) in the Vasicek
model.
Next we wish to compute the variance Var(r(t)). For this, let h(t, r) = (g(t, r))2 =
e2βt r2 . Applying the Ito Lemma on h(t, r), we have

∂h ∂h 1 ∂2 h
dh = dt + dr + (dr)2
∂t ∂r 2 ∂2 r
= 2βe2βt r2 (t)dt + 2e2βt r(t)dr + e2βt (dr)2

= 2βe2βt r2 (t)dt + 2e2βt r(t)((α − βr(t))dt + σ r(t)dW(t)) + e2βt σ2 r(t)dt,

where the last term in the above equality follows on account of dW(t)dW(t) =
dt, dW(t)dt = 0, dtdt = 0. Simplifying the above relation, we get

dh = (2α + σ2 )e2βt r(t)dt + 2σe2βt (r(t))3/2 dW(t)

βt
= (2α + σ2 )eβt g(t, r)dt + 2σe 2 (g(t, r))3/2 dW(t),

which on equivalent integration form yields

, t , t
βs
βs
h(t, r) = r (0) + (2α + σ )
2 2
e g(s, r(s))ds + 2σ e 2 (g(s, r(s)))3/2 dW(s).
0 0

Taking expectation, and using the fact that the expectation of the Ito integral
is zero, we get
362 Financial Mathematics: An Introduction

, t
E(h(t, r)) = r (0) + (2α + σ )
2 2
eβs E(g(s, r))ds
0
, t
α βs
= r (0) + (2α + σ )
2 2
eβs (r(0) + (e − 1))ds
0 β
2α + σ2
α (2α + σ2 )α 2βt
= r2 (0) + (r(0) − )(eβt − 1) + (e − 1).
β β 2β2
We have used (11.12) for E(g(s, r)) in the second last equality. Therefore

−2βt 2 2α + σ2 α −βt −2βt (2α + σ2 )α

E(r (t)) = e
2
r (0) + (r(0) − )(e − e ) + (1 − e−2βt ).
β β 2β 2

To complete the discussion

Var(r(t)) = E(r2 (t)) − (E(r(t)))2
σ2 ασ2
= r(0)(e−βt − e−2βt ) + 2 (1 − 2e−βt + e−2βt ).
β 2β
α ασ2
Observe that when t → ∞, then E(r(t)) → and Var(r(t)) → 2 .
β 2β
The CIR model has a mean reverting property just like Vasicek model. Here,
r(t) ≥ 0 almost surely, and as soon as r(t) = 0 the model SDE (11.10) reduces to a
pure positive drift αdt with zero volatility. This forces the interest rate back into
the positive zone. It is more instructive to compare the CIR model to the Vasicek
model. The following matters stand out in particular.
(i) In the Vasicek model, interest rates can be negative. In the CIR model, negative
interest rates are impossible.
(ii) In the Vasicek model, the volatility (the term associated with dW(t) in the
SDE) of the short-term interest rate is a constant. In the CIR model, the
volatility of the short-term interest rate increases as the short-term interest
rate increases.
(iii) As in the Vasicek model, the short-term interest rate in the CIR model ex-
α
hibits mean reversion with the same mean .
β

11.8 Partial Differential Equation

The Vasicek model and the CIR model discussed above are examples of simplest
fixed income market models that begin with the SDE of the type
 Partial Differential Equation 363

dr(t) = δ(t, r(t))dt + γ(t, r(t))dW(t). (11.13)

Since r(t) is driven by only single SDE, these models are refer to one factor
short rate models.
Let the short rate r(t) be driven by one factor model under risk neutral measure
(
P as follows.
(
dr(t) = δ(t, r(t))dt + γ(t, r(t))dW(t), (11.14)
( is the Brownian motion under P.
where W(t) (
Since r(t) is a solution of SDE (11.14), it must be a Markov process. This
implies that the bond price at time t of a ZCB with maturity time T and face
value 1 can be written as some function of r(t) and t. That is,
:T
( −
B(t, T) = E(e t
r(u)du
/Ft ) = g(t, r),
for some function g(t, r) of two variables t and r, and filtration {Ft , t ≥ 0} to which
the process {r(t), t ≥ 0} is adapted.
In trading on these bonds, one wants to be sure that there are no arbitrage
opportunity. According to the first fundamental theorem of asset pricing (see,
Chapter 3), for no arbitrage, a risk neutral probability measure P ( must exist
such that the discounted bond prices D(t)B(t, T) are martingale under P, ( where
:t
{D(t), t ≥ 0} is a discount process given by e− 0 r(s)ds . We apply the following three
standard steps of the Feynman-Kac theorem (Chapter 9) to determine the partial
differential equation (PDE) describing the dynamics of ZCB price.
Step 1. D(t)B(t, T) = D(t)g(t, r) is a martingale.
Step 2. Observe that
d(D(t)B(t, T)) = −r(t)D(t)g(t, r)dt + D(t)dg(t, r)
1
= D(t)(−r(t)g + gt + δ(t, r(t))gr + γ2 (t, r(t))grr )dt
2
(
+D(t)γ(t, r(t))gr dW(t).
Step 3. Set dt coefficient equal to zero.
Thus we get that the bond price B(t, T) = g(t, r) is a solution of the following
PDE.
1
gt (t, r) + δ(t, r)gr (t, r) + γ2 (t, r)grr (t, r) = rg(t, r), (11.15)
2
with the terminal condition B(T, T) = g(T, r) = 1, for all r.
We present below two simplest scenarios when the bond price can be computed
explicitly.
364 Financial Mathematics: An Introduction

Example 11.8.1 Find the value of a ZCB when the short rate is governed by the
SDE dr = 0.
Solution As discussed above, the value of the ZCB B(t, T) = g(t, r), is the solution
of (11.15). Since dr = 0 implies r is a constant, (11.15) reduces to gt = rg. A
solution is given by
ln(g(t, r)) = c + rt,
where c is a constant. Using the boundary condition g(T, r) = 1 yields c = −rT.
Thus, B(t, T) = e−r(T−t) , which is in agreement with our knowledge about bond
value when interest rate is a constant.


The next example is a particular case of Ho-Lee model in which the short term
interest rate can be described by the SDE
(,
dr = θ(t)dt + σdW

for some appropriate function θ(t). Comparing the above expression with (11.14),
we have δ(t, r(t)) = θ(t) and γ(t, r(t)) = σ.
Example 11.8.2 Find the value of a ZCB when the short rate is governed by the
SDE
(
dr = adt + σdW, a is a constant.
Solution The given SDE is similar to the Ho-Lee model SDE with θ(t) = a. Thus,
δ(t, r(t)) = a and γ(t, r(t)) = σ. The bond value of a ZCB is B(t, T) = g(t, r) is a
solution of (11.15); which, in the considered case, becomes
1
gt + agr + σ2 grr = rg.
2
Assume that g(t, r) = e−r(T−t) A(t, T) is a solution of the above PDE. Then, by
means of classical calculus, we have
1
(rA + At ) − aA(T − t) + σ2 A(T − t)2 = rA,
2
that is,
1
dlnA = ((T − t)a − (T − t)2 σ2 )dt.
2
Integrating and thereafter using the boundary condition that B(T, T) = g(T, r) =
A(T, T) = 1, we get the constant of integration zero. Consequently,
 Partial Differential Equation 365

(T − t)2 1
lnA(t, T) = − a + (T − t)3 σ2 .
2 6
We thus have an explicit formula for B(t, T) = e−r(T−t) A(t, T), where A(t, T) can be
computed using an immediate above expression.


We now go back to the interest rate models studied in sections 11.6 and 11.7.
Recall the Vasicek model. We want to find the bond price of a ZCB with face
value 1 and maturity time T when the dynamics of the market spot rate r(t) is
governed by the Vasicek model SDE (11.6). For this we need to solve the PDE
(11.15) under Vasicek model. Comparing (11.6) with (11.13), we have, δ(t, r(t)) =
α − βr(t) and γ(t, r(t)) = σ. We assume that a solution of (11.15) is of the form

g(t, r) = e−rC(t,T)+A(t,T) , t ∈ [0, T],

where A(t, T) and C(t, T) are non-random functions to be determined; although

we skip the details as to how we guess this form of solution. Then,

gt = (At − rCt )g , gr = −Cg , grr = C2 g.

Note that, for notational convenience, we have avoided writing the variables in all
functions. Thus, from (11.15), we get

1
(At − rCt − (α − βr)C + σ2 C2 − r)g = 0.
2
For this to hold for all values of r, we must have

Ct = βC − 1
1
At = αC − σ2 C2 .
2
Also, the terminal condition g(T, r) = 1 in (11.15) reduces to A(T, T) = 0, C(T, T) =
0. It is easy to solve the first equation to get
1
C(t, T) = + keβt .
β

e−βT
Using the terminal condition C(T, T) = 0, we get k = − . Thus,
β
366 Financial Mathematics: An Introduction

1
C(t, T) = (1 − e−β(T−t) ).
β
Substituting the function C(t, T) in the second equation and thereafter solving
the resultant first order PDE with terminal condition A(T, T) = 0, we can get the
function A(t, T). After some work (which we leave for the readers to complete) we
can see that
( 1 σ2 − αβ)((T − t) − C) σ2 C2
A(t, T) = 2 − .
β2 4β
Once the bond price B(t, T) is known then the bond yield Y(t, T) can be determined
by using
1
Y(t, T) = − ln(B(t, T))
T−t
1
= (r(t)C(t, T) − A(t, T)).
T−t
Next, suppose we wish to find the bond price B(t, T) when the spot rate r(t)
follows the dynamics of the CIR model. Recall the CIR model SDE (11.10). Com-
paring it with (11.13), we have, δ(t, r(t)) = α − βr(t) and γ(t, r(t)) = σ r(t).
Again we assume that the solution of the PDE (11.15) is of the form
g(t, r) = e−rC(t,T)+A(t,T) , t ∈ [0, T],
where A(t, T) and C(t, T) are non-random functions to be determined. In CIR
model case, the PDE (11.15) becomes
1
(At − αC + (−Ct + βC + σ2 C2 − 1)r)g = 0.
2
The above equation to hold for all r is possible only when the following hold.
1
Ct = σ2 C2 + βC − 1
2
At = αC.
The first of the two equations is a Riccati equation and can easily be solved by
a standard technique to get C(t, T). Once through with it, we can use the C(t, T)
in the second equation to solve it for A(t, T). In this solution procedure we also
need to use the terminal conditions A(T, T) = 0 and C(T, T) = 0. Moreover, the
1
bond yield is given by Y(t, T) = (r(t)C(t, T) − A(t, T)).
T−t
For both the Vasicek model and the CIR model, the ZCB yield is an affine
function of short rate r(t). Such models are therefore also called the affine yield
models.
 Summary and Additional Notes 367

11.9 Summary and Additional Notes

• Though we have not covered all types of bonds in the chapter but some other
important ones we would like to take note in brief. One such bond is called a
callable bond which is a fixed rate bond where the issuer has the right but not
the obligation to repay the face value of the security at a pre-agreed value prior
to the final original maturity of the security. Callable bonds can be viewed as
a combination of noncallable bonds and an option to call a bond. The writer
of the call option is the holder of the bond and the buyer of the call is the
stockholder of the issuing corporation. The call option gives the issuer the right
to call the bond at a fixed strike price any time before bond maturity, after an
initial protection period. There is a good literature available on the pricing of
such bonds. We cite here few of them [16, 54] for readers to explore. Opposite
of these bonds are puttable bond. The other types of bonds are floating rate
bonds, index linked bonds, foreign bonds or sovereign bonds, perpetual bonds,
to name a few.
• We all can note that most of the term structure operations and the dynamic
we’ve carried out used the prices of zero coupon bonds as inputs. However, in
reality, vast majority of traded bonds do have coupons. We have seen in Section
11.2 that the pricing of the coupon bonds can be done through the pricing of
ZCBs. It is the converse that can interest us that how can we infer ZCBs prices
(or yields) from coupon bonds prices (or yields). The method generally used
to achieve this is called bootstrapping. Some textbooks, like [120], explain this
method. We also refer to a good paper [35] to know more about the method
and its technicalities.
• We have talked about the plain vanilla interest rate swaps in Section 11.5.
However, there are other types of swaps which could be of some interest to
readers, like currency swaps and commodity swaps. These swaps help the com-
panies to hedge against the risk of unanticipated interest rates in the market.
For instance, if a buyer company is engaged in a commodity (say oil) swap over
a swap tenor on a predetermined notional principal, then, despite the changes
in oil price and forward interest rate on it, the company has to pay fixed rate
to the other floating rate counter-party as per the terms of the swap, over the
period of the swap. Currency swaps also has a fairly large market. For more on
swaps and their economics, we refer the readers to [46, 87, 127, 133].
• Researchers around the world have made efforts on determining the dynamics
of interest rates [60, 66, 72, 142]. But no single model appears to be valid in all
markets across the world. Besides the two first generation models that we had
368 Financial Mathematics: An Introduction

discussed in Sections 11.6-11.8, there are several other term structure models
studied in the literature under different set ups. Prominent among them are the
binomial lattice model of Black-Derman-Toy [13] and a continuous-time Heath-
Jarrow-Morton (HJM) model [61]. The HJM model evolves the yield curve in
terms of the forward rates. This model is significant so much so that every term
structure model driven by the Brownian motion is the HJM model including
the Vasicek model and the CIR model. For details on how the model dynamics
and its evaluation can be described we urge the interested readers to go through
Chapter 10 in [121]. The continuous-time models of Black-Dermon-Toy and the
Black-Karasinski [14] instead assume that lnr(t) is normally distributed. The
theory of one-factor models has also been extended to two-factor model and
multi-factor models which involve several factors, like short term rate, long
term rate, etc. The main difficulties in the latter models lies in their calibra-
tion which requires lot of data to determine all the parameters in the model
under consideration. There are several other models, like Heston model [62],
which assume the stochastic volatility, besides stochastic short-term returns,
governed by two SDEs. Certain studies are available on www which try to cre-
ate appropriate model for a specific market of a specific country. The literature
in the context is extremely vast to be quoted here but we encourage readers to
take a look at some excellent texts and web pages on interest rate models.
• Another significant point to bring forth is the calibration of the interest rate
models. Calibration of interest rate models under the risk neutral measure
typically entails the availability of some derivatives such as swaps, caps or
swaptions. The primary tool used to estimate the parameters in a system of
stochastic differential equations are Bayesian method, Cholesky decomposition,
principal component analysis, Monte Carlo simulation, various regression tech-
niques, interpolations and splines, generating function, and machine learning.
We refer to few texts [69, 142, 126] in this regard.

11.10 Exercises
Exercise 11.1 Find the spot rate for a 1-year ZCB trading at Rs 92 with a face
value of Rs 100.
Exercise 11.2 A 1-year ZCB with face value Rs 100 is currently selling for Rs
85. What is the interest rate after 6 months if the investment for 6 months in zero
coupon bonds gives a continuously compounding annual return of 12%.
(Hint: Find y(0.5) using B(0, 1) er = Fe−y(0.5)/12 ).
 Exercises 369

Exercise 11.3 A coupon bond with a face value Rs 100 makes coupon payments
of Rs 1.50 every three months. What is the coupon rate?
Exercise 11.4 What asset is consider to be riskier, a 3-month ZCB bond or a
30-year coupon bond maturing in 3 months?
Exercise 11.5 If the 6-month spot rate is 3% and the 1-year spot rate is 5%, then
find the price of a 1-year bond with a 8% annual coupon rate payable semiannually
and a face value Rs 100.
(Hint: Use B(0, 1) = Fce−r1 + (Fc + F)e−2r2 and note c = 0.04).
Exercise 11.6 You intend to purchase a 10-year, Rs 100 face value bond that
pays Rs 5 coupon every 6 months. If the required continuous compounding return
is 5% per annum, how much should you be willing to pay for the bond ?
(Hint: Example 11.2.6 with coupon value).
Exercise 11.7 Compute the price of a 5-year coupon bond with an 6% annual
coupon payable semiannually, the annual bond yield is 8%, and the face value of
bond is Rs 100.
(Hint: Example 11.3.4 with n = 10, c = 0.03, Ym = 0.04).
Exercise 11.8 What is the yield to maturity on a 10-year ZCB with a face value
of Rs 100 that is selling for Rs 65? (Hint: Use B(0, T) = Fe−TYm to compute YTM
Ym ).
Exercise 11.9 A bond with 20 years remaining to maturity is selling for Rs 120
has a 9% annual coupon paid semiannually. If the face value of the bond is Rs
100, calculate the bond yield to maturity.
Exercise 11.10 Calculate the YTM on a 1-year coupon bond with annual coupon
Rs 10 and the face value Rs 100. The bond is selling for Rs 95.
(Hint: Example 11.3.5 with coupon rate replace by coupon value).
Exercise 11.11 Which among the following security has the higher YTM?
(a)A 1-year ZCB with a face value of Rs 100 that is selling at Rs 90.
(b)A 10-year bond with a 5% coupon and a face value of Rs 100 that is selling at
par.
(Note: A bond selling at par means that it is selling at full face value).
Exercise 11.12 Compute the forward rate when the 6-month spot rate is 3%, and
the 1-year spot rate is 5%.
370 Financial Mathematics: An Introduction

Exercise 11.13 A 3-month ZCB is selling for Rs 95, a 6-month ZCB is selling
for Rs 92, a 9-month ZCB is selling for Rs 88 and a 1-year ZCB selling for Rs
85. All four bonds have a face value of Rs 100. Compute the spot rates for the
four periods. Also, compute the second, third and fourth periods forward rates.
(Hint: Use Definition 11.3.1 with F = 100 and then (11.3)).
Exercise 11.14 For ZCBs, each bond with face value Rs 1, the following bond
prices per annum are observed.
B(0,1) B(0,2) B(0,3) B(0,4) B(0,5)
0.9654 0.9173 0.8735 0.8115 0.7855
For each maturity year, compute the bond yields and the 1-year implied forward
rate.
Exercise 11.15 The ZCB continuously compounded annual yields are observed
as follows.
Y(0,1) Y(0,2) Y(0,3) Y(0,4) Y(0,5)
0.03 0.035 0.04 0.045 0.05
For each maturity year, compute the ZCB prices and and the 1-year implied for-
ward rate. What are the simple rate annual yields on the bonds?
(Hint: Simple rate annual yield y(0, T) should be such that 1 + y(0, T) = eY(0,T) ).
Exercise 11.16 Suppose a loan of Rs 100000 will be taken 3 months from now
(date of borrow). It is expected that the loan will be repaid 1 year from now (repay-
ment date). An FRA is designed with guaranteed 6% annual simple interest rate
for 1 year on Rs 100000. The actual interest rate is 4% on the date of borrow.
Determine the settlement of the FRA if the settlement occurs on the date the loan
is (i) borrowed, (ii) repaid.
(Hint: If the FRA is settled at the time the money is borrowed, payments will be
less than when the same is settled on the date of repayment because the borrower
has time to earn interest on the FRA settlement.)
Exercise 11.17 Suppose that the risk neutral probabilities are equal to 0.5 in
every state. If the short-term rates (in percentage per month) are described by the
following table, then find the prices of a 3-month ZCB with face value Rs 100 (take
a one month step).
0.0101
0.0098 0.0091
0.0095 0.0088 0.0081
 Exercises 371

Exercise 11.18 A discrete-time model is used to model both the price of a non-
dividend paying stock and the short-term interest rate. The stock is selling for Rs
100 and the annual simple interest rate is 5%. After 1 year, only two states, upstate
and downstate, are observed in the economy. The stock prices are Rs 110 (upstate)
and Rs 95 (downstate) and the annual short interest rates are 6% (upstate) and
4% (downstate). Find the price of 2-year ZCB with face value Rs 100.
(Hint: Compute risk neutral probability measure p∗ using 1 period binomial lattice
model for stock. Find B(0, 2) using B(0, 2) = 1+r
1
(p∗ Bu (1, 2) + (1 − p∗ )Bd (1, 2)).)
Exercise 11.19 Compute the spot rate binomial lattice model for 6 months with
upward and downward parameters 1.25 and 0.9 respectively. Using this lattice,
compute the price of a 4-months ZCB with face value 100. A European call option
is written on the above bond. If the strike price of the call is Rs 85 and expiration
time 3 months then find the price of the call.
Exercise 11.20 An American put option is written on the 3-year ZCB with face
value Rs 100 governed by the following short rate lattice.

0.08
0.064 0.059
0.052 0.047 0.041
0.042 0.038 0.032 0.026

If the strike price of put is Rs 110 and expiration 2 years, then calculate the put
price.
Exercise 11.21 Suppose the following market data for ZCBs with a maturity
payoff of Rs 1 is given.

Maturity (years) Price Volatility (σ)

1 0.945 0%
2 0.885 10%

Calibrated the data on a 2-period interest rate binomial lattice with upstate annual
2σ
interest re and downstate annual interest r. Calculate r.
 1 1 
Note : B(0, 2) = B(0, 1) + .
2(1 + ru ) 2(1 + rd )
Exercise 11.22 Given a current short rate of 10%, upward and downward param-
eters of u = 1.15 and d = 0.95. Generate a 3-period binomial lattice of short-term
rate. Using this lattice, determine the value of a 2-period coupon bond with face
372 Financial Mathematics: An Introduction

value Rs 100. The coupons are paid at every period at 4% on the bond face value.
(Hint: Take B(0, 3) = 104 for all four nodes at third period. Move backward and
at each preceding node add a coupon of 4 to the node value to get the actual value
(bond price) of that node).
Exercise 11.23 Suppose that the short rate process {r(t), t ≥ 0} follows the fol-
lowing SDE
dr = θdt + βdW,
where θ > 0 and β > 0 are constants. Is the process mean reverting? Justify.
Exercise 11.24 Suppose the short rate process {r(t), t ≥ 0} follows the Vasicek
interest rate model with α = 0.01, β = 0.1, σ = 0.02, and the short rate is 10%.
Find the 2-year ZCB price of face value Rs 1. What is the bond yield?
Exercise 11.25 Suppose that the short rate is currently 3% and its volatility mea-
sure is 0.9% per annum. What happens to the volatility measure when the short
rate increases to 5.5% in (i) Vasicek’s model, (ii) the CIR model?
12
Optimal Trading Strategies

12.1 Introduction
The concept of portfolio optimization and diversification has played a key role in
the development and understanding of financial markets. The major breakthrough
came in 1952 with the publication of Harry Markowitz’s theory of portfolio se-
lection [90]. The theory, popularly referred to as mean-variance portfolio theory,
provided an answer to investors fundamental question: How should an investor
allocate funds among the possible investment choices? Thus, the major interest
for the investor is to have a balance between the total risk of the portfolio and its
expected return. Investors generally set their priority in terms of minimizing total
risk and maximizing return of the portfolio. On the lines of Markowitz’s, Sharpe
[119] proposed capital asset pricing model (CAPM) which introduces the notions
of systematic risk and specific risk. We have already studied Markowitz’s theory
and CAPM in Chapter 5. An interesting question which every investor faces apart
from holding the shares is about how to acquire them. This is where optimal trad-
ing strategy enters into the picture. The goal of optimal trading strategies is to
formulate a mathematical approach which tries to answer some of those questions
that arise during the phase of implementation of investments.
With the established positions, fund managers or the institutional investors
need to re-balance their portfolios frequently, either to include new stock picks,
sell stocks that are out of favor, or to improve the risk/reward characteristics of the
portfolio. This generates huge orders that must be executed in a fixed time horizon.
The execution costs associated with such orders can be substantial. Numerous
studies have shown that these costs typically comprise the largest quantity of the
fund tracking error. This is hardly a problem for an individual investor as trading
volumes are normally small and he/she hardly worry about the execution costs in
acquiring or selling out of a portfolio. On the other hand, the quantities traded
374 Financial Mathematics: An Introduction

by institutional investors and portfolio managers comprise a large fraction of the

average daily volume of many stocks. They need to estimate the execution costs
associated with the transactions to keep a track of their actual returns. Optimal
trading strategies give us a way to estimate the transaction costs and how to
acquire these portfolios with minimal cost and risk.
Financial transaction cost consists of fixed and variable components. In the next
section we will first explain each one of them and then a brief introduction to types
of orders that are executed in the market. Once the framework of transaction cost
and market structure is ready, we will move forward with the execution algorithms
that helps in minimizing the execution cost.

12.2 Introduction to Market Structure

We first explain the components of the transaction cost involved in trading the
large size/block orders and then taking forward from there to types of orders
traded by the traders in the market.
Execution cost has nine different transaction cost components namely - com-
mission, fees, taxes, spreads, investment delay, price appreciation, market impact,
timing risk and opportunity cost. We may refer to Fig.12.1 in this regard. Some of
the components are fixed costs like commission, fees and taxes and they are inde-
pendent of the implementation strategy. The variable cost components like price
appreciation (natural price movement of the stock), market impact (movement
in stock prices caused by particular orders or trades), timing risk (uncertainties
surrounding orders) and opportunity cost (foregone profit of not being able to im-
plement investment decisions) are of higher interest to us while formulating our
strategies. These arise due to buying or selling pressure exerted into the market
by the prospective sellers or buyers. The price of the security is decided by the law
of supply and demand in the market. Now, demanding more of a good or service
will lead to price appreciation and extra premium has to be paid on top of the
purchase price to attract additional suppliers of the good into the market. Same
goes for the supplier too.
Market impact is defined as the movement in the price of the stock caused
by a particular trade. It is the difference between the stock’s price trajectory
with the order and what the price trajectory would have been had the order
not been submitted into the market. Since we cannot simultaneously measure
these two occurrences, market impact is sometime referred as the “Heisenberg
uncertainty principle of finance” [77]. Because of this limitation it is measured as
the difference between the stock price at the beginning of trading and the order’s
 Introduction to Market Structure 375

Taxes

Commissions

Investment Delays

Spread + Price Aprreciation

Market Impact

Opportunistic Cost + Timing Risk

Fig. 12.1. Transaction Cost Components

average execution price. It causes investors to pay premiums to complete the buy
orders and provides discounts to complete the sell orders. Market impact is caused
by two primary reasons. These are (i) Supply Demand Imbalance (liquidity needs)
and (ii) Information Leakage.
In efficient market, where the price of every security fully reflects all available
information and hence is equal to its true investment value, prices adjust contin-
uously to ensure that buying demand equals selling demand. As investors seek
to buy shares, they are required to raise their price to attract additional sellers
into the market. Further, due to the immediacy needs, investors often have orders
that are larger than the quantity of shares available at the best market quote.
To achieve immediate execution (market order), it is often necessary to eat into
the limit order book, making each successive transaction more expensive. This
shift in price of shares is temporary and market soon returns to equilibrium posi-
tion where demand equals supply. This type of market impact is called temporary
market impact.
Every time an order is released to market, it conveys information regarding the
investment and trading intentions of investors which in turn change the beliefs of
other investors for a long time. This causes the market to believe that the future
376 Financial Mathematics: An Introduction

prices will be different than originally expected or there is a change in the stock’s
intrinsic value. This brings a quick price adjustment causing a jump or drop in
the price that remains permanent. This type of market impact is called permanent
market impact.
A portfolio manager while minimizing market impact may increase timing risk
(market risk) of his transaction. To explain clearly, take a simple example of
buying a stock whose price in the market is currently Rs 100. If we need to buy
a large quantity of this stock, we can place a market order to immediately buy
the stock. Then we might drive the price of the stock up so that the average
cost of buying is say Rs 102 or Rs 103. An obvious preventative step to reduce
temporary as well as permanent market impact is to trade more slowly, i.e. break
the large trade into a number of small trades and trade over a longer interval.
In this case, market risk exists where the price of the stock might move in the
opposite direction (increases) due to trades by other market participants in the
same interval. Thus, trading strategy of an investor is the decision of how fast does
one want to trade and it depends on the balance one wishes to strike between the
execution cost (due to market impact) and the market risk.
Thus, trading for institutional investors is not about making profits. It is about
rapid execution with minimal execution costs. Therefore, the aim of the ‘optimal
trading strategies’ is to formulate a mathematical approach to address the ques-
tions and issues that arise during the implementation phase of the investment
cycle- like: How do we estimate trading cost? How long should execution take?
How should the order be sliced? How do we choose an alternate strategy like to
trade aggressively or passively? etc. To begin with, in the following, we describe
the types of the order available to the traders while trading large/small orders in
the market.
Orders are trade instructions. They specify what traders want to trade, whether
to buy or sell, how much, when and how to trade, and, most important, on what
terms. Thus, orders are the fundamental building blocks of the trading strategies.
To trade effectively, a trader must specify exactly what he/she wants. An order
submission strategy is the most important determinant of success of a trader.
Therefore, the proper order used at the right time can make the difference between
a good trade, a costly trade and no trade at all [59].
Traders indicate their willingness to buy or sell by making bids or offers re-
spectively. Traders quote their bids and offers when they arrange their own trades.
Otherwise they use orders to convey their bids and offers to the brokers or auto-
mated trading systems that arrange their trades.
The highest bid price in a market is the best bid. The lowest offer price is the
 Various Types of Orders in the Market 377

best offer (or, equivalently best ask). Traders also call them the market bid and
the market offer because they are the best prices available in the market. The
prices at which orders fill are trade prices. The difference between the best bid
and the best ask is the bid/ask spread.
A market is said to be liquid when traders can trade without significant adverse
affect on execution price. An order offers liquidity or equivalently supplies liquidity
if it gives an opportunity to the other traders to trade large size quickly at a low
cost. Both buyers and sellers can offer liquidity. Buyers offer liquidity when their
bids give other traders opportunity to sell, sellers offers liquidity when their offers
give other traders opportunity to buy.
Markets and traders treat orders differently, depending on whether they are
agency orders or proprietary orders. Agency orders are orders that brokers repre-
sent as agents for their clients. Proprietary trading (also “prop trading”) occurs
when a firm trades stocks, bonds, currencies, commodities, their derivatives, or
other financial instruments, with the firm’s own money as opposed to its cus-
tomers’ money, so as to make a profit for itself.
In the subsequent section we will define various types of orders that traders
send to the market once they decide upon how much to execute, their time limit,
and the side of the trade (i.e. buy or sell).

12.3 Various Types of Orders in the Market

A market order is an instruction to trade securities at the price currently available
in the market. Market orders usually fill quickly, but sometimes at inferior prices.
Traders who want to be certain about their order execution, are sometimes known
as impatient traders. The execution of a market order depends on its size and on
the liquidity currently available in the market. Small size market orders usually
fill immediately with little or no effect on the prices.
The bid/ask spread (also known as bid/offer or buy/sell spread) for securities
(such as stocks, futures contracts, options, or currency pairs) is the difference be-
tween the prices quoted (either by a single market maker or in a limit order book)
for an immediate sale (ask) and an immediate purchase (bid). Generally market
order traders pay the bid/ask spread. The bid/ask spread is the price impatient
traders pay for immediacy. Impatient traders buy at the ask price and sell at the
bid price. The spread is the compensation dealers and limit order traders receive
for offering immediacy. The trader initiating the transaction is said to demand
liquidity, and the other party (counter party) to the transaction supplies liquidity.
Such traders are called liquidity demanders and liquidity suppliers respectively.
378 Financial Mathematics: An Introduction

Liquidity demanders place market orders and liquidity suppliers place limit or-
ders. For a round trip (a purchase and sale together) the liquidity demander pays
the spread and the liquidity supplier earns the spread. The size of the bid-offer
spread in a security is one measure of the liquidity of the market and of the size
of the transaction cost.
Large market orders are more difficult to execute than smaller ones. Traders
willing to take the other side of a large trade are often hard to find market order.
The reason could be due to possibility of informed trading. However to attract
buyers (or sellers) impatient trader often move prices. Large buyers increase bid
prices of their order to encourage sellers to sell to them and vice versa. The
premiums that large buyers pay and the discounts that large sellers offer are price
concessions.
Let us explain the market impact using an example. Let us assume that the
currently available trades in the market are 500 shares at level first with ask
price Rs 100, and 300 shares at level second with ask price Rs 101. Now trader
wants to buy 700 shares. Therefore he/she has to look towards the ask side of the
market i.e. if sellers are available or not. Hence he will buy 500 shares at Rs 100
and rest 200 shares at Rs 101, therefore the impact price at which he will buy is
(500 × 100 + 200 × 101)/700 = 100.2857. Similarly, if in the market 300 shares are
available at first level with bid price Rs 99, further if trader buys 200 shares at
Rs 100 and sell 200 shares at Rs 99 then he has paid Rs 1 as bid ask spread for
the respective trade.
The price, at which market orders trade, depend on current market conditions.
Since market conditions can change quickly, traders who use market orders are at
risk of trading at worse prices than what they expected. This risk in literature is
termed as execution price uncertainty. Execution price uncertainty is due to quote
changes that may occur between the submission of an order and its execution, and
to the unpredictable price concessions that may be required to fill large orders.
Thus, those traders who are concerned about the execution price risk may prefer
to submit limit orders.
A limit order is an instruction to trade at best price available, but only if it
is no worse than the limit price specified by the trader. For buy orders, the trade
price must be at or below the limit price. For sell orders, the price must be at or
above limit price.
In continuously trading markets, a broker (or a exchange) will attempt to trade
a newly submitted limit order as soon as it arrives. If no trader is immediately
willing to take the opposite side at an acceptable price, the order will not be
traded. Instead, it will stand as an offer to trade until someone is willing to trade
 Various Types of Orders in the Market 379

at its limit price, until it expires, or until the trader who submitted it cancels it.
Standing limit orders are placed in a file called a limit order book.
The probability that a limit order will trade depends on its limit price. If the
limit price of a buy order is too low, the order will not trade. Likewise, if a sell
limit price is too high, the order will not trade. Buy limit orders with high prices
and sell limit orders with low prices are aggressively priced. Aggressively prices
limit orders are easiest limit orders to fill.
Traders classify limit orders with limit prices at which they are placed relative
to the market. The market is the range of prices bounded above by the best offer
(lowest price) and below by the best bid (highest price). A marketable limit order
is an order that the broker can execute immediately when a trader submits it. The
limit price of a marketable limit buy order is at or above the best offer. The broker
therefore can manage to buy immediately from the seller quoting the best offer.
Moreover, marketable limit orders are like market orders, except that they limit
the price concessions that brokers can make to fill them. Marketable limit orders
with very high limit buy prices or very low sell prices are essentially market orders.
Traders use marketable limit orders instead of market orders to limit execution
price uncertainty and to limit what they will pay for liquidity.
Limit buy orders that stand at the best bid, and limit sell orders that stand at
the best offer, are at the market. The traders who submit these orders make the
market. To summarize, marketable limit orders are the most aggressively prices
limit orders. Traders who submit standing limit orders offer liquidity to the other
traders. Their limit orders give others the opportunity to trade when they want
to trade. In particular, sell limit orders are call options that give other traders
an opportunity to buy when they want to buy. Buy limit orders, likewise are put
options that give other traders opportunity to sell when they want to sell. The
option strike price is limit price.
The other type of orders are stop orders. A stop instruction stops an order from
executing until price reaches a stop price specified by the trader. Traders attach
stop instructions to their orders when they want to buy only after price rises to
the stop price or sell only after price falls to the stop price. Orders with the stop
instructions are called stop orders.
Traders most commonly use stop orders to stop their losses when prices move
against their positions. For example, suppose that a trader buys 100 sugar future
contracts at Rs 100 each. To limit the potential loss on this position, he/she may
issue a market sell order for 100 contracts with a stop price of Rs 90. If the sugar
drops to or below Rs 90, his/her broker will immediately try to sell 100 contracts
at the best price then available in the market. Traders often calls such orders stop
380 Financial Mathematics: An Introduction

loss orders. The price at which a stop order executes may not be the stop price.
In the above example, if sugar falls quickly from Rs 97 to Rs 88, his/her broker
may be able to sell the 100 contracts at Rs 87.5.
When a trader attach a stop instruction to a limit order, they must specify
two prices. The stop price indicates when the limit order becomes active, and the
limit price indicates the terms upon which a trade may be arranged. The combined
order is stop limit order. With stop limit order, trader do not need to monitor the
market, and thus are free to attend to other business.
Stop orders accelerate price changes. Prices often change because traders on
one side of the market demand more liquidity than what is available. When these
price changes activate the stop orders, it contributes to the one-sided demands
for liquidity. Stop orders accelerate price changes by adding buying pressure when
prices are rising and selling pressure when prices are falling. They demand liquidity
when it is least available. Traders claim that stop orders add momentum to the
market. Traders who pursue momentum trading strategies buy when prices are
rising and sell when price are falling. They basically take the advantage of stop
orders to create momentum in the market. On the other hand, contrarian traders
employ the opposite trading strategy. They buy when prices are falling and sell
when prices are rising. They therefore stabilize prices when they trade.
A trailing stop order is entered with a stop parameter that creates a moving or
trailing activation price, hence the name. This parameter is entered as a percentage
change or actual specific amount of rise (or fall) in the security price. Trailing stop
sell orders are used to maximize and protect profit as a stock’s price rises and limit
losses when its price falls. Trailing stop buy orders are used to maximize profit
when a stock’s price is falling and limit losses when it is rising.
For example, a trader has bought stock ABC at Rs 10 and immediately places
a trailing stop sell order to sell ABC with a Rs 1 trailing stop. This sets the stop
price to Rs 9. After placing the order, ABC doesn’t exceed Rs10.00 and falls to
a low of Rs 9.01. The trailing stop order is not executed because ABC has not
fallen Rs 1 from Rs 10. Later, the stock rises to a high of Rs 15 which resets the
stop price to Rs 14. It then falls to Rs 14 (Rs 1 from its high of Rs 15) and the
trailing stop sell order is entered as a market order.
Another type of order is good till cancelled order. A good-till-cancelled (GTC)
order is an order to buy or sell a security at a specific or limit price that lasts
until the order is completed or cancelled. A GTC order will not be executed until
the limit price has been reached, regardless of how many days or weeks it might
take. Investors often use GTC orders to set a limit price that is far away from the
current market price. Some brokerage firms may limit the time a GTC order can
 Execution Models 381

remain in effect and may charge more for executing this type of order. An uptick is
when the last (non-zero) price change is positive, and a downtick is when the last
(non-zero) price change is negative. Any tick sensitive instruction can be entered
at the trader’s option, for example buy on downtick, although these orders are
rare.
So far in this section, we have discussed various types of orders that traders
execute. In making the decision of how much to trade depends upon the market
conditions like price and volume available in the market. Thus traders generally
use technical indicators to calculate the price and volume change in the previous
trading interval using historical data to forecast for the next trading interval. Now
a days traders prefer to use algorithmic trading strategies that learn the market
conditions and based on the favorable trading rules/circumstances schedule the
trade for the current trading interval. The traders while trading large blocks of
trade face market impact on their execution price.
In the following sections, we describe some models which provide quantitative
trading strategies that are helpful to the traders to minimize the market impact.
These models are refered as execution models.

12.4 Execution Models

Optimal execution characterizes the minimal cost realization of an investment
decision in a financial market. An investment decision concerns the change in
portfolio position of an investor through the acquisition or liquidation of listed
assets. For an investor wishing to buy (sell) a given volume at a particular price
there must exist a counterparty (or number of counterparties) willing to sell (buy)
the same volume. In a listed financial market the buyers and sellers post orders for
the assets that they wish to acquire or liquidate. For an investor wishing to buy
(sell) at a particular price there must be enough counter liquidity at the specified
price. If there is insufficient counter liquidity then the investor must choose to
either transact at a worse price where there is sufficient counter liquidity or over
a period of time in the hope that market dynamics attract more counter liquidity
at an acceptable price. The performance of the trade execution, on completion of
the given transaction, may then be measured and benchmarked. Performance can
be measured using a simple process of taking the initial price at the start of the
transaction, referred to as the arrival price, multiplied by the total volume that
was exchanged, and comparing this against the weighted average price received
multiplied by the volume exchanged over the duration of the transaction. The
382 Financial Mathematics: An Introduction

difference between the theoretical benchmark price and the actual price received
is the implementation shortfall (Perold [107]).
The first quantitative description of estimating and minimizing market impact
has been given by Bertsimas and Lo [11] in 1998. Here, given a fixed block S̄
of shares to be executed (acquired/liquidated) within a fixed time interval [0,T],
and given price dynamics that captures price impact, the objective is to find an
optimal sequence of trades (as a function of the state variables) that will minimize
the expected cost of executing S̄ shares within time T. For this, Bertsimas and
Lo [11] considered various price processes and employed dynamic programming
technique to develop optimal trading strategies for a risk neutral investor. We will
discuss their models and also limitations of these models.
Another major contribution in this area is due to Almgren and Chriss [2],
who observed that market impact and timing risk associated with the execution
of trades are conflicting concepts and hence it is not possible to simultaneously
minimize both. They proposed a mean variance based objective function for their
model and obtained the static strategies where the optimal trading path was
determined in advance of trading. Further, a model is called static if there is no
serial correlation among prices, or no change in the market perception of the price
of the stock. In general, we expect the optimal strategy to be dynamic because
optimal trading path cannot be determined priori of trading, as one needs to
update the optimal strategy based on observed state variables at each time point.
We shall not discuss the details of Almgren and Chriss’ model here and shall refer
to [2] for further study in this regard. However, taking motivation from [2], we
shall incorporate variance terms in the orginal models of Bertsimas and Lo [11]
and present a static approximation technique for the same (Khemchandani et al.
[76]).
The problem statement for obtaining optimal strategy that minimizes the ex-
ecution cost can be formulated mathematically as follows: Consider an investor
who wishes to buy a large block S̄ of shares of some stock over a fixed time in-
terval [0, T]. We divide the interval [0, T] into N subintervals of length τ = T/N
and define the discrete times tk = kτ, for k = 0, 1, . . . , N. Thus t0 = 0 and tN = T.
Here it is understood that the trading begins from the first period. Thus at t1 , S1
shares are purchased at price P1 . In general, at tk , Sk shares are purchased at price
Pk , k = 1, . . . , N. Let Wk be the number of shares that remain to be purchased at
time tk . The investor’s objective is to minimize expected cost of buying S̄ shares
for a given level of risk, say V ∗ . If we use variance of cost of buying S̄ shares as a
measure of timing risk [2], [64], then investor’s problem is represented as follows.
 Execution Models 383

* 
N
Min E1 k=1 Pk Sk
{Sk }
subject to
* 
Var1 N
k=1Pk Sk ≤ V ∗
*N
k=1 Sk = S̄
Sk ≥ 0 (k = 1, 2, . . . , N).
Here E1 and Var1 respectively represent the expected cost and risk which
trader will face at the starting of the interval assuming that he/she has traded
{S1 , S2 , . . . , SN } shares in the intervals {T1 , T2 , . . . , TN } respectively. The constraint
Sk ≥ 0 implies shares purchased while Sk ≤ 0 implies shares sold. The problem
to sell S̄ shares in time [0, T] is a symmetric problem. We need to maximize the
objective function, i.e. maximize the revenue generated by liquidating S̄ shares.
We can reformulate the above problem as

N  
N 
Min E1 Pk Sk + λVar1 Pk Sk
{Sk }
k=1 k=1
subject to
*N
k=1 Sk = S̄
(12.1)
Sk ≥ 0 (k = 1, 2, . . . , N),

where λ ≥ 0 is a parameter. The parameter λ has a financial interpretation, and it

is a measure of risk-aversion, that is, how much we penalize execution risk relative
to expected cost. The only way traders can incur a lower trading cost is through a
strategy with a higher quantity of risk and vice versa. For a risk neutral investor
λ = 0, while for a risk averse investor λ > 0.
Depending upon the law of motion for price process {P(t), t ≥ 0}, optimal strate-
gies can be static or dynamic. Static strategy is one where optimal trading path is
determined in advance of trading. Dynamic strategy is one where optimal trading
path cannot be determined priori of trading and one needs to update the optimal
strategy based upon the market information at each time period. Bertsimas and
Lo [11] and Huberman et al. [64] have used stochastic dynamic programming to
solve the above problem for various law of motion for price process. Some of their
models are discussed in more detail in the following sections.
384 Financial Mathematics: An Introduction

12.5 Static Optimal Strategies for Risk Neutral Investor:

A Dynamic Programming Approach
To provide a concrete illustration of the dynamic programming approach, we
assume that the investor is risk neutral (λ = 0) and the price in the absence of
our trade is given by arithmetic random walk. The impact cost is a linear function
of trade size so that a purchase of Sk shares may be executed at the prevailing
price Pk−1 plus an impact premium of θSk , θ > 0. Therefore, the law of motion for
Pk may be expressed as

Pk = Pk−1 + θSk + k , θ > 0, (12.2)

where k s are assumed to be a zero-mean independent and identically distributed

(i.i.d) random shocks, which provide randomness to the underlying price process

E[k |Sk , Pk−1 ] = 0, Var[k |Sk , Pk−1 ] = σ2 .

This model is same as the basic model of Bertsimas and Lo [11] who derived the
optimal solution of (12.1) for risk neutral investor under price process (12.2), using
dynamic programming. The basic requirements for any dynamic programming
problem are the state of the environment at time tk , the control variables, the
randomness, the objective function, and the law of motion. In our context, the
state at time tk , for k = 1, . . . , N consists of the price Pk−1 realized at the previous
period, and Wk , the number of shares remain to be purchased at time tk . The state
variables give all the information the investor requires in each period tk to make
decision regarding the control. The control variable at time tk is the number of
shares Sk purchased. The objective function is given by

N 
Min E1 Pk Sk ,
{Sk }
k=1

and the law of motion for state variable Wk is represented as

Wk = Wk−1 − Sk−1 , W1 = S̄, WN+1 = 0. (12.3)

Here we ignore the non-negativity constraint Sk ≥ 0 which is discussed later in

this section.
The dynamic programming algorithm is based on ‘the principle of optimality’
which states that, “An optimal policy has the property that, whatever the initial
 Static Optimal Strategies for Risk Neutral Investor: A Dynamic Programming Approach 385

state and control are, the remaining decisions must constitute an optimal policy
with regard to the state resulting from the first decision”. Therefore for every tk ,
k = 1, . . . , N, the sequence {S∗k , S∗k+1 , . . . , S∗N } must still be optimal for the remaining
*
program Vk (Pk−1 , Wk ) = Et [ N i=k Pi Si ]. This property is summarized by the Bellman
equation, which relates the optimal value of the objective function in period tk to
its optimal value in period tk+1

Vk (Pk−1 , Wk ) = Min Ek [Pk Sk + Vk+1 (Pk , Wk+1 )]. (12.4)

{Sk }

By starting at the end (time tN ) and applying the Bellman equation (12.4)
and the law of motion for Pk (12.2) and Wk (12.3) recursively, the optimal value
function can be derived as a function of the state variables. The Bellman equation
(12.4) for time tN is given by

VN (PN−1 , WN ) = Min EN [PN SN ] = (PN−1 + θWN )WN . (12.5)

{SN }

Since this is the last period and WN+1 = 0 by (12.3), there is no choice but to
execute the entire remaining order WN , hence the optimal trade size S∗N = WN . In
the next to last period tN−1 , the Bellman equation is

VN−1 (PN−2 , WN−1 ) = Min EN−1 [PN−1 SN−1 + VN (PN−1 , WN )] (12.6)

{SN−1 }

Substituting the law of motion for PN−1 from (12.2) and WN from (12.3), we get

VN−1 (PN−2 , WN−1 ) = Min EN−1 [(PN−2 + θSN−1 + εN−1 )SN−1

{SN−1 }

+ VN (PN−2 + θSN−1 + εN−1 , WN−1 − SN−1 )]

Using the fact that EN−1 [εN−1 ] = 0, and using the right-hand side of (12.5), the
above equation can be expressed as explicit function of SN−1 as

VN−1 (PN−2 , WN−1 ) = Min [θS2N−1 − θSN−1 WN−1 + θWN−1

2
+ PN−2 WN−1 ].
{SN−1 }

This can be minimized by taking its derivative with respect to SN−1 , and solving
for its zero, yielding
S∗N−1 = WN−1 /2,
 3 
VN−1 (PN−2 , WN−1 ) = WN−1 PN−2 + θWN−1 .
4
Continuing through backward recursion, we get
386 Financial Mathematics: An Introduction

S∗N−k = WN−k /(k + 1), (12.7)

 k+2 
VN−k (PN−k−1 , WN−k ) = WN−k PN−k−1 + θWN−k . (12.8)
2(k + 1)
We encourage the readers to work out (12.7) and (12.8). When we finally reach
the beginning of the program (time t1 ), we get

S∗1 = W1 /N,
 N+1 
V1 (P0 , W1 ) = W1 P0 + θW1 .
2N
Substituting the initial condition W1 = S̄ from (12.3) into above equation gives

S∗1 = S̄/N, (12.9)


N 

θS̄2 1
V1 (P0 , W1 ) = E1 Pk Sk∗ = P0 S̄ + 1+ .
2 T
k=1

Using (12.9), (12.7) and (12.3), we find that

S∗1 = S∗2 = . . . = S∗N = S̄/N.

The optimal execution strategy is simply to divide the total order S̄ into N equal
parts and trade them at regular intervals. This strategy is called a “naive” strategy.
This simple trading strategy comes from the fact that the price impact θSk does
not depend on either the prevailing price Pk−1 or the size of the unexecuted order
Wk . Hence, the price impact function is same in each period and independent from
one period to the next.

12.6 Static Optimal Strategies for A Risk Averse

Investor: A Dynamic Programming Approach
Bertsimas and Lo [11] minimized only expected cost of execution and ignored the
risk associated with execution strategy. Almgren and Chriss [2], and Huberman
and Stanzl [64] defined risk as the variance of cost of execution and used a mean-
variance objective function. But, when we use mean-variance objective function, it
no longer remains additive-separable. Thus, the dynamic programming technique
cannot be used to compute the optimal strategy for a risk averse investor. For
 Static Optimal Strategies for A Risk Averse Investor: A Dynamic Programming Approach 387

static models, Huberman and Stanzl [64] have shown that an equivalent dynamic
program exists and one can find the optimal strategy for risk averse investors. We
used the same result to derive the optimal strategy for risk averse investor for
basic model of Bertsimas and Lo.
For a risk averse investor, the objective function is given by

N  
N 
Min E1 Pk Sk + λVar1 Pk Sk .
{Sk }
k=1 k=1

The law of motion for state variables Pk and Wk are as before in (12.2) and
(12.3) respectively. The Bellman equation for a risk averse investor is given by
!
Vk (Pk−1 , Wk ) = Min Ek [Pk Sk + Vk+1 (Pk , Wk+1 )] + λVark [Pk Sk + Vk+1 (Pk , Wk+1 )] .
{Sk }
(12.10)
As before, the Bellman equation at time tN is given by
!
VN (PN−1 , WN ) = Min EN [PN SN ] + λVarN [PN SN ]
{SN }

= (PN−1 + θWN )WN + λWN2 σ2 .

In the next to last period tN−1 , the Bellman equation is

VN−1 (PN−2 , WN−1 ) = Min EN−1 [PN−1 SN−1 + VN (PN−1 , WN )]

{SN−1 }
!
+ λVarN−1 [PN−1 SN−1 + VN (PN−1 , WN )] .

Substituting the law of motion for PN−1 from (12.2) and WN from (12.3) and
simplifying, we get a quadratic function explicit in SN−1 which can be minimized
by solving for zero of its derivative as before. The best execution strategy and
optimal value function can be obtained by recursively solving. Here, they are
given by
S∗T−k = bk WT−k ,
VT−k = PT−k−1 WT−k + (ak + λσ2 )WT−k
2
,
for k = 0, 1, . . . , T − 1, where


θ
ak = θ 1 − , a0 = θ, (12.11)
4(ak−1 + λσ2 )
388 Financial Mathematics: An Introduction

θ
bk = 1 − , b0 = 1. (12.12)
2(ak−1 + λσ2 )

The optimal strategies for risk neutral, risk averse and risk seeking investors
are compared using a numerical example. The results are shown in Fig 12.2 and
Fig 12.3. Risk averse investor follows an aggressive strategy trading heavily in the
initial phases. Risk seeking investor on the other hand follows passive strategy,
waiting for the price to fall and trades at the end. Risk-neutral investor follows
the naive strategy.

9000

8000

7000

6000
No. of Shares traded (S)

5000

4000

3000

2000

λ=0
1000
λ=1e−5
λ=−1e−5
0
0 2 4 6 8 10 12 14 16 18 20
Time Interval (T)

Fig. 12.2. Comparison of strategies for risk neutral, risk averse and risk seeking investors for Model 1.
S represents the number of shares acquired in each period. The black bold curve corresponds to a risk
neutral investor (λ = 0). The black dash curve corresponds to risk averse investor (λ = 1 × 10−5 ), while
the black dotted curve corresponds to risk seeking investor (λ = −1 × 10−5 ).
 Dynamic Models 389

4
x 10
10
λ=−1e−5
9
λ=1e−5
8 λ=0
No. of Shares traded (S)
7

0
0 5 10 15 20 25
Time Interval (T)

Fig. 12.3. Comparison of strategies for risk neutral, risk averse and risk seeking investors for Model
1. We represents the number of unexecuted shares. The black bold curve corresponds to a risk neutral
investor (λ = 0). The black dashed curve corresponds to risk averse investor (λ = 1 × 10−5 ), while the
black dotted curve corresponds to risk seeking investor (λ = −1 × 10−5 ).

12.7 Dynamic Models

The strategies obtained so far were static where the optimal solution S∗ was de-
termined in advance of trading. It was based on the observation that static opti-
mization procedures lead to globally optimal trading strategies in absence of mar-
ket information or serial correlation among price process. If we introduce market
information variable, the optimal strategy will no longer be static. The dynamic
programming approach for risk neutral investor, for various dynamic models, given
by Bertsimas and Lo is discussed in the sequel.

Linear Price Impact with Market Information Model

If we assume that price Pk at time tk is a function of market information which
changes over time, then the number of shares traded at time tk also depends on the
value of market information at time tk . Therefore, the optimal strategy obtained
is dynamic where one can not determine the strategy in advance of trading. The
following model illustrates the concept of dynamic strategy.
Let Xk be a serially-correlated state variable which affects the execution price Pk
linearly. The price impact function is again assumed to be linear in Sk . Therefore,
the law of motion for execution price Pk is represented as:

Pk = Pk−1 + θSk + γXk + k , θ > 0, (12.13)

390 Financial Mathematics: An Introduction

Xk = ρXk−1 + ηk , ρ ∈ (−1, 1), (12.14)

where k and ηk are independent white noise processes with mean 0 and variances
σ2ε and σ2η , respectively. The presence of Xk in the law of motion for Pk captures the
potential impact of changing market conditions or of private information about
the security, and γ represents its importance for the price process Pk . For example,
Xk might be the return on the S&P 500 Index, a common component in the prices
of most equities. Using dynamic programming technique [11] the best-execution
strategy can be obtained and is given by

S∗N−k = WN−k /(k + 1) + δk XN−k (12.15)

where δk is the coefficient of XN−k which can be computed recursively, see [11] for
details.
In contrast to the case of a linear price-impact function with no information,
the best execution strategy (12.15) varies over time as a linear function of the
remaining shares WN−k and the information variable XN−k . The first term of (12.15)
is simply the naive strategy of dividing the remaining shares WN−k at time N − k
evenly over the remaining k + 1 periods. The second term of the (12.15) is an
adjustment that arises from the presence of serially correlated information XN−k .
Note that the number of shares traded at time tk is a function of Xk , and therefore
the optimal strategy obtained is dynamic in nature.
If ρ = 0, the term δk vanishes then the naive strategy is an optimal solution.
If ρ > 0, then δk is positive, implying that the positive realizations of XN−k in-
creases the number of shares purchased at time tN−k . Similarly, if ρ < 0, positive
realizations of XN−k decreases the number of shares purchased at time tN−k as δk
is negative.
Though the strategy obtained in (12.15) is dynamic in nature, the above model
suffers from a number of drawbacks that are mentioned below.
• Prices Pk are assumed to follow arithmetic random walk in (12.13) implying a
positive probability for negative prices.
• Price impact and information have only permanent effect on prices whereas
several recent empirical studies suggest that some combination of permanent
and temporary impact exists in the market.
• The percentage price impact - as a percentage of the execution price - is a
decreasing function of the price level, which is counterfactual.
To overcome above limitations, Bertsimas and Lo [11] suggested the linear per-
centage temporary price impact model.
 Dynamic Models 391

Linear-percentage Temporary Price Impact Model

In this model, we assume that the execution price at time tk comprises of two
components, the no-impact price P̃k and the price impact ∆k :

Pk = P̃k + ∆k .

The no impact price may be viewed as the price which would prevail in the
market in the absence of any market impact. To ensure that prices do not go
negative we assume geometric Brownian motion for the price dynamics P̃k and is
defined as

P̃k = P̃k−1 exp(Zk )

where Zk are i.i.d. normal random variables with mean uz and variance σ2z .
The price impact ∆k captures the effect of trade size Sk on the transaction price,
hence

∆k = (θSk + γXk )P̃k ,

Xk = ρXk−1 + ηk ,

where ηk is white noise with mean 0 and variance σ2η . We set Xk to be an AR(1)
(auto-regressive with lag 1) process. The parameters θ and γ measure the sensitiv-
ity of price impact to trade size and market conditions. The closed form solution
can be obtained using dynamic programming [11] as

S∗N−k = δwk WN−k + δxk XN−k + δ1k , (12.16)

where δwk , δxk and δ1k are fixed coefficients. We are skipping the details here but
interested readers are encourage to see [11].
The linear percentage temporary price impact model resolves a number of
problems discussed with the aforementioned models.
• First, P̃k is guaranteed to be non-negative and Pk is also guaranteed to be non
negative under mild restriction on ∆k .
• Second, by separating the transaction price Pk into a no-impact price compo-
nent P̃k and the impact component ∆k , the price impact of a trade is temporary,
moving the current transaction price but having no effect on future prices.
• Third, the percentage price impact increases linearly with the trade size.
392 Financial Mathematics: An Introduction

• Fourth, the linear percentage temporary price impact law of motion implies
a natural decomposition of execution costs, decoupling market microstructure
effects from price dynamics. The objective function can be separated into two
terms:

8
N   N  
N 
Min E1 Pk Sk = Min E1 P̃k Sk + E1 ∆k Sk .
{Sk } {Sk }
k=1 k=1 k=1

The first term is the no-impact cost of execution and second term is the total
impact cost. This decomposition is precisely the one proposed by Perold [107]
in his definition of implementation shortfall, but now applied to executing S̄.

12.8 Limitations of Dynamic Programming

The dynamic programming technique provides a general approach to find the
optimal strategy, nevertheless, it has several important limitations.
(i) Inability to impose constraints
The most important limitation of dynamic programming approach to obtain opti-
mal strategy is the inability to impose constraints in the dynamic models. In most
practical applications, there will be constraints on the kind of execution strategies
that institutional investors can follow. For example, if a block of shares is to be
purchased within time [0, T], it is very difficult to justify selling the stock during
this time period even if such sales are warranted by the best-execution strat-
egy. For a buy program, imposing a simple constraint like Sk ≥ 0, (k = 1, . . . , N)
increases the feasible intervals for calculating S∗N−k exponentially in k.
We know that for k = 0 the optimal control is S∗T = WT . At the next stage,
k = 1, the optimal control S∗T−1 is computed by minimizing a quadratic function
of ST−1 subject to the constraints 0 ≤ ST−1 ≤ WT−1 . The solution is given by
⎧
⎪
⎪ 0, if a1 WT−1 + b1 XT−1 < 0,
∗
⎪
⎨
ST−1 = ⎪
⎪ a W
1 T−1 + b X
1 T−1 , if 0 < a1 WT−1 + b1 XT−1 < WT−1 ,
⎪
⎩ WT−1 , if a1 WT−1 + b1 XT−1 > WT−1 .
This partitions the range of WT−1 into three intervals, over each interval
there is a different optimal control S∗T−1 . For example, if N = 20, there are
320 = 3, 486, 784, 401 intervals at the last stage of the dynamic program [11].
The discussion pertaining to the complexity of imposing constraint to execution
problem with linear price impact with information model for N = 3, have been
summarized in [55].
 Limitations of Dynamic Programming 393

Further, the closed form solution obtained above does not impose non-negativity
constraint and it is possible that the solution given by optimal strategy sells the
the stock and buys it again in the time interval [0, T]. This limitation of dynamic
programming can be overcome by quadratic programming approach discussed in
Section 12.6. Other types of constraint like shrinking portfolio constraint, partici-
pation rate constraint, tax-motivated constraint are also possible [77], but difficult
to work out with dynamic programming approach.

(ii) Risk averse investors

For a dynamic programming problem, the objective function must be additive-
separable. When we use mean-variance objective function, it no longer remains
additive-separable. As discussed above, one can find the optimal strategy for risk
averse investors for static models but not for dynamic models.

(iii) Curse of dimensionality

We have assumed that the price follows an arithmetic random walk or geometric
Brownian motion, while the market information is an autoregressive process of
order one. These assumptions may not hold true in real markets where more
general law of motions are observed. It may not be possible to obtain close form
solutions for general dynamics of price and market information. Therefore, some
numerical techniques have been suggested to solve such problems [11]. The most
common technique is to discretize the state and control space and use grid search
to find the optimal solution at each stage. As the number of levels for state and
control space increases, the computational requirement increases dramatically due
to which it becomes practically infeasible to solve the problem in the time frame
available to a trader. This problem is called curse of dimensionality.
Let us consider an example to understand the computational requirement of
discretization. Suppose that 100, 000 shares, currently trading at Rs 50, must be
executed over the next 20 periods. Assuming that price will be within Rs 45 to
Rs 55, with a minimum price variation of 0.125, there are 80 possible price states.
We can assume that shares are traded in a lot size of 100. Hence, there are 1, 000
lots possible. suppose that information variable can also be discretized into 10
states. Then at each step we will need to compute the optimal value function for
(1000 ∗ 80 ∗ 10) values of the state and control variables. With 20 periods this
implies a total of 16 million evaluations of the optimal value-function. Assuming
that computation of an optimal value function takes 10−6 seconds, 160 seconds
are required to find the optimal strategy which is huge as compared to the time
available to investor to make rational decision. If more refined discretization is
394 Financial Mathematics: An Introduction

needed, the computational complexity increases dramatically. For example, if we

allow the trade size to vary in increments of 1 share instead of 100 shares, and if the
information variable takes on 100 discrete values instead of 10, the computational
demands increases by a factor of 1, 000, i.e, 16, 000 seconds or roughly 4 hours.
The following table summarizes the situations or types of models which can/can
not be handelled by the dynamic programming approach.

Static models Dynamic models

Imposing constraint Yes No
Risk neutral investor Yes Yes
Risk averse investor Yes No
Table 12.1. Models which dynamic programming can solve

12.9 Quadratic Programming Approach for Optimal

Execution: Static Models
The prominent limitations of dynamic programming technique are the inability to
impose constraints and inclusion of risk in the objective function. These limitations
can be overcome by using an approximation method where we determine the
static solution and make it dynamic artificially by recomputing the static strategy
whenever there is a change in the market information. For a linear price impact
function, expectation and variance of implementation shortfall computed at any
time instant is a quadratic functions of control variables. Therefore the static
optimization problem becomes a quadratic programming problem which can be
solved using any standard optimizer like interior point algorithm. We can easily
impose any kind of constraints on this static optimization problem.
For a static model, this approximation method always gives optimal solutions
that are in line with the solution obtained with Dynamic Programming. For dy-
namic models, we show that this approximation method is very close to the opti-
mal strategy. In fact, for the unconstrained problem of risk-neutral investor, the
approximation method gives exact optimal solutions.
We will discuss the approximation method in more detail in the following sec-
tions and try to compare the results with those of optimal strategies computed
using dynamic programming.
 Quadratic Programming Approach for Optimal Execution: Static Models 395

We begin the discussion first with static models.

Arithmetic Random Walk with Linear Price Impact Model: A Static
model

Let Sk be the number of shares acquired in period tk at price Pk . The price

process is same as that is mentioned in Section 12.4

Pk = Pk−1 + θSk + k , θ > 0, E[k |Sk , Pk−1 ] = 0, (12.17)

where k is assumed to be a zero-mean (i.i.d.) random process and P0 is initial

price of stock. To illustrate the approach, we divide T into 2 intervals. Later, we
will generalize for N periods. For a risk neutral investor, the problem for two time
periods is

Min E1 [P1 S1 + P2 S2 ] (12.18)

{S1 ,S2 }

subject to
S1 + S2 = S̄
S1 , S2 ≥ 0.

Using (12.17), the price P1 and P2 can be computed as

P1 = P0 + θS1 + 1
P2 = P1 + θS2 + 2
= (P0 + θS1 + 1 ) + θS2 + 2

Therefore,

E1 [P1 S1 + P2 S2 ] = E1 [(P0 + θS1 + 1 )S1 + ((P0 + θS1 + 1 ) + θS2 + 2 )S2 ]

= P0 S1 + θS21 + P0 S2 + θS1 S2 + θS22
= P0 (S1 + S2 ) + θ(S21 + S1 S2 + S22 )

Thus, the problem (12.18) is equivalent to

Min CT S + ST QS
S
subject to
eT S = S̄
(12.19)
S ≥ 0,
396 Financial Mathematics: An Introduction

where
    
P0 θ θ/2 S
C= , Q= , S= 1 .
P0 θ/2 θ S2

Generalizing above problem for N periods, we have the following quadratic

programming problem,

Min CT S + ST QS
S
subject to
eT S = S̄
(12.20)
S ≥ 0,

where
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
⎜⎜ P0 ⎟⎟ ⎜⎜ θ θ/2 . . . θ/2 ⎟⎟ ⎜⎜ S1 ⎟⎟
⎜⎜ P ⎟⎟ ⎜⎜ θ/2 θ . . . θ/2 ⎟⎟ ⎜⎜ S ⎟⎟
⎜⎜ 0 ⎟⎟ ⎜⎜ ⎟⎟ ⎜⎜ 2 ⎟⎟
⎜ ⎟ ⎜
C = ⎜⎜ .. ⎟⎟ , Q = ⎜⎜ .. . . . ⎟⎟ , S = ⎜⎜⎜ .. ⎟⎟⎟ .
⎟
⎜⎜ . ⎟⎟ ⎜⎜ . .
. . .
. . ⎟⎟⎟ ⎜⎜ . ⎟⎟
⎜⎝ ⎟⎠ ⎜⎝ ⎠ ⎜⎝ ⎟⎠
P0 θ/2 θ/2 . . . θ SN

We encourage the readers to see that the optimal solution of above QPP will
be naive strategy, that is,
⎛ ⎞
⎜⎜ S̄/N ⎟⎟
⎜⎜ S̄/N ⎟⎟
⎜⎜ ⎟⎟
S = ⎜⎜⎜ .. ⎟⎟⎟
∗
⎜⎜ . ⎟⎟
⎜⎝ ⎟⎠
S̄/N

which is same as one obtained using dynamic programming.

For a risk averse investor, the optimization problem is

N  
N 
Min E Pk Sk + λVar Pk Sk
Sk
k=1 k=1
subject to
*N
k=1 Sk = S̄
(12.21)
S1 , . . . , SN ≥ 0 .

For two period problem,

 Quadratic Programming Approach for Optimal Execution: Static Models 397

Var[P1 S1 + P2 S2 ] = S21 Var[P1 ] + S22 Var[P2 ] + 2S1 S2 Cov[P1 , P2 ]

where

Var[P1 ] = Var[P0 + θS1 + 1 ]

= Var[1 ]
= σ2 .

Var[P2 ] = Var[P1 + θS2 + 2 ]

= Var[P0 + θS1 + 1 + θS2 + 2 ]
= Var[1 ] + Var[2 ] + 2Cov[1 , 2 ]
= 2σ2 .

Cov[P1 , P2 ] = Cov[P1 , P1 + θS2 + 2 ]

= Cov[P1 , P1 ] + Cov[P1 , θS2 ] + Cov[P1 , 2 ]
= Var[P1 ] + 0 + 0
= σ2 .

Therefore, Var[P1 S1 + P2 S2 ] = ST ΣS where

 2   
σ σ2 S1
Σ= 2 2 , S = .
σ 2σ S2

In general for N periods, we have

Var[Pk ] = kσ2 , k = 1, . . . , N,

Cov[Pi , P j ] = Var[Pi ] = iσ2 , i < j.

*N
Therefore, we get Var[ k=1 Pk Sk ] = ST ΣS where
⎛ 2 2 2 ⎞ ⎛ ⎞
⎜⎜ σ σ . . . σ ⎟⎟ ⎜⎜ S1 ⎟⎟
⎜⎜ σ2 2σ2 . . . 2σ2 ⎟⎟ ⎜⎜ S ⎟⎟
⎜⎜    ⎟
⎟⎟ ⎜⎜ 2 ⎟⎟
Σ = ⎜⎜ .. .. . . .. ⎟⎟ , S = ⎜⎜⎜ .. ⎟⎟⎟ .
⎜
⎜⎜ . . . . ⎟⎟⎟ ⎜⎜ . ⎟⎟
⎜⎝ ⎜⎝ ⎟⎠
2⎠
σ 2σ . . . Nσ
2 2
SN
398 Financial Mathematics: An Introduction

Thus, for a risk averse investor with risk aversion parameter λ, we have the
following quadratic programming problem

Min CT S + ST QS + λST ΣS
S
subject to
eT S = S̄
S ≥ 0.

Using a numerical example, the results of QP approach and DP are compared.

The following parameter values are used

θ = 5 × 10−5 , N = 20, σ = 0.125, λ = 0.00005.

The price process is simulated assuming k ∼ N(0, σ2 ). The results are shown in
Table 12.2. The solution for two techniques are exactly same as expected.

12.10 Quadratic Programming Approach for Optimal

Execution: Dynamic Models
Almgren and Chriss [2] proved that optimal strategies are piecewise static i.e.
trades are statical up to an event like change in market information, and then re-
acts explicitly to the outcome of the event. Therefore, we follow a static strategy
till a new information arrives, and then recompute the strategy using a new in-
formation. In particular, if we assume that market information is updated at each
period, then we need to recompute the strategy at each period using the updated
information. We can make the static optimization method artificially dynamic by
the following method.
• Solve QPP1 which depends on initial price and market information and get the
optimal strategy, {S(1) 1
, S(1)
2
, . . . , S(1)
N
}.
(1)
• Implement S1 in the first period, and reformulate a new QPP, say QPP2 which
depends on latest price and market information available at the end of first
period. Let {S(2)
2
, S(2)
3
, . . . , S(2)
N
} be the optimal solution of QPP2 .
• In the second period, implement S(2) 2
and reformulate QPP using updated in-
formation and so on.
In general, follow an optimal strategy till information is updated, and then
reformulate the quadratic program. This method is illustrated in the following
 Quadratic Programming Approach for Optimal Execution: Dynamic Models 399

Period P1 S1 W1 P2 S2 W2
1 50.75666 16215 100000 50.75666 16215 100000
2 51.22827 13596 83785 51.22827 13596 83785
3 51.81408 11403 70189 51.81408 11403 70189
4 52.32833 9565 58786 52.32833 9565 58786
5 52.5864 8028 49221 52.5864 8028 49221
6 53.07228 6740 41193 53.07228 6740 41193
7 53.50411 5664 34453 53.50411 5664 34453
8 53.7376 4764 28789 53.7376 4764 28789
9 53.97918 4013 24025 53.97918 4013 24025
10 54.1704 3388 20012 54.1704 3388 20012
11 54.29048 2868 16624 54.29048 2868 16624
12 54.50313 2439 13756 54.50313 2439 13756
13 54.53384 2085 11317 54.53384 2085 11317
14 54.89655 1796 9232 54.89655 1796 9232
15 54.9577 1564 7436 54.9577 1564 7436
16 55.04097 1381 5872 55.04097 1381 5872
17 55.23632 1240 4491 55.23632 1240 4491
18 55.30066 1138 3251 55.30066 1138 3251
19 55.34233 1073 2113 55.34233 1073 2113
20 55.29029 1040 1040 55.29029 1040 1040
Table 12.2. Arithmetic random walk with linear permanent market impact model. {P1, S1, W1} is the
solution using dynamic programming. {P2, S2, W2} is the solution using quadratic programming. The
two strategies are exactly same.

section and the results are compared with those of the dynamic programming
approach.

(i) Linear Price Impact with Information Model

As discussed earlier let Xk be the information variable which is modeled using
AR(1) process. The law of motion for execution price Pk is given by
Pk = Pk−1 + θSk + γXk + k , θ > 0 (12.22)

Xk = ρXk−1 + ηk , ρ ∈ (−1, 1) (12.23)

where k and ηk are independent white noise process with mean 0 and variances
σ2 and σ2η , respectively. At time t1 , initial price P0 and market information X1 are
400 Financial Mathematics: An Introduction

known. Therefore, for a risk neutral investor, the optimization problem is defined
as

N 
Min E1 Pk Sk
Sk
k=1
subject to
*N
k=1 Sk = S̄
(12.24)
S1 , . . . , SN ≥ 0.
*N
We can express E1 [ k=1 Pk Sk ] as a quadratic function (see [55]) of S1 , . . . , SN as

N 
E1 Pk Sk = CT S + ST QS ,
k=1

where C is N × 1 matrix, S is N × 1 matrix and Q is N × N matrix given by

⎛ ⎞ ⎛ ⎞
⎜⎜ P0 + γX1 ⎟⎟ ⎜⎜ S1 ⎟⎟
⎜⎜ P0 + γX1 (1 + ρ) ⎟
⎟ ⎜⎜ S ⎟⎟
⎜⎜ ⎟⎟ ⎜⎜ 2 ⎟⎟
C = ⎜⎜ ⎜ . ⎟
⎟ , S = ⎜⎜ . ⎟⎟ ,
⎜⎜ .. ⎟
⎟⎟ ⎜⎜ .. ⎟⎟
⎜⎝ ⎟⎠ ⎜⎜ ⎟⎟
⎝ ⎠
P0 + γX1 (1 + ρ + . . . + ρN−1 ) SN

⎛ ⎞
⎜⎜ θ θ/2 . . . θ/2 ⎟⎟
⎜⎜ θ/2 θ . . . θ/2 ⎟⎟
⎜⎜ ⎟⎟
Q = ⎜⎜⎜ .. . . . ⎟
.. . . .. ⎟⎟⎟⎟ .
⎜⎜ .
⎜⎝ ⎟⎠
θ/2 θ/2 . . . θ

We urge the readers to work out the above details. Therefore, a risk neutral
investor solves the following quadratic programming problem (QPP1 ) at time t1

Min CT S + ST QS
S
subject to
eT S = S̄
(12.25)
S ≥ 0.

Let {S(1)
1
, S(1)
2
, . . . , S(1)
N
} be optimal strategy for optimization problem (12.25).
(1)
The investor acquires S1 shares at time t1 . At the beginning of time t2 the price
 Quadratic Programming Approach for Optimal Execution: Dynamic Models 401

and information parameters are updated. Let P1 and X2 be price and information
available at time t1 . Moreover, the number of shares to be acquired at time t1
are W2 = S̄ − S(1)
1
. So, the investor reformulates the quadratic program (QPP2 ) as
follows

Min CT S + ST QS
S
subject to
eT S = W2
(12.26)
S ≥ 0,

where C is (N − 1) × 1 matrix, Q is (N − 1) × (N − 1) matrix and S is (N − 1) × 1

matrix given by
⎛ ⎞ ⎛ ⎞
⎜⎜ P0 + γX2 ⎟⎟ ⎜⎜ S2 ⎟⎟
⎜⎜ P0 + γX2 (1 + ρ) ⎟
⎟ ⎜⎜ S ⎟⎟
⎜⎜ ⎟⎟ ⎜⎜ 3 ⎟⎟
C = ⎜⎜ ⎜ . ⎟
⎟ , S = ⎜⎜ . ⎟⎟ ,
⎜⎜ .. ⎟
⎟⎟ ⎜⎜ .. ⎟⎟
⎜⎝ ⎟⎠ ⎜⎜ ⎟⎟
⎝ ⎠
P0 + γX2 (1 + ρ + . . . + ρN−2 ) SN

⎛ ⎞
⎜⎜ θ θ/2 . . . θ/2 ⎟⎟
⎜⎜ θ/2 θ . . . θ/2 ⎟⎟
⎜⎜ ⎟⎟
Q = ⎜⎜⎜ .. . . . ⎟
.. . . .. ⎟⎟⎟⎟ .
⎜⎜ .
⎜⎝ ⎟⎠
θ/2 θ/2 . . . θ

Let {S(2)
2
, S(2)
3
, . . . , S(2)
N
} be the optimal solution of (12.26). Now, the investor acquires
(2)
S2 shares at time t2 and update the remaining number of shares to be purchased
at time t3 to W3 = W2 − S(2) 2
. Using latest price and information available at
time t2 , investor reformulates the quadratic program. Therefore, at each step, a
new optimization problem is solved. At the last period, the investor acquires the
remaining number of shares.
For N = 5, the results of all quadratic programs are shown in Table 12.3. The
following parameter values are used for calculation

θ = 5 × 10−5 , ρ = 0.50, γ = 5.0, σ2 = (0.125)2 , σ2η = 0.001 (12.27)

k and ηk are assumed to be drawn from normal distribution.

If we follow a similar method for a static model like the basic model of Bertsimas
and Lo, the number of shares to be traded in time period t2 obtained from QPP1
402 Financial Mathematics: An Introduction

Time QPP1 QPP2 QPP3 QPP4 QPP5

t1 22307
t2 20424 22421
t3 19482 19600 19550
t4 19011 18189 18199 18457
t5 18776 17483 17523 17264 17264
Table 12.3. Static strategy at each time instant for N = 5 for dynamic model. The ith column denotes
the solution of (QPPi ). The number of shares to be traded in period t2 obtained from (QPP1 ) is not equal
to that obtained from (QPP2 ). Since, we assume that market information is updated at each period, the
diagonal gives the final strategy.

will be same as that obtained from QPP2 . This can be seen in Table 12.4. The
following parameter values are used for calculation

θ = 5 × 10−5 , N = 5, λ = 1 × 10−5 , σ = 0.125.

Time QPP1 QPP2 QPP3 QPP4 QPP5

t1 20740 0 0 0 0
t2 20244 20244 0 0 0
t3 19876 19876 19876 0 0
t4 19631 19631 19631 19631 0
t5 19509 19509 19509 19509 19509
Table 12.4. Static strategy for N = 5 for a static model. The ith column denotes the solution of (QPP1 ).
The number of shares to be traded in each period tk are same for all QPP’s.

The results of dynamic programming approach and static approximation ap-

proach for a risk neutral investor are compared for various cases. The parameters
used are given in (12.27).
In what follows we use the abbreviations DP and QP for the dynamic program-
ming and the quadratic programming respectively.
Case 1 Unconstrained DP and Unconstrained QP
The strategies given by two techniques are exactly the same as depicted in Fig.
12.4.
 Quadratic Programming Approach for Optimal Execution: Dynamic Models 403

Unconstrained DP and Unconstrained QP

9000
S
1
S
8000 2

7000

6000
Shares Traded (S)
5000

4000

3000

2000

1000

0
0 2 4 6 8 10 12 14 16 18 20
Time Interval (T)

Fig. 12.4. Best execution strategy without constraints for N = 20 using DP and QP. The black dotted
curve represents DP strategy while the black bold curve represents QP strategy. The two curve coincides.

Case 2 Unconstrained DP and Constrained QP

In Fig 12.5, the optimal strategy using DP is never negative and the two strate-
gies are exactly same. In Fig. 12.6, the two strategies follow the same trend. The
strategy from DP at time t3 , t22 , t23 , t24 are negative while it is zero from QP ap-
proach. These figures clearly indicate that static approximation procedure gives
very similar strategy to dynamic programming.

Unconstrained DP and Constrained QP

9000
S1
8000
S
2
7000
Shares Traded (S)

6000

5000

4000

3000

2000

1000

0
0 2 4 6 8 10 12 14 16 18 20
Time Interval (T)

Fig. 12.5. Best execution strategy for N = 20 using DP and QP. The black dotted curve represents
DP strategy without constraints while the black bold curve represents QP strategy with non-negativity
constraints. S1 is never negative and the two curve coincides.
404 Financial Mathematics: An Introduction

Unconstrained DP and Constrained QP

10000
S1
S
8000 2

6000

4000

Shares Traded
2000

−2000

−4000

−6000

−8000
0 5 10 15 20 25 30
Time Insterval (T)

Fig. 12.6. Best execution strategy for N = 30 using DP and QP. The black dotted curve represents
DP strategy with no constraints while the black bold curve represents QP strategy with non-negativity
constraint. The effect of imposing non-negativity constraint is clearly visible. The strategies follow same
trend, but QP strategy becomes zero as DP strategy goes negative.

Case 3 Constrained DP and Constrained QP

Constrained DP is solved for N = 3 with the following parameters,

S̄ = 10000, θ = 5 × 10−5 , ρ = 0.50, γ = 5.0, σ2 = (0.125)2 , σ2η = 0.013

and the results are shown in Table 12.5. The static approximation procedure gives
strategy similar to one given by dynamic programming without non-negativity
constraint, but the optimal strategy with non-negativity constraints is different.
The optimal strategy by dynamic programming with non-negativity constraints
for time t1 takes into account that there is a finite probability that strategy at time
t2 goes negative. Therefore, optimal solution trades more at time t1 as compared
optimal solution without constraint.
For a risk averse investor, the quadratic programming problem at time period
t1 will be

Min CT S + ST QS + λST ΣS
S
subject to
eT S = S̄
S≥0
 Quadratic Programming Approach for Optimal Execution: Dynamic Models 405

Time DP without constraint QP with constraint DP with constraint

t1 2162 2162 3643
t2 861 861 120
t3 6977 6977 6237
Total 10000 10000 10000
Table 12.5. Comparison of best execution strategy for unconstrained DP, constrained QP and con-
strained DP.

where the matrices C and S are defined in (QPP1 ) above, and Σ is variance-
covariance matrix of price Pk are discussed in [55]. As before, we reformulate the
quadratic program at each time period to get a dynamic strategy. Fig. 12.7 shows
the best execution strategy with non-negativity constraints obtained using static
approximation procedure. As λ is increased, the strategy becomes more and more
aggressive, that is, trade more in the beginning to reduce the risk. For lower values
of λ, the strategy follows a similar trend as market information variable Xk (black
bold line with ).

4
0x 10 2 4 Effect6 of risk Aversion
8 parameter
10 12on the strategy
14 16 18 20
8 0.08
λ=0.001
7 λ=0.0001 0.06
λ=0.00001
6 λ=0.000001 0.04
Shares Purchased (S)

5
0.02
4
0
3
−0.02
2

1 −0.04

0 −0.06
0 2 4 6 8 10 12 14 16 18 20
Time Interval (T)

Fig. 12.7. Effect of λ on best execution strategy

(ii) Linear-Percentage Temporary Price Impact Model

In this model, the no-impact price P̃k is modelled using the geometric Brownian
motion and market information is modeled using AR(1) process:
406 Financial Mathematics: An Introduction

P̃k = P̃k−1 exp(Zk ) , (12.28)

Xk = ρXk−1 + ηk , (12.29)

where Zk are i.i.d. normal random variables with mean uz and variance σ2z and ηk
is white noise with mean 0 and variance σ2η . The execution price Pk at time tk is
comprised of two components, the no-impact price P̃k and the price impact ∆k

Pk = P̃k + ∆k , (12.30)

∆k = (θSk + γXk )P̃k . (12.31)

At time t1 , initial price P̃0 and market information X1 are known. Thus, a risk
neutral investor solves the following quadratic programming problem with Q and
C defined in [55] at time t1

Min CT S + ST QS
S
subject to
eT S = S̄
(12.32)
S ≥ 0,

where
⎛ ⎞ ⎛ ⎞
⎜⎜ qP̃0 (1 + γX1 ) ⎟⎟ ⎜⎜ S1 ⎟⎟
⎜⎜ 2 ⎟ ⎜⎜ S ⎟⎟
⎜⎜ q P̃0 (1 + γρX1 ) ⎟⎟⎟ ⎜⎜ 2 ⎟⎟
C = ⎜⎜ ⎜ .. ⎟⎟ , S = ⎜⎜⎜ .. ⎟⎟⎟ ,
⎜⎜ ⎟⎟ ⎜⎜ . ⎟⎟
⎜⎝ . ⎟⎟ ⎜⎝ ⎟⎠
⎠
qN P̃0 (1 + ρN−1 γX1 ) SN

⎛ ⎞
⎜⎜ θqP̃0 0 . . . 0 ⎟⎟
⎜⎜ ⎟
⎜⎜ 0 θq2 P̃0 . . . 0 ⎟⎟⎟
⎜
Q = ⎜⎜ .. .. . . .. ⎟⎟⎟⎟ .
⎜⎜ . . . . ⎟⎟
⎜⎝ ⎠
0 0 . . . θq P̃0
N

In this model Q is a diagonal matrix. This is due to the fact that we consider
only temporary market impact. The number of shares traded at time t1 does not
affect the future price. Therefore, the execution price P2 for shares traded at time
 Quadratic Programming Approach for Optimal Execution: Dynamic Models 407

t2 is independent of number of shares traded at time t1 . As before, market infor-

mation is updated at each time period, so we need to reformulate the quadratic
program at each period.
The strategy obtained using DP approach and QP approach can be compared
using numerical example with following parameters that were used by Bertsimas
and Lo [11].
√
µz = 0, σz = 0.02/ 13, θ = 5 × 10−7 , γ = 0.01,

ρ = 0.50, η ∼ N(0, 1 − ρ2 ), σ2η = 1 − ρ2 .

For N = 5, the strategies for all the quadratic programs are shown in Table
12.6.

Time QPP1 QPP2 QPP3 QPP4 QPP5

t1 28703.59 0 0 0 0
t2 24075.95 21608.49 0 0 0
t3 16223.69 18198.27 18053.14 0 0
t4 12415.06 15646.57 15251.64 16267.66 0
t5 18581.7 18581.7 15350.2 13770.54 15367.11

Table 12.6. Dynamic strategy for N = 5 for Model 3. The ith column denotes the solution of (QPPi ).
The number of shares to be traded in period t2 obtained from (QPP1 ) is not equal to that obtained from
(QPP2 ). For each period, we use the latest market information, and solve the corresponding QP and
follow this strategy till new information comes. Since, we assume that market information is updated
at each period, the diagonal gives the number of shares to trade in that period.

In Fig. 12.8, the two strategies obtained using DP and QPP approach are
exactly same as the optimal strategy using DP is never negative whereas in Fig.
12.10, the two strategies differ.
408 Financial Mathematics: An Introduction

4
x 10
3
S1
S2
2.5

Shares Traded (S)

2

1.5

0.5

0
1 2 3 4 5 6 7 8 9 10
Time Interval (T)

Fig. 12.8. Best execution strategy for N = 10 for Model 3 (Linear percentage temporary price impact)
using DP and QP. The black dotted curve represents DP strategy with no constraints. The black bold
curve represents QP strategy with non-negativity constraint. Since S1 is never negative, the two curve
coincides.

0.5
X

−0.5

−1

−1.5

−2

−2.5
1 2 3 4 5 6 7 8 9 10

Fig. 12.9. Market information, AR(1) process

4
x 10
2
S1
S2

1.5

1
Shares Traded

0.5

−0.5

−1
0 2 4 6 8 10 12 14 16 18 20
Time Interval (T)

Fig. 12.10. Best execution strategy for N = 10 for Model 3 (Linear percentage temporary price impact)
using DP and QP. The black dotted curve represents DP strategy no constraints. The black bold curve
represents QPP strategy with non-negativity constraint.
 Summary and Additional Notes 409

1.5
X
1

0.5

−0.5

−1

−1.5

−2
0 2 4 6 8 10 12 14 16 18 20

Fig. 12.11. Market information, AR(1) process

12.11 Summary and Additional Notes

• Section 12.2 introduces the basics of market structure. Here various components
of transaction cost are described. Further Section 12.3 presents various types
of orders in the market are introduced. Readers may refer to Haris [59] and,
Kissell and Glantz [77] for more details in this regard.
• The problem of optimal execution is presented in Section 12.4. The whole
discussion in this chapter centers around the study of this problem. There are
two major contributions in this area. These are due to Bertsimas and Lo [11]
and Almgren and Chriss [2]. We have presented the dynamic programming
approach of Bertsimas and Lo [11] in detail, both for risk neutral and risk
averse trader. While static models are discussed in Sections 12.5 and 12.6, the
dynamic models are presented in Section 12.7. Section 12.8 also discussed the
limitations of dynamic programming.
• A quadratic programming approach for optimal execution is presented in Sec-
tion 12.9 and Section 12.10. The contents of these two sections are based on
Gupta and Choudhary [55] and Khemchadani et al [76].
• Algorithmic trading in general, and optimal trading strategies in particular
have now become very important in trading scenario. In this chapter we have
illustrated the applications of dynamic programming and quadratic program-
ming for determining optimal trading strategies. Various machine learning
methodologies have also found applications in this area. One of these is rein-
forcement learning. The readers may refer to Sutton and Barto [132], Kaelbling
410 Financial Mathematics: An Introduction

and Littman [73] for basics of reinforcement learning. Nevmyvaka et al. [101]
provides a good application of reinforcement learning in the area of optimal
trading strategies.
• This chapter presents optimal trading strategies in the context of a single asset
only. But a more realistic case in that of a portfolio. Bertsimas et al. [12], and
Aitsahlia et al. [1] discuss optimal execution for portfolios.
• Similar to the concept of efficient frontier in the Markowitzs’ model of portfolio
optimization, Kissell and Glantz [77] introduced the concept of efficient trading
frontier in the context of optimal trading strategies.

12.12 Exercises
Exercise 12.1 Compute the associated transaction cost of a buy order for 100,000
shares of ABC if the price at the beginning of trading was Rs 50 per share and the
average execution price was Rs 52 per share and all shares were executed.
Exercise 12.2 Consider the set of parameters,

S̄ = 10000, θ = 5 × 10−5 , ρ = 0.50, γ = 5.0, σ2 = (0.125)2 , σ2η = 0.013

Plot the efficient frontier for various values of λ lies between 0 and 1 using all the
price process mentioned in the chapter.
13
Credit Risk Management

13.1 Introduction
Credit risk is the risk of a trading partner for not meeting its obligations in full
on the due date or at any time thereafter. Such an event is called a default, hence
another term for credit risk is default risk . There are three well known types
of credit risk: default risk/counterparty risk, credit spread risk and downgrade
risk. There are several ways through which credit risk can be taken into account
within commercial and retail banking activities. To name some, deposits, mortgage
lending and credit cards are part of retail banking, whereas commercial banking
includes loans, letters of credit and asset finance.
For example, in a credit card issuance financial service/bank agrees to pay
retailers for purchases made by credit card holders in exchange for an unsecured
promise to repay the card balance. Thus, any given debt has a short expected life,
as when card holders pay off some of their balances and then buy new goods, new
exposures roll in to place of old ones.
Financial services/banks need a mechanism to quantify the risk factors relevant
for an obligor’s ability and willingness to pay. Credit scoring has been used as
a tool in modern banking to access the risk factors due to the large number of
applications received on a daily basis and the increased regulatory requirements for
banks. In this chapter, we shall present some of the basic concepts with emphasis
on statistical and data mining algorithms, and their application to credit scoring.
Risk management is a core activity conducted by banks, insurance and in-
vestment companies, or any financial institution that evaluate risk due to losses.
Losses can result from either counter-party default, or from a decline in market
value stemming from the credit quality migration of an issuer or counter-party.
There are two primary types of models that attempt to describe default pro-
cesses in the credit risk literature. These are (i) structural models and (ii) reduced
412 Financial Mathematics: An Introduction

form models. Structural models have been pioneered by Black-Scholes and Merton.
The basic idea, common to all structural type models is that a company defaults
on its debt if the value of the assets of a company falls below a certain default
point. For this reason, these models are also known as firm-value models. In these
models it has been demonstrated that default can be modeled as an option and,
as a result, researchers have been able to apply the same principles used for option
pricing to the valuation of risky corporate securities.
The second group of credit models, known as reduced form models, are more
recent. These models, most notably the Jarrow-Tunbull and Duffie and Singleton
models, do not look inside the firm. Instead, they model directly the likelihood
of default or downgrade. Apart from modeling the current probability of default,
some researchers attempt to model a ‘forward curve’ of default probabilities. This
can be used to price instruments of varying maturities.
In recent past, managing credit risk at portfolio level has influenced many
researchers. In order to utilize the advantage of the credit portfolio one should
know the risk of portfolio and the factors that affect the portfolio risk profile. Thus
portfolio managers are interested in knowing the effect of changing the portfolio
mix, how would risk-based pricing at the individual contract and the portfolio
level be influenced by the level of expected losses and credit risk capital?
Traditionally used tools for assessing and optimizing market risk assume that
the portfolio return-loss is normally distributed. With this assumption, the two
statistical measures, mean and standard deviation, could be used to balance return
and risk. However, to cope with skewed return-loss distributions, conditional value-
at-risk (CVaR) has been introduced as the risk measure. This measure is also
known as mean excess loss, mean shortfall, or tail VaR. By definition, (1 − β)-
CVaR is the expected loss exceeding (1 − β)-value-at-risk (VaR), i.e. it is the mean
value of the worst β × 100% losses. For instance, at (1 − β) = 0.95, or equivalently
β = 0.05, CVaR is the average of the 5% worst losses.
In this chapter we focus on the quantification of credit risk. We will discuss
different approaches to estimate the probability that a company will default. In the
subsequent sections we will explain how a bank or other financial institution can
estimate its loss given that the default has occurred. Last part of the chapter also
covers credit rating migration and default correlation, and application of machine
learning techniques in the credit scoring scenario.
 Basic Terminology 413

13.2 Basic Terminology

To begin with we shall discuss some basic definitions that would help us to under-
stand the quantification of credit risk and procedure of calculating the probability
that company would default.
A valuation method is used to estimate the attractiveness of an investment
opportunity. Discounted cash flow (DCF) analysis uses future free cash flow pro-
jections and discounts them (most often using the weighted average cost of capital)
to arrive at a present value, which is used to evaluate the potential for investment.
If the value arrived at through DCF analysis is higher than the current cost of
the investment, the opportunity may be a good one.
Definition 13.2.1 (Discounted Cash Flow) Let, for i = 1, 2 . . . n, CFi denote
the cash flow for the i-th period (say year) and r be the continuing compounded
interest rate. Then the discounted cash flow (DCF) is defined as

CF1 CF2 CFn

DCF = + + ... + .
(1 + r) 1 (1 + r) 2 (1 + r)n

Definition 13.2.2 (Credit Premium) The credit premium is the discounted

value of cash flows, when there is probability of default. It is given by
CF1 q1 CF2 q2 CFn qn
CP = + + ... + ,
(1 + r) 1 (1 + r) 2 (1 + r)n

where qi denotes the probability that the counter party is solvent (not refutable) at
the ith period and CFi and r are defined as above.
The credit/default premium is paid by companies with lower grade bonds or
by individuals with poor credit. As an illustration, companies with poor financial
state will tend to compensate investors for the additional risk by issuing bonds
with high yields. Individuals with poor credit must pay higher interest rates in
order to borrow money from the bank.
Definition 13.2.3 (Credit Spread) Credit spread is the yield spread, or differ-
ence in yield between different securities, due to different credit quality. The credit
spread reflects the additional net yield an investor can earn from a security with
more credit risk relative to one with less credit risk. The credit spread of a particu-
lar security is often quoted in relation to the yield on a credit risk-free benchmark
security or reference rate, typically either U.S. Treasury bonds or LIBOR.
414 Financial Mathematics: An Introduction

For instance, the credit spread could be the difference between yields on gov-
ernment bonds and those on single A-rated industrial bonds. A company must
offer a higher return on their bonds because their credit is worse in comparison
to government bonds.
Definition 13.2.4 (Recovery Rate: Loss Given Default) Recovery rate is
defined as the proportion of the “claimed amount” received in the the event of a
default Loss-given-default is the percentage we expect to lose when default occurs.
Obviously R=1-LGD.
When default occurs, a portion of the value of the portfolio can usually be
recovered. Because of this, a recovery rate is always considered when evaluating
credit losses. It represents the percentage value which we expect to recover, given
default.
Definition 13.2.5 (Credit Exposure).It is the maximum loss that a portfolio
experience at the time of default taken with a certain confidence level.
Definition 13.2.6 (Default Probability) Probability of default is the likelihood
that a loan will not be repaid.
A bank assigns to every customer a default probability (DP); a loss fraction
called the loss given default (LGD); describing the fraction of the loans exposure
expected to be lost in case of default, and the exposure at default (EAD) expected
to be lost in the considered time period. The loss of any obligor is then defined by
a loss variable L̃ = EAD × LGD × L with L = 1D × P(D) = DP, where D denotes the
event that the obligor defaults in a certain period of time (most often one year),
1D is the characteristic function of D and P(D) denotes the probability of D.
The task of assigning a default probability to every customer in a bank credit
portfolio is far from being easy. There are essentially two approaches to compute
default probabilities: calibration of default probabilities from ratings and calibra-
tion of default probabilities from market data. We intend to discuss these in the
subsequent sections.

13.3 Expected Default Losses on Bonds

The first step in estimating default probabilities from bond prices is to calculate
the expected default losses on corporate bonds of different maturities. This in-
volves comparing the price of a corporate bond with the price of a risk-free bond
that has the same maturity and pays the same coupon. The usual assumption is
 Expected Default Losses on Bonds 415

that the present value of the cost of defaults equals the excess of the price of the
risk free bond over the price of the corporate bond.
Probability of Default Assuming no Recovery
Let y(T) be the yield on a T year corporate zero-coupon bond, and y∗ (T) be
the yield on a T year risk free zero-coupon bond. Let D(T) be the probability that
corporation will default between time zero and time T.
∗
The value of a T year risk free zero coupon bond with principal 100 is 100e−y (T)T
while the value of similar corporate bond is 100e−y(T)T . The expected loss from
∗
default is therefore 100(e−y (T)T − e−y(T)T ).
If we assume that there is no recovery in the event of default, the calculation
of D(T) is relatively easy. There is a probability D(T) that the corporate bond
will be worth zero at maturity and a probability 1 − D(T) that it will be worth
Rs 100. The value of the bond is therefore {[D(T) × 0] + [(1 − D(T)) × 100]e−y∗(T)T }.
The yield on the bond is y(T), so that

100e−y(T)T = 100(1 − D(T))e−y∗(T)T .

The above equation gives

e−y∗(T)T − e−y(T)T
D(T) =
e−y∗(T)T
i.e.
D(T) = 1 − e−[y(T)−y∗(T)]T .

Probability of Default Assuming Recovery

Suppose that the claimed amount of the bond is the no-default value of the
bond. In the event of a default, the bondholder receives a proportion R of the
bond’s no-default value. While in no default case, the bondholder receives Rs 100.
The bond’s no-default value is 100(e−y∗(T)T ) and the probability of default is D(T).
The value of bond is therefore

{[D(T) × 100Re−y∗(T)T ] + [(1 − D(T)) × 100]e−y∗(T)T }.

The yield on the bond is y(T), hence

100e−y(T)T = {[D(T) × 100Re−y∗(T)T ] + [(1 − D(T)) × 100]e−y∗(T)T },

which on simplification gives

 416 Financial Mathematics: An Introduction

1 − e−[y(T)−y∗(T)]T
D(T) = .
1−R
The above estimation of default probability is based on the following set of
assumptions
(i) Amount claimed in the event of default equals the no-default value of the bond.
(ii) Claim made in the event of default equals the bond’s face value plus accrued
interest.
(iii) Probabilities are calculated on zero coupon bonds.
Over the time, the bonds are liable to move from one rating category to an-
other. This is sometime referred as credit rating migration. Rating agencies gen-
erate, from historical data, a rating transition matrix whose entries correspond to
percentage probability of a bond moving from one rating to another during a cer-
tain period of time. Next we present an approach termed CreditMetrics proposed
for estimating risks associated with default.

13.4 CreditMetrics Analysis

JP Morgan first published and well publicized CreditMetrics model in 1997 [99].
CreditMetrics computes the forward distribution of values of the loan portfolio,
given that the transition probabilities are estimated from historical data, and that
the correlation between the behavior of the rating classes are directly obtained
from the stock market data.
CreditMetrics has several aims; two of which are the creation of a benchmark
for measuring credit risk and the increase in market liquidity. If the former aim is
achieved then it becomes possible to measure risks systematically across instru-
ments and, at the very least, to make relative value judgments. From this follows
the second aim. Once instruments, and in particular the risks associated with
them, are better understood they would appear less scary to investors and hence
would help in promoting liquidity.
The CreditMetrics data set can be freely downloaded from www.jpmorgan.com.
This website also contains detailed study of the CreditMetrics methodology. The
data set consists of four data types: yield curves, spreads, transition matrices and
correlations. The CreditMetrics yield curve data set consists of the risk-free yield
to maturity for major instruments. It contains yields for maturities of one, two,
three, five, seven, ten and thirty years. For each credit rating, the data set gives
the spread above the riskless yield for each maturity. It is generally observed that
 CreditMetrics Analysis 417

riskier the bond higher the yield; thus the yield on the BBB bond is higher than
that on the AA bond which is in turn higher than the risk-free yield. Thus higher
yield for risky bonds is compensation for the possibility of not receiving future
coupons or the principal.
In the CreditMetrics framework, the transition matrix has its entries as the
probability of a change of credit rating at the end of a given time horizon, for
example the probability of a upgrade from AA to AAA might be 5.5%. The
time horizon for the CreditMetrics data set is one year. Unless the time horizon
is very long, the highest probability is typically for the bond to remain at its
initial rating. In the risk-free yield,the spreads and the transition matrix, contain
sufficient information for the CreditMetrics method to derive distributions for the
possible future values of a single bond. However, when it comes to examine the
behavior of a portfolio of risky bonds, we must consider whether there is any
relationship between the re-rating or default of one bond to another. In other
words, are bonds issued by different companies or governments are some sense
correlated? This is where the CreditMetrics correlation data set comes in. This
data set gives the correlations between major indices in many countries.
The CreditMetrics methodology is about calculating the possible values of a
risky portfolio comprising of bonds at some time in future (the time horizon) and
estimating the probability of occurrence of default. There are many credit rat-
ing agencies who compile data on individual companies or countries and estimate
the likelihood of default. The most famous of these are Standard & Poor’s and
Moody’s. These agencies assign a credit rating or grade to firms as an estimate
of their creditworthiness. Standard & Poor’s rate businesses as one of AAA, AA,
A, BBB, BB, B, CCC or Default. Moody’s use Aaa, Aa, A, Baa, Ba, B, Caa,
Ca, C. Both these companies also have finer grades within each of these primary
categories. The Moody’s grades are described in the Table 13.1 below.

Transition matrices, refer to Table 13.2, measure rating movements over time.
The diagonal of a transition matrix represents the share of ratings that remain
unchanged during the course of the reference period of one year. For instance,
84.67% of issuers rated in the ‘A’ category at the beginning of 2009 were still
rated ‘A’ by year-end, compared with only 69.34% of issuers in the ‘B’ category
(see Table 13.2). Here NR refers to nondefault rate. The probabilities are based
on historical data and therefore real world probabilities.
For the illustration of the concepts we shall be using the transition matrix
provided by Standard & Poor’s as on 15th March 1996, see Table 13.3.
The credit rating agencies continuously gather data on individual firms and de-
418 Financial Mathematics: An Introduction

The Meaning of Moody’s Ratings

Aaa -Bonds of best quality. Smallest degree of risk.
-Interest payments protected by a large or stable margin.
Aa -High quality. Margin of protection lower than Aaa.
A -Many favorable investment attributes.
-Adequate security of principal and interest.
-May be susceptible to impairment in the future.
Baa -Neither highly protected nor poorly secured.
-Adequate security for the present.
-Lacking outstanding investment characteristics.
-Speculative features.
Ba Speculative elements.
-Future not well assured.
B Lack characteristics of a desirable investment.
Caa Poor standing.
-May be in default or danger with respect to principal or interest.
Ca High degree of speculation. Often in default.
C Lowest-rated class.
-Extremely poor chance of ever attaining any real investment standing.
Table 13.1. Moody’s Categories

Initial Rating AAA AA A BBB BB B CCC/C D NR

2009
AAA 87.65 8.64 0 0 0 0 0 0 3.7
AA 0 86.17 15.96 0.64 0.21 0 0 0 7.02
A 0 0.36 84.67 7.74 0.43 0.29 0 0.21 6.3
BBB 0 0 2 83.71 5.94 0.8 0.2 0.53 6.81
BB 0 0 0 3.09 72.95 11.48 0.6 0.7 11.18
B 0 0 0.16 0 2.29 69.34 8.42 10.14 9.65
CCC/C 0 0 0 0 0 6.32 27.37 48.42 17.89
Table 13.2. One year (2009) transition matrix of percentage probabilities
 Credit VaR and Credit CVaR 419

Initial Rating AAA AA A BBB BB B CCC D

1996
AAA 90.81 8.33 0.68 0.06 0.12 0 0 0
AA 0.7 90.65 7.79 0.64 0.06 0.14 0.02 0
A 0.09 2.27 91.05 5.52 0.74 0.26 0.01 0.06
BBB 0.02 0.33 5.95 86.93 5.3 1.17 1.12 0.18
BB 0.03 0.14 0.67 7.73 80.53 8.84 1 1.06
B 0 0.11 0.24 0.43 6.48 83.46 4.07 5.2
CCC 0.22 0 0.22 1.3 2.38 11.24 64.86 19.79
Table 13.3. One year (1996) transition matrix of percentage probabilities

pending on the information, grade/regrade a company according to well-specified

criteria. A change of rating is called a migration and has an important effect on the
price of bonds issued by the company. Migration to a higher rating will increase
the value of a bond and decrease its yield, since it is seen as being less likely to
default. Clearly there are two stages involved in modeling risky bonds under the
credit-rating scenario. First, we must model the migration of the company from
one grade to another and second we must price bonds taking into account the mi-
gration factor. The main objective of the CreditMetrics framework is to produce
a consistent forward distribution of changes in value of the corporate bond for an
arbitrary one year horizon.

13.5 Credit VaR and Credit CVaR

CreditMetrics/CreditVaR are methodologies based on the estimation of the for-
ward distribution of the changes in value of a portfolio of loan and bond type
products at a given time horizon, usually 1 year. The changes in value are related
to eventual migrations in credit quality of the obligor, both up and downgrades,
as well as default.
Unlike to market-VaR, Credit-VaR posses two new challenging difficulties.
First, the portfolio distribution is far from being normal, and second, measur-
ing the portfolio effect due to credit diversification which is much more complex
than for market risk. While it is legitimate to assume normality of the portfolio
changes due to market risk, it is no longer valid for the case for credit returns
which are by nature highly skewed and fat-tailed. Indeed, there is limited upside
420 Financial Mathematics: An Introduction

to be expected from any improvement in credit quality, while there is substantial

downside consecutive to downgrading and default. The percentile levels of the dis-
tribution cannot be any longer estimated from the mean and variance only. The
calculation of VaR for credit risk requires simulating the full distribution of the
changes in portfolio value. This, together with lack of historical data to estimate
credit correlations, posses significant modeling challenges compared to market risk
modeling.
To measure the effect of portfolio diversification we need to estimate the cor-
relations in credit quality changes for all pairs of obligors. But these correlations
are not directly observable. CreditMetrics/CreditVaR base their evaluation on the
joint probability of asset returns, which itself results from strong simplifying as-
sumptions on the capital structure of the obligor, and on the generating process
for equity returns. This is clearly a key feature of CreditMetrics/CreditVaR on
which we will elaborate in the sequel.
The first step is to specify a rating system, with rating categories, together
with the probabilities of migrating from one credit quality to another over the
credit risk horizon. The details have been discussed in the previous section. We
next present an example that explain the steps in calculating CreditVaR.

CreditVaR of the Bond

Let the problem be to find CreditVaR for a senior unsecured BBB rated bond
maturing exactly in 5 years and paying an annual coupon of 6%. Following are
the steps involved in calculating Credit-VaR.

Step 1: Specify the transition matrix.

The bond issuer has currently a BBB rating, 13.4 shows the probabilities esti-
mated by Standard & Poors for a BBB issuer to be, in 1 year from now, in one
of the 8 possible states, including default. Obviously, the most probable situation
is for the obligor to stay in the same rating category, i.e. BBB, with a probability
of 86.93%. The probability of the issuer defaulting within 1 year is only 0.18%,
while the probability of being upgraded to AA is also very small, i.e. 0.33%. Such
transition matrix is produced by the rating agencies for all initial ratings. Default
is an absorbing state, i.e. an issuer who is in default stays in default.
Step 2: Specify the credit risk horizon.
The risk horizon is usually 1 year, and is consistent with the transition matrix
shown in Table 13.4. But this horizon is purely arbitrary, and is mostly dictated
by the availability of the accounting data and financial reports processed by the
 Credit VaR and Credit CVaR 421

rating agencies.
Step 3: Specify the forward pricing model.
The valuation of a bond is derived from the zero-curve corresponding to the
rating of the issuer. Since there are seven possible credit qualities, seven “spread”
curves are required to price the bond in all possible states, all obligors within
the same rating class being marked-to-market with the same curve. The spot zero
curve is used to determine the current spot value of the bond. The forward price of
the bond in 1 year from now is derived from the forward zero-curve, 1 year ahead,
which is then applied to the residual cash flows from year one to the maturity of
the bond. Table 13.4 gives the 1-year forward zero-curves for each credit rating.

Category Year 1 Year 2 Year 3 Year 4

AAA 3.60 4.17 4.73 5.12
AA 3.65 4.22 4.78 5.17
A 3.72 4.32 4.93 5.32
BBB 4.1 4.67 5.25 5.63
BB 5.55 6.02 6.78 7.27
B 6.05 7.02 8.03 8.52
CCC 15.05 15.02 14.03 13.52
Table 13.4. Source: CreditMetrics, JP Morgan. One year forward zero curve for each credit rating (%)

The 1-year forward price of the bond, if the obligor stays BBB, is then
6 6 6 106
VBBB = 6 + + 2
+ 3
+ = 107.55 .
1.0410 (1.0467) (1.0525) (1.0563)4
If we replicate the same calculations for each rating category we obtain the
values shown in Table 13.5
Step 4: Derive the forward distribution of the changes in portfolio value.
The distribution of the changes in the bond value, at the 1-year horizon, due
to an eventual change in credit quality is shown in the Table 13.6. The first
entry in the last column of Table 13.6 is obtained as 109.37 − 107.55 = 1.82. The
other enterers in this column are obtained in a similar manner. This distribution
exhibits long downside tails. The first percentile of the distribution of
V, which
corresponds to CreditVaR at the 99% confidence level is 83.64 − 107.55 = −23.91.
It is much lower if we compute the first percentile assuming a normal distribution
for
V. In that case CreditVaR at the 99% confidence level would be only −7.43.
422 Financial Mathematics: An Introduction

Year-end rating Value ($)

AAA 109.37
AA 109.19
A 108.66
BBB 107.55
BB 102.02
B 98.10
CCC 83.64
Default 51.13
Table 13.5. Source: CreditMetrics, JP Morgan. One year forward values for a BBB bond

Let the mean m, and the variance σ2 , of the distribution for

V are (from the
Table 13.5)

m = mean(
V)

= pi
Vi
i
= 0.02% × 1.82 + 0.33% × 1.64 + . . . + 0.18% × (−56.42)
= −0.46 .
σ = Var(
V)
2

= pi (
Vi − m)2
i
= 0.02%(1.82 + 0.46)2 + 0.33%(1.64 + 0.46)2 + . . . + 0.18%(−56.42 + 0.46)2
= 8.95 ,

and σ = 2.99. The first percentile of a normal distribution N(m, σ2 ) is (m − 2.33σ),

i.e. -7.43.
Credit-VaR for a Loan or Bond Portfolio

First, consider a portfolio composed of two bonds with an initial rating of

BB and A, respectively. Given the transition matrix shown in Table 13.3, and
assuming no correlation between changes in credit quality, we can derive easily
the joint migration probabilities shown in Table 13.7. Each entry is simply the
product of the transition probabilities for each obligor. For example, the joint
probability that obligor one and obligor two stay in the same rating classes is
80.53% × 91.05%=73.32% where 80.53% is the probability that obligor one keeps
 Credit VaR and Credit CVaR 423

Year-end rating Probability p (%) Forward price: V (Rs)

V(Rs)
AAA 0.02 109.37 1.82
AA 0.33 109.19 1.64
A 5.95 108.66 1.11
BBB 86.93 107.55 0
BB 5.30 102.02 -5.53
B 1.17 98.10 -9.45
CCC 0.12 83.64 -23.91
Default 0.18 51.13 -56.42
Table 13.6. Distribution of the bond values, and changes in value of a BBB bond, in 1 year

his current rating BB, and 91.05% is the probability that obligor two stays in
rating class A.
However, this table is not very useful in practice when we need to assess the
diversification effect on a large loan or bond portfolio. Indeed, the actual correla-
tions between the changes in credit quality are different from zero. Correlations
are expected to be higher for firms within the same industry or in the same region,
than for firms in unrelated sectors. In addition, correlations vary with the relative
state of the economy in the business cycle. If there is a slowdown in the economy,
or a recession, most of the assets of the obligor will decline in value and quality,
and the likelihood of multiple defaults increases substantially. The contrary hap-
pens when the economy is performing well as default correlations go down. Thus,
we cannot expect default and migration probabilities to stay stationary over time.
There is clearly a need for a structural model that bridges the changes of default
probabilities to fundamental variables whose correlations stay stable over time.
Furthermore, for the sake of simplicity, CreditMetrics/CreditVaR have chosen the
equity price as a proxy for the asset value of the firm that is not directly ob-
servable. This is another strong assumption in CreditMetrics that may affect the
accuracy of the method. CreditMetrics estimates the correlations between the eq-
uity returns of various obligors, the model then infers the correlations between
changes in credit quality directly from the joint distribution of equity returns.
The proposed framework initially developed by Merton [93] is the option pricing
approach to the valuation of corporate securities. The firm’s assets value, Vt , is
assumed to follow a standard geometric Brownian motion, i.e.
  √ !
Vt = V0 exp µ − σ2 /2 t + σ tZt , (13.1)

with Zt ∼ N(0, 1), µ and σ2 being respectively the mean and variance of the in-
424 Financial Mathematics: An Introduction

AAA AA A BBB BB B CCC Default

A 0.09 2.27 91.05 5.52 0.74 0.26 0.01 0.06
BB
AAA 0.03 0 0 0.03 0 0 0 0 0
AA 0.14 0 0 0.13 0.01 0 0 0 0
A 0.67 0 0.02 0.61 0.4 0 0 0 0
BBB 7.73 0.01 0.18 7.04 0.43 0.06 0.02 0 0
BB 80.53 0.07 1.83 73.32 4.45 0.6 0.2 0.01 0.05
B 8.84 0.01 0.2 8.05 0.49 0.07 0.02 0 0
CCC 1 0 0.02 0.91 0.06 0.01 0 0 0
Default 1.06 0 0.02 0.97 0.06 0.01 0 0 0
Table 13.7. Joint migration probabilities (%) with zero correlation for two issuers rated BB and A

stantaneous rate of return on the assets of the firm. Vt is lognormally distributed

with expected value E(Vt ) = V0 exp{µt} at time t. It is further assumed that the
firm has a very simple capital structure, financed only by equity St and a single
zero-coupon debt instrument maturing at time T with face value F and current
market value Bt . In this framework, default only occurs at maturity of the debt
obligation when the value of the assets is less than the promised payment F to the
bond holders.
Merton’s model is extended by CreditMetrics to include changes in credit qual-
ity as illustrated in Fig. 13.1. This generalization consists of slicing the distribution
of asset returns into bands in such a way that, if we draw randomly from this dis-
tribution, we reproduce exactly the migration frequencies shown in the transition
matrix mentioned in Table 13.7. Fig. 13.1 shows the distribution of the normalized
assets rates of return, 1 year ahead with mean zero and variance one. The credit
rating thresholds correspond to the transition probabilities in Table 13.3 for a
BB rated obligor. The right tail of the distribution on the righthand side of ZAAA
corresponds to the probability for the obligor of being upgraded from BB to AAA,
i.e. 0.03%. Then, the area between ZAA and ZAAA corresponds to the probability
of being upgraded from BB to AA, etc. The left tail of the distribution, on the left-
hand side of ZCCC , corresponds to the probability of default, i.e. 1.06%. Table 13.8
shows the transition probabilities for two obligors rated BB and A, respectively,
and the corresponding credit quality thresholds. This generalization of Merton’s
model is quite easy to implement. It assumes that the log-returns over any period
of time are normally distributed with mean 0 and variance 1, and it is the same
for all obligors within the same rating category.
 Credit VaR and Credit CVaR 425

Fig. 13.1. Generalization of the Merton model to include rating changes

Rated-A obligor Rated-BB obligor

Rating in 1 year Prob. (%) Z(σ) Prob. (%) Z(σ)
AAA 0.09 3.12 0.03 3.43
AA 2.27 1.98 0.14 2.93
A 91.05 -1.51 0.67 2.39
BBB 5.52 -2.30 7.73 1.37
BB 0.74 -2.72 80.53 -1.23
B 0.26 -3.19 8.84 -2.04
CCC 0.01 -3.24 1.00 2.30
Default 0.06 1.06
Table 13.8. Transition probabilities and credit quality thresholds for BB and A rated obligors

If pDe f denotes the probability for the BB-rated obligor defaulting, then the
critical asset value VDe f is such that

pDe f = P[Vt ≤ VDe f ],

which can be translated into a normalized threshold ZCCC , such that the area in
the left tail below ZCCC is pDe f . Indeed, according to (13.1), default occurs when
Zt satisfies
426 Financial Mathematics: An Introduction

   
ln(VDe f /V0 ) − µ − σ2 /2 t
pDe f = P √ ≥ Zt (13.2)
σ t
   
ln(V0 /VDe f ) + µ − σ2 /2 t
= P Zt ≤ − √ (13.3)
σ t
= Φ(−d2 ), (13.4)
 
ln(Vt /V0 ) − µ − σ2 /2 t
where the normalized return r = √ is N(0, 1). The ZCCC is
σ t
simply the threshold point in the standard normal distribution corresponding to
a cumulative probability of pDe f . Then the critical asset value VDe f , which triggers
default, is such that ZCCC = −d2 , where
 
ln(V0 /VDe f ) + µ − σ2 /2 t
d2 = √ ,
σ t
and is also called distance-to-default .
Accordingly ZB is the threshold point corresponding to a cumulative probability
of being either in default or in rating CCC, i.e., pDe f + pCCC , etc. Further, since
asset returns are not directly observable, CreditMetrics/ CreditVaR chose equity
returns as a proxy equivalent to assuming that the firm’s activities are all equity
financed. Now, for the time being, assume that the correlation between asset rates
of return is known and is denoted by ρ, which is assumed to be equal to 0.20 in
our example. The normalized log-returns on both assets follow a joint normal
distribution:

8
1 −1
f (rBB , rA ; ρ) = exp [r − 2ρrBB rA + rA ] .
2 2
2π (1 − ρ2 ) 2(1 − ρ2 ) BB

We can then easily compute the probability for both obligers of being in any
combination of ratings, like that they remain in the same rating classes BB and
A, respectively:
, 1.37 , 1.98
P(−1.23 < rBB < 1.37, −1.51 < rA < 1.98) = f (rBB , rA ; ρ)drBB drA = 0.7365.
−1.23 −1.51

If we implement the same procedure for the other 63 combinations we obtain

Table 13.9. We can compare Table 13.9 with Table 13.7, the latter being derived
assuming zero correlation, to notice that the joint probabilities are different.
To be more specific, consider two obligors whose probabilities of default are
P1(PDe f 1 ) and P2(PDe f 2 ), respectively and their asset return correlation is ρ. The
 Credit VaR and Credit CVaR 427

Second company (A)

First Company BB AAA AA A BBB BB B CCC Def Total
AAA 0 0 0.03 0 0 0 0 0 0.03
AA 0 0.01 0.13 0 0 0 0 0 0.14
A 0 0.04 0.61 0.01 0 0 0 0 0.67
BBB 0.02 0.35 7.1 0.2 0.02 0.01 0 0 7.69
BB 0.07 1.79 73.65 4.24 0.56 0.18 0.01 0.04 80.53
B 0 0.08 7.8 0.79 0.13 0.05 0 0.01 8.87
CCC 0 0.01 0.85 0.11 0.02 0.01 0 0 1
Def 0 0.01 0.9 0.13 0.02 0.01 0 0 1.07
Total 0.09 2.29 91.06 5.48 0.75 0.26 0.01 0.06 100
Table 13.9. Joint rating probabilities (%) for BB and A rated obligors when ρ = 0.2

events of default for obligors 1 and 2 are denoted as DEF1 and DEF2 , respectively,
and P(DEF1,DEF2) is the joint probability of default. Then, it can be shown that
the default correlation is
P(DEF1, DEF2) − P1.P2
corr(DEF1, DEF2) = .
P1(1 − P1)P2(1 − P2)
The joint probability of both obligors defaulting is, according to Merton’s
model,
P(DEF1, DEF2) = Pr[V1 ≤ VDe f 1 , V2 ≤ VDe f 2 ] ,
where V1 and V2 denote the asset values for both obligors at time t, and VDe f 1
and VDe f 2 are the corresponding critical values which trigger default. It is further
equivalent to
P(DEF1, DEF2) = Pr[r1 ≤ −d12 , r2 ≤ d22 ] = N2 (−d12 , −d22 , ρ),
where r1 and r2 denote the normalized asset returns for obligors 1 and 2, respec-
tively, and d12 and d22 are the corresponding distance to default. N2 (x, y, ρ) denotes
the cumulative standard bivariate normal distribution and ρ is the correlation
coefficient between x and y.
CreditMetrics does not itself answer the question regarding the pricing of credit
risk and its underlying modeling. Forward zero curves are supplied as an input to
the framework that could account for the drawback of CreditMetrics approach.
However, forward curves are the result of an estimation process of the term struc-
ture of the credit spreads. This leads to an integrating framework which provides
an idea about credit portfolio diversification.
428 Financial Mathematics: An Introduction

The second drawback of CreditMetrics is the assumption of credit homogeneity

within a credit rating class. The entire framework relies on the choice of a rating
scheme. Each class defines a class for systematic credit risk and two obligors
are supposed to have perfectly the same credit behavior if in the same credit
rating class. The methodology is good but it does not allow for good evaluation
of nonlinear instruments such as credit derivatives, swaps etc.
In this league of work, KMV is quite attractive mainly for the methodology re-
quired to estimate consistent correlations between obligors of different categories.
KMV’s model does not start from any transition matrix between rating categories,
therefore making no initial assumption on the common behavior of two counterpar-
ties belonging to the same rating class. Using a variant of Merton’s model, KMV
uses a structural approach to define the “distance-to-default category” (DD) to
which the counterparty belongs. Then, relying on a huge default database, KMV
is able to relate this DD category to an “expected default frequency” (EDF) for
each issuer. Details of KMV method could be read from [96, 147].
CreditPortfolioView is a multi-factor model which is used to simulate the joint
conditional distribution of default and migration probabilities for various rating
groups in different industries, for each country, conditional on the value of macroe-
conomic factors like the unemployment rate, the rate of growth in GDP, the level
of long-term interest rates, foreign exchange rates, government expenditures and
the aggregate savings rate. CreditPortfolioView is based on the casual observation
that default probabilities, as well as migration probabilities, are linked to the econ-
omy. When the economy worsens both downgrades as well as defaults increase.
It is contrary when the economy becomes stronger. In other words, credit cycles
follow business cycles closely. Since the state of the economy is, to a large extent,
driven by macroeconomic factors, CreditPortfolioView proposes a methodology
to link those macroeconomic factors to the default and migration probabilities
[96, 147].
Although these models have different calculation techniques and parameters,
studies have shown that they in fact represent a remarkable consensus in their
underlying frameworks and financial intuition. However, before these models can
deliver on their promise, they must not only prove to be conceptually sound and
empirically valid; they must also be well integrated with a bank day-to-day credit
risk management. Merton-based models, using asset correlation derived from eq-
uity data, might be more accurate for publicly traded companies, while actuarial
models, using default-rate volatility based on historical experience, might be more
accurate for illiquid asset classes or small business portfolios.
In the Table 13.10 we summarize the analytic model data requirements for the
 Credit Risk Optimization with CVaR Criteria 429

aforementioned models.

Model Input Output

Default and migration probabilitiesEcono. capital
(transition matrices) (both expected loss and
unexpected loss)
JP Morgans CreditMetrics Credit spreads and yield curves Return distribution
Pair-wise correlation Loss percentiles
Recovery rates
Credit exposures
Default rates Econo. capital
Default rate volatility Loss distribution
CSFPs CreditRisk+
Recovery rates Loss percentiles
Credit exposures
Expected default frequencies Econo. capital
Credit spreads Sharpe ratio
KMVs Portfolio Manager Pair-wise correlation Mis-pricing
Credit exposures Optim. benefits
Return distribution
Macroeco. variables Econo. capital
Default and migration history return distribution
McKinsey & Co.s Credit
Credit spreads Loss percent
Recovery rates
Credit exposures
Table 13.10. Input-Output of various credit risk models

13.6 Credit Risk Optimization with CVaR Criteria

In earlier section we studied VaR that seems to provide an efficient solution and
provide answer to the question: what is the maximum loss with the confidence
level (1 − β) × 100% over a given time horizon? Thus, its calculation reveals that
the loss will exceed VaR with likelihood β × 100%, but no information is provided
on the amount of the excess loss, which may be significantly large. Mathemat-
ically, VaR has serious limitations. In case of a finite number of scenarios, it is
a nonsmooth, nonconvex, and multi extremum function [91] (with respect to po-
430 Financial Mathematics: An Introduction

sitions), making it difficult to control and optimize. Also, VaR has some other
undesirable properties, such as the lack of sub-additivity.
By contrast, CVaR is considered a more consistent measure of risk than VaR.
CVaR supplements the information provided by VaR and calculates the quantity
of the excess loss. Since CVaR is greater than or equal to VaR, portfolios with a
low CVaR also have a low VaR. Under quite general conditions, CVaR is a convex
function with respect to positions [140], allowing the construction of efficient opti-
mization algorithms. CVaR has been compared with the widely accepted VaR risk
performance measure for which various estimation techniques have been proposed,
see for example [39, 109].
Bucay and Rosen [22] applied CreditMetrics methodology to a portfolio of cor-
porate and sovereign bonds issued in emerging markets. They estimated the credit
risk of the undertaking portfolio by taking into account both defaults and credit
migrations. On the similar lines, we utilize CVaR optimization routine developed
in Chapter 6 for the bond portfolio discussed in [3].
The test portfolio has been compiled by a group of financial institutions to
access the state-of-the-art of portfolio credit risk models. The portfolio consists of
197 emerging markets bonds, issued by 86 obligors in 29 countries. The date of the
analysis was October 13, 1998 and the mark-to-market value of the portfolio was
8.8 billion USD. Most instruments are denominated in US dollars but 11 fixed rate
bonds are denominated in seven other currencies; DEM, GBP, ITL, JPY, TRL,
XEU and ZAR. Bond maturities ranged from a few months to 98 years and the
portfolio duration is approximately five years.
Let x = (x1 , x2 , . . . , xn ) be obligor weights (positions) expressed as multiples
of current holdings, b = (b1 , b2 , . . . , bn ) be future values of each instrument with
no credit migration (benchmark scenario), and y = (y1 , y2 , . . . , yn ) be the future
(scenario-dependent) values with credit migration. The loss due to credit migra-
tion for the portfolio is defined as f (x, y) = (b − y)T x. The CVaR optimization
problem is formulated as
Min n Φ(x) = Φ(z, α)
x∈X⊂R

where X is the feasible set in Rn . For the definition of Φ please refer to Chapter
6. This set could also contain constraints pertaining mean return constraint, box
constraints on the positions of instruments, etc. However to make content simple
and easy to read we have not discussed them in optimization routine. Interested
reader should read [3, 91]. They have approximated the performance function
using scenarios y j , j = 1, . . . , J, which are sampled with the density function p(y).
As discussed in Chapter 6, minimization of the CVaR function Φ(x) could be
 Credit Risk Optimization with CVaR Criteria 431

reduced (approximately) to the following linear programming problem


J
Min α+ν zj
x,α,z
j=1

subject to
f (x, y j ) − α ≤ z j ( j = 1, 2, . . . , J)
eT x = 1
(13.5)
zj ≥ 0 ( j = 1, 2, . . . , J)
li ≤ xi ≤ ui (i = 1, 2, . . . , n) ,
1
where ν = , where β is the confidence level.
(1 − β)J
Also, if (x∗ ; α∗ ; z∗ ) is an optimal solution of the optimization problem (13.5),
then x∗ is an approximation of the optimal solution of the CVaR optimization
problem, the function Φ(z∗ ; α∗ ) equals approximately the optimal CVaR, and α∗ is
an approximation of VaR at the optimal point. Thus, by solving problem (13.5)
we can simultaneously find approximations of the optimal CVaR and the corre-
sponding VaR.
Further, in [3, 91], two sets of additional constraints, mentioned below, in the
optimization of CVaR are considered to evaluate the performance against bench-
mark case. These set of constraints are
(i) no short positions allowed, and the positions can be at most doubled in size.
(ii) positions, both long and short, can be at most doubled in size.
The first constraint simply means that li = 0 and ui = 2, i.e. 0 ≤ xi ≤ 2
and the second constraint implies that −2 ≤ xi ≤ 2. Further, authors [3] have
considered that the re-balanced portfolio should*nmaintain *the future expected
n
value, in absence of any credit migration, i.e. i=1 bi xi = i=1 bi . After adding
these set of constraints the result of the optimization, in the case of no short
positions (no short), and in the case of both long and short positions (long and
short), are presented in Table 13.10.
Table 13.10 shows that the two risk measures, VaR and CVaR, are significantly
improved after the optimization. When no short positions are allowed, VaR and
CVaR reduce by about 60%. For example, at (1 − β) = 0.99, we lowered CVaR to
559 million from the original 1320 million USD. By allowing both short and long
positions slightly improves reductions, but not significantly. Thus, it has been
observed that risk measures could be reduced by about 60% with the multiple
obligor optimization.
432 Financial Mathematics: An Introduction

Case β VaR VaR (%) CVaR CVaR (%)

Original 0.900 340 - 621 -
0.950 518 - 824 -
0.990 1,026 - 1,320 -
0.999 1,782 - 1,998 -
No Short 0.900 163 52 279 55
0.950 239 54 359 56
0.990 451 56 559 58
0.999 699 61 761 62
Long 0.900 149 56 264 58
and 0.950 226 56 344 58
Short 0.990 433 58 542 59
0.999 680 62 744 63
Table 13.11. VaR, CVaR (in Millions of USD) and corresponding VaR and CVaR reductions (in %)
for the Multiple Obligor Optimization.

Table 13.11 shows that the two risk measures, expected loss and standard de-
viation, also are dramatically improved when we minimized CVaR. For example,
in case of both long and short positions, the expected loss and standard deviation
are reduced about 50%. The corresponding position weights for the original twelve
largest risk contributors are presented in Table 13.12.

Case β Expected Expected Standard Standard

Loss Loss (%) Deviation Deviation (%)
Original - 95 - 232 -
No Short 0.90 50 47 107 54
0.95 51 46 109 53
0.99 60 37 120 48
0.999 63 34 126 46
Long 0.90 42 56 105 55
and 0.95 44 54 107 54
Short 0.99 53 44 118 49
0.999 58 39 124 47
Table 13.12. Expected Loss (in Millions of US-D), Standard Deviation (in Millions of US-D) and
Corresponding Reductions (in %) for the Multiple Obligor Optimization
 Credit Scoring and Internal Rating 433

Obligor Original No Short Short and Long

Brazil 1 0.08 0.18
Russia 1 0 0.09
Venezuela 1 0 -0.41
Argentina 1 0.35 0.47
Peru 1 0 -0.38
Colombia 1 0.89 1.00
Morocco 1 0.02 0.12
RussiaIan 1 0 -2.00
MoscowTel 1 1.52 1.99
Romania 1 0.45 1.33
Mexico 1 0.94 0.90
Philippines 1 1.04 0.94
Table 13.13. Positions for the Multiple Obligor Optimization (Minimization of CVaR with (1 − β) =
0:99).

This idea behind using the CVaR optimization framework is to simultaneously

optimize two closely related risk measures: CVaR and VaR. Although with the
CVaR as a performance function, the optimization leads to reduction of all other
aforementioned risk measures e.g. CVaR, VaR, the expected loss, and the standard
deviation. From a bank perspective, this approach looks quite attractive. The bank
should have reserves to cover expected loss and capital to cover unexpected loss.

13.7 Credit Scoring and Internal Rating

So far we have considered the CreditMetrics, credit VaR and CVaR to analyze
and estimate risks associated with credit default by risk departments of the banks
and financial firms. However, some companies run a credit risk department where
the job is to assess the financial health of their customers, and extend credit
accordingly. The credit scoring model forms an essential part of the aforementioned
framework in order to grant credit to clients. Thus, credit scoring can be defined as
a technique that helps credit providers decide whether to grant credit to customers
by segregating ‘good’ and ‘bad’ customers in terms of their creditworthiness.
The aim of a credit scoring model is to build a single aggregated risk indicator
for a set of risk factors. Hence it is a tool that aid decision whether to grant
credit to a new applicant. The other type of decision is how to deal with existing
customers i.e if an existing customer wants to increase his credit limit should
434 Financial Mathematics: An Introduction

the firm agree to that? If the customer starts to fall behind in repayments what
actions should the firm take? Techniques that help with these decisions are called
behavioural scoring.
In order to grant loan to customers following are the steps considered in the
lending process:
(i) Solicitation: Either the firm solicits applications for loans, for instance via
advertising, or a client approach to firm with a new request.
(ii) Information gathering: Information about the applicant via various means
including interviews, visits, review of financial data or accounts, possible use
of credit reference agencies, ratings agencies or other available data has been
gathered to rate the client.
(iii) Recommendation: On the basis of information gathered and consideration of
why the client wants the money an internal rating is assigned and a lending
decision is recommended. This is then reviewed, perhaps by a loan committee
in the case of a corporate loan or branch staff in the case of a retail exposure.
(iv) Closing administration: Any collateral is perfected, documentation is finalized
and signed and funds are made available.
(v) Monitoring: The performance of the obligator and their condition are re-
assessed periodically. One aim of internal ratings is to allow many applications
from different kinds of corporates in different countries to be assessed on an
equitable basis. It also allows the bank to set break-even spreads for internal
ratings classes which encourage lending to good counterparties.
The analogue of internal ratings for retail exposures is credit scoring. Credit
scoring can be formally defined as a statistical (or quantitative) method that is
used to predict the probability that a loan applicant or existing borrower will
default or become delinquent [95]. This helps to determine whether credit should
be granted to a borrower [100]. Credit scoring can also be defined as a system-
atic method for evaluating credit risk that provides a consistent analysis of the
factors that have been determined to cause or affect the level of risk [135]. Data
for a system should have discriminating power that could be based on certain
parameters like gross income, age, number of years spent in the current position,
gender etc. This can then be used both reactively to decide on whether to accede
to a mortgage application, for instance and proactively to solicit application for a
particular type of credit card.
Credit scoring has many benefits that accrue not only to the lenders but also
to the borrowers. For example, credit scores help to reduce discrimination as
scoring models provide an objective analysis of a customers credit worthiness.
 Credit Scoring and Internal Rating 435

With the help of the credit scores, financial institutions are able to quantify the
risks associated with granting credit to a particular applicant in a shorter time.
Further, credit scores can help financial institutions determine the interest rate
that they should charge their customers and to price portfolios [6]. Higher-risk
customers are charged a higher interest rate and vice versa. Based on the customers
credit scores, the financial institutions are also able to determine the credit limits
to be set for the customers [106, 118]. This further help financial institutions to
manage their accounts more effectively and profitably.
Credit scores are also used as a basis to adjust premiums. Generally, customers
with bad credit scores have a higher chance of failing insurance claims compared
with customers with good credit scores. Therefore, the former are charged a higher
premium. Further, the credit information is used to assess a customers account-
ability and performance under the conditions of an insurance policy.
The methods generally used for credit scoring are based on statistical pattern
recognition techniques. Historically, discriminant analysis and linear regression
were the most widely used techniques for building scorecards. Later the logit and
probit models were suggested in Martin [88] and Ohlson [105]. All these models
belong to the class of generalized linear models (GLM) and could also be inter-
preted using a latent (score) variable. Their core decision element is a linear score
function (graphically represented as a hyperplane in a multidimensional space)
separating successful and failing companies. The company score is computed as a
value of that function. In the case of the probit and logit models the score is via
a link function which is directly transformed into a probability of default (PD).
The major disadvantage of these popular approaches is the enforced linearity of
the score and, in the case of logit and probit models, the prespecified form of the
link function (logit and Gaussian) between PDs and the linear combination of
predictors.
Some of the techniques that have been previously used, but rather infrequently,
to construct credit scoring models include genetic algorithm, k−nearest neighbor,
linear programming, and expert systems Thomas et al. [135]. In recent years,
new techniques have been increasingly used to construct credit scoring models.
In particular, the decision tree approach has become a popular technique for
developing credit scoring models because the resulting decision trees are easily
interpretable and visualized. Further, neural networks are also commonly used. All
the methods and techniques mentioned above can be considered as an important
data mining techniques for predictive modeling.
436 Financial Mathematics: An Introduction

13.8 Construction of Credit Scoring Models

The methodology for constructing credit scoring models generally involves the
following process. First, a sample of previous customers is selected and classified
as good or bad depending on their repayment performance over a given period
(for simplicity, only a dichotomy is used here). Next, data are compiled from
loan applications, personal and/or business credit records, and various sources if
available (e.g. credit bureau reports). Finally, statistical (or other quantitative)
analysis is performed on the data to derive a credit scoring model. The model
comprises weights to apply to the different variables (or attributes) in the data and
a cut-off point. The sum of the weights applied to the variables for an individual
customer constitutes the credit score. The cut-off point determines whether this
customer should be classified as “good” or “bad”. The probability associated with
this classification can also be generated. Different models can be constructed for
different segments of the data (e.g. for different products).
The concept of incurred cost associated with the probabilities of repayment
of loans will be used to illustrate some of the methods. For simplicity, assume
that the population of loans consist of two groups or classes G and B that denote
loans that (after being granted) will turn out to be good or bad in the future,
respectively. Good loans are repaid in full and on time. Bad loans are subject to
the default.
Usually the class/group sizes are very different, so that for the probability that
a randomly chosen customer belongs to group G, denoted as pG , one has pG > pB .
Let x be a vector of independent variables (also called the measurement vector)
used in the process of deciding whether an applicant belongs to group G or B. Let
the probability that an applicant with measurement vector x belongs to group G
be p(G/x), and that of B be p(B/x). Let the probability p(x/G) indicate that a good
applicant has measurement vector x. Similarly, for bad applicants the probability
is p(x/B). The task is to estimate probabilities p(·/x) from the set of given data
about applicants which turn out to be good or bad and to find a rule for how to
partition the space X of all measurement vectors into the two groups AG and AB
based on these probabilities, so that in AG would be the measurement vectors of
applicants who turned out to be good and vice versa.
It is usually not possible to find perfect classification as it may happen that
the same vector is given by two applicants where one is good and the other is
bad. Therefore it is necessary to find a rule that minimizes the cost of a bank
providing credit connected with the misclassification of applicants. Let cG denote
the costs connected with misclassifying a good applicant as bad and cB the costs
 Logistic Regression 437

connected with classifying a bad applicant as good. Usually cB > cG , because costs
incurred due to misclassifying a bad customer are financially more damaging than
cost associated with the former kind of error. If applicants with x are assigned to
B, the expected cost is
class * * cB p(B/x) and the expected loss for the whole sample
is cB x∈AB p(B/x)p(x) + cG x∈AG p(G/x)p(x), where p(x) is a probability that the
measurement vector is equal to x. This is minimized when, into group G, such
applicants are assigned who have their group of measurement vectors AG = {x :
cB
cB p(B|x) ≤ cG p(G/x)} which is equivalent to AG = {x : p(G/x) ≥ }. Without
cG + cB
loss of generality, the misclassification costs can be normalized to cG + cB = 1.
In this case, the rule for classification is to assign an applicant with x to class
G if p(G/x) > cB and otherwise to class B. An important task is to specify the
cost of lending errors and to accurately specify the optimal cutoff-score for credit
scoring, as banks have to choose the optimal trade-off between profitability and
risk. Credit policies that are too restrictive may ensure minimal costs in terms of
defaulted loans, but the opportunity costs of rejected loans may exceed potential
bad debt costs and thus profit is not maxmized. On the other hand, policies that
are too liberal may result in high losses from bad debt.
In the sequel, we would discuss three algorithms which are very popular in the
domain of credit scoring. These are logistic regression, Fisher’s linear discrimi-
nant analysis and support vector machines. Our presentation here is motivated
by Thomas et al. [135].

13.9 Logistic Regression

The traditional linear regression approach tries to find the best linear combination
w0 + w1 X1 + w2 X2 + . . . + wn Xn of characteristics which explains the probability
of default. Here (X1 , X2 , . . . , Xn ) is the set of n random variables that describe
the information available on an applicant for credit from the application form.
Let pi (i = 1, 2, . . . , n) be the probability that the tth applicant in the sample has
defaulted. Then our aim in linear regression is to find w0 , w1 , . . . , wn such that the
linear combination w0 +w1 xi,1 +. . .+wn xi,n best approximates the probability of de-
fault pi , i.e., pi = w0 +w1 xi,1 +. . .+wn xi,n for all i. For the ith applicant, the observed
data is xi,1 , xi,2 , . . . , xi,n for the random variables X1 , X2 , . . . , Xn respectively.
The regression approach to linear discrimination is not very satisfactory. This
is because the linear combination w0 + w1 xi,1 + . . . + wn xi,n could take any value
in (−∞, ∞, but the probability of default pi takes values between 0 and 1 only.
Therefore it would be better if we could modify the usual linear regression ap-
438 Financial Mathematics: An Introduction

proach so that the linear combination w0 + w1 xi,1 + . . . + wn xi,n best approximates

some
 function
 of pi which could take wider range of values. One such function is
pi
ln which results in logistic regression. Thus in logistic regression we aim
1 − pi
to find w0 , w1 , . . . , wn so that the linear
 combinationw0 + w1 xi,1 + . . . + wn xi,n best
pi pi
approximates the function ln for all i. Since takes values between
  1 − pi 1 − pi
pi
0 and ∞, ln takes values between −∞ and ∞. Thus for all i, we take
1 − pi
 
pi
ln = w0 + w1 xi,1 + . . . + wn xi,n . (13.6)
1 − pi
(i)
eZ 1
Taking exponentials on both sides of (13.6) we have pi = = ,
1+e Z(i)
(1 + e−Z(i) )
T (i)
ew x
where Z = (i)
, w = (w0 , w1 , w2 , . . . , wn )T . Therefore the logistic regression
(1 + ewT x(i) )
approach has the assumption that pi is of this specific form. The function f (u) =
1
is called the logistic function.
(1 + e−u )
If we assume that the distribution of the characteristics values of the ‘good’
and of the ‘bad’ is multivariate normal then we can show that the assumption of
logistic regression
 is satisfied. In this case we can obtain a close form expression
pi *
for ln in terms of the vectors µG , µB and the matrix . Here µG is the
1 − pi
* of means among the ‘goods’, µB is the vector of means among the ‘bads’
vector
and is the variance-covariance matrix for the random vector (X1 , X2 , . . . , Xn ).
In classical regression, it is simple to calculate the vector of coefficients w by
employing the method of least squares. But the determination of the vector w
in logistic regression is not that simple. We need to use the maximum likelihood
approach to get estimates for these coefficients. This involves an application of
the Newton-Raphson method to solve the equation that arise.
In conclusion the usual linear regression and logistic regression are very similar
in approach but very different conceptually. The linear regression tries to fit the
probability p of defaulting  by alinear combination of the attributes while logistic
p
regression tries to fit ln by a linear combination of attributes. The ad-
1−p
vantage of logistic regression is that it does not require linearity of relationship
 Fisher’s Linear Discriminant Analysis 439

between the independent variables and the dependent variable, and it does not re-
quire the assumption of normality. But it assumes that the independent variables
p
are linearly related to logit (i.e. ln 1−p ) of the dependent variable.

13.10 Fisher’s Linear Discriminant Analysis

Discriminant analysis which was studied by Fisher as early as 1936 is an alter-
native to logistic regression. It assumes that the explanatory variables follow a
multivariate normal distribution and have a common variance-covariance matrix.
This method is used to classifying observations in two classes. In the context of
credit scoring, the two groups are those classified by the lender as ‘goods’ and
‘bads’ and the characteristics are the applicant’s form details and the credit bu-
reau information.
Let Y = w1 X1 + w2 X2 + . . . + wn Xn be any linear combination of characteristics
X = (X1 , X2 , . . . , Xn ). The most natural measure of separation seems to be the
difference between the mean values of two different * groups of ‘good’ and ‘bads’ in
the sample. Thus it makes sense to find wi with ni wi = 1 such that the difference
between E(Y|G) and E(Y|B) is maximum. But it is also important to see how
closely each of the two groups cluster together when we discuss their separation.
Therefore the aim should be to minimize the distance within each group and
maximize the distance between different groups using some suitable discriminant
function.
Fisher suggested that, under the assumptions as stated above, an appropriate
measure of separation is
square of distance between sample means of two groups
M= .
(sample variance of each group)1/2
Here we have division by the square root of the variance to make the measure
scale independent. Thus if Y is changed to CY, the measure M will not change.
Let mG and mB be the vector of sample means of ‘goods’ and ‘bads’ and S be
the common sample variance. For Y = w1 X1 + w2 X2 + . . . + wn Xn , the separating
measure M is given by
wT (mG − mB )w
M= . (13.7)
(wT Sw)1/2
The above equation follows because E(Y|G) = wT mG , E(Y|B) = wT mB , and
Var(Y) = (wT Sw). To maximize M we compute ∇w M and equate it to zero. This
gives
440 Financial Mathematics: An Introduction

(mG − mB )w.(wT Sw) = wT (mG − mB )w.Sw,

i.e. (mG − mB )w = M.Sw, giving

Mw = S−1 (mG − mB )w

Therefore solving generalize eigenvalue problem leads to eigen vector w directly

proportional to S−1 (mG − mB ). This relation can be used to find weights wi which
maximize M. *
We then compute the score s(x) = i wi xi , where xi are explanatory variables.
We use the score s(x) to discriminate between the two groups. The normality
assumption only becomes important if significance tests are to be undertaken. The
advantages of the linear discriminant analysis (LDA) method are that it is simple,
it can be very easily estimated and it actually works very well. The theoretical
disadvantage of linear discriminant analysis is that it requires normally distributed
data but the credit data are often non-normal (and categorized). But if one is not
too much concerned about significance tests then it can be used for most of data.

13.11 Support Vector Machines

Support vector machine is a classifier technique, first proposed by Vapnik [141].
This method involves three elements. A score formula which is a linear combination
of features selected for the classification problem, an objective function which
considers both training and test samples to optimize the classification of new
data, an optimizing algorithm for determining the optimal parameters of training
sample objective function.
Credit applicants are assigned to good or bad risk classes according to their
records of defaulting. Each applicant is described by a high dimensional input
vector of situational charateristics and with an associated class label. A statisti-
cal model which maps the inputs to the labels can decide whether a new credit
applicant should be accepted or rejected by predicting the class label given the
new inputs. Support vector machines (SVM), from the statistical learning theory,
can build such models from the data, requiring but extremely weak prior assump-
tions about the model structure. Furthermore, SVM divide a set of labeled credit
applications into subsets of “typical” and “critical” patterns. The correct class
label of a typical pattern is usually very easy to predict, even with the linear
classification methods. Such patterns do not contain much information about the
classification boundary. The critical patterns (the support vectors) contain the
less trivial training samples.
 Support Vector Machines 441

The advantages of the method are that, in the nonparametric case, SVM re-
quires no data structure assumptions such as normal distribution and continuity.
SVM can perform a nonlinear mapping from an original input space into a high
dimensional feature space and this method is capable of handling both continuous
and categorical predictions. The weaknesses of this method are that, it is difficult
to interpret unless the features interpretable and the standard formulations do
not contain any specification of business constraints.
Given a training set of instance-label pairs (xi , yi ), i = 1, 2, . . . , m where xi ∈ Rn
and yi ∈ {+1, −1}, SVM finds an optimal separating hyperplane with the maximum
margin by solving the following optimization problem
1 T  m
(LP) Min w w+C ξi
w,b 2 i=1
subject to
yi (wT xi + b) + ξi − 1 ≥ 0
ξi ≥ 0
where C is a penalty parameter on the training error and ξi is the non-negative
slack variable. Introducing the Karush KuhnTucker (KKT) [98] condition for the
optimum constrained function, then LP is transformed to the dual Lagrangian
LD(α).

m
1
m
LD(α) Max αi − αi α j yi y j < xi .x j >
α
i=1
2 i=1
subject to
*m
i=1 αi yi = 0
0 ≤ αi ≤ C (i = 1, 2, . . . , m) .
To find the optimal hyperplane, a dual Lagrangian LD(α) must be maximized
with respect to non-negative αi . The solution αi for the dual optimization problem
determines the parameters w∗ and b∗ of the optimal hyperplane. Thus, the optimal
hyperplane
⎛ m decision function ⎞ f (x) = sign(w∗T x + b∗ ) can be written as f (x) =
⎜⎜ ⎟⎟
sign ⎜⎜⎝ yi α∗i < xi , x > +b∗ ⎟⎟⎠ . The nonlinear SVM maps the training samples from
i=1
the input space into a higher-dimensional feature space via a mapping function
φ. In the dual Lagrange, the inner products are replaced by the kernel function
< φ(xi ), φ(x j ) >= K(xi , x j ), and the nonlinear SVM dual Lagrangian LD(α) is
similar with that in the linear generalized case i.e.
442 Financial Mathematics: An Introduction


m
1
m
Max αi − αi α j yi y j K(xi , x j )
α
i=1
2 i=1
subject to
*m
αi yi = 0
i=1
0 ≤ αi ≤ C (i = 1, 2, . . . , m)

Followed by the steps described in the linear generalized case, we obtain decision
function of the following form
⎛ m ⎞
⎜⎜ ⎟⎟
f (x) = sgn ⎜⎜⎝ yi αi < φ(x), φ(xi ) > +b ⎟⎟⎠
∗ ∗

i=1
⎛ m ⎞
⎜⎜  ⎟⎟
= sgn ⎜⎜⎝ yi α∗i K(x, xi ) + b∗ ⎟⎟⎠ . (13.8)
i=1

Proper kernel parameters setting can improve the SVM classification accuracy.
With the RBF kernel, there are two parameters C and γ to be determined in the
SVM model. The grid search approach [63] is an alternative to finding the best C
and γ when using the radial basis function (RBF) kernel function. In addition to
the proper parameters setting, feature subset selection can improve the SVM clas-
sification accuracy. To pursue even small improvement in credit scoring accuracy,
many methods have been investigated in the last decade. Artificial Neural Net-
works (ANNs), genetic algorithms (GA) are the most commonly soft computing
method used in credit scoring modelling [134, 135].

13.12 Drawbacks
Although credit scoring has significant benefits, its limitations should also be
noted. One major problem that can arises when constructing a credit scoring
model is that the model may be built using a biased sample of customers who
have been granted credit [58]. This may occur because applicants (i.e. potential
customers) who are rejected are not included in the data for constructing the
model. Hence, the sample is biased (i.e. different from the general population) as
good customers are too heavily represented. The credit scoring model built using
this sample may not perform well on the entire population since the data used
to build the model is different from the data that the model will be applied to.
The second problem that can arise when constructing credit scoring models is the
 Summary and Additional Notes 443

change of patterns over time. The key assumption for any predictive modeling is
that the past can predict the future [10]. In credit scoring, this means that the
characteristics of past applicants who are subsequently classified as good or bad
creditors can be used to predict the credit status of new applicants. Sometimes,
the tendency for the distribution of the characteristics to change over time is so
fast that it requires constant refreshing of the credit scoring model to stay relevant.
One of the consequences of credit scoring is the possibility that end-users become
so reliant on the technology that they reduce the need for prudent judgment and
the need to exercise their knowledge on special cases. In other instances, end-
users may unintentionally apply more resources than necessary to work the entire
portfolio. This could run into the risk of a self-fulfilling prophecy [84]. In the U.S., a
new industry has emerged that is dedicated to help borrowers improve their credit
scores by rearranging finances [136], rather than obeying the simple rule: pay your
bills on time and keep your debt low. Such score-polishing actions could potentially
distort the patterns of credit default. Despite the limitations highlighted above,
there is no doubt that credit scoring continues to be a major tool in predicting
credit risk in customer lending. It is envisaged that organizations using credit
scoring appropriately will gain important strategic advantage and competitive
edge over their rivals.

13.13 Summary and Additional Notes

• This chapter seeks to give an overview of the techniques that are used and
being developed to forecast the financial risk involved in lending to customers.
We have sought to give a fairly comprehensive biography of the literature of
the credit scoring models. Credit scoring are some of the most important fore-
casting techniques used in the retail and customer finance areas. As a pure
forecasting tool as opposed to a decision making one, credit scoring has mainly
been used as a way of forecasting future bad debt in order to set aside appro-
priate provisioning.
• There are number of different ways of estimating the probability that a com-
pany will default during a particular period of time in the future. One involves
bond prices, another involves historical data and a third involves equity prices.
The default probabilities backed out from bond prices are risk-neutral proba-
bilities. The probabilities backed out from historical data are real-world prob-
abilities. Real world probabilities should be used for scenario analysis and the
calculation of credit VaR. Risk-neutral probabilities should be used for valuat-
444 Financial Mathematics: An Introduction

ing credit sensitive instrument. In general risk-neutral default probabilities are

significantly higher than real-world probabilities.
• A Credit VaR measure can be defined as the credit loss that with a certain
probability will not be exceeded during a certain time period. It can be defined
to take into account only losses arising from defaults. Alternatively, it can be
defined so that it reflects the impact of both defaults and credit rating changes.

13.14 Exercises
Exercise 13.1 Let A = {(2, 0), (0, 0)} and B = {(4, 4)} be two sets in R2 . It is
claimed that the line y = 3 is the hard margin classifier. Verify this claim analyt-
ically.
Exercise 13.2 Let x = (0, 2, −3)T and (−2, 1, 4)T . Determine ||(x)+ ||1 , ||(y)+ ||1 and
||(x + y)+ ||1 and ||(x)+ + (y)+ ||1 . Show that the polynomial kernel of degree 2 is a
Mercer kernel.
Exercise 13.3 Let A = {(0, 0), (2, 0), (0, 2)} and B = {(0, −1), (1, 0), (0, 1)} be two
datasets with class labels +1 and -1, respectively.
1. Use a polynomial kernel of degree 2 to find a soft margin classifier.
2. Write the optimization problem (both the primal and dual forms) if a Gaussian
kernel with σ = 1 is employed for finding the soft margin classifier.
Exercise 13.4 Perform the simulation of Merton’s Model for Credit-VaR for a
Bond portfolio when the correlation 0.4.
14
Monte Carlo Simulation

14.1 Introduction
Monte Carlo methods (or Monte Carlo simulations) are a class of computational
algorithms that rely on repeated random sampling to compute their results. The
use of Monte Carlo simulations in quantitative finance or specifically in financial
engineering for valuation of options was probably first suggested by Phelim Boyle
in 1977 [19]. Ever since then the Monte Carlo simulation has become an essential
tool in pricing the derivative securities and also in risk management. These ap-
plications have, in turn, stimulated research into new Monte Carlo methods and
renewed interest in some older techniques. This chapter develops the Monte Carlo
methods and study their applications in finance. It also uses simulation, which is
a fictitious representation of reality, as a vehicle for presenting models and ideas
from financial engineering. The subject of Monte Carlo methods can be viewed as
a branch of “experimental mathematics” in which one uses random numbers to
conduct experiments. Typically experiments/simulations are carried on systems
using anywhere from hundreds to billions of random numbers.
A Monte Carlo method is a numerical method based on random sampling that
can be used to solve a mathematical or statistical problem. It provides approximate
solutions to a variety of mathematical problems by performing statistical sampling
experiments. These methods derive their collective name from the fact that Monte
Carlo, the capital of Monaco, has many casinos and casino roulette wheels.
One of the most important uses of Monte Carlo methods is in evaluating diffi-
cult multi-dimensional integrals for which we have very few methods for computa-
tion. Note that these methods only provide an approximation of the actual value.
The attempt to minimize this error is the reason for many different Monte Carlo
methods. For example, a model of a random process that produces (or mimic)
traffic movements on a particular highway is a simulation. Now suppose, we are
446 Financial Mathematics: An Introduction

interested to know the average number of cars passing through that highway or
the probability that the waiting time is more than 5 minutes, then in these cases
we need the Monte Carlo simulation.
A Monte Carlo simulation uses repeated sampling to determine the properties
of some statistical phenomenon. The Monte Carlo simulation technique has for-
mally existed since the early 1940s, where it had its applications in research into
nuclear fusion. One important feature of Monte Carlo simulation is that it is pos-
sible to estimate the order of magnitude of estimation error, in terms of statistical
confidence intervals. Thus, Monte Carlo simulation includes the distribution of
the random variables, analysis of the output and efficiency of the simulation.
In this chapter, we focus on generating random numbers and variables, and
Monte Carlo techniques, and show their applications in the area of finance. We
shall be referring mainly to Glasserman [49] and Jackel [69].

14.2 Generating Random Numbers

The essential feature common to all Monte Carlo computations is that at some
point we have to substitute for a random variable a corresponding set of actual
values, having the statistical properties of the random variable. The values that
we substitute are called random numbers on the grounds that they could well have
been produced by a suitable random process.
The efficient ways to obtain random numbers using arithmetic operation is sug-
gested by John von Neumann in 1940’s. There are two types of random numbers
namely pseudo-random numbers and quasi-random numbers. The random num-
ber sequences generated in a deterministic way is called pseudo-random numbers.
Any standard random number generator produces a series of uniformly distributed
numbers on the interval [0, 1]. In software MATLAB, the rand(n, m) function gen-
erates uniformly distributed pseudo-random numbers in an (n×m) matrix. A good
generator will have a very large repeat cycle nearly 109 . Note that, the sequence of
numbers generated by the algorithm is not random but deterministic in the sense
that the numbers generated by the algorithm reproduce independence and contain
no any discernible information on the next value. This property is often referred
to as lack of predictability. In other words, the number generated are spread out
evenly across the interval [0, 1].
We input a number, generally called a random number seed, to start with the
generation of random numbers sequence. Then the set of mathematical operations
are performed on the random number seed. The resultant generated random num-
 Generating Random Numbers 447

bers are called pseudo-random numbers. These random numbers are tested with
rigorous statistical tests to ensure that the numbers are random.
The most common method used for pseudo-random number generation is a
recursive technique called the linear congruence generator or Lehmner generator.
It is defined by the recursive formula

xn+1 = (Axn + C) (mod M),

where the integers A, C and M are parameters that can be adjusted for convenience
and to ensure the desired nature of the sequence of pseudo-random numbers, and
mod stands for modulus operation. This generator is initiated with a seed x0 .
In general, the random integers so generated should be mapped into the interval
[0, 1]. For a linear congruence generator taking possible values x ∈ {0, 1, . . . , M−1},
it is suggested to use u = x/M or (x+0.5)/M to get values which are approximately
uniformly distributed in the unit interval [0, 1]. Note that, there are many tests
that can be applied to determine whether the hypothesis of independent uniform
variables is credible.
As an illustration, let x0 = 4, A = 81, C = 35 and M = 256 (8 bits). This
generates the sequence

x1 = (81 × 4 + 35) (mod 256) = 359 mod 256 = 103

x2 = (81 × 103 + 35) (mod 256) = 8378 mod 256 = 186
x3 = (81 × 186 + 35) (mod 256) = 15101 mod 256 = 253 .

Therefore, the corresponding ui s are u1 = 103/256 = 0.4023, u2 = 0.726, u3 =

0.98828.
The above repetition is inevitable for a linear congruential generator. There
are at most M possible numbers after reduction mod M and once we arrive back
at the seed the sequence is destined to repeat itself. In the example above, the
sequence cycles after 256 numbers. The length of one cycle, before the sequence
begins to repeat itself again, is called the period of the generator.
Quasi-random numbers are numbers selected from a quasi-random sequence.
Quasi-random sequence is a sequence of n-tuples that fills n-space more uniformly
than uncorrelated random points, sometimes also called a low-discrepancy se-
quence. Quasi-random numbers are useful in computational problems such as
quasi-Monte Carlo integration. Although the ordinary uniform random numbers
and quasi-random sequences both produce uniformly distributed sequences, there
is a major difference between the two. A uniform random generator on [0, 1) pro-
duces outputs so that each trial has the same probability of generating a point on
448 Financial Mathematics: An Introduction

equal subintervals, for example [0, 1/2) and [1/2, 1). Therefore, it is possible for
n trials to coincidentally all lie in the first half of the interval, while the (n + 1)th
point still falls within the other half with probability 1/2. This is not the case
with the quasi-random sequences, in which the outputs are constrained by a low-
discrepancy requirement that has a net effect of points being generated in a highly
correlated manner (i.e. the next point “knows” where the previous points are).
Illustration 14.2.1: Value of pi using Monte Carlo simulation
We consider a unit circle within a square with sides equal to 2 (see Fig. 14.1).

x2 + y2 = 1

−1 0 1 x

−1

Fig. 14.1. Value of pi by Monte Carlo simulation

Now if we pick a random point (x, y) where x and y are between −1 and 1,
the probability that this random point lies inside the unit circle is given by the
proportion between the area of the unit circle and the square. Thus
Area of the unit circle π
P(x2 + y2 < 1) = = .
Area of the square 4
Therefore if we pick a random point N times in the square and observe that M
times the point lies inside the unit circle, the probability that a random point lies
inside the unit circle is given by
M
P̃(x2 + y2 < 1) = ,
N
where P̃ indicates that it is a discrete distribution because M and N are integers.
But if N becomes very large, then as a consequence of the central limit theorem,
the above two probabilities will become equal. This gives
 Generating Random Variables 449

4M
π= .
N
Therefore to use Monte Carlo simulation, we generate a large number of random
(x, y) positions in the square, i.e. x and y are between −1 and 1. We next determine
which of these positions are inside the circle. Each time it is inside the circle, we
add ‘one’ to the counter. Hence by the methodology of Monte Carlo simulation
we have
Number of points inside circle
π≈4 .
Total number of point generated in a square

14.3 Generating Random Variables

In last section we have learned to generate pseudo-random numbers from the
uniform distribution U[0, 1]. In this section we use this knowledge to learn how
to generate a general distributed random numbers. One of the simplest methods
of generating random samples from a distribution with cumulative distribution
function (CDF) F(x) = P(X ≤ x) is based on the inverse of the CDF. The CDF is
an increasing function, however it is not necessarily continuous. Thus we define the
generalized inverse F− (u) := in f {x : F(x) ≥ u}, 0 ≤ u ≤ 1. The Fig. 14.2 illustrates
its definition. If F is continuous, then F− (u) = F−1 (u).
Theorem 14.3.1 (Inversion Method) Let U ∼ U[0, 1] and F be its CDF.
Then F− (U) has the CDF F.
Proof. It is easy to see (e.g. see, Fig. 14.2) that F− (U) ≤ x is equivalent to
U ≤ F(x). Thus for U ∼ U[0, 1]

P(F− (U) ≤ x) = P(U ≤ F(x)) = F(x) ,

which shows that F is the CDF of X is F− (U).



Inversion method can be summarized as follows
generate U ∼ U[0, 1]
Return X ← F− (U).
Illustration 14.3.1: Generation of Exponential Distribution
The exponential distribution with rate λ > 0 has the CDF

FX (x) = 1 − exp(−λx) for x ≥ 0 .

450 Financial Mathematics: An Introduction

Fig. 14.2. Illustration of the definition of the generalized inverse F− of a CDF F

Thus
1
F−X (u) = F−1
λ (u) = − ln(1 − u) .
λ
Therefore to generate the given exponential distribution, we first generate U ∼
U[0, 1] and then use the inversion method. This is because if U ∼ U[0, 1] then
1
− ln(1 − U) is exponential distribution with rate λ > 0. As U and 1 − U have
λ
1
the same distribution, we can use − ln(U) as well. The details are summarized
λ
as follows.
generate U ∼ U[0, 1]
ln(U)
X←−
λ
ln(1 − U)
(or Return X ← − ).
λ
Illustration 14.3.2: Generation of a Discrete Distribution
Let X be a discrete random variable with possible values ai (i = 1, 2, . . . , n). We
assume that a1 < a2 < . . . < an . The cumulative distribution function of X is given
by
 Generating Random Variables 451

q0 = 0 ,
i
qi = P(X = a j ) = FX (ai ) (i = 1, 2, . . . , n) .
j=1

Therefore to generate the given discrete distribution we employ the following steps.
Step 1. Generate U ∼ U[0, 1].
Step2. Find k ∈ {1, 2, . . . , n} such that qk−1 < U ≤ qk .
Step 3. Set X = ak .
The inversion method is a very efficient tool for generating random numbers.
However very few distributions possess CDF whose (generalized) inverse can be
evaluated efficiently. For example, CDF of a Gaussian distribution is not even
available in the closed form. Note however that the generalized inverse of the CDF
is just one possible transformation and that there might be other transformations
that yield the desired distribution. An example of such a method is the Box-Muller
method for generating Gaussian random variables.
Illustration 14.3.3: Box-Muller Method for Generating Normal Distri-
bution
This algorithm generates a sample form the standard bivariate normal distri-
bution, each component of which is thus a univariate standard normal variate.
This algorithm is based on the below given lemma.
Lemma 14.3.1 Let X1 , X2 ∼ N(0, 1) and independent, i.e. Z = (X1 , X2 ) ∼
N(0, I2 ), I2 being a (2 × 2) identity matrix. Then
(i) R = X12 + X22 is exponential distributed with mean 2, i.e. P(R ≤ x) = 1 − e−x/2 .
(ii)√Given R, the point (X1 , X2 ) is uniformally distributed on the circle of radius
R centered at the origin.

(iii) R and θ = tan−1 XX21 are independent random variables.

The above lemma suggests that to generate (X1 , X2 ), we may√first generate

R and then choose a point uniformally from the circle of radius R. In view of
Illustration 14.3.1, we may generate R by setting R = −2ln(U1 ) with U1 ∼ U[0, 1].
Next we generate a random point on the circle. For this we generate a
random angle θ between 0 and 2π. This is achieved by setting θ = 2πU2 ,
U2 ∼ U[0,
√ 1]. Therefore
√ the corresponding point on the circle has the coordi-
nates ( R cos θ, R sin θ). The details of the algorithm are as follows.
generate U1 , U2 independent from U[0, 1]
R ← −2ln(U1 )
452 Financial Mathematics: An Introduction

√ θ ← 2πU2 √
X1 ← R cos(θ), X2 ← R sin(θ)
return X1 , X2 .
The idea of transformation methods, like the inversion method, is to generate
random samples from a distribution other than the target distribution and to
transform them such that they come from the desired target distribution. In many
situations we cannot find such a transformation in closed form. In these cases we
have to find other ways of correcting for the fact that we sample from the wrong
distribution. The next two algorithms present two such ideas namely the rejection
sampling and importance sampling [102].
1. Acceptance- Rejection Sampling
In the sequel we would describe acceptance-rejection Monte Carlo method.
The acceptance-rejection method, introduced by Von Neumann, is among the
most widely applicable mechanisms for generating random samples. This method
generates samples from a target distribution by first generating candidates from
a more convenient distribution and then rejecting a random subset of generated
candidates. The rejection mechanism is designed so that the accepted samples are
indeed distributed according to the target distribution.
The basic idea of rejection sampling is to sample from an instrumental distri-
bution and reject samples that are “unlikely” under the target distribution.
Assume that we want to have a sample from a target distribution whose density
f is known to us. The simple idea underlying rejection sampling (and other Monte
Carlo algorithms) is the rather trivial identity
, f (x) , 1
f (x) = 1du = 10≤u≤ f (x) du .
0 0

Thus f (x) can be interpreted as the marginal density of a uniform distribution

on the area under the density f (x), i.e. {(x, u) : 0 ≤ u ≤ f (x)}. The Fig. 14.3
illustrates this idea. It suggests that we can generate a sample from f by sampling
from the area under the curve.
Illustration 14.3.4: Sampling from a Beta distribution
The Beta(a, b) distribution (a, b ≥ 0) has the density
Γ(a + b) a−1
f (x) = x (1 − x)b−1 , 0 < x < 1,
Γ(a)Γ(b)
: +∞
where Γ(a) = ta−1 e−t dt is the Gamma function. For a, b > 1 the Beta(a, b)
0
(a − 1)
density is unimodal with mode . The Fig. 14.3 shows the density of
(a + b − 2)
 Generating Random Variables 453

Beta(3, 5) distribution. It attains its maximum of 1680/729 ≈ 2.305 at x = 1/3.

Sampling from the area under the curve (dark gray) corresponds to sampling
from the Beta (3, 5) distribution. We use a uniform distribution of the light gray
rectangle as proposal distribution. Empty circles denote rejected values, filled
circles denote accepted values.

Fig. 14.3. Sampling from a Beta(3, 5) distribution

Using the above identity we can draw sample from Beta(3, 5) by drawing sample
from a uniform distribution on the area under the density {(x, u) : 0 ≤ u ≤ f (x)}
(the area shaded in dark gray in Fig. 14.3). We will sample from the light gray
rectangle and only keep the samples that fall in the area under the curve.
Mathematically speaking, we sample independently X ∼ U[0, 1] and U ∼
U[0, 2.4]. We keep the pair (X, U) if U < f (X), otherwise we reject it. The
conditional probability that a pair (X, U) is kept if X = x is P(U < f (X)/
X = x) = P(U < f (x)) = f (x)/2.4. As X and U were drawn independently we
can rewrite our algorithm as
Draw X from U[0, 1] and accept X with probability f (X)/2.4, otherwise reject X.
The method proposed in the above example is based on bounding the density of
the Beta distribution by a box. Whilst this is a powerful idea, it cannot be directly
applied to other distributions, as the density might be unbounded or have infinite
support. However we might be able to bound the density of f (x) by Mg(x), where
g(x) is a density that we can easily sample from.
454 Financial Mathematics: An Introduction

Algorithm (Rejection Sampling) Given two densities f, g with f (x) < Mg(x) for
all x, we can generate a sample from f by employing following steps.
Step 1. Draw X from density g.
f (X)
Step 2. Accept X as a sample from f with probability , otherwise go back
Mg(X)
to Step 1. This step is equivalent to the following step.
Generate U ∼ U[0, 1]. If U ≤ f (x)/Mg(x), accept X, otherwise go to Step 1.
The above algorithm can be justified mathematically as detailed below. We
have
, :
f (x) f (x) dx
P(X ∈ X and is accepted) = g(x) dx = X ,
X Mg(x) M
and thus
1
P(X is accepted) = P(X ∈ X and is accepted) = .
M
X is the domain of random variable X. This yields
P(X ∈ X and is accepted)
P(x ∈ X/X is accepted) =
P(X is accepted)
:
X
f (x) dx/M
=
1/M
,
= f (x) dx .
X

Thus the density of the values accepted by the algorithm is f (·).

Remark 14.3.1 If we know f only up to a multiplicative constant, i.e. if we only
know π(x), where f (x) = Cπ(x), we can carry out rejection sampling using

π(X)
Mg(X)

as probability of accepting X, provided π(x) < Mg(x) for all x.

Illustration 14.3.5: Rejection sampling from N(0, 1) distribution using
a Cauchy proposal
Assume we want to sample from the N(0, 1) distribution with density
 2
1 −x
f (x) = √ exp , − ∞ < x < ∞,
2π 2
 Generating Random Variables 455

using the Cauchy distribution with density

1
g(x) = , − ∞ < x < ∞,
π(1 + x2 )

  The smallest M we can choose such that f (x) ≤

as instrumental distribution.
√
Mg(x) is M = 2π exp −1 2
.
The Fig. 14.4 illustrates the results. As before, filled circles correspond to ac-
cepted values whereas open circles correspond to rejected values. Sampling from
the area under the density f (x) (dark gray) corresponds to sampling from the
N(0, 1) density. The proposal g(x) is a Cauchy(0, 1).

Fig. 14.4. Rejection sampling from N(0, 1) distribution using a Cauchy proposal

Note that it is impossible to do rejection sampling from a Cauchy distribution

using a N(0, 1) distribution as instrumental distribution. This is because there is
no M ∈ R such that
 2
1 1 −x
<M√ exp ,
π(1 + x )
2
2πσ2 2

as the Cauchy distribution has heavier tails than the Gaussian distribution.
2. Importance Sampling
In rejection sampling we have compensated for the fact that we sampled from
the instrumental distribution g(x) instead of f (x) by rejecting some of the values
proposed by g(x). Importance sampling is based on the idea of using weights to
correct for the fact that we sample from the instrumental distribution g(x) instead
of the target distribution f (x).
Importance sampling is based on the identity
456 Financial Mathematics: An Introduction

, , ,
f (x)
P(X ∈ X) = f (x) dx = g(x) dx = g(x) w(x) dx ,
X X g(x) X

for all g(·), such that g(x) > 0 for (almost) all x with f (x) > 0.
Suppose we are interested in estimating
, , ,
f (x)
α = E[h(X)] = h(x) f (x) dx = h(x) g(x) dx = h(x) w(x) g(x) dx .
X X g(x) X

We can have the following algorithm based on the principle of importance

sampling.
Algorithm (Importance Sampling)
Choose g such that supp( f h) ⊂ supp(g), where supp is support set i.e. supp(g) =
{x : g(x) > 0}. Follow the below given steps.
Step 1. For i = 1, . . . , n: generate Xi from density g.
f (Xi )
Step 2. Set w(Xi ) = .
g(Xi )
*n *n
w(Xi ) h(Xi ) w(Xi ) h(Xi )
Step 3. Return either µ̂ = i=1
*n or µ̃ = i=1
.
i=1 w(Xi ) n
In simulation, certain values of the input random variables have more impact on
the parameter being estimated than others. If these influential values are generated
by sampling more frequently, then the variance of estimator can be reduced. This
is possible by choosing another distribution to the random variables. This method
of importance sampling is called a variance reduction technique.
To look this differently, we observe that the error associated with a Monte
Carlo estimate is proportional to σ and inversely proportional to the square root
of the number of trials. Hence, there are only two ways of reducing the error in
a Monte Carlo estimate. These are either to increase the number of trials or to
reduce the variance. We introduce importance sampling techniques that reduce σ
and there by improve the efficiency of each trial. This is done by changing the
probability measure which is a standard tool in financial mathematics.
Let (Ω, F , P) be a probability space. Let X be a random variable with pdf f (x)
under P. Suppose we want to estimate θ = E[h(x)] under P. Let g(x) be another
pdf of the random variable X under Q. Then
, ,
θ = EP [h(x)] = h(x) f (x) dx = h(x) g(x) dQ(x) = EQ [h(x)] ,
 Generating Random Variables 457

f (x)
where dQ(x) = g(x) dx. Note that, EP [θ] = EQ [θ] where as VarQ [θ] < VarP [θ].
Hence, we should generate the samples of the random variable X using pdf g(x)
instead of f (x).
A common pitfall of importance sampling is that the tails of the distributions
matter. While g(x) might be roughly the same shape as f (x), serious difficulties
arise if g(x) gets small much faster than f (x) out in the tails. In such a case,
though it is improbable (by definition) that we will realize a value xi from the far
f (xi )
tails of g(x), if we do so then the Monte Carlo estimator will take a jolt for
g(xi )
such an improbable xi which may be orders of magnitude larger than the typical
f (x)
values that we see.
g(x)
Illustration 14.3.6: Geometric Distribution with Parameter p.
This is a discrete distribution which describes the number of independent trials
necessary to achieve a single success with the probability of a success on each trial
is p. We know that the probability mass function is
p(i) = p(1 − p)i , (i = 1, 2, . . .) ,
and the cumulative distribution function is
F(x) = P(X ≤ x) = 1 − (1 − p)[x] , x ≥ 0 ,
where [x] denotes the integer part of x. We wish to output an integer value of x
which satisfies the inequalities
F(x − 1) < U ≤ F(x) .
Solving these inequalities for integer x,we obtain
1 − (1 − p)x−1 < U ≤ 1 − (1 − p)x
(1 − p)x−1 > 1 − U ≥ (1 − p)x
(x − 1) ln(1 − p) > ln(1 − U) ≥ x ln(1 − p)
ln(1 − U)
(x − 1) < ≤x.
ln(1 − p)
We should therefore choose the smallest integer
 for X which  is greater
 than or
ln(1 − U) ln(1 − U) −E
equal to or equivalently, X = 1 + or 1 + where we
ln(1 − p) ln(1 − p) ln(1 − p)
write −ln(1 − U) = E, an exponential distributed random variable with parameter
1. In MATLAB, the geometric random number generators is called geornd.
458 Financial Mathematics: An Introduction

14.4 Monte Carlo Techniques

There are many considerations when using Monte Carlo techniques to perform
approximations. One of the chief concern is to get as accurate an approximation
as possible. Thus, for each method it is advisable to discuss their associated error
statistics. In this section we shall discuss Monte Carlo methods, namely, crude
Monte Carlo method and stratified sampling method [86, 110] in the context of
evaluation of integrals.
Consider the evaluation of following integral
, b
I= f (x) dx .
a

We generate N random sample values, xi , (i = 1, 2, . . . , N) in [a, b]. Next find

f (xi ), ( i = 1, 2, . . . , N) and obtain

b−a
N
IN = f (xi ) .
N i=1
It is well known that IN is unbiased, i.e. its expectation is I and, because of law
of large numbers, IN → I with probability 1 as n → ∞. The approximation error
is of order !(n−1/2 ). This follows from central limit theorem, or can be viewed
√
through the mean square error (MSE) E(IN − I)2 = VarIN = √σn .
We next describe stratified sampling method. Consider again the following in-
tegral
, b
I= f (x) dx .
a
The basic principle of stratified sampling method is to divide the interval [a, b]
into n subintervals and then to perform a crude Monte Carlo method on each
subinterval. Thus we write
, b , a1 , a2 , b
f (x) dx = f (x) dx + f (x) dx + . . . + f (x) dx ,
a a a1 an−1

and then apply crude Monte Carlo method in each of the integrals of the right
hand side. The reason we might use this method is that instead of finding variance
in one go, we can find the variance by adding up the variances of each subinterval.
This method is generally helpful when function is step like or has periods of flat.
Thus advantage of stratified sampling method is that we get to split the curve
into parts that could have certain advantageous properties when evaluating them
on their own.
 Applications in Finance 459

14.5 Applications in Finance

In this section, we present the applications of Monte Carlo simulation in option
pricing and risk management. Monte Carlo methods are ideal for pricing options
where the payoff is path dependent or options where the payoff is dependent on a
basket of underlying assets (rather than just a single asset).
The first application to European call option pricing was by Phelim Boyle in
1977 [19]. As with other option pricing techniques Monte Carlo methods are used
to price options using what is essentially a three step process. The three steps are
as follows
Step 1. Calculate potential future prices of the underlying asset(s).
Step 2. Calculate the payoff of the option for each of the potential underlying
price paths.
Step 3. Discount the payoffs back to today and average them to determine the
expected price.
The first step in using Monte Carlo methods is to generate (a large number of)
potential future asset prices. This is done by selecting an appropriate (stochastic)
model for the time evolution of the underlying asset(s) and then simulating the
model through time.
For example, the standard model for evolution of equity prices is given by the
Wiener process which is described as
  
σ2
S(
t) = S(0) exp µ −
t + (σ (
t)) .
2

Here S(0) is the stock price today, S(

t) is the stock price at a (small) time
into the future,
t is a small increment of time, µ is the expected return, σ is
the expected volatility,  is (random) number sampled from the standard normal
distribution.
Typically many thousands, if not tens of thousands, of simulated paths must
be generated to enable an accurate option price to be calculated. The more paths
that are generated the longer the simulation will take to be performed and hence
the longer the time taken to price the option.
Consider Monte Carlo simulation of European call option on a stock whose
current value is S(0). Let S(t) be the stock price at time t. The option expires at
time T and the strike price is K. We assume constant interest rate r and the stock
price follows a lognormal distribution with volatility parameter σ.
We know that, the price of the option is its discounted expected value. The
return on this simulation is the discounted difference between the terminal value
460 Financial Mathematics: An Introduction

of the stock and the strike price. Then, we repeat this many a times, averaging
the discounted returns to estimate the present value of the option.
We present the steps involved as follows.
Step 1. Generate several random price paths.
Step 2. Calculate the associated exercise value for each path.
Step 3. Find the average payoff.
Step 4. Find discounted average payoff.
Now we present the Monte Carlo algorithm for this option pricing. By using
random sampling of normal distribution and then using the Black-Scholes model,
the Monte Carlo algorithm can be described as follows.
For i = 1:N
E(i) = normalrnd(0, 1);
√
S(i) = S(0) exp((r − 0.5 σ2 )T + σ TE(i));
V(i) = exp (−rT Max (S(i) − K, 0)) ;
end
1 
N
ÂN = V(i).
N i=1

Note that estimator ÂN will be unbiased and consistent estimator since E(ÂN ) = A
and ÂN → A with probability 1 as n → ∞.
In the above mechanism, the payoff is determined by the terminal stock price
S(T) and does not otherwise depend on the evolution of S(t) between times 0 and
T. Each simulated path of the underlying asset consists of two points S(0) and
S(T).
In order to obtain a more accurate approximation to sampling we divide the
time interval [0, T] into smaller subintervals, we simulate the payoff of a derivative
security which may depend explicitly on the values of underlying asset at multiple
time points. We present the algorithm as follows. For 0 = t0 < t1 < . . . , tn = T,
√
S(i + 1) = S(i) exp((r − 0.5 σ2 )(ti+1 − ti ) + σ ti+1 − ti E(i)) .
We next illustrate the Monte Carlo method for interest rate model.
Illustration 14.5.1: Interest Rate Model Consider
√
drt = α(b − rt )dt + σ rt dWt , t ∈ [0, T] , (14.1)
where α and b are positive constants. If r0 > 0, then rt will never be negative. If
2αb ≥ σ2 , then rt remains strictly positive for all t, at most surely.
 Applications in Finance 461

A simple Euler discretization of (14.1) suggests simulating rt at times 0 = t0 <

t1 < . . . < tn = T by setting
+ √ 
rti+1 = rti + α(b − rti )(ti+1 − ti ) + σ r+ti ti+1 − ti Zi+1

with Zi , (i = 1, 2, . . . , n) i.i.d random variables each having N(0, 1). Here r+ti =
Max(rti , 0).
Using probability theory, observe that for 0 < u < t, given ru , rt is distributed
as  √ 2
rt = c Z + λ + χ2d−1
where Z ∼ N(0, 1), χ2 is chi-square distribution with d − 1 degrees of freedom and
 
σ2 1 − e−α(t−u) 4bα 4 ru α e−α(t−u) ru e−α(t−u)
c= ; d= ; λ = = .
4α σ2 σ2 (1 − e−α(t−u) ) c
We now present the details of algorithm.
Case 1: d > 1.
For i = 1, 2, . . . , n − 1 
σ2 1 − e−α(ti+1 −ti ) 4 rti αe−α(ti+1 −ti )
Step 1. ci = , λi =
4α c
Step 2. Generate Zi ∼ N(0, 1)
Step 3. Generate Xi ∼ χ2d−1
 √ 2
Step 4. rti+1 ← c Zi + λi + Xi
end

Case 2: d ≤ 1.
For i = 1, 2, . . . , n − 1 
σ2 1 − e−α(ti+1 −ti ) 4 rti αe−α(ti+1 −ti )
Step 1. ci = , λi =
4α c
Step 2. Generate N ∼ Poisson (λ/2)
Step 3. Generate Xi ∼ χ2d+2N
Step 4. rti+1 ← cXi
end
Illustration 14.5.2: Portfolio Optimization
Let us consider a portfolio consisting of two stocks A and B. Let SA (0) and
SB (0) respectively denote the process of these stocks at time t = 0. The prices of
462 Financial Mathematics: An Introduction

these stocks at time t are denoted by SA (t) and SB (t). At time t = 0, the portfolio
value V(0) is given by
V(0) = nA SA (0) + nB SB (0) ,
where the portfolio consists of nA units of stock A and nB units of stock B. Also
let us suppose SA (t) and SB (t) follow the geometric Brownian motion with drift µA
and volatility σA and drift µB and volatility σB respectively. Further, let SA (t) and
SB (t) be independent. We are interested in finding the estimate of probability that
the value of portfolio drops by more than 15% at maturity date T using Monte
Carlo simulation.  
V(T)
The desired probability is θ = P ≤ 0.85 . The steps for estimating θ are
V(0)
as follows.
For i = 1 to N; generate Xi = (S(i)
A
(T), S(i)
B
(T)).
Compute ⎧
⎪
⎪ nA S(i) (T) + nB S(i) (T)
⎪
⎨ 1, if A B
≤ 0.85
I(Xi ) = ⎪
⎪ nA SA (0) + nB SB (0)
⎪
⎩ 0, if otherwise
end.
Set θ̂N = N1 (I(X1 ) + I(X2 ) + . . . + I(XN )) .
The estimator θ̂N will be unbiased and consistent estimator since E(θ̂N ) = θ and
θ̂N → θ with probability 1 as n → ∞.

14.6 Summary and Additional Notes

• In this chapter we discussed linear congruential generators to generate pseudo
random numbers. However there are few other generators available in the litera-
ture. One of these is the mixed congruential generator which is a generalization
of the linear congruential generators.
• Most common random numbers that have been talked about in the financial
simulations are random numbers generated from uniform distribution and nor-
mal distribution. However other random numbers that have been generated
from non-uniform continuous distributions using inverse transform are expo-
nential random numbers, Cauchy random numbers, Pareto random numbers,
extreme values and logistic random numbers. Some of them are very useful for
financial applications that involve fat tail distributions.
 Exercises 463

• For most of the distributions MATLAB have readily available functions to

generate random numbers with those distributions. For e.g. such as normal:
normrnd(µ, σ, 1, n), Students t: trnd(ν, 1, n), exponential: exprnd(λ, 1, n), uni-
form: unifrnd(a, b, 1, n) or rand(1, n) if a = 0, b = 1, Weibull: weibrnd(a, b, 1, n),
gamma: gamrnd(a, b, 1, n), Cauchy: a+b∗trnd(1, 1, n), Poisson: poissrnd(λ, 1, n),
binomial: binornd(m, p, 1, n),.
• There are variance reduction techniques available in the literature to improve
the speed and efficiency of a simulation. Some of these are antithetic sampling,
control variates, variance reduction by conditioning and stratified sampling.
We may refer to Glasserman et al [50] in this regard.
• Quasi-Monte Carlo methods are purely deterministic, numerical analytic meth-
ods in the sense that they do not even attempt to emulate the behavior of inde-
pendent uniform random variables, but rather cover the space in d-dimensions
with fewer gaps than independent random variables would normally admit.
Quasi Monte Carlo frequently generates estimates superior to the Monte Carlo
methods in many problems of low or intermediate effective dimension. If the
dimension is large, but a small number of variables determine most of the vari-
ability in the simulation, then we might expect Quasi Monte-Carlo methods to
continue to perform well. The common method to generate Quasi Monte Carlo
numbers are the Halton sequence [56], the Sobol sequence [123].
• In this chapter we have attempted to cover some of the most basic aspects of
Monte Carlo methods in quantitative finance. For further details the readers
may refer to Glasserman [49] and Jackel [69].

14.7 Exercises
Exercise 14.1 Consider the linear congruential generator xn+1 = Axn +C (modM)
with M = 64. Generate the pseudo-random numbers with A = 4, C = 1 and x0 = 2.
Determine the period of the generator starting with seed x0 = 2 and with seed
x0 = 3.
Exercise 14.2 Consider the uniformly distributed random variable X on the in-
terval [0, 1]. Find a function of X which is uniformly distributed on the interval
[0, 2]. Generate the sequence of random numbers of uniformly distributed on the
interval [0, 2].
Exercise 14.3 Using inverse transform method, develop an algorithm for simu-
lation of Poisson distribution with parameter λ.
464 Financial Mathematics: An Introduction

, 1
e−x dx, by the crude Monte Carlo method. Also, eval-
4
Exercise 14.4 Evaluate
0
uate it by acceptance-rejection method.
Exercise 14.5 Consider a multivariate normal random vector * X = (X1 , X2 , . . . , Xn )
having mean vector (µ1 , µ2 , . . . , µn ) and covariance matrix of order n×n matrix.
We wish to generate
* this n-dimensional random vector.
* The procedure involves a

decomposition of into factors such that A A = . Suppose Z = (Z1 , Z2 , . . . , Zn )
is a vector of independent standard normal random variable, the required n-
dimensional random vector is given by X = (µ1 , µ2 , . . . , µn ) + ZA .
Exercise 14.6 Consider a Vasicek interest rate model

drt = α(b − rt )dt + σdWt , t ∈ [0, T],

where α, b and σ are positive constants. Prove the algorithm for the exact sim-
−α(ti+1 −ti )rti
ulation at time 0 = + t0 < t1 < . . . < tn = T is given by rtn = e +
 
b 1 − e−α(ti+1 −ti ) + σ 2α
1
(1 − e−2α(ti+1 −ti ) ) Zi+1 where Zi is the standard normal sam-
ple.
(Hint: refer to Chapter 11.)
15
MATLAB Codes for Selected Problems
in Finance

15.1 Introduction
MATLAB has a very useful toolbox, namely Financial Toolbox, which contains
many standard programs of finance. The aim of presenting the MATLAB codes
in this book is to encourage readers to attempt writing codes for similar problems
in financial mathematics, taking these codes as building blocks.
% at the beginning of any statement in MATLAB code stands for comment
lines, which are added in the codes to facilitate understanding of programs.

15.2 Binomial Lattice Model for European and American

Call Option/Put Option
% Binomial Lattice Model for European and American call
% option/put option with generic input parameter values.
% Symbols have their usual meanings.

% Starting the code.

% Generic input parameter values.

S0 = input(’S0 = ’);
u = input(’u = ’);
d = input(’d = ’);
T = input(’T = ’);
k = input(’k = ’); % Strike price
r = input(’r = ’);
466 Financial Mathematics: An Introduction

R = 1 + r;

% For instance, we can input

% S0 = 100; u = 1.3; d = 0.9; T = 3; k = 110; R = 1.1.

%risk neutral probability measure is p

p = (R-d)/(u-d);

% Finding of binomial lattice for stock market movements.

BinomialLattice = zeros(T,T);
r = 0; % power of d in the binomial lattice
for i =T:-1:1 % row generation
r = 1;
% i+j = T+1 => j= T+1-i % constructing a lower triangle
for j = T+1-i:T % column generation
BinomialLattice(i,j) = S0*uˆ(T-i)*dˆ(r-1);
r = r+1;
end
end disp(’Binomial lattice for stock movement ’);
disp(BinomialLattice);

% Asking for American or European call/put options.

fprintf(’\n 1 : for American options

\n 2 : for European options \n ’);
choice_A_E = input(’Your choices ’);
switch choice_A_E
case 1
fprintf(’\n 1 : for call option \n 2 : for put option ’);
choice_p_c = input(’your choice ’);
c0 = zeros(T,T);
c0(:,T) = max(BinomialLattice(:,T) - k, 0);
if choice_p_c == 1
for j = T-1:-1:1 % j = column traversal
for i = T:-1:2 % i = row traversal
if i+j >=T+1 % construct a lower triangle
 Binomial Lattice Model for European and American Call Option/Put Option 467

c0(i,j) = max(BinomialLattice(i,j) - k,
(1/R)*(p*c0(i-1,j+1)+(1-p)*c0(i,j+1)));
end
end
end
disp(’American call option binomial lattice mm’);
disp(c0);
else
p0 = zeros(T,T);
p0(:,T) = max(k - BinomialLattice(:,T), 0);
for j = T-1:-1:1 % j = column traversal
for i = T:-1:2 % i = row traversal
if i+j>=T+1
p0(i,j) = max(k - BinomialLattice(i,j),
(1/R)*(p*p0(i,j+1)+(1-p)*p0(i-1,j+1)));
end
end
end
disp(’American put option binomial lattice ’);
disp(p0);
end
break;
case 2
fprintf(’\n 1 : for call option \n 2 : for put option ’);
choice_p_c = input(’your choice ’);
c0 = zeros(T,T);
c0(:,T) = max(BinomialLattice(:,T) - k, 0);
if choice_p_c == 1
for j = T-1:-1:1 % j = column traversal
for i = T:-1:2 % i = row traversal
if i+j >=T+1 % construct lower triangle
c0(i,j) =(1/R)*(p*c0(i-1,j+1)+(1-p)*c0(i,j+1));
end
end
end
disp(’European call option binomial lattice ’);
disp(c0);
else
468 Financial Mathematics: An Introduction

p0 = zeros(T,T);
p0(:,T) = max(k-BinomialLattice(:,T),0);
for j = T-1:-1:1 % j = column traversal
for i = T:-1:2 % i = row traversal
if i+j>=T+1
p0(i,j) =(1/R)*(p*p0(i,j+1)+(1-p)*p0(i-1,j+1));
end
end
end
disp(’European put option binomial lattice ’);
disp(p0);
end
break;
otherwise
disp(’Unknown Choice.’);
end

15.3 CRR Model for European and American Call

Option/Put Option
% CRR model for European and American call option/put option
% with generic input parameter values.
% Symbols have their usual meanings.

% Starting the code.

% Generic input parameter values.

S0 = input(’S0 = ’);
T = input(’T = ’); % Expiration time
n = input(’n = ’); % no of periods
k = input(’k = ’); % Strike price
r = input(’r = ’);
sigma = input(’Sigma = ’) % volatility of the stock price.

% For example we can take

%S0 = 100; T = 4; n = 2; k = 110; r = 0.05; sigma = 0.24.
 CRR Model for European and American Call Option/Put Option 469

deltat = T/n;
R = exp(r*deltat);
u = exp(sigma*sqrt(deltat));
d = 1/u;
p = (R-d)/(u-d); % risk neutral probability measure.

% Finding of binomial lattice for stock market movements.

BinomialLattice = zeros(T,T);
r=0; % power of d in the binomial lattice
for i = T:-1:1 % row generation
r = 1;
% i+j = T+1 =>j= T+1-i % create lower triangle display
for j = T+1-i:T % column generation
BinomialLattice(i,j) = S0*uˆ(T-i)*dˆ(r-1);
r = r+1;
end
end disp(’Binomial lattice for stock movement ’);
disp(BinomialLattice);

% Asking choice for American or European call/put option.

fprintf(’\n 1 : for American options

\n 2 : for Europen options \n ’);
choice_A_E = input(’Your choices ’); switch choice_A_E

case 1
fprintf(’\n 1 : for call option \n 2 : for put option ’);
choice_p_c = input(’your choice ’);
c0 = zeros(T,T);
c0(:,T) = max(BinomialLattice(:,T) - k,0);
if choice_p_c == 1
for j = T-1:-1:1 %j = column traversal
for i = T:-1:2 %i = row traversal
if i+j >=T+1 % create lower triangle
470 Financial Mathematics: An Introduction

c0(i,j) = max(BinomialLattice(i,j) - k,
(1/R)*((1-p)*c0(i-1,j+1)+(p)*c0(i,j+1)));
end
end
end
disp(’American call option binomial lattice mm’);
disp(c0);
else
p0 = zeros(T,T);
p0(:,T) = max(k-BinomialLattice(:,T),0);
for j = T-1:-1:1 %j = column traversal
for i = T:-1:2 %i = row traversal
if i+j>=T+1
p0(i,j) = max(k - BinomialLattice(i,j),
(1/R)*(p*p0(i,j+1)+(1-p)*p0(i-1,j+1)));
end
end
end
disp(’American put option binomial lattice ’);
disp(p0);
end
break;
case 2
fprintf(’\n 1 : for call option \n 2 : for put option ’);
choice_p_c = input(’your choice ’);
c0 = zeros(T,T);
c0(:,T) = max(BinomialLattice(:,T)-k,0);
if choice_p_c == 1
for j = T-1:-1:1 % j = column traversal
for i = T:-1:2 % i = row traversal
if i+j >=T+1 % To make lower triangle
c0(i,j) =(1/R)*((1-p)*c0(i-1,j+1) + (p)*c0(i,j+1));
end
end
end
disp(’European call option binomial lattice ’);
disp(c0);
else
 Black Scholes Formula for Dividend and Non Dividend Paying Stock 471

p0 = zeros(T,T);
p0(:,T) = max(k - BinomialLattice(:,T),0);
for j = T-1:-1:1 %j = column traversal
for i = T:-1:2 %i = row traversal
if i+j>=T+1
p0(i,j) = (1/R)*((p)* p0(i,j+1) + (1-p)*p0(i-1,j+1));
end
end
end
disp(’European put option binomial lattice ’);
disp(p0);
end
break;
otherwise
disp(’Unknown Choice.’);
end

15.4 Black Scholes Formula for Dividend and Non

Dividend Paying Stock
% Black Scholes formula for dividend and non dividend paying stock.
% Symbols have their usual meanings.

% Starting the code.

s0 =input(’Enter the stock price at t = 0 : S0 = ’);

k = input(’Enter the strike price for underlying asset ’);
r = input(’Enter the risk free rate of interest ’);
T = input(’Enter the expiry time ’);
sigma=input(’Enter the volatility for underlying asset ’);

% For example, s0 = 100; k = 125; r = 0.05;

% T = 3/12; sigma = 0.24.

CH = 1;

while CH == 1
472 Financial Mathematics: An Introduction

fprintf(’1 : for Dividend \n 2 : non Dividend \n ’);

choice = input(’ choice ’);
switch choice % for Dividend and non dividend
case 1
disp(’Data for dividend paying stocks ’);
dtime = input(’Enter the time in array ’);
damount = input(’Enter the dividend amount in array ’);
rd = -r*dtime;
erd = exp(rd);
ds = damount.*erd;
sa = s0 - sum(sum(ds));
lsa = (log(sa/k) + (r + (sigma * sigma)/2)*T);
d1 = lsa/(sigma * sqrt(T));
d2 = d1 - sigma * sqrt(T);

fprintf(’1 : call option \n2 : put option \n ’);

ch = input(’ choice for put or call ’);
if ch == 1 % for call option
c0 = sa*normcdf(d1) - k * exp(-r*T) * normcdf(d2);
fprintf(’Value of call option = %4.4f\n’,c0);
end
if ch == 2 % for put option
p0 = k * exp(-r*T) * normcdf(-d2)- sa * normcdf(-d1);
fprintf(’Value of put option = %4.4f\n’,p0);
end
if ch ˜= 1 && ch ˜= 2 % ˜ means not equal to
disp(’Invalid choice’);
end
case 2 % for non dividend
lso = (log(s0/k) + (r + (sigma. * sigma)/2)*T);
d1 = lso/(sigma * sqrt(T));
d2 = d1 - sigma * sqrt(T);
fprintf(’1 : call option \n 2 : put option \n ’);
ch = input(’ choice for put or call ’);
if ch == 1 % call option
c0 = s0 * normcdf(d1) - k * exp(-r*T) * normcdf(d2);
fprintf(’\n Value of call option = %f\n’,c0);
end
 Minimum Variance Portfolio 473

if ch == 2 % put option
c0 = s0 * normcdf(d1)- k * exp(-r*T) * normcdf(d2);
p0 = c0 - s0 + k * exp(-r*T);
fprintf(’\n Value of put option = %4.4f\n’,p0);
end
if ch ˜= 1 && ch ˜= 2
disp(’Invalid choice’);
end
end
CH = input(’1-continuing option pricing else any other key: ’);
end

15.5 Minimum Variance Portfolio

% ri is the column vector of return over the given periods.
% It can be generated either by historical data from the market
% or by simulation.
% Here it is done by simulation i.e.,
% ri is randomly generated from standard normal distribution.

% R = [r1 r2 ... rn] returns matrix.

% We take n = 10, the number of assets.

% We take 30 periods of return.

R = randn(30,10); % returns of assets over period of time.

C = cov(R); % Covariance matrix of returns.
e = ones(10,1);

pC = pinv(C); % Finding inverse of the covariance matrix.

w = pC*e/(e’*pC*e); % Weights of assets in portfolio.

disp(’Minimum variance portfolio is ’);

disp(w);
474 Financial Mathematics: An Introduction

15.6 Markowitz Efficient Frontier

% ri is the column vector of return over the given periods,

% which can be generated either by historical data from the
% market or by simulation.
% Here it is taken by simulation randomly generated
% from standard normal distribution.
% R = [r1 r2 ... rn] returns matrix;
% Here we have taken n = 10, number of assets,
% and there are 30 periods of return.

% Start the code.

% ri = input(’Enter the return vector of the asset i ’);

R = randn(30,10); % Returns of assets over period of time;

C = cov(R); % Covariance matrix of the returns.
e = ones(10,1);

pC = pinv(C); % Finding inverse of the covariance matrix.

wmin = pC*e/(e’*pC*e);

disp(’Minimum variance portfolio is ’);

disp(wmin);

disp(’Expected return of each assets over the period ’);

meanR = mean(R);
disp(meanR);

mu1 = meanR * wmin;

disp(’Expected return at minimum variance portfolio ’);
disp(mu1);
fprintf(’Enter the expected return greater than %f’,mu1);
mu2 = input(’ mu2 = ’);

% We calculate the portfolio for the given mu2.

 Marowitz Efficient Frontier: User Specified Data 475

A = [mu2 meanR * pC * e ; 1 e’ * pC * e];

B = [meanR * pC * meanR’ mu2; e’ * pC * meanR’ 1];

D = [meanR * pC * meanR’ meanR * pC * e ; e’*pC * meanR’ e’ * pC * e];

w2 = (det(A)*(pC * meanR’) + det(B) * (pC * e))/(det(D));

disp(’Another portfolio ’);

disp(w2);

% Using the two fund theorem, we find the efficient frontier.

port = zeros(10,1);
effr = zeros(10,1);
effsigma = zeros(10,1);
for i=1:11
lambda =(i-1)/10 ;
port = (1-lambda)* wmin + (lambda*w2);
effr(i,1)= meanR * port;
effsigma(i,1) = port’* C * port;
end plot(effsigma, effr);
xlabel(’Sigma’);
ylabel(’Return’);

15.7 Marowitz Efficient Frontier: User Specified Data

% The program calls in-built MATLAB function ‘quadprog’
% for solving a Markowitz quadratic program model.

% Symbols have their usual meanings.

% Start the code.

% Set-up
if ˜exist(’quadprog’)
msgbox(’The optimization toolbox is required to run
this demo.’, ’Product dependency’)
476 Financial Mathematics: An Introduction

return
end

n = input(’Enter the number of assets = ’);

str = [’Enter the expected returns of ’num2str(n)’
Companies in 1 x ’num2str(n)’ matrix form = ’];
returns = input(str);

str1 = [’Enter the standard deviation in 1 x ’num2str(n)’

matrix form = ’];
STDS = input(str1);

str2 = [’Enter the correlation matrix in ’num2str(n)’ x ’num2str(n)’

matrix form = ’];
correlations = input(str2);

% Convert to variance-covariance matrix.

% Convert the standard deviation and correlation.

covariances = corr2cov(STDS, correlations);

% Compute and plot efficient frontier for 20 portfolios.

% The MATLAB financial toolbox in built function ‘portopt’
% output the mean-variance efficient frontier with user-specified
% covariance and returns.

portopt(returns, covariances, 20)

% Randomize assets weights.

% Randomly generate the asset weights for 1000 portfolios,
% a uniformly distributed 1000 x n matrix of
% pseudo-random numbers in (0,1).

rand(’state’, 0)
weights = rand(1000, n);
 CAPM with Generic Input Parameters 477

% Normalized the weights so that their sum is 1.

total = sum(weights, 2); % sum of square of weights.

total = total(:, ones(n,1));
weights = weights. / total;

% Compute expected returns and risks for each portfolio,

% using a MATLAB financial toolbox in-built
% function ‘portstats’

[portRisk,portReturn]=portstats(returns,covariances,weights);

% Now plot the efficient frontier.

hold on
plot(portRisk, portReturn, ’.r’)
title(’Mean-Variance efficient frontier for random portfolios’)
hold off

15.8 CAPM with Generic Input Parameters

% First we generate returns of 20 companies,

% for last one year (12 months) using simulation
% Also generate return for market portfolio for last one year.
% But one can easily generalize the program.

% Start the code.

% MR = market return,
% it can taken from historical data of any market.
% Data collections/genration

MR = randn(52,1); % Random weekly returns for market;

R = randn(52,20); % weekly returns of assets a1...a20 for one year;
MMR = mean(MR);
478 Financial Mathematics: An Introduction

% C(i) = Covariance(MR,R(i)) of the i-th asset.

C = zeros(20,1);

for i = 1:20;
MERR = mean(MR. * R(:,i));
C(i,1) = MERR - MMR * mean(R(:,i));
end

VMR = var(MR);

% Calculating the beta coefficient.

% beta(i) = beta coefficient of i-th asset.

% beta(i) = C(i)/var(MR);

beta = zeros(20,1);
for i = 1:20
beta(i) = C(i,1)/VMR;
end

% Calculating the return of each asset form CAPM

% r_i = risk free return + beta_i *(E(M)- risk free return);
% rf = input(’Enter the risk free interest rate per week ’);

rf = 0.005; % Assumed per week risk-free-return.

r = zeros(size(beta));

for i = 1:length(beta)
r(i) = rf + beta(i)*(MMR-rf);
end

fprintf(’\tAseets \t beta \t\t asset_return_pm \n’);

for i =1 :20
fprintf(’\t a%d \t %f \t\t %f\n’,i, beta(i),r(i));
 Geometric Brownian Motion using Simple Random Walk 479

end

15.9 Geometric Brownian Motion using Simple Random

Walk

% The Geometric_brownian(N,r,sigma,T)
% simulates a Geometric Brownian motion on [0,T].

% Recall: If X is a Brownian motion then exp(X) is a

% geometric Brownian motion.

function [Y] = geometric_brownian(N,r,sigma,T)

% N = input(’number of step =’);

N = 1000;
T = input(’Enter length of interval =’);
r = input(’Enter interest rate = ’);
sigma = input(’Enter diffusion coefficient = ’);

t = (0:1:N)’/N; % t is the column vector [0 1/N 2/N ... 1]

W = [0; cumsum(randn(N,1))]/sqrt(N);
% cumsum function is running sum of normal N(0,1/N) variables.

t = t * T;
W = W * sqrt(T);
X = (r-(sigmaˆ2)/2)* t + sigma * W;
Y = exp(X);

plot(t,Y);
title([’Sample Path of Geometric Brownian motion with
diffusion coefficient=’num2str(sigma),’Interest rate=’num2str(r)])
xlabel([’Time’])
480 Financial Mathematics: An Introduction

15.10 Binomial Lattice for Bond Price and Interest Rate

Modeling
% Binomial lattice model for bond price & Interest rate modeling.
% Symbols have their usual meanings.

% Start the code.

% Generic input parameter values.

r00 = input(’r00 = ’);

u = input(’u = ’);
d = input(’d = ’);
T = input(’T = ’);
k = input(’k = ’); % Strike price

% For example, r00 = 0.07; u = 1.3; d = 0.9; T = 5; k = 90;

% Finding binomial lattice for interest rate modeling

IntLattice = zeros(T,T);

r=0; % power of d in the Binomial lattice

for i = T:-1:1 % row generation
r = 1;
% i+j = T+1 =>j= T+1-i
% it will make a lower triangle matrix

for j = T+1-i:T % column generation

IntLattice(i,j) = r00*uˆ(T-i)*dˆ(r-1);
r = r+1;
end
end disp(’Short interest rate dynamics’)
disp(IntLattice);

% Finding bond price which is the underlying

% asset for call/put options.
 Binomial Lattice for Bond Price and Interest Rate Modeling 481

p = zeros(T+1,T+1);
x = 1; % We want find in percentage later will be converted
p(:,T+1) = x;
for j = T:-1:1 % j = column traversal
for i = T+1:-1:2 %i = row traversal
if i+j >=T+2 % To make a lower triangle
p(i,j) =(1.0/(1 + IntLattice(i-1,j)))*
(0.5*p(i-1,j+1)+0.5*p(i,j+1));
end
end
end

disp(’Underlying assets bond price ’) disp(p)

% Calculation of Call option where underlying assets is bond price.

C = zeros(T,T);
for i = T+1:-1:2
C(i-1,T) = max(p(i,T)*100-k,0);
end
for j = T-1:-1:1 % j = column traversal
for i = T:-1:2 % i = row traversal
if i+j >=T+1 % To make a lower triangle
C(i,j) =(1.0/(1 + IntLattice(i,j)))*
(0.5*C(i,j+1)+0.5*C(i-1,j+1));
end
end
end

disp(’Binomial lattice for call price ’);

disp(C);

% Calculation of Put option where underlying assets is bound price.

P = zeros(T,T);
for i = T+1:-1:2
P(i-1,T) = max(k - p(i,T)*100, 0);
end
482 Financial Mathematics: An Introduction

for j = T-1:-1:1 % j = column traversal

for i = T:-1:2 % i = row traversal
if i+j >=T+1 % construct a lower triangle
P(i,j) =(1.0/(1 + IntLattice(i,j)))*
(0.5*P(i-1,j+1)+0.5*P(i,j+1));
end
end
end
disp(’Binomial lattice for put option ’);
disp(P);

15.11 European/American Call Options via Monte Carlo

Method
% S(i) Stock price is consider as Geometric Brownian motion.
% Simulate the solution of Black-Scholes model
% stochastic differential equation:
% ds(t) = mu * s(t)dt + sigma * s(t) * dw(t).
% mu = r = expected return = drift rate.

% Symbols have their usual meanings.

% Start the Code

% Generic input parameter values.

S0 = input(’S0 = ’);
T = input(’T = ’);%Expiration time
k = input(’k = ’);
r = input(’r = ’);
sigma = input(’Sigma = ’) % volatility of stock price.
N = input(’No of simulations’);

% For example, we can take, S0 = 100; T = 4; k = 110;

% r = 0.05; sigma = 0.24; N = 100;
E = zeros(N,1);
S = zeros(N,1);
V = zeros(N,1);
 European/American Call Options via Monte Carlo Method 483

fprintf(’\n 1 : for European options \n 2 : for American options \n ’);

choice_A_E = input(’Your choices ’); switch choice_A_E
case 1
fprintf(’\n 1 : for call option \n 2 : for put option ’);
choice_p_c = input(’your choice ’);
if choice_p_c == 1
for i=1:N
E(i) = normrnd(0,1);
S(i) = S0*exp((r-(1/2)*sigmaˆ2)*T + sigma*sqrt(T)*E(i));
V(i) = (exp(-r*T))* max(S(i) - k,0);
end
disp(’call price ’);
c0 =(1/N) * sum(V);
disp(c0)
else
for i=1:N
E(i)= normrnd(0,1);
S(i)= S0*exp((r-(1/2)*sigmaˆ2)*T + sigma*sqrt(T)*E(i));
V(i)=(exp(-r*T))* max(k - S(i),0);
end
p0 =(1/N) * sum(V);
disp(’call price ’);
disp(p0)
end
case 2
fprintf(’\n 1 : for call option \n 2 : for put option ’);
choice_p_c = input(’your choice ’);
if choice_p_c == 1
for i = 1:N
E(i) = rand(0,1);
S(i) = S0*exp((r-(1/2)* sigmaˆ2)* T + sigma*sqrt(T)*E(i));
V(i) = (exp(-r*T))* max(S(i)- k,0);
end
c0 =(1/N) * sum(V)
else
for i=1:N
E(i) = normrnd(0,1);
S(i) = S0*exp((r-(1/2)* sigmaˆ2)* T + sigma*sqrt(T)*E(i));
484 Financial Mathematics: An Introduction

V(i) = (exp(-r*T))* max(k - S(i),0);

end
p0 =(1/N)*sum(V)
end
end
References

1. F. Aitsahila, Y. -C. Sheu and P. M. Pardalos, Optimal execution of time-

constrained portfolio transaction,in Computation Methods in Financial Engi-
neering, Part 1 (J. Erricos, J. Kontoghiorghes, B. Rustem, P. Winker, editors),
pp. 95–102, 2008.
2. R. Almgren and N. Chriss, Optimal execution of portfolio transactions, Jour-
nal of Risk, Vol. 3(2), pp. 5–39, 2000.
3. F. Anderson, H. Mausser, D. Rosen, and S. Uryasev, Credit risk optimization
with conditional value at risk criterion. Mathematical Programming, Vol. 89,
pp. 273–291, 2000.
4. D. Applebaum, Levy Processes and Stochastic Calculus, Cambridge University
Press, 2nd edition, 2009.
5. P. Artzner, F. Delbaen, J. M. Eber and D. Heath, Coherent measures of risk,
Mathematical Finance, Vol. 9(3), pp. 203–228, 1999.
6. R. B. Avery, R. W. Rostic, P. S. Calem and G. B. Canner, Credit scoring:
statistical issues and evidence from credit-bureau files, Real Estate Economics,
Vol. 28(3), pp. 523–547, 2000.
7. M. Bartholomew-Biggs, Nonlinear Optimization with Financial Applications,
Kluwer Academic Publishers, 2005.
8. M. Baxter and A. Rennie, Financial Calculus: An Introduction to Derivative
Pricing, Cambridge University Press, 1996.
9. A. Ben-Tal, T. Margalit and A. N. Nemirovski, Robust modeling of multistage
portfolio problems, in High Performance Optimization (H. Frenk, K. Roos,
T. Terlsky, and S. Zheng editors), Kluwer Academic Publishers, pp. 303–328,
2002. http://www2.isye.gatech.edu/∼nemirovs/rob portf.pdf.
10. M. J. A. Berry and G. S. Linoff, Mastering Data Mining: The Art and Science
of Customer Relationship Management, John Wiley & Sons, 2000.
11. D. Bertsimas and A. Lo, Optimal control of execution costs, Journal of Fi-
nancial Markets, Vol. 1, pp. 1–50, 1998.
12. D. Bertsimas, P. Hummel and A. Lo, Optimal control of execution costs for
portfolios, Computing in Science and Engineering, Vol. 1(6), pp. 40–53, 1999.
486 Financial Mathematics: An Introduction

13. F. Black, E. Derman and W. Toy, A one-factor model of interest rates and its
application to treasury bond options, Financial Analysts, Vol. 22, pp.33–39,
1990.
14. F. Black and P. Karasinski, Bond and option pricing when short rates are
lognormal, Financial Analysts, Vol. 47, pp. 52–59, 1991.
15. F. Black and M. Scholes, The pricing of options and corporate liabilities,
Journal of Political Economy, Vol. 81, pp. 637–659, 1973.
16. M. E. Blume and D. B. Kein, The valuation of callable bonds, Wharton Fi-
nancial Institution Center,
http://finance.wharton.upenn.edu/∼rlwctr/papers/8914.PDF
17. E. Bouyé, V. Durrleman, A. Nikeghbali, G. Riboulet and T. Roncalli, Copulas
for finance: A reading guide and some applications, 2000.
http://thierry-roncalli.com/download/copula-survey.pdf
18. E. Bouyé, V. Durrleman, A. Nikeghbali, G. Riboulet and T. Roncalli, Copulas:
an open field for risk management, 2003.
http://www.thierry-roncalli.com/download/copula-rm.pdf
19. P. Boyle, Options: a Monte Carlo approach, Journal of Financial Economics,
Vol. 4, pp. 323–338, 1977.
20. P. P. Boyle and T. Vorst, Option replication in discrete time with transaction
costs, Journal of Finance, Vol. 47, pp. 271-293, 1992.
21. D. Brigo and F. Mercurio, Interest Rate Models - Theory and Practice: With
Smile, Inflation and Credit, Springer Finance, 2nd edition, 2006.
22. N. Bucay and D. Rosen, Credit risk of an international bond portfolio: a case
study, ALGO Research Quarterly, Vol. 2(1), pp. 9–29, 1999.
23. X. Cai, K. -L. Teo, X. Yang and X. Y. Zhou, Portfolio optimization under a
minimax rule, Management Science, Vol. 46, pp. 957–972, 2000.
24. A. J. G. Cairns, Interest Rate Models, Princeton University Press, 2004.
25. M. Capinski and T. Zastawniak, Mathematics for Finance: An Introduction
to Financial Engineering, 2nd edition, Springer 2010.
26. M. M. Chawla, On the bionomial tree method for the valuation of options,
International Journal of Applied Mathematics, Vol. 20(2), pp. 235–251, 2007.
27. M. M. Chawla, On mean variance portfolio optimization, International Jour-
nal of Applied Mathematics, Vol. 21(3), pp. 473–494, 2008.
28. Y.-W. Chen and C.-J. Lin, Combining SVMs with various feature selec-
tion strategies, 2005. Available from http://www.csie.ntu.edu.tw/ cjlin/ pa-
pers/features.pdf.
29. U. Cherubini, E. Luciano and W. Vecchiato, Copula Methods in Finance,
John Wiley & Sons, 2004.
 References 487

http://www.elan.com.mx/biblioteca/FINANCE%20MATERIAL%
20CM/Wiley%20Copula%20Methods%20in%20Finance.pdf.
30. V. K. Chopra and W. T. Ziembia, The effect of errors in means, variances,
and co-variances on optimal portfolio choices, Worldwide Asset and Liability
Modeling (W. T. Ziembia and J. M. Mulvey editors), Cambridge University
Press, 1998.
31. N. A. Chriss, Black-Scholes and Beyond Option Pricing Models, McGraw Hill,
1997.
32. G. Cornuejols and R. Tütüncü, Optimization Methods in Finance, Cambridge
University Press, 2007.
33. J. C. Cox, S. A. Ross and M. Rubinstein, Option pricing: a simplified approach,
Journal of Financial Economics, Vol. 7(3), pp. 229-263, 1979.
34. J. C. Cox, J. E. Ingersoll and S. A. Ross, A theory of the term structure of
interest rates, Econometrica, Vol. 53, pp. 385–407, 1985.
35. R. Deaves and M. Parlar, A generalized bootstrap method
to determine the yield curve, Macmaster University, 1999.
http://www.stats.uwo.ca/events/Colloquium/W09/pdf/Rodrigo.pdf
36. X. -T. Deng, Z. -F. Li and S. -Y. Wang, A minimax portfolio selection strategy
with equilibrium, European Journal of Operational Research, Vol. 166, pp.
278–292, 2005.
37. D. Dentcheva and A. Ruszczynski, Portfolio optimization with stochastic dom-
inance constraints, Journal of Banking and Finance, Vol. 30, pp. 438–451,
2006.
38. E. Derman, I. Kani, and N. A. Chriss, Goldman Sachs–quantitative strategies
research notes, implied trinomial trees of the volatility smile, 1996.
http:// www.ederman.com/new/docs/gs-implied-trinomial-trees.pdf.
39. D. Duffie and J. Pan. An overview of Value-at-Risk, Journal of Derivatives,
Vol. 4, pp. 7–49, 1997.
40. P. Eizek and K. Komorad, Implied trinomial trees, SFB 649 Discussion Paper
2005-007,
http://edoc.hu-berlin.de/series/sfb-649-papers/2005-7/PDF/7.pdf.
41. E. J. Elton, M. J. Gruber, S. J. Brown and W. Goetzmann, Modern Portfolio
Theory Investment Analysis, 8th edition, John Wiley & Sons, 2009.
42. P. Embrechts, A.J. McNeil and D. Straumann, Correlation and dependency
in risk management: properties and pitfalls, Risk Management: Value at Risk
and Beyond ( M. Dempster and H.K. Moffatt editors), Cambridge University
Press, 2000.
43. T. W. Epps, Pricing Derivative Securities, World Scientific Publishing, 2000.
488 Financial Mathematics: An Introduction

44. C.I. Fábián, G. Mitra and D. Roman, Processing second order stochastic dom-
inance models using cutting plane representations, Mathathematical Program-
ming (Ser A), Vol. 130(1), pp. 33–57, 2011.
45. A. Fensterstock, Credit scoring and the next step, Business Credit, Vol. 107(3),
pp. 46–49, 2005.
46. D. Filipović, Interest Rate Models, http://www.mathematik.uni-
muenchen.de/∼filipo/ZINSMODELLE/zinsmodelle1.pdf
47. P. C. Fishburn, Utility Theory for Decision Making, John Wiley & Sons, 1970.
http://www.dtic.mil/cgi-bin/GetTRDoc?AD=AD0708563.
48. M. Garman and S. Kohlhagen, Foreign currency option values, Journal of
International Money and Finance, Vol. 2, pp. 231–237, 1983.
49. P. Glasserman, Monte Carlo methods in Financial Engineering, Springer,
2003.
50. P. Glasserman, P. Heidelberger and P. Shahabuddin, Variance reduction tech-
niques for estimating Value-at-Risk, Management Science, Vol. 46(10), pp.
1349–1364, 2000.
51. C. Goffman, Introduction to Real Analysis, A Harper International Edition,
1966.
52. A. J. Goldman and A. W. Tucker, Linear Equalities and Related Systems,
Priceton University Press, pp. 53–97, 1956.
53. M. Grigoriu, Stochastic Calculus: Applications in Science and Engineering,
Birkhäuser, 2002.
54. L. Güntay, N.R. Prabhala and H. Unal, Callable bonds
and hedging, Wharton Financial Institution Center,
http://fic.wharton.upenn.edu/fic/papers/02/0213.pdf
55. N. Gupta and A. Chaudhary, Optimal trading strategies, M.Tech. Disserta-
tion, Department of Mathematics, IIT Delhi, 2009.
56. J. H. Halton, On the efficiency of certain quasi-random sequences of points in
evaluating multi-dimensional integrals, Numerical Mathematics, Vol. 2, 84–90,
1960.
57. H. Han and X. Wu, A fast numerical method for the Black-Scholes equation
of American options SIAM Journal on Numerical Analysis, Vol. 41, pp. 2081–
2095, 2003.
58. D. J. Hand, Modelling consumer credit risk, IMA Journal of Management
Mathematics, Vol. 12(1), pp. 139–155, 2001.
59. L. Haris, Trading and Exchanges Market Microstructure for Practitioners, Ox-
ford University Press, 2003.
 References 489

60. M. Haugh, Continuous-time short rate models, Term structure models: IEOR
E4710, 2010. http://www.columbia.edu/∼mh2078/cts shortrate models.pdf
61. D. Heath, R. Jarrow and A. Morton, Bond pricing and the term structure of
interest rates: a new method for contingent claim valuation, Econometrica,
Vol. 60, pp. 77–105, 1992.
62. S. Heston, A closed form solution for options with stochastic volatality and
applications to bond and currency options, Review of Financial Studies, Vol.
6, pp. 327–343, 1993.
63. C. W. Hsu, C. C. Chang and C. J. Lin, A practical guide to support vector
classification, 2003.
http://www.csie.ntu.edu.tw/∼cjlin/papers/guide/guide.pdf.
64. G. Huberman and W. Stanzl, Optimal liquidity trading, Review of Finance,
Vol. 9, pp. 165-200, 2005.
65. J. C. Hull, Options, Futures and Other Derivatives, 8th edition, Prentice Hall,
2012.
66. J. Hull and A. White, Pricing interest rate derivative securities, Review of
Financial Studies, Vol. 3, pp. 573–592, 1990.
67. P. J. Hunt and J. E. Kennedy, Financial Derivatives in Theory and Practice,
John Wiley & Sons, 2004.
68. R. Jarrow and A. Rudd, Option Pricing, Irwin Professional Publishers, 1983.
69. P. Jäckel, Monte Carlo Methods in Finance, John Wiley & Sons, 2002.
70. N. Ju and Zhong, An approximate formula for pricing American options.
http://www.rhsmith.umd.edu/faculty/nju/JZ99JD.pdf
71. T. Joro and P. Na, Portfolio perference evaluation in mean-variance-skewness
framework, European Journal of Operation Research, Vol. 175, pp. 516–542,
2006.
72. O. Josef, Models of interest rate evolution-Vasicek and CIR models, Journal
of Applied Mathematics, Vol. 2(3), pp. 117–125, 2009.
73. L. Kaelbling, M. Littman and A. Moore, Reinforcement learning: a survey,
Journal of Artificial Intelligence Research, Vol. 4, pp. 237–285, 1996.
74. P. Kall, Stochastic Linear Programming, Springer, 1976.
75. I. Karatzas and S. E. Shreve, Brownian Motion and Stochastic Calculus, 2nd
edition, Springer, 1991.
76. R. Khemchandani, N. Gupta, A. Chaudhary and S. Chandra, Optimal exe-
cution with weighted impact functions: a quadratic programming approach,
Optimization Letters, 2012. DOI 10.1007/s11590-012-0441-4
77. R. Kissell and M. Glantz, Optimal trading strategies: quantitative approaches
for managing market impact and trading risk, AMACON, 2003.
490 Financial Mathematics: An Introduction

78. F. C. Klebaner, Introduction to Stochastic Calculus with Applications, 3rd edi-

tion, Imperial College Press, 2011.
79. H. Konno and K. Suzuki, A mean-variance-skewness optimization model,
Journal of Operation Research Society of Japan, Vol. 38, pp. 135–187, 1995.
80. H. Konno and H. Yamazaki, Mean- absolute deviation portfolio optimization
model and its application, Management Science, Vol. 37, pp. 519–531, 1991.
81. H. Leland, Option pricing and replication with transactions costs, Journal of
Finance, Vol. 40, pp. 1283–1301, 1985.
82. D. Leisen and M. Reimer, Binomial models for option valuation - ex-
amining and improving convergence, Discussion Paper no. B 309, 1995.
http://www.smartquant.com/references/OptionPricing/option31.pdf
83. H. Levy, Stochastic Dominance and expected utility: survey and analysis,
Management Science, Vol. 38, pp. 555–593, 1992.
84. P. Lucas, Score updates, Collections and Credit Risk, Vol. 7(10), pp. 22–25,
2002.
85. D. G. Luenberger, Investment Science, Oxford University Press, 1998.
86. J. S. Liu, Monte Carlo Strategies in Scientific Computing, Springer, 2001.
87. R. L. McDonald, Derivatives Markets, 2nd edition, Addison Wesley Pearson,
2006.
88. D. Martin, Early warning of bank failure: A logit regression approach. Journal
of Banking and Finance, Vol. 1, pp. 249-276, 1977.
89. R. Mansini, W. Ogryczak and S. Uryasev, Conditional value at risk and re-
lated linear programming model for portfolio optimization, Annals Operations
Research, Vol. 152, pp. 227–256, 2007.
90. H. M. Markowitz, Portfolio selection, Journal of Finance, Vol. 7, pp. 77–91,
1952.
91. H. Mausser and D. Rosen, Beyond VaR: from measuring risk to managing
risk, ALGO Research Quarterly, Vol. 1(2), pp. 5–20, 1998.
92. A. Mele, Lecture Notes in Financial Economics, 2011,
http://www.antoniomele.org/files/fin eco.pdf.
93. R. Merton, Option pricing when underlying stock returns are discontinuous,
Journal of Financial Economics, Vol. 3, pp. 125–144, 1976.
94. R. C. Merton, Theory of rational option pricing, Bell Journal of Economics
and Management Science, Vol. 4, pp. 141–183, 1973.
95. L. J. Mester, What’s the point of credit scoring?, 1997.
http://www.phil.frb.org/research-and-data/publications/business-
review/1997/september-october/brso97lm.pdf.
 References 491

96. C. Michel, D. Galai and R. Mark, A comparative analysis of current credit

risk models, Journal of Banking and Finance, Vol. 24, pp. 59–117, 2000.
97. T. Mikosch, Elementary Stochastic Calculus (with Finance in View), World
Scientific Publication, 1998.
98. O. L. Mangasarian. Nonlinear Programming, SIAM, Philadelphia, 1994.
99. J. P. Morgan, CreditMetrics, Technical Document, 1997.
http://www.ma.hw.ac.uk/∼mcneil/F79CR/CMTD1.pdf
100. J. Morrison, Introduction to survival analysis in business, The Journal of
Business Forecasting Methods and Systems. Vol. 23(1), pp. 18–22, 2004.
101. Y. Nevmyvaka, Y. Feng and M. Kearns, Reinforcement learning for optimized
trade execution, Proceedings of International Conference on Machine Learning,
2006.
102. H. Niederreiter, Random Number Generation and Quasi-Monte Carlo Meth-
ods, SIAM, 1998.
103. W. Ogryczak and A. Ruszczynski, From stochastic dominance to mean-risk
models: semideviations as risk measures, European Journal of Operation Re-
search, Vol. 116, pp. 33–50, 1999.
104. W. Ogryczak and A. Ruszczynski, On consistency of stochastic dominance
and mean-semideviations models, Mathematical Programming, Vol. 89, pp.
217–232, 2001.
105. J. Ohlson, Spring Financial ratios and the probabilistic prediction of
bankruptcy, Journal of Accounting Research, Vol. 18, pp. 109-131, 1980.
106. S. Park, Solving the mystery of credit scoring models, Business Credit, Vol.
106(3), pp. 43–47, 2004.
107. A. F. Perold, The implementation shortfall: paper versus reality, Journal of
Portfolio Management, Vol. 14, pp. 4–9, 1988.
108. S. Perrakisa and J. Lefoll, Option pricing and
replication with transaction costs and dividends,
http://www.smartquant.com/references/OptionPricing/option2.pdf
109. M. Pritsker, Evaluating value at risk methodologies, Journal of Financial
Services Research, Vol. 12(2/3), pp. 201–242, 1997.
110. C. P. Robert and G. Casella, Monte Carlo Statistical Methods, 2nd edition,
Springer, 2004.
111. R. T. Rockafellar and S. Uryasev, Optimization of conditional value-at-risk,
The Journal of Risk, Vol. 2, pp. 21–41, 2000.
112. R. T. Rockafellar, Coherent approaches to risk in optimization
under uncertainty: tutorials in operational research, INFORMS,
492 Financial Mathematics: An Introduction

pp. 38–61, 2007. http://www.math.washington.edu/∼rtr/papers/rtr-

RiskTutorial INFORMS2007.pdf
113. S. Roman, Introduction to the Mathematics of Finance: From Risk Manage-
ment to Options Pricing, Springer, 2004.
114. S. M. Ross, An Introduction to Mathematical Finance: Options and Other
Topics, Cambridge University Press, 1999.
115. M. Rubinstein, Implied binomial tree, Journal of Finance, Vol. 49(3), pp.
771–818, 1994.
http://www.haas.berkeley.edu/groups/finance/WP/rpf232.pdf
116. A. Ruszczy’nski and R.J. Vanderbei, Frontiers of stoschastically nondomi-
nated portfolios, Econometrika, Vol. 71, pp. 1287–1297, 2003.
117. P. A. Samuelson, Proof that properly anticipated prices fluctuate randomly,
Industrial Management Review, Vol. 6, pp. 41 -50, 1965.
118. A. L. Sandler S. E. McGinn and J. L. F. Barloon, Lending scrutiny of credit
score-based underwriting systems, ABA Bank Compliance (March/April) Vol.
38, 2000.
119. W. F. Sharpe, Capital asset prices - a theory of market equilibrium under
conditions of risk. Journal of Finance Vol. 13(3), pp. 425-442, 1964.
120. W. F. Sharpe, G. J. Alexander and J. V. Bailey, Investments, Prentice Hall
International, 1998.
121. S. E. Shreve, Stochastic Calculus for Finance I: The Binomial Asset Pricing
Model, Springer, 2004.
122. S. E. Shreve, Stochastic Calculus for Finance II: Continuous Time Models,
Springer, 2004.
123. I. M. Sobol, On the distribution of points in a cube and the approx imate
evaluation of integrals, USSR Computational Mathematics and Mathematical
Physics, Vol. 7, pp. 86–112, 1967.
124. M. S. Sodhi, LP modelling for asset liability management: a survey of choices
and simplifications, Operations Research, Vol. 53, pp. 181–196, 2005.
125. H. M. Soner, S. E. Shreve and J. Cvitanic, There is no nontrivial portfolio
for option pricing with transaction costs, The Annals of Applied Probability,
Vol. 5, pp. 327-355, 1995.
126. J. B. Sousa, M. L. Esquivelb and R. M. Gaspar, Machine
learning Vasicek model calibration with Gaussian processes,
http://www.dm.fct.unl.pt/sites/www.dm.fct.unl.pt/files/preprints/2010/
4 10.pdf
127. P. Söderlind, Lecture Notes in Finance 2 (MiQE/F, MSc course at UNISG),
 References 493

2011. http://home.datacomm.ch/paulsoderlind/Courses/OldCourses/
Fin2MiQEFAll.pdf
128. S. Sriboonchitta, W. K. Wong, S. Dhompongsa and H. T. Nguyen, Stochastic
Dominance and Applications to Finance, Risk and Economics, CRC Press,
2009.
129. M. C. Steinbach, Markowitz revisted: mean variance models in financial port-
folio analysis SIAM Review, Vol. 43(1), pp. 31–85, 2001.
130. R. E. Steuer, Y. Qi and M. Hirschberger, Portfolio optimization: new ca-
pabilities and future methods, Zeitschrift fr Betriebswirtschaft, Vol. 76, pp.
199–219, 2006.
131. W. Sun, A. C. Fan, L. -W. Chen, T. Schouwenaars and M. A. Albota, Op-
timal Rebalancing Strategy Using Dynamic Programming for Institutional
Portfolios, Available at SSRN 2004.
132. R. Sutton and A. Barto, Reinforcement Learning: An Introduction, MIT
Press, Cambridge, 1998.
133. California Debt and Investment Advisory Commission, Understanding inter-
est rate swap math pricing, 2007.
http://www.treasurer.ca.gov/cdiac/publications/math.pdf
134. L. C. Thomas, A survey of credit and behavioral scoring forecasting financial
risk of lending to consumers, International Journal of Forecasting, Vol. 16, pp.
163–167, 2000.
135. L. C Thomas, D. Edelman and J. N. Crook, Credit Scoring and its Appli-
cation, SIAM Monograph on Mathematical Modelling and its Computation,
2002.
136. Y. Tian, A modified lattice approach to option pricing, Journal of Futures
Markets, Vol. 13(5), pp. 563–577, 1993.
137. H. Timmons, The cracks in credit scoring: as loan defaults rise, worries grow
about how creditors pick borrowers, Business Week (November) 25, 2002.
138. R. S. Tsay, Analysis of Financial Time Series, John Wiley & Sons, 2005.
139. R. H. Tütüncü and M. Koenig, Robust asset allocation, Annals of Operation
Research, Vol. 132, pp. 157–87, 2004.
140. S. Uryasev and R.T. Rockafellar, Optimization of conditional Value-at-Risk,
Research Report 99-4, ISE Dept., University of Florida, 1999.
141. V. N. Vapnik, The Nature of Statistical Learning Theory, Springer, 1995.
142. J. R. Varma, Bond valuation and the pricing of interest rate options in India,
The ICFAI Journal of Applied Finance, Vol. 2(2), pp. 161–176, 1996.
143. O. A. Vasicek, An equilibrium characterization of the term structure, Journal
of Financial Economics, Vol. 15, pp. 177–188, 1977.
494 Financial Mathematics: An Introduction

144. P. -C. G. Vassiliou, Applied Stochastic Finance, Vol. 2, John Wiley & Sons,
2012.
145. F. de Weert, Exotic Options Trading, John Wiley & Sons, 2008.
146. G. West, Interest Rate Derivatives: Lecture Notes, 2010.
http://www.finmod.co.za/ird.pdf.
147. P. Wilmott, Introduces Quantitative Finance, John Wiley & Sons, 2007.
148. L. Y. Yu, X. D. Ji and S. Y. Wang, Stochastic programming models in
financial optimization: a survey, Advance Modeling and Optimization, Vol. 5,
pp. 1-26, 2003.
149. R. Zagst, Interest-Rate Management, Springer, 2002.
150. X. Y. Zhou and D. Li, Continuous time mean- variance selection: a stochastic
LQ framework, Applied Mathematics and Optimization, Vol. 42, pp. 19–33,
2000.
Index

above par, 16 bond price, 335

acceptance-rejection sampling, 452 bond yield, 339
adapted process, 254 Borel σ-field, 218, 249
admissible portfolio, 7 Borel sets, 218, 249
affine yield model, 366 bounded variation, 278
American call/put options, 50 Box-Muller method, 451
American option, 12 Brownian bridge, 233
annuity, 21 Brownian motion, 229
arrival price, 381 Brownian motion with drift and
asset weights, 131 volatility, 233
at par, 16 bull spread, 59

bear spread, 59 capital asset pricing model, 129, 149

below par, 16 capital gain, 17
Bernoulli process, 223 capital market line, 150, 153
Beta distribution, 452 CAPM as a pricing formula, 158
beta of asset, 156 CAPM as factor model, 159
beta of portfolio, 157 CIR interest rate model, 359
binomial lattice for term structure, coherent risk measures, 169, 204
353 comparison principle, 9
binomial lattice model, 61 complete market models, 93
binomial model, 239 conditional expectation, 243, 245
binomial process, 223 conditional value-at-risk, 193
Black-Scholes formula, 113, 215 cone, 204
Black-Scholes formula for dividend contingent claims, 10
paying stock, 114 continuous compounding, 18
Black-Scholes PDE, 327 continuous time martingale, 267
bond, 334 contrarian traders, 380
bond options, 355 convex cone, 204
496 Financial Mathematics: An Introduction

copula, 211 effective rate, 20

counterparty risk, 411 efficient frontier, 145
coupon bonds, 16 equities, 16
coupon rate, 336 equivalent measure, 314
covariance stationary, 221 estimation of VaR, 192
credit exposure, 414 European call option, 49
credit premium, 413 European call/European put option,
credit risk, 16, 411 12
credit scoring, 411 European put option, 50
credit scoring model, 436 exchange rate, 215
credit spread, 413 execution cost, 374
credit spread risk, 411 exercise price, 12
Credit-VaR, 419 exotic options, 96
CreditMetrics, 416 expected return of the portfolio, 3
CRR model, 100 expected shortfall, 193
crude Monte Carlo method, 458 expected tail loss, 193
cumulative distribution function, 218 expiration date, 12
exponential distribution, 227, 240,
449

default probability, 414

default risk, 16, 411 F -events, 217
delivery date, 11, 29 F -measurable sets, 217
delta hedging, 121 face value, 16
delta-gamma hedging, 122 feasible region, 175
derivative securities, 10 Feynman-Kac theorem, 326
diffusion, 298 filtration, 243, 244
discount factor, 19 filtration in a continuous time, 253
discounted cash flow, 413 filtration in a discrete time, 253
discounted wealth process, 266 finite dimensional distribution, 221
discrete time martingale, 259 first fundamental theorem of asset
distance-to-default, 426 pricing, 75
dividend, 17 first order stochastic dominance, 205
divisibility, 5 first variation, 278
Doob’s martingale, 275 fixed income securities, 16
downgrade risk, 411 forward contract, 10, 29
downside or safety-first risk measures, forward price, 11, 29
168 forward price formula, 31
 Index 497

forward price formula with carrying instantaneous short rate, 348

costs, 33 interest rate swap, 350
forward price formula with dividend, inversion method, 449, 450
33 investable assets, 183
forward rate, 346 Ito process, 300
forward rate agreement (FRA), 347 Ito’s calculus, 277
forward rate curve, 348 Ito-Doeblin formula, 303, 305
futures contract, 39
futures pricing, 42
L1 -risk measure, 170
L1 -risk models, 169
gambler’s ruin problem, 243 L2 -risk models, 169
Gaussian process, 224, 231, 238 L∞ -risk measure, 175
geometric Brownian motion, 234, 303 L∞ -risk models, 169
geometric distribution, 457 Langevin equation, 305
Girsanov theorem, 318 law of one price, 9
Goldman-Tucker theorem, 76 Lehmner generator, 447
good-till-cancelled, 380 Levy process, 238
Greeks, 118 LIBID, 349
growth factor, 17 LIBOR, 349
linear congruence generator, 447
linear discriminant analysis, 437
heat equation, 329 Lipschitz condition, 296
hedging problem, 66 liquidity, 6
Heisenberg uncertainty principle, 374 logarithmic return, 104
historical volatility, 105 logistic regression, 437
hitting time, 268 lognormal distribution, 235, 304
long call, 51
long forward contract/position, 11, 29
implementation shortfall, 382 long position on stock, 6
implied interest rate, 23 long put, 51
implied volatility, 125 loss given default, 414
importance sampling, 452, 455 lower Riemann sum, 286
independent increment, 222, 226
independent process, 222
inflation indexed bond, 16 margin account, 29, 40
inflation risk, 16 market impact, 374
initial value problem, 296 market interest rate, 23
498 Financial Mathematics: An Introduction

market order, 375 nominal interest rate, 16

market portfolio, 150 nominal value, 16
market price, 40 non-dominated point, 145
market price of risk, 321 numeraire, 81
marking to market, 29, 39, 40
Markov process, 223, 224, 227, 231,
234, 238
one fund theorem, 151
Markov property, 223, 224, 227, 231,
opportunity cost, 374
234
optimal trading strategies, 376
Markov times, 268
option pricing, 50
Markowitz bullet, 142
optional sampling theorem, 268, 270
Markowitz curve, 141
optional times, 268
Markowitz efficient frontier, 145
options on dividend paying stock, 89
martingale, 243, 244
order statistic, 192
maturity, 16
mean of the portfolio, 132 Ornstein-Uhlenbeck process, 238, 286,
mean reverting, 357 305
measurable space, 217
MIBOR, 349
minimum variance point, 143 p-variation, 278
minimum variance set, 143 par value, 16
mixed congruential generator, 462 parameter space, 220
modified duration, 338 pay-off curve of the option, 59
moment based risk measures, 168 PDE for bond price, 363
momentum trading strategies, 380 periodic compounding, 17
money market, 22 permanent market impact, 376
money market account, 16 perpetuity, 22
Monte Carlo methods/simulations, Picard’s method, 296
445 Poisson process, 227, 238
Monte Carlo simulation of European portfolio, 3, 131
call option, 459 portfolio optimization problem, 129,
multi asset portfolio optimization, 139 132
multi period binomial lattice model, portfolio risk, 129
68 price appreciation, 374
price formula with continuous
dividend, 34
natural filtration, 253 pricing American options, 84
no arbitrage principle, 8 pricing European option, 109
 Index 499

primal-dual linear programming pair, risk of an asset, 130

75 risk seeking stochastic dominance,
principal, 16 207
probability measure, 217 robust optimization techniques, 168
prop trading, 377
pseudo-random numbers, 446
pure discount bonds, 16 (σ, µ)-graph, 132
put option, 10 σ-field, 216, 248
put-call parity, 52 σ-field generated by the event, 216
σ-field generated by two proper
subsets, 216
quadratic variation, 280, 283 σ-measurable random variable, 256
quantile based risk measures, 168 sample path, 220
quantitative concept of risk, 203 sample space, 216
quasi-random numbers, 446, 447 scaling property, 232
second fundamental theorem of asset
pricing, 75, 323
Radon-Nikodym derivative, 314 second order stationary, 221
Radon-Nikodym derivative process, security market line, 157
318 self financing, 93
Radon-Nikodym theorem, 314 short call, 51
random differential equation, 297 short forward contract/position, 11,
random number seed, 446 29
random process, 219 short position on stock, 6
random variable, 218 short put, 51
rejection sampling, 452 short selling, 6
replicating portfolio, 93 simple interest rate, 17
repo, 351 simple random walk, 103
return of an asset, 2, 129 single period model, 2
return of the portfolio, 3 solvency, 7
Riemann integral, 285 spot price, 40
Riemann-Stieltjes integral, 286 spot rate, 340
risk attitudes, 200 spread, 59
risk neutral pricing formula, 66, 324 standard deviation, 3
risk neutral probability, 65 state space, 220
risk neutral probability martingale, 81 stationary increment, 222, 226
risk neutral probability measure, 75, stochastic calculus, 277
109, 264, 315 stochastic differential equation, 5, 297
500 Financial Mathematics: An Introduction

stochastic dominance, 205 trailing stop order, 380

stochastic integral equation, 297 trivial σ-field, 216
stochastic orders, 206 two assets portfolio optimization, 133
stochastic process, 215, 219 two fund theorem, 147
stockholder, 17
stop orders, 379
stop-loss-order, 207 unbounded variation, 283
stopping time, 268 underlying, 10
stopping times, 268 unit bonds, 23
stratified sampling method, 458 upper Riemann sum, 285
strict sense stationary stochastic
process, 221
strong solution, 298 value of a forward contract, 37
strong stationary stochastic process, value of the option at expiration, 51
221 value-at-risk, 187
subadditive risk measure, 191 variance of the portfolio, 132
submartingale, 259 variance reduction technique, 456
supermartingale, 259 Vasicek model, 356
support vector machine, 437
swap, 29
symmetric property, 232
weak solution, 298
symmetric random walk, 225
weak stationary, 221
wealth process, 266
tail event occurs, 194 wide sense stationary process, 223
tail VaR, 193 Wiener process, 229, 238
temporary market impact, 375
term structure of interest, 339
time inversion property, 232 yield curve, 341
time series, 220 yield to maturity, 339
time value of money, 16
timing risk, 374
total σ-field, 216 zero coupon bonds, 16, 334