Discrete Time Modelling and Derivative Securities

Version 221106
Contents

1 Single-step asset pricing models 4

1.1 Single-step binomial tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Option pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 General derivative securities . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4 Two underlying securities . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.5 The trinomial model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.6 General single-step model . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
1.7 Auxiliary lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

2 Multi-step binomial model 48

2.1 Two-step binomial tree example . . . . . . . . . . . . . . . . . . . . . . . . 48
2.2 Partitions and information . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.3 Martingale properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.4 The Cox–Ross–Rubinstein model . . . . . . . . . . . . . . . . . . . . . . . 65
2.5 Delta hedging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

3 Multi-step general models 74

3.1 Partitions and conditioning . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.2 Properties of conditional expectation . . . . . . . . . . . . . . . . . . . . . 78
3.3 Filtrations and martingales . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.4 Trading strategies and arbitrage . . . . . . . . . . . . . . . . . . . . . . . . 85
3.5 General multi-step model . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.6 The Fundamental Theorems of Asset Pricing . . . . . . . . . . . . . . . . . 99
3.7 Forwards and futures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3.8 Complex derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

4 American options 113

4.1 Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.2 Stopping times and optimal exercise . . . . . . . . . . . . . . . . . . . . . . 120
4.3 Hedging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.4 The Doob decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Introduction
These lecture notes cover the material of the Discrete Time Modelling and Derivative Se-
curities module, one of the three components of the Certificate stage of the on-line MSc in
Mathematical Finance, offered by the University of York.

View the introductory video.

To access the video links you need to first log into https://vle.york.ac.uk/webapps/login/
with your University of York username and password.

The module is divided into four parts:

ˆ Part 1:
Lecture notes: Chapter 1, Exercises 1, Coursework Assignment 1

ˆ Part 2:
Lecture notes: Chapter 2, Exercises 2, Coursework Assignment 2

ˆ Part 3:
Lecture notes: Chapter 3, Exercises 3, Coursework Assignment 3

ˆ Part 4:
Lecture notes: Chapter 4, Exercises 4, Coursework Assignment 4

The text is accompanied by videos and Excel files, which explain various ideas, examples and
arguments. The videos can be accessed by links like the one above.

The following textbooks are recommended as auxiliary sources:

ˆ M.Capinski and E.Kopp, Discrete Models of Financial Markets, Cambridge University

Press 2012

ˆ M.Capinski and T.Zastawniak, Mathematics for Finance, 2nd ed., Springer 2011.

3
Chapter 1

Single-step asset pricing models

In this chapter we explore the simplest option pricing models. We assume that there is a
single trading period, from the present date (time 0) to a fixed future time T . We exam-
ine mathematical models for the behaviour of the prices of one or more underlying stocks,
beginning with the simplest case of a single stock whose price at time T takes one of just
two possible values. This analysis is extended to a model with two stocks, models with three
possible prices for each stock, and finally to a general pricing model with d > 1 stocks, each
with m possible outcomes for its price at T.
The principal economic requirement, arising from the assumption that the financial mar-
kets being modelled are efficient is that, provided all market participants share the same
information about the random evolution of the stock price, an investor in this market should
not be able to make a profit without incurring some level of risk. This assumption is cast in
mathematical terms and underlies the methods we develop to establish how to determine the
value of financial instruments whose price depends only on the price of the underlying stock
at time T : the so-called European derivative securities, whose study is the principal topic of
this and the next two chapters.

1.1 Single-step binomial tree

The most important feature of financial markets is the uncertainty arising from the fact that
the future prices of various financial assets cannot be predicted. We begin by examining
the simplest possible setting and construct a ‘toy model’ whose two dimensions, time and
uncertainty, are illustrated by means of the simplest possible setup.
We take time as discrete and reduced to just two time instants: the present time 0 and
future time T . To simplify the notation and emphasise the fact that we are now dealing with
just one step we shall write 1 instead of T when we have in mind a single step of length T.
The uncertainty is reflected by the number of possible scenarios, which in this section is
minimalist, namely, just two: ‘up’ or ‘down’. This deceptively simple model reflects many of
the key features of much more general pricing models, and we examine it in detail.

4
Assets
Assume we are given some asset, which is usually thought of as a stock and is customarily
referred to, in defiance of proper grammatical usage, as the underlying since its price will
determine the prices of the securities we wish to study.
The current price S(0) > 0 of the underlying is known, while its future price S(1) is not,
but we can consider all possible future prices and the probabilities of attaining them. This
is performed formally by first choosing a non-empty set Ω, called the sample space. The
elements of this set are related to outcomes of some random experiment (either specific,
tossing a coin, or quite vague, the future state of the economy) and we call them scenarios
or states. In these lecture notes Ω will always be a finite set.
Just for the current section we take a two-element set Ω = {u, d} and we let the future
price of our asset be a function

S(1) : Ω → (0, +∞).

In general, a function defined on Ω is called a random variable. Since S(1) takes just two
values, so does the return
S(1) − S(0)
KS = ,
S(0)
which determines the price
S(1) = S(0)(1 + KS ).
We shall define S(1) by first deciding the shape of KS . Assume that

KS (u) = U,
KS (d) = D,

where
−1 < D < U
to avoid negative prices. Then

S(1, u) = S(0)(1 + U ),
S(1, d) = S(0)(1 + D).

It is convenient to introduce the following notational convention: S u = S(1, u), S d = S(1, d),
which gives the random variable of stock prices as
 
Su S(0)(1 + U )
S(1) = = .
Sd S(0)(1 + D)

The probabilities are given by prescribing a number p from the interval (0, 1) so that

P ({u}) = p,
P ({d}) = 1 − p.

5
Here we are careful with notation since P assigns numbers to subsets of Ω, but the following
simplified notation is quite useful: P (ω) = P ({ω}) for any ω ∈ Ω. Furthermore, the proba-
bility P is assumed to be zero for the empty set and 1 for the whole of Ω, and so it is defined
for all subsets of Ω, called events, satisfying additionally the so-called additivity property:

P (A ∪ B) = P (A) + P (B)

for any A, B ⊂ Ω such that A ∩ B = ∅. (The choice of such A, B in our case is rather
limited!)
We also assume that there is an accompanying non-random asset, called the money
market account, with deterministic prices A(0), A(1), where

A(1) = A(0)(1 + R)

for some R > 0.

View a video illustrating this setting.

Investment
Suppose we have at our disposal a certain sum of money, which we invest by purchasing a
portfolio of x units of the above risky asset and y units of the money market account. We
build a portfolio (x, y) with initial value given by

V(x,y) (0) = xS(0) + yA(0).

The value of this portfolio at time 1 is

V(x,y) (1) = xS(1) + yA(1),

that is, 
xS(0)(1 + U ) + yA(0)(1 + R)
V(x,y) (1) = .
xS(0)(1 + D) + yA(0)(1 + R)
We assume that the market is frictionless: we do not impose any restrictions on the
numbers x, y, so that unlimited short-selling is allowed: at time 0 we can borrow a share,
sell it, and purchase some other asset, while at time 1 we buy the share back to return it
to the owner. The assets are assumed to be arbitrarily divisible, meaning that x, y can
take arbitrary real values. We do not impose any bounds on x, y, thus assuming unlimited
liquidity in the market. Finally, there are no transaction costs involved in trading, that is, the
same stock price applies to long (buy: x > 0) and short (sell: x < 0) positions. Moreover,
risk-free investment (y > 0} and borrowing (y < 0) both use the same rate of return R.

Watch a video on the notation for portfolios.

6
 Our most important modelling assumption is the No Arbitrage Principle (NAP), which
asserts that trading cannot yield riskless profits. In general, the form of this assumption
depends crucially on the class of trading strategies that we allow in the market. In the
present context, with no intervening trading dates, this is trivial, since a ‘strategy’ consists
of a single portfolio chosen at time 0.
Definition 1.1
A portfolio (x, y), chosen at time 0, is called an arbitrage opportunity if

(1) V(x,y) (0) = 0,

(2) V(x,y) (1) ≥ 0 with probability 1,

(3) V(x,y) (1) > 0 with positive probability.

For our choice of Ω, condition (3) reduces to saying that we have a strict inequality for
at least one of the two possible scenarios.
Assumption 1.2 (No Arbitrage Principle)

Arbitrage opportunities do not exist.

This has an immediate consequence for the relationship between the returns on the risky
and riskless assets:
Theorem 1.3
The No Arbitrage Principle implies that

D < R < U.

Proof

See video with a numerical example of the proof.

If R ≤ D, that is,
S(0)(1 + R) ≤ S d < S u ,
we take x = 1 and y = −S(0)/A(0). At time 1

V(x,y) (1) = S(1) − S(0)(1 + R) ≥ 0,

which is strictly positive in the ‘up’ state, that is, with probability p > 0.
If R ≥ U , that is,
S d < S u ≤ S(0)(1 + R),

7
we take the opposite position: x = −1 and y = S(0)/A(0). Then

V(x,y) (1) = −S(1) + S(0)(1 + R) ≥ 0,

and this is strictly positive with probability 1 − p > 0.

In each case we have constructed a portfolio that leads to arbitrage. Hence neither
inequality can hold and the theorem is proved.

The converse implication is also true.

Theorem 1.4
The condition
D<R<U
implies that an arbitrage opportunity does not exist.

Proof
Suppose (x, y) is a portfolio with zero initial value. By the definition of portfolio value

V(x,y) (0) = xS(0) + yA(0) = 0,

so
S(0)
y = −x .
A(0)
We compute the terminal value with y as determined above:

V(x,y) (1) = xS(1) + yA(0)(1 + R)

= x(S(1) − S(0)(1 + R)),

thus, for any positive x,


x[S u − S(0)(1 + R)] = xS(0)(U − R) > 0
V(x,y) (1) =
x[S d − S(0)(1 + R)] = xS(0)(D − R) < 0

under the given assumption that D < R < U. If x < 0, the above signs are reversed
and they remain opposites. So (x, y) cannot be an arbitrage opportunity for any choice
of x.

Definition 1.5
Suppose that an asset X has value X(1) at time 1. Its discounted value (discounted

8
 to time 0) is given by
X̃(1) = X(1)(1 + R)−1 .
Moreover, for the price X(0) at time 0 we define the discounted value simply as

X̃(0) = X(0).

For the stock S we have

S̃(1) = S(1)(1 + R)−1 ,

S̃(0) = S(0)

and define the discounted stock price process S̃ = {S̃(0), S̃(1)}. Clearly, Ã(1) = Ã(0) =
A(0). Similarly, the discounted value process becomes

Ṽ(x,y) (1) = V(x,y) (1)(1 + R)−1 = xS̃(1) + yA(0),

Ṽ(x,y) (0) = V(x,y) (0) = xS(0) + yA(0),

yielding

Ṽ(x,y) (1) − Ṽ(x,y) (0) = xS̃(1) + yA(0) − (xS(0) + yA(0))

= x[S̃(1) − S̃(0)],

so the change of the discounted portfolio value results exclusively from the change of the
discounted value of the risky asset.

1.2 Option pricing

A European call option provides profit for the holder of a long position if the price of the
underlying at time 1 is above a predetermined level K, the strike price or exercise price,
but excludes the possibility of losses. We define the payoff of a call as

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

where we use the notation f + (ω) = max{0, f (ω)} for any function f on Ω and ω ∈ Ω. (The
reader should take care not to confuse the return KS on the stock S with the strike price K
of the option.) The effect of holding a call option is the right (without an obligation) to buy
the underlying at the expiry (or exercise) time 1 for a price that cannot exceed K since the
option will pay the difference if the asset is more expensive. If the underlying can be bought
in the market for less than K at time 1, the option is simply not exercised.

9
 The right to sell the asset for at least K at the expiry time 1 is called a European put
option and its payoff is of the form

K − S(1) if S(1) < K
P (1) =
0 otherwise
= (K − S(1))+ .
To avoid trivial cases, we assume that the strike price K satisfies
S(0)(1 + D) < K < S(0)(1 + U ).
Then it is clear from the No Arbitrage Principle that options cannot be issued for free at
time 0 since the payoffs are always non-negative and have positive probability of being strictly
positive.
Finding ‘rational’ (that is, arbitrage free) prices C(0) (resp. P (0)) for a call (resp. put)
option in a given market model is one of the major tasks of mathematical finance. The initial
price C(0) (resp. P (0)) is called the option premium.
Example 1.6

Let U = 0.3, D = −0.2, and R = 0.2. Moreover, let S(0) = 100 and A(0) = 100.
We consider a put and a call option with expiry time 1 and strike price K = 110. Since
S u = 130 and S u = 80, the call payoff will be
 u 
S − K in the u scenario 20 in the u scenario
C(1) = = ,
0 in the d scenario 0 in the d scenario

and the put payoff will be

 
0 in the u scenario 0 in the u scenario
P (1) = d = .
K − S in the d scenario 30 in the d scenario

We now look for a portfolio (x, y) (consisting of the underlying asset and a position in
the money market account) with the same value at time 1 as the option,
V(x,y) (1) = C(1).
It is called a replicating portfolio. This gives a pair of simultaneous equations
xS(0)(1 + U ) + yA(0)(1 + R) = S(0)(1 + U ) − K,
xS(0)(1 + D) + yA(0)(1 + R) = 0,
which are easily solved:
S(0)(1 + U ) − K
x= ,
S(0)(U − D)
(1 + D)(S(0)(1 + U ) − K)
y=− .
A(0)(U − D)(1 + R)
10
 Example 1.7

Given the data from Example 1.6, let us compute the replicating portfolio (x, y) for
the call option. The above simultaneous equations become

130x + 120y = 20,

80x + 120y = 0.

Solving this system, we obtain the positions

2 4
x= and y = −
5 15
in the replicating portfolio.
We can verify that the final value of the portfolio does indeed match the call
payoff in both scenarios:
2 4
V( 2 ,− 4 ) (1) = × S(1) − × A(1)
5 15 5 15
2 4

5
× 130 − 15 × 120 in the u scenario
= 2 4
× 80 − 15 × 120 in the d scenario
 5
20 in the u scenario
= .
0 in the d scenario

Video

We claim that the price of the option should be the initial value of the replicating portfolio.
This is based on the intuition that two portfolios with identical values at time 1 should have
the same values at time 0. The justification involves the No Arbitrage Principle: in the case
of different initial values, at time 0 we buy the cheaper asset and sell the more expensive
one, and at time 1 we keep the profit. Doing this requires trading in options, which leads us
to extend the notion of a portfolio in the market extended by the option.
We extend the notion of a portfolio by including the position in options: the extended
portfolio is now a triple (x, y, z), where x represents the number of units of stock S, y the
risk-free position (the money market account A), and z the number of units of the call option
C held (z can be negative, which corresponds to taking a short position by writing and selling
z options).
The definition of the portfolio value is also extended in a natural way:

V(x,y,z) (0) = xS(0) + yA(0) + zC(0),

V(x,y,z) (1) = xS(1) + yA(1) + zC(1).

11
Bear in mind that the second equation asserts equality of random variables! This allows us
to formulate our key assumption in a similar way as before.
Definition 1.8
We say that (x, y, z) is an (extended) arbitrage opportunity if

(1) V(x,y,z) (0) = 0,

(2) V(x,y,z) (1) ≥ 0 with probability 1,

(3) V(x,y,z) (1) > 0 with positive probability.

Assumption 1.9 (No Arbitrage Principle)

Extended arbitrage opportunities do not exist.

Later on we shall always assume that the notion of a portfolio is similarly extended to
cover all securities traded in the marker.
Theorem 1.10
The No Arbitrage Principle implies that the price of a European call with strike price K
is the value at time 0 of the replicating portfolio (x, y), that is,

C(0) = xS(0) + yA(0).

Proof
Suppose that C(0) > xS(0) + yA(0). Then, if we sell the expensive asset (option)
and buy the cheap one (portfolio (x, y)), we obtain a positive balance C(0) − (xS(0) +
yA(0)) at time 0 To construct an arbitrage opportunity we invest this amount in the
money market account in addition to yA(0) already held, so our extended portfolio is
1
(x0 , y 0 , z 0 ) = (x, (C(0) − xS(0)) , −1)
A(0)

and its current value is zero,

V(x0 ,y0 ,z0 ) (0) = x0 S(0) + y 0 A(0) + z 0 C(0)

1
= xS(0) + (C(0) − xS(0)) A(0) − C(0)
A(0)
= 0.

12
At time 1, in each scenario the value of this portfolio is

V(x0 ,y0 ,z0 ) (1) = x0 S(1) + y 0 A(1) + z 0 C(1)

= xS(1) + (C(0) − xS(0))(1 + R) − C(1)
= (C(0) − (xS(0) + yA(0)))(1 + R)

since by replication C(1) = xS(1) + yA(0)(1 + R). But C(0) − (xS(0) + yA(0)) > 0
by our assumption, so we have a strictly positive outcome in each scenario at zero
initial cost, more than is needed to contradict the No Arbitrage Principle.
On the other hand, if C(0) < xS(0) + yA(0), we can take the opposite position
to construct an arbitrage opportunity.

Example 1.11
With the data given in Example 1.6 for the market model and a call option, we com-
puted the replicating portfolio in Example 1.7 to be
 
2 4
(x, y) = ,− .
5 15

Then, according to Theorem 1.10, we can find the time 0 call price as
2 4 40
C(0) = xS(0) + yA(0) = × 100 − × 100 = ≈ 13.33.
5 15 3

Video

Remark 1.12
The fact that the probabilities of the stock movements do not appear in the pricing
formula is a paradox. One might expect that a call option where the probability of
the up movement of the stock is high should cost more as it is more likely to be
exercised, but this is not the case. This surprising effect arises from the fact that the
No Arbitrage Principle and replication do not directly involve the probabilities of up or
down movements of the stock.

Example 1.13
An example of placing bets is quite illuminating. To determine the pay-out ratios a
bookmaker follows the proportion of bets placed, and not the ‘true’ probabilities based
on his own perceptions of likelihoods. So if 1000 bets are on team A and 2000 on

13
 team B, then the ratios are 3/1 for A and 1.5/1 for B. The bookmaker is safe and
cannot lose money.
On the other hand, in casinos the pay-outs for roulette depend on the probabilities
being 34/1 since there are 34 equally likely fields. The house here can go bankrupt.
In reality, in both cases the ratios are slightly unfavourable to the players, so the
bookmaker has a steady guaranteed income and the casino has an income in the long
run.

1.3 General derivative securities

A derivative security, also called a contingent claim, is a security whose payoff depends
on the prices of some underlying asset S (henceforth simply called the underlying), which
can be a stock, an index, an exchange rate, an interest rate or any other quantity that can
be traded or otherwise observed in the market. Within this chapter, a derivative security is
represented by a random payoff of the form

H(1) = h(S(1))

for some payoff function h : (0, ∞) → R. (In fact, any random variable X : Ω → R can be
written in such form.) We simply write h(S u ) = X(u) and h(S d ) = X(d), and extend h to
the whole of R arbitrarily. (Here we make use of the fact that S u 6= S d .)
Example 1.14
For example, the call option has the payoff function of the form

h(x) = max{x − K, 0} = (x − K)+ ,

where a+ = max{a, 0}.

For a put option, the payoff function is

h(x) = max{K − x, 0} = (K − x)+ .

For a straddle we have

h(x) = |x − K|.

Replication of H(1) gives the system of linear equations

xS u + yA(0)(1 + R) = h(S u ),
xS d + yA(0)(1 + R) = h(S d ).

14
Solving this linear system for x and y, we get

h(S u ) − h(S d )
x= ,
Su − Sd
h(S ) u u d
h(S u ) − xS u S u −S d
(S u − S d ) − h(SS u)−h(S
−S d
) u
S
y= =
A(0)(1 + R) A(0)(1 + R)
1 h(S )S − h(S u )S d
d u
= .
A(0)(1 + R) Su − Sd

Definition 1.15
The delta ∆H of the derivative security H is the underlying asset position in the
replicating portfolio, given by

h(S u ) − h(S d )
∆H = .
Su − Sd

The formulae for the replicating portfolio can be written as

h(S u ) − h(S d )
x= ,
S(0)(U − D)
(1 + U )h(S d ) − (1 + D)h(S u )
y= .
A(0)(U − D)(1 + R)

We claim that the derivative price H(0) is the present value of this portfolio, namely H(0) =
V(xH ,yH ) (0) = xH S(0) + yH A(0), which becomes

h(S u ) − h(S d ) (1 + U )h(S d ) − (1 + D)h(S u )

H(0) = + . (1.1)
U −D (U − D)(1 + R)

The proof of the following theorem is exactly the same as in the case of the call options
above. We need only replace C(1) by H(1) = h(S(1)). The details are left to the reader.
Theorem 1.16
The No Arbitrage Principle implies that the price of the security with payoff H(1) =
h(S(1)) has the form (1.1).

Example 1.17
Considering the same data and the same call option as in Examples 1.6 and 1.7, we

15
 can now use formula (1.1) to compute the time 0 call price once again. With

h(x) = max{x − K, 0} = max{x − 110, 0}

being the payoff function for the call, we get

h(S u ) − h(S d ) (1 + U )h(S d ) − (1 + D)h(S u )

C(0) = +
U −D (U − D)(1 + R)
h(130) − h(80) 1.3h(80) − 0.8h(130)
= +
0.5 0.5 × 1.2
20 − 0 1.3 × 0 − 0.8 × 20
= +
0.5 0.5 × 1.2
≈ 13.33,

which is of course consistent with C(0) computed in Example 1.11.

Now, we can write the right-hand side of (1.1) as

 
u 1+D 1
H(0) = h(S ) − +
(U − D)(1 + R) U − D
 
d 1+U 1
+ h(S ) −
(U − D)(1 + R) U − D
 
u R−D −1 R−D
= h(S ) d
(1 + R) + h(S ) 1 − (1 + R)−1
U −D U −D
U −R R−D
since U −D
=1− U −D
. Putting
R−D
q= ,
U −D
we have
H(0) = (1 + R)−1 (qh(S u ) + (1 − q)h(S d )). (1.2)
Since D < R < U , we have 0 < R − D < U − D, so that 0 < q < 1 and we can
interpret q as the probability of an ‘up’ move and 1 − q as the probability of a ‘down’ move.
This defines a probability Q on Ω = {u, d} with values q and 1 − q at u and d. Moreover,
observe that
 
−1 R−D U −R
S(0) = (1 + R) S(0)(1 + U ) + S(0)(1 + D)
U −D U −D
= (1 + R)−1 (qS u + (1 − q)S u )

Conversely, if S(0) = (1 + R)−1 (qS u + (1 − q)S u ), then q = R−D

U −D
. This gives rise to the
following definition.

16
 Definition 1.18
Let Q be a probability on Ω = {u, d} with values q = Q(u) and 1 − q = Q(d). We
say that Q us a risk-neutral probability on Ω = {u, d} if 0 < q < 1 and

S(0) = (1 + R)−1 qS u + (1 − q)S d .

(1.3)

For any random variable X on Ω = {u, d},

EQ (X) = qX(u) + (1 − q)X(d)
is called the risk-neutral expectation of X. Observe that (1.3) can be written as
S(0) = (1 + R)−1 EQ (S(1)).
Moreover, the argument leading to (1.2) proves the following result.
Theorem 1.19
The price of a derivative security is the expected value of the discounted payoff with
respect to the risk-neutral probability:

H(0) = EQ (1 + R)−1 h(S(1)) .

Example 1.20
Numerical computations for option pricing using the risk-neutral probability in a single-
step binomial model can be found in the accompanying Excel file. The input data and
call option are the same as in Examples 1.6, 1.7, 1.11, and 1.17

Remark 1.21
Expectation can be defined for any random variable X : Ω → R and for any probability
on Ω. In the case when the probability used is the original probability P (called
physical or real probability, in contrast to the risk-neutral one) the subscript P is
often dropped, so that

E(X) = pX(u) + (1 − p)X(d).

Consider the discounted underlying asset price and compute the risk-neutral expectation
S(0)(1 + U ) S(0)(1 + D)
EQ (S(1))
e = q + (1 − q)
 1+R  1+R 
R − D S(0)(1 + U ) U − R S(0)(1 + D)
= +
U −D 1+R U −D 1+R
= S(0) = S(0).
e

17
Saying that EQ (S(1))
e = S(0)
e means that the sequence S(0),
e S(1)
e is a (one-step) martingale
under Q. Accordingly, we often call Q a martingale probability.
The reason why Q is referred to as ‘risk-neutral’ becomes clear when we calculate the
expected return on the stock under this probability: it turns out to be equal to the return R
on the risk-free asset.
Theorem 1.22
The expected return under Q on the underlying is equal to the risk-free return if and
only if Q is a martingale probability:

EQ (KS ) = R.

Proof

Recall that the return on the underlying is KS = S(1)−S(0)

S(0)
, which equals U in the
R−D
up-state and D in the down state. Since q = U −D , we have

R−D U −R
EQ (KS ) = qU + (1 − q)D = U+ D = R.
U −D U −D
R−D
On the other hand, if qU + (1 − q)D = R, then solving this for q gives q = U −D
.

1.4 Two underlying securities

Here we discuss the possibility of modelling the prices of two underlying assets in a binomial
model. This is a very natural step towards reality, where we face a large number of stocks
traded in the market. Even when we restrict to a single stock, we may wish to add an
international perspective: if the stock is traded on some foreign market, we should include in
our model the home currency on this market. Finally, some derivative securities may involve
many underlying assets.
The initial prices of our two underlyings will be denoted by S1 (0) and S2 (0). Their future
prices S1 (1) and S2 (1) are determined by random returns, denoted by K1 , K2 and assumed
to be defined on the probability space Ω = {u, d} for simplicity, with obvious modifications
if a general space is needed. So we have

Siu = Si (0)(1 + Ki (u)),

Sid = Si (0)(1 + Ki (d)),

i = 1, 2. In general, we do not expect both returns to agree with the ‘up’ and ‘down’
meaning of the elements of Ω = {u, d}. For instance, we may have K1 (u) > K1 (d),

18
but K2 (u) < K2 (d). However we always assume distinct values: w the natural notation
Ki (u) = Ui , Ki (d) = Di for i = 1, 2, we assume that U1 6= D1 and U2 =
6 D2 .
Example 1.23
There are real numbers a 6= 0, and b such that K1 = aK2 + b. To show this we need
to solve the following two equations for a and b:

U1 = aU2 + b,
D1 = aD2 + b.

This gives
U1 − D1
a= ,
U2 − D2
D1 U2 − D2 U1
b= .
U2 − D2

We can interpret this result in terms of the covariance of K1 and K2 . Given a random
variable X, we obtain the centred (mean-zero) random variable Xc = X − E(X) by subtract-
ing its expectation. So the covariance of X and Y can be written as cov(X, Y ) = E(Xc Yc ).
Normalising this, we obtain the correlation coefficient
cov(X, Y )
ρX,Y = ,
σX σY
since σX = σXc and similarly for Y. Note that we always have |ρX,Y | ≤ 1. The linear
relationship between K1 and K2 shows that they are perfectly correlated, that is |ρK1 ,K2 | = 1.
We can regard S2 as a derivative security with S1 as the underlying; a function h such
that S2 (1) = h(S1 (1)) can be easily found.
Denote by Q1 the risk-neutral probability determined by S1 , that is Q1 = (q1 , 1 − q1 ) with
R − D1
q1 = .
U1 − D1
We know that the no-arbitrage price of the derivative security S2 is given by
S2 (0) = EQ1 (S̃2 (1)). (1.4)
Introduce the risk-neutral probability Q2 given by S2 , that is, with
R − D2
q2 = ,
U2 − D2
and by the martingale property for S̃2 (i.e. the equality of expectations) we have
S2 (0) = EQ2 (S̃2 (1)). (1.5)

19
From (1.4) and (1.5) we get

EQ1 (K2 ) = EQ2 (K2 ) = R.

This gives q1 U2 + (1 − q1 )D2 = R, which implies q1 = q2 . We have thus proved the following
result.
Theorem 1.24
No-arbitrage implies Q1 = Q2 .

The converse is also true.

Theorem 1.25
If Q1 = Q2 then the market composed of a risk-free asset and two stocks does not
admit arbitrage.

Proof
Seeking a candidate for arbitrage, suppose that for some portfolio we have

xS1 (0) + yA(0) + zS2 (0) = 0

and
V(x,y,z) (1) = xS1 (1) + yA(1) + zS2 (1) ≥ 0.
Take the expectation with respect to Q1

EQ1 (V(x,y,z) (1)) = xEQ1 (S1 (1)) + yA(1) + zEQ1 (S2 (1))
= xEQ1 (S1 (1)) + yA(1) + zEQ2 (S2 (1))
= xS1 (0)(1 + R) + yA(0)(1 + R) + zS2 (0)(1 + R)
= 0.

In general, a non-negative random variable with zero expectation must be zero almost
surely. Here this expectation is a weighted average of non-negative numbers with
positive weights, so each of the numbers must be zero, and arbitrage is excluded.

Example 1.26

Consider two risky assets with up/down returns U1 = 20%, D1 = −10% and U2 =
30%, D2 = −20%, respectively. Suppose that R = 10%. We can compute the

20
 risk-neutral probability of the ‘up’ scenario for each these two risky assets:

R − D1 0.1 − (−0.1) 2
q1 = = = ,
U1 − D1 0.2 − (−0.1) 3
R − D2 0.1 − (−0.2) 3
q2 = = = .
U2 − D2 0.3 − (−0.2) 5

According to Theorem 1.25, since q1 6= q2 , we know that there is an arbitrage oppor-

tunity.

This video shows how to construct an arbitrage opportunity.

Remark 1.27
What the above results show is that the binomial model does not allow sufficient
freedom to include two arbitrary stocks in the model: under the No Arbitrage Principle,
we find that their risk-neutral probabilities are identical and the stocks are perfectly
correlated. Thus, in order to construct models which reflect real markets more closely
we are forced to increase their complexity. First we shall increase the number of possible
random outcomes.

1.5 The trinomial model

An immediate generalisation of the binomial model is the trinomial one, where at time 1 the
stock price has three possible values. To describe this we introduce an intermediate return
M in addition to U, D considered before. So the rate of return for the stock is a random
variable of the following form:
 u
 K = U with probability pu ,
K= K m = M with probability pm ,
 d
K = D with probability pd ,

with pu , pm , pd ∈ (0, 1) such that pu + pm + pd = 1, assuming that Ω = {u, m, d}. The the
expectation of a random variable X on Ω takes the form

E(X) = pu X(u) + pm X(m) + pd X(d).

We assume that the three returns are different, and without loss of generality we may further
assume D < M < U.

21
 A derivative security will be a random variable H(1) = h(S(1)) as in the binomial model,
and again we note that any random variable X : Ω → R is of this form by defining h on the
three (different!) values of S(1) to agree with those of H(1) and extending it arbitrarily.
If R is outside the range of stock returns, the model makes no sense since it is always
preferable to either invest risk free or in the stock; in other words, the model admits arbitrage
and, as in the binomial case, the converse is also true.
Theorem 1.28
No-arbitrage is equivalent to D < R < U.

Proof
The proof of the implication from the left to right is the same as in the binomial case
(Theorem 1.3).
Conversely, a candidate (x, y) for an arbitrage opportunity must satisfy

S(0)
V(x,y) (0) = 0, xS(0) + yA(0) = 0, y = −x .
A(0)

We investigate the terminal value

V(x,y) (1) = xS(1) − xS(0)(1 + R) = xS(0)(KS − R).

When x > 0, we have


 xS(0)(U − R) > 0 in the ‘u’ state
V(x,y) (1) = xS(0)(M − R) in the ‘m’ state ,
xS(0)(D − R) < 0 in the ‘d’ state


with the opposite inequalities when x < 0. Hence (x, y) cannot be an arbitrage: if
R lies between U and D, then the three values above cannot all have the same sign
when x 6= 0, and they are all zero when x = 0. Therefore the model is arbitrage free.

In real life stock prices change in a discrete way, quite frequently, but the movements
are limited to the so-called ticks. So for example, given S(0) = 30 the possible values after
one time step are 30.50 or 29.50, say. However, the price may not change at all, hence
introducing the medium return is quite realistic.
Unfortunately, this leads to mathematical complications since replication (using only risk-
free investment and a single underlying) of random payoffs at time 1 is no longer always
possible. To see this, we try to replicate a random payoff H(1) with three values H u , H m , H d ,

22
which requires finding a portfolio (x, y) with V(x,y) (1) = H(1), that is,

xS u + yA(0)(1 + R) = H u ,
xS m + yA(0)(1 + R) = H m ,
xS d + yA(0)(1 + R) = H d ,

obtaining a system of three equations in two variables which in general has no solution. Thus
typically a replicating portfolio for H(1) does not exist.

The no-arbitrage interval

Since replication with V(x,y) (1) = H(1) is impossible in general, we relax the condition
somewhat.
Definition 1.29
A portfolio (x, y) super-replicates the option with payoff H(1) if

V(x,y) (1) ≥ H(1).

The holder of such a portfolio is protected against the option payoff. Super-replication is
satisfactory from the point of view of the trader who sold the option, for whom the option
payoff is a liability. This is a source of risk and super-replication hedges (i.e. eliminates) this
risk completely. Whatever happens, the trader will have sufficient funds to cover the amounts
due to the option holder. Therefore, the party selling the option would willingly accept the
following price:
Definition 1.30
The super-replication option price H super is the minimal initial value of a super-
replicating portfolio:

H super = inf {V(x,y) (0) : V(x,y) (1) ≥ H(1)}.

(x,y)∈R2

Example 1.31

Let S(0) = 30, U = 20%, M = 10%, D = −10%, R = 5%, and H = (S(1) − 32)+
(a call option with strike price 32). The least expensive super-replicating portfolio is
x = 0.4444, y = −11.4286 and the super-replication price is H super = 1.9048.
For details, see the Excel file, in which the Solver has been used to compute the
solution.

We should also analyse the problem from the point of view of the option buyer. Suppose
that the buyer wishes raise cash to buy the option by selling a portfolio (x, y) at time 0.

23
Then, at time 1, the buyer will receive the option payoff H(1) and will have to pay V(x,y) (1)
to close the short position in the portfolio. There will be no risk to the buyer as long as
V(x,y) (1) ≤ H(1). This leads to the following definition of a sub-replicating portfolio.
Definition 1.32
A portfolio (x, y) sub-replicates the option with payoff H(1) if

V(x,y) (1) ≤ H(1).

The highest amount V(x,y) (0) that can be raised by selling a sub-replicating portfolio to
buy the option is the so-called sub-replication price, formally defined as follows.
Definition 1.33
The sub-replication option price H sub is the maximal initial value of a sub-replicating
portfolio:
H sub = sup {V(x,y) (0) : V(x,y) (1) ≤ H(1)}.
(x,y)∈R2

Example 1.34

Let S(0) = 30, U = 20%, M = 10%, D = −10%, R = 5%, and H = (S(1) − 32)+
(a call option with strike price 32). The most expensive sub-replicating portfolio is
x = 0.1667, y = −4.2857 and the sub-replication price is H sub = 0.7143.
For details, see the Excel file, in which the Solver has been used to compute the
solution.

Remark 1.35
In fact, H sub is a maximum and, similarly, H super is a minimum (and not just a
supremum or infimum) since we are optimising the linear function V(x,y) (0) of x, y over
a closed subset of the plane, so that the supremum and infimum are attained at some
elements of the set.

If, given two portfolios (x, y) and (x, y), we have

V(x,y) (1) ≤ H(1) ≤ V(x,y) (1)
then V(x,y) (0) ≤ V(x,y) (0) since otherwise we could initiate an obvious arbitrage strategy. This
implies that H sub ≤ H super since the inequality will be preserved after taking the supremum
over all (x, y) and then the infimum over all (x, y). Thus an interval of possible prices emerges.
Definition 1.36
We call the open interval (H sub , H super ) the no-arbitrage interval.

24
 The next Proposition justifies the terminology.
Proposition 1.37

If a ∈ (H sub , H super ), then taking H(0) = a does not lead to an (extended) arbitrage
opportunity.

Proof
Let (x, y, z) be such that

V(x,y,z) (0) = xS(0) + yA(0) + zH(0) = 0

and suppose that

V(x,y,z) (1) = xS(1) + yA(1) + zH(1) ≥ 0.

When z = 0, this cannot be an arbitrage opportunity by the No Arbitrage Principle

in the model without the option. We may therefore assume that z = 1 or z = −1
without loss of generality (divide both sides by |z|). Suppose z = 1, so

H(0) = −xS(0) − yA(0) = V(−x,−y) (0)

and
−xS(1) − yA(1) ≤ H(1),
hence (−x, −y) is sub-replicating. Thus

V(−x,−y) (0) ≤ H sub < H(0),

contradicting our assumption that V(x,y,z) (0) = 0. Similarly, when z = −1, we find
that (x, y) is super-replicating, yielding the contradiction V(x,y) (0) ≥ H super > H(0).

Next, consider the possibility of using risk-neutral probabilities, which in the binomial
model served as an alternative to replication for pricing. We use Q to denote a probability
measure on the three-element sample space Ω = {u, m, d}. It can be represented by a triple
qu , qm , qd ≥ 0 with qu + qm + qd = 1. The definition of a risk-neutral probability is similar to
that for two-element sample space.
Definition 1.38
We say that Q is a risk-neutral probability on Ω = {u, m, d} if qu , qm , qd ∈ (0, 1),

qu + qm + qd = 1. (1.6)

25
 and
S(0) = (1 + R)−1 (qu S u + qm S m + qd S d ). (1.7)

For any random variable X on Ω = {u, m, d}, we write

EQ (X) = qu X(u) + qm X(m) + qd X(d)

for the risk-neutral expectation of X.

Equality (1.7) is equivalent to EQ (S(1)) = S(0)(1 + R), with in turn is equivalent to

qu U + qm M + qd D = R. (1.8)

We have two linear equations (1.7) and (1.8) with three unknowns qu , qm , qd . Such a system
may have many solutions, potentially giving infinitely many risk-neutral probabilities. Of
course, we still have to identify which of these solutions take all their values in the interval
(0, 1). The next proposition will show that in a pricing model consistent with the No Arbitrage
Principle there are always some that do.
Example 1.39

Let U = 20%, M = −1%, D = −10% and R = 5%. With qm = 0.8, we get

qu = 0.268 and qd = 1 − qu − qm = −0.068. This does not represent a probability.
See the Excel file to rectify the example.

Proposition 1.40
If the trinomial pricing model is consistent with the No Arbitrage Principle, then there
is a probability Q that is non-zero on each scenario u, m, d such that the underlying
stock prices satisfy
S(0) = EQ [S̃(1)].

Proof
To find such a probability note that the absence of arbitrage (Theorem 1.28) ensures
that S d < S(0)(1 + R) < S u , or, equivalently, S̃ d < S(0) < S̃ u . Now, either S̃ m >
S(0) or S̃ m ≤ S(0). In the first case let A = 21 (S̃ u + S̃ m ) and let B = S̃ d ; in
the second case take A = S u and B = 12 (S̃ m + S̃ d ). In both cases we obtain B <
S(0) < A, so there is a c ∈ (0, 1) such that S(0) = cA + (1 − c)B. Hence either
S(0) = c 21 (S̃ u + S̃ m ) + (1 − c)S̃ d or S(0) = cS̃ u + (1 − c) 12 (S̃ m + S̃ d ). In each case
we can find the required probability measure Q = (qu , qm , qd ): in the first, we take
qu = qm = 2c , qd = 1 − c, and in the second case we use qu = c, qm = qd = 12 (1 − c).

26
 In the binomial model the risk-neutral probability is unique and gives the price of any
derivative security as the discounted expectation of the payoff with respect to the risk-
neutral probability. Since here we may have many such probabilities, we cannot expect the
same result, but we now show that the risk-neutral probabilities determine a range of option
prices consistent with the No Arbitrage Principle. Let

I = {(1 + R)−1 EQ (H(1)) : Q is a risk-neutral probability}.

Note that I is an interval since the set of risk-neutral probabilities is obviously convex and
the map
(qu , qm , qd ) → (1 + R)−1 EQ (H(1))
is linear.
Proposition 1.41
Each number in the interval I is a price of H consistent with the No Arbitrage Principle.

Proof
Let a ∈ I, write H(0) = a, so that H(0) = (1 + R)−1 EQ (H(1)) for some risk-neutral
probability Q. Consider a portfolio (x, y, z) in the extended market with V(x,y,z) (0) = 0
and V(x,y,z) (1) ≥ 0. Compute the expectation of discounted V(x,y,z) (1) with respect
to Q:

EQ ((1 + R)−1 V(x,y,z) (1)) = xEQ ((1 + R)−1 S(1)) + yA(0) + zEQ ((1 + R)−1 H(1))
= xS(0) + yA(0) + zH(0) = 0.

But zero expectation of a non-negative discrete random variable forces this random
variable to be zero, hence V(x,y,z) (1) = 0.

Finally, we show that the two ‘no-arbitrage intervals’ coincide.

Theorem 1.42
Let H be a European derivative security in the single-period, single-stock trinomial
market model, and assume that H sub < H super . Then

I = (H sub , H super ).

Proof
First, we show that I is contained in [H sub , H super ]. We assume A(0) = 1 without loss
of generality.

27
 Suppose the (x, y) is a sub-replicating portfolio, that is,

xS(1) + y(1 + R) ≤ H(1).

Take any risk-neutral probability Q and compute the expectation of both sides:

EQ (xS(1) + y(1 + R)) ≤ EQ (H(1)).

By linearity and the martingale property EQ (S(1)) = S(0)(1 + R), we have

EQ (xS(1) + y(1 + R)) = xS(0)(1 + R) + y(1 + R),

which implies
xS(0) + y ≤ (1 + R)−1 EQ (H(1)).
Take the supremum over all such portfolios

sup(xS(0) + y) ≤ (1 + R)−1 EQ (H(1)),

so that
H sub ≤ (1 + R)−1 EQ (H(1)).
In the same manner we can prove that H super ≥ (1 + R)−1 EQ (H(1)) by considering
super-replicating portfolios (x, y) such that

xS(1) + y(1 + R) ≥ H(1),

which implies
xS(0) + y ≥ (1 + R)−1 EQ (H(1)).
Taking the infimum over all super-replicating portfolios (x, y) we conclude that

H super = inf(xS(0) + y) ≥ (1 + R)−1 EQ (H(1)).

Hence I is a sub-interval of [H sub , H super ].

Next, for notational convenience assume that Ω = {ω1 , ω2 , ω3 } and let

A = {(S(1, e ωi )) : i = 1, 2, 3} ⊂ R2 .
e ωi ), H(1,

Let C denote the closed convex hull of A, that is, the set of all, not necessarily strict,
convex combinations of these points. Since A has three elements, the closed set C is
either a triangle or a line segment in R2 . Since D < M < U , the three values S(1, e ωi )
for i = 1, 2, 3 are different from one another. This means that if C is a triangle, then
none of its sides is vertical, and if C is a line segment, then it is not vertical either.

28
 First, suppose that the three points of A are not collinear, so that C is a triangle
2
in R . The No Arbitrage Principle ensures that S(0) lies strictly between the smallest
and largest value of S(1,
e ωi ) for i = 1, 2, 3. This follows since S(0)(1 + D) < S(0)(1 +
R) < S(0)(1 + U ). Thus the intersection of the triangle C with the vertical line
through (S(0), 0) is a non-empty line segment with endpoints (S(0), a), (S(0), b), say.
Clearly, (a, b) = {u : (S(0), u) ∈ int(C)}, where int(C) denotes the interior of C.
Moreover, we can express any (S(0), u) in int(C) as a strictly convex combination of
the points of A, that is,
3
X 3
X
(S(0), u) = ci (S(1,
e ωi ), H(1,
e ωi )), where ci > 0, ci = 1.
i=1 i=1

These three positive numbers define a probability measure Q on Ω such that qu = c1 ,

qm = c2 , qd = c3 . The equality
3
X
S(0) = ci S(1,
e ωi ) = EQ [S(1)]
e = (1 + R)−1 EQ [S(1)]
i=1

shows that Q is a risk-neutral probability, which, together with the equality

3
X
u= ci H(1,
e ωi ) = EQ [H(1)]
e = (1 + R)−1 EQ [H(1)],
i=1

shows that u ∈ I. Hence (a, b) ⊂ I. We have therefore shown that

(a, b) ⊂ I ⊂ [H sub , H super ].

Now observe that, since the triangle C has no vertical edges, there is a non-vertical
straight line l with gradient x and intercept y passing through the point (S(0), b) such
that A is contained in the half-plane below this line. In fact, this is a line containing
one side of the triangle (see Figure 1.1, where Pi = (S(1, e ωi ), H(1,
e ωi ))). In other
words, H(1)
e ≤ xS(1)
e + y or, equivalently,

H(1) ≤ xS(1) + y(1 + R),

which means that (x, y) is a super-replicating portfolio. Because the line passes through
(S(0), b), we have
b = xS(0) + y ≥ H super .
Similarly, there is also a non-vertical straight line, again containing one side of the
triangle, with gradient x0 and intercept y 0 that passes through (S(0), a) such that A is

29
contained in the half-plane above this line (in Figure 1.1 this would be the line passing
through P1 and P3 ), so that H(1)
e ≥ x0 S(1)
e + y 0 . It follows that

H(1) ≥ x0 S(1) + y 0 (1 + R),

so that (x0 , y 0 ) is a sub-replicating portfolio. Since the line passes through (S(0), a),
we have
a = x0 S(0) + y 0 ≤ H sub .
Hence we have shown that (a, b) contains (H sub , H super ), so that, I =
(H , H super ). This completes the proof in the case when C is a triangle.
sub

Finally, if C were a line segment, we already know that it would not be ver-
tical, so the line would have intercept x and gradient y. Since all three points
Pi = (S(1,
e ωi ), H(1,
e ωi )) would lie on this line, we would have xS(1)
e + y = H(1),
e so

xS(1) + y = H(1),

meaning that (x, y) would be both a sub-replicating and a super-replicating portfolio.

This, in turn, would means that

H sub ≥ xS(0) + y ≥ H super ,

contrary to the assumption that H sub < H super . Hence C cannot, in fact, be a line
segment in these circumstances.

Figure 1.1: Diagram in the proof of Theorem 1.42.

30
 Remark 1.43
It should be clear from the above proof that this result also holds for any single-step,
single-stock model with n ≥ 3 possible outcomes at time 1. The convex hull is then
a polygon with at most n vertices in the plane, and otherwise the proof is essentially
the same.

Special case: H super = H sub

Finally, we need to discuss the case when sub- and super-replicating prices meet. In the
trinomial model this is unusual, but it may take place.
Proposition 1.44

If H sub = H super then there exists a replicating portfolio.

Proof
In the definitions of H sub and H super we can take min and max since the sets are
closed. So there are portfolios (xsub , y sub ) sub-replicating H(1) and (xsuper , y super )
super-replicating H(1) such that

xsub S(0) + y sub = xsuper S(0) + y super ,

hence
(xsub − xsuper )S(0) = y super − y sub .
Moreover,
xsub S̃(1) + y sub ≤ H̃(1) ≤ xsuper S̃(1) + y super ,
so
(xsub − xsuper )S̃(1) ≤ y super − y sub = (xsub − xsuper )S(0).
Hence if xsub − xsuper > 0, then S̃(1) ≤ S(0), that is, U, M, D ≤ R, contradicting
the No Arbitrage Principle. Similarly, if xsub − xsuper < 0, then S̃(1) ≥ S(0), that
is, U, M, D ≥ R, also contradicting the No Arbitrage Principle. It follows that the
two portfolios are the same, (xsub , y sub ) = (xsuper , y super ), and thus form a replicating
portfolio.

This shows the existence of a replicating portfolio when H sub = H super . Next we show
that such a portfolio is unique.
Proposition 1.45
The replicating portfolio is unique.

31
 Proof
Suppose that (x1 , y1 ) and (x2 , y2 ) replicate the claim H. This means hat

x1 S̃(1) + y1 = H̃(1) = x2 S̃(1) + y2 .

Their initial values must be the same,

x1 S(0) + y1 = x2 S(0) + y2 ,

so
(x1 − x2 )S̃(1) = (y2 − y1 ) = (x1 − x2 )S(0).
As a result, x1 = x2 because S(1) is a random variable with at least two distinct
values. This, then, implies that y1 = y2 .

Finally we observe that in the current circumstances the claim H has a single price H(0)
consistent with the No Arbitrage Principle.
Proposition 1.46

The existence of a replicating portfolio for H implies that there is a unique price H(0)
that does not allow an extended arbitrage opportunity, and this price is given by

H(0) = EQ (H̃(1))

for any risk-neural probability Q in the trinomial model.

Proof
If a portfolio (x, y) replicating the claim H exists, then the price H(0) of the claim
must be equal to the initial value of the portfolio,

H(0) = xS(0) + y,

or else an extended arbitrage opportunity would exist. This proves the uniqueness
of H(0).
Now let Q be a risk-neutral probability, so that S(0) = EQ (S̃(1)). Since the
replicating portfolio satisfies xS̃(1) + y = H̃(1), we have

H(0) = xS(0) + y = EQ (xS̃(1) + y) = EQ (H̃(1)),

completing the proof.

Going back to the proof of Theorem 1.42, this can be seen very clearly. When C is a line

32
segment (that is, the three points of A are collinear), the intersection of C and the vertical
line through (S(0), 0) is a single point. This point has coordinates (S(0), H(0)), where H(0)
is the unique price of the contingent claim.
We shall show in Theorem 1.63 that the existence of a replicating portfolio for every
derivative security ensures not only that the price of each derivative security is unique, but
also implies uniqueness for the risk-neutral probability.

Replication with two stocks

An alternative way of approaching the replication problem and obtaining a unique price is
to ‘complete the market’ by adding a third variable in our three equations. In our trinomial
model we may assume that, in addition to the money market account, there are two stocks
S1 , S2 with returns K1 , K2 such that
K1u = U1 , K2u = U2 ,
K1m = M1 , K2m = M2 ,
K1d = D1 , K2d = D2 ,
for the three scenarios in Ω = {u, m, d}. The role of the indices u, m, d is to indicate the
scenario and not necessarily the sizes of the returns. The stock prices are then given by
S1 (1) = S1 (0)(1 + K1 ),
S2 (1) = S2 (0)(1 + K2 ).
For replication of a claim H in this model we need a portfolio (x1 , x2 , y) such that
x1 S1u + x2 S2u + yA(0)(1 + R) = H u ,
x1 S1m + x2 S2m + yA(0)(1 + R) = H m ,
x1 S1d + x2 S2d + yA(0)(1 + R) = H d .
In matrix form:
AX = H,
where      
S1u S2u A(0)(1 + R) x1 Hu
A =  S1m S2m A(0)(1 + R)  , X =  x2  , H =  Hm  .
S1d S2d A(0)(1 + R) y Hd
Depending on the coefficients, this linear system has either no solution, or a unique one, or
infinitely many. In particular, when the determinant of A is non-zero, we can invert A and
obtain a unique solution:
X = A−1 H.
The solution gives the option price as H(0) = x1 S1 (0) + x2 S2 (0) + yA(0). A model which
has a unique solution for every choice of random payoff H is called complete. We rephrase
this as a general definition.

33
 Definition 1.47
A market model is complete if every random payoff H(1) can be replicated.

We now seek a risk-neutral probability so that both discounted stock prices satisfy the
martingale condition, that is,
EQ (K1 ) = R,
EQ (K2 ) = R,
which can be written as
qu U1 + qm M1 + qd D1 = R,
(1.9)
qu U2 + qm M2 + qd D2 = R.
A system of three linear equations (the third one is qu + qm + qd = 1) with three variables
can be solved, apart from some singular cases. However, the question for us is whether the
solution represents a genuine probability measure, that is, if the three numbers qu , qm , qd
belong to the interval (0, 1). We illustrate this with examples.
Example 1.48
For an example where Q is not a genuine probability, consider stocks S1 , S2 with
U1 = 30%, M1 = 5%, D1 = −20%, U2 = −10%, M2 = 0%, D2 = 15%, R = 5%.
Solving the above equations, we find qu = 1, qm = −1, qd = 1.
This model allows arbitrage: If S1 (0) = 100, S2 (0) = 50, A(0) = 1, the portfolio
x1 = −0.5, x2 = −2, y = 150 is an arbitrage opportunity: V(x1 ,x2 ,y) (0) = 0, while
V(x1 ,x2 ,y) (1) ≥ 2.5 in each scenario. Note that here the two stocks do not go ‘up’
together.
The calculations can be found in the Excel file.

This example illustrates the following First Fundamental Theorem of Asset Pricing, which
will be proved more generally in the next section.
Theorem 1.49
When Ω is finite, there is no arbitrage if and only if there exists a risk-neutral probability
measure, with positive probabilities for each scenario.

Consider some further examples:

Example 1.50

Let H(1) = (S1 (1) − X)+ , which is a call option on stock 1. Let R = 5%, and let
S1 (0) = 30, A(0) = 1, X = 32, U1 = 20%, M1 = 10%, D1 = −10%. Suppose
now that S2 (0) = 30 and the movements of this stock are given by U2 = −5%,
M2 = 15%, D2 = 10%. We can find a replicating strategy, which provides the option

34
 price H(0) = 1.6402. To do this we simply solve the matrix equation
     
36 28.5 1.05 x1 4
 33 25.5 1.05   x2  =  1 
27 33 1.05 y 0

and compute H(0) = x1 S1 (0) + x2 S2 (0) + y. See the Excel file for details.
There is a unique risk-neutral probability Q and

H(0) = (1 + R)−1 EQ (H(1)) = 1.6402,

see the Excel file.

This seemingly positive result requires some critical thinking. We have obtained a unique
price of H, but this was obtained as a result of completing the market by introducing a second
underlying security. The option price therefore depends on this choice, which contradicts
common sense. Why should some other stock influence the price of a call option written
on stock 1? So, in fact, we still have a range of prices, this time related to the variety of
securities which can be chosen to complete the market. To see a possible use of the above
ideas suppose that a derivative security H written on S1 is traded, its market prices available.
Now we can choose H as the additional security, making the market complete, and we may
price other derivatives using the unique risk-neutral probability. In this way a statement, often
found in the financial literature, that ‘the market chooses the risk-neutral probability’, gains
substance.
Example 1.51

With U2 = 10%, M2 = 5%, D2 = −20% and the remaining data as before, the
replication problem can be solved with H(0) = −1.2698 which is nonsense (a negative
price for a positive payoff security gives arbitrage). See Excel file.
Here there is no risk-neutral probability: the equations (1.9) have the unique
solution qm = 2, qu = −0.83 which does not yield a probability measure. See Excel
file.

To price options we need both conditions: we need no-arbitrage to have replication,

and completeness in order to obtain a unique price. This leads to the following Second
Fundamental Theorem of Asset Pricing.
Theorem 1.52
A market model is arbitrage-free and complete if and only if there exists a unique
risk-neutral probability Q.

As with the first fundamental theorem, this result will be formulated and proved in a more

35
general setting in the next section. Our two simple examples (binomial and trinomial models)
serve to illustrate the main features of the general theory we now proceed to develop.

1.6 General single-step model

Assume given a riskless asset A with A(0) = 1 (we keep this assumption throughout this
section to simplify notation), accruing interest at rate R, together with d assets with prices
S1 (n), . . . , Sd (n), which we consider at times n = 0, 1. We assume that for i = 1, . . . , d the
prices Si (0) are known (non-random) and that at time 1 each Si (1) is a (positive) random
variable defined on a finite sample space Ω = {ω1 , . . . , ωM }, with a probability measure P
defined on each subset A of Ω by
X
P (A) = P (ω),
ω∈A

where P (ωj ) = pj ∈ (0, 1) for j = 1, . . . , M, and M

P
j=1 pj = 1. We take P to be the physical
(real-world) probability on Ω. A random variable is now a function

X: {ω1 , . . . , ωM } → R,

with expectation given by

M
X
E(X) = pj X(ωj ).
j=1

The random vector of prices (S1 (1), . . . , Sd (1)) consists of d random variables, each
taking at most M different values. In fact, we shall assume that altogether M different
values of this random vector appear at time 1. This ensures that all M elements of Ω are
‘needed’. As a consequence of this natural non-degeneracy assumption, in the same way as
in the binomial or trinomial model, any random variable X: Ω → R can be written in the
form X = h(S1 (1), . . . , Sd (1)) for some function h: Rd → R.
These data constitute a finite single-step market model. The notion of a portfolio
must be adjusted to incorporate d stocks: it is a (d + 1)-vector (x, y) = (x1 , . . . , xd , y)
whose coordinates indicate the stock holdings and risk-free investment, respectively. The
initial value of the portfolio (x, y) is
d
X
V(x,y) (0) = xi Si (0) + yA(0),
i=1

while its final value is the random variable

d
X
V(x,y) (1) = xi Si (1) + yA(1).
i=1

36
Without loss of generality, we can take A(0) = 1 for simplicity. Then A(1) = 1 + R, where R
is the riskless rate.
Note that fixing the initial value V (0) and the vector x = (x1 , . . . , xd ) of stock holdings
of a portfolio automatically determines the bond holding y. This is true, in particular, for
arbitrage opportunities, whose definition extends as follows to the current setting.
Definition 1.53
An arbitrage opportunity is a portfolio (x, y) = (x1 , . . . , xd , y) such that

(1) V(x,y) (0) = 0,

(2) V(x,y) (1) ≥ 0, with probability 1,

(3) V(x,y) (1) > 0 with positive probability.

The first fundamental theorem of asset pricing, which we now formulate and prove in this
setting, gives an equivalent condition for the absence of arbitrage.

First Fundamental Theorem

We consider a simple version of this key result, applicable to the single-step model (the
general multi-step case will be considered later). First we note what is meant by a risk-
neutral probability measure in this context.
Definition 1.54
A sequence Q = (q1 , . . . , qM ) of numbers is called a risk-neutral probability measure
for Ω = {ω1 , . . . , ωM } and stock prices {(Si (0), Si (1)) : i = 1, . . . , d} if

q1 , . . . , qM ∈ (0, 1),
M
X
qj = 1,
j=1

and
EQ (Si (1)) = (1 + R)Si (0)
for all i = 1, . . . , d, where EQ denotes the expectation with respect to Q, and R
denotes the riskless interest rate. Such Q is often called a (single-step) martingale
probability for the stock price processes {Si : i = 1, . . . , d}.

37
Theorem 1.55 (First Fundamental Theorem: single step)

In a finite single-step market model there is no arbitrage opportunity if and only if there
exists a risk-neutral probability measure.

Proof
Recall that we assume A(0) = 1. First suppose that there is a risk-neutral prob-
ability Q. Consider any portfolio (x, y) = (x1 , . . . , xd , y) with V(x,y) (0) = 0 and
V(x,y) (1) ≥ 0 (otherwise this portfolio could not be an arbitrage opportunity). Com-
pute
d
X 
EQ (V(x,y) (1)) = EQ xi Si (1, ωj ) + y(1 + R)
i=1
d
X
= xi EQ (Si (1)) + y(1 + R)
i=1
d
X
= (1 + R) xi Si (0) + y(1 + R)
i=1
= (1 + R)V(x,y) (0) = 0.

This implies that V(x,y) (1) = 0 with probability 1 since zero expectation of a non-
negative random variable forces this random variable to be zero with probability 1. It
follows that there is no arbitrage opportunity.
For the converse implication, suppose that there is no arbitrage opportunity. De-
fine the discounted gains of a portfolio (x, y) = (x1 , . . . , xd , y) to be the random
variable
X d  S (1) 
i
Gx = xi − Si (0) .
i=1
1+R
Note that y does not appear since the discounted gains of a risk-free position are zero.
The set
W = {Gx : x ∈ Rd }
of discounted portfolio gains is a vector space. It can be identified with a subspace
of RM by writing Gx = (Gx (ω1 ), . . . , Gx (ωM )).
We claim that if Gx (ω) ≥ 0 for all ω, then Gx (ω) = 0 for all ω. To see this,
take x = (x1 , . . . , xd ) such that Gx (ω) ≥ 0 for all ω. Ensure that the portfolio (x, y)
has zero initial wealth V(x,y) (0) = 0 by supplementing the xi by a suitable y. The

38
discounted final value of this portfolio is
d
1 X
Ve(x,y) (1) = xi Si (1) + y
1 + R i=1
d
1 X
= xi Si (1) + y − V (0)
1 + R i=1
d
X h S (1) i
i
= xi − Si (0)
i=1
1+R
= Gx ≥ 0.

Hence to avoid arbitrage Ve(x,y) (1) = Gx must be zero.

The risk-neutral probabilities we seek belong to the following subset of RM :
n M
X o
A = Q = (q1 , . . . , qM ) : qj = 1, qj ≥ 0 .
j=1

We shall show that there is a vector Q in A with positive coordinates that is orthog-
onal to the subspace W. The coordinates of such a vector give the required probabili-
ties qj since orthogonality means that the Euclidean inner product of (q1 , . . . , qM ) and
(Gx (ω1 ), . . . , Gx (ωM )) vanishes for each Gx in W , that is,
M
X
EQ (Gx ) = qj Gx (ωj ) = 0
j=1

for all Gx . Now, for any i = 1, . . . d, we can take xi = 1 and xk = 0 when k 6= i, so

that Gx = S1+Ri (1)
− Si (0), which implies EQ (Si (1)) = S(0)(1 + R).
The existence of such a Q = (q1 , . . . , qM ) is a consequence of the following lemma,
which we shall also use later when the multi-step version of the theorem is presented.
Lemma 1.56
Suppose that A ⊂ Rn is convex and compact, and W is a vector subspace of Rn
disjoint from A. Then there is a z ∈ Rn suchPthat hz, ai > 0 for all a ∈ A,
and hz, wi = 0 for all w ∈ W, where hx, yi = ni=1 xi yi is the Euclidean inner
product in Rn .

For the proof of this lemma, see page 45.

To see that this is sufficient to prove the theorem with n = M and A, W as above,
note that A is obviously convex. It is compact as it is closed and bounded (Heine–Borel

39
 theorem!). Since Gx ≥ 0 implies Gx =P0, it follows that A∩W = ∅. Now Lemma 1.56
provides a z = (z1 , ..., zM ) such that M j=1 zj aj > 0 for all a = (a1 , ..., aM ) ∈ A and
PM
j=1 zj Gx (ωj ) = 0 for all Gx in W. It follows that zj > 0 for each j = 1, . . . , M
since we can take (a1 , ..., aM ) in A with aj = 1 (and so with the remaining coordinates
necessarily zero). Hence, for each j = 1, . . . , M, we can take
1
qj = PM zj > 0,
i=1 zi
PM
so that j=1 qj = 1 and
M
X
qj Gx (ωj ) = 0
j=1

for all Gx in W. This completes the proof of the theorem.

Remark 1.57
We have shown that for finite single-step models the existence of a risk-neutral proba-
bility is equivalent to the absence of arbitrage in the model. We shall show later that
this remains true for multi-step finite market models, in general. Our proof will depend
heavily on the single-step case, and crucially on the finiteness of the market model,
that is, on the sample space Ω being finite.

Next we turn our attention to the question of completeness.

Second Fundamental Theorem

The trinomial model is an example of an incomplete market model, the defining feature
being the lack of replicating portfolios for some derivative securities. The binomial model,
on the other hand, makes it possible to replicate all derivative securities and is an example
of a complete market model. We re-state the definition of completeness more generally.
Recall that our market model contains d risky securities, whose prices we denote, for ease of
notation, as a random vector S(1) = (S1 (1), . . . , Sd (1)). The value at time 1 of a portfolio
(x, y) is the random variable V(x,y) (1) = hx, S(1)i + yA(1), where h·, ·i denotes the inner
product in Rd .
Definition 1.58
A single-step arbitrage-free market model is complete whenever for each derivative
security H = h(S(1)) there exists a replicating portfolio (x, y), that is, a portfolio
such that V(x,y) (1) = h(S(1)).

40
 Any random variable H is a derivative security in our market model since its value is
completely determined by the price vector S(1) and it can be written in the form H = h(S(1))
for some h : Rd → R, which is a consequence of the non-degeneracy assumption made at
the beginning of this section. In particular, for any ω ∈ Ω, the random variable 1{ω} is a
derivative security.
We now show that a complete market model is characterized by the existence of a unique
risk-neutral probability.
Theorem 1.59 (Second Fundamental Theorem: single step)

A single-step arbitrage-free market model is complete if and only if there is exactly one
risk-neutral probability.

Proof
No arbitrage implies the existence of a martingale probability, so we need to show that
we cannot have more than one such probability in a complete market. Suppose we
have two martingale probabilities Q1 6= Q2 . For any ω ∈ Ω, the random variable 1{ω}
is a derivative security. Since the market model is complete, there exists a replicating
portfolio (x, y) such that

hx, S(1)i + y(1 + R) = 1{ω} .

Take the expectation with respect to Qk , k = 1, 2:

EQk (hx, S(1)i + y(1 + R)) = EQk (1{ω} ) = Qk (ω)

and note that the left-hand side is independent of k since

EQk (hx, S(1)i + y(1 + R)) = hx, EQk (S(1))i + y(1 + R)

= hx, S(0)i(1 + R) + y(1 + R)

so that
Q1 (ω) = Q2 (ω).
But ω ∈ Ω was arbitrary, so the probabilities Q1 and Q2 coincide.
For the converse implication, assume Q is the unique risk-neutral probability and
suppose there is a derivative H which cannot be replicated. Our goal will be to
construct a martingale probability Q1 different from Q.
Consider the set of values at time 1 of all portfolios:

W = {V(x,y) (1) : (x, y) ∈ Rd+1 }.

Since none replicates H, we can conclude that H ∈

/ W. We treat W as a vector
M M
subspace of R by identifying R with the set of all random variables on Ω =

41
{ω1 , . . . , ωM }. The subspace W contains the vector 1 = (1, . . . , 1) because if x = 0
1
and y = 1+R then V(x,y) (1) = 1 for all scenarios. Now equip RM with the inner
PM
product hw, viQ = j=1 wj vj Q(ωj ) and use the following version of Lemma 1.56
with the subspace W and the compact convex set A = {H}.
Lemma 1.60
Suppose A ⊂ Rn is convex and compact, and W is a vector subspace of Rn
disjoint from A. Then there is a z ∈ Rn such that hz, aiQ > 0 for all a ∈ A,
and hz, wiQ = 0 for all w ∈ W, where hx, yiQ = di=1 xi yi qi , where qi > 0 for
P
all i = 1, . . . , d, is an inner product in Rn .

For the proof of this lemma, see page 47.

From the lemma we obtain an RM -vector Z = (z1 , . . . , zM ) such that hZ, HiQ
> 0 and hZ, V iQ = 0 for V ∈ W. The inequality implies that Z 6= 0, while, taking
V = 1 ∈ W , we have hZ, 1iQ = M
P
j=1 zj Q(ωj ) = 0, hence EQ (Z) = 0, where the
expectation is taken under Q.
Now we define a new measure, which will turn out to be a martingale probability
different from Q:  zj 
Q1 (ωj ) = 1 + Q(ωj )
2a
for each j = 1, . . . , M , where a = max{|z1 |, . . . , |zM |} > 0. This is a probability
measure since
M M M
X X X zj
Q1 (ωj ) = Q(ωj ) + Q(ωj )
j=1 j=1 j=1
2a
1
= 1+ EQ (Z) = 1.
2a
Clearly −a ≤ zj for all j, so the number 1 + zj /2a is always positive and for at least
one coordinate it is different from 1 since Z is non-zero, thus Q1 6= Q. Finally, we
check the martingale property of each Si (for i ≤ d) under Q1 :
M
X zj
EQ1 (Si (1)) = Si (1, ωj )(1 + )Q(ωj )
j=1
2a
M
1 X
= EQ (Si (1)) + zj Si (1, ωj )Q(ωj )
2a j=1
1
= EQ (Si (1)) + hZ, Si (1)iQ
2a
= Si (0)(1 + R).

42
 The final identity follows because Si (1) belongs to W (take xi = 1, xk = 0 for k 6= i,
and y = 0, which gives V(x,y) (1) = Si (1)), so that hZ, Si (1)iQ = 0. This completes
the proof.

Consider a derivative security with payoff H(1) and time 0 price H(0). The notion of an
extended portfolio and extended arbitrage opportunity can be adapted to the general single-
step setting as follows. An extended portfolio is a triple (x, y, z), where x = (x1 , . . . , xd )
represents the positions in the d stocks, y the risk-free position (the money market account
A), and z the number of units of the derivative security H. The values of the extended
portfolio at times 0 and 1 are
d
X
V(x,y,z) (0) = xi Si (0) + yA(0) + zH(0),
i=1
d
X
V(x,y,z) (1) = xi Si (1) + yA(1) + zH(1).
i=1

Definition 1.61
We say that and extended portfolio (x, y, z) is an (extended) arbitrage opportunity
if

(1) V(x,y,z) (0) = 0,

(2) V(x,y,z) (1) ≥ 0 with probability 1,

(3) V(x,y,z) (1) > 0 with positive probability.

We can now apply the fundamental theorems to option pricing in a single-step model in
which there is no extended arbitrage opportunity. Suppose the payoff H(1) can be replicated
by a portfolio (x, y) = (x1 , . . . , xd , y), that is,

V(x,y) (1) = H(1).

It follows that
H(0) = V(x,y) (0)
since otherwise we would have an extended arbitrage opportunity by a similar argument as
in the proof of Theorem 1.10. The expectation with respect to a risk-neutral probability Q

43
(which exists by the first fundamental theorem) gives

EQ (V(x,y) (1)) = EQ (hx, S(1)i + y(1 + R))

d
X
= xi EQ (Si (1)) + y(1 + R)
i=1
d
X 
= (1 + R) xi Si (0) + y
i=1
= (1 + R)V(x,y) (0) = (1 + R)H(0),

where we have used the martingale property for all stock prices. We obtain
1
H(0) = EQ (H(1)).
1+R
If the model is complete, then every derivative security can be replicated, which means that
the price H(0) of every derivative security is given by this formula with expectation taken
under the unique (by the second fundamental theorem) risk-neutral probability Q.
Example 1.62
In a trinomial model, consider a basket option with payoff

H(1) = max(S1 (1), S2 (1)).

Let S1 (0) = S2 (0) = 60 and R = 5%, and let U1 = 20%, M1 = 10%, D1 = −10%
and U2 = −5%,
M2 = 15%, D2 = 10%. There is a unique risk-neutral probability
7 1 4
Q = 18 , 6 , 9 , and the option price is H(0) ≈ 65.56.

See video for details.

Finally, we note that the sub- and super-replicating prices coincide in complete models.
Theorem 1.63
Completeness implies that H sub = H super .

Proof
If (x, y) = (x1 , . . . , xd , y) is a replicating portfolio, V(x,y) (1) = H(1), and (x, y) is any
super-replicating portfolio, V(x,y) (1) ≥ H(1), then

V(x,y) (1) ≥ V(x,y) (1).

44
 The No Arbitrage Principle implies

V(x,y) (0) ≥ V(x,y) (0)

so taking the infimum over super-replicating portfolios, we have

inf {V(x,y) (0)} ≥ V(x,y) (0).

(x,y) super-replicating

But the replicating price belongs to the set of all super-replicating prices,

V(x,y) (0) ∈ {V(x,y) (0) : (x, y) super-replicating},

so the infimum is attained and

H super = inf {V(x,y) (0)} = V(x,y) (0).

(x,y) super-replicating

The argument for sub-replicating portfolios is similar and leads to

H sub = sup {V(x,y) (0)} = V(x,y) (0),

(x,y) sub-replicating

where the supremum is taken over all sub-replicating portfolios.

1.7 Auxiliary lemmas

Lemma 1.56
Suppose A ⊂ Rn is convex and compact, and W is a vector subspace of Rn disjoint
n
from A. Then there is a z ∈ RP such that hz, ai > 0 for all a ∈ A, and hz, wi = 0
n
for all w ∈ W, where hx, yi = i=1 xi yi is the Euclidean inner product in Rn .

Proof
Consider the set
A − W = {a − w : a ∈ A, w ∈ W }.
Since A and W are disjoint, 0 ∈ / A − W.
Next, we note that A − W is convex since both A and W are. In more detail,
if x, y ∈ A − W, x = a − v, y = b − w, then with α ∈ [0, 1], αx + (1 − α)y =
[αa + (1 − α)b] − [αv + (1 − α)w] ∈ A − W. Thus A − W is convex.
Consider the set B = {|x| : x ∈ A − W }. We claim that it is closed. To show this

45
consider a convergent sequence bn ∈ B with bn → b. Then there exists a sequence
xn ∈ A − W such that |xn | = bn , and we can write xn = an − wn for some an ∈ A and
wn ∈ W. The sequence an has a convergent subsequence ank → a since A is compact.
The sequence wnk is bounded since otherwise |ank − wnk | = bnk would go to infinity.
So selecting a further subsequence gives a convergent sequence wnkm → w. Now by
the continuity of x 7→ |x|, we have |a − w| = b, so b ∈ B. Hence B is closed.
Take z to be the element of A − W closest to 0, so that for any other element c
of A − W we have |c| ≥ |z| > 0 (recall that 0 ∈ / A − W ). The existence of z follows
from the fact that the set B = {|x| : x ∈ A − W } is closed, does not contain 0, and
so has the smallest positive element, which is the Euclidean norm of some element of
A − W , which we take as z.
Consider the triangle (in Rn ) with vertices 0, z, c for any c in A − W . We shall
prove that hz, ci ≥ hz, zi = |z|2 > 0. The idea is best explained by a picture, see
Figure 1.2. We take a vector y parallel to c such that z and y − z are perpendicular.
Then
|z| |c|
hz, ci = |z| |c| cos α = |z| |c| = |z|2 ,
|y| |y|
so it is sufficient to show that |c| ≥ |y|. Suppose that |c| < |y|. The line segment
joining c and z is contained in the convex set A−W since both z and c belong to A−W .
But the line intersects the interior of a ball with the centre at 0 and radius |z| . So this
segment contains a point with distance to 0 strictly smaller than |z|, a contradiction
with the definition of z.
This gives hz, a − wi = hz, ai − hz, wi ≥ hz, zi for any a ∈ A and w ∈ W.
However, W is a vector space, so nw ∈ W , hence hz, ai − nhz, wi ≥ hz, zi for
each n, which is impossible unless hz, wi = 0. This, in turn, yields hz, ai ≥ hz, zi > 0,
completing the proof.

y
0  c

Figure 1.2: Picture used in the proof of Lemma 1.56.

46
Lemma 1.60
Suppose A ⊂ Rn is convex and compact, and W is a vector subspace of Rn disjoint
from A. Then there is a z 0 ∈ Rn such that hz 0 , aiQ > 0 for all a ∈ A, and hz 0 , wiQ = 0
for all w ∈ W, where
X n
hx, yiQ = xi yi qi
i=1

with qi > 0, i = 1, . . . , n is an inner product in Rn .

Proof
By Lemma 1.56, we have z ∈ Rn with ni=1 zi ai > 0, ni=1 zi wi = 0 for any a =
P P
∈ A and w =P(w1 , . . . , wn ) ∈ W. Take zi0 = zi /qi , which satisfies the
(a1 , . . . , an ) P
claims since i=1 zi0 yi qi = i=1 zi yi for any yi .

47
Chapter 2

Multi-step binomial model

Having described single-step pricing models, we are ready to consider models with a finite
number of consecutive trading dates. We shall see that these models behave essentially as a
succession of one-step models, so that our earlier analysis can be applied repeatedly to yield
explicit pricing formulae.
For a simple binomial model we now describe how the use of risk-neutral probabilities
implies the martingale property of discounted stock prices and thus of the discounted value
process of any self-financing strategy. In this model pricing techniques for European options
using replicating portfolios or risk-neutral probabilities are compared and an explicit pricing
formula for the European call is then derived.

2.1 Two-step binomial tree example

Most of the essential features of a general multi-step model can be seen in a simple example
with two time steps, where we avoid cumbersome notation related to a general case. Here
we take time to be 0, T, 2T, and we simplify the notation by just specifying the number of
a step, ignoring its actual length T . We wish to build a model of stock prices S(0), S(1),
S(2), where S(0) is a given number and the remaining quantities are random variables. The
construction consists of two repetitions of the single-step binomial tree dynamics. In the
single-step case the core of the construction was the model of the rate of return with two
possible values, which can be thought of as related to a single toss of a coin. Intuitively, for
many steps we can use more coins, or toss one coin repeatedly. We write the results obtained
in a two-step experiment as a pair built of symbols u and d. This gives the sample space as

Ω = {uu, ud, du, dd}.

48
The probability on Ω will be defined by multiplying probabilities for single coin tosses. De-
noting the probability of an up movement in each single step by p, we have

P (uu) = p2 ,
P (ud) = P (du) = p(1 − p),
P (dd) = (1 − p)2 ,

using the simplified notation P (ω) rather than P ({ω}). Given the probabilities of single-
element sets, we can find the probability of any subset of Ω as
X
P (A) = P (ω).
ω∈A

Next we consider the returns K1 , K2 as random variables on Ω defined for n = 1, 2 by


U with probability p,
Kn =
D with probability 1 − p.

The value of Kn is assumed to depend only on the nth element of ω and it acts in the same
way as in the single-step binomial model, with K1 determined by the first step, K2 by the
second. So

K1 (uu) = K1 (ud) = U,
K1 (du) = K1 (dd) = D,
K2 (uu) = K2 (du) = U,
K2 (ud) = K2 (dd) = D.

Observe that for all real numbers x1 , x2 we have

P (K1 = x1 and K2 = x2 ) = P (K1 = x1 ) × P (K2 = x2 )

written more formally as

P ({ω : K1 (ω) = x1 } ∩ {ω : K2 (ω) = x2 })

= P ({ω : K1 (ω) = x1 }) × P ({ω : K2 (ω) .= x2 }) (2.1)

This means that K1 , K2 are independent.

The stock prices are defined as

S(1) = S(0)(1 + K1 ),
S(2) = S(1)(1 + K2 ) = S(0)(1 + K1 )(1 + K2 ).

49
We introduce convenient self-explanatory notation where a superscript determines both the
number of steps and the position on the tree:
 u
S = S(0)(1 + U ),
S(1) =
S d = S(0)(1 + D),
 uu
 S = S(0)(1 + U )2 ,
S(2) = S ud = S du = S(0)(1 + U )(1 + D),
 dd
S = S(0)(1 + D)2 .
Similar notation will also be used below for other process.
We also assume a risk-free asset, represented by the money market account determined
by riskless return R. We have

A(n) = A(0)(1 + R)n , n = 1, 2

by a straightforward extension of the familiar single-step scheme.

Example 2.1

Let S(0) = 100, U = 20%, D = −10%, and p = 0.6. Then we have a concrete
recombining (i.e. such that S ud = S du ) tree of stock prices:
S(0) S(1) S(2)
144
0.6

120
0.6
0.4
100 108
0.6
0.4
90
0.4
81
Taking A(0) = 100 and R = 5%, we have A(1) = 105 and A(2) = 110.25.

Trading strategies
Example 2.2
Continuing Example 2.1, suppose we have at our disposal a cash amount of 50 and
wish to invest in the market by buying one share of stock. For the first portfolio the
stock and risk-free positions will be denoted x(1) and y(1) respectively. This is slightly
surprising as one might expect to indicate time by 0 here, that is, the time when this
portfolio is created. However, it is an established notational tradition to label the
portfolio by the number of the time step over which the portfolio is held (step 1 here).

50
 Taking x(1) = 1, with V (0) = 50 we have a short risk-free position

V (0) − x(1)S(0) 50 − 1 × 100

y(1) = = = −0.5.
A(0) 100

At time 1 we consider the two possible outcomes of the first step:

Case 1: ‘up’
The value of our portfolio is now

V u = x(1)S u + y(1)A(0)(1 + R) = 67.5.

The next portfolio is denoted by (xu (2), y u (2)). We decide to reduce our stock position
to 0.8, which determines the risk-free component since we assume no inflow or outflow
of funds, so
67.5 = 0.8 × 120 + y u (2) × 105
to give y u (2) = −0.285. We can find the final values of our investment:

V uu = 0.8S uu + (−0.285)A(0)(1 + R)2 = 85.275,

V ud = 0.8S ud + (−0.285)A(0)(1 + R)2 = 56.475.

Case 2: ‘down’
In this case the value of the portfolio becomes

V d = x(1)S d + y(1)A(0)(1 + R) = 37.5.

With cheap stock we decide to buy more, so we set xd (2) = 1.2. This implies

37.5 = 1.2 × 90 + y d (2) × 105

with y d (2) = −0.671. The final values are

V du = 1.2 × S du + (−0.671) × A(0)(1 + R)2 = 55.575,

V dd = 1.2 × S dd + (−0.671) × A(0)(1 + R)2 = 23.175.

Note that the tree of portfolio values is not recombining: V ud 6= V du .

In general, the value of a strategy is a process given by

V(x,y) (0) = x(1)S(0) + y(1)A(0),

V(x,y) (1) = x(1)S(1) + y(1)A(1),
V(x,y) (2) = x(2)S(2) + y(2)A(2),

51
with the self-financing condition

V(x,y) (1) = x(2)S(1) + y(2)A(1)

representing the fact that at the rebalancing moment n = 1, when (x(2), y(2)) are selected,
we only use the funds available then, and we use them all. We then say that (x, y) is a
self-financing strategy.

Pricing
As in the single-step model, the basis for pricing is the No Arbitrage Principle. We start with
the definition of an arbitrage strategy in the current context.
Definition 2.3
We say that (x, y) = {(x(n), y(n))}n=1,2 is an arbitrage strategy if

(1) V(x,y) (0) = 0,

(2) V(x,y) (1) = x(2)S(1) + y(2)A(1) with probability 1 (that is, (x, y) is a self-
financing strategy),

(3) V(x,y) (2) ≥ 0 with probability 1,

(4) V(x,y) (2) > 0 with positive probability.

Then the No Arbitrage Principle takes the following form.

Assumption 2.4
There is no arbitrage strategy.

A derivative security of European type has payoff of the form H = h(S(2)) for some
real-valued function h defined on [0, ∞). Such a security is called path-independent. A
path-dependent version has payoff H = h(S(1), S(2)) for some h : [0, ∞) × [0, ∞) → R.
The prices of such a derivative security form a sequence H(n), n = 0, 1, 2, with H(2) = H.
As for the single-step model, we now formulate an extended version of the No Arbitrage
Principle, which allows us to determine these derivative security prices.
First, we need to adapt the concept of extended strategy, by which we shall now
understand a triple (x, y, z), where (x, y) = {(x(n), y(n))}n=1,2 is a two-step strategy as
above describing the positions in stock and the money market, and z ∈ R is a position in the
derivative security H.

52
 Remark 2.5
Note that, in contrast to x(n) and y(n), the position z in the derivative security H
does not depend on the time step n. Once a position z in H is created at the time 0,
the same position z is held up to the final time 2. We say that a static position is
held in H.

The values of such an extended strategy are

V(x,y,z) (0) = x(1)S(0) + y(1)A(0) + zH(0) = V(x,y) (0) + zH(0),

V(x,y,z) (1) = x(1)S(1) + y(1)A(1) + zH(1) = V(x,y) (1) + zH(1),
V(x,y,z) (2) = x(2)S(2) + y(2)A(2) + zH(2) = V(x,y) (2) + zH(2).

We say that (x, y, z) is an extended self-financing strategy whenever (x, y) is self-

financing, that is,
V(x,y) (1) = x(2)S(1) + y(2)A(1).

Definition 2.6
We say that (x, y, z) is an extended arbitrage strategy whenever

1. V(x,y,z) (0) = 0,

2. (x, y) is a self-financing strategy (that is, (x, y, z) is an extended self-financing

strategy),

3. V(x,y,z) (2) ≥ 0 with probability 1,

4. V(x,y,z) (2) > 0 with positive probability.

Now the extended No Arbitrage Principle takes the following form.

Assumption 2.7

There is no extended arbitrage strategy (x, y, z).

Example 2.8

Let S(0) = 100, U = 20%, D = −10%, and R = 5%. Consider a call option with
exercise price K = 100 at time 2. The payoff of the option is

 (S(0)(1 + U )2 − K)+ = 44 if ω = uu,
+
C(2, ω) = (S(0)(1 + U )(1 + D) − K) = 8 if ω = ud or ω = du,
(S(0)(1 + D)2 − K)+ = 0 if ω = dd.


53
It is a path-independent derivative security, by which we mean that

C(2, ud) = C(2, du).

We work out the replicating strategy (x, y) by moving backwards in time, following
the scheme depicted in Figure 2.1. Replication means that the value of the replicating
strategy at time 2 should match the option price, V(x,y) (2) = C(2). Thus, in the ‘up’
case we get the system

V uu = xu (2)S uu + y u (2)A(0)(1 + R)2 = 44,

V ud = xu (2)S ud + y u (2)A(0)(1 + R)2 = 8,

with solution

xu (2) = 1,
y u (2) = −0.907,

and in the ‘down’ case we solve

V du = xd (2)S du + y d (2)A(0)(1 + R)2 = 8,

V dd = xd (2)S dd + y d (2)A(0)(1 + R)2 = 0,

to find

xu (2) = 0.296,
y u (2) = −0.218.

By the self-financing condition for the replicating strategy, we get the values if the
strategy at time 1:

V u = xu (2)S u + y u (2)A(0)(1 + R) = 24.762,

V d = xd (2)S d + y d (2)A(0)(1 + R) = 3.810.

Next, replicating these values, we get

V u = x(1)S u + y(1)A(0)(1 + R) = 24.762,

V d = x(1)S d + y(1)A(0)(1 + R) = 3.810,

so

x(1) = 0.698,
y(1) = −0.562,

54
 resulting in
V (0) = x(1)S(0) + y(1)A(0) = 13.605.
This, then, gives the option price

C(0) = V (0) = 13.605

by the extended No Arbitrage Principle. To see this requires just a moment of reflection.
Suppose that C(0) < 13.605 (bearing in mind the rounding error), in which case we
buy the option and sell the strategy investing the difference risk free. Replication at
the exercise time ensures that the value of the strategy will cancel with the option
payoff and a profit will remain. In case of the opposite inequality, we take the opposite
action.
In a similar manner, we can also get the option prices at time 1 by the extended No
Arbitrage Principle for the single-step subtrees with root at node u and d, respectively:

C u = V u = 24.762,
C d = V d = 3.810.

This video shows a similar example for a put option.

uu uu uu
u u u
ud ud ud
d d d
dd dd dd

Figure 2.1: Diagram in Example 2.8.

Note that the values of the strategy constructed above form a recombining tree, but
this does not have to be the case if the payoffs of the derivative security for ud and du are
different (i.e. a path-dependent security).
Recall that in the single-step model a convenient alternative was to use risk-neutral prob-
abilities to compute the price of a derivative security. This can readily be done at each step
of the above procedure, but our goal is to give the story extra flavour by developing some
theory, which can be generalised to more general settings.

55
2.2 Partitions and information
For our next topic it is in fact better to extend the time scale to three steps. This gives the
corresponding sample space as

Ω = {uuu, uud, udu, udd, duu, dud, ddu, ddd}.

The probability will be defined in a similar way as before,

P (uuu) = p3 ,
P (uud) = P (udu) = P (duu) = p2 (1 − p),
P (udd) = P (dud) = P (ddu) = p(1 − p)2 ,
P (ddd) = (1 − p)3 .

We have to modify the definition of K1 , K2 and add another return K3 with value U
for triples ending with u and value D for triples ending with d, so that Kn depends only
on the nth element of ω, just as in the two-step binomial model. However, the notion of
independence is now somewhat more involved, as clarified by the following proposition.
Proposition 2.9

K1 , K2 , K3 are independent, which by definition means that we have condition (2.1)

not only for the pair of indices 1,2, but for the pairs of indices 1,3 and 2,3 too, and
also the following condition for the triple of indices 1,2,3:
3
\  Y3
P {ω : Ki = xi } = P ({ω : Ki = xi }).
k=1 k=1

Proof
The verification for each pair of indices is the same as in the two-step case. To verify
the above identity for the triple of indices, note that by definition both sides are 0
unless each xi takes one of the two values U, D. There are 8 possible cases, grouped
according to the number of times U occurs: for U U U we have p3 on the left by the
definition of P, and this equals the RHS. Similarly for U U D, U DU or DU U we obtain
p2 (1 − p) on each side, for U DD, DU D, DDU it is p(1 − p)2 , and for DDD we
obtain (1 − p)3 .

The third stock price takes the form

 uuu

 S = S(0)(1 + U )3 ,
 uud
S = S udu = S duu = S(0)(1 + U )2 (1 + D),
S(3) = S(2)(1 + K3 ) =

 S ddu = S dud = S udd = S(0)(1 + U )(1 + D)2 ,
 ddd
S = S(0)(1 + D)3 .
56
 Example 2.10
Within the numerical scheme of Example 2.1, we have the picture
S(0) S(1) S(2) S(3)
172.8
0.6

144
0.6
0.4
120 129.6
0.6 0.6
0.4
100 108
0.6
0.4 0.4
90 97.2
0.6
0.4
81
0.4
72.9

Example 2.11
Consider the following random variable

H = (S(3) − 100)+

which is nothing but a call payoff with strike K = 100 and exercise time 3. If we own
such a security, we might be interested in the expected payoff, so we compute

E(H) = 72.8 × P ({uuu}) + 29.6 × P ({uud, udu, duu})

+ 0 × P ({udd, dud, ddu}) + 0 × P ({ddd})
= 28.512.

Next we analyse the future possible developments and their impact on our perception. It
is important to emphasise that we are now performing a ‘what if’ analysis. We do not travel
in time, but consider hypothetically all possible future scenarios with the purpose of revising
our point of view and possibly responding with some appropriate action.
Position at step 1
Case ‘up’
Let us prepare ourselves for the case of an increase in the stock price after one step. This
means that the collection of available scenarios will be reduced to the triples beginning with u,
the set of which we denote by Bu = {uuu, uud, udu, udd}. Knowledge that the first step is
‘up’ means that this set will play the role of Ω and a probability is produced by adjusting the

57
original probabilities to ensure that the new probability of Bu is 1. For ω ∈ Bu we put

P (ω)
Pu (ω) = .
P (Bu )

Observe that u is the first element of each ω ∈ Bu and, according to the definition of P, we
have P (Bu ) = p and this is the first factor producing P (ω), so it will cancel. For instance

P (udu) p(1 − p)p

Pu (udu) = = = (1 − p)p.
P (Bu ) p
1
P
Of course Pu (Bu ) = P (Bu ) ω∈Bu P (ω) = 1. Observe that for A ⊆ Bu ,

P (A ∩ Bu )
Pu (A) = = P (A|Bu ),
P (Bu )

which the conditional probability of A given Bu .

Example 2.12
Next we compute the expectation of H in this new probability space Bu . We consider
the values of H corresponding to all four elements of Bu and write the expectation as
X
H(ω)Pu (ω) = 72.8 × 0.62 + 29.6 × 2 × 0.6 × 0.4 + 0 × 0.42 = 40.416,
ω∈Bu

which would be good news for the option holder.

The sum in this example is the conditional expectation of H given Bu . We introduce the
following two alternative pieces of notation for the conditional expectation:
X
E(H|Bu ) = E(H|S(1) = 120) = H(ω)Pu (ω) = 40.416.
ω∈Bu

Case ‘down’
In the case of a down move at the first step we introduce, similarly, the set Bd of all still
remaining scenarios (all those beginning with a d) and the adjusted probabilities for ω ∈ Bd :

P (ω)
Pd (ω) = .
P (Bd )

Clearly, for any A ⊂ Bu , we have the conditional probability

P (A ∩ Bd )
Pd (A) = = P (A|Bd ),
P (Bd )

58
and the expectation of H in this situation is

E(H|Bd ) = E(H|S(1) = 90) = 29.6 × p2 = 10.656,

a bit disappointing for anyone who owns such a security

The two cases considered decompose all scenarios into two groups. Mathematically, we
have a so-called partition of Ω, meaning that Ω = Bu ∪ Bd , and the two components are
disjoint, which motivates the following general definition.
Definition 2.13

S P = {Bi } of subsets of Ω is a partition of Ω if Bi ∩ Bj = ∅ for i 6= j and

A family
Ω = i Bi .

The partition defined at the first step will be denoted by

P1 = {Bu , Bd }.

We put together these two cases defining a random variable with two competing notations,
equivalent since the partition P1 is fully determined by S(1),

40.416 if ω ∈ Bu ,
E(H|P1 )(ω) = E(H|S(1))(ω) =
10.656 if ω ∈ Bd .

Position at step 2
Consider all possible price movements in the first two steps. There are four cases, which
can be described by specifying a partition

P2 = {Buu , Bud , Bdu , Bdd }

of Ω into four disjoint sets:

Buu = {uuu, uud},

Bud = {udu, udd},
Bdu = {duu, dud},
Bdd = {ddu, ddd}.

Each of these is equipped with a probability defined in a similar way as before, for example,
for any ω ∈ Bud ,
P (ω)
Pud (ω) = .
P (Bud )
Observe the effect of the cancellation mentioned above and notice for example that P (Bud ) =
p(1 − p), so Pud (udu) = p and Pud (udd) = 1 − p, and in each case we have a familiar single-
step binomial tree.

59
 As before, we compute the expected value of H in each of these cases, that is, the
conditional expectations

E(H|Buu ) = 55.52,
E(H|Bud ) = 17.76,
E(H|Bdu ) = 17.76,
E(H|Bdd ) = 0,

hence 
 55.52 if ω ∈ Buu ,
E(H|P2 )(ω) = 17.76 if ω ∈ Bud ∪ Bdu ,
0 if ω ∈ Bdd .


Remark 2.14
We can see that the conditional expectation given Bud is equal to the conditional
expectation given Bdu . This means that we could employ the alternative, more intuitive
notation E(H|S(2)). To this end note that the partition generated by the values of
S(2) is different from the partition related to the history of the movements over the
first two steps, introduced above. The partition generated by S(2) consists of just
three elements: Buu , Bud ∪ Bdu , Bdd , and the random variable E(H|S(2)) takes three
values. The fact that the values E(H|Bud ), E(H|Bdu ) coincide gives the equivalence
of these two approaches in this instance, but this does not have to be the case in
general.
The actual values of S(2) are irrelevant for defining the conditional expectation
since they do not appear in the computations. What matters is the partition related
to these values.

Position at step 3
Given the knowledge of all three steps, we will know the actual payoff. The above analysis
can formally be performed for the partition

P3 = {Buuu , . . . , Bddd }

of Ω into eight single-element parts with no randomness and the obvious conclusion that

E(H|P3 ) = H.

We are in a position to give the general definition.

Definition 2.15
For any random variable X and any event B such that P (B) 6= 0, the conditional

60
 expectation of X given B is defined as
X
E(X|B) = X(ω)P (ω|B),
ω∈Ω

1
where P (ω|B) = P (B) P (ω) for ω ∈ B and P (ω|B) = 0 otherwise.
Given a partition P = {Bi }, the above definition applied to each Bi gives a
function defined for each ω in Ω, constant on each Bi ,

E(X|P) : Ω → R

by assigning
Bi 3 ω → E(X|Bi ).
This random variable is called the conditional expectation of X with respect to the
partition {Bi }.

The properties of E(X|P) are discussed in a general setting in the next chapter. Here we
continue the analysis of binomial trees.
Example 2.16
Now we consider the 4-step binomial tree.

This video discusses the partitions in this case.

2.3 Martingale properties

In the single-step model, a martingale was simply a process with constant expectation. With
many steps this notion is much more sophisticated.

Discounted stock prices

We work with the three-step case as in the previous section, and introduce discounted stock
prices
S(n)
e = S(n)(1 + R)−n .
R−D
Recall that in the single-step model we changed the probability to q = U −D
and obtained,
for the probability Q = {q, 1 − q} defined on {u, d}, that

Su Sd
EQ (S̃(1)) = q + (1 − q) = S(0).
1+R 1+R

61
In the three-step model the probability Q is defined on subsets of the 8-element probability
space Ω = {uuu, uud, . . . , ddd} in the same way as P was defined by means of p. For
example, Q(udu) = q 2 (1 − q) and then Q({ω : S(1)(ω) = S u }) = Q(Bu ) = q. Taking
a trivial partition P0 = {Ω} and noting that conditional expectation with respect to P0
coincides with the expectation, we have
1
EQ (S̃(1)|P0 ) = EQ (S̃(1)|Ω) = (q S̃ u + (1 − q)S̃ d ) = EQ (S̃(1)) = S(0).
Q(Ω)

Next, we compute EQ (S̃(2)|P1 ), which requires finding EQ (S̃(2)|Bu ) and EQ (S̃(2)|Bd ).

Now
1 1
EQ (S̃(2)|Bu ) = (q 2 S uu + q(1 − q)S ud )
Q(Bu ) (1 + R)2
1
= (qS u (1 + U ) + (1 − q)S u (1 + D)) since Q(Bu ) = q
(1 + R)2
1
= S u (1 + R) since qU + (1 − q)D = R
(1 + R)2
1
= S u,
1+R
and similarly
1
EQ (S̃(2)|Bd ) = S d.
1+R
Putting these together, we have

EQ (S̃(2)|P1 ) = S̃(1).

The next step is to find EQ (S̃(3)|P2 ). We just consider one of the four cases in detail:

1 1
EQ (S̃(3)|Buu ) = (q 3 S uuu + q 2 (1 − q)S uud )
Q(Buu ) (1 + R)3
1
= (qS uu (1 + U ) + (1 − q)S uu (1 + D)) since Q(Buu ) = q 2
(1 + R)3
1
= S uu (1 + R) since qU + (1 − q)D = R
(1 + R)3
1
= 2
S uu ,
(1 + R)

with the conclusion that

EQ (S̃(3)|P2 ) = S̃(2).

62
Filtration
Observe that the partitions P0 , P1 , P2 , P3 considered above are such that for each n = 1, 2, 3
and each B ∈ Pn there is an A ∈ Pn−1 satisfying B ⊂ A. This represents the fact that A
carries more information than B. As time passes, we gain more information.
Example 2.17
For example, Bud ⊂ Bu . If we happen to be in Bu , we know that the first step was
‘up’, but have no knowledge of the second (or third) step. However, if we have landed
in Bud , we have more information: we know not only that the first step was ‘up’, but
also that the second step was ‘down’ (but still have no knowledge of the third step).

We are ready for two general definitions.

Definition 2.18
A sequence (Pn )n=0,...,N of partitions of Ω is called a filtration whenever for each
n = 1, . . . , N and each B ∈ Pn there is an A ∈ Pn−1 such that B ⊂ A.

Definition 2.19
A sequence (M (n))n=0,...,N of random variables defined on Ω is a martingale for the
filtration (Pn )n=0,...,N and probability P if

EP (M (n)|Pn−1 ) = M (n − 1).

We can conclude that the discounted stock prices follow a martingale with respect to Q
for n = 0, 1, 2, 3. Later we shall see that of course this is true for any number of steps.

Strategies and pricing

Suppose that we invest V (0), choosing the stock positions x(n), n = 1, 2, 3 with x(1) decided
at time 0, so constant, x(2) decided at time 1, thus being a random variable with two values
xu (2), xd (2), and x(3) decided at time 2 on the basis of the information provided by the
stock prices. The risk-free position is chosen according to the self-financing principle

V (0) − x(1)S(0)
y(1) = ,
A(0)
V(x,y) (1) − x(2)S(1)
y(2) = ,
A(1)
V(x,y) (2) − x(3)S(2)
y(3) = ,
A(2)

63
where we should remember that the last two lines represent equalities between random vari-
ables.
Proposition 2.20

The discounted strategy values Ṽ(x,y) (n) form a martingale with respect to the filtration
P0 , P1 , P2 , P3 and risk-neutral probability Q.

Proof
First
1 
EQ (Ṽ(x,y) (1)|P0 ) = q [x(1)S u + y(1)A(1)]
1+R

+ (1 − q) x(1)S d + y(1)A(1)


1
= x(1)[qS(0)(1 + U ) + (1 − q)S(0)(1 + D)]
1+R
+ y(1)A(0)
= x(1)S(0) + y(1)A(0)
= V (0).

Next, to find EQ (Ṽ(x,y) (2)|P1 ) we compute

1 1  2 u
EQ (Ṽ(x,y) (2)|Bu ) = q [x (2)S uu + y u (2)A(2)]
(1 + R)2 Q(Bu )
+ q(1 − q)[xu (2)S ud + y u (2)A(2)]
1 1
= 2
xu (2)[qS uu + (1 − q)S ud ] + y u (2)A(1)
(1 + R) 1+R
since Q(Bu ) = q
1 1
= 2
xu (2)S u (1 + R) + y u (2)A(1)
(1 + R) 1+R
1
= (xu (2)S u + y u (2)A(1))
1+R
= Ṽ u (1),

having inserted y(2). In addition to the single-step martingale property, the key points
here are that xu (2) is constant on Bu , with the same value in each of the two cases,
and the self-financing property, as captured by the form of the risk-free component.
Clearly, similar analysis can be performed for Bd and in all cases at time 2. So, for
n = 0, 1, 2 we have shown that

EQ (Ṽ(x,y) (n + 1)|Pn ) = Ṽ(x,y) (n).

64
 Corollary 2.21
As an immediate consequence, we have the martingale property of the process of option
prices since they are values of the replicating strategy.

The final property of the discounted value process we shall need is that for n = 0, 1, 2, 3,

EQ (Ṽ(x,y) (n)) = V (0).

This will imply, in particular, that

1
C(0) = EQ ((S(3) − K)+ ).
(1 + R)3
We already have this result for n = 1, and we only analyse the case n = 2 to see the general
pattern, leaving the rest to the reader:
1
EQ (Ṽ(x,y) (2)) = [q 2 V uu + q(1 − q)V ud + (1 − q)qV du + (1 − q)2 V dd ]
(1 + R)2
1
= [q(qV uu + (1 − q)V ud ) + (1 − q)(qV du + (1 − q)V dd )]
(1 + R)2
and here

qV uu + (1 − q)V ud = qxu (2)S uu + qy u (2)A(2) + (1 − q)xu (2)S ud

+ (1 − q)y u (2)A(2)
= xu (2)[qS uu + (1 − q)S ud ] + y u (2)A(2)
= xu (2)S u (1 + R) + y u (2)A(1)(1 + R)
= V u (1 + R) by the self-financing property,
qV du + (1 − q)V dd = V d (1 + R) similarly,

so that
1
EQ (Ṽ(x,y) (2)) = [qV u + (1 − q)V d ]
1+R
= EQ (Ṽ(x,y) (1))
= V (0).

2.4 The Cox–Ross–Rubinstein model

We generalise the above analysis by extending the scheme to N steps. We assume that
trading occurs at finitely many specified dates. Denote the time interval between trading
dates by h, so that trading occurs at times t = nh, n = 1, 2, . . . , N.

65
 We introduce the sample space for the N -step model exactly as it was done for n = 2, 3.
Namely, we take Ω = {u, d}N , the set of all sequences of length N consisting of characters
u, d (such sequences will be called paths). In keeping with the above cases, the probability
of a path ω in Ω will be
P (ω) = pk (1 − p)N −k ,
where k is the number of occurrences of u and N − k the number of occurrences of d
N


in ω. Note that there are k such paths, so the probability of the set of such paths is
N
p (1 − p)N −k .

k
k
For each n = 1, . . . , N the binomial splittings induce a partition of Ω into 2n disjoint sets:
Pn = {Buu...u , ..., Bdd...d }, where each set in the partition has an n-tuple subscript consisting
of u’s and d’s, so that in each set all paths ω have their first n entries in common. We can see
(as is obvious in the earlier examples) that for any n = 1, . . . , N , each set in the partition Pn
is a disjoint union of two sets in Pn+1 , just like for N = 3 we have Bu = Buu ∪ Bud . The
sequence of partitions P0 , P1 , . . . , PN is clearly a filtration.
Fix two numbers −1 < D < U to represent the random returns in each step defined by

U with probability p,
Kn =
D with probability 1 − p.

Again, Kn acts only on the nth entry of the sequence defining the path ω. We note that
these returns are independent, in the sense described in the next remark.
Remark 2.22

The random variables {Kn : n = 1, 2, . . . , N } defined on Ω = {u, d}N are indepen-

dent. By this we mean that for any subset {i1 , . . . , ik } of indices from {1, . . . , N }
and any real x1 , . . . , xk , we have
k
\  Yk
P {Kij = xj } = P (Kij = xj ).
j=1 j=1

This is again easy to see from the definition of the Kn (see also Proposition 2.9).

Assume that S(0) is given, while

S(n) = S(n − 1)(1 + Kn )

for n = 1, . . . , N. This means, in particular, that the value of S(n, ω) for a path ω ∈ Ω
depends only on the number of ‘up’ and ‘down’ movements (that is, the number of u and d
symbols) in the first n steps of the path ω, but not on the actual order of these movements
(symbols). In this case we say that prices follow a recombinant binomial tree.
In addition to holding stock, we can invest in the money market account, which we take
as a security manufactured by a series of single risk-free deposits, with return R per period,

66
assumed constant throughout. So we have a sequence of numbers, known in advance, with
A(0) given (we shall frequently take A(0) = 1 for simplicity), and

A(n) = A(n − 1)(1 + R)

for n = 1, . . . , N. To avoid arbitrage we assume that D < R < U , as we did in the single-step
case.
As in the cases of one, two, or three steps, the risk-neutral probability Q is obtained
by replacing p by
R−D
q=
U −D
in each step.
The key property observed for a three-step model can easily be extended.
Theorem 2.23
The sequence of discounted stock prices in the binomial model is a martingale with
respect to the filtration P0 , P1 , . . . , PN and risk-neutral probability Q.

Proof
The key argument, repeated in each case above, is concerned with examining the
two next step prices at each node and computing their expectation exactly as for
a single step. This can be summarised in the following way: for any B ∈ Pn at
n = 0, 1, 2, . . . , N − 1,

Q(S(n + 1) = S(n)(1 + U )|B) = q,

Q(S(n + 1) = S(n)(1 + D)|B) = 1 − q,
R−D U −R
and since q = U −D
and 1 − q = U −D
, we obtain

EQ (S(n + 1)|Pn ) = S(n)(1 + U )q + S(n)(1 + D) (1 − q)

R−D U −R
= S(n)(1 + U ) + S(n)(1 + D)
U −D U −D
= S(n)(1 + R),

which gives the result upon discounting.

The above observations lead us to the following theorem.

67
 Theorem 2.24
In the binomial model the discounted prices of a derivative security with payoff H(N )
are given by the recursive relations

H̃(n − 1) = EQ (H̃(n)|Pn−1 )

for n = 1, 2, . . . , N , that is, the discounted option price process H̃(n) is a martingale
with respect to the filtration P0 , P1 , . . . , PN and risk-neutral probability Q.
In particular, the initial price is given by

H(0) = (1 + R)−N EQ (H(N )).

Proof
Take any n = 1, . . . , N and consider the single-step submodels at each node at time
n − 1. The formula in Theorem 1.19 applies for each such single step model, and can
be written as
H(n − 1) = EQ (1 + R)−1 H(n)|Pn−1 .

Dividing both sides by (1 + R)n−1 gives

H̃(n − 1) = (1 + R)n−1 H(n − 1) = EQ (1 + R)−n H(n)|Pn−1 = EQ (H̃(n)|Pn−1 ).

Now, using the tower property for expectations, we obtain

EQ (H(n − 1)) = EQ EQ (1 + R)−1 H(n)|Pn−1 = EQ (1 + R)−1 H(n)

for each n = 1, 2, . . . , N , hence

 
H(0) = EQ (H(0)) = EQ (1 + R)−1 H(1) = · · · = EQ (1 + R)−N H(N ) .

Example 2.25
We can use Theorem 2.24 to price a European call expiring at time 2 in a two-step
binomial model. Let S(0) = 100, U = 10%, D = −10%, and R = 5%, and let
the strike price for the call be K = 90. Then H u (1) = 24.29, H d (1) = 6.43,
H(0) = 18.88. See the Excel file for details.

If the random variable H(N ) has the following form

H(N ) = h(S(N )),
then we can write
H(N ) = h S(0)(1 + U )Y (1 + D)N −Y ,

68
where Y is a random variable giving the number of upward movements in N steps along each
path, which has the binomial distribution
 
N k
Q{Y = k} = q (1 − q)N −k for k = 0, 1, . . . , N
k
under the risk-neutral probability. Hence
 
H(0) = EQ (1 + R)−N H(N )
N  
−N
X N k
q (1 − q)N −k h S(0)(1 + U )k (1 + D)N −k .


= (1 + R)
k=0
k

Example 2.26 (The CRR formula)

In particular, for the European call option with strike price K this summation begins
at the first integer k for which the payoff is non-zero:
N  
−N
X N k
H(0) = (1 + R) q (1 − q)N −k (S(0)(1 + U )k (1 + D)N −k − K),
k=l
k

where
l = min{k : S(0)(1 + U )k (1 + D)N −k > K}.
This formula can be written in a more concise form. First, we write it as
N  
−N
X N k
H(0) = (1 + R) S(0) q (1 − q)N −k (1 + U )k (1 + D)N −k
k=l
k
N  
−N
X N k
− (1 + R) K q (1 − q)N −k .
k=l
k

The second term can be written as

−(1 + R)−N K(1 − Fq (l − 1))
where Fq is the cumulative distribution function of the binomial distribution,
n  
X N k
Fq (n) = q (1 − q)N −k .
k=0
k

The first term can be rearranged as

N   k  N −k
X N 1+U 1+D
S(0) q (1 − q) .
k=l
k 1+R 1+R

69
 Write
1+U
q1 = q
1+R
so that
1 + R − q − qU (1 − q) + (R − qU )
1 − q1 = =
1+R 1+R
and, since qU + (1 − q)D = R, we have

1+D
1 − q1 = (1 − q) .
1+R
The first term above then takes the form
N  
X N k
S(0) q1 (1 − q1 )N −k = S(0)(1 − Fq1 (l − 1))
k=l
k

using the distribution function again. Finally,

C(0) = S(0)(1 − Fq1 (l − 1)) − (1 + R)−N K(1 − Fq (l − 1)).

This is the Cox–Ross–Rubinstein (CRR) formula for the initial price of a European
call option. (Many texts prefer to state this formula in terms of the complementary
binomial distribution Ψ, where
N  
X Nk
Ψ(l, N, r) = 1 − Fr (l − 1) = rk (1 − r)N −k
k=l
k

for 0 < r < 1.)

The coefficient at S(0) has a clear meaning: it is the probability that the option will be
exercised, computed by means of the modified probability q1 .
In the second term the exercise price is discounted, which is logical since the exercise
price K applies at time N, but the formula is concerned with the price at time 0. The coef-
ficient at the discounted K gives the martingale probability that the option will be exercised.
Example 2.27

In the Excel file we use the CRR formula to price a call option. Let N = 90, S(0) = 100,
U = 0.06%, D = −0.02%, and R = 0.02%, and let the strike price be K = 100.
Then l = 23 and the CRR formula gives C(0) = 1.7837.

70
2.5 Delta hedging
Example 2.28
Let us put ourselves in a position of an option writer who has sold a call with exercise
time N = 3 for the no-arbitrage price resulting from the binomial model with U =
20%, D = −10%, S(0) = 100 with R = 10%, with an exercise price of K = 105. The
price can easily be computed using the CRR formula: C(0) = 23.31.
We are exposed to some risk. For instance, if the stock keeps going up in each period,
we will have to pay the holder of the option the amount

max (S uuu (3) − K, 0) = max 100 × 1.23 − 105, 0 = 67.80.

Of course, if the stock goes down we will keep the amount 23.31 we have cashed when
selling the option.
To make our position safe we embark on hedging, which boils down to replicating
the option by means of a suitable strategy. However, we do not have to construct the
whole tree of replicating portfolios by working backwards from the payoff. This would
be simple in the case considered here, but for large trees, with many time instants, it
could become unduly complicated. Instead, we work step-by-step from the beginning,
following the path of actual stock changes and not bothering to analyse all scenarios.
Our first portfolio requires finding the call prices at time 1, for which we again use
the CRR formula: 
33.94
C(1) =
9.04
and then computing the delta coefficient, (that is, the ratio of one-step price changes
in option and stock) which, as we have seen in the single-step case, gives our stock
position x(1) in the replicating portfolio:
33.94 − 9.04
x(1) = = 0.83,
120 − 90
meaning that we have to buy this fraction of a share. This enables us to complete the
portfolio by taking
y(1) = 23.31 − 0.83 × 100 = −59.70,
which means borrowing this amount of money to have zero initial balance.
Now suppose the stock goes up. The new delta based on two option prices at time 2
is
48.55 − 14.91
xu (2) = = 0.93,
144 − 108
which means we have to purchase more stock (where we use our notational convention,
exceptionally adding the time variable for greater clarity). The corresponding money

71
market position is

(xu (2) − x(1)) S u (1)

y u (2) = y(1) − = −71.07.
1+R
Note that some texts express this as the cash value of the money market position, that
is, as (1 + R) y u (2).
Next, suppose the stock goes down, and we have
24.60 − 0
xud (3) = = 0.76,
129.60 − 97.20
meaning that it is time to sell the fraction 0.93 − 0.76 = 0.17 of a share, with the
result that our money market position becomes

y ud (3) = −55.45.

At the next time instant the option will be exercised. Suppose the stock goes up. We
have S udu = 129.60. We owe the option holder the amount 129.60 − 105 = 24.60. We
owe the bank (negative money market position) 55.45 × 1.13 = 73. 80. So, altogether
we have to pay 98. 40. We own 0.76 of a share worth 129.60, which we sell and get
0.76 × 129.60 = 98.40 as required to cover the payment.
It is easy to see that by following any other path and performing the required tasks
similarly, the cash balance will be zero after closing all positions irrespectively of the
path, which is no surprise since pricing was based on replication, and here we have just
tested the scheme.
This video shows delta hedging the option for a different scenario.

Remark 2.29
It is worth noting that the stock holding x(1) is not the coefficient at S(0) in the CRR
formula.

Example 2.30

In a the three-step binomial tree model, let S(0) = 100, U = 25

Example 2.31

%, D = −5%, and R = 5%, and consider the arithmetic Asian call option with

72
exercise time 3 and payoff
 +
S(1) + S(2) + S(3)
H(4) = −K ,
3

where K = 105 is the strike price.

This video shows how to compute the option prices at all nodes of the tree.

This video shows how to hedge the option along a given scenario (delta hedging).

In this video we find the entire replicating strategy (for all scenarios). Here is
the accompanying Excel file.

73
Chapter 3

Multi-step general models

Building on our experience with binomial models, we now formulate a general market model
with finite sample space and finitely many discrete time steps. Beginning with a single-stock
model, we review the definitions of basic concepts encountered for binomial models, such as
partitions of the sample space, conditional expectations with respect to partitions or random
variables, and filtrations and martingales, and derive some of their properties. We then
describe predictable self-financing trading strategies and formulate the No Arbitrage Principle
for general finite models. Finally, moving on to a multi-stock model, the two Fundamental
Theorems of Asset Pricing, describing the absence of arbitrage and market completeness
in terms of the existence and uniqueness of a risk-neutral probability, are proved for finite
multi-step models.

3.1 Partitions and conditioning

Assume given a finite sample space Ω = {ω1 , ω2 , . . . , ωM }, which remains fixed throughout
this chapter. Fix a probability P on Ω with P ({ω}) > 0 for all ω in Ω. We shall write
P
P(ω) in place of P ({ω}) for ease of notation. The additivity principle means that P (A) =
ω∈A P (ω) for any A ⊂ Ω, and we insist that P (Ω) = 1.

Definition 3.1
For any non-empty B ⊂ Ω, we define a new probability PB (ω) = P (ω)/P (B) if ω ∈ B
and 0 otherwise. Thus PB (B) = 1, and for any A ⊂ Ω,

P (A ∩ B)
PB (A) = .
P (B)

We call this the conditional probability of A given B and denote it by P (A|B).

74
 Definition 3.2
As in Definition 2.15, for any random variable X and any non-empty event B ⊂ Ω,
the conditional expectation of X given B is defined by
X
E(X|B) = X(ω)PB (ω).
ω∈Ω

Let 1B be the indicator of the set B. We have

X X P ({ω} ∩ B)
X(ω)PB (ω) = X(ω)
ω∈Ω ω∈Ω
P (B)
1 X
= X(ω)P (ω)
P (B) ω∈B
1 X
= X(ω)1B (ω)P (ω)
P (B) ω∈Ω
1
= E(1B X),
P (B)

so that
1
E(X|B) = E(1B X). (3.1)
P (B)
Recall Definition 2.13: a partition of Ω is a (finite) family P = {Bi } of pairwise disjoint
subsets of Ω whose union is Ω.
Definition 3.3
Applying the above construction to each Bi gives a random variable E(X|P) : Ω → R
with constant value
1
E(X|Bi ) = E(1Bi X)
P (Bi )
on Bi for each i. We say that E(X|P) is the conditional expectation of X with
respect to the partition P. Gluing these together, we have
k
X 1
E(X|P) = E(1Bi X)1Bi . (3.2)
i=1
P (Bi )

75
Remark 3.4
A partition P can be expanded to a field F by forming all possible (finite!) unions of
elements of P and adding the empty set:
( k )
[
F = {∅} ∪ Bj : B1 , . . . , Bk ∈ P .
j=1

This procedure can be reversed. The minimal non-empty elements of a field F are its
atoms, which together form a partition of Ω, yielding the original field by the above
expansion.

Example 3.5
Consider the probability space

Ω = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9}

with uniform probability

1
P (ω) = for each ω ∈ Ω,
10
a random variable X : Ω → R such that

X(ω) = ω 2 for each ω ∈ Ω,

and the partition P = {{0, 1, 2, 3, 4}, {5, 6}, {7, 8, 9}} .

In the video we find the conditional expectation E(X|P).

Definition 3.6
If X is an arbitrary random variable on Ω and Y is a random variable with possible
values y1 , y2 , . . . , yk such that the sets Bi = {Y = yi } are non-empty for all i, then
the conditional expectation E(X|Y ) of X given Y is the conditional expectation
of X with respect to the partition
PY = {B1 , . . . , Bk },
that is,
E(X|Y ) = E(X|PY )
for i = 1, 2, . . . , k .

76
 Remark 3.7
The idea is that we decompose Ω into pairwise disjoint parts Bi on which Y is constant.

It follows that
1
E(X|Y ) = E(X|Bi ) = E(1{Y =yi } X)
P ({Y = yi })
on Bi for each i = 1, 2, . . . , k.
Example 3.8
Let
Ω = {1, 2, 3, 4, 5, 6, 7, 8, 9}
with uniform probability
1
P (ω) = for each ω ∈ Ω,
9
and let X, Y : Ω → R be random variables such that

X(ω) = ω 2 and Y (ω) = (ω − 3)(ω − 7) for each ω ∈ Ω.

In the video we find the conditional expectation E(X|Y ).

Observe that if Y is constant on a subset of Ω, then E(X|Y ) is also constant on that

subset, and that the values of E(X|Y ) depend on the subsets on which Y is constant and
not on the actual values of Y . For random variables Y and Z defining the same partition,
we shall always have the same conditional expectations
E(X|Y ) = E(X|Z).

Example 3.9
The random variable Z : Ω → R defined on the same probability space as in Exam-
ple 3.8 by
Z(ω) = |ω − 5| for each ω ∈ Ω
defines the same partition as the random variable Y in Example 3.8,
PZ = PY .
Therefore
E(X|Y ) = E(X|Z).

77
3.2 Properties of conditional expectation
First, note that taking the conditional expectation is a linear operation.
Proposition 3.10
For any random variables X, Y , any numbers a, b ∈ R and any partition P, we have

E(aX + bY |P) = aE(X|P) + bE(Y |P).

Proof
This is evident from the definition and linearity of the usual expectation:
X 1
E(aX + bY |P) = E(1Bi (aX + bY ))1Bi
i
P (B i )
X 1 X 1
= a E(1Bi X)1Bi + b E(1Bi Y )1Bi
i
P (B i ) i
P (Bi )
= aE(X|P) + bE(Y |P).

where P = {Bi } .

Proposition 3.11 (tower property)

The expectation of the conditional expectation is the same as the expectation:

E(E(X|P)) = E(X). (3.3)

Proof

The video illustrates the idea behind this proof with a concrete choice of a
probability space Ω, random variable X, and partition P.

78
With P = {Bi }, we have
X
E(E(X|P)) = E(X|Bi )P (Bi ) definition of expectation
i
X 1
= E(X1Bi )P (Bi ) definition of conditional expectation
i
P (Bi )
!
X
= E X 1Bi additivity of expectation
i
X
= E(X) since P is a partition, so 1Bn = 1,
n

as required.

Remark 3.12
This result can be interpreted as follows: the average of averages is the same as the
overall average.

Property (3.3) can be generalised.

Example 3.13

First consider an example using the random variables S(1) and S(2) specified for the
stock prices in the binomial model. The partition P1 defined by S(1) consists of two
sets Bu , Bd , and the partition P2 determined by S(1) and S(2) of four Buu , Bud ,
Bdu , Bdd (as defined in the previous chapter). Notice that Bu = Buu ∪ Bud and
Bd = Bdu ∪ Bdd . We say that P2 is finer than P1 . In practice this means that S(2)
together with S(1) carry more information than just S(1).

Definition 3.14
Given two partitions P and P0 , we say P0 is finer than P if each element of P can be
represented as a union of some elements of P0 . Equivalently, we say that P is coarser
than P0 .

Proposition 3.15 (tower property)

Consider partitions P and P0 such that P0 is finer than P. Then the following formula
holds
E(E(X|P0 )|P) = E(X|P)
This is known as the tower property or the iterated expectations property.

79
Proof
Let P = {Bi } and P0 = {Cj } . Since P0 is finer than P, for each i, j we have either
Cj ⊂ Bi , in which case 1Cj 1Bi = 1Cj , or Cj ∩ Bi = ∅, in which case 1Cj 1Bi = 0. It
follows that for each i
!
X 1
E (E(X|P0 )1Bi ) = E E(1Cj X)1Cj 1Bi
j
P (Cj )
 
X 1
= E E(1Cj X)1Cj 
j:C ⊂B
P (C j )
j i
X 1

= E(1Cj X)E 1Cj
j:Cj ⊂Bi
P (Cj )
  
X X
= E(1Cj X) = E  1Cj  X  = E(1Bi X).
j:Cj ⊂Bi j:Cj ⊂Bi

The above calculation resembles (and extends) the proof of Proposition 3.11. It follows
that
X 1
E(E(X|P0 )|P) = E(E(X|P0 )1Bi )1Bi
i
P (Bi )
X 1
= E(1Bi X)1Bi = E(X|P).
i
P (Bi )

Proposition 3.16
Suppose that P is a finer partition than the partition PY defined by a random variable Y .
Then
E(Y X|P) = Y E(X|P).

Remark 3.17
This property is often referred to as ‘taking out what is known’. The partition P
being a finer than PY means that if we know which event in P has occurred, then we
know the value of Y .

80
 Proof
For any event Bi in P the random variable Y is constant on Bi . Hence E(Y X1Bi )1Bi =
E(X1Bi )Y 1Bi , so
X 1
E(Y X|P) = E(Y X1Bi )1Bi
i
P (Bi )
X 1
= Y E(X1Bi )1Bi = Y E(X|P).
i
P (B i )

Proposition 3.18
If X and Y are independent random variables, then

E(X|Y ) = E(X) and E(Y |X) = E(Y ).

Proof
Let PY = {Bi } be the partition generated by Y . Then the random variables X and 1Bi
are independent, so E(X1Bi ) = E(X)E(1Bi ) for each i. It follows that
X 1 X 1
E(X|Y ) = E(X1Bi )1Bi = E(X)E(1Bi )1Bi
i
P (Bi ) i
P (Bi )
X 1 X
= E(X) E(1Bi )1Bi = E(X) 1Bi = E(X),
i
P (Bi ) i

with the same proof for the second identity.

3.3 Filtrations and martingales

Example 3.19

Consider two random variables X and Y , where X takes values {x1 , . . . , xk } and Y
takes values {y1 , . . . , yl }. First construct the partition PX defined by X, then decom-
pose each element Bi = {X = xi } according to the values taken by Y on Bi :
l
[
Bi = Bij , Bij = Bi ∩ {Y = yj }.
j=1

This defines a partition of Ω, denoted by PX,Y = {Bij }, which is finer than PX .

81
 We can proceed in this way, introducing more random variables. Let X1 , . . . , XN
be a sequence of random variables and denote by Pk the partition defined by
X1 , . . . , Xk .f or any k = 1, . . . , N . We also take P0 = {Ω} to be the trivial par-
tition. Thus, we have a sequence of partitions P0 , P1 , . . . , PN with Pk finer than Pk−1
for each k = 1, . . . , N .
The tower property has the following consequence: for any j, k = 0, 1, . . . , N
such that j < k,
E(E(X|Pk )|Pj ) = E(X|Pj )
so if we denote Y (i) = E(X|Pi ) for i = 0, . . . , N , then

E(Y (k)|Pj ) = Y (j)

for any j, k = 0, 1, . . . , N such that j < k. This sequence gives an example of a

martingale, generalising what we observed in the binomial setting.

This motivates the general definition a martingale with respect of a filtration. In applica-
tions the role of a martingale will usually be played by the stock prices S(n) or by the value
process V (n) of a given trading strategy. In our applications the time horizon N will remain
finite.
Definition 3.20
A sequence of partitions P0 , . . . , PN of Ω such that Pk is finer than Pk−1 for each
k = 1, . . . , N is called a filtration.

Definition 3.21
A sequence M (0), . . . , M (N ) of random variables is a martingale with respect to a
filtration P0 , . . . , PN if
E(M (k)|Pj ) = M (j)
for any j, k = 0, 1, . . . , N such that j < k.

In terms of ‘information’ carried by a filtration, this property can be read as follows: given
some (partial) information, the expected value of some quantity (price) in the future is the
value of the quantity at that time. In other words, the values M (n) encapsulate all the
predictive power of the information available at time n.
If M (n) is a martingale with respect to Pn then Pn is a finer partition than PM (n)
since M (n) is the conditional expectation with respect to Pn . We then say that M (n) is
Pn -measurable.

82
 Proposition 3.22

If E(M (n + 1)|Pn ) = M (n) for each n = 0, . . . , N − 1, then M is a martingale.

Proof
This follows from the tower property. Let m > n. Then

E(M (m)|Pn ) = E(E(M (m)|Pm−1 )|Pn )

= E(M (m − 1)|Pn ) = . . . = E(M (n + 1)|Pn ) = M (n)

after m − n steps.

Proposition 3.23
Let X be a random variable and let P0 , . . . , PN be a filtration. Then the sequence

E(X|P0 ), E(X|P1 ), E(X|P2 ), . . . , E(X|PN )

is a martingale.

Proof
For the sequence E(X|Pn ) to be a martingale, it must satisfy the condition

E(E(X|Pm )|Pn ) = E(X|Pn ),

for n < m, which holds by the tower property because Pm is a finer partition than Pn .

A martingale is a mathematical model of a fair game of chance: on average we expect

our wealth to remain the same after each round of the game as the following proposition
shows. It says that constant expectation is a necessary condition for a martingale.
Proposition 3.24

If M (n), n = 0, 1, . . . , N is a martingale (with respect to a filtration Pn , n = 0, . . . , N ,

then
E(M (0)) = E(M (1)) = · · · = E(M (N )).

Proof
If M (0), M (1), . . . , M (N ) is a martingale, then M (n) = E(M (n+1)|Pn ) for each n =
0, . . . , N − 1. Applying expectation to both sides of this equality, we have
E(M (n)) = E(E(M (n + 1)|Pn )) = E(M (n + 1))

83
 for each n, as required.

Example 3.25

This video shows an example where the reverse implication does not hold, hence
E(M (0)) = E(M (1)) = · · · = E(M (N )) does not necessarily mean that
M (0), M (1), . . . , M (N ) is a martingale.

Definition 3.26
We say that a sequence M (0), M (1), . . . , M (N ) of random variables is a submartin-
gale (respectively, a supermartingale) with respect to a filtration P0 , P1 , . . . , PN if
the partition of M (n) is coarser than Pn and

E(M (n + 1)|Pn ) ≥ M (n) (respectively, ≤ M (n))

for each n = 0, 1, . . . , N − 1.

The condition that the partition of M (n) should be coarser than Pn , i.e. that M (n)
is measurable with respect to Pn , simply means that M (n) is known at time n. The
partition Pn represents our knowledge at time n, and if Pn is known, then M (n) is known.
In defining martingales we did not need to state explicitly the condition that the partition
of M (n) should be coarser than Pn because it follows directly from the identity E(M (n +
1)|Pn ) = M (n).
Definition 3.27
We can also define an increasing sequence of fields (Fn )n≤N , again called a filtration,
where F0 = {∅, Ω} and for n ≥ 1, the field Fn is generated by the first n elements
of the sequence, that is, by M (1), . . . , M (n). The field generated by a single random
variable X is of the form FX = {X −1 (B) : B ⊂ R, B a Borel set}, and we take Fn
to be the smallest field containing FM (k) for all k = 1, . . . , n. In this way we ensure
that (M (n))n≤N is adapted to (Fn )n≤N , which means that for all n, M (n) is Fn -
measurable.

In applications, where the sequence of random variables in question is often the sequence
{S(n) : n = 0, 1, . . . , N } of stock prices, it will be convenient to assume that FN = F. The
latter is 2Ω , so to ensure equality we have to reduce Ω if the atoms of FN are not single-
element sets by picking one element in each to form the new sample space, again denoted
by Ω for simplicity.

84
3.4 Trading strategies and arbitrage
We return to the case when the martingales are related to stock prices. Assume that
S(0), S(1), . . . , S(N ) is a sequence of stock prices for all trading dates in some pricing
model with time horizon N. By a (trading) strategy we mean a sequence of portfolios. At
each step our decision on how to adjust the previous portfolio depends on the information
available at this time. We are ready for the general definitions.
Definition 3.28
A (trading) strategy is a sequence of pairs of random variables showing, for n =
1, 2, . . . , N, the number of shares x(n) of stock and the number y(n) of money market
account units chosen at time n − 1 and held until time n, and such that (x(n), y(n)) is
Fn−1 -measurable, for each n = 1, 2, . . . , N . In general, a sequence of random variables
satisfying this measurability requirement is called predictable.

Definition 3.29
The value process of a strategy is a sequence defined for n = 1, . . . , N by

V(x,y) (n) = x(n)S(n) + y(n)A(n),

together with the initial investment

V(x,y) (0) = x(1)S(0) + y(1)A(0).

The next definition summarises the requirement that the funds available at a given time
are used for building the next portfolio, that is, no inflows or outflows of funds are allowed.
Definition 3.30
A strategy (x, y) is self-financing whenever

V(x,y) (n) = x(n + 1)S(n) + y(n + 1)A(n)

for all n = 1, . . . , N − 1.

For any sequence X(n), denote its increment by

∆X(n) = X(n) − X(n − 1)

for n = 1, . . . , N. We can write the self-financing condition in the following equivalent form,
which will be useful later.

85
 Proposition 3.31
A strategy is self-financing if and only if

∆V(x,y) (n) = x(n)∆S(n) + y(n)∆A(n)

for each n = 1, . . . , N .

Proof
This is immediate since the strategy is self-financing if an only if V(x,y) (n − 1) =
x(n)S(n − 1) + y(n)A(n − 1) for each n = 2, . . . , N .

Given the initial wealth, all that matters are the decisions concerning the number of shares
held, with the money market position following as a consequence of self-financing.
Proposition 3.32

Given a predictable sequence x(n), n = 1, . . . , N, and initial wealth V (0), there is a

unique predictable sequence y(n), n = 1, . . . , N, such that (x, y) is a self-financing
strategy and V (0) is its initial value, that is, V (0) = V(x,y) (0).

Proof
Given V (0) and x(1), let
1
y(1) = (V (0) − x(1)S(0)).
A(0)
Then, for each n = 1, . . . , N − 1, let
1
y(n + 1) = y(n) + (x(n) − x(n + 1))S(n).
A(n)
Observe that y(n + 1) is Fn -measurable since all the ingredients on the right are.
Moreover,

x(n + 1)S(n) + y(n + 1)A(n)

 
1
= x(n + 1)S(n) + y(n) + (x(n) − x(n + 1))S(n) A(n)
A(n)
= x(n)S(n) + y(n)A(n) = V(x,y) (n)

for each n = 1, . . . , N − 1, hence (x, y) is a self-financing strategy. Finally,

V (0) = x(1)S(0) + y(1)A(0) = V(x,y) (0).

86
 Example 3.33
In a 3-step binomial three model with

U = 20%, D = −10%, R = 10%

and
S(0) = 100, A(0) = 1
take the predictable process

x(1) = 4.5 xu (2) = −1.2 xuu (3) = 3.3

xud (3) = −2.5
xd (2) = 2.6 xdu (3) = −1.3
xdd (3) = 0.5

and let
V (0) = 10.5.

This video shows how to construct the unique predictable process y(1), y(2), y(3)
such that (x, y) is a self-financing strategy with

V (0) = V(x,y) (0).

Next, we focus on a key property of the discounted values of a strategy. For the multi-
step binomial model, we described in the previous chapter how to construct a probability Q
on the sample space Ω built for that model such that the discounted stock prices form a
Q-martingale. This provides the basis for the assumption made in the following theorem.
In the next section we shall show that this assumption is equivalent to the No Arbitrage
Principle for general finite models.
Theorem 3.34
Suppose that there is a probability Q such that EQ (S̃(n + 1)|Fn ) = S̃(n) for all
n = 0, 1, . . . , N − 1. Then the discounted values of a self-financing strategy form a
martingale with respect to Q.

Proof
By definition,
1
Ve(x,y) (n) = V(x,y) (n).
A(n)

87
 Since S(n)
e = S(n)/A(n) is a martingale, that is, EQ (S(n
e + 1)|Fn ) = S(n)
e for each
n = 0, 1, . . . , N − 1, we can proceed as follows:

EQ (Ve(x,y) (n + 1)|Fn ) = EQ (x(n + 1)S̃(n + 1) + y(n + 1)|Fn )

by the definition of the value of a strategy
= x(n + 1)EQ (S̃(n + 1))|Fn ) + y(n + 1)
since the strategy is predictable
= x(n + 1)S̃(n) + y(n + 1)
since S(n)
e is a martingale
1
= (x(n + 1)S(n) + y(n + 1)A(n))
A(n)
1
= (x(n)S(n) + y(n)A(n))
A(n)
by the self-financing condition
= Ve(x,y) (n).

Let us emphasise that this result does not use any particular model of stock prices. We only
used the martingale property of the discounted stock prices together with the self-financing
and predictability conditions.
We conclude the section by giving a general multi-step version of the No Arbitrage Prin-
ciple. First we define what is meant by arbitrage opportunities in our multi-step single-stock
model based on a finite sample space Ω.
Definition 3.35
A predictable strategy (x(n), y(n)), n = 0, 1, . . . , N, is an arbitrage opportunity if

(1) V(x,y) (0) = 0,

(2) (x, y) is a self-financing strategy

(3) V(x,y) (n) ≥ 0 with probability 1 for all n = 1, . . . , N ,

(4) there is an n = 1, . . . , N such that V(x,y) (n) > 0 with positive probability.

Assumption 3.36 (No Arbitrage Principle)

We assume that arbitrage opportunities do not exist in the market model.

88
 Remark 3.37
As we have seen before, for extended markets the definition of arbitrage strategy has
to be modified to include positions in the relevant derivative securities.

The next proposition provides an apparently weaker, but in fact equivalent, version of the
notion of arbitrage, which is useful in some applications.
Proposition 3.38

If a market model admits a predictable strategy (x(n), y(n)), n = 0, 1, . . . , N such

that

(1) V(x,y) (0) = 0,

(2) (x, y) is a self-financing strategy,

(3) V(x,y) (N ) ≥ 0 with probability 1,

(4) V(x,y) (N ) > 0 with positive probability,

then it also admits an arbitrage opportunity (as in Definition 3.35).

Proof

This video illustrates the main idea of the proof in the simple case when N = 2.

Suppose we have a strategy (x, y) satisfying the assumptions of the proposition, and
suppose that it does not satisfy Definition 3.35, so its value is negative, V(x,y) (n) < 0
on B for some n = 1, . . . , N and some B ∈ Pn . Suppose that n is the largest
integer with this property. Clearly, n < N by condition (3) in the Proposition. In any
scenario ω ∈ B the strategy (x, y) generates some profit at time n + 1 since the value
becomes non-negative after being negative at n. We shall exploit this to construct an
arbitrage opportunity in the sense of Definition 3.35.
We construct a new strategy (x∗ , y ∗ ) as follows. We begin by constructing x∗ .
We do not invest in stock on any trading date other than n:

x∗ (k) = 0 for all k = 1, . . . , n and k = n + 2, . . . , N.

At time n we employ the stock position in original strategy for any scenario in B and
do not invest in stock in scenarios outside B:

∗ 0 on Ω \ B.
x (n + 1) =
x(n + 1) on B.

89
Clearly, the process x∗ (k), k = 1, . . . , N, is predictable. We know (Proposition 3.32)
that there is a unique predictable sequence y ∗ (k), k = 1, . . . , N, such that (x∗ , y ∗ ) is
self-financing and V(x∗ ,y∗ ) (0) = V(x,y) (0) = 0. We have
y ∗ (1) = Ṽ(x,y) (0) − x(1)S̃(0)
and
y ∗ (k) = y ∗ (k − 1) + (x∗ (k − 1) − x∗ (k))S̃(k − 1) for each k = 2, . . . , N.
It follows that
y ∗ (k) = 0 for each k = 1, . . . , n.
Next,
y ∗ (n + 1) = 0 on Ω \ B
and
y ∗ (n + 1) = y ∗ (n) + (x∗ (n) − x∗ (n + 1))S̃(n)
= −x(n + 1)S̃(n) on B,
so
V(x∗ ,y∗ ) (n + 1) = 0 on Ω \ B
and
Ṽ(x∗ ,y∗ ) (n + 1) = x∗ (n + 1)S̃(n + 1) + y ∗ (n + 1)
= x(n + 1)S̃(n + 1) − x(n + 1)S̃(n)
= Ṽ(x,y) (n + 1) − y(n + 1) − x(n + 1)S̃(n) on B.
From the self-financing condition we have
y(n + 1) + x(n + 1)S̃(n) = V(x,y) (n).
Inserting this in the above expression, we further obtain
Ṽ(x∗ ,y∗ ) (n + 1) = Ṽ(x,y) (n + 1) − Ṽ(x,y) (n) > 0 on B
since V(x,y) (n + 1) ≥ 0 and V(x,y) (n) < 0 on B.
Next, for each k = n + 2, . . . , N , we have x∗ (k) = 0, that is, all wealth in the
strategy is in the money market account at these time steps. This implies that
Ṽ(x∗ ,y∗ ) (N ) = Ṽ(x∗ ,y∗ ) (n + 1).
Since we have shown that Ṽ(x∗ ,y∗ ) (n − 1) ≥ 0 with probability 1 and Ṽ(x∗ ,y∗ ) (n − 1) > 0
with positive probability, this proves that (x∗ , y ∗ ) is an arbitrage opportunity (in the
sense of Definition 3.35). By the No Arbitrage Principle, this is impossible, completing
the proof.

90
3.5 General multi-step model
The multi-step models we have discussed so far all involve a market with a single underlying
stock. We now analyse derivative securities defined on an underlying market containing several
risky assets (stocks) and a single money market account. As we saw for trinomial models,
where market completeness requires two underlying stocks, pricing and hedging derivatives in
such a model proceeds much like in the single-stock case, and it is often useful to treat the
vector of stock prices as a single entity. With this proviso, much of theory we have developed
so far will generalise to this more general setting quite simply, and we begin by reviewing the
main concepts.
We still assume that time runs in a discrete way, just as in the binomial model. With h
being the length of the time step, we use n = 0, 1, . . . , N to indicate the time instant nh.
The basic market consists of the money market account and d risky assets (stocks) with prices
S1 (n), . . . , Sd (n), which comprise the components of the Rd -valued random variable S(n)
and are discrete strictly positive random variables defined on the finite sample space Ω =
{ω1 , . . . , ωM }. The dynamics of the single risk-free asset (the money market account) is again
determined by the risk-free rate R, which is assumed constant, so A(n) = A(0)(1 + R)n . A
portfolio is a vector (x, y) = (x1 , . . . , xd , y) with coordinates representing positions in the
corresponding securities. A strategy is a sequence of portfolios
(x(n), y(n)) = (x1 (n), . . . , xd (n), y(n)), n = 1, . . . , N.
These are Rd+1 -valued random vectors, except for the initial portfolio, which is determined
at time 0, so is a deterministic vector. We shall write (x, y) for this sequence.
Later, this market will be extended by adding derivative securities, and the portfolio
vectors will contain additional coordinates zk , reflecting the positions in these derivatives.
The value of a strategy at time n is
d
X
V(x,y) (n) = xj (n)Sj (n) + y(n)A(n), for n = 1, 2, . . . , N,
j=1
d
X
V(x,y) (0) = xj (1)Sj (0) + y(1)A(0).
j=1

We shall again assume throughout that the strategies we consider are predictable, so
the random vectors (x(n + 1), y(n + 1)) are Fn -measurable. This condition reflects the
convention that the portfolio at time n + 1 is constructed at time n using the funds and the
information about stock prices available at that time.
The key ideas developed for the single-stock model also apply in this more general model.
A strategy is self-financing if
d
X
V(x,y) (n) = xj (n + 1)Sj (n) + y(n + 1)A(n) for n = 0, 1, . . . , N − 1.
j=1

91
 We state the version of the No Arbitrage Principle which applies in the model as follows.
Definition 3.39
A predictable strategy (x, y) is an arbitrage opportunity in the underlying market if
its value process satisfies

(1) V(x,y) (0) = 0,

(2) (x, y) is a self-financing strategy

(3) V(x,y) (n) ≥ 0 with probability 1 for all n,

(4) there is an n = 1, . . . , N such that V(x,y) (n) > 0 with positive probability.

For the market extended by adding some derivative securities, the definition is similar;
cf. Definitions 1.8 and 1.61.
Assumption 3.40 (No Arbitrage Principle)

Arbitrage opportunities do not exist in any market model.

As a consequence, we prove the following property, known as the law of one price.
Theorem 3.41 (Law of One Price)

If H, H 0 are two securities with the same value at time N,

H(N ) = H 0 (N ),

then for each n = 0, 1, . . . , N their values agree:

H(n) = H 0 (n).

Proof
This is routine: if for one ω0 , at some time n0 we have H(n0 , ω0 ) < H 0 (n0 , ω0 ), say,
then we do nothing in scenarios other than ω0 , and in scenario ω0 do nothing up to
time n0 , and then take the position xH (n0 , ω0 ) = 1, xH 0 (n0 , ω0 ) = −1 in the two
securities, and y(n0 , ω0 ) = H 0 (n0 , ω0 ) − H(n0 , ω0 ) in the money market to obtain an
arbitrage opportunity.

In particular, put-call parity can be obtained as a consequence of the law of one price.

92
 Proposition 3.42 (Put-call parity)

For any n = 0, 1, . . . , N, we have the following relation between the call and put values

C(n) − P (n) = S(n) − K(1 + R)−(N −n) .

Proof
Since the call and put payoffs are

C(N ) = max {S(N ) − K, 0} ,

P (N ) = max {K − S(N ), 0} ,

It follows that
C(N ) − P (N ) = S(N ) − K.
The law of one price implies that

C(n) − P (n) = S(n) − K(1 + R)−(N −n) .

Remark 3.43
Note that put-call parity does not require using any particular model and can be justified
directly by the No Arbitrage Principle via the Law of One Price.

In the First Fundamental Theorem below we shall prove that the No Arbitrage Principle is
equivalent to the existence of a martingale probability, that is, a probability Q with Q(ω) > 0
for all ω in Ω and such that for each j = 1, . . . , d, the discounted stock price process
{Sej (n) : n = 0, 1, . . . , N } is a Q-martingale.
The following definition is both intuitive and useful.
Definition 3.44
The gains process G(x,y) = (G(x,y) (n))n ≤ N generated by the strategy (x, y) is defined
by

G(x,y) (0) = 0,
n X
X d 
G(x,y) (n) = xj (k)∆Sj (k) + y(k)∆A(k) for n = 1, . . . , N.
k=1 j=1

En equivalent condition for self financing can be formulated in terms of the gains process.

93
 Proposition 3.45

The strategy (x, y) is self-financing if and only if

∆V(x,y) (n) = ∆G(x,y) (n)

for each n = 1, . . . , N.

Proof
The proof is virtually the same as that of Proposition 3.31.

Remark 3.46
Observe that
G(x,y) (n) = V(x,y) (n) − V(x,y) (0)
for any self-financing strategy (x, y).

In self-financing predictable strategies the risk-free positions are determined by the stock
positions and the initial value of the strategy, as already observed in the case of a single
stock, see Proposition 3.32.
Theorem 3.47
Given a predictable sequence x(n) = (x1 (n), . . . , xd (n)), n = 1, . . . , N, and initial
wealth V (0), there is a unique predictable sequence y(n), n = 1, . . . , N, such that
(x, y) is a self-financing strategy and V(x,y) (0) = V (0).

Proof
This is proved by induction. First, let

V (0) − x1 (1)S1 (0) − · · · − xd (1)Sd (0)

y(1) = ,
A(0)

which ensures that V(x,y) (0) = V (0). Thus y(1) is non-random, as is necessary for
y(n) to be predictable. We have therefore determined V(x,y) (1) uniquely,

V(x,y) (1) = x1 (1)S1 (1) + · · · + xd (1)Sd (1) + y(1)A(1),

and the initial induction step is complete. Next, for the induction step, suppose that
Fn−1 -measurable y(n) has already been constructed for some n = 1, . . . , N − 1. This
gives
V(x,y) (n) = x1 (n)S1 (n) + · · · + xd (n)Sd (n) + y(n)A(n),

94
 which is Fn -measurable (since so are S1 (n), . . . , Sd (n)). We define y(n + 1) in such a
way that the self-financing condition is guaranteed:

V(x,y) (n) − x1 (n + 1)S1 (n) − · · · − xd (n + 1)Sd (n)

y(n + 1) = .
A(n)

Then y(n + 1) is Fn -measurable since for j = 1, 2, . . . , d both the xj (n + 1) and Sj (n)

are. This completes the induction step.

Recall the definition of discounted prices. For any asset X we put

X(i)
X(i)
e = .
(1 + R)i
Then the successive differences are
X(i) X(i − 1)
∆X(i)
e =
i
− .
(1 + R) (1 + R)i−1
In particular we obtain
A(i)
A(i)
e = = A(0), hence ∆A(i)
e = 0.
(1 + R)i

Definition 3.48
The discounted gains process will be denoted by
n X
X d n
X
G
ex (n) = xj (i)∆Sej (i) = hx(i), ∆S(i)i,
e
i=1 j=1 i=1

where we use the notation ha, bi = dj=1 aj bj for the usual inner product of vectors
P

in Rd . The vector ∆S(i)

e has the discounted price increments ∆Sej (i) as its components.

As the discounted increment in A is zero, there is now no dependence on y, only on x.

We now verify that for self-financing strategies the sum of the initial investment and
discounted gains gives the discounted value process.
Theorem 3.49
If a strategy (x, y) is self-financing then

Ve(x,y) (n) = V(x,y) (0) + G

ex (n).

95
 Proof
This is an immediate consequence of Proposition 3.45.

Definition 3.50
A strictly positive adapted process is called a numeraire.

Example 3.51

For example, A(n) = (1 + R)n is a numeraire.

A numeraire can be regarded as the price process of a benchmark security against which
the price changes of other securities are compared.
Proposition 3.52

Given an adapted sequence Z(n) > 0 for n = 0, 1, . . . , N (that is, a numeraire), with
(x, y) and (S, A) as above, the strategy (x,y) is self-financing for the price process
(S, A) if and only if it is self-financing for the price process (ZS1 , . . . , ZSd , ZA).

Proof
Write V(x,y) (n) = hx(n), S(n)i , where, by abuse of notation, we write
S = (S1 , ..., Sd , A) and x = (x1 , ..., xd , y) as vectors in Rd+1 . Dropping the subscript
(x, y) for simplicity, the self-financing condition becomes

∆V (n) = hx(n), ∆S(n)i

for each n = 1, . . . , N . This is the same as saying that

h∆x(n), S(n − 1)i = 0

for each n = 1, . . . , N since

∆V (n) = hx(n), S(n)i − hx(n − 1), S(n − 1)i

= hx(n), S(n) − S(n − 1)i + hx(n) − x(n − 1), S(n − 1)i
= hx(n), ∆S(n)i + h∆x(n), S(n − 1)i .

Changing S to ZS = (ZS1 , . . . , ZSd , ZA), we obtain

h∆x(n), Z(n − 1)S(n − 1)i = Z(n − 1) h∆x(n), S(n − 1)i = 0.

Conversely, if h∆x(n), Z(n − 1)S(n − 1)i = 0, then h∆x(n), S(n − 1)i = 0.

96
 The last proposition means that the self-financing property of trading strategies is invari-
ant under a change of numeraire. It also shows that the converse of Theorem 3.49 holds.
Indeed, if a strategy is self-financing for the discounted prices, then it is also self-financing for
the original prices since the two price processes differ by multiplication by a strictly positive
adapted process (in fact, a deterministic one).
To discuss martingale properties in general we introduce a filtration of fields Fn as before.
Then we have the familiar definition.
Definition 3.53
A probability Q is called risk-neutral (or a martingale probability) if the discounted
stock prices are martingales

EQ (Sej (n + 1)|Fn ) = Sej (n) all j,

where for each n, Fn is generated by Sj (k), k ≤ n, j = 1, . . . , d (i.e. the smallest field

for which all Sj (k), k ≤ n, are measurable).

We shall prove that under a risk-neutral probability Ve(x,y) (n) is a martingale as well, thus
generalising Theorem 3.34. To this end we formulate the appropriate theorem using general
notation since it can be used in various applications.
Theorem 3.54
If Mj (n) are martingales and Hj (n), j = 1, . . . , d, are predictable processes, then the
process X(n) defined by

X d
n X n
X
X(n) = X(0) + Hj (k)∆Mj (k) = X(0) + hH(k), ∆M(k)i,
k=1 j=1 k=1

where X(0) is an arbitrary real number, is a martingale.

This right-hand side of the above equation is sometimes called a martingale transform
or discrete stochastic integral.
Proof
First note that
d
X
X(n + 1) − X(n) = Hj (n + 1)∆Mj (n + 1)
j=1
d
X
= Hj (n + 1) [Mj (n + 1) − Mj (n)] .
j=1

97
The theorem is formulated for any probability space, hence we use the general symbol
for expectation. We take the conditional expectation of the above equality to get

E(X(n + 1) − X(n)|Fn )
X d 
= E Hj (n + 1) [Mj (n + 1) − Mj (n)] Fn
j=1
d
X  
= E Hj (n + 1) [Mj (n + 1) − Mj (n)] Fn
j=1
d
X
= Hj (n + 1)E(Mj (n + 1) − Mj (n)|Fn )
j=1
= 0,

where we have used the property of conditional expectations referred to as ‘taking out
what is known’. The process H is predictable, so the value at time n + 1 is measurable
with respect to Fn . The final result E(X(n + 1) − X(n)|Fn ) = 0 proves that X is a
martingale.

We also need the following result.

Theorem 3.55
For M (n) to be a martingale it is sufficient that
N
X 
E H(n)∆M (n)) = 0
n=1

for each predictable process H(n).

Proof
Fix an k = 0, . . . , N − 1, take any A ∈ Fk , and put

1A for n = k + 1,
H(n) =
0 otherwise.

Then the sum reduces to just one term

N
X
H(n)∆M (n) = 1A ∆M (k + 1).
n=1

98
 By assumption,
N
X 
E H(n)∆M (n) = E(1A ∆M (k + 1))
n=1
= E(1A [M (k + 1) − M (k)]) = 0.

Since this holds for an arbitrary A ∈ Fk , it follows that

E(M (k + 1)|Fk ) = M (k).

3.6 The Fundamental Theorems of Asset Pricing

We remain with the multi-step, multi-stock model described above, and prove the two key
results for arbitrage (or risk-neutral) pricing of European derivative securities in a discrete-
time setting, linking the absence of arbitrage to the existence and uniqueness of risk-neutral
probability.
We begin by distinguishing between two basic types of European derivative securities,
examples of which we have already encountered.
Definition 3.56
A European derivative security confers the right to obtain at time N a payoff H(N )
of a form specified at the initial trading date 0. We say that H is

ˆ path independent if H(N ) = h(S(N )) for some h : Rd → R,

ˆ path dependent if H(N ) = h(S(0), S(1), . . . , S(N )), where the market model
has time set T = 0, 1, . . . , N , and h : R(N +1)d → R.

(We do not need impose any regularity on h since F = 2Ω .)

Call and put options are clearly path independent.

Throughout this section we assume that the filtration (Fn )N
n=0 generated by the stock
prices satisfies FN = F = 2Ω .
Definition 3.57
We say that two probabilities P1 , P2 are equivalent (written as P1 ∼ P2 ) if they have
the same sets of probability zero.

Recall that the sample space Ω = {ω1 , . . . , ωM } is finite, we assume that P is defined
on F = 2Ω , and that we take P (ω) > 0 for all ω. In this setting, to ensure the equivalence

99
P ∼ Q, any risk-neutral probability must have Q(ω) > 0 for all ω. It is worth mentioning
that Q is defined on the field 2Ω due to the standing assumption that FN = F.

First Fundamental Theorem

Theorem 3.58
The absence of arbitrage is equivalent to the existence of a risk-neutral probability Q
equivalent to P such that all discounted price processes Sej (n), j = 1, . . . , d, are
martingales with respect to Q.

Proof
Assume first the existence of a risk-neutral probability Q. Take any predictable strategy
(x, y). By Theorem 3.54, its discounted value process Ve(x,y) is a martingale under Q.
Suppose that V(x,y) (0) = 0 and V(x,y) (n) ≥ 0 for each n, which are necessary conditions
for an arbitrage strategy. Then by the martingale property of Ve(x,y) under Q, for each n
we have
EQ (Ve(x,y) (n)) = EQ (Ve(x,y) (0)) = EQ (V(x,y) (0)) = 0,
so after multiplying both sides by (1 + R)n
X
EQ (V(x,y) (n)) = Q(ω)V(x,y) (n, ω) = 0.
ω∈Ω

Since all terms in the sum are non-negative and Q(ω) > 0 for each ω ∈ Ω, it follows
that V(x,y) (n, ω) = 0 for each ω ∈ Ω, and this shows that it is impossible to construct
an arbitrage opportunity.
For the converse implication, assume the absence of arbitrage. Let x =
{(x1 (n), . . . , xd (n)) : n = 1, . . . , N } be a predictable sequence. The probability space
Ω = {ω1 , . . . , ωM } is finite, so a random variable can be regarded as an element of
the Euclidean vector space RM . This means that the discounted gains process G ex has
the form
Gex (n) = (G ex (n, ωM )) ∈ RM
ex (n, ω1 ), . . . , G
for each n, and
ex (N ) : x is predictable} ⊂ RM
W = {G
is a vector subspace of RM since it is clearly closed under linear combinations.
We claim that if G ex (N, ω) ≥ 0 for all ω, then G ex (N, ω) = 0 for all ω. Take
V (0) = 0 and complete x with y(n) so that (x, y) is a self-financing predictable
strategy such that Ve(x,y) (0) = V (0) (see Theorem 3.47). By Theorem 3.49, Ve(x,y) (n) =
V (0) + Gex (n), and so G ex (N, ω) = 0 for all ω, or else there would be an arbitrage
opportunity by Proposition 3.38.

100
 To conclude the proof of the theorem, recall that we are working under the No
Arbitrage Principle and that we have identified W as a vector subspace of RM . We
can apply the Separation Lemma (Lemma 1.56), which says that if A ⊂ RM is convex
and compact and W is a vector subspace of RM disjoint from A, then there exists a
z = (z1 , . . . , zM ) ∈ RM such that hz, ai > 0 for all a ∈ A, and hz, wi = 0 for all
w ∈ W (with h·, ·i denoting the Euclidean inner product in RM ). For our application
we take
n M
X o
A = a = (q1 , . . . , qM ) : qi = 1, qi ≥ 0
i=1

and the above W. Define

zi
Q(ωi ) = PM > 0.
i=1 zi

Using the definition of the inner product in RM , we obtain

M
X
hz, wi = zi G
ex (N, ωi ) = 0
i=1

so
M
X
Q(ωi )G
ex (N, ωi ) = EQ (G
ex (N )) = 0.
i=1

But the stock holdings x were taken as an arbitrary predictable (vector-valued) process,
and we can fix j = 1, . . . , d and take x(n) = (0, . . . , 0, xj (n), 0 . . . , 0) for each n =
1, . . . , N, where xj (n) is a real-valued predictable process. Then the above identity
reduces to X N 
EQ xj (n)∆Sj (n) = 0.
e
n=1

By Theorem 3.55, the Sej (n) are martingales for j = 1, . . . d since the xj (n) are arbitrary
predictable processes.

101
 Remark 3.59
The First Fundamental Theorem describes the ‘economic’ no-arbitrage axiom in purely
mathematical terms, namely the existence of an equivalent martingale measure. Our
proof contains the crucial additional assumption that the sample space Ω has finitely
many elements (and we can therefore give each of them positive probabilities). This
naturally begs the question whether this restriction can be removed. In discrete-time
models this does turn out to be the case, although the proof uses considerably more
advanced concepts in functional analysis, which are beyond our scope (see R. J. Elliott
and P. E. Kopp, Mathematics of Financial Markets, Springer 2005).

Even in the relatively simple setting of a finite market model, the above theorem does not
tell us whether or not the risk-neutral probability Q is unique. In fact, we have already seen
in the one-step trinomial model that this need not be the case. Our next result characterises
the pricing models with unique risk-neutral probability.

Second Fundamental Theorem

We formulate the main principle of pricing by arbitrage. It is based on the No Arbitrage
Principle but takes into account the fact that the notion of arbitrage is relative to the market
since the notion of portfolio depends on the choice of the securities traded.
Definition 3.60
The payoff of a derivative security is a random variable H measurable with respect to
the field FN generated by the underlying security prices up to time N. The prices of
H form a sequence H(n) such that H(N ) = H and the market extended by H is free
of arbitrage. (Recall that we assume FN = 2Ω by adjusting Ω if necessary.)

The basic market of underlying securities consists of d stocks and the risk-free asset.
The extended market includes some specified additional derivative securities.
Definition 3.61
A derivative security can be replicated if there exists a self-financing predictable strat-
egy (x, y) such that
H(N ) = V(x,y) (N ).

Definition 3.62
The market is complete if every derivative security can be replicated.

We have exactly the same result as in the single-step case.

102
Theorem 3.63
An arbitrage-free market model is complete if and only if there is exactly one risk-neutral
probability Q ∼ P.

Proof
This proof is very similar to that in the single-period case. As in the First Fundamental
Theorem we shall make crucial use of the finiteness of Ω.
First, assume the No Arbitrage Principle and completeness. A risk-neutral prob-
ability exists by the First Fundamental Theorem, so all we have to do is to prove
uniqueness. Suppose we have two risk-neutral probabilities Q1 6= Q2 . Fix A in FN and
consider a derivative security with payoff
H(N ) = 1A .
By completeness, there exists a self-financing predictable strategy (x, y) such that
V(x,y) (N ) = 1A .
Taking expectations with respect to each risk-neutral measure, we can see that
EQi (V(x,y) (N )) = EQi (1A ) = Qi (A) i = 1, 2.
But the discounted values of the strategy form a martingale under each Qi , so the
expectation is constant in time
(1 + R)−N EQi (V(x,y) (N )) = EQi (Ve(x,y) (N )) = V(x,y) (0).
The right-hand side does not depend on i, so Q1 (A) = Q2 (A) (both are equal to
V(x,y) (0)(1 + R)N ). Hence Q1 = Q2 , as A ∈ FN was arbitrary.
For the converse implication, assume that there exists a unique risk-neutral proba-
bility Q ∼ P. By the First Fundamental Theorem, there are no arbitrage opportunities.
To prove completeness, suppose that a certain derivative security with payoff H(N )
cannot be replicated, and show that this allows us to construct another risk-neutral
probability, which would be a contradiction. Consider the set of real random variables
defined on Ω by
W = {V(x,y) (N ) : (x, y) is self-financing and predictable}.
This is a vector subspace of RM with H(N ) ∈ / W. Investing risk free, we can obtain a
deterministic final value so 1 = (1, . . . , 1) ∈ W. Equip RM with the inner product
M
X
hz, wiQ = zi wi Q(ωi ).
i=1

103
By Lemma 1.60 with A = {H(N )}, there exists a z ∈ RM such that hz, H(N )iQ > 0
and hz, wiQ = 0 for each w ∈ W , which means that z 6= 0 and hz, 1iQ = 0. Write
a = maxi=1,...,M |zi | > 0, and define a new probability
 zi 
Q1 (ωi ) = 1 + Q(ωi ) for i = 1, . . . , M.
2a
We have proved (see the proof of the single-period Second Fundamental Theorem)
that Q1 is a probability equivalent to P (since Q1 (ωi ) > 0 for each i = 1, . . . , M ) and
Q1 6= Q.
It remains to show that Q1 is a risk-neutral probability. To this end take a
predictable sequence x(n), and use it to create a self-financing strategy (x,y) with
initial value V (0) = 0 (see Theorem 3.47). We use the fact that the discounted values
of the strategy form a Q-martingale, so the expectation is constant in time:
M
X zi e
EQ1 (Ve(x,y) (N )) = Q(ωi )(1 + )V(x,y) (N )
i=1
2a
M M
X 1 X
= Q(ωi )Ve(x,y) (N ) + zi V(x,y) (N )Q(ωi )
i=1
2a(1 + R)N i=1
1
= EQ (Ve(x,y) (N )) + hz, V(x,y) (N )iQ
2a(1 + R)N
1
= V (0) + hz, V(x,y) (N )iQ
2a(1 + R)N
= 0,

since hz, wi = 0 for all w ∈ W. Because V(x,y) (0) = V (0) = 0, we have

Ve(x,y) (N ) = Gex (N ), so EQ1 (G ex (N )) = 0. Because x is an arbitrary predictable
process, Theorem 3.55 can be applied to show, just like at the end of the proof of
Theorem 3.58, that the Sej are martingales under Q1 for each j = 1, . . . , d, so that Q1
is a risk-neutral probability.
We have obtained a contradiction with the uniqueness of the risk-neutral proba-
bility, so every derivative security can be replicated, that is, the model is complete.

Example 3.64
We take the single-step trinomial model with
U = 30%, M = 20%, D = −10%
and
R = 10%.

104
 This video constructs several risk neutral probabilities in this setting, as well as
an example of a derivative security which cannot be replicated.

Consequences for option pricing

We give a summary of the consequences of the Fundamental Theorems.
Theorem 3.65
If the underlying market (d stocks and money market) does not admit arbitrage and
a derivative security with payoff H(N ) is replicable by a strategy (x, y), then for the
extended market to be arbitrage-free we must have

H(n) = V(x,y) (n)

for each n = 0, 1, . . . , N .

Proof
To see this, note that if the equality failed at any time, an arbitrage opportunity could
be constructed. The investment would be started at that time, the option sold and
the portfolio bought (or vice versa). The profit would be maintained since at the final
moment the option payoff is identical to the portfolio value.

Example 3.66
In Example 3.64 we have seen that the derivative security with payoff

 10 in the u scenario
H(1) = 0 in the m scenario
10 in the d scenario


cannot be replicated.

The video shows another way to verify this, based on Theorem 3.65.

Then, employing the fact that the stock prices and also the values of strategies are

105
martingales with respect to the unique risk-neutral probability, we obtain

Ve(x,y) (n) = EQ (Ve(x,y) (N )|Fn )

= EQ (H(N
e )|Fn )

inserting the payoff as a result of the replication assumption. By definition

Ve(x,y) (n) = H(n)

e = (1 + R)−n H(n)

so
1
H(n) = EQ (H(N )|Fn ).
(1 + R)N −n
In particular (for n = 0),
H(0) = (1 + R)−N EQ (H(N )).

3.7 Forwards and futures

In this section we use the general theory to investigate simple properties of forward and future
contracts.

Forward contracts
One of the simplest examples of a derivative security is a forward contract. This contract
is a binding agreement between two parties to buy or sell the underlying asset at a given
future time N for a specified delivery price K. The party committed to buying the asset holds
a long position, and the party committed to selling the asset holds a short position. The
payoff at time N for the long position is

Hlf (N ) = S(N ) − K

and brings profit if S(N ) > K. For the short position the payoff is

Hsf (N ) = K − S(N ).

In a complete arbitrage-free model the price of such a contract at any time n = 0, . . . , N

can easily be computed:

H
e lf (n) = EQ (H e ) − K|F
e lf (N )|Fn ) = EQ (S(N e n ) = S(n)
e − K(1 + R)−N .

When discounting is removed, we obtain a formula for the value of a forward contract at
time n. However, this formula can be proved without referring to any model.

106
Proposition 3.67
The No Arbitrage Principle implies that the time n price of a long forward contract
with delivery price K is

Hlf (n) = S(n) − K(1 + R)−(N −n) .

Proof
We build a strategy (x, y) such that V(x,y) (N ) = S(N ) − K by taking x(n) = 1,
1
y(n) = − A(0) K(1 + R)−N for each n = 0, . . . , N (this is a static strategy, that is,
the positions are the same for each n). Hence, by the No Arbitrage Principle, at each
time n we must have Hlf (n) = V(x,y) (n). This means that

Hlf (n) = V(x,y) (n) = S(n) + y(n)A(n) = S(n) − K(1 + R)−(N −n) .

Definition 3.68
We define the forward price F (n, N ) as the delivery price that gives zero value to a
contract entered into at time n, with delivery at time N.

Corollary 3.69
No-arbitrage implies that the forward price at time n = 0, . . . , N is

F (n, N ) = S(n)(1 + R)N −n .

Proof
From H(n) = 0 we get the result at once.

Example 3.70
Let
S(0) = 100
R = 5%
and consider a forward contract exchanged at time 0 with delivery time
N = 10.
Suppose that the long position holder would like to close it at time n = 6.
The video shows how closing the long forward position can be archived.

107
Futures
In addition to stock prices, markets also quote so-called futures prices. The principal difference
between them and forwards is that futures are quoted on an exchange, whereas forwards are
traded ‘over-the-counter’ (OTC), by mutual agreement between the two parties concerned.
The exchange seeks to minimise the risk of default, and investors are required to hold a
margin account with the exchange, which must be adjusted daily to reflect the fluctuations
in futures prices.
We restrict our attention to discrete-time situations.
Definition 3.71
Use f (n, N ) to denote the futures price at time n = 0, 1, . . . , N − 1 for delivery
at time N of an underlying asset S(n). This provides a series of derivative securities
called futures defined by means of the following conditions:

ˆ f (N, N ) = S(N ).

ˆ f (n, N ) is Fn -measurable, where Fn is the field generated by the stock prices

S(1), . . . , S(n).

ˆ At time n a long futures position generates the cash flow

f (n, N ) − f (n − 1, N ), n = 1, . . . , N.

ˆ A time n a short futures position generates the cash flow

f (n − 1, N ) − f (n, N ), n = 1, . . . , N.

ˆ It costs nothing to enter any futures position at any time.

The cash flows, called marking-to-market, are credited/debited to the margin ac-
count.

Theorem 3.72
If the interest rate is constant, then

f (0, N ) = F (0, N ).

Proof
We claim that the payoff S(N ) − K of a long forward contract with K = f (0, N ) and
delivery time N can be replicated at zero initial cost in a market with futures contracts.

108
If this is so, then K must be equal to the forward price F (0, N ) or else an arbitrage
opportunity would arise. This means that f (0, N ) = F (0, N ).
For N = 1 the claim is obvious. Let us verify it for N = 2. Consider a portfolio
consisting of the following two futures contracts:
1
1+R
long futures initiated at time 0,
R
1+R
long futures initiated at time 1.

The marking-to-market cashflow of the first contract consists of two payments:

f (1, 2) − f (0, 2) at time 1, and S(2) − f (0, 2) at time 2. The first payment can
be invested risk free, and becomes [f (1, 2) − f (0, 2)] (1 + R) at time 2. For the sec-
ond futures contract, there is a single marking-to-market payment S(2) − f (1, 2) at
time 2. Multiplying these time 2 amounts by the corresponding positions in these
futures contracts and adding up, gives the following total:
1 R
([f (1, 2) − f (0, 2)] (1 + R) + S(2) − f (1, 2)) + (S(2) − f (1, 2))
1+R 1+R
= S(2) − f (0, 2).

It means that this portfolio of futures replicates the payoff S(2) − f (0, 2), proving the
claim for N = 2.
It remains to extend the above argument to verify the claim for an arbitrary
N > 0. In this case, consider the following portfolio consisting of N positions in
futures contracts:
1
(1+R)N −1
long futures initiated at time 0,
R
(1+R)N −1
long futures initiated at time 1,
R
(1+R)N −2
long futures initiated at time 2,
.. ..
. .
R
(1+R)2
long futures initiated at time N − 2,
R
1+R
long futures initiated at time N − 1.

It is left to the reader to verify that the cash flow of this portfolio invested risk free up
to time N does indeed replicate the payoff S(N ) − f (0, N ).

Remark 3.73
The key point in this argument is that the future interest rate R is known in advance,
which allows us to specify the size of the futures positions initiated at times 0, 1, . . . , n.

109
 Example 3.74

Suppose that R = 20%, N = 2, and the market follows the scenario

S(0) = 100.00, S(1) = 110.40, S(2) = 115.12.

The video illustrates the proof of Theorem 3.72 by replicating a forward contract
with futures in this scenario.

3.8 Complex derivatives

Call and put options are often called (plain) vanilla options as they have no special features
and are regarded by market traders as very familiar. More complex options are often called
exotic, although this term is not used with much precision.
Vanilla options can be combined in various ways to form trading strategies whose payoffs
are suited to the investors’ various needs. We offer some simple examples, reflecting different
investment needs.
Example 3.75
Spreads are portfolios using options of the same type, some in long and some in short
positions. A popular example is the butterfly spread, which is designed for investors
who believe that large fluctuations in the price of the underlying stock are unlikely in
[0, N ]. One buys calls with strikes K1 and K3 > K1 and at the same time sells two calls
at strike K2 = 12 (K1 + K3 ). The initial outlay is thus I(0) = C1 (0) + C3 (0) − 2C2 (0),
which is positive, as the call price is a convex function of the strike. The payoff is
0 outside the strike interval (K1 , K3 ], S(N ) − K1 on (K1 , K2 ] and K3 − S(N ) on
(K2 , K3 ]. An investor who believes that the stock price will not vary much from X2
is thus likely to profit from this position and is well protected against large unforeseen
price changes.

Example 3.76
A bottom straddle consists of long positions in both a call and a put with common
strike K and expiry date N. This is a simple example of a combination, using options
of different types on the same underlying. We buy a call and a put, both with strike
K and expiry N. The payoff is |S(N ) − K|, against an initial outlay of C(0) + P (0).
Thus the holder will lose money if the price of the stock stays close to K but can
gain substantially if the final price differs significantly from K. This investment thus

110
 suits an investor who expects large price changes, by taking options with strike close
to S(0).

Path-dependent options
Here we provide a brief glimpse of the variety of derivatives that can be constructed from the
basic building blocks we have analysed
Recall, that a European derivative security H is path-dependent if

H(N ) = h(S(0), S(1), . . . , S(N )).

We consider some common path-dependent options below.

Example 3.77
Barrier options are instruments that are designed to hedge against movements in
the stock price crossing a given level Y , seen as a ‘barrier’. If the stock price reaches
(or crosses) Y, a knock-out barrier option becomes worthless. If S(n) reaches Y
from below, we have an up-and-out option, and otherwise down-and-out. On the
other hand, a knock-in option only comes into existence once the barrier is reached,
although the option premium has been paid at the outset.
For a single barrier Y , we can devise eight European-style barrier options by using
the terms call/put, up/down, and in/out in combination. In addition, one can develop
two-sided barrier options, using both an upper and a lower barrier, with different
outcomes depending on whether activation (or deactivation) occurs when either or
both barriers are breached.
Further complications can arise in practice when the option holder may be given
a rebate when the option is knocked out, or, as for Parisian options, the option is
knocked out only once the barrier has been breached for a specified length of time.
Barrier options are attractive to a buyer seeking to minimise the cost of option
premia by excluding scenarios she considers unlikely, while accepting the risk that they
may occur. For example, a down-and-out call should be cheaper than an ordinary one,
since the option is knocked out if S(n) reaches Y from above, irrespective of whether
the stock price recovers to above Y by time N. The writer of an up-and out call, on
the other hand, limits her potential liabilities in the event that the stock rises sharply.

This video shows pricing and replicating an down-and-in call. Here is the ac-
companying Excel file.

111
 Example 3.78
Lookback options come in several guises, using the maximum Mlb or minimum mlb of
the underlying stock price over a preset ’lookback period’. We have the floating strike
variety, where the strike becomes a random variable. Thus the lookback call allows
one to buy the stock at the minimum value achieved in the lookback period, hence
the payoff is (S(N ) − mlb )+ . For the put the holder uses the maximum instead, with
payoff (Mlb − S(N ))+ . There are also fixed strike versions, with payoffs (Mlb − K)+
and (K − mlb )+ , and even partial lookbacks, using a preset fraction λMlb or λmlb
instead. Clearly, all these are designed to ‘eliminate regret’ over lost opportunities, and
their premia will reflect this.

Example 3.79

The Asian call option, in an N -step model has payoff H(N ) = (A(N ) − K)+ ,
where A(N ) is either the arithmetic average N1 N
P
QN i=1 S(i) or the geometric average
1/n
[ i=1 S(i)] of the stock prices, together with a fixed strike K. Clearly, these are
path-dependent options. They are usually traded as of European type, to avoid early
exercise occurring before the averaging period is completed, thus losing the protection
it can provide.
To see why one may wish to use such options in certain situations, consider a
fixed strike K = 1.5, N = 6 and A(N ) as the arithmetic average. Suppose that the
price evolution of the underlying S is given by

n 1 2 3 4 5 6
S(n) 1.5 1.6 1.7 1.8 1.6 1.4

Then H(6) = 0.1, whereas for a European call C(6) = 0. The Asian option may be
preferred by a company which regularly buys some asset and wants to hedge against
high prices, for example, to guard against price spikes near the trading horizon. It can
buy a series of vanilla call options with different maturities or a single Asian call on the
average price. The latter is usually satisfactory from the point of view of securing the
company’s business.

Pricing formulae have been extended to cover many other exotic options, including
choosers, where the option holder can decide at an intermediate time 0 < N1 < N whether
he wishes to treat his option as a put or a call, and compound options, where the underlying
is not a stock but an option with earlier expiry.

112
Chapter 4

American options

As we recall, a European option confers the right to a random payoff at a prescribed time N
in the future. By contrast, an American derivative security allows the holder to exercise the
corresponding right at any time n ≤ N , so for an American call we can obtain the payoff
(S(n) − K)+ on demand at any time n ≤ N. When exercised, such an option is terminated.
The payoff depends on the values of S(n) for all n up to the moment of exercise, hence a
representation of the payoff similar to that of the European case is not possible in general.
Even in the simple binomial setting, a direct analogue of the CRR formula is not possible
as the payoff depends on the holder’s choice of the exercise date. This requires so-called
stopping times to describe the optimal exercise time for the option.
We discuss optimal stopping times in a general discrete-time model with finitely many
trading dates, where we assume that we are given a finite probability space (Ω, F, P ) equipped
with a filtration of fields Fn (see Definition 3.27) generated by a process S(n) representing
the underlying security prices. We illustrate the resulting pricing techniques for single-stock
models numerically throughout in the binomial three model where, as usual, the space Ω is
taken in the form {u, d}N , that is, an element ω is an N -tuple of symbols u, d corresponding
to the returns U, D of the price of the underlying.

4.1 Pricing
Consider the binomial model and assume that the option holder’s choice of exercise date is
made in order to maximise the amount received. At each time she faces the choice between
exercising immediately and postponing this till later. The sum of money given by the payoff
can readily be seen at each time, in each scenario, being a known function of the current
stock price. Valuing the alternative, that is, the choice not to exercise poses a problem and
depends on assumptions about the future behaviour of the stock. This makes it natural to
seek to solve the pricing problem by means of backward induction, while taking into account
the additional choice of whether to exercise or not at each node of the binomial tree. The
method is best illustrated through an example which is sufficiently complex to reveal the

113
various possible scenarios while remaining very simple computationally. Thus, all notions and
theorems in this chapter will be illustrated by a concrete example of a single-stock model on
a 5-step binomial tree.
Example 4.1
We begin with the specification of the dynamics of the underlying security. The initial
price is S(0) = 100, the returns are U = 15%, D = −10%, so the tree of the prices
is as shown in Figure 4.1.
Consider a put option with expiry date N = 5 and exercise price K = 100. Since
the American option can be exercised at any time prior to the expiry, it is necessary to
compute the immediate payoff of the option at each node of the tree. We denote it
by I(n) = (K − S(n))+ . The results are shown in Figure 4.2.
These calculations can be found in the Excel file.

Figure 4.1: Stock prices S(n) in Example 4.1.

Figure 4.2: Put payoff I(n) in Example 4.1.

Pricing will be performed in a similar fashion as for European claims, starting from the
expiry time and moving backwards. The value of the American put at time n will be denoted
by P (n), so
P (5) = I(5).

114
Example 4.2
We continue Example 4.1, and our first serious goal is to compute the prices at time
n = 4. At this time the holder of the option has a choice between exercising and
waiting till the final moment.
The decision about ‘waiting’ now depends on the value of the European option
with exercise date one step later. This can easily be computed by means of replication,
or, equivalently, by using the risk-neutral probability q. Suppose the risk-free return is
R = 5%, so q = UR−D −D
= 0.6. At each state at time n = 4 we compute the discounted
expected value of the payoff available at time 5, obtaining the numbers as in Figure 4.3.
In the two lowest states at time n = 4 it is better to exercise immediately rather
than wait and the values of the option are the corresponding payoffs (of course we mean
here the positive numbers, zeros are irrelevant from this point of view). In the middle
state ‘waiting’ has positive value while the immediate payoff is zero. Figure 4.4 shows
the results. The highlighted umbers are those where immediate exercise prevailed.
See the Excel file for details.

Figure 4.3: Discounted expected payoff at time 4 in Example 4.2.

Figure 4.4: American put prices at time 4 in Example 4.2.

We can summarise the above in a concise way. The value of the European put at time 4

115
is the random variable
1 1
EQ (I(5)|F4 ) = EQ (P (5)|F4 ).
1+R 1+R
Therefore, taking the benefit of immediate exercise, if profitable,
 
1
P (4) = max I(4), EQ (P (5)|F4 ) .
1+R
It is convenient to divide both sides by (1 + R)4 , which will lead to the discounted prices
satisfying n o
P (4) = max I(4), EQ (P (5)|F4 ) .
e e e

The prices P (4) computed above are the true market values of the American put, encapsulat-
ing the rights belonging to the holder: exercise immediately or wait and keep the option. As
such they represent the possible cash flow at time 4, since the option can be sold. Therefore
they form a basis for the next (backward) induction step, moving from time 4 to 3. Applying
the same argument to this step, we obtain
n o
Pe(3) = max I(3),
e EQ (Pe(4)|F3 ) .

Example 4.3

We continue Example 4.2. The numerical results for P (3) are shown in Figure 4.5.
Note that in the DDU state the exercise payoff is 6.85, which is lower than the value
of ‘waiting’.
See the Excel file.

Figure 4.5: American put prices at time 3 in Example 4.3.

In general we have shown how to develop the following backwards induction procedure,
which has the form
Pe(5) = I(5),
e
n o
Pe(n − 1) = max I(n e − 1), EQ (Pe(n)|Fn−1 )

116
for the discounted values.
Example 4.4
This allows us to complete Example 4.3. The resulting in the numbers given in Fig-
ure 4.6, where in particular we find the time 0 price of the American put in question:
P (0) = 4.51. All numbers here are rounded to two decimal points.
See the Excel file for the full computation details.

Figure 4.6: American put prices in Example 4.4.

We have constructed an example of a general mathematical object.

Definition 4.5
A sequence Y (n) of random variables is the Snell envelope of the sequence Z(n),
n = 1, 2, . . . , N, adapted to Fn if

Y (N ) = Z(N ),
Y (n − 1) = max{Z(n − 1), E(Y (n)|Fn−1 )}.

The notion is defined in a general setting, but in applications we shall use it for binomial
trees, with expectation with respect to the risk-neutral probability.
So, as we have seen, Y (n) = Pe(n) is the Snell envelope of Z(n) = I(n),e n = 1, . . . , 5
with respect to the risk-neutral probability Q and the filtration generated by the partitions of
the probability space of scenarios in a binomial tree.
Directly from the definition, we can see that Y (n − 1) is Fn−1 -measurable (that is, Y is
adapted), and
Y (n − 1) ≥ E(Y (n)|Fn−1 ),
(these two conditions by definition mean that Y is a supermartingale), as well as

Y (n − 1) ≥ Z(n − 1),

117
so the Snell envelope Y of Z is a supermartingale dominating Z. This turns out to be the
key to characterising Snell envelopes.
Theorem 4.6
The Snell envelope Y of Z is the smallest supermartingale dominating Z. In other
words, if T (n) is a supermartingale satisfying T (n) ≥ Z(n), then Y (n) ≤ T (n).

Proof
We have already observed that Y is a supermartingale and that it dominates Z. Given
any supermartingale T with T (n) ≥ Z(n) for all n, we have, in particular,

T (N ) ≥ Z(N ) = Y (N ).

We shall prove by backward induction that this property holds for all n. Suppose, for
the induction hypothesis, that this is true for some k, that is, T (k) ≥ Y (k). This
immediately implies
E(T (k)|Fk−1 ) ≥ E(Y (k)|Fk−1 ).
The sequence T is a supermartingale, so

T (k − 1) ≥ E(T (k)|Fk−1 ).

Putting these two inequalities together, we get

T (k − 1) ≥ E(Y (k)|Fk−1 ).

In addition, as T dominates Z, we have T (k − 1) ≥ Z(k − 1), so

T (k − 1) ≥ max{Z(k − 1), E(Y (k)|Fk−1 )} = Y (k − 1)

(the last equality is the definition of the Snell envelope), which completes the induction
step and so the proof.

The discounted American option price

Y (n) = Pe(n)

is the Snell envelope of the discounted payoff

Z(n) = I(n).
e

118
Example 4.7

In the Excel file the values of Z(n) and Y (n) are computed for the American put in
our 5-step numerical example (see Example 4.4).

Proposition 4.8

The prices CA (n) of an American call and CE (n) a European call on the same un-
derlying S(n) and with the same delivery date N and strike price K are equal to one
another,
CA (n) = CE (n)
for each n = 0, . . . , N .

Remark 4.9
Here we consider an underlying asset that pays no dividends. The equality between
American and European call prices would, in general, fail otherwise.

Proof
Observe that
1 +

CE (n) = EQ (S(N ) − K)
(1 + R)N −n
1
≥ N −n
(EQ (S(N )) − K)+
(1 + R)
 +
K
= S(n) −
(1 + R)N −n
≥ (S(n) − K)+

for each n = 0, . . . , N − 1. Here the first inequality holds because EQ (f (X)) ≥

f (EQ (X)) for any convex function f (x) and any random variable X. The second
inequality holds because (x − a)+ ≥ (x − b)+ whenever a ≤ b.
Now we proceed by backwards induction. For the payoff at time N, we have

CA (N ) = CE (N ) = (S(N ) − K)+ .

Suppose that
CA (n) = CE (n)

119
 for some n = 1, . . . , N . Then
 
1
+
CA (n − 1) = max (S(n − 1) − K) , EQ (CA (n))
1+R
 
+ 1
= max (S(n − 1) − K) , EQ (CE (n))
1+R
max (S(n − 1) − K)+ , CE (n − 1)

=
= CE (n − 1),

completing the proof.

4.2 Stopping times and optimal exercise

Example 4.10

Continuing Example 4.7, suppose we bought the American put for P (0) = 4.51. The
problem we are facing at all times is the decision whether to exercise this option or
not. (Refer to Figure 4.6 for the option prices computed at each stage.)
Consider first the strategy of exercising at the earliest possible time when the
option price is equal to the available payoff.
Of course, we do not exercise at time 0 since the payoff is 0. Let Bu (resp. Bd ) be
the set of all paths beginning with u (resp. d), and similarly define Buu , Bud , Bdu , Bdd ,
and so on.

ˆ Suppose the stock goes down in the first step, that is, consider ω ∈ Bd . We
exercise the option at time 1, cashing 10, which, as we saw, is higher than the
expected profit from waiting.

ˆ Suppose the stock goes up in the first step, so let ω ∈ Bu . Here we distinguish
three cases:

– If ω = udddu or ω = udddd, we exercise at time 4, obtaining 16.17.

– If ω = uuddd, ω = ududd, or ω = uddud, we exercise at time 5 receiving
3.59.
– For the remaining paths, we do not exercise the option at all and receive
zero. Or, in other words, we exercise at time 5, where the payoff is zero.

We have defined a random variable τ assigning to each ω the time when we exercise

120
 the option: 
 1 if ω ∈ Bd ,
τ (ω) = 4 if ω ∈ Buddd ,
5 otherwise.


This is an example of an important general notion, which we define next.

Definition 4.11
A random variable τ : Ω → {1, 2, . . . , N } is called a stopping time if {ω : τ (ω) ≤
n} ∈ Fn for all n.

The condition means that the decision to stop at time n will be based on the information
available at that moment. This is even clearer in the following equivalent formulation.
Proposition 4.12

τ : Ω → {1, 2, . . . , N } is a stopping time if and only if {ω : τ (ω) = n} ∈ Fn for all n.

Proof
Suppose the condition holds for τ and fix n ≤ N. Since Fk ⊂ Fn for all k ≤ n, it
follows that n
[
{ω : τ (ω) ≤ n} = {ω : τ (ω) = k} ∈ Fn .
k=1

This means that τ is a stopping time.

Conversely, given a stopping time τ and an n ≤ N, we have
n−1
\
{ω : τ (ω) = n} = {ω : τ (ω) ≤ n} ∩ {ω : τ (ω) > k} ∈ Fn
k=1

since Fk ⊂ Fn when k ≤ n.

Example 4.13
Going back to Example 4.10, we can define a natural modification of the process of
the option values related to early exercise. These values fluctuate with time when
we observe them along a given scenario. For example, if ω = udddu, we have the
following sequence of prices P (n, ω), n = 0, 1, 2, 3, 4, 5 :

4.51, 1.23, 2.94, 6.94, 16.17, 3.59.

In this scenario our strategy tells us to exercise at time 4. Imagine that we keep the

121
 money we have cashed, so the sequence is modified to become

4.51, 1.23, 2.94, 6.94, 16.17, 16.17.

For this particular ω, we have τ (ω) = 4, and we’ve left the sequence unchanged for
n ≤ 4, replacing the subsequent values by the value at time 4, so for n ≥ 4 we have
P (n, ω) = P (τ (ω), ω).

This motivates the next definition.

Definition 4.14
For any sequence of random variables X(n) and any stopping time τ , we define the
stopped process Xτ (n) by

Xτ (n) = X(min{τ, n}),

or, more explicitly, by


X(n, ω) if n ≤ τ (ω),
Xτ (n, ω) =
X(τ (ω), ω) if n > τ (ω).

In particular, Xτ (N ) = X(τ ). The stopped process Xτ inherits all the important proper-
ties of X.
Proposition 4.15

If X is adapted, then Xτ is also adapted. If X is a martingale (resp. supermartingale),

then Xτ is also a martingale (resp. supermartingale).

Proof
First we observe that
n
X
Xτ (n) = X(0) + 1{j≤τ } (X(j) − X(j − 1)) (4.1)
j=1

for each n = 0, 1, . . . , N .
The sequence 1{j≤τ } is predictable since

{j ≤ τ } = Ω \ {j > τ } = Ω \ {τ ≤ j − 1} ∈ Fj−1 .

Hence Xτ is adapted since all the terms on the right in (4.1) are Fn -measurable. So
if X is a martingale, then Xτ is a martingale transform, so Xτ is also a martingale by
Proposition 3.54.

122
 Next, again using the representation of Xτ (n) given in (4.1), we compute the
conditional expectation of the increment:

E(Xτ (n + 1) − Xτ (n)|Fn ) = E(1{n+1≤τ } (X(n + 1) − X(n))|Fn )

= 1{n+1≤τ } E(X(n + 1) − X(n)|Fn ),

where we used the ‘take out what is known’ property of the conditional expectation
(1{n+1≤τ } is Fn -measurable). Therefore the random variables E(Xτ (n+1)−Xτ (n)|Fn )
and E(X(n + 1) − X(n)|Fn ) have the same sign: they are either both greater than
or equal to zero, or both less than or equal to zero. Hence if X is a supermartingale,
that is,
E(X(n + 1) − X(n)|Fn ) = E(X(n + 1)|Fn ) − X(n) ≤ 0
then Xτ is a supermartingale.

Remark 4.16
A submartingale is an Fn -adapted sequence satisfying E(X(n + 1)|Fn ) ≥ X(n) for
each n. The above proof also shows that if X is a submartingale, then the stopped
process has the same property.

We seek a criterion for optimality of the exercising strategy. Since the sum of money
generated by such strategies can be random (in our special strategy the decision and the
outcome depends on ω), their comparison is difficult. Random variables are functions, and
functions are rarely comparable. For this reason we need to associate a single number with
each exercising strategy.
A natural candidate for an optimality criterion is to maximise the mathematical expecta-
tion of the payoff obtained at the exercise time. If the moment at which we exercise is denoted
by τ, the money received in a particular scenario ω is the payoff I(τ (ω)). These amounts of
money emerge at different time instants, so we should discount them to make them compa-
rable (or, equivalently, invest them in the money market account to hold until time N ). We
choose the risk-neutral probability Q for the purpose of computing the expectation.
Example 4.17
Returning to Example 4.13, to find the expected value of all discounted payments,
note that we receive 10 for ω ∈ Bd , 16.17 for ω ∈ Buddd , and 3.59 for ω = uuddd,
ω = ududd, or ω = uddud, so that
 
1 10 16.17 3.59
EQ τ I(τ ) = (1−q) +q(1−q)3 4
+3q 2 (1−q)3 = 4.51,
(1 + R) 1+R (1 + R) (1 + R)5
(4.2)
which, remarkably, is the money we paid for the option. (We show the results up to

123
 two decimal points, so the equality is approximate. But we shall prove that the equality
is exact, in general.)

As we shall see, this is the best we can get. Such a claim could be verified numerically
in our example since we have to deal with a finite number of possible strategies, but we will
prove it in general.
First we give a formal definition of an optimal exercise strategy.
Definition 4.18
A stopping time τ is optimal if

E(Z(τ )) = max {E(Z(ν)) : ν is a stopping time} . (4.3)

Analysing exercising strategies in general, consider again the discounted payoff Z(n) =
I(n) of the American put option. The Snell envelope Y of Z gives the value of our option
e
at each trading date. If it satisfies Y (n) > Z(n), then it is not reasonable to exercise at
time n, since we would be replacing a security worth Y (n) by a smaller amount Z(n). So,
a sensible exercise time τ should satisfy the condition Y (τ ) = Z(τ ). If such an opportunity
emerges, then from the economic point of view we should grab it as soon as possible due to
the time value of money. Below we prove that this intuition gives a correct solution.
Proposition 4.19
The random variable τmin defined by

τmin = inf{n : Y (n) = Z(n)},

where inf(∅) = N, is an optimal stopping time. In addition,

Y (0) = E(Z(τmin )). (4.4)

Proof
First note that τmin is a stopping time since
n−1
\
{τmin = n} = {Y (k) > Z(k)} ∩ {Y (n) = Z(n)}
k=0

and each of the sets on the right belongs to Fn since the random variables involved
are suitably measurable. For instance, for k < n,

{Y (k) > Z(k)} = (Y (k) − Z(k))−1 ((0, ∞)) ∈ Fk .

124
 We will show that the stopped Snell envelope Yτmin is a martingale. Using the
representation (4.1) of the stopped process, we have

Yτmin (n + 1) − Yτmin (n) = 1{n+1≤τmin } (Y (n + 1) − Y (n)).

If {τmin ≥ n + 1} then Y (n) > Z(n). By the definition of the Snell envelope,

Y (n) = max {Z(n), E(Y (n + 1)|Fn )} = E(Y (n + 1)|Fn ).

We insert this on the right-hand side above and compute the conditional expectation

E(Yτmin (n + 1) − Yτmin (n)|Fn )

= 1{n+1≤τmin } E(Y (n + 1) − E(Y (n + 1)|Fn )|Fn ) = 0,

which proves that Yτmin is a martingale.

Martingales have constant expectation, so for all n,

E(Yτmin (n)) = E(Yτmin (0)) = Yτmin (0) = Y (0),

the initial value being non-random. In particular, E(Yτmin (N )) = Yτmin (0), but by the
definition of τmin , we have E(Yτmin (N )) = E(Zτmin (N )). Putting all this together, we
get Y (0) = E(Zτmin (N )) as claimed in (4.4).
Finally, we show that τmin is optimal. Take any stopping time ν. We know that Yν
is a supermartingale, so the expectation decreases with time, hence Y (0) = Yν (0) ≥
E(Yν (N )). By the definition of the stopped process, E(Yν (N )) = E(Y (ν)), which is
greater than or equal to E(Z(ν)) by the definition of the Snell envelope. So

Y (0) ≥ max{E(Z(ν)) : ν is a stopping time}.

But (4.4) holds, so Y (0) is equal to one of the expectations on the right. Therefore,
we must have equality, that is, (4.3) holds, which completes the proof.

Remark 4.20
Optimality of the early exercise strategy follows from simple economic considerations. If
the option holder chooses not to exercise when the payoff is higher than the expectation
of the future possibilities, then she effectively chooses an asset which is worth less than
the payoff, so suffers a loss (which is the profit of the other party, the seller of the
option). This will be seen clearly in the hedging argument presented below.

Putting together (4.4) and optimality of τmin (which means that (4.3) holds), we have

125
the following important conclusion concerning the price of the option:

Y (0) = max{E(Z(ν) : ν is a stopping time} (4.5)

(Y (0) is the price P (0) if we consider the Snell envelope of a sequence of put option payoffs).
Analysing the optimal exercise strategy in our leading example, path by path, we can see
that before we exercise, the prices follow a martingale. After we exercise (i.e. stop), the
sequence becomes constant, so it is obviously a martingale. Thus we have found a simple
and elegant criterion for optimality, however with little practical meaning since verification of
the martingale condition of the stopped process is not straightforward.
Theorem 4.21
A stopping time τ is optimal if and only if Z(τ ) = Y (τ ) and Yτ (n) is a martingale.

Proof
Suppose that Z(τ ) = Y (τ ) and Yτ (n) is a martingale, which implies (by the constant
expectation property) that

Y (0) = E(Yτ (N )) = E(Y (τ )) = E(Z(τ )),

the last equality holding by the definition of the stopped process. The last equality
holds because Y (τ ) = Z(τ ) by our hypothesis. Hence Formula (4.5) means that
E(Z(τ )) = max{E(Z(ν) : ν is a stopping time}, so τ is optimal as claimed.
Conversely, if τ is optimal, then by (4.3) and (4.5) we have Y (0) = E(Z(τ )). By
the definition of the Snell envelope, we have E(Z(τ )) ≤ E(Y (τ )). But we know that
Yτ is a supermartingale, so E(Y (τ )) = E(Yτ (N )) ≤ Y (0). Putting these together, we
get
Y (0) = E(Z(τ )) ≤ E(Y (τ )) = E(Yτ (N )) ≤ Y (0),
so E(Z(τ )) = E(Y (τ )). In addition, we know that Z(τ ) ≤ Y (τ ), so Z(τ ) = Y (τ ).
It remains to prove that Yτ is a martingale. To this end it is sufficient to show
that Yτ (n) = E(Y (τ )|Fn ). We know that Yτ is a supermartingale, hence Yτ (n) ≥
E(Yτ (N )|Fn ) = E(Y (τ )|Fn ). Then to show that the random variables Yτ (n) and
E(Y (τ )|Fn ) are equal, it is sufficient to show that their expectations are equal:

E(Yτ (n)) = E(E(Y (τ )|Fn )).

The right-hand side is E(Y (τ )) by the tower property The left-hand side is the same,

126
 as the following chain of inequalities shows:

Y (0) ≥ E(Yτ (n)) since Yτ is a supermartingale

≥ E(E(Yτ (N )|Fn )) as above
= E(Y (τ )) by the tower property since Yτ (N ) = Y (τ )
= E(Z(τ )) as established above
= Y (0) by (4.5) since τ is optimal.

Example 4.22
In the Excel file we check each single-step subtree for the martingale property of the
discounted price process Y (y) in our 5-step numerical example (see Example 4.17).
The stopped process Yτ is computed in the Excel file along each path in the 5-step
binomial tree (in the file we use the notation 1 = u, 0 = d, the yellow fields indicate
the payoff at the stopping time τ , blue fields indicate the nodes where Y satisfies the
martingale property, and purple fields correspond to nodes such that τ ≤ n, where the
Yτ (n) = Y (τ ).

Example 4.23
In the binomial model with

U = 4%, D = −5%, R = 2%,

S(0) = 100, A(0) = 1

consider an American put with expiry time 4 and strike price K = 100.

The video shows how to price the option and compute the optimal exercise
time τmin . Here is the Excel file used in the video.

4.3 Hedging
Example 4.24
We keep working within the framework of our 5-step binomial example; see Exam-
ple 4.22. Suppose we have written and sold the American put, where S(0) = 100,
U = 15%, D = −10%, K = 100, R = 5%. Having cashed the price of 4.51, we are
exposed to some risk.

127
 To make our position secure we can construct a replicating strategy based on
taking a position in the underlying and completing the portfolio with a position in the
money market account. Our initial position x(1) in stock is determined by the delta
(see Definition 1.15), computed as

Pu − Pd 1.23 − 10 ∼
x(1) = u d
= = −0.350644,
S −S 115 − 90
where superscripts denote the prices resulting from the up and down movements of the
stock in the first step. This number is negative, which means short selling the stock.
The money market cash position y(1) is then

y(1) = 4.51 + 0.350644 × 100 ∼

= 39.58,

so this amount is invested risk free for one period.

Consider the case where the stock goes down to 90. The option holder can
exercise, cashing 10. Or the holder can wait, the value of this alternative being 8.92
(which can be seen in the Excel file by replacing the payoff 10 by 0 in cell G23). This
is the price of the American put option with stock price 90 and 4 steps to expiry.
Therefore the option holder must exercise or otherwise an arbitrage would emerge: the
option writer would receive a free gift of 1.08 since for further hedging the amount of
8.92 is sufficient. Once he option is exercised, it ceases to exist and the hedging stops
here. If the option holder wishes to gamble on a further decline of stock prices, this
can be achieved by buying a new put at time 1 for 8.92. So the option is exercised
and we have to pay 10, which this comes from liquidating our portfolio (x(1), y(1)).
We have to close the short stock position paying −0.350644 × 90 ≈ 31.56. We have
39.58 × (1 + 5%) ≈ 41.56 in the money market account, so we break even.
Now suppose the stock goes up to 115. The payoff is zero, so the option will not
be exercised. Our portfolio is worth 41.56 −0.350644 × 115 = 1.23, which matches
the put price. We are in the same situation as at time 0: suppose we sold the option
for 1.23 at time 1 with stock price 115. We can construct the hedging portfolio in the
same way as above:
0.2 − 2.94 ∼
x(1) = = −0.10,
132 − 104
y(1) = 39.6 × (1 + 5%) + ((−0.35) − (−0.10)) × 115 = 12.2,

and the value of this portfolio agrees with the option price,

−0.10 × 115 + 12.2 = 1.23.

An interesting path is uddd. Here at time 4 we have the payoff 16.17 with the
value of waiting for the final, time 5, payoff being 11.40, so again the option holder

128
 exercises. Our position allows us to settle similarly as before: we have (see the Excel
file)
66.13 × (1 + 5%) − 0.64 × 84 = 16.17.
By inserting 0 in J37, the same amount is provided by the book-keeping formula in
Q37.

Example 4.25

In this video we continue Example 4.23 to compute the hedging strategy. The
Excel file used in the video is included.

4.4 The Doob decomposition

The sequence of prices Y (n) is a supermartingale, as we know. In particular, their expected
values decrease. To make such a sequence a martingale one should somehow compensate this
‘leakage of value’ by adding some suitable quantities. We shall show that such compensation
is always possible, and, moreover, that it is unique within a certain family of processes.
Theorem 4.26 (Doob decomposition)

If Y is a supermartingale, then there exist a martingale M and a predictable non-

decreasing process A with
A(0) = 0
such that
Y (n) = M (n) − A(n)
for each n = 0, 1, . . . , N .
This decomposition, called the Doob decomposition, is unique in the above
classes of processes M and A, with A given by the recursive formula

A(n) = A(n − 1) − E(Y (n) − Y (n − 1)|Fn−1 )

for n = 1, 2, . . . , N .

129
Proof
To find A(1), which, as an F0 -measurable random variable, must be a constant, we
take the expectation on both sides of the target relation:

Y (1) = M (1) − A(1)

to find

A(1) = −E(Y (1)) + E(M (1))

= −E(Y (1)) + M (0) if M is going to be a martingale
= −E(Y (1)) + Y (0) since Y (0) = M (0) − A(0) and A(0) = 0.

We must then put

M (1) = A(1) + Y (1)

= Y (1) − E(Y (1)) + Y (0).

Next, to find A(2) satisfying

Y (2) = M (2) − A(2)

we take the conditional expectation of this target identity with respect to F1 , recalling
that A is to be predictable, to get

A(2) = −E(Y (2)|F1 ) + E(M (2)|F1 )

= −E(Y (2)|F1 ) + M (1) if M is going to be a martingale
= −E(Y (2)|F1 ) + A(1) + Y (1) since Y (1) = M (1) − A(1).

Rearranged, this gives a formula for A(2):

A(2) = A(1) − E([Y (2) − Y (1)]|F1 ).

This suggests a general recursive formula: for n ≥ 1 write

A(n) = A(n − 1) − E([Y (n) − Y (n − 1)]|Fn−1 ) (4.6)

with the form of M implied by our goal

M (n) = A(n) + Y (n).

This was an educated guess based on the desired properties, and it is time for a
rigorous argument. First note that A is increasing since the supermartingale property
of Y gives
E(Y (n) − Y (n − 1)|Fn−1 ) ≤ 0.

130
 Next, we use induction to prove predictability: A(1) is constant, so F0 -measurable,
and if A(n − 1) is Fn−2 -measurable, it automatically is Fn−1 -measurable. With the
second term on the right of (4.6) also being Fn−1 -measurable, we find that A(n) is
Fn−1 -measurable.
It remains to see that M is a martingale:

E(M (n)|Fn−1 )
= E([A(n) + Y (n)]|Fn−1 )
= A(n) + E(Y (n)|Fn−1 )
= A(n − 1) − E([Y (n) − Y (n − 1)]|Fn−1 ) + E(Y (n)|Fn−1 )
= A(n − 1) + Y (n − 1)
= M (n − 1).

For uniqueness, suppose on the contrary that we have another such decomposition,
with Y (n) = M 0 (n) − A0 (n), so that

Y (n) = M (n) − A(n)

= M 0 (n) − A0 (n).

Then
M (n) − M 0 (n) = A(n) − A0 (n),
and note that for n = 0 we have

M (0) = M 0 (0) = Y (0)

since
A(0) = A0 (0) = 0.
Next, take the conditional expectation on both sides with respect to Fn−1 :

E(M (n) − M 0 (n)|Fn−1 ) = E(A(n) − A0 (n)|Fn−1 ),

which gives
M (n − 1) − M 0 (n − 1) = A(n) − A0 (n).
So, the equality M (n − 1) = M 0 (n − 1) implies thatA(n) = A0 (n), which yields
M (n) = M 0 (n). By induction on n, this proves uniqueness.

We can see that the role of A is to add to Y what is needed to obtain a martingale
(A(n) ≥ 0 since it starts from 0 and is non-decreasing), which justifies the commonly used
term compensator: the increasing process A ‘makes up for’ the lack of the martingale

131
property in Y.
Corollary 4.27
If X is a submartingale, then it can be written uniquely as

X(n) = M (n) + A(n)

with a martingale M and a predictable non-decreasing process A such that A(0) = 0.

Proof
It is sufficient to note that Y = −X is a supermartingale. The Doob decomposition
of Y gives Y = M − A, so −Y = −M + A, and so −M is the desired martingale for
the decomposition of X.

Going back to the hedging problem, in the notation for the put option we consider the
Doob decomposition of the sequence of discounted option prices Y (n) = Pe(n), so

Pe(n) = M (n) − A(n)

for some martingale M and increasing predictable A. In the binomial model (or in general, in
any complete model) we can replicate any Fn -measurable random variable given at any time n.
We choose to replicate M (N ), that means we find a self-financing strategy (x(n), y(n)), n =
1, . . . , N, such that
Ve(x.y) (N ) = M (N ).
When writing and selling the option, we receive the price P (0) = Y (0) = V(x,y) (0), so we
can start such a hedging strategy. The process Ve(x,y) (n) is a martingale, so

Ve(x,y) (n) = EQ (Ve(x,y) (N )|Fn ) = EQ (M (N )|Fn ) = M (n).

This implies
Pe(n) = Ve(x,y) (n) − A(n).
The value of our strategy for all n makes our position secure since the value of the strategy
dominates the option payoff at all times

Ve(x,y) (n) ≥ Pe(n) = Y (n) ≥ Z(n) = I(n).

e

The holder of the option may exercise if I(n) > 0, but this is not sufficient since if P (n) >
I(n), then he is better off by selling the option.
If P (n) = I(n) and A(n) = 0, the inequality Ṽ(x,y) (n) ≥ I(n) ˜ is then an equality and
the financial result for the writer is zero if the option is exercised.
If A(n) > 0, the inequality becomes strict, Ṽ(x,y) (n) > I(n), ˜ which means additional
profit, equal to the value of the compensator, for the writer if the option is exercised.

132
 Therefore the following exercise strategy is of interest: we do not exercise as long as Y is
a martingale. If A becomes non-zero, the process Y ceases to be a martingale, so we should
exercise just before that instant. At a given trading date we are able to tell whether A is
going to become positive at the next step since A is predictable. Of course, if A never gets
positive, being zero up to time N , the exercise time has to be the trading horizon:

N if A(N ) = 0,
τmax =
min{n : A(n + 1) 6= 0} otherwise.

As it turns out, this exercise strategy is optimal, but not very relevant from the practical
point of view.
Theorem 4.28
The stopping time τmax is optimal.

Proof
First, directly from the definition, τmax is a stopping time since the set {τmax = n} is
determined by A(0), A(1), . . . , A(n + 1), which are Fn -measurable.
Next, again by the definition, A(τmax ) = 0, so Yτmax (n) = Mτmax (n), and Yτmax is
a martingale.
According to Theorem 4.21, to complete the proof it is sufficient to show that
Y (τmax ) = Z(τmax ). Fix an arbitrary ω and denote by j the time instant τmax (ω).
Insert the Doob decomposition into the definition of the Snell envelope

Y (τmax (ω), ω) = Y (j, ω)

= max(Z(j, ω), E(Y (j + 1)|Fj )(ω))
= max(Z(j, ω), E(M (j + 1) − A(j + 1)|Fj )(ω))
= max(Z(j, ω), M (j, ω) − A(j + 1, ω))
= max(Z(τmax (ω), ω), M (τmax (ω), ω) − A(τmax (ω) + 1, ω))

since M is a martingale and A is predictable. Next, by the definition of τmax , we have

A(τmax (ω), ω) = 0 and A(τmax (ω)+1, ω) > 0. Since M (τmax (ω), ω) = Y (τmax (ω), ω),
we have M (τmax (ω), ω) − A(τmax (ω) + 1, ω) < Y (τmax (ω), ω). So Y (τmax (ω), ω) =
max(Z(τmax (ω), ω), M (τmax (ω), ω) − A(τmax (ω) + 1, ω)) = Z(τmax (ω), ω) for each
ω ∈ Ω.

In practical terms, the irrelevance of τmax follows from its comparison with a much simpler
strategy based on τmin . For exercise to be truly profitable the payoff should be strictly greater
than the value of waiting. In this case, if τmin (ω) = n, then

Y (n) = Z(n) > E(Y (n + 1)|Fn )

133
and, moreover, Y (k) > Z(k) for all k ≤ n − 1, which implies Y (k) = E(Y (k + 1)|Fk ).
Consequently,

A(k + 1) = A(k) − (E(Y (k + 1)|Fk ) − Y (k))

= A(k) = 0

and in particular A(n) = 0, but

A(n + 1) = A(n) − (E(Y (n + 1)|Fn ) − Y (n))

= Y (n) − E(Y (n + 1)|Fn ) > 0.

This gives τmax (ω) = n and so here, with strict inequalities above, both stopping times
coincide: τmax = τmin . This also proves that τmax ≥ τmin .
Example 4.29
See the Excel file for a numerical example, in the same setting as Example 4.24.

Example 4.30

Continuing Example 4.25, this video shows how to construct the martingale and
compensator by a different method than presented above. The optima stopping
time τmax is also computed. Here is the Excel file used in the video.

134