Discrete Time Finance

MATH 5320M
School of Mathematics

c University of Leeds

Term 2, 2020/21
Contents

1 Single-Period Market Models 2

1.1 The most elementary market model . . . . . . . . . . . . . . 3
1.2 A general single period market model . . . . . . . . . . . . . 13

2 Multi-Period Market Models 35

2.1 General Model Specifications . . . . . . . . . . . . . . . . . . 35
2.2 Properties of the general multi period market model . . . . . 49
2.3 Conditional expectation . . . . . . . . . . . . . . . . . . . . 54
2.4 Risk neutral probability measures: arbitrage and pricing . . 57
2.5 Exotic options . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.6 The Binomial Asset Pricing Model . . . . . . . . . . . . . . 66

3 Investment 74
3.1 Single Period Investment . . . . . . . . . . . . . . . . . . . . 74
3.2 Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.3 Mean-Variance Analysis . . . . . . . . . . . . . . . . . . . . 91

1
Chapter 1

Single-Period Market Models

Single period market models are the most elementary market models. The
beginning of the period is usually denoted by the time t = 0 and the end of
the period by time t = 1. At time t = 0 stock prices, bond prices,possibly
prices of other financial assets or specific financial values are recorded and
the financial agent can choose his investment, often a portfolio of stocks
and bond. At time t = 1 prices are recorded again and the financial agent
obtains a payoff corresponding to the value of his portfolio at time t = 1.
Single period models are unrealistic in a way, that in reality trading takes
place over many periods, but they allow us to illustrate and understand
many of the important economic and mathematical principles in Financial
Mathematics without being mathematically to complex and challenging.
We will later see, that more realistic multi period models can indeed be
obtained by the concatenation of many single period models. Single period
models are therefore the building blocks of more complicated models.

Single period market models are the atoms of Financial

Mathematics.

Within this chapter, we assume that we have a finite sample space

Ω := {ω1 , ω2 , ..., ωk }.

We think of the samples ωi as possible states of the world at time t = 1.

The prices of the financial assets we are modeling in a single period model,

2
depend on the state of the world at time t = 1 and therefore on the ωi ’s.
The exact state of world at time t = 1 is unknown at time t = 0. We cannot
foresee the future. We assume however that we are given information about
the probabilities of the various states. More precisely we assume that we
have a probability measure P on Ω with P(ω) > 0 for all ω ∈ Ω. This
probability measure represents the beliefs of the agent. Different agents may
have different beliefs and therefore different P’s. However in the following
we choose one agent who is in a way a representative agent.

1.1 The most elementary market model

The most elementary but still interesting market model occurs when we
assume that Ω contains only two states. We denote these two states by
ω1 = H and ω2 = T . We think of the state at time t = 1 as determined by
the toss of a coin, which can result in Head or Tail,

Ω = {H, T }.

The result of the coin toss is not known at time t = 0 and is therefore
considered as random. We do not assume that the coin is a fair coin,
i.e. that H and T have the same probability, but that there is a number
0 < p < 1 s.t.

P(H) = p, P(T ) = 1 − p.

We consider a model, which consists of one stock and a money market

account. If we speak of one stock, we actually mean one type of stock, for
example Coca Cola, and agents can buy or sell arbitrary many shares of
this stock. For the money market account we think of a savings account.
The money market account pays a deterministic (non random) interest rate
r > 0. This means that one pound invested into the money market account
at time t = 0 yields a return of 1 + r pounds at time t = 1. The price of
the stock at time t = 0 is known and denoted by S0 . The price of the stock
at time t = 1 depends on the state of the world and can therefore take the

3
two values S1 (H) and S1 (T ), depending whether the coin toss results in H
or T . It is not known at time t = 0 and therefore considered to be random.
S1 is a random variable, taking the value S1 (H) with probability p and the
value S1 (T ) with probability 1 − p. We define

S1 (H) S1 (T )
u := , d := .
S0 S0

We assume that 0 < d < u. This means that the stock price can either
go up or down, but in any case remains positive. The stock can then be
represented by the following diagram:

Su
p8 0
p ppppp
p
ppp
ppp
S0 NNN
NNN 1−p
NNN
NNN
N&
S0 d
To complete our first market model we still need trading strategies. The
agents in this model are allowed to invest in the money market account
and the stock. We represent such an investment by a pair (x, φ) where x
gives the total initial investment in pounds at time t = 0 and φ denotes
the numbers of shares bought at time t = 0. Given the investment strategy
(x, φ), the agent then invests the remaining money x − φS0 in the money
market account. We assume that φ can take any possible value, i.e. φ ∈ R.
This allows for example short selling as well as taking arbitrary high credits.
At the end of this section we will give some remarks on the significance of
these assumptions.
The value of the investment strategy (x, φ) at time t = 0 is clearly x,
the initial investment. The agent has to pay x pounds in order to buy the
trading strategy (x, φ). Within the period, meaning between time t = 0
and time t = 1 the agent does nothing but waiting until time t = 1. The
value of the trading strategy at time t = 1 is given by its payoff. The payoff
however depends on the value of the stock at time t = 1 and is therefore

4
random. In fact it can take two values:

V (x, φ)(H) = (x − φS0 )(1 + r) + φS1 (H)

if the coin toss results in H or

V (x, φ)(T ) = (x − φS0 )(1 + r) + φS1 (T )

if the coin toss results in T . We combine these two equations in the following
definition.
Definition 1.1.1. The value process of the trading strategy (x, φ) in our
elementary market model is given by (V0 (x, φ), V1 (x, φ)) where V0 (x, φ) = x
and V1 is the random variable

V1 (x, φ) = (x − φS0 )(1 + r) + φS1 .

An essential feature of an efficient market is that if a trading strategy can

turn nothing into something, then it must also run the risk of loss.
Definition 1.1.2. An arbitrage is a trading strategy that begins with no
money, has zero probability of losing money, and has a positive probability
of making money.
This definition is mathematically not precise. It does not refer to the specific
model we are using, but it gives the basic idea of an arbitrage in words. A
more mathematical definition is the following:
Definition 1.1.3. A trading strategy (x, φ) in our elementary market model
is called an arbitrage, if
1. x = V0 (x, φ) = 0 (i.e. the trading strategy needs no initial investment)
2. V1 (x, φ) ≥ 0 (i.e. there is no risk of losing money)
3. E(V1 (x, φ)) = pV1 (x, φ)(H) + (1 − p)V1 (x, φ)(T ) > 0 (i.e. a strictly
positive payoff is expected).

5
 A mathematical model that admits arbitrage cannot be used for analysis.
Wealth can be generated from nothing in such a model. Real markets some-
times exhibit arbitrage, but this is necessarily fleeting; as soon as someone
discovers it, trading takes actions that remove it. We say that a model is
arbitrage free, if there is no arbitrage in the model. To rule out arbitrage
in our elementary model we must assume that d < 1 + r < u, as we will see
now:

If d ≥ (1 + r), then the following strategy is an arbitrage:

• begin with zero wealth and at time zero borrow S0 from the money
market in order to buy one share of the stock.
Even in the worst case of a tail on the coin toss, i.e. S1 = S0 d, the stock
at time one will be worth S0 d ≥ S0 (1 + r), enough to pay off the money
market debt and the stock has a positive probability of being worth strictly
more since u > d ≥ 1 + r, i.e. S0 u > S0 (1 + r).

If u ≤ 1 + r, then the following strategy is an arbitrage:

• sell one share of the stock short and invest the proceeds S0 in the money
market
Even in the best case for the stock, i.e. S1 = S0 u the cost S1 of replacing it
at time one will be less than or equal to the value S0 (1 + r) of the money
market investment, and since d < u ≤ 1 + r, there is a positive probability
that the cost of replacing the stock will be strictly less than the value of the
money market investment.

We have therefore shown:

No arbitrage ⇒ d < 1 + r < u.

The converse is also true:

d < 1 + r < u ⇒ No arbitrage.

6
The proof of this is left as an exercise. It will also follow from the coming
discussion in Section 1.2. However, from this we get our first proposition:
Proposition 1.1.4. The elementary single period market model discussed
above is arbitrage free, if and only if d < 1 + r < u.
Let us now introduce other financial assets into our elementary market
model:
Definition 1.1.5. A European call option is a contract which gives its
buyer the right (but not the obligation) to buy a good at a future time T
for a price K. The good, the maturity time T and the strike price K are
specified in the contract.
Definition 1.1.6. A European put option is a contract which gives its
buyer the right (but not the obligation) to sell a good at a future time T
for a price K. The good, the maturity time T and the strike price K are
specified in the contract.
We will consider such European options in all of our financial market
models, which we are going to discuss in this lecture. European call and
put options are frequently traded on financial markets. A central question
will always be:

What price should such a European call/put option have?

Within our elementary market model we do not have too many choices.
First, we assume that the good is the stock, and second that the maturity
time is T = 1, the end of the period. This is the only nontrivial maturity
time. The owner of a European call option can do the following:
• if the stock price S1 at time 1 is higher than K, buy the stock at time
t = 1 for the price K from the seller of the option and immediately sell
it on the market for the market price S1 , leading to a profit of S1 − K.
• if the stock price at time 1 is lower than K, then it doesn’t make sense
to buy the stock for the price K from the seller, if the agent can buy
it for a cheaper price on the market. In this case the agent can also do
simply nothing, leading to a payoff of 0.

7
This argumentation shows, that a European call option is equivalent to an
asset which has a payoff at time T = 1 of
max(S1 − K, 0).
This payoff is what the option is worth at time t = 1. Still the question is,
how much the option is worth at time t = 0? We will answer this question
in the remaining part of this section, by applying the replication principle.
To do this, we consider a more general option, which is of the type h(S1 ),
where h : R → R is a function. Please note that since S1 is random, h(S1 )
is also random. The European call option from above, is then given by
choosing the function h(x) := max(x − K, 0). (What function should be
taken for a European put option?) There are many other possible choices
for h leading to different options. We will discuss some of them in a later
section. The replication principle says the following:

Replication principle: If it is possible to find a trading strategy

which perfectly replicates the option, meaning that the trading
strategy guarantees exactly the same payoff as the option at
maturity time, then the price of this trading strategy must coincide
with the price of the option.

What would it be, if the replication principle did not hold? Assume that
the price of the option is higher than the price of a replicating strategy.
Then, with zero initial investment, one could sell the option and buy the
replicating strategy. Since one earns more from selling the option than
paying for the replication strategy, one has a positive amount of money at
hand at time t = 0. This money can then be used to invest into the savings
account (or another riskless asset). At maturity, one may have to pay the
obligation from the option, but the replicating strategy which one owns will
pay exactly for this obligation. On the other hand, one obtains a strictly
positive payoff at maturity time of the option from the money invested in
the savings account. This is an arbitrage.
If the price of the option were lower than the price of the replicating
strategy then a similar strategy as above, where the option is bought and the
replicating strategy is sold short would lead to an arbitrage. We therefore

8
see that under the assumption that there is no arbitrage in the market the
price of the replicating strategy is the only possible price for the option. Let
us formulate these ideas more precisely in the elementary model.
Definition 1.1.7. A replicating strategy or hedge for the option h(S1 )
in the elementary single period market model is a trading strategy (x, φ)
which satisfies V1 (x, φ) = h(S1 ), which is equivalent to
(x − φS0 )(1 + r) + φS1 (H) = h(S1 (H)), (1.1)
(x − φS0 )(1 + r) + φS1 (T ) = h(S1 (T )). (1.2)
The following proposition follows from the argumentation above:
Proposition 1.1.8. Let h(S1 ) be an option in the elementary single period
market model, and let (x, φ) be a replicating strategy for h(S1 ), then x is the
only price for the option at time t = 0 which does not introduce arbitrage.
One way to find a price for an option is therefore to look for a replicat-
ing strategy and take the initial investment for this replicating strategy as
the price. How to find the replicating strategy and does it always exist?
Equations (1.1) and (1.2) represent a system of two linear equations with
two unknown variables x and φ. If we manage to solve it, the result will be
the sougth for replicating strategy. Solving for φ is easy. We subtract (1.2)
from (1.1) and obtain
h(S1 (H)) − h(S1 (T ))
φ= . (1.3)
S1 (H) − S1 (T )
We can now substitute this values for φ and solve for x. The formula (1.3)
is often called the Delta hedging formula.
Although we have computed the replicating strategy, we would like to
take an opportunity to introduce in this simple setting a method that is
suitable for more complicated models. Let us rewrite equations (1.1) and
(1.2) in a different form:
 
1 1
x+φ S1 (H) − S0 = h(S1 (H)), (1.4)
1+r 1+r
 
1 1
x+φ S1 (T ) − S0 = h(S1 (T )). (1.5)
1+r 1+r
9
Now we define
1+r−d
p̃ := . (1.6)
u−d
It follows from the assumption d < 1 + r < u that 0 < p < 1. We have
1+r−d
1 − p̃ = 1 −
u−d
u − d − (1 + r − d)
=
u−d
u−1−r
= .
u−d
For this choice of p̃ we have
1+r−d
1 + u−1−r
u−d S0 u u−d S0 d
(p̃S1 (H) + (1 − p̃)S1 (T )) =
1+r 1+r
(1 + r − d)u + (u − (1 + r))d)
= S0
(u − d)(1 + r)
= S0 .
Let us multiply equation (1.4) by p̃ and equation (1.5) by 1 − p̃. By ding
them we obtain
 1
 1 


x+φ p̃S1 (H)+(1−p̃)S1 (T ) −S0 = p̃ h S1 (H) +(1−p̃)h S1 (T ) ,
1+r 1+r
which by the choice of p̃ and the equation above is equivalent to


1


x= 1+r p̃ h S1 (H) + (1 − p̃)h S1 (T ) . (1.7)

Either from (1.3) or from (1.7) we see that under the condition u > d (we
assumed it at the very beginning) we can always find a replicating strategy
for an option in our elementary single period market model. Models which
have this property are called complete. We will soon see, that there are
also models which are not complete and where the technique of pricing by
the replication principle does not work.
It is interesting to note, that the price x for the option computed above,
does not depend on the probabilities p and 1 − p. In particular it does not

10
coincide with the discounted expectation of the payoff of the option using
the probability measure P, i.e. in general
 
1 1 


x 6= EP h(ST ) = p h S1 (H) + (1 − p)h S1 (T ) .
1+r 1+r
The latter equation is only true if p = p̃ and hence also 1 − p = 1 − p̃,
or if h(S1 (H)) = h(S1 (T )) i.e. the payoff of the option is deterministic
(non random). This however is hardly the case in reality. On the other
hand, if we define another measure P̃ on the underlying probability space
Ω = {H, T } by
P̃(H) = p̃, P̃(T ) = 1 − p̃
then by taking expectations under the measure P̃ instead of P we conclude
from equation (1.7) that
 
1
x = EP̃ h(ST ) . (1.8)
1+r

The measure P̃ is often called a risk neutral measure, since under this
measure the option price depends only on the expectation of the payoff, not
on its riskiness. Such measures will play a major role in the following lecture.
As we will see risk neutral measures, or equivalent martingale measures how
they are also called, will enable us to compute prices for options in complete
and incomplete markets. Notice that formula (1.3), that is the first pricing
method we described, is inherently related to the replication principle which
is not applicable in incomplete markets.
We devote the rest of this section to examples.
Example 1.1.9. Assume the parameters in our elementary market model
are the following: r = 13 , S0 = 1, u = 2, d = 12 , p = 43 . We want to compute
the price of a European call option with the strike price K = 1 and the
maturity time t = 1. In this case
1 + 31 − 1
2 5
p̃ = =
2 − 12 9

11
and we obtain the price of the option x as
 
1 5 4 15
x= · (2 − 1) + 0 = .
1 + 13 9 9 36
Again, the example shows that the value of the probability p is completely
irrelevant for the computation of the price of the option. Only the risk
neutral probability p̃ matters. We can also verify that the number computer
with p instead of p̃ is different:
 
1 3 1 9
1 · (2 − 1) + 0 = .
1+ 3 4 4 16

Example 1.1.10. Using the same parameters as in the previous example,
we compute the price of a European put option, which has the following
payoff
h(S1 ) := max(K − S1 , 0).
The price of the European put is then given by
 
1 5 4 1 6 1
x= · 0 + · (1 − ) = =
1 + 31 9 9 2 36 6

There is an interesting relationship between the European call consid-
ered in Example 1.1.9 and the European put considered in Example 1.1.10.
Clearly
15 1 1 1
− = =1−
36 6 4 1 + 13
or in words
1
price of call − price of put = S0 − K. (1.9)
1+r
This relationship not only holds for the special parameters chosen in the
Examples but is true in general whenever the underlying model is arbitrage
free. Formula (1.9) is called the Put-Call parity.

12
1.2 A general single period market model
We will now consider a general single period market model, in which an
agent is allowed to invest in a money market account (i.e. savings account)
and a finite number of stocks S 1 , ..., S n . The price of the i-th stock at time
t = 0, resp. t = 1, is denoted by S0i , resp. S1i . The money market account is
modeled in exactly the same way as in section 1.1. The prices of the stocks
at time t = 0 are known but the prices at time t = 1 are not known at time
t = 0 and are considered to be random. We assume that the world at time
t = 1 can take up one of k states ω1 , ..., ωk which we all put together into a
set Ω, i.e.
Ω := {ω1 , ..., ωk }. (1.10)

We assume that there is a probability measure P defined on Ω. This measure

tells us about the likelihood P(ωi ) of the world being in the the i-th state
at time t = 1 (as seen from time t = 0). The stock prices S1i can therefore
be considered as random variables

S1i : Ω → R.

Then S1i (ω) denotes the price of the i-th stock at time t = 1 if the world is
in state ω ∈ Ω at time t = 1. For technical reasons, we assume that each
state at time t = 1 is possible, i.e.
P(ω) > 0 for all ω ∈ Ω.
Notice that we can always remove from Ω the states of the world that have
probability of ocurring equal zero (i.e. they are impossible) and conse-
quently we obtain a model that satisfies the above requirement.
Let us now formally define the trading strategies which our agents are
going to use.
Definition 1.2.1. A trading strategy for an agent in our general single
period market model is a pair (x, φ), where x is the initial total investment
at time t = 0 and φ = (φ1 , ..., φn ) ∈ Rn is an n-dimensional vector specifying
the number of shares φi of the i-th stock that we own at t = 0.

13
 Given a trading strategy (x, φ) as above, we always assume that the rest
of the money:
X n
x− φi S0i
i=1
is invested in the money market account. As in Section 1.1 we define the
value process corresponding to a trading strategy.
Definition 1.2.2. The value process of the trading strategy (x, φ) in our
general single period market model is given by (V0 (x, φ), V1 (x, φ)), where
V0 (x, φ) = x and V1 (x, φ) is the random variable
Pn i i

Pn i i
V1 (x, φ) = x − i=1 φ S0 (1 + r) + i=1 φ S1 . (1.11)

It is often useful to consider an additional process, the so called gains

process G(x, φ), which in a single period market model consists only of
one random variable which is defined by
Pn i i

Pn i i
G(x, φ) := x − i=1 φ S0 r+ i=1 φ ∆S .

In this equation ∆S i represents the change in price of the i-th stock, i.e.

∆S i := S1i − S0i . (1.12)

As the name indicates, G represents the gains (or losses) the agent obtains
from his investment. A simple calculation gives

V1 (x, φ) = V0 (x, φ) + G(x, φ). (1.13)

Note that this is an equation of random variables, meaning that this equa-
tion holds in any possible state the world might attend at time t = 1, i.e.
for all ω ∈ Ω. Equation (1.13) says that any change in the value of the
trading strategy must be due to a gain or loss in the investment and not,
for example, due to the addition of funds from outside sources.

14
 It is often convenient to study the prices of the stocks in relation to the
money market account. For this reason we introduce the discounted stock
prices Ŝti defined as follows:

Ŝ0i := S0i ,
1 i = 1, . . . , n.
Ŝ1i := S1i ,
1+r

We also define the discounted value process corresponding to the trading

strategy (x, φ) via

V̂0 (x, φ) := x
n
X n
X
i
V̂1 (x, φ) := (x − φ S0i ) + φi Ŝ1i
i=1 i=1

as well as the discounted gains process Ĝ(x, φ) via

Pn i i
Ĝ(x, φ) := i=1 φ ∆Ŝ (1.14)

with ∆Ŝ i = Ŝ1i − Ŝ0i . The verification of the following two equations is left
as an exercise:
Vt
V̂t = (1.15)
Bt
for t ∈ {0, 1} with B0 = 1 and B1 = 1 + r as well as
V̂1 (x, φ) = V̂0 (x, φ) + Ĝ(x, φ). (1.16)
Example 1.2.3. We consider the following model featuring two stocks S 1
and S 2 as well as the states Ω = {ω1 , ω2 , ω3 }. The prices of the stocks at
time t = 0 are given by S01 = 5 and S02 = 10 respectively. At time t = 1 the
prices depend on the state ω and are given by the following table:
ω 1 ω2 ω3
60 60 40
S11 9 9 9
40 80 80
S12 3 9 9

15
We assume that the interest rate r is equal to 91 . Let us consider a trading
strategy (x, φ), with φ = (φ1 , φ2 ) ∈ R2 . Then
1
V1 (x, φ) = (x − φ1 · 5 − φ2 · 10)(1 + ) + φ1 S11 + φ2 S12
9
and depending on the state of the world:
1 60 1 40 2
V1 (x, φ)(ω1 ) = (x − 5φ1 − 10φ2 )(1 + ) + φ + φ
9 9 3
1 60 1 80 2
V1 (x, φ)(ω2 ) = (x − 5φ1 − 10φ2 )(1 + ) + φ + φ
9 9 9
1 40 1 80 2
V1 (x, φ)(ω3 ) = (x − 5φ1 − 10φ2 )(1 + ) + φ + φ
9 9 9
The increments ∆S i are given by the following table
ω1 ω2 ω3
5 5
∆S 1
3 3 − 95
10
∆S 2 3 − 10
9 − 10
9
and the gains process G by
1 5 10
G(x, φ)(ω1 ) = (x − 5φ1 − 10φ2 ) + φ1 + φ2
9 3 3
1 5 10
G(x, φ)(ω2 ) = (x − 5φ1 − 10φ2 ) + φ1 − φ2
9 3 9
1 5 10
G(x, φ)(ω3 ) = (x − 5φ1 − 10φ2 ) − φ1 − φ2
9 9 9
The discounted prices of the stock at time t = 1:
ω 1 ω2 ω3
Ŝ11 6 6 4
Ŝ12 12 8 8
The discounted value process at time t = 1:
V̂1 (x, φ)(ω1 ) = (x − 5φ1 − 10φ2 ) + 6φ1 + 4φ2
V̂1 (x, φ)(ω2 ) = (x − 5φ1 − 10φ2 ) + 6φ1 + 8φ2
V̂1 (x, φ)(ω3 ) = (x − 5φ1 − 10φ2 ) + 4φ1 + 8φ2

The increments of the discounted stock prices ∆Ŝ i :

16
 ω1 ω2 ω3
∆Ŝ11 1 1 −1
∆Ŝ12 2 −2 −2

The discounted gains process Ĝ:

Ĝ(x, φ)(ω1 ) = 1φ1 + 2φ2

Ĝ(x, φ)(ω2 ) = 1φ1 − 2φ2
Ĝ(x, φ)(ω3 ) = −1φ1 − 2φ2


Given the definition of the wealth process (1.11) in the general single
period market model, the definition of an arbitrage in this model looks
almost the same as in Definition 1.1.2:
Definition 1.2.4. A trading strategy (x, φ) in our general single period
market model is called an arbitrage, if
1. x = V0 (x, φ) = 0
2. V1 (x, φ) ≥ 0
Pk
3. E(V1 (x, φ)) = i=1 P(ωi )V1 (x, φ)(ωi ) >0
Here V1 (x, φ) is given by equation (1.11).
The following remark is often very helpful:
Remark 1.2.5. Condition 3. in Definition 1.2.4 is equivalent to:
3’. There exists ω ∈ Ω such that V1 (x, φ)(ω) > 0.
The definition of an arbitrage can also be formulated with the discounted
value process or with the discounted gains process. The following proposi-
tion gives us such a statement.
Proposition 1.2.6. A trading strategy (x, φ) in the general single period
market model is an arbitrage if and only if one of the following two equivalent
conditions hold:

17
 • 1. x = V̂0 (x, φ) = 0
2. V̂1 (x, φ) ≥ 0
3. E(V̂1 (x, φ)) > 0 or equivalently: there exists ω ∈ Ω s.t. V̂1 (x, φ)(ω) >
0.
• 1. x = V̂0 (x, φ) = 0
2. Ĝ(x, φ) ≥ 0
3. E(Ĝ(x, φ)) > 0 or equivalently: there exists ω ∈ Ω s.t. Ĝ(x, φ)(ω) >
0.
We are now coming back to the subject of risk neutral measures which
we shortly mentioned in Section 1.1.
Definition 1.2.7. A measure P̃ on Ω is called a risk neutral probability
measure for our general single period market model if
1. P̃(ω) > 0 for all ω ∈ Ω,
2. EP̃ (∆Ŝ i ) = 0 for i = 1, . . . , n.
Another way of formulating the second condition in Definition 1.2.7 is:
 
1
EP̃ i
S1 = S0i .
1+r
Notice that when Ω consists of only two elements and there is only one
tradable stock (n = 1) Definition 1.2.7 is consistent with what we called
a risk neutral measure in Section 1.1. We also mentioned before that risk
neutral measures are closely related to arbitrage and pricing of options. The
following theorem, apparently contrary to its title, addresses first of these
questions:
Theorem 1.2.8. Fundamental Theorem of Asset Pricing: In the
general single period market model, there are no arbitrages if and only if
there exist a risk neutral measure.
The proof of this theorem is essentially geometric and requires some
preparation. First of all it is very useful to think of random variables on Ω

18
as vectors in the k dimensional euclidean space Rk . This is possible by the
following identification:

X ⇔ (X(ω1 ), X(ω2 ), ..., X(ωk ))> ∈ Rk

The identification means that every random variable can be interpreted as

a vector in Rk and, vice versa, every vector in Rk defines a random variable
on Ω. We can therefore identify the set of random variables on Ω with the
set Rk . A probability measure Q on Ω can as well be identified with a vector
in Rk . The identification is formally identical to the one above:

Q ⇔ (Q(ω1 ), Q(ω2 ), ..., Q(ωk )) ∈ Rk

The difference to the situation above, when we identify random variables

with vectors, is that probability measures yield vectors (X1 , ..., Xk ) ∈ Rk
with the following two properties:
1. Xi ≥ 0 for all i = 1, ..., k
2. ki=1 Xi = 1.
P

These two properties follow from the properties Q(ω) ≥ 0 for all ω ∈ Ω,
Q(Ω) = 1 and Q(A ∪ B) = Q(A) + Q(B) for disjoint sets A and B, which
every probability measure has to satisfy. The subset of Rk consisting of
the vectors with properties 1. and 2. above is often called the standard
simplex in Rk . Although every vector in Rk defines a probability measure,
we can identify the set of probability measures on Ω with the standard
simplex in Rk .
In the following we will always use this identification, writing X for the
vector representing the random variable X and Q for the vector representing
the probability measure Q. Using this interpretation we can for example
write the expectation value of a random variable with respect to a prob-
ability measure Q on Ω as a scalar product in the euclidean space Rk as
follows:
EQ (X) = ki=1 X(ωi )Q(ωi ) =< X, Q >,
P

19
where < ·, · > denotes the standard scalar product in Rk . Let us now
consider the following set:

W = {X ∈ Rk |X = Ĝ(x, φ) for some trading strategy (x, φ)}. (1.17)

One should think of the elements of W as the possible discounted values at

time t = 1 of trading strategies starting with an initial investment x = 0.
Note that W is a linear subspace of Rk . Next we consider the set

A = {X ∈ Rk |X ≥ 0, X 6= 0}. (1.18)

This is the nonnegative orthant in Rk , since by X ≥ 0 for a vector X we

mean that each coordinate of X is greater or equal to zero. So with the help
of the above notation we can write the no-arbitrage property differently:

no arbitrage ⇔ W ∩ A = ∅. (1.19)

The elements in W ∩ A are exactly the discounted values of the arbitrages

at time t = 1. Let us now consider the orthogonal complement of W which
is given by

W⊥ = {Y ∈ Rk | < X, Y >= 0 for all X ∈ W}. (1.20)

Furthermore we define
Pk
P + = {X ∈ Rk | i=1 Xi = 1, Xi > 0}. (1.21)

This set can be identified with the set of probability measures on Ω which
satisfy property 1. from Definition 1.2.7. We have the following lemma:
Lemma 1.2.9. A measure P̃ is a risk neutral probability measure on Ω if
and only if P̃ ∈ P + ∩ W⊥ .
Proof. Since there is ”if and only if” in the statement of the lemma we have
to prove two assertions:

20
(a) If a measure P̃ is a risk neutral probability measure on Ω then P̃ ∈
P + ∩ W⊥ .
(b) If P̃ ∈ P + ∩W⊥ then the measure P̃ is a risk neutral probability measure
on Ω.
We start with (a). It is indicated by saying: assume first that P̃ is a risk
neutral probability measure on Ω. Then by property 1. in Definition 1.2.7
P̃ ∈ P + . Using property 2. in Definition 1.2.7 as well as the definition of
the discounted gains process Ĝ(x, φ) in (1.14) we have for X = Ĝ(x, φ) ∈ W
k
! k
X X  
i i i i
< X, P̃ >= EP̃ (Ĝ(x, φ)) = EP̃ φ ∆Ŝ = φ EP̃ ∆Ŝ = 0
i=1 i=1 | {z }
=0

and therefore P̃ ∈ W⊥ . Together with the first part this gives P̃ ∈ P + ∩ W⊥ .

On the other hand if P̃ is an arbitrary vector in P + ∩ W⊥ (we are now
proving (b)!) then P̃ defines a probability measure satisfying condition 1.
in Definition 1.2.7. For a given i = 1, .., k consider the trading strategy
”buy one share of the asset i”, i.e. the strategy (x, φ) with x = S0i and
φ = (0, ..., 0, 1, 0, ...0) with the sole 1 is in the i-th position. The discounted
gains process of this strategy clearly satisfies Ĝ(x, φ) = ∆Ŝ i . By definition
Ĝ(x, φ) ∈ W and since P̃ ∈ W⊥ we have
 
i
0 =< Ĝ(x, φ), P̃ >= EP̃ ∆Ŝ .

Therefore P̃ also satisfies condition 2. in Definition 1.2.7, which completes

the proof.
Definition 1.2.10. We denote by M = W⊥ ∩ P + the set of risk neutral
measures.
We are now ready to prove Theorem 1.2.8.
Proof of Theorem 1.2.8. Notice that in the statement of the theorem there
is ”if and only if”, so we have to prove two assertions as in Lemma 1.2.9.
Assume first that the no arbitrage condition holds. Let us define the set
A+ = {X ∈ A|hX, Pi = 1}

21
It is a closed, bounded and convex subset of RK . Since A+ ⊂ A it follows
from (1.19) that
no arbitrage ⇒ W ∩ A+ = ∅.
By the separating hyperplane theorem, there exists a vector Y ∈ W⊥ , s.t.
hX, Y i > 0 for all X ∈ A+ . (1.22)
For each i = 1, ..., k define the vector X i as a vector in Rk whose i-th
coordinate is equal to 1/P(ωi ) and the remaining components are zero. Then
1
hX i , Pi = P(ωi ) = 1
P(ωi )
and hence X i ∈ A+ . Denoting with Yi the i-th component of Y it then
follows from (1.22), that
1
0 < hX i , Y i = Yi
P(ωi )
and therefore Y (ωi ) = Yi > 0 for all i = 1, .., k. Let us now define Q by
Y (ωi )
Q(ωi ) = .
Y (ω1 ) + ... + Y (ωk )
Clearly, Q ∈ P + . Since Q is merely a scalar multiple of Y and W⊥ is a
vectorspace it follows that Q ∈ W⊥ . Therefore
Q ∈ P + ∩ W⊥
and by Lemma 1.2.9 we have that Q is a risk neutral measure on Ω. We
have therefore shown, that the condition no arbitrage implies the existence
of a risk neutral measure.
Let us now show the converse. We assume there exists a risk neutral
measure Q. Let (x, φ) be an arbitrary trading strategy. As in the proof of
Lemma 1.2.9
k
! k
X X  
i i i i
EQ (Ĝ(x, φ)) = EQ φ ∆Ŝ = φ EQ ∆Ŝ = 0.
i=1 i=1

If we assume that Ĝ(x, φ) ≥ 0 then the last equation clearly implies that
Ĝ(x, φ)(ω) = 0 for all ω ∈ Ω. Hence by Proposition 1.2.6 there cannot be
any trading strategy that is an arbitrage.

22
Example 1.2.11. We continue with example 1.2.3. Recall the increments
of the discounted prices:
ω1 ω2 ω3
∆Ŝ 1 1 1 −1
∆Ŝ 2 2 −2 −2

By definition of W and Ĝ(x, φ) = φ1 ∆Ŝ 1 + φ2 ∆Ŝ 2 it therefore follows that

 1 2
 
 φ + 2φ 
1 2  1 2
W=  φ − 2φ φ ,φ ∈ R .
1 2
−φ − 2φ
 

Note that for any vector X ∈ W we have X1 + X3 = 0, where Xi once again

denotes the i-th coordinate of the vector X. On the other hand, whenever
we have a vector X ∈ R3 s.t. X1 + X3 = 0, we can choose φ1 = 12 (X1 + X2 )
and φ2 = 41 (X1 − X2 ) and obtain
  1
φ + 2φ2
 
X1
X =  X2  =  φ1 − 2φ2  .
X3 −φ1 − 2φ2
Hence
W = {X ∈ R3 |X1 + X3 = 0}.
It is then easy to see that the orthogonal complement of W is given by
W⊥ = {Y ∈ R3 |Y = (λ, 0, λ)T , λ ∈ R},
which clearly implies that W⊥ ∩P + = ∅ and therefore there is no risk neutral
measure for this model. By Theorem 1.2.8 there must be an arbitrage
strategy in the model. By comparing W and A we can see that
W ∩ A = {X ∈ R3 |X1 = X3 = 0, X2 > 0}.
Let us compute an arbitrage strategy: we start with any positive number
X2 > 0. Since  
0
 X2  ∈ W ∩ A
0

23
we know that there must be a trading strategy (x, φ) s.t
 
0
Ĝ(x, φ) =  X2  .
0
We compute it by solving the following system of linear equations:
φ1 + 2φ2 = Ĝ(x, φ)(ω1 ) = 0
φ1 − 2φ2 = Ĝ(x, φ)(ω2 ) = X2
−φ1 − 2φ2 = Ĝ(x, φ)(ω3 ) = 0
where the last equation above is redundant (when multiplied by −1 it is
identical to the first one). The solution is:
X2 X2
φ1 = , φ2 = − .
2 4
These numbers tell us how many shares we ought to buy in order to obtain
the arbitrage. We still need to know how much money we have to invest in
the money market account. This is easy. Since the arbitrage starts with an
initial total investment of zero, we must invest
X2 X2
0 − φ1 S01 − φ2 S02 = − · 5 − (− ) · 10 = 0
2 4
in the money market. The arbitrage we computed is therefore a strategy
which only invests in the risky assets, i.e the stocks. 
We now come back to the question: what should the price of an option
in our model be? In Section 1.1 we considered options of the type h(S1 ),
where h is a payoff profile, a function of a single stock S1 at time t = 1. In
our general model we have more than one stock and the payoff profiles may
look more complicated. For this reason we generalize our definition of an
option. We call this more general product contingent claim.
Definition 1.2.12. A contingent claim in the general single period mar-
ket model is a random variable X on Ω representing a payoff at time t = 1.
To price a contingent claim, we may follow the same approach as taken
in Section 1.1 and apply the replication principle.

24
Proposition 1.2.13. Let X be a contingent claim in our general single
period market model, and let (x, φ) be a hedging strategy for X, i.e. a
trading strategy which satisfies V1 (x, φ) = X. The only price of X which
complies with the no arbitrage principle is V0 (x, φ), which by definition is
equal to x.
The proof of this proposition follows along the same argumentation as in
Section 1.1. A crucial difference to the elementary single period model as
discussed in Section 1.1 is however that in the general single period market
model a replicating strategy might not exist. This can happen when there
are more effective sources of randomness than there are stocks to invest in.
Let us consider the following example that represents an elementary version
of a so called stochastic volatility model.
Example 1.2.14. We consider the following market model. It consists
of two tradeable assets, one money market account Bt and one stock St
(t = 0, 1), as well as third object which we call the volatility v. The volatility
determines whether the stock price can make big jumps or small jumps.
In this model the volatility is assumed to be random, or in other words
stochastic. Such models are called stochastic volatility models. To be
a bit more precise, we assume that our state space consists of 4 states,
Ω := {ω1 , ω2 , ω3 , ω4 }.
and that the volatility is given by
(
h, if ω = ω1 , ω2
v(ω) :=
l, if ω = ω3 , ω4
Here 0 < l < h < 1 and l stands for low volatility whereas h stands for high
volatility. The stock price S1 is then modeled by
(
(1 + v(ω))S0 , if ω = ω1 , ω3
S1 (ω) :=
(1 − v(ω))S0 , if ω = ω2 , ω4
where S0 denotes the initial stock price. The stock price can therefore jump
up or down as in the first elementary single period market model from
Section 1.1. The difference to that model is that the amount by which it

25
jumps is itself random determined by the volatility. Finally, the money
market account is modeled as before by
B0 = 1, B1 = 1 + r.
Let us now consider a digital call in this model, i.e.
(
1, if S1 > K
X=
0 otherwise
Assume that the strike price K satisfies
(1 + l)S0 < K < (1 + h)S0 .
Then a nonzero payoff is only possible if the volatility is high and the stock
jumps up. This is the case if and only if the state of the world at time t = 1
is given by ω = ω1 . The contingent claim X can therefore alternatively be
written as (
1, if ω = ω1
X(ω) =
0 if ω = ω2 , ω3 , ω4
Let us check if there exists a replicating strategy for this contingent claim,
i.e. a trading strategy (x, φ) satisfying
V1 (x, φ) = X.
Using the definition of V1 (x, φ) and our vector notation for random variables,
the last equation is equivalent to the following system of linear equations:
     
1+r (1 + h)S0 1
 + φ ·  (1 − h)S0  =  0 
 1+r     
(x − φS0 ) · 
 1+r   (1 + l)S0   0 
1+r (1 − l)S0 0
We rearrange the terms and write equivalently:
    

1+r h−r 1
 + φS0 ·  −(h + r)
 1+r    0
x· 1+r   l−r
= 
 0
1+r −(l + r) 0

26
Solving the second equation (the second component) for x yields:
1
x= φS0 (h + r).
1+r
On the other hand, solving the fourth equation (the fourth component) for
x gives:
1
x= φS0 (l + r).
1+r
Since S0 (h + r) 6= S0 (l + r), both equations can only hold together if φ = 0.
This however implies that x = 0 and so we get 0 = 1 in the first equation.
This of course isn’t true. The conclusion is that there is no trading strategy
(x, φ) which replicates X. The heuristic explanation is that there is a source
of randomness in the volatility which cannot be hedged since the volatility
is not tradeable. The mathematical explanation is just that the system of
linear equations above has no solution. 
To take account of this difficulty we introduce the following definition.
Definition 1.2.15. A contingent claim X is called attainable, if there
exists a trading strategy (x, φ) which replicates X, i.e. satisfies V1 (x, φ) =
X.
For attainable contingent claims the replication principle applies and it
is clear how to price them, namely by the total initial investment needed for
a replicating strategy. There might be more than one replicating strategy,
but it follows again from the no arbitrage principle, that the total initial
investment for replicating strategies is unique.
In equation (1.8) we established a way to use a risk neutral measure
to compute the price of an option in the elementary single period market
model. This approach works fine in the general single period market model
as long as attainable contingent claims are considered.
Proposition 1.2.16. Let X be an attainable contingent claim and P̃ be an
arbitrary risk neutral measure. Then the price x of X at time t = 0 defined
via a replicating strategy can be computed by the formula

1


x = EP̃ 1+r X . (1.23)

27
Proof. Let (x, φ) be a replicating strategy of X, i.e. X = V1 (x, φ). It follows
1
from the equality V̂1 (x, φ) = 1+r V1 (x, φ) that
1
X = V̂1 (x, φ)
1+r
and from Definition 1.2.7
 
1  
EP̃ X = EP̃ V̂1 (x, φ)
1+r
 
= EP̃ x + Ĝ1 (x, φ)
k
!
X
= x + EP̃ φi ∆Ŝ i
i=1
k
X  
i i
=x+ φ EP̃ ∆Ŝ
i=1 | {z }
=0
= x.

Remark 1.2.17. Proposition 1.2.16 tells us, in particular, that for all risk
neutral measures for the model we get the same value when computing the
expectation in equation (1.23).
The following example shows that the situation changes dramatically if
the contingent claim is not attainable.
Example 1.2.18. Let us compute the set of risk neutral measures for our
stochastic volatility model from Example 1.2.14. For simplicity we assume
r = 0. In this case the discounted processes and the original processes
coincide. We have

v(ω) · S0 , if ω = ω1 , ω3
∆Ŝ(ω) :=
−v(ω) · S0 , if ω = ω2 , ω4
or in the vector notation
 
h
 −h 
∆Ŝ = S0 · 
 l .


−l

28
For a trading strategy (x, φ) the discounted gains process is given by

Ĝ(x, φ) = φ · ∆Ŝ

and hence the vectorspace W is one dimensional, spanned by the vector

∆Ŝ, i.e.  

 h 

−h
  
W = span    .

 l 
−l
 

The orthogonal complement of W is then given by

      

 q 1 * q 1 h + 

q q −h
 
⊥ 2 2
    
W =  
  q3  ,  l  = 0 .
    

 q 3 
q4 q4 −l
 

On the other hand (q1 , q2 , q3 , q4 )> ∈ P + if and only if q1 + q2 + q3 + q4 = 1

and q1 , q2 , q3 , q4 > 0. Since the set of risk neutral measures is given by
M = W⊥ ∩ P + we find that
 
q1
 ∈ M ⇔ 0 < q1 , q2 , q3 , 0 < q4 = 1 − (q1 + q2 + q3 ),
 q2 

 q3  and h(q1 − q2 ) + l(q3 − (1 − (q1 + q2 + q3 )) = 0.
q4
The set of risk neutral measures in our stochastic volatility model is there-
fore given by:
  

 q 1 

q q > 0, q + q + q < 1
 
2 i 1 2 3

M=   
 h(q1 − q2 ) = l(1 − (q1 + q2 + 2q3 )) 

 q3 
1 − (q1 + q2 + q3 )
 

Clearly, this set is not empty and we can conclude that our stochastic
volatility model is arbitrage free. In particular, as it is easy to see, for
every 0 < q1 < 1 there exists Q ∈ M s.t. Q(ω1 ) = q1 . Let us now

29
compute the (discounted) expectation of a Digital call under a measure
Q = (q1 , q2 , q3 , q4 )> ∈ M:

EQ (X) = q1 · 1 + q2 · 0 + q3 · 0 + q4 · 0 = q1 .

Since q1 is arbitrary, except that it has to lie between zero and one, we
see that we obtain many different values as discounted expectation under a
risk neutral measure, in fact every value x which satisfies 0 < x < 1. The
situation is therefore completely different than in Proposition 1.2.16 because
as we showed in Example 1.2.14 the contingent claim X is not attainable.

Let us now consider a general contingent claim X, attainable or not.
Definition 1.2.19. We say that a price x for the contingent claim X com-
plies with the no arbitrage principle, if the extended model, which
consists of the original stocks S 1 , ..., S n and an additional asset S n+1 whose
prices satisfy S0n+1 = x and S1n+1 = X is arbitrage free.
The additional asset S n+1 defined in the above proposition may not be
interpreted as a stock, since it can take negative values if the contingent
claim takes negative values. For the general arbitrage and pricing theory
developed so far positiveness of asset prices was however not essential.
In view of the previous example the following proposition might be sur-
prising. It says that whenever one uses a risk neutral measure to price a
contingent claim by formula (1.23) one obtains a price which complies with
the no arbitrage principle. Even if prices differ when using different risk
neutral measures.
Proposition 1.2.20. Let X be a possibly unattainable contingent claim and
P̃ a risk neutral measure for our general single period market model. Then
 
1
x = EP̃ X (1.24)
1+r
defines a price for the contingent claim at time t = 0 which complies with
the arbitrage principle.

30
Proof. By the Fundamental Theorem of Asset Pricing (Theorem 1.2.8) it is
enough to show that there exists a risk neutral measure for the correspond-
ing model which is extended by the asset S n+1 as in Definition 1.2.19. By
assumption P̃ is a risk neutral measure for the original model, consisting
of the stocks S 1 , ..., S n , i.e. P̃ satisfies 1. and 2. in Definition 1.2.7 for
i = 1, ..., n. For i = n + 1 the second condition translates into
 
1
EP̃ (∆Ŝ n+1 ) = EP̃ X −x
1+r
 
1
= EP̃ X −x
1+r
= x−x=0

Hence P̃ is a risk neutral measure for the extended model.

In the situation of Example 1.2.18 this proposition applied to the Digital
call says that any price between zero and one is a price which does not
allow arbitrage and can therefore be considered as fair. This non unique-
ness of prices is a serious problem which has not been completely resolved
until today. We will later discuss a few methods which try to address this
issue. For now let us characterize the models, in which the problem of non
uniqueness of prices does not occur:
Definition 1.2.21. A financial market model is called complete, if for any
contingent claim X there exists a replicating strategy (x, φ). A model which
is not complete is called incomplete.
By Proposition 1.2.16 the issue of computing prices in complete market
models is completely solved. But how to recognize complete models? The
following proposition gives us a criterion for completeness.
Proposition 1.2.22. Assume a general single period market model consist-
ing of stocks S 1 , ..., S n and a money market account modeled on the state
space Ω = {ω1 , ..., ωk } is arbitrage free. Then this model is complete if and

31
only if the k × (n + 1) matrix A given by

1 + r S11 (ω1 ) · · · S1n (ω1 )

 
 1 + r S11 (ω2 ) · · · S1n (ω2 ) 
 
 · · · 
A=  ·

 · · 

 · · · 
1 n
1 + r S1 (ωk ) · · · S1 (ωk )

has the rank k, i.e. rank(A) = k.

Proof. The model is complete if one can find a replicating strategy (x, φ)
for any claim X. In other words, the model is complete if for any X ∈ Rk
there exists a solution to the system of linear equations:
Pn
1 + r S11 (ω1 ) · · · S1n (ω1 ) x− iSi
     
i=1 φ 0 V1 (x, φ)(ω1 )

 1 + r S11 (ω2 ) · · · S1n (ω2 ) 


 φ1  
  V1 (x, φ)(ω2 ) 


 · · · 


 ·  
= · 
.

 · · · 


 ·  
  · 

 · · ·   ·   · 
1 n
1 + r S1 (ωk ) · · · S1 (ωk ) φn V1 (x, φ)(ωk )

This shows that computation of a replicating strategy for the contingent

claim X is equivalent to solving the equation AZ = X, where Z is an n + 1
dimentional vector:
x − ni=1 φi S0i
 P 

 φ1 

 · 
Z= .

 · 

 · 
φn
By standard linear algebra the matrix A has the rank k if and only if for
every X ∈ Rk the equation AZ = X has a solution Z ∈ Rn+1 . Hence the
statement of the proposition follows immediately.
Example 1.2.23. We have already seen that the stochastic volatility model
discussed in Examples 1.2.14 and 1.2.18 is not complete. Another way to
see this is by using Proposition 1.2.22. The matrix A in this case has the

32
form  
1+r (1 + h)S0
 1+r (1 − h)S0 
A=
 1+r

(1 + l)S0 
1+r (1 − l)S0
The rank of this matrix is 2 which is not equal to k = 4. Hence it doesn’t
have full rank, and the model is incomplete. 
Proposition 1.2.22 presents a method to determine whether a model is
complete without computing replicating strategies. Now, if the model is
incomplete, is there a method to determine whether a specific contingent
claim is attainable without trying to compute the replicating strategy? Yes,
there is. The following proposition shows how.
Proposition 1.2.24. The contingent claim X is attainable if and only if
1


EQ 1+r X takes the same value for all Q ∈ M.
The proof of this result can be found in Pliska ([1], page 23). We omit the
proof here. An important consequence of this proposition is the following
theorem:
Theorem 1.2.25. Under the assumption that the model is arbitrage free, it
is complete if and only if M consists of only one element, i.e. there is only
one risk neutral measure.
Proof. Since the model is arbitrage free, it follows from Theorem 1.2.8 that
there is at least one risk neutral measure, i.e. M 6= ∅. Assume first that
there is only one risk neutral measure. Then the condition in Proposition
1.2.24 is trivially satisfied for all contingent claims X and so the market
model is complete.
Assume now the market model is complete and there are two risk neutral
measures Q1 and Q2 in M. For each i = 1, .., k consider the contingent claim
X i given by 
i 1 + r, if ω = ωi
X (ω) =
0, otherwise

33
Since the model is complete, X i is an attainable contingent claim. It follows
from Remark 1.2.17 that
   
1 i 1 i
EQ1 X = EQ2 X .
1+r 1+r
But clearly,
   
1 i 1 i
Q1 (ωi ) = EQ1 X , Q2 (ωi ) = EQ2 X ,
1+r 1+r
which together implies that Q1 (ωi ) = Q2 (ωi ). Therefore Q1 = Q2 and we
have hence shown, that M consists of only one risk neutral measure.

34
Chapter 2

Multi-Period Market Models

2.1 General Model Specifications

The two most important new features of multi period market models as
compared to single period market models are:

• Agents can buy and sell assets not only at the beginning of the trading
period, but at any time t out of a discrete set of trading times t ∈
{0, 1, 2, ..., T }. t = 0 is the beginning of the trading period, t = T is
the end.
• Agents can gather information over time, since they can observe prices.
For example they can make their investments at time t = 1 dependent
on the prices of the asset at time 1. These are unknown at time t = 0
and couldn’t be used in order to choose the investment at time t = 0.

These two aspects need special attention. We have to model trading as a

dynamic process as opposed to the static trading approach in single period
market models, and we have to take care about how the level of information
evolves over time. The second aspect leads to the probabilistic concept of
σ-algebras and filtrations. We still assume that we work on a finite state
space Ω on which a probability measure P is defined.
Definition 2.1.1. A collection F of subsets of the state space Ω is called a
σ-algebra if the following conditions hold:
1. Ω ∈ F

35
 2. If F ∈ F then F c = Ω \ F ∈ F
3. If Fi ∈ F for i ∈ N, then ∞
S
i=1 Fi ∈ F.

In our case, where the state space is finite, the third condition can be relaxed
to condition
3’. If F, G ∈ F, then F ∪ G ∈ F.
There are a few properties that can be easily derived from the above
definition:
• ∅ ∈ F,
• If F, G ∈ F, then F ∩ G ∈ F.
To see the first one, notice that by 1. A = Ω ∈ F. By virtue of 2. one
gets Ac ∈ F, but the compliment of the full space is an empty set, so
∅ = Ac ∈ F. Take now F, G ∈ F. It can be easily checked that
F ∩ G = (F c ∪ Gc )c .
This reformulation will be very useful. By 2. we obtain F c , Gc ∈ F. Further,
their sum F c ∪ Gc is an element of F (see 3’.). Applying 2. once more, we
have (F c ∪ Gc )c ∈ F, which completes the proof of the second statement.
The idea is that a σ algebra models a certain level of information. The
elements of a σ-algebra F are subsets of the state space Ω. Given that F
is chosen to model the level of information of an agent or an observer of
any kind, then it is assumed that this agent can distinguish between two
sets F and G which belong to F, but not between the actual elements of
F or the actual elements of G. In a way one could say, that if the agent
looks at the state space Ω his resolution is not high enough to recognise
the actual states ω but only to see the sets belonging to the σ-algebra F.
One can think of the states ω as atoms, and as the sets contained in the
σ-algebra as molecules, which are built from the atoms. The agent can only
see the molecules, but not the atoms. The larger the σ-algebra, the higher
the resolution is, the more information is in the σ-algebra.

A σ-algebra on a finite state space can be decomposed into a so called

partition:

36
Definition 2.1.2. A partition of a σ-algebra F is a collection of sets
{Ai : i ∈ I}, where I is some index set, such that:
1. Ai 6= ∅ for i ∈ I.
2. Ai ∈ F for i ∈ I.
3. EverySset F ∈ F can be written as a union of some of the Ai ’s, i.e.
F = j∈J Aj with J ⊂ I.
4. Sets Ai are pairwise disjoint, i.e. Ai ∩ Aj = ∅ whenever i 6= j.
The set I can be thought of as a set of integers 1, 2, . . . , N for some
N . Therefore, a partition is a collection of sets A1 , A2 , . . . , AN satisfying
conditions 1.-4. of Definition 2.1.2. It can be easily deduced that all sets Ai
in a partition satisfy a certain minimality condition: if A ∈ F and A ⊂ Aj
for some j, then A = Aj , since otherwise it could not be written as a union
of some of the Ai ’s (a contradiction with 3). It is not hard to show that
given a σ-algebra F, a partition for this σ-algebra always exist and is in
fact unique.
Example 2.1.3.
1. If Ω is a finite state space, then the power set P(Ω) is a σ-algebra.
This is the largest possible σ-algebra on Ω, but beware: If Ω is not
finite, then the power set of Ω is not a σ-algebra! The partition of F
consists of sets {ω} for all ω ∈ Ω.
2. The trivial σ-algebra corresponding to a state space Ω is the σ-algebra
F = {∅, Ω}. This clearly satisfies the conditions in Definition 2.1.1.
The trivial σ-algebra is the smallest σ-algebra on Ω. Here, the partition
is composed from one set Ω.
3. On a state space Ω = {ω1 , ω2 , ω3 , ω4 } consisting of four elements, the
following is a σ-algebra:
F = {∅, {ω1 , ω2 }, {ω3 , ω4 }, Ω}.
This can be easily verified. A partition of this σ-algebra, is given by
the two sets A1 = {ω1 , ω2 } and A2 = {ω3 , ω4 }.

37
 4. For a collection Fi ⊂ Ω for i ∈ I, one can show that there is always
the smallest σ-algebra which contains the sets Fi . This σ-algebra is
denoted by σ (Fi : i ∈ I) and is called the σ-algebra generated by the
sets Fi .

Let us now consider a random variable X : Ω → R (we can think of a con-
tingent claim). It is F-measurable if its values depend only on information
in F. This is a general mathematical definition:
Definition 2.1.4. The random variable X : Ω → R is called F-measurable,
if for every closed interval [a, b] ⊂ R the preimage under X belongs to F,
i.e.
X −1 ([a, b]) ∈ F.
It is rather difficult to imagine and check. However, we can provide a
simpler reformulation in our case of finite state space Ω. It is based on the
observation that the set of values of X is finite.
Proposition 2.1.5. Let X : Ω → R be a random variable and (Ai )i∈I a
partition of the σ-algebra F. The random variable X is F-measurable, if
and only if X is constant on each set of the partition, i.e. there exist cj ∈ R
for all j ∈ I such that

X(ω) = cj for all ω ∈ Aj .

Proof. Assume (Ai )i∈I is a partition of the σ-algebra F and that X is F-

measurable. Let j ∈ I be an index and ω ∈ Aj be arbitrary. Define
cj := X(ω). Since X is F-measurable, we have X −1 (cj ) ∈ F and hence by
remarks following the definition of a σ-algebra, we have

6 X −1 (cj ) ∩ Aj ∈ F.
∅=

Since clearly X −1 (cj ) ∩ Aj ⊂ Aj , we have by the aforementioned minimality

property of the sets contained in the partition that

X −1 (cj ) ∩ Aj = Aj .

38
But this means nothing else than X(ω) = cj for all ω ∈ Aj and hence X is
constant on Aj . By varying j we obtain such cj ’s for all j ∈ I.
Now assume vice versa, that X : Ω → R is a function which is constant
on all sets Aj belonging to the partition and that the cj are given as in the
statement of the proposition. Let [a, b] be a closed interval in R and define

C := {cj |j ∈ I and cj ∈ [a, b]}.

S
Since i∈I Ai = Ω no other elements then the cj occur as values of X.
Therefore
[ [
−1 −1 −1
X ([a, b]) = X (C) = X (cj ) = Aj .
j|cj ∈C j|cj ∈C

Finally, due to property 3. of a σ-algebra we have

[
Aj ∈ F.
j|cj ∈C

In the proposition above, it is not assumed that the cj ’s all have different
values. In fact, some of them may coincide.
Example 2.1.6.
1. If Ω is a finite state space and F = P(Ω) is a power σ-algebra, then
every random variable X : Ω → R is F-measurable.
2. A random variable that is measurable with respect to a trivial σ-algebra
is constant, i.e. its set of values consists of exactly one point.
3. On a state space Ω = {ω1 , ω2 , ω3 , ω4 } consider a σ-algebra F given by
the partition A1 = {ω1 , ω2 } and A2 = {ω3 , ω4 }. Then every random
variable that is F-measurable satisfies

X(ω1 ) = c1 , X(ω2 ) = c1 , X(ω3 ) = c2 , X(ω4 ) = c2 ,

for some numbers c1 , c2 .

39
 4. On a finite state space Ω consider a random variable X : Ω → R. A
σ-algebra generated by X is given by a partition
 −1
X (c) : c is a value of X .
This partition is finite, because X can only attain a finite number of
values. Moreover, F is the smallest σ-algebra with respect to which X
is measurable (check it!).

The main idea behind the notion of a filtration is to model the devel-
opment of information in time. A filtration is a sequence of σ-algebras.
We assume that information can only increase in time, i.e. we never forget
anything!
Definition 2.1.7. A sequence (Ft )0≤t≤T of σ-algebras on Ω is called a fil-
tration, if Fs ⊂ Ft whenever s < t.
We interpret the σ-algebra Ft as the information which is available to
an agent or an observer at time t. In particular F0 represents the informa-
tion which is available at the beginning of the period before anything else
has started. This is, in a way, the initial information. Usually we know
everything about the present state of the market, i.e. the prices of assets.
Therefore, we will be assume that F0 is a trivial σ-algebra F0 = {∅, Ω}.
To model a dynamic random behaviour of some quantities (e.g. as-
set prices), we introduce the concept of stochastic processes, which are
sequences of random variables in time.
Definition 2.1.8. A family (Xt ), where 0 ≤ t ≤ T , consisting of random
variables is called a stochastic process. If (Ft )0≤t≤T is a filtration, the
stochastic process (Xt ) is called (Ft )-adapted if for all t we have that Xt
is Ft -measurable.
We shall often write “process” instead of stochastic process. If it is clear
from the context which filtration is meant, we shall also write “adapted”
instead of (Ft )-adapted. The stochastic process (Xt ) is therefore adapted, if
and only if its time t component Xt depends only on the information which
is available at time t, i.e. Ft .

40
Example 2.1.9. An ideal example of a stochastic process is the evolution
of stock prices. We denote the stock price at times 0 ≤ t ≤ T with St . Since
these prices are not known at time 0, based on the initial information, St
are assumed to be random variables for t ≥ 1:
St : Ω → R+ .
At time t however we know the price of St (i.e. it is a part of the information
available to us) and we therefore assume that St is Ft -measurable. We
therefore model the stock price evolution as an adapted process (St ), where
the filtration (Ft ) is otherwise specified. 
In the above example we discussed the requirements for the price process
so that is complies with the information structure we have. It is usually
assumed that the only information we get comes from the prices themselves.
Below, we present a general method how to obtain a filtration, such that
the process in question is measurable with respect to this filtration, and the
filtration represents the information flow which is induced by observing the
stochastic process.
Definition 2.1.10. Let (Xt )0≤t≤T be a stochastic process on (Ω, P). Define

FtX := σ Xu−1 ([a, b]) : 0 ≤ u ≤ t, a ≤ b ∈ R

(2.1)

This is the smallest σ-algebra which contains all the sets Xu−1 ([a, b]) where
0 ≤ u ≤ t and a ≤ b. Clearly FsX ⊂ FtX and (FtX )0≤t≤T is a filtration. It fol-
lows immediately from the definition that (Xt )0≤t≤T is (FtX )0≤t≤T -adapted.
(FtX )0≤t≤T is called the filtration generated by the process X.
Let us make the above definition more understandable. First, (2.1) reads
as
FtX = {the information generated by X0 , X1 , . . . , Xt }.
We know how to extract the information carried by one random variable
(see Example 2.1.9). Here, we have t + 1 random variables. Let Gi be the σ-
algebra generated by Xi , i = 0, 1, . . . , t. Then FtX is the smallest σ-algebra
containing G0 , G1 , . . . , Gt . It means that we put all sets from collections
G0 , . . . , Gt into one bag called G and look for a smallest collection of sets
(Ai )i∈I , for some set of indexes I, that satisfies

41
 1. Ai 6= ∅ for i ∈ I.
2. EverySset G ∈ G can be written as a union of some of the Ai ’s, i.e.
G = j∈J Aj with J ⊂ I.
3. Sets Ai are pairwise disjoint, i.e. Ai ∩ Aj = ∅ whenever i 6= j.
Actually, we can put into G solely the partitions of G0 , . . . , Gt . Notice that
above conditions resemble those from Definition 2.1.2. Indeed, it is true
that FtX is the σ-algebra whose partition is given by (Ai )i∈I .
Let us illustrate these concepts in the following example which models
the stock price evolution over two periods.
Example 2.1.11. The following diagram describes the evolution of a single
stock over the period of two time steps:

S2 = 9 ω1
3
4 rrr
rr9
rr
rrr
S1B = 8L
LLL 1
 LL4L
 LLL
2  %
5 
 S2 = 6 ω2




S0 =: 5
::
::
::
:: 3
::5 S2 = 6 ω3
::
:: 2
7 rrr
rr9
:: rr
: rrr
S1 = 4L
LLL 5
LL7L
LLL
%
S2 = 3 ω4

The underlying probability space is given by Ω = {ω1 , ω2 , ω3 , ω4 }. The num-

bers at the arrows indicate the probabilities by which this move happens.
At time t = 0 the stock price is known, and there is only one possible value
S0 = 5. The σ-algebra F0S is therefore the trivial σ-algebra. At time t = 1

42
the stock can take two possible values. The following is easy to verify:


 Ω if a ≤ 4 and 8 ≤ b
{ω1 , ω2 } if 4 < a and 8 ≤ b

S1−1 ([a, b]) =
{ω , ω } if a ≤ 4 and b < 8
 3 4


∅ if a < 4 and b < 4.
These are all cases which can occur, and therefore

F1S = {∅, {ω1 , ω2 }, {ω3 , ω4 }, Ω}.

The observing agent at time t = 1 is therefore able to decide, whether the

true state of the world belongs to the set {ω1 , ω2 } or to the set {ω3 , ω4 },
but in neither case is he able to decide, whether it is ω1 or ω2 in the first
case, or ω3 or ω4 in the second case. However he learns the true state
of the world at time t = 2. The reason for this is as follows: F2S must
contain the sets F1 := S2−1 ([9, 10]) = {ω1 }, F2 := S2−1 ([6, 9]) = {ω1 , ω2 , ω3 },
F3 := S2−1 ([5, 6]) = {ω2 , ω3 } and F4 := S1−1 ([6, 9]) = {ω1 , ω2 } and since
F2S is a σ-algebra it must therefore also contain: {ω1 } = F4 \ F1 , {ω2 } =
F3 ∩ F4 , {ω3 } = F2 \ F4 and {ω4 } = F2c .
Let us now use the second method to obtain the above filtration. First,
if t = 0, F0S is generated by S0 , so it is trivial. For t = 1 denote by G0
a partition of the σ-algebra generated by S0 , and by G1 a partition of the
σ-algebra generated by S1 :

G0 = {Ω},

G1 = {ω1 , ω2 }, {ω3 , ω4 } .

Notice that the partition G1 is finer than the partition G0 (every set in the
partition G0 can be obtained as a union of some sets from the partition G1 ).
S
The partition of F1 is then equal to {ω1 , ω2 }, {ω3 , ω4 } and

F1S = {∅, {ω1 , ω2 }, {ω3 , ω4 }, Ω}.

Let G2 be a partition of the σ-algebra generated by S2 :


G2 = {ω1 }, {ω2 , ω3 }, {ω4 } .

43
We cannot use the same argument as above: G2 is not finer than G1 . How-
ever, let us put into one bag all sets from G0 , G1 and G2 :

{ω1 , ω2 }, {ω3 , ω4 }, {ω1 }, {ω2 , ω3 }, {ω4 }, Ω .
Now, following the algorithm, we need to find the smallest collection of
disjoint sets (Ai ) such that every set above can be represented as a union
of some of them. Obviously, this collection is given by

{ω1 }, {ω2 }, {ω3 }, {ω4 } ,

which implies that the partition of F2S is equal to {ω1 }, {ω2 }, {ω3 }, {ω4 }
and F2S is a power σ-algebra. 
In order to specify our market model, we still have to define what trading
strategies the agents in our model are allowed to use in order to trade on
the market. As in Chapter 1 we assume that the market consists of a money
market account (e.g. a bond) denoted by B, which evolves deterministically
as Bt+k = Bt (1 + r)k , where r is the interest rate (we can assume that
B0 = 1 without loss of generality), and n stocks S 1 , . . . , S n , which are
assumed to be stochastic processes on the underlying state space. A trading
strategy is then given by an Rn+1 valued stochastic process (φt )t=0,...,T whose
components are denoted φ0t , φ1t , ..., φnt and where φit denotes the number of
shares of the i-th stock hold at time t and φ0t the number of bonds at time
t.
Definition 2.1.12. The value process corresponding
to the trading strat-
egy φ = (φt )0≤t≤T −1 is the stochastic process Vt (φ) 0≤t≤T , where

n
X
Vt (φ) = φ0t Bt + φit Sti , t = 0, . . . , T.
i=1

It is unreasonable to allow all trading strategies. In particular, the choice

of investment should be based on the available information and the capital
in possession.
Definition 2.1.13. A trading strategy φ = (φt )0≤t≤T is called (Ft )-adapted,
if the stochastic process (φt )0≤t≤T is adapted to the filtration (Ft )0≤t≤T .

44
Definition 2.1.14. A trading strategy φ = (φt )0≤t≤T is called self financ-
ing, if for all t = 0, ..., T − 1

n
X n
X
φ0t Bt+1 + φit St+1
i
= φ0t+1 Bt+1 + φit+1 St+1
i
. (2.2)
i=1 i=1

An agent who invests according to a self-financing portfolio, invests

money into the strategy in time t = 0 and in the following moments only
rebalances his portfolio, neither withdrawing any money out of the market,
nor investing new money into the market. The agent chooses his investments
at time t and does not change his portfolio in the time interval [t, t+1). The
left-hand side of equation (2.2) represents the value of the agents portfolio
immediately before time t + 1, whereas the right-hand side represents the
value of his portfolio immediately after time t + 1, when rebalancing of his
portfolio has taken place. The self financing assumption says that these two
values must be equal, and this means that neither money is withdrawn nor
new money is invested.
We are now able to set up our general multi period market model:
Definition 2.1.15. A general multi period market model is given by
the following data:
1. A probability space (Ω, P) with the filtration (Ft )t=0,...,T .
2. A money market account (Bt ), which evolves according to Bt = (1+r)t .
3. A collection of n financial assets whose prices are given by stochastic
processes
(St1 ), ..., (Stn ) t=0,...,T ,

which are assumed to be (Ft )-adapted.

4. A set T of self financing and (Ft )-adapted trading strategies.
In practical applications the filtration Ft is often taken to be generated
by the prices of assets, so that it represents only the information that can
be learnt from the market. Following Definition 2.1.10 we know how to

45
construct such a filtration for one stochastic process (n = 1). If n > 1, it
is easiest to consider the approach following Definition 2.1.10 and to
make
1 n
use of partitions. If the filtration (Ft ) is generated by (St ), ..., (St ) t=0,...,T ,
then
Ft = {the information generated by Sji , i = 1, . . . , n, j = 0, . . . , t}.
As before we put all sets from the partitions of σ-algebras generated by Sji ,
i = 1, . . . , n, j = 0, . . . , t into one bag denoted by G. We find the smallest
collection of sets (Ai )i∈I , for some set of indexes I, that satisfies
1. Ai 6= ∅ for i ∈ I.
2. EverySset G ∈ G can be written as a union of some of the Ai ’s, i.e.
G = j∈J Aj with J ⊂ I.
3. Sets Ai are pairwise disjoint, i.e. Ai ∩ Aj = ∅ whenever i 6= j.
We claim that this collection is a partition for Ft (check it!).
Example 2.1.16. Consider a diagram given in Example 2.1.11. The un-
derlying probability space is Ω = {ω1 , ω2 , ω3 , ω4 }. The physical measure is
encoded in the diagram above the arrows:
2 3 2 1
· ,
P(ω1 ) = P(ω2 ) = · ,
5 4 5 4
3 2 3 5
P(ω3 ) = · , P(ω4 ) = · .
5 7 5 7
The filtration generated by the asset prices has been determined as
F0 = {∅, Ω},
F1 = {∅, {ω1 , ω2 }, {ω3 , ω4 }, Ω},

F2 = ∅,
{ω1 }, {ω2 }, {ω3 }, {ω4 },
{ω1 , ω2 }, {ω1 , ω3 }, {ω1 , ω4 }, {ω2 , ω3 }, {ω2 , ω4 }, {ω3 , ω4 },
{ω1 , ω2 , ω3 }, {ω1 , ω2 , ω4 }, {ω1 , ω3 , ω4 }, {ω2 , ω3 , ω4 },
Ω .


46
 It is often useful to represent the self financing condition in terms of the
(multi period) gains process.
Definition 2.1.17. Assume we are given a general multi period market
model as described in Definition 2.1.12. The increment process (∆Sti )1≤t≤T
corresponding to the i-th stock is defined by

∆Sti := Sti − St−1

i
, t = 1, ..., T. (2.3)

The increments of the money market account are given by

∆Bt = Bt − Bt−1 = (1 + r)t−1 r = Bt r, t = 1, . . . , T.
Given a trading strategy φ, the corresponding gains process (Gt (φ))0≤t≤T
is given by G0 (φ) = 0 and

t−1
X n X
X t−1
Gt (φ) = φis ∆Bs+1 + φis ∆Ss+1
i
, t = 1, . . . , T. (2.4)
s=0 i=1 s=0

As in the single period market models, it is worthwhile to consider dis-

counted counterparts of the price process, gains process and value process:
Definition 2.1.18. Given a general multi period market model, the dis-
counted prices are given by

Sti
Ŝti = , (2.5)
Bt

the discounted gains by

∆Ŝti = Ŝti − Ŝt−1

i
, (2.6)

the discounted value process corresponding to a trading strategy φ =

(φt )0≤t≤T by
Vt (φ)
V̂t (φ) = , (2.7)
Bt

47
and the discounted gains process corresponding to a trading strategy
φ = (φt )0≤t≤T by Ĝ0 (φ) = 0 and

n X
X t−1
Ĝt (φ) = φis ∆Ŝs+1
i
, t = 1, . . . , T . (2.8)
i=1 s=0

The self financing condition (2.2) states that changes of the amount of
capital invested are caused only by gains or losses incurred by investments
in the market. Below is a formal statement of this intuition:
Proposition 2.1.19. An adapted trading strategy φ = (φt )0≤t≤T is self fi-
nancing, if and only if any of the two equivalent statements holds
1. Vt (φ) = V0 (φ) + Gt (φ), for all 0 ≤ t ≤ T .
2. V̂t (φ) = V̂0 (φ) + Ĝt (φ), for all 0 ≤ t ≤ T .
Proof. Exercise !

48
2.2 Properties of the general multi period market model
In this section we redefine the general concepts of financial mathematics,
such as arbitrage, hedging etc. in the framework of multi period market
models. This is also a good repetition.
Definition 2.2.1. A trading strategy φ = (φt )0≤t≤T in a general multi period
market model is called an arbitrage if
1. V0 (φ) = 0
2. VT (φ) ≥ 0
3. E(VT (φ)) > 0
As in a single period market model condition 3. can be replaced by:
3’. there exists ω ∈ Ω such that VT (φ)(ω) > 0.
Furthermore one can use the discounted value process or the discounted
gains process in order to express the arbitrage conditions.
We come back to the notion of a contingent claim and a hedging (repli-
cating) strategy:
Definition 2.2.2. A contingent claim in a multi period market model
is an FT -measurable random variable X on Ω representing a payoff at the
terminal time T . A hedging strategy for X is a trading strategy φ ∈ T
s.t.
VT (φ) = X,
i.e. the terminal value of the trading strategy is equal to the payoff of the
contingent claim.
In a multi period market model it is reasonable to look for the price
of a contingent claim in any time t < T . This certainly includes t = 0.
Therefore, instead of of talking about a price of the contingent claim we
shall talk about a price process of the contingent claim. We start with
an application of the replication principle. If X is a contingent claim and
φ is the hedge for X then Vt (φ) is a proper price process of X. We will see
later (Proposition 2.4.5) that this is the only price that does not introduce
arbitrage opportunities in the market.

49
Example 2.2.3. Let us consider the two period model from Example 2.1.11:

S2 = 9 ω1
3
4 rrr
rr9
rr
rrr
S1B = 8L
 LLL 1
 LL4L
 LLL
2  %
5 
 S2 = 6 ω2




S0 =: 5
::
::
::
:: 3
::5 S2 = 6 ω3
::
:: 2
7 rrr
rr9
:: rr
: rrr
S1 = 4L
LLL 5
LL7L
LLL
%
S2 = 3 ω4
We assume that the interest rate is equal to zero, i.e. Bt = 1 for t = 0, 1, 2.
We consider a digital call in this model, which pays 1, if the stock price at
time t = 2 is greater or equal than 8, i.e.
(
1, S2 (ω) ≥ 8,
X(ω) =
0, S2 (ω) < 8.

In this model the contingent claim X pays off 1 if and only if the state of
the world is ω1 : (
1, ω = ω1 ,
X(ω) =
0, otherwise.
How can one find a self financing hedging strategy for X, i.e. a self financing
strategy satisfying

V2 (φ)(ω) = φ02 (ω) + φ12 (ω)S2 (ω) = X(ω).

Since there are four states of the world, these are four equations for the eight
variables φ02 (ωi ) and φ12 (ωi ). The solution may not be unique. However,

50
because of the self financing condition, we have

φ02 (ω) + φ12 (ω)S2 (ω) = φ01 (ω) + φ11 (ω)S2 (ω)

so the same equations hold for the φ01 (ω)’s and φ11 (ω)’s:

φ01 (ω) + φ11 (ω)S2 (ω) = X(ω).

Therefore if we use self financing trading strategies the terminal share hold-
ings φT does not play any role: e.g. we can simply choose the φ02 (ω)’s and
φ12 (ω)’s equal to the φ01 (ω)’s and φ11 (ω)’s.
The trick now is to decompose the model into single period market mod-
els. In this case we have three, one starting at time t = 0 going until time
t = 1 with price S0 = 5, one starting at time t = 1 with price S1 = 8 and
one starting at time t = 1 with price S1 = 4. Let us first solve the hedging
problem in the t = 1 model with S1 = 8:

S =9
2
rrrr9
rrr
rrr
S1 = 8L
LLL
LLL
LLL
%
S2 = 6
This model is in fact an elementary single period market model, as in Section
1.1 and therefore the Delta hedging formula (1.3) applies. Denoting the still
unknown hedging strategy with φ = (φt )0≤t≤2 we find
1−0 1
φ11 (ω) = = , for ω = ω1 , ω2 .
9−6 3
In this case, the rest of the money available at time t = 1, i.e. V1 (φ) − 13 · 8
is invested in the money market account. Let us compute V1 (φ). Since φ is
a hedge, we have
 
1 1
V1 (φ)(ω1 ) − · 8 + · 9 = 1
3 3
 
1 1
V1 (φ)(ω2 ) − · 8 + · 6 = 0.
3 3
51
These equations are satisfied for V1 (φ)(ω1 ) = 32 = (V1 (φ)(ω2 ). Equality of
the value process under the two states ω1 and ω2 is no coincidence, it is a
must, because V1 (φ) has to be F1S measurable (it results from Definition
2.1.12). We then have
2 1
φ01 (ω) = − · 8 = −2, for ω = ω1 , ω2 .
3 3
Let us now come to the hedging problem at time t = 1 when the stock price
is S1 = 4:
S9 2 = 6
r
rrrrr
r
rrr
S1 = 4L
LLL
LLL
LLL
%
S2 = 3
The state of the world is here either ω3 or ω4 . There is no chance that the
digital call will pay off anything else than zero. The hedging strategy for
this payoff is easy, invest zero in the money market account and invest zero
in the stock. We therefore have

φ01 (ω) = 0 = φ11 (ω), for ω = ω3 , ω4 .

Let us now consider the hedging problem at time t = 0 with price S0 = 5:

S =8
1
rrrr9
rrr
rrr
S0 = 5L
LLL
LLL
LLL
%
S1 = 4
Here we don’t have to hedge the digital call, since it pays off at time t = 2,
but we have to find a hedge for the contingent claim that pays off V1 (φ) at
t = 1: (
2
3, if ω = ω1 , ω2 ,
V1 (φ) =
0, if ω = ω3 , ω4 .

52
The reason why we do this, is because if we hedge this contingent claim in
the one period from time t = 0 to time t = 1, and then follow the strategy
we computed at time t = 1, we are done. Solving this hedging problem is
easy again. We can apply the delta hedging formula once more in order to
obtain
2
1 −0 1
φ0 (ω) = 3 = , ω = ω1 , ω2 , ω3 , ω4 .
8−4 6
In order to hedge we must have
 
1 1 2
V0 (φ)(ω) − · 5 + · 8 = , ω = ω1 , ω2
6 6 3
 
1 1
V0 (φ)(ω) − · 5 + · 4 = 0, ω = ω3 , ω4 .
6 6
1
Both equations are satisfied for V0 (φ)(ω) = 6 for all ω ∈ Ω. This yields
1 1 4
φ00 (ω) = − ·5=− , ω = ω1 , ω2 , ω3 , ω4 .
6 6 6
The price process of the digital call is given by Vt (φ), t = 0, 1, 2, where φ
is the above trading strategy. The price at t = 0 is equal to 61 . The price
at t = 1 depends on the state of the market. If S 1 = 8, it is equal to 23 . If
S 1 = 4, it is equal to 0. The price at t = 2 is simply equal to the payoff of
the digital option under consideration.
We can present the hedging strategy in a concise way in a table:
t=0 t=1
ω1 ω2 ω 3 ω4
φ0 − 64 −2 −2 0 0
1 1 1
φ1 6 3 3 0 0
It is now easy to check that this an adapted and self-financing trading
strategy. 
Due to the replication principle (and above example) we know a method
to price contingent claims for which there exists a hedging strategy. As
in general single period market models there may be contingent claims for
which there is no hedging strategy in T . We have then analogous definitions
as in the single period case.

53
Definition 2.2.4. A contingent claim X is called attainable in T , if there
exists a trading strategy φ ∈ T which replicates X, i.e. satisfies VT (φ) = X.
Definition 2.2.5. A general multi period market model is called complete,
if and only if every contingent claim is attainable. A model that is not
complete is called incomplete.
Example 2.2.6. A multi period version of the stochastic volatility model
in Example 1.2.14 is an incomplete multi period market model. 

2.3 Conditional expectation

In order to extend the notion of risk neutral probability measure to the
multi period case, we need the concept of conditional expectation. This is
due to the dynamic character of the multi period model.
Definition 2.3.1. Let (Ω, P) be a finite probability space, X a random vari-
able and G a σ-algebra of sets of Ω. Denoting the unique partition of G
with (Ai )i∈I the conditional expectation E(X|G) of X with respect to G
is defined as the random variable which satisfies
P
E(X|G)(ω) = x x P(X = x|Ai ), ω ∈ Ai . (2.9)

Here
P({X = x} ∩ Ai )
P(X = x|Ai ) =
P(Ai )
denotes the conditional probability of the event {ω|X(ω) = x} given the
event Ai .
It is worth to mention that E(X|G) is a random variable. It has the
same value on each set of the partition, i.e. E(X|G)(ω) = E(X|G)(ω 0 ) if
there exists i ∈ I such that ω, ω 0 ∈ Ai . Notice, however, that for ω’s in
different sets of the partition values can be different. Until now it seems
that the conditional expectation is a purely artificial construct. Not at all.
The intuitive meaning is the following: the conditional expectation gives
the average of values of X (weighted by their probabilities) on each set Ai

54
of the partition. It is somehow ”the expected value” of X on each set of
the partition of G. It follows from Proposition 2.1.5 that E(X|G) is a G-
measurable random variable. The verification of the following identity is
left as an exercise: for any G ∈ G one has
P P
ω∈G X(ω)P(ω) = ω∈G E(X|G)(ω)P(ω). (2.10)

Example 2.3.2. In the context of Example 2.1.11, we want to compute

E(S2 |F1S ), where, as we have already shown,

F1S = {∅, {ω1 , ω2 }, {ω3 , ω4 }, Ω}.

Sets A1 = {ω1 , ω2 } and A2 = {ω3 , ω4 } form a partition of F1S . We have

2 3
P({ω ∈ A1 s.t. S2 (ω) = 9}) P(ω1 ) 5 · 4 3
P(S2 = 9|A1 ) = = = 2 =
P(A1 ) P({ω1 , ω2 }) 5
4
2 1
P({ω ∈ A1 s.t. S2 (ω) = 6}) P(ω2 ) 5 · 4 1
P(S2 = 6|A1 ) = = = 2 =
P(A1 ) P({ω1 , ω2 }) 5
4
P({ω ∈ A1 s.t. S2 (ω) = 3}) P(∅)
P(S2 = 3|A1 ) = = =0
P(A1 ) P({ω1 , ω2 })
P({ω ∈ A2 s.t. S2 (ω) = 9}) P(∅)
P(S2 = 9|A2 ) = = =0
P(A2 ) P({ω3 , ω4 })
3 2
P({ω ∈ A2 s.t. S2 (ω) = 6}) P(ω3 ) 5 · 7 2
P(S2 = 6|A2 ) = = = 3 =
P(A2 ) P({ω3 , ω4 }) 5
7
3 5
P({ω ∈ A2 s.t. S2 (ω) = 3}) P(ω4 ) 5 · 7 5
P(S2 = 3|A2 ) = = = 3 = .
P(A2 ) P({ω3 , ω4 }) 5
7
A more concise way of writing the above equalities is the following:
2 3
P(A1 ∩ {S2 = 9}) P(ω1 ) 5 · 4 3
P(S2 = 9|A1 ) = = = 2 =
P(A1 ) P({ω1 , ω2 }) 5
4
..
.
3 5
P(A2 ∩ {S2 = 3}) P(ω4 ) 5 · 7 5
P(S2 = 3|A2 ) = = = 3 = .
P(A2 ) P({ω3 , ω4 }) 5
7

55
It then follows from equation (2.9), that for ω ∈ A1
3 1 33
E(S2 |F1S )(ω) = 9 · +6· +3·0=
4 4 4
and for ω ∈ A2
2 5 27
E(S2 |F1S )(ω) = 9 · 0 + 6 · +3· = .
7 7 7
This is enough to define the conditional expectation
(
33
S 4, if ω = ω1 , ω2 ,
E(S2 |F1 ) = 27
7, if ω = ω3 , ω4 .


The following properties of the conditional expectation are very impor-
tant and useful. The proofs are omitted.
Proposition 2.3.3. Assume we have a probability space (Ω, P) and σ-
algebras G, G1 and G2 of subsets of Ω. Assume furthermore that G2 ⊂ G1 .
Then
1. If X : Ω → R is a random variable then

E(X|G2 ) = E(E(X|G1 )|G2 ) (2.11)

2. If Y : Ω → R is a G-measurable random variable, then

E(Y X|G) = Y E(X|G). (2.12)

3. If G = {∅, Ω} is the trivial σ-algebra, then

E(X|G) = E(X) (2.13)

56
2.4 Risk neutral probability measures: arbitrage and
pricing
Using the concept of conditional expectation, we can now define a risk
neutral probability measure in the multi period framework.
Definition 2.4.1. A measure Q on Ω is called a risk neutral probability
measure for a general multi period market model if
1. Q(ω) > 0 for all ω ∈ Ω.
2. EQ (∆Ŝti |Ft−1 ) = 0 for i = 1, .., n and for all 1 ≤ t ≤ T .
Another way of formulating the second condition is:
 
1 i
EQ St+1 Ft = Sti , 0 ≤ t ≤ T − 1.
1+r

The Fundamental Theorem of asset pricing is valid in the framework of

multi period market models. The proof is similar to the one we presented for
the single period market model. Here, we only prove the implication ”risk
neutral measure ⇒ no arbitrage”. The proof of the inverse implication relies
on the separating hyperplane theorem as in the single period case, but one
has to cope with some technicalities.
Theorem 2.4.2. Fundamental Theorem of Asset Pricing. The fol-
lowing statements are equivalent in the framework of the multi period market
model:
1. There are no arbitrages in the set T of self-financing and adapted trad-
ing strategies.
2. There exists a risk neutral probability measure.
Proof. As we said, we shall only prove the implication: 1. =⇒ 2. Assume
that Q is a risk neutral measure for a general multi period market model.
We shall prove that the model is arbitrage free by contradiction. Assume
therefore that there exists an arbitrage strategy φ ∈ T . Such a strategy
satisfies:

57
 1. V̂0 = 0.
2. ĜT ≥ 0.
3. EQ (ĜT ) > 0.
On the other hand using the definition of the discounted gains process as
well as property 2. of a risk neutral measure and properties 1., 2. and 3. of
the conditional expectation we obtain
n Xt−1
!
X
EQ (ĜT ) = EQ φis ∆Ŝs+1
i

i=1 s=0
n
XXt−1
= EQ (φis ∆Ŝs+1
i
)
i=1 s=0
Xn Xt−1
= EQ (EQ (φis ∆Ŝs+1
i
|Fs ))
i=1 s=0
Xn Xt−1
= EQ φis (EQ (∆Ŝs+1
i
|Fs ))
| {z }
i=1 s=0 =0
= 0.
This contradicts (3.), which implies that the market model is arbitrage
free.
Let us now come back to pricing of contingent claims. Recall that it
is easy to find prices of attainable contingent claims using the replication
principle. The following proposition gives a hint how to use risk neutral
measures for this task (the proof is left as an exercise; it is practically
shown in the proof of Theorem 2.4.2).
Proposition 2.4.3. If φ is an adapted and self-financing trading strategy
and Q is a risk neutral measure, then


V̂s (φ) = EQ V̂t (φ)|Fs , 0 ≤ s ≤ t ≤ T.
In particular,
 
VT (φ)
Vt (φ) = Bt EQ Ft , t = 0, 1, . . . , T.
BT
58
 Let X be an attainable contingent claim. By virtue of Proposition 2.4.3
we can find the price process for X without computiation of the hedging
(replicating) strategy. Indeed, the price ar t is equal to
 
X
Bt EQ Ft .
BT
Consider now a general contingent claim X, attainable or not. As in the
single period market model we introduce the following definition:
Definition 2.4.4. We say that an adapted stochastic process (Xt )t=0,...,T −1
is a price process for the contingent claim X which complies with the
no arbitrage principle, if there is no arbitrage strategy in the extended
model, which consists of:
• the original stocks Bt and (St1 ), ..., (Stn ), and
• an additional asset given by Stn+1 = Xt for 0 ≤ t ≤ T −1 and STn+1 = X.
Clearly, we can formulate a useful result whose proof is very easy and
left as an exercise.
Proposition 2.4.5. Let X be a contingent claim in a multi period market
model. If φ ∈ T be a hedging strategy for X, then the only price of X at time
t which complies with the no arbitrage principle is Vt (φ). In particular, the
price at the beginning of the trading period at time t = 0 is the total initial
investment in the hedge.
As in the single period market model, we can easily distinguish between
complete and incomplete markets. However, the proof is rather difficult as
we skip it.
Theorem 2.4.6. Assume that the multi period market model is arbitrage
free. Then it is complete if and only if there exists exactly one risk neutral
probability measure.
Due to Fundamental Theorem of Asset Pricing, the market model is
arbitrage free if and only if there exists a risk neutral measure. If the
set M of risk neutral measures consists of one element, then the model is

59
complete. If the set M is bigger, then the model is incomplete, i.e. there
exists an unattainable contingent claim. To sum it up, there are only three
possibilities:
• M = ∅ if and only if there is an arbitrage.
• M has one element if and only if the model is arbitrage free and com-
plete.
• M is infinite if and only if the model is arbitrage free but incomplete.
The set M cannot consist of more than one element and be finite (check it
using convex combinations of risk neutral measures).
Notice that in the definition of completeness we do not assume that the
model is arbitrage free. So there is a big class of models that allow arbitrage
and are complete or incomplete. Hence, we cannot use risk neutral measures
to classify them. Fortunately, these models are of no interest to us during
this lecture.
Let us now move to incomplete models. If X is not attainable, we cannot
use the replication principle. However, the arbitrage pricing concept is still
in effect and yields
Proposition 2.4.7. Let X be a possibly unattainable contingent claim and
Q a risk neutral measure for a general multi period market model. Then
 
1
Xt = Bt EQ BT X Ft (2.14)

defines a price process for the contingent claim X which complies with the
no arbitrage principle.
The proof of the above result is almost the same as the proof of Propo-
sition 1.2.20.
Example 2.4.8. Let us come back to Example 2.2.3 and consider again
the digital call. As before we assume that the interest rate is equal to zero,
i.e. r = 0. We already know that an arbitrage free price price is given by
x = 16 , and that this price is in fact unique, because the contingent claim

60
can be hedged. However, alternatively, let us compute the price by using
risk neutral measures. In order to do this, we first compute a set of risk
neutral measures. Let Q = (q1 , q2 , q3 , q4 )> be the risk neutral measure we
are looking for. The following two conditions for the qi are due to the fact
that Q is a probability measure and satisfies condition 1. in Definition 2.4.1:

q1 + q2 + q3 + q4 = 1, q1 , q2 , q3 , q4 > 0.

Note that since the interest rate is equal to zero prices agree with the dis-
counted prices. Due to condition 2. of Definition 2.4.1 we obtain

5 = EQ (Sˆ1 |F0 ) = EQ (Sˆ1 ) = q1 · 8 + q2 · 8 + q3 · 4 + q4 · 4.

Substitution of q4 = 1 − q1 − q2 − q3 into this equation gives

5 = 4 + 4(q1 + q2 ),

which is equivalent to
1
q 1 + q2 = . (2.15)
4
This also implies that
3
q 3 + q4 = . (2.16)
4
Recall that the partition of F1S is given by two sets A1 = {ω1 , ω2 } and
A2 = {ω3 , ω4 }. Using the definition of the conditional expectation, we see
that for ω ∈ A1 , we have
q1 q2
8 = EQ (Ŝ2 |F1S )(ω) = 9 · +6 = 36q1 + 24q2 .
q1 + q2 q1 + q2
This gives
9q1 + 6q2 = 2. (2.17)
Furthermore for ω ∈ A2 we have
q3 q4
4 = EQ (Ŝ2 |F1S )(ω) = 6 · +3 = 8q3 + 4q4 .
q3 + q4 q 3 + q4
This gives
2q3 + q4 = 1. (2.18)

61
Now we have four linear equations (2.14),(2.15),(2.16), and (2.17), and four
unknown variables q1 , q2 , q3 and q4 :
3 1
9q1 + 6q2 = 2, 2q3 + q4 = 1, q3 + q4 = , q 1 + q2 = .
4 4
The unique solution of this system is
 >
> 1 1 1 1
Q = (q1 , q2 , q3 , q4 ) = , , , . (2.19)
6 12 4 2
We check that this solution satisfies the inequalities:

q1 , q2 , q3 , q4 > 0.

We therefore see that the market model is arbitrage free. By virtue of

Theorem 2.4.6 it is also complete.
Let us now compute the risk neutral measure using the trick that we
already benefited in Example 2.2.3, namely the decomposition of our multi
period market model into single period market models. We have three
embedded single period market models: one starting at time t = 0 going
until time t = 1 with price S0 = 5, one starting at time t = 1 with price
S1 = 8 and one starting at time t = 1 with price S1 = 4. Let us first find a
risk neutral measure (q1∗ , 1 − q1∗ )T in the t = 1 model with S1 = 8:

S =9
92
q1∗ rrrr
rrr
rrr
S1 = 8L
LLL 1−q ∗
LLL 1
LLL
%
S2 = 6
This model is in fact an elementary single period market model, as in Section
1.1 and therefore the risk neutral measure is given by the formula (recall
that r = 0):
8 · (1 + 0) − 6 2
q1∗ = = .
9−6 3

62
The single period model at time t = 1 when the stock price is S1 = 4 is
given by
S9 2 = 6
q2∗ rrrr
rrr
rrr
S1 = 4L
LLL 1−q ∗
LLL 2
LLL
%
S2 = 3
The risk neutral measure (q2∗ , 1 − q2∗ )T is defined by the formula
4 · (1 + 0) − 3 1
q1∗ = = .
6−3 3
Similarly, at time t = 0 with price S0 = 5 we have

S =8
91
q0∗ rrrr
rr r
rrr
S0 = 5L
LLL 1−q ∗
LLL 0
LLL
%
S1 = 4
where
5 · (1 + 0) − 4 1
q0∗ = = .
8−4 4
We put the computed values into the diagram (see Figure 2.1). Let us
compute the probabilities of ω’s as in Example 2.1.16:
1 2 1 1 1 1
Q(ω1 ) = · = , Q(ω2 ) = · = ,
4 3 6 4 3 12
3 1 1 3 2 1
Q(ω3 ) = · = , Q(ω4 ) = · = ,
4 3 4 4 3 2
which is exactly the result obtained in (2.19). Remember that after compu-
tations one always has to check whether all probabilities are greater than
zero!
In order to compute the price of the digital call X = 1{S2 |≥8} we observe
that X(ω) is 1 for ω = ω1 and 0 for all other ω’s. The only price at time 0

63
 S2 =9 ω1
2
sss9
s
sss
3

sss
S1B = 8K
 KKK 1
 KK3K
 KKK
1   %
4 
 S2 = 6 ω2



S0 =8 5
88
88
88
88 34
88 S2 =6 ω3
88 1
sss9
88 s
sss
3
88
 sss
S1 = 4K
KKK 2
KK3K
KKK
%
S2 = 3 ω4

Figure 2.1: Diagram with the risk neutral measure Q.

that complies with the no arbitrage principle is equal to

 
X 1
EQ = q 1 · 1 + q 2 · 0 + q 3 · 0 + q 4 · 0 = ,
(1 + r)2 6
since r = 0. 

64
2.5 Exotic options
The options we have considered so far, were of the type h(ST ), where h is
a pay off function depending on the terminal stock price ST , but not on
the stock prices at times strictly before T . Such options are called path
independent. However there are also path dependent options. Here are
some examples:
1. Asian option: An Asian option is in a way a European call option on
the average stock price. There are two kinds which are mainly traded:
(a) arithmetic average option:
T
!+
1 X
X= St − K (2.20)
T +1 t=0

(b) geometric average option:

 1
! T +1 +
YT
X= St − K (2.21)
t=0

2. Barrier options: Barrier options are options which are activated or

deactivated if the stock price hits a certain barrier. As an example we
consider so called ”knock out” barrier options. There are two types of
these:
(a) Down and Out call:
X = (ST − K1 )+ · 1{min0≤t≤T St >K2 } (2.22)
This option has the same payoff as a European call with strike K1 ,
but only if the stock price remains over the barrier level of K2 on
the time interval [0, T ]. Otherwise it pays off 0.
(b) Down and In call:
X = (ST − K1 )+ · 1{min0≤t≤T St ≤K2 } (2.23)
This option has the same payoff as a European call with strike K1 ,
but only if the stock price falls below the barrier level of K2 on the
time interval [0, T ]. Otherwise it pays off 0.

65
 3. Look back option: A look back option is an option on a maximum
or a minimum of the stock price on the interval [0, T ]. As an example
consider call options:
(a) Call option on maximum:
 +
X = max St − K
0≤t≤T

(b) Call option on minimum:

 +
X = min St − K
0≤t≤T

The main difference of the path dependent options and a standard Euro-
pean options is that their payoff does not only depend on the terminal stock
price ST but on the whole path that the stock price has taken in the interval
[0, T ]. Notice that our pricing rules are ready to deal with path dependent
options, because all contingent claims above depend on asset prices up to
time T and hence they are FT measurable. Therefore, one simply has to
compute the expected value of BXT under the risk neutral measure.

2.6 The Binomial Asset Pricing Model

The most important multi period discrete time market model is the bino-
mial asset pricing model, which we discuss in this section. This model is
often called the Cox-Ross-Rubinstein model and abbreviated to CRR
model. The binomial market model is the concatenation of several elemen-
tary single period market models, as discussed in section 1.1. We assume
here that we have one stock (St ) and one money market account (Bt ), but a
generalization to the case of more then one stock is possible. At each point
in time when the stock price is St , there are two possible outcomes for the
stock price in the next period: uSt , or dSt , u, d > 0. Typically u > 1 and
0 < d < 1. In order to guarantee that the model is arbitrage free we require
that d < 1 + r < u. The probability of an up move (the probability at the
arrow) is assumed to be the same 0 < p < 1 for each period, and is assumed

66
t
S27 = S0 u2 ω1
p oooo
o
ooo
ooo
S1 =
C S0Ou OOO
 OO1−p
 OOO
OO'

p 
 S2 = S0 ud ω2



S0 7
77
77
77
77 1−p
77 S27 = S0 du ω3
77 oo
77 pooo
77 ooo
 ooo
S1 = S0Od
OOO
OO1−p
OOO
OO'
S2 = S0 d2 ω4

Figure 2.2: Two period binomial asset pricing model

to be independent of all previous stock price movements. A sample two

period binomial model in presented in Figure 2.2.
Definition 2.6.1. A stochastic process (Xt )1≤t≤T on a probability space
(Ω, P) is called a Bernoulli process with parameter p (0 ≤ p ≤ 1), if the
random variables X1 , X2 , X3 , ..., XT are independent and

P(Xt = 1) = 1 − P(Xt = 0) = p, t = 0, 1, . . . , T.

The Bernoulli counting process (Nt )1≤t≤T is defined via

Nt (ω) := X1 (ω) + X2 (ω) + ... + Xt (ω), ω ∈ Ω.

Using the Bernoulli counting process, the stock price process in the bino-
mial model is defined via a deterministic initial value S0 and for 1 ≤ t ≤ T

St (ω) := S0 uNt (ω) dt−Nt (ω) , ω ∈ Ω. (2.24)

67
The idea behind this construction is that the Bernoulli process (Xt ) deter-
mines the up and down movements of the stock. The stock price moves
up at time t if Xt (ω) = 1 and moves down if Xt (ω) = 0. The Bernoulli
counting process (Nt ) counts the up movements. Before and including time
t, the stock price has moved up Nt times, and it has moved down t − Nt
times. Assuming that the stock price can only move up resp. down by the
factors u resp. d we obtain equation (2.19). Why is this model called the
binomial model? The reason is that for each t the random variable Nt has
a binomial distribution with parameters (p, t), i.e.

t


P(Nt = k) = k pk (1 − p)t−k .

The distribution of the stock price at time t is then

t


P(St = S0 uk dt−k ) = k pk (1 − p)t−k , k = 0, 1, ..., t. (2.25)

It can be shown, that the filtration (FtS )0≤t≤T generated by (St ) is the same
as the filtration (FtX )0≤t≤T generated by the Bernoulli process, where F0X
is the trivial σ-algebra by definition. The money market account (Bt ) is as
usual assumed to be defined via B0 = 1 and

Bt = (1 + r)t . (2.26)

Definition 2.6.2. The binomial market model with parameters p, u, d

(0 < p < 1, u ≥ d > 0) and time horizon [0, T ] on a probability space (Ω, P)
is the multi period market model consisting of one stock and one money
market account, where the stock price evolution is described by equation
(2.19) and the money market account by (2.21). The filtration is assumed
to be the filtration (FtS )0≤t≤T generated by (St ).
Given a binomial market model, the constituting Bernoulli process (Xt )
can be regained from the stock prices as follows: Clearly Xt = Nt − Nt−1 .

68
Then
St S0 uNt dt−Nt
=
St−1 S0 uNt−1 dt−1−Nt−1
= uNt −Nt−1 d1−(Nt −Nt−1 )
= uXt d1−Xt

u if Xt (ω) = 1
= .
d if Xt (ω) = 0
Is the binomial market model arbitrage free?
Proposition 2.6.3. The binomial market model is arbitrage free if and only
if d < 1 + r < u. Moreover, if it is arbitrage free, then it is complete.
The above proposition can be easily proved with the methods developed
in Example 2.4.8. We can also use:
Proposition 2.6.4. Under the assumption d < 1 + r < u a probability
measure Q on Ω is a risk neutral measure for the binomial market model
with parameters p, u, d and time horizon [0, T ] is and only if the following
three conditions hold:
1. X1 , X2 , X3 , ..., XT are independent under the measure Q
2. 0 < q := Q (Xt = 1) < 1 for all 1 ≤ t ≤ T
3. q u−1−r d−1−r



1+r + (1 − q) 1+r =0
where (Xt ) is the Bernoulli process corresponding to the model.
Condition 3. here, is the same as in Section 1.1 and is equivalent to

1+r−d
q= u−d . (2.27)

Therefore under the assumption d < 1 + r < u the binomial model is

arbitrage free: once we specify q and the measure Q is fully defined by
conditions 1. and 2. of Proposition 2.6.4. As the value for q is also unique,

69
we obtain from Theorem 2.4.6 that the binomial model is also complete.
We will assume from now on that

d < 1 + r < u.

It seems that the binomial model grows very large if the number of peri-
ods T is very large. In fact, in general one could obtain 2T different leaves
of the binomial tree. Sure? Look at the Figure 2.3. This representation is
very useful for pricing of a certain type of contingent claims, namely path
independent contingent claims. These are the claims whose payoffs de-
pend only on the value of ST , i.e. there are of the form h(ST ) for some
function h. Perfect examples are: call and put options, binary call and put
options. But beware! You cannot use the representation from Figure 2.3
to price path dependent options such as Asian, barrier or lookback options.
In fact, we only have T + 1 possible values for the stock price at time T .
Let us now see how certain options can be evaluated using the bino-
mial market model above. To simplify the computation we shall impose an
assumption that is often used in practice:

d = u−1 . (2.28)

Then
St = S0 u2Nt −t . (2.29)

Consider a European call with the payoff:

X = (ST − K)+ .

The price of this contingent claim at time t = 0 can be computed via the
formula  
1
V0 = EQ X .
(1 + r)T
Using Proposition 2.6.4, this is the same as
T  
1 X T k
V0 = T
q (1 − q)T −k max(0, S0 uk dT −k − K).
(1 + r) k
k=0

70
 ;
ww
www
ww
ww
4
: S0 u CC
ttt CC
t CC
ttt CC
tt CC
!
{=
3
; S 0 u JJ {
vv JJ {
vv JJ {{
vvv JJ {{{
vv J$ {{
2 3
S
< 0 u HH S: 0 u dC
yyy HH ttt CC
CC
yy HH tt CC
y HH tt
yy H# tt CC
C!
{=
S0 u E 2
S; 0 u dJ
|> EE v JJ {{
{
||
| EE vv JJ {{
|| EE
EE vvv JJ
JJ {{
v {{
|| " vv $
S0 B S0 udH 2 2
S0 u dC
BB yy< HH tt: CC
BB yy HH tt CC
BB y HH t CC
B yy HH ttt CC
yy # t C!
{=
S0 d E 2
S; 0 ud J {
EE v JJ {{
EE vv JJ {{
EE vvv JJ {{
E" v J
vv J$ {{
S 0 d2 H S: 0 udC 3
HH tt CC
HH t CC
HH ttt CC
HH tt CC
# t C!
S 0 d3 J {{=
JJ {{
JJ {{
JJ {{
JJ
$ {{
S 0 d4 G
GG
GG
GG
GG
#

Figure 2.3: Recombination of the binomial model

71
We have that
 u k K
k T −k
S0 u d −K >0⇔ >
d S0 dT  
u K
⇔ k log > log
d S0 dT
 
log S0KdT
⇔k>
.
log ud

We define k̂ as the smallest integer k such that this inequality is satisfied.

If there are less than k̂ upward jumps, there is no chance that the option
will pay off anything. We can therefore write
k̂−1  
1 X T k
V0 = T
q (1 − q)T −k 0
(1 + r) k
k=0
T  
1 X T k T −k k T −k


+ q (1 − q) S 0 u d − K
(1 + r)T k
k=k̂
T  
S0 X T k
= T
q (1 − q)T −k uk dT −k
(1 + r) k
k=k̂
T  
K X T k
− T
q (1 − q)T −k
(1 + r) k
k=k̂

T   k  T −k
X T qu (1 − q)d
= S0
k 1+r 1+r
k=k̂
T  
K X T k
− T
q (1 − q)T −k
(1 + r) k
k=k̂
T   T  
X T k K X T k
= S0 q̂ (1 − q̂)T −k − T
q (1 − q)T −k ,
k (1 + r) k
k=k̂ k=k̂
qu
where q̂ = 1+r . We have therefore proved:

72
Proposition 2.6.5. (CRR-option pricing formula) An arbitrage free
price at time t = 0 for the European call option X = (ST − K)+ in the
binomial market model with parameters u = d−1 ,r, T is given by
T   T  
X T k K X T k
V0 = S0 q̂ (1 − q̂)T −k − T
q (1 − q)T −k
k (1 + r) k
k=k̂ k=k̂

where
1+r−d qu
q= , q̂ = , (2.30)
u−d 1+r
and k̂ is the smallest integer such that
 
log S0KdT
k̂ >
.
log ud
Though simple by its construction, the binomial market model is very
powerful. If one chooses the number of time periods very large and the
units of time very small, then one can approximate trading in continuous
time. The CRR-option pricing formula then approximates the Black-Scholes
option pricing formula (details are given in Elliot, page 50-55) and in the
Continuous Time Finance module, next semester.

73
Chapter 3

Investment

3.1 Single Period Investment

In this section we return to the study of the general single period market
model which consists of one money market account and n stocks, based on
the underlying state space Ω = {ω1 , ..., ωk } and probability measure P. In
the first chapter we were mainly interested in the construction of this model
and the computation of prices of financial derivatives such as options and
contingent claims. Now we want to study a different question, which is by
no means less important:

What is the optimal way to invest money into the market?

The answer of this question naturally depends on the choice of the model,
and in this section we will stick to our general single period market model
from the last section of the first chapter. We will later consider the same
question in multi-period models.
Before we answer this question we have to specify what exactly we mean
by optimality, i.e. what is our measure for the performance of a trading
strategy. In order to define such a performance measure three fundamental
characteristics of financial markets, or more precisely of the agents trading
on the markets, have to be included:

1. Agents prefer higher payoffs to lower payoffs

74
This idea is so innate to financial markets, that it does not need any discus-
sion. However we cannot judge trading strategies purely by this character-
istic. As we saw in the previous sections, the payoffs of financial assets are
generally modeled as random variables. Assume now, we have two trading
strategies, the performance of which we want to compare. It is then possi-
ble that in one state of the world the first trading strategy yields a higher
payoff and in another state of the world the second trading strategy yields
a higher payoff. In order to compare the performances of the two trading
strategies, we must somehow take an average over the states of the world,
but this is nothing else then taking expectations.

2. Agents assess trading strategies by average performance

The second characteristic leads directly to the use of expected value,

which fits in very well into our probabilistic models. There is however a
third characteristic which has to be included. To illustrate this, consider the
following simple example: Assume an agent is offered to choose between two
alternatives. If he chooses the first one he will be paid 10 million pounds. If
he chooses the second one, a fair coin will be tossed. If the coin shows head,
he will be paid 20 million pounds instead of the 10 million, but if the coin
shows tail, he will be paid nothing. What would the agent choose? If the
agent is not yet a billionaire, he would probably go for the first alternative,
which gives him 10 million for sure and financial safety for the rest of his
life. Denoting the payoff of the first alternative with X 1 and the payoff
of the second alternative with X 2 we see that if we simply judge the two
payoffs by its expectation, the agent would be indifferent between the two
alternatives:
1 1
E(X 1 ) = 10 million = · 0 + · 20 million = E(X 2 ).
2 2
The reason why the agent would prefer the first alternative is, that he is
risk averse. His idea is, that 10 million will bring him financial safety and
the possibility of an upper class lifestyle and that in fact for him 10 million
or 20 million wouldn’t be such a big difference, but 10 million or nothing
would make a huge difference. We therefore have found our third and last
characteristic, which we want to include into our model.

75
 3. Agents are risk averse

The concept of risk aversion can be included in our model by the use of
utility functions.
Definition 3.1.1. A continuously differentiable function u : R+ → R is
called a risk averse utility function if it has the following three proper-
ties:
1. limx→0 u0 (x) = +∞ and limx→∞ u0 (x) = 0
2. u is strictly increasing in the way that u0 (x) > 0 for all x ∈ R
3. u is strictly concave in the way that u(λx + (1 − λ)y) > λu(x) + (1 −
λ)u(y).
Third condition in Definition 3.1.1 is equivalent to u00 (x) < 0 for all
x ∈ R+ if the function if twice differentiable. Let us now consider a random
variable X which is interpreted as a random payoff. Fixing a utility function
u we will measure the performance of this payoff by

Pk
E(u(X)) = i=1 P(ωi )u(X(ωi )).

This way of measuring the performance of a payoff incorporates all three

of the characteristics we have discussed above. The first one is reflected
by the fact that the utility function is strictly increasing and the second
one is reflected by usage of the expected value. The third characteristic
is reflected by the concavity of the utility function. In our little example
where the agent is offered to choose between the two alternatives, we have
1 1
E(u(X 1 )) = u(10 million ) = u( · 0 + · 20 million )
2 2
1 1
> · u(0) + · u(20 million )) = E(u(X 2 )).
2 2
The investor would therefore decide for the first alternative. This an exam-
ple of what is also called 2nd order stochastic dominance.

76
 It is important to note, that the expectation in our performance measure
has to be taken under the original measure P which represents the traders
beliefs, and not any risk neutral measure. The value E(u(X)) crucially de-
pend on the choice of the utility function u. It is clear that whenever an
agent wants to compare a set of payoffs, he has to use the same utility func-
tion for this evaluation. However this utility function is like a personal trait
of the agent and different agents may have different utility functions, such
as some agents are more risk averse than others. Here are some examples
of risk averse utility functions:
• Logarithmic utility: u(x) = ln(x)
• Exponential utility*: u(x) = −e−λx , λ > 0.
• Power utility: u(x) = γ1 xγ , γ ∈ (0, 1).
√
• Square root utility: u(x) = x (it is an example of the power utility
function).
Some of the utility functions above, clearly do not satisfy the Definition
3.1.1 of a utility function in the strict sense. *Notice for example that the
exponential utility function has a different range than the one we used in
Definition 3.1.1. The limits in the first property of the same definition are
different as well. However, utility functions are used to rank alternative
investments and the actual value of the utility of a specific outcome is not
meaningful on its own. For this reason, any utility function of the form
αU (x) + β would provide the same ranking of outcomes as U (x) and is
called equivalent to U (x). For example 1 − e−λx is equivalent to −e−λx .
Exponential utility functions are very often used by economists and are very
useful in continuous time models. All the results proved in this section, are
also valid for exponential utility functions with some appropriate simple
modifications.
Let us now try to formalize the risk-aversion property of a utility function.
This is done by the Arrow-Pratt coefficients of risk aversion.
00
Coefficient of Absolute Risk Aversion: a(x) = − UU 0 (x)
(x)

77
Notice that due to the last two properties of a risk averse utility function, the
above coefficient is always positive. In fact the only reason for the presence
of the first derivative in the coefficient is to ensure the same coefficient of
absolute risk aversion for all equivalent utility functions. In an investment
problem, absolute risk-aversion measures the actual amount in pounds that
an investor will choose to hold in risky assets, given a certain wealth level
x.
Similarly the coefficient of relative risk aversion is defined.
00
Coefficient of Relative Risk Aversion: r(x) = − xU (x)
U 0 (x)

In financial terms, the coefficient of relative risk-aversion measures the per-

centage of wealth held in risky assets, for a given wealth level x.
It is clear that different utility functions exhibit different risk aversion
properties. For example, the exponential utility function has a Constant
Absolute Risk Aversion coefficient (CARA) a(x) = λ. It also has
an Increasing Relative Risk Aversion (IRRA) r(x) = xλ. On the
other hand the logarithmic utility function has a Decreasing Absolute
Risk Aversion (DARA) a(x) = 1/x and a Constant Relative Risk
Aversion (CRRA) coefficient r(x) = 1.
Using the concept introduced above, the question of finding the optimal
investment into our market now translates into the question of finding a
trading strategy (x, φ) s.t. E(u(V1 (x, φ)) achieves an optimal value. This
problem is called the optimal portfolio problem. The optimal value will
naturally depend on the initial investment x. A higher initial investment
will naturally result in a higher expected utility, so we have to take the
initial investment as a parameter for our problem.
Definition 3.1.2. A trading strategy (x, φ∗ ) is called a solution for the
optimal portfolio problem with initial investment x and utility function u, if
E(u(V1 (x, φ∗ )) = max E u(V1 (x, φ))


(3.1)
φ

It is not clear whether a solution to the optimal portfolio problem exists.

That depends in general on the particular model and the chosen utility
function.

78
 An interesting relationship is the relationship between the existence of a
solution to the optimal portfolio problem and the existence of an arbitrage.
Assume for the moment, that there exists an arbitrage (0, ψ) in our model.
Then for every trading strategy (x, φ) we would have

V1 (x, φ + ψ)(ω) = V1 (x, φ)(ω) + V1 (0, ψ)(ω) ≥ V1 (x, φ)(ω),

where (x, φ+ψ) is the trading strategy given by the sum of (x, φ) and (0, ψ),
i.e. the trading strategy which buys φi + ψ i shares of the i-th stock. By
the definition of an arbitrage, this strategy needs an initial investment of x.
That the inequality above holds is also a result of (0, ψ) being an arbitrage,
but we can say even more, the inequality above is strict for at least one ω.
Therefore for any risk averse utility function u we have

E(u(V1 (x, φ + ψ))) > E(u(V1 (x))).

Hence in the presence of an arbitrage for every strategy (x, φ) there is an-
other strategy with same initial investment which does perform strictly
better. Consequently, if there is an arbitrage in the market then there is no
solution to the optimal portfolio problem or in other words:
Proposition 3.1.3. If there exists a solution to the optimal portfolio prob-
lem then there can not exist an arbitrage in the market.
We already know by the fundamental theorem of asset pricing that the no
arbitrage property is equivalent to the existence of a risk neutral measure.
Such a risk neutral measure can be explicitly computed from a solution to
the optimal portfolio problem as the following proposition shows.
Proposition 3.1.4. Let (x, φ∗ ) be a solution to the optimal portfolio problem
with initial wealth x and utility function u, then the measure Q defined by
P(ω)u0 (V1 (x, φ∗ )(ω))
Q(ω) = , ω ∈ Ω,
E(u0 (V1 (x, φ∗ ))
is a risk neutral probability measure.

79
Proof. That Q is indeed a probability measure follows from Q(ω) > 0 for
all ω (recall that u0 (x) > 0 for all x) and
k k
X X P(ωi )u0 (V1 (x, φ∗ )(ωi ))
Q(ωi ) =
i=1 i=1
E(u0 (V1 (x, φ∗ ))
k
1 X
= P(ωi )u0 (V1 (x, φ∗ )(ωi ))
E(u0 (V1 (x, φ∗ )) i=1
1
= 0 ∗
E(u0 (V1 (x, φ∗ ))
E(u (V1 (x, φ ))
= 1.

It remains to prove that Q satisfies condition 2. of Definition 1.2.7. We

know that the function

φ 7→ E(u(V1 (x, φ))), φ ∈ Rn ,

is differentiable and attains its maximum at a point φ∗ ∈ Rn . Therefore all

partial derivatives of this function must vanish at the point φ∗ . This leads
to the following equations:
∂  1 1 n n


0= E u B1 (x + φ ∆Ŝ (ωi ) + ... + φ ∆Ŝ (ωi )) , j = 1, ..n,
∂φj φ=φ ∗

(3.2)
since using the well known identities V1 (x, φ) = B1 V̂1 (x, φ) and V̂1 (x, φ) =
x + Ĝ(x, φ) we have

E(u(V1 (x, φ)) = E(u(B1 (x + φ1 ∆Ŝ 1 + ... + φn ∆Ŝ n )). (3.3)

Computing derivatives in (3.2) gives:

k
X
0= P(ωi )u0 (B1 (x + φ∗1 ∆Ŝ 1 (ωi ) + ... + φ∗n ∆Ŝ n (ωi ))) · B1 · ∆Ŝ j (ωi )
i=1
k
X
= B1 · P(ωi )u0 (V1 (x, φ∗ )) · ∆Ŝ j (ωi ), j = 1, . . . , n.
i=1

80
Since B1 = 1 + r > 0 this is equivalent to
k
X
0= P(ωi )u0 (V1 (x, φ∗ )) · ∆Ŝ j (ωi ), j = 1, . . . , n.
i=1

This implies that for j = 1, . . . , n

k
X
j
EQ (∆Ŝ ) = Q(ωi )∆Ŝ j (ωi )
i=1
k
X P(ωi )u0 (V1 (x, φ∗ )(ωi ))
= 0 ∗
· ∆Ŝ j (ωi )
i=1
E(u (V1 (x, φ ))
k
1 X
= 0 ∗
P(ωi )u0 (V1 (x, φ∗ )(ωi )) · ∆Ŝ j (ωi )
E(u (V1 (x, φ )) i=1
= 0.

Notice that risk neutral measures pop up both in pricing and replication,
and in optimal portfolio selection. This hints that these problems are deeply
interconnected. We will pursue this issue after the example which shows how
optimal portfolios can be computed explicitly in our general single period
market model.
Example 3.1.5. Suppose that n = 2, k = 3, r = 19 , Ŝ01 = 6, Ŝ02 = 10 and
that the discounted price process is given by the following table:
ω 1 ω2 ω3
Ŝ11 6 8 4
Ŝ12 13 9 8
We wish to use the exponential utility function with λ = 1, i.e. u(x) =
1 − exp(−x). Since additive constants does not play a role in our opti-
mization process we shall use an equivalent but simpler utility function
u(x) = − exp(−x).

81
 In order to compute the optimal trading strategy for an initial investment
x, we have to compute the maximum of the function
U : R2 → R
(φ1 , φ2 ) 7→ E(− exp(−(V1 (x, (φ1 , φ2 )).
Using equation (3.3), computing the expectation and setting the partial
derivatives with respect to φ1 and φ2 equal to zero, we obtain the following
two equations:
   
10 10
0 = P(ω1 ) exp − (x + 0φ1 + 3φ2 · ·0
9 9
   
10 1 2 10
+ P(ω2 ) exp − (x + 2φ − φ · ·2
9 9
   
10 10
+ P(ω3 ) exp − (x − 2φ1 − 2φ2 · · (−2),
9 9

   
10 10
0 = P(ω1 ) exp − (x + 0φ1 + 3φ2 · ·3
9 9
   
10 10
+ P(ω2 ) exp − (x + 2φ1 − φ2 · · (−1)
9 9
   
10 1 2 10
+ P(ω3 ) exp − (x − 2φ − 2φ · · (−2).
9 9
These two equations have to be solved for φ1 and φ2 . Although numerically
possible this may not be easy. We therefore propose another technique in
the following. 
As you could see the naive technique to solve the optimal portfolio prob-
lem can be computationally difficult. There is however a more efficient
technique to solve the problem. This technique involves risk neutral mea-
sures and the Lagrange multiplier method. The idea of the technique is as
follows: Decompose the optimal portfolio problem into two subproblems:
Step 1: Compute the maximizer V1 of the function V 7→ Eu(V ), where
V is allowed to take values is a certain feasible set.

82
 Step 2: Compute a trading strategy which has the maximizer V1 com-
puted in step 1 as value at time t = 1.
The trading strategy arising from the second step is then an optimal
portfolio. Step 2 is in fact a replication or hedging problem and we know
that such a problem poses certain difficulties in an incomplete market model.
Therefore, let us assume first that our model is complete. In this case there
is a unique risk neutral measure which we denote by Q (see Theorem 1.2.25).
Definition 3.1.6. We define the set of attainable wealths from initial in-
vestment x > 0 by

1


Wx := W ∈ Rk |EQ 1+r W =x .

It follows from Proposition 1.2.16 that if W ∈ Wx then there exists a

trading strategy (x, φ) s.t. W = V1 (x, φ) and hence the name is justified.
First step of our optimization process can be stated as the following con-
strained optimization problem:

maximize E(u(W ))
subject to W ∈ Wx .

This problem can be solved by the Lagrange multiplier method. To do this

we consider the Lagrange function

1



L(W, λ) := E(u(W )) − λ EQ 1+r W −x . (3.4)

A solution of the above constrained optimization problem may then be

computed by setting all partial derivatives with respect to Wi = W (ωi ) and
λ equal to zero. In the definition of the Lagrange function in (3.4) we have
to compute expectations with respect to two different measures P and Q.
In order to avoid this, we introduce the so called state price density
Q(ω)
L(ω) := . (3.5)
P(ω)

83
Then for any random variable Y we have
EQ Y = EY L.
Using the state price density we can write the Lagrange function as
follows:
k   
X 1
L(W, λ) = P(ωi ) u(W (ωi )) − λ L(ωi ) W (ωi ) − x
i=1
1 + r

Computing the partial derivatives with respect to Wi = W (ωi ) and setting

them equal to zero gives
L(ωi )
u0 (W (ωi )) = λ . (3.6)
1+r
Taking the expected value of both sides of the above equality we obtain:
λ = E((1 + r)u0 (W )). (3.7)
Let us denote the inverse function of u0 (x) with I(x), i.e. I(x) is the
function s.t. u0 (I(x)) = x = I(u0 (x)). Such a function exists, since u0 (x) is
strictly decreasing by property 3. of a utility function. It is easy to check
that
1. I is strictly decreasing,
2. limx→0 I(x) = ∞,
3. limx→∞ I(x) = 0.
Applying I to equation (3.6) we obtain
 
W (ω) = I λ L(ω)
1+r . (3.8)

Last equation would give the solution of the constrained optimization prob-
lem if we knew the correct value of λ. Equation (3.7) does not help us in
computation of λ since it involves the so far unknown W . We know however,
that W must satisfy  
1
EQ W =x
1+r
84
and substituting the expression from equation (3.8) into last equation we
obtain
1 L



EQ 1+r I λ 1+r = x. (3.9)

If we solve this equation for λ and plug the solution into (3.8), we get the
solution of our constrained optimization problem. Such a λ always exists
and is in fact
 uniquely
 determined
 by equation (3.9) because the function
1 L(ω)
h(λ) := EQ 1+r I λ 1+r is strictly decreasing, continuous and satisfies
limλ→0 h(λ) = ∞ and limλ→∞ h(λ) = 0.
Example 3.1.7. Suppose the utility function in the optimal portfolio prob-
lem is given by u(x) = ln(x). In this case u0 (x) = x1 and therefore I(x) = x1
as well. Hence, the optimal wealth in equation (3.8) is given by
 L  1+r
W =I λ = . (3.10)
1+r λL
We remind that L > 0. Equation (3.9) gets the following form:
!
1  L 
EQ I λ = x.
1+r 1+r

The solution can be easily computed and is given by

1
λ= .
x
Substituting the last expression in equation (3.10) we obtain
1+r
W =x . (3.11)
L
This is the optimal wealth in the general complete single period market
model where logarithmic utility of the form u(x) = ln(x) is used. In order to
compute the optimal trading strategy we would now have to find a hedging
strategy for W . This strategy depends on the specific model. We have
discussed methods how to find replicating strategies in the previous sections.

85
 Let us now compute the optimal expected utility, i.e. the expected utility
of the optimal wealth we have just calculated. Plugging expression for the
optimal wealth W from (3.10) into u(x) gives
1+r
u(W ) = ln(x) + ln .
L
Computing the expectation yields
 
1+r
E(u(W )) = ln(x) + E ln .
L

Let us now consider the optimal portfolio problem for a market model
which is incomplete. The crucial difference to the previous case where we
studied a complete model is that one has to be more careful when defining
the set of attainable wealths. In a complete market model every contingent
claim is attainable and in defining the set of attainable wealths one has
only to take the initial investment into account. Doing the same thing in
an incomplete market it might be that the optimal wealth computed is not
attainable at all. The remark concluding Definition 3.1.6 does not hold in
the incomplete market case. We have therefore to modify Definition 3.1.6
in the way that Wx only contains wealths which are indeed attainable by
trading strategies, which start with an initial investment of x. We know
from Proposition 1.2.24 that a contingent claim X in a possibly incomplete
1
market model is attainable if and only if the value of EQ ( 1+r X) is the same
for all Q ∈ M. The natural generalization of Definition 3.1.6 for the case of
an incomplete single period market model is therefore:
Definition 3.1.8. We define the set of attainable wealths from initial in-
vestment x > 0 in a possibly incomplete market model by

1


Wx := W ∈ Rk |EQ 1+r W = x for all Q ∈ M .

The first step in the optimal portfolio problem is then formally the same
as the first step in the complete market case:

86
 maximize E(u(W ))
subject to W ∈ Wx .

The real problem lies here in the set of constraints determining the set
Wx . We already know that in general in an incomplete market model there
exists an infinite number of risk neutral measures leading to an infinite
number of constraints. The Lagrange multiplier method which we used to
solve the optimal portfolio problem before works only for a finite number
of constraints. Therefore we have to reduce the number of constraints to a
finite number. This is possible as the following lemma shows.
Lemma 3.1.9. Consider a general single period, possibly incomplete, mar-
ket model. Then there exist finitely many probability measures Qi ∈ M,
i = 1, ..., l such that the space Wx from Definition 3.1.8 is given by

1


Wx := W ∈ Rk |EQi 1+r W = x for all i = 1, ..., l .

Notice a mysterious bar over M in the statement of the lemma. It stands

for the closure of M, i.e. the set M together with its boundary. If we char-
acterize M by a system of equalities and inequalities, as we can always do,
all the inequalities are strict. They are basically derived from the condition
that the probability of each ω must be greater than 0 and smaller than 1.
Now, we obtain M by changing all the strict inequalities > and < to ≥ and
≤.
We omit the proof of this lemma. The main reason why this lemma
holds is that the space M can be expressed as the intersection of the finite
dimensional vectorspace W⊥ and the set P + and can therefore be written
as a convex combination of a finite number of fixed risk neutral measures.
The optimal portfolio problem above then translates into the problem

maximize E(u(W ))
 
W
subject to EQi = x, i = 1, ..., l.
1+r

87
This problem is a constrained optimization problem with a finite number of
constraints and can therefore be solved by the Lagrange multiplier method
in the same way as before. Defining the Lagrange function L via
l    
X Li W
L(W, λ) = E(u(W )) − λi E −x
i=1
1+r

i (ω)
with Li (ω) = QP(ω) and λ = (λ1 , ..., λl )T we obtain as the solution of the
first step in the optimal portfolio problem
P 
l Li (ω)
W (ω) = I i=1 λi 1+r .

To obtain the unknown Lagrange multiplier one now has to solve the l
equations

λ1 L1 +...+λl Ll



E Li I 1+r = (1 + r)x, i = 1, . . . , l.

Solutions of this system of equations exist and are unique by the same
reasons as in the complete case.

3.2 Returns
This section acts as a introduction for the famous Capital Asset Pricing
Model that will be discussed in the following section. We shall set up some
notation regarding the returns of trading strategies and obtain a few key
relations.
As in the previous section we use the notation L for the state price
density
Q(ω)
L(ω) = P(ω) .

Here Q is a fixed risk neutral measure. It is not important whether the

model is complete or not, but in the case of incompleteness we choose one

88
risk neutral measure Q for the rest of the section. The return of the i-th
stock in our model is defined by the random variable

S1i −S0i
Ri = S0i
, i = 1, . . . , n.

The return of the money market account is given by the deterministic in-
terest rate, i.e.
B1 −B0
R0 = B0 = r.

Most computations in our single period market model can be done by using
the returns instead of the actual stock prices. For example the gains process
can be written as
n
! n
X X
i i 0
G(x, φ) = x − φ S0 B0 R + φi S0i Ri . (3.12)
i=1 i=1

The verification of the last equation is left as an exercise.

PnTo simplify the notation in the following we denote the investment x −
i i
i=1 φ S0 in the money market account implemented by the trading strat-
egy (x, φ) with φ0 . A trading strategy can then be represented by an n + 1-
dimensional vector φ = (φ0 , φ1 , ..., φn ). It follows from the second property
in Definition 1.2.7 of a risk neutral measure that

EQ (Ri ) = r, i = 1, . . . , n.

This can be seen as follows:

 i i
EQ (S1i ) − S0i S0i (1 + r) − S0i rS0i

S − S
EQ (Ri ) = EQ 1 i 0 = = = = r.
S0 S0i S0i S0i
Here we used the fact that S0i is a number (not a random variable).
We denote the mean return of the i-th stock by
i
R = E(Ri ).

89
Let us now consider the covariance of the returns with the state price den-
sity. We have

cov(Ri , L) = E(Ri L) − E(Ri )E(L)

= EQ (Ri ) − E(Ri )
i
= r−R.
i
The difference R −r is called the risk premium. Normally this is positive,
because investors insist that the expected returns of the risky securities are
higher than the returns of riskless securities such as money market accounts.
Let us now consider the return of a portfolio, or more precisely the return
induced by the value of a trading strategy:
V1 (x, φ) − V0 (x, φ)
R(x, φ) = .
V0 (x, φ)
In the following we will omit the dependence of R(x, φ) and V (x, φ) on the
trading strategy (x, φ) and just write R and V . Using the definition of V1
and (3.12) we obtain
n
φ0 X φi S0i i
R= r+ R.
V0 i=1
V 0

It follows from the properties of the covariance and the previous discussion
on the Ri that
R − r = −cov(R, L) (3.13)
where R = E(R) is the expected return of the portfolio. We leave the details
as an exercise.
Let us now fix two scalars a and b with b 6= 0 and assume that the
contingent claim a + bL is attainable. In this case there exists a trading
strategy (x0 , φ0 ) s.t. the corresponding value process Vt0 = Vt (x0 , φ0 ) for
t = 0, 1 satisfies
V10 = a + bL.
Denoting with R0 the return corresponding to V 0 we obtain

V00 (1 + R0 ) = a + bL.

90
Solving this equation for L gives
V00 (1 + R0 ) − a
L= .
b
Substituting this into equation (3.13) gives
V00
cov(R, L) = cov(R, R0 ). (3.14)
b
Here R still corresponds to an arbitrary trading strategy. We can therefore
rewrite equation (3.13) as
V00
R − r = − cov(R, R0 ).
b
For the particular case of R = R0 we obtain
0 V00
R −r =− var(R0 ).
b
V00
We can now use the last equality in order to substitute for b in (3.14),
which gives
cov(R,R0 ) 0
R−r = var(R0 ) (R − r). (3.15)

Formula (3.15) holds whenever R0 is the return process of the contingent

claim a+bL and the latter is attainable. We will later see the importance of
this formula in connection with the Capital Asset Pricing Model (CAPM).
0
The ratio cov(R,R )
var(R0 ) is sometimes called the beta of the trading strategy
corresponding to R with respect to the trading strategy corresponding to
R0 . The formula says, that the risk premium is proportional to its beta with
respect to a linear transformation of the state price density.

3.3 Mean-Variance Analysis

We now come to a central issue in single period market models, the so called
Mean-Variance approach to portfolio analysis, which goes back mainly to
Markowitz. Harry Max Markowitz (born August 24, 1927) is (still alive!) an

91
influential economist at the Rady School of Management at the University of
California, San Diego. Formerly at the RAND Corporation, Markowitz won
the Nobel Prize in 1990 while a professor of finance at Baruch College of the
City University of New York. He is best known for his pioneering work in
modern portfolio theory, studying the effects of asset risk, correlation and
diversification on expected investment portfolio returns. The main idea
of his results is that when agents are offered two financial assets having
the same expected return they would choose the asset that has the lower
variance , i.e. is less risky.

Problem 1

minimize V ar(R)
subject to E(R) = ρ
R is a portfolio return

where R is a return induced by the trading strategy and ρ is a specified

scalar. We assume ρ ≥ r. If ρ = r then the minimal value of the problem
is zero, because we can generate the return r by investing in the money
market account, which is a non risky investment having variance 0. Let
us recall that the return of the value process corresponding to a trading
strategy φ = (φ0 , ..., φn ) can be written as
n
φ0 X φi S i
0 i
R= r+ R.
V0 i=1
V 0

It therefore follows that an arbitrary payoff R is a return corresponding to

the value process of a trading strategy if and only if it can be written as
n
X
R = (1 − F1 − ... − Fn )r + Fi R i ,
i=1

where
φi S0i
Fi = .
V0

92
If there are no short-sales the number Fi can be interpreted as the fraction of
the total initial investment which is invested in the i-th stock. Furthermore
we define the covariance matrix C = (Cij ) by
Cij = Cov(Ri , Rj ), i, j = 1, ..., n.
With this notation we have V ar(R) = F > CF where F = (F1 , ..., Fn )> ∈ Rn .
Then we can rewrite the minimum variance problem as follows:

Problem 2

minimize F > CF
n
X i
subject to (1 − F1 − ... − Fn )r + Fi R = ρ.
i=1

This is a quadratic programming problem which can now be solved by the

Lagrange multiplier method. In order to obtain the actual trading strategy
from the Fi one has set φi = FSi ·x
i , where x denotes the initial investment.
0

Example 3.3.1. Let us consider a model with two risky assets and a money
market account. For simplicity we assume that the interest rate r is equal
to zero. The expected returns, variance and covariances of the risky assets
are assumed to be
1 2
R = 1, R = 0.90,
V ar(R1 ) = 0.10,
V ar(R2 ) = 0.15,
Cov(R1 , R2 ) = −0.1.
The covariance matrix C is therefore given by
 
0.1 −0.1
C= .
−0.1 0.15
We want to solve the problem
minimize F > CF
subject to F1 + 0.90F2 = 0.956

93
Consider the Lagrange function
L(F1 , F2 , λ) = 0.1F12 + 0.15F22 − 0.2F1 F2 + λ(F1 + 0.9F2 − 0.956).
Setting the partial derivatives of L equal to zero we obtain
0 = 0.2F1 − 0.2F2 + λ
0 = 0.3F2 − 0.2F1 + 0.9λ
0 = F1 + 0.9F2 − 0.956.
Solving this system, we get F1 = 0.56,F2 = 0.44 and λ = −0.0232. This
result is very interesting. From a naive point of view one would expect that
all investments into the risky assets go into the first asset, since it has both
a higher expected return and a lower variance. The reason why it is not
so is that there is a negative correlation between the two assets. It is no
cheating: the variance of the optimal portfolio is equal to:
0.1 · (0.56)2 + 0.15 · (0.44)2 − 0.2 · 0.56 · 0.44 = 0.0112
A second interesting point is that the optimal portfolio does not make use
of the money market account. Does this has to do with the fixed expected
return? Someone would ask if this would be the case if we require
F1 + 0.90F2 < 0.956,
or
F1 + 0.90F2 > 0.956.
You can try the following expected returns for the portfolio: R̄ = 0.5 and
R̄ = 1.5. In the first case, solving the optimisation problem will give F1 =
0.19 and F2 = 0.19 (notice that F1 + F2 < 1), while in the second case
F1 = 0.693 and F2 = 0.8763 (notice that F1 + F2 > 1). We will see in the
next pages what is so special about the return 0.956.

For the rest of this section we impose the following assumption:

The market model is complete, i.e. there exists one risk neutral
measure Q.

94
We will also assume that Q 6= P and denote by L the state price density
Q(ω)
L(ω) = .
P(ω)
In order to obtain a classical result, which is called the Capital Asset
Pricing Model, we will translate the mean variance problem into an op-
timal portfolio problem which we can solve with the methods from Section
3.1. We will do it in two steps. Consider first the following problem:

Problem 3

minimize V ar(V1 )
subject to E(V1 ) = x(1 + ρ)
V0 = x.

In this problem the constraints identify the set of all time t = 1 values
corresponding to trading strategies that can be implemented with a total
initial investment of x and which have a mean of x(1 + ρ). We claim that
this problem is equivalent to the original minimum variance problem. In
fact, V̂1 is a solution of the latter problem if and only if R̂ = (V̂1x−x) is a
solution of the original mean variance problem. We leave the details as an
exercise.
Let us now introduce a fourth problem which under special choice of
the parameter β is equivalent to the previous one, but has the form of a
standard optimal portfolio problem:

Problem 4
 
1
maximize E − V12 + βV1
2
subject to V0 = x.

The function u(x) = − 12 x2 + βx is not a utility function in the strict

sense as defined in Definition 3.1.1. However, its second derivative is clearly
negative and the function is therefore concave. One can show that in this

95
special case this is enough in order to make the methods discussed in Section
3.1 work. We have u0 (x) = −x + β and by coincidence I(x) = −x + β as
well. We solve equation (3.9) from Section 3.1 to obtain the value for the
Lagrange multiplier:
(x(1 + r) − β)(1 + r))
λ=− .
EQ (L)
Substituting this value into equation (3.8) we obtain the optimal wealth
β
L
V̂1 = EQ (L) − L + x(1 + r) . (3.16)
EQ (L) EQ (L)
Then (there is no mistake, try to find all the differences)
β
1
E(V̂1 ) = EQ (L) − 1 + x(1 + r) . (3.17)
EQ (L) EQ (L)
We now want this optimal wealth to satisfy the first constraint of the Prob-
lem 3, i.e.
E(V̂1 ) = x(1 + ρ). (3.18)
Using (3.17) this translates into
β
1
EQ (L) − 1 + x(1 + r) = x(1 + ρ).
EQ (L) EQ (L)
Since by our assumptions Q 6= P, we have EQ L > 1 (this is also left as an
exercise). Solving the above equation for β gives
x[(1 + ρ)EQ (L) − (1 + r)]
β= . (3.19)
EQ (L) − 1
With this choice of β the solution V̂1 of Problem 4 satisfies the constraints
in Problem 3. We will now show that it is in fact a solution of Problem 3. If
V1 is any other random variable which satisfies the constraints in Problem
3 then in particular
E(V1 ) = x(1 + ρ) = E(V̂1 )
and therefore
   
1 2 1 2
E − V̂1 + β V̂1 ≥ E − V1 + βV1 .
2 2
96
It is equivalent to    
1 2 1 2
E V̂ ≤E V
2 1 2 1
and, consequently, to
V ar(V̂1 ) ≤ V ar(V1 ).
On the other hand, reversing the argument above, we can see that a
solution of problem 3 is also a solution of Problem 4. These two problems
are therefore equivalent, provided β and ρ are related by equation (3.19).
Now we come to the most complicated part of this section leading to the
famous results of CAPM. Substituting β from equation (3.19) into equation
(3.16) gives
x[(1 + ρ)EQ (L) − (1 + r)] L
V̂1 = (EQ (L) − L) + x(1 + r)
(EQ (L) − 1)EQ (L) E (L)
  Q
x(1 + r) x(ρ − r) x(1 + r)
= + − (EQ (L) − L)
EQ (L) − 1 EQ (L) − 1 EQ (L)(EQ (L) − 1)
L
+ x(1 + r)
EQ (L)
 
x(1 + r)EQ (L) − x(1 + r) x(ρ − r)
= + (EQ (L) − L)
EQ (L)(EQ (L) − 1) EQ (L) − 1
L
+ x(1 + r)
EQ (L)
 
x(1 + r) x(ρ − r) L
= + (EQ (L) − L) + x(1 + r)
EQ (L) EQ (L) − 1 EQ (L)
x(1 + r)L x(ρ − r) x(1 + r)L
= x(1 + r) − + (EQ (L) − L) +
EQ (L) EQ (L) − 1 EQ (L)
x(ρ − r)
= x(1 + r) + (EQ (L) − L).
EQ (L) − 1

97
The equivalence of Problem 3 and Problem 1 implies

V̂1 − x x(ρ − r)
R̂ = =r+ (EQ (L) − L)
x EQ (L) − 1
r(EQ (L) − 1) ρEQ (L) − rEQ (L) ρ−r
= + − L
EQ (L) − 1 EQ (L) − 1 EQ (L) − 1
ρEQ (L) − r ρ−r
= − L.
EQ (L) − 1 EQ (L) − 1
We have therefore proved the following proposition:
Proposition 3.3.2. Let R̂ be the return of a trading strategy which has the
minimum variance among all other portfolios with expected returns equal to
ρ. Then

ρEQ (L) − r ρ−r

R̂ = − L.
EQ (L) − 1 EQ (L) − 1

The interesting point in the last proposition is that the solution R̂ of

the mean variance problem is an affine function of the state price density.
Recall now the formula (3.15) from the preciding section which states a
relationship between the expected return of such an affine function and an
arbitrary return. Clearly the solution R̂ of the mean variance problem is
attainable and we therefore have proved the following theorem, which is a
main result in the Capital Asset Pricing Model (CAPM):
Theorem 3.3.3. (CAPM) If R̂ is a solution to the mean variance problem
(Problem 1) for ρ ≥ r and if R is the return corresponding to an arbitrary
trading strategy, then

Cov(R,R̂)
E(R) − r = var(R̂)
(E(R̂) − r).

Proof. The proof of Theorem 1.4.1 follows directly from the discussion above
and formula (3.15).

98
 The following Lemma has important consequences for the interpretation
of the CAPM.
Lemma 3.3.4. Let R̂ be a solution to the mean variance problem with
parameter ρ > r. Now let ρ̃ be a different parameter for the expected return.
Then choosing
ρ̃ − ρ
γ=
r−ρ
the portfolio return R̃ = γr + (1 − γ)R̂ is a solution for the mean variance
problem with parameter ρ̃.
The proof of this lemma follows by performing some algebra with the
formula in Proposition 3.3.2. The important consequence of Lemma 3.3.4
is that one has to solve the mean variance portfolio problem only for one
parameter ρ and one gets solutions for all other parameters by investing
according to a combination of the money market account and this fixed
solution of the mean variance problem. This is called the mutual fund
principle:
Proposition 3.3.5. (Mutual Fund Principle) Suppose we fix a portfolio
whose return is a solution to the mean variance portfolio problem (Problem
1) corresponding to some mean return ρ. Then the solution to the mean
variance problem for any other mean return is obtained by a portfolio con-
sisting of investments in the riskless money market account and the fixed
portfolio.
The mutual fund principle has the following consequence. Suppose that
everyone is a mean-variance optimizer as described above and that everyone
agrees on the probabilistic structure, i.e. associates the same subjective
probabilities to the states of the world (has the same objective measure
P). Then every agent can solve his individual mean-variance problem by
investing in a single fixed asset, the mutual fund from above, and the money
market account. The fraction of money invested in the mutual fund and
the fraction of the money invested in the money market account depend on
the agents parameters, as in Lemma 3.3.4, but effectively, everyone buys
the same fund. The question is, what is that fund? This question is easy to

99
answer. It is clear that the total investment, i.e. the sum of all investments
of all agents must add up to the market portfolio, i.e. the totality of all
shares of all stocks traded on the market. Now if all agents invest into the
same risky fund, then this risky fund can only be the market portfolio. In
most applications of the capital asset pricing model, the market portfolio is
assumed to be a solution of the mean variance problem and its return can
then substituted for R̂ in the equation in Theorem 3.3.3.
Pricing: So far it is not clear why the CAPM is referred to as a pricing
model. Assume we have a contingent claim X with the maturity time t = 1
for which we want to compute a price x at time t = 0. The return of the
contingent claim is
X −x
R= .
x
Substituting this into the equation of Theorem 3.3.3 gives

E(X) − x Cov(X, R̂)

−r = (E(R̂) − r)
x x · V ar(R̂)
Solving for x gives

Cov(X,R̂)
E(X) − V ar(R̂)
(E(R̂) − r)
x= .
1+r

How does this pricing formula relate to the one we developed in the previ-
ous section? Prices computed using the capital asset pricing model are in
general not arbitrage free prices. The prices include the subjective prob-
ability measures P, which is a handicap, since this measure is in general
neither known, nor can it be computed. It is also unrealistic to assume
that all agents have the same subjective probabilities and that all agents
are mean-variance minimizers. We should however see it as a complement
or alternative to the methods discussed in the previous section.

100
Bibliography

[1] Pliska, S. R.: Introduction to Mathematical Finance - Discrete Time

Models, Blackwell, 1997.
[2] Shreve, S. E.: Stochastic Calculus for Finance I: The Binomial Asset
Pricing Model, Springer, 2004.
[3] Elliott R. J. and Kopp, P. E.: Mathematics of Financial Markets,
Springer, 1999.
[4] Luenberger D. G.: Investment Science, Oxford University Press,
1998.

101