Risk Neutral Valuation Lecture Notes

∗
Giuseppe Di Graziano

October 12, 2014

∗
Draft version, do not distribute. Please send any comment about typos and errors to
peppe.digraziano@gmail.com

1
1 Discrete Time Models

1.1 No Arbitrage

Loosely speaking, an arbitrage is an opportunity to make a risk-free profit. Although

arbitrages may exists for short periods of time, if markets are efficient, arbitrage oppor-
tunities will not last for long. In real markets they may last only a fraction of a second.
In this course we shall assume, unless otherwise specified, that markets are arbitrage
free. The no arbitrage assumption will allow us to develop a consistent pricing and
hedging theory. Furthermore, knowing the arbitrage free price of a contract and the
corresponding hedging policy will allow us to recognise an arbitrage and will give us a
strategy to take advantage of it.
More formally, assume that the market consists of M risky assets (e.g. stocks, options,
forward contracts, etc) whose price we will denote by Ait with i ∈ {1, . . . , M } and
one risk-free (deterministic) asset A0t ≡ Bt = ert which we will call the money market
account. We shall assume that the risk-free interest rate r is deterministic and market
participants will have r as reference rate for both lending and borrowing transactions.
A portfolio or trading strategy is a collection of units invested in tradable assets over
the relevant time horizon. More precisely, let ψt be the amount (positive or negative)
invested in the risky asset Ai at time t. If we indicate by τ the set of all possible
trading periods, an hedging strategy will be the sequence of vectors:

ψt0 , . . . , ψtM


{t∈τ }
. (1)

The portfolio value at a given time t, say πt , will be given by

M
X
πt ≡ ψti Ait . (2)
i=0

An arbitrage
exists, if it is possible to find a portfolio (the arbitrage portfolio)

ψt1 , . . . , ψtM {t∈τ } and a time t such that

1. π0 ≤ 0

2. πt ≥ 0 a.s.

2
 3. P (πt > 0) > 0,

that is, a strategy with zero or negative initial value (in the latter case, somebody is
”paying” us to implement the strategy) which is non negative at time t and has a strictly
positive probability to have a positive value at time t. For example a combination of
one stock, one option and the money market account which costs zero at t = 0 and has
a positive value at time t = T in say, one market scenario out of 10, would constitute
an arbitrage.
A portfolio is said to be self-financing if

M
X M
X
πt ≡ ψti Ait = i
ψt−1 Ait , (3)
i=0 i=0

for all t ≥ 1 or equivalently

πt = πt−1 + ψt−1 · (At − At−1 ). (4)

where we have used the “dot” product between vectors ψt−1 and At − At−1 . All port-
folios considered in this course will be self-financing unless otherwise specified.

1.2 Forward Contracts

A Forward Contract is an obligation to buy or sell a given asset, e.g. a stock S, at

a pre-determined price (strike) K at a future date T . The pay-off of the contract, seen
from the point of view of the buyer, at expiry T is given by

VT = (ST − K). (5)

Since forwards contracts are regularly traded before the expiry, two questions arise:

1. What is the value Vt of such a contract for t ≤ T ?

2. How can you hedge the contract by investing in a simple (static) portfolio com-
prising the asset S and the money market account B?

3
In the following sections we shall assume that S is a non dividend paying asset and
that interest rates are constants and equal to r.

1.2.1 Hedging Approach

We shall show that the t-value Vt of the forward contract is equal to

Vt = St − Ke−r(T −t) . (6)

We shall prove the claim for t = 0. It is straightforward to extend the result to a

generic t ≤ T .
Assume first that the market value of a forward contract with maturity T and strike
K is equal to

Ṽ0 = V0 + x > S0 − Ke−rT . (7)

It is possible to realise a risk-less profit by executing the sequence of operations below.

At time zero:

1. Sell the forward contract (i.e. agree to sell S at time T for K) and receive Ṽ0 in
exchange.

2. Borrow an amount of cash equal to Ke−rT .

3. Buy the asset and pay for it S0 using part of the cash raised with the first two
transactions.

4. Deposit the excess cash x into the money market account.

The net cash flows of the transactions above will be equal to zero.
At expiry T :

1. Deliver the asset to the counter-party of the forward contract and receive K in
exchange.

4
 2. Repay the debt plus interest, i.e. K.

3. Withdraw an amount equal to xerT from the money market account.

The net cash flow of the transactions above at time T is equal to xerT . We started
with a zero cost portfolio and ended up with a strictly positive profit. It is clear that
the price Ṽ0 allows for arbitrage.
More formally, let ψt be the unit invested in the forward contract Ṽt , ∆t be the holdings
of St and φt the amount invested (i.e. cash deposited or borrowed) in the money market
account Bt at time t. Consider the portfolio ψ0 = −1, ∆0 = 1 and φ0 = −Ke−rT + x.
The value of the portfolio π0 at time zero is equal to

π0 = ψ0 Ṽ0 + ∆0 S0 + φ0 B0 = 0. (8)

Using the self-financing property of the portfolio we have that at time T , the value of
the portfolio is equal to

πT = ψ0 ṼT + ∆0 ST + φ0 BT (9)
= −(ST − K) + ST − K + xerT (10)
= xerT . (11)

Similarly, an arbitrage opportunity arises if

Ṽ0 ≡ V0 − x < S0 − Ke−rT . (12)

The following strategy will allow us to realise a risk-less profit.

At time zero:

1. Buy the forward contract (i.e. buy S forward) and pay Ṽ0 in exchange.

2. Borrow the stock from the market and sell it for S0 .

3. Invest Ke−rT + x in the money market account

5
The net cash flows of the transactions above will be equal to zero.
At expiry T :

1. Withdraw K + xerT from the money market account.

2. Receive the asset S from the counter-party of the forward contract and pay K in
exchange.

3. Give back the borrowed stock.

Again, we started with a zero cost portfolio and ended up with a profit. In terms of
our arbitrage portfolio, let ψ0 = 1, ∆0 = −1 and φ0 = Ke−rT + x. The value of the
portfolio π0 at time zero is given by

π0 = ψ0 Ṽ0 + ∆0 S0 + φ0 B0 = 0. (13)

The value of the portfolio at time t = T was equal to

πT = ψ0 ṼT + ∆0 ST + φ0 BT (14)
= (ST − K) − ST + K + xerT (15)
= xerT , (16)

which is again strictly positve.

If we assume that the market is arbitrage free, then we must conclude that the value
at time zero of a forward contract must be equal to

V0 = S0 − Ke−rT . (17)

1.2.2 Risk Neutral Pricing Approach

In the previous section we showed how to calculate and hedge a simple contingent
claim (a forward contract) by constructing a replicating portfolio which is arbitrage
free. The price of such a claim can be calculated using a probabilistic approach. In
fact, it is often easier to derive the price of a claim using the probabilistic approach.

6
Once the no arbitrage price is known, it will be also easier to find the hedging strategy
for the claim.
In order to do that we need to introduce the fundamental theorem of asset pricing. In
the simple one period framework, the theorem reduces to:

Theorem 1 There is no arbitrage if and only if there exists a measure Q, which we

shall refer to as the Risk Neutral Measure, such that

E Q [VT BT−1 ] = V0 , (18)

for any (admissible) asset VT .

For example we must have that

E Q [ST BT−1 ] = S0 . (19)

The price of a forward contract with VT = (ST − K) can be derived as follows

V0 = E Q [VT BT−1 ]
= E Q [(ST − K)BT−1 ]
= E Q [ST BT−1 ] − KBT−1
= S0 − Ke−rT .

In particular, the Forward Price F (0, T ) at time zero for a forward contract expiring
at time T , is the strike K such that V0 = 0. From the result above, it is straightforward
to see that the forward price is equal to

F (0, T ) = S0 erT . (20)

1.3 The One-Period Binomial Model

In this section we shall show how to price a simple option using the one period binomial
model.

7
A Call Option gives the buyer the right (but not the obligation) to buy a given asset,
e.g. a stock S, at a pre-determined (strike) price K at a future date T . The pay-off of
the contract at expiry T is given by

CT ≡ (ST − K)+ . (21)

Let’s assume that there asset S follows the binomial dynamics:

STu

S0

STd

The option dynamics will be then given by

CTu

C0 =?

CTd

where CTu = (STu − K)+ and CTd = (STd − K)+ .

1.3.1 Hedging Approach

The basic idea is to find a (non-trivial) portfolio consisting of an investment in the

option C, the risky asset S and the money market account B with zero value at time
t = 0 and expiry T . Equivalently, we would like to find a portfolio consisting of an
investment in S and B such that its value it is equal to the pay-off of the option C for
all possible values of S at time T . Without loss of generality we shall set the expiry
T = 1.
In the simple set up we are analyzing, the hedging problem can be reduced to solving

8
a simple system of two equations in two unknowns, namely ∆0 i.e. the investment in
the risky asset and φ0 i.e. the investment in the money market account at inception.
More precisely, we need to solve the system of equations

(
C1u = ∆0 S1u + φ0 B1
(22)
C1d = ∆0 S1d + φ0 B1 ,

where we have used the self-financing property (3). It is straightforward to see that
the system above admits the solution

C1u −C1d
(
∆0 = S1u −S1d
,
(23)
φ0 = (C1d − ∆0 S1d )B1−1 .

Imagine you are trader and you have sold a call option to an investor. If you buy the
portfolio (∆0 , φ0 ) constructed above, you will be perfectly hedged at t = 1. In other
words, any money you may have to pay to investors if the option expires in the money
(i.e. has a positive pay-out) will be perfectly off-set by a gain in the hedging portfolio
(∆0 , φ0 ). Note also that the ∆0 is reminiscent of the first derivative of the option price
with respect to the price of the underlying asset. We shall encounter this quantity
again in the continuous time setting.
If we rule out the possibility of an arbitrage, then it follows the value at time t = 0
of the hedging portfolio, say p0 , must be equal to the initial price of the option C0 .
Hence, if we know the hedging strategy (∆0 , φ0 ) we can calculate the initial price of
the option as follows

C0 = ∆0 S0 + φ0 . (24)

1.3.2 Risk Neutral Pricing Approach

As for forward contracts, we can calculate the price of a contingent claim using the
fundamental theorem of asset pricing. The first step is to derive the risk neutral
probabilities explicitly. This can be done by solving the equation in q:

S0 = E Q [S1 B1−1 ] = [qS1u + (1 − q)S1d ]B1−1 . (25)

9
It is straightforward to see that the risk neutral probabilities are given by

S0 B1 − S d
q = (26)
Su − Sd
S1u − S0 B1
1−q = . (27)
Su − Sd

In order to ensure that probabilities above are non negative, the nodes of the tree must
satisfy S1d < S0 B1 < S1u .
Once the risk neutral probabilities have been calculated, the price of the option follows
immediately from a simple application of the fundamental theorem of asset pricing

C0 = E Q [C1 B1−1 ] = [qC1u + (1 − q)C1d ]B1−1 . (28)

A natural question to ask is whether the price we obtained using (28) is consistent with
the price we obtained under the hedging approach (24). To show that this is indeed
the case, consider

C0 = ∆0 S0 + φ0
= ∆0 S0 + (C1d − ∆0 S1d )B1−1
= ∆0 (S0 B1 − S1d ) + C1d B1−1
 
d
S0 B1 − S1d
  
u S0 B1 − S1
= C1 u d
+ C1 1 − u B1−1
S1 − S1d S1 − S1d
= [qC1u + (1 − q)C1d ]B1−1 ,

which proves that the two approaches are consistent and lead to the same result under
no-arbitrage.
Last but not least note that as far as derivatives, and in general, asset pricing is
concerned, real world probabilities are irrelevant. Even if we knew with certainty the
true probability distribution of the asset, it wouldn’t affect asset pricing. The key
for understanding this apparently counter-intuitive concept is to look at the hedging
approach. In order to hedge the simple securities presented so far in the course, all
we had to do was to borrow or lend money via the money market account and have

10
a long or short investment in the stock. Moreover, we were able to perfectly replicate
contingent claims regardless of the final value of the underlying asset. It is natural thus
to expect that the “real world” distribution of the asset will not come to play in our
pricing formulae. All we have to know is the current price of the asset and the interest
rate (both market observable) 1 .

1.4 Multi-Period Binomial Model

The model presented in the previous section is clearly too simplistic to be used in
practical applications. Fortunately the binomial model can be easily extended to a
multi-period set up. In what follows we shall discuss a special case of the binomial
approach called Cox Ross and Rubinstein (CRR) model. The main features of the
model are the following

1. The node of the tree are recombining (lattice model).

2. “Up” and “Down” probabilities are constant across the various nodes of the tree.

3. All time steps between nodes are equal to ∆t.

The first property makes sure that the branches of the tree do not “explode”, making
pricing and risk management difficult. The possible outcomes in a non recombining
N -steps tree are equal to 2N , whereas they are equal to N + 1 in a recombining tree.
The second and third properties allow us to derive very simple formulae for the pricing
of derivatives contracts. More importantly, we shall show that for a specific choice of
the ”up” and ”down” values of the asset in the tree, the CRR (discrete) asset price
process converges, as the number of steps increases and time interval decreases, to a
geometric Brownian motion (GBM), which is the foundation of continuous time pricing
and hedging.
Consider a fixed time horizon T and subdivide T in N steps of equal size ∆t, i.e. let
∆t = T /N . Let u and d be two real numbers satisfying the constraint

d < er∆t < u. (29)

The asset price Sn,i at time tn ≡ n∆t, where n, is given by

1
The future values of the asset can be chosen arbitrarily as long as S1d < S0 B1 < S1u

11
 Sn,i = S0 ui dn−i . (30)

The graph below shows the binomial model for N = 2.

S0 u2

S0 d

S0 S0 ud = S0 du

S0 u

S0 d2

It is easy to see that the risk neutral probabilities in this model are equal to

B∆t − d
q = , (31)
u−d
u − B∆t
1−q = , (32)
u−d
(33)

for all nodes of the tree. The distribution of the asset S at a given time step n will be
binomial and can be calculated explicitly using the formula

 
i n−i n i
q (1 − q)n−i


Q Sn = S0 u d = (34)
i

where Q (Sn = x) indicates the risk neutral probability of Sn taking value x.

12
1.4.1 Pricing Options Using the CRR Model

The CRR model is a powerful tool for pricing options. As we shall see, it provides
a methodology to price not only call and put options, but a much larger class of
derivatives claims depending on the underlying assets. The CRR is particularly useful
when no closed formula is available for a given option in the corresponding continuous
time GBM set up.
Let’s consider a generic pay-off VT = Ξ(ST ). This is an European style pay-off because
it depends solely on the terminal value of S. The initial price V0 of the claim will be
equal to

N  
X N i
V0 = E Q
[Ξ(ST )BT−1 ] = i N −i
Ξ(S0 u d ) q (1 − q)N −i . (35)
i=0
i

If Ξ(ST ) = (ST − K)+ for example, V0 will be the price of a call option. It is instructive
to simplify formula (35) in the case a call option (a similar result can be obtained for
put option). We shall see that the formula will look surprisingly similar to Black and
Scholes formula. Formally let C0 represent the price at time zero of a call option and
let a ≡ inf{i : S0 ui dN −i > K}, then

N  
+ N
X
−rT i N −i
C0 = e (S0 u d − K) q i (1 − q)N −i
i=0
i
N  
−rN ∆t
X N i
= e i i
(S0 u d − K) q (1 − q)N −i
i=a
i
N
X N  N  
−r∆t i −r∆t N −i −r∆t
X N i
= S0 (uqe ) (de (1 − q)) − Ke q (1 − q))N −i
i=a
i i=a
i
N
X N   N
X N  
= S0 q̄ i (1 − q̄)N −i − Ke−rT q i (1 − q)N −i
i=a
i i=a
i
= S0 Q1 − Ke−rT Q2 ,

where we have defined q̄ ≡ uqe−r∆t ,

13
 N  
X N
Q1 ≡ q̄ i (1 − q̄)N −i , (36)
i=a
i

and

N  
X N
Q2 ≡ q i (1 − q)N −i . (37)
i=a
i

Q2 is the (binomial) probability of the underlying asset being greater than the strike at
maturity. Q1 is a similar measure but with slightly modified risk neutral probabilities.
We shall see that the Black and Scholes formula derived in a continuous time setting
resemble closely the formula above. Also, we shall show that as N → ∞ and ∆t → 0,
the distribution of the CRR asset process S converges to the distribution of a geometric
Brownian motion. Moreover, option prices under the binomial model converge to the
corresponding GBM prices.
So far we have taken the shape of the CRR tree as given. Is there an intuitive way to
choose u and d? A convenient choice is the following

√
u = eµ∆t+σ√∆t
d = eµ∆t−σ ∆t ,

where µ ∈ R and σ ∈ R+ are constants representing the drift and volatility of the
asset respectively.
We would like to derive the limit distribution of the process S as the time step decreases,
i.e. for N → ∞ (or equivalently ∆t → 0). In order to do so, define the random variable
XN to be the number of up moves in the first N steps. The asset process can be thus
be written as

14
 (N )
ST = S0 uXN dN −XN
 √ √ 
= S0 exp XN (µ∆t + σ ∆t) + (N − XN )(µ∆t − σ ∆t)
 √ 
= S0 exp µT + ∆tσ(2XN − N )
!
√ XN − 1/2N
= S0 exp µT + T σ 1
√
N
 √ 2
= S0 exp µT + T σZN

where we have defined

XN − 1/2N
ZN ≡ 1
√ . (38)
2
N

XN can be represented as a sum of i.i.d. random variables. In particular

N
X
XN = i , (39)
i=1

where

(
1 if Si /Si−1 = u with prob q
i = (40)
0 if Si /Si−1 = d with prob 1 − q.

Each i has a Bernoulli distribution with mean and variance

E Q [i ] = q (41)
V arQ [i ] = q(1 − q), (42)

15
respectively. It follows that XN has the following mean
N
X
Q Q
E [XN ] = E [ i ] (43)
i=1
N
X
= E Q [i ] (44)
i=1
= N q, (45)

and variance

N
X
Q Q
V ar [XN ] = V ar [ i ] (46)
i=1
N
X
= V arQ [i ] (47)
i=1
= N q(1 − q). (48)

In order to make further progresses, we need to find a simple approximation for the
risk neutral probability q. This will allow us to derive a simple explicit expression for
the mean and variance of XN and ultimately ZN . Consider the Taylor expansion of u,
d and B∆t in their respective exponents up to terms of order ∆t

√
u = eµ∆t+σ ∆t
√ 1
≈ 1 + µ∆t + σ ∆t + σ 2 ∆t,
2

√
d = eµ∆t−σ ∆t
√ 1
≈ 1 + µ∆t − σ ∆t + σ 2 ∆t,
2

and

B∆t = er∆t
≈ 1 + r∆t.

16
Substituting the expansions of u, d and B∆t into (31) and (32) , we obtain

1 1 √ (r − µ) − 1/2σ 2
+q ≈∆t (49)
2 2 σ
1 1 √ (r − µ) − 1/2σ 2
1−q ≈ − ∆t . (50)
2 2 σ

Using the results above, the mean of XN can be approximated as follows

E Q [XN ] = N q (51)
1 1√ √ (r − µ) − 1/2σ 2
= N+ N N ∆t (52)
2 2 σ
1 1 √ √ (r − µ) − 1/2σ 2
= N+ N T . (53)
2 2 σ

It is then straightforward to see that the mean of ZN is equal to

E Q [XN ] − 1/2N √ (r − µ) − 1/2σ 2

E Q [ZN ] = √ = T . (54)
1/2 N σ

Similarly the the variance of XN is given by

V arQ [XN ] = N q(1 − q)

2
(r − µ) − 1/2σ 2

1 1
= N − N ∆t
4 4 σ

and

V Q [XN ]
V arQ [ZN ] =
1/4N
2
(r − µ) − 1/2σ 2

= 1 − ∆t .
σ

17
Note that the mean of ZN does not depend on N . Moreover that variance of ZN → 1
as ∆t → 0 (or equivalently N → ∞). Using the Central Limit Theorem (ZN is a
rescaled, mean adjusted sum of the i.i.d. random variables i ), we can prove that the
distribution of ZN converges to the distribution of normal random variable with mean
m and standard deviation v given by

√ (r − µ) − 1/2σ 2
m = T (55)
σ
v = 1. (56)

Let Z = limN →∞ ZN and Y be N (0, 1). Then Z admits the following representation

Z = m + Y. (57)

We have now all the ingredients to derive the limiting distribution of the asset process
(N )
ST

(N )
ST ≡ lim ST
N →∞
 √ 
= S0 exp µT + T σZ
 √ 
= S0 exp µT + T σ(m + Y )
 
1 2
= S0 exp (r − σ )T + σWT
2

√
where we have defined WT ≡ Y T . We have just shown that as N → ∞, the
dynamics of the asset S under the CRR model converge to a GBM which form the
basis of continuous time pricing and the Black and Scholes framework.

1.4.2 Martingales and The Fundamental Theorem of Asset Pricing

In a multi-period model, the version of the fundamental theorem of asset pricing in-
troduced in section 1.2.2 needs to be generalised slightly.

18
Theorem 2 There is no arbitrage if and only if there exists a measure Q, the Risk
Neutral Measure, with respect to which discounted asset prices are martingales.

Remember that a process M is called a Martingale (relative to (Q, Fn )) if

1. M is adapted,

2. E Q [| Mn |] < ∞, for all n,

3. E Q [Mn | Fn−1 ] = Mn−1 a.s. (n ≥ 1).

Is the CRR model free of arbitrage then? The answer is of course yes. . In order
to prove the claim, it is sufficient to show that the process Mn ≡ Sn Bn−1 is a (Q, Fn )
martingale. For simplicity, we shall assume that {F}n≥0 is the filtration generated by
the process S.
Consider a generic time 1 ≤ n ≤ N and let i be the series of i.i.d. random variables
defined in (40). Observe that

E Q [Mn | Fn−1 ] = E Q [Sn Bn−1 | Fn−1 ]

−1 n 1−n −r∆t
= E Q [Sn−1 Bn−1 u d e | Fn−1 ]
−1 Q n 1−n −r∆t
= Sn−1 Bn−1 E [u d e | Fn−1 ]
−1 Q n 1−n −r∆t
= Sn−1 Bn−1 E [u d ]e
= Sn−1 Bn−1 [qu + (1 − q)d]e−r∆t
−1

−1
= Sn−1 Bn−1 [q(u − d) + d]e−r∆t
−1
= Sn−1 Bn−1 [er∆t − d + d]e−r∆t
−1
= Sn−1 Bn−1
= Mn−1 .

Furthermore note that for each n, the process S is positive and bounded, which implies
that for all n

E Q [| Sn |] < ∞. (58)

It follows that CRR model is arbitrage free.

19
1.4.3 Hedging in the CRR Model

Hedging in the CRR model is no more difficult than hedging in the simple one period
binomial model. The CRR model can in fact be decomposed in a series of one period
binomial models.
Start from the penultimate time step and for each node consider the only two possible
outcomes. More precisely, consider node i and time t = N − 1 where i ∈ {0, . . . , N − 1}

SN,i

SN −1,i

SN,i+1

and the corresponding option prices

CN,i

CN −1,i =?

CN,i+1

Solving a system of equations similar to the one encountered in the one period model,
we can derive the hedging ratios as follows

CN,i − CN,i+1
∆N −1,i =
SN,i − SN,i+1
−1
φN −1,i = [CN,i+1 − ∆N −1,i SN,i+1 ]B∆t

By construction the hedging portfolio above will be equal to the option pay-off CN no
matter whether the stock price goes up or down. Next, we need to derive the option
price CN −1,i which can be done using the martingale property of discounted asset (and
thus option) prices, i.e.

20
 −1 Q
CN −1,i = B∆t E [CN | SN −1 = SN −1,i ]
−r∆t
= e [qCN,i + (1 − q)CN,i+1 ].

The steps above need to be carried out for all i ∈ {0, . . . , N − 1}. Next, move back on
time step and repeat the exact same calculations. Summing up, we need to run the
following algorithm:
For k = 1 to N
For i = 0 to N − k

CN −k+1,i − CN −k+1,i+1
∆N −k,i =
SN −k+1,i − SN −k+1,i+1
−1
φN −k,i = [CN −k+1,i+1 − ∆N −k,i SN −k+1,i+1 ]B∆t
CN −k,i = e−r∆t [qCN −k+1,i + (1 − q)CN −k+1,i+1 ].

end
end

21
2 Continuous Time Models

2.1 Brownian Motion

Brownian motion is one of the fundamental concepts in stochastic calculus and math-
ematical finance. It is a stochastic process W satisfying:

• W0 = 0.

• Wt is continous in t.

• For all 0 = t0 < t1 < · · · < tm , the increments Wt1 − Wt0 , Wt2 − Wt1 , . . . ,
Wtm − Wtm−1 are independent.

• Given any positive t and s the random variable Wt+s −Wt is normally distributed
with mean zero and variance s.

In what follows we shall analyse a few properties of the Brownian motion which will
be key in the construnction of stochastic integrals and the derivation of Ito’s formula.
Partition the time interval [0, t] as follows Π ≡ {0 = t0 < t1 , . . . tn = t} and let

| Π |≡ max ti+1 − ti ,
0≤i≤n−1

be the largest interval in the partition.

The quadratic variation of a process X over a fixed interval, say [0, t], is defined as

n−1
X
[X, X]t = lim (Xti+1 − Xti )2 . (59)
|Π|↓0
i=0

Note that the quadratic variation is defined path-wise and will in general be a ran-
dom variable for fixed t or a stochastic process if seen as a function of t. We will
show however, that in the case of Brownian motion, the quadratic variation [W, W ]t is
deterministic and equal to t.

Lemma 3 The sampled quadratic variation of the Brownian motion [W, W ]nt over the
interval [0, t] converges to t in L2 , more precisely

22
 lim E ([W, W ]nt − t)2 = 0.
|Π|↓0

Here we have defined the sampled quadratic variation, i.e. the quadratic variation
corresponding to a given partition, as

n−1
X
[W, W ]nt ≡ (Wti+1 − Wti )2
i=0

Proof. Consider first the expected sampled quadratic variation

" n−1 #
X
E[W, W ]nt = E (Wti+1 − Wti )2
i=0
n−1
X
= E[(Wti+1 − Wti )2 ]
i=0
n−1
X
= (ti+1 − ti )
i=0
= t.

The expected sampled quadratic variation is thus independent of partition under con-
sideration. What about the variance?

23
 V ar([W, W ]nt ) = E ([W, W ]nt − E[W, W ]nt )2
= E ([W, W ]nt − t)2
n−1
!2
X
2
= E (Wti+1 − Wti ) − t
i=0
n−1
!2
X
= E (Wti+1 − Wti )2 − (ti+1 − ti )
i=0
n−1
X
E (Wti+1 − Wti )4 − 2(ti+1 − ti )(Wti+1 − Wti )2 + (ti+1 − ti )2
 
=
i=0
n−1
X
E 3(ti+1 − ti )2 − 2(ti+1 − ti )2 + (ti+1 − ti )2
 
=
i=0
Xn−1
= 2 (ti+1 − ti )2
i=0
n−1
X
≤ 2|Π| (ti+1 − ti ),
i=0
= 2|Π|t

where we have used the fact that increments of the Brownian motion on non overlapping
intervals are independent. Now, letting | Π | go to zero (or equivalently letting n ↑ ∞)
it is easy to see that V ar([W, W ]nt ) ↓ 0. Note also that the second line of the derivation
above implies L2 convergence.
We shall summarise the result in differential form as follows
d[W, W ]t = (dWt )2 = dt.

The quadratic variation of the Brownian motion will play a key role in the construction
of Ito’s integral. Another important result which we shall need in the derivation of Ito’s
formula is the following:

Lemma 4 The quantity [W, T ]nt defined as

n−1
X
[W, T ]nt ≡ (Wti+1 − Wti )(ti+1 − ti ),
i=0

24
converges to 0 in L1 , i.e.

lim | [W, T ]nt − 0 |= 0.

|Π|↓0

Proof. Consider the following

n−1
X
| [W, T ]nt | ≤ | (Wti+1 − Wti )(ti+1 − ti ) |
i=0
n−1
X
= | (Wti+1 − Wti ) | (ti+1 − ti )
i=0
n−1
X
≤ max | (Wti+1 − Wti ) | (ti+1 − ti )
0≤i≤n−1
i=0
≤ max | (Wti+1 − Wti ) | t
0≤i≤n−1

which tends to 0 as | Π |↓ 0 because Brownian paths are continuous.

We shall summarize this result by writing

d[W, T ]t = dWt dt = 0.

The last important result we shall need in the derivation of Ito’s formula is the following

Lemma 5
n−1
X
[T, T ]t ≡ lim (ti+1 − ti )2 ≡ lim [T, T ]nt
|Π|↓0 |Π|↓0
i=0

Proof.

n−1
X
[T, T ]nt = (ti+1 − ti )2 (60)
i=0
n−1
X
≤ |Π| (ti+1 − ti ) (61)
i=0
= | Π | t. (62)

25
Letting | Π |↓ 0 we obtain the desired result.
In differential form, we shall write

d[T, T ]t = (dt)2 = 0.

2.2 Stochastic Integrals

Let ∆t be an Ft adapted process such that

Z t
E[ ∆2u du] < ∞ (63)
0

for all t. We would like to make sense of integrals of the type

Z t
It ≡ ∆u dWu (64)
0

where Wt is a Brownian motion adapted to Ft . The stochastic integral above is often

called Ito’s integral and it is a stochastic process if seen as a function of t. Does such an
integral exists and in what sense? The integral does not exists in the Rienman-Stieltjes
sense because the path of the Brownian motion are nowhere differentiable. However,
the stochastic integral can be defined as the limit (in L2 ) of simpler integrals.
Changing the notation slightly to emphasise that It is a stochastic process, let 0 ≤ t ≤ T
and consider the partition Π ≡ {0 = t0 < t1 < · · · < tn = T }. A simple process
∆nt is a piece-wiese constant stochastic process. In particular, let ∆nt be an Ft adapted
process which is constant for all t ∈ [tj , tj+1 ) and j = 0, . . . , n − 1. For t ≤ T , let’s
define the stochastic integral with respect to the simple process above as

Z t k−1
X
Itn ≡ ∆nu dWu = ∆tj (Wtj+1 − Wtj ) + ∆tk (Wt − Wtk ). (65)
0 j=0

Lemma 6 The process Itn is a Martingale.

Proof. It is straightforward to see that E[| Itn |] < ∞. Moreover Itn is adapted because
both Wt and Itn are Ft adapted. It remains to prove that

26
 E[Itn | Fs ] = Isn , (66)

for all 0 ≤ s < t. For simplicity of exposition, we shall limit our attention to the set
s ∈ {0 = t0 , . . . , tn−1 }.
For i < k consider

" k−1 #
X
E[Itn | Fti ] = E ∆tj (Wtj+1 − Wtj ) + ∆tk (Wt − Wtk ) | Fti
j=0
k−1
X
= E[∆tj (Wtj+1 − Wtj ) | Fti ] + E[∆tk (Wt − Wtk ) | Fti ].
j=0

For j ≥ i we have that

E[∆tj (Wtj+1 − Wtj ) | Fti ] = E[∆tj E[(Wtj+1 − Wtj ) | Ftj ] | Fti ] (67)
= E[∆tj (E[Wtj+1 | Ftj ] − Wtj ) | Fti ] = 0 (68)

where we have used the tower property of conditional expectations, the fact that the
Brownian is a martingale with respect to Ft and the adaptness of ∆t . Similarly we
have that

E[∆tk (Wtt − Wtk ) | Fti ] = E[∆tk (E[Wt | Ftk ] − Wtk ) | Fti ] = 0. (69)
The calculations above imply that
" i−1 #
X
E[Itn | Fti ] = E ∆tj (Wtj+1 − Wtj ) | Fti = Itni . (70)
j=0

The proof for a general time s is similar. Hence, the stochastic integral for simple
integrand processes is a martingale.
Another fundamental property of stochastic integrals is the so called Ito’s isometry.
This property is very important because it will be used to prove the existence of
stochastic integrals for (more) general integrands.

27
Lemma 7 Let Itn be defined as in (64); the following equality holds

"Z 2 #
t Z t 
E[(Itn )2 ] =E ∆nu dWu =E (∆nu )2 du . (71)
0 0

Proof. In order to simplify the notation let

Dj = Wtj+1 − Wtj j = 0, . . . , k − 1
Dk = Wt − Wtk .

From elementary algebra it follows that

k
X X
(Itn )2 = (∆ntj )2 Dj2 + 2 ∆nti ∆ntj Di Dj .
0 0≤i<j≤k

Note that for i < j, using the tower property of conditional expectations and the fact
that ∆n and W are adapted, we have

E[∆nti ∆ntj Di Dj ] = E[E[∆nti Di ∆ntj Dj | Fj ]]

= E[∆nti Di ∆ntj E[Dj | Fj ]]
= E[∆nti Di ∆ntj ]E[Dj ]
= 0

where we have used the fact that

E[Dj | Fj ]] = E[Dj ] (72)

= E[Wtj+1 − Wtj ] (73)
= E[Wtj+1 ] − E[Wtj ] = 0 + 0. (74)

From the calculations above and the linearity of the expectation operator, it follows
that

k
X
E[(Itn )2 ] = E[(∆ntj )2 Dj2 ]. (75)
j=0

28
Using again the tower property, the terms on the right hands side of the summation
can be simplified as follows

E[(∆ntj )2 Dj2 ] = E[E[(∆ntj )2 Dj2 | Fj ]]

= E[(∆ntj )2 E[Dj2 | Fj ]]
= E[(∆ntj )2 E[(Wtj+1 − Wtj )2 | Fj ]]
= E[(∆ntj )2 E[(Wtj+1 − Wtj )2 ]]
= E[(∆ntj )2 ]E[(Wtj+1 − Wtj )2 ]
= E[(∆ntj )2 ](tj+1 − tj )

Finally, substituting the result above into (75) and exchanging the oder of summation
and expectation, we obtain the desired result, i.e.

k
X Z t 
E[(Itn )2 ] = E[(∆ntj )2 ](tj+1 − tj ) = E (∆nu )2 du .
0 0

We have thus succedeed in proving that Ito’s isometry holds when the integrand is a
simple process. What about the quadratic variation of such stochastic integrals?

Theorem 8 Z t 
n n
(∆nu )2 du = E (Itn )2 .
 
E([I , I ]t ) = E (76)
0

In other words, the quadratic variation of the Ito’s integral is equal to its variance.

Proof. We can sub-divide each interval [tj , tj+1 ) of the partition Π in smaller sub-
intervals [si,j , si+1,j ) for i = 0, . . . , m − 1 such that s0,j = tj and sm,j = tj+1 . For
simplicity, but without loss of generality, assume that t = T = Tn .

29
 n−1 m−1
X X
n n
[I , I ]t = lim (Isni+1,j − Isni,j )2
m↑∞
j=0 i=0
n−1 m−1
X X
2
= (∆ntj )2 lim Wsi,j+1 − Wsi,j
m↑∞
j=0 i=0
n−1
X
= (∆ntj )2 (tj+1 − tj )
j=0
n−1 Z tj+1
X
= (∆ntj )2 du
j=0 tj
Z t
= (∆nu )2 du
0

Is it possible to generalise the notion of stochastic integral to more general integrands?

The answer is of course yes. In what follows, we shall consider the class of adapted
integrand ∆t satisfying property (63). Note that the process ∆ may in general be
discontinuous.

Lemma 9 Given a process ∆ satisfying property (63) and a time t, there exists a
sequence of simple processes ∆n converging to ∆ in the sense below

Z t 
lim E (∆nu 2
− ∆u ) du = 0. (77)
n↑∞ 0

The Ito’s integral for such a process will be defined as

Z t Z t
It ≡ ∆u dWu = lim ∆nu dWu .
0 n→∞ 0

where the limits holds in L2 .

Proof. A full proof of the lemma above is beyond the scope of this notes and we shall
present only a sketch of the main steps. In particular, we shall take for granted that
given ∆u , an approximating sequence satistisfying (77) actually exists.

30
We need to show that if we construct a series of stochastic integrals Itn starting from
the simple processes ∆n , this sequence converges to something. We shall define this
’something’ to be the stochastic integral for ∆. Note that the integrals
Z t
n
It ≡ ∆nu dWu ,
0
2
form a Cauchy sequence in L (Ω, F, P). Using Ito’s isometry we can show that for any
 > 0, there exists an integer N , such that for all n, m ≥ N ,

Z t Z t 2
E (Itn − Itm )2 = E ∆nu dWu ∆m
 
− u dWu
0 0
Z t 2
= E (∆nu − ∆m
u )dWu
0
Z t 
= E (∆nu − ∆m 2
u ) du < .
0

The calculations above show that the sequence Itn is Cauchy and converges in L2 (Ω, F, P).

Note also that the proof above and (76) imply that
Z t 2 Z t 
 2 2
E It = E ∆u dWu = E ∆u du ,
0 0
i.e., Ito’s isometry holds for general integrands. Last but not least, the process It is a
martingale.
Summary of Important Properties

• The Ito’s integral It for general adapted, square integrable processes (i.e. pro-
cesses satisfying condition (63) )is well defined.
• It is a martingale.
• Ito’s isometry holds: Z t 
E[It2 ] =E ∆2u du .
0

• The expected quadratic variation of It is equal to its variance, i.e.

Z t 
E([I, I]t ) = E (∆u ) du = E[It2 ].
2
0

31
2.3 Ito’s Formula

Ito’s formula allows us to study the behaviour of functions of the Brownian motions
and other more general processes. It is a very powerful tool in stochastic calculus and
derivatives pricing. We shall make use of Ito’s formula extensively in our analysis of
the celebrated Black and Scholes model.

Lemma 10 Let f (t, w) be a real-valued function with continuous first and second
derivative and let Wt be standard Brownian motion. The following holds:

Z T Z T Z T
1
f (T, WT ) = f (t, Wt )+ ft (u, Wu )du+ fw (u, Wu )dWu + fww (u, Wu )dt. (78)
t t 2 t

where 0 ≤ t ≤ T and ft , fw , fww indicate the partial derivative of f (t, w) with respect
to t and w.

It is convenient, when doing calculations, to write formula above in differential form.

The following expression does not have a mathematical meaning of its own and should
only be consider a short-cut.

1
df (t, Wt ) = ft (t, Wt )dt + fw (t, Wt )dWt + fww (t, Wt )dt. (79)
2
Proof. We shall present a heuristic proof of Ito’s formula. Consider the stochastic
Taylor expansion of f (t, Wt )

1 1
df (t, Wt ) = ft (t, Wt )dt + fw (t, Wt )dWt + fww (t, Wt )(dWt )2 + ftt (t, Wt )(dt)2
2 2
+ ftw (t, Wt )dtdWt + O(3, 3)

where O(3, 3) indicate terms of oder 3 and above in t and Wt . Using the equalities

d[W, W ]t = (dWt )2 = dt
d[T, T ]t = (dt)2 = 0
d[W, T ]t = dtdWt = 0.

32
it is easy to see that terms O(3, 3) must all be equal to zero and equation (80) simplifies
to (79).
A simple extension of ito’s formula will allow us to deal with functions f (t, Xt ) of
general Ito’s processes.

Lemma 11 Let f (t, w) be a real-valued function with continuous first and second
derivative and let Xt an Ito’s process, i.e. a stochastic process of the form

Z t Z t
Xt = X0 + ∆u dWu + φu dBu , (80)
0 0

where Bt is an adapted process of finite variation. Here ∆t and Bt satisfy

Z T Z T 
E ∆2t dt + | Bt | dt < ∞.
0 0

The following holds

Z T Z T Z T
1
f (T, XT ) = f (t, Xt ) + ft (u, Xu )du + fw (u, Xu )dXu + ∆2u fww (u, Xu )du,
t t 2 t
(81)
where as before 0 ≤ t ≤ T and ft , fw , fww indicate the partial derivative of f (t, w) with
respect to t and w respectively.

Proof. The proof of Ito’s formula for general Ito processes is similar to the one
encountered for the Brownian motion. The first step is to perform a Taylor expansion
of f (t, Xt )

1 1
df (t, Xt ) = ft (t, Xt )dt + fw (t, Xt )dXt + fww (t, Xt )(dXt )2 + ftt (t, Xt )(dt)2
2 2
+ ftw (t, Xt )dtdXt + O(3, 3).

Note that Xt can be express in diffential form as

33
 dXt = ∆t dWt + φt dBt . (82)

Now, the quadratic variation d[X, X]t of the process Xt is equal to

d[X, X]t = (dXt )2 = (∆t dWt + φt dBt )2

= ∆t (dWt )2 + φ2t (dBt )2 + ∆t φt dBt dWt
= ∆2t dt

where we have used the fact that the quadratic variation d[B, B]t = (dBt )2 of a fi-
nite variation process is equal to zero and the covariation d[B, W ]t = dBt dWt of the
Brownian motion with a finite variation process is also equal to zero

0
dBt = Bt dt
0
(dBt )2 = (Bt )2 (dt)2 = 0
0
dBt dWt = Bt (dtdWt ) = 0.

Similar calculations show that the term dtdXt also vanishes. Substituting the results
above into the Taylor expansion of f (t, Xt ) we obtain the desired result.
Another powerful tool is the stochastic version of the integration by parts (IBP) formula
which we shall report without proof.

Lemma 12 Let Xt and Yt be two Ito processes. The following is true

Z t Z t
Xt Yt = X 0 Y0 + Xu dYu + Yu dXu + [X, Y ]t (83)
0 0
which in differential form reads

d(Xt Yt ) = Xt dYt + Yt dXt + dXt dYt , (84)

where as usual we have used the short-cut d[X, Y ]t = dXt dYt .

34
2.4 Applications of Ito’s lemma and IBP formula

2.4.1 Black and Scholes Stock Dynamics

Consider the following stochastic differential equation (SDE)

dSt = µSt dt + σSt dWt , (85)

for some real µ and positive σ. The process above is often used to model stock dynamics.
Note that (85) is a stochastic differential equation because St appears on both sides of
the equal sign. We can use Ito’s formula to find a solution to the SDE. Consider the
function

f (w) = ln w.

Using Ito’s formula to calculate df (St ) and noting that ft (w) = 0, fw (w) = 1/w and
fww (w) = −1/w2 , we have that

1 1 1
df (St ) = d ln St = dSt − (dSt )2
St 2 St2
1 1 1
= (µSt dt + σSt dWt ) − (µSt dt + σSt dWt )2
St 2 St2
1 1
= µdt + σdWt − 2
(µSt2 (dt)2 + 2σµdtdWt + σ 2 St2 dWt )
2 St
 
1
= µ − σ 2 dt + σdWt ,
2

which in integral form can be written as

Z t  Z t
1 2
ln St = ln S0 + µ − σ du + σdWu
0 2 0
 
1 2
= ln S0 + µ − σ t + σWt .
2
Finally exponentiating both side of equality

35
   
1 2
St = S0 exp µ − σ t + σWt . (86)
2

2.4.2 Vasicek Model

Vasicek model, often use in interest rate theory modelling, provides us with an inter-
esting application of the integration by parts formula. Consider the SDE

dXt = (α − βXt )dt + σdWt .

The standard way to solve the SDE above is to apply the IBP formula to the product
eβt Xt

d(eβt Xt ) = eβt ((α − βXt )dt + σdWt ) + βeβt Xt + deβt dXt (87)
= eβt (αdt + σdWt ) . (88)

Multiplying both side of the equality eβt and integrating after some re-arrangment we
obtain

Z t
−βt α
1 − e−βt + σ e−β(t−u) dWu .


Xt = X0 e +
β 0

2.5 The Black and Scholes Model

Despite its limitations, Black and Scholes is still fundamental to understand the pricing
and hedging of financial derivatives in continuous time. Assume that the dynamics of
the underlying asset S satisfy the SDE

dSt = µSt dt + σdWt ,

under the physical measure P. As usual µ is a real number, σ is a positive real number
and Wt is a standard Brownian motion.

36
2.5.1 Hedging Approach

Assume we have bought a claim which pays CT = Φ(ST ) at maturity, for some pay-off
function Φ(x). How do we go about hedging the claim? What is the value of such a
claim at time before expiry?
Similarly to the discrete time framework, we need to form a portfolio consisting of ∆t
units of the asset and φt units of the money market account with value equal to the
option price (say Ct ) for any t before expiry. In order words, we shall impose that

πt ≡ ∆t St + φt Bt = Ct , (89)
for all 0 ≤ t < T . Again we shall restrict our attention to self-financing portfolios. In
continous time a portfolio is self-financing if it satisfies

Z t Z t
πt = π0 + ∆u dSu + φu dBu , (90)
0 0

for all 0 ≤ t < T . Equations (89) and (90) imply that the initial price of the option
will be given by

C0 = π0 = ∆0 S0 + φ0 , (91)
and Z T Z T
C T = π0 + ∆u dSu + φu dBu .
0 0

Pricing and hedging the option amounts to finding the hedging ratios ∆t and φt for all
0 ≤ t < T . If we assume for now that we have found a portfolio (∆0 , φ0 ) such that
(91) holds, in oder for πt to equal Ct or equivalently for Xt ≡ Ct − pit to hold for all t,
we must have

dXt = dCt − dπt = 0.

Assuming that Ct = C(t, St ) and C(t, S) ∈ C 2 i.e. the price of the option is a function
of the current time and asset price with twice differentiable, continuous derivatives, we
can apply Ito’s formula to dXt

37
 dXt = dC(t, St ) − ∆t dSt − φt dBt
 
1 2 2
= Ct (t, St ) + σ St CSS (t, St ) dt + CS (t, St )dSt − ∆t dSt − rφt Bt dt.
2
By choosing

∆t = CS (t, St ), (92)
we can eliminate the asset risk impact from the portfolio Xt . Moreover from (89) and
the equation above, it follows that

φt = (C(t, St ) − CS (t, St )St )Bt−1 , (93)

which in turns implies

 
1 2 2
dXt = Ct (t, St ) + σ St CSS (t, St ) − rC(t, St ) + rSt CS (t, St ) dt (94)
2
In oder for dXt = 0 and XT = 0, the following PDE needs to be satisfied

(
Ct (t, S) + 12 σ 2 S 2 CSS (t, S) + rSCS (t, S) − rC(t, S) = 0
(95)
C(T, S) = Φ(S)

for all S and 0 ≤ t < T .

The price of the option C(t, S) can be calculated by solving the PDE above analytically
(if possible) or numerically. Once the price of the option has been calculated, the
hedging policy can be obtained formulae (92) and (93).

2.5.2 Feynman-Kac Theorem

In the case of call and put options it is possible to solve the Black and Scholes PDE
above explicitly by reducing it, via some variable transformation, to the heat equation.
However, it is easier as well as instructive to solve the pricing problem via a probabilistic
approach and show that the solution is also the solution of the Black and Scholes PDE.

38
Feynman-Kac theorem establishes a connection between the PDE and the probabilistic
representation of the option price.

Theorem 13 Let F (t, x) solve the following PDE

(
Ft (t, x) + 12 σ 2 x2 Fxx (t, x) + kxFx (t, x) − kF (t, x) = 0
C(T, x) = Φ(x)
and let

dXt = kXt dt + σXt dWt

where Wt is a standard Brownian motion, k, σ ∈ R. If

Z T
2

−kt
E e σXt Fx (t, Xt ) dt < ∞, (96)
0

then
F (t, Xt ) = E e−k(T −t) Φ(XT ) | Ft .
 

Proof. In order to prove the claim, let’s apply the IBP formula and Ito’s formula to
the product e−kt F (t, Xt )

d(e−kt F (t, Xt )) = −ke−kt F (t, Xt )dt + e−kt dF (t, Xt )

 
−kt −kt 1 2 2
= −ke F (t, Xt )dt + e Ft (t, Xt ) + σ Xt Fxx (t, Xt ) + kXt Fx (t, x) dt
2
−kt
+ e σXt Fx (t, x)dWt
 
−kt 1 2 2
= e Ft (t, Xt ) + σ Xt Fxx (t, Xt ) + kXt Fx (t, x) − kF (t, Xt ) dt
2
−kt
+ e σXt Fx (t, x)dWt .

The dt terms in the expression above is equal to 0 by assumption. Integrating and

taking expectations on both side of the equality, we obtain

39
 Z T 
E e−kT F (T, XT ) | Ft = e−kt F (t, Xt ) + E −ku
 
σe Xu Fx (u, x)dWu | Ft .
t

The last term on the RHS of the equality is equal to zero because the stochastic integral
Z t
It = σe−ku Xu Fx (u, x)dWu
0
is a Martingale (note condition (96)). Rearranging and using the terminal condition,
we finally obtain

F (t, Xt ) = E e−k(T −t) Φ(XT ) | Ft

 

We can now make use of Feynman-Kac theorem to solve the Black and Scholes PDE.
Setting k = r and amending the notation slightly, we have that

C(t, St ) = E Q e−r(T −t) Φ(ST ) | Ft ,

 
(97)

where
dSt = rSt dt + σSt dW̃t , (98)

and W̃t is a Q Brownian motion.

In other words, the Black and Scholes PDE, implies that the risk neutral drift rate
of the underlying process must be given by r. We shall prove that this are the risk
neutral dynamics, i.e. Q is the risk neutral measure and under Q all discounted asset
price processes (including St ) are Martingales.
In order to derive the price C(t, St ) of the option we can either solve the expectation
(97) or the Black and Scholes PDE. For general pay-offs, one approach may be better
than the other, but in the case of calls and puts they are both relatively simple and lead
to close form solutions. Before deriving Black and Scholes formula explicitly however
we shall introduce the concept change of measure.

2.5.3 Change of Measure and Girsanov Theorem

Two measures P and Q are said to be equivalent if they agree on a set of probability
zero, i.e. if A is a set such that P(A) = 0 then Q(A) = 0 and vice-versa. Whenever

40
two measures are equivalent, we can define one as a functional of the other.

Theorem 14 Let P and Q be equivalent, then there exists a random variable Z, such
that

P({Z > 0}) = 1,

E P [Z] = 1
and

Q(A) = E P [1{A} Z]
 
P 1
P(A) = E 1{A}
Z

for all A ∈ F.
The reverse is also true, i.e. if such Z exists, then the measures P and Q defined above
are equivalent.

The theorem above, known as Radon-Nykodim theorem, also implies that if P and Q
are equivalent, then

E Q (X) = E P [XZ]
and  
P 1 Q
E (X) = E X
Z
where X is an integrable random variable.
How do the dynamics of the Brownian motion and in general the asset dynamics St
change when we change the measure? What is the relationship between the physical
measure P and the risk neutral measure Q? Girsanov’s theorem will help us answering
those questions.

41
Theorem 15 Let Wt be a Brownian motion with respect to (P, Ft ) and let θt be an
Ft ) adapted process such that

  Z T 
1 2
E exp θ du < ∞.
2 0 u

Define the process

1 t 2
 Z Z t 
dQ
Lt ≡ |F = exp − θ du − θu dWu ,
dP t 2 0 u 0

and

Q(A) = E P [1{A} Lt ]

for all A ∈ F. Note that Lt is a (P, Ft ) Martingale.

Then the process

Z t
W̃t = Wt + θu du
0

is a (Q, Ft ) Brownian motion. Moreover, given a stochastic process Xt and 0 ≤ s < t,

the following equality holds

E P [Xt Lt | Fs ]
E Q [Xt | Fs ] = (99)
Ls

Intuitively, changing
R t measure is equivalent to shifting the mean of Brownian motion
from zero to − 0 θu du. If Wt is Brownian motion under P, under Q it will be a
Brownian motion (a Q Brownian motion to be precise) minus a drift term.
The following theorem makes use of Girsanov’s theorem to give a characterisation of
the notion of no-arbitrage.

Theorem 16 If there is no arbitrage, then there exists a probability measure Q (the

Risk Neutral Measure) equivalent to P (the Physical or Real Word Measure) under
which discounted asset prices are Martingales. In particular the process

42
 S̃t ≡ St e−rt ,

is a (Q, Ft ) Martingale. Moreover, the change of measure density is given by

 
dQ 1 2
Lt ≡ |F = exp − θ t − θWt ,
dP t 2
with
µ−r
θ= .
σ

Proof. Under P a simple application of the IBP formula shows that the dynamics of
S̃t are given by

dS̃t = (µ − r) S̃t dt + σ S̃t dWt .

Making use of Girsanov’s theorem, we can derive the dynamics of Wt under Q

µ−r
dWt = dW̃t − θdt = dW̃t − dt,
σ

where W̃t is a Q Brownian motion. The dynamics of S̃t under the risk neutral measure
will then be

dS̃t = σ S̃t dW̃t

or equivalently

Z t
S̃t = S0 + σ S̃u dW̃u
0

It follows that S̃t is a Martingale under the risk neutral measure Q.

From the results above,it is easy to see (the proof is left as an exercise) that the process
St has the following Q dynamics

43
 dSt = rSt dt + σSt dW̃t ,
which are the same dynamics we found using the PDE approach in conjunction with
Feynman-Kac theorem.

2.6 Black and Scholes Formula

We are now ready to calculate the price of a call option (deriving the price of a put op-
tion is left as an exercise). We shall solve the problem using the probabilistic approach
rather than solving the PDE. As mentioned before however, the two approaches are
equivalent and lead to the same result. Remember that the pay-off of a call option
with strike K and maturity T is equal to

Φ(S) = (S − K)+ .

The price of a call option at any time t prior to expiry will be the given by

C(t, St ) = EtQ (ST − K)+ e−r(T −t)

 

= EtQ (ST − K) 1{ST >K} e−r(T −t)

 

= e−r(T −t) EtQ [ST 1{ST >K} ] − Ke−r(T −t) Qt ({ST > K})
≡ P1 − P2 .

Using the properties of Brownian motion, the second term of the RHS of the equal sign
reduces to

P2 ≡ Ke−r(T −t) Qt ({ST > K})

1 2
= Ke−r(T −t) Qt ({St e(r− 2 σ )(T −t)+σ(WT −Wt ) > K})
ln(K/St ) − (r − 21 σ 2 )(T − t)
 
−r(T −t)
= Ke Qt {W̃T − W̃t > }
σ
ln(St /K) + (r − 12 σ 2 )(T − t)
 
−r(T −t)
= Ke N √ }
σ T −t

44
where N(·) is the cumulative distribution function of a normal-(0, 1) random variable.
We can solve the first expectation in a similar way.

P1 ≡ e−r(T −t) EtQ [ST 1{ST >K} ]

1 2 (T −t)+σ(W
= EtQ [St e 2 σ T −Wt )
1{ST >K} ]

It is convenient and instructive to use a change of measure to calculate the expectation

above. Set

 
dQ̂ 1 2
Λt ≡ |F = exp − σ t + σ W̃t
dQ t 2
and let

Ŵt = W̃t − σt

be a Q̂-Brownian motion. Using formula (99), we can use the change of expectation to
simplify P1 as follows

ΛT
P1 = St EtQ [ 1{ST >K} ]
Λt
= St EtQ̂ [1{ST >K} ]
= St Q̂t ({ST > K})
ln(K/St ) − (r + 21 σ 2 )(T − t)
 
= St Q̂t {ŴT − Ŵt > }
σ
ln(St /K) + (r + 12 σ 2 )(T − t)
 
= St N √ .
σ T −t

In summary the Black and Scholes price of a call option is given by

−r(T −t)
C(t, St ) = St N(d+
t ) − Ke N(d−
t )

where

45
 +/− ln(St /K) + (r + / − 12 σ 2 )(T − t)
dt = √ .
σ T −t

46