TEACHING NOTE 00-04:

GIRSANOV’S THEOREM IN DERIVATIVE PRICING

Version date: October 14, 2003 C:\Classes\Teaching Notes\TN00-04.doc

The purpose of this teaching note is to present Girsanov’s Theorem as it applies to

derivative pricing and to provide some conjunction of the mathematical theory with the financial
economic theory. Girsanov’s Theorem appears to be a somewhat complex rule that, at best,
provides limited use if one already understands derivative pricing. Indeed it is quite possible to
obtain a solid understanding of derivative pricing without encountering Girsanov’s Theorem.
Nowhere did it appear in the literature on derivatives for many years, in spite of many
exceptionally powerful advances developed by financial economists. Yet Girsanov’s Theorem
provides much of the mathematical rigor that underlies derivative pricing. Unfortunately
Girsanov’s Theorem provides virtually no intuition and is rarely, if ever, integrated with any
form of economic discussion of what is really going on. In short, mathematicians approach the
derivative pricing problem in their own ways, one being the application of Girsanov’s Theorem,
and financial economists approach it other ways. That these two groups arrive at the same
answer is not surprising, but they speak different languages and travel different routes. Here we
shall try to bring them together a little. The discussion is admittedly more mathematical than
financial, but that is because the readership is likely to be more financial than mathematical and
is more in need of seeing the mathematics and being reminded of when the mathematicians are
saying something that a financial economist recognizes.
To price a derivative instrument such as an option, we can resort to any of several
methods. One of the more formal mathematical approaches is the application of Girsanov’s
Theorem. In short, what we do is alter the probability distribution of the stock return such that it
follows a stochastic process known as a martingale. A martingale, among other things, is a
process without a drift.1 In that case, we can obtain the value today of the derivative by
determining the expected future value of the derivative, where the expectation is arrived at by
using the altered probability distribution. In doing it this way, we avoid the problem of having to

1
In addition a martingale must be finite and each realization must be independent of the previous one.
D.M. Chance, TN00-04 1 Girsanov’s Theorem in Derivative Pricing
solve a differential equation, though we can still arrive at the solution in this manner by solving
the differential equation.
First, however, let us identify why we are doing this. One of the results of mathematical
finance is that any two stochastic processes that are both martingales when the same probabilities
are applied can be related to each other via a simple transformation. For example, given one
martingale, Xt, and another Yt, the relationship
t
Yt = Y0 + ∫ λ jdX j ,
0

represents one martingale in terms of the other. Accordingly, this is referred to as the martingale
representation theorem. It is important to note, however, and often overlooked in the
mathematical finance literature that these two processes cannot simply be any two martingales.
They must be uniquely linked. Let me explain what this means.
Suppose we consider X and Y, which fluctuate randomly and have expected returns of
zero. Nonetheless, X and Y are quite independent. X may be the price of a gas company in
Warsaw, Poland while Y might be the price of corn in Madison County, Iowa. There is no
amount of mathematical wizardry that can convert the price of a gas company in Warsaw to the
price of corn in Iowa, even though both may be martingales given the probability distribution of
states of the global economy. The martingale representation theorem seems to say that this can
be done. But what is missing can be easily seen by remembering the binomial model.
Suppose X can go up to X+ or down to X-. Now we want to model Y. Intuitively we
might specify that Y can go up to Y+ or down to Y-. If we allow only two states of nature and
jointly consider X and Y in the same model, then whenever X is X+, then Y is Y+ and whenever
X is X-, then Y is Y-. X and Y are perfectly related to each other. Knowledge of the value of
one reveals the value of the other. In a very loose sense X and Y are thought of as perfectly
correlated, though correlation is a specific type of relationship that does not have to hold to make
this point. Obviously when the price of corn is up, the price of the Polish gas company can be up
or down. There is probably little relation between our X and Y. Furthermore, we do not require
such extreme examples to make this point. X could be the price of Microsoft while Y is the price
of Exxon. Though both share a common relationship, as determined by general stock market
movements, there is no way that we can completely determine the value of one from the other.

D.M. Chance, TN00-04 2 Girsanov’s Theorem in Derivative Pricing

So if Microsoft and Exxon were martingales, we could not relate one to the other as the
martingale representation theorem seems to suggest.
But if X is the price of Microsoft and Y is the price of a derivative, such as a call option,
on Microsoft, then we may be able to connect the two. Indeed we can, since the price of the
latter is completely determined by the price of the former. In the binomial sense, we can easily
see this point. The option payoff at expiration for a given outcome is completely driven by the
stock price in that outcome.2 Thus, when we move to a continuous time framework and attempt
to employ the martingale representation theorem, we require that one martingale be completely
determined by the other.
How we use the martingale representation theorem is that if the price of a stock, as
indicated by the X variable above, is a martingale, then by the transformation above, we can turn
it into the Y variable, provided that Y is completely determined by X. If Y is the option on X,
then we have replicated Y with X. It should be apparent that λ is likely to represent a certain
number of shares of X held to replicate Y.
Those who have studied option pricing theory from a financial economics perspective,
however, will find this result a bit disconcerting, because you know that one must hold bonds as
well as shares to replicate an option. The martingale representation theorem gives us only the
condition under which the uncertainty in Y can be captured by the uncertainty in X. More
formally, if we differentiate Yt with respect to Xt,, we obtain
dYt = λ t dX t .

This is the result we need and the one that clearly indicates that the uncertainty in y is driven by
the uncertainty in X. As we know, we must holds bonds as well as stock to replicate an option.
But the martingale representation theorem does tell us that we need to find a stochastic
process λt such that we are holding just the right amount of X. In addition we must hold a
certain number of bonds, such that we replicate Y. Formally,
Yt = λ t X t + θ t B t ,

where Bt is the value of the bonds, and which accrues value by the factor erdt. Note that λ and θ
are stochastic processes. They are indexed by t and are determined as the stock evolves along its

2
Of course the exercise price also determines the option payoff, but it is not random so it causes us no problems.
D.M. Chance, TN00-04 3 Girsanov’s Theorem in Derivative Pricing
stochastic paths. But they need be known only at t.3 They change as we move through time but
are completely determined once we know the new value of the stock price. Financial economists
know this as nothing more than the option’s delta changing. Indeed delta is what we are
currently designating as λ.
Mathematical finance formally tells us that if the stock can be transformed into a
martingale, we can find a stochastic process such that the stock can be transformed into another
martingale, which replicates the option. Once we have replicated the option, we can price it
using the stock price, its number of holdings, the bondholdings, the interest rate, etc., or in other
words, known values. We can obtain this by finding the expectation of the future value of this
process.
Before we proceed with the mathematics, let us return to one final point. As noted above,
mathematicians seem to suggest that we can relate one martingale to the other without
considering that these two martingales might be totally unrelated. We cannot replicate the price
of corn in Iowa with the price of a gas company in Warsaw. The problem lies in the advantage
that financial economists have over financial mathematicians. Financial economists study
general equilibrium., which is the study of how markets reach equilibrium where outcomes are
determined by the interaction of preferences and expectations to form demand, supply and,
ultimately, prices and holdings of assets. General equilibrium models like the Capital Asset
Pricing Model, as flawed as they may be, nonetheless, bring us an appreciation for the fact that
markets are complex combinations of individuals and assets. Assets do not exist in isolation and,
as such, stochastic processes of single assets, which attempt to model outcomes, make us forget
everything else that is going on in the market at that time. Derivative pricing theory is
considered partial equilibrium. The general equilibrium process is taken as determined
externally. Without recognizing the general equilibrium process, mathematicians might have us
believe that the price of corn in Iowa could be hedged with the price of a gas company in
Warsaw.
Now the one case where the mathematicians would be correct is when there is a single
source of uncertainty driving the two variables. That is why, as we shall see, that one random
variable can be expressed in terms of another. In financial applications, one random variable is

3
Mathematicians call this property previsibility.
D.M. Chance, TN00-04 4 Girsanov’s Theorem in Derivative Pricing
often a derivative of the other. But it is misleading to state the martingale representation theorem
as generally as most mathematicians do. They are often thinking of a world in which an option
and its underlying, perhaps with a risk-free bond, are the only assets that trade.
Now let us proceed to learn the mathematics of how derivative prices are found. Our first
step is to learn how to change the drift, i.e., expected return of a random variable.
Introducing the Radon-Nikodym Derivative by Changing the Drift for a Single Random
Variable
Let us first begin by examining the process of changing a probability distribution for a
general random variable. We are given a random variable x, which is simply a single unknown
outcome and not a stochastic process. We shall take x as distributed normally with mean µ and
variance σ2. The probability density of x is
2

dP(x )
 x −µ 
1 −  /2
f (x ) = e  σ 
= ,
σ 2π dx
where the probability distribution function is P(x). Now suppose that we wanted to change the
location of this probability distribution. Specifically, we wish to shift its mean by an amount γ.
Then the density we want is
2
 x −µ − γ 
1 −  /2
f (x ) = e σ 
.
σ 2π
In other words, we change the mean but not the variance. We need not specify if ( is positive or
negative; we could shift the mean upward or downward. Let us call this new density fQ(x) and
the new distribution function Q(x), so fQ(x) = dQ(x)/dx. Let us see how we can make this
change whereby P(x) becomes Q(x).
Look at the expression in the exponent above, -((x - µ - γ)/σ)2/2 and note that it equals
-(1/2σ2)(x2 - 2xµ - 2γx + 2µγ + µ2 + γ2). If we compare this to what we started with, -((x -
µ)/σ)2/2 = -(1/2σ2)(x2 - 2xµ + µ2), we see that we need to multiply the original expression by
something. That is,
2
 x −µ 

e ( )
1 −
σ 
 /2
− − 2 γx + 2 γµ + γ 2 / 2 σ 2
e 
σ 2π
2
 x −µ − γ 
1 −
σ 
 /2
= e  .
σ 2π
D.M. Chance, TN00-04 5 Girsanov’s Theorem in Derivative Pricing
Let us now designate this new multiplier as

φ(x )e = (2 γx −2 γµ− γ )/ 2 σ .
2 2

Since f(x) = dP(x)/dx, then f(x)dx = dP(x). What we want is fQ(x)dx = dQ(x), where we have a
new probability measure Q(x). Since f(x)N(x) = fQ(x), then f(x)dxN(x) = fQ(x)dx. Then
f Q (x )
= φ(x ) .
f (x )
It follows that
dQ(x )
= φ(x ) .
dP(x )
Our multiplier N(x) can be thought of as an adjustment that converts one probability measure,
P(x), into another Q(x). We must be careful, however, in that one cannot just arbitrarily multiply
one measure by some other factor, for the resulting measure should be of the same type, here the
normal distribution, as the one we started with, though now having a different mean.
In some cases our random variable will be standard normal, meaning that µ = 0 and σ =
1. In that case,

φ(x ) = e xγ − γ
2
2
.
Note that in any case if γ > 0, the mean is shifted upward and if γ < 0 the mean is shifted
downward.
This special function N(x), which we noted can be expressed as dQ(x)/dP(x), is a
derivative itself and is called the Radon-Nikodym derivative. For this derivative to exist, it is
necessary that the function Q(x) and the function P(x) be considered equivalent probability
measures. What this means is that if an event is possible under one measure, then it is possible
under the other measure. In other words, events that cannot occur in the first place, cannot be
made possible by simply changing the probability measure. Likewise events that can occur in
the first place, cannot be made impossible by changing the probability measure.
Changing the Drift for a Continuous Time Stochastic Process
For our applications in finance, the random variable we deal with is often a stochastic
process. In many cases, the random variable will be a Brownian motion, Wt, such that
t
Wt = ∫ dWu ,
0

D.M. Chance, TN00-04 6 Girsanov’s Theorem in Derivative Pricing

where we know that W0 = 0 and the increments are distributed with mean zero and variance dt.4
The density of Wt is5
2
1  wt 
1 −  
f (Wt ) = e 2 t 
.
2 πt
We wish to change this Brownian motion into another one that has a new probability
measure Q. We shall shift it by an amount γt. In later applications we shall see that γt will
become an extremely simple function of t, but for now let us leave it unspecified.
So what we want is a new Brownian motion that has a mean of γt. What will accomplish
this trick is to designate:
t t
1
∫ γ udWr − 2 ∫ γ udu
2

φt = e 0 0
.

In order for this transformation to be possible, we must impose a constraint on the behavior of γu.
Specifically, we require that
 ∫ γ 2udu 
t

E e 0  < ∞ .
 
 
This is called the Novikov condition. In simple terms it means that the variation in (u must be
finite. For all of our applications, the Novikov condition will be met.6
If these requirements are met, then Nt can be shown to be a martingale. Let us first apply
Itô’s Lemma on φt:
∂φ t 1 ∂ 2 φt
dφ t = dWt + dWt2 .
∂Wt 2 ∂Wt 2

First we find the partials:

4
Recall that specifically, dWt = ε t dt where εt is a standard normal random variable.
5
The following specification is because Wt has a mean of zero, since it starts off at W0 = 0, and a variance of t. The
term is parentheses is the value of the random variable minus its mean divided by its variance.
6
As we shall ultimately see, ( will be a very simple function that in any rational financial market will have finite
variation.
D.M. Chance, TN00-04 7 Girsanov’s Theorem in Derivative Pricing
 ∂  0∫ γ udWr − 2 0∫ γ udu 
1 2 t t
∂φ t
= e 
∂Wt ∂Wt  
∂ t 1t 2 
= φt ∫
 0 γ u dWr − ∫ γ u du 
∂Wt 20 
= φt γ t

∂ 2 φt
= 0.
∂Wt2
So7
t t
1
∫ γ udWr − 2 ∫ γ udu
2

dφt = φ t γ t dWt = e 0 0
γ t dWt .
Now let us consider the value at t = 0:
0 0
1
∫ γ udWr − 2 ∫ γ udu
2

φ0 = e 0 0
γ t dWt = e 0 = 1 .
Now we have
t t

∫ dφ
0
µ = ∫ φ u γ u dWu = φ t − φ0 , so
0

t t
φ t − 1 = ∫ φµ γ u dWu , and φ t = 1 + ∫ φu γ u dWu .
0 0

The second term on the right-hand side is a known martingale. Its expectation is zero so E(φt) =
1 = φ0. Consequently, φt is a martingale.
Thus, we can now be certain that Girsanov’s Theorem applies. Our Brownian motion can
be transformed as follows:
t
WtQ = Wt − ∫ γ u du ,
0

where WtQ is a Weiner process under a new probability measure Q such that
dQ
= φ.
dP
Remember that N is the Radon-Nikodym derivative.

7
The result on the third line above is obtained from the result on the second line above by differentiating the stochastic
integral at the end point.
D.M. Chance, TN00-04 8 Girsanov’s Theorem in Derivative Pricing
 We shall ultimately need the Weiner differential, dWtQ. It is obtained as follows:
t
WtQ = Wt − ∫ γ u du,
0
t
W − Wt = − ∫ γ u du implies
t
Q

dW − dWt = − γ t dt, so
t
Q

dWtQ = dWt − γ t dt.

Now let us step back and think about how this is important for our purposes. We shall
want to convert our stock price process to a martingale. This will remove the drift and permit us
to price the option by evaluating its expected future value under the new probability measure.
When we remove the drift, what we are doing is removing the risk premium and the risk-free
rate. If we knew the value of the risk premium, it would be no problem: we would simply
subtract it out. But we do not know what the risk premium is. We do not know how much of the
stock’s expected return to remove. We do know, however, that if we remove just enough that the
stock return is a martingale, then we require no discounting whatsoever. So the trick is to change
the probability distribution so that the stock return is a martingale. Here is where the finance
ends and the math takes over. Girsanov’s Theorem tells us how to change a probability
distribution to leave it the same type of distribution with the same variance but with a different
drift. What we have just seen above is that the Brownian motion process, Wt, can be changed
such that it is a martingale. Since the stock price process is a simple transformation of the
Brownian motion process, it should be easy to transform it as well into a martingale.
We have seen above that we are subtracting a function γt. This means that γ can
potentially change with t. We are somewhat lucky here, because for our purposes γt is a very
simple function of t: γt = γt.8 Now notice what we obtain for our Radon-Nikodym derivative:
t t
1 t t
∫ γdWr − 2 ∫ γ du
2 1
r ∫ dWr − γ 2 ∫ du
2 0 2
φt = e =e = e γWt − γ t 2 .
0
0 0

In other words, if we multiply the density function of Wt by Nt as specified above, we should

obtain the density function for a new Brownian motion, which we shall call WtQ, in which the
mean has been shifted by γt. Let us see. Given,

8
If this were not the case, we would see that our method would not work later on.
D.M. Chance, TN00-04 9 Girsanov’s Theorem in Derivative Pricing
 e − (Wt ) , and
1
f (Wt ) =
2

2 πt
dQ = e γWt − γ
2
t2
,
dP
we obtain by multiplication:

e −(Wt / 2 t + γWt − γ t 2 )
dQ 1
f (Wt )
2 2
=
dP 2 πt
2
 Wt − γ t 
1 − (1 2 ) 
= e  t 
.
2 πt
This is the density for a Brownian motion with its zero mean shifted by -γt. So
WtQ = Wt − γt
To recap, we have that Wt is a Brownian motion under the probability measure P, such
that
E P (Wt ) = W0 = 0
E P (WtQ ) = E P (Wt − γt ) = E P (Wt ) − γt
= 0 − γt.
The first statement defines that Wt is a Brownian motion under P. The second statement says
that under P, WtQ is not a Brownian motion. Its expectation, -γt, is not zero, except at t = 0, and
varies with t. But WtQ is a Brownian motion under Q:
E Q (WtQ ) = W0Q = 0 .
This statement follows since Wt and WtQ both start at a value of zero. Under Q, Wt, is not a
Brownian motion because:
E Q (Wt ) = E Q (WtQ + γt ) = E Q (WtQ ) + γt = γt .
When we say that some random process, such as Wt and WtQ, is a Brownian motion
under a given measure, we are saying that the probabilities of its possible paths are assigned such
that its central property, a constant expectation of zero is preserved. When the probabilities are
changed such that the process no longer has a zero expected value, it is no longer the same thing.
But another process, can and in this case, does have the property of a Brownian motion under the
new probability measure.
Changing the Drift of a Stock Price Process

D.M. Chance, TN00-04 10 Girsanov’s Theorem in Derivative Pricing

 In TN96-04 and TN00-03 we obtained the familiar stochastic process for a stock:
dS t S t = αdt + σdWt .

If we change Wt such that now Wt = WtQ + γt, then we substitute its differential, dWt = dWtQ +
γdt, into the above stochastic differential equation to obtain
dSQt SQt = αdt + σ(dWtQ + γdt )
= (α + σγ )dt + σdWtQ .
So if we change the probability measure for Wt, which is the probability that drives St, we are
now working with the above stochastic process and a new set of probabilities. But have we
converted StQ to a martingale? Not yet. If, however, we specify that γ = - α/σ, the drift becomes
zero, leaving us with
dSQt SQt = σdWtQ ,

which is clearly a martingale. From this result, we can assign an obvious interpretation to γ, an
interpretation we have already arrived at. First, ignoring the minus sign, the expression α in the
numerator is the expected return. The denominator is clearly the risk. Thus, γ, is the return over
the risk, a kind of risk-return tradeoff. It is somewhat more natural, however, to specify γ in a
slightly different manner:
α−r
γ=− ,
σ
which, when substituted back into the stochastic differential equation, gives us

dSQt SQt = rdt + σdWtQ ,

Now we have a more natural interpretation of γ. Again, ignoring the minus sign, the numerator
is the expected return minus the risk-free rate, or the risk premium. The denominator is the risk.9
In financial economics, this ratio is the risk premium per unit of risk and is sometimes called the
market price of risk. It reflects the risk-return tradeoff, i.e., the additional expected return
necessary to induce investors to assume risk.10

9
It is even more natural that instead of defining γ as - (µ - r)/σ, we define it as (µ - r)/σ and then change the sign such
that instead of adding γ to α, we are subtracting it. In that way, we appear to be subtracting a risk premium, a more sensible way
to describe what is happening.
10
In the study of general market equilibrium, such as the Capital Asset Pricing Model, the appropriate measure of risk
is not the standard deviation, but the systematic risk, also known as β. The derivatives pricing framework, however, takes the
D.M. Chance, TN00-04 11 Girsanov’s Theorem in Derivative Pricing
 Now the adjustment is, more or less, just a subtraction of the risk premium (per unit of
risk). But the beauty of it all is that we never had to obtain the risk premium. By converting the
process to a martingale, we removed the risk premium automatically.
For the log process, recall that its drift, µ, is equal to α - σ2/2, so then11
µ + σ2 2 − r
γ=− .
σ
But if you have been paying attention, you should note that it appears we no longer have
St in the form of a martingale. After all, its new expected return is
dSQt S t = rdt ,
which is certainly non-zero. Clearly the stock now drifts upward at the risk-free rate. Now we
seem to have a problem, but a slightly different spin on things saves the day.
First we should be comforted in knowing that by removing the risk premium, we have
taken out the most difficult part of the problem. We ought to be able to solve the option pricing
problem from what we now know. Indeed that is the case. Financial economists have long
known that if we change the stock’s expected return to the risk-free rate, we can then evaluate
the expected option payoff under the assumption that the stock price is randomly generated by
the standard stochastic differential equation with a drift set at the risk-free rate. Economists then
go on to explain that everything we need to know about how investors feel about risk is
impounded into the stock price. It is not necessary to reflect any effect of investors’ risk
preferences in the option price. Consequently, we can proceed to evaluate the option as if the
expected return on the stock is the risk-free rate. This approach is often called risk neutral
pricing. What we have done is equivalent to the well-known procedure of taking a short position
in an option, hedging it with a long position in units of the underlying asset, thereby eliminating
the risk, followed by setting the return on this hedged portfolio to the risk-free rate. From there
we obtain a partial differential equation whose solution is the option pricing model.
But how do we salvage our approach, which now leaves us with St no longer a martingale?
The trick lies in recognizing that we can work with the discounted value of St. In other
words, say we start off at time 0 with a value of S0. Then at time t, we have St. But suppose we

general equilibrium as given and as fully reflected in the price of the underlying asset. Risk is measured in isolation and σ is the
only measure of risk recognized within option pricing theory.
11
Let us emphasize that this is not a new definition for γ. It simply expresses γ in terms of the log return.
D.M. Chance, TN00-04 12 Girsanov’s Theorem in Derivative Pricing
transform our stock price into its discounted value, Ste-rt. Now let us look at some of our
previous results. Recall from TN00-03 that the solution to the stochastic differential equation,
giving us St in terms of S0 is
S t = S0 e µt +σWt .
Suppose that we adjust St to its discounted value, Ste-rt. Then substituting into the above
equation:
S t e − rt = S0 e (µ−r )t +σWt .

Right now, however, we are under the original probability measure. Substituting WtQ + γt for
Wt, we obtain

S t e − rt = S0 e (µ −r )t +σ (Wt ).
Q
+ γt

Noting that we defined γ as - (µ + σ2/2 - r)/σ and substituting this result, we obtain
2 Q
S t e − rt = S0 e −σ t 2+σWt .
You may wish to look back in TN00-03 where we used this result for St, along with the
density function for a normally distributed Wt to obtain the expected future stock price:

E[S t ] = S0 e (µ+ σ 2 )t .
2

If we follow that same procedure here for Ste-rt, we obtain

[ ]
E S t e − rt = S0 e − σ
2
t 2 σ2 t 2
e = S0 .
The absence of a positive expected return shows that the discounted price is a martingale.
Let us do one more thing. We shall apply Itô’s Lemma to the discounted stock price.
First for simplicity of notation, let us designate St* as Ste-rt. Now applying Itô’s Lemma to St*,
we obtain:
∂S*t ∂S* 1 ∂ 2S*t
dS*t = dWt + t dt + dWt2 .
∂Wt ∂t 2 ∂Wt2
Using the above result,
S t e − rt = S0 e (µ−r )t +σWt ,
we can obtain the partial derivatives:
∂S*t ∂ 2S*t ∂S*t
= S*tσ, = S*tσ 2 , = S*t (µ − r ) .
∂Wt ∂Wt 2
∂t

D.M. Chance, TN00-04 13 Girsanov’s Theorem in Derivative Pricing

Substituting and noting that dWt2 = dt, we obtain the following stochastic differential equation
for St*:

= (µ − r + σ 2 2 )dt + σdWt
dS*t
*
St
= (α + r )dt + σdWt .

We see that once we have taken the risk-free rate into effect in defining the underlying variable,
we no longer can account for the risk-free rate in the drift. But still we do not have a martingale.
Recall that to obtain a martingale, we substitute dWtQ + γt where γ = - (α - r)/σ into the above,
giving us
dS*t
*
= σdWtQ ,
St
which is clearly a martingale.
To summarize, we adjust the drift of the stock price process by changing the probabilities
such that we obtain a martingale. What we ultimately have done is gone ahead and discounted
the stock price and then worked with the discounted stock price to change its probability
measure, leaving us with a martingale. Then we can easily evaluate the option by applying the
probability distribution of the discounted stock price to the option payoff. In this way the option
price is its expected payoff at expiration without discounting.
If we do the one, discounting the stock price, without the other, changing the measure, we
have technically not completed the process. But as it turns out, we can get away with changing
the measure without discounting the stock price. Recall that dSt/St = αdt + σdWt. Substitute
dWtQ + γdt and γ = - (α - r)/σ and we obtain:
dS t
= rdt + σdWtQ ,
St
We can work with this model and do the discounting later, i.e., after we have evaluated the
expected option payoff at expiration. That is because whether we discount before we have
performed the expectation or after, we have not altered the fundamental process or results of
taking expectations other than by the simple linear adjustment, e-rt. While mathematicians would

D.M. Chance, TN00-04 14 Girsanov’s Theorem in Derivative Pricing

prefer that we convert the stock price to a martingale, requiring that we do the discounting
beforehand, financial economists would prefer to do the discounting afterwards. That is because
the latter approach is more in line with the intuition provided by economic theory: the price of
any asset is its expected future value, discounted to the present at an appropriate rate. That in
this case the appropriate rate is the risk-free rate is quite intuitive. The risk has been removed via
the risk-free hedge, or alternatively, can be viewed as fully imbedded into the price of the
underlying asset and, therefore, cannot legitimately be incorporated again. Moreover, if the risk
is either not present or removed, investors’ risk preferences are irrelevant to the valuation
process. In that case, one might just as well use the simplest form of risk preferences: risk
neutrality, wherein investors discount future values at the risk-free rate.
Finally, we might just simply say that if the price of the underlying asst is given, any two
investors regardless of their feelings about risk will value the option in the same manner.
Consequently, we can treat both investors as though they had the simplest risk preferences of all,
risk neutrality.

References

Baxter, M. and A. Rennie. Financial Calculus. Cambridge: Cambridge University Press

(1996), Ch. 3.

Dothan, M. U. Prices in Financial Markets. New York: Oxford University Press (1990),
Chs. 5, 6.

Duffie, D. Dynamic Asset Pricing Theory, 2nd. ed. Princeton: Princeton University
Press (1996), Ch. 6, Appendix D.

Duffie, D. Security Markets: Stochastic Models. Boston: Academic Press (1988), Chs.
21.

Epps, T. W. Pricing Derivative Securities. Singapore: World Scientific (2000), Ch. 3.

Karatzas, I. and S. E. Shreve. Brownian Motion and Stochastic Calculus, 2nd. ed. Berlin:
Springer-Verlag (1991), Chs. 3, 5.

Musiela, M. and M. Rutkowski. Martingale Methods in Financial Modelling. Berlin:

Springer-Verlag (1997), Appendix B.

Neftci, S. An Introduction to the Mathematics of Financial Derivatives, 2nd. ed. San

Diego: Academic Press (2000), Chs. 14, 15.
D.M. Chance, TN00-04 15 Girsanov’s Theorem in Derivative Pricing
 Nielsen, L. T. Pricing and Hedging Derivative Securities. Oxford, U.K.: Oxford
University Press (1999), Ch. 2.

D.M. Chance, TN00-04 16 Girsanov’s Theorem in Derivative Pricing