Sei sulla pagina 1di 21

3 Random walk model of asset prices

The term asset is used to collectively refer to stocks, commodities (such as gold, oil,
wheat, etc.), options, currency, etc. The goal of this course is not to explicitly predict
how asset prices will vary as a function of time. If this were possible, the wise investor
would keep it a secret, and happily make as much money as they could regardless of the
overall market performance. In reality, it is not possible to predict with certainty, how
asset prices will vary as a function of time. Figure 4 shows the closing value of the FT-SE
100 stock index on trading days between 1 April 2008 and 21 October 2008. Over this
time period the general market trend is downwards, but there is considerable variability:
on some days the market goes up on some days it goes down. The overall trend in the
data in Figure 4 is for the stock index to go downwards, but on any given day it seems
impossible to predict whether the index is going to go up or go down. In essence there are
random variations which combine with the overall market trend to determine the index
as a function of time. One data point is plotted for each trading day in Figure 4. Similar
fluctuations would also be seen if we plotted the index value with a much shorter time
interval (for example once every minute over the course of a trading day). These random
fluctuations appear to be an inherent characteristic of asset values. In section 1.2 it was
argued that the value of an asset is determined by the market forces of supply and demand
and that the ultimate value is the price agreed by the buyer and the seller. The random
fluctuations are the results of different buyer/seller pairs having slight differences in their
view of the the asset value.
The model we will use in this course is given by the following stochastic differential
equation
dS = S (σdX + µdt) . (7)
Here dS is the change in the asset price (i.e. share price) in a small time step dt. Also, S
is the current asset price, µ is a measure of the average rate of growth of the asset price
(known as the drift), σ is called the volatility of the share (which measures the standard
deviation of any random variation in the share price), and dX is a random number taken
from a normal distribution. The mean of the normal distribution that dX is taken from
is equal to zero and the variance of the normal distribution is √equal to dt. Since the
variance is equal
√ to dt, then the standard deviation is equal to dt. Therefore, we can
write dX = φ dt where φ is a random number taken from a normal distribution with
mean zero and standard deviation equal to one.
Therefore, the above model describes how the asset price S, changes from S to S + dS,
during a time step from t to t + dt. We will ultimately examine this model in the limit
dt → 0.
Note that if σ = 0 then the above equation simplifies to
dS = Sµdt. (8)
Therefore, using separation of variables,
dS
Z Z
= µdt.
S

18
7000

6000
FT−SE 100 Index

5000

4000

3000
0 50 100 150
Time (days since 1 April 2008)

Figure 4: Closing values of the FT-SE 100 index call between 1 April 2008 and 21 October 2008

Therefore,

ln S = µt + c,

where c is a constant of integration. Therefore,

S = So eµ(t−to ) ,

where the initial condition S = So at time t = to is used. This equation is formally


equivalent to the case of money earning continuously compounded interest in a bank.
For the case of bank interest, however, the interest rate r is always positive. The drift
coefficient µ in equation (8) can be either positive, negative, or even zero. Therefore, we
see that µ measures the non-random increase in the share price if µ > 0 (or non-random
decrease in the share price if µ < 0). Of course, the case with σ = 0 is not realistic, since
the share price S varies randomly.
Example: Suppose S = 10, t = 1, σ = 0.1, µ = 0.01 and dt = 0.01. Then dX
must be a random number selected from a normal distribution of mean 0 and standard

19
deviation equal to 0.1. Of course dX is random and cannot be predicted, but suppose it
was equal to 0.03. (Note that dX could be anything since it is random; this is just an
example of a number it could be!) Then dS = S (σdX + µdt), which in this case gives
dS = 0.031. So in this time step, t increases from 1 to 1 + dt = 1.01, and S increases from
10 to 10 + dS = 10.031.
For the next time step, we now have S = 10.031, t = 1.01, σ = 0.1, µ = 0.01 and dt =
0.01. Again dX is a random number selected from a normal distribution of mean 0 and
standard deviation equal to 0.1. Once again dX cannot be predicted because it is random.
But suppose the next value for dX was equal to −0.02. Then dS = S (σdX + µdt),
which in this case gives dS = −0.0190589. So in this time step, t increases from 1.01 to
1 + dt = 1.02, and S changes from 10.031 to 10.031 + dS = 10.0119411. etc.
The repeated iteration of equation (7) to calculate the value S(tnew ) = S(told ) + dS
where tnew = told +dt is known as a random walk. Since dX is a random variable, repeated
independent walks starting from the same initial position will likely give different predicted
values of S at any later time. Thus, equation (7) cannot be used to make predictions
about future asset values. Nevertheless by performing many repeated random walks, it
is possible to make some qualitative statements. Consider the following simple MatLAB
program:

s(1) = 100
mu = 0.1
sigma = 0.01
dt = 1/365
for i = 1:365
dX = randn*sqrt(t);
dS = s(i)*(mu*dt + sigma*dX);
s(i+1) = s(i)+dS;
end
plot(s)

In the above we are assuming that time is measured in years and dt = 1/365 represents
one day. The value of µ and σ√are specified, as is the value of S on day 1. We have
explicitly assumed that dX = φ t where φ is a normal random variable with mean zero
and standard deviation one. In MatLAB we use the in-built function ‘randn’ to calculate
the required values of φ at each iteration. After 365 days (one year) we then plot S as a
function of time. Using this program we can then easily perform repeated random walks
to determine the influence of varying the various parameters.
Figures 5, 6 and 7 show repeated random walks for different values of σ. In each case
we start with S = 100 and we perform 365 time steps with dt = 1/365. The drift rate is
held fixed at µ = 0.1. In Figure 5 we have set σ = 0.01, in Figure 6 we have set σ = 0.1
and in Figure 7 we have set σ = 0.3. The solid lines in these figures represent the repeated
random walk. The dotted line (which is the same in all figures) represents the exponential
growth that would be expected if σ = 0. Observe in Figure 5 that when σ is relatively
small (in this case σ = 0.01) the different realizations of the random walk all lie close to

20
150.0

125.0

100.0
S

75.0

50.0
0.0 100.0 200.0 300.0 400.0
Time (days)

Figure 5: Random walks using equation (10) with σ = 0.01.

exponential growth curve. As σ is increased to 0.1 in Figure 6 and to 0.3 in Figure 7 we


see that the different realizations of the random walk start to strongly diverge from each
other. In some cases the random walk greatly exceeds the expected exponential growth
due to drift along, and in some cases the random walk falls well below. The volatility σ of
an aspect is a direct measure of risk. The larger the volatility, the greater the uncertainty
of the investment. From these figures it is clear that high risk investments can lead to
much greater earnings than low risk investments, but they can also lead to much greater
losses. These figures certainly indicate that when the volatility σ is ‘large’, it is impossible
to make predictions about asset values in the future.
This module assumes that asset values can be modelled using the stochastic differential
equation (7). The results of figures 5-7 clearly show the effects of randomness, thus
indicating that this equation cannot be used to make predictions about future asset values.
Why then is this equation so important? First, in a statistical sense, it has been shown
that equation (7) generally does a very good job of modelling the historical data of most
asset prices (better agreement is achieved for stocks, options and commodities, than for
currencies). Second, and more significantly, we can use equation (7) as the foundation for

21
150.0

125.0

100.0
S

75.0

50.0
0.0 100.0 200.0 300.0 400.0
Time (days)

Figure 6: Random walks using equation (10) with σ = 0.1.

more sophisticated models, and these models can be used to completely eliminate risk.
The formation of these models is the primary goal of this module. Before, we proceed to
examining these more sophisticated models, we first note that equation (7) requires the
specification of two fundamental parameters µ and σ. These parameters will depend on
the particular asset which is being modelled, and they can be estimated by looking at the
historic data of the asset prices. In particular, suppose that we know the value of S at
n + 1 discrete, and equally spaced, points in time, which we denote by So , S1 , S2 , . . . , Sn
with the subscripts denoting sequential points in time. Assuming that dX is normally
distributed then we estimate
n−1
1 X Si+1 − Si
µ ≈ m̄ =
ndt i=0 Si

and
n−1  2
2 2 1 X Si+1 − Si
σ ≈ σ̄ = − m̄dt (9)
(n − 1)dt i=0 Si

22
150.0

125.0

100.0
S

75.0

50.0
0.0 100.0 200.0 300.0 400.0
Time (days)

Figure 7: Random walks using equation (10) with σ = 0.3.

The accuracy of these estimates will depend on the number of points in the time series,
as well as the size of the time step dt. It has been assumed that µ and σ in equation (7)
are constant, but, in reality these parameters will be slowly varying functions of time.

23
3.1 The size of random fluctuations
The starting point for our modelling of asset prices is the stochastic differential equation

dS = S (σdX + µdt) . (10)

The drift term µdt corresponds to the deterministic rise or fall in the value of S with
time in the absence of any random fluctuations. The σdX term accounts for the influence
of random fluctuations. The relevant question is ‘Which is more important?’: the deter-
ministic drift of S with respect to time, or the influence of random fluctuations. This
question is relatively straightforward to answer. If µdt >> σdX then deterministic forces
will dominate. If σdX >> µdt then random fluctuations will dominate. If the terms
µdt and σdX are approximately of the same magnitude, then, the value of the asset S
will depend on both deterministic and random forces. Historical evidence suggests that
stock prices are indeed determined by a combination of deterministic and random forces
and thus we would like to have an appropriate balance in the sizes of the µdt term and
the σdX term. Moreover, we would like the parameter µ to uniquely characterise the
deterministic drift (growth or fall) of S and σ to uniquely characterise the influence of
random fluctuations. Ultimately this means that we will need to comment on the relative
size of dt versus dX in order to ensure that equation (10) is a realistic model of the time
evolution of asset prices.
In the previous section we made the statement that dX can be modelled as a normally √
distributed random variable with zero mean and variance equal to dt (i.e. dX = φ dt
where φ is a random variable from a standardised normal distribution with mean 0 and
variance 1). This is indeed the appropriate choice of scaling for dX as a function of dt,
but it is instructive to try to convince ourselves, why this scaling is appropriate. In this
section we will use computer simulations to justify this scaling. In the following section
we will use more rigorous mathematical methods to justify this √ scaling. The key point
is that if we can successfully argue that dX is proportional to dt, then we can convert
the stochastic differential equation (10) for S into a fully deterministic equation (i.e. the
Black Scholes equation).
Our objective is to use equation (10) to perform a large sequence of random walks
and then to compare these random walks to each other in order to see if any meaningful
information can be obtained. In particular, we shall keep µ and σ fixed and iterate
equation (10) from some initial time t = t0 until some final time t = t1 by performing
n iterations with fixed dt = (t1 − t0 )/n. Because equation (10) can be seen as a leading
order Taylor series expansion for S as a function of time, we expect that our results will
become most meaningful in the limit as dt → 0 (or n → ∞). In equation (10) we will set

dX = φdtk (11)

where φ is a random variable from a standardised normal distribution with mean 0 and
variance 1, and we will then examine the influence of varying k for values of k > 0.
The following MatLAB program can be easily modified to perform our comparison:

24
s(1) = 100
mu = 0.1
sigma = 0.3
n = 100000
dt = 1/n
k = 0.5
for j = 1:500 % perform 500 random walks

for i = 1:n
dX = randn*dt^k;
dS = s(i)*(mu*dt + sigma*dX);
s(i+1) = s(i)+dS;
end

ss(j) = s(n+1) % save the final value of each random walk

end

plot(ss)

The above program was used to integrate from t = 0 to t = 1 using a time-step of


dt = 10−5 with µ = 0.1 and σ = 0.3 fixed for different values of k. Figures 8, 9 and 10
show results for k = 1, k = 1/2 and k = 1/4 respectively. In each case, 500 separate
random walks were performed assuming an initial value of S(0) = 100, and the value of
S at time t = 1 is then plotted. Since we are interested in the limit where dt << 1 it
follows that dt1 << dt1/2 << dt1/4 . Consequently, it would be expected that the results
for k = 1 will be least influenced by the effects of random fluctuations, while the results
for k = 1/4 will be most strongly influenced. This is indeed the case. Figure 8 shows that
when k = 1, the share price at time t = 1 is virtually deterministic and independent of
random fluctuations. We consider this an unsatisfactory outcome since we would expect
random fluctuations to play an important role in determining the value of S. For the case
of k = 1/4 in Figure 10, we also obtain unsatisfactory results with the value of S varying
from approximately 0, to essentially infinity. Thus, for k = 1/4, we cannot say anything
meaningful about the predicted value of S at time t = 1. The case of k = 1/2 shown
in Figure 9, is however, satisfactory. The results certainly show the influence of random
fluctuations, and while each individual random walk is just as likely as any other, we can
with a certain level of confidence make meaningful predictions about the value of S at
time t = 1. For example we can state that there is a 60% (estimated) chance that the
value of S at time t = 1 will lie in the interval 100 < S < 150. Further random walks can
be performed for different values of k, and in the limit as dt → 0 it would be observed
that all results for k > 1/2 are similar (and thus unsatisfactory) to those in Figure 8
and results for k < 1/2 are similar (and thus unsatisfactory) to those in Figure 10. The
distinguished limit of k = 1/2 in equation (11) is the only admissible value in order for

25
112.0

111.0
S

110.0

109.0
0.0 100.0 200.0 300.0 400.0 500.0
Walk number

Figure 8: A sequence of random walks using equation (10) with dX = φdt.

equation (10) to meaningfully model the variation of stock prices.

3.2 Black-Scholes equation: derivation (part 1)


The rest of this module is concerned with how to work out how much to pay for an option,
or how much to sell it for if it is to be traded prior to the expiry date. We shall derive the
Black-Scholes equation, which will be the basis of all such calculations. We will derive
the fair price for an option, though a trader will always try to do better than that! This
equation gives the basis for all trading in options and is used extensively in all Investment
Banks.

In the derivation of the Black-Scholes equation we will consider holding a portfolio


made up of one option (it can be a call or a put, or in fact any type of option) which has
value V , plus δ lots of the underlying share S. Thus the portfolio that we hold has total
value α where

α = V + δS,

26
300.0

200.0
S

100.0

0.0
0.0 100.0 200.0 300.0 400.0 500.0
Walk number

Figure 9: A sequence of random walks using equation (10) with dX = φdt1/2 .

i.e. α equals the total value of the portfolio, V equals the value of the option, S equals
the value of the underlying share, and δ is the number of shares in the portfolio. e.g.
a portfolio like this might consist of one European Call Option where the underlying is
Rolls Royce plus δ lots of Rolls Royce shares where δ is a constant.
The value of the option V will vary with time t and also with the underlying share
price S. Thus V = V (S, t). The equation that we will derive below will be valid for any
function V which is a function of S where S is determined by the lognormal random walk
given by equation (10) and time t. Strictly speaking V does not have to represent the
value of an option, but this is a helpful starting point.
The following derivation is mathematically informal. See the recommended textbooks
for notes on how to make the following mathematical argument more rigorous.
We consider a small time step dt over which the time varies from t to t + dt, and the
underlying share price varies from S to S + dS. During this small time interval, the value
of the option will vary from V to V + dV , where

dV = V (S + dS, t + dt) − V (S, t).

27
20000.0

15000.0

10000.0
S

5000.0

0.0
0.0 100.0 200.0 300.0 400.0 500.0
Walk number

Figure 10: A sequence of random walks using equation (10) with dX = φdt1/4 .

We take δ to be constant over the time step t to t + dt, but δ may vary from one time
step to another.
We can use Taylor’s theorem (it is called Ito’s Lemma in the case of using an expansion
with a random variable) to find

∂V ∂V 1 ∂2V ∂2V 1 ∂2V


V (S + dS, t + dt) = V (S, t) + dS + dt + dS 2 2 + dSdt + dt2 2 + ...
∂S ∂t 2 ∂S ∂S∂t 2 ∂t
where the derivatives are all evaluated at (S, t). This is just the standard formula for
Taylor’s theorem in two variables. Rearranging slightly gives

∂V ∂V 1 ∂2V ∂2V 1 ∂2V


V (S + dS, t + dt) − V (S, t) = dS + dt + dS 2 2 + dSdt + dt2 2 + ...
∂S ∂t 2 ∂S ∂S∂t 2 ∂t
But dV = V (S + dS, t + dt) − V (S, t), therefore

∂V ∂V 1 ∂2V ∂2V 1 ∂2V


dV = dS + dt + dS 2 2 + dSdt + dt2 2 + ...
∂S ∂t 2 ∂S ∂S∂t 2 ∂t

28
But dS = S (σdX + µdt). Substituting this into the above equation gives

∂V ∂V 1 ∂2V
dV = S (σdX + µdt) + dt + (S (σdX + µdt))2
∂S ∂t 2 ∂S 2

∂2V 1 2 ∂2V
+S (σdX + µdt) dt + dt + ...
∂S∂t 2 ∂t2
Rearranging gives
∂2V
 
∂V ∂V ∂V 1
dV = SσdX + dt Sµ + + (dX)2 S 2 σ 2 2
∂S ∂S ∂t 2 ∂S

∂2V ∂2V 2
∂2V ∂2V
   
2 1 2 2 2∂ V
+dtdX S σµ 2 + Sσ + dt S µ + 2Sµ + 2 + ...
∂S ∂S∂t 2 ∂S 2 ∂S∂t ∂t

But from the previous section, we know we can write dX = φ dt where φ is a random
number taken from a normal distribution with mean zero and standard deviation equal
to one. Hence,
√ for dt small, even though dX √is random, we know the size of dX will be of
the order of dt i.e. dX is proportional to dt. Using this we can examine the relative
size of terms in the above equation when dt is small. Retaining terms only up to the size
of the order of dt, gives
∂2V
 
∂V ∂V ∂V 1
dV = SσdX + dt Sµ + + (dX)2 S 2 σ 2 2 + ...
∂S ∂S ∂t 2 ∂S

All higher-order terms (denoted by +...) will be smaller and hence we will neglect. (These
neglected terms are of size (dt)3/2 and smaller, where dt is small.)
But α = V + δS. Therefore

dα = dV + δdS.

Using the above formula for dV , as well as dS = S (σdX + µdt), gives

∂2V
 
∂V ∂V ∂V 1
dα = SσdX + dt Sµ + + (dX)2 S 2 σ 2 2 + ... + δS (σdX + µdt)
∂S ∂S ∂t 2 ∂S
Rearranging gives
∂2V
   
∂V ∂V ∂V 1
dα = SσdX + δ + dt Sµ + + Sµδ + (dX)2 S 2 σ 2 2 + ...
∂S ∂S ∂t 2 ∂S

Justification that (dX)2 can be replaced by dt

29
As a rule of thumb, we can replace (dX)2 by dt in all such equations whenever we see
it, and this will give the correct result when used. We now give an informal demonstration
of this fact. We consider a typical time interval that we wish to solve the option problem
for and denote this as a time interval from 0 to t. We break this time interval up into n
equal time steps each of time dt, where n is some integer. We consider the time at each
time step to be tj where
jt
tj = ,
n
where j is an integer which varies from 1 to n. Note that when j = n then jt/n = t i.e.
the end of that time interval. We choose
t
dt = ,
n
so that each time step is of length dt as required. i.e.
t1 = nt , t2 = 2t
n
= t1 + dt, t3 = 3t
n
= t2 + dt, ..., tn = t.
We let dX (tj ) correspond to the value of the random variable at time tj . We let dX (tj )
be a random variable selected from a normal√distribution with mean zero and variance
equal to dt (i.e. standard deviation equal to dt).
If dX is a random variable selected from a normal distribution with mean zero and
variance equal to dt, then we define the quantity expectation of the random function
g (dX) as
Z ∞
E [g(dX)] = g(x)f (x)dx,
−∞

where
x2
 
1
f (x) = √ exp − .
2πdt 2dt
(Those students who have taken modules in Statistics will recognise this, but it doesn’t
matter if you don’t. We don’t need to consider at all what the expectation corresponds
to - we may think of it as just being an integral. In fact, f (x) is the probability density
function of the normal distribution with mean zero and variance equal to dt, and the
expectation tells us the mean of the resulting function.)
We use this definition to calculate
 !2 
X n
E (dX (tj ))2 − t  .
j=1

This equals
 !2 
n
X n
X
E (dX (tj ))2 − 2t (dX (tj ))2 + t2 
j=1 j=1

30
by multiplying out the brackets. This equals
2
nE (dX)4 + n (n − 1) E (dX)2 − 2tnE (dX)2 + t2 E [1] .
    

This is because each dX (tj ) comes from the same normal distribution for
 every value
of j i.e. E (dX (tj ))2 = E (dX (ti ))2 for all integers i and j; also E (dX (tj ))4 =
 

E (dX (ti ))4 for all integers i and j.


 

The first two terms in the above equation come from the term
n
!2
X
(dX (tj ))2 .
j=1

When squaring this and expanding out the brackets, we get n terms of the form

(dX (tj ))4 ,

which has given us the nE (dX)4 term above, and we get n (n − 1) terms of the form
 

(dX (tj ))2 (dX (ti ))2


2
where i 6= j, which gives us the n (n − 1) E (dX)2

term in the above equation.
In doing this we have written E (dX) = E (dX(tj )) for all j, and E (dX)2 =
4 4
   

E (dX(tj ))2 for all j. Also note that we have used


 

2
E (dX(tj ))2 (dX(ti ))2 = E (dX(tj ))2 E (dX(ti ))2 = E (dX)2
      

which can be proved using the fact the the random numbers dX(ti ) and dX(tj ) are selected
at different times (i.e. at times ti and tj ) and hence are independent from each other.
To explain the above steps in more detail, we will consider the following simple case
when n = 3. Then
 !2 
X3 X3
E (dX (tj ))2 − 2t (dX (tj ))2 + t2 
j=1 j=1

equals
h 2 i
E (dX(t1 ))2 + (dX(t2 ))2 + (dX(t3 ))2 − 2t (dX(t1 ))2 + (dX(t2 ))2 + (dX(t3 ))2 + t2 .


This equals

E (dX(t1 ))4 + (dX(t2 ))4 + (dX(t3 ))4 + 2(dX(t1))2 (dX(t2 ))2 + 2(dX(t1 ))2 (dX(t3 ))2


+2(dX(t2 ))2 (dX(t3 ))2 − 2t (dX(t1 ))2 + (dX(t2 ))2 + (dX(t3 ))2 + t2 .
 

31
This equals
2
3E (dX)4 + 6 E (dX)2 − 6tE (dX)2 + t2 E [1] .
    

We note that this is of the correct form for n = 3. This can then be generalised for all n
in this manner.
We now evaluate our expression
2
nE (dX)4 + n (n − 1) E (dX)2 − 2tnE (dX)2 + t2 E [1] .
    

To do this we note that


Z ∞
E [1] = f (x)dx = 1,
−∞

Z ∞
2
x2 f (x)dx = dt,

E (dX) =
−∞

and
Z ∞
E (dX)4 = x4 f (x)dx = 3 (dt)2 ,
 
−∞

using integration by parts.


Substituting these results into the above formula gives that
 !2 
Xn
(dX (tj ))2 − t  = n 3 (dt)2 + n (n − 1) (dt)2 − 2tn (dt) + t2 .

E
j=1

But dt = t/n, so
 !2 
n  2 !  2
2t2
 
X 2 t t t
E (dX (tj )) − t  = n 3 + n (n − 1) − 2tn + t2 = .
j=1
n n n n

We let n → ∞ in the above expression. Therefore


 !2 
n
X 2 2t2
E  (dX (tj )) − t  = →0
j=1
n

as n → ∞. Since the left-hand side of the above expression corresponds to an integral


over the square of a quantity, the only way that this integral can be zero, in the limit of
n → ∞, is if
n
X
(dX (tj ))2 = t
j=1

32
as n → ∞. This is called the mean square limit.
We now define the stochastic integral as
Z t n
X
h(t)dX = lim h(tj )dX(tj )
0 n→∞
j=1

for a function h(t). While this looks complicated, it isn’t. It just says that we can think
of the integral on the left-hand side as an infinite sum of lots of very small areas under
the curve.
Therefore,
Xn Z t
2
lim (dX(tj )) = (dX)2 ,
n→∞ 0
j=1

by choosing h(t) to be equal to dX(t). Therefore,


Z t
(dX)2 = t.
0
We can rewrite this equation as
Z t Z t
2
(dX) = dt.
0 0
Therefore,
Z t
(dX)2 − dt = 0.

0
2
Therefore, (dX) is equal to dt when written under an integral.
We now return to the equation
∂2V
   
∂V ∂V ∂V 1
dα = SσdX + δ + dt Sµ + + Sµδ + (dX)2 S 2 σ 2 2 + ...
∂S ∂S ∂t 2 ∂S
and integrate throughout from 0 to t giving
Z t Z t     2

∂V ∂V ∂V 1 2 2 2∂ V
dα = SσdX + δ + dt Sµ + + Sµδ + (dX) S σ + ...
0 0 ∂S ∂S ∂t 2 ∂S 2
We can replace (dX)2 in this integral with dt. Therefore,
Z t Z t     2

∂V ∂V ∂V 1 2 2∂ V
dα = SσdX + δ + dt Sµ + + Sµδ + dtS σ + ...
0 0 ∂S ∂S ∂t 2 ∂S 2
Hence, removing the integral, we get
∂2V
   
∂V ∂V ∂V 1
dα = SσdX + δ + dt Sµ + + Sµδ + dtS 2 σ 2 2 + ...
∂S ∂S ∂t 2 ∂S
Thus we have simply replaced (dX)2 in the equation by dt. The above justification of
this has been rather informal. For more details see ’Paul Wilmott Introduces Quantitative
Finance’ pages 119 to 137.

33
3.3 Black-Scholes equation: derivation (part 2)
So we can simply replace (dX)2 by dt, as demonstrated in the previous section.
Therefore, by collecting together the dt terms, we get

1 2 2 ∂2V
   
∂V ∂V ∂V
dα = SσdX + δ + dt Sµ + + Sµδ + S σ + ...
∂S ∂S ∂t 2 ∂S 2

So far we have not chosen δ. If we choose


∂V
δ=− ,
∂S
then
∂2V
 
∂V 1
dα = dt + S 2 σ2 2 + ...
∂t 2 ∂S

Remember that the neglected terms, denoted by +..., are of size (dt)3/2 or smaller.
Dividing through by dt gives

dα ∂V 1 ∂2V
= + S 2 σ 2 2 + ...
dt ∂t 2 ∂S

Therefore, the neglected terms, denoted by +..., are now of size (dt)1/2 or smaller.
Taking the limit dt → 0, we see that all the extra terms, denoted by +... in the
previous equation, all tend to zero, since the largest of them are proportional to (dt)1/2 ,
and (dt)1/2 → 0 as dt → 0. Hence

dα ∂V 1 ∂2V
= + S 2 σ2 2 .
dt ∂t 2 ∂S
This is remarkable, since we have eliminated all the random dX’s from the above equa-
tion. Thus if we know the right-hand side at any time step, then α varies in a non-random
way according to the above equation over that time step.

Before we proceed further, let’s summarise what we have shown so far. We have shown
that a portfolio α = V + δS has a non-random evolution over a time-step dt given by

dα 1 ∂2V ∂V
= σ2 S 2 2 +
dt 2 ∂S ∂t
where the value of the option is V (S, t), S is the value of the underlying and t is time.
We also take E to be the strike price, σ to be the volatility, r to be the interest rate, and
T to be the expiry time. The above equation requires us to choose
∂V
δ=− .
∂S

34
Note that this is often instead re-written as α = V − ∆S with
∂V
∆= ,
∂S
so that
δ = −∆.
Either way, this is commonly called a Delta Hedging strategy.

Now if the money α is instead invested in a bank then it would increase by



= rα
dt
over the same time period, according to our interest rate equation. If these two equations
for dα/dt are not equal to each other, then it would be possible to make a risk free profit
i.e. it would be possible to borrow from the bank and invest in the portfolio, or sell the
portfolio and put the money in the bank (depending on which expression for dα/dt is the
greater), making a risk free profit which is not subject to any random effects. This is not
considered possible, since the stock market would automatically adjust the value of the
underlying S by increasing or decreasing its value to remove this possibility of risk free
investment. This is called the principle of no arbitrage. This says that there is no such
thing as a risk free profit; or that the stock exchange acts by continuously repositioning
the price of shares S to remove any risk free profit. There are traders on the Stock
Exchange who look for arbitrage opportunities, where risk free profits can be made, but
these opportunities close up very quickly if they ever form. Prices on the stock market
effectively move continuously to prevent arbitrage (i.e. preventing risk free profit).
Therefore, by the principle of no arbitrage, these two expressions for dα/dt must be
equal to each other, and hence we have
dα 1 ∂2V ∂V
= σ2 S 2 2 + = rα.
dt 2 ∂S ∂t
But α = V − ∆S. Therefore
1 2 2 ∂2V ∂V
σ S + = r (V − ∆S) ,
2 ∂S 2 ∂t
where
∂V
∆= .
∂S
Rearranging this gives the Black-Scholes equation
1 2 2 ∂2V ∂V ∂V
σ S 2
+ − rV + rS = 0.
2 ∂S ∂t ∂S
This is a partial differential equation for the value of the option V (S, t).

35
3.4 Black Scholes equation: brief discussion
Note that even though we have considered a Delta Hedging strategy (i.e. α = V − ∆S
where ∆ = ∂V /∂S) in the derivation of the Black-Scholes equation, the Black-Scholes
equation is independent of α. Therefore it is a valid equation for V even if the portfolio
you own is not this Delta Hedging portfolio e.g. this equation for V is valid even if you
just own one option of value V and do not own any shares in the underlying; or, in fact,
if you own any other portfolio containing V . This equation will always describe V (S, t).
The Black-Scholes equation can be solved to find the value of the option V at time
t when the value of the underlying S is known. The value of S varies randomly though,
and V is dependent on S. But since we always know the value of the underlying S at the
current time, this equation tells us the value of the option V at that time. The equation
also tells us how future values of V will vary with the share price S of the underlying.
As a bi-product of the previous derivation of the Black-Scholes equation, we have also
shown that the Delta-Hedging portfolio α = V − ∆S is a risk free investment, with value
varying according to

= rα.
dt
For this we must repeatedly change ∆ so that it equals ∂V /∂S, by buying or selling shares
in the underlying. Note that ∆ needs to be a constant over each time-step (since δ was
assumed constant over a time step in the derivation and δ = −∆). Therefore, you would
tend to update the value of ∆ on a regular basis e.g. on a daily basis. This is called
dynamic hedging. Also, note that ∆ is called the Delta of the portfolio.
Also note that the Black-Scholes equation is independent of the drift µ, and also in-
dependent of the random variable dX.

3.5 Solution of the Black-Scholes equation for European options

The Black-Scholes equation can be simplified. Substitute


!
(k − 1) x (k + 1)2 τ
V = Eu (x, τ ) exp − −
2 4

(where S = Eex , t = T − 2τ /σ 2 and k = 2r/σ 2) into the Black-Scholes equation. This


can then be rearranged to show that u (x, τ ) satisfies the diffusion equation i.e.
∂u ∂2u
= 2.
∂τ ∂x
(Try this: Problem sheet 4. Question 1. Warning: very lengthy calculation!)
Note that the share price 0 ≤ S < ∞. Therefore −∞ < x < ∞ since S = Eex . Also,
the Black-Scholes equation is valid for t ≤ T , since we only want to solve the equation for

36
the times before the expiry date (t < T ) and at the expiry date (t = T ). Therefore we
wish to solve the diffusion equation for τ ≥ 0 since t = T − 2τ /σ 2 .
The diffusion equation has the solution
Z ∞  
1 1 2
u= √ u0 (s) exp − (x − s) ds
2 τ π −∞ 4τ
where u = u0 (x) at τ = 0. (Show this: Problem sheet 4. Question 3.) Therefore the Black-
Scholes equation can be solved by just doing an integral! (But only for European options:
later in this module we will look at American options which are more complicated.)
Usually this integral must be solved numerically (e.g. trapezium rule or Simpson’s rule).
The function u = u0 (s) is found from the pay-off function for the European option.
(See Problem sheet 4. Question 2.) This can be done as follows:
If the pay-off function for an option is g(S) where S is the value of the underlying,
then V = g(S) when t = T since T is the time of expiry (the expiry date) i.e. the pay-off
function is just the value of the option at the expiry time. But when t = T then τ = 0
since t = T − 2τ /σ 2 . Therefore the pay-off function g(S) satisfies the equation
 
1
V (S, T ) = g(S) = Eu(x, 0) exp − (k − 1)x
2
since
!
2
(k − 1) x (k + 1) τ
V (S, t) = Eu (x, τ ) exp − −
2 4

(i.e. we have put τ = 0 and t = T into the above formula).


But u(x, 0) = u0 (x) since u = u0 (x) at τ = 0. Therefore,
 
1
g(S) = Eu0 (x) exp − (k − 1)x .
2
Therefore, rearranging this expression gives
g(S = Eex )
 
1
u0 (x) = exp (k − 1)x
E 2
where g(S) is the pay-off function, and since S = Eex . Therefore
g(S = Ees )
 
1
u0 (s) = exp (k − 1)s .
E 2
Example.
For a European vanilla call g(S) = max(S − E, 0) since this is the pay-off function.
Therefore for a European vanilla call
max(Eex − E, 0)
 
1
u0 (x) = exp (k − 1)x
E 2

37
since S = Eex . Therefore, replacing x with s gives

max(Ees − E, 0)
 
1
u0 (s) = exp (k − 1)s .
E 2

Simplifying gives
 
s 1
u0 (s) = max(e − 1, 0) exp (k − 1)s ,
2

or equivalently
     
1 1
u0 (s) = max exp (k + 1)s − exp (k − 1)s , 0
2 2

by multiplying through by the exponential.

Therefore the value of a European vanilla call can be found from calculating
Z ∞        
1 1 1 1 2
u= √ max exp (k + 1)s − exp (k − 1)s , 0 exp − (x − s) ds
2 τ π −∞ 2 2 4τ

(which can be solved numerically), and then substituting that result for u into
!
(k − 1) x (k + 1)2 τ
V = Eu (x, τ ) exp − −
2 4

gives the value of the option V .


Alternately the above integral can be rearranged into the form
V = SN(d1 ) − E exp(−r(T − t))N(d2 ) (see problem sheet 4, question 3) where
Z q  
1 1 2
N(q) = √ exp − s ds,
2π −∞ 2

log(S/E) + (r + σ 2 /2)(T − t)
d1 = √
σ T −t
and
log(S/E) + (r − σ 2 /2)(T − t)
d2 = √ .
σ T −t
Here ’log’ is the natural logarithm, or equivalently ’ln’.
The above expression in terms of N(d1 ) and N(d2 ) is a common way of rewriting the
integral.

38

Potrebbero piacerti anche