Вы находитесь на странице: 1из 9

MSC FINANCIAL ENGINEERING

PRICING I, AUTUMN 2010-2011


LECTURE 3: THE BINOMIAL MODEL, II
RAYMOND BRUMMELHUIS
DEPARTMENT EMS
BIRKBECK

We continue our discussion of binomial models with the following


topics: the backwards induction algorithm to compute prices, the valuation of American options, calibration to real stock-price data and,
finally, as a first step towards continuous-time finance the continuous
time limit of the model.
1. Backwards induction
Once values are given for the binomial models parameters u, d, r and
N option prices can in principle be computed using (??). If we would
implement this formula on a computer then, for large values of N , we
might encounter problems due to the exponential growth of some of
the binomial coefficients1. A better idea is to use backwards induction,
which basically amounts to applying the one-period valuation formula
at each node of the binomial tree.
Let us call Vn,j the value of the derivative at the (n, j)-th node of
the tree, where n is the time, and j the number of down-moves in
the stock-price. Specifically, at the (n, j)-th node, we have that the
stock-price equals
(1)

unj dj S0 =: Sn,j ,

while the derivatives price is


(2)

Vn (Sn,j ) = Vn (unj dj S0 ) =: Vn,j .

If we find ourselves at node (n, j), then at time n + 1 the stock-price


will either go up to uSn,j = Sn+1,j or go down to dSn,j = Sn+1,j+1 :
(3)

Sn,j

%
&

Sn+1,j
.
Sn+1,j+1

1One can for example show that for large and even N , the middle binomial coefficient is
approximatively

 r
2 2N
N

'
,
1
N
N +1
2
cf. e.g. http://planetmath.org/encyclopedia/AsymptoticsOfCentralBinomialCoefficient.html.
1

R. BRUMMELHUIS

Correspondingly, for the derivative,


(4)

Vn,j

%
&

Vn+1,j
.
Vn+1,j+1

Consequently, when holding the derivative at node (n, j) we will be


holding an instrument which at time n + 1 will have a pay-off of either
Vn+1,j (when the stock-price goes up from Sn,j to Sn+1,j ) or Vn+1,j+1
(when the stock-price goes down to Sn+1,j+1 ). Applying the arguments
of the one-period binomial model, we find that the value at node (n, j)
of derivative is therefore
1
(5)
Vn,j =
(qVn+1,j + (1 q)Vn+1,j+1 ) .
1+r
At time N , the value of the option is simply its pay-off, which is a given
function of SN . Therefore, VN,j is known, for all j = 0, 1, . . . , N . For
example, for the case of a call,
(6)

VN,j = max(SN,j K, 0) = max(uN j dj S0 K, 0).

Starting from n = N , we can then use (5) to traverse the tree backwards
in time, to finally arrive at V0,0 = V0 (S0 ), according to the scheme

(7)

VN,j
VN,j+1
VN,j+1
VN,j+2

&
%

VN 1,j

&
%

VN 1,j+1

&
%

VN 2,j ,

etc. This procedure is called backward induction. It can easily be easily implemented on a computer, and circumvents potential numerical
problems coming from large binomial coefficients.
2. Pricing American options
A small variation on the backward induction method will give us a
simple algorithm for pricing American options binomial models. To fix
A
ideas, let us consider an American put with strike K, and let Vn,j
be
the value of this put in node (n, j) of the binomial tree. If we are still
holding the put at expiry, n = N , then clearly
(8)

A
VN,j
= max(K SN,j , 0), j = 0, . . . , N.

Next, suppose we are holding the put at node (n, j). There are then
two possibilities: either we exercise the put, giving us an immediate
pay-off of
(9)

max(K Sn,j , 0),

or we do not exercise the put, in which case we will be holding in an


A
A
asset paying off either Vn+1,j
or Vn+1,j+1
at time n + 1, depending on

PRICING I, LECTURE 3

whether the stock goes up or down. The value, at node (n, j), of keeping
the American option alive (not exercised) till n + 1 is therefore

1
A
A
(10)
qVn+1,j
+ (1 q)Vn+1,j+1
,
1+r
which might be called the continuation value of the put. The American
puts value at (n, j) is then simply the greatest of (9) and (10):



1
A
A
A
Vn,j = max
qVn+1,j + (1 q)Vn+1,j+1 , max(K Sn,j , 0) ,
1+r
the maximum of the continuation value and the value upon exercise.
The same backwards induction procedure as before, starting off from
A
, the puts value at time 0.
(8), will then yield V0,0
The exercise region of the put is, by definition, the set of nodes
A
(n, j) where the puts value is its exercise value, that is, where Vn,j
=
max(K Sn,j , 0). Equivalently, we can think of it as the set of pairs
(n, Sn,j ) such that VnA (Sn,j ) = max(K Sn,j , 0). In the continuous
time-limit which we will consider below, when the individual binomial
time steps correspond to time-intervals of length T /N and N ,
this set would tend to a region in the (t, S)-plane of the form
{(t, S) : 0 t T, 0 S S (t)},
bounded by the graph of some function S (t) of time t, called the
exercise boundary of the put. Intuitively, the put will be exercised at
time t if and only if the stock-price at t is less than S (t). It hasnt
proven possible to find an explicit analytic expression for this exercise
boundary.
3. Calibration of binomial models
Suppose we would want to use the binomial model to price a real-life
European option on some real-life asset S. How would we choose the
models parameters, in particular u and d (since they determine q and
therefore the deru=ivative prices)? Suppose the maturity of the option
is T and that t = 0 corresponds to today. We divide the time-interval
[0, T ] up into N equally sized subintervals [tj , tj+1 ], where
tj = j

T
= jt, j = 0, . . . , N 1.
N

where
t :=

T
.
N

Let
log(Stj /Stj1 )
be the (ex-post) log-return of the stock over the period [tj1 , tj ]. From
a time tj1 perspective, this will be a random variable, whose value
will be realized at time tj . We will make the modelling assumption:

R. BRUMMELHUIS

The log(Stj /Stj1 ) are iid random variables, with mean t and
variance 2 t, where and are constants.
Observe that we assume that both mean and variance of these logreturns are proportional t, the time elapsed. We stress that we are
making an assumption here, which should (and has been) be tested
econometrically. Likewise for the independence assumption, or the assumption that and 2 are constants. The values of these constants
will have to be estimated from price data.
From a mathematical modelling perspective, the assumption of mean
and variance being proportional to time elapsed is convenient, because
it will allow us to analyze the large N - (or small t -) limit of the
binomial model below using the the central limit theorem.
We now set up a binomial model Sj = Xj Sj1 , with Xj iid Bernouilli,
which will model the stock-price Stj at time tj . It is convenient to
introduce

log u, probability p ,
(11)
Yj := log Xj =
log d, probability 1 p ,
the log-return in our binomial model. To determine u and d, we simply
require that the mean and variance of the log-return in the binomial
model coincide with the mean and variance of Yj = log(Stj /Stj1 ). This
leads to
E(Yj ) = p log u + (1 p) log d,
2
and
var(Yj ) = (p p2 )(log u)2 + ((1 p) (1 p)2 )(log d)2 +
2p(1 p) log u log d.
Equating these to t respectively 2 t gives us a system of 2 equations
in three unknowns: log u, log d and p, which is one too many. However,
we already know that pricing options, the value of p doesnt matter. If
we artificially set p = 0.5, our system of equations becomes:
 1
(log u + log d) = t
2
1
(log u log d)2 = 2 t,
4
with solution:

(12)

log u = t + t
log d = ut t.

Hence
(13)

u = et+

, d = et

For S0 we of course take todays stock price. Finally, we should have


to calibrate the risk-free interest rate. If we agree from now on that r
designates the (continuously compounded) interest rate in % per year,
2using e.g.. that var(Y ) = E(Y 2 )(E(Y ))2 ; one can of course also directly apply the definition
j
j
j
of variance.

PRICING I, LECTURE 3

and that time is measured in units of one (financial) year, then the
continuously compounded interest over a single period [tj , tj+1 ] will be
ert . We therefore replace r by ert in all formulas established previously
for the binomial model. In particular, the (single-period) risk-neutral
probability for an up-move in the stock price will be given by

ert et t

.
(14)
q=
et+ t et t
This completes our calibration.
As a preparation for the continuous time-limit, we now look at the
leading-term approximation in this expression for q as t
0. It turns
out that it will be sufficient to consider terms to order t, treating
higher powers as negligible. Taylors series ex = 1 + x + 21 x2 + . . . for
the exponential gives that, after some calculation3


t + (r 12 2 )t + O (t)3/2

q =
2 t + O ((t)3/2 )
1 r 12 2
(15)
+
t + O(t).
=
2
2
We will now use this approximations to study the large-N limit of our
calibrated binomial model.
4. Continuous time-limit and a first encounter with Black
and Scholes
We want to distinguish between the binomial models gross returns
Xj or log-returns Yj = log Xj when considered with the objective
probability p and with the risk-neutral q. To do so, we will put a hat
over any random variable representing a return or a price when we
considering it with respect to the risk-neutral probability. Explicitly,

cj = u probability q
(16)
X
d probability 1 q .
bj . Since, working with the risk-neutral
and similarly for Ybj = log X
4
probability ,
bN X
bN 1 X
b1 S0 = eYb1 ++YbN S0 ,
SbN = X
3Note that
et+

=
=

et t

1
(1 + t + t + (t + t)2 +
2



1
1 + t + t + 2 t + O (t)3/2 ) ,
2

and similarly for


(when approximating the denominator).
4The analogous formula holds with the unhatted variables also holds; only the probabilities
would change.

R. BRUMMELHUIS

our option pricing formula (??) can be written as


 

1
Yb1 ++YbN
(17)
V0 (S0 ) =
EQ V S0 e
,
(1 + t)N
where V (ST ) is the Europeans options final pay-off at maturity. The
occurrence of a sum Yb1 + + YbN of a large number of iid random
variables naturally leads to the idea of trying to apply the Central
Limit Theorem:
Theorem 4.1. (Central Limit Theorem or CLT) Let the random variables Zi be iid (=independent and identically distributed), with mean
a, variance b, then
Z1 + + ZN N a

w W N (0, 1).
(18)
b N
The meaning of the limit w is that the probability distribution
function of the left hand side converges to the probability distribution
of a normal N (0, 1)-random variable. In particular, this implies that,
for any reasonable function g:
 

Z
Z1 + + ZN N a
dz
2

(19)
E g

g(z)ez /2 .
2
b N

To apply the CLT, we compute the mean and variance of Ybj , the
hat indicating that we use the risk-neutral probability q. Using the
approximation (15), we find that Eq (Ybj ) = q log u + (1 q) log d is
approximatively equal to





r 12 2
r 12 2
1
1
+
t
(t
+

t)
+

t
(t t)
2
2
2
2

= (r 21 2 )t + O (t)3/2 .
(20)
A similar (but slightly longer) calculation shows that

(21)
var(Ybj ) = 2 t + O (t)3/2 .
We therefore find, with a the mean of Ybj and b2 its variance, that
Yb1 + + YbN (r 21 2 )T + O(N 1/2 )
Yb1 + + YbN N a

=
,
b N
T + O(N 1/2 )
where we used that t = T /N and hence N (t)3/2 = O(N 1/2 ). By the
CLT, this random variable has a probability distribution which tends
to that of a standard5 normal random variable. Since the terms of
order O(N 1/2 ) tend to 0 as N tends to infinity, we can conclude that
Yb1 + + YbN (r 12 2 )T

w W N (0, 1),
T
5i.e. mean 0, variance 1

PRICING I, LECTURE 3

and therefore
Yb1 + + YbN

1
(r 2 )T + T W
2

1 2
2
N ((r )T ), T .
2

The enables us to replace


the expectations

part of (17) by the expec
1 2
tation of V S0 exp r 2 )T + T W , with W N (0, 1). As to
the discount factor in front, using the classical result that

x N
1+
ex ,
N
together with t = T /N , we find that
1
erT ,
N
(1 + rt)
and we arrive at the following valuation formula:
Theorem 4.2. (Continuous time limit of the binomial model) If N
, then the price at time 0 of a European option in the binomial model
(calibrated as above) tends to:
Z 
 1 2 dw

1 2
rT
(22)
V0 (S0 ) = e
V S0 e(r 2 )T + T w e 2 w .
2

This result is due to Black and Scholes, who derived it by other


means, as we will see in the next section. The present demonstration
by taking the limit of a sequence of binomial models is due to Cox, Ross
and Rubinstein [3], who also introduced these binomial models. If we
take V (S) = max(S K, 0) we would find, after some computation, the
well-known Black and Scholes formula for the fair price of a European
call - cf. the next lecture.
5. Remark on calibration
Note that, to first order in t, the variance of the log-returns with
respect to the risk-neutral probability is the same as their real-life
variance. Since, in the limit of large N , the mean return disappears
from our pricing formula, the only thing which matters is the stocks
volatility, . We can exploit this by setting = 0, leading to the
simpler calibration formulas
u = e
and

, d = E

ert e t
,
q=
e t e t
used e.g. in Bingham and Kiesel [1] below. See also Wilmott, Howison
and Dewynne [3] for alternative calibrations. These references also

R. BRUMMELHUIS

contain some worked examples. The book [4] is entirely devoted to


binomial models and their use in finance.
6. Variations on the binomial model
The binomial model can be further developed in various directions.
For example, we can, at each node (n, j) let the next time-steps up
or down move depend on the value of the stock-price at that node,
effectively replacing u, d by a node-dependent un,j , dn,j . More generally,
we could make u and d themselves random at each of the nodes. These
possibilities correspond in continuous time models to, respectively local
volatility and stochastic volatility, which we will encounter later on
during these lectures.
References to Lecture 3
[1] Bingham, N. H. and Kiesel, R., Risk-Neutral Valuation: Pricing
and Hedging of Financial Derivatives, Springer Finance, 2-nd edition
(2004), chapter 4, in particular sections 4.5, 4.6.
[2] Cox, J. , Ross, A. and Rubinstein, M., Option pricing, a simplified
approach. J. Fin. Econ. 7 (1979), 229 - 263.
[3] Wilmott, P., Howison, S. and Dewynne, J., The Mathematics of
Financial Derivatives, Cambridge University Press (1999), Chapter 10.
[4] Hoek, John van der, Elliott, Robert J. Binomial Models in Finance,
Springer Finance (2006).
Exercises to lecture 3
Exercise 3.1 Explain how you can exploit the independence of option
prices of the actual (or physical) probability p of an up move, to be
able to set = 0 in the calibration procedure of section 3 of this
Lecture; cf. also section 5.
(Hint: change p so as to achieve that the expectation using p of the
one-period return becomes 0.)
Exercise 3.2 We consider 3-month maturity options written on a
stock which today is worth 100. Suppose that the stocks volatility is
20% p.a. (per annum or yearly) and that the risk-free rate is 2% p.a.
(a) Calibrate and construct a three-period risk-neutral binomial tree
for the stock. (Be careful how you scale back volatility to the monthly
level.)
(b) Use the tree to price a European call with a strike of 95.
(c) Use the tree to price an at-the-money American put on the stock
with a strike of 100.

PRICING I, LECTURE 3

Exercise 3.3 Continuing with the set-up of the previous exercise,


suppose that it is known that in 2 months time the stock will pay out
a dividend of 5. Now find the price at time 0 of an American call on
the stock with a strike of 100. Will it be optimal to exercise at the
ex-dividend date or not?
Exercise 3.4 Suppose you now want to construct a binomial tree to
price options on a stock which at time t + 1 pays out a dividend of St ,
where St is the stock-price at t, t = 0, 1, 2, , N 1.
(a) Show that the risk-neutral probability of an up-move is now
1+rd
q=
.
ud
(b) For q to be a bona-fide probability, we need that 0 q 1.
Show that if q is outside of this range, then the market-model presents
arbitrage opportunities, already in the one-period framework.
(c) By repeating the arguments of lecture 2 (replicating portfolio or
hedging), show that the q found in part (a) is indeed the probability
one has to use to price options in a binomial tree.
Exercise 3.5 Suppose that the stock from Exercise 3.2 now also pays
out a dividend in each period, at an annual the rate of 5%. By constructing a 3-period binomial tree, price an American call on the stock
with a strike of 100.
Exercise 3.6 Consider a risk-neutral binomial tree modelling a stockprice with annual volatility . We assume that the tree is calibrated
using the procedure of section 5 of this Lecture; in particular,

ert e t
,
q=
e t e t
and

u = e t , d = e t .
We know if if V (S, t) is the price of an option at a node of the tree
(that is , t = n t for some n, S = Sn,j for some j), then
(23)
n 


o

1
V (S, t) =
qV Se t , t + t + (1 q)V Se t , t + t .
1 + rt
Assume that V (S, t) is as many times differentiable as needed. By letting t 0, show that V (S, t) satisfies the Black and Scholes equation:
V
1 2 2 2V
V
+ S
+ rS
= rV.
2
t
2
S
S

(Hint: Using Taylor series,


expand (23) to order 2 in t, treating S
as fixed; the terms of order t should automatically cancel, and those
of order t should combine to give the Black and Scholes equation.)

Вам также может понравиться