Beruflich Dokumente
Kultur Dokumente
Kaushik Amin
is with L.&man Brothers in New York.
This article provides a selective description of state variables. Monte Carlo methodsor numerical methods
numerical techniquesfor option valuation when the under- for solving partial di@zrential equations are almost com-
lying model is based on certain kinds of dijiision processes. pletely ignored.
It emphasizes models based on discretiring the underlying In the interestof clarity, thefocus is on spec@ pro-
processon a finite state spaceat discretepoints in time and cessesto illustrate the basic principles, forgoing technical
describes techniquesfor building both path-dependent and details that may cloud the basic concepts.The major focus
path-independent discrete-time models that can approxi- is on describing models that can be realistically implemented
mate a given continuous-timeprocessfor the underlying in practice.
I
n most option pricing models, an American
option’s price is basedon the formula: Price = E
[ 1
Payoff (T)
e&j-r(u)du
(2)
Price = Max, E
[ 1
Payoff (t)
,& r(u)du
(1)
where T is the maturity date of the option.
I will take Equations (1) and (2) as given and
not delve into the theory of option valuation. Our
where z is any early exercise strategy (stopping objective is to study methods that can compute the
time);’ E is the expectation over some probability solutions to (1) and (2) when these equations cannot
distribution described by a diffusion process; Payoff be solved in closed form.
(t) is the state-dependent cash flow to the option if it Numerical methods to solve (1) and (2) can be
is exercised at date t; and r(u) is the continuously classifiedinto two types: 1) probabilistic methods, and
compounded spot interest rate at date u. 2) methods that require the numerical solution to
At the terminal date, the option price equals partial differential equations. The probabilistic meth-
the exercise value. The discount factor e6 r(u)du rep- ods compute the solution by constructing an approxi-
resents the time t value of a $1 investment at time 0 mating discrete-time process with discrete states at
that is continuously rolled over at the spot interest each discrete date such that the option value from
rate at each date. It is stochastic in models with this processis an approximate solution to (1) or (2). I
stochastic interest rates and may depend both on the will focus on some of these methods.
time and the state since the spot interest rate may be The solutions to Equations (1) and (2) can also
a function of the state. be written as solutions to an elliptic partial differential
If the option is European, (1) is replaced by equation (hereafter, PDE) with a free-boundary con-
the simpler equation: dition for (1) and a fixed boundary condition for (2).
I first describe the approximation of jointly The increments to the processes due to the dt term
Iognormal (or normal) processes that have constant are hereafter called the drifts of the two processes, and
volatilities and correlation parameters, perhaps the the terms due to the Brownian motions are called the
simplest type of path-independent models. For illus- dispersion terms.
tration, first consider a process with two state vari- From a numerical perspective, it is much easier
ables. As I later show, it is easy to construct models to rewrite the equations in terms of orthogonal
with an arbitrary number of state variables. Brownian motions. Each Brownian motion can be
separately approximated by a simple discrete-time pro-
Model with ‘Iho Lognormal Processes
cess without worrying about the other Brownian
Consider a model with two jointly lognormal motion. Therefore, we redefine (W,, W,) to be
and correlated state variables (say, Stock 1 and Stock 2 orthogonal and rewrite Equation (4) as:
I
tion (1) is satisfied. Further, from (8), it is apparent
that the discrete-time process cannot have Iarge jumps “c XI (ih), k X2 (ih)
i=O i=O
because increments are of the order d-h . Therefore,
our discrete-time process will converge in distribution
to the desired process in (5). Correspondingly, option Therefore, to compute the value of (S,, S2) at
In particular, we do not need to know the sequence Log [Sj(t + h)] - Log [Sj(t)] =
(or the path) by which the different values of (Xl, XJ
are realized at each of the prior dates. Therefore, the
model is path-independent. [~j - ~ ~J?] h + ~ in Oji Xi(t)
A simple figure illustrates the notion of path- i=l
38 OPTIONPRICINGTREES SUMMER1995
ESHIBIT 2 cross-product of the states due to each dimension.
EVOLUI’IONOFTwo STATEVARIABLES
wm FOUR This state space is numerically feasible for realistic
POSSIBLEINCKEMINCS Two PERIODS problems with up to five state variables and thirty time
ATEACHDA-ITOVE:R
stepsand can be used in practice.
(B)
(AB)
What if the volatility function is not lognormal
with a constant volatility? The simple discretizations
(A.0
given above will not yield path-independent models.
63 (AD). UK)
Path-dependent models, however, can be quite use-
ful, as we will show. In some special cases,it is possi-
ble to use additional tricks to make the model path-
e (D) (B.D)
independent.
CASE 1: SINGLE STATE
VOLATILITY FUNCTION A DETERMINISTIC FUNC-
VARIABLE;
where nr is the number of time steps up to time t and we are interested in the evolution of S(t) up to
(which now depends on the time variation in the some date T. Discretize the time interval into N peri-
volatility). 6, (i) and hi are the volatility and size of the ods of duration h = T/N. Instead of building a tree,
time step over the ifh interval. fix an arbitrary state space at each discrete date with
The trick used above dependson the concept of points equally spaced in the space dimension where
time change in probability theory Essentially, when the spacing is “sufficiently” close.4 Such a state space
volatility is high, we slow the clock by taking smaller is shown in Exhibit 3 where the state points are equal-
time steps, and when volatility is low, we speedup the ly spacedA units apart.
clock by takmg larger time steps.Under the new clock, At the terminal date of the option, the option
it appearsas if the volatility is constant, and we can use value at each of the points on the grid can be deter-
our earlier methods for approximating the process.But, mined as a function of the state and time. Given the
as is apparent from this analogy,we cannot extend this value of the option at the terminal date, we can
trick easily if there are multiple state variablessince the work backward through time on the state space grid,
time change (variable step size) for one variable may computing the option values by applying Equations
not be the sameasthat for the other statevariables. (6) and (7).
The important feature is that the time change Supposethe option prices for each of the points
is deterministic, i.e., we speed up the clock and slow on the grid at date t + h have already been computed.
it down according to a predetermined scheme. Now, consider the grid at date t. To apply (6) and (7),
Otherwise, at any point in the future, we would not we have to approximate the distribution of the incre-
know how much time had elapsed, and therefore we ments to the processfrom date t to date t + h.
would not be able to discount appropriately (if the Consider a discretization of the continuous-
interest rate is not constant, we would not know what time processgiven by:
rate to use) or know if we are at the maturity date.3
Even though there is a corresponding concept when S(t + h) - S(t) = CL(t, S (t)) h + CF(t, s (t)) &
the volatility can also be a function of the state, imple-
menting a tree-based model of the type discussed probability = 0.5 State A
above is difficult and requires augmenting the state
spaceby an additional state variable. = CL(t,S 6))h - o (t, s (t)) dii
Ho, Stapleton, and Subrahmanyam [1993] have
developed some techniques for time-dependent probability = 0.5 State I3 (14)
any given date t depends on the cumulative value of a of N is given in Exhibit 5. The model is feasible with
(u, T) for u < t, the dependence of a (u, T) on the only fifteen to twenty time steps.
state makes the cumulative drift path-dependent; i.e., Even with so few time steps, it is possible to
the sequence in which the different values of X(t) are compute accurate option prices with maturities up to
realized affects the forward rates at any future date. ten years.6 Since the path-dependent tree includes a
Correspondingly, even if the dispersion term is large number of nodes, it is very dense. So, if we
path-independent, the model will be path-dependent. choose the points at each time step carefully, the dis-
Therefore, a model of the form (21) necessarily crete-time distribution approximates the continuous-
requires a path-dependent discrete-time model. time limiting distribution quite well.
Since our model is path-dependent, the tree One scheme for choosing the state points is to
representing the evolution of the term structure of choose linearly increasing or decreasing time steps
forward interest rates is “not recombining” as in such that the ratio of the largest time step to the
Exhibit 3. This tree is represented in Exhibit 4. Note smallest time step is between 2 and 3. The efficacy of
that the state (A, B) corresponding to state A at date this approach is illustrated in Exhibit 6. This figure
h and state B at date 2h is not the same as state (B, plots the terminal cumulative distribution function
A), which corresponds to state B at date h and state A obtained using the discrete-time path-dependent
at date 2h. model with ten time steps (dotted line) and the theo-
With a two-point distribution for X(t) as in retical cumulative distribution function (solid line).
Example 3, the number of states after time step N is The ratio of the maximum to the minimum time step
used to generate this figure is 3.
EXHIBIT 4 For simplicity, I have plotted a distribution that
STATE SPACEFORA THREE-PERIOD PATH-DEPENDENT approximates Brownian motion at some future date
IC~ODELWTII Two POSSIBLEINCRE~WWS(A AND B) AT and have scaled the axes in terms of the standard devi-
EACH DATE ation of the Brownian motion. Therefore, the graph is
independent of any fixed maturity date.
It is apparent from the graph that the cumula-
tive density function is approximated extremely well
by the scheme suggested above. With more compli-
cated schemes, the approximation can be improved
further. With these schemes, option prices are typical-
<
ABA ly within 1% or 2% of their limiting values.
AB For a description of implementing path-depen-
dent models with multiple state variables, see Amin
ABB and Bodurtha [ 19951.
0
BAA
III. SUMMARY
NBA
< I have described a large class of methods for
BAB implementing multivariate option pricing models
BBA based on diffusion processes. The path-independent
BB and state space models demonstrated here are very
intuitive and simple to implement. The path-depen-
< BBB dent model is more complicated and requires signifi-