Professional Documents
Culture Documents
July, 2001
1. ASSETS AND DERIVATIVES. Assets of all sorts are traded in
financial markets: stocks and stock indices, foreign currencies, loan contracts
with various interest rates, energy in many forms, agricultural products,
precious metals, etc. The prices of these assets fluctuate, sometimes wildly.
As an example, Figure 1 shows the price of IBM stock within a single day.
The picture would look more or less the same across a month, a year, or a
decade, though the axis scales would be different.
If you could anticipate the price fluctutations to any significant extent,
then you could clearly make a great amount of money very quickly. The
fact that many people are trying to do exactly that makes the fluctuations
essentially unpredictable for practical purposes. A fundamental principle
of finance, the efficient market hypothesis [9] asserts that all information
available to anyone anywhere is instantly expressed in the current price,
as market participants race to be the first to profit from new information.
Thus successive price changes may be considered to be uncorrelated random
variables, since they depend on as-yet unrevealed information. This principle
is the subject of intensive analytical testing and some controversy [7], but is
an excellent approximation for our purposes.
Although the directions of the price motions are completely unpredictable,
statistics can tell us a lot about their expected size. Figure 2 shows the dis-
tribution of percentage changes in IBM stock price across half hour time
intervals. We can identify a typical size of the fluctuations, about half of
one percent in this example. Since the fluctuations are uncorrelated and
have mean near zero, this typical size is the single most important statisti-
cal quantity that we can extract from the price history. We may additionally
ask about the form of this distribution, for example, whether or not it is a
Gaussian. Again, this is the subject of active research [10].
∗
To appear in American Mathematical Monthly
1
Robert Almgren/July 2001 Financial Derivatives and PDEs 2
95
94.5
94
93.5
93
10 12 14 16
Figure 1: Price of one share of IBM stock, on November 16, 1999; the x-
axis is time of day. These are prices at which trades actually occured; this
picture contains 5400 data points. The fastest oscillations, on scales of a few
seconds, represent “bounce” between bid and ask prices. But complicated
structure is clear on all higher scales, and continues across decades.
300
250
Number of events
200
150
100
50
0
−0.02 −0.01 0 0.01 0.02
Fractional price change
Figure 2: Fractional price change in IBM stock price across half-hour time
intervals, for 1999 (about 3000 data values). Although the direction of the
changes is unpredictable, we can still identify a characteristic size of the
changes, about half a percent in this example.
Robert Almgren/July 2001 Financial Derivatives and PDEs 3
Options trade in marketplaces, just like the assets on which they are
based. Figure 3 shows market prices of 117 call and put options on IBM
stock on the same day as in Figure 1. The expiration dates range from a
few weeks to six months into the future (longer-term options also exist), and
the strike prices of the options range over about a factor of two above and
below the current stock price. The prices shown are daily closing values;
throughout the day the options fluctuate just as actively as the stock itself.
The first attempt to explain option prices was by Poincaré’s student
Louis Bachelier in 1900 [1], [3]. He proposed that the “correct” value of an
option was the expected value of its payoffs, and by introducing a specific
probabilistic model for the underlying price motion, he was able to calculate
this expectation and compare his results with market prices. In his formu-
lation, the option holder and the writer both took on risk associated with
fluctuations about this mean value.
The surprising fact is that, under suitable assumptions about the statis-
tics of the asset price motion, the risk can be eliminated by following a
suitable hedging strategy. The value of the option is then uniquely deter-
mined. Again, the value is obtained by computing an expectation, which
can be carried out by solving a partial differential equation, though the in-
terpretation is quite different than in Bachelier’s model. This observation in
1973 by Black, Scholes, and Merton [2], [11] led to the development of large
options exchanges—and to the 1997 Nobel Memorial Prize for Merton and
Scholes.
60
40
Calls
20
Apr 2000
0
60 Jan 2000
80 100
120 140 Dec 1999
160 Nov 1999 T
K
60
40
Puts
20
0
Apr 2000
Figure 3: Options on IBM stock, closing prices on Nov. 16, 1999. The x-axis
is the strike price K, at which the option holder may eventually buy or sell
the stock; the dashed line at about 95 shows the closing price of the stock
itself. The y-axis is the expiration date T . The z-axis is the price at which
the option defined by parameters (K, T ) could be either bought or sold at
the Chicago Board of Options Exchange. Call options are more valuable
for lower strike prices, and conversely for puts. The Black-Scholes theory
largely explains the structure of these prices.
Robert Almgren/July 2001 Financial Derivatives and PDEs 6
which depends on the unknown change δS. But we can eliminate the first-
order dependence by choosing
That is, if the investor is able to compute the function V (S, t), then she can
compute its derivative with respect to S and artificially implement a trading
strategy that at first order tracks the same risks.
Assuming this strategy has been implemented, we then have
¡ ¢ ¡ ¢
δ V − Π = Vt − rC δt + 12 VSS δS 2 + · · · , (3)
where the higher-order terms in δt and δS are small if the time interval and
the corresponding price changes are small. This change is still an uncertain
quantity, since we do not know δS 2 .
However, we argued in the introduction that we know much more about
the size of the changes δS than about their direction. As a consequence,
it may be that δS 2 is effectively deterministic, as long as we average over
enough small steps. Let us consider a time interval ∆t that is small compared
with the overall lifetime of the option, yet large compared with the time
intervals δt at which we are able to trade. Let us take ∆t = N δt, and
denote the small price changes by δSj for j = 1, . . . , N .
Now, for the first time, we write down a specific probabilistic model for
the price changes. We suppose that
√
δSj = a δt + b δt ξj , (4)
→ b2 ∆t
Note that the “drift coefficient” a has disappeared, hence does not affect
the value of the option. The only coefficients that appear are b and r, the
size of the motions and the risk-free interest rate, respectively. This partial
differential equation must be satisfied by the value of any derivative security
depending on the asset S.
The PDE (7) is linear: two options are worth twice as much as one
option, and a portfolio consisting of two different options has value equal to
the sum of the individual options.
The PDE is backwards parabolic. Thus, terminal values V (S, T ) must be
specified at some future time T , from which values V (S, t) can be determined
for t < T . Typically, the value of an option is known at expiration, and this
equation is solved to determine its values for earlier times. In addition,
boundary conditions may arise from features of the option specification.
The option price may also be calculated by probabilistic methods [6]. In
this equivalent formulation, the discounted price process e−rt S(t) is shifted
into the “risk-free measure” using the Girsanov theorem, so that it becomes
a martingale. The option price V (S, t) is then the discounted expected value
of the payoff Λ(S) in this measure, and the PDE (7) is obtained as the back-
wards evolution equation for the expectation. Our derivation has followed
the original reasoning of Black and Scholes, although the probabilistic view
is more modern and can more easily be extended to general market models.
We have achieved the remarkable conclusion that the market value of
the option V (S, t) is uniquely determined by its boundary conditions and
Robert Almgren/July 2001 Financial Derivatives and PDEs 10
by the parameters of the probabilistic model for the underlying asset price
motion. Let us summarize the assumptions that went into this model:
• The asset S can be bought and sold. This is essential for us to construct
a suitable hedging portfolio. Thus the model cannot be applied to
risk factors that don’t exist in a market. For example, home heating
costs depend on outside temperature. You are at the mercy of this
risk, unless you can find a way to buy and sell temperature (this is
now possible, precisely by using hedging portfolios composed of energy
futures).
• Assets can be bought and sold with no transaction costs. In practice,
trading costs are substantial, including both the bid-ask spread on
small amounts and liquidity limits on large amounts. This constrains
the frequency with which new values of D can be taken and means
that risk cannot be eliminated completely.
• The market parameters r and b are known. The interest rate r is dif-
ferent for different investors (usually the rate taken is the “overnight”
rate for short-term cash deposits between major banks), but does not
have too large an effect on the result. However, the computed value
V is very sensitive to the input value of b, and this value is very diffi-
cult to estimate empirically. In practice, option prices quoted on the
exchanges are usually used to determine appropriate “implied” values
for the parameters of the price process.
• The asset price follows a Brownian motion. This model follows from
the assumptions that (a) price is a continuous function of time, (b) its
increments are independent random variables, even when viewed on
arbitrarily short time intervals, and (c) variance is finite and constant.
As a consequence, the price changes ∆S across intermediate time in-
tervals have a Gaussian distribution. Although these assumptions are
very plausible, the results are not consistent with empirical observa-
tions either of the asset prices themselves [10] or of option prices.
Constructing improved models for asset price motion and for option
pricing is a subject of active research.
An especially popular choice is the lognormal model b(S, t) = σS, so
δSj = a(S, t) δt + σ S ξj .
That is, the percentage size of the random changes in S, rather than the
absolute sizes, are assumed to be constant as S varies. The parameter σ is
Robert Almgren/July 2001 Financial Derivatives and PDEs 11
√
called the volatility, and σ ∆T is the expected size of changes across a time
interval ∆T . For this model, the Black-Scholes equation is
Vt + 21 σ 2 S 2 VSS + rSVS − rV = 0.
40
30
Call option price
20
S
10
0
60 80 100 120 140 160
Strike price K
American options. Let us again suppose that the option price V (S, t)
is a smooth function, and denote by δV the change in value across a short
time interval, δV = V (t + δt) − V (t). Earlier, we argued that δV was given
by the expression on the right-hand side of (1).
Now, however, the reasoning needs to be modified. We argued previ-
ously that ∆(V − Π) = r(V − Π)∆t, for if the equality were violated in
either direction, there would be a strategy available to the option holder
that would guarantee a profit. One side of that strategy depended on short-
ing the option. But for an American option, shorting it means giving the
counterparty the decision whether to exercise, and it is unlikely that he will
do it in a way that is optimal for the holder of the option. Thus (6) needs
to be modified to ¡ ¢ ¡ ¢
∆ V − Π ≤ r V − Π ∆t,
and the PDE (7) becomes
Vt + 21 b2 VSS + rSVS − rV ≤ 0. (8)
Of course, a partial differential inequality is not sufficient to determine
V (S, t). The second piece of information needed is that
V (S, t) ≥ Λ(S, t), (9)
precisely because the option may be exercised at any time: if it ever hap-
pened that V < Λ, then a risk-free profit could be made by purchasing the
option for V and immediately exercising it to collect Λ.
One further statement is required to determine V uniquely. At each
moment, the option value arises purely from the possibility either to exercise
immediately or to hold and get the exercise value at a later time. This yields
the third constraint:
At each (S, t), at least one of (8) and (9) is an equality. (10)
The three pieces (8)–(10) constitute a linear complementarity or obstacle
problem [12]. The (S, t)-plane may be divided into two pieces, an exercise
region in which (9) is an equality and a hold region in which (8) is an equal-
ity. Assuming everything is regular, the boundary between these regions is a
curve S∗ (t), the optimal exercise boundary. The option holder should exer-
cise when the price S(t) crosses this level. The motion of the boundary can
be modeled as a free boundary problem, mathematically similar to the Ste-
fan problem of freezing/melting. Solutions may be obtained numerically by
adding obstacle features to a finite-difference code or by solving an integral
equation for the motion of the boundary.
Robert Almgren/July 2001 Financial Derivatives and PDEs 14
Other exotics. Almost any contract between consenting parties that could
possibly be written is probably being traded somewhere right now. Here is
a brief list of possible extensions, any one of which may include American
features as well.
• The payoff function may have a more complicated form. For a digital
option, Λ(S) is a step function: it pays a fixed amount if S > K
at expiration, zero if S < K. By putting together piecewise linear
payoffs, one obtains spreads, straddles, collars, etc. These are valued
in the same way as the vanilla assets, namely, by solving the PDE with
appropriate terminal data.
• A barrier option loses its value if the underlying asset price crosses a
specified level (which may be either above or below the current price)
at any time before expiration. It is valued by adding a homogeneous
Dirichlet boundary condition to the diffusion equation (7).
• Asian options depend in some way on a time average of the price. For
example, one may have the right to acquire the asset for its average
price over a specified time period. These are valued by adding an
additional state variable. The option price V (S, I, t) also depends on
the value of the running average I, and the Black-Scholes equation (7)
becomes a degenerate diffusion equation in two variables [13]. Lookback
options depend on the maximum or minimum asset price in a time
window and are priced by similar means.
In addition to changing the definition of the option, one may make the
market model more realistic. For example, transaction costs are important
in real life, preventing the hedging strategy of Section 2 from being imple-
mented in continuous time.
Finally, a more realistic statistical model may be used for the price pro-
cess itself. Two prominent models for incorporating realistic statistics are
stochastic volatility [4] and jump models such as variance gamma [8]. Both
of these incorporate fat tails, hence are capable of correcting the systematic
mispricings mentioned at the end of Section 2.
Robert Almgren/July 2001 Financial Derivatives and PDEs 15
In this note, we have reviewed the basic theory of pricing and hedging
options, and a few of its extensions. Besides the practical importance of the
calculations, the development of such quantitative models represents a real
advance in our ability to think rationally about risk and hedging. Although
the classical framework of Brownian motion and Black-Scholes hedging has
been developed to a high degree of refinement, there still remains a lot to do
in the area of more realistic stochastic models, not to mention the connection
of these ideas with traditionally less quantitative areas such as insurance and
risk management.
References
[2] F. Black and M. Scholes. The pricing of options and corporate liabilities.
J. Political Econ., 81:637–654, 1973.
[5] J. C. Hull. Options, Futures, and Other Derivatives. Prentice Hall, 4th
edition, 2000.