✐

916 VII. The Influence of Mathematics

. 2004. World poverty: causes and pathways. In Annual that these instruments are generally regarded as being
World Bank Conference on Development Economics 2003: riskless.
Accelerating Development, edited by F. Bourguignon and No arbitrage, the second fundamental principle, sim-
B. Pleskovic, pp. 159–96. New York: World Bank and ply says that it is impossible to make money without
Oxford University Press.
taking risk. It is sometimes called the “no free lunch”
Debreu, G. 1959. Theory of Value. New York: John Wiley.
principle. In this context, “making money” is defined
Diamond, J. 1997. Guns, Germs and Steel: A Short History
of Everybody for the Last 13,000 Years. London: Chatto &
to mean making more money than could be obtained
Windus. by investing in a riskless government bond. A simple
Durlauf, S. N., and H. Peyton Young, eds. 2001. Social application of the principle of no arbitrage is that if one
Dynamics. Cambridge, MA: MIT Press. changes dollars into yen and then the yen into euros
Evans, G., and S. Honkapohja. 2001. Learning and Expecta- and then the euros back into dollars, then, apart from
tions in Macroeconomics. Princeton, NJ: Princeton Univer- any transaction costs, one will finish with the same
sity Press. number of dollars that one started with. This forces a
Fogel, R. W. 2004. The Escape from Hunger and Premature simple relationship between the three foreign exchange
Death, 1700–2100: Europe, America, and the Third World.
(FX) rates:
Cambridge: Cambridge University Press.
FX$,€ = FX$,� FX�,€ . (1)
Fudenberg, D., and E. Maskin. 1986. The folk theorem
in repeated games with discounting or with incomplete Of course, occasional anomalies and exceptions to
information. Econometrica 54(3):533–54. this relationship can occur, but these will be spotted
Landes, D. 1998. The Wealth and Poverty of Nations. New by traders. The exploitation of the resulting arbitrage
York: W. W. Norton.
opportunity will quickly move the exchange rates until
Ramsey, F. P. 1928. A mathematical theory of saving. Eco-
the opportunity disappears.
nomic Journal 38:543–49.
One can roughly divide the use of mathematics in
Samuelson, P. A. 1947. Foundations of Economic Analysis.
Cambridge, MA: Harvard University Press. finance into four main areas.
. 1958. An exact consumption loan model with or with-
Derivatives pricing. This is the use of mathematics
out the social contrivance of money. Journal of Political
Economy 66:1002–11. to price securities (i.e., financial instruments), whose
value depends purely upon the behavior of another
asset. The simplest example of such a security is a
VII.9 The Mathematics of Money call option, which is the right, but not the obligation,
Mark Joshi to buy a share for a pre-agreed price, K, on some
specified future date. The pre-agreed price is called
1 Introduction the strike. The pricing of derivatives is heavily reliant
upon the principle of no arbitrage.
T&T note: check
The last twenty years have seen an explosive growth Risk analysis and reduction. Any financial institution word spacing
before press.
in the use of mathematics in finance. Mathematics has has holdings and borrowings of assets; it needs to
made its way into finance mainly via the application of keep careful control of how much money it can lose
two principles from economics: market efficiency and from adverse market moves and to reduce these risks
no arbitrage. as necessary to keep within the owners’ desired risk
Market efficiency is the idea that the financial mar- profiles.
kets price every asset correctly. There is no sense in Portfolio optimization. Any investor in the markets
which a share can be a “good buy,” because the mar- will have notions of how much risk he wants to take
ket has already taken all available information into and how much return he wants to generate, and most
account. Instead, the only way that we have of dis- importantly of where he sees the trade-off between
tinguishing between two assets is their differing risk the two. There is, therefore, a theory of how to invest
characteristics. For example, a technology share might in shares in such a way as to maximize the return at
offer a high rate of growth but also a high probability a given level of risk. This theory relies greatly on the
of losing a lot of money, while a U.K. or U.S. govern- principle of market efficiency.
ment bond would offer a much smaller rate of growth, Statistical arbitrage. Crudely put, this is using mathe-
but an extremely low probability of losing money. In matics to predict price movements in the stock mar-
fact, the probability of loss is so small in the latter case ket, or indeed in any other market. Statistical arbi-
 ✐

VII.9. The Mathematics of Money 917

trageurs laugh at the concept of market efficiency, The prices of shares appear to fluctuate randomly,
and their objective is to exploit the inefficiencies in but often with a general upward or downward ten-
the market to make money. dency. It is natural to model them by means of a Brown-
ian motion with an extra “drift term.” This is what Black
Of these four areas, it is derivatives pricing that has and Scholes did, except that it was the logarithm of the
seen the greatest growth in recent years, and which share price S = St that was assumed to follow a Brown-
has seen the most powerful application of advanced ian motion Wt with a drift. This is a natural assumption
mathematics. to make, because changes in prices behave multiplica-
tively rather than additively. (For example, we measure
2 Derivatives Pricing inflation in terms of percentage increases.) They also
assumed the existence of a riskless bond, Bt , grow-
2.1 Black and Scholes ing at a constant rate. To put these assumptions more
formally:
Many of the foundations of mathematical finance were
log S = log S0 + µt + σ Wt , (2)
laid down by Bachelier (1900) in his thesis; his math-
rt
ematical study of brownian motion [IV.24] preceded Bt = B0 e . (3)
that of Einstein (see Einstein (1985), which contains his Notice that the expectation of log S is log S0 + µt, so it
1905 paper). However, his work was neglected for many changes at a rate µ, which is called the drift. The term
years and the great breakthrough in derivatives pricing σ is known as the volatility. The higher the volatility,
was made by Black and Scholes (1973). They showed the greater the influence of the Brownian motion Wt ,
that, under certain reasonable assumptions, it was pos- and the more unpredictable the movements of S. (An
sible to use the principle of no arbitrage to guarantee investor will want a large µ and a small σ ; however,
a unique price for a call option. The pricing of deriva- market efficiency ensures that such shares are rather
tives had ceased to be an economics problem and had rare.) Under additional assumptions such as that there
become a mathematics problem. are no transaction costs, that trading in a share does
The result of Black and Scholes was deduced by ex- not affect its price, and that it is possible to trade con-
tending the principle of no arbitrage to encompass the tinuously, Black and Scholes showed that if there is no
idea that an arbitrage could result not just from static dynamic arbitrage, then at time t, the price of a call
holdings of securities, but also from continuously trad- option, C(S, t), that expires at time T must be equal to
ing them in a dynamic fashion depending upon their BS(S, t, r , σ , T ) = SΦ(d1 ) − Ke−r (T −t) Φ(d2 ), (4)
price movements. It is this principle of no dynamic
with
arbitrage that underpins derivatives pricing.
log(S/K) + (r + σ 2 /2)(T − t)
In order to properly formulate the principle, we have d1 = √ (5)
σ T −t
to use the language of probability theory.
An arbitrage is a trading strategy in a collection of and
assets, the portfolio, such that log(S/K) + (r − σ 2 /2)(T − t)
d2 = √ . (6)
σ T −t
(i) initially the portfolio has a value of zero; Here, Φ(x) denotes the probability that a standard nor-
(ii) the probability that the portfolio will have a nega- mal random variable has value less than x. As x tends
tive value in the future is zero; to ∞, Φ(x) tends to 1, and as x tends to −∞, Φ(x) tends
(iii) the probability that the portfolio will have a posi- to 0. If we let t tend to T , we find that d1 and d2 tend
tive value in the future is greater than zero. to ∞ if ST > K (in which case log(ST /K) > 0) and to −∞
if ST < K. It follows that the price C(S, t) converges to
Note that we do not require the profit to be certain; max(ST − K, 0), which is the value of a call option at
we merely require that it is possible that money may expiry, just as one would expect. We illustrate this in
be made with no risk taken. (Recall that the notion of figure 1.
making money is by comparison with a government There are a number of interesting aspects to this
bond. The same is true of the “value” of a portfolio: it result that go far beyond the formula itself. The first
will be considered positive in the future if its price has and most important result is that the price is unique.
increased by more than that of a government bond.) Using just the hypothesis that it is impossible to make a
 ✐

918 VII. The Influence of Mathematics

50 2.2 Replication

40 The fundamental idea underlying the Black–Scholes

B lack– Scholes price

proof and much of modern derivatives pricing is

30 dynamic replication. Suppose we have a derivative Y
that pays an amount that depends on the value of the
20 share at some set of times t1 < t2 < · · · < tn , and sup-
pose that the payout occurs at a certain time T tn .
10 This can be expressed in terms of a payoff function,
f (t1 , . . . , tn ).
0 The value of Y will vary with the share price. If, in
60 68 76 84 92 100 108 116 124 132 140
addition, we hold just the right number of the shares
Share price
themselves, then a portfolio consisting of Y and the
Figure 1 The Black–Scholes price of a call option struck at shares will be instantaneously immune to changes in
100 for various maturities. The value decreases as maturity
the share price, i.e., its value will have zero rate of
decreases, with the bottom line denoting a maturity of zero.
change with respect to the share price. As the value
of Y will vary with time and share price, we will need to
riskless profit, along with some natural and innocuous continuously buy and sell shares to maintain this neu-
assumptions, we discover that there is only one possi- trality to share-price movements. If we have sold a call
ble price for the option. This is a very strong conclu- option, then it turns out that we will have to buy when
sion. It is not just the case that the option is a bad deal the share price goes up and sell when it goes down;
if traded at a different price: if a call option is bought so these transactions will cost us a certain amount of
for less or sold for more than the Black–Scholes price, money.
then a riskless profit can be made. Black and Scholes’s proof showed that this sum of
A second fact, which may seem rather paradoxical, is money was always the same and that it could be com-
that µ, the drift, does not appear anywhere in the Black– puted. The sum of money is such that by investing it
Scholes formula. This means that the expected behav- in shares and riskless bonds, one can end up with a
ior of the share’s future mean price does not affect the portfolio precisely equal in value to the payoff of Y no
price of the call option; our beliefs about the probabil- matter what the share price did in between.
ity that the option will be used do not affect its price. Thus if one could sell Y for more than this sum of
Instead, it is the volatility of the share price that is money, one would simply carry out the trading strategy
all-important. from their proof and always end up ahead. Similarly, if
As part of their proof, Black and Scholes showed that one can buy Y for less, one does the negative of the
the call option price satisfied a certain partial differ- strategy and always ends up ahead. Both of these are
ential equation (PDE) now known as the Black–Scholes outlawed by the principle of no arbitrage, and a unique
equation, or BS equation for short: price is guaranteed.
∂C ∂C ∂2C The property that the payoff of any derivative can be
+ rS + 12 σ 2 S 2 = r C. (7) replicated is called market completeness.
∂t ∂S ∂S 2
This part of the proof did not rely on the derivative
2.3 Risk-Neutral Pricing
being a call option: there is in fact a large class of
derivatives whose prices satisfy the BS equation, dif- A curious aspect of the Black–Scholes result, mentioned
fering only in boundary conditions. If one changes vari- above, is that the price of a derivative does not depend
ables, setting τ = T − t and X = log S, then the BS equa- upon the drift of the share price. This leads to an alter-
tion becomes the heat equation [I.3 §5.4] with an native approach to derivatives pricing theory called
extra first-order term which can easily be removed. This risk-neutral pricing. An arbitrage can be thought of as
means that the value of an option behaves in a similar the ultimate unfair game: the player can only make
way to time-reversed heat: it diffuses and spreads out money. By contrast, a martingale [IV.24 §4] encapsu-
the farther back one gets from the option’s expiry and lates the notion of a fair game: it is a random process
the more uncertainty there is about the value of the whose expected future value is always equal to its cur-
share at time T . rent value. Clearly, an arbitrage portfolio can never be
 ✐

VII.9. The Mathematics of Money 919

a martingale. So if we can arrange for everything to be of their payoff; this makes them into martingales by
a martingale, there can be no arbitrages, and the price construction.
of derivatives must be free of arbitrage. Everything is now a martingale and there can be no
Unfortunately, this cannot be done because the price arbitrage. Of course, this merely shows that the price
of the riskless bond grows at a constant rate, and is is nonarbitrageable, rather than that it is the only non-
therefore certainly not a martingale. However, we can arbitrageable price. However, work by Harrison and
carry out the idea for discounted prices: that is, for Kreps (1979) and by Harrison and Pliska (1981) shows
prices of assets when they are divided by the price of that if a system of prices is nonarbitrageable, then there
the riskless bond. must be an equivalent martingale measure. Thus the
In the real world, we do not expect discounted prices pricing problem is reduced to classifying the set of
to be martingales. After all, why buy shares if their equivalent martingale measures. Market completeness
mean return is no better than that of a bond that car- corresponds to the pricing measure being unique.
ries no risk? Nevertheless, there is an ingenious way of Risk-neutral evaluation has become such a pervasive
introducing martingales into the analysis: by changing technique that it is now typical to start a pricing prob-
the probability measure [III.73 §2] that one uses. lem by postulating risk-neutral dynamics for assets
If you look back at the definition of arbitrage, you rather than real-world ones.
will see that it depends only on which events have zero We now have two techniques for pricing: the Black–
probability and which have nonzero probability. Thus, Scholes replication approach, and the risk-neutral
it uses the probability measure in a rather incomplete expectation approach. In both cases, the real-world
way. In particular, if we use a different probability mea- drift, µ, of the share price does not matter. Not surpris-
sure for which the sets of measure zero are the same, ingly, a theorem from pure mathematics, the Feynman–
then the set of arbitrage portfolios will not change. Two Kac theorem, joins the two approaches together by stat-
measures with the same sets of measure zero are said ing that certain second-order linear partial differen-
to be equivalent. tial equations can be solved by taking expectations of
A theorem of Girsanov says that if you change the diffusive processes.
drift of a Brownian motion, then the measure that you
derive from it will be equivalent to the measure you 2.4 Beyond Black–Scholes Terri: Tim would
like to keep the
had before. This means that we can change the term µ. heading as it is.
OK?
A good value to choose turns out to be µ = r − 12 σ 2 . For a number of reasons, the theory outlined above is
With this value of µ, one has not the end of the story. There is considerable evidence
that the log of the share price does not follow a Brown-
E(S/Bt ) = S/B0 (8) ian motion with drift. In particular, market crashes
for any t, and since we can take any time as our start- occur. For example, in October 1987 the stock market
ing point, it follows that S/Bt is a martingale. (The extra fell by 30% in one day and financial institutions found
− 12 σ 2 in the drift comes from the concavity of the coor- that their replication strategies failed badly. Mathemat-
dinate change to log-space.) This means that the expec- ically, a crash corresponds to a jump in the share price,
tation has been taken in such a way that shares do and Brownian motion has the property that all paths
not carry any greater return, on average, than bonds. are continuous. Thus the Black–Scholes model failed to
Normally, as we have mentioned, one would expect an capture an important feature of share-price evolution.
investor to demand a greater return from a risky share A reflection of this failure is that options on the same
than from a bond. (An investor who does not demand share but with differing strike prices often trade with
such compensation is said to be risk neutral.) However, different volatilities, despite the fact that the BS model
now that we are measuring expectations differently, we suggests that all options should trade with the same
have managed to build an equivalent model in which volatility. The graph of volatility as a function of the
this is no longer the case. strike price is normally in the shape of a smile, dis-
This yields a way of finding arbitrage-free prices. playing the disbelief of traders in the Black–Scholes
First, pick a measure in which the discounted price pro- model.
cesses of all the fundamental instruments, e.g., shares Another deficiency of the model is that it assumes
and bonds, are martingales. Second, set the discounted that the volatility is constant. In practice, market activ-
price process of derivatives to be the expectations ity varies in intensity and goes through some periods
 ✐

920 VII. The Influence of Mathematics

when share prices are much more volatile and others methods are most effective for low-dimensional prob-
when they are much less so. Models must therefore be lems, and so research is devoted to trying to make them
corrected to take account of the stochasticity of volatil- effective in a wider range of cases.
ity, and the prediction of volatility over the life of an One method that is less affected by dimensionality is
option is an important part of its pricing. Such models Monte Carlo evaluation. The basis of this method is very
are called stochastic volatility models. simple: both intuitively and (via the law of large num-
If one examines the data on small-scale share move- bers) mathematically, an expectation is the long-run
ments, one quickly discovers that they do not resem- average of a series of independent samples of a random
ble a diffusion. They appear to be more like a series variable X. This immediately yields a numerical method
of small jumps than a Brownian motion. However, if for estimating E(f (X)). One simply takes many inde-
one rescales time so that it is based on the number pendent samples Xi of X, calculates f (Xi ) for each one
of trades that have occurred rather than on calendar and computes their average. It follows from the cen-
time, then the returns do become approximately nor- tral limit theorem [III.73 §5] that the error after N
mal. One way to generalize the Black–Scholes model is draws is approximately distributed as a normal distri-
to introduce a second process that expresses trading bution with variance equal to N −1/2 times the variance
time. An example of such a model is known as the vari- of f (X). The rate of convergence is therefore dimen-
ance gamma model. More generally, the theory of Lévy sion independent. If the variance of f (X) is large, it
processes has been applied to develop wider theories may still be rather slow, however. Much effort is there-
of price movements for shares and other assets. fore devoted by financial mathematicians to developing
Most generalizations of the Black–Scholes model do methods of reducing the variance when one computes
not retain the property of market completeness. They high-dimensional integrals.
therefore give rise to many prices for options rather
than just one. 2.6 Vanilla versus Exotics

Generally, a simple option to buy or sell an asset is

2.5 Exotic Options
known as a vanilla option, whereas a more complicated
Many derivatives have quite complicated rules to deter- derivative is known as an exotic option. An essential
mine their payoffs. For example, a barrier option can be difference between the pricing of the two is that one
exercised only if the share price does not go below a cer- can hedge an exotic option not just with the underlying
tain level at any time during the contract’s life, and an share, but also by trading appropriately in the vanilla
Asian option pays a sum that depends on the average options on that share. Typically, the price of a deriva-
of the share price over certain dates rather than on the tive will depend not just on observable inputs, such as
price at expiry. Or the derivative might depend upon the share price and interest rates, but also on unob-
several assets at once, such as, for example, the right servable parameters, such as the volatility of the share
to buy or sell a basket of shares for a certain price. price or the frequency of market crashes, which cannot
It is easy to write down expressions for the value of be measured but only estimated.
such derivatives in the Black–Scholes model, either via When trading exotic options, one wishes to reduce
a PDE or as a risk-neutral expectation. It is not so easy to dependence upon these unobservable inputs. A stan-
evaluate these expressions. Much research is therefore dard way to do this is to trade vanilla options in such
devoted to developing efficient methods of pricing such a way as to make the rate of change of the value of the
options. In certain cases it is possible to develop ana- portfolio with respect to such parameters equal to zero.
lytic expressions. However, these tend to be the excep- A small misestimation of their value will then have little
tion rather than the rule, and this means that one must effect on the worth of the portfolio.
resort to numerical techniques. This means that when one prices exotic options,
There is a wealth of methods for solving PDEs and one wishes not just to capture the dynamics of the
these can be applied to derivatives-pricing problems. underlying asset accurately but also to price all the
One difficulty in mathematical finance, however, is that vanilla options on that asset correctly. In addition, the
the PDE can be very high dimensional. For example, if model will predict how the prices of vanilla options
one is trying to evaluate a credit product depending change when the share price changes. We want these
on 100 assets, the PDE could be 100 dimensional. PDE predictions to be accurate.
 ✐

VII.9. The Mathematics of Money 921

The BS model takes volatility to be constant. How- 3.2 Value-at-Risk

ever, one can modify it so that the volatility varies with
the share price and over time. One can choose how it One method of limiting an institution’s risks in deriva-
varies in such a way that the model matches the market tives trading is to place a limit on the amount it can
prices of all vanilla options. Such models are known as lose with a given probability over a specified period of
local volatility models or Dupire models. Local volatility time. For instance, one might consider the losses at a
models were very popular for a while, but have become 1% level over ten days, or at a 5% level over one day.
less so because they give a poor model for how the This value is called Value-at-Risk or VAR.
prices of vanilla options change over time. To compute VAR one has to build up a probabilistic
model of how the portfolio of derivatives might change
Much of the impetus behind the development of
in value over the time period. This requires a model
the models we mentioned in section 2.4 comes from
of how all the underlying assets can move. Given this
the desire to produce a model that is computationally
model, one then builds up the distribution of possible
tractable, prices all vanilla options correctly, and pro-
profits and losses over the given time period. Once one
duces realistic dynamics for both the underlying assets
has this distribution one simply reads off the desired
and the vanilla options. This problem has still not been
percentile.
wholly solved. There tends to be a trade-off between
The issues involved in modeling the changes for VAR
realistic dynamics and perfect matching of the vanilla
computation are quite different from those for deriva-
options market. One compromise is to fit the market
tives pricing. Typically, a VAR computation is done over
as well as possible using a realistic model and then
a very short time period, such as one or ten days, unlike
to superimpose a local volatility model to remove the
the pricing of an option, which deals with a long time
remaining errors.
frame. Also, one is not interested in the typical path for
VAR, but instead one focuses on the extreme moves.
3 Risk Management In addition, since it is the VAR of an entire portfolio
that matters, one has to develop an accurate model
3.1 Introduction of the underlying assets’ joint distributions: the move-
ment of one underlying asset could magnify the price
Once we have accepted that it is impossible to make movement of another, or it could act as a hedge.
money in finance without taking risk, it becomes impor- There are two main approaches to developing a prob-
tant to be able to measure and quantify risks. We wish abilistic model for computing VAR. The first, the histor-
to measure accurately how much risk we are taking and ical approach, is to record all the daily changes over
decide whether we are comfortable with that level of some time period, for example two years, and then
risk. For a given level of risk, we want to maximize our assume that the set of changes tomorrow will be identi-
expected return. When considering a new transaction, cal to one of the sets of changes we have recorded. If we
we will want to examine how it affects our risk lev- assign equal probability to each of those changes, then
els and returns. Certain transactions may even reduce we get an approximation to the profit and loss distribu-
our risk while increasing our returns if they cancel out tion, from which we can read off the desired percentile.
other risk. (A risk that can be canceled out by other Note that as we are using a day’s change for all assets
risks that have a tendency to move in the opposite simultaneously, we automatically get an approximation
direction is called diversifiable.) to the joint distribution of all the asset prices.
The control of risk becomes particularly important A second approach is to assume that asset price
when dealing with portfolios of derivatives, which are movements come from some well-known class of dis-
often of zero value initially but which can very quickly tributions. For example, we could assume that the logs
change value. Placing a limit on the value of the con- of the asset price movements are jointly normal. We
tracts held is therefore not of much use, and controls would then use historical data to estimate the volatil-
based on deal sizes are complicated by the fact that ities and the correlations between the various prices.
often many derivatives contracts largely cancel each The main difficulty with this approach is obtaining
other out; it is the residual risk that one wishes to robust estimates of the correlations given a limited
control. amount of data.
 ✐

922 VII. The Influence of Mathematics

4 Portfolio Optimization A consequence of the model is the distinction be-

tween diversifiable risk and undiversifiable risk. While
4.1 Introduction
investors are compensated for taking undiversifiable,
The job of a fund manager is to maximize the return or systematic, risk via higher expected returns, diver-
on the money invested while minimizing the risk. If we sifiable risk does not carry a risk premium. This is
assume that markets are efficient, then there is no point because one can cancel out diversifiable risk by holding
in trying to pick shares that we believe to be underval- appropriate combinations of other assets. Therefore, if
ued as we have assumed that they do not exist. A corol- it carried a risk premium, investors could receive extra
lary is that just as no shares are good buys, no shares return without taking any risk.
are bad buys. In any case, over half the shares in the Much of the current research in this area is directed
market are owned via funds and therefore under the at trying to find more accurate models for the joint dis-
control of fund managers. Therefore, the average fund tribution of returns, and at finding techniques that esti-
manager cannot expect to outperform the market. mate the parameters of such returns. A related problem
It may seem that this does not leave much for fund is the “equity premium puzzle,” which is that the excess
managers to do, but in fact it leaves two things. return on investing in shares is much higher than the
model predicts for reasonable levels of risk aversion.
(i) They can attempt to control the amount of risk
they are taking. 5 Statistical Arbitrage
(ii) For a given level of risk, they can maximize their
We only briefly mention statistical arbitrage as it is a
expected return.
rapidly changing area that is shrouded in secrecy. The
To do these things requires an accurate model of the fundamental idea in this area is to squeeze informa-
joint distribution of asset prices over the longer term, tion out of asset price movements that the market has
and a quantifiable notion of risk. not already acted on. It therefore contradicts the prin-
ciple of market efficiency, which says that all available
4.2 The Capital Asset Pricing Model information is already encoded in the market price.
One explanation is that it is the action of taking such
Portfolio theory has been in its modern form for
arbitrages that makes the market efficient.
longer than derivatives pricing. As an area, it relies
less on stochastic calculus and more on economics. Further Reading
We briefly review the key ideas. The best-known model
Bachelier, L. 1900. La Théorie de la Spéculation. Paris:
for modeling portfolio returns is the capital asset pric-
Gauthier-Villars.
ing model (or CAPM), which was introduced in the Black, F., and M. Scholes. 1973. The valuation of options
1950s by Sharpe (see Sharpe 1964), and is still ubiqui- and corporate liabilities. Journal of Political Economy 81:
tous. Sharpe’s model built on earlier work of Markowitz 637–54.
(1952). Einstein, A. 1985. Investigations on the Theory of the Brown-
The fundamental problem in this area is to assess ian Movement. New York: Dover.
what portfolio of assets, generally shares, an investor Harrison, J. M., and D. M. Kreps. 1979. Martingales and
arbitrage in multi-period securities markets. Journal of
should hold in order to maximize returns at a given
Economic Theory 20:381–408.
level of risk. The theory requires assumptions to be Harrison, J. M., and S. R. Pliska. 1981. Martingales and
made about the joint distribution of share returns, e.g., stochastic integration in the theory of continuous trading.
joint normality, and/or about the risk preferences of Stochastic Processes and Applications 11:215–60.
investors, e.g., that they only care about the mean and Markowitz, H. 1952. Portfolio selection. Journal of Finance
variance of returns. 7:77–99.
Under these assumptions, the CAPM yields the result Sharpe, W. 1964. Capital asset prices: a theory of market
equilibrium under conditions of risk. Journal of Finance
that every investor should hold a multiple of the “mar-
19:425–42.
ket portfolio,” which is essentially a portfolio consist-
ing of everything traded in appropriate quantities to
achieve maximum diversification, together with a cer- VII.10 Mathematical Statistics
tain amount of the risk-free asset. The relative amounts Persi Diaconis
are determined by the investor’s risk preferences.