10.1098/rspa.2004.

1386

R EVIEW PAPER

A survey of mathematical finance

By David H o b s o n
Department of Mathematical Sciences, University of Bath,
Bath BA2 7AY, UK ([email protected])
and Department of Operations Research and Financial Engineering,
Princeton University, Princeton, NJ 08540, USA

Received 5 January 2004; accepted 2 August 2004; published online 5 October 2004

Finance is one of the fastest growing areas in mathematics. In some senses it is not
a discipline in its own right, but rather an application area in which mathematicians
with backgrounds in probability theory, statistics, optimal control, convex and func-
tional analysis and partial differential equations can bring to bear experiences and
results from their own fields to problems of real world interest.
In this survey we begin with the simplest possible financial model, and then give
an account of the Black–Scholes option pricing formula, in which the key ideas are
the replication of option pay-offs and pricing under the risk-neutral measure. Then
we move on to discuss other important problems in finance, including the general
theory for semi-martingale price processes, pricing in incomplete markets, interest-
rate models and credit risk. The emphasis is on techniques and methodologies from
stochastic processes.
Keywords: derivative pricing; Black–Scholes; incomplete markets;
stochastic calculus; martingale measures

1. Preamble
Despite the comparatively recent origins of the subject, mathematical finance is one
of the most important application areas of mathematics today. Three decades ago
the subject barely registered as a research area, but when in the early 1970s Fisher
Black, Myron Scholes and Robert Merton linked the well-developed notions of Brow-
nian motion and Itô calculus to the problems of derivative pricing and hedging, a
new and vibrant discipline was created. The celebrated Black–Scholes option-pricing
formula (the discovery and development of which earned Nobel prizes in 1997 for
Scholes and Merton, Black having died a couple of years previously) revolution-
ized the finance industry, facilitating the subsequent rapid expansion in the trading
of financial derivatives. The growth in volume of trading of these instruments has
been matched by the growth of mathematical finance as a research endeavour. This
has helped create new topics for mathematical inquiry, reinvigorating many exist-
ing areas, and developing bridges between previously unconnected subjects. Now
many mathematics departments in the United Kingdom and throughout the world

Proc. R. Soc. Lond. A (2004) 460, 3369–3401

c 2004 The Royal Society
3369
3370 D. Hobson

are developing research and teaching programmes in finance, and the output of these
programmes, both in terms of the research and the graduates, provides an important
resource for the City of London and elsewhere.
Mathematician’s Brownian motion was first introduced by Bachelier (1900), who
was motivated by an attempt to model the fluctuations of asset prices and to price
derivatives. Although he was the first researcher to characterize Brownian motion
and his work was well known to Kolmogorov and Doob, the impact of his work was
not recognized by the finance community for many years. (His name is, however, hon-
oured by the main international mathematical finance society.) Indeed it was much
later that Samuelson (1965) suggested using exponential Brownian motion to model
stock prices. In the exponential Brownian model the proportional price changes are
generated by a Brownian motion. Over a small time-interval the proportional price
changes are Gaussian random variables with a variance proportional to the length of
the interval, and price changes over disjoint intervals are uncorrelated. The exponen-
tial Brownian model reflects the limited liability (non-negativity) property of share
prices and, while it is not appropriate for all financial assets in all market conditions,
it remains the reference model against which any alternative dynamics are judged.
It was in a model with exponential Brownian assets that Black & Scholes (1973)
constructed a replicating portfolio and with it proposed a ‘fair’ price for a financial
derivative. (A derivative security or contingent claim is a financial instrument whose
pay-off is derived from, or contingent upon, the behaviour of some other underlying
asset. For example, a call option on a stock or share gives the option holder the
right, but not the obligation, to purchase one unit of the stock at a pre-specified
price called the strike.) Their ideas were quickly advanced by Merton (1973). The
key insight was that if it was possible to replicate the pay-off of the derivative as
the gains from trade from a dynamic, self-financing hedging strategy, then the initial
fortune required to finance that strategy was exactly the arbitrage-free price for the
option. Furthermore, since all the risks associated with the option were removed by
hedging, the price is independent of the risk preferences of the agent.
This argument was developed into a mathematical theory by Harrison & Kreps
(1979) and Harrison & Pliska (1981). These authors emphasized the central role
of probability theory and martingales (a martingale is a random process which is
as likely to go up as it is down, on average) and it is their stochastic theory that
we explain here, and which provides the foundation for much of the subsequent
development of the subject. Their key conclusion is that option prices are given
by expectations—but not expectations with respect to the real world or physical
measure. Instead prices are expectations with respect to the risk neutral measure
under which the discounted price of the underlying asset is a martingale.
In this survey we concentrate on the problems of derivative pricing. We begin with
an analysis of option pricing in the simplest possible one-period binomial model, the
conclusions from which—including the fact that there is a unique, preference inde-
pendent, fair option price—are subsequently mirrored in the Black–Scholes world.
We then investigate the extent to which the Black–Scholes model can be generalized
without destroying these key features.
When all options can be priced via replication, the model is complete. Otherwise
the model is incomplete. In this situation there is no universal scheme for pricing
options. Instead we compare and contrast some of the possible alternatives, and this
topic is the main theme of the article. In particular we discuss in some simple but

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3371

canonical settings how options can be priced and hedged under various investment
criteria.
No survey of mathematical finance can cover all areas of the subject in equal depth,
and any summary inevitably reflects the background and interests of the author. The
fact that this article stresses stochastic methods for derivative pricing in complete
and incomplete markets is a case in point. In the final few sections we cover, briefly,
some of the other important topics in finance, including interest-rate models and
credit risk.

2. Derivative pricing: a first pass

Consider the following model of a financial market. There is a single risky asset whose
price is given by (Xt )0
t
T and a risk-less bank account. The market in these assets
is perfect, by which we mean that there are no transaction costs or taxes, the risky
and risk-less assets can be bought in arbitrary quantities and agents are price takers.
A derivative security, or contingent claim, is a financial security whose value is
contingent upon the value of the risky asset. For example, a call option (with strike
K and maturity T ) gives the holder the right, but not the obligation, to buy one
unit of the risky asset at time T for price K. If XT > K then the option holder can
exercise this right, and (perhaps by selling the asset) make a profit of (XT − K),
whereas if XT
K, the option matures worthless. At maturity the call option is
worth (XT − K)+ .
The fundamental problem in mathematical finance is to give a fair price for the
random pay-off of a derivative security given a stochastic model for the behaviour of
the underlying.

(a) The simplest case: the binomial model

Suppose X0 = x and that, at time T , XT takes one of the values xu or xd where
u > d. (More formally we let Ω = {ωu , ωd } and define XT (ωu ) = xu, XT (ωd ) = xd
and we suppose 0 < P({ωu }) < 1.) There is also a bank account which pays a fixed
and constant rate of interest r over the period [0, T ] so that one unit invested in the
account at time 0 is worth R = (1 + r) at time T . We assume R ∈ (d, u) to prevent
simple arbitrages.
The problem is to price a derivative security which pays off hu = h(xu) in a year
when the price has moved ‘up’, and hd = h(xd) otherwise.
Suppose we can find θ, φ which solve
hu = θxu + φR, (2.1)
hd = θxd + φR. (2.2)
Then the agent is indifferent between receiving the derivative and holding an initial
portfolio of θ units of risky asset and investing φ units in the bank. Hence the time-
zero fair value for the option is C = θx + φ, the cost of financing the strategy
implicit in the right-hand-sides of (2.1) and (2.2). This is our first example of pricing
by arbitrage; if the derivative trades at any price other than C, then there are risk-
free profits to be made, either by selling the derivative and purchasing the portfolio
(θ, φ) or by following the reverse strategy. Since this cannot happen in any sensible
market—there would be infinite demand for the derivative if it traded for a price

Proc. R. Soc. Lond. A (2004)

3372 D. Hobson

below C, and infinite supply if it traded above C—the derivative must trade for the
arbitrage-free price C.
In this simple binary model the values of θ and φ can be calculated from (2.1) and
(2.2). We find θ = (hu − hd )/(x(u − d)) and φ = (uhd − dhu )/(R(u − d)), so that an
expression for the derivative price is


1 R−d u−R
hu + hd . (2.3)
R u−d u−d
There are two key observations to be made in this simple model which will inspire
our future analysis.
The first is that the key to option pricing is the concept of replication: the fact
that the fair price is determined by a trading strategy which creates the same pay-off
as the option. In the binomial model it is always possible to find θ and φ to solve
(2.1) and (2.2), so that replication is possible for all contingent claim pay-offs h.
The second key observation relates to the concept of martingale pricing. If we write
q = (R − d)/(u − d), then q ∈ (0, 1) and the derivative price (2.3) can be written as
1 1
{qhu + (1 − q)hd } = Eq [h(XT )],
R R
so that the option price is the discounted expected pay-off of the option, where
the expectation is taken with respect to the risk-neutral probabilities (q, 1 − q). The
probability q has the special property that the expected value of the discounted asset
price under the probabilities (q, 1 − q) is the initial value; i.e. q satisfies
1
x= (qxu + (1 − q)xd).
R
The discounted asset price is a martingale if we take expectations using the q proba-
bilities. Note that we have completed a full analysis of the problem without reference
to the probabilities of the various events under the real-world measure P.
Rather than focusing on the measure or probabilities, we can consider instead the
state-price density. Let p = P({ωu }) and define ζ via ζ0 = 1 and
q 1 (R − d)
ζT (ωu ) = = ,
pR pR (u − d)
(1 − q) 1 (u − R)
ζT (ωd ) = = .
(1 − p)R (1 − p)R (u − d)
Then (ζt Xt )t=0,T is a martingale, and the fair price of the option is E[ζT h(XT )].
The above model, which is essentially due to Cox et al . (1979), can be made more
realistic by extending it to cover several time-steps. (Indeed, since a random walk
converges to Brownian motion, the suitably scaled limit will be the continuous-time
model of the next section.) The contingent claim pricing problem can be solved by
backward induction and the derivative price is precisely the discounted expected
pay-off where the probabilities have been modified to make the discounted prices of
traded assets into martingales.
Note that if it is possible for the risky asset to take on more than two price values
at the end of the time-step, then the replication argument fails. For example, in a
trinomial model in which XT may take the values xu, xR, xd say, then the analogue

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3373

to (2.1) and (2.2) is a triple of simultaneous equations in two unknowns for which
there is no solution in general. Conversely, there are many choices of probabilities
which make the price process into a martingale.

(b) The Black–Scholes model: pricing and hedging

We now consider the derivative pricing problem in continuous time. Following
Samuelson (1965), the model is based on a Brownian motion or a Wiener process
Wt . The stochastic process Wt does not have finite variation and so the standard
rules of calculus do not apply. Instead we use stochastic calculus. For a very brief
introduction to the key concepts, see the appendix, or one of the many introductory
(Mikosch 1998; Steele 2001) or more specialist texts (Revuz & Yor 1998; Rogers &
Williams 2000).
We suppose that we have a perfect frictionless model (as before, zero transaction
costs, zero taxes and dividends, the same interest rate for both borrowing and lend-
ing, agents as price takers) in which trading takes place in continuous time. The
economy consists of a single risky asset with price process (Xt )0
t
T which follows
an exponential Brownian motion, and a bank account which pays a constant rate of
interest, r. The dynamics for the risky asset are specified under the physical measure
P and are exogenous to the model. This reflects the fact that agents are taken to
be small investors, and their actions do not affect the market price. The risky asset
price and the value of R0 units of currency invested in the bank account are given
by
Xt = X0 exp{σWt + (ν − 12 σ 2 )t}, Rt = R0 exp{rt},
or, in differential notation (using Itô’s formula (A 2))
dXt = Xt (σ dWt + ν dt), dRt = rRt dt. (2.4)
Here the parameters σ > 0, ν and r (respectively the volatility and drift of the risky
asset and the interest rate) are taken to be constants. The value of monies invested
in the bank account, Rt , obeys standard Newtonian calculus and the ordinary dif-
ferential equation for Rt in (2.4) might more usually be written dR(t)/dt = rR(t).
We use the form dRt = rRt dt as an analogy to a stochastic differential equation,
and to remind us that in a more complicated model the interest rate may itself be
stochastic. We call the asset with price Rt a bond.
Our goal, as in the binomial model, is to consider the wealth process which results
from holding a portfolio consisting of θt units of the risky asset and φt units of the
bond. The elements of the portfolio θt and φt must be chosen based on information
available at time t. We assume this information set or filtration is generated by the
price process Xt , which means in our current context that it is the Brownian filtration
generated by Wt . The value of the portfolio is then given by
Vt = θt Xt + φt Rt . (2.5)
We further assume that the dynamics of the portfolio value satisfy
 t  t
V t = V0 + θs dXs + φs dRs (2.6)
0 0
or, in differential notation,
dVt = θt dXt + φt dRt . (2.7)

Proc. R. Soc. Lond. A (2004)

3374 D. Hobson

It should be emphasized that (2.7) is not obtained by taking the Itô derivative of the
products in (2.5). Instead it is postulated as a modelling assumption, motivated by
the situation in discrete time. See the remarks in § 3 for a further discussion of this
issue.
A value process Vt which satisfies (2.7) is said to be self-financing. The term
captures the idea that no inputs or outputs of cash are needed to create Vt ; instead
all fluctuations in value come from the investment in the risky asset and bond.
Further, if Vt solves (2.5), then, once θt has been chosen, φt is determined via the
relationship φt Rt = Vt − θt Xt . In particular, we do not need to model φ explicitly;
φt merely represents the number of bonds we can buy with the cash surplus after
we purchase θt units of Xt . Sometimes we write V θ to stress the dependence of the
self-financing value process on the strategy θ, or V v,θ if we also wish to stress the
starting wealth. It follows that we can rewrite (2.7) as
dVtθ = θt (dXt − rXt dt) + rVtθ dt, (2.8)
which, given the stochastic dynamics of Xt is equivalent to
dVtθ = θt Xt σ(dWt + λ dt) + rVtθ dt, (2.9)
where λ = (ν − r)/σ is the Sharpe ratio of the risky asset. It turns out to be much
more convenient to work with the Sharpe ratio λ rather than the drift ν, so that ν
will not be mentioned again.
Consider now the problem of pricing a contingent claim with non-negative pay-off
h(XT ) at time T .
Define a super-replicating strategy to be a pair (v, θ) such that the wealth process
V v,θ , defined via V0v,θ = v and V v,θ , solves (2.8), satisfies Vtv,θ  0 and VTv,θ  h(XT ),
P-almost surely. A replicating strategy has VTv,θ = h(XT ). The key idea is that, if
there exists a super-replicating strategy for initial wealth v, then an agent would
be at least as happy to receive initial fortune v and to follow trading strategy θ, as
to receive the option. Hence the no-arbitrage principle gives us that v is an upper
bound on the fair price of the claim.
Consider X̃t = R0 Xt /Rt . We will use a superposed tilde to denote a discounted
quantity. We have
   
R0 X t R0 Xt dXt
dX̃t = d = − r dt ,
Rt Rt Xt
which in our case can be simplified to
dX̃t = X̃t σ(dWt + λ dt). (2.10)
Now consider the discounted process = Ṽtθ R0 Vtθ /Rt . If V is self-financing, then V θ
θ

solves (2.8) and, in terms of discounted quantities,

dṼtθ = θt dX̃t , (2.11)
or equivalently, dṼtθ
= θt X̃t σ(dWt + λ dt). The simplicity of this equation shows the
advantage we gain from switching to discounted variables. Now suppose V v,θ is the
value process associated with a replicating strategy (v, θ). Then
 T
R0 R0 v,θ
h(XT ) = VT = ṼTv,θ = v + θt dX̃t (2.12)
RT RT 0
P-almost surely.

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3375

Suppose, for a brief moment, that λ = 0 and X̃t is a martingale. Then we can take
T
expectations in (2.12) and, provided that 0 θt dX̃t is a true martingale and not just
a local martingale, we can deduce a value for v. This value represents the replication
price for the contingent claim.
Now remove the assumption that λ = 0, so that the discounted price is not a mar-
tingale. Suppose, however, that we can find a new probability measure Q, equivalent
to P, such that the stochastic integral in (2.12) is a martingale under Q. Then the
identities in (2.12) hold Q-almost surely and taking expectations under Q we have
the formula  
Q R0
v=E h(XT ) . (2.13)
RT
This gives us the fair price of the option. The measure Q is a computational device,
but it is extremely powerful in that it leads us to the option price.
Motivated by the above analysis, our goal is to find a measure Q under which the
price process is a martingale, or to use a language more familiar to economists, to
find a state-price density process ζt such that ζt Xt is a martingale.
Define the change of measure density Zt via
Zt = exp(−λWt − 12 λ2 t)
and let Q and ζt be given by
R0
Q(A) = E[ZT IA ] and ζt = Zt . (2.14)
Rt
The probability measure Q is then equivalent to P and, by the Cameron–Martin–
Girsanov formula (see the appendix), W Q defined via WtQ = Wt + λt is a Q-Brownian
motion. Hence (see (2.10)), dX̃t = σ X̃t dWtQ and X̃ is a Q-martingale. Alternatively,

d(ζt Xt ) = d(Zt X̃t ) = (σ − λ)(ζt Xt ) dWt

so that ζt Xt is a P-martingale. The above result is a example of the simple proposition
that for any process Yt , we have that Ỹt is a (local) martingale under Q if and only
if ζt Yt is a (local) martingale under P.
Now suppose that V v,θ is the value process of a super-replicating strategy for
h(XT ). Then, from (2.11), Ṽ v,θ is a local Q-martingale. Furthermore, V v,θ , and
hence Ṽ v,θ , is non-negative, and we conclude that Ṽ is a Q-supermartingale. Thus

v  EQ [ṼTv,θ ]  EQ [R0 h(XT )/RT ] = E[ζT h(XT )].

In particular E[ζT h(XT )] is a lower bound on the fair price of the derivative.
If E[ζT h(XT )] = ∞, then there is no super-replicating strategy corresponding to
a finite initial price. Henceforth we exclude this case.
Now we want to show that there is a super-replicating strategy with initial fortune
v = E[ζT h(XT )]. Define the martingale
 
R0
Π̃t = EQ
t h(X T ) ,
RT
where Et denotes expectation given information available at time t. Observe that
Π̃t  0, and Π̃T = R0 h(XT )/RT Q-almost surely (and hence P-almost surely, since

Proc. R. Soc. Lond. A (2004)

3376 D. Hobson

P and Q are equivalent). By the Brownian martingale representation theorem (recall

that the filtration is generated by Wt ) we can write any Q-martingale Π̃t as a stochas-
tic integral with respect to the Q-Brownian motion W Q . We have
 t  t
Π̃t = v + ψs dWsQ = v + θsΠ dX̃s , (2.15)
0 0

where θtΠ = ψt /σ X̃t and dX̃s = X̃s σ dWsQ . Then Πt defined via Πt = Rt Π̃t /R0 sat-
isfies Π0 = v, Πt  0 and ΠT = h(XT ), P-almost surely, with dynamics
dΠt = θtΠ (dXt − rXt dt) + rΠt dt.
Hence Πt defines the value process of a self-financing, super-replicating (and indeed
replicating) strategy with initial value v = EQ [R0 h(XT )/RT ] and it follows that v is
the fair price for the derivative. The associated hedging strategy is given by θtΠ .
Note that, in exact parallel with the binomial model, the key ideas are the repli-
cation of the option pay-off and the idea of finding a change of measure under which
the discounted price process is a martingale. That measure is then used for pricing.
The Sharpe ratio λ in the original model is irrelevant for pricing (as is the drift), and
instead volatility σ is the crucial parameter. The fact that we price the option by
replication means that an agent who sells the option for its fair price can remove all
the risk via a hedging strategy. This explains why the risk preferences of the agent
do not enter into the pricing formula.
To date we have identified the fair price of the option, but not the replicating
strategy θtΠ . To do this in general we need to know how to represent a martingale as
a stochastic integral in a Brownian filtration. This can be done by Clark’s theorem,
which is a special case of Malliavan calculus. Alternatively, for pay-offs which are
a function of XT alone (or perhaps a function of XT and a small number of other
path-dependent state variables—see the examples below) we can exploit the Markov
property to give an explicit form for the hedging strategy θ.
Suppose the option pay-off depends only on the value of the underlying asset
at time T . By the Markov property we can represent the time-t value Vt of the
contingent claim via  
Q Rt
Vt = V (Xt , t) = Et h(XT ) . (2.16)
RT
Recall that dXt = σXt dWtQ + rXt dt. Then, by Itô’s formula, assuming that V is
sufficiently smooth,
dVt = V
(Xt , t) dXt + 12 V

(Xt , t)(dXt )2 + V̇ (Xt , t) dt
= V
(Xt , t)Xt σ dW Q + [V
(Xt , t)rXt + 12 V

(Xt , t)σ 2 Xt2 + V̇ (Xt , t)] dt.
Conversely, if V is self-financing, then from (2.9)
dVt = θt Xt σ dWtQ + Vt r dt.
If V is the value function of a self-financing replicating portfolio, then these represen-
tations must be almost surely identical, and for (almost every) path realization we
must have θt = V
(Xt , t) (for Lebesgue almost surely all t ∈ [0, T ]). Further, when
we equate finite variation terms we find that the value function must solve
LV = 0 subject to V (x, T ) = h(x), (2.17)

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3377

where
Lf (x, t) = rxf
(x, t) + 12 σ 2 x2 f

(x, t) + f˙(x, t) − rf (x, t). (2.18)
The partial differential equation (PDE) (2.17) for V can be shown to be equivalent
to the stochastic pricing formula (2.16) using the Feynman–Kac formula and is some-
times called the Black–Scholes pricing PDE. The hedging strategy θt = V
(Xt , t) is
known as the delta hedge.

(c) Vanilla and exotic options

In the setting of the Samuelson–Black–Scholes exponential Brownian motion model
for option pricing we have shown that it is possible to derive a unique fair price for
contingent claims. The key mathematical tools that we used were Itô’s formula, the
Cameron–Martin–Girsanov change of measure and the Brownian martingale repre-
sentation theorem. In later sections we discuss in more detail the class of admissible
trading strategies and the extent to which the conclusions of the above analysis are
robust to changes in the underlying model. We also consider the impact that the
failure of the model assumptions has on hedging and pricing. However, in the rest of
this section we assume that the model holds and investigate the implications for the
pricing of some common traded options.
The advantage of working with a simple model, albeit an overly simplistic one, is
that it gives insights into the behaviour of derivative prices which might be hidden
in a more realistic situation. For example, it allows us to investigate the comparative
statics of the option price and to understand how prices depend on the key parameters
such as volatility (Bergman et al . 1996; Renault & Touzi 1997; Hobson 1998). The
true test of a model is partly how well it explains option prices in the market (but,
as Figlewski (2002) argues, one does not need the full power of the Black–Scholes
call pricing function for that), and partly how well the theoretical hedges perform.

(i) Call options

Traditionally the first, simplest and most widely traded options are put and call
options. A call option with maturity T and strike K has pay-off (XT − K)+ . The
time-t price of the call option is
V (Xt , t) = e−r(T −t) EQ
t [(XT − K) ] = Xt Φ(d+ ) − Ke
+ −r(T −t)
Φ(d− ),
where Φ is the cumulative normal distribution function and
ln(Xt /Ke−r(T −t) ) ± σ 2 (T − t)/2
d± = √ .
σ T −t
The delta-hedging strategy is given by θt = Φ(d+ ).
A put option gives the holder the right to sell the risky asset for price K. Since
(XT − K) = (XT − K)+ − (K − XT )+ , there is a put–call parity result, namely that
the price of a call option minus that of a put option equals Xt − e−r(T −t) K.

(ii) American options

If a claim is European in style, then it is exercised at a fixed predetermined time T .
American style options can be exercised at any (stopping) time τ up to the, possibly

Proc. R. Soc. Lond. A (2004)

3378 D. Hobson

infinite, maturity T . The price becomes (see Myneni 1992)

ess sup EQ [e−rτ h(Xτ )],
τ
T

where the ess sup is taken over all stopping times τ with t
τ
T . If
h(x) = (x − K)+ (an American call), then provided there are no dividends it is never
optimal to exercise the option early and the American call has the same price as a
European call. However, for an American put option with h(x) = (K − x)+ , the
benefits of the convexity of the pay-off can sometimes be outweighed by the losses
associated with the fact that the undiscounted prices increase on average over time
and the pay-off function is decreasing. The pricing problem becomes an optimal
stopping problem in which the optimal exercise strategy has to be determined.
One fruitful approach to this problem is to consider it as a dynamic programming
problem. The martingale optimality principle allows us to write down a Hamilton–
Jacobi–Bellman (HJB) equation. The pricing function solves V (x, t)  h(x) and
LV = 0 on It = {x : V (x, t) > h(x)}, where, as before,

Lf = 12 σ 2 x2 f

+ rxf
− rf + f˙,
together with a smooth fit condition on ∂It . This is a free boundary problem for
which there is no closed-form solution. It is related to the Stefan problem from fluid
dynamics (Friedman 2000).
The natural explanation for the European/American nomenclature would be that
options of appropriate style were traded in the relevant geographical markets. How-
ever, there is no strong evidence for this proposition. (Instead there is an anecdote
which claims that the adjectives were coined by an American researcher who wanted
to appropriate the more sophisticated and challenging option for his own continent.)
Whatever the origins of the terminology, it began a trend for naming options after
regions or countries—Asia, Bermuda, Paris, Russia and Israel each have an option
named after them.
Puts and calls have simple pay-offs and are sometimes called vanilla options in
honour of the most basic flavour of ice cream. Options with more complicated pay-
offs are said to be exotic.

(iii) Barrier options

An example of an exotic option is an option whose pay-off is contingent upon both
the value of the underlying at maturity and the value of the maximum price attained
by the underlying over some period. For example, a knock-out call option has pay-off
(XT − K)+ I{X̄T
B} , where X̄T is the maximum price attained by the underlying and
B is the barrier level. The option becomes worthless if ever the underlying exceeds the
barrier. In the Black–Scholes model there are closed-form expressions for the prices
and associated hedging strategies for barrier options which involve the cumulative
normal distribution function.
In practice, barrier options can be difficult instruments to hedge. The classical
delta hedge can involve very large positions, especially when the underlying asset is
near the barrier and the time to maturity is small. In these cases practical issues
tend to dominate—for example, it can be useful to hedge using the call as well as the
underlying (see Andersen & Andreasen 2000; Brown et al . 2001)—and an alternative

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3379

pricing rule and hedging strategy is needed, perhaps aiming to super-replicate the
pay-off rather than aiming to replicate exactly.
Barrier options are closely related to digital and lookback options. A digital option
pays one if ever the underlying crosses the barrier, while the pay-off of a lookback is
contingent upon the maximum price attained by the underlying over the lifetime of
the option. In the Black–Scholes model there are formulae for all of these (see, for
example, Goldman et al . 1979).

(iv) Asian options

T
An Asian fixed-strike call has pay-off (AT − K)+ , where AT = (1/T ) 0 Xu du. (Of
course this is an idealized mathematical version of the real contract, which is based
upon discrete averaging.) Asian options are options on the average rate and were
introduced partly to meet the need for commodity producers who sold their output
at a constant rate over time, and partly to negate the effects of price manipulation.
The Asian pricing problem is to calculate the distribution of
 T
2
ÂT = eσWs +(r−σ /2)s ds
0

in such a way that it is possible to give a simple representation formula for the price of
the Asian call. In general there are no closed-form solutions but the pricing problem
motivated several attempts to give a stochastic characterization of the distribution
(see Geman & Yor 1993), as well as various ideas for the pricing of Asian options
via Monte Carlo methods (with carefully chosen control variates (see Rogers & Shi
1995) or PDEs (Večer 2001)).

(v) Passport options

The passport option (introduced by Hyer et al . (1997)) is an example of an exotic
option which was not widely traded, but which generated some novel research prob-
lems in mathematics. In the symmetric passport option problem the aim is to evaluate
sup EQ [(G̃θt )+ ],
|θt |
1

where G̃θ is the discounted gains from trade using a self-financing strategy θ. In
particular
 T
θ
G̃t = g + θs dX̃s .
0
It turns out (see, for example, Andersen et al . 1998) that the optimal strategy is to
take θs = − sgn(G̃s ). Moreover, the price is related via the Skorokhod problem and
local times to that of a lookback option (Henderson & Hobson 2000; Delbaen & Yor
2002).

(d ) Numéraires
We saw in the analysis of the Black–Scholes model that it is convenient to work
with discounted prices. This switch can be described as a change of numéraire from
cash to bond, and the fundamental and very sound economic principle upon which

Proc. R. Soc. Lond. A (2004)

3380 D. Hobson

it is based is that the prices of contingent claims should not depend on the units in
which they are denominated.
As well as cash and bond, it is sometimes useful to use a risky asset, or the
gains from trade of a portfolio of risky assets as numéraire (see Geman et al . 1995;
Gourieroux et al . 1998). For example, consider pricing an exchange option (Margrabe
1978) with pay-off (XT − YT )+ , where the price processes Xt and Yt are given by
correlated Brownian motions. Then a change of numéraire from cash to Yt reduces
the pricing problem to that of pricing a standard call in the Black–Scholes model on
the single underlying Xt /Yt .
In general the form of a martingale measure Q depends on the choice of numéraire
N (see Branger 2004), and for clarity one should consider the pair (NT , QN ). Alter-
natively, we can fix attention on the state-price density
N0 dQN
ζT = ,
NT dP
which is numéraire independent.

(e) Optimal consumption and investment problems

Consider an agent who can trade in a market as in § 2 b. Suppose that, rather
than trying to price a derivative, the aim of this agent is to maximize his utility of
terminal wealth, or alternatively to maximize his utility of consumption over time.
Let U : R+ (or R) → R be an increasing (to reflect the fact that agents prefer more
to less) and concave (to reflect the law of diminishing marginal returns) utility func-
tion. Examples include power-law utilities U (x) = x1−R /(1 − R), for R > 0, logarith-
mic utility U (x) = ln x and exponential utility U (x) = −e−x , together with various
other less tractable families such as
U (x) = κ−1 (1 + κx − 1 + κ2 x2 ), κ > 0.
The classical Merton problem (Merton 1969) is to find the optimal trading strategy
which maximizes the expected utility of terminal wealth EU (VT ), where VT is given
by (2.6). In the Black–Scholes model there is a full solution to this optimal control
problem. In the primal approach it is possible to write down an HJB equation for the
value function of an agent, and then, at least for the case of power law, logarithmic
and exponential utilities, to conjecture the form of the solution. In simple cases a
standard verification theorem gives the result that indeed we have a solution of the
HJB equation, and the optimal strategy. (In less simple cases the solution of the HJB
equation may only exist in the sense of a viscosity solution (see Duffie et al . 1997).)
There is an alternative approach, called the dual method, which gives very powerful
insights (see Karatzas (1989) for a survey). The problem is to maximize the expected
utility of terminal wealth VT subject to the wealth satisfying a budget constraint. If
we write this in Lagrangian form
E[U (VT ) − µ(ζT VT − v)]
and introduce the Legendre transform Û (y) = supv {U (v) − vy} of U , then
E[U (VT ) − µ(ζT VT − x)]
µv + EÛ (µζT ), (2.19)

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3381

with equality when U
(VT ) = µζT almost surely. This inequality holds for all admis-
sible strategies, and all (positive) Lagrange multipliers so we have

sup E[U (VT )]
inf {µv + EÛ (µζT )}. (2.20)

VT µ

Further, in standard cases (when the asymptotic elasticity of utility is less than one
(Kramkov & Schachermayer 1999)), there is no duality gap and there is equality
between the expressions in (2.20). The optimal solution given by a target wealth VT∗
and a Lagrange multiplier µ∗ is such that VT∗ = I(µ∗ ζT ), where I is the inverse to
the derivative of U . (In fact µ∗ is the value of the Lagrange multiplier such that
E[ζT VT∗ ] = E[ζT I(µ∗ ζT )] = v.)
In the analysis of the Merton problem for the Black–Scholes model presented here,
the dual problem is simpler than the primal problem, since the minimization takes
place over a single real-valued Lagrange multiplier rather than a random-variable
valued space of terminal wealths. If we think of the dual problem, then it is natural
to look for utilities whose Legendre transform Û takes a simple form. For example,
consider the class of dual functions given by Û

(y) = Ay q−2 for q ∈ R and A a
positive constant. The class of associated utility functions is exactly the class of
hyperbolic absolute risk aversion utilities, which includes the power, logarithmic and
exponential utilities as special cases (see Merton 1990, p. 137).
Instead of aiming to maximize expected utility of terminal wealth it is also natural
to consider agents who wish to maximize expected discounted utility of consumption
over time. Let the wealth process be described by the equation
dVt = θt dXt + (r − θt Xt ) dt − ct dt,
where ct is the consumption rate. Then the problem facing the agent is to maximize
 ∞ 
E U (t, ct ) dt , (2.21)
0

or more especially to determine optimal investment and consumption pairs (θt , ct )t0 .
Again this problem can be attacked via primal or dual methods.
It should be noted that (2.21) is an unsatisfactory formulation in a couple of ways.
Firstly, (2.21) does not arise as the continuous time limit of a realistic situation
in which consumption occurs in discrete lumps and, secondly, the value function
depends only on the marginal distributions of the consumption process (ct )t0 , and
not on the joint distribution. Duffie & Epstein (1992) introduced stochastic differen-
tial utilities to address this second issue.

(f ) The successes and failures of the Black–Scholes model

The Black–Scholes model has the property that it is possible to define a unique
fair price, the replication price, for any contingent claim. This price is given as the
discounted expectation of the option pay-off under the unique risk-neutral or mar-
tingale measure. The model can be extended to include dividends and to other types
of underlyings, such as forwards, futures, indices and foreign exchange rates. Above
all, the Black–Scholes model has provided a language for the pricing of derivatives
and a reference against which modifications of the model can be compared.

Proc. R. Soc. Lond. A (2004)

3382 D. Hobson

In principle, in the Black–Scholes paradigm the option-pricing problem is solved,

and the solution given in (2.13), but on occasion it may be difficult to evaluate
this stochastic expression and give an analytic pricing formula. Instead practitioners
sometimes resort to solving the PDE (2.17), or approximate the price via Monte
Carlo simulation or even solve a multi-period extension of the Cox–Ross–Rubinstein
binomial model. In such cases the issue is to execute any of these approaches effi-
ciently and accurately, particularly in high dimensions.
Unfortunately, the assumptions of the Black–Scholes model are never satisfied, a
theme we return to in § 4. (It is clear that something must be wrong since the traded
prices of different derivatives are frequently, by which we mean invariably, consistent
with different values of the volatility parameter.) Continuous trading is impossible,
there are taxes, interest rates differ for borrowing and lending, agents are never price
takers and face a bid/ask spread, and the prices of underlyings never quite follow
exponential Brownian motion with constant known parameters. Understanding and
accommodating some of these market frictions and imperfections is one of the main
remaining goals of mathematical finance and one of the subjects of the remaining
sections.

3. The general theory

Our aim in this section is to review the analysis we gave in the Samuelson–Black–
Scholes exponential Brownian case and to consider the extent to which the results
and conclusions generalize to a wider class of models. At first sight it might appear
that such generalizations are issues of idle mathematical curiosity. In fact, since
the assumptions of the Black–Scholes model clearly fail in practice, it is crucial to
understand which results are robust to model mis-specification. Our brief survey is
based on the discussion in Schachermayer (2003), and the reader who wishes to learn
more about the background to the ‘théorie générale’ is referred to that very readable
text.
We begin with a filtered probability space (Ω, F, (Ft )0
t
T∞ , P), where T∞ is a
fixed horizon time which is greater than the maturity of any options of interest.
We suppose that the discounted price process of the risky asset X̃t is a (locally
bounded, ‘continue à droite, limite à gauche’ (cádlág)) semi-martingale which is
adapted to the filtration Ft satisfying the usual conditions. The filtration Ft captures
the information available at time t. The process X̃t may be vector valued, although
our notation will not emphasize this. We have chosen to work with discounted price
processes (in part this is just a choice of numéraire), so that although there is a bank
account in the model, it does not appear in the analysis.
Already the perceptive reader may be wondering whether it is necessary to assume
that the price X̃t of the risky asset is a semi-martingale. This assumption is very
convenient because the well-developed theory of stochastic integration is based upon
semi-martingales. Further, according to theorem 7.2 of Delbaen & Schachermayer
(1994), if the model is to be consistent with no arbitrage, then the price process
must be a semi-martingale, at least when the set of admissible trading strategies is
sufficiently large. This rules out certain candidate families of models very quickly.
For example, fractional Brownian motion is not a semi-martingale. Rogers (1997)
gives a direct proof that fractional Brownian motion models admit arbitrage. On

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3383

the other hand, we cannot take models which are too simple: if the discounted price
process is of finite variation then there is also arbitrage.
Our first task is to define the class of admissible portfolios and the associated value
functions. Let θt be an adapted process which represents the purchases of the risky
asset and define Ṽ θ , the associated self-financing value process with initial wealth
Ṽ0 , via  t
Ṽtθ = Ṽ0 + θs dX̃s . (3.1)
0
As before, the investment φ in the bank account is implicit rather than explicit.
The integral on the right-hand-side of (3.1) is an Itô stochastic integral. In one
sense the choice of the Itô integral is arbitrary—we could equally use the Stratanovich
integral, for example, provided we include all the appropriate correction terms. But
in another sense the Itô stochastic integral is the only stochastic integral which makes
economic sense. To see this observe that, if the portfolio θt is a simple (piecewise
constant) strategy, then the discounted gains from trade from investment in the risky
asset are given by
G̃t = θti (X̃ti+1 ∧t − X̃ti ∧t ).
i
In particular the gains process is obtained by multiplying the increments of the
price process by the number of units of risky asset held at the beginning of the
relevant time-interval. The Itô integral shares this non-anticipatory property—it is
the integral of the integrand against the forward increments of the integrator.
We now define an admissible strategy as an adapted portfolio process θt for which
the associated value function is such that the Itô stochastic integral
 T
θt dX̃t
0

is well defined and defined via (3.1) is bounded below: Ṽtθ  −M for some con-
Ṽtθ
stant M . This definition is sufficient to rule out doubling strategies, but does not
prevent suicide strategies.
The key idea which underpinned pricing in the Black–Scholes model was the notion
of an equivalent martingale measure. In general it is too much to expect the under-
lying to become a martingale under a change of measure, and all we really need is
that the discounted traded asset process, and hence the discounted wealth process,
becomes a local martingale. We have the following tautological but important def-
inition: a measure Q, equivalent to P, under which the discounted asset price is a
local martingale is called an equivalent local martingale measure.
Before we discuss option pricing in general we would like to know whether the
model we have makes economic sense, and in particular whether it is consistent with
no arbitrage. (If there are arbitrage opportunities in the model— loosely described
to be ways of making profits at zero risk—then the model is unsustainable. Some or
indeed all agents would want to follow these profit making strategies and the current
market prices would not survive in equilibrium.) It turns out that the ‘right’ concept
to work with is the idea of ‘no free lunch with vanishing risk’ (NFLVR). Roughly
speaking there is a free lunch with vanishing risk if, when you look at the class of
contingent claims which can be replicated by an admissible portfolio, and then look
at the limits of sequences of such claims, there is a limit random variable which

Proc. R. Soc. Lond. A (2004)

3384 D. Hobson

is non-negative almost surely and positive with positive probability. The key result
is due to Delbaen & Schachermayer (1994, corollary 1.2), but see also Harrison &
Pliska (1981) for the finite case, and also Kreps (1981) and Delbaen & Schachermayer
(1998).
Theorem 3.1 (first fundamental theorem of asset pricing). Suppose X̃ is
a locally bounded semi-martingale. Then there exists an equivalent local martingale
measure if and only if the model satisfies NFLVR.
This theorem is one of the triumphs of the theory of mathematical finance in
the abstract semi-martingale setting. It was clear that one side of the if-and-only-
if condition should be the existence of an equivalent (local) martingale measure,
since this is a powerful assumption from which many natural and useful properties
follow easily. Thus the difficult part of the theorem involved finding the appropriate
definitions of admissible strategy and no arbitrage which would give the martingale
measure condition an economically meaningful interpretation.
Since we want to work with economically meaningful models, we assume that
the model satisfies NFLVR. Hence, we are entitled to assume that there exists an
equivalent local martingale measure. Set ZT = dQ/dP and Zt = Et [ZT ]. Then Zt
and Zt X̃t are both P-local martingales.
In the general setting we say that a pair (v, θ) is a super-replicating strategy for
H if the strategy is admissible and if the associated value process Ṽ v,θ satisfies
(3.1) subject to Ṽ0θ = v and ṼTv,θ  H̃, the discounted pay-off of the claim. Then, by
the same analysis as before, if (v, θ) is a super-replicating strategy, then Zt Ṽtθ is a
P-supermartingale and
v  E[ZT ṼTθ ]  E[ZT H̃].
Hence E[ZT H̃] is a lower bound on the replication price of the option.
This raises the question as to whether there is a super-replicating strategy for
the option with initial wealth v. In the one-dimensional Brownian context we have
seen how the Brownian martingale representation theorem can be used to produce
a replicating strategy. In general it is not always the case that this is possible. The
condition under which replicating strategies can be found for all options can again
be related to a condition on the equivalent martingale measures, and is again given
in Delbaen & Schachermayer (1994).
Theorem 3.2 (second fundamental theorem of asset pricing). Every
bounded claim can be replicated if and only if there is only one equivalent local
martingale measure.
This is the subject of the next section.

4. Incomplete markets
Our analysis of the Samuelson–Black–Scholes model relied on two results from the
theory of stochastic processes and Brownian motion. Firstly, the Cameron–Martin–
Girsanov theorem guarantees the existence of an equivalent martingale measure Q
under which the discounted price process Xt is a martingale (or, equivalently, the
existence of a state-price density ζt with the property that ζt Rt and ζt Xt are mar-
tingales). Secondly, the Brownian martingale representation theorem says that any
random variable whose value is known at time T can be written as its expected value

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3385

plus a stochastic integral against Brownian motion. In the Black–Scholes market set-
ting this translates into the result that any option pay-off can be written as the price
plus the gains from trade from a dynamic investment strategy in the underlying
asset.
In the previous section we saw that the existence of a martingale measure is related
to the question of whether a model makes economic sense. In this section we consider
the role of the martingale representation theorem, and especially the situation in
which it is no longer possible to write every claim as the terminal value of a trading
strategy.
Recall that RT , which we no longer assume to be deterministic, is the value of
R0 units of cash invested in the bank account. We say that a contingent claim H is
replicable if it can be written
  T 
RT
H= v+ θs dX̃s
R0 0
for an admissible trading strategy θ, or equivalently if the option pay-off can be repli-
cated via a dynamic hedging strategy. In this case there is a unique fair replication
price for the option  
Q R0
v=E H = E[ζT H],
RT
where Q is any martingale measure and ζT is the related state-price density. An
option which can be replicated in this way is said to be redundant in the sense that
adding the option to the (perfect frictionless) economy has no impact since its pay-off
can be created synthetically through dynamic trading. If every claim is redundant,
then the market is complete.
In an incomplete market it is not possible to replicate every contingent claim.
For such claims there is no replication price, and the Black–Scholes theory we have
introduced has nothing to say about the fair price of the option. Instead we have
reached what Hakansson (1979) calls the ‘catch 22 of option pricing’: the claims we
can price are redundant, and the claims that are not redundant we cannot price.
The problem facing economists (and financial mathematicians) is to determine a
method for pricing non-redundant options which is consistent with the Black–Scholes
methodology for those derivatives that can be replicated.
It is clear that, if there is more than one state-price density, then there exists a
claim for which it is possible to define more than one price (via expectation) and
hence that that option cannot be replicated. The converse is also true, so that, if
there exists a unique equivalent local martingale measure, then the model is complete
and every claim can be replicated. This is the second fundamental theorem of asset
pricing.
Incompleteness can arise from many sources, for example, transaction costs
(Hodges & Neuberger 1989; Davis et al . 1993; Soner et al . 1995), jump models
(Merton 1976; Bardhan & Chao 1996), constraints on the trading strategies (Soner
& Touzi 2001; Cvitanić & Karatzas 1993) or stochastic volatility (Hull & White 1987;
Heston 1993; Fouque et al . 2000) and to some extent the best approach to pricing
and hedging must depend on the context. However, fundamentally, one has to answer
the question of how to price and hedge a contingent claim H which is completely
independent of the remainder of the model. Our goal is to analyse two simple models
which exhibit incompleteness.

Proc. R. Soc. Lond. A (2004)

3386 D. Hobson

(a) Non-traded assets

As a first and simple example of an incomplete market (see Davis 2000; Henderson
& Hobson 2002a, b; Henderson 2002) consider an economy with a deterministic bond
Rt = R0 ert and a single risky asset with dynamics
dXt
= σ(dWt + λ dt) + r dt. (4.1)
Xt
For simplicity we assume that all parameters σ, λ and r are constants. All contingent
claims on X can be replicated. Now introduce a second risky asset Yt with price
process
dYt = at dWt
+ bt dt, (4.2)
where W
is correlated to W with dWt dWt
= ρ dt. Suppose that Y is not traded
and consider the problem of pricing a contingent claim H = H(YT ).
The situation we are trying to model is one where an agent has a random endow-
ment H whose pay-off depends on an asset Y , but that asset cannot be used for
hedging. This may be because of legal reasons (consider an executive who receives
compensation in the form of stock options, but who is contractually forbidden from
actively trading in stock on his own company (Henderson 2003) or simply liquidity
issues (trading in the asset Y may be so thin as to make hedging with Y impractical).
However, the agent can use the correlated asset X for hedging.
The Black–Scholes theory tells us that for pricing purposes we should switch to a
martingale measure under which the discounted prices of traded assets are martin-
gales, but it does not tell us how to determine the drifts on non-traded assets.

(b) Stochastic volatility models

Consider a market consisting of a bond paying constant rate of interest r and a
single risky asset with price process Xt . Suppose that under the physical measure P
the dynamics of the risky asset are given by
dXt
= σ(Yt , t)(dWt + λ(Yt , t) dt) + r dt, (4.3)
Xt
where the process driving the volatility is an autonomous diffusion process
dYt = a(Yt , t) dWt
+ b(Yt , t) dt, (4.4)
where W
is correlated to the Brownian motion W . The problem is to price an option
with pay-off H = H(XT ).
Stochastic volatility models were introduced to model the empirical fact that his-
torical time-series for volatility reveals patterns which indicate that volatility changes
randomly over time. Examples include modelling the volatility σ(Yt , t) as a shifted
Ornstein–Uhlenbeck process (Stein & Stein 1991), a square-root or Cox–Ingersoll–
Ross process (Hull & White 1988; Heston 1993) and an exponential Brownian motion
(Hull & White 1987). There are also jump models for Y ; see, for example, the model
popularized by Barndorff-Nielsen & Shephard (2000).
Which model of stochastic volatility should one choose? A good model should
be tractable, realistic (for example, a shifted Ornstein–Uhlenbeck process can go

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3387

negative which is an undesirable property) and it should be straightforward to esti-

mate the parameters. Moreover, as well as providing a fit to historic price data, the
model should also have the ability to explain option-price smiles both over strike and
over maturity. Finally, the model should give superior hedging performance to the
Black–Scholes model.

(c) Incomplete markets and martingale measures

It is clear from the form of the models in both the non-traded asset and the
stochastic volatility cases that these models are incomplete. In a frictionless diffusion
model, the rule of thumb is that a model is incomplete if the number of sources of
randomness is greater than the number of traded assets.
We begin by describing the space of equivalent martingale measures. It is conve-
nient to introduce a Brownian motion B, which is independent of W and such that
Wt
= ρWt + ρ̄Bt , where ρ̄2 = 1 − ρ2 . Define
  t  t  t  t 
1 1
Zt = exp − λu dWu − λ2u du − ξu dBu − ξu2 du . (4.5)
0 2 0 0 2 0

Provided that E[ZT ] = 1 we can define a (local) martingale measure Q via a process
similar to (2.14) (see Frey 1997). (The first moment condition guarantees that Q is
a probability measure). Then ζt = e−rt Zt is a state-price density and ζt Xt is a P
(local) martingale. Under Q,
 t  t
WtQ = Wt + λu du and BtQ = Bt + ξu du
0 0
are Brownian motions.
t Note that the change of drift on Wt is enforced by the require-
ment that Wt + 0 λu du is a martingale, whereas the change of drift on Bt is unde-
termined. The classξ of changes of measure is parametrized by the process ξ, and we
write Qξ and (W Q , B Q ) ≡ (W ξ , B ξ ) to emphasize this.
ξ

It remains to check that Qξ is equivalent to P, and hence that there exists an

equivalent (local) martingale measure and thus there is no arbitrage. The task of
checking that a general stochastic exponential such as (4.5) is a true martingale is
a difficult one (the Novikov condition rarely applies), but in the Markovian setting
other approaches have recently been developed (see Hobson & Rogers 1998; Wong
& Heyde 2004) which reduce to checking that certain processes are non-explosive.
It remains to decide if the model is complete. By the (multidimensional) Brownian
martingale representation theorem, given the measure Qξ , the discounted option
pay-off R0 HT /RT can be written as a stochastic integral with respect to the two-
dimensional Brownian motion (Wtξ , Btξ ):
 T  T
R0 ξ ξ
HT = v + ψt dWt + χξt dBtξ
RT 0 0
 T ξ  T
ψt
=v+ dX̃t + χ̄ξt dBtξ .
0 X̃σ(Yt , t) 0

The first two terms correspond to the initial wealth and discounted gains from trade,
respectively, of a dynamic hedging strategy involving investments in the traded asset
and bank account. However, it is not possible to trade on the second asset and in
general the claim cannot be replicated.

Proc. R. Soc. Lond. A (2004)

3388 D. Hobson

5. Option pricing in incomplete markets

In a complete market the fair prices of options are uniquely determined by the
replication price. These prices can be calculated as the discounted expected values
under the equivalent martingale measure. In an incomplete market there is no unique
fair price and no universal pricing algorithm. Instead there are several alternative
methodologies which have been proposed as pricing mechanisms.
The first approach is to finesse the problem by writing down the dynamics of assets
under a pricing measure. This approach bypasses the physical measure. A second and
related idea (see, for example, Heston 1993) is to choose (essentially arbitrarily) a
market price of risk for the non-traded assets. For example, the Föllmer & Schweizer
(1990) minimal martingale measure corresponds to a choice of a zero market price
of risk for the non-traded Brownian motion, or equivalently in our setting ξ = 0.
Another idea which has sometimes been exploited in the stochastic volatility liter-
ature (see Scott 1987) is to assume that there is a call option which is liquidly traded.
The introduction of a second traded asset completes the market. Hence, given the
traded price of a call option it is possible to price and hedge any other contingent
claim. Of course, this approach does not explain the price of the original traded call.
This idea has been extended by Dupire (1994) to create an elegant (though not very
robust) theory for the pricing of exotic options. Suppose that calls with all possible
maturities and strikes are traded on the market. Then, under the assumption that
the price process possesses the Markov property, it is possible to infer the dynamics
of the underlying process. In this approach prices for vanilla options are taken from
the market and then used to give prices for path-dependent exotic options. For a
more robust version of the idea see Brown et al . (2001).
The remaining approaches we shall discuss all acknowledge the incompleteness of
the market and price options accordingly. Respectively, they involve pricing via a
hedging criteria, super-replication pricing, minimal distance martingale measures,
convex risk measures and utility indifference pricing.

(a) Hedging criteria

In an incomplete market, perfect hedging is impossible. Instead one might aim
to minimize some functional of the hedging error. Föllmer & Sondermann (1986)
suggest minimizing
E[(H − VTv,θ )2 ]

over initial wealths v and trading strategies θ. The resulting optimal values are
the mean–variance price and hedge, respectively. It turns out that in markets with
(2) (2)
zero interest rates v = E[HζT ], where ζT is the variance-optimal state-price den-
sity which is independent of the choice of derivative H (see Schweizer 1996). For
extensions of this idea see Gourieroux et al . (1998) on stochastic interest rates and
Grandits & Krawczyk (1998) on Lp norms on the hedging error.
An alternative criterion is proposed by Föllmer & Leukert (2000). They propose
minimizing the shortfall E[(H − VTv,θ )+ ]. This overcomes the disadvantage of the
quadratic hedging condition which penalizes super-replication, but at the cost of
tractability.

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3389

(b) Super-replication pricing

In the discussion on the complete market we introduced the idea of super-
replication. In an incomplete market we can use the same notion to define the super-
replication price as the smallest initial fortune which is needed to super-hedge the
option pay-off with probability one. The super-replication price can be thought of as
an extreme hedging criteria in which the agent is not willing to accept any risk.
The super-replication price is the supremum of the possible prices which are con-
sistent with no arbitrage. As such, it often gives a price which is unrealistically high.
In the non-traded-assets model the super-replication price of a call option on Y is
infinite (Hubalek & Schachermayer 2001), while in a stochastic volatility model the
super-replication price of a call on X is the cost of buying one unit of the underlying
(Frey & Sin 1999).
A key alternative characterization of the super-replication price is given in
El Karoui & Quenez (1995) (see also Delbaen & Schachermayer 1994; Föllmer &
Kramkov 1997; Föllmer & Kabanov 1998), as

sup EQ [R0 H/RT ],

Q

where the supremum is taken over the set of martingale measures. Thus the super-
replication price is the price under the worst case martingale measure.

(c) Minimal distance martingale measures

Rather than choosing a state-price density arbitrarily, one approach is to choose
the state-price density which is smallest in an appropriate sense. Given a convex
function f : R+ → R the problem is to minimize E[f (ζT )] over choices of state-price
density. When interest rates are deterministic and f is homogeneous, this minimiza-
tion problem is equivalent to finding the minimal distance martingale measure, the
(local) martingale measure Q which minimizes
E[f (ZT )], (5.1)
where ZT = dQ/dP. (Some care is needed in this minimization procedure as the
optimizing element may not itself belong to the class of equivalent martingale mea-
sures.)
As we pointed out earlier the class of martingale measures depends on the choice of
numéraire. However, since incomplete markets involve unhedgeable risks, the choice
of almost any pricing criterion involves a decision about the units to be used to
measure these risks. It seems most natural to use cash for this purpose. Alterna-
tively, if we minimize E[f (ζT )], then the problem is numéraire independent, and this
is another argument for focusing on the state-price density. To date, however, the
mathematical literature has concentrated on the problem of minimizing (5.1). In any
case, for the examples we consider, interest rates are deterministic and there is no
distinction between the problems of determining the minimal distance state-price
density and the minimal distance martingale measure for a cash numéraire.
The problem of finding minimal distance measures has been studied by many
authors, but see especially Goll & Rüschendorf (2001), who give various characteri-
zations which determine the optimal Q in terms of f . One minimal distance measure

Proc. R. Soc. Lond. A (2004)

3390 D. Hobson

which has been the subject of particular attention in the literature (for example,
Rouge & El Karoui 2000; Frittelli 2000) is the minimal entropy martingale measure.
Consider now our canonical models of incomplete markets. Suppose, following
Hobson (2004), that we have a representation of the mean–variance trade-off process
of the form
 T  T  T  T
1 1
λ2 du = c + ηu (dWu + λu ) du + χu dBu + χ2u du. (5.2)
2 0 0 0 2 0

Note that this is an identification of random variables and not of processes, and that
the solution consists of a constant c and integrands η and ξ. This equation can be
viewed as an example of a backward stochastic differential equation (BSDE) (see
Mania et al . 2003). BSDEs provide a general framework for many characterization
problems in finance (El Karoui et al . 1997).
Now consider f (z) = z ln z, and E[f (ZTξ )] for martingale measures ZTξ given by
(4.5). We have
  ξ 
ξ ξ Qξ dQ
E[ZT ln ZT ] = E ln
dP
and, using the representation (4.5),
 ξ  T  T  T  T
dQ 1 1
ln =− ξ
λu dWu + λ du −
2 ξ
ξu dBu + χ2 du
dP 0 2 0 0 2 0
 T  T  T
1
=c+ (ηu − λu ) dWuξ + (χu − ξu ) dBuξ + (χu − ξu )2 du,
0 0 2 0
(5.3)
where we have used (5.2) and the fact that, under Qξ , W ξ given by dWtξ = dWt +λt dt
and B ξ given by dBtξ = dBt + ξt dt are Brownian motions. Then, assuming that the
stochastic integrals in (5.3) are true martingales we have
 T 
ξ ξ 1 Qξ
E[ZT ln ZT ] = c + 2 E (χu − ξu ) du  c,
2
0

with equality for ξ = χ. Hence the problem of finding the minimal entropy martingale
measure reduces to finding the solution of (5.2). More generally, (5.2) is the special
case, corresponding to q = 1, of a more general formula which covers distance metrics
of the form f (x) = xq /(q(q − 1)).
In the non-traded assets model described in § 4 a the left-hand side of (5.2) is
constant, and there is a trivial solution corresponding to η ≡ 0 ≡ χ. (In this case
all the minimal distance measures are identical and equal to the Föllmer–Schweizer
minimal martingale measure.) Alternatively, in the stochastic volatility model, if ρ is
constant and Y is an autonomous diffusion, then there is a stochastic representation
of the solution to (5.2) given in Hobson (2004).
Once a minimal distance martingale measure Q∗ has been identified, it can be
used for pricing in the sense that we can define the option price to be
 
Q∗ R0
E H = E[ζT∗ H],
RT

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3391

where ζT∗ is the state-price density associated with the pricing measure Q∗ . The
resulting prices are linear in the number of units of claim sold, and as we shall see
later they are related to the marginal price of the claim for a utility maximizing
agent. Further, if we can solve the analogue of (5.2) for a variety of q, then we can
begin to compare option prices under different martingale measures (see Henderson
et al . 2003).

(d ) Convex risk measures

Coherent risk measures were introduced by Artzner et al . (1999), in an attempt to
axiomatize measures of risk (and also to prove that Value-at-Risk was ‘incoherent’).
In order to be consistent with the rest of this section we talk about coherent pricing
measures for claims rather than measures of risks.
Let H ∈ H be a contingent claim. Then φ : H → R is a coherent pricing measure
if it has the properties:

subadditivity φ(H1 + H2 )
φ(H1 ) + φ(H2 ),

positive homogeneity for λ  0, φ(λH) = λφ(H),
monotonicity H1
H2 =⇒ φ(H1 )
φ(H2 ),
translation invariance φ(H + m) = φ(H) + m.

The idea is that φ represents the amount of compensation which an agent would
demand in order to agree to sell the claim H (or the size of the reserves he should
hold if he has outstanding obligations amounting to H). The key result of Artzner
et al . (1999) is that, at least for finite sample spaces, there is a representation of a
coherent pricing measure of the form

φ(H) = sup EQ [H],

Q∈Q

where Q is a set of measures. For example, the super-replication price is obtained by

taking the set Q to be the set of all martingale measures.
Subsequently, Föllmer & Schied (2002) introduced the notion of a convex risk
measure. Convex risk measures attempt to model situations in which the ask price
of a claim depends on the number of units sold. The subadditivity and positive
homogeneity properties are replaced by a convexity property; for µ ∈ [0, 1],

φ(µH1 + (1 − µ)H2 )
µφ(H1 ) + (1 − µ)φ(H2 ).

Convex pricing measures are associated with a pricing mechanism which is nonlinear
in the number of units of the claim. Again there is a representation of a convex
pricing measure of the form

φ(H) = sup {EQ [H] − α(Q)},

Q∈P

where now P is the set of all probability measures, and α is a penalty function.
For example, to recover the super-replication price we may take α(Q) = 0 if Q is a
martingale measure, and α(Q) = ∞ otherwise.

Proc. R. Soc. Lond. A (2004)

3392 D. Hobson

(e) Utility indifference pricing

Utility indifferent option prices (Hodges & Neuberger 1989) can be considered as
a dynamic version of the notion of a certainty equivalent price in economics. The
utility indifference (ask) price is the unique price p at which the agent is indifferent
(in the sense that his expected utility under optimal trading is unchanged) between
not selling the claim and receiving p now in return for agreeing to make the random
pay-out H at time T .
Consider the problem with k units of the claim. (We take k to be positive if the
agent is buying units of claim, and k negative if the agent is short the contingent
claim.) Assume that initially the agent has wealth v and zero endowment of the
claim. Define
u(v, k) = sup E[U (VT + kHT )],
where the supremum is taken over attainable terminal wealths which satisfy the
budget constraint E[ζT VT ]
v for all state-price densities ζT . The utility indifference
price p(k) is then the solution to
u(v, 0) = u(v − p(k), k).
Note that, if the claim can replicated, then p(k) = kE[ζT H] for any state-price
density ζT .
In order to solve for the utility indifference price we need to solve the agent’s utility
maximization problem both with and without the claim. In the absence of the claim,
the problem is the classical Merton problem in an incomplete market. By analogy
with (2.20) we have an inequality, which holds for all state-price densities, of the
form
sup E[U (VT )]
inf inf {µv + EÛ (µζT )}, (5.4)
VT µ ζT

where Û is the Legendre transform of U . There is quality in (5.4) (subject to reg-

ularity conditions) if U
(VT0 ) = µ0 ζT0 for some optimal target wealth VT0 , Lagrange
multiplier µ0 and state-price density ζT0 (the superscript zero corresponds to zero
units of the claim). Note that if Û is a power law, then ζT0 corresponds to a minimal
distance state-price density.
In the case with the option (see Cvitanić et al . 2001), we have
U (VT + kH) − µ(ζT VT − v) = U (VT + kH) − µζT (VT + kH) + µv + µkζT H

Û (µζT ) + µ(v + kζT H)
and then
u(v, k) = sup E[U (VT + kH)]
inf inf E[Û (µζT ) + µ(v + kζT H)]. (5.5)
VT µ ζT

It follows that, if µ0 and ζT0 are as above,

u(v − kE[ζT0 H], k)
E[Û (µ0 ζT0 ) + µ0 (v − kE[ζT0 H]) + kµ0 ζT0 H]
= u(v, 0)
= u(v − p(k), k),
and the bid price for k units satisfies p(k)
kE[ζT0 H].

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3393

With further work, and under further assumptions (see Henderson & Hobson
2002a; Hobson 2003; Hugonnier et al . 2004) it is possible to show that for positive
claims
p(k)
lim = E[ζT0 H],
k↓0 k

so that the marginal bid price is the discounted expected pay-off under a minimal
distance state-price density. For small claim amounts it is also possible to consider
the total price as an expansion in k (see Henderson & Hobson 2002b; Henderson
2002).
As an explicit example in the stochastic volatility model suppose r = 0 and U (v) =
−e−v so that Û (y) = y ln y. Then, when we take the infimum over µ we find that

inf inf E[Û (µζT ) + µ(v + kζT H)] = exp − 1 − v − inf {kE[ζT H] + E[ζT ln ζT ]}
µ ζT ζT

and the option price becomes (see Delbaen et al . 2002)

p(k) = inf {kE[ζT H] + E[ζT ln ζT ]} − inf {E[ζT ln ζT ]}. (5.6)
ζT ζT

The problem of minimizing the entropy was discussed in § 5 c, but in general the
problem of finding the first infimum in (5.6) is hard. There are, however, explicit
solutions in the non-traded asset model (see Henderson & Hobson 2002a).
The expression in (5.6) shows that the utility indifference price for exponential
utility corresponds to a convex risk measure. Note that exponential utility is unique
in that wealth factors out of the problem, to leave option prices which are independent
of wealth. This is a necessary condition for a risk measure.

6. Interest-rate modelling
To date we have concentrated on markets in which the underlying is a risky asset
which can be modelled by a diffusion process. Now we want to consider an interest-
rate market in which the characteristics of the traded assets are different. Three
canonical texts on the subject are Musiela & Rutkowski (1997), Björk (1998) and
Cairns (2004).
Consider a frictionless market in which there is a bank account and a family
of zero-coupon bonds. A zero-coupon bond with maturity date T (a T -bond) is a
contract which guarantees to make a unit payment to the holder at time T . A T -
bond makes no intermediate payments and is typically a mathematical ideal rather
than a genuinely traded instrument. Let the time-t price of the T -bond be denoted
by p(t, T ), and then p(T, T ) = 1.
From the bond prices it is possible to deduce the instantaneous forward rates
f (t, T ) which solve f (t, T ) = −(∂/∂T ) ln p(t, T ) or equivalently

 T 
p(t, T ) = exp − f (t, s) ds ,
t

and the instantaneous short rate rt = f (t, t). The assumption is that the bank
account pays the instantaneous short rate as a stochastic rate of interest, and if
so this is equivalent to investing in a portfolio of ‘just maturing’ bonds. Given the
relationships between the short-rate, the bond prices and the forward prices we can
choose to model any of these.

Proc. R. Soc. Lond. A (2004)

3394 D. Hobson

(a) Short-rate models

Models based on the short rate provide an important subclass of interest-rate
models. We suppose that the short rate rt follows dynamics (under P)
drt = σ(t, rt )(dWt + λ(t, rt ) dt).
Examples include taking rt to be a shifted Ornstein–Uhlenbeck process (Vasicek
1977) or the sum of squares of OU processes (Cox et al . 1985). In a short-rate model
a zero-coupon bond plays the role of a derivative which is to be priced.
In the light of our previous discussion it is useful to know if the model is arbitrage-
free and complete. In fact the discounted price of the traded asset is
R0
Rt = R0 ,
Rt
which is constant under any equivalent measure. Thus there exist equivalent mar-
tingale measures and every equivalent measure is an equivalent martingale measure.
To put this another way, if we fix an equivalent measure Q, then we can define bond
prices via

 T  
Q
p(t, T ) = E exp − rs ds Ft ,
t

but these prices are not the only ones consistent with no arbitrage.
We return to the problem we faced in the previous section: how do we choose an
appropriate measure Q. The two most popular solutions are to finesse the issue by
writing down the dynamics under Q, or to choose a market risk premium γt , whence,
under Q,
drt = σ(t, rt )(dWt + (λ(t, rt ) − γt ) dt).
Given a martingale measure Q we can price bonds and more complicated derivatives
such as options on bonds and interest-rate swaps, and in simple cases we can often
find analytical formulae for these quantities. However, these instruments cannot be
replicated, although, as in a stochastic volatility model, once it is assumed that one
bond is traded, all other zero-coupon bonds with shorter maturity can be hedged
through dynamic trading in that bond.

(b) Forward-rate models

Short-rate models have the feature that the entire interest-rate market is governed
by a single explanatory variable. It is possible to overcome this drawback, perhaps
by including other interest rates, such as the long rate, in the model. However, short-
rate models have largely been supplanted in the academic literature and the industry
by a paradigm shift in which the fundamental modelling objects have become the
forward rates. This leads to interesting new mathematics, not least because the state
variable is now a yield curve which is an infinite-dimensional object.
The method we outline was first proposed by Heath et al . (1992). Let W be a
d-dimensional Brownian motion and suppose that for each fixed T the forward rates
satisfy
df (t, T ) = σ(t, T )(dWt + α(t, T ) dt). (6.1)

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3395

The initial condition {f (0, T )}T 0 can be specified by the initial market of bond
prices and forward rates.
When we switch to the martingale measure Q, under which the discounted traded
quantities (the discounted T -bonds) are martingales, we find that the forward rates
satisfy
  T  
Q
df (t, T ) = σ(t, T ) dWt + σ(s, T ) ds dt
t

and that, although the no-arbitrage conditions fix the drifts in (6.1), there is almost
complete freedom in modelling the volatility structure. Once the volatility coefficients
have been specified under P or Q the market is complete and any derivative can be
priced and replicated using d zero-coupon bonds as hedging instruments.

(c) Market models

A more tractable alternative to the class of forward-rate models are the market
models of Miltersen et al . (1997) and Brace et al . (1997). Instead of concentrating
upon the unobservable forward rates a market model takes quoted interest rates, such
as LIBOR, as the fundamental modelling objects. Moreover, these key objects are
assumed to have a lognormal distribution. One of the main benefits of this assumption
is that it is possible to derive closed-form expressions for simple derivatives such as
caps and floors.

7. Credit and default risk

Financial risks occur in many forms. To date in this article we have been concerned
with market risk—the adverse effects of changes in the values of underlying assets
or interest rates on the market value of a portfolio. But there are other risks facing
agents in financial markets, including credit risk, the risk that a counterparty will fail
to meet its obligations. Given the recent high profile failures of Enron and WorldCom,
these risks have claimed a prominent position in the market psyche.
In a fairly general setting the issue of credit risk can be synthesized into the pricing
of bonds issued by a company. In this case the valuation problems inherent in interest-
rate products are compounded by the risk of default by the issuing company.
There are two main classes of models for credit risk. The first class of models,
called structural models, was introduced by Merton (1974) in an attempt to model
default via a microeconomic description of the assets and liabilities of the firm. The
firm defaults the first time that the assets fall below some threshold. If the assets
are described by a diffusion process, then this means that default is a predictable
event, and it follows that credit spreads of very-short-term bonds should be close to
zero. Unfortunately, this property is not a feature of credit data. There have been
various attempts to modify the class of structural models to overcome this failing,
for instance, by making the price process a jump diffusion (Zhou 2001), or allowing
for imperfect information (Duffie & Lando 2001).
The second class of credit risk models is the reduced-form or intensity based mod-
els. In this class credit events are specified exogenously and default arrives according
to a Poisson process with intensity γt . These models are somewhat arbitrary, but
they provide a good match to data, they are flexible and tractable, and they can be

Proc. R. Soc. Lond. A (2004)

3396 D. Hobson

made to fit smoothly into an interest-rate framework. For example, if default events

happen at rate γt , then the probability of no default by time t is
  t 
exp − γu du
0

and the value of a T -bond (assuming zero recovery on default) is given by


 T 
Q
E exp − (ru + γu ) du ,
0

where expectations are taken with respect to an equivalent martingale measure.

The above descriptions have concentrated on the modelling of default events for
a single company, but one of the main problems in credit is the pricing of portfolios
of corporate debt, in which case it is necessary to model correlated and dependent
default. Schönbucher (2003) gives a full review of credit modelling.

8. Final thoughts
Mathematical finance is concerned with the related problems of quantifying risk,
pricing risk and mitigating the impact of risk via hedging. In general we think of
these risks as arising from changes in the prices of underlying assets—stock prices,
exchange rates, interest rates—which are specified exogenously to the model. (But
one can ask where these prices come from (see, for example, Bick 1987; Cox et al .
1985), and what, if any, are the rational explanations of bubbles and market crashes.)
Given the prices of underlyings, the beautiful Black–Scholes–Merton theory gives
powerful insights into the way derivatives are priced, and leads us to the conclusion
that in perfect markets the prices of derivatives are fully determined.
In imperfect markets option prices are not fully determined. Market imperfections
can arise in many ways, some of which we have discussed in the article above, and
the first challenge facing mathematicians is to model these imperfections in a way
which is amenable to analysis. In some markets, such as energy or weather derivatives
(Brody et al . 2002), exponential Brownian motion is a poor descriptor of the price
process. In some markets liquidity issues mean that delta hedging is infeasible (Cetin
et al . 2004). In some markets agents may have differential information (Amendinger
et al . 1998; Föllmer et al . 1999). In all markets the ways that agents interact and
their relative market power (Cvitanić & Ma 1996; Platen & Schweizer 1998; Bank
& Baum 2004) can have a fundamental impact. These problems require careful and
sympathetic modelling.
The second challenge facing financial mathematics is to the relate the conclusions
from these models to real-world financial practice. This means that questions of
model fit and parameter estimation become crucial, together with an acknowledge-
ment that often the behaviour of agents is as much influenced by factors outside
the model, such as tax considerations or regulatory issues, as the predictions of a
sophisticated mathematical theory.
This review paper was written in partial fulfilment of the conditions of the Adams Prize (2003)
awarded to the author by the University of Cambridge and St John’s College, Cambridge. The
author is pleased to acknowledge the support of the EPSRC via an Advanced Fellowship, and to
thank the referees for pointing out various errors and for suggesting several clarifications. The
remaining errors and biases are the responsibility of the author.

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3397

Appendix A. Stochastic calculus

In this section we review, briefly, the essentials of stochastic calculus that are needed
for the derivation of the Black–Scholes formula. Standard texts on Itô processes
include Revuz & Yor (1998) and Rogers & Williams (2000), or for more basic treat-
ments motivated solely by the applications to finance consider Mikosch (1998) or
Steele (2001).
If Zt = f (Wt , t) then Itô’s formula (Rogers & Williams 2000, § IV.32.8) tells us
that (provided the various derivatives exist)
 t  t


Zt = Z0 + f (Ws , s) dWs + [ 12 f

(Ws , s) + f˙(Ws , s)] ds, (A 1)
0 0

where the first integral is an Itô stochastic integral and the second is Lebesgue–
Stieltjes. Sometimes it is convenient to abbreviate this expression to a stochastic
differential equation,
dZt = df (Wt , t) = f
(Wt , t) dWt + [ 12 f

(Wt , t) + f˙(Wt , t)] dt, (A 2)
but this differential version should be interpreted via the stochastic integral repre-
sentation (A 1). Itô’s formula can be extended to cover functions of semi-martingales
Zt = f (Yt , t) and to functions of more than one stochastic variable Zt = f (Yt1 , Yt2 , t).
The Cameron–Martin–Girsanov theorem (Rogers & Williams 2000, § IV.38.5) says
that if (Ω, F, P) is the canonical probability space supporting a Brownian motion W
(such that the filtration Ft satisfies the usual conditions), and if (Zt )0
t
T defined
via  t  t 
1
Zt = exp ηs dWs − 2
ηs ds (A 3)
0 2 0

is a uniformly integrable martingale, then Q, defined via

dQ
= ZT , (A 4)
dP FT
s
is equivalent to P and, under Q, Bs = Ws − 0 ηu du is a Brownian motion. Moreover,
the converse is also true, in the sense that if Q is equivalent to P, then Q has a
representation via (A 4) and (A 3).
The Brownian martingale representation theorem (Rogers & Williams 2000,
§ IV.36.5) says that, if Mt is a martingale with respect to a filtration Ft generated
by a Brownian motion Wt , then Mt can be written
 t
M t = M0 + ψs dWs
0

for some integrand ψ.

References
Amendinger, J., Imkeller, P. & Schweizer, M. 1998 Additional logarithmic utility of an insider.
Stoch. Process. Applic. 75, 263–286.
Andersen, L. & Andreasen, J. 2000 Static barriers. Risk Mag. 13(9), 120–122.

Proc. R. Soc. Lond. A (2004)

3398 D. Hobson

Andersen, L., Andreasen, J. & Brotherton-Ratcliffe, R. 1998 The passport option. J. Computat.
Finance 1, 15–36.
Artzner, P., Delbaen, F., J-M., Eber & Heath, D. 1999 Coherent measures of risk. Math. Finance
9, 203–228.
Bachelier, L. 1900 Théorie de la spéculation. Ann. Ecole Norm. Sup. 17, 21–86.
Bank, P. & Baum, D. 2004 Hedging and portfolio optimisation in financial markets with a large
trader. Math. Finance 14, 1–18.
Bardhan, I. & Chao, X. 1996 On martingale measures when asset returns have unpredictable
jumps. Stoch. Process. Applic. 63, 35–54.
Barndorff-Nielsen, O. & Shephard, N. 2000 Non-Gaussian Ornstein–Uhlenbeck-based models
and some of their uses in financial economics. J. R. Stat. Soc. B 63, 1–42.
Bergman, Y. Z., Grundy, B. D. & Wiener, Z. 1996 General properties of option prices. J. Finance
51, 1573–1610.
Bick, A. 1987 On the consistency of the Black–Scholes model with a general equilibrium frame-
work. J. Financ. Quantit. Analysis 22, 259–275.
Björk, T. 1998 Arbitrage theory in continuous time. Oxford University Press.
Black, F. & Scholes, M. 1973 The pricing of options and corporate liabilities. J. Polit. Econ. 81,
637–654.
Brace, A., Gatarek, D. & Musiela, M. 1997 The market model of interest rate dynamics. Math.
Finance 7, 127–155.
Branger, N. 2004 Pricing derivative securities using cross entropy: an economic analysis. Int. J.
Theor. Appl. Finance 7, 63–82.
Brody, D. C., Syroka, J. & Zervos, M. 2002 Dynamical pricing of weather derivatives. Quantitat.
Finance 2, 189–198.
Brown, H., Hobson, D. G. & Rogers, L. C. G. 2001 Robust hedging of barrier options. Math.
Finance 11, 285–314.
Cairns, A. J. G. 2004 Interest rate models: an introduction. Princeton University Press.
Cetin, U., Jarrow, R. & Protter, P. 2004 Liquidity risk and arbitrage pricing theory. Finance
Stochast. 8, 311–342.
Cox, J., Ross, S. & Rubinstein, M. 1979 Option pricing: a simplified approach. J. Financ. Econ.
7, 229–263.
Cox, J., Ingersoll, J. & Ross, S. 1985 A theory of the term structure of interest rates. Econo-
metrica 53, 385–408.
Cvitanić, J. & Karatzas, I. 1993 Hedging contingent claims with constrained portfolios. Ann.
Appl. Probab. 3, 652–681.
Cvitanić, J. & Ma, J. 1996 Hedging options for a large investor and forward-backward SDEs.
Ann. Appl. Probab. 6, 370–398.
Cvitanić, J., Schachermayer, W. & Wang, H. 2001 Utility maximisation in incomplete markets
with random endowment. Finance Stochast. 5, 259–272.
Davis, M. H. A. 2000 Optimal hedging with basis risk. (Available at http://www.ma.ic.ac.uk/
˜mdavis/docs/basisrisk.pdf.)
Davis, M. H. A., Panas, V. G. & Zariphopoulou, T. 1993 European option pricing with trans-
action costs. SIAM J. Control Optim. 31, 470–493.
Delbaen, F. & Schachermayer, W. 1994 A general version of the fundamental theorem of asset
pricing. Math. Annln 300, 463–520.
Delbaen, F. & Schachermayer, W. 1998 The fundamental theorem of asset pricing for unbounded
stochastic processes. Math. Annln 312, 215–250.
Delbaen, F. & Yor, M. 2002 Passport options. Math. Finance 12, 299–328.
Delbaen, F., Grandits, P., Rheinländer, T., Samperi, D., Schweizer, M. & Stricker, C. 2002
Exponential hedging and entropic penalties. Math. Finance 12, 99–123.

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3399

Duffie, D. & Epstein, L. 1992 Stochastic differential utility. Econometrica 60, 353–394.
Duffie, D. & Lando, D. 2001 Term structures of credit spreads with incomplete accounting
information. Econometrica 69, 633–664.
Duffie, D., Fleming, W., Soner, H. M. & Zariphopoulou, T. 1997 Hedging in incomplete markets
with Hara utility. J. Econ. Dynam. Control 21, 753–782.
Dupire, B. 1994 Pricing with a smile. Risk Mag. 7(5), 18–20.
El Karoui, N. & Quenez, M.-C. 1995 Dynamic programming and pricing of contingent claims in
an incomplete market. SIAM J. Control Optim. 33, 29–66.
El Karoui, N., Peng, S. & Quenez, M.-C. 1997 Backward stochastic differential equations in
finance. Math. Finance 7, 1–71.
Figlewski, S. 2002 Assessing the incremental value of option pricing theory relative to an infor-
mationally passive benchmark. J. Derivatives 10(Fall), 80–96.
Föllmer, H. & Kabanov, Y. 1998 Optional decomposition and Lagrange multipliers. Finance
Stochast. 2, 69–81.
Föllmer, H. & Kramkov, D. 1997 Optional decompositions under constraints. Probab. Theory
Relat. Fields 109, 1–25, 1997.
Föllmer, H. & Leukert, P. 2000 Efficient hedging: cost versus shortfall risk. Finance Stochast. 4,
117–146.
Föllmer, H. & Schied, A. 2002 Robust preferences and convex risk measures. In Advances in
finance and stochastics. Springer.
Föllmer, H. & Schweizer, M. 1990 Hedging of contingent claims under incomplete information.
In Applied stochastic analysis (ed. M. H. A. Davis & R. J. Elliott). New York: Gordon and
Breach.
Föllmer, H. & Sondermann, D. 1986 Hedging of contingent claims under incomplete information.
In Contributions to mathematical economics (ed. W. Hildenbrand & A. MasColell), pp. 205–
223. Amsterdam: North-Holland.
Föllmer, H., Wu, C.-T. & Yor, M. 1999 Canonical decomposition of linear transformations of
two independent Brownian motions motivated by models of insider trading. Stoch. Process.
Applic. 84, 137–164.
Fouque, J.-P., Papanicolaou, G. & Sircar, K. R. 2000 Derivatives in financial markets with
stochastic volatility. Cambridge University Press.
Frey, R. 1997 Derivative asset analysis in models with level-dependent and stochastic volatility.
CWI Q. 10, 1–34.
Frey, R. & Sin, C. 1999 Bounds on European option prices under stochastic volatility. Math.
Finance 9, 97–116.
Friedman, A. 2000 Free boundary problems in science and technology. Not. Am. Math. Soc. 47,
854–861.
Frittelli, M. 2000 The minimal entropy martingale measure and the valuation problem in incom-
plete markets. Math. Finance 10, 35–92.
Geman, H. & Yor, M. 1993 Bessel processes, Asian options and perpetuities. Math. Finance 3,
349–375.
Geman, H., El Karoui, N. & Rochet, J.-C. 1995 Changes of numéraire, changes of probability
measure and option pricing. J. Appl. Probab. 32, 443–458.
Goldman, B., Sosin, H. & Gatto, M. A. 1979 Path-dependent options: buy at the low, sell at
the high. J. Finance 34, 1111–1127.
Goll, T. & Rüschendorf, L. 2001 Minimax and minimal distance martingale measures and their
relationship to portfolio optimisation. Finance Stochast. 5, 557–581.
Gourieroux, C., Laurent, J. P. & Pham, H. 1998 Mean–variance hedging and numéraire. Math.
Finance 8, 179–200.
Grandits, P. & Krawczyk, L. 1998 Closedness of some spaces of stochastic integrals. Sém. Probab.
32, 73–85.

Proc. R. Soc. Lond. A (2004)

3400 D. Hobson

Hakansson, N. 1979 The fantastic world of finance: progress and the free lunch. J. Financ.
Quantit. Analysis 14, 717–734.
Harrison, J. M. & Kreps, D. M. 1979 Martingales and arbitrage in multiperiod securities markets.
J. Econom. Theory 20, 381–408.
Harrison, J. M. & Pliska, S. R. 1981 Martingales and stochastic integrals in the theory of
continuous trading. Stoch. Process. Applic. 11, 251–260.
Heath, D., Jarrow, R. & Morton, A. 1992 Bond pricing and the term structure of interest rates.
Econometrica 60, 77–106.
Henderson, V. 2002 Valuation of claims on non-traded assets using utility maximisation. Math.
Finance 12, 351–373.
Henderson, V. 2003 The impact of the market portfolio on the valuation, incentives and optimal-
ity of executive stock options. (Available at http://www.orfe.princeton.edu/%7Evhenders/
OptCompRF.pdf.)
Henderson, V. & Hobson, D. G. 2000 Local time, coupling and the passport option. Finance
Stochast. 4, 69–80.
Henderson, V. & Hobson, D. G. 2002a Substitute hedging. Risk Mag. 15(5), 71–75.
Henderson, V. & Hobson, D. G. 2002b Real options with constant relative risk aversion. J. Econ.
Dynam. Control 27, 329–355.
Henderson, V., Hobson, D. G., Howison, S. & Kluge, T. 2003 A comparison of q-optimal option
prices in a stochastic volatility model with correlation. Oxford Financial Research Centre,
University of Oxford. Preprint no. 2003-MF-02.
Heston, S. L. 1993 A closed-form solution for options with stochastic volatiltiy and applications
to bond and currency options. Rev. Financ. Stud. 6, 327–343, 1993.
Hobson, D. G. 1998 Volatility mispecification, option pricing and superreplication by coupling.
Ann. Appl. Prob. 8, 193–205.
Hobson, D. G. 2003 Real options, non-traded assets and utility indifference prices. (Preprint.)
University of Bath, UK.
Hobson, D. G. 2004 Stochastic volatility models, correlation, and the q-optimal measure. Math.
Finance 14, 537–556.
Hobson, D. G. & Rogers, L. C. G. 1998 Complete models with stochastic volatility. Math.
Finance 8, 27–48.
Hodges, S. D. & Neuberger, A. 1989 Optimal replication of contingent claims under transaction
costs. Rev. Fut. Mark. 8, 223–239.
Hubalek, F. & Schachermayer, W. 2001 The limitations of no-arbitrage arguments for real
options. Int. J. Theor. Appl. Finance 4, 361–363.
Hugonnier, J., Kramkov, D. & Schachermayer, W. 2004 On the utility based pricing of contingent
claims in incomplete markets. Math. Finance (In the press.)
Hull, J. & White, A. 1987 The pricing of options on assets with stochastic volatilities. J. Finance
42, 281–299.
Hull, J. & White, A. 1988 An analysis of the bias in option pricing caused by a stochastic
volatility. Adv. Futures Opt. Res. 3, 29–61.
Hyer, T., Lipton-Lipschitz, A. & Pugachevsky, D. 1997 Passport to success. Risk Mag. 10(9),
127–131.
Karatzas, I. 1989 Optimization problems in the theory of continuous trading. SIAM J. Control
Optim. 27, 1221–1259.
Kramkov, D. & Schachermayer, W. 1999 The asymptotic elasticity of utility functions and
optimal investment in incomplete markets. Ann. Appl. Prob. 9, 904–950.
Kreps, D. M. 1981 Arbitrage and equilibrium in economies with infinitely many commodities.
J. Math. Econ. 8, 15–35.
Mania, M., Santacrocce, M. & Tevzadze, R. 2003 A semimartingale BSDE related to the minimal
entropy martingale measure. Finance Stochast. 7, 385–402.

Proc. R. Soc. Lond. A (2004)

 A survey of mathematical finance 3401

Margrabe, W. 1978 The value of an option to exchange one asset for another. J. Finance 33,
177–186.
Merton, R. C. 1969 Lifetime portfolio selection under uncertainty: the continuous time case.
Rev. Econ. Statist. 51, 247–257.
Merton, R. C. 1973 Theory of rational option pricing. Bell J. Econ. Mngmt Sci. 4, 141–183.
Merton, R. C. 1974 On the pricing of corporate debt: the risk-structure of interest rates. J.
Finance 29, 449–470.
Merton, R. C. 1976 Option pricing when the underlying stock returns are discontinuous. J.
Financ. Econ. 3, 125–144.
Merton, R. C. 1990 Continuous time finance. Oxford: Blackwell.
Mikosch, T. 1998 Elementary stochastic calculus with finance in view. Singapore: World Scien-
tific.
Miltersen, K., Sandmann, K. & Sondermann, D. 1997 Closed form solutions for term-structure
derivatives with lognormal interest rates. J. Finance 52, 409–430.
Musiela, M. & Rutkowski, M. 1997 Martingale methods in financial modelling. Springer.
Myneni, R. 1992 The pricing of the American option. Ann. Appl. Prob. 2, 1–23.
Platen, E. & Schweizer, M. 1998 On feedback effects from hedging derivatives. Math. Finance
8, 67–84.
Renault, E. & Touzi, N. 1997 Option hedging and implied volatilities in a stochastic volatility
model. Math. Finance 7, 399–410.
Revuz, D. & Yor, M. 1998 Continuous martingales and Brownian motion, 3rd edn. Springer.
Rogers, L. C. G. 1997 Arbitrage with fractional Brownian motion. Math. Finance 7, 95–105.
Rogers, L. C. G. & Shi, Z. 1995 The value of an Asian option. J. Appl. Probab. 32, 1077–1088.
Rogers, L. C. G. & Williams, D. 2000 Diffusions, Markov processes and martingales, vols 1, 2.
Cambridge University Press.
Rouge, R. & El Karoui, N. 2000 Pricing via utility maximisation and entropy. Math. Finance
10, 259–276.
Samuelson, P. A. 1965 Rational theory of warrant pricing. Indust. Mgmt Rev. 6, 13–31, 1965.
Schachermayer, W. 2003 Introduction to the mathematics of financial markets. In Lecture notes
on Probability Theory and Statistics, Saint Flour Summer School 2000, pp. 111–177. Springer.
Schönbucher, P. 2003 Credit derivatives pricing models: models, pricing and implementation.
Wiley.
Schweizer, M. 1996 Approximation pricing and the variance optimal martingale measure. Ann.
Probab. 24, 206–236.
Scott, L. O. 1987 Option pricing when the variance changes randomly, theory, estimation and
an application. J. Financ. Quantit. Analysis 22, 419–438.
Soner, H. M. & Touzi, N. 2001 Superreplication under gamma constraints. J. Control Optim.
41, 404–424.
Soner, H. M., Shreve, S. E. & Cvitanić, J. 1995 There is no nontrivial hedging portfolio for
option pricing with transaction costs. Ann. Appl. Probab. 5, 327–355.
Steele, J. M. 2001 Stochastic calculus and financial applications. Springer.
Stein, E. M. & Stein, J. C. 1991 Stock-price distributions with stochastic volatility: an analytic
approach. Rev. Financ. Stud. 4, 727–752.
Vasicek, O. 1977 An equilibrium characterisation of the term structure. J. Financ. Econ. 5,
177–188.
Večer, J. 2001 A new PDE approach for pricing arithmetic Asian options. J. Computat. Finance
4, 105–113.
Wong, B. & Heyde, C. C. 2004 On the martingale property of stochastic exponentials. J. Appl.
Probab. 41, 654–664.
Zhou, C. 2001 The term structure of credit spreads with jump risk. J. Bank. Finance 25, 2015–
2040.

Proc. R. Soc. Lond. A (2004)