SIAM J. APPL. MATH.

c 2000 Society for Industrial and Applied Mathematics


Vol. 61, No. 1, pp. 232–272

THE FEEDBACK EFFECT OF HEDGING IN ILLIQUID MARKETS∗

PHILIPP J. SCHÖNBUCHER† AND PAUL WILMOTT‡
Abstract. This paper analyzes the influence of dynamic trading strategies on the prices in
financial markets. After a thorough discussion of the modeling issues involved we derive the modi-
fication of the stochastic process of the underlying asset that follows from the presence of dynamic
trading strategies. We analyze the nonlinear effects and the feedback from prices to trading strategy.
The pricing, hedging, and replication of options in the context of illiquid markets is discussed and a
nonlinear partial differential equation for an option replication strategy is derived. Finally the effects
of one of the most popular trading strategies—Put-option replication—on the price of the underlying
asset are illustrated using numerical simulations.

Key words. option pricing, illiquid markets, market manipulation, portfolio insurance, feedback
trading

AMS subject classifications. 90A09, 90A60

PII. S0036139996308534

1. Introduction. The aim of this paper is to provide a simple framework for

modeling the dynamics of prices in illiquid financial markets and for analyzing the
effects of dynamic trading strategies in such markets. We focus on financial markets
that are influenced by dynamic trading strategies, i.e., trading strategies which pre-
scribe the number of shares to be held in the portfolio as a function f (S, t) of the share
price S and of time t. Dynamic trading strategies of this type have enjoyed increasing
popularity in the past decades not least because of their central importance in option
pricing theory.
In perfectly liquid markets one can find a dynamic trading strategy that repro-
duces the movement of the price of a given derivative asset by judiciously trading
the underlying asset and cash. Finding this trading strategy is the key step in the
derivation of most option pricing equations.1 In theory, any option in a complete and
perfectly liquid market can be replicated by following the appropriate trading strategy.
Since 1973 trading strategies of this type have enjoyed tremendous popularity not
only in the theory but also in the practice of finance. A writer of an option can use a
dynamic trading strategy to hedge his exposure, an arbitrageur can exploit mispricing
in traded options by selling an overpriced option and replicating it with a dynamic
trading strategy, and an investor can synthetically create the option with the payoff
that suits her needs best.
∗ Received by the editors August 26, 1996; accepted for publication (in revised form) June 7, 1999;

published electronically July 5, 2000. The research of the first author was supported by the Deutsche
Forschungsgemeinschaft SFB 303 at the University of Bonn and the Deutscher Akademischer Aus-
tauschdienst (DAAD). The research of the second author was supported by the Royal Society. Some
of the results of this paper have appeared in Hedging in illiquid markets: Nonlinear effects, in
ICIAM/GAMM 95: Proceedings of the Third International Congress on Industrial and Applied
Mathematics, Hamburg, 1995: Issue 3: Applied Stochastics and Optimization, O. Mahrenholtz, K.
Marti, and R. Mennicken, eds., Special Issue: Zeitschrift für Angewandte Mathematik und Mechanik
(ZAMM), 1996, pp. 81–84.
http://www.siam.org/journals/siap/61-1/30853.html
† University of Bonn, Department of Statistics, Adenauerallee 24-42, 53113 Bonn, Germany

(P.Schonbucher@finasto.uni-bonn.de).
‡ University of Oxford, Mathematical Institute, 24-29 St. Giles, Oxford OX1 3LB, United Kingdom

and Imperial College, Department of Mathematics ([email protected]).

1 Black and Scholes [5] and Merton [24] introduced this method, which has been used extensively

since then; see, e.g., Hull [15] and, for exotic options, Wilmott, Howison, and Dewynne [31].
232
 FEEDBACK IN ILLIQUID MARKETS 233

The application of dynamic trading strategies has not been restricted to option
pricing and hedging: Portfolio managers use dynamic trading strategies to insure
themselves against movements in the share price. This strategy is called portfolio
insurance and has been the subject of extensive analysis (see Leland [20], Brennan
and Schwartz [8], and Bookstaber and Langsam [7]).
More trading strategies have been derived in the investment analysis literature
(for the continuous time case see, e.g., Merton [24]). These portfolio investment rules
give the investor’s investment a strategy that maximizes his expected utility. Chartist
trading strategies form another class of trading strategies that is popular in practice.
The effects of these trading strategies can also be analyzed within a slight modification
of the framework presented here.
Given a trading strategy f (S, t), a change in the share price by dS will mean the
portfolio has to be adjusted by df , and df shares are bought in the market. But, by
being trades in the share, these orders themselves may influence the price S on which
they are conditioned and cause a further price movement dS  . This price movement
will then trigger more dynamic trading. This feedback effect is one of the main
concerns of this paper. We will investigate how these trading strategies influence the
stochastic process of the price of the underlying share and how the stability of the
price process is affected by the presence of these trading strategies.
Usually it is assumed that the effect of individual trading on the asset price is
too small to be of any importance and is neglected when the strategy is derived. This
seems justified if the market in question has many participants and a high degree of
liquidity, which is typically true for modern financial markets. On the other hand,
the trading strategies are very often implemented on a large scale, and the liquidity
of the financial markets is sometimes very limited. In the case of the October 1987
stock market crash some empirical studies (e.g., Furbush [13]) and even the official
report of the investigations carried out by the Brady commission [27] suggest that
portfolio insurance trading helped to aggravate the effects of the crash. But even in
noncrash environments increasing volatility has been attributed to the presence of
dynamic trading strategies.
Given the evident importance of dynamic trading strategies it is surprising that
their effect on the price of the underlying asset has received relatively little attention
in the theoretical literature. The early work by Genotte and Leland [14] and Jarrow
[17, 16] addressed the problem of market manipulation and program trading in a
discrete-time framework. Genotte and Leland analyzed the effects of program trading
in a rational expectations framework with special focus on the effects of differential
between investors. They showed that liquidity can be very low if uninformed investors
mistake program trading for informed trading. Jarrow is concerned with the existence
of opportunities for market manipulation. He gave conditions under which these
manipulations are not possible and went on to price a Call option in a two-period
binomial model. Because they are restricted to discrete time these papers cannot
capture the dynamic effects we would like to model in this paper.
In continuous time, Bick [3, 4] showed how the canonical security price process of
the Black–Scholes model, the lognormal Brownian motion, can be derived in a general
equilibrium model with price-taking agents. This model is a general equilibrium model
with complete markets and continuous trading by all agents. Thus any new derivative
security is obsolete, and therefore it will be neither traded nor hedged or replicated;
the agents have no need to implement the option replication strategies we want to
study here. In general for equilibrium models with complete markets, derivative
234 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

security pricing is reduced to finding consistent prices, prices that are consistent with
the already achieved equilibrium.
In contrast to this full general equilibrium approach in complete markets we
believe that—although some traders (usually large financial institutions) have con-
tinuous trading opportunities—many investors cannot trade continuously. To these
investors the markets are incomplete and a new option opens new investment oppor-
tunities to them. This new option is issued by a financial institution (with continuous
trading opportunities) and sold to the investors who cannot trade continuously. The
financial institution then hedges its risk by replicating the issued options with a con-
tinuous trading strategy. In this situation the introduction of a new security does
create a significant amount of replication trading. In our study we want to investigate
the effects of this replication trading.
Besides general equilibrium theory, market microstructure theory is another area
of economic theory that is relevant for the dynamic modeling of markets. Market
microstructure is concerned with the analysis of the mechanisms by which prices are
set in markets. Special attention is paid to the strategic interaction of traders in the
markets that results from a particular price forming mechanism and to the way in
which private information is incorporated in the prices. As the collapse of option
trading in illiquid markets results from the large trader’s strategic exploitation of
his market power, market microstructure theory can be of great use here. A good
introduction to the theory can be found in O’Hara [26]. Unfortunately almost all
models are set in discrete time with the exception of Back [1, 2], who transferred the
strategic trading model by Kyle [19] into continuous time.
In parallel research to this paper Frey and Stremme [12] and Frey [10, 11] devel-
oped a continuous-time model similar to the one presented in this paper. Frey and
Stremme [12] analyzed the relative hedging performance of super-replication strate-
gies and naive Black–Scholes trading in this environment. Frey [10, 11] then derived a
partial differential equation for perfect replication trading strategies and option pric-
ing in feedback markets, giving existence and uniqueness results. Frey’s focused on
the option pricing in illiquid markets, while this paper’s main concern are the price
dynamics in the underlying share: feedback effects and the possible discontinuities
(crashes) in the share price. In a later work and using a similar framework, Sircar and
Papanicolaou [30] focused on the numerical solution of the feedback option pricing
equations and discuss the differences to the “classical” option pricing. We will argue
later on why there may be fundamental difficulties with options in illiquid markets
and why option replication trading, on which all these pricing equations are based,
may not be optimal. In many cases derivative security markets are not viable if the
market for the underlying asset is illiquid. Nevertheless, we give the derivation of the
option replication formula for illiquid markets as it was first presented in Schönbucher
[28] and Schönbucher and Wilmott [29].
The rest of the paper is organized as follows. In the following section we set up
the model we used to analyze trading strategies in illiquid markets. We focus on the
new modeling aspects that are introduced by the illiquidity of the markets, especially
the price forming mechanism and manipulation opportunities. We also derive the
parameters of the stochastic process of the price both with and without the influence
of a large trader.
In the next section we give a short overview of the classical theory of the dynamic
trading strategies in liquid markets and derive the canonical example of a dynamic
trading strategy that replicates the payoff of a European Put option. To show the new
 FEEDBACK IN ILLIQUID MARKETS 235

features introduced by illiquidity, we then analyze the difference between the paper
value and the real value of a portfolio in illiquid markets and give a price manipulation
strategy for the large trader in this section.
The following section looks at options in illiquid markets and manipulation strate-
gies are pointed out that will lead to a collapse of option trading in many cases. We
then look at different scenarios for payoff replication in illiquid markets and derive
the partial differential equations that are satisfied by the value process of these payoff
functions.
In the next section we illustrate the price effects of hedging in illiquid markets
using the example of Put-option replication trading. Special emphasis is laid on the
destabilizing effect of feedback trading and the possibility of jumps in the price process
(i.e., crashes) that are induced by some dynamic trading strategies. The numerical
solution for the probability density function is used to illustrate these effects. In the
conclusion the main results of the paper are summarized.
2. Modeling markets with dynamic trading.
2.1. Traded securities and definitions. There are two investment opportu-
nities: the share2 and the risk-free bond. The price process of the share is denoted
by St and the price process of the risk-free bond is Bt . We assume that the market
for the share is illiquid while the market for the bond is perfectly liquid.
In classical financial theory the stochastic processes for the share price and the
bond price are directly specified, and using this specification derivative securities are
priced or investment decisions are made. In contrast to this approach we will have to
derive the stochastic law for the share price process later on, taking demand, supply,
and dynamic trading into account. The illiquidity of the share market forces us to go
one step further back in the modeling approach. We only assume that the share price
process can be written as a diffusion process of the form

(1) dSt = µ(St , t)dt + σ(St , t)dWt ,

where dW is the increment of a standard Brownian motion and the exact form for the
functions µ(S, t) and σ(S, t) will be derived later on from a more fundamental model
of the share market.
The Black–Scholes [5] lognormal random walk can be represented in this frame-
work as

(2) µ(S, t) = µ̃S and σ(S, t) = σ̃S,

where µ̃ and σ̃ are constants.

As the market for bonds is usually very much more liquid than equity markets,
we need not worry about price effects of our trading and can directly prescribe the
following price process for the risk-free bond:3

(3) dBt = rBt dt or Bt = ert ;

the value of the risk-free bond grows at the risk-free interest rate r, and its initial
value is normalized to B0 = 1.
2 The
model is not restricted to equity shares: S can denote the price process of any traded asset.
3 There
is no loss of generality: we could even take the risk-free bond as numeraire and fix its
price at Bt ≡ 1.
236 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

2.2. Market participants and information. There are two types of market
participants in the model: a large trader and a large population of small traders,
representing the rest of the market. All market participants can trade continuously
without transaction costs. New information continuously arrives at the market in
the form of a Brownian motion Wt . In a more realistic model one would take a
Markovian state vector instead of a Brownian motion; we chose the Brownian motion
for simplicity.
2.3. The small traders. We do not want to model the small traders’ invest-
ment problem explicitly but rather restrict ourselves to modeling the result of it: the
markets reaction to buy and sell orders, and the price dynamics in absence of the
large trader. Therefore we only model the aggregate behavior of the small traders in
this market. Given a share price S, a value W of the Brownian motion, and time t,
the small traders want to hold (on aggregate)

(4) D(S, W, t)

shares in their portfolio. This is the aggregate demand for shares in this market.
Similarly we write the supply,4 i.e., the total number of shares available in the market,
as S(S, W, t). The definitions of the equilibrium (8) and (9) will use only the excess
demand X (S, W, t) in the market:

(5) X (S, W, t) = D(S, W, t) − S(S, W, t).

We make the following assumptions:

• There are no other factors affecting the decision of the small traders, all
necessary stochastic influences are contained in W . Specifically, the small
traders are unaware of the large trader’s presence and of his trading strategy.5
• The excess demand depends only on current values of S, W , and t.
• X (S, W, t) is twice continuously differentiable in S and W and once continu-
ously differentiable in t.
• The excess demand curve X (S, W, t) has a negative slope with respect to the
share price:
∂
(6) X (S, W, t) < 0
∂S
for all S, W, t. This common assumption in economic theory (see, e.g., Mas-
Colell, Winston, and Green [22]) means that, everything else being equal, the
small traders would like to hold more shares in their portfolio if the shares are
cheap (low price S) and fewer if the shares are expensive (high S). Several
justifications can be given for this assumption: The small traders’ budget
constraint will mean that they simply cannot afford to buy many shares if
the price is high. Second, assuming that the small investors believe in a
fundamental value of the share, they will want to hold more shares if they
believe them to be undervalued. Finally, only with a negative slope of the
demand curve will the price be moved up by a buy order of the large trader
4 Supply in the share market changes only if there is a new issue of shares, but in commodities it

is an important factor.
5 This assumption serves to avoid strategic trading as a reaction to the large trader’s trades.

Otherwise we would have to find the full game-theoretic equilibrium between the large trader and
the small traders.
 FEEDBACK IN ILLIQUID MARKETS 237

and down by a sell order of the large trader. This effect is the common
reaction of prices to large transactions in most exchanges (see, e.g., Board
and Sutcliffe [6]). With a positive slope this relationship would be inverted.
• For each W and t there is a S such that

(7) X (S, W, t) = 0.

This S is called the (undisturbed) equilibrium price.

The interpretation of X (S, W, t) as an excess demand function serves to motivate
the special form of the slope of X (S, W, t) and links the model to standard models in
economic theory. This interpretation is not essential to the model: one can also regard
X (S, W, t) together with the price forming mechanism (8) and (9) as a functional
description of the market’s reaction to trades and nothing else.
2.4. The large trader. The large trader follows a trading strategy

f (S, t),

which could be any dynamic trading strategy. We assumed that this strategy does
not explicitly depend on the information arrival process W (except indirectly through
S). This assumption can be easily dropped but to get the feedback effects we aim
to model a large trader who conditions trades mainly on the price. Again we assume
sufficient differentiability in f .
The demand generated by this trading strategy is f (S, t); it is added to the small
traders’ demand to form the total demand. To be able to do this we must assume
that the small traders are unaware of the presence of the large trader in the market;
otherwise we would expect X (S, W, t) to condition on the large trader’s strategy f
itself. Even though we will continue to speak of the large trader, one can also think of
f (S, t) as a trading strategy followed collectively by another group of small traders.
2.5. Share prices. In the absence of the large trader, the equilibrium price is
defined as the price S which solves

(8) X (S, W, t) = 0,

given the excess demand function (5), time t, and the information process W .
With the large trader the equilibrium condition defining the share price S becomes

(9) X (S, W, t) + f (S, t) = 0,

where the additional demand due to the large trader’s trading strategy is added to
the small traders’ excess demand.
In the demand/supply interpretation, S is the price at which demand equals
supply. If the price S  in the market were lower than S, there would be surplus
demand because of the slope of the demand curve. Surplus demand represents people
who want to buy but cannot. They will bid the price up, toward the equilibrium
price. Similarly, the excess supply for S  > S (representing people wanting to sell
at S  ) makes the price drop, again toward the equilibrium price S. At S there is no
unsatisfied demand or supply, and because of the dynamics the market equilibrium
is stable.6 Note that this argument supposes that the excess demand function is
negatively sloped; for a positive slope the equilibrium would still exist but it would
6 This argument for equilibrium also holds in the presence of f , the demand of the large trader.
238 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

be unstable. We will encounter locally positively sloped excess demand functions

later on.
Disequilibrium is obviously possible but given the speed of the flow of information
in these markets and the large number of professionals on the stock markets, a full
equilibrium in stocks and flows in modern financial markets is a good approximation.
This does not mean that these markets are static; in our model both demand and
supply can change in time because of the stochastic parameter W .
Example 1 (lognormal random walk as price process). As an example, consider
the following specification of the excess demand which is linear in S and exponential
in W:

  
(10) X (S, W, t) = a S0 exp (µ̃ − 12 σ̃ 2 )t + σ̃W − S .

Then X (S, W, t) = 0 leads to S = S0 exp{(µ̃ − 12 σ̃ 2 )t + σ̃W } and

(11) dS = µ̃Sdt + σ̃SdW,

S follows the lognormal random walk of (2), and a is a parameter measuring the
liquidity of the market with larger a representing higher liquidity.
Example 2 (Brownian motion as price process). If the excess demand function
is linear in both S and W ,

X (St , Wt , t) = a(S0 + Wt − St ),

then the resulting price process is simply a Brownian motion

dSt = dWt .

Here S0 is the initial share price, and again a measures liquidity and S0 is the initial
share price.
Example 3 (general process as price process). In a framework with a general
state vector X instead of just a Brownian motion W as input, one can construct to
any arbitrary process xt an excess demand function which yields xt as equilibrium
price process: Take x as first component of the state vector X and simply choose the
excess demand

X (S, X, t) = a(x − S).

Thus any price process can be supported in the excess-demand framework.

2.6. The price mechanism in detail. The intuition behind the definition of
the equilibrium price by (8), (9) is as follows: Both large and small traders can submit
a full demand function, i.e., the number of shares they want to buy (or sell) conditional
on the price at which trading will take place. The large trader, for example, submits
the orders df (S, t) = f (S, t) − f (S− , t−), which gives his demand as a function of
the new price S; dX (S, W, t) functions similarly for the small traders. In reality
such a strategy can be approximated by mixing limit orders with different limits.
This mechanism is therefore similar to the order book as it is found in many stock
exchanges. The sequence of events is as follows:
– The new signal dW arrives.
– The small traders submit their orders

dX (S, W, t) = X (S, W, t) − X (S− , W− , t−).

 FEEDBACK IN ILLIQUID MARKETS 239

– Simultaneously the large trader submits an order

df (S, t) = f (S, t) − f (S− , t−).

– The equilibrium price is determined. It is the price S at which buy orders

and sell orders balance (this is a common price mechanism in many exchanges7 ),
dX (S, W, t) + df (S, t) = 0.
– At this price trading takes place.
If the market was in equilibrium before (at t− we had X (S− , W− , t−)+f (S− , t−) =
0), then it will be in equilibrium at t too: X (S, W, t)+f (S, t) = 0. Thus an equilibrium
in the flow variables (dX and df ) will sustain an equilibrium in the stock variables (X
and f ).

2.7. The stochastic process of the price. Instead of solving (8) or (9) di-
rectly, one can derive the stochastic differential equation satisfied by S from these
equations. The stochastic differential equation will show the dynamics of S much
more clearly than the implicit definition in the equilibrium equations. We are looking
for a stochastic differential equation of the form (1)

dS = µ(S, t)dt + σ(S, t)dW,

which is implicitly defined by (9):

X (S, W, t) + f (S, t) = 0.

From (9) we have

0 = dX (S, W, t) + df (S, t)

∂ ∂2
= X (S, W, t) + 12 σ 2 (S, t) 2 X (S, W, t)
∂t ∂S

∂2 1 ∂
2
+ σ(S, t) X (S, W, t) + 2 X (S, W, t) dt
∂S ∂W ∂W 2
∂ ∂
+ X (S, W, t)dS + X (S, W, t)dW
∂S ∂W
 
∂ ∂2 ∂
+ f (S, t) + 12 σ 2 (S, t) 2 f (S, t) dt + f (S, t)dS.
∂t ∂S ∂S

2
For the cross-derivative term ∂S∂∂W f in Itô’s formula we needed the instantaneous
covariation between the share price S and the Brownian motion W . This is dSdW =

7 For other price setting systems (especially market-maker systems) the model will have to be

modified accordingly at this point.

240 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

σ(S, t)dt. We can substitute (1) for dS and reach


∂ ∂ ∂2
0= X (S, W, t) + f (S, t) + 12 X (S, W, t)
∂t ∂t ∂W 2
 2 2

2 1 ∂ 1 ∂ ∂2
+ σ (S, t) 2 2 X (S, W, t) + 2 2 f (S, t) + σ(S, t) X (S, W, t)
∂S ∂S ∂S ∂W
 
∂ ∂
(12) +µ(S, t) X (S, W, t) + f (S, t) dt
∂S ∂S
   
∂ ∂ ∂
(13) + σ(S, t) X (S, W, t) + f (S, t) + X (S, W, t) dW.
∂S ∂S ∂W

For this equality to hold both the stochastic component (in (13)) and the locally
deterministic component (in (12) and the lines above (12)) must equal zero. From
(13) follows that σ(S, t) must satisfy
∂
∂W X (S, W, t)
(14) σ(S, t) = − ∂ ∂
,
∂S X (S, W, t) + ∂S f (S, t)

otherwise the increments of the right-hand side could not be equal to zero. This and
the fact that the dt-component must be zero yield
 2
1 ∂ ∂ 1 ∂
µ(S, t) = − ∂ X (S, W, t) + f (S, t) + X (S, W, t)
∂
∂S X (S, W, t) + ∂S f (S, t)
∂t ∂t 2
∂W 2

2  
∂
X (S, W, t) ∂2 ∂2
+ 12 ∂W
X (S, W, t) + f (S, t)
∂
∂S X (S, W, t) + ∂
∂S f (S, t) ∂S 2 ∂S 2


∂
∂W X (S, W, t) ∂2
(15) − ∂ ∂
X (S, W, t) .
∂S X (S, W, t) + ∂S f (S, t)
∂S ∂W

In the absence of the large trader’s trading strategy f (S, t), this reduces to
∂
X (S, W, t)
(16) σ(S, t) = − ∂W
∂
∂S X (S, W, t)

and

2
1  ∂ X (S, W, t) + 1 ∂ X (S, W, t)
µ(S, t) = − ∂
∂S X (S, W, t)
∂t 2
∂W 2

2
∂
1 ∂W X (S, W, t) ∂2
(17) + X (S, W, t)
2 ∂
∂S X (S, W, t) ∂S 2


∂ 2
∂W X (S, W, t) ∂
− ∂
X (S, W, t) .
∂S X (S, W, t)
∂S ∂W
 FEEDBACK IN ILLIQUID MARKETS 241

In fact we have given drift µ and volatility σ of the share price in terms of both S
and W , i.e. we have given µ(S, W, t) and σ(S, W, t) instead of just µ(S, t) and σ(S, t)
as in (1). But W is an implicit function W (S) of the share price S by (9). Thus (14)
to (17) can indeed be written in the form of (1).
For (14) to (17) to make sense we need for the derivative of the excess demand with
∂
respect to the price to not be zero: ∂S X (S, W, t) = 0. Economically this restriction
means that the excess demand does react to price changes. Otherwise it would not
be possible to find an equilibrium by adjusting the price and the market would be
completely illiquid.
Finally, we can use (16) and (17) to regain the lognormal random walk in Exam-
ple 1. There,

∂ 1 2 ∂
X (S, W, t) = aS0 σ̃e(µ̃− 2 σ̃ )t+σ̃W , X (S, W, t) = −a,
∂W ∂S
∂2 1
2 (µ̃− 2 σ̃ 2 )t+σ̃W ∂ 1 2
2 X (S, W, t) = aS0 σ̃ e , X (S, W, t) = aS0 (µ̃ − 12 σ̃ 2 )e(µ̃− 2 σ̃ )t+σ̃W ,
∂W ∂t
and thus by (16) and (17)
1 2
(18) σ(S, W, t) = S0 σ̃e(µ̃− 2 σ̃ )t+σ̃W
= σ̃S

and
     
1 1
µ̃− 2 σ̃ 2 t+σ̃W
1
µ̃− 2 σ̃ 2 t+σ̃W
(19) µ(S, W, t) = aS0 (µ̃ − 12 σ̃ 2 )e + aS0 12 σ̃ 2 e = µ̃S.
a

2.8. Manipulation and arbitrage. Limited liquidity in the markets changes

some of the properties of classical arbitrage opportunities as well as opening the door
to market manipulation strategies. Both have to be precluded in a consistent model.8
Market manipulation. Market manipulation in its classical sense refers to trad-
ing strategies that deliberately move the price to gain a risk-free profit.9 In section
3.4 we discuss possible price manipulation strategies, but within the model as it is
presented here they do not lead to risk-free profits (as long as there are no derivative
assets). This is not necessarily always the case.
For instance, assume the prices react with a delay to trades, and that a buy is
followed by a rise in the price and a sell is followed by a drop in the price. In this
situation a manipulator could buy large amounts of shares at the initial price S, the
price would react after this (delay!) and move up to S  > S; then the manipulator
would sell shares at the higher price S  , after which the prices would move back
down to S. This round-trip trade generated a risk-free profit for the manipulator of
S  − S > 0 per share.10 The key feature that allowed this manipulation strategy is
that the price adjusted with a delay to the trade. This enabled the manipulator to
buy at the low price before it moved up, and to sell at the high price before it moved
down. Jarrow [17] shows that to prevent manipulation strategies the price adjustment
mechanism must not exhibit delays.
8 For anecdotal evidence on price manipulation and short squeezes on NASDAQ see Morgenson
[26].
9 See Jarrow [17, 16] for a more detailed analysis of market manipulation.
10 This profit would even increase with the size of the trade.
242 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

Another much discussed possibility of market manipulation is caused by dif-

ferences in the information among the investors. An informed investor could, by
strategically releasing information to uninformed investors, influence the price with-
out having to actually trade. We exclude this possibility in our framework. Here,
the only possible influence an investor can have on the price is through trading the
share.11
Jarrow [16] gives a sufficient condition that precludes profitable market manip-
ulation in a discrete-time economy: if the price mechanism does not depend on the
history of the orders but only on the current positions of the traders, then there are
no profitable manipulation strategies.
Market corners and short squeezes. A special form of market manipulation
is the short squeeze or market corner. Here the manipulator tries to exploit a situation
in which many traders have incurred short positions in the share. In practice a short
position is incurred by borrowing shares from another investor and selling them on
the market. When the lender wants shares back (“calling in the shorts”), the short
seller has either to find another lender or to buy back the share in the market. Under
normal conditions the seller can find another lender and thus keep the short position
for as long as desired, but not in a cornered market.
In a situation with large short positions in the share a manipulator can buy up
the whole supply of the share and lend some of those to the short sellers; the short
sellers sell these shares on the market, the manipulator buys these shares as well
and ends up with a position consisting of all available shares in his or her portfolio
plus claims on the shares lent to the short sellers. The manipulator has cornered
the market. Next the manipulator calls in the shorts, knowing that the short sellers
can buy the shares they need only from the manipulator: The manipulator holds the
total supply of shares in his or her portfolio. Being in this powerful position, the
manipulator can dictate the price, and the short sellers must buy. This is called a
short squeeze.
Because they require very large positions in the share, market corners and short
squeezes are very difficult to implement. Nevertheless, especially in the 1920s and
1930s, they have been implemented in a number of cases, and even recently there
have been a number of alleged market corners.12 Short squeezes and market corners
can be excluded in the model by limiting the positions of the traders.
Arbitrage opportunities. An arbitrage opportunity arises whenever the same
thing is traded at two different prices. The “same thing” can be a derivative and its
replicating strategy or an index and the portfolio of shares it is composed of, or shares
of one company that are traded on different exchanges simultaneously. In illiquid
markets arbitrage does not mean unlimited profits (as in the fully liquid case) but
the feature of “money for nothing” remains. A simple example illustrates the point:
Assume the share is traded in two different exchanges at prices S and S  , respectively.
If these prices differ (S < S  , say) this would mean an arbitrage opportunity: buy
the share in the cheaper market and sell it where the price is higher, cashing in the
difference. In perfectly liquid markets this would generate infinite profits as the size
of the trade goes to infinity.

11 Differential information may still be present; we only exclude the possibility of influencing the

price without trading. A more thorough treatment of markets with differential information can be
found in O’Hara [26].
12 See Jarrow [16, p. 311] for examples and references.
 FEEDBACK IN ILLIQUID MARKETS 243

If the markets involved are illiquid, taking the arbitrage would move the prices
together and the arbitrage opportunity would disappear as it was taken. Buying in
the cheaper market will push up the price in this market while selling at the higher
price would lower that price. There will only be limited scope for taking the arbitrage
here.
Summing up, in illiquid markets absence of arbitrage is not the only necessary
condition for a market model to be consistent; we also need absence of manipulation
strategies and market corners.
3. Dynamic trading strategies. In the previous section we set up a market
model that describes how market prices react to a trading strategy f . Now we take a
closer look at the trading strategy itself, and the peculiarities introduced by illiquidity
of the markets. We have to generalize the classical concepts of the value process of a
trading strategy and the concept of a self-financing trading strategy to illiquid mar-
kets. Furthermore, we take a look at examples of trading strategies used in practice,
with a special emphasis on option replication trading strategies.
We need these results in later sections to derive and analyze results about option
replication trading strategies, but the concepts introduced hold in more general sit-
uations than the model of section 2. Therefore, we present them with more general
notation.13
We use the following notation:
• ft denotes the holdings in the share,
• ct denotes the holdings in the bond,
• Yt denotes the (paper) value of the portfolio, and
• Yt denotes the real value of the portfolio.
A dynamic trading strategy is defined by fully specifying the processes ft and ct as
functions of the information available. We assume that the trading is nonanticipatory
and f and c are predictable processes.
The value process. At any time t the (paper) value of the portfolio is

(20) Yt = ft St + ct Bt ,

the sum of the value of the shares and the bonds in the portfolio, valued at the current
market prices. This definition only holds if the markets are perfectly liquid. If the
market for shares is illiquid, St will not be the price at which one will be able to sell
the ft shares in the portfolio, and Yt will only represent paper wealth. Therefore, we
also define the liquidation value or real value Yt of the portfolio in illiquid markets.
This is the value of the portfolio if the share position were to be unwound in the next
instant. The real value of a portfolio is discussed in section 3.3.
Self-financing trading strategies. Self-financing trading strategies are trading
strategies that do not require cash inflows or outflows after they have been set up.
Buying and selling of shares is financed by selling or buying of bonds in the same
portfolio.
In a quasi-discrete approximation the time sequence of events is as follows:
1. At time t− := t − dt the portfolio contains ft− shares and ct− bonds and has
a value of Yt− = ft− St− + ct− Bt− .
13 For intuition, the reader can still think of all processes as functions of the state variables (t, S, W )

(i.e., ft is really f (t, S(t)), etc.) but the definitions also hold for general adapted stochastic processes
that are not Markovian, for both trading strategies ft , ct and price processes St , Bt .
244 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

2. Then (between t− and t) the prices of the securities change to the new prices
St = St− +dSt and Bt = Bt− +dBt . The mechanism by which the changes in price are
achieved is left open in standard finance; we modeled this explicitly in the previous
section.
3. This causes a change in value in the portfolio of ft− dSt on the shares and
ct− dBt on the bonds held.
4. Trading also takes place, but it takes place at the new prices St and Bt ,
respectively. The number of shares bought is dft , costing St dft , and the number of
bonds bought is dct , costing Bt dct .
5. After trading has taken place the new value of the portfolio is Yt = ft St + ct Bt .
This can be written as

dYt = Yt − Yt− = ft St − ft− St− + ct Bt − ct− Bt−

= ft St − ft− St + ft− St − ft− St− + ct Bt − ct− Bt + ct− Bt − ct− Bt−
= St dft + ft− dSt + Bt dct + ct− dBt .

A self-financing strategy satisfies St dft + Bt dct = 0: the cash-flow generated by selling

(buying) one security exactly finances the buying (selling) of the other security. The
value process is then

(21) dYt = ft− dSt + ct− dBt .

Furthermore, for any given share trading strategy f one can find the corresponding
strategy c in the bond that makes the joint strategy self-financing.
Although the definition of the value of a portfolio will be modified in illiquid
markets, the concept of a self-financing trading strategy will not change. The key
property is absence of external inflow/outflow of funds, which is unaffected by the
question whether S is influenced by the large trader or not. All that counts is that S
is the price at which trading takes place.
3.1. Trading strategies: Some examples. In modern financial markets many
dynamic trading strategies are in use. In this section we will give some examples which
will illustrate the concept of a trading strategy. As mentioned before we will mainly
restrict ourselves to trading strategies that can be written as a function f (S, t) of
the share price S and time t, but some more general trading strategies (e.g., some
technical trading strategies) can be considered in the same framework.
Buy and hold f (S, t) = 1. This strategy involves holding one share and no trading
at any time.
Limit orders. A sell order with a constant limit S̄ can be considered as a trading
strategy: The strategy is to sell one share if the share price moves above or to S̄. In
this case the function f (S, t) has the following form: Until (and including) the time
of the sale, H(S̄ − S) is the Heaviside function with a step from 1 to 0 at S = S̄.
This means holding on to the share until S = S̄. After the sale it is f (S, t) = 0: do
nothing. Similarly, a buy order with limit S̄ is f (S, t) = H(S̄ − S) before the buy and
f (S, t) = 1 after the buy.
Trading through the barrier. Here f (S, t) = H(S−S̄). In words: When S > S̄ hold
one share in your portfolio, when S < S̄ have no shares in the portfolio. Whenever
the share price crosses S̄ from above, the share is sold; when it crosses from below, a
share is bought.
Portfolio-optimization trading strategies. An example for these strategies can be
found in Merton [24, p. 97 ff.], who analyzes the optimum investment rule for an
 FEEDBACK IN ILLIQUID MARKETS 245

investor with constant absolute risk aversion utility function when the security price
process is lognormal. The optimum investment strategy in this case is always to hold
a fixed proportion of one’s wealth in the share. Combined with the resulting wealth
processes, this gives an investment strategy of the form f (S, t) = c0 e−c1 t S −c2 , where
c0 , c1 , c2 are positive constants depending on the specification of the model. More
complex rules can be derived for other specifications of the investor’s preferences
or investment opportunities by setting up the respective optimization problem and
solving the corresponding Bellman equation. The derivation of these trading strategies
is now standard in finance theory (see, e.g., Bick [4, 3] for an application and Duffie
[9] for general reference).
Chartist trading strategies. Simple chartist trading strategies can be written with
an extended state vector x, which includes chart indicators like moving averages or
running maxima/minima, etc. The trading strategy can be described as f (S, x, t) = 1
if the chart theory gives a “buy” signal for the state x, share price S, and time t
(the chartist wants to be long the share), and f (S, x, t) = −1 if the chart theory
recommends selling. For instance, let x be a moving average of the share price. Then
a simple chartist trading strategy would be f (S, x) = 2H(S − x) − 1; i.e., if the share
price has crossed the moving average and is above it now (S ≥ x), then the chartist
expects it to rise further and wants to be long in the share: f (S, x) = 1; and if the
share price is below the moving average S < x, then he wants to be short in the share:
f (S, x) = −1.

3.2. Option replication in liquid markets. Using the definitions of the pre-
vious sections we can derive more sophisticated trading strategies that replicate the
payoff of derivative securities in liquid markets. These dynamic trading and hedg-
ing strategies are now standard in option pricing theory; see, for instance, Wilmott,
Howison, and Dewynne [31] or Duffie [9].
As a canonical example we will consider the replication of a European Put option.
Assume the share price follows the diffusion process given in (1):

dS = µ(S, t) dt + σ(S, t) dW.

We assume there is a self-financing trading strategy (f, c) that replicates the payoff of
a European Put option, i.e., a strategy whose value process Y at expiry T coincides
with the payoff of the option almost surely. To prevent arbitrage opportunities, the
value process Y (S, t) = Sf (S, t) + Bc(S, t) of this trading strategy must then be equal
to the market price P (S, t) of the Put option at all times before expiry, too. Otherwise
a risk-free profit could be made by buying cheaper and selling the more expensive of
the two equivalent investments: “option” and “trading strategy.”14
Because the trading strategy is self-financing we have

dY (S, t) = f (S, t)dS + rc(S, t)B(t)dt.

Substituting from Y = f S + cB for cB yields

dY (S, t) − rY (S, t)dt = f (S, t)dS − f (S, t)rSdt.

14 We assumed that in this trading strategy f and c can be written as functions f (S, t) and c(S, t)

of share price S and time t.

246 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

From the no-arbitrage condition, P (S, t) = Y (S, t) follows then with Itô’s lemma:
 
∂ ∂ ∂2
P (S, t)dS + P (S, τ ) + 12 σ(S, t)2 2 P (S, t) − rP (S, t) dt
∂S ∂t ∂S
= f (S, t)dS−f (S, t)rSdt.

For the increments with respect to dS to cancel we need

∂
(22) f (S, τ ) = P (S, τ ).
∂S
The number of shares to be held in the portfolio has to be equal to the partial deriva-
tive of the option value with respect to S, the so-called Delta of the option. Following
this strategy the stochastic terms in the option price process and the portfolio value
process cancel exactly, and the share half of the trading strategy is sufficient to hedge
the stochastic component of the change in value of the option. The bond investment
strategy is necessary only to make the strategy self-financing and to reproduce the
nonstochastic components of the change in value of the option. Therefore, practic-
tioners often ignore the bond half of the trading strategy and implement only the
share part.
After substituting (22) for f , we reach the following condition to cancel the other
integrals (with respect to dt):

∂ ∂ ∂2
(23) 0= P (S, t) − rP (S, t) + rS P (S, t) + 12 σ(S, t)2 2 P (S, t).
∂t ∂S ∂S
This is the well-known Black–Scholes partial differential equation: The price of the
Put option P (S, t) must satisfy the Black–Scholes partial differential equation, subject
to the final condition that the option price must be the option payoff at expiry:
P (S, T ) = (E − S)+ , and appropriate boundary conditions15 at S = 0 and S → ∞.
Specifying the lognormal random walk (2) for the share price process

dS = µ̃S dt + σ̃S dW,

Black and Scholes [5] found an explicit solution to this partial differential equation, a
formula for the value P (S, t) of an European Put option depending on the share price
S and time t. It is

(24) P (S, t) = Ee−r(T −t) N (−d2 ) − S N (−d1 ),

where

log(S/E) + (r + σ̃ 2 /2)(T − t)
d1 = ,
σ̃(T − t)1/2
d2 = d1 − σ̃(T − t)1/2 .

Here E denotes the exercise price and T the expiry date of the option, and N (x) is
the cumulative normal distribution.
15 See Wilmott, Howison, and Dewynne [31] for details.
 FEEDBACK IN ILLIQUID MARKETS 247

In their derivation Black and Scholes assumed that the share price follows a
lognormal diffusion process of the form (2), absence of transaction costs, continuous
trading and perfectly liquid makets. 16
For the European Put equation (24) and (22) give the replication trading strategy

f (S, t) = N (d1 ) − 1.

This is the trading strategy we are going to analyze in more detail later on. Typical
plots of f are given in Figure 1. This figure shows the delta of a European put at
various times before expiry; the steeper the curve, the closer the option is to expiry.
3.3. Real value, paper value, and the trading of large blocks. In illiquid
markets there is a distinction between the paper value and the real value of a portfolio.
Given a portfolio of ft shares and ct bonds at time t the paper value of the portfolio
was defined in (20) as

(20) Yt = ft St + ct Bt ,

the value of the portfolio at the current market prices. We define the real value of
the portfolio f, c as the value of the c bonds plus the liquidation value of the f shares
using the optimal instantaneous liquidation strategy. We will derive this liquidation
strategy below:
Let the share price St− at t− := t − dt be defined by (9):

(25) X (St− , Wt− , t−) + ft− = 0.

If the large trader decides in t to liquidate his holdings in the share, i.e., places an
order of dft = ft − ft− = −ft− , then the new price St is determined by

(26) X (St , Wt , t) = 0

and at this price the transaction is executed, giving a revenue of St ft− to the large
∂
trader. But, because of the slope ∂S X (S, W, t) < 0 of the excess demand function
the new price St will be much lower than St− , and the selling of the large trader has
depressed the market price to St < St− . This will always happen and the real value
of the portfolio must be lower than the paper value.
But by choosing a suitable liquidation strategy the large trader can minimize the
loss due to the adverse price effect of his trades. Assume the shares are not sold as
one block but in two halves: the first half will then be sold at S  , which is defined by

(27) X (S  , Wt , t) + 12 ft = 0,

and the second half will be sold at St from equation (26). Again because of the slope
of X we have St− > S  > St , and the first half of the shares was sold at a better
price than St . This procedure can be repeated to show that the best strategy for
liquidation of a portfolio is to sell off the shares in a rapid sequence of small orders.
Call the solution of (9) for the share price S̃(f ; Wt , t); i.e., S̃ is defined by

(28) X (S̃(f ; Wt , t), Wt , t) + f = 0 ∀f

16 The liquidity assumption is in fact contained in assuming the stochastic process (2) for S

independent of any trading.

248 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

∂
and by the inverse function theorem ∂f S̃(f : W, t) > 0. Then the real value of ft
shares is
 ft
(29) S̃(f  ; Wt , t) df  ,
0

1
 ft
with an average liquidation price of S̄t = ft 0
S̃(f  ; Wt , t) df  , and the real value of
the portfolio is
 ft
(30) Yt = S̃(f  ; Wt , t) df  + ct Bt = S̄t ft + ct Bt ,
0

which is still less than the paper value Yt . With this specification the real value of a
portfolio cannot be influenced by self-financing trades in the share: by (29), buying
another df shares adds S̃df to the real value of the shares in the portfolio, but this
is also the amount of money that has to be paid for these shares. This money must
come from selling bonds in a self-financing portfolio, which means that the net effect
of the trade on the real value is zero.
In reality the continuous selling of shares will take some time and cannot be done
instantaneously, but within this model it can be done arbitrarily quickly. The path
of the Brownian motion W is continuous in time and X is continuous, too. Therefore
this liquidation value Yt can be approached in the limit.
Obviously this strategy can also be used to buy a large block of shares instead
of selling it. A round trip (i.e., buying and then selling a block of shares at the
same time) using these continuous strategies is cost free, because the real value of the
portfolio is not affected.
An upper bound for the profitability of a round trip must be zero; otherwise
the large trader could make infinite profits by doing round-trip trades. Therefore
the liquidation strategy described above is optimal; it achieves the upper bound on
profitability.
3.4. A price manipulation strategy. If aware of the equilibrium relation (9),
the large trader is able to move the price to any level desired by buying or selling
the appropriate amount of shares. The large trader is even able to achieve this with
vanishing cost and risk: Assume the large trader wants to move the price from St at
t to S ∗ at T . Given the value of WT at T , the number of shares necessary to reach
the target price is defined by

(31) f ∗ = −X (S ∗ , WT , T ),

which is only a function f ∗ (W ) of the value of W at time T . Using the continuous

buying strategy of the previous section, the large trader will buy f ∗ shares at T − dt
at a cost of
 f∗
− S̃(f  ; WT −dt , T − dt) df  ,
0

hold these shares over T , and sell them at T + dt for

 f∗
S̃(f  ; WT +dt , T + dt) df  .
0
 FEEDBACK IN ILLIQUID MARKETS 249

Because W is continuous, in the limit WT −dt ≈ WT , thus the large trader knows f ∗
already at T − dt and the strategy can be executed. The profit/cost of this strategy
is
 f∗  
1
S̃(f  ; WT +dt , T + dt) − S̃(f  ; WT −dt , T − dt) df  = O dt 2 ,
0

which will go to zero as the strategy is executed quickly enough. There are several
possible criticisms of this manipulation strategy.
First, the quasi-continuous buying and selling of shares may not be possible in
reality. The manipulator will have to pay higher prices when buying and lower prices
when selling the shares, and therefore there will be some cost to the round trip. This
is true to some extent; on the other hand it has been observed that in practice brokers
execute large block transactions in exactly this way: they are split up into smaller
pieces and then these smaller pieces are sold on the market. Furthermore, delayed
publication rules for large block transactions make this execution strategy for large
blocks easier and cheaper. Many exchanges use such rules.17
Second, this manipulation is easily spotted; one might therefore expect manipu-
lations of this kind to be banned in practice. This ban would have to be imposed by
an independent observer (e.g., the SEC) who observes the prices and punishes manip-
ulators. It is true that this particular manipulation strategy is easily identified, but
unfortunately there are more sophisticated trading strategies that cannot be identified
from the outside.
As an example18 consider the Brownian motion Example 2. The excess demand
function is simply linear:

X (St , Wt , t) = a(Wt − St ),

leading to a Brownian motion as unmanipulated price process: dSt = dWt . Now

consider the trading strategy

S ∗ − St
dft = dt,
T −t
which, when added to the excess demand, will lead to the manipulated price process

S ∗ − St
dSt = dt + dWt .
a(T − t)

This is known as a Brownian bridge from S0 to S ∗ at T . It is well known19 that

a Brownian motion and a Brownian bridge are indistinguishable if the target point
of the Brownian bridge is normally distributed with mean zero and variance T . In
our situation a regulator will have to decide from one single sample path whether
this path was a Brownian motion (normal market behavior) or a Brownian bridge to
its endpoint (the market has been manipulated). This is an impossible task without
knowing the target of the manipulator beforehand.

17 See, e.g., Board and Sutcliffe [6] for block transactions and publication rules at the London

Stock Exchange.
18 This example is adapted to our modeling environment from Back [1]. The discrete-time original

is due to Kyle [19].

19 See Karatzas and Shreve [18].
250 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

The reason for this is that the manipulator can hide his manipulation trades
(which move the price by an amount√of order dt) successfully behind the normal mar-
ket movements (which are of order dt). Obviously similar manipulation strategies
could be derived in other markets. This manipulation strategy will involve price ex-
posure for a limited period of time, but again, within the model it can be performed
arbitrarily quickly.
Thus in our model a price manipulation in the share market can be performed
by the large trader. In reality this manipulation will probably not be cost free, but
manipulation remains possible (even if it is not profitable). We will see below that in
the presence of derivative securities, price-manipulation strategies may become very
profitable.
4. Option pricing.
4.1. Options in illiquid markets. If there are derivative securities that condi-
tion on the prices in illiquid markets some new problems arise. Typically the market
for the option will collapse because of the possibility of manipulation in the share
price by the large trader. At expiry (as at any other time) the share price is now
basically a decision variable of the large trader which can be freely manipulated and
thus the trader can influence the payoff of the derivative as desired.
The cash payoff of the derivative security is denoted by P1 (ST ) and the phys-
ical payoff by P2 (ST ). Let us consider the example of a European Call option:
A European Call option with exercise E and cash settlement would be defined by
P1 (ST ) = (ST − E)+ and P2 (ST ) = 0, while the Call with physical delivery would
be defined by P1 (ST ) = −EH(S − E) and P2 (ST ) = H(S − E). (Because of the
difference between real value and paper value we have to distinguish between cash
and physical settlements of the option.)
4.1.1. Cash settlement. Options with cash settlement are characterized by
P2 (ST ) = 0. As shown in the preceding subsection, the large trader can manipulate
the share price at time T , cost- and risk-free, to any value desired. Therefore the
large trader will manipulate ST to

arg min P1 (ST )

ST

if short on the option and to arg maxST P1 (ST ) if long on the option.20 In the case
of the Call option the manipulator would manipulate the share price into the out-of-
the-money area if short on the option and would manipulate the share price as far
into the money as possible if long on the option.
There is no need for him to hedge his position because the manipulation enables
him to avert any suboptimal outcome. The danger of this manipulation will pre-
vent any nonmanipulator from trading the option and the market for the option will
collapse. The only derivative security that is still viable in this environment is one
with P1 (ST ) = c a constant payoff. But this is a bond and does not really qualify as
derivative security.

20 For a function g(x) and a set M the arg min is defined as the x ∈ M that minimizes g(x)

(assuming it is unique):

x∗ = arg min g(x) ⇔ g(x∗ ) ≤ g(x) ∀ x ∈ M.

M

The arg max is defined analogously.

 FEEDBACK IN ILLIQUID MARKETS 251

4.1.2. Physical settlement. Similarly the large trader can manipulate the pay-
off of Call options with physical settlement. First consider the case when the large
trader has a long position in the Call option and the unmanipulated share price is
below the exercise price of the option. Here it would not be sensible for the manipu-
lator to manipulate the price into the money, because then it would be necessary to
buy the share at the exercise price (which is higher than the price at which it could
have been bought before the manipulation).
If the unmanipulated share price were above the exercise price a manipulator
with a long position would not manipulate the share price either, because pushing
the share price out of the money would destroy the real value of the option and any
manipulation within the money would not change the option’s payoff.
Unfortunately the large trader has an incentive to manipulate the share price if
there is a short position in the option. If the option threatens to expire in the money,
the large trader can manipulate the share price down and thus avoid having to deliver
the share at price E.
If the unmanipulated share price were below the exercise price the manipulator
could push up the share price just above the exercise price and then sell the shares
already bought to push the price up to the holders of the option at the exercise price.
The exercise price in this case will be above the average price the manipulator paid
for the shares to push the price up.21
Generally, given the unmanipulated share price Su and the payoff functions P1 (S)
and P2 (S), a large trader with a position of fo options will not manipulate the share
price at expiry iff

(32) arg max Y  (fo P2 (S), Su ) + fo P1 (S) = Su ∀Su , fo ;

S

i.e., the real value Y  of the physical payoff plus the cash payoff cannot be increased
by manipulating the share price from Su for all unmanipulated share prices and for
all positions of the large trader. The bond and the share itself satisfy this equation
but most options do not.
One might argue that the small traders will recognize these manipulation strate-
gies and therefore still exercise the option (even if the large trader has manipulated
the price out of the money), because the real value of the share is more than the
exercise price E. This argument has several problems.
First, why should the holders of the option exercise it (which is equivalent to
buying at the exercise price E), if they could also buy at the manipulated price in the
market instead?
Second, the “real value” of the share that the holders of the option get by exercis-
ing can be realized only after the large trader has unwound his manipulation position.
This may take some time and waiting for it will definitely involve some risk. The
real value of shares in a manipulated market depends on the manipulator’s strategy
(unless one is prepared to sell at the manipulated prices).
Third, if the small traders are aware of the manipulation and can anticipate the
large trader’s behavior, other strategic considerations come to mind. They will know
beforehand at which level S ∗ the share price will end up at time T . They will try to
profit from that by buying the share if its price before T is below S ∗ and selling it if
its price is above S ∗ , with the intention of covering the position at price S ∗ at time
21 This manipulation assumes that the number of shares required to push the price up to the

exercise price is approximately equal to the number of shares the manipulator will have to deliver.
252 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

T . Ironically this will relieve the large trader of the need to manipulate the market,
because this trading strategy of the small traders will move the price to S ∗ at T itself.
The large trader’s threat to manipulate the market suffices to encourage enough small
traders to jump the bandwagon and do the work themselves.
Finally, it should be pointed out that in this model we do not allow the small
traders to react to the large trader’s trading strategy, as is implicitly suggested in
the arguments above. This is incorporated in the fact that the small traders’ excess
demand X (S, W, t) is independent of the large trader’s trading strategy f (S, t). Thus
the model as it is presented here assumes that the small traders are either unaware of
the large trader’s actions or are unable to condition their actions on the large trader.
In principle it is possible to build a full model of strategic interaction between large
and small traders in the stock market along the lines of current market microstructure
research and game theory (see, e.g., Back [1, 2] for a very promising first step in this
direction). These models tend to become highly complicated even in simple two-period
discrete-time frameworks. Here these questions lead away from the aim of this paper.
4.1.3. Consequences for option pricing. The arguments presented above
apply not only to European Call options but in principle to any derivative security
whose payoff depends on the share price.
Large traders who are aware of their market power will manipulate the share
price to their favor. If there is a cost to the share price manipulation, the large
trader will take this cost into account when choosing a manipulation strategy, but will
still manipulate (maybe to a different extent). Because of these manipulations the
derivatives lose their character as risk-management instruments; their payoff becomes
largely independent of the risk they are supposed to hedge.
If the small traders are aware of the presence and market power of the large trader
the risk of manipulation will make them unwilling to trade derivatives any more, and
the option market will collapse. Option pricing in the classical sense is no longer
possible; the large trader will value the option according to optimal manipulation
strategy, and the small traders will do the same. The notion of pricing by using a
replicating trading strategy loses its power if the price process itself becomes subject
to manipulation.
The option payoff manipulations are always possible in models that take the
market’s reaction function to the large trades as given. The model presented here, as
well as the models by Jarrow [17, 16], Frey and Stremme [12], Frey [10, 11], and Sircar
and Papanicolaou [30] all suffer from the possibility of these payoff manipulations.
Several ways around this problem have been implicitly suggested in the literature.
For example, Jarrow [17, 16] assumes that the price moving power of the large trader
is limited to a small environment of the equilibrium price and gives bounds beyond
which the large trader may not move the price. Unfortunately these bounds lack
further justification, and within the bounds the large trader will still manipulate as
much as possible.
Another possibility (e.g., Back [2] and Frey [11]) is to assume perfectly liquid
markets at expiry of the option, which is usually done by assuming a fixed final
“true” value of the share. At this price unlimited amounts of the share can be traded
without manipulation. Although it is clearly unrealistic this assumption is beneficial if
information aspects of the model are to be studied because it allows us to analyze the
incorporation of information into the price over time; i.e., “Given that some traders
are informed and know the true value beforehand, how quickly does the price converge
to its ‘true’ value?” (This is the case in most market microstructure models, e.g., Kyle
 FEEDBACK IN ILLIQUID MARKETS 253

[19] or Back [1, 2].) For option pricing alone, though, there is no reason why at the
expiry date of the option the share price should be perfectly liquid, and illiquid before.
For exotic options (Asians, barriers, etc.) this approach fails completely.
4.2. Payoff replication in illiquid markets. Despite the very “destructive”
results of the previous section, there are some cases in which the “pricing by replica-
tion” paradigm is still valid in illiquid markets.
4.2.1. Price takers’ pricing. One is taking the point of view of a small trader
to whom the markets are still liquid. Consider an illiquid market that is influenced
by a large trader (or a large group of small traders) following a given trading strategy
f , and a small trader wants to know how to price an option given the large trader’s
modification of the price process. For some reason the large trader is not able to
hold the option or to manipulate the payoff, either because the option is only traded
over the counter (OTC) or because the large trader is really a large group of small
traders who happen to follow the same trading strategy but who cannot coordinate
for strategic market manipulation.
The smallness of the small trader means that he or she can basically trade any
amount of shares on his or her scale without influencing the price, that the price
process from (14) and (15) can be taken as given, and that the paper value and real
value of the portfolio coincide. Thus the argument of section 3.2 can be repeated,
whence we will arrive at the following modification of the pricing partial differential
equation (23):
∂ ∂ 1 ∂2
(33) 0= P (S, t) + rS P (S, t) + σ 2 (S, t) 2 P (S, t) − rP (S, t),
∂t ∂S 2 ∂S
where the volatility is defined by (14) as
∂
∂W X (S, W, t)
(14) σ(S, t) = − ∂ ∂
.
∂S X (S, W, t) + ∂S f (S, t)

Final and boundary conditions apply in the usual way. This case is a simple general-
ization of the classical Black–Scholes argument. See the appendix for an asymptotic
analysis of this case in the framework of Example 1.
4.2.2. Paper value replication. The other case is the case when the large
trader is aware of the market’s illiquidity but abstains from manipulation and wants
to find a self-financing trading strategy that replicates a given payoff function. This
may be the case when the large trader has some exogenous exposure to hedge (without
a counterparty that can be defrauded by manipulation). Now the trading strategy f
is not given but has to be found as part of the problem.
Assume the paper value payoff at time T that is to be replicated is given by F (S),
and that there exists a self-financing trading strategy f, c whose final paper value is
F (S). We assume that this trading strategy f (S, t), c(S, t) and its paper value process
Y (S, t) can be written as functions of S and t. Furthermore, the price process S is
continuous and given by (15) and (14).
Then the argument used to derive the option pricing trading strategy in section
3.2 carries through as before. Again we reach (33), this time for Y (S, t), the initial
paper value of the self-financing trading strategy:
∂ ∂ 1 ∂2
(34) 0= Y (S, t) + rS Y (S, t) + σ 2 (S, t) 2 Y (S, t) − rY (S, t),
∂t ∂S 2 ∂S
254 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

∂
Now we have to substitute for σ(S, t) from (14) using f (S, t) = ∂S Y (S, t) to reach
(35)

2
∂
∂ ∂ 1 ∂W
X (S, W, t) ∂2
0= Y (S, t) + rS Y (S, t) + Y (S, t) − rY (S, t).
∂t ∂S 2 ∂
∂S
X (S, W, t) + ∂2
∂S 2
Y (S, t) ∂S 2

By solving (35) with final condition Y (S, T ) = F (S), one can find the trading strategy
to replicate the paper value payoff F (S). The nonlinearity in (35) is introduced
by the feedback effects between the share price process and the trading strategy.22
This partial differential equation was proposed for option valuation by the authors in
[28] and [29]; equivalent equations have been derived and analyzed by Frey [10, 11]
(existence and uniqueness) and Sircar and Papanicolaou [30] (numerical solution). An
asymptotic analysis of this equation can be found in Appendix B2.
Again it should be pointed out that Y (S, t) above does not represent the value of
the option paying off F (S) at T . There is the possibility of manipulation as explained
in the previous section, and the value function and the payoff are given in paper value
and not in real value.
5. The influence of hedging. We now analyze in detail the effects that can
arise when dynamic trading strategies are used in illiquid markets using the exam-
ple of Put-replication trading in the market setup of Example 1. There are many
small hedgers that all replicate a European Put option in an illiquid market with
linear price reaction to the trades. The small traders are unaware of the illiquid-
ity of the market and we group the small trader’s strategies to one “large trader.”
Using this example we can show several points. First, we demonstrate again how
to derive a modified price process; second, we see that the volatility of the price
process is increased; and finally, in this example the nature of the price process is
changed dramatically (by the introduction of discontinuities or jumps in the price
process).
If there were no feedback effects, the share price process would be a lognormal
random walk,
(36) dS  = µ̃S  dt + σ̃S  dW,
where S  denotes the undisturbed price process:
S  := S0 exp{(µ̃ − 21 σ̃ 2 )t + σ̃W }.
The partial derivatives of X are
∂X ∂2X ∂X ∂X ∂2X
= aσ̃S  , 2 = aσ̃ 2 S  , = −a, = a(µ̃ − 12 σ̃ 2 )S  , and = 0.
∂W ∂W ∂S ∂t ∂S∂W
Under the influence of the program trader this price process is modified according
to (14) and (15) to become dS = µ(S, t)dt + σ(S, t)dW with
aσ̃S 
(37) σ(S, t) = ∂
a − ∂S f (S, t)
and
 
1  ∂ 1 a2 σ̃ 2 S 2 ∂2
(38) µ(S, t) = aµ̃S + f (S, t) + f (S, t) .
∂
a − ∂S f (S, t) ∂t ∂
2 (a − ∂S f (S, t))2 ∂S 2

22 Equation (35) may seem to be three-dimensional, but by the equilibrium relation (9) W can be

expressed in terms of S and f = ∂Y

∂S
.
 FEEDBACK IN ILLIQUID MARKETS 255

5.1. The importance of Put-option replication. We chose the Put-option

replication trading strategy for several reasons:
European Put (and Call) options are (after Futures contracts) the most popular
derivative securities by far. The net effect of the increasing popularity of the basic
Put and Call options is an increase in hedging of short Put and Call positions or,
equivalently, of replication trading of long positions in Puts and Calls in the funda-
mental market. As argued in the introduction, this is due to the fact that the buyers
of these options are typically not able to implement a dynamic trading strategy that
would replicate these options; being end-users of these risk management instruments
they use the options to insure themselves against fundamental risks they previously
left unhedged. The issuers of these options, on the other hand, are usually financial
institutions that want to hedge against the risk from writing the option. They are
able to implement the respective hedging trading strategy and do so.
Therefore, even though there is the short position of the issuer to offset the long
position of the investor, the investor will not implement a replication trading strategy
because he or she is already hedged, but the issuer will implement a trading strategy
to hedge the short position: the issuer does not have any fundamental risk that is
offset by the short position.
The trading strategy for Call option replication is equivalent to the trading strat-
egy for Put-option replication: It is fC (S, t) = N (d1 ) for the Call as opposed to
fP (S, t) = N (d1 ) − 1 for the Put. Therefore analyzing Put replication is sufficient to
capture this motivation of replication trading.
The European Put-option replication trading strategy is often used by fund man-
agers as Portfolio Insurance trading strategy.23 Following such a trading strategy
offers better protection against downward movements in the share price than a stop-
loss order, as the exposure is dynamically adjusted. McMillan [23] ranks portfolio
insurance trading among the three most important automatized trading strategies.
(The other two are index arbitrage and the execution of large block transactions.)
Finally, as shown above, replication strategies are closely connected to option
pricing issues. From analyzing the effect of this strategy in illiquid markets it is a
small step to analyzing the option pricing in these markets.
5.2. Scalings and parameters. To analyze the behavior of the modified price
process (37) and (38) we make the variables and equations dimensionless.
We let
∗
S = S0 S ∗ , S  = S0 S  , t = T t∗ , X = N X ∗ , f = N ψf ∗ .
Here S0 is a typical value of the share price, a constant; T is a time scale, the time to
maturity of an option or the time horizon of a portfolio insurer; and the dimensionless
quantity ψ is a small number and denotes the size of the portfolio relative to the scale
of the excess demand N . The asterisk denotes a dimensionless variable. No further
scalings will be necessary and so we henceforth drop the asterisks.
Thus we reach the dimensionless form of the modified equilibrium condition
(39) α(S  − S) + ψf = 0,
where the (dimensionless) parameter α is
aS0
α= .
N
23 See, e.g., Leland [20, 21], Brennan and Schwartz [8], Bookstaber and Langsam [7], and McMillan
[23].
256 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

The liquidity of a market can be defined according to the value of α:

A liquid market is a market in which α is large. An illiquid market

is a market in which α is small.

In the undisturbed equilibrium with appropriate choice of scalings α is also equal

to the price elasticity of demand which is defined as (S/D)(∂D/∂S). Large α means
in this context that a small relative change in the price causes a large relative change
in demand. α−1 is the demand elasticity of the price. Genotte and Leland [14] found
that α can vary enormously between markets; they found values ranging from 0.05 to
14 depending on the informational structure of the market.
Now define
ψN
= ψα−1 = ,
aS0
the product of the size of the portfolio (as a fraction of the total demand) and the
demand elasticity of the price. The dimensionless parameter will play a central role
in the further analysis. It measures the size of the influence of the trading strategy f
on the market and is typically small. Thus (39) becomes
(40) (S  − S) + f = 0.
With these scalings the modified price process ((14) and (15)) becomes
dS = µdt + σdW
with µ and σ given by

1  ∂ 1 σ̃ 2 S 2 ∂2f
(41) µ= µ̃S + f+
1− ∂
∂S f
∂t 2 (1 − ∂S f )2 ∂S 2
∂

and
σ̃S 
(42) σ= ∂
.
1 − ∂S f
Again, (41) and (42) reduce to the lognormal random walk for = 0, confirming the
choice of scalings.
5.3. The influence of hedging. A typical form of the f of the trading strategy
replicating a European Put is displayed in Figure 1. This figure shows the delta of a
European Put option at t = 1.0, 0.5, 0.1, and 0.01 before expiry. As the expiry date
of the option is approached f goes towards a step function with step from −1 to 0 at
the exercise price. This corresponds to the payoff diagram of a European Put whose
derivative is exactly this step function.
Recalling (40) the equilibrium condition can be written as
(43) −S  = −S + f.
The effect of the trading strategy is a small (i.e., of order ) perturbation added to
the original demand function. Far from expiry the right-hand side of (43) is simply
−S since f is small. Close to expiry the f term becomes important and the shape of
the demand curve alters dramatically, becoming as shown in Figure 2. Figure 2 shows
the right-hand side of (43) for the Put-replicating strategy (with = 1 to magnify the
effects).
 FEEDBACK IN ILLIQUID MARKETS 257

∂
Fig. 1. The Put-option replication strategy: f (S, t) = ∂S
P (S, t; E, T, r, σ), with r = 0.1,
σ 2 = 0.2, E = 10 for T − t = 1, 0.5, 0.1, and 0.01.

Fig. 2. The sum of the original linear demand function and the extra demand f (S, t) due to
put replication.

As expiry approaches, the demand curve’s slope becomes less and less negative in
an interval around the exercise price until it is positively sloped for times very close
to expiry. This signifies unstable equilibria there.
Examination of the drift µ and variance σ of the modified price process (41) and
258 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

(42) yields the same result: both have a term of the form

∂
1− f (S, t)
∂S
in the denominator. This is the negative of the derivative of the total excess demand
function X + f . When this derivative becomes equal to zero the excess demand
function has zero slope. If the demand curve has zero slope it is locally impossible to
find a market clearing price with small adjustments of the price; the price adjustments
have to be of a larger order of magnitude. But this point is also where µ and σ
approach infinity.
∂f ∂f
But even when ∂S < −1 , a positive ∂S has the effect of increasing both σ and the
∂f
absolute value of µ: the market becomes more volatile. If conversely ∂S is negative,
its effect is to decrease the volatility of the market. If f is regarded as the trading
strategy to replicate a derivative security V , the relation f (S, t) = ∂V
∂S (S, t) yields that:

Replication of derivative securities V with positive gamma

∂2
Γ = ∂S 2 V (S, t), i.e., with concave payoff profile, destabilizes

the market of the underlying security.

Long positions in put and call options have positive gamma.

Geometrically, (43) can be interpreted as follows: For a given t, the market equilib-
ria are defined as the points where the randomly moving horizontal line −S  intersects
the function −S + f .
As long as −S + f has no local minimum or maximum the equilibrium is unique
for any value of S  . But because f approaches a step function as expiry is approached,
−S + f must become multivalued at some point. From this point on the time axis
there are three possible equilibria of which two are stable and the one between them
is unstable.
Let us now consider further what happens in these situations (Figure 3). The
horizontal line denotes the left-hand side in the equilibrium condition (43) and the
curve is the corresponding right-hand side for t close to expiry T . There are four
possible situations depending on the value of S  at that point in time.
1. Here S  is outside the critical region where there are several equilibria. The
equilibrium E1 is unique.
2. In this case there are three feasible equilibria: E1 , E2 , and E3 . E1 and E3
correspond to a negatively sloped demand function, they are stable. E2 corresponds
to a positive slope in the demand function. As argued previously, this is an unstable
equilibrium.
3. This is the limiting case between cases 1 and 2. There are exactly two
equilibria, E1 and E2 . E1 is a classical stable equilibrium while E2 is stable only if
the next step of I on its random walk is down (dS  < 0). Should dS  > 0 be the case
then E2 will not exist any more and the only attainable equilibrium is just above E1 .
The market will move very quickly down to that new equilibrium: a jump downward24

24 Note that the difference between the horizontal line and the curve is exactly the negative of the

excess demand function (i.e., the excess supply function). If the market price was at E2 , then falling
prices would mean more supply, which would push the price even further down. This self-feeding
wave of selling would only stop when E1 is reached. This jump would be an arbitrage opportunity
to an investor who is aware of the presence and nature of the feedback trading. We excluded such
strategic exploitation of the large trader’s trading.
 FEEDBACK IN ILLIQUID MARKETS 259

Fig. 3. The intersection of the supply and demand curves. There are four cases depending on
the relative position of the intersection and the maxima and minima.

occurs and there is a discontinuity in the price. We will have to apply jump boundary
conditions on the stochastic process here.
4. The fourth case is the equivalent to case 3 but with opposite signs. If dS  < 0,
then E1 disappears and the market will jump up to E2 . Again we need jump boundary
conditions for the price process.
Thus there are four different points of interest: the two extrema from which the
jumps come and the two points to which the price jumps (see Figure 4). Jumps25 can
occur from point B to point D and from point C to point A. These points are defined
as follows:
S = B(t) and S = C(t) are the local extrema of the demand function, B is the
local minimum, and C is the local maximum. The definitions of these points are
∂ ∂f
(44) [X + f ](B, t) = −1 + (B, t) = 0,
∂S ∂S
∂ ∂f
(45) [X + f ](C, t) = −1 + (C, t) = 0,
∂S ∂S
and
∂2 ∂2
2 [X + f ](B, t) ≥ 0 ≥ [X + f ](C, t).
∂S ∂S 2
25 Deterministic jumps are not allowed in classical models of assets since they lead to arbitrage

opportunities. As mentioned before, the small traders are assumed to be unaware of the presence
and the effects of the large trader, and therefore they cannot exploit this arbitrage.
260 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

Fig. 4. The positions of the four points A, B, C, and D.

S = A(t) and S = D(t) are defined as

A ≤ B ≤ C ≤ D,
(46) [X + f ] (A, t) = [X + f ](C, t),
(47) [X + f ] (D, t) = [X + f ](B, t).

The positions of these points change in time. A typical graph for a Put-replicating
trading strategy is shown in Figure 5. For any continuous time-dependent trading
strategy all four boundary points have to arise at the same point. This point is
characterized as a saddle point of the full excess demand function X + f , which
obviously satisfies the conditions (44) to (47). In Figure 5 this point is marked as I.
Subsequently the points will constantly maintain the order A < B < C < D and fan
out as shown.
The Put-replicating strategy shown in Figure 1 approaches a step function as
expiry approaches. B and C disappear together at the exercise price and expiry date.
Points A and D approach E − and E + , respectively, resulting in a curve looking
like a tulip lying on its side. Note that all the boundary curves are confined to an
area of order ( × 2 ) around (E, T ).

6. The forward equation. The appropriate instrument to analyze the new

stochastic process (41) and (42) is the probability density function (PDF) p(S, t). The
PDF gives the probability density of the share price being at S at time t subject to
an initial distribution. It can be viewed as the result of a large number of Monte-
Carlo simulations of sample paths of the stochastic process, where roughly speaking
the height p(S, t) of the PDF at a given point (S, t) denotes the proportion of sample
paths that had share price S at time t.
 FEEDBACK IN ILLIQUID MARKETS 261

Fig. 5. The time-dependent positions of the four points A, B, C, and D for the Put-replicating
strategy.

The PDF satisfies the Kolmogorov forward equation

∂p ∂ 1 ∂2
(48) (S, t) = − (p(S, t) µ(S, t)) + (p(S, t) σ 2 (S, t)) =: (Lp)(S, t).
∂t ∂S 2 ∂S 2
Here L denotes the forward equation operator as defined. Frequently used initial
conditions are either a delta function if the starting point of the stochastic process is
known with certainty, or constant if all starting points are equally likely.
The boundary conditions for the PDF on the points ABCD are derived in the
appendix; here we just observe two facts:
Points B and C act locally as absorbing boundaries. As soon as the price process
hits either B or C the price jumps to the (distant) points D and A, but once it has
jumped away it cannot jump back. The jumps can only go in one direction. Thus B
and C act locally as “sinks,” and probability mass drains away through them.
Conversely, there is a “source” of probability mass at points A and D. The price
process cannot only reach A and D via normal diffusion but also by jumping from C
and B.
The precise strength of these effects is derived in Appendix A.
7. Numerical results. In this section we present numerical results for the so-
lution of the forward equation in three cases. In all cases we chose the parameters as
2
follows (unless otherwise stated): E = 10, r = 0.1, µB = 0.2, σB = 0.2, and = 1. (
was chosen large to illustrate the effects.)
The first example is the evolution of the probability density function in the absence
of any feedback. Thus = 0. In Figure 6 we show a three-dimensional plot of the PDF
against asset price and time. The starting condition is a delta function at S = 10. As
262 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

0, 9

0, 8

0, 7

0, 6

pd
0 ,5

0 ,4

0, 3

0, 2

0, 1

12
0,0 0

11 ,8
1 1,6
0 ,0 1

11 ,4
11 ,2
0 ,0 1

11
10,8
0,02

10,6
10 ,4
t
0 ,03

1 0,2
10
S
9 ,8
0,03

9,6
9,4
0,04

9 ,2
0 ,04

9
8 ,8
8,6
0 ,0 5

8,4
8,2
0 ,0 6
8

Fig. 6. The PDF p(S, t) for an asset with no feedback. The initial data is a delta function.

time increases the curve flattens out as a lognormal density function. In Figure 7 we
show a contour map of this same function.
The first nontrivial case is a time-invariant trading strategy that is strong enough
to give rise to the four new boundaries A, B, C, and D. This may perhaps model
the effect of a large number of investors hedging options with the same strike but
each with a different expiry date, such that the overall effect of these hedgers is to
add a time-independent function to the demand. Thus, as an example of this, we
have chosen f to be the delta of an option that will expire after a certain fixed time
interval. This is the simplest case to consider since the coefficients in the forward
equation are time independent.
The time-independent addition to the demand, f , was taken to be that from
a Put-replicating trading strategy with time fixed at t = T − 0.05, shortly before
expiry. The initial price is assumed to be known: S = 11; thus the PDF contains a
delta-function at S = 11. This case helps to visualize the effects of the boundaries.
The second case is simple Put replication with only one investor moving the
market. The trading strategy is the replication strategy for a European Put option.
This is one of the most popular trading strategies used by portfolio managers and of
direct relevance for option pricing. The addition to the demand due to this one large
investor is now time dependent.
A finite-difference theta-scheme (θ = 12 ) was used for the numerical calculations.
In the time-dependent cases the time step was automatically adjusted around the
boundaries A, B, C, and D.

7.1. Time-independent trading strategy. Time-invariant hedging schemes

are used to maintain the general performance of the portfolio without the need to
satisfy any precise conditions at fixed points in time (such as the potential liabilities
from writing an option). Here we will use the delta of a European Put at time 0.05
before expiry.
 FEEDBACK IN ILLIQUID MARKETS 263

Fig. 7. The contour plot of Figure 6.

Figure 8 shows the full development of the PDF as time proceeds. At t = 0

the PDF is a delta function at S = 11, meaning that the price at t = 0 is known
to be 11. Later the PDF spreads out. Figure 9 shows the contour plot for the
PDF. These figures clearly show the “corridor” between asset values of 9.6 and 10.3
(approximately) which the asset price can never reach. This region is that between
the two points previously labelled B and C. Since the replication strategy is time
independent this corridor does not change shape. Even though the starting value for
the asset (S = 11) is above this region the asset can still reach values less than 9.6 by
reaching the barrier C and jumping across to the point A. For more realistic values of
this corridor is very narrow and, away from the corridor, is effectively only a small
perturbation to the usual lognormal PDF as shown in Figures 6 and 7.
7.2. Put-replication trading. The more interesting case is the time-dependent
Put-replicating trading and the development in the area around the “tulip curve.”
The PDF is shown in Figure 10; the corresponding contour plots are shown in
Figure 11.
Since this replication strategy is genuinely time dependent the corridor that we
saw in the above example is now the tulip shape shown in Figure 5.
The asset price starts off at S = 10 and then evolves. The effect of the replication
strategy is felt immediately but the tulip curve itself does not appear until about
t = 0.03. At this time the barren region appears around the exercise price (10),
which the asset price avoids. This zone is shown most clearly in the contour plot of
Figure 11.
After expiry of the replicated option, t = 0.04, the barren zone disappears and
all values of S are attainable.
264 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

0,9

0,8

0,7

0 ,6

pd
0 ,5

0 ,4

0 ,3

0,2

0 ,1

0 , 00

0 ,0 1

0 , 02
14,00
1 3,30

t
12 ,60

0 ,0 3
11,90

11,20

0 , 04
s
10 ,5 0

9,80

9,10

0 ,0 5
8,40

7,70

7 ,00

Fig. 8. The PDF p(S, t) for a time-independent trading strategy. The initial data is a delta
function.

Fig. 9. The contour plot of Figure 8.

8. Conclusions. The influence of trading strategies and modern portfolio man-

agement has been the subject of much discussion but, apart from empirical studies,
little theoretical research. In this paper we presented a way to formally incorporate
trading strategies into the stochastic process followed by the underlying asset.
Many trading strategies that are used today are derived from replicating strategies
 FEEDBACK IN ILLIQUID MARKETS 265

1 ,0 0

0, 90

0, 80

0, 70

p(S,t)
0, 60

0, 50

0, 40

0, 30

0, 20

0, 10

0, 00

12 ,00
11 ,8 0
11,60
11 ,4 0
0,00

11,2 0
0,01

1 1,00
10,8 0
0,01

10,6 0
0,02

1 0,40
1 0,20
0,02

1 0,00
0 ,0 3

9,80
t

9,60
0,0 3

9,40
0,04

9,2 0
9,00
0 ,04

8,80
0 ,0 5

8,60
8,40
0 ,0 5

8 ,2 0
0 ,0 6
8,00

Fig. 10. The PDF p(S, t) for Put replication; the initial data is a delta function.

for derivative securities. This class of trading strategies is also central to the theory of
option pricing. We analyzed the difficulties encountered if one tries to build an option
pricing theory in illiquid markets and argued that in many cases the possibility of
market manipulation will make option trading impossible. On the other hand it will
frequently still be possible to find a trading strategy that replicates a given payoff.
We gave the partial differential equations that have to be solved to find these trading
strategies for the scenarios of replication by a small trader in the presence of a large
trader, and replication of a paper value payoff by a large trader.
In the second part of the paper we analyzed the effects of feedback trading in a
special case, option replication. These trading strategies are very popular in practice
both among option traders and among portfolio managers. We found that the securi-
ties whose replication destabilizes the market can be described by their payoff profile
and its gamma.
The effects are especially strong in markets with low liquidity and can even induce
discontinuities in the price process. Such price discontinuities are not allowed in
classical models of asset prices since they lead to arbitrage. Our model does allow
such arbitrage and we justify this in several ways. First, our model is deliberately
nonclassical, being a first attempt to put a real-world phenomenon into a theoretical
framework. Second, such effects as we describe do occur in practice and traders with
a knowledge of the positions of other market players and their hedging requirements
can take advantage of this knowledge. Third, it may be possible to remove certain
arbitrage opportunities by incorporating elasticity in the response of the market price
to large trades. The arbitrage only occurred because we did not let the small traders
react fully rationally to the large trader’s trading, otherwise taking the arbitrage
the small traders would have eliminated it. Modeling this requires a full strategic
interaction model between large and small traders, a complication we sought to avoid
in this paper. The latter point will be the subject of future work.
The numerical calculation of some examples showed the typical influences that
266 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

Fig. 11. The contour plot of Figure 10.

trading strategies can have on the price process.

Appendix A. Local analysis of the jump boundaries. To analyze the local
behavior of the price process and the PDF at the jump boundaries we will proceed in
the following steps:
First we show in a trinominal discretization of the model the behavior of the
stochastic process at a jump boundary. To apply these results to our case we then
analyze the leading-order behavior of the diffusion coefficients and of the PDF at the
jump boundary. Combining these results will yield the boundary conditions at the
jump boundaries and the source conditions at the jump targets.
A.1. The trinomial approximation. Consider a general diffusion process with
infinitesimal parameters µ and σ, which may be functions of S and t:

(49) dS = µdt + σdW.

To approximate the continuous case we consider the restriction of the diffusion process
of equation (49) onto a discrete mesh in the (S, t) domain with spacing δS in S and
δt in t.
Starting from state (S, t) the process can either move up by δS to (S + δS, t + δt),
move down by δS to (S − δS, t + δt), or stay at the same level to become (S, t + δt).
These three events have probabilities α, β, and (1 − α − β), respectively.
For this representation of the diffusion to be a good approximation we need to
use the right values for α and β. Remembering the definitions of the infinitesimal
parameters µ and σ of the diffusion µ(S, t) = limh0 h1 E{S(t+h)−S(t)} and σ 2 (S, t) =
limh0 h12 E{(S(t + h) − S(t))2 }, we have to require the same properties from our
 FEEDBACK IN ILLIQUID MARKETS 267

discrete process:

µ δt = E{S(t + δt) − S(t)}

(50) = δS (α − β),
2
σ δt = E{(S(t + δt) − S(t))2 }
(51) = (δS)2 (α + β).

Equations (50) and (51) determine α and β uniquely:

 
1 δt δt 2
(52) α= µ+ σ ,
2 δS (δS)2
 
1 δt δt 2
(53) β= − µ+ σ .
2 δS (δS)2

Because µ and σ 2 are generally functions of S and t, α and β will depend on S and
t, too. They are evaluated at the points where the arrows start. Being probabilities,
α, β, and (1 − α − β) must lie between 0 and 1, which can be achieved by a suitable
choice of δt.
Let us now analyze the behavior of the PDF at a jump boundary point. The jump
boundary consists of two separate points: the point from which the jump occurs (B
or C) and the point to which the jump goes (D or A). As the process immediately
jumps away from B or C the probability of finding the process there is actually zero;

(54) p(S = B, t) = 0 = p(S = C, t).

Next consider the case when the process is at B − δS at time t. By the definition of
the PDF finding the process here has probability p(B − δS, t). Here a jump occurs
with probability α: the process moves up by δS to hit B and then it jumps to D.
Thus the probability of finding the process at D at time t + δt is

p (D, t + δt) + αp(B − δS, t),

where p is the probability of finding the process at D if there were no jumps.26

Therefore the jumps contribute a source of strengths

1 1
(55) αp(B − δS, t) and βp(C + δS, t)
δt δt

per time step at points D and A, respectively. We normalized the contributions per
time step by dividing by δt.

A.2. Local behavior around B and C. To find the precise strength of the
sources at A and D we need to analyze the local behavior of µ, σ, and p around the
points B and C. We omit most of the details and give the intuition instead.
The local behavior of µ and σ is given by expanding (41) and (42) while using

26 Thus p just contains contributions from diffusion steps from D, D − δS, and D + δS according

to the discrete-time approximation p (D, t + δt) = (1 − αD − βD )p(D, t) + αD−δS p(D − δS, t) +

βD+δS p(D + δS, t).
268 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

the defining properties (44) and (45) of the points B and C. The result is at B

2
1 1 σ̃S 
(56) µ(B − δS, t) = 2 ,
(δS)3 ∂2
∂S 2
f

2
2 1 σ̃S 
(57) σ (B − δS, t) = ,
(δS)2 ∂2
∂S 2
f

where all functions are evaluated at B. The result at C is exactly the same only with
the functions evaluated at C and approximating at C + δS.
To reach the local behavior of p around B and C we go back to the forward
equation (48). Two local analyses are needed: one around the boundary points B
and C, and another at the point where the jump boundaries first appear.
It can be shown using the behavior of µ and σ around the points B and C that
near these points the forward equation becomes

∂ ∂ ∂2
cS̄ 4 p = 3p − 3S̄ p + S̄ 2 2 p,
∂t ∂ S̄ ∂ S̄
where c is (locally) a positive constant, and S = B + S̄ (or S = C + S̄ at the other
2
∂3
point). From this it follows that ∂∂S̄ p = 0 and ∂∂S̄ 2 p = 0 at B and C, and that ∂S 3p

is either finite or has a logarithmic singularity.

This particular singular solution can be eliminated by the second local analysis in
the region of the point of the first appearance of the boundaries, i.e., where the “stem”
of the tulip in the tulip curve meets the “blossom.” Here the singularities in µ and σ
are an order higher than (56) and (57) because we are at a saddle point of the demand
function. We can show that here the first three partial derivatives of the PDF with
respect to S are zero. From this and the properties of diffusion equations it follows
that the singular solution can be eliminated. This is confirmed by the numerics that
found (using the above trinomial grid) only finite source strengths at A and D.
Thus close to the boundaries we have the expansion of p:

1 ∂3
p(B − δS, t) = − (δS)3 3 p(B, t) + O((δS)4 ).
6 ∂S
Analogously close to C we have

1 ∂3
p(C + δS, t) = (δS)3 3 p(C, t) + O((δS)4 ).
6 ∂S
A.3. Putting it together. Substituting the above expressions for p and (56)
and (57) into the source strength approximations (55) we reach that

2
1 1 1 σ̃S  ∂3
(58) αp = p(B, t)
δt δS 8 ∂2
∂S 2
f ∂S 3

is the source strength at D (caused by jumps from B), and

2
1 1 1 σ̃S  ∂3
(59) βp = p(C, t)
δt δS 24 ∂2
∂S 2
f ∂S 3
 FEEDBACK IN ILLIQUID MARKETS 269

is the source strength at A (caused by jumps from C).

Finally we have to notice that the grid function 1/δS converges27 to the Dirac-
delta function at δS → 0, to get the source strengths at D and A as

2
1 σ̃S  ∂3
(60) δ(S − D)QD =δ(S − D) p(B, t),
8 ∂2
∂S 2
f ∂S 3

2
1 σ̃S  ∂3
(61) δ(S − A)QA =δ(S − A) p(C, t).
24 ∂2
∂S 2
f ∂S 3

This means that the forward equation becomes

∂
p(S, t) = (Lp)(S, t) + δ(S − D)QD + δ(S − A)QA ,
∂t

where the sources δ(S − D)QD + δ(S − A)QA have been added to the usual diffusion
behavior generated by the operator L.

Appendix B. Asymptotic analysis for small
. This appendix contains an

asymptotic analysis as → 0 of the option price replication equations (33) and (35)
and the probability density function (48) in the linear market model of Example 1
and section 5. In this section we ignore the jumps that were analyzed in Appendix A,
and we assume sufficient regularity in f to permit a regular asymptotic expansion;
specifically, we need the derivatives of f to be bounded. This will, for example, be
given if f is the aggregate result of many small traders where each is following a
slightly different strategy, and it is also the case for option replication strategies if one
excludes the region of the jumps.
The volatility of the modified price process is given by (42),

1
σ= ∂
σ̃(S − f ),
1− ∂S f

and the drift is given by (41),

  
1 ∂f 1 ∂2f
µ= ∂
µ̃S + −f + + σ2 2 .
1− ∂S f
∂t 2 ∂S

Here we used that S  = S − f by (40). We use the expansion

 2  3
1 ∂f 2 ∂f 3 ∂f
∂
=1+ + + + ···
1− ∂S f
∂S ∂S ∂S

to reach the expansion for the volatility

      2   
∂f 2 ∂f ∂f 3 ∂f ∂f 4
σ ≈ σ̃ S+ S −f + S −f + S −f + O( ) .
∂S ∂S ∂S ∂S ∂S

27 This is because the (discrete) integral of 1/δS equals one for all δS and the limit of 1/δS is zero

everywhere except at the point of the singularity.

270 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

For σ 2 the expansion is

     
∂f ∂f ∂f
σ 2 ≈ σ̃ 2 S2 + S − f 2S + 2 S −f 3S −f
∂S ∂S ∂S
    
3 ∂f ∂f ∂f 4
+ S −f (S + 1) − f + O( ) .
∂S ∂S ∂S

The expansion for the drift follows similarly:

 2

∂f ∂f 1 2 2∂ f
µ ≈ µ̃S + + µ̃S + 2 σ̃ S −f
∂t ∂S ∂S 2
  
∂f ∂f 3 ∂2f ∂f 2
2∂ f
+ 2 + µ̃S + σ̃ 2 S 2 − f − σ̃ f + O( 3 ).
∂t ∂S 2 ∂S 2 2 ∂S ∂S 2

Here we already see that to the leading order O(1) in the modified price process
coincides with the undisturbed price process

dS = µ̃Sdt + σ̃SdW + O( ).

B.1. The probability density. This result carries through to the Kolmogorov
forward equation of the probability density of the share price. This equation is

∂p ∂ 1 ∂2
+ (µp) + (σ 2 p) = 0
∂t ∂S 2 ∂S 2

with the appropriate initial conditions. Using a regular asymptotic expansion for
p(S, t)
2
p(S, t; ) = p0 (S, t) + p1 (S, t) + p2 (S, t) + O( 3 )

and equating terms of the same order yields the leading-order equation

∂p0 ∂ 1 ∂2
+ (µ̃Sp0 ) + (σ̃ 2 S 2 p0 ) = 0,
∂t ∂S 2 ∂S 2

where p0 has to satisfy the imposed boundary conditions. Note that this is the Kol-
mogorov forward equation for the density of the share price in the undisturbed case.
For the first-order correction p1 we reach the following partial differential equation:

∂p1 ∂ 1 ∂2
+ (µ̃Sp1 ) + (σ̃ 2 S 2 p1 )
∂t ∂S 2 ∂S 2
   
1 ∂2 2 ∂f
=− σ̃ S − f 2Sp 0
2 ∂S 2 ∂S
  
∂ ∂f ∂f 1 2 2 ∂2f
− + µ̃S + σ̃ S − f p0
∂S ∂t ∂S 2 ∂S 2

with zero boundary conditions.

 FEEDBACK IN ILLIQUID MARKETS 271

B.2. The pricing equations. For (33) we expand the option price in a regular
asymptotic expansion
2
P (S, t; ) = P0 (S, t) + P1 (S, t) + P2 (S, t) + O( 3 )

and reach analogously as leading-order equation the classical Black–Scholes partial

differential equation,

∂P0 ∂ 1 ∂2
+ rS P0 − rP0 + σ̃ 2 S 2 2 P0 = 0,
∂t ∂S 2 ∂S
which has to be satisfied with the usual boundary conditions. The first-order correc-
tion term satisfied a similar partial differential equation

∂P1 ∂ 1 ∂2
+ rS P1 − rP1 + σ̃ 2 S 2 2 P1
∂t ∂S 2 ∂S
  2
1 ∂f ∂
= − σ̃ 2 S S −f P0 ,
2 ∂S ∂S 2
with zero boundary conditions.
The solution of (35) is expanded in the same way:
2
Y (S, t; ) = Y0 (S, t) + Y1 (S, t) + Y2 (S, t) + O( 3 ),

and again we reach as leading-order equation the classical Black–Scholes partial dif-
ferential equation,

∂Y0 ∂ 1 ∂2
+ rS Y0 − rY0 + σ̃ 2 S 2 2 Y0 = 0,
∂t ∂S 2 ∂S
which has to be satisfied with the usual boundary conditions. The first-order correc-
tion term satisfied a similar partial differential equation:

∂Y1 ∂ 1 ∂2
+ rS Y1 − rY1 + σ̃ 2 S 2 2 Y1
∂t ∂S 2 ∂S
 2
 2
1 ∂ ∂ ∂
= − σ̃ 2 S S 2 Y0 − Y0 Y0 ,
2 ∂S ∂S ∂S 2
with zero boundary conditions. The difference to the first-order correction above
is that the delta ∂Y
∂S of the leading-order solution takes the place of the trading
0

strategy f .
Acknowledgments. The authors thank the participants of the ECMI 1993 con-
ference in Kaiserslautern and the ICIAM 1995 conference in Hamburg and two anony-
mous referees for helpful comments and suggestions.

REFERENCES

[1] K. Back, Insider trading in continuous time, Rev. Financial Stud., 5 (1992), pp. 387–409.
[2] K. Back, Asymmetric information and options, Rev. Financial Stud., 6 (1993), pp. 435–472.
[3] A. Bick, On the consistency of the Black-Scholes model with a general equilibrium framework,
J. Financial Quantitative Anal., 22 (1987), pp. 259–275.
272 PHILIPP J. SCHÖNBUCHER AND PAUL WILMOTT

[4] A. Bick, On viable diffusion price processes of the market portfolio, J. Finance, 45 (1990),
pp. 673–689.
[5] F. Black and M. Scholes, The pricing of options and corporate liabilities, J. Political Econ-
omy, 81 (1973), pp. 637–54.
[6] J. Board and C. Sutcliffe, The Effects of Trade Transparency in the London Stock Exchange:
A Summary, Special Paper 67, Financial Markets Group, London School of Economics,
January 1995.
[7] R. Bookstaber and J. A. Langsam, Portfolio insurance trading rules, J. Futures Markets, 8
(1988), pp. 15–31.
[8] M. J. Brennan and E. S. Schwartz, Time-invariant portfolio insurance strategies, J. Finance,
43 (1988), pp. 283–299.
[9] D. Duffie, Security Markets: Stochastic Models, Academic Press, San Diego, 1988.
[10] R. Frey, Asset Price Volatility and Option Hedging in Imperfectly Elastic Markets, Ph.D.
thesis, University of Bonn, Germany, 1995.
[11] R. Frey, it Perfect option replication for a large trader, Finance ans Stochastics, 2 (1998), pp.
115–142.
[12] R. Frey and A. Stremme, Market volatility and feedback effects from dynamic hedging, Math-
ematical Finance, 7 (1997), pp. 351–374.
[13] D. Furbush, Program trading and price movement: Evidence from the October 1987 market
crash, Financial Management, 18 (1989), pp. 68–83.
[14] G. Genotte and H. L. Leland, Market liquidity, hedging and crashes, American Economic
Rev., 80 (1990), pp. 999–1021.
[15] J. Hull, Options, Futures and Other Derivative Securities, Prentice-Hall, Englewood Cliffs,
NJ, 1989.
[16] R. A. Jarrow, Market manipulation, bubbles, corners and short squeezes, J. Financial Quan-
titative Anal., 27 (1992), pp. 311–336.
[17] R. A. Jarrow, Derivative securities markets, market manipulation and option pricing theory,
J. Financial Quantitative Anal., 29 (1994), pp. 241–261.
[18] I. Karatzas and S. E. Shreve, Brownian Motion and Stochastic Calculus, Springer-Verlag,
Berlin, Heidelberg, New York, 1991.
[19] J. A. Kyle, Continuous auctions and insider trading, Econometrica, 53 (1985), pp. 1315–1335.
[20] H. E. Leland, Who should buy portfolio insurance?, J. Finance, 35 (1980), pp. 581–594.
[21] H. E. Leland, Option pricing and replication with transaction costs, J. Finance, 15 (1985),
p. 1283.
[22] A. Mas-Colell, M. D. Whinston, and J. R. Green, Microeconomic Theory, Oxford Univer-
sity Press, Oxford, UK, 1995.
[23] L. G. McMillan, McMillan on Options, John Wiley, New York, Chichester, Brisbaine,
Toronto, Singapore, 1996.
[24] R. C. Merton, Continuous Time Finance, Blackwell Publishers, Oxford, UK, 1992.
[25] G. Mortenson, One day soon. . . , Forbes, July 29, 1996; also available online from
http://www.forbes.com/forbes/072996/5803072a.htm
[26] M. O’Hara, Market Microstructure Theory, Blackwell Publishers, Oxford, UK, 1995.
[27] Presidential Task Force on Market Mechanisms, Report of the Presidential Task Force
on Market Mechanisms, U.S. Government Printing Office, Washington, DC, Fed. Sec. L.
Rep. (CCH) Special Report No. 1267, 1988.
[28] P. J. Schönbucher, Option Pricing and Hedging in Finitely Liquid Markets, Master’s thesis,
Mathematical Institute, University of Oxford, UK, September 1993.
[29] P. J. Schönbucher and P. Wilmott, Hedging in illiquid markets: Nonlinear effects, in
ICIAM/GAMM 95: Proceedings of the Third International Congress on Industrial and
Applied Mathematics, Hamburg, 1995: Issue 3: Applied Stochastics and Optimization,
O. Mahrenholtz, K. Marti, and R. Mennicken, eds., Special Issue: Z. Angew. Math. Mech.,
1996, pp. 81–84.
[30] K. R. Sircar and G. Papanicolaou, General Black-Scholes Models Accounting for Increased
Market Volatility from Hedging Strategies, Working Paper, Department of Mathematics,
Stanford University, Stanford, CA, June 1996.
[31] P. Wilmott, S. Howison, and J. Dewynne, Option Pricing: Mathematical Models and Com-
putation, Oxford Financial Press, Oxford, UK, 1993.