Annals of Operations Research 37(1992)101-123 101

THE DYNAMICS OF SPECULATIVE BEHAVIOUR

Carl CHIARELLA
School of Finance and Economics, University of Technology Sydney, P.O. Box 123,
Broadway 2007, Sydney, N.S.W. Australia

Abstract
A number of recent empirical studies cast some doubt on the random walk theory
of asset prices and suggest these display significant tr~.nsitory components and complex
chaotic motion. This paper analyses a model of fundamentalists and chartists which can
generate a number of dynamic regimes which are compatible with the recent empirical
evidence.

1. Introduction

One of the cornerstones of the modem theory of finance is the view that asset
prices exhibit random walk behaviour. This view is important in empirical finance
because it is the theoretical underpinning of the large number of efficient market
studies which focus on the forecastibility of asset returns. It is also important in
theoretical finance because it is the basis of the stochastic price mechanisms assumed
in many of the key theoretical models in finance, e.g. the optimal portfolio rules
of Merton [24], the intertemporal capital asset pricing model of Merton [25] and
models for the pricing of contingent claims, Black and Scholes [4].
The impressive statistical evidence in favour of market efficiency discussed
for example by Fama [9] has been taken as support for the random walk model and
until fairly recently financial economists have been contented with this view as the
explanation of the time series behaviour of observed asset prices.
Recent empirical studies, however, have led to some questioning of the basic
tenets of the efficient market model, or at least of the view that it suggests for asset
price dynamics. For a start, there is a large number of studies reporting various
anomalies relating excess returns to a variety of factors such as firm size, market
sentiment, day of the week, month of the year, etc. Much of the recent anomalies
literature is surveyed by Keim [18]. Lo and MacKinlay [21], employing tests based
upon variance estimators find that a random walk model is generally not consistent
with the stochastic behaviour of weekly stock returns and also find significant
positive serial correlation for weekly and monthly holding period returns. Their
results also indicate that mean reverting fads models as suggested by Shiller and
Perron [33] and Summers [35] cannot account for the departures from the random

© J.C. Baltzer AG, Scientific Publisliing Company

lO2 C. Chiarella, The dynamics of speculative behaviour

walk model. Poterba and Summers [26] find evidence that transitory components
account for a large fraction of the variance in common stock returns. Fama and
French [10] find evidence of large predictable variation in 3 - 5 year stock retum
variances. A number of other empirical studies provide evidence that stockmarket
returns contain a sizeable predictable component, e.g. French et al. [13] and Campbell
[6]. Then there is a large and growing literature on the excess volatility of asset
prices, starting with Shiller [32] and recently surveyed by West [36], that puts into
question the view that asset prices are determined solely from rationally expected
fundamental values, which according to Samuelson [30] should follow a random
walk.
The stock market crash of 1987, and more recent gyrations, have caused
financial economists to ponder more deeply the applicability of the random walk/
efficient markets paradigm and have thereby given impetus to the quest for a fresh
look at models of how the stock market in particular, but speculative markets in
general, work. Leland [20] suggests a number of competing theories for stock
market behaviour, one of which is firmly rooted within the traditional efficient
markets paradigm, others, however, are based either on a perceived breakdown of
the market mechanism or on a consideration of different classes of investors.
Significantly, despite the dominance of the random walk/efficient markets
paradigm in the academic literature, some major investment institutions continue to
devote a considerable amount of valuable resources to technical analysis and there
are even claims that it is possible to "beat the market" with such techniques, e.g.
Pruitt and White [28].
The above considerations suggest that there is a need to consider some altemative
models of asset price dynamics (i.e. alternative to the view that the price reflects
only the rationally expected fundamental value of the asset). What would we require
of an alternative model to reconcile it with the various strands of literature we have
cited? Firstly, we would require that the model generate a significant transitory
component around an equilibrium which reflects the rationally expected value of
the asset. Secondly, the model must allow for the incorporation of chartists, a group
which bases its market actions on an analysis of past price trends. Since this group
seems to be an important part of real markets it is important to determine what
effect its activity has on the behaviour of assets prices and whether the behaviour
of a model incorporating chartists comes closer to explaining some of the empirical
results cited earlier.
In this paper we analyse a model of asset price dynamics containing
fundamentalists (who are forming rational expectations on the fundamental value
of the asset) and chartists (who are reacting to price trends). A number of models
of fundamentalists and chartists have been analysed in the literature. We cite in
particular the works of Zeeman [37], Beja and Goldman [3], Frankel and Froot [12]
and Bowden [5]. The model we analyse is a development of the model of Beja and
Goldman. These authors posit demand for the asset as having a fundamental component
and a speculative (chartist) component. The fundamental demand is the aggregate
 C. Chiarella, The dynamics of speculative behaviour 103

demand that the Walrasian auctioneer would face if a Walrasian auction were
conducted in the economy at each point in time. The speculative (chartist) demand
is assumed to be a function of the difference between the expected return on the
asset (measured by the chartists' estimate of price trend) and the return on some
alternative asset (e.g. government bonds). We argue in the next section that the
speculative demand function must have a certain linear form. We also assume that
asset prices adjust with finite speed to aggregate asset demand and that chartists
form their estimate of the price trend by an adaptive expectations scheme on past
price changes. A very rich form of price dynamics becomes possible when the
nonlinear speculative demand function is combined with the dynamic elements of
the model. The equilibrium price level may be stable or unstable. In the stable case
prices will converge back to the equilibrium; this case would correspond to the
price dynamics of much of the received theory. In the unstable case prices will not
converge back to the equilibrium after a shock but rather will tend to a stable limit
cycle around the equilibrium.
The stability/instability of the equilibrium as well as the nature of the fluctuations
which occur in the case of instability depend upon the interaction between the
strength of fundamental demand, the strength of speculative demand and the speed
with which the chartists revise their estimate of the trend. We show that a discrete
time version of the model is capable of generating chaotic price behaviour which
is consistent with the recent empirical evidence of Scheinkman and LeBaron [31]
that US stock returns exhibit chaotic behaviour.
Ours is an aggregative descriptive model in that we do not attempt to model
the asset market's microstructure via consideration of the intertemporal optimising
behaviour of the two groups of economic actors in our story. Rather we summarise
the microstructure details in a small number of structural parameters and concentrate
on an analysis of the price dynamics that the interaction of fundamentalists and
chartists generates. Genotte and Leland [15] develop a microstructure model of the
stock market containing groups of informed and uninformed investors all of whom
maximise expected utility of terminal wealth over a single period. Their model
generates an aggregate excess demand function for stock which has the same nonlinear
shape as the one we encounter in our model; however, it seems difficult within the
framework of their model to derive the nonlinear differential equations of price
dynamics which arise in our model. Ideally, we would like to be able to construct
a model of the market microstructure detail which can also generate the nonlinear
dynamic behaviour that we focus on in this paper, however, in our belief such an
enterprise is beyond the techniques currently available in dynamic optimisation
theory.

2. A model of fundamentalists and chartists

Instead of assuming that all transactions are executed at equilibrium prices

as in the typical Walrasian story, we follow Beja and Goldman (BG) and consider
104 C. Chiarella, The dynamics of speculative behaviour

a market where transactions and price adjustments occur simultaneously. The basic
premise of BG, which we will adopt, is that price changes are brought about by
aggregate demand of investors and that the speed of price adjustment is finite. This
assumption may be expressed mathematically as

P(t) = Dr(. ), (1)

where P(t) denotes the log of the asset price at time t and D~(.) investors' excess
demand for the security at time t. We leave the arguments of D I unspecified for the
moment.
Following BG we decompose the excess demand function as

Dt(.) = DO(.) + dr(.), (2)

where D°is the fundamentalists' demand which we take to be the aggregate demand
that a Walrasian auctioneer would face if at time t he or she were to conduct a
Walrasian auction in the economy and all transactions were executed at equilibrium
prices. The second component of aggregate excess demand, dr('), is the chartists"
demand which is dependent on "the state of the market" in a way to be made more
explicit below. By definition this demand is the difference between actual excess
demand and the fundamental demand. Such demand arises from the fact that since
transactions and price adjustments are occurring simultaneously some investors (i.e.
chartists) will base decisions not only on the state of the economic environment (i.e.
preferences, endowments, etc.) but also on the state of the market (i.e. price trends).
We refer the reader to sections I and II of BG for a fuller description of the market
environment that we are assuming here.
Letting W(t) denote the logarithm of the price that clears fundamental demand
at time t (i.e. D°(W(t))= 0), we write the fundamentalists demand as the linear
function

Do(.) = a(W(t) - P(t)), (3)

where a is a positive constant. In accordance with the Walrasian paradigm, the
asset's underlying equilibrium price at time t, W(t), depends upon traders' preferences,
endowments and information.
Turning now to the chartists' demand, we let g(t) denote the chartists' assessment
of the current trend in P(t), and g(t) be the return available on investments in non-
equilibrium trading of alternative securities (e.g. ~(t) could refer to stocks and g(t)
to bonds). The chartists' demand is assumed to be a nonlinear function of the
anticipated return differential V ( t ) - g(t). In particular we write

d,(.) = h(~(t) - g(t)), (4)

where the function h has the properties
 C. Chiarella, The dynamics of speculative behaviour 105

h'(x) > 0, for all x, (5.i)

h(0) = 0; (5.ii)

there exists an X* such that h"(x) < 0 (> 0) for all x > x*(< x*), (5.iii)

lim h'(x) = 0. (5.iv)

X.----)+ ¢~

These properties imply the general shape shown in fig.1.

L
~-g

Fig. 1. The chartists' demand function.

One explanation for the assumed shape of the chartists' demand function would
posit each chartist as seeking to allocate a fixed amount of wealth between the
speculative asset (risky) and bonds (riskless) so as to maximise intertemporal utility
of consumption. We know from Merton [25] that in such a situation the demand
for the risky asset is proportional to the difference in expected retums V - g, but
is also bounded above and below by the constraint that all available wealth may be
invested in the risky asset at one extreme ( V - g >> 0) and all in bonds at the other
106 C. Chiarella, The dynamics of speculative behaviour

extreme ( V - g << 0). For a single chartist the demand function would be piece-wise
linear, but adding together the demand functions of many chartists with differing
wealth, time horizons and preferences will result in an aggregate speculative demand
function as shown in fig.1. Since the total amount of wealth available to chartists
is assumed to be bounded this explanation would impose upper and lower bounds
on the chartists' demand function, i.e. to the conditions in eq.(5) we would also add,
(v) h(+oo) = hu and h(-oo) = hi" As a technical matter the condition (v) turns out to
not be essential to the dynamic behaviour we discuss in later sections, rather the
crucial conditions are (iii) and (iv) which ensure that the elasticity of speculative
demand reduces in magnitude as V and g deviate from one another.
Another explanation for the shape assumed for the chartists' demand function
could be developed along lines similar to Leland [20] who divides investors into
information based (with whom our fundamentalists may be identified) and information
free (equivalent to our chartists). Whereas Leland has the information free investors
adjusting their demand as a function of the level of stock prices, it also seems
plausible to argue that such demand would equally react to the difference ( V - g )
representing the relative expected growth potential of stocks. These investors are
hedging/rebalancing between stocks and bonds not because of information relating
to fundamentals but because they wish to modify their risk positions as their expected
wealth (which fluctuates with V - g ) varies.
In order to close the model we need to specify the manner in which chartists
form their estimate of V. One of the simplest assumptions we can make is that V
is taken as an exponentially declining weighted average of past price changes,
which may be expressed mathematically as the first order differential equation
V(t) = c(P(t) - V(t)), (6)
where c (0 < c < oo) is the speed with which chartists adjust their estimate of the
trend to past price changes. Alternatively the quantity 't" (= 1/c) may be viewed as
the time lag in the information of expectations and it can be shown that in a loose
sense V ( t ) = P ( t - ~ ) . Note that in the limit c---> oo ('r --> 0) chartists would be
forming their estimate of the price trend based only on the most recent price change,
i.e. V(t) = P(t). This limit is of interest because as it is approached we find that
chartists' estimate of the trend can exhibit large sudden changes and asset prices can
change direction abruptly.
Equations (1) to (6) imply that the dynamics of asset prices are driven by

P = a ( W - e ) + h ( v - g), (7)
and
fg = - a c P - c V + c h ( v - g) + acW, (8)
which constitutes a system of second order nonlinear differential equations. The
following two sections give an analysis of the type of dynamic behaviour we would
expect to observe if asset prices were driven by a system such as (7) and (8).
 C. Chiarella, The dynamics of speculative behaviour 107

3. Dynamic analysis
In order to make the analysis of the model tractable we shall assume that W
and g are constant. The equilibrium of the differential system (7) and (8) is given
by
= 0 and P = W + h(O - g)/a. (9)
Introducing the price deviation p = P - P we may put the differential system
(7) and (8) in the form
p = - ap + k(o/), (lOa)
vflt = - a p - gt + k( gt), (lOb)
where
k(~t) = h( ~ - g) - h ( - g ) . (11)

The equilibrium of the differential system (lO) is now at the origin ~ = 1/7 = O.
Our analysis of the nonlinear differential system (lO) proceeds in a number
of steps. Firstly, we perform a local linear analysis around the equilibrium to
determine the regions of stability and instability; this part of the analysis reproduces
the results of BG. Secondly, we apply theorems from the theory of nonlinear
differential equations to determine whether limit cycle motion is born at the boundary
of the stable and unstable regions. Thirdly, we apply some computational techniques
to analyse the characteristics of the limit cycle behaviour, in particular how it is
dependent on the underlying parameters of the model such as the relative strengths
of fundamentalist and chartist demand and the speed of adjustment of chartists'
estimate of the trend. Finally, we give a geometric discussion which enables us to
locate the limit cycle on a phase plane; such a constuction is useful as it allows us
to easily analyse the special case when chartists revise their estimate of the trend
infinitely rapidly and it also allows us to link our model to the catastrophe theory
model of Zeeman (op. cir.)

3.1. LOCALLINEAR ANALYSIS

The Jacobian matrix of the differential system (10) calculated at the equilibrium
point turns out to be

-a b ]
J=-a/'¢ (b-1)/'r ' (12)

where we have defined b = h'(-g).

Note first of all that det(J) = a/r> 0, which rules out saddle point behaviour
so that locally either prices are converging to the equilibrium (local stability) or
diverging from the equilibrium (local instability). To determine which case prevails
we need to examine trace (J). In fact trace (J) = - a + (b - 1)/at. Thus the equilibrium
is locally stable if
108 C. Chiarella, The dynamics of speculative behaviour

"c> ( b - l ) / a = 'r*. (13)

The equilibrium becomes locally unstable as chartists give more weight to

more recent price changes. The condition (13) is the one obtained by BG, who also
go on to discuss the dynamic charateristics of the linearised system.

3.2. EMERGENCE OF LIMIT CYCLE BEHAVIOUR

In the case that the equilibrium is locally unstable asset price time paths
diverge away from the equilibrium in either monotonic or oscillatory fashion. In a
linear model containing a unique equilibrium, such as that of BG, there are no
endogenous factors within the model to constrain the divergence of the price time
paths, which must therefore become unbounded. There are a number of ways in
which we could build a model to endogenously constrain the divergence of the price
path. One way would be to construct a linear model with multiple equilibria so that
the price paths diverging from the unstable equilibrium point would be attracted to
another equilbrium point which is stable; the model of Frankel and Froot [12] is
in this mould, Another approach is to construct a nonlinear model which has the
property of being capable of generating limit cycle motion around the unstable
equilibrium point; i.e. self-sustaining oscillating paths which persist for a wide
range of arbitrary changes in parameter values. Our model is in this mould.
One way of detecting the presence of limit cycle motion is to apply the Hopf
bifurcation theorem, discussed for example by Lorenz [22]. To this end we must
first calculate the eigenvalues of the Jacobian matrix J, namely

~1,2 = [ - a ( 1 - ~ / r ) + {aZ(1-v*/r) 2 - 4a/r } ltz]/2. (14)

At 'r= 'r* the eigenvalues have the properties that (a) they are pure imaginary, and
(b) /)Re(/~)/8"r=-a/2'r* < 0, where Re(A) denotes the real part of ~. By the Hopf
bifurcation theorem these two properties imply that a limit cycle emerges as "t-passes
through v*. Further analysis is required to determine the stability of the limit cycle
and its characteristics.

3.3. ANALYSIS OF THE NONLINEAR MODEL

The dynamic analysis up to this point has relied upon the linearisation of the
differential system (10), to proceed any further we need to undertake a nonlinear
analysis. To this end we eliminate p between eqs. (10) to obtain a nonlinear second
order differential equation for V, namely,

"c~ + ( a e - k ' ( gt)+k'(O))fg + a V = O, (15)

where e = "r- 7:*.

 C. Chiarella, The dynamics of speculative behaviour 109

Equation (15) has the same mathematical structure as the model of monetary
dynamics analysed in Chiarella [7] where by application of some theorems from the
theory of nonlinear differential equations (in particular Olech's theorem and the
Levinson-Smith theorem) it can be shown that for e > 0 the equilibrium point of
eq. (15) is globally asymptotically stable, whilst for e < 0 a unique stable limit
cycle exists. It should be pointed out that in order to obtain these results we assume
that the inflection point of the function h referred to in eq. (5) occurs at - g .
We would like to go beyond merely demonstrating that a limit cycle exists
when the equilibrium of the model is locally unstable, we would also if possible
like to see how some of the model's underlying parameters affect the limit cycle
fluctuations. We achieve this aim by applying the method of averaging (for a
discussion of which see Andronov et al. [2]) which gives a representation of a first
approximation to the limit cycle solution and relates it to the model's underlying
parameters. The basic idea of the method of averaging is to start from the observation
that when e = 0 (i.e. at the bifurcation point) and the nonlinear effects are ignored
(in which case k'() = constant), the solution to (15) is
gt(t) = Asin(f2t+ 0), (16)

where f~2= a/r. In this last equation the amplitude A and the phase angle 0 are
constants depending upon the initial conditions. The method of averaging seeks a
solution to the nonlinear differential eq. (15) when e ~ 0 in the form
gt(t) = A(t)sin[f~t + 0(t)], (17)

where now the amplitude A and the phase angle 0 become time dependent. Of
course up to this point no real simplification has been achieved, we have merely
replaced the problem of finding ~(t) with the problem of finding A(t) and 0(t).
However, Andronov et al. [2] outline a systematic technique for approximating A(t)
and O(t).
Applying this technique to eq. (15) we find that a first order approximation
to the amplitude function A(t) is given by the solution of the differential equation
= A(K(A)-ae/2)/z, (18)
where K(A) is defined by
21r
1
K(A ) = -£ ~[ k'(Asin y)-k'(O)]cos 2 ydy.
o
The function K(A) has been analysed by Chiarella [7], who shows that it has the
qualitative features illustrated in fig. 2.
By studying the qualitative behaviour of the differential eq. (18) we are able
to determine how the amplitude of the asset price fluctuations evolves.
For e > 0 the term K(A) -ae/2 < 0 for all A, hence A < 0 for all A and so A
declines steadily to zero as illustrated in fig. 3(a). This picture merely reconfirms
110 C. Chiarella, The dynamics of speculative behaviour

a~/2

-b/2

Fig. 2. The function K.

(a) e > 0 (b) ~<0

Fig. 3. Dynamic behaviour of the amplitude.

what we already know from the previous analysis, namely that in this case the
equilibrium at the origin is locally asymptotically stable. For e > 0 the equation
K(A) - a e / 2 = 0 has one root ,4. It is clear that A > 0 (< 0) for A < .4 (> A) and hence
from any initial value the amplitude converges to ,4. It should be pointed out that
the method of averaging is valid for e small.
 C. Chiarella, The dynamics of speculative behaviour 111

3.4. GEOMETRIC ANALYSIS

The geometric analysis consists of constructing the phase plane for the nonlinear
differential system (10). To achieve this end we need to sketch the curves p = k(~)/a
and p = (k(~¢) - ~z)/a. The former curve is monotonic increasing as shown in fig.
4 whilst the latter is monotonic decreasing for b < 1 and has two turning points for

~p

p = k($)/a

;+
p = (k(@)-@)/a

b>l

Fig. 4. Geometric analysis of the limit cycle.

b > 1 which is the case sketched in fig. 4. It is a simple matter to determine the
direction arrows referring to the differential equations. We see from fig. 4 that p
and 9, are oscillating around the equilibrium point and we know from our earlier
analysis that when b > 1 + aT the oscillations are being attracted towards the limit
cycle shown.
Examining fig. 4 we see that there are four phases in the fluctuations of p
and Vz. During the first phase (A to B) chartists' excess demand is greater than
fundamentalists' excess demand and so p is rising, also, chartists' estimate o f the
trend lags the actual trend so lg is rising. During the second phase (B to C) chartists'
112 C. Chiarella, The dynamics of speculative behaviour

excess demand is still greater than fundamentalists' excess demand and so p continues
to rise, on the other hand the chartists' estimate of the trend is now overshooting
the actual trend and so g¢ starts to decline. During the third phase (C to D) chartists'
excess demand is driven below fundamentalists' excess demand since ~ has been
falling during the previous phase, however chanists' estimate of the trend is still
overshooting the actual trend and so vcontinues to decline. During the fourth phase
(D to A)chartists' excess demand remains less than the fundamentalists' excess
demand since V has been falling in the previous phase; p continues to decline, on
the other hand the decline in ~ has led to the chartists' estimate of the trend to lag
the actual trend and so vstarts to rise. The cycle then reenters the first phase where
the rise in V from the previous phase causes chartists' excess demand to exceed
fundamentalists' excess demand and so the cycle repeats itself.

4. Speculation and dynamic behaviour

We observe first of all that if b < 1 then the equilibrium is locally, and in fact
globally, asymptotically stable and so asset prices converge to the equilibrium.
Recalling that b measures the slope of the chanist demand function at V = 0 this
result tells us that if chanist demand is sufficiently low then the model remains
stable and asset prices will converge to equilibrium given by eq. (9).
If the rate of change of chartist demand is sufficiently large that b > 1, then
"r* > 0 and if chartists are revising their estimate of the trend sufficiently rapidly that
"r < "r* then local instability of the equilibrium occurs with prices converging to the
limit cycle described in the previous section. Asset prices will fluctuate steadily
around the equilibrium until there is some change in the parameters of the model.
In this way the model is able to generate the significant transitory and predictable
fluctuations around the equilibrium which are suggested by the empirical studies
cited in the introduction. The model is also consistent with Cootner's [8] reflecting
barriers model and our analysis could be seen as providing more detail on the
dynamics of this construct. It is worth pointing out that Anderson [ 1] provides some
empirical evidence in favour of Cootner's reflecting barrier model.
It is of interest to inquire how changes in the parameters of the model will
affect the amplitude of the fluctuations.
Consider first of all the effect of varying a, the slope of the fundamental
demand function. Since ae = a z - ( b - 1 ) we see from fig. 2 that the effect of increasing
a is to yield a lower value of ,4. Hence a ceteris paribus increase in fundamental
demand leads to a decrease in the amplitude of the asset price fluctuations.
Next, consider the effect of changes in b, the slope of the speculative demand
function at equilibrium. The outcome of this change is not clear cut. Referring to
fig. 2 we see that on the one hand the horizontal lines at a e / 2 and - b / 2 both drop
(by the same amount) which tends to increase ,4. On the other hand the K(A)
function also declines which tends to decrease ,4. Since the method of averaging is
 C. Chiarella, The dynamics of speculative behaviour 113

only valid for small e the shifts being considered occur over the flatter part of the
K function so that the first effect is more likely to dominate and the amplitude of
the fluctuations will increase.
Finally consider the effect of changes in z, the average time lag in the chartists'
formation of the trend estimate. A ceteris paribus decrease in ~" (i.e. speculators
react more quickly to price changes) makes ae more negative, hence from fig. 2 a
larger value of A will emerge. Thus, the amplitude of asset price fluctuations
increases as speculators react more quickly to asset price changes.

~p

L
\ << I

A ~L'

Fig. 5. The limit cycle as "r --~ 0.

Most interesting, however, is the effect on the fluctuations when "r ~ 0, this case
being most conveniently analysed by use of the phase diagram in fig. 5. Figure 5 is
constructed by observing from the differential eq. (10b) for V that as z---> 0, then
V ~ o0 except in a small neighbourhood of the curve p = (k(v) - V)/a (denoted LIL).
114 C. Chiarella, The dynamics of speculative behaviour

~qJ

'1- -- -- 9-- --

i I
!

I I I
!
t
i I
I
! !
i
I I
I 1
I
I I I
| I I
I
. . . . 1

t ~ I

Fig. 6. Time paths for p and g.

 C. Chiarella, The dynamics of speculative behaviour 115

Thus, from any initial value (e.g. I in fig. 5) trajectories are swept almost instantaneously
to the curve LE. We use the double arrowheads to represent this instantaneous
adjustment. Once on LE, trajectories move along this curve to reach either A or C
from where they are again swept instantaneously to either B or D respectively. The
cycle ABCD then repeats itself indefinitely. The corresponding time paths for grand
p are displayed in fig. 6. The chartists' trend estimate gr displays large jumps and
asset price p displays sharp changes in direction at the jump points A --~ B, C -~ D.
This is the type of dynamic behaviour we might observe when the asset market is
particularly volatile as in the recent past. It could be argued from the viewpoint of
the model presented here that the accumulation of a series of adverse "news" events
on the US economy (mounting trade deficit, doubts about continuing lines of credit
from foreign financial institutions to the US economy) has caused chartists to
considerably lower their value of "cand thereby place greater weight on most recent
price changes in their estimate of the trend.
We should also point out that in the z= 0 limit the model has a mathematical
structure which is very similar to Zeeman's catastrophe theory model of stock
exchange behaviour.

5. The effect of a noisy environment

The model as we have discussed it up to this point is entirely deterministic.

However, we would never expect observed prices to follow the smooth periodic
paths which the model generates in the locally unstable region. Rather the smooth
paths of the deterministic model will be perturbed by the effect of random influences.
A number of random influences may be identified. Firstly, the theory of equilibrium
prices in perfect markets argues that the asset's notional equilibrium value W(t)
follows a random walk; see Samuelson [30]. Secondly, the mechanism describing
the adjustment of observed prices will contain a randomly fluctuating component
reflecting the unpredictable noisy elements in the market's behaviour, e.g. the arrival
of "news". Thirdly, there is the randomness arising from the fact that the instabilities
giving birth to the periodic motion can cause the market system to be extremely
sensitive to changes in initial conditions so that the dynamics of asset prices exhibit
deterministic chaos. The latter type of randomness will be discussed in the next
section. In this section we shall analyse the impact on asset price dynamics of the
first type of randomness when the notional equilibrium price follows a random
walk. It is also possible in this framework to analyse the influence of the second
type of randomness due to market noise but the technical details of this are lengthy
and shall be studied in a later paper.
Specifically we shall assume that W(t) is driven by the stochastic differential
equation
dW = crdz, (21)
where z is a standard Wiener process with zero drift and unit variance per unit time
116 C. ChiareUa, The dynamics of speculative behaviour

and iris constant. For the purposes of manipulating the differential equations describing
the dynamics of asset prices it is convenient to write eq. (21) in the form
I/¢= o-J, (22)
where i is a white noise process. This notation is not strictly mathematically correct
but is common in a large literature on the analysis of nonlinear stochastic dynamical
systems upon which we will need to draw, in particular Stratonovich [34], Khas'minskii
[19] and Roberts and Spanos [29]. These authors are however careful to switch back
to the mathematically more correct Ito form of the stochastic differential equation
at the appropiate point in the analysis.
In the now stochastic environment the dynamics of asset prices are governed
by eqs. (6) (for ~ , (7) (for P) and (22) (for W). Rearranging these in much the same
way as we did to obtain eq. (16) in section 3 but now taking account of the
stochastic term introduced through eq. (22) we find that ~ is driven by the second
order nonlinear stochastic differential equation

g/ +[ae + k ' ( O ) - k'( ~)]flt / ~ + a~t / "¢ = o'2/~. (23)

Equation (23) is identical to eq. (16) apart from the white noise term on the right-
hand side. The effect of this noise term is to give to the limit cycle path of the
deterministic model a probability distribution. We would like to know as much as
possible about this distribution and how its behaviour is altered by changes in the
intensity of the noise term tr and in the parameters a, b and ar which characterise
our economy of fundamentalists and chartists.
Defining ¢, = ~ we may replace eq. (23) by the system of first order Ito
stochastic differential equations
d ~ = ~ dt, (24a)
d~ = - v (gt,~) dt + o-dz/r, (24b)
where v (~,~) = [ae + k'(O) - k'(gt)]~/," + a~/~. For every Ito stochastic differential
system such as (24) it is possible to write down a partial differential equation known
as the F o k k e r - P l a n c k - Kolmogorov equation (FPKE) which describes the transition
probability density function F(~, ~, t; Vo, ~o, to) from an initial point gto, ¢o, to to
the current point ~, ~, t; see for example Gardiner [14]. For eq. (24) the FPKE
assumes the form
~F/~t = - ~ [ ~ F ] / ~ + ~[v (~, ~?)F]/~(~ + (o2/2"r)~2Fl~d? 2. (25)

Ideally, we would like to solve eq. (25) for F, however, analytical solutions to the
FPKE are known for only a small number of special cases of which (25) is not one.
We are therefore forced to resort to some approximate method of analysis and the
one that has been developed in the literature on nonlinear stochastic differential
equations is a stochastic version of the method of averaging which is used to obtain
a probabilistic description of the approximate amplitude of the limit cycle fluctuations.
 C. Chiarella, The dynamics of speculative behaviour 117

For e < 0 and --0 the stochastic averaging method seeks a solution to (23) in
the form of eq. (18). Application of the stochastic averaging techniques as outlined
by Roberts and Spanos [29] reveal that the amplitude of the fluctuations, A, satisfies
the lto stochastic differential equation
dA = {A[K(A) - ae/2]/'c + ~r), ~2/2A }dt + ~(o-lr)lt2d;I,, (26)

where ~ is a standard Wiener process.

The FPKE for the conditional probability distribution F(A,t) associated with
eq. (26) has the form
~F/~t = ( cr D ~ /2 )~2F/~A 2 - ~[g(A )F]/~A, (27)
where g(A) = {A[K(A) - a e / 2 ] / v + ~o'f~2/2A}. Whilst it is not possible to find an
explicit solution for the time varying distribution tile stationary distribution is readily
found to be
F~(A) = CA exp [ - 2()'a~) -1 fA {ae/2 - K(A)} dA], (28)
where C is a normalising constant. In order to reduce (28) to a form on which we
can perform some qualitative analysis we need to make some reasonable approximation
to the A { a e / 2 - K ( A ) } term. We recall that the method of stochastic averaging is
valid for e -- 0, whilst from fig. 2 we see that for e = 0, K(A) may be approximated
by a concave down parabola. Thus for A in the neighbourhood o f A it is not
unreasonable to make the approximation

a e / 2 - K(A) = (ae/2)(1 - (A//~)2). (29)

Using this approximation we find that

Fs(A) = 2 N - l a exp [(oc2 - a412)/t~], (30)

where a = A/A, 8 = -2o'er/(eA-2) and N = (%tS/2)l/2erfc[ - 1/(28)1/2]. Stratonovich [34]

has shown that for the distribution (30)
E(A) -- (1 - 8/8) ,~, (31a)
and
E([A-,~] 2) -- (t5/4)/~2, (31 b)
where the approximation is valid provided that t5 is relatively small, which will be
so if the intensity of the noise term (tr) is sufficiently small. Hence, an increase in
the intensity of the noise term decreases the mean amplitude of the fluctuations and
increases the variance. By checking the signs of the appropriate partial derivatives
it is not difficult to show that a ceteris paribus increase in the strength of the
fundamental demand, a, leads to a decrease in the mean amplitude of the fluctuations
and an increase in the variance.
The previous stochastic analysis is valid for 'r close to 'r*. A stochastic analysis
of the ~- --> 0 limit would also be o f interest, especially since this is more likely
the operating mode of many asset markets at the current time. The techniques
118 C. Chiarella, The dynamics of speculative behaviour

expounded by Grasman [17] would allow us to attempt such an analysis however,

the technical details would considerably lengthen this paper and so we defer this
analysis. The characteristic of the motion is that the jumps A --->B and C --->D in
fig. 5 now occur at random times and the object of the stochastic analysis is to relate
the probability distribution of the inter-jump times to the structural parameters of
the model.

6. The emergence of chaotic motion

Up to this point the analysis has revealed that the model of chartists and
fundamentalists can exhibit two dynamic regimes. Namely, convergence to the
Walrasian equilibrium price (when the equilibrium is locally asymptotically stable)
and limit cycle fluctuations around the equilibrium price (when the equilibrium is
locally asymptotically unstable). We would like to know whether the dynamic
interaction of fundamentalists and chartists can also exhibit that third type of dynamic
regime known as chaos, which has recently been found to be important in dynamic
modelling of physical and biological systems and empirical evidence of which has
been found in financial time series, as reported in the introduction. The chief
characteristic of chaotic motion is that a deterministic system can generate a time
series that seems entirely random and that the time history of paths starting from
nearby initial points will look entirely different.
The mathematical expression of the model we have analysed is a continuous
time two dimensional differential system, however, to generate chaotic motion in
a continuous time system at least three dimensions are required. So in order to have
the possibility of generating chaos in continuous time the model would have to be
extended by at least one further differential equation, perhaps from a more detailed
modelling of the market for the alternative asset. An alternative approach, which
we adopt, would be to consider a discrete time version of the model, since it is
possible to generate chaos in even a one dimensional discrete time system.
To obtain a discrete time system we approximate the time derivatives in (10)
by first differences across successive units of time. The discretisation of the differential
equation system (10) under this approximation yields

P~+i = (1 - a)pt + k(v~), (32a)

v/,+ l = - ( a R ) p t + (1-1/z')N, + k(N,)/'r. (32b)

Equation (32) represents a two dimensional nonlinear difference equation. The

analysis of the chaotic regimes of two dimensional discrete time systems would
require a far more detailed analysis than space permits us here. We are, however,
able to make our basic point by considering the one dimensional sub-case obtained
when t--* 0. In this case (32b) reduces to

ap, = - ~, + k ( v , ). (33)
 C. Chiarella, The dynamics of speculative behaviour 119

Eliminating Pt between (32a) and (33) we find that ~t satisfies the one dimensional
difference equation

~+ 1 = F(~,), (34)

where F(Ig) = H-I(G(~)) with H ( ~ ) = k(v,/) -Vr and G ( ~ ) = k ( ~ ) + (a - 1)~.

~F
I'

(a) 0 < b < ( 1 - a )

p f

B ...... Cff ~'ff¢

/'F E
J
f

t"

(b) 1 - a / 2 < b <1

Fig. 7. The map F.

The qualitative features of the functions H, G and F have been analysed by

Chiarella [7]. A number o f sub-cases are possible; however, we shall limit
discussion to the subcase a < 1 when it is found that F is monotonic increasing for
0 < b < 1 - a (see fig. 7(a)) and has two turning points for 1 - a < b < 1 (see
fig. 7(b)). Since F ' ( 0 ) = G'(O)/H'(O) it is not difficult to show that the origin is
12o C. Chiarella, The dynamics of speculative behaviour

locally asymptotically stable for 0 < b < 1 - a/2 and locally asymptotically unstable
for 1 - a / 2 < b < l .
It is the unstable region which is of interest to our discussion on the possibility
of chaotic motion. Referring to fig. 7(b) and constructing the standard cobweb type
motions we see that all trajectories are eventually trapped in the region ABCDEFA.
Since the origin is unstable, all trajectories will be attracted to either a periodic or
chaotic motion. In the region under discussion the map F is bimodal and over the
section ABCDEFA may be reasonably well approximated by a cubic. May [23] has
examined the dynamics of a general cubic map and shown that it can exhibit the
transition through period doubling to chaos familiar from the quadratic map. The
significance of the quadratic map lies in its universal characteristic, namely that all
unimodal maps ultimately exhibit the period doubling to chaos characteristics for
appropriate parameter values; see Feigenbaum [11]. Whilst it is suspected that a
universal result exists for bimodal maps a proof does not seem to have been given
so far. Hence, we cannot assert definitely that map F undergoes period doubling to
chaos as the parameter b increases, though this is strongly suggested by May's
result.

7. Closing remarks

We have seen that the model of fundamentalists and chartists is capable of

generating sustained fluctuations around the Walrasian random walk equilibrium
price. The fluctuations are either noisy limit cycle motions or chaotic motion. The
picture we have drawn up is certainly compatible with the empirical studies finding
evidence of non-random walk behaviour, of significant transitory components and
chaotic dynamics. Of course it should be possible to construct many models generating
the stylised dynamic facts emerging from the empirical studies. Whether models of
fundamentalists and chartists such as the one presented here provide a better framework
for understanding observed financial time series will probably be a matter of some
debate and hopefully will become a subject of empirical study.
The view we have presented here would justify the use of chartism, though
many down-town practitioners would probably claim that they do not need an
academic to confirm their belief in the charts. However, the model also tells us why
we would not expect chartism to have a one hundred percent success rate. As we
have seen the model can exhibit three basic dynamic regimes which are (i) convergence
to the Walrasian equilibrium, (ii) limit cycle motion around the Walrasian equilibrium,
and (iii) chaotic motion around the Walrasian equilibrium, and one or other of these
regimes will be operative depending on the relationship between the structural
parameters a, b and "r. We would expect chartism to be successful in the limit cycle
regime, however the arrival of news of sufficient magnitude could cause a change
in the structural parameters that would switch the system into one of the other
regimes where chartism would not be successful. This process of regime switching
is undoubtedly occurring incessantly and unpredictably in real markets.
 C. Chiarella, The dynamics of speculative behaviour 121

A model exhibiting the three dynamic regimes we have delineated could also
contain the seeds of an explanation to the apparent excessive volatility of stock
prices, the study of which was initiated by Shiller [32] and has subsequently spawned
an enormous literature. These empirical studies tacitly assume a theoretical framework
in which only the first dynamic regime is a possibility. However, once we admit
into the theoretical framework the other two regimes in which prices fluctuate
around the Walrasian equilibrium then the observed volatility may not be excessive
at all.
A number of directions for further research suggest themselves. Firstly, there
is a need to model the microstructure of the asset market in some detail. We have
already alluded to the difficulty of this approach in the introduction. However, a
partial solution would be to adopt the approach of Genotte and Leland [15] within
the framework of an overlapping generations model. It may then be possible to
obtain a nonlinear map for price dynamics which exhibits the three basic dynamic
regimes and which is dependent on microstructure parameters such as agents'
preferences.
We have cited empirical literature which finds evidence of nonlinear dynamic
behaviour. An important next step would be to attempt to apply the rapidly developing
nonlinear time series techniques, e.g. Priestly [27], to fit some nonlinear models to
stock price data. The nonlinear maps inherent in such fitted models could be compared
with the nonlinear maps of the theoretical model discussed in this paper. Such
models could also be used to take a fresh look at the excess volatility studies.
An essential element of the traditional random walk/efficient market studies
is an understanding of the time-variance relationship of successive price changes
of the theoretical model over differing observation periods. If the nonlinear modelling
framework is to proceed to the empirical stage we will need to obtain some knowledge
of the time-variance relationship implied by nonlinear stochastic models of the
type discussed in section 4.
Our story has given a passive role to g, the return on the alternative asset.
A more fully specified model would also need to consider the dynamics of the
market for the alternative asset. We would expect to find the various dynamic
regimes occurring in this market as well. Mathematically we would now find ourselves
in the world of coupled nonlinear oscillators. However, it would probably be best
to await more confirming empirical evidence of nonlinear behaviour before attempting
the difficult task of modelling such linkages.

References
[1] S.C. Anderson, Evidence on the reflecting barriers model: New opportunities for technical analysis?,
Financial Analysts J. (May-June 1989)67-71.
[2] A.A. Andronov, A.A. Vitt and S.E. Khaikin, Theory of Oscillators (Pergamon, Oxford, 1966).
[3] A. Beja and M.B. Goldman, On the dynamic behaviour of prices in disequilibrium, J. Finance
35(1980)235-248.
122 C. Chiarella, The dynamics of speculative behaviour

[4] F. Black and M. Scholes, The pricing of options and corporate liabilities, J. Political Econ. 81(1973)
637-659.
[5] R.L Bowden, Predictive disequilibria and the short run dynamics of asset prices, Austral. J. Manag.
15(1990)65-87.
[6] J.Y. Campbell, Stock returns and the term structure, J. Financial Econ. 18(1987)373-400.
[7] C. Chiarella, The Elementsof a Nonlinear Theory of Economic Dynamics, Lecture Notes in Economics
and Mathematical Systems, Vol. 343 (Springer, 1990).
[8] P.H. Cootner, Stock prices: Random vs, systematic changes, Ind, Manag. Rev. (Spring 1962)24-45.
[9] E.F. Fama, Foundations of Finance (Basic Books, New York).
[10] E.F. Fama and K.R. French, Permanent and temporary components of stock prices, J. Political Econ.
96(1988)246-273.
[11] M.J. Feiganbaum, Universal behaviour in nonlinear systems, Essay No. 7, in: Nonlinear Dynamics
and Turbulence, ed. G.I. Barenblatt et al. (Pitman, 1983).
[12] J.A. Frankel and K.A. Froot, Understanding the US dollar in the eighties: The expectations of
chartists and fundamentalists, Economic Record, Special Issue on Exchange Rates and the Economy
(1986) 24-38.
[13] K.R. French, G.W. Schwert and R.F. Stambaugh, Expected stock returns and volatility, L Financial
Econ. 19(1987)3-30.
[14] C.W. Gardiner, Handbook of Stochastic Methods (Springer, 1985).
[15] G. Genotte and H. Leland, Market liquidity, hedging and crashes, Amer. Econ. Rev. 80(1990)
999-1021.
[16] M.B. Goldman and A. Beja. Market prices vs. equilibrium prices: Returns' variance, serial correlation,
and the role of the specialist, J. Finance 34(1979)595-607.
[17] J. Grasman, Asymptotic Methods for Relaxation Oscillations and Applications (Springer, 1987).
[18] D.B. Keim, The CAPM and equity return regularities, Financial Analysts J. (May-June, t986)
19-34.
[19] R.Z. Khas'minskii, The behaviour of a self-oscillating system acted upon by a slight noise, P.M.M.
(English transl.) 27(1963)1035-1044.
[20] H. Leland, Black Monday: The catastrophe theory, Risk 1(10)(1988)6-13.
[21] A.W. Lo and A.C. MacKinlay, Stock market prices do not follow random walks: Evidence from
simple specification test, Rev. Finincial Studies 1(1988)41-66.
[22] H.W. Lorenz, Nonlinear Dynamical Economics and Chaotic Motion, Lecture Notes in Economics
and Mathematical Systems, Vol. 334 (Springer, 1989).
[23] R.M. May, Nonlinear problems in ecology and resource management, in: Chaotic Behaviour of
Deterministic Systems, ed. G. loos (North-Holland, 1983).
[24] R.C. Merton, Optimum consumption and portfolio rules in a continous-time model, J. Econ. Theory
3(1971)373-413,
[25] R.C. Merton, An intertemporal capital asset pricing model, Econometrica 41(1974)867-887,
[26] J.M. Poterba and L.H. Summers, Mean reversion in stock prices: Evidence and implications, NBER
Working Paper 2343 (1987),
[27] M.B. Priestly, Nonlinear and Nonstationary Time Series Analysis (Academic Press, 1988).
[28] S.W. Pruitt and R.E. White, The CRISMA trading system: Who says technical analysis can't beat
the market, J. Portfolio Manag. (Spring 1988) 55-58.
[29] J.B. Roberts and P,D. Spanos, Stochastic averaging: An approximate method of solving random
vibration problems, Inter. J. Non-Linear Mechanics 21(1986)111-134.
[30] P.A. Samuelson, Proof that properly anticipated prices fluctuate randomly, Ind. Manag. Review
6(1965)41-49.
[31] J.A. Scheinkrnan and B. LeBaron, Nonlinear dynamics and stock returns, J. Business 62(1989)
311-337.
[32] R.J. Shiller, Do stock prices move too much to be justified by subsequent changes in dividends?,
Amer. Econ. Rev. 71(1981)421-436.
 C. Chiarella, The dynamics of speculative behaviour 123

[33] R.J. Shiller and P. Perron, Testing the random walk hypothesis: Power versus frequency of observation,
Econ. Lett. 18(1985)381-386.
[34] R.L. S~'atonovich, Topics in the Theory of Random Noise, Vols. 1 and 2 (Gordon and Breach, New
York, 1967).
[35] L.H. Summers, Does the stock market rationally reflect fundamental values?, J. Finance 41(1986)
491-600.
[36] K.D. West, Bubbles, fads and stock price volatility: a partial evaluation, J. Finance 43(1988)
639- 660.
[37] E.C. Zeeman, On the unstable behaviour of stock exchanges, J. Math. Econ. 1(1974)39-49.