Nonlinear Dynamics in Economics and Finance

Author(s): Jose A. Scheinkman

Source: Philosophical Transactions: Physical Sciences and Engineering, Vol. 346, No. 1679,
Mathematics of Nonlinear System (Feb. 15, 1994), pp. 235-250
Published by: Royal Society
Stable URL: [Link]
Accessed: 14-05-2018 04:08 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@[Link].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
[Link]

Royal Society is collaborating with JSTOR to digitize, preserve and extend access to
Philosophical Transactions: Physical Sciences and Engineering

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
 Nonlinear dynamics in economics and finance
BY JosE A. SCHEINKMAN

Department of Economics, University of Chicago, 1126 E. 59th Street,

Illinois 60637, U.S.A.

The apparent success of 'chaos' in the physical sciences has had inevitable
repercussions on economics. In this paper I describe theoretical models that show
that complicated deterministic dynamics may arise even in the most standard
economic environment, and some of the attempts to evaluate empirically the
importance of these nonlinearities in economics and finance. There seems to be no
evidence that 'deterministic chaos' can adequately describe economic data, but some
evidence of a role for nonlinearities.

1. Introduction

The apparent success of 'chaos' in the physical sciences, has had inevit
repercussions on economics, and some recent theoretical work has focused on t
of nonlinearities in economic dynamics. This is, in a sense, a revival of earlier
in the study of economic fluctuations (Kaldor 1940; Hicks 1950; Goodwin 1951)
regarded the market mechanism as dynamically unstable and tried to
economic fluctuations as the output of nonlinear deterministic dynamical syste
However, at least since the early 1960s, the profession had largely switched to a
on linear (really log-linear) models where exogenous stochastic shock
unforecastable changes in technology, in monetary or in fiscal policy) w
transformed, through the economy's propagation mechanism, into low-order l
stochastic difference equations, that in turn generated cyclic processes that mim
actual business cycles.
There seem to have been at least two reasons that led to the dominance of the
linear stochastic difference equations approach. The first one was the fact that the
nonlinear systems seemed incapable of reproducing some of the most obvious aspec
of economic time series. At best, such models were able to produce periodic motion
and the examination of the spectra of economic time series showed the absence of th
spikes that characterize periodic motion. This objection is now widely known to be
invalid in view of results (Sakai & Tokumaru 1980) that show that determinis
systems can generate spectra that would exactly reproduce those of random system
The second reason was the relative empirical success of the models based on lin
stochastic difference equations as well as the lack of evidence of any gains with th
introduction of nonlinearities. Examination of asset prices, that economists h
long understood to result from many of the same forces that govern output,
seemed to favour stochastic linear models. For the most part a random walk seeme
to adequately describe stock returns over periods at least as long as a week (Fa
1970). The development of new algorithms that in principle could detect whet
time series that look 'random' were actually the product of low-dimensional
Phil. Trans. R. Soc. Lond. A (1994) 346, 235-250 ? 1994 The Royal Society
Printed in Great Britain 235

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
236 J. A. Scheinkman

deterministic dynamics, has led economists to re-examine the adequacy of the

stochastic linear models.
The interest on nonlinear dynamics goes beyond the theoretical question
concerning the intrinsic stability of market economies. Predicting asset returns o
the basis of past patterns has been for a long time an active industry, and 'chaotic
dynamics' seemed to give it new respectability. Economists have long been sceptica
of the possibility of survival for long periods of stable laws that would allo
'technical traders' to exploit profit opportunities. This is a result of viewing mark
prices as resulting from the actions of economic agents, many of whom are constan
trying to find profit opportunities. In fact, as discussed below, asset prices may,
principle, display chaotic behaviour while, at the same time, no possibility f
'arbitrage' is present.
In this paper I will first describe some of the theoretical models that show that
complicated deterministic dynamics may arise even in the most standard economic
environment. Though asset returns data lead one naturally to consider econom
were a certain amount of 'uncertainty' is present, in these stochastic economies th
may still be an important role for nonlinearities, since the conditional distribution
future state variables is, in general, a nonlinear function of the current values of t
state variables. The deterministic models can thus still be used to try to understan
the economic forces that may be responsible for part of the apparent uncertainty
Another entirely different matter is whether these potential nonlinearities
needed to explain actual economic data. Although the most successful efforts
confront dynamic equilibrium models with data involve parametrizations where, i
the absence of shocks, fluctuations will be absent, there is much left to be explain
in the cyclical behaviour of modern economies and in asset prices (Murphy et
1989). In an attempt to assess the importance of nonlinearities, some research
have examined economic data using measures developed in the mathematics
physics literature such as the correlation dimension and Liapunov exponents,
using statistical techniques based on these measures. In what follows I will disc
some applications to economic time series. There is no evidence that 'determini
chaos' can adequately describe economic data, but some evidence of a role fo
nonlinearities. However, economic forces limit the form of the nonlinearities and,
particular, there has been no successful demonstration that one may produc
profitable trading strategies making use of these nonlinearities.

2. Equilibrium models
In modelling economic dynamics we must confront the fact that the agents whose
behaviour are generating the dynamics are themselves observing and trying to
forecast the actual dynamics. Perhaps this is simpler to understand in the context of
stock prices. Suppose that a vector of stock prices follows a dynamics Pt+l = f(Pt, Pt-l,
..., Pt-j) and that given a particular history of prices (Pt, Pt-1, ... , Pt-j) the rate of return
as of time t, p+l/pt of an asset i is superior to that of other assets (for simplicity we
assume that no dividends are being paid to holders of the asset between t and t+ 1).
Speculators that become aware of the dynamics would buy asset i at t and plan to
sell it at t + 1 to profit from this opportunity. This demand by speculators would by
itself pressure the price of asset i at t upward, destroying the original dynamics.
Though stock speculators may not have available the most sophisticated tools for
reconstructing the dynamics from market observations, the continuous observation
Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
 Nonlinear dynamics in economics and finance 237

of market prices must lead them to discover at least some of the profit opportunities
and their attempts to profit from the uncovered patterns should alter the dynamics.
Of course this does not rule out the possibility that profitable trading patterns may
survive for short periods.
Though there is a general agreement among economists that, as in this example,
agents' forecasts of future values of the state variables and the effect of these
forecasts on the future values of these variables must be modelled explicitly, there
is much less consensus on how to accomplish this. At one extreme are the 'rational
expectations' or 'perfect foresight' dynamic equilibrium models. Here one tries to
derive aggregate behaviour from assumptions on the tastes of individual consumers
and technologies available to producers as well as the market structure and by
postulating that the economic agents in the model completely understand the
structure of the model. In the stock price dynamics discussed above, this approach
will require that the law of motion f be such that when agents forecast future prices
using f, the demand for assets equals the supply. One should notice that the word
equilibrium is used here to mean market clearing. The actual price path Pt is certainly
not constant and may in fact exhibit very complicated dynamics. An important
aspect of this 'perfect foresight equilibrium' is the consistency between the law of
motion of price that agents perceive and the actual law. This consistency insures that
there are no incentives for agents to revise their forecasting rule.
It is, of course highly unlikely that, if prices follow complicated trajectories, agents
can learn to forecast perfectly future prices or dividends, even after a long string of
observations. The same property that makes chaotic systems look as if they are
random - their sensitive dependence to initial conditions - also makes the task for
forecasting future values extremely difficult. An outside observer (economist) may
very well decide to treat the problem using statistical techniques. The fact that he
is mistakenly treating the data as if arising from a random system will not, by itself,
invalidate his statistical methods. If the observer allows for a sufficient number of
parameters, he can in fact, with enough data, approximate well the true law
motion. If the forecaster uses some simple statistical model, e.g. a linear model, t
forecasts will, even in the limit, still display error but in any case the forecastin
activity of the economist does not alter the dynamics. In actual economies, howev
agents forecast, at each time t, future prices and rates of return and make buying
selling decisions. The aggregate decisions, in turn, affect the rates of return of ass
that each agent is trying to learn. In order to define the price dynamics we m
postulate the mechanism used by agents to learn about the actual dynamics.
Though some progress has been made concerning equilibrium models with so
special learning rules in the presence of exogenous shocks (cf. Bray 1982; Marcet &
Sargent 1987; Guesnerie & Woodford 1992), much less is known in the contex
deterministic but complicated dynamics. In any case, rational agents facing t
problem of forecasting a variable that follows 'chaotic' dynamics, may very
treat it as if the future values of the variables they are trying to predict are at
in part random. In this case we may have to model the economy as if uncertainty
present. From this viewpoint, it may be less crucial to decide whether '
uncertainty is present, say through random shocks to the future dividends of th
stock in the example discussed above, or whether uncertainty is a result of ag
limitations in forecasting. (This has a flavour of the results in the 'sunspot' litera
surveyed in Guesnerie & Woodford (1992).) It should be noted, however, that in b
cases the realizations of the 'random variables' are likely to affect the futu
Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
238 J. A. Scheinkman

evolution of the state variables. If, for instance, prices of certain assets turn out to
be higher than expected at a certain date, agents consumption and savings decisions
will be affected. This in turn is likely to affect the future price of the assets. In other
words the dynamics of such an economy is given, at best, by a system such as:
Pt+l = g(Ptq,t) as opposed to a deterministic system in which the observables are
subject to random noise.

3. A dynamic Robinson Crusoe economy

The simplest dynamic equilibrium model is one that describes an economy with a
fixed technology and homogeneous agents. The homogeneity of agents insures that
this economy is essentially a one agent economy much like the Robinson Crusoe
economy familiar from economic textbooks. This is certainly not realistic, but will
allow me, without spending an inordinate amount of space, to raise some of the issues
that I want to discuss.
To describe fully such an economy we will need to specify the productio
possibilities and how consumers evaluate the different consumption streams, i.e.
utility that consumers obtain from consumption sequences. Assume that there is
fixed set of n capital goods and a single consumption good. Consumers are suppos
to discount future consumption using a discount factor 0 < 8 < 1. That is,
representative agent attributes to a sequence {ct}to the utility t,"o tCt.
production possibilities are the triples (x, y, c) representing the combinations of
of period capital stocks y eR and c consumption amounts c eR that can be obtaine
from an initial capital stock vector xeRn at any period t. I will assume that there
exists a compact convex set B c Rn, containing (0,0), and a C2, strictly con
function v defined on B, with values on the non-negative reals, such that a t
(x, y, c) is feasible if and only if (x, y) eB and, 0 < ct < v(x, y). Further, v is assu
to be non-decreasing in its first n coordinates and non-increasing in its la
coordinates. The real number v(x, y) measures the maximum utility flow achieva
in the period if the initial vector of capital stocks is x and the final vector is y.
monotonicity imposed on v corresponds to the natural hypothesis that in order t
achieve a fixed final vector of capital stocks, it is easier to start with a higher cap
stock and that it requires consumption sacrifice to accumulate capital.
compactness of B will keep the possible capital stock paths {xt}t?O in a boun
set. (It would be more natural to assume only that {yeRn I there exist x eR w
(x, y)eB} is compact. This, however, would add no generality.) The convexity
and concavity of v correspond to the absence of increasing returns. This absence
increasing returns is necessary to show that the behaviour of this economy can b
mimicked by a market economy as discussed in the next section. With this notat
one may write the problem that the representative consumer solves as:
Problem (P):
Max 8 3tv(xt, xt+) s.t. (xt, xt+,)B and x 0 > 0 given.
O0t<oo

I will write V(x) for the value of problem (P) when xo = x. No

V(x) > -oo, for any x such that there exists a sequence {xt}t7o
xo = x. This will be true, in particular, if (x, 0) eB. The function
equation:
V(x) = max {v(x,y) + V(y)}. (1)
(x, y)eB

Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
 Nonlinear dynamics in economics and finance 239

4. Complicated dynamics
The economy described in the previous section is equivalent to a standard convex
dynamic optimization problem. None the less dynamics of the state vector xt can be
remarkably complicated. In fact, the following result by Boldrin & Montrucchio
(1986) states that one may arbitrarily choose the dynamics associated with problems
such as (P):
Proposition 4.1. Let X be a compact, convex set in Rn, and f :X-X be a C2 map.
Then there exists convex and compact B c X x X, such that for any x eX, there exists
y eX such that (x, y)eB, and smooth strictly concave function v:B- R+, increasing in
the first n coordinates and decreasing in the last n coordinates, and de (0, 1) such that
xt+ = f(xt) defines the unique solution to:

Max E Jtv(xt, xt+) s.t. (x,xt+) eB and xo O given.

Ot<oo

Further, v and 8 can be chosen such that if V(x) denotes

the function V :X- R is C2 and strongly concave.

5. Examples
The proof of Proposition 4.1 is constructive and o
candidate dynamics f, a function v and a discount f
solves P. Proposition 4.1 shows that complicated dyn
with the standard assumptions of competitive models,
into the economic logic that leads to the optimal behav
this one would like to start with certain parametrized
parameters can, in principle, be matched to data on act
for certain parameter values the resulting optimal poli
would also help in judging the likelihood that chao
economies. Since solving problem P explicitly is no
strategy is to show that, under certain 'natural' assump
policy function f is chaotic.
All fully worked out examples in the literature on eco
from purely deterministic models as the one des
Woodford's (1990) survey and references therein) inv
n = 1, even though intuition indicates that multidimen
to give rise to complicated dynamics. When n = 1, typi
guarantee the existence of a period three cycle for the
hence topological chaos. For concave and smooth v,
condition for a period three cycle (x,,x2,x3) in the
ie {1,2, 3}, if x4 = x1 and x5 = x,2 the associated Eule

v (xi)Xi+1) +dX (Xi+l Xi+2) = 0 (2)

Though equation (2) has been used to show that chaotic dynamics m
context of some parametrized examples, the parameter values that are
not reasonable. (For instance setting B = {(x,y)eR+ :x 1, y 1, yy <
= (1-y)P (x-yy)" with 0 < y < 1, 0 < c < 1, and 0 < ca+fi 1.) In
discount rate 8 is seldom above 0.3, which indicates a rate of inter
Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
240 J. A. Scheinkman

per period. (If the utility function v is strongly concave, for 8 sufficiently close to 1,
all optimal trajectories must be convergent (Scheinkman 1976). Hence complicated
dynamics require, other things held equal, a relatively large discount factor.
However, other forms of non-convergent behaviour, e.g. periodic orbits have been
shown to appear in similar examples at much more reasonable value for d.) More
reasonable examples may be obtainable in the presence of many capital goods, but
none have yet been produced. (Several types of 'market imperfections' can also be
used to generate complicated dynamics. Another alternative is to deal with
'overlapping generations' models (cf. Boldrin & Woodford 1990).)

6. Price dynamics
As stated, the Boldrin-Montrucchio result deals with the trajectory of capital
stocks in a centrally planned economy. One interest here is in the behaviour of
market economies and in fact much of the available economic data, specially high
frequency data, refer to prices (and sometimes quantities traded) of assets. A
decentralized version of our artificial economy can, however, be constructed, in a
way that is similar to the usual textbook Robinson Crusoe story. In this section I
discuss how this decentralized economy can be constructed. More precise statements
can be found in Appendix A.
The assumed convexity of the set B and concavity of the utility function v is
enough to allow us to characterize a solution to Problem P as resulting from an
equilibrium of a dynamic economy. In this economy at each time t there are three
sets of markets open. In the first set of markets, capital goods that are used as inpu
in the production at t are traded in exchange for the consumption good produced a
t. In the second set of markets capital goods that are produced at t are exchanged f
the consumption good produced at t. There is also an idealized stock market that I
discuss below.
A profit maximizing firm is assumed to own the technology B. At each period t,
firm buys capital goods produced at t- l from consumers, and uses them to produ
the consumption good and new capital goods that it sells to consumers. The f
takes prices as given. Notice that the firm faces a purely static problem of pr
maximization and we need not make any assumptions concerning how the fir
managers forecast future prices.
The problem for the representative consumer involves, on the other han
interaction across periods. At each point in time he has to decide how many capit
goods of each kind he wants to have next period to sell to firms. This, in turn, requ
the consumer to forecast prices that will prevail in the future. Since I hav
normalized the technology in such a way that capital goods last only one period, t
be able to discuss the prices of long lived assets, I will introduce an extra a
market. One share of the firm entitles its owner in period t to receive a dividend
equals the total profit the firm realizes at t. These shares are also traded in
competitive market in exchange for consumption goods.
A representative consumer takes as given the sequence of future prices in all th
sets of markets as well as the sequence of future profits that the firm will pay t
stockholders. A representative consumer chooses at each t > 0, an amount c
consume, an amount xt+1 of capital stock to carry into the next period as well as
amount 6t+x of shares to the profits of the firm. In each period t, a consumer's cho
must satisfy a budget constraint that states that the consumption and acquisition
Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
 Nonlinear dynamics in economics and finance 241

assets at a period t must be financed by the sale of assets or dividends received. The
objective of the consumer is, as before, to maximize tco &tct.
An equilibrium is a sequence of prices for capital goods inputs and outputs, and the
firm's share, today and at each future period, such that if consumers forecast these
future prices then their actions will ensure that in all three sets of markets, demand
equals supply at every t. This notion of equilibrium makes strong requirements
concerning the consumers foresight and missing is any explanation of how consumers
would arrive at this forecast, but, at this point, I merely want to show that
complicated dynamics may arise even in an economic world in which drastic
simplifying assumptions have been made.
As shown in Appendix A, whenever the optimal solution to P is interior, the
equilibrium prices of inputs pi satisfy pT = V'(xt) and hence,
pt+1 Vf( l(p)))-
Since V is strongly concave, the dynamics of pT is equivalent to that of xt.
dynamics of the vector of input prices pi can be arbitrarily complicated. Th
result holds for either the price of capital goods output at t, or for the dynam
share price. Hence the Boldrin-Montrucchio result can be used to establis
a very simple competitive economy we may obtain, as an equilibrium, co
paths for the capital stocks, prices of capital goods or stock prices.
These simple dynamic economies do, however, impose other, very
restrictions on asset prices. The rate of return at time t of an asset is equal t
plus dividends at t + 1 divided by its price at t. If, at time t asset 1 has a hi
of return than asset 1', a consumer that buys asset 1' at time t can obtain a
consumption at t+l, without lowering his consumption at any future
buying asset I instead of 1'. If two assets have a different rate of return, a
will not hold any of the asset with the lower rate of return. Hence, in equil
all assets that exist in positive amounts must have the same rate of return.
asset prices can exhibit complicated dynamics, asset dividends must ad
equalize returns.
The equality of ex-post returns, which obviously does not hold even
approximation, can be readily used to reject these purely deterministic m
discussed above, in these economies consumers are assumed to have perfect f
concerning future prices and dividends. In these perfect foresight economie
able to show that equilibrium prices can follow very complicated traject
rates of return must be equalized across all assets. As we argued above, it is
unlikely that, if prices follow complicated trajectories, consumers can
perfectly future prices or dividends, and in fact these economies may l
exogenous stochastic shocks are present. Further, the particular realizations
'random variables' are likely to affect the future evolution of the state vari
for instance, prices of the capital goods that are used as inputs in the pr
process turn out to be higher than expected at a certain date, consum
probably try to save some of their unexpected income. This in turn will aff
supply of capital goods next period and the distribution of future output. In
words the dynamics of such an economy is given, at best, by a system
Xt+l =f(xt,,t), as opposed to a deterministic system in which the observ
subject to random noise. (If the technology set B (or the function v) is s
each t to an independently and identically distributed shock At, then the so
problem (P) is given by a stochastic difference equation such as xt+l =f(
Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
242 J. A. Scheinkman

we further assume that enough markets exist then we may decentralize the economy
much in the same manner as above. In this case asset prices will also follow
stochastic difference equations (cf. Lucas 1978).)
The presence of uncertainty does not, however, eliminate the role for nonlinearities.
The same forces that create nonlinearities in the deterministic economies I discussed
above, can also create nonlinear dependence when randomness is present. (Benhabib
& Nishimura (1989) illustrate this point.) Suppose the state variables follow a system
such as xt+1 = f(xt, it) where at is a random variable, and f is nonlinear. Estimating
a linear system from the data may lead to an exaggerated view of the role of
uncertainty. Another entirely different matter is whether these potential non-
linearities are needed to explain actual economic data. The next section explains
some tools developed to answer this question and discusses some empirical results.

7. Statistical tools

The earlier efforts in applying the ideas of chaotic dynamics to uncover no

dependence in economic data (cf. Scheinkman 1985; Brock 1986) consisted sim
using certain tools developed in the mathematics and physics literature in a r
direct way. The application of these techniques to economics present se
problems. First, the time series in economics, with perhaps a few except
finance, tend to be much shorter than it seems necessary to obtain good esti
(Ramsey & Yuan (1987) contains a discussion of the statistical properties of
dimension calculation with small data-sets. Ruelle (1989) makes some important
admonitions concerning dimension calculation.) Secondly, as I argued above, it is
unlikely that the economic time series of interest are generated by purely
deterministic systems. Further, the uncertainty is likely to affect the dynamics itself
as opposed to merely affecting the observations. None the less the earlier work
suggested that nonlinearities may be present in certain economic time series and
inspired the development of asymptotic distribution theory for some statistics
related to the correlation dimension (cf. Brock et al. 1987). (There are of course many
other statistical techniques designed to detect the presence of nonlinearities that
have been applied to economic time series (cf. Engle 1982; Hinich 1982; Tsay 1986).
However, I focus here exclusively on methods related to nonlinear dynamics.)
Let y1, y2, ..., be a sequence of vectors in Rp. For each y > 0, let

Cm r(m- 1) -i<j<m (3)

where 0(a) = 0 if a < 0, and 0(a) = 1 if a > 0. Here, yi-yjl
Intuitively Cm(y) denotes the fraction of the first m vecto
of each other. For each y, if the limit exists, let
C(y) = lim Cm(y). (4)
m?-yo

The quantity C(y) indicates the fraction of all vectors that are within y of ea
(Recall that the correlation dimension of {Yt}o0 is defined as d = lim,0 log C
y.) If x, x2, ... is a sequence of real numbers, for N > 1 each zN = (x, xt+
will be called an N-history. For each m, let
2 N-1
CN(y) = ( 1 [O(7-Ixo+k-xj+l)], (5)
Ph\il. Traes. -. Soc. L<jom k=oA
Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
 Nonlinear dynamics in economics and finance 243

and, if the limit exists, CN(y) = lim CN(y). (6)

mn--oo

Clearly, CN(y) is the fraction of the first m, N-histories that are within y of each other
and similarly for CN(y).
If each xt is an observation of independently and identically distributed (IID)
random variables then one should expect that, for m large,

Cm(y) = [Cl(y)]N. (7)

In fact, it can be shown (see Appendix B for
positive numbers VN m, that can be comput
an observation of IDD random variables, then

WN(y) = V/m[CN(y)- (C(y))N]/VN, m (8)

is asymptotically distributed as a normal distribution with mea

variance. This distribution free statistics can be used in testing for th
nonlinearities. A key point in establishing (8) is the recognition th regn
U-statistic in the sense of Hoeffding (1948). (Originally U-statistics w
the case where yy, Y2, ..., are IID. A symmetric function h :RN R is a
, = Eh(y, ..., YN). Corresponding to the kernel h there is a U-statis

U(yl, ... Ym) = (N E h(y, , . .,YiN)

where the summation is over all (N) combinations of N distinct el

from { 1, ..., m}.) Several results concerning central limit theorem for
in the literature. Modern treatments of the theory of U-statistics ca
Sefling (1980) and Denker & Keller (1983). The fact that CN(y) is U-sta
be used to show asymptotically normal behaviour of other statisti
correlation dimension, including estimates of the slope of log [CN(y)]
Frequently one is interested in finding nonlinear dependence on
particular models fitted to the data. In many macroeconomic tim
example, low-order autoregressive models are known to yield a
analyses of foreign exchange rates, ARCH models (cf. Engle 1982) wer
(1989) to pre-filter the data. In practice one can proceed as in Scheink
(1989a, b) to examine the distribution of the estimated residuals. Firs
estimated and a set of residuals is generated. These residuals are rand
and data-sets are then reconstructed using the estimated model. In ea
sets one re-estimates the model and measures the Wm(y) statistics on
This 'bootstrap' like procedure is then used to determine the significa
of the statistics in the original residuals. Another possibility is to dete
on the variance of the estimator caused by the fact that the residual
Results of this type can be found in Brock et al. (1990).

8. Empirical results
The statistics discussed in the preceding section as well as metho
the mathematics and physics literature, have been used to detect evid
or at least for the presence of nonlinearities, in economic time series
Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
244 J. A. Scheinkman

In macroeconomics, time series are simply too short and noisy. Most macro-
economic data have a quarterly or at most monthly frequency. There are a few
economic time series that have been produced for a very long period but, in these
cases, there is usually strong evidence against stationarity. Statistics such as (8)
above, as well as others, have been used to detect evidence for nonlinearity in
economic time series including U.S. industrial production and unemployment series.
Financial time series seem immune to many of the problems that plague
macroeconomic data. First, prices of assets are observed rather directly and this
avoids many of the measurement issues concerning macroeconomic time series.
Secondly, data are sampled at high frequencies. Nonetheless the application of tests
developed in the mathematics and physics literature have not led to any convincing
evidence for deterministic chaos.
Attempts to use statistics based on the correlation dimension, such as (8), to fi
evidence for the presence of nonlinearities have met with more success. Scheinkm
& LeBaron (1989a) found strong evidence to reject the hypothesis of IID innovatio
in weekly and daily returns on the value weighted portfolio from CRSP. (The Cen
for Research in Securities Prices at the University of Chicago.) Potential sour
these results were several calendar anomalies that had been detected earlier in stock
returns including monthly by Ariel (1987) and weekly by French (1980). LeB
(1988) found that accounting for these anomalies did not alter the reported reje
of the hypothesis of IID innovations.
Fama (1965) following a suggestion of Mandelbrot, found some evidence that l
absolute price changes in certain stock prices tended to be followed by other lar
absolute price changes. The ARCH model and its GARCH variant (cf. Engle 1
Bollerslev 1986) provide parametric versions that capture conditional hetero
skedasticity, that is the fact that the variance of the distribution of price chan
conditional on past prices, is not constant. Hsieh (1990) discusses the literature a
presents further evidence indicating the presence of conditional heteroskedasticit
weekly and daily stock returns. To examine whether these models could
accommodate the observed departures from IID innovations, LeBaron (1991) looked
at GARCH residuals of the weekly returns on the CRSP value weighted portfolio. The
estimates based on the whole sample (1962-86) showed that the normalized GARCH
residuals continued to show significant departures from IID innovations. However, if
the sample is divided in two halves, one cannot reject the IID hypothesis on the
residuals of the second half. This result raises questions concerning the stationarity
of the return series, which is frequently an implicit assumption.
The departure from IID residuals would seem to indicate that a prediction of
returns superior to the forecasts implied by the random walk model could be made.
LeBaron (1988) tried locally weighted regressions and was unable to beat the random
walk prediction in a convincing manner. His results seem to indicate that the
departures from the random walk may occur principally through the forecastability
of the second or higher moments. Mayfield & Mizrach (1989) also examined the
question of predictability in intra-day data. They examined data on the S & P 500
stocks average, sampled once every twenty seconds and concluded that predictions
superior to random walk predictions could not be made even five minutes ahead.

Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
 Nonlinear dynamics in economics and finance 245

9. Conclusion

In this paper I surveyed the impact of some of the recent developme

mathematics of nonlinear deterministic time evolutions in economics. From a
theoretical viewpoint it has helped economists understand that, at least in prin
some of the apparent randomness in economic time series may be the result of
presence of nonlinearities. Eventually this may prove to be a fruitful appr
explaining the observed instability of economic time series, but this must wait
development of more 'realistic' theoretical models. Examination of asset re
indicates the presence of nonlinear dependence in several financial time
though persuasive evidence to favour chaos has not been uncovered.
The potential use of dynamical systems methods to predict future asset
seemed to give new respectability to 'technical trading'. The published literatur
date does not support this view though there is, of course, the possibilit
methods of dynamical systems may be used to quickly detect temporary p
that may appear in asset price series. These results do not surprise most econo
There is a lot of resources dedicated to finding strategies that yield extraordin
returns. As such opportunities are detected, the effect on prices of the co
actions of agents trying to make profits tends to destroy the existing pattern
some point in time, nonlinearities allow for strategies that generate abnormal p
market forces should eventually lead to a change in the dynamics of stock pri

Appendix A
In this appendix I state more formally the results discussed in ?5. If one assu
that the technology set B is 'sufficiently productive', the postulated convexity
the convexity of B and the concavity of the function v) is enough (cf. McKenzie
to guarantee that if the sequence xt, t > 0 solves (P), there exists a sequenc
such that, for each t, and each (x, y) eB

dtv(xt, xt+l) + qt+l 't+l --qt ' xt >- 3tv(x, y) + qt+, y - qt x

Further, qt/8t is a subgradient of the concave function V at xt. (With the assum
we have made, it suffices to add the hypothesis that f(X) has a non-empty int
and that x0 is in the interior of X. Here, f:X-+X is the candidate optimal
function. McKenzie (1976) contains much weaker hypothesis.)
This result can be used to give a 'decentralized' interpretation to our econom
profit maximizing firm owns the technology. At each period t the firm buys
goods produced at t- 1 from consumers, and uses them to produce the consum
good as well as new capital goods that it sells to consumers. I denote all period t
in terms of period t consumption good. Write pt for the price of the capital go
is used as input to the production process at t and p? for the price of the capita
that are produced at t. The firm takes the prices as given, and chooses amo
feasible production plans the one that maximizes profits. That is, the firm sol
Max{v(x,y)+p 'y-p 'x} s.t. (x,y)eB. (10)
Since v is strictly concave, there is a unique (x, y) that solves
as (z'(p,po),z?(p',pt)). Let 7r(p},p ) denote the maximal valu
One share of the firm entitles its owner in period t to receiv
the total profit the firm realizes at t. These shares are traded i
Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
246 J. A. Scheinkman

and let st denote its time t price, after the dividend is paid. Again this price is in terms
of time t consumption good.
The consumer takes as given the sequence of prices (p,pio, St) for t > 0. He also
takes as given the sequence of profits 7t that the firm will pay to stockholders, at each
t > 0. The representative consumer chooses at each t > 0, an amount ct to consume,
an amount xt+l of capital stock to carry into the next period as well as an amount
Ot+1 of shares to the profits of the firm. In each period t, the consumer's choices must
satisfy the budget constraint

c~ +p?'?t+1
ct +8~(01+ -0~) < <p Pt
+ pto Xt+l+8t(Ot+1-t) 'xttt+tt
+ 7t (l
. (I)
l)
The right-hand side of ( 1) defines the income of the consumer at t
sale of capital goods to the firm and of dividends that are pr
stockholdings. The left-hand side gives us his total expenditu
acquisition of capital goods for future sales and changes in his stoc
s = {St}t=o, 7T = (7Tt}o0, etc. The objective of the representative co

Problem Q(s, 7r,p, p0)

Max Stct
0<t<oo

s.t. (11)andx0 0 given, 00= 1, xt>O and t>0. (12)

Let ct(s,7r,pi,po), xt(s, r,pi,pp) and Ot(s, r,pi,p?) for t 0 denote
Problem Q(s, T, pi, pO), if it exists.
An equilibrium is a triple of non-negative sequences, pi,po, s, such
t 0,

t(8, 7T, p,p = 1, (13)

Xt(s, 7T, pi, po) = zi(pi, po), (14)
Xt+l(S, 7T, p, p) = z?(pt,pt), (15)
ct(, 7r, pi,p?) = v(zi(pi,pt), z(p,p) ). (16)
Equation (13) guarantees the clearing of the stock market. Equation (14) sta
the quantity of capital demanded by the firms for current production
quantity offered by the representative consumer. Equation (15) state
amount of capital goods produced by the firm equals the amount that the
desires to buy. Finally equation (16) guarantees that the consumption go
is in equilibrium. In words, an equilibrium is a sequence of prices for capit
inputs and outputs, and the stock, today and at each future period, su
consumers forecast these future prices then their actions will ensure that m
clear at these prices at every t.
Though an equilibrium involves an infinite number of equations, it
characterize an equilibrium in the Robinson Crusoe economy by making us
'support property' (9). In fact, one may prove:
Proposition A 1. Suppose xt solves (P) and qt are the support prices defin
Then the sequences pi, po and s given by,

Pt = /t, (17)
p = qt+l//t, (18)
St = V(xt)- (qt/8t) xt- 7(qt/8t, qt+l/8t), (19)
Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
 Nonlinear dynamics in economics and finance 247

form an equilibrium. In this equilibrium, for each t > 0, xt(s, 7T, pi, p0) = xt, and ct(s, Tr,
p, 0) = v(Xt,Xt+l).
Proof. From (9), it is clear that zi(qt/5t, qt+1/8t) = xt, and zo(qt/,t, qt+l/8t) = xt+,. It
suffices then to show that ct = v(xt, xt+1), xt and Ot = 1 for t > 0 solve P(s, 7r, pi, p).
Note that
iPt+ = Pt, (20)
8(7Tt+1 + t+i) = st. (21)
Since, p'/edOV(xt+1), and V > 0, Pt' xt+i V(xt+
it follows that limt , 6t(p" xt+1 + st) = 0. This c
the optimality of ct = v(xt, xt+1), xt and Ot = 1.
This result allows us to relate equilibrium price
obtained in Proposition 4.1. If xt is an interior p
will satisfy
Pt+1 = V(f(V 1())) (22)

and this dynamics is equivalent to t

Further, from (19), we have that if xt

st = V(xt)- (qt//t) Xt-r7T(qt/ /xt, q

Since V is strongly concave, if n = l, w

St+1 = g(st),

and the dynamics of st is equivalent t

show, as in Takens (1983), that 'gene
equivalent to observations of xt. Th
determines the 'observable' cannot be
combination of Propositions 4.1 and P
in very simple dynamic economies p
These simple dynamic economies do, ho
The rate of return between t and t +
at t+ 1 divided by its price at t. Equ
of the stock equals 1/S. Similarly (20)
j, j = 1,
= 2,
~ V...,
", n,
, ' at t, pay p and sells it
]gt+l
1/8. The fact that the rate of retu
of numeraire, the units in which p
the price at which the firm sells th
would no longer hold, except in v
of return would fluctuate over tim
return, at each time t, is the same
a consequence of the consumer's o
return, a consumer will not hold
Though asset prices can exhibit c
to equalize returns.
Appendix B
This appendix contains further
theory for some statistics relat
exposition in Brock et al. (1990).
Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
248 J. A. Scheinkman

Let F be the common distribution of the xts and,

c= JO(y - u-vl)dF(u)dF(v), (23)

(23)

K = [F(u + )-F(u- y)]2dF(u). (24)

Notice that K > C2, and Dechert (1988) showed that unle
constant then K > C2. I can now state:

Proposition B 1. If {xt}tO1 is IID and if K > C2, then for N > 2, as m -> 0

/m [C(y) - (C (y))N]/ -]>N(O, 1), (25)

where

VN = 4[N(N-2) C2N2(K- C2) +KN _ C2N]

N-1

+8 E [C2j(KN-j-C2N-2j) -NC(2N-2)(K-C2)]. (
j=l

Proof. Notice that CmN(y)-[Ci(7)]N = g[CN(y), C(y)], where g(u,v) = u

Since \/mG(Xm) is asymptotically normal if /mXm ->N(O, ) and G has non
gradients (this uses what is known in statistics as the delta method) it suffices to
that

/m ([CN(y), Cl (y)] - [ON, C]) -N(O, N ). (27)

More precisely from (27) it follows that (cf. Serfling 1980, ch. 3),

/m {C(y)- [C (y)]N} ->N(o, 0o),

where - = LL-au
uug '] Zov]
ovj -auu {1 NCN- {1 ,NCN-}.
In turn, (27) follows if one can show that for any pair (A1, A2),

V/m {(A1 CN() - A2 C (y) -1 ON - A C} ->N(O, -(A1, A2)). (28)

In fact in this case, the entries of Z are given by, 1l = 0r(1,0), s22 = ((0, 1), ' 12
21 = [O(1, 1)--(1, 0)--(0, 1)].
To show that (28) holds, note that for any pair A1, A2 and for any rea
histories z,, 2, ..., Zm we can write,

A1CN(y) 2C(y) = [2/m(m--)] h(,zj), (29)

l i<j<m

N-1

where h(zi, zj) = A1 I O(y-Ixi+k-Xj+kl)+A 2 0( 7-lxi--xj

k=O

and zi = (x, x+, ..., xi+N_). The function h is symmetric, i.e. h(6l, 62) = h(62, 61) f
any pair of vectors (1,62)e R2N. Hence A1CN(y)+A2C(y) is a U-statistic,
furthermore, even though two arbitrary histories zt and ZT are not in gene
independent, they will be so if It-rT >N. Hence the theorems of Sen (1963)
Denker & Keller (1983) apply and since Eh(zi,zj) = AlCN+A2C if li-jl
\/m {A1CN( ) + C() + () CN - A2 C} is asymptotically normal. The formula fro
Sen (1963) can be used to compute the variance o(A1, A2) and to obtain (26).
Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
 Nonlinear dynamics in economics and finance 249

In the above proposition the constants C and K involve the actual distribution F.
However, let
am m
1-- 2 0(7y- Ix- xtl), (30)
s=1 t=l

and Km6y- \,=

n Km -' m1\6 -,-X\), (31)
0(YO(Y-IXr-Xsl)O(Y-IXs-Xtl),
r=l s=l t=l

and V, m equals the right-hand side of (26) wh

We have then:

Corollary B 1. Under the conditions of Proposition B 1, as m - oo,

/ [CmN()-(Cm(,r))N]/ VN, m N(O, 1).

Proof. Cm and Km are V-statistics (cf. Serfling 1980) which converge almost sur
to C and K respectively. Thus VN m converges a.s. to V and by Slutsky's theorem
Proposition B 1 above we have the result.

References

Ariel, R. A. 1987 A monthly effect in stock returns. J. Financial Economics 18, 161-174.
Benhabib, J. & Nishimura, K. 1989 Stochastic equilibrium oscillations. Int. Economic Re
85-101.

Boldrin, M. & Montrucchio, L. 1986 On the indeterminacy of capital accumulation paths. J

Economic Theory 40, 26-39.
Boldrin, M. & Woodford, M. 1990 Fluctuations and chaos: a survey. J. Monetary Economics
189-222.

Bollerslev, T. 1986 Generalized autoregressive conditional heteroskedasticity. J. Econometrics

307-328.

Bray, M. 1982 Learning, estimation and the stability of rational expectations. J. Econom
26, 318-339.
Brock, W.A. 1986 Distinguishing random and deterministic systems: an expanded version.
Department of Economics, University of Wisconsin, Madison.
Brock, W. A., Dechert, W. A. & Scheinkman, J. A. 1987 A test for independence based on the
correlation dimension. Department of Economics, University of Wisconsin, University of
Houston, University of Chicago.
Brock, W. A., Dechert, W. A., Scheinkman, J. A. & LeBaron, B. 1990 A test for independence
based on the correlation dimension. University of Wisconsin, Madison.
Denker, L. & Keller, G. 1983 On U-statistics and von Mises statistics for weakly dependent
processes. Zeitung Wahrscheinklichkeitstheorie verw. Gebiete 64, 505-522.
Engle, R. 1982 Autoregressive conditional Heteroskedasticity with estimates of the variance of
the U.K. inflations. Econometrica 50, 987-1007.
Fama, E. 1965 The behaviour of stock market prices. J. Business 38, 34-105.
Fama, E. 1970 Efficient capital markets: review of theory and empirical work. J. Finance, May,
pp. 383-417.
French, K. 1980 Stock returns and the weekend effect. J. Financial Economics 8, 55-69.
Goodwin, R. M. 1951 The nonlinear accelerator and the persistence of business cycles.
Econometrica 19, 1-17.
Guesnerie, R. & Woodford, M. 1991 Endogenous fluctuation. Mimeo University of Chicago.
Hicks, J. 1950 A contribution to the theory of the trade cycle. Oxford: Clarendon Press.
Hinich, M. 1982 Testing for gausianity and linearity of stationary time series. J. Time Series
Analysis 3, 169-176.
Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]
250 J. A. Scheinkman

Hoeffding, W. 1948 A class of statistics with asymptotically normal distribution. Ann. math.
Statistics 19, 293-325.
Hsieh, D. 1989 Testing for nonlinear dependence in foreign exchange rates. J. Business, July, pp.
339-368.

Hsieh, D. 1990 Chaos and nonlinear dynamics: application to financial markets. Working Pap
Fuqua School of Business, Duke University.
Kaldor, N. 1940 A model of the trade cycle. Economic J. 50, 78-92.
LeBaron, B. 1988 Nonlinear puzzles in stock returns. Ph.D. thesis, University of Chicago.
LeBaron, B. 1991 Nonlinear econometrics for chaos: empirical results. University of Wiscon
Madison.

Lucas, R. E. Jr 1978 Asset prices in an exchange economy. Econometrica 46, 1429-1445.

Marcet, A. & Sargent, T. 1987 The fate of systems with 'adaptive' expectations. Mimeo, Stanf
University.
Mayfield, E. S. & Mizrach, B. 1989 On determining the dimension of real-time stock price data.
Working Paper 187, Department of Economics, Boston College.
McKenzie, L. W. 1976 Turnpike theory. Econometrica 44, 841-865.
Murphy, K., Shleifer, A. & Vishny, R. 1989 Building blocks of market clearing business cycle
models. Mimeo, NBER.
Ramsey, J. & Yuan, K. 1987 The statistical properties of dimension calculation using small data
sets. Mimeo, New York University.
Ruelle, D. 1990 Deterministic chaos: the science and the fiction. Proc. R. Soc. Lond. A 427,
241-248.

Sakai, H. & Tokumaru, H. 1980 Autocorrelations of a certain chaos. IEEE Trans. Acoustic Spe
Signal Process 28, 588-590.
Scheinkman, J. A. 1976 On optimal steady states of n-sector growth models when utility
discounted. J. Economic Theory 12, 11-30.
Scheinkman, J. A. 1985 Distinguishing deterministic from random systems: and examinati
stock returns. Manuscript prepared for the Conference on Nonlinear Dynamics, Paris.
Scheinkman, J. A. & LeBaron, B. 1989a Nonlinear dynamics and stock returns. J. Business, J
pp. 311-337.
Scheinkman, J. A. & LeBaron, B. 1989b Nonlinear dynamics and GNP data. In Economic
complexity: chaos, sunspots, bubbles and nonlinearities (ed. W. Barnet, J. Geweke & K. Shell).
Proc. 4th Int. Symp. Economic Theory and Econometrics. Cambridge University Press.
Sen, P. 1963 On the properties of U-statistics when the observations are not independent. I.
Calcutta statist. Ass. Bull. 12, 69-92.
Serfling, R. 1980 Approximation theorems of mathematical statistics. New York: Wiley.
Takens, F. 1983 Distinguishing deterministic and random systems. In Nonlinear dynamics and
turbulence (ed. G. Borenblatt, G. Loss & D. Joseph), pp. 315-333. Boston: Pitman.
Tsay, R. 1986 Nonlinearity tests for time series. Biometrika 73, 461-466.

Phil. Trans. R. Soc. Lond. A (1994)

This content downloaded from [Link] on Mon, 14 May 2018 [Link] UTC
All use subject to [Link]