Toward Agent-Based Models for

Investment
J. Doyne Farmer
McKinsey Professor
Santa Fe Institute
Santa Fe, New Mexico

Although agent-based models are not yet ready for practical investment application, they
can yield powerful insights about market behavior, particularly in regard to the second
order inefficiencies that create profit-making opportunities. When practical use of
agent-based models becomes possible (perhaps within the next five years), their
effectiveness will cause securities prices to change.

s far as I know, no one currently uses agent- statement by George Soros that gold prices are going
A based models for investment. In the context of
this presentation, an agent-based model involves a
to rise can have a significant effect on the price of
gold. Predicting prices is more like telling peoples
model for price formation and a model for agent fortunes. If a fortune-teller says you will take a long
behavior, including the models information inputs trip and have a terrible accident and you have any
and how the inputs behave, which could involve faith in her forecast, you will stay home. If you do
learning. From an investors point of view, agent- stay home and no terrible accident occurs, was she
based models represent a new direction that may or right? The point is that forecasts regarding humans,
may not result in more practical investment models. who have free will and can alter their behavior, can
From an academics point of view, agent-based mod- be self-invalidating. This point is the idea underlying
els are fascinating new tools that make it possible to the theory of efficient markets: If someone discovers
investigate problems previously out of reach. The a pattern, it diminishes as it is exploited. If enough
models presented here do not provide practical tools, people discover the pattern, it will disappear.
but they do offer a taste of several new directions that Informational efficiency (the lack of arbitrage) is
may or may not turn out to have practical value in a central idea of modern finance. Anybody who has
the future. spent time building market models knows that mar-
One of the things that attracted me to financial kets are arbitrage efficient at some level. Building
prediction is the possibility of forecasting the behav- models that make good market predictions is not
ior of people. People have free willyou never know easy. But markets cannot be perfectly efficient; other-
what they will do. Of course, we make predictions wise, why would so many smart people waste their
about people in our day-to-day lives. We all know time investing in them? Milton Friedman originally
people whose behavior is pretty consistent and pointed out the paradox of efficient markets: If ratio-
whose responses are predictable. But making quan- nal speculation makes markets efficient, then because
titative predictions about people in general is a real the market is efficient, no profits can be made and all
challenge, and little success has been apparent so far. the rational speculators should leave, thus causing
Because prices involve people, predicting prices the market to revert to an inefficient state. The theory
is different from predicting the weather. When the of efficient markets is inherently self-contradictory.
weatherman says it is likely to rain, this prediction Finding a self-consistent theory to replace it is an
has no effect on whether it will rain. In contrast, a important problem that demands a solution. As a
physicist, I would state the situation as follows: At
Authors note: This presentation is partially based on research done in first order, markets are arbitrage efficient, but at sec-
collaboration with John Geanakoplos, Shareen Joshi, and Paul Melby. ond order, they are not. For markets to function, there

Copyright 2001, Association for Investment Management and Research. Reproduced

and republished from Benchmarks and Attribution Analysis with permission from the
2001, AIMR Association for Investment Management and Research. All Rights Reserved. www.aimr.org 61
Developments in Quantitative Investment Models

have to be small inefficiencies to motivate investors. rience in the field of chaotic dynamics. I developed
These small inefficiencies are what sustain the ecol- nonlinear time-series models that could take advan-
ogy of the market. All the action, all the interesting tage of low-dimensional chaos and, in some cases,
behavior of markets, is in the second order. This point give improved forecasts for such problems as fluid
is clearly understood by practitioners and should turbulence, ice ages, and sunspot cycles.
also be understood by economic theorists. Prediction Company uses a purely empirical
Finance is like any other industry. Profit-making approach to market forecasting. The models are not
opportunities, or inefficiencies, are created by the based on a fundamental, reasoned understanding of
evolving needs of customers. Specialization is driven how markets work; instead, they are based on a
by economies of scale in digesting and understand- search for patterns in historical data. The fundamen-
ing information. One must regard the market as an tal premise is that patterns that existed in the past will
ecology of highly specialized, heterogeneous agents, repeat themselves in the future. These patterns are
linked together by their relationships to each other usually rather subtle. The market is pretty efficient in
and the flows of money that these relationships that extremely few persistent simple patterns exist. It
imply. is not quite so efficient with respect to complex pat-
The story I am going to tell is about how the terns, and if you look carefully and comprehensively,
demands of investors create inefficiencies. In this you may find things that very few have seen before.
case, inefficiencies arise because investors demand The big problem is that once you allow for complex
liquidity, thus driving a need for market makers. patterns, the number of possibilities is enormous and
Market makers have risk aversion, which means that careful statistical testing becomes essential.
once they acquire positions, they have to off-load Feature extraction is a key aspect of Prediction
them. But risk aversion means that their actions are Companys models. In the extraordinarily noisy envi-
somewhat predictable. For example, if they off-load ronment of the financial world, statistical modeling
their positions gradually, they cause trends in prices. methods that blindly generate models directly from
Trends in prices are exploited by technical traders, data do not work well. The problems are too much
who analyze patterns in past prices and trade accord- noise, the limited size of datasets, and the fact that
ingly. Thus, we see how the demand for liquidity can techniques have a hard time finding patterns that are
real and persistent rather than ephemeral random
sustain a population of technical traders, even
fluctuations. Throwing away distracting information
though the technical traders are not providing liquid-
is crucial.
ity directly. Money flows from liquidity demanders
The same principles apply in many other prob-
to technical traders. This pattern does not mean the
lems that require cognition, such as image recogni-
market is not pretty efficient; it just means that this
tion. Suppose you want to design a machine that can
approximate efficiency comes about only through a
recognize objects and interpret visual images. You
web of interactions among heterogeneous players,
might naively try to do it by operating directly on the
who are all part of an interconnected market ecology.
pixels of an image. But this approach has not been
But I am getting ahead of my story.
successful. We know the brain does not work this
way, as was dramatically illustrated in a famous set
Fundamental vs. Empirical Models of experiments by two neuroscientists, Hubel and
My own experience at Prediction Company and else- Weisel, who studied vision in spider monkeys. They
where illustrates the difference between agent-based gave the monkeys a variety of different visual stimuli
models, which are a fundamental approach to mod- and recorded the output of selected neurons in the
eling, and empirical models, which are a strictly prac- visual cortex of the monkeys. They showed that the
tical, data-driven approach.1 Norman Packard, Jim image processing in the cortex happens in stages. The
McGill, and I were the senior founding partners of first several stages are for pre-processing, to encode
Prediction Company. When we founded Prediction the information from the retina into features that are
Company, we did not know anything about financial easier to interpret. Neurons are specialized to
markets. We were experts in time-series modeling, respond to features, such as edges, lines, motion in a
specific direction, or differences between the right
which is an empirical approach to building models
and left eye. It appears that these features evolved to
for data that are ordered in time. I came to be involved
facilitate the process of decomposing an image into
in time-series forecasting because of my earlier expe-
understandable features and properties, such as
1
Prediction Company acts as a fund advisor and technology devel-
plants, animals, trees, and objects and their move-
opment group for a proprietary, fully automated trading operation ment with respect to each other. Irrelevant informa-
of Warburg Dillon-Read, which is an affiliate of the Union Bank of tion is discardedindeed, most of the bits of
Switzerland. information in the image are thrown away so that the

62 www.aimr.org 2001, AIMR

 Toward Agent-Based Models for Investment

brain can focus all its attention on the small subset of process by installing appropriate learning algorithms
information that is likely to be useful. and training the robot. Alternatively, the engineer
Finding the right signal-processing feature can use her fundamental knowledge of physics.
detectors to make sense out of a financial time series Without ever watching a ball get hit, she can con-
is not easy, because we have no a priori knowledge struct a model for the balls motion in terms of the
of what these features are. Evolution at Prediction laws of gravity, air friction, and kinematics. She could
Company has proceeded on a timescale of years use physics to develop an algorithm for the robot to
rather than millennia, driven by researchers formu- parse the image from its video camera eyes, estimate
lating and testing new feature detectors and accumu- the velocity and position of the ball, and use New-
lating lore that is brought to each successive tons laws to tell the robot how to forecast where the
modeling problem. This area is where we spend most ball will go and how to move in order to catch it.
of our time and where we believe we gain most of Prediction Companys approach is like the
our edge. We originally based our search for good human approach, and agent-based modeling is like
features on the technical-trading literature, which the approach of the engineer discussed above. This
has evolved over roughly a hundred years. As you analogy is apt because it illustrates the strengths and
might expect, we found that most technical-trading weaknesses of each method. In a situation with lots of
rules are useless by themselves, but we do not use historical data and little fundamental understanding,
them by themselves. Instead, we just regard them as the empirical method is superior. But it has the disad-
pre-processing steps, signal filters that separate the vantage that when no historical data are available
data into features, such as 15-day trend or maxi- (e.g., if conditions in the market are fundamentally
mum under-5-day smoothing. Although standard
different from what they have been in the past), the
technical-trading rules formed the starting point for
empirical method will fail. The fundamental method
our original models, the later models also involve
is harder to develop, but once developed, it can be
many signal filters (not based on technical-trading
powerful in dealing with situations that are outside
rules) that we discovered through a lot of hard work
normal experience. For predicting the trajectories of
and trial and error.
baseballs, I doubt that the fundamental method can
This approach involves a great deal of human
beat the empirical method. But for other tasks, such
intervention, and it is not cut-and-dried science but
as predicting the motion of satellites, the fundamental
rather an art form with statistical validation as its
method is clearly superior. (In fact, for the baseball
critic. But it is still almost entirely empirical. Although
problem, I am willing to bet that most engineers
we may structure our models based on ideas about
would opt for a mixture of the two methods.)
what makes markets tick, the models are still mainly
based on fitting these signal-processing filters to data The pros and cons of these two methods were
in concert with standard function-approximation brought home to me by an experience I had while I
algorithms, such as linear regression and artificial was in graduate school. Norman Packard and I devel-
neural nets. Ideas about how markets work drive the oped a system for beating roulette by treating it as a
choice of the inputs and shape the way the data are physical system, that is, measuring the position and
presented to the learning algorithms. But we cannot velocity of the ball after it is released by the croupier
say that the models reflect a fundamental under- and taking advantage of the fact that bets can be
standing of what people are doing and how the mar- placed until a few seconds before the ball exits the
ket works. Rather, the trading models are merely track. We built concealable computers (which was
complex, unconscious, stimulus-response machines. quite difficult in 1976) and used them to time the
(The trading is completely automated, with no motion of the ball and predict its likely landing point.
human intervention.) The early version of the hardware and software
One can make an analogy to the prediction of a implemented expressions based solely on physics. As
physical phenomenon, such as the trajectory of a a result, the program had the advantage of being
baseball. Suppose an engineer is given the task of extremely small and could be fit in the 3,000 bytes
training a robot to catch baseballs. There are two (not megabytes or kilobytes!) of memory that were
approaches: empirical or fundamental. The empirical available to store the program and the 128 bytes of
approach mimics how the ball is caught. A human RAM that were available for scratch space. The dis-
learns to catch baseballs through a process of trial and advantage was that the parameters of the physical
error, by seeing a lot of baseballs get hit and forming solution were not orthogonal, so it was difficult to
a largely unconscious set of mental maps, with distinguish the effect of a change in one parameter
empirical rules for forecasting the trajectory of the (the fall-off velocity of the ball) from a change in the
ball and responding to it. The engineer can mimic this other parameter (the deceleration rate of the ball).

2001, AIMR www.aimr.org 63

Developments in Quantitative Investment Models

Only I and one other physicist were able to reliably trage efficiency. On one hand, this decline in correla-
master the manual fitting procedure. tion can be seen as a vindication of the idea of
A few years later, when we had a lot more mem- arbitrage efficiency, illustrating how slowly market
ory and much faster computers, we revisited the efficiency is achieved. On the other hand, it can be
problem and solved it with an empirical method. We interpreted as illustrating how slowly the achieve-
formatted the data in a natural way that made use of ment of market efficiency occurs. This signal has
the fact that the wheel is a circle, and we fit a cubic persisted for at least 23 years and is not dead yet. The
spline to the data. This change made the algorithm statistical significance of the signal is so strong that it
easy to use. No knowledge of physics was required. is easily detected with only a year of data, and anyone
But this approach was not as accurate as the funda- who had found this signal a long time in the past
mental method when the parameters were properly could have made a great deal of money by now. If a
fittedbecause the empirical solution was not as risk premium is associated with this signal, the risk
constrained. factors must be extremely difficult to detect and are
Agent-based models have the potential to give not reflected in the performance of the model, which
us a more fundamental understanding of how mar- has overwhelming statistical significance. The point
kets work and better tools to predict them. They will is that although efficiency eventually makes itself felt,
have the advantage over empirical models in that it is very slow in arriving.
they can be applied to previously unseen market The second signal, plotted in Panel B, is even
conditions. Like the development of physics, achiev- more surprising. The signal begins in the early 1980s
ing this understanding will be hard fought and will and grows through time to the present. This increase
take a long time to develop. Success will depend on seems to be inconsistent with arbitrage efficiency. I
developing good agent models, either by hard-wir- do not expect this signal to continue to get stronger,
ing the behaviors of the agents or, more likely, by but it illustrates that new inefficiencies can be gener-
mimicking their learning process. The advent of prac- atedmarkets do not always become more efficient
tical agent-based models is still at least several years with the passage of time.
away. Of course, as soon as profitable models are
developed, they will begin to make markets more
efficient and eventually will invalidate their own
Simplifying Agent-Based Models
forecasts. In the far future, financial speculation After I left Prediction Company, I came to the Santa
likely will be an escalating arms race of agent-based Fe Institute, where I have returned to my roots in
models attempting to out-anticipate the aggregate complex systems theory. This time around, however,
behavior of each other. I am thinking about markets and having fun trying
to model representative agents and their interactions.
As a physicist, I naturally think about market partic-
Arbitrage Efficiency Takes Time ipants as being similar to molecules, with trading
The possibility of earning a profit in a market envi- (which necessarily involves an interaction between
ronment is largely attributable to the fact that the agents) being similar to a molecular collision. Of
progression toward market efficiency is slow and that course, the problem with this analogy is that these
new inefficiencies are created in an ongoing manner molecules are rather complicated; they are people,
by the needs and foibles of market participants. Fig- not ping-pong balls, and their behavior can be quite
ure 1 shows some declassified data from one of complex.
Prediction Companys models. We think about our The key to good agent-based modeling is to cap-
models as being composed of signals, which are ture the agent behaviors as realistically as possible
highly processed versions of a related cluster of data with a minimum of assumptions. Unfortunately, in
inputs. This figure illustrates the performance of two many peoples minds, agent-based modeling has
different signals as a function of time. The y-axis plots come to be associated with highly complicated mod-
the percent correlation of each of these signals with els in which each agent has his own customized
the following forward return of the market; the data kitchen sink. Such models are so complicated that
are smoothed to make the picture easier to view. they are almost as hard to understand as real markets.
Panel A shows that in the mid-1970s, this signal had In order to enhance understanding, and to ever have
almost a 15 percent correlation with the future move- a hope of using models for prediction (and fitting
ments of the market (by market, I mean a universe of their parameters to real data), it is important to keep
liquid U.S. stocks hedged with respect to a long list agent-based models simple.
of risk factors). During the 23-year period that is
shown, the average strength of the signal slowly Walrasian Price Formation. Most academic
declines. This result both illustrates and rebuts arbi- treatments of price formation are based on the Wal-

64 www.aimr.org 2001, AIMR

 Toward Agent-Based Models for Investment

Figure 1. Strength of Two Proprietary Predictive Signals, 197598

A. Signal 1
Correlation (%)

15

10

0
1/14/75 6/17/81 9/18/87 1/19/94 10/29/98

B. Signal 2
Correlation (%)
10

0
1/14/75 6/17/81 9/18/87 1/19/94 10/29/98

rasian mechanism of price formation, which was equilibrium under appropriate circumstances. Rea-
inspired by the 19th century Paris Bourse. Price set- sonable demand properties should exist: Buying
ting is done by an auctioneer. The auctioneer does not should drive up the price, and selling should drive it
trade but rather polls those who want to trade for the down. The inventory of all the agents should remain
quantity they are willing to buy or sell at any given bounded for reasonable agent behaviors. No agent
price. Transactions take place only when the auction- should have to sacrifice utility. That is, markets are
eer has found a price at which supply equals not a philanthropic business. An agent, at least a
demandthat is, when the market clears. speculator, is not expected to stay in the market if he
This trading mechanism is still used for the Lon- is losing money. The price-formation rule should be
don gold fix, where market participants trade twice a as simple as possible and, within reason, should
day, and the London Metals Exchange, where they match the statistical properties of the data.
trade four times a day, but most markets no longer
operate this way. The Walrasian price-setting mecha- Market-Making Model. The following model is
nism has become unpopular for many reasons. The simple. It is designed to explore a few basic qualita-
most important reasons are that price setting is a tive properties of markets and prices. In particular, in
lengthy process and liquidity takes time to accumu- terms of the effects on price formation, what is the
late. As a result, transactions are made only at discrete, difference between the classic Walrasian call market
widely spaced intervals. But there are other reasons (with an auctioneer and market clearing) and contin-
as well: Agents do not like to reveal their demand uous markets, with competitive market makers? We
functions because they are worried that others will argue that market making necessarily induces tem-
use this information to pick them off. Indeed, most poral structure in prices. The particular market-mak-
agents do not even know their full demand functions ing rule we study causes trending. These side effects
and do not want to take the trouble to formulate make it possible to sustain a population of technical
themthey just know how much they want to buy or traders. The presence of the technical traders
sell and what price they are willing to pay. Another increases arbitrage efficiency, but it also causes prices
problem is that a Walrasian call market is a difficult to make larger deviations from fundamental val-
environment for liquidity providers and thus may ues. This model is much too simple to be a good
actually result in more-volatile prices. predictor for real prices or to provide a practical
model for practitioners. But it provides an example
Price-Formation Rule. For simplicitys sake, in of how simple agent-based models can be built to
the agent-based model described next, we have set understand theoretical problems.
some requirements on the price-formation rule. Given This model is based on an order-driven market.
observed agent behaviors, prices should converge to Traders place ordersin this case, market orders

2001, AIMR www.aimr.org 65

Developments in Quantitative Investment Models

and the auctioneer is replaced by a market maker. at what they think is the correct valuejust take it as
Unlike the auctioneer, the market maker provides given externally. For convenience, the assets per-
liquidity by buying and selling. The agents do not ceived fundamental value will change over time
know at what price the orders will be filled, so the according to a logarithmic random walk following
transactions automatically take place out of equilib- the equation v(t + 1) = v(t) + (t), where v(t) is the
rium. Using an order-driven market solves the prob- logarithm of the value. Thus, we just assume that
lems of a Walrasian market, but other problems pop there is some positive number that the value investor
up in their place. Prices can differ from Walrasian perceives as being associated with the asset, which
prices, which for various sound theoretical reasons changes randomly with time. The value investors
should be good prices. As we will see, this difference position at any given time is given by the function
can mean that prices automatically contain trends.
For the sake of simplicity, we begin with a single xv(t + 1) = c[v(t) p(t)].
asset and assume a single representative market Figure 2 plots price versus time for a scenario in
maker. We allow market orders only and synchro- which the only two agents present in the market are
nize actions so that trades take place at t, t + 1, and so the value investor and the market maker. The price
on. We use the following price-formation rule: roughly tracks the value, oscillating around it but not
tracking it perfectly. Figure 2 shows that the model
1
p ( t + 1 ) p ( t ) = ---

i
N

=1
i
(t ) X ( t ) , (1) satisfies the criterion that the prices converge to equi-
librium under normal conditions. If the value of is
increased, the price will track the value even more
where closely, and if is lowered, it will track less closely.
i = market order of agent i If is lowered all the way to zero, it will not track at
p = log price all. That is, when = 0, the price and the value both
i(t + 1) = xi(t + 1) xi(t) just make random walks, but they do so without
= liquidity maintaining any particular relation to each other. We
X = market-maker position have already learned something: Market-maker risk
= market-maker risk aversion aversion is important for keeping the price close to
xi = position of agent i its equilibrium value. (If we had used the Walrasian
The change in the logarithm of the price is pro- price mechanism, it would always be the case that
portional to the sum of the net order imbalance. The price = value, so the value is the equilibrium price.)
first term is the sum of the orders that are placed by The same relationship is also true of bounded inven-
the agents at time t, and the second term is the order tories: When > 0, inventories are bounded, but
placed by the market maker, which is always a frac- when = 0, they become unbounded. This pattern is
tion of the market makers current position. The not surprising, because unbounded market-maker
variable X, the market makers total position, is the inventory means unbounded risk and the parame-
total amount that supply and demand are out of ter is there to take the market makers risk aversion
balance in the market. The constant of proportional- into account.
ity, 1/ , can be thought of as liquidity, and it deter- One question that any trader will immediately
mines the amount that an order of a given size will ask is: Who is making profits? To address this ques-
move the price (a net order of size will change the tion, we sweep the parameters and c. (Remember,
price by a factor of e). Any particular family of agent the parameter c sets the scale of the value investors
behaviors can be studied by specifying a set of func- trading.) For high values of and small values of c,
tions xi(t) and iterating the resulting equations. the value investor makes a net profit on average, and
Behavioral Models. Now, I will introduce a for the opposite parameters, the market maker makes
simple behavioral model involving four classes of a net profit on average. The boundary between the
investors: Market makers, which I have already intro- two regions is shown in Figure 3.
duced, fundamental (or value) investors, technical We can probably safely assume that neither
traders (or chartists), and liquidity demanders (or player would be happy letting the other make profits
noise traders). at his expense. Thus, if the market maker is making
Value investors simply take a position based on net profits, the value investor will adjust c to increase
the perceived fundamental value of an asset. The his profits (and thereby decrease the market makers
more underpriced the asset, the larger the position profits), and similarly, the market maker can adjust
that value investors take in the asset. For the purposes . This struggle should always result in a solution
of this model, I do not care how the investors arrive that is along the boundary curve shown in Figure 3.

66 www.aimr.org 2001, AIMR

 Toward Agent-Based Models for Investment

Figure 2. Price and Value for a Risk-Neutral Market Maker

Price/Value
1.2

1.1

1.0

0.9

0.8
Price

0.7

0.6
Value
0.5

0.4

0.3
0 500 1,000 1,500 2,000
Time

But this is not the whole story. If you look closely value investors). When > 1, however, it starts out
at Figure 2, you will notice that even though changes flat and then rolls off. This observation corresponds
in value are completely random, changes in the price to the effect of risk-averse market making, which
are not. This occurs because the market makers makes prices trend. As increases, this roll-off
actions introduce trends into the prices. Once the
market maker acquires a position, because of her risk
aversion, she has to get rid of it. By selling a fraction Figure 3. Profitability Constraints
at each time step, she will unload the position a bit
at a time. This behavior causes a trend in prices. Any
risk-averse behavior on the part of the market maker 1.0
will result in a temporal structure of some sort in
prices.
0.8
To provide a better look at the temporal structure
in this case, Figure 4 shows what is called the power Value Investor Profits
spectrum of the returns. The power spectrum is the 0.6
square of the Fourier transform and can be thought
of as the amount of power, or variance, present at
each frequency. If the returns were completely ran- 0.4
Market-Maker
dom the power spectrum would be white, which Profits
would mean all frequencies were equally well repre-
0.2
sented (i.e., the power spectrum would be constant).
Instead, the power spectrum decreases at high fre-
0
quencies. When = 0, the power spectrum is pretty 0 0.2 0.4 0.6 0.8 1.0
flat throughout most of its range (with a small rise
coming from the mean-reverting behavior of the

2001, AIMR www.aimr.org 67

Developments in Quantitative Investment Models

Figure 4. Power Spectrum of Price Returns By playing with the parameters in the previous
example, we showed that a point exists at which
Power
everything is efficient. What remains to be shown is
1.00 whether the market will go there on its own. To find
the answer to that question, we must first introduce
some new elements into the modelnamely, rein-
vestment and credit limits. Both of these elements
have their own side effects.

0.10 Credit Limits. Real traders have credit limits

(i.e., there is a wealth-dependent limit to the size of
the positions they can take). If a trader tries to take a
position that is too big with respect to her credit limit
and wealth, her position will be capped. If the agents
wealth decreases, she may have to sell some assets to
stay within the limit. This limit introduces a nonlin-
0.01
earity into the agents strategies that has several
0.001 0.010 0.100 effects on the price dynamics of the market. It can
Frequency cause prices to deviate even more from fundamental
= 0.0 = 0.5 values and, in extreme cases, can even make prices
= 0.2 = 0.8 oscillate on their own.
Another key element is reinvestment. As traders
make profits, they change the scale of their trading
accordingly. This behavior means that the c parame-
becomes sharper, because the power spectrum is
ters are not fixed; rather, they are functions that
nearly white at low frequencies and the timescale of
depend on wealth and credit limits.
the trending is shorter. When we put all of these elements together, we
find that the market does indeed tend toward the
Trend Followers. The fact that the market
point where everything is efficient. This point is
maker causes a temporal structure in prices creates
somewhat noisy; there are fluctuations around it, but
an opportunity for technical traders. For example, each agents wealth stays more or less fixed. At this
assume this technical trader is a trend follower. A point, there is equilibrium among the market maker,
trend follower holds positive positions when the value investor, and trend follower.
price trend is up and negative positions when the
trend is down. For example, consider an exponential Dynamic Market Efficiency. The following
moving average: example shows how the progression toward effi-
ciency can be much more complicated and can
x(t) = cs(t) increase rather than decrease volatility. We will now
s(t + 1) = ar(t) + (1 a)s(t), (2) study an entirely different set of agent behaviors.
There are two groups. One group consists of seasonal
where traders that alternately buy and sell. Imagine that
x(t) = position they are doing this because they need to (for example,
s(t) = signal because they are farmers whose crops are harvested
c = scale at a particular time of year). They are willing to trade
r(t) = price return to reduce risk or to gain liquidity and do so even if
a = inverse timescale they lose money on average. The other group consists
This trend-following rule is not necessarily the best of technical traders. The technical traders are trying
to exploit the first group. Unlike the previous techni-
possible trading rule to use, but it is sufficient for the
cal traders, however, we allow many different tech-
trader to extract profits, at least for a reasonable
nical trading strategies, which we will make up at
choice of the parameter a, which sets the time scale random, and let the market select the most successful
for trend following, and c, which sets the trading size. ones.
In fact, for a reasonable choice of a, by playing with c To keep things simple, assume the technical trad-
for the value investor and c for the trend follower, we ers base their strategies on only the signs of the most
can find parameter values at which everyone just recent price changes. To construct a particular strat-
breaks even. The market is efficient. egy, list each possible set of signs for past price

68 www.aimr.org 2001, AIMR

 Toward Agent-Based Models for Investment

changes, which will be called a pattern, and ran- sonal traders have almost all the wealth, and the price
domly assign either +1 or 1 as the output corre- simply alternates up and down as they trade. After
sponding to each pattern. The output will be some time passes, the technical traders who have
interpreted as being proportional to the position of been successful start to make money, and as they
the strategy. For example, suppose that the pattern is make money, they rise in importance. They begin
(+,,+) and the corresponding output is +1. That set placing bigger orders because their wealth is growing
of conditions means that if the last three price as a result of their trading profits. As this response
changes were first up, then down, then up, the strat- happens, the oscillating pattern slowly dampens
egy would give a buy signal. With a buy signal, the because the price impact of their trades cancels the
agent will invest a fixed fraction of his wealth in the impact of the seasonal traders.
asset (and similarly on the short side for a sell signal). When time equals 5,000 (the bottleneck seen in
Now, suppose we choose a bunch of traders and Figure 5), a point is reached at which the market seems
randomly assign each of them two strategies (where to be highly efficient. The original oscillating pattern
each strategy is an entire list of responses to every of prices is now virtually gone, cancelled out by the
possible pattern). We give all of them a little money action of the technical traders. Just as they are about
to get started and turn them loose. to award Nobel Prizes for the theory of market effi-
Figure 5 shows a simulation of this situation, ciency, however, something odd happens: Prices sud-
illustrating the long-term evolution toward market denly go crazy and start oscillating again. This effect
efficiency. The beginning of the simulation is on the is caused by the technical traders, who have gained
left side (at time 0). Panel A shows a snapshot of the enough wealth to exert a greater influence on prices
price every 101 time steps. Panel B shows the total than the seasonal traders. Whereas the sequence of
wealth of all the traders. At the beginning, the sea- signs had previously been always alternating

Figure 5. Long-Term Evolution Toward Market Efficiency

24
A. Price
Price (100)
23

22
24

21
23

20
22

19
21
0 5,000 10,000 15,000
20 Time

19
B. Total Capital
0 Capital
Total 5,000 10,000 15,000
Time
20
15 B. Total Capital
Total Capital
10
205
150
10 0 5,000 10,000 15,000
Time
5
0
0 5,000 10,000 15,000
Time

2001, AIMR www.aimr.org 69

Developments in Quantitative Investment Models

between plus and minus, there are now sometimes not replicate all of them accurately. These models are
two plus signs or two minus signs in a row. This simple, but we have illustrated certain basic points.
pattern invokes a set of previously unused patterns in We see how the market mechanism matters: Using an
the strategies of each agent. Prices start changing in a order-based market with market makers as liquidity
chaotic, volatile manner again. No longer is the pat- providers can result in patterns in prices that sustain
tern a simple periodic oscillation; rather, the pattern trend followers. When we introduce a large number
is much more random and is driven by a complex set of different strategies and let the market select them,
of interactions among all the agents. The cash flows the path toward efficiency can be surprising because
shift again. Although the market might be getting of the dynamics of the strategies interacting with each
more efficient, it is certainly not getting less volatile. other.
Strategies that had been unprofitable when the sea- Agent-based modeling of markets is still in its
sonal players were dominating may now become infancy. I predict that as the models get better, they
profitable and begin to accumulate wealth. will begin to be useful for real problems. The work I
This scenario happens in the absence of noisy have presented here is still some distance from that
inputs. The simulation is completely deterministic. point. If I needed an investment strategy now, I
The statistical properties of prices continue to change, would certainly use the empirical approach. But to
even tens of thousands of iterations later, as the feed- really understand how markets work, we have to use
ing relationships of who is exploiting whom shift agent-based models. Within five years, people may
around. There is a rich and slowly evolving ecology
be trading money with agent-based models. Time
of agents, with shifting interactions. Market effi-
will tell. The degree to which fortune-tellers affect the
ciency takes a long time to happen.
future depends on their credibility. Thus, as agent-
based models achieve some successes, market partic-
Future Research ipants will start to pay attention, and prices will
As I said in the introduction, at this stage, none of change in response. Agent-based models are a brand
these models is of practical use yet. Although we can new tool, and the market for them is still far from
replicate some of the properties of real prices, we do efficient.

70 www.aimr.org 2001, AIMR