Lecture Notes in Economics

and Mathematical Systems 602

Founding Editors:
M. Beckmann
H.P. Künzi

Managing Editors:
Prof. Dr. G. Fandel
Fachbereich Wirtschaftswissenschaften
Fernuniversität Hagen
Feithstr. 140/AVZ II, 58084 Hagen, Germany
Prof. Dr. W. Trockel
Institut für Mathematische Wirtschaftsforschung (IMW)
Universität Bielefeld
Universitätsstr. 25, 33615 Bielefeld, Germany

Editorial Board:
A. Basile, A. Drexl, H. Dawid, K. Inderfurth, W. Kürsten
Norman Ehrentreich

Agent-Based
Modeling
The Santa Fe Institute
Artificial Stock Market
Model Revisited

123
Dr. Norman Ehrentreich
RiverSource Investments, LLC
262 Ameriprise Financial Center
Minneapolis, MN 55474
USA
[email protected]

ISBN 978-3-540-73878-7 e-ISBN 978-3-540-73879-4

DOI 10.1007/978-3-540-73878-7

Lecture Notes in Economics and Mathematical Systems ISSN 0075-8442

Library of Congress Control Number: 2007937522

c 2008 Springer-Verlag Berlin Heidelberg



This work is subject to copyright. All rights are reserved, whether the whole or part of the material
is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication
of this publication or parts thereof is permitted only under the provisions of the German Copyright
Law of September 9, 1965, in its current version, and permission for use must always be obtained
from Springer. Violations are liable to prosecution under the German Copyright Law.

The use of general descriptive names, registered names, trademarks, etc. in this publication does
not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.

Production: LE-TEX Jelonek, Schmidt & Vöckler GbR, Leipzig

Cover-design: WMX Design GmbH, Heidelberg

Printed on acid-free paper

987654321

springer.com
Meinen Eltern gewidmet.
Foreword

When the original Santa Fe Institute (SFI) artificial stock market was
created in the early 1990’s, the creators realized that it contained many
interesting new technologies that had never been tested in economic
modeling. The authors kept to a very specific finance message in their
papers, but the hope was that others would pick up where these papers
left off and put these important issues to the test. Tackling the com-
plexities involved in implementation has held many people back from
this, and many parts of the SFI market remain unexplored. Ehren-
treich’s book is an important and careful study of some of the issues
involved in the workings of the SFI stock market.
As Ehrentreich’s book points out in its historical perspective, the
SFI market was intended as a computational test bed for a market
with boundedly rational learning agents replacing the standard setup
of perfectly rational equilibrium modeling common in economics and fi-
nance. These agents exhibit reasonable, purposeful behavior, but they
are not able to completely process every aspect of the world around
them. This can be viewed much more as a function of the complex-
ity of the world, rather than the computational limitations of agents.
In a financial world out of equilibrium, optimal behavior would re-
quire knowledge of strategies being used by all the other agents, an
information and computational task which seems well out of reach of
any trader. The SFI market’s main conclusion was that markets where
agents were learning might not converge to traditional simple rational
expectations equilibria. They go to some other steady state in which
a rich set of trading strategies survives in the trader population. In
this steady state the market demonstrates empirical signatures that
are present in most financial time series.
VIII Foreword

This book is an excellent reference to both the learning, and em-

pirical literature in finance. It stresses the difficult empirical facts that
are out of reach of most traditional financial models including per-
sistent volatility and trading volume, and technical trading behavior.
However, Ehrentreich’s main mission is to dig deeply into the SFI mar-
ket structure to understand what is actually going on. Computational
economic models can often be explored at three levels. There is sort
of a big picture level where concepts such as rational expectations and
bounded rationality are explored. There is also the very low level where
researchers discuss the nuts and bolts of different modeling languages.
In between these sits a region where many of the computational learn-
ing technologies are implemented. This is where technologies such as
genetic algorithms, classifier systems, and neural networks drive much
of what is going on. This is Ehrentreich’s area of exploration, and it is
critically important to agent-based modelers since one needs to know
the sensitivity of the higher level results to changes in the learning
structures used beneath them.
The SFI market uses two learning mechanisms extensively: the ge-
netic algorithm, GA, and classifier system. Both of these are devel-
opments of John Holland, one of the SFI market coauthors. The GA
is a type of general evolutionary learning mechanism, and it is used in
both computer science and economics. Its properties have been studied,
but it is still not completely understood. In computer science it is often
studied in difficult optimization problems. These are problems with well
defined objectives, and are quite different from the more open ended co-
evolutionary problems in economics where agents are competing with
each other. The classifier system is an interesting learning structure
that allows agents to dynamically find relevant states in the world
around them. For example, actions might be conditioned on whether
a stock is currently priced above a certain multiple of dividends. The
classifier has the power to endogenously slice up a stream of empirical
information into states of the world. Very few learning mechanisms are
able to do this. With this generality comes a lot of model complexity,
and many implementations of the classifier seem computationally un-
wieldy. They also involve many implementation questions that need to
be explored.
In several chapters Ehrentreich explores some of the more impor-
tant aspects of the SFI classifier implementations. He shows that the
SFI classifier is sensitive to certain design characteristics. Under dif-
ferent assumptions about evolution the classifier system behaves very
differently from the original SFI model. Ehrentreich carefully modifies
 Foreword IX

and explores his own operation on mutating trading strategies. Using

this modified mutation causes a situation in which the SFI market is
much more likely to converge to the rational expectations equilibrium,
and the rich technical trading dynamic does not emerge. The results in
the original SFI market are clearly sensitive to how mutation is imple-
mented. The book goes on to do a comparative study between mutation
operators. A key issue is how many technical trading related rules are
evolved, and whether the system is likely to generate lots of technical
rules by chance in the evolutionary process. The modified mutation op-
erator does not generate many of these rules, so they never really get a
foot in the door of trading activity. The SFI structure facilitates their
formation, but it is possible this could be driven more by genetic drift
than selection. The original SFI studies never really answered these
questions, and it only looked at trading strategy formation in an indi-
rect level by looking at aggregate numbers. This was a clear weakness.
Ehrentreich does some careful checks to see if technical rules are adding
value at the agent level. It appears that they are, so many of the SFI
indirect conclusions are sound.
The dynamics of wealth was never part of the original SFI market.
It is an interesting omission that the SFI market never really considered
long term wealth in a serious way in its implementation. This is strange
since many arguments about efficient markets thrive on the relative
dynamics of trader wealth. Ehrentreich concludes that this is a complex
problem, and there may be difficulties with some of the other studies
that try to tag a wealth dynamic onto the SFI market. In my opinion
this is one of the biggest limitations of the actual SFI market structure.
This book is an important piece of work for understanding the dy-
namics of models with interacting learning agents. I think researchers
in the future will find it critical in helping them to understand the
dynamics of evolutionary learning models. Most importantly, it sets
an important standard for doing careful internal experiments on these
markets and the learning mechanisms inside them.

Brandeis University, Waltham, MA

September 2007 Blake LeBaron
Preface

The road of science is filled with surprises. When embarking on a sci-

entific journey, we probably have a specific destination in mind, but
we never know whether the road will take us there nor what places we
may encounter along the way.
This trip was no exception. Before anyone starts reading this trav-
elogue, I think that I should briefly mention a few places that I visited,
but decided to pass over while writing this book. I originally aimed at
converting the well-known Santa Fe Institute Artificial Stock Market
(SFI-ASM) into a two stock version to study portfolio decisions of in-
dividual investors. My early forays into this unknown territory yielded
some interim results, but until now they are still waiting to be further
examined.
Instead, my road took a sudden and unexpected turn. One of the
most important findings of the original SFI-ASM was the emergence
of technical trading for faster learning speeds. Yet a thorough analysis
of the agent’s learning algorithm suggested that this might have been
caused by an ill-designed mutation operator. For a couple of years,
many tests confirmed this supposition. For instance, even though tech-
nical trading rules emerged in the original SFI-ASM, they were rarely
acted upon. Most importantly, though, was that agents with an alter-
native mutation operator discovered the homogeneous rational expec-
tations equilibrium, a result that found immediate approval by neoclas-
sically inclined economists.
I traveled a long way down this road. Since I considered the ex-
istence of technical trading to be an empirical fact of financial mar-
kets, I tried to unearth the necessary ingredients to reintroduce it into
my model. Nothing that I devised, neither social learning nor explicit
herding mechanisms, succeeded in that endeavor. There was, however,
XII Preface

another surprise waiting behind the supposedly final turn of my jour-

ney. One newly designed test showed a slight superiority of technical
trading rules in the original model. A side-trip all the way down to
population genetics finally proved that my agents were committing a
mistake by deciding to ignore technical trading rules. Again, parts of
my prior research were discarded, and a new chapter was written ex-
plaining why I and previous researchers went wrong in interpreting the
simulation results. I hope that this chapter will prove most useful for
any research involving genetic algorithms. My prior belief that technical
trading was an artificially introduced model artifact had also caused me
to visit some previous studies about wealth levels. I was able to show
that the SFI-ASM was not designed to address any questions related
to wealth. Fortunately, this part was unaffected by the breakdown of
the initial motivation to look into the wealth generation process.
A long journey with such detours was certainly not easy. I could not
have arrived at the final destination without the tremendous support
and encouragement that I have found along the way. Above all, I wish to
thank my parents Werner and Ellinor Ehrentreich, for without them,
I would not have had the opportunity to embark on this journey. I
would also like to thank Reinhart Schmidt for letting me choose my
destination and for giving me the freedom to follow my own path.
Among the numerous friends, colleagues, and conference participants
who have contributed in many ways are Manfred Jäger, Ulrike Neyer,
Ralf Peters, Martin Klein, Heinz-Peter Galler, Joseph Felsenstein, Alan
Kirman, and James Stodder. Of course, this book would not have been
finished without the contributions by Blake LeBaron. Not only did
he play a major role in the creation of the model that I set out to
extend, then critiqued, and finally confirmed, he also often helped and
clarified many questions that I was pondering. Many thanks also go
to Lars Schiefner, Doris Storch, and Klaus Renger, especially for their
help during the final stages of this project. Last, but not least, I thank
Tanya Novak for her patience and help, especially for her proofreading.
Nonetheless, I absolve her from all remaining mistakes and typos and
credit them to my cats, Zina and Francesco, who stubbornly insisted
on their input by jumping on the keyboard.
I now hope that the reader will find it useful to visit the places that
I have found worthwhile to mention in this book.

Minneapolis, MN
September 2007 Norman Ehrentreich
Contents

Part I Agent-Based Modeling in Economics

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 The Rationale for Agent-Based Modeling . . . . . . . . . . . . 5

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 The Representative Agent Modeling Approach . . . . . . . . . 7
2.2.1 Avoiding the Lucas-Critique . . . . . . . . . . . . . . . . . . . 8
2.2.2 Building Walrasian General Equilibrium Models . . 9
2.2.3 Representative Agents and the Fallacy of
Composition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.4 Expectation Formation in Markets with
Heterogeneous Investors . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Rational Expectations and Disequilibrium Dynamics . . . . 13
2.4 The Economy as an Evolving Complex Adaptive System 14
2.5 Some Methodological Aspects of Agent-Based Simulations 16

3 The Concept of Minimal Rationality . . . . . . . . . . . . . . . . . 19

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 Economic, Bounded, and Situational Rationality . . . . . . . 21
3.3 Situational Analysis, Minimal Rationality, and the
Prime Directive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.4 Minimal Rationality and the Phillips-Curve . . . . . . . . . . . 26

4 Learning in Economics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Definitions of Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.3 Rationality-Based Learning Models . . . . . . . . . . . . . . . . . . . 31
4.4 Biologically Inspired Learning Models . . . . . . . . . . . . . . . . . 32
XIV Contents

4.4.1 Learning Through Replicator Dynamics . . . . . . . . . 34

4.4.2 Learning Through Genetic Algorithms . . . . . . . . . . 36
4.4.3 Learning Through Classifier Systems . . . . . . . . . . . . 46

5 Replicating the Stylized Facts of Financial Markets . . 51

5.1 Efficient Markets and the Efficient Market Hypothesis . . 51
5.1.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.1.2 Random Walks or Martingales? . . . . . . . . . . . . . . . . 53
5.1.3 Tests for Market Efficiency . . . . . . . . . . . . . . . . . . . . . 55
5.2 Stylized Facts of Financial Markets . . . . . . . . . . . . . . . . . . . 56
5.2.1 Non-Normal Return Distributions . . . . . . . . . . . . . . 56
5.2.2 Volatility Clustering of Returns . . . . . . . . . . . . . . . . 60
5.2.3 High and Persistent Trading Volume . . . . . . . . . . . . 64
5.2.4 Existence of Technical Trading . . . . . . . . . . . . . . . . . 65
5.3 Alternative Market Hypotheses . . . . . . . . . . . . . . . . . . . . . . 70
5.3.1 The Fractal Market Hypothesis . . . . . . . . . . . . . . . . . 70
5.3.2 The Coherent Market Hypothesis . . . . . . . . . . . . . . . 71
5.3.3 The Adaptive Market Hypothesis . . . . . . . . . . . . . . . 72
5.3.4 The Interacting-Agent Hypothesis . . . . . . . . . . . . . . 75
5.4 Agent-Based Computational Models of Financial Markets 75
5.4.1 Allocative Efficiency with Zero-Intelligence Traders 76
5.4.2 Models with a Random Communication Structure 79
5.4.3 Models of Chartist-Fundamentalist Interactions . . . 83
5.4.4 Many-Strategy Models with Learning . . . . . . . . . . . 85

Part II The Santa Fe Institute Artificial Stock Market Model

Revisited

6 The Original Santa Fe Institute Artificial Stock

Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.2 The Marimon-Sargent Hypothesis and the SFI-ASM . . . . 92
6.3 An Overview of SFI-ASM Versions . . . . . . . . . . . . . . . . . . . 93
6.4 The Basic Structure of the SFI-ASM . . . . . . . . . . . . . . . . . 94
6.4.1 Trading Rules and Expectation Formation . . . . . . . 95
6.4.2 Learning and Rule Evolution . . . . . . . . . . . . . . . . . . . 99
6.4.3 Other Programming Details and Initialization of
Model Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.5 The Homogeneous Rational Expectations Equilibrium . . 102
6.6 The Marimon-Sargent Hypothesis Refined . . . . . . . . . . . . . 103
6.7 Simulation Results of the SFI-ASM . . . . . . . . . . . . . . . . . . . 104
 Contents XV

6.7.1 Time Series Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . 104

6.7.2 Forecast Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.8 A Potential Problem: A Biased Mutation Operator . . . . . 108

7 A Suggested Modification to the SFI-ASM . . . . . . . . . . . 113

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.2 An Unbiased Mutation Operator . . . . . . . . . . . . . . . . . . . . . 114
7.3 Simulation Results with the Modified SFI-ASM . . . . . . . . 115
7.3.1 Trading Bit Behavior . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.3.2 Time Series Properties . . . . . . . . . . . . . . . . . . . . . . . . 118
7.4 Robustness of the Zero-Bit Solution . . . . . . . . . . . . . . . . . . 121
7.4.1 Stochastic versus Periodic Dividends and the
Classifier System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.4.2 Dependence on Other Parameter Values . . . . . . . . . 122
7.4.3 Generalization or Consistent Trading Rules? . . . . . 123

8 An Analysis of Wealth Levels . . . . . . . . . . . . . . . . . . . . . . . . 127

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
8.2 Wealth Levels in the SFI-ASM: An Economic(al)
Explanation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
8.3 Previous Studies Based on Wealth Levels
in the SFI-ASM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
8.3.1 Financial Markets Can Be at Sub-Optimal
Equilibria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
8.3.2 Technical Trading as a Prisoner’s Dilemma . . . . . . . 131
8.4 Wealth Levels in the SFI-ASM: Alternative Explanations 133
8.4.1 Risk-Premium, Taxation, and Two Benchmark
Wealth Levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
8.4.2 Average Stock Holdings and Wealth Levels . . . . . . 135
8.4.3 Activated Rules and Rule Selection . . . . . . . . . . . . . 139
8.5 A Verdict on Wealth Analysis in the SFI-ASM . . . . . . . . . 145

9 Selection, Genetic Drift, and Technical Trading . . . . . 147

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
9.2 Technical Trading and the Aggregate Bit Level . . . . . . . . . 148
9.3 The Zero-Bit Solution: Some Disturbing Evidence . . . . . . 150
9.4 Random Genetic Drift in Genetic Algorithms . . . . . . . . . . 152
9.5 The Neutralist–Selectionist Controversy . . . . . . . . . . . . . . . 154
9.6 Fitness Driven Selection or Genetic Drift? . . . . . . . . . . . . . 157
9.6.1 Selection or Genetic Drift in the Modified
SFI-ASM? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
XVI Contents

9.6.2 Selection or Genetic Drift in the Original

SFI-ASM? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
9.6.3 Genetic Drift, Fitness Gradient, and Population
Size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
9.7 The Effect of Mutation on Genetic Drift . . . . . . . . . . . . . . 162
9.7.1 Genetic Drift, Mutation, and Crossover Only in
SFI Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
9.7.2 Genetic Drift, Mutation, and Crossover Only in
Bit-Neutral Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
9.7.3 An Equilibrium Analysis of Genetic Drift and
Mutation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
9.7.4 A Final Assessment of the Two Mutation
Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
9.8 Detection of Emergence of Technical Trading . . . . . . . . . . 172
9.8.1 Predictability in the Price Series . . . . . . . . . . . . . . . . 172
9.8.2 Trading Bits and Fitness Values . . . . . . . . . . . . . . . . 173
9.8.3 Equilibrium Bit Frequencies and Bit Fixations . . . 176
9.9 An Evolutionary Perspective on Technical Trading . . . . . 177

10 Summary and Future Research . . . . . . . . . . . . . . . . . . . . . . 181

Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
11.1 Timing in the Stock Market . . . . . . . . . . . . . . . . . . . . . . . . . 187
11.2 Fundamental and Technical Trading Bits . . . . . . . . . . . . . . 189

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
 Part I

Agent-Based Modeling in Economics

1
Introduction

What I cannot create, I do not understand.

Richard P. Feynman

In addition to deduction and induction, simulation is sometimes seen

as a third methodology for doing research. Even though simulation does
not prove theorems, it can enhance our understanding of complex phe-
nomena that have been out of reach for deductive theory. Tesfatsion de-
fines agent-based simulations as the computational study of economies
that are modeled as evolving systems of autonomous interacting agents
[421]. In the last decade, they have become a widely accepted tool for
studying decentralized markets.
A major advantage is that agent-based models allow the removal
of many restrictive assumptions that are required by analytical models
for tractability. For instance, all investors could be modeled as het-
erogeneous with respect to their preferences, endowments, and trading
strategies.
Among the numerous agent-based simulations of financial markets
[247, 268, 81], the Santa Fe Institute Artificial Stock Market (SFI-ASM)
is one of the pioneering models and thus, probably the most well-known
and best studied. It was created by a group of economists and com-
puter scientists at the Santa Fe Institute in New Mexico to test whether
artificially intelligent agents would converge to the homogeneous ratio-
nal expectations equilibrium or not. The original SFI-ASM has been
described in a series of papers [331, 11, 330, 244].
In agent-based simulations, replication of existing models from
scratch is an important, but often neglected step. Axelrod empha-
sizes that without this outside confirmation, possible erroneous results
4 1 Introduction

based on programming errors or misinterpretation of simulation data

can go undetected [13]. Skepticism towards computational models is
sometimes voiced since their results are seen as counterintuitive and
incomprehensible because the computer program remains an impene-
trable “black box”. Judd considers the black box criticism to be more a
result of poor exposition of computational work and a lack of sufficient
sensitivity analyses. Since third party replications need to open that
black box widely, that criticism can effectively be addressed [208].
This book starts by presenting the reasons that led to the adoption
of the agent-based programming approach in economics. Since one of
these reasons, the desire to replace the representative agent approach
with heterogeneous agents, leads to a break down of rational expec-
tation formation, chapter 3 contrasts several concepts of agent ratio-
nality. When modeling less than fully rational agents, researchers have
to equip their agents with learning algorithms which are discussed in
chapter 4.
The replication of selected stylized facts of financial markets through
agent-based simulations of financial markets is the focus of chapter 5. It
starts by introducing the Efficient Market Hypothesis (EMH). Much of
our empirical knowledge about financial markets, often summarized as
stylized facts, stems from attempts to either prove or disprove the EMH.
Several competing market hypotheses that strive to better explain the
stylized facts than the EMH are subsequently discussed. In the final
part of chapter 5, a selection of agent-based models of financial markets
is briefly introduced.
The main part of this book analyzes a particular Java-replication
of the SFI-ASM model which was originally programmed in Objective-
C. The goal is to assess whether the SFI-ASM’s result of emergent
technical trading is robust to changes in the model design. To this
end, the simulation results are supplemented by theoretical analyses
of certain model features. The replication results of the reprogrammed
Java-version are presented in chapter 6. A Markov chain analysis of
the original mutation operator delivers the motivation to develop an
alternative mutation operator in chapter 7. Because of the differences
in simulation results, chapter 8 needs to reexamine the model structure
with respect to wealth accumulation. The final chapter analyzes and
compares the two mutation operators in more detail. By interpreting
the insights gained from the theoretical analysis of the two mutation
operators, chapter 9 concludes by offering an explanation as to why
the validity of the EMH is entirely consistent with the simultaneous
existence of technical trading rules.
2
The Rationale for Agent-Based Modeling

Imagine how hard physics would be if electrons could think.

Murray Gell-Man1

2.1 Introduction

The ascent of powerful and affordable microcomputers and the avail-

ability of huge economic data sets have sparked the development of a
rather recent branch in economic research. The rapidly growing field
of computational economics is a broad concept and encompasses many
different areas. An eclectic and not exhausting overview of all these
different activities2 can be found in the “Handbook of Computational
Economics” by Amman et al. [8]. Since a precise definition of compu-
tational economics is not being offered by its editors, Riechmann goes
as far as to suggest that every economist who uses a computer for more
than mere typewriting engages in computational economics [356].
But if we emphasize the aspect of computability of economic prob-
lems, i.e., problems that allow for numerical results, the roots of com-
putational economics could easily be dated back before computers were
actually used by economists.3 Several contributions in this handbook
1
Attributed to Gell-Mann, 1969 Nobel laureate in physics and co-founder of the
Santa Fe Institute, cited by Page [329].
2
Among those are, for instance, the numerical computation of Nash equilibria, de-
terministic and stochastic simulations, or numerical dynamic programming prob-
lems.
3
Nagurney [316], as well as Kendrick [217], mention the early contributions to
computational economics in the 1950ies by Koopmans, Samuelson, or Solow. In
finance, the work by Markowitz about efficient portfolio diversification needs to
be mentioned at this point [288, 289].
6 2 The Rationale for Agent-Based Modeling

focus on algorithms and numerical methods for finding Nash-equilibria

or the solutions to dynamic nonlinear systems of equations, yet some of
them were developed even before computers were. The access to com-
puter technology just reduced the sometimes prohibitive computational
costs of these algorithms and allowed them to be used for practical pur-
poses.
Within computational economics, the field of agent-based modeling
or simulation (ABM, ABS), sometimes also called microscopic simula-
tion (MS, Levy et al. [248]) or agent-based computational economics
(ACE, Tesfatsion [420]), seems currently to be the most rapidly grow-
ing discipline. This is acknowledged, for instance, by the appearance
of a second volume of the “Handbook of Computational Economics”
which will be solely dedicated to this approach [209].
In the ABS approach, model economies are built from the bottom
up, i.e., they consist of many autonomous and interacting agents. It
had its first breakthrough with the influential models by Schelling [375,
376] about endogenous neighborhood segregation.4 These models are
populated by two types of agents who only care about the composition
of their own small neighborhood. In particular, they do not tolerate
more than a certain fraction of agents of the other type in their vicinity,
however, they do not care about integration or segregation on the city
level. Unsatisfied agents are allowed to move to a neighborhood that
they are happy with. Schelling showed that for a wide range of the
agents’ neighborhood tolerance parameter, an initially integrated city
emerges to an almost completely segregated entity.5 Thus, Schelling
demonstrated how stable macrobehavior may emerge from strictly local
motives on the agents’ level, a macrobehavior that would be hard to
predict by exclusively looking at individual motives. Nowadays, this
phenomenon is known as emergence, a key concept of the theory of
complex adaptive systems.
The agent-based approach is currently considered by many re-
searchers as the latest revolution in economic methodology. However,

4
A comparison of the different versions by Schelling is given by Pancs and Vriend
[332]. The most important distinction is the dimensionality of neighborhoods. In
his 1969 model, a neighborhood is defined only in one dimension, i.e., agents popu-
late a line, while his later models use two dimensional lattices. A two-dimensional
web-based simulation example in NetLogo by Wilensky [436] can be found at
http://ccl.northwestern.edu/netlogo/models/Segregation.
5
Pancs and Vriend [332] extend the framework by asking how individual prefer-
ences may alter the outcome. Surprisingly, the results are very robust to changes
in preferences. Even if all agents strictly prefer perfect integration, neighborhood
segregation will still occur.
 2.2 The Representative Agent Modeling Approach 7

in order to understand and properly evaluate this shift in methodol-

ogy, Bankes points out that one should not focus on the progress in the
computer sciences which made this whole development possible, but on
the inappropriateness of traditional methods which made it necessary
[22].
The deficiencies of analytical modeling can only be sought in its
assumptions. Thus, the first part of this chapter discusses some of
these and contrasts them with the requirements of agent-based mod-
eling. The main culprits from an agent-based point of view are the
widespread assumption of representative agents and that of rational
expectations. First, it is obvious that fictitious representative agents
in macroeconomic models are incapable of generating emergent phe-
nomena. Financial markets would also be characterized by the absence
of any trading activity. Secondly, fully rational and perfectly informed
agents have essentially no ability to exercise free choice. Despite some
suggestive rhetoric, individuals can follow only one, i.e., the rational
course of action [235].
The agent-based modeling approach, on the other hand, requires
neither of these two assumptions. The ability to cope with heteroge-
neous and boundedly rational agents makes it a perfect tool to study
decentralized markets. Instead of a reductionist approach, agent-based
models treat the economy as an evolving complex adaptive system con-
sisting of many heterogeneous and interacting agents. The development
of artificial financial markets has thus become a major application for
the agent-based paradigm.6

2.2 The Representative Agent Modeling Approach

The notion of representative agents appeared in the economic literature

already in the late 19th century. Edgeworth used the term ‘representa-
tive particular ’ [104, p. 109], while Marshall introduced a ‘representative
firm’ in his Principles of Economics [290].7 However, only after Lucas
[261] had published his article about econometric policy evaluation—
the famous Lucas-critique—they became the dominant macroeconomic
6
According to [227], scientific paradigms share two essential characteristics: First,
their achievements must have enough novelty to attract a permanent group of
scientists away from competing modes of scientific activity. Secondly, their open-
endedness must allow for addressing many different kinds of problems. Whether
the agent-based modeling approach already is or might become the next paradigm
in the economic sciences is left to the reader to be answered.
7
A discussion of the origins of representative agents can be found in [172].
8 2 The Rationale for Agent-Based Modeling

approach. Today’s representative agent models are characterized by

an explicitly stated optimization problem of the representative agent,
which can be either a consumer or a producer. The derived individual
demand or supply curves are then, in turn, used for the corresponding
aggregate demand or supply curves.
A series of papers in which the rationale for using representative
agent models is convincingly set forth does not exist according to
[172]. From the “large set of introductions, paragraphs, and parenthet-
ical asides”, however, Hartley identifies several motives for their use.
They were thought to avoid the Lucas critique, to provide rigorous
microfoundations to macroeconomics, and to help build powerful Wal-
rasian general equilibrium models.

2.2.1 Avoiding the Lucas-Critique

Before [261], macroeconomic models were often defined in terms of

three vectors: yt being the set of endogenous variables, xt the vector
of exogenous forcing variables, and ǫt the set of random shocks. Fixed
parameters are subsumed in a vector θ.

yt+1 = F (yt , xt , θ, ǫt ) . (2.1)

Lucas, however, criticized that it is likely that some of the the parame-
ters contained in θ change due to a shift of policy regime λ. Aggregate
quantities and prices might react differently than predicted since agents
may change their behavior in a way which is not captured in the ag-
gregate equations. For instance, agents could adapt their expectations
about future inflation rates or, in the case of rational expectations,
change them even before an anticipated policy shift is implemented.
An attempt to exploit a potential trade-off between unemployment and
the inflation rate through an expansionary monetary policy may thus
be foiled. Taking the Lucas-critique into account, equation (2.1) should
be rewritten as
yt+1 = F (yt , xt , θ(λ), µ, ǫt ) , (2.2)
where θ(λ) contains regime dependent parameters, while µ is thought
to consist of truly invariable taste and technology parameters.
While Lucas offered no solution to this fundamental problem, rep-
resentative agent models were soon to be thought of offering an easy
escape from it. Going beyond simple aggregate relationships and an-
alyzing the economy at a deeper level than before, macroeconomists
pretended to know the structural equations from which the aggregate
 2.2 The Representative Agent Modeling Approach 9

supply and demand curves are derived [371]. If a policy change is an-
nounced or implemented, the representative agent simply recalculates
his optimization problem, given his objective function and budget con-
straints. This approach also satisfied the desire for a microfoundation
in the new classical sense since behavior is derived from a utility max-
imization problem.8
Hartley argues though that it is impossible to identify truly in-
variable taste and technology parameters [172]. Acknowledging this,
economists should realize that the Lucas critique imposes a standard
that no macroeconomic model can probably ever fulfil. Since represen-
tative agent models suffer from the same deficiencies as old style Key-
nesian macroeconomic models, the justification for their use is greatly
undermined.

2.2.2 Building Walrasian General Equilibrium Models

The desire to build Walrasian general equilibrium models of the econ-

omy provides another strong motivation for using representative agents.
Not only are economists interested in the existence of equilibria within
these types of models, they should furthermore be unique and stable.
The Arrow-Debreu framework as the modern embodiment of Walrasian
models, however, is far too complex to be solved for millions of het-
erogeneous consumers and firms. Using a representative agent instead
makes it easy to find the competitive equilibrium allocation for a model
economy [370].
However, for Walrasian models to be true, their structural assump-
tions have to be true. False structural assumptions lead to false conclu-
sions.9 Since real individuals are obviously heterogeneous with respect
to their preferences and cognitive abilities, the assumption of a rep-
resentative agent cannot be considered structural since it is not true.
Thus, it must be superficial, i.e., irrelevant to the underlying structure
of the economy. When using representative agents instead of hetero-
geneous individuals in a Walrasian model, the modeler must believe

8
According to [172], there is another view of what constitutes an appropriate mi-
crofoundation. Keynesian models, for instance, backed up their macroeconomic
relationships with some explanatory story which is thought to be entirely suffi-
cient.
9
Hartley contrasts the requirement of true structural assumptions in Walrasian
equilibrium models with Friedman’s famous dictum that the realism of assump-
tions is irrelevant [172]. For Friedman, the validity of a theory is based on how
well its prediction match reality, no matter how realistic its assumptions.
10 2 The Rationale for Agent-Based Modeling

that the actual world would not look very different from the model if
populated by identical clones [172, p. 66].
There is mounting evidence, however, that this is not the case. First
of all, representative agent models are usually characterized by a com-
plete absence of trade and exchange in equilibrium [225], one of the
most basic activities in a market economy.10 For instance, in the classi-
cal CAPM [389, 252, 313], there is no trade after agents have completed
their initial portfolio-diversification.

2.2.3 Representative Agents and the Fallacy of Composition

It has long been known in economics that what is true for individual
agents may not hold for the aggregate economy. This phenomenon is
called the fallacy of composition [61, 172]. Together with its logical
counterpart, the fallacy of division, it highlights the tension between
micro- and macroeconomics. The economy as a whole is formed by
many consumers and firms whose interactions may cause emergent be-
havior at the macroeconomic level. Correct policy recommendations for
individual economic units may not work for the aggregate economy or
vice versa. For instance, in times of recession, a profit maximizing firm
is likely to lay off workers in order to survive, while a similar action
by the government as an aggregate player will aggravate the economic
downturn.
Representative agent models usually commit the fallacy of com-
position by ignoring valid aggregation concerns. Kirman, for instance,
provides a graphical example based on [203] in which the representative
agent disagrees with all individuals in the economy [225]. Policy rec-
ommendations based on the representative agent, a common practice
in today’s macroeconomics, are illegitimate in this case.
A rigorous treatment of this logical fallacy can be found in the liter-
ature on exact aggregation.11 Gorman [160] was the first who derived
general conditions under which the aggregation of individual prefer-
ences is possible. He showed that aggregate demand is dependent on
income distribution unless all agents have identical homothetic utility
functions. Only when their Engel curves are parallel and linear, a redis-
tribution of income will leave the aggregate demand unaffected. Other
authors [412, 203, 249] derived similar conditions for exact aggregation,
10
In the literature, there are as many no-trade theorems [306, 12, 422] as attempts
to solve this apparent contradiction with economic reality. These attempts are
usually characterized by a relaxation of the assumption of strict homogeneity of
market participants [83, 230, 428].
11
An introduction to the problem of exact aggregation can be found in [161].
 2.2 The Representative Agent Modeling Approach 11

the least restrictive ones given in [249]. In any case, those conditions
are still so special that no economist would ever consider them to be
plausible [225]. In the unlikely event of them being satisfied, no-trade
situations would be the result.
For apparent reasons, the literature on exact aggregation has largely
been ignored by representative agent modelers. In summarizing this
body of research, Kirman concludes that the reduction of a group of
heterogeneous agents to a representative agent is not just an analytical
convenience, but is “both unjustified and leads to conclusions which are
usually misleading and often wrong”. Hence, “the ‘representative’ agent
deserves a decent burial, as an approach to economic analysis that is
not only primitive, but fundamentally erroneous” [225, p. 119].

2.2.4 Expectation Formation in Markets with Heterogeneous

Investors

In asset pricing models with homogeneous investors, asset prices typ-

ically reflect the discounted expected payoffs and follow a martingale,
i.e., today’s expectation of next period’s price Et [pt+1 ] just equals the
current price pt [365]. This martingale property of asset prices implies
that the expectations of the representative investor satisfy the law of it-
erated expectations: Et [Et+1 (pt+2 )] = Et [pt+2 ], that is, his expectation
today of tomorrow’s expectation of future payoffs equals his current
expectation of these future payoffs [6]. Since all investors are alike,
individual and average expectations of future asset prices coincide.
Yet, when giving up the concept of a representative investor, the
knowledge of average expectations of future payoffs becomes impor-
tant for an individual investor. With differential private information
and public information, an individual’s expectation and the average
expectation about future payoffs are likely to diverge. Furthermore,
for average expectations, the law of iterated expectations is not satis-
fied anymore. Thus, according to Keynes [218], real financial markets
with heterogeneous agents resemble more a beauty contest in which
the competitors have to choose the six prettiest faces from a hundred
photographs, the winner being the one whose choice most nearly cor-
responds to the average preferences of the other competitors. Instead
of choosing the face that one considers prettiest, participants devote
their intelligence to anticipate “what average opinion expects the aver-
12 2 The Rationale for Agent-Based Modeling

age opinion to be” with some participants forming even higher order
beliefs about other participants’ beliefs.12
Asset pricing models with higher order beliefs, however, are prone to
an excess reliance on public signals [6]. Instead of always tracking the
fundamental value, those models may exhibit bubbles and sunspots,
i.e., equilibria that are different from the fundamental equilibrium sim-
ply because agents believe that a purely extrinsic random variable, a
so-called sunspot13 variable, has an effect on the equilibrium alloca-
tion. Bubbles and self-fulfilling sunspot equilibria constitute a kind of
market created uncertainty [418] or behavioral uncertainty [334], which
is caused by the direct or indirect interaction between heterogeneous
market participants.
The prevailing hypothesis in expectation formation since the early
seventies is that of rational expectations [315]. It assumes that agents
have perfect knowledge of the underlying market equilibrium equations
and that they use these equations to determine their rational expec-
tations forecast. As a result, the agents’ subjective expectations are
identical to the objective mathematical expectations given the avail-
able information set. Any errors in expectations must be random with
a mean of zero. In other words, a consistent bias in expectations is
inconsistent with the presumed rationality of the agents.
When forming their expectations of future payoffs, investors with
rational expectations do not stop with third or fourth order beliefs
about other beliefs. Their perfect deductive logic requires them to do
an infinite regress about the expectations of every other investor. Only
for homogeneous expectations among investors are these expectations
well defined. The expectations of heterogeneous investors, on the other
hand, that their expectations remain indeterminate under the Ratio-
nal Expectations Hypothesis [334, 11]. Thus, heterogeneous investors
cannot use deductive logic in forming their expectations about future
payoffs. Perfect rationality and rational expectations in particular can-
not be well defined in analytical asset pricing models with many het-
erogeneous investors.

12
A formalization of Keynes’ beauty contest can be found in [334, p. 276–281].
Further discussions of expectation formation among heterogeneous investors are
[425, 33].
13
Instead of “sunspots”, the terms “animal spirits” [218] or “self-fulfilling prophe-
cies” are sometimes used in the literature. The term “sunspot” is used to point out
the arbitrariness of the variable. A model solution depends on a sunspot variable
only because everyone believes in its importance. Reviews on sunspot equilibria
and endogenous fluctuations can be found in [29, 14].
 2.3 Rational Expectations and Disequilibrium Dynamics 13

2.3 Rational Expectations and Disequilibrium Dynamics

If perceptions of the environment, including perceptions about the be-

havior of other people, were left unrestricted, then economic models in
which peoples’ actions depend on their perceptions remain indetermi-
nate and can produce many possible outcomes. Such models are useless
as instruments for generating predictions as Sargent [372] pointed out.
Because of the requirement of consistency of beliefs, the Rational Ex-
pectations Hypothesis has proved fruitful in restricting the number of
possible outcomes of economic models. Yet, there are still many ratio-
nal expectations models with multiple equilibria, e.g. [284, 364], and
the problem of indeterminacy remains.
Rational expectations models also implicitly assume that agents al-
ready have acquired all relevant information such as the various prob-
ability distributions they face. The process of gathering and processing
this information to optimally choose their perceptions and actions is
intentionally left out from the analysis.14 Lucas [262] thus emphasizes
that the Rational Expectations Hypothesis is just an equilibrium con-
cept and not a behavioral assumption.15 It omits any description of
information perception, information acquisition, and information pro-
cessing. As an equilibrium concept, it can at best describe how a system
might eventually behave if it ever settles down to a situation in which
all agents have successfully acquired all necessary information [372]. It
can neither describe whether the system would actually arrive at one
specific equilibrium nor tell us anything about the out-of-equilibrium
dynamics of the system. These questions are out of the scope of rational
expectations models and cannot be answered without any additional
concepts.
A basic proposition of the Rational Expectation Hypothesis is that
expectations are unbiased estimates of the true underlying stochas-
tic process. However, if information acquisition is costly, agents will
14
This concept of rationality in economics is classified by Simon [395] as substan-
tive as opposed to his own concept of procedural rationality. Simon used bounded
rationality and procedural rationality as synonyms since the focus is on the rea-
soning process itself rather than on identifying a goal-maximizing action in a
given situation.
15
Much of the criticism from behavioral economists and psychologists is directed
towards the lacking behavioral foundations of the Rational Expectations Hypoth-
esis [397]. According to [140], however, this criticism is misdirected because it is
often raised disregarding the predictive qualities of the respective rational expec-
tations models. In his famous dictum, the descriptive accuracy of assumptions is
of no importance; they even have to be descriptively false. What matters is only
the accuracy of predictions.
14 2 The Rationale for Agent-Based Modeling

only slowly adapt their expectations to changes in their economic en-

vironment. During this adjustment process, their expectations will be
systematically in error which seems to violate the definition of rational
expectations. [368] point out though, that in fact it does not. Ratio-
nal expectation models such as [260] assume that agents learn about
changes of a stochastic process by the beginning of the next period.
While the length of this period is never explicitly defined, it is im-
plicitly understood as whatever time it takes for complete learning to
occur. Thus, if economists are about to address the short-run phenom-
ena during the adjustment process towards a new long-run rational
expectations equilibrium, they will have to create evolutionary models
of the economy.

2.4 The Economy as an Evolving Complex Adaptive

System

The evolutionary character of the economy consisting of many indepen-

dent and interacting agents led Holland [183], from the Santa Fe Insti-
tute in New Mexico, to describe it as an adaptive nonlinear network
(ANN), a term which has largely been replaced by the notion of complex
adaptive systems (CAS). The origins of this multidisciplinary concept
are in mathematics and the natural sciences [430, 20, 322]. However,
when trying to identify the distinguishing properties of complex adap-
tive systems, one realizes that no commonly agreed upon definition
of complexity exists.16 It ranges from simple statements about com-
plicated behavior to exploding computational or informational require-
ments for solving certain problems. Simon [398], for instance, points out
that complexity can arise from sheer numbers. In statistical mechanics,
for instance, physicists study systems made of large numbers of inter-
acting, i.e., colliding and energy exchanging, particles.17 While these
particles are often assumed to be homogeneous, the ‘particles’ in social
or economic interactions—consumers, investors, firms, private or gov-
ernmental organizations—are not. Contrary to elementary particles in
physics, economic agents have preferences and a free will. They thus en-
gage in decision-making which increases the degree of complexity found
in social systems. However, local interactions between particles is not
a sufficient condition to form a complex system. In his introduction to
16
A discussion of these various definitions can be found, for instance, in [146] in the
first issue of the journal Complexity and on [361].
17
[248] point out that one of the first applications of the microscopic simulation
method was for nuclear fission in the 1950’s.
 2.4 The Economy as an Evolving Complex Adaptive System 15

dynamic complex systems, Bar-Yam emphasizes that the distinguish-

ing property of thermodynamic and complex systems in physics is the
concept of interdependency [23]. In an economic context, interdepen-
dencies arise because the decisions made by one agent usually affect
other agents in the economy.
The “Santa Fe Perspective” of complexity has been defined by [186]
and by [10, p. 3-4] who list six features of the economy which may cause
traditional linear mathematical methods to fail when studying CAS.
1. There are dispersed parallel interactions between many heteroge-
neous agents.
2. There is no global entity that controls the agent interactions. In-
stead of a centralized control mechanism, agents compete with each
other and coordinate their actions.
3. The economic system has many hierarchical levels of organizations
and interactions. Lower levels serve as building blocks for the next
higher level.
4. Behaviors, strategies, and products are continuously adapted as
agents learn and accumulate experience.
5. Perpetual novelty leads to new markets, new behaviors, and tech-
nologies. Niches emerge and are filled.
6. The economy works far away from any optimum or global equilib-
rium with constant possibilities for improvement.
Because of these difficulties for traditional analytical methods, agent-
based modeling has now become a major tool to study complex adap-
tive systems. The predecessors of agent-based simulation of CAS in
economics—cybernetics, catastrophe theory, and chaos theory—share
a common thread of nonlinear relationships and an interest in discon-
tinuities.18 Recessions, market bubbles, or crashes are just simple ex-
amples of such discontinuities in economic systems. The most famous
representative of the cybernetics approach in economics has been Jay
Forrester [133, 134, 135] who created the World3 model, the founda-
tion for the influential Club of Rome report “The Limits to Growth”
[295].19 Forrester’s World3 is an aggregate model of the world economy
consisting of a system of nonlinear difference equations. Even though
feedback loops were modeled, learning and, thus, shifting aggregate
relationships were completely ignored in the World3 model.
18
For a comparison of these four concepts see [193, 361]. Horgan [193] also fires a
harsh criticism at complexity theory and all its predecessors.
19
A technical documentation of the World3 model is found in [296]. There is a 20
and 30 year update of the original study available [294, 297]. Critiques of the
theoretical foundations were formulated by Nordhaus [323, 324].
16 2 The Rationale for Agent-Based Modeling

An economic example of catastrophe theory can be found in [444]

where stock market crashes are seen as a result of too many chartist
traders relative to fundamentalist traders. Chaos theory was popular-
ized in [149] and is characterized by a sensitive dependence on initial
conditions and a deterministic system behavior which appears to be
stochastic. It remains a lively area in economic research [335, 336, 54,
55] and agent-based complexity models are nowadays often tested to
see whether they also exhibit chaotic behavior [266, 145].
While sharing the thread of nonlinearities and discontinuities with
its predecessors, the main theme of complexity theory is that of emer-
gence. The complicated behavior at the aggregate level is explained as
the result of the interactions of constantly adapting agents whose mi-
crobehaviors are qualitatively different. Loosely speaking, emergence
is the appearance of some patterns, structures, or properties at the
macro level of a complex adaptive system which are not peculiar to
the constituting parts.20 Emergence is also an example of the fallacy of
composition (see section 2.2.3) and is also closely related to the concept
of “self-organized criticality” [20].

2.5 Some Methodological Aspects of Agent-Based

Simulations

The ability to study high-dimensional complex adaptive systems in

economics through agent-based simulations, however, comes at a cost.
Agent-based simulations not only require inductive methods on the
level of the agents, but also require inductive reasoning on the level of
the modeler. Unlike most analytical models, an agent-based simulation
does not produce theorems and existence proofs. It usually generates
time series of state variables both on the agent level and on the macro
level. However, one simulation run is just one particular realization of
the potential infinite set of all possible realizations. To gain an under-
standing of the model behavior, researchers have to analyze the gen-

20
A commonly accepted formal definition of emergence in CAS does not yet exist
to my knowledge. Bankes [22] observed that the declaration of emergent behavior
seems to be at the discretion of the researcher who intuitively relies on visual ob-
servations of differences between the micro- and macrolevel. Thus, he calls for an
operational definition of emergence such as the specification of some standardized
threshold level of micro- and macrolevel differences above which emergence can
be declared.
 2.5 Some Methodological Aspects of Agent-Based Simulations 17

erated time series with econometric methods, and general conclusions

can only be made by means of inductive reasoning.21
Pheby [337] explains that induction is ampliative, i.e., its conclu-
sions go beyond the available evidence. It involves reasoning from spe-
cific statements towards more general ones, while deductive arguments
move from general to more specific statements.22 Judd [208] then points
to a trade-off between deductive and inductive methods in economics.
One can either achieve logical purity with deductive methods in low-
dimensional models of the economy while sacrificing realism, or analyze
more realistic high-dimensional models and accept numerical impreci-
sion.23 The real issue is not whether to approximate, but where. In com-
paring the strengths and weaknesses of each approach, Judd concludes
that computational models with inductive methods and deductive rea-
soning in economic theory are not substitutes, but rather complements
[208]. Pheby [337] also asserts that the problem associated with induc-
tion is not so much one of rationality, but one of reliability.
The reliability issue is also connected with another objection that is
often raised in connection with computational methods in economics.
The black box criticism is the argument that the results of computa-
tional models are counterintuitive and incomprehensible because the
computer program generating the results is only seen as an impenetra-
ble black box. For Judd, this objection points to a more general problem
of how computational studies are presented. Opening the black box by
making the source code publicly available is one way to address this
problem, extensive sensitivity analysis is another. The current fragmen-
tation of the agent-based community in economics with respect to pro-
gramming languages and simulation platforms may limit the impact of

21
Axelrod [13] views simulation as a third research method in science. Like deduc-
tion, it starts with explicit assumptions, but it does not prove theorems. The
simulated data are then analyzed by means of indexinduction induction. But un-
like induction, these data come from a specified rule set and not from real world
observations.
22
According to Popper [341], the long-standing debate about the problem of induc-
tion in science was initiated by David Hume in the eighteenth century. Because
past evidence tells us only about past events, to base expectations about future
events on them is simply irrational. Popper then claimed to have solved Hume’s
induction problem by positing his principle of falsification as a direct antithesis
to induction. In economics, the tension between inductive and deductive methods
culminated in the famous “Methodenstreit” (conflict of methods) between Gus-
tav Schmoller, a representative of the Younger Historical School in Germany, and
Carl Menger as a member of the Austrian School of Economics.
23
A similar trade-off between rigor and relevance in economics is presented in [291].
18 2 The Rationale for Agent-Based Modeling

making source codes available.24 But as with the lack of sufficient sen-
sitivity analyses and standardized testing methods for computational
models, it should not be a severe obstacle for the usefulness of agent-
based simulation in economics.

24
The Santa Fe Artificial Stock Market, for instance, was first programmed in
Objective-C. Later, it was reprogrammed to utilize the SWARM-simulation li-
brary. Wilpert [437] ported the model to Borland-C++, while Ehrentreich [106]
reprogrammed the market under Java and the RePast-library. Other artificial
stock markets use mathematical program packages such as GAUSS [268, 269] or
Matlab [240].
3
The Concept of Minimal Rationality

I have had my results for a long time: but I do not yet know
how I am to arrive at them.
Karl Friedrich Gauss

3.1 Introduction

For a long time economists have felt that the idealization of rationality
for modeling purposes needed theoretical justification. The first evo-
lutionary arguments for rationality appeared during the 1930’s in the
works of Schumpeter [382] or in the famous marginalism controversy
by Harrod [171] who compared new business procedures to “mutations
in nature”. Alchian [3] later claimed that competition will lead to a
survival of firms with positive profits, whether or not they consciously
maximize them or not.1
The debate whether forces of competition will assure that only the
“fittest”, i.e., most rational actors survive in the market was further
spurred on by Friedman’s famous “as-if” argument [140]. He claimed
that individuals who do not behave completely rationally, for instance,
by not consciously maximizing expected returns, could still be consid-
ered as if they do. Business behavior that is consistent with the maxi-
mization of returns hypothesis, even though it may only be by chance,
will lead to prospering firms, while “natural selection” will weed out
those firms whose behavior contradicts the maximization hypothesis.

1
A discussion of the marginalist controversy can be found in [310]. Evolutionary
ideas in economics, however, can be traced back even further than this. Hodgson
[180] gives a comprehensive account of the history of evolutionary thought in
economics.
20 3 The Concept of Minimal Rationality

Instead of creating economic models which are based on more re-

alistic assumptions of individual behaviors, the arguments above sug-
gest that one can easily take a shortcut by outright assuming ratio-
nal behavior. This is also in line with Friedman’s other famous claim
that the realism of an assumption does not matter as long it yields
correct predictions. This argument was also used by Becker [28], who
showed that negatively inclined market demand curves can result both
from rational and irrational behavior. In a way, this could be labeled
as emergent rationality. Rationality on the macro-level, i.e., coherence
with the aggregate results predicted by models populated by rational
agents, would come about in spite of irrationality on the micro-level.
Emergent rationality as a property of a complex adaptive system is
also investigated in [69], who formally interpret both the Efficient Mar-
ket Hypothesis and the Rational Expectations Hypothesis as an exam-
ple of emergent behavior in the context of an artificial stock market.
Evidence that indexstrategic complementarity “strategic complemen-
tarity” and “strategic substitutability” are important in determining
aggregate outcomes is provided in [125]. Strategic substitutability de-
scribes a situation in which an increase in the action of one individual
results in a decrease in the action of other individuals. A minority of ra-
tional agents can be sufficient to generate aggregate rationality. Under
strategic complementarity, i.e., a situation in which an increase in i’s
action causes increases in the actions of other agents, a small amount
of individual irrationality may suffice to generate large deviations from
the predictions of rational models. Herding behavior in financial mar-
kets is an example for strategic complementarity. The possibility of
strategic complementarity, however, is a strong argument against the
unreflective use of rational agents.
The generality of the competition argument for assuming rational
agents was critically examined by Winter [438] who argued that in
evolution not the best conceivable, but the best available individual
survives. Similarly, Blume and Easley illustrate that the link between
fitness, i.e., a quantitative measure of success in a specific environment,
and rationality in repeated financial markets may be weak and that
rationality does not necessarily produce fit rules [41]. In general, the
competition argument for rationality turns out to be highly conditional,
and counterexamples are easily found [124, 98, 34]. The as-if argument
was critiqued by Conlisk [79] and Sunder, for whom it is unsatisfactory
since scientific models serve not only to predict, but also to convey an
understanding to others: “Understanding of phenomena is crucial to
 3.2 Economic, Bounded, and Situational Rationality 21

science; prediction without understanding does not build science [413,

p. 4].2
Despite vague references to competition and natural selection, how-
ever, the specific mechanism by which aggregate and/or individual ra-
tionality would come about had largely been left unanswered and re-
mained a black box concept. Starting with the contribution by Nelson
and Winter [319], the now burgeoning literature on modern evolution-
ary economics tries to open this black box by rigourously modeling
dynamic processes of change in the economy. In its broadest sense,
Witt characterizes evolutionary economics “by its interest in economic
change and its causes, in the motives and the understanding of the in-
volved agents, in the processes in which change materializes, and in
the consequences” [440, p. xiii]. A good starting point for evolutionary
economics is the works of [439, 49, 101].
Evolutionary economics is primarily concerned with the equilibrium
selection problem [284, 435, 364]. Its other focus is to model how agents
learn to become ‘rational’ in the economic sense [40, 137, 175]. The
primary concern with equilibrium selection, however, is criticized by
Börgers [48]. Since there is mixed empirical evidence whether and under
which situations agents behave rationally or not, he urges first studying
how rationality comes about. He emphasizes that this logically precedes
the question of equilibrium selection.

3.2 Economic, Bounded, and Situational Rationality

Learning in an evolutionary context is often considered to be an adap-

tive process enabling agents to become rational. The prevailing concept
of rationality in economics is an operational refinement of the former
popular term self-interest and includes perception, preference, and pro-
cess rationality [292]. Perception rationality assumes that agents be-
have as if they process information to form perceptions and beliefs
by using Bayesian statistical principles. Preference rationality requires
the acceptance of some well-defined axiom sets, for instance, the ax-
ioms of Herstein and Milnor [178] and Savage [374]. Typically, these

2
An extensive discussion and critique of Friedman’s evolutionary argument can be
found in the chapter “Optimization and Evolution”, in [180, ch. 13]. Celebrating
the fiftieth anniversary of Milton Friedman’s ‘The methodology of positive eco-
nomics’, the 2003 annual meeting of the Allied Social Science Associations devoted
a complete symposium to it. The papers delivered at that session, e.g., [353, 273],
are contained in a special issue of the Journal of Economic Methodology.
22 3 The Concept of Minimal Rationality

axioms require a complete, continuous, and transitive preference order-

ing which should be invariant to framing and independent of irrelevant
alternatives. Finally, process rationality in economics implies the max-
imization of an objective function subject to some constraints. I will
refer to this rationality concept as economic rationality.3
Even though this is a widely accepted perception of rationality, at-
tacks on it from within the economic profession or from other fields do
not cease. One strand of criticism is directed towards its axiomatic foun-
dations. By now, there is abundant empirical evidence that real decision
makers often deviate from these rationality axioms [4, 109, 212, 258].4
The competing concept of bounded rationality in economics primar-
ily criticizes the neoclassic notion of process rationality and calls for re-
placing the maximization principle through satisficing procedures.5 Its
founder, Herbert A. Simon, also refers to it as procedural rationality. He
argues that economics has largely been preoccupied with the results of
rational choice rather than processes of choice and thus, defines rational
behavior when it is the outcome of appropriate deliberation [395, 396].
A central feature of bounded rationality is non-optimizing procedures
such as simple search heuristics and decision rules [388].
However, I follow Langlois in defining rationality in a situational
context [231, 233]. The method of situational analysis goes back to
Max Weber, Alfred Schütz, and most recently, Karl Popper .6 It “in-
sists that agents act not optimally but merely reasonably under the
circumstances” [231, p. 693]. While the concept of economic rational-
ity emphasizes solely the agent’s preferences as the basis for his best
course of action, situational analysis focuses on the circumstances in
which these preferences are embedded.7

3
The term “economic rationality” might not be optimal since it is by no means
“economical.” It always requires the use of all our cognitive and computational
capabilities—sometimes even beyond that—to perform simple tasks for which
simple heuristics would suffice. In this respect, procedural or situational rational-
ity, to be defined below, are more economical concepts of rationality.
4
A survey about violations of the basic postulates of economic rationality can be
found in [64].
5
Herbert A. Simon revived the Scottish word ”satisficing” (=satisfying) to denote
problem solving and decision making that sets an aspiration level, searches until
an alternative is found that is satisfactory by the aspiration level criterion, and
elects that alternative [394, Part IV ].
6
For a discussion of Weber’s and Schütz’s contributions to rational choice theory,
see [379].
7
A typical definition of economic rationality by Hirshleifer [179] illustrates the
lack of any situational components. “In the light of one’s goals (preferences),
if the means chosen (actions) are appropriate, the individual is rational; if not,
 3.2 Economic, Bounded, and Situational Rationality 23

Langlois agrees with the proponents of bounded rationality that

there are many situations in which it would be quite unreasonable to
solve foot-long Lagrangians to determine an optimal response [231].
The culprit, however, is not the optimization approach as such, but
its misapplication to ill-defined situations. Langlois thus criticizes that
behaviorists threw out the baby with the bath water when the opti-
mization approach was completely replaced with ad-hoc rules about
satisficing behavior. Thus, not the logic of the situation, but ad-hoc
imposed behavioral rules determine an agent’s behavior. In cases in
which models of bounded rationality offer a theory from where these
rules come from, they become an instance of situational analysis.8 The
same can be said about neoclassical models when optimization is rea-
sonable under the given circumstances. Thus, situational analysis com-
bines both approaches to rationality.
It was pointed out in chapter 2 that the assumption of heterogene-
ity leads to indeterminate expectations of fully rational agents. Does
this mean that heterogeneous agents must form their expectations in
an irrational way? Barro, for instance, dubbed the appropriation of
the term rational by the rational expectations revolution as one of its
cleverest features since opponents of this approach were then “forced
into the defensive position of either being irrational or of modeling oth-
ers as irrational, neither of which are comfortable positions for most
economists” [25, p. 179]. Simon [396], however, points out that the
economic notion of rationality does not correspond to any classical cri-
terion of rationality.9 The label “rational expectation models” provides
a rather unwarranted legitimization. In his view, a more neutral name
such as “consistent expectations” would have been the better choice.10

irrational. ‘Appropriate’ here refers to method rather than result.” Note that
rationality is solely defined in terms of one’s preferences, not in terms of the
specific environment the decision maker faces.
8
Langlois [232] points out that models in the field of New Institutional Economics
are able to explain the evolution of rules. Norms and conventions emerge through
a process of repeated games. Eventually, agents will discover “evolutionary stable
strategies.” These strategies are simple rules which then become institutionalized.
9
A dictionary definition simply asserts that being rational means “agreeable to
reason, reasonable, sensible”. In psychology, an appropriate deliberation process
is viewed as a key ingredient of rationality. It is in this tradition that Herbert
Simon used bounded rationality and procedural rationality as synonyms since
the focus is on the reasoning process itself rather than on identifying a goal-
maximizing action in a given situation.
10
However, the term consistent expectations equilibria has now been claimed by
Hommes and Sorger [191] and Sögner and Mitlöhner [406]. It is an informationally
less demanding concept than the Rational Expectations Hypothesis. Agents do
24 3 The Concept of Minimal Rationality

Since economic rationality may imply unlimited cognitive and com-

putational capabilities on the agent’s part, it thus seems to be the
limiting extreme in a spectrum of possible degrees of rationality. On
the other end of that spectrum, one would find “perfect” irrational be-
havior. This, however, would probably be far from being random since
a ‘perfectly irrational’ agent would first figure out a rational or optimal
solution to a problem and then, for some pathological reason, systemat-
ically choose to do the opposite (if an opposite action can be defined).
Random behavior, on the other hand, may not be as irrational as one
would think. A situation that somehow eludes analytical investigation,
mixed strategies over all possible actions might be the best an agent can
do. For practical purposes, the spectrum of behaviors that a modeler
could chose from is most likely limited between pure random behavior
and perfect economic rationality.
From these different levels of rationality, which one should he choose
though? This book espouses the notion of minimal rationality [70],
which is an approach where the agent’s actions typically fall between
randomness and perfection. Instead of positing unbounded rationality,
why not investigate how much intelligence or rationality is necessary to
explain the observed data? With a few notable exceptions, e.g., [150,
356, 413, 123], this principle has seldom been applied to economics.

3.3 Situational Analysis, Minimal Rationality, and the

Prime Directive

The principle of minimal rationality is the antithesis to what Rubin-

stein [362] referred to as the Prime Directive in financial economics:
“Explain asset prices by rational models. Only if all attempts fail, re-
sort to irrational investor behavior.” He laments that the burgeoning
behavioralist literature seems to have lost all constraints to adhere to
this directive by always seeking an explanation for a financial anomaly
in systematic irrational behavior.
Unfortunately, Rubinstein does not explicitly state whether he sub-
sumes models with lesser degrees of rationality into the group of ratio-
nal models or not. However, since he defines rationality as the adher-
ence to the Savage axioms [374], it is implied that investors maximize
expected utility using unbiased subjective probabilities. Rubinstein’s
insistence on the maximization principle thus implies that he does not

not need to know the underlying market equilibrium equations. They only need
to have beliefs which turn out to be consistent with actual observations.
3.3 Situational Analysis, Minimal Rationality, and the Prime Directive 25

grant the rationality status to models with lesser rationality. Hence, it

can be assumed that in his view the Prime Directive and the principle
of minimal rationality are indeed irreconcilable positions.11
Friedman [140] points out that the validity of a hypothesis is not
a sufficient criterion for choosing among alternative hypotheses.12 If
several hypotheses perform equally well for predictive purposes, Fried-
man’s instrumentalist answer does not help in deciding which level of
rationality should be preferred. Since descriptive validity of the model’s
assumptions does not matter either in his view, what is then a sufficient
criterion for choosing among alternative levels of agents’ rationality?
The Prime Directive asserts always picking the (most) rational model.
The philosophical basis for the preferential treatment of (economic)
rationality is sought by Rubinstein in the long cultural and scientific
tradition dating back to the ancient Greeks who spoke of “reason” as
the guide to life. However, it has already been mentioned that the defi-
nition of rationality in economics is rather new and much more specific
than it used to be in the scientific tradition invoked by Rubinstein.
Proponents of the bounded rationality approach to economics often
invoke a principle known as Occam’s razor to justify their deviance
from full rationality. This principle is attributed to the Franciscan monk
William of Occam (ca. 1285–1349, sometimes also known as William of
Ockham) and is nowadays usually stated as follows: “Of two competing
theories or explanations, all other things being equal, the simpler one is
to be preferred.”13 Scientific parsimony, for which Occam’s razor stands,
tells us to prefer the postulate that men are merely reasonable over
the postulate that they are supremely rational [396]. However, in his
Nobel Prize lecture, Simon [397] points out that Occam’s razor has a
double edge. Even though satisficing models allow for weaker rationality
assumptions, optimization models can be stated more briefly than the
former. The two edges thus cut in opposite directions.
The philosophical justification for the principle of minimal rational-
ity, however, is not sought in an application of Occam’s razor.14 Min-
imal rationality is seen as a straightforward application of situational

11
Rubinstein, however, allows for investor heterogeneity by allowing differences of
opinion and uncertainties about other investors’ characteristics.
12
Friedman also claims that if there is one hypothesis that is consistent with the
available evidence, then there is an infinite number that are consistent (p. 9).
13
In Occam’s own words it reads as “Pluralitas non est ponenda sine neccesitate”,
which translates into English as “Causes are not to be multiplied beyond neces-
sity.”
14
Karl Popper points out that the application of Occam’s razor may be used inap-
propriately, which may lead to bad problem reductions [341, p. 308].
26 3 The Concept of Minimal Rationality

analysis combined with the economic principle. Since human agents

have only limited resources, e.g., memory, time, and computational
abilities, efficient reasoning strategies are the result of minimizing the
involved costs in decision making given a certain aspiration level. Solv-
ing foot-long Lagrangians for complicated decisions under time pres-
sure might not be reasonable given the fact that simple heuristics might
lead to satisfactory results. The economic principle contains the idea
of achieving the most with a given amount of input or minimizing the
input level to achieve a given output level. If we view cognitive and
computational abilities as scarce resources that the agent economizes
on, an analysis of the circumstances lead the agents to employ just
as much intelligence and reasoning as necessary to perform the task
they face. Because of the economic principle, situational analysis and
minimal rationality are two concepts that necessitate one another.15

3.4 Minimal Rationality and the Phillips-Curve

It is surprising that the final breakthrough of the Rational Expectations

Hypothesis seemingly adhered to the principle of minimal rationality.
Even though the Rational Expectation Hypothesis had been developed
by John F. Muth [315] as early as 1961, the rational expectations rev-
olution was triggered only by the influential contributions of Robert
E. Lucas [260, 261]. Its power of persuasion was derived from the dis-
cussion revolving around the famous Phillips curve, for which rational
expectations turned out to be the appropriate level of rationality.16
The Phillips-curve discourse was opened by Arthur W. Phillips
[339], who found a negative relation between wage inflation ŵ and un-
employment U , a result that was verified and refined by Lipsey [254].
The Phillips curve received its political explosiveness from Samuelson
[367], who took productivity growth into account and substituted wage
inflation with price inflation p̂. Since price inflation is an exogenous
parameter that can largely be controlled by the monetary authorities,
a society, in principal, should be able to choose from any of the combi-
nations on this modified Phillips curve.
Both Phillips and Lipsey argued that the negative slope of the
Phillips curve is caused by nominal wage pressure due to excess de-
mand or supply on the labor market, the unemployment rate being
15
There is, however, a possible infinite regress in determining the optimal decision
procedures. A solution to this problem is suggested by Lipman [253] in an article
with the telling name “How to decide how to decide how to . . .” .
16
An overview of this discussion can be found in [368, 142].
 3.4 Minimal Rationality and the Phillips-Curve 27

a proxy for the latter. However, instead of purely negotiating nomi-

nal wages, employers and employees bargain over real wages w/p for a
contract period. Since the nominal wages are usually fixed during the
term of the wage contract, both sides have to form expectations about
the future price level within the contract period. In surveying the lit-
erature on the Phillips curve, though, Santomero and Seater conclude
that the early authors did not consider expectations as a reason to
include price changes as an explanatory variable [368]. In hindsight,
one realizes that this first stage of the Phillips curve discussion was
characterized by an implicit assumption of static inflation expectations
E [pt+1 ] = pt . Since it was obviously incapable of accounting for mount-
ing empirical evidence of shifting Phillips curves and loops in the time
series, static inflation expectations turned out to be an unsatisfactory
level of rationality on the agents’ part.
It was Edmund Phelps [338] and Milton Friedman [141] who drew
attention to the role of inflation expectations and posited the natural
rate hypothesis. Search frictions and other market imperfections would
determine an equilibrium rate of unemployment—the natural rate of
unemployment—that is independent of nominal phenomena such as
the price level or expected changes in the price. Since this natural rate
hypothesis is inconsistent with a long-run trade-off between inflation
and unemployment, it implies a vertical long-run Phillips curve. If the
the monetary authorities increase the inflation rate, agents would soon
begin to adjust their inflation expectations. Thus, instead of moving
along a stable, negatively-sloped Phillips curve, the short-run Phillips
curve shifts upwards, and after completion of the expectations adjust-
ment process, the economy ends up with a higher inflation rate at the
same equilibrium unemployment level.
While Friedman did not specify the exact mechanism for adjusting
inflation expectations, the following second stage in the Phillips curve
discussion typically employed some kind of higher order autoregressive
inflation expectations
T


E [pt+1 ] = cn pt−n . (3.1)
n=0

Negatively-sloped Phillips curves emerged only because agents were

slow to adapt their inflation expectations to the actual change.17 For

17
For static inflation expectation, i.e., for c0 = 1 and all other cn = 0 in equation
3.1, agents do not adapt their inflation expectations at all.
28 3 The Concept of Minimal Rationality

positive inflation rates, individuals’ inflation expectations will thus con-

sistently be downwardly biased. While the use of extrapolative expec-
tations

E [pt+1 ] = pt + α (pt − pt−1 ) (3.2)

= (1 + α)pt − αpt−1
= c0 pt + c1 pt−1

or adaptive expectations [62, 321]]

E [pt+1 ] = E [pt ] + α (pt − E [pt ]) (3.3)

T


= α(1 − α)n pt−n
n=0

certainly increased the level of rationality in expectation formation,

these concepts are oriented completely backwards. Agents do not con-
sider any information they might have about future events affecting
the inflation rate. Neglecting this information is certainly less rational
than using it. The systematic downward bias of autoregressive inflation
expectations contradicts the rationality assumption of economic agents
since, for rational expectations, errors in expectation must be random
with a mean of zero. In other words, rational expectations forecasts are
unbiased estimates of the true underlying stochastic inflation process.
Under rational expectations, agents are using all available information.
If they correctly anticipate a monetary expansion through the central
bank, the expansion itself will have no economic effects. Monetary pol-
icy is effective only to the extent that it deceives economic agents.
The development of the Rational Expectations Hypothesis thus
closely followed the principle of minimal rationality. Only when re-
sorting to fully rational individuals, the models were able to explain
the observed empirical relationships between unemployment and price
inflation. It is thus not surprising that the assumption of rational ex-
pectations became the standard assumption in economic theory. It is,
however, likely that in other problem fields, lesser degrees of agent
rationality may suffice to equally explain the empirical facts. The prin-
ciple of minimal rationality would then prescribe the use of the model
that requires the least amount of rationality.
4
Learning in Economics

You can always count on Americans to do the right thing —

after they’ve tried everything else.
Winston Churchill

4.1 Introduction

Replacing fully rational economic agents with boundedly or minimally

rational agents leads to the necessity of learning. Another motivation to
endow agents with learning routines is to predict unique outcomes when
a model possesses multiple equilibria. Being less than perfectly rational
or omniscient does not mean that agents are irrational. Instead, they act
according to the knowledge and skills they are equipped with. Learning
agents then have the potential to gradually discover an optimal or at
least satisfactory course of action.
After defining learning in more detail, this chapter briefly discusses
the rationality-based Bayesian learning model. At the center of inter-
est, however, are biologically inspired learning models. While replicator
dynamics are extensively used in game theory, genetic algorithms or
classifier systems are widely used in agent-based simulations.

4.2 Definitions of Learning

Modeling learning processes requires us to define learning more pre-

cisely. In the psychological literature, a behaviorist approach is usu-
ally taken. Burrhus F. Skinner, perhaps the most famous among the
behaviorists, defines learning as a “change in probability of response”
[399, p. 199]. Kimble set forth a widely accepted definition according
30 4 Learning in Economics

to which “learning is a relatively permanent change in a behavior po-

tentiality which occurs as a result of reinforced practice” [219, p. 82].
Bower and Hilgard add that it is necessary that this change “cannot
be explained on the basis of the subject’s native response tendencies,
maturation, or temporary states (such as fatigue, drunkenness, drives,
and so on)” [50, p. 11].
The emphasis on change in response probability or behavior poten-
tiality indicates that learning may not be translated into behavior un-
til some time after the learning has taken place [177]. Because of this,
learning may not be immediately observable. According to Brenner,
this specification does not matter in an economic context. Economists
are not interested in the study of learning processes themselves, it is
only their consequences, i.e., changes in economic behavior, that are of
interest to them [52].
Through the principle of reinforcement, i.e., rewards or punishments
that are associated with certain actions, individuals identify successful
behaviors and attach higher activation probabilities to them. Failing
behaviors will be less likely to be acted upon in the future. Brenner,
however, considers this reinforcement learning, or conditioning as it was
originally named by psychologists, as too restrictive in an economic con-
text. While humans share this simple way of learning with animals, we
are, in addition, able to reflect on our actions and their consequences.
The psychology literature labels this as cognitive learning. Thus, Bren-
ner’s learning definition includes “any cognitive or non-cognitive pro-
cessing of experience that leads to a direct or latent change in eco-
nomic behaviour, or to a change of cognitive pattern that influences
future learning processes” [52, p. 3]. Because our cognitive resources
are scarce and thus, cannot be applied to every situation, Brenner [53]
argues that many of our actions are a result of reinforcement learn-
ing. He considers the effect of cognitive learning on behavior, however,
stronger than that of reinforcement learning.
The simultaneous existence of at least two kinds of learning pro-
cesses implies different ways of how to model them. Brenner [53] no-
tices an increasing number of learning models in economics in the past
few years and points out that there are many possible ways to classify
them.1 Based on their origin, he distinguishes them as psychology-based
models such as reinforcement learning, rationality-based models such
as Bayesian or least-square learning, or models inspired from computer
science and biology such as evolutionary algorithms or neural networks.
He remarks, however, that different learning models are usually applied
1
Other surveys on learning theories in economics can be found in [400, 405].
 4.3 Rationality-Based Learning Models 31

in different economic fields and, therefore, might be classified accord-

ingly. While Bayesian and least-square learning dominate in macroeco-
nomics, game theorists seem to prefer fictitious play, replicator dynam-
ics, and other adaptive learning models. Experimental economists tend
to employ reinforcement learning, fictitious play, and indexlearning di-
rection theory learning direction theory, while evolutionary algorithms
and genetic programming are prominent in agent-based computational
models. Besides the fact that mathematical economists are limited in
their choice to analytically treatable models, these different preferences
within the economics profession mainly seem to reflect historical devel-
opments rather than inevitable design decisions.

4.3 Rationality-Based Learning Models

In the beginnings of modeling learning in economics, normative ap-
proaches such as Bayesian or least-square learning have been the most
dominant. These rationally-based models assume that agents behave
optimally, first by taking the best action regarding the available infor-
mation, and second, by rationally updating their prior opinion as soon
as new evidence becomes available.2 Descriptions of both learning types
can be found in [52, 53]. In the case of Bayesian learning, individuals
start out with an initial set of hypotheses H about the situation they
face. Each hypothesis h ∈ H is characterized by a list of subjective
probabilities P (e|h) with which mutually exclusive events
 e in the set
of possible events E occur, i.e., for each hypothesis e∈E P (e|h) = 1.
In period 0, agents typically assign to each hypothesis h an equal prob-
ability p(h, 0). If better information is available, these so-called prior
probabilities might differ from each other. The set of possible events
has to be complete and complementary, i.e., h∈H p(h, t) = 1 must
hold for all periods t.
As agents learn about new evidence in the next period, they update
their initial subjective probabilities associated with each hypothesis
according to
P (e(t)|h) · p(h, t)
p(h, t + 1) =  . (4.1)
h∈H P (e(t)|h) · p(h, t)

Through this equation, an observed event increases the probability of

that hypothesis for which the occurrence of this event is most likely.
2
Blume and Easley, however, consider the term “rational learning” as a poorly cho-
sen euphemism. In their view, Bayesian learning does not deserve this normative
judgement [42].
32 4 Learning in Economics

The
 normalization through the denominator ensures that the condition
h∈H p(h, t) = 1 is fulfilled for each period. As time proceeds, the
probability p(h, t) for the correct hypothesis should converge to 1, while
the probabilities for the other hypotheses should approach zero. Given
these updated probabilities for each hypothesis, agents chose an action
that maximizes their expected utility.
While Brenner points out that the basic mechanism of Bayesian
learning roughly corresponds to the psychological notion of cognitive
learning, he admits that it lacks empirical or experimental justification.
It is very doubtful whether people are indeed able to do the necessary
calculations that Bayesian updating requires. He also refers to the psy-
chology literature which finds that people tend not to simultaneously
consider many competing hypotheses [52]. Experimental evidence such
as Monty Hall’s three-door anomaly, too, suggests that agents consis-
tently deviate from the rational Bayesian solution [139].

4.4 Biologically Inspired Learning Models

Evolutionary algorithms such as evolutionary strategies [351] and ge-

netic algorithms [182] were initially developed to optimize technical
systems.3 They are designed to mimic natural evolutionary processes
since they have been very successful in creating well-adapted species
to changing environments. Evolution is sometimes simply character-
ized as “transmission with adaptations”, whereby adaptations simply
refer to differences. John Holland defines adaptation more precisely as
any “process whereby a structure is progressively modified to give better
performance” [182, p. xiii]. Since these adapted structures may be as
diverse as protein molecules, human brains, or investment rules, adap-
tation is a rather universal principle applicable to many different fields.
Because learning theories are concerned with how behavioral de-
cision rules are modified, evolutionary algorithms were soon thought
to describe individual and social learning processes [51, 355, 401]. The
connection is usually made by analogy [354]. For instance, the genotype
in biology is thought to correspond to behavioral decision rules. The
process of mutation of a gene could be translated into the creation of
new decision rules through experiments or unintended mistakes. New
decision rules can also be the result of communication between agents
who derive them by combining previously successful strategies. This
would correspond to the concept of crossover. Contrary to mutation
3
An overview of the history of evolutionary computation can be found in [93].
 4.4 Biologically Inspired Learning Models 33

and crossover, replication, or reproduction as it is often referred to, is

variety preserving and can be interpreted as the imitation of successful
decision rules. Selection is a variety restricting process of choosing in-
dividuals for reproduction, while fitness is the quality that is selected
for [162]. As such, fitness only represents a propensity for evolution-
ary success, which is a retrospective measure of the relative increase
or decrease in the whole population of possible solutions or decision
rules.4 As Hodgson [180] points out, even though fitness implies an ex-
pectation of success, it does not necessarily means survival. The fittest
solution can go extinct while bad solutions have a positive probability
of survival.5
It has repeatedly been pointed out that the analogies between learn-
ing and evolutionary algorithms are not isomorphic. Brenner [51] and
Slembeck [401] emphasize that in spite of similar structure, evolution-
ary algorithms and learning processes possess crucial differences. First,
individuals are likely to improve their decision rules less stochastically
than suggested by the genetic operators , crossover, and selection, which
are essentially stochastic processes. Brenner finds that motivational as-
pects cannot be adequately captured by evolutionary algorithms. Sec-
ond, they generally have no memory, thus, past experiences of individ-
uals are hard to model.6 Third, the definition of a fitness function is
easier and more objective than the way individuals attribute success
to certain decision rules. In spite of these concerns, evolutionary algo-
rithms continue to be widely used to model learning processes. A per-
fect analogy of evolutionary algorithms with learning processes might
not be necessary as long as the results adequately describe changes in
human behavior.7

4
In population genetics , fitness is often defined as reproductive success which is
only an ex post measure. Since species or members of a specific species do not come
with a fitness function attached to them, evolutionary biology has tremendous
difficulties in defining fitness. Lewontin admits that “although there is no difficulty
in theory in estimating fitnesses, in practice the difficulties are insuperable. To
the present moment no one has succeeded in measuring with any accuracy the net
fitness of genotypes for any locus in any environment in nature” [250, p. 236].
5
In evolutionary algorithms, this depends on the specific implementation of the
selection operator. Elitism, for instance, retains some of the fittest individuals in
each generation [308].
6
Even though most evolutionary algorithms have the Markov property, past ex-
periences could be implicitly captured in an appropriate definition of the fitness
function.
7
The application of “Darwinism” to economics and the use of biological analogies,
however, is vehemently defended by Hodgson [181]. He finds that most criticisms
of biological analogies are unfounded and that Darwinism involves a basic philo-
34 4 Learning in Economics

4.4.1 Learning Through Replicator Dynamics

In the marginalism debate, Alchian [3] and Friedman [140] suggested

that irrational behavior would be driven out of the market. In a way,
they must have considered markets to act as efficient evaluation and
selection mechanisms. In modern evolutionary economics, their compet-
itive argument could be modeled through what is now known as repli-
cator equations8 [383]. Replicator equations are based on Ronald A.
Fisher’s [128] Fundamental Theorem of Natural Selection (also known
as Fisher’s Law) which formalized the notion of the survival of the
fittest in evolutionary biology.9 Consider a population of n competing
“species” which are characterized by their genetic fitness f . In evolu-
tionary economics the “species” may simply be taken as a synonym for
competing technologies, strategies, or trading rules. In the marginal-
ism debate, the equivalent to a species is a group of firms exhibiting
the same level of rationality. The higher their degree of rationality, the
higher their fitness level. The overall population of firms  is described
n
by a vector of relative frequencies x = (x1 , x2 , . . . , xn ), i=1 xi = 1
with which these types of firms exist. For each firm i, its next period
share within the total population calculates as
fi (t)
xi (t + 1) = xi (t) (4.2)
f (t)
with
n


f (t) = xj (t)fj
j=1

sophical commitment to detailed, cumulative, and causal explanations. While

Darwinism has by now made substantial inroads into economics, more than 100
years earlier, Thorstein Veblen [429] expressed dismay that economics had not
become an evolutionary science yet.
8
The term replicator equations for Fisher’s selection equation was suggested by
Schuster and Sigmund [383], who pointed out the close relationship to the concept
of replicators which Richard Dawkins [90] proposed in 1976. Replicators refer to
a collection of molecules capable of either producing a copy of themselves, or
replicating by using raw materials either from themselves or from other sources.
Schuster and Sigmund also showed that Fisher’s selection equation is a special
case of the famous predator-prey equations ( Lotka-Volterra equations).
9
The term “survival of the fittest” had originally been coined by Herbert Spencer,
who anticipated many of Darwin’s evolutionary ideas. Darwin, who first used
the expression “struggle for existence”, adopted this expression only later, i.e.,
in his sixth edition of “The Origin of Species”. Spencer became known as the
father of “social Darwinism”, which is widely seen today as a misapplication of
evolutionary ideas to the social realm.
 4.4 Biologically Inspired Learning Models 35

being the average fitness of the whole population.10 The evolutionary

process described by this kind of replicator dynamics (4.2) maximizes
the average fitness of the population since firms with above average
fitness prosper, while the share of below average fitness firms shrinks.
In the limit, only the most rational type of firm in the initial population
survives.
However, this result holds only for constant fitness functions, i.e., if
fitness is independent of the frequency distribution in the population
and not otherwise influenced by a changing environment. Most often,
and possibly in the marginalism debate as well, fitness is time and
frequency-dependent and equation 4.2 should be rewritten as

fi (x, t)
xi (t + 1) = xi (t) , i = 1, . . . , n. (4.3)
f (x, t)
In this case, multiple equilibria are possible, and the mean fitness of the
population may or may not increase [234]. That is, under the conditions
likely to exist in competitive markets, the formalization of Friedman’s
argument with replicator dynamics shows that evolutionary market
forces do not ensure emerging rationality as an inevitable result.11
Replicator equations are often used to model social learning in which
agents learn from others by imitating other, more successful agents
[377, 378]. However, replicator equations predetermine the benchmark
level of efficiency to that of the most efficient firm contained in the
initial population [102]. Since replicator dynamics account for selection
only and contain no element of discovery or invention, a superior be-
havior missing in the initial population of behavioral rules cannot be
found. Novelty can be included with the so-called selection-mutation
equation, sometimes also referred to as the Fisher-Eigen equation.12
This is an extension of simple replicator dynamics by which novelty
is accounted for by adding a mutation term to equation 4.2.13 Since
replicator dynamics and the selection-mutation equation allow for a

10
The continuous time
 version
 of the replicator equation is more widely known and
reads as ẋi = xi fi (t) − f¯ . However, since computer simulations can only evolve
in discrete time steps, the discrete versions of any formula will usually be given.
11
In a more detailed analysis, Blume and Easley show on the one hand that market
selection favors profit maximizing firms. The long-run behavior of evolutionary
market models, however, is not consistent with equilibrium models based on the
profit-maximization hypothesis [43].
12
This equation is named after two evolutionary biologists, Ronald A. Fisher from
Great Britain, and Manfred Eigen from Germany.
13
More details are given in [52].
36 4 Learning in Economics

rigorous mathematical treatment, its domain is mainly in evolution-

ary game theory. In simulation based approaches, genetic algorithms,
classifier systems, and neural networks prevail.

4.4.2 Learning Through Genetic Algorithms

Genetic algorithms (GAs) were developed in the 1970ies by John Hol-

land [182] at about the same time that evolutionary strategies were
designed by Ingo Rechenberg [351]. The main difference between these
two approaches is that GAs were designed to work on binary coded solu-
tions (so-called genetic individuals) while evolutionary strategies work
directly with real numbers. While the issue of binary versus real-valued
representation caused a heated debate between Holland and Rechen-
berg1973, this distinction currently seems to be vanishing. As we will
see, the GA in the SFI-ASM uses a combination of binary represen-
tation for the condition parts and a real valued representation for the
forecast parameters. Brenner considers it a puzzle why modelers have
not moved to use evolutionary strategies, which, he claims, seem more
suited for economic problems [53]. The reason for this might be bet-
ter understood from a historical perspective, possibly reflecting a path
dependency. Classic textbooks on GAs have been written by Goldberg
[154] and Mitchell [308], while Dawid [89] and Riechmann [356] have
written newer textbooks with a special focus on learning in an economic
context.
The basic element in a GA is a population or gene pool consisting of
different genetic individuals. Since the gene pool in GAs most often con-
tains only haploid genetic individuals, it has become commonplace in
the GA literature to refer to these genetic individuals as chromosomes
or, less often, as genomes or genotypes.14 Chromosomes are made up
of genes and the position i of a gene on a chromosome is known as its
locus. A binary gene has two allele values, for instance, 0 and 1. The
cardinality of a gene refers to the number of different alleles that a gene
can take. A gene with cardinality 3 has three alleles, e.g., 0, 1, and 2.
14
In population genetics , one usually distinguishes between haploid, diploid, or
polyploid species. Haploid species have only one copy of each chromosome while
diploid species, the most common case in nature, possess two copies of each chro-
mosome. For haploid individuals with only one chromosome, the terms “genetic
individual”, “chromosome”, “genotype”, and “genome” coincide, yet one should
be aware that they have slightly different meanings for diploid and polyploid
species in evolutionary biology. For a definition of these terms in the context of
population genetics see, for instance, [174, 80]. In the following, the terms genetic
individual, chromosome, genotype, and genome will be used interchangeably.
 4.4 Biologically Inspired Learning Models 37

In most GA representations, all genes have the same cardinality and a

common allele set.
GAs are inherently parallel. Since a GA has to work on a population
of several genetic individuals, in each time step several new solutions
are simultaneously created and then tested for their performance. Little
synchronization is needed and parallelization by means of distributing
the numeric calculations onto multiple processors can be easily accom-
plished.
By now, there have been many different variants of GAs developed,
some of which are are discussed in [112]. Most of them descended from
the so-called canonical GA developed by Holland [182] and described
in [154]. The basic structure of the canonical GA is as shown in figure
4.1.

BEGIN
randomly create initial population P0
WHILE NOT stopping condition DO
t := t + 1
Evaluation of Pt−1
Selection from Pt−1 and Reproduction into Pt
Recombination (Crossover) on Pt
Mutation on Pt
END
END
Fig. 4.1. Basic structure of the canonical GA.

The canonical GA is a non-overlapping populations approach in

the sense that the offspring generation created through selective repro-
duction, mutation, and crossover completely replaces the parent gen-
eration. In the canonical GA, the selection probability Πi of genetic
individual i to be reproduced into the next generation is proportional
to its relative fitness
fi
Πi = N , (4.4)
j=1 fj

N being the population size. In other words, i’s expected number of

offspring equals its relative fitness in the population.
In contrast to the canonical GA, generation gap methods implement
an overlapping generations such that the parent and offspring gener-
ations compete with each other.15 The “generation gap” parameter
15
Generation gap methods are discussed in [373].
38 4 Learning in Economics

G refers to the amount of overlap between parents and offspring. For

G = 1, we have a complete generational replacement as in the canoni-
cal GA. A generation gap of G = 0.2 would replace 20% of the parent
generation, while the so-called “steady-state GA” usually replaces only
one or two individuals, i.e., G = 1/N or G = 2/N .

Social Versus Individual GA Learning

There is an ongoing debate about whether GAs represent learning on
the social or individual level. Riechmann, for instance, conceives of GA
learning only as a way of social learning. “As social learning always
means learning from others, there simply is no GA learning by single,
isolated agents” [354, p. 226]. Whether GA learning corresponds to
social or individual learning depends on the way it is implemented.
When constructed as social learning, each agent is represented as one
solution string. The population of agents and the population of possible
solutions coincide in this case. It is also popular to model social learning
by having a common pool of behavioral strategies to which all agents
have equal access [241]. When all agents are equipped with a rule set of
their own, and agents do not communicate about their favorite rules,
GA learning takes place only on the individual level [9, 244]. Vriend
[431] has demonstrated in a standard Cournot oligopoly game that
social or individual GA learning may yield sharply different results.
He showed that profit maximizing firms that used a social GA learning
algorithm converged to the Cournot-Nash output level while individual
learning moved the firms to the competitive Walrasian output level.
The reason for this divergence is the so-called spite effect in social
learning, i.e., agents may choose an action because it hurts others more
than it hurts themselves. In spite of having the same objective function,
social learning firms compare their single production rule to those of
other firms, all of which are active in the same period. Production rules
of individual learners, on the other hand, compete with each other in
the learning process, but not in the Cournot market. Since an individual
firm uses only one of its rules in any given period, the generated payoff
of that rule is not influenced by the payoffs of other production rules
in the firm’s rule set. While there is still a spite effect in the market, it
does not affect the learning process.
When employing learning algorithms, one has to be aware of whether
the aggregate outcome can be a result of a possible spite effect. The de-
cision about whether one should choose an individual or social learning
algorithm is not a matter of convention or computational convenience,
but should depend on the specific problem to be solved. Vriend also
 4.4 Biologically Inspired Learning Models 39

points the possibility of hybrid approaches in which elements of both

learning variants are combined.16

Binary Encoding and Decoding

Historically, the binary encoding of objects as genetic individuals has

been the predominant approach in GAs. Consider a binary chromosome
of length L that represents a possible action or solution to an optimiza-
tion problem. Such an action might refer to quantities such as how much
to produce or to invest. This binary representation for the GA and the
use of real-valued numbers outside the GA requires a constant encoding
and decoding of the object variables.17 The following is a short example
of how to decode a binary chromosome. Let us assume that a choice
variable can take on values in the interval [Dmin = −10, Dmax = 40].
Consider, for instance, a chromosome i with a length of 16 bits

Si = { 0 1 1 0 1 0 1 0 1 1 0 0 1 0 1 1}.

A direct transformation into the decimal system would result in a nu-

merical value of

B(Si ) = 214 + 213 + 211 + 29 + 27 + 26 + 23 + 21 + 20 = 27, 339.

The numerical decimal value of i is then determined as

Dmax − Dmin
V (Si ) = Dmin + × B(Si ) (4.5)
2L − 1
60
= −20 + × 27, 339
65, 535
≈ 5.03.

The maximum attainable precision for an object variable is given by

Dmax − Dmin
. (4.6)
2L − 1
To achieve higher numerical precision, longer binary chromosomes have
to be used.

16
A combination of social and individual learning would correspond to migration
between different populations. In population genetics this is analyzed with so-
called island or stepping-stone models [126].
17
Often, the adherents of the competing school of evolutionary strategies criticized
the necessity of constantly coding and decoding the choice variables [351].
40 4 Learning in Economics

Instead of the pure binary coding as described above, Gray codes18

are frequently used in GA applications. Gray coded numbers are char-
acterized by the “adjacency property”, i.e., adjacent integers differ only
in one single bit and have a Hamming distance of one. In pure binary
representations, adjacent integers may lie many bit flips apart, which
makes it less likely for a single bit flip to cause only small changes. Holl-
stien found that a GA with Gray code representation worked slightly
better for optimizing functions than a pure binary encoding [188]. He
attributed this to the adjacency property since it improves the chances
that mutation, i.e., the flipping of a single bit, causes only incremen-
tal improvements. Although fewer mutations may cause large changes,
those which do may cause even bigger changes than the flipping of the
most significant bit in a pure binary representation.19

Evaluation

The definition of an objective function f : Ax → R, Ax being the

state space of the optimization problem, is perhaps the most impor-
tant element in any evolutionary algorithm. It evaluates the quality of
genetic individuals and thus, specifies the direction towards which the
GA should try to improve them. Generally, this definition is dependent
on the specific optimization problem. In an economic context, usually
some costs are to be minimized or profits to be maximized. If the ob-
jective function produces negative values or has to be minimized, it has
to be mapped to a fitness function Φ : Ax → R+ . Often, an additional
scaling function is applied.20
It is sometimes helpful to imagine the objective function as defin-
ing a fitness landscape over the underlying state space. Adaptation
could then be visualized as a process of “hill-climbing” through minor
modifications towards “peaks” of high fitness on this fitness landscape
[215]. Flat fitness landscapes with a single peak, so-called “needle-in-
a-haystack” landscapes, or random fitness landscapes, are notoriously
hard to solve, yet they are often used to analyze the performance of
different evolutionary algorithms.21
18
Gray codes are named after Frank Gray, who developed them in 1953 for use in
shaft encoders. More on the history and the use of Gray codes can be found in
[173, 144, 343].
19
An overview of different encoding and decoding functions for GAs is given in [94].
20
More information on fitness scaling and possible mappings of objective values
into fitness functions can be found in [162].
21
The analogy between the fitness function and a landscape is due to Sewall Wright
who considered forces other than selection in his shifting balance theory, in par-
 4.4 Biologically Inspired Learning Models 41

Unlike with simple function optimization or many combinatorial

problems such as the travelling salesman problem, many GAs work
with state dependent fitness functions.22 Riechmann [355] defines such
GAs as economic GAs. A state dependent fitness function means that
the fitness of an agent does not only depend on his strategy, but also
on the strategies of all other agents.23 Actually, state dependent fitness
functions are even more common in biological evolution where species
are constantly adapting to each other. Stuart Kauffman , for instance,
speaks of coupled fitness landscapes and species that constantly coe-
volve [215]. Any step of one species on its fitness landscape deforms the
landscape for another species which will react accordingly. This is also
known as an evolutionary arms race.

Selection and Reproduction

The genetic operator of selection chooses genetic individuals for repro-

duction, i.e., chosen solutions will create identical offspring who become
members of the next population. Conceptually, selection for reproduc-
tion can be divided into two different steps. First, a selection algorithm
assigns each individual a real-valued expected reproduction value, i.e.,
its expected number of offspring. The distribution of expected offspring
could be visualized as a “wheel of fortune” where the selection algo-
rithm determines the size of the section assigned to each individual. In
a second step, a sampling algorithm chooses an integer number of ac-
tual offspring for each parent. The two most used sampling algorithms
are roulette wheel (RW) and stochastic universal sampling (SUS).
There are numerous selection schemes. One of the most popular
selection schemes is proportional selection [182] where an individual’s
i selection probability ΠRW (i) is proportionate to its relative fitness,
i.e.,
Φ(i)
ΠRW (i) = N , (4.7)
j=1 Φ(j)

N being the population size. Since fitness proportionate selection re-

quires positive fitness values, it is sometimes necessary to add a positive

ticular, genetic drift, as a means of moving a population across valleys in the

fitness surface [442].
22
GAs with state dependent functions are analyzed in more detail in [89].
23
Riechmann [356] also points out that the term “strategy” is used differently in
game theory and in the theory of economic GA learning. Whereas in game theory,
“strategy” refers to a set of contingent actions, it might refer simply to any myopic
action in GA learning.
42 4 Learning in Economics

constant C to all raw fitness values Φ(i). This, however, causes a prob-
lem since simple fitness proportionate selection is not scaling invariant,
i.e., adding an identical offset C tends to equalize the selection proba-
bilities.

Table 4.1. The property of scaling invariance for fitness proportionate selec-
tion means that adding an identical offset—in this case 100 in column four—to
the raw fitness values Φ(i) of three genetic individuals almost equalizes their
selection probabilities ΠRW (i).

Roulette Wheel 1 Roulette Wheel 2

ind. Φ(i) ΠRW (i) (%) Φ(i) + 100 ΠRW (i) (%)
1 0.01 0.93 100.01 33.22
2 0.07 6.54 100.07 33.24
3 0.99 92.52 100.99 33.54

The problem of scaling invariance for fitness proportionate selection

is illustrated in table 4.1. For the second roulette wheel in columns
four and five, an offset of 100 is added to the raw fitness values given
in column two. When using the raw fitness values, the worst genetic
individual has a very small selection probability of less than 1 %, while
the best individual is chosen almost 93 % of the time. When an offset
of 100 is added, the selection probabilities are 33.22 % and 33.54 %,
i.e., the worst and the best genetic individual have almost an equal
probability of being selected. Various approaches to dealing with scal-
ing invariance are discussed in [154, 308]. Sigma scaling, for instance,
remedies the problem by also taking into account the mean and stan-
dard deviation of fitness values in a population.
In contrast to fitness proportionate selection, the selection probabili-
ties in ranking-based selection [163] depend only on the rank ordering of
the individuals in the current population. Ranking eliminates the need
for fitness scaling, since the selection intensity is maintained even if the
 4.4 Biologically Inspired Learning Models 43

fitness values converge to a very narrow range, as is often the case as

the population evolves. Other popular selection procedures are tourna-
ment selection [38], and Boltzmann selection [271]. Additional selection
mechanism are discussed in [131]. Most of these selection schemes are
analyzed and compared in [39, 167].

Recombination (Crossover)
Since crossover involves the exchange of genetic material between two
parents, it is often called a sexual genetic operator. By selecting two
chromosomes from the mating pool with high fitness values as parents,
genetic material from both chromosomes is exchanged to create two
new offspring. In the canonical GA, the offspring replace their parents,
and simple one point crossover is used.

parent 1 parent 2
0 1 0 0 0 1 1 0 0 1 1 0 1 1 1 0 0 1 1 0 1 1 0 1 1 1 0 0 0 1 0 0

0 1 0 0 0 1 0 1 1 1 0 0 0 1 0 0 0 1 1 0 1 1 1 0 0 1 1 0 1 1 1 0
offspring 1 offspring 2

crossover point crossover point

Fig. 4.2. Example of one-point crossover with a random break point after
the fifth bit.

In one point crossover, a break point between 1 and L−1 is randomly

chosen. Contiguous bit segments that begin or end at the break point
are exchanged between the two parents. One-point crossover, however,
is rarely used in practice [47].

parent 1 parent 2
0 1 0 0 0 1 1 0 0 1 1 0 1 1 1 0 0 1 1 0 1 1 0 1 1 1 0 0 0 1 0 0

crossover bit mask crossover bit mask

0 0 1 0 0 1 0 1 1 1 0 1 1 0 1 0 0 0 1 0 0 1 0 1 1 1 0 1 1 0 1 0

0 1 1 0 0 1 1 1 1 1 1 0 0 1 0 0 0 1 0 0 1 1 0 0 0 1 0 0 1 1 1 0
offspring 1 offspring 2

Fig. 4.3. Uniform crossover with a randomly generated crossover mask.

The class of k-point crossover operators uses several break points

while the uniform crossover operator sets k at its maximum L − 1.
44 4 Learning in Economics

Uniform crossover has been shown to be superior in most cases to one-

point and two-point crossover [415]. In a first step, a binary mask of
length L is randomly created. Whenever the value in this crossover
mask is 0, the corresponding bit from parent one is copied to the first
offspring, and the second offspring receives the corresponding bit from
parent two. For value 1 in the crossover mask, the bit exchange is re-
versed. Other crossover operators are, for instance, shuffle crossover,
and punctuated crossover. For the travelling salesman problem, spe-
cial crossover operators such as PMX (partially mapped crossover) or
OMX (ordinal mapped crossover) have been designed. These and other
crossover operators are discussed in [182, 154, 47].
In a social learning context, recombination is often interpreted as
communication or information exchange between agents. If modeled as
individual learning, crossover might be seen as individuals experiment-
ing by combining some of their successful strategies.

Mutation

Mutation refers to the creation of a new solution from only one par-
ent through spontaneous allele changes. It is an important element of
any evolutionary algorithm to prevent premature convergence. Unlike
crossover, mutation is able to introduce non-existing alleles at previ-
ously homogeneous bit positions. If, for instance, the most significant
bit is set in none of the trading rules, yet the optimal solution would
require it to be set, a GA without mutation could never reach the op-
timum. Mutation aims at maintaining a sufficiently diverse population
since without it, the final outcome of a GA is a uniform population
with no diversity at all.
In mutation, each parent bit is flipped with mutation probability
ΠM . Mutation probabilities are usually chosen to be very low, usu-
ally in the magnitude of 10−2 − 10−3 . Originally intended only as a
background operator [154], recent empirical and theoretical research
demonstrates improvements in the GA when strengthening the role of
mutation as a search operator [17]. Fogarty [130] empirically finds that
GAs with initially bigger mutation rates which exponentially decreases
over time do better than fixed mutation probabilities, a finding that is
theoretically confirmed by Bäck [16].
As with crossover, a mutated offspring replaces its parent in the
canonical GA. In both social and individual learning contexts, muta-
tion is usually interpreted as a metaphor for either experimentation or
discovery through unintended mistakes.
 4.4 Biologically Inspired Learning Models 45

The Schema Theorem and the Building Block Hypothesis

The Schema Theorem and the Building Block Hypothesis developed by

Holland [182] are attempts to theoretically explain why GAs work in
practice. For a binary representation, a schema ξ is a ternary template
consisting of the symbols 0, 1, and #, where # is known as the don’t
care symbol.24 The defining positions of a schema are all non-# signs,
and its order o(ξ) is the number of all its defining positions. The dis-
tance between the first and last defining position of schema ξ is called
its defining length δ(ξ), while members of a schema are called instances.
For instance, the schema ξ = {1##010} is of order four and has a defin-
ing length of five. Its instances are {100010, 101010, 110010, 111010}.
The Schema Theorem, sometimes also called the Fundamental The-
orem of GAs [154], derives a lower bound for the number of instances of
a schema in the next generation, given the number of instances in the
current population Nξ (t), fitness proportionate selection, and complete
generational replacement:

fξ (t)
Nξ (t + 1)|Nξ (t) ≥ Nξ (t) [1 − Dc (ξ)][1 − Dm (ξ)]. (4.8)
f (t)
The observed fitness fξ (t) is given by the mean fitness of all instances
belonging to schema ξ in generation t, while f is the mean fitness of the
whole population at t. Dc (ξ) and Dm (ξ) denote upper bounds for the
disruptive effects on schema membership by crossover and mutation.25
These disruptive effects have to be analyzed specifically for each opera-
tor. For instance, one-point crossover with crossover probability Πc can
disrupt the schema membership of an offspring only if its cross point
falls within the defining region of a schema. For chromosomes of length
L, the upper bound for the disruptive effect of one-point crossover on
schema membership is then given by Πc δ(ξ)/(L − 1). Similarly, for mu-
tation in which each gene is altered with mutation probability Πm ,
schema membership is only disrupted if the mutation operator alters
at least one of the defining positions of the chromosome. Since, by def-
inition, the probability that none of the defining positions of a schema
is altered is (1 − Πm )o(ξ) , the membership disruption Dm (ξ) caused by
24
A more general analysis for alphabets with higher cardinality can be found in
[347].
25
Note that if we neglect these disruptive terms, i.e., if we have a GA without
crossover and mutation, the evolution of the GA population is completely driven
by the dynamics of the selection operator. The inequality equation 4.8 then re-
duces to the already known replicator dynamics equation 4.3 on page 35.
46 4 Learning in Economics

mutation is 1 − (1 − Πm )o(ξ) . For Πm ≪ 1, this is approximated by

Πm o(ξ). That is, for one point crossover and point mutation, equation
4.8 can be written as

 
fξ (t) δ(ξ)
Nξ (t + 1)|Nξ (t) ≥ Nξ (t) 1 − Πc [1 − Πm o(ξ)] . (4.9)
f (t) L−1
The Schema Theorem thus shows how low order schemata with
small defining length and above average fitness will quickly propa-
gate through a population. Holland and Goldberg [182, 154] call these
schemata building blocks, which, through recombination, are likely to
form more complex solutions with potentially higher fitness. They refer
to this conjecture as the Building Block Hypothesis. However, while the
Schema Theorem is easy to prove, Radcliffe doubts that the Building
Block Hypothesis can ever be proved [347].
Overall, the significance of the Schema Theorem and of the Building
Block Hypothesis remains highly debated [352, 92]. One of the short-
comings of the schema theorem is, for instance, that it only considers
the disruptive effects of crossover and mutation [89]. In addition, the
notion of implicit parallelism [182], i.e., the idea that GAs can pro-
cess many different schemata at once, and its relation to the cardinal-
ity of the representation remains highly disputed. Implicit parallelism
and the supposed optimality of binary representation is also discussed
in [154, 347]. An alternative set of principles for understanding the
working of a GA—the evolutionary progress principle, the genetic re-
pair hypothesis, and the mutation-induced speciation by recombination
principle—is suggested by Beyer [31].

4.4.3 Learning Through Classifier Systems

Classifier systems are machine learning systems that derive their name
from their ability to actively classify their environment into recurrent
patterns. Classifier systems in the tradition of the Michigan-approach
were developed by John Holland and Judy Reitman at the University of
Michigan [187], while those in the tradition of the Pitt-approach were
developed by Smith at the University of Pittsburgh [403].26 Both are
a synthesis of expert systems as described in [301, 434] and Holland’s
genetic algorithms. In contrast to expert systems, though, classifier
26
For a comparison of both styles, see [402]. The following description is based on
Michigan-style classifier systems since they are more popular and have undergone
more development.
 4.4 Biologically Inspired Learning Models 47

systems do not need to be initialized with a fixed rule for every possible
situation. Instead, classifier systems are flexible in that a GA adapts
the rule sets to a changing environment. Today, it is common to speak
of learning classifier systems (LCS) instead of classifier systems. The
general structure of LCS and their interactions with the environment
are shown in figure 4.4.

Environment

5 1 6

action

Effectors Detectors Credit Allocation

(e.g., bucket brigade)

Message Board

internal messages external messages rewards

1.65 0101011100

matching
4

Conflict Resolution
2 Classifier List
(rule selection)
condition part action part strength active
{01##01###0,0.95,1.25} 01##01###0 0.95 1.25 ¥
or 3 ##1100##10 2.25 0.78 –
{0##101##00,1.65,1.33} 01##00#100 3.50 0.33 –
?? : : : :
0##101##00 1.65 1.33 ¥

LCS Genetic Algorithm

Fig. 4.4. Structure of a Learning Classifier System (LCS).

The state of the environment is sent to the LCS with a set of detec-
tors. These detectors post standardized messages about the environ-
ment to a message board with which a list of classifiers are matched.
Classifiers are condition/action rules of the form
48 4 Learning in Economics

if (condition(s) fulfilled), then (take action).

Because of this particular structure, LCS are also called “condition-
action” classifier systems [185].
The environment is often described with a binary representation, yet
higher cardinality alphabets are also possible [402]. For instance, the
first bit of the message {01100} could mean that an asset price is below
its long-term average while the second bit indicates that last period’s
trading volume exceeded a certain threshold level. Similarly, the last
three bits are Boolean descriptors of other pre-defined environmental
conditions.
For such a binary representation of the environment, the condition
part of a classifier is a ternary list {0, 1, #}, with # being a wildcard
sign which indicates that the particular condition encoded by the bit
on that position is ignored. Matching occurs by a comparison of each
position between the condition part and the external messages issued by
the detectors. For instance, the condition parts {#11#0} and {0##00}
match the above message, but not the condition parts {1###0} or
{##0#1}. Detailed rules, i.e., those with only a few wildcard signs
in their condition part, are called specific. General rules with many
wildcard signs, on the other hand, match many different environmental
conditions. By favoring more specific classifiers over general ones, LCS
can implement a default hierarchy in which specific exemptions override
more general situations.27
While the rule base of an expert system is formulated such that each
possible situation is matched by only one rule, in LCS, several rules usu-
ally match and are marked as active. Activated classifiers compete with
each other for execution based on their strength, which is a function of
their past performance. Classifiers that have been proven to be highly
successful, for instance, by being profitable, have a greater chance of be-
ing activated than less successful classifiers. The chosen classifier posts
its message, for instance, a numeric value of how much of a commodity
to produce, sell, or buy, to the message board which will then be exe-
cuted by the effectors. Credit allocation is an environmental feedback
mechanism that rewards profitable classifiers and punishes failing ones
by increasing or decreasing their strength.
While credit allocation is a learning mechanism that identifies and
rewards good rules from the set of existing classifiers, a GA is an integral
part of any LCS to generate new classifiers. The GA is invoked on a
27
It is also possible, however, to favor general rules over specific rules by punishing
each bit other than # with some associated cost. Instead of default hierarchies,
one could then model an agent’s complexity aversion.
 4.4 Biologically Inspired Learning Models 49

regular basis to create new offspring through selection, crossover, and

mutation. In contrast to the canonical GA, which performs a complete
generational replacement of the parent population, the offspring in LCS
replace only some of the worst classifiers. The best solutions are thus,
usually preserved in LCS.
The credit allocation process as described above is not particulary
difficult since there are only single classifiers that are directly respon-
sible for past actions. The feedback provided in terms of reward or
punishment directly increases or decreases a classifier’s strength. Most
descriptions of LCS, however, refer to the bucket brigade algorithm
when explaining the apportionment of credit.
In order to understand the bucket brigade algorithm, it is necessary
to note that rules can be coupled to act sequentially. A rule is said to be
coupled (or chained) to another if its action part generates a message
that will potentially activate another rule in the next period. Coupled
rule sequences are crucial for handling complex situations, for devel-
oping plans and strategies, and for modeling causal relations in the
environment. Obviously, stage setting rules that contributed to prof-
itable future actions of other rules should be similarly rewarded than
the final rule which receives its reward directly from the environment.
The bucket brigade algorithm does exactly this. It derives its name
from an analogy to an old-fashioned fire brigade whose members are
handing buckets of water to each other. In the bucket brigade algo-
rithm, every suppliant classifier that activates another classifier is di-
rectly paid by the latter. The consuming classifier, i.e., an activated
and chosen classifier, pays its suppliant classifiers by ‘investing’, i.e.,
distributing its strength to all of its suppliant classifiers. If the action
turns out to be successful, its strength value will be replenished by the
environmental reward, otherwise not. In this way, the bucket brigade
algorithm strengthens verified classifier chains and weakens less suc-
cessful rule sequences.
More detailed descriptions of LCS and the bucket brigade can be
found in [185, 184].
5
Replicating the Stylized Facts of Financial
Markets

... to be able to imitate reality is a form of understanding ....

Benoit B. Mandelbrot and Richard L. Hudson [280, p. 19]

5.1 Efficient Markets and the Efficient Market

Hypothesis

The Efficient Market Hypothesis (EMH) has been the cornerstone of

finance for more than thirty years [392]. In 1978, Jensen claimed that
no other hypothesis in economics had more empirical support [201].
He also acknowledged, though, that as econometric sophistication in-
creased, more and more inconsistencies arose that could no longer be
ignored. The years following this statement saw the ascent of behavioral
economics, a new approach to financial economics that further amassed
anomalies in apparent contradiction to the EMH. The interpretation
of the empirical results, however, is often impeded by the fact that the
EMH may be expressed in a number of ways, some of which are not
equivalent and are differentiated only by subtle technicalities [245, 84].

5.1.1 Definitions

The EMH is often stated in Fama’s words, namely that a “market in

which prices always ‘fully reflect’ available information is called effi-
cient” [115, p. 383]. Since only news is not yet available and, by def-
inition, not predictable, the EMH holds that prices should randomly
fluctuate in the same way as the news finally materializes. By defin-
ing “available information”, Fama distinguishes three levels of market
efficiency. For the weak form version of the EMH, the information set
52 5 Replicating the Stylized Facts of Financial Markets

consists exclusively of past price and return information. The semi-

strong version additionally includes all publicly available information
at time t, e.g., balance sheets information, income statements, or an-
nouncements of dividend changes. Finally, the strong form version of
the EMH asserts that all public and private information is fully re-
flected in market prices. In strong form efficient markets, even insiders
would be unable to make an excess profit from their private knowledge.1
Fama’s original definition, however, neglects the costs of information
acquisition. If information acquisition is costly, perfect informational
market efficiency is impossible to achieve, as Grossman and Stiglitz
show [164]. If markets were perfectly efficient, informed traders could
not earn a return on their investment in information acquisition, and
markets would eventually collapse.2 In the sequel to the original article,
Fama [116] adopts a weaker, but more realistic definition of market
efficiency suggested by Jensen [201]. A market is efficient with respect
to information set Φ if it is impossible to make economic profits by
trading on the basis of that information set. Economic profits are risk-
adjusted returns net of all costs. That is, prices reflect information
only to the point where the marginal benefits of acquiring information
do not exceed the marginal costs of doing so. Informational efficiency
is a benchmark which is impossible to achieve under the presence of
information costs. Markets should, rather, be thought of as possessing
an equilibrium degree of inefficiency.
When discussing market efficiency, financial economists usually re-
fer to informational efficiency. Yet this begs the question whether in-
formational efficiency implies a Pareto-efficient resource allocation in
the economy. Cuthbertson claims that if stock prices always fully re-
flect fundamental values, then the market will optimally allocate funds
among competing firms [84]. Stiglitz [411], on the other hand, seriously
challenges this view with a variety of arguments. He shows that the view
of efficiency in the finance literature is neither necessary nor sufficient
for Pareto optimality in the economy. The recent stock market bubble

1
In his 1991 article [116], however, Fama relabels the taxonomy of information sets
which he developed in 1970. The first category is now called “tests for return pre-
dictability”. Semi-strong tests of market efficiency are relabeled as event studies,
while for strong-form tests, he suggests the term “tests for private information”.
2
A critique and solution to Grossman and Stiglitz’s information paradox is given by
Helliwg [176]. In an informationally efficient market, investors who do not engage
in costly information acquisition derive information from the market price even
before any transaction has taken place. In a dynamic setting, however, investors
who actively acquire information are able to use it before it is revealed through
prices.
 5.1 Efficient Markets and the Efficient Market Hypothesis 53

in the 1990’s with excessive investments in “new economy” firms, fol-

lowed by significant underinvestment thereafter, is just another indica-
tion that the stock market does not always allocate resources efficiently.
Since the EMH is perfectly suited to rationalize even “irrationally ex-
uberant” asset prices, widespread belief in it may even obstruct the
resource allocation function of capital markets.
A recent refinement of the notion of market efficiency deserves men-
tion at this point though. Paul Samuelson [366] proposed a dictum for
the stock market in which he argued that markets exhibit considerable
micro efficiency while being macro inefficient as a whole. He considers
the EMH to be more applicable for individual stocks than for stock
market indices. Jung and Shiller [210] tested this hypothesis by analyz-
ing dividend-price ratios for all 49 U.S. firms that existed in the period
from 1926–2001. They do so by running a regression of future multi-year
dividend changes on current dividend-price ratios and testing whether
the dividend-price ratio predicts these changes. Efficient market theory
predicts a negative slope of minus one when plotting dividend growth
over dividend-price ratios. Jung and Shiller find a negative coefficient
for their regression analysis which supports Samuelson’s idea of effi-
ciency on the level of individual stocks. Firms with lower dividend-price
ratios did indeed have higher subsequent dividend growth. When con-
structing a stock market index from the 49 stock and doing the same
regression, the regression coefficient becomes positive, which indicates
that the EMH is less appropriate for the stock market as a whole.3

5.1.2 Random Walks or Martingales?

The idea that security prices fluctuate randomly was formulated as
early as 1900 by Louis Bachelier in his doctoral thesis “Théorie de
la spéculation” [15].4 In it, he developed the idea that the prices of
French government bonds resembled a random walk.5 A random walk
price series could be described by
3
Samuelson’s notion of micro efficiency and macro inefficiency of the stock market
adds another twist to the discussion on aggregation issues. Often, the argument
is put forth that irrationalities on the micro-level would cancel each other out,
thus, causing aggregate behavior that is consistent with the assumption of a fully
rational representative agent. Interestingly, the direction in Samuelson’s argument
is reversed, i.e., the amount of rationality and efficiency is higher on the micro-
level.
4
An English excerpt of Bachelier’s dissertation can be found in [82]. An introduc-
tion to his life and work is given in [416].
5
Bachelier, however, did not use the term “random walk”. Probably the first time
the notion of a random walk emerged was in a written conversation between Karl
54 5 Replicating the Stylized Facts of Financial Markets

pt+1 = pt + ǫt (5.1)

with ǫt ∼ iid(0, σǫ2 ), i.e., a sequence of independently and identically

distributed random disturbances with zero mean and equal variances.6
Bachelier’s results did not receive any attention until an increasing
number of studies in the 1950’s, some of which are reprinted in [82],
rediscovered that the behavior of various asset prices can be approxi-
mated by the random walk model.7
The first rigorous arguments explaining why security prices fluctu-
ate randomly were given by Paul Samuelson [365] and Benoit Man-
delbrot [277]. They showed that in competitive markets with rational
risk-neutral investors, asset prices follow martingales. A stochastic dis-
counted price process pt would satisfy the martingale property if

Et [p̃t+1 |Φt ] = pt (5.2)

holds, Φt being a given information set.8 The tilde sign ˜ is used to

indicate random variables which are not yet realized in period t. Ac-
cording to 5.2, the best predictor of next period’s asset price is simply
today’s price. The martingale model implies that (p̃t+1 − pt ) is a “fair
game”, i.e.,
Et [(p̃t+1 − pt )|Φt ] = 0. (5.3)
Equation 5.3 says that increments in value, i.e., price changes adjusted
for dividend income, are unpredictable, given the information set Φt .
This means that all available information is fully reflected in current

Pearson and Lord Rayleigh in 1905 in the journal Nature [333, 349]. These three
letters can be downloaded from http://www.e-m-h.org/Pear05.pdf.
6
Fama [115] notes that the terminology in finance is loose at this point. First
of all, expected price changes can be non-zero such that we have a “random
walk with drift”. Secondly, if one period returns are iid, prices will not follow a
random walk since the distribution of price changes is dependent on the price
level. Third, instead of simple white noise ǫt ∼ iid(0, σǫ2 ) disturbances, Gaussian
white noise ǫt ∼ iidN (0, σǫ2 ) is often assumed [59]. Besides these variations, three
versions of the random walk hypothesis are identified in the finance literature: The
“independently and identically distributed returns” version, the “independent
returns” version, and the “uncorrelated returns” version [65].
7
For a more detailed account of the early history of efficient markets, see [115, 245,
100].
8
It is common in the EMH literature to speak of prices as following martingales.
We will adopt this somewhat imprecise but convenient usage by specifying that
prices should be understood to include reinvested dividends. Since prices can
easily be converted to returns and vice versa, equation 5.2 could also be written as
Et [p̃t+1 |Φt ] = pt [1 + E(r̃t+1 |Φt )], with r̃t+1 = (p̃t+1 − pt )/pt being next period’s
return of the asset.
 5.1 Efficient Markets and the Efficient Market Hypothesis 55

prices, and that it is impossible to make excess profit by trading on the

basis of Φt .
The random walk model of speculative asset prices is only a spe-
cial case of the more general class of martingale models. Contrary to
the random walk case, martingale models do not require the stochas-
tic disturbances ǫt to be iid, and are, thus, less restrictive.9 Many of
the stylized facts of financial asset prices such as volatility clustering,
fat tails, and other deviations in the return distribution from a normal
distribution are, thus, again compatible with the notion of an efficient
market.10 By assuming the less restrictive martingale model, these styl-
ized facts can be seen as a mirror image of the statistical properties of
the unobservable news arrival process.

5.1.3 Tests for Market Efficiency

Market efficiency is often equated with the assumption of rational ex-

pectations and the martingale property. A litmus test for market effi-
ciency is whether it can be shown that consistently higher risk-adjusted
returns than a simple buy-and-hold of the market portfolio can be
earned [362]. Roll [359], however, observed that it is extremely difficult
to profit from even the most severe violations of market efficiency.
Dependent on the size of the information set, several types of tests
for market efficiency have been developed. The first weak-form effi-
ciency tests primarily used past return data [216, 114] before expand-
ing their scope to other publicly available information such as divi-
dend yields, earning-price ratios, or financial accounting data [118].
The semi-strong form of the EMH is often tested with event studies
which measure the adjustment speed of asset prices to new informa-
tion [117]. Event studies average the cumulative performance of stocks
9
Equation 5.2 can equivalently be rewritten as pt+1 = pt +ǫt , which looks identical
to the random walk model in equation 5.1. The only difference lies in the more
restrictive assumptions for the random shocks ǫt in the random walk case. While
the random walk model implies that the first four moments of the price or return
distributions—mean, variance, skewness , and kurtosis—are independent of Φt ,
the martingale property only states that the mean of the price distribution is
independent of the information set.
10
Violations of the Gaussian price increments assumption discredit the modern asset
pricing models or Black and Scholes’ options pricing formula [37] more than the
EMH, since the former all rely on the standard deviation as a well defined measure
of risk. If asset returns were to follow stable Paretian distributions with no well-
defined risk measure [276], the risk-return calculations in modern asset pricing
models would become obsolete. Even Mandelbrot and Hudson [280] acknowledge
that the martingale property is usually not contradicted by empirical evidence.
56 5 Replicating the Stylized Facts of Financial Markets

over a time windows covering a specified number of periods before and

after an event. The cumulative abnormal stock returns are obtained by
adjusting for market-wide movements in security prices. Ball finds, for
instance, that the risk-adjusted abnormal returns after the announce-
ment of an earning announcement are systematically non-zero which
would be inconsistent with market efficiency [21].
In one of the first tests for the profitability of private information,
Jaffe rejects strong-form market efficiency [196]. When surveying more
recent contributions in this field, Fama [116] concludes that by now vi-
olations of strong market efficiency are well established, a finding that
has been confirmed for the German stock market [380]. Two classic
overviews over the different tests on market efficiency are the contri-
butions by Fama [115, 116]. A textbook on informationally efficient
capital markets is due to Sapusek [369] who also provides empirical
data on market efficiency for the German and Austrian stock market.
Even though the area of weak and semi-strong market efficiency is
highly contentious, the basic belief in the EMH is not shaken [100]. This
is demonstrated by the fact that deviations from it are still referred to
as “anomalies”.

5.2 Stylized Facts of Financial Markets

Much of our knowledge about the stylized facts of financial markets

stems from permanent attempts to either prove or disprove the EMH or
the classical asset pricing models that build on the EMH. Since it would
be hard to define which empirical phenomena qualify as a stylized facts,
an exhaustive description is impossible at this point. Instead, priority
will be given to those empirical regularities that can be addressed by
means of agent-based models of financial markets.

5.2.1 Non-Normal Return Distributions

At the time when the tenets of the EMH were formulated, logarithmic
asset returns r△t (t) = ln p(t + △t) − ln p(t) were usually thought to
be normally distributed. At the same time, however, Mandelbrot [276]
presented convincing evidence that this assumption cannot be upheld.
In his long-term analysis of cotton prices, he showed that for various
choices of △t, ranging from one day up to one month, extreme events
occur much more often than implied under the Gaussian hypothesis.
Subsequent empirical studies unanimously confirmed this finding of
leptokurtosis for virtually all financial asset return distributions, i.e.,
 5.2 Stylized Facts of Financial Markets 57
0,030

0,025
relative frequency

0,020

0,015

0,010

0,005

0,000
-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5
sigma

Daily DAX-Returns (1/1986-12/2005) Normally Distributed Returns

Fig. 5.1. Comparison of daily DAX returns for the period 01/03/1986 −
12/30/2005 with normally distributed returns. The negative 9.47σ-event on
October 16, 1989 was cut off. The daily mean return at σ = 0 equals 0.027%,
the standard deviation is 0.014 (implying an annual variance of 23.0%), the
skewness is negative with −0.463, and the excess kurtosis is determined as
5.84.

the kurtosis11 as a distribution’s fourth moment

N
1
(rj − r̄)4
κ= (5.4)
N σ4
j=1

is always larger than three. A comparison of daily DAX returns with

normally distributed returns in figure 5.1 illustrates that real finan-
cial data have more probability mass in the tails and in the center.
In other words, they are fat-tailed and high-peaked. The kurtosis for
the daily DAX returns in the analyzed period is 8.7, significantly more
than the expected kurtosis of three for a normal distribution. Within
the N = 5, 040 daily observations of DAX returns, three +5σ, two −5σ,

11
Kurtosis derives from the Greek word kyrtos for “curved” and is a measure of
how tall or squat a curve is. A normal distribution is mesokurtotic with a kur-
tosis of three. Thin-tailed distributions with a kurtosis less than three are called
platykurtotic. Excess kurtosis is defined as κ − 3.
58 5 Replicating the Stylized Facts of Financial Markets

three −6σ, and one −9σ events occurred. Under the Gaussian assump-
tion of asset returns, this would have been extremely unlikely. The oc-
currence of extreme events in figure 5.1 is also in line with Mandelbrot’s
observation that the positive tails contain systematically fewer events
than the negative tails. This particular feature of empirical return dis-
tributions is captured by a distribution’s third moment skewness
N
1
(rj − r̄)3
S= . (5.5)
N σ3
j=1

A skewness smaller than zero indicates that the negative tail of the
distribution has more probability mass than the positive tail.
As a joint test of sample skewness S and sample kurtosis κ, the
Jarque-Bera test statistic [197, 30]
 2
S κ−3
JB = N + ∼ χ2 (2) (5.6)
6 24
of 7313.4 with a p-value of 0.000 clearly rejects the null hypothesis of
normality.12
The leptokurtosis of return data is explained by the mixture of dis-
tributions hypothesis [74] which asserts that the returns are sampled
from a mixture of distributions with different conditional variances. The
stable Paretian hypothesis advanced by Mandelbrot, on the other hand,
claims that returns distributions are best characterized by a different
class of distributions. In addition to being non-Gaussian, Mandelbrot
found that the shape of daily, weekly, and monthly return distributions
of cotton prices appeared to be similar under different time scales [276].
This invariance is called scaling and it is characterized by power law be-
havior in the tails of a distribution, i.e., extreme events are distributed
according to f (r) ∼ |r|−(1+α) .13 The stable Paretian hypothesis states
12
Other popular tests for normality are, for instance, the Kolmogorov-Smirnov test,
the Shapiro-Wilk W -test, and the Anderson-Darling test. For a review of normal-
ity tests, see [86].
13
Power-law scaling was first discovered by Vilfredo Pareto (1848 − 1923) in the
distribution of wealth. Pareto found that in a population, the proportion of in-
dividuals with wealth W above a certain threshold is determined according to
a power law f (W ) ∼ CW −(1+α) . The power law exponent α was estimated by
Pareto to be around −3/2, f (W ) being the probability density function, and C
a positive constant. Scaling and power law distributions are also very familiar
to physicists because they appear near critical phase transitions from ordered to
disordered states. In fact, much of the work on scaling behavior in finance has
been done by physicists in the econophysics-movement. For an introduction into
econophysics, see [283].
 5.2 Stylized Facts of Financial Markets 59

that return distributions belong to a general class of stable Paretian

distributions14 with a characteristic exponent α strictly smaller than
two. Paretian distributions generally have no closed analytical form, but
can be expressed with their characteristic function f (q), which is the
Fourier-transform of the probability distribution, q being the Fourier
transformed variable:

  
 iδq − γ|q|α 1 + iβ q tan π α if α = 0
ln f (q) =
|q| 2 (5.7)
 iδq − γ|q| 1 + iβ 2 q ln |q| if α = 1
π |q|

Stable Paretian distributions are described by four parameters: i) a

location parameter δ for the mean, ii) a scale parameter γ > 0, iii)
a parameter for asymmetry ( skewness) β, and iv) the characteristic
exponent α. Special cases of Paretian distributions for which closed-
form solutions are known are the Gaussian (α = 2, β = 0) and the
Cauchy (α = 1, β = 0) distribution.15 In his 1963 study, Mandelbrot
[276] estimated the characteristic exponent α for cotton prices to be
around 1.7, similar to what Fama reported for the Dow Jones Industrial
Average [113].
While data availability and computational limitations allowed Man-
delbrot only to investigate small data sets, the recent analysis of huge
high frequency data sets suggests a more complicated scaling behavior
of return distributions.16 Mantegna and Stanley, for instance, noted
that stable Lévy-type behavior does not apply to the most extreme
parts of return distributions and suggested instead a truncated Lévy
process whose tails exponentially decay after some threshold value
[282]. Gopikrishnan et al. [158] reported that return distributions slowly
converge to Gaussian behavior as △t gets larger. They also note that
the scaling properties for small △t break down when they reshuffle
their data set, thereby destroying all possible time dependencies in the
14
Stable Paretian distributions are also called fractal, L-stable, or Lévy distribu-
tions. This general class of distributions was discovered by Paul Lévy. They are
by construction stable, i.e., invariant under addition. When adding two Lévy ran-
dom variables, the functional form, and hence, the shape of the distribution, is
maintained (stable). For more information on stable Paretian distributions see,
for instance, [276, 113, 114, 265, 346, 381].
15
An implication of the stable Paretian hypothesis is that the second and higher
moments do not asymptotically converge to a finite value, i.e., they do not exist
[265].
16
When analyzing rare events, sample size matters a lot. While Mandelbrot ana-
lyzed approximately 2,000 data points, modern high frequency data sets consist
sometimes of up to 106 − 107 data points [282, 158, 159].
60 5 Replicating the Stylized Facts of Financial Markets

returns. Therefore, they conclude that scaling behavior is caused by

time dependencies in the returns.
Traditional parametric and non-parametric estimation methods usu-
ally provide a poor fit to the extreme tails of return distributions [147].
Instead of trying to fit the complete return distribution, many stud-
ies now adopt the methods from extreme value theory and exclusively
focus on extreme values rather than the whole data set when estimat-
ing the power law exponent. Empirical studies using this methodology
now find the characteristic exponent α to be outside the stable Lévy-
regime. For the German stock market, Lux [265] finds α in the interval
(2,4). Gopikrishnan et al., too, estimate 2 < α < 4 for the S&P-500
index for high frequency data [158]. On longer time scales, the charac-
teristic exponent for the Hang-Seng and the NIKKEI index was even
greater. Thus, while scaling behavior is usually confirmed, most studies
agree that the return distributions scale outside the stable Lévy-regime
[2, 259, 165].
The exact distribution of financial returns and its causes remains an
open question. Various alternative return distributions are suggested,
some of which are mentioned in [267]. Since knowledge about the correct
probability distribution is crucial for many financial applications such
as portfolio selection, derivative pricing, or value-at-risk estimations,
research into the appropriate distributional model will continue.

5.2.2 Volatility Clustering of Returns

The first stylized fact of empirical return distributions stated that the
frequency of extreme events is larger than under the Gaussian hypoth-
esis. Volatility clustering as the second stylized fact notes that these
large returns often tend to emerge in clusters, i.e., large returns tend
to be followed by large returns of either sign.17 This is also known as
volatility persistence. The volatility clusters for the daily DAX returns
for the period 01/1986 − 12/2005 are displayed in figure 5.2.
Volatility persistence means that volatility is serially correlated and,
therefore, partially predictable. Another way to visualize this is to plot

17
Fat tails and clustered volatility are closely connected and not unique to financial
time series. In a study of hydrology, Mandelbrot coined the term “Noah-effect” for
the first phenomenon, referring to extreme precipitation in the biblical account
of Noah [281]. The phenomenon of clustered volatility, i.e., persisting high or
low levels in rivers, was labeled by them as the “Joseph-effect”, drawn from the
familiar biblical story of Joseph interpreting the Pharaoh’s dream to mean seven
years of feast followed by seven years of famine.
 5.2 Stylized Facts of Financial Markets 61

0,10

0,05
log. returns

0,00

-0,05

-0,10

-0,15
1 / 86
1 / 87
1 / 88
1 / 89
1 / 90
1 / 91
1 / 92
1 / 93
1 / 94
1 / 95
1 / 96
1 / 97
1 / 98
1 / 99
1 / 00
1 / 01
1 / 02
1 / 03
1 / 04
1 / 05
Fig. 5.2. Daily DAX-returns for the period 01/03/1986 − 12/30/2005.

the autocorrelation of absolute and squared returns, which both mea-

sure volatility. While the autocorrelation coefficients of raw returns fluc-
tuate around zero, those for absolute and squared returns are usually
much higher and indicate long-memory processes.18 The autocorrela-
tion coefficients for the first two hundred lags of raw, absolute, and
squared daily DAX returns are plotted in figure 5.3 and are represen-
tative of most other financial return series.
Theoretical models that explain the possible causes of volatility per-
sistence are rare. One of the few exceptions is the preference-based asset
pricing model by McQueen and Vorkink [293] which generates low-
frequency conditional volatility on the basis of agents’ state-dependent
risk-aversion and sensitivity to news. There are, however, many struc-
tural models of return generating processes that abstract from specific
economic relationships. The existence of clustered volatility suggests
abandoning the assumption of linear return generating processes such
as Autoregressive (AR), Moving Average (MA), or ARMA processes
and focusing on non-linear processes.
18
While most older studies agreed on insignificant autocorrelation for raw returns
[114], Lo [255] provides evidence of time-varying small, but significant, serial price
correlation.
62 5 Replicating the Stylized Facts of Financial Markets
Autocorrelation of raw, absolute, and squared daily DAX returns
0,3
absolute returns
0,25
squared returns
Autocorrelation

0,2
raw returns
0,15
0,1
0,05
0
-0,05
1 51 101 151 Lag 201

Fig. 5.3. The first 200 autocorrelations for the raw, squared, and absolute
daily DAX returns (01/03/1986 − 12/30/2005).

Threshold Autoregressive (TAR) models [424] are additive non-

linear return generating processes since they jump between several lin-
ear stochastic processes based on certain threshold levels. This class of
models is thus locally linear, yet non-linear over the whole state space.
Another more popular approach to model serial correlation of volatility
was proposed by Engle [110]. He introduced the Autoregressive Condi-
tionally Heteroskedastic ARCH(q) model

rt = µt + σt ǫt (5.8)

q
2
σt = a0 + ai ǫ2t−i , (5.9)
i=1

in which the variance of an error ǫt is conditional on lagged variances.

The error terms ǫt are normally distributed ǫt ∼ N (0, 1), µt is the
expected return in period t, and as parameter restrictions a0 > 0 and
ai ≥ 0 for
q all i = 1, . . . , q apply. For a stationary return processes like
above, i=1 ai < 1 must hold. In this case, the unconditional variance
σ 2 exists and is a constant
a
σ2 = 0 . (5.10)
1 − qi=1 ai
Using equations 5.9 and 5.10, the conditional variance for an ARCH(1)-
process  
σt2 = σ 2 + a1 ǫ2t−1 − σ 2 (5.11)
is mean-reverting around the unconditional variance σ 2 . Equation 5.11
also illustrates that unexpected large changes will result in a higher con-
ditional volatility, i.e., clustered volatility. Simple ARCH(1) models are,
 5.2 Stylized Facts of Financial Markets 63

however, unable to fully capture volatility persistence, and higher order

ARCH models are computationally hard to estimate. Bollerslev [45]
thus suggested the Generalized ARCH GARCH(p,q) model in which
equation 5.9 for the conditional variance is replaced by
q

p


σt2 = a0 + ai ǫ2t−i + 2
bj σt−j . (5.12)
i=1 j=1

For σt2 to be non-negative, the restrictions a0 > 0, ai , bj ≥ 0 for all

i, j = 1, . . . , q, p must apply, and for stationary GARCH(p,q) processes
q

p


ai + bj < 1 (5.13)
i=1 j=1

must hold. In contrast to ARCH(1) models, low-order models in the

GARCH class often describe return distributions with fat-tail behavior
and volatility persistence reasonably well. Fitting a GARCH(1,1) model
for the daily DAX -returns yields

σt2 = 5.33 · 10−6 + 0.119ǫt−1 + 0.858σt−1

2

for the conditional variance.19 Pagan, however, points out that the
residuals from estimated GARCH models are still non-Gaussian and
leptokurtotic [328]. In addition, Mikosch and Stărică [302] discuss some
theoretical properties of GARCH(p,q) processes that are not in line
with real financial data. For instance, stochastic long-term memory
processes exhibit hyperbolic decays in their autocorrelation functions
(ACF), yet the decline in the theoretical ACF for all GARCH(p,q) pro-
cesses is exponential, thus, indicating only short-memory processes.20
Mikosch and Stărică have also recently questioned the validity of
GARCH models to describe stochastic volatility and argued that long-
range dependence effects might be only spurious. They detected that
structural changes, for instance, due to business cycles, cause a viola-
tion of the stationarity assumption by shifting unconditional variances
[304, 305].
Various other refinements for the conditional variance have been
suggested. In asymmetric GARCH models such as the Exponential
19
All parameter estimates are highly significant and the stationarity condition is
fulfilled.
20
Note, however, that Fractionally Integrated ARCH (FIGARCH) models [18] are
able to model long-run dependence in stochastic volatility. For a technical dis-
tinction between short- and long-term dependence, see [303].
64 5 Replicating the Stylized Facts of Financial Markets

GARCH [318], volatility depends not only on the size, but also on the
sign of past shocks. Asymmetric GARCH models thus allow incorpo-
rating the leverage effect, i.e., the empirically observed asymmetric re-
actions of market volatility to positive and negative shocks. For reviews
on GARCH time series models, see [46, 251]. An empirical comparison
of different volatility forecast models can be found in [169, 168].
The continuing dependence on Gaussian innovations with time-
varying variances by GARCH type models is challenged by Mandel-
brot and Hudson [280]. As an alternative, they suggest the Multifractal
Model of Asset Returns (MMAR) [279]. The MMAR combines frac-
tional Brownian motion with the concept of multifractal trading time,
i.e., a random distortion of clock time.21 Rather than being an au-
toregressive process, the MMAR is a cascade model which generates
price series by randomly splitting up a complete interval in smaller
sub-intervals, thereby increasingly roughening the price series by suc-
cessive interpolations. In the limit, the generated price series reproduce
the main features of financial prices: fat tails in the return distribu-
tions, scale consistency, and long memory in volatility. In addition,
the MMAR can generate either martingales or long memory behavior
in log prices. First empirical investigations of the MMAR hypothesis
show promising results [63, 314, 44], yet the assumption of multifractal
time in an economic context is still being questioned [307].

5.2.3 High and Persistent Trading Volume

In the world of a representative agent, fully rational investors who share
the same information and expectations would obviously never trade
with each other. More surprising is the no-trade theorem by Milgrom
and Stokey who showed that information asymmetry alone cannot in-
duce trading [306]. Given the same prior beliefs and an ex-ante Pareto-
efficient allocation, the possibility of asymmetric information arrival
would make a risk-averse rational trader believe that a potential trad-
ing partner is better informed than himself. Under these circumstances,
no one would ever engage in any trading activity.
To reconcile the spate of trading activity in real financial markets
with no-trade theorems, several solutions have been suggested. For
Black [36], the missing ingredient is noise trading, i.e., he believes that
investors often trade on noise as if it were information, for instance,
by following certain technical trading rules.22 As long as these trading
21
More on the concept of multifractal time can be found in [278]. A similar concept
of time relativity in financial markets is espoused by Montassér [311].
22
The noise-trader approach to finance is also advanced in [96, 393, 98].
 5.2 Stylized Facts of Financial Markets 65

strategies are uncorrelated, high trading volumes imply that irrational

trades cancel each other out and asset prices are close to fundamental
values. However, Shleifer [392] remarks that the latter hinges on the as-
sumption of uncorrelated trading strategies, something that is at times
violated in a world with mass media.23 Similar to noise traders, liquid-
ity traders or life cycle investors may act without regard to expected
profits. Other theoretical explanations for ongoing trading activities are
sought in dropping the assumption of common prior beliefs [213, 428],
or in dropping the assumption of efficient and complete markets by
integrating various insurance motives [360].
Besides being positive, transaction volume exhibits many other in-
teresting features. Lobato and Velasco, for instance, have shown it
to be a persistent long memory process [257]. Furthermore, empiri-
cal data show that trading volume is characterized by a positive au-
tocorrelation coefficient which only slowly decays, and that a positive
cross-correlation between volatility of returns and trading volume ex-
ists [417, 136, 328]. This raises the question about the informational
content of volume data and whether it may be exploited for stock price
prediction.
According to the mixture of distributions hypothesis [74], return
volatility and trading volume are jointly dependent on a latent driving
process. Theoretical support and empirical evidence for Clark’s thesis
that the conditional variance of log prices is a function of transaction
volume is provided in [111]. Other studies investigating the relationship
between transaction volume and asset prices are [417, 143, 57, 311]. A
survey of the relationship between price changes and trading volume
can be found in [214].

5.2.4 Existence of Technical Trading

The idea that asset prices and trading volumes tend to form certain
geometric patterns that can be profitably exploited by forecasting fu-
ture price movements is known as technical trading. Charting, as it is
sometimes called, had been popular even before the disclosure of finan-
cial balance sheet information sparked the development of fundamental
analysis as its main competitor. Its origins can be found in the works
of Charles H. Dow (1851–1902) and William P. Hamilton (1867–1929),
who published a series of 252 editorials in the Wall Street Journal at

23
Coordination on similar trading strategies because of media-induced herding be-
havior would constitute a strategic complementarity [125] which can lead to large
deviations from the outcome predicted by rational models.
66 5 Replicating the Stylized Facts of Financial Markets

the turn of the century. These editorials contained the basic tenets of
what became later known as the “Dow Theory” [320]. Another school of
technical analysis, Elliott-wave theory, claims that market prices follow
several superimposed market trends of different degrees, all of which
are rooted in cyclical investor psychology [107, 108, 342].
Since technical analysis is based on the assumption that markets
exhibit not even the lowest degree of market efficiency, it is anathema
to many academics. Malkiel [274], for instance, compares the scien-
tific merits of chart reading to those of alchemy.24 Jegadeesh [199], too,
points out that academic concerns against technical analysis run deeper
than simple “linguistic barriers” [256]. He identifies its weak founda-
tions, i.e., the missing plausible explanations why one should expect
market patterns to repeat, as the main culprit.

The Use of Technical Trading in Financial Markets

The academic studies that deal with the extent to which technical trad-
ing is used in real markets mainly investigate the foreign exchange mar-
ket. Frankel and Froot, for instance, attribute the speculative bubble
in the period from 1981–85 to a shift from fundamental trading strate-
gies towards technical ones [138]. Questionnaire evidence from forex
traders about their use of technical analysis was gathered by Allen
and Taylor [7, 419]. Both studies found that the relative importance
of charting strategies diminishes as the investment horizons increase.
For horizons ranging from intra-day to one week, up to 90% of the
respondents relied more on technical than fundamental analysis when
forming exchange rate expectations, a finding that is reversed for time
horizons exceeding one year. It is also reported that around 2% of
traders appeared to never use fundamental analysis, independent of
their investment horizons. Menkhoff confirms these results for a differ-
ent place (Germany) and a different time and concludes “that technical

24
According to Popper’s demarcation criterion citePopper1959, however, technical
analysis principally follows a scientific methodology. It produces falsifiable propo-
sitions that can be refuted by empirical observations. Given the questionable track
record of many technical trading systems, their survival surprises many EMH ad-
herents. One possible answer will be given at the end of chapter 9. Another
plausible answer to this mystery is immunization strategies. A website dedicated
to the Dow Theory, for instance, states the following: “To define Dow Theory one
must be open to the concept that there are specific rules that must be accepted with-
out question, while at the same time being ever aware that the theory is dynamic
in its evolution” (http://www.dowtheoryproject.com/define.php). It is hard too
conceive of a more blatantly unscientific statement than this.
 5.2 Stylized Facts of Financial Markets 67

currency analysis should be interpreted not as either a marginal phe-

nomenon or representing secondary information or a self-eliminating
or second-best strategy, but possibly as a kind of self-fulfilling prophecy”
[298, p. 307].25 Even though academic studies investigating the impor-
tance of technical trading in the stock market do not abound, anecdotal
evidence and the mere existence of countless books and newsletters on
technical stock market trading point to a widespread use, too. Ques-
tionnaire evidence gathered by Shiller following the stock market crash
in October 1987 confirms that many traders were influenced by tech-
nical analysis considerations [390].

The Profitability of Technical Trading Rules

The main body of academic research on technical analysis focuses

on its profitability. While the sixties and seventies were marked by
a widespread consensus among academics that technical trading net of
transaction costs did not generate systematic returns greater than a
simple buy-and-hold strategy [115, 202], the “intellectual dominance”
of the EMH had become less universal by the beginning of the new
millennium [275].
Aside from reported weak form inefficiencies for obvious candi-
dates such as Latin American or Asian emerging stock markets26
[348, 309, 387, 441], return predictability has been found even for West-
ern European equity markets, for instance, in [300] for the Italian stock
market, and in [132] for the Spanish stock market. For the US, Brock
et al. [56] tested two popular technical trading strategies, the moving-
average oscillator and the trading-range break, on daily Dow Jones
Industrial Average data from 1897 to 1986. The generated buy (sell)
signals were followed by returns that were generally higher (lower) than
“normal” returns, indicating that these technical rules have predictive
power. Kwon and Kish extended the analysis of Brock et al. by includ-
ing volume data and supported their basic conclusion [229, 228]. They
did find, however, a gradual weakening in profit potential, implying
that the market became more efficient for the period under considera-
25
Other empirical studies based on questionnaire evidence in the foreign exchange
market are [263, 299, 71, 72, 325, 1].
26
The literature on technical trading profitability in the foreign exchange market
is even more voluminous than for stock markets. Many studies have found prof-
itable trading rules around central bank interventions [237, 317, 363]. Because of
these interventions, foreign exchange markets are likely to be governed by other
rules than equity markets and will thus be ignored in the discussion of profitable
technical trading rules.
68 5 Replicating the Stylized Facts of Financial Markets

tion. Ready [350], however, considers the apparent profitability of the

technical trading rules used by Brock et al. to be a spurious result of
data snooping.
Instead of using ad hoc specifications of technical trading rules, other
authors ask whether it is possible to derive “optimal” trading rules by
means of artificial intelligence techniques. Allen and Karjalainen used
genetic programming to derive new technical trading rules from build-
ing blocks—certain moving averages and maxima and minima of past
prices—and evaluated them against a buy-and-hold strategy [5]. Using
daily S&P 500 index data from 1928–1995, they found little evidence
of economically significant technical trading rules. Even though the
trading rules were able to detect periods with overall positive or neg-
ative returns, the economic profits minus the transaction costs were
not consistently higher than a simple buy-and-hold strategy. Allen and
Karjalainen attributed the limited forecasting ability to positive low-
order serial correlation in stock index returns, since the introduction of
a one-day delay to trading signals caused the positive returns to vanish.
Jasic and Wood [198] derive neural network predictions for daily
returns of the S&P 500, DAX, TOPIX and FTSE stock market in-
dices. Unlike Allen and Karjalainen, they find strong evidence of high
and consistent predictability even in the presence of plausible transac-
tion cost assumptions. Other studies that use neural networks on stock
market data are [127, 443]. Another genetic programming approach is
by Kaboudan [211]. In summarizing the current discussion about the
profitability of technical trading rules, one can only conclude that the
debate is far from being settled.

The Effect of Technical Trading on Asset Prices

More theoretical in nature are studies which focus on the effect that
the use of technical trading rules may have on the aggregate market
outcome. The surge of articles in the field of behavioral finance may be
seen partly as a response to the failure of explaining asset price based
on pure fundamentals.
Price-to-price feedback models, as discussed in [391], are closely
related to trend-following trading strategies. As asset prices go up,
the success of some investors attracts the attention of other investors,
thereby fueling the expectations of future price increases and investor
demand. As long as the feedback mechanism is uninterrupted, unrea-
sonably high asset prices can be sustained over extended periods of
time, yet eventually, the bubble has to burst. However, the causal di-
rection, whether pre-existing technical trading strategies initiate trends
 5.2 Stylized Facts of Financial Markets 69

as precursors of boom and crash periods or vice versa, is not always

clear. The same can be said for momentum or over- and under-reaction
theories of the stock market [200, 24, 88, 192]. Overreaction, i.e., the
tendency of individuals to overweight recent information and to under-
weight prior data, implies that extreme stock price movements will be
followed by subsequent adjustments. DeBondt and Thaler [95] remark
that this hypothesis implies a violation of weak-form market efficiency.
Certain technical trading rules should, therefore, be able to exploit the
observed patterns as was shown in [200].
Multi-agent systems often address the effect of technical trading
rules on asset prices by modeling the population dynamics of differ-
ent trader groups, usually depicted as fundamentalists and chartists.
Instead of two groups, agents may endogenously switch between these
two groups, thus, influencing the importance of certain trading strate-
gies. Lux [266], for instance, developed a multi-agent model of the for-
eign exchange market in which fundamentalists and chartists interact.
Chartists have either a “bullish” or “bearish” attitude and observe
both the behavior of their competitors as well as price movements.
The transition probabilities between the two chartist subgroups reflect
mimetic contagion and actual price movements, while the transition
probabilities from and to the fundamentalist group are a function of the
profitability of the existing trading strategies. As the ratio of chartists
to fundamentalists constantly changes, the simulated model exhibits
optimistic and pessimistic waves. More interestingly, though, are the
statistical properties of the simulated price series. They reproduce the
stylized fact of leptokurtotic return distributions which become less
pronounced when returns are calculated using increasing time inter-
vals.
A similar population dynamic is adopted by Goldbaum [152]. He
varies an intensity of choice parameter with which agents switch be-
tween a fundamental and a technical trading rule as a result of per-
ceived differences in expected profits. Goldbaum finds that high values
of this parameter lead to large population shifts. Asset price movements
reflect the market impact of technical traders rather than fundamen-
tal values, and an overall pattern of repeating price bubbles and col-
lapses emerges. Similar multi-agent systems that try to explain asset
price behavior as a result of the population dynamics are described in
[264, 151, 269, 73, 121].
70 5 Replicating the Stylized Facts of Financial Markets

5.3 Alternative Market Hypotheses

Many of the stylized facts of financial markets discussed are in stark

contradiction to the random walk version of the EMH. As a result,
various other market hypotheses have been postulated, some of which
will be briefly discussed in the following section.

5.3.1 The Fractal Market Hypothesis

An alternative market hypothesis to the EMH was proposed by Peters

[335, 336]. The key ingredient of his Fractal Market Hypothesis (FMH)
is the assumption that markets consist of traders with different invest-
ment horizons. Depending on the specific proportion of these investor
groups, a market can either be in a stable or an unstable state. The
FMH asserts that the market is stable when investors cover a large
number of investment horizons. Long-term investors stabilize the mar-
ket by offering liquidity to short term investors. If an event renders the
validity of long-term fundamental information questionable, long-term
investors either stop participating in the market or become short term
investors themselves. As the overall investment horizon of the market
becomes more uniform, the market approaches its unstable state. The
assumption of different investment horizons implies that, contrary to
the EMH, prices do not always reflect all available information. Short
term investors, for instance, may value information differently than
long-term investors. In times with more uniform investment horizons,
information may thus only be partially reflected in asset prices.
The FMH derives its name from the prevalent fractal structure when
the market is in its stable state. Because of the differing investment
horizons, a market does not have a characteristic time scale. That is,
when plotting the distributions of daily, weekly, or monthly returns,
they all look similar. This feature of self-similarity on different (time)
scales is one of the defining properties of simple fractals.27
The observed fat tails in the return distributions are caused by ex-
treme price variations in the market’s unstable state. Since high peaks
and fat tails are not compatible with the assumption of Gaussian re-
turn distributions, Mandelbrot proposed the more general class of sta-
ble Paretian distributions for stock returns [276]. Within this class, the
normal distribution is only a special case, i.e., when the characteristic
27
Financial market charts are, however, better characterized as being self-affine than
self-similar. Note that a random walk is also self-similar, but it is not fractal since
its fractal dimension is an integer (and not fractional). For geometric fractals, self-
similarity and scaling are spatial and not temporal properties.
 5.3 Alternative Market Hypotheses 71

exponent α, a measure of the degree of tail fatness, equals two. The

stable Paretian hypothesis of stock returns asserts that α is strictly
less than two. Empirically, the characteristic exponent α was, indeed,
almost always found to be in the interval (1, 2). Besides fat-tailedness,
this implies the existence of a well-defined mean, but second or higher
moments do not exist [265].
Peters points out that the inverse of the characteristic exponent α is
equal to the Hurst exponent H, which is a measure of trend persistence
[336]. For Hurst exponents 0.5 < H < 1, which correspond to α ∈ (1, 2),
time series are said to be persistent or trend-reinforcing. Furthermore,
by applying rescaled range (R/S) analysis [87], Peters detects periodic
or nonperiodic components (cycles) in the returns of the Dow Jones
Industrial Average.

5.3.2 The Coherent Market Hypothesis

A second market hypothesis with connection to chaos theory was pro-
posed by Vaga [426]. Like the FMH, the Coherent Market Hypothesis
(CMH) is based on the premise that markets can shift between stable
and unstable regimes. The CMH applies the well-known Ising model
of ferromagnetism to the stock market. Instead of modeling the forma-
tion of clusters with the same magnetic orientation, the Ising model
in a social context is reinterpreted to describe group formation and
group behavior. The two most important parameters are one for mar-
ket sentiment or crowd behavior, and another for fundamental bias.
Based on the particular combination of these two parameters, the mar-
ket could be in one of four possible states. To be in the stable random
walk state—the EMH case—, the crowding parameter must be be-
low a certain threshold level, and aggregate fundamental views should
not have a bullish or bearish bias. As the crowd parameter increases
above a critical threshold, the model becomes instable, and the prob-
ability distribution of stock market fluctuations becomes bimodal. If
the fundamental bias is not very pronounced, the stock market is in
its chaotic state where large swings can be triggered by small news.
In periods with considerable crowd behavior and strong fundamental
bias, the stock markets are in either the coherent bull or coherent bear
state. Vaga defines coherent behavior as a state of macroscopic order in
a complex system, i.e., when a large number of freely interacting parts
temporarily align with each other.28
28
Obviously, this perception of coherent market behavior is close to emergent mar-
ket behavior. The latter, however, is more general since it does not exclusively
refer to correlated behavior.
72 5 Replicating the Stylized Facts of Financial Markets

Vaga claims that predictive powers are greatest in coherent market

periods. If investors miss being fully invested in these periods, they will
underperform the market in the long run. Unlike the FMH, Vaga con-
structed the CMH to assist in investment decisions through better mar-
ket timing. He describes technical signals which presumably indicate
the transition between different market states, yet his own investment
record compared to a buy-and-hold strategy remains ambiguous. This,
however, does not disprove the CMH since the transition signals might
have been inadequate. Steiner and Wittkemper [410], for instance, im-
plement the CMH in a portfolio optimization model and generate the
transition signals through artificial neural networks. A simulation with
out-of-sample data resulted in consistently higher positive annual re-
turns compared to the market portfolio. The EMH allows for consis-
tently higher profits only as a fair compensation for risk-taking. Yet the
outperformance of Steiner and Wittkemper’s model was achieved with
only 41% of the risk of a buy-and-hold strategy of the market portfolio.

5.3.3 The Adaptive Market Hypothesis

The Adaptive Market Hypothesis (AMH) is a new behavioral market
hypothesis that was recently suggested by Lo [255]. He considers the
AMH as a new version of the classical EMH, yet derived from evo-
lutionary principles. Many of the deviations from economic rational-
ity that have been pointed out by the behavioral finance literature—
overconfidence, overreaction, loss aversion, herding, hyperbolic dis-
counting, and other behavioral biases—are in Lo’s view consistent with
an evolutionary perspective on individuals adapting to a changing en-
vironment via simple heuristics. From an evolutionary perspective, a
financial market can be seen as a co-evolving ecology of trading strate-
gies. The strategies would correspond to biological species, and the
total capital commanded by a strategy is analogous to the population
of that species. New strategies are constantly created, thereby changing
the profitability of pre-existing strategies, in some cases even replacing
them or driving them to become extinct [122].
In such a dynamically changing world, it should come as no sur-
prise that the heuristics pertaining to an outdated environment may
be ill-advised in the context of a new environment. Those heuristics
might seem irrational, but labeling them as maladapted is more ap-
propriate. Lo views the classical EMH as a steady state situation with
fixed environmental conditions, whereas the AMH describes constantly
adapting and co-evolving groups of market participants or strategies.
At the present stage, the AMH is more of a qualitative concept that
 5.3 Alternative Market Hypotheses 73

still needs to be operationalized. Nonetheless, Lo derives a number of

concrete implications.
First, the relationship between risk and returns is unlikely to be
stable over time. As a result, the equity risk premium and aggregate risk
preference are not immutable constants, but time-varying and possibly
path-dependent. Natural selection affects the relative importance of
different participating investor types who may have differing degrees of
risk aversion. The recent burst of the technology bubble probably led
to a vastly different population of active investors whose preferences
might have changed, too.
A second implication of the AMH is the occasional existence of prof-
itable arbitrage opportunities. As they are exploited, they disappear.
An example provided by Farmer [120] demonstrates how slowly the
achievement of market efficiency may come about.29 Figure 5.4 shows
the gradual decline in correlation between a model generated signal and
the future market movement. During the 23-year period that is shown,
the quality of the signal has declined, but not vanished.

Fig. 5.4. Decline in magnitude of a market inefficiency. Correlation between a

trading signal and the following forward return of a specific US stock market
segment over time. The signal was created by a model used by Prediction
Company for actual trading purposes. Source: [120, p. 65].

Lo also points out that a changing environment leads to the constant

creation of new profit opportunities. Figure 5.5, which again is taken
29
Farmer is a co-founder of Prediction Company, a Santa Fe-based firm that devel-
ops and markets financial forecasting systems. A lively and entertaining account
of Prediction Company and its two founders, John Doyne Farmer and Norman
Packard, is given in [26]. Farmer is a trained physicist whose views of finance were
shaped by conversations with traders and academics. He claimed that Prediction
Company made highly significant profits which should have been impossible ac-
cording to the EMH [119].
74 5 Replicating the Stylized Facts of Financial Markets

from [120], may be interpreted as an example of a new inefficiency on

the market.

Fig. 5.5. Emergence of a new market inefficiency. Correlation between a

trading signal and the following forward return of a specific US stock market
segment over time. The signal was created by a model used by Prediction
Company for actual trading purposes. Source: [120, p. 65].

Over a longer time horizon, the emergence and disappearance of

arbitrage opportunities should result in a cyclical behavior of market
efficiency. Lo illustrates this conclusion by plotting the first-order auto-
correlation coefficients for monthly returns of the S&P Composite Index
over a 132-year period. They are at time zero, but fluctuate wildly with
peaks up to 50%. According to the random walk hypothesis, these serial
correlations should have been zero all the time. Finally, Lo concludes
that the key to survival in a changing environment is innovation. And
while profit or utility maximization may be important aspects in finan-
cial markets, survival is the main driving force in market evolution, the
only objective that really matters to market participants.
While the label “Adaptive Market Hypothesis” seems new, the ideas
expressed by it have been around for a while. Chen and Yeh [69], for
instance, explore the possibility whether the EMH and the Rational Ex-
pectation Hypothesis can be explained as emergent properties in an ar-
tificial stock market consisting of co-evolving heterogeneous agents. The
importance of financial innovation in a changing environment is also
emphasized by Markose et al. who invoked the Red Queen principle30
to explain market efficiency as an emergent behavior of co-evolving
agents [287]. The Red Queen Principle is known in evolutionary biol-
ogy and describes the idea that a species needs continuing development
30
The Red Queen Principle is named after a famous passage in Lewis Carol’s book
“Alice Through the Looking Glass”. “‘Well, in our country’ said Alice, still pant-
ing a little ‘you’d generally get to somewhere else if you ran very fast for a long
time as we’ve been doing.’ ‘A slow kind of country!’ said the Red Queen. ‘Now
here, you see, it takes all the running you can do, to be in the same place’”.
 5.4 Agent-Based Computational Models of Financial Markets 75

just in order to maintain its fitness relative to other species with which
it is co-evolving [427]. It is often referred to as an evolutionary arms
race between competing species. One of the first to study Red Queen
effects in economics was Robson [357], and Markose further discusses
Red Queen examples in an economic context [286].

5.3.4 The Interacting-Agent Hypothesis

The EMH essentially assumes that the return distribution simply mir-
rors the non-observable news arrival process. As researchers acknowl-
edge more and more that stock returns were not normally distributed,
the random walk hypothesis was replaced with the less demanding mar-
tingale model. The random shocks no longer needed to be indepen-
dently distributed, thus, allowing for deviations from normality in the
higher moments of the return distribution and for volatility clustering.
The Interacting Agent Hypothesis (IAH) by Lux and Marchesi
[268, 269], on the other hand, assumes that the news arrival process is
again Gaussian distributed. In their agent-based simulation of a stock
market, however, Lux and Marchesi find that the time series of returns
exhibits both fat tails and volatility dependence.31 Thus, they con-
clude that these statistical properties appear as emergent phenomena
as a consequence of the market interactions of heterogeneous agents. It
is the trading process itself that transforms and magnifies exogenous
news, modeled as white noise, into realistic fat-tailed return distribu-
tions with clustered volatility. This feature has been verified by Chen
at al. [67]. Many other agent-based simulations of financial markets ex-
hibit the same emergent characteristics of fat-tailed return distributions
with clustered volatility, some of which will be briefly introduced in the
following sections. The IAH is thus a serious competitor to the EMH in
explaining how the observed return characteristics come about, yet an
empirical test to discriminate between the two hypotheses is hard to
conceive of. The Heterogeneous Market Hypothesis (HMH) proposed
by Hommes [189] rests on a argument similar to the IAH.

5.4 Agent-Based Computational Models of Financial

Markets
Even though the field of heterogeneous agent models in economics and
finance is still relatively young, categorizing the different models has al-
ready become quite difficult. The first problem arises in deciding which
31
Their artificial stock market model is presented in section 5.4.3 on page 83.
76 5 Replicating the Stylized Facts of Financial Markets

type of models to include or not. A pragmatic approach is to distin-

guish between those models that can be analytically handled and those
that cannot and need to be studied by means of computer simulations.
The first group of analytical heterogeneous agent models are surveyed
in Hommes [190], while LeBaron [243] focuses on agent-based com-
putational models.32 I will follow this distinction and concentrate on
computational models.
The second problem arises in defining the criteria according to how
these models should be categorized. One could easily arrange them ac-
cording to which stylized facts of financial markets they replicate, to
what kind of learning or market clearing mechanism they employ, or to
the degree of agent-heterogeneity.33 As with probably any categoriza-
tion scheme, there will be overlap and possibly models that do not fit
in the proposed scheme.
The guiding criterion in the following selection of artificial financial
markets is the principal of minimal rationality. I shall assume that
agents with random demands exhibit lower degrees of rationality than
those who either learn their optimal demands over the course of time
or derive them by optimizing a given utility function.

5.4.1 Allocative Efficiency with Zero-Intelligence Traders

A natural starting point when discussing various artificial markets from

the viewpoint of minimal rationality are the “zero-intelligence” (ZI)
traders by Gode and Sunder [150].34 In order to determine how much
of the performance in terms of allocative efficiency can be attributed
to market structure and how much to human intelligence, Gode and
Sunder designed an experiment in which the results of a control group
of human traders were compared with two types types of ZI machine
traders.
32
Other surveys on agent-based computational models in finance can be found in
[236, 238, 248].
33
LeBaron [239] discusses all the different design issues when building agent-based
financial markets. These building blocks, for instance, agents, trading and market
clearing mechanism, and the kind of available securities, could all certainly serve
as classification criteria.
34
To my knowledge, Gode and Sunder are currently the only ones to use the term
“minimal rationality economics” as keywords and when referring to their research
interests. Minimally rational traders are also studied by Farmer et al [123]. Re-
lated to the zero-intelligence approach is Schweitzer’s concept of Brownian agents
[384, 386, 385]. Brownian agents possess intermediate complexity and are mini-
malistic in the sense that they act on the simplest set of rules without deliberative
actions.
 5.4 Agent-Based Computational Models of Financial Markets 77

An arbitrary commodity was traded in a double auction market

consisting of six buyers and six sellers. If a buyer’s bid (quantity and
price) was matched or crossed by a seller’s offer, a trade was executed
between the two. In the case of crossing bids and offers, the transac-
tion price pi for unit i of the commodity equaled the earlier of the
two. For each buyer, individual demand functions were implemented
by defining a sequence of redemption values vi , i = 1, 2, . . . , n for each
unit i of the commodity. Since the redemption values were private in-
formation, the aggregate demand function was unknown to buyers and
sellers. Individual supply functions for sellers were similarly defined by
imposing a private sequence of production costs ci . Hence, profits for
buyers equaled vi − pi , and for sellers pi − ci .
The ZI traders were simple computer programs that randomly gen-
erated either bids or offers. Since the ZI traders had no utility functions,
there was no need to memorize or learn anything. It seems appropriate
then to label them as having zero intelligence. The ZI traders come
in two variations. Budget-constrained ZI-C buyers were not allowed to
bid above their redemption value while ZI-C sellers were not allowed
to sell below their production costs ci . Hence, the random bids where
independently, identically, and uniformly distributed over the interval
[1, vi ], while the random offers were restricted to the interval [ci , 200].
ZI-U traders were not required to trade within a budget constraint.
The bids and offers for these unconstrained ZI-U traders were then
uniformly distributed across the entire range of possible trading prices
[1, 200].
Gode and Sunder designed the experiment in such a way as to sepa-
rate the performance differences that are due to market discipline from
those that are a result of human intelligence and profit maximization.
It can be seen from figure 5.6 that the transaction prices for ZI-
C traders are much less volatile than for ZI-U traders and converge
to the equilibrium price (depicted by a horizontal line). Since both
trader types possess no intelligence, this difference is attributable to
market discipline imposed by the budget constraint on ZI-C traders.
The price series in markets with human traders rapidly converges to
the equilibrium price and exhibits almost no volatility. The difference in
markets with ZI-C traders reflects the contribution of human rationality
and profit maximization on the outcome.
The allocative efficiency of each market is assessed by dividing the
total profits earned by the maximum attainable profit, i.e., the sum
of consumer and producer surplus [404]. Compared to humans and
budget-constrained ZI traders, the efficiency of ZI-U markets is always
78 5 Replicating the Stylized Facts of Financial Markets

Fig. 5.6. Typical results from one of Gode and Sunder’s experiments. Top: ZI-
U traders. Middle: ZI-C traders. Bottom: Human traders. Supply and demand
schedules for this experiment are shown on the left. Source: [76, p. 6], redrawn
from [150, p 127].

lowest. Extra-marginal units, i.e., those beyond the equilibrium point

and therefore difficult to trade by budget-constrained traders, were
also traded, albeit at a loss, thus accounting for the lower efficiency. It
varies, depending on the demand and supply schedules, between 50%
and 90%. The allocative efficiency in human markets averaged 97.9%,
and that for ZI-C markets was about 98.7%. Even though Gode and
Sunder did not perform any statistical tests, this difference in efficiency
 5.4 Agent-Based Computational Models of Financial Markets 79

hardly seems statistically significant. Gode and Sunder concluded that

imposing market discipline on random-unintelligent behavior is enough
to improve the efficiency from the baseline level to that attained by
human traders.
While Cliff and Bruten [75] do not question that the structure of
a double auction market is responsible for achieving high levels of al-
locative efficiency, they show that Gode and Sunder’s claim that the
convergence of transaction prices to the theoretical equilibrium price
is a consequence of market discipline is incorrect. By using probability
density functions, they show that the mean transaction price is close
to the equilibrium price only when the gradients of linear supply and
demand curves are of the same magnitude and of opposite signs. For
all other cases, the expected difference between transaction prices in
ZI-C markets and the equilibrium price can be determined in advance
by a probabilistic analysis. Simulation results supported these find-
ings. Hence, Cliff and Bruten conclude that ZI-Q traders do not have
sufficient rationality to exhibit the equilibrating tendencies of human
traders in double auction markets.
In a related paper, Cliff and Bruten [76] developed “zero-intelligence-
plus” (ZIP) traders that use a first-order adaptive mechanism to adjust
their bids and offers. By increasing or decreasing their profit margin
based on the latest entered bid or offer, Cliff and Bruten implicitly
establish some kind of utility function which agents do not maximize,
but improve upon. Simulations with different types of supply and de-
mand schedules showed that these improved ZIP traders were able to
slowly converge to the equilibrium price. A more extensive discussion
of Gode and Sunder’s zero-intelligence traders can be found in [103]. A
new model applying the zero-intelligence approach to financial markets
has recently been presented by Farmer et al. [123].

5.4.2 Models with a Random Communication Structure

A series of models with little rationality on the agents’ part was devel-
oped by researchers from the field of econophysics [81, 194, 195, 408,
407]. These models have in common that the decision processes leading
to the individual demands are not explicitly modeled, and individual
demands are assumed to be a random variable. Cont and Bouchaud
[81] argue that it is highly unlikely that the individual demands in
real markets are independent. Therefore, they add a random communi-
cation structure, allowing agents to share information and coordinate
their actions. Random communication structures, even though differ-
80 5 Replicating the Stylized Facts of Financial Markets

ently specified in most models, are a distinctive feature shared in this

class of models.

The Cont-Bouchaud Percolation Stock Market Model

In Cont and Bouchaud’s model [81], an agent i is linked to any other
agent j with probability pij = p = c/N . The links could be consid-
ered as communication channels over which agents exchange ideas and
coordinate their actions. The average number of agents to whom a
trader is connected to is given by (N − 1)p, N being the number of
agents. This linking structure and the linking probabilities are similar
to those of percolation models which are used in physics to describe
how liquids percolate through porous objects.35 Percolation systems in
physics are characterized by a critical percolation threshold pc below
which no percolation can occur. If the probability with which the con-
stituting elements are linked is in the vicinity of the threshold level pc ,
the system is near its critical point at which its state variables can be
described through power laws. This feature makes percolation theory
attractive to financial modeling since return distributions and other
variables exhibit power law behavior. Percolation models in economics
may help in understanding how local influences “percolate” through
an entire economy, causing macroeconomic variables to fluctuate, even
though the links are randomly and uniformly distributed [129].
All agents that are linked to each other form a coalition that acts
in unison. That is, they all either buy or sell one unit of stock, or
stay inactive. Aggregate excess demands are transformed into asset
price changes via a standard linear price adaptation rule. The numer-
ical value of the parameter c is crucial for the model behavior since
it controls the willingness of agents to form clusters which could be
interpreted as mutual funds or investors following specific trading rules
in real markets. Whether a cluster acts as a seller, buyer, or stays in-
active is also randomly determined and independent of cluster size. At
the percolation threshold c = 1, the probability density for the cluster
size distribution decreases asymptotically as a power law, while for val-
ues for c slightly smaller than one, the cluster size distribution is cut
off by an exponential tail.
Given the random demand and supply of stock, a naive market
model would most likely give rise to normally distributed asset re-
35
Cont and Bouchaud depart from the general percolation structure of nearest-
neighbor percolation on lattices by assuming that all agents have a positive prob-
ability of being connected with each other. This creates infinitely long interactions
instead of the usual local interactions [408].
 5.4 Agent-Based Computational Models of Financial Markets 81

turns. The endogenous formation of trading clusters slightly below the

percolation threshold of c = 1, however, mimics non-sequential herding
processes through communication processes.36 Since the analytically
derived asset price distribution is characterized by fat tails and excess
kurtosis as in real asset returns, Cont and Bouchaud’s model suggests
that herding behavior in financial markets is responsible for this styl-
ized fact. Since their model is exogenously forced slightly below the
critical point, it is not surprising to have such power law behavior in
the state variables. Cont and Bouchaud thus point out that it would be
interesting to know whether a modified model would endogenously con-
verge to the critical region, as the concept of self-organized criticality
by Bak et al. [20] suggests.

Ising-Models of Stock Markets

Other interesting models in this model class reproduce the positive

cross-correlation between volatility and trading volume [194] and vola-
tility clustering [195]. Traders are represented by the nodes of an L × L
square lattice where the links depict their connectivity, i.e., each trader
is connected to its four nearest neighbors. Individual demands are again
determined without recourse to a utility function. Agents do, however,
take a budget constraint into account. Before reaching a final decision,
a trader i repeatedly inquires about the temporary decisions Si,j ∈
{−1, 0, +1} of his four nearest neighbors and calculates an activation
signal

Yi = Ji,j Sj + Aνi , (5.14)
i,j

where νi is an idiosyncratic noise term and Ji,j a measure of mutual

influence.37 Agent i arrives at his final decision
36
Cont and Bouchaud and most other authors in the field of econophysics tacitly
bypass the subtleties contained in the concept of herding. In this model class,
herding is simply considered as a clustering process. Cluster formation could also
be the result of “spurious herding”, i.e., when groups face similar decision prob-
lems and information sets and end up acting similarly [35]. In economics, however,
it has become standard to view herding as a process that can lead to systematic
sub-optimal decisions by entire populations [99]. For instance, investors may re-
verse their initial investment decision after having observed other investors and
intentionally herd by imitating their decision.
37
The specification of the Ji,j terms determines the system’s behavior. Ji,j = 1
yields the well-known Ising-model which is used in physics to describe magnetiza-
tion processes. In Iori’s model, low levels of idiosyncratic noise would cause agents
to give up their “degrees of freedom” and align themselves to large “magnetic”
82 5 Replicating the Stylized Facts of Financial Markets

 −1 : Yi (t) ≤ ξi (t)
Si (t) = 0 : −ξi (t) < Yi (t) < ξi (t) (5.15)

+1 : Yi (t) ≥ ξi (t),

where ξi (t) denotes individual threshold activation levels. This mecha-

nism could be interpreted as a trade friction such as a transaction cost,
which causes some agents to be inactive. A market maker determines


aggregate demand D(t) = Si (t)
i:Si (t)>0

and supply

Z(t) = Si (t),
i:Si (t)<0

and announces a new stock price

 
 a
D(t)+Z(t)
D(t) L2
P (t + 1) = P (t) . (5.16)
Z(t)
While Cont and Bouchaud’s market maker has a symmetric reaction
function to order imbalances, Iori’s price adjustment rule is asymmet-
ric. Finally, the next period’s individual threshold levels are adjusted
according to
P (t)
ξi (t + 1) = ξi (t) . (5.17)
P (t − 1)
Numerical simulations were run for L = 100 agents and for different
parameter values. Neither linking probabilities of zero, the independent
agent scenario, nor probabilities close to the critical point, the Cont-
Bouchaud percolation case, yielded satisfactory price and volume data.
Only for the Ising case, i.e., when agents are linked to their nearest
neighbors with a probability of one, volatility clustering of returns and
a positive cross-correlation between price volatility and trading volume
emerged. These characteristics arose purely from communication and
imitation among traders, even in the absence of an aggregate exogenous
shock, thus supporting the Interacting Agent Hypothesis by Lux and
Marchesi [268]. The asymmetric price adjustment rule and the endoge-
nous adjustment of individual threshold activation levels ξi , however,
seem critical in achieving these interesting results.
clusters. They would choose identical buying/selling decisions and thus, would
cause large stock price fluctuations. Note that setting Ji,j to one with probability
p, and to zero with probability 1 − p, leads to Cont and Bouchaud’s percolation
model.
 5.4 Agent-Based Computational Models of Financial Markets 83

Other examples in the indexpercolation model class of percolation

and Ising-type models of stock markets are [66, 409] and Raberto et al.
[344, 345] who developed the Genoa artificial stock market.
While the assumptions of random individual demands seem to re-
quire only minimally sophisticated agents, Bhamra studies rational
profit maximizing agents in an Ising-like model [32]. In it, he finds
that profit-maximization implies that agents should, to some extent,
imitate each other’s behavior. Thus, the use of the Ising-type mod-
els to describe agent-agent interactions can be justified on economic
grounds.

5.4.3 Models of Chartist-Fundamentalist Interactions

Models investigating the dynamics from two [98] or three investors

types [85, 97] are not new to the finance literature. While relaxing the
assumption of investor homogeneity, the restriction to only a few strat-
egy types, often a chartist and a fundamental trading rule, allowed
those models to be analytically solved. There are, however, several
models of chartist-fundamentalist interaction that require simulation
techniques in order to fully describe their dynamics, some of which will
be discussed below.38
Instead of postulating the simultaneous existence of two different
trader types who stick to their initial trading rules, computational
models of chartist-fundamental interaction often model the endogenous
switching between trading strategies. An example in this line of research
is the artificial stock market by Lux [266], which is further analyzed in
[268, 269, 67].
The market is populated by chartists and fundamental traders. Fun-
damentalists are supposed to know the asset’s fundamental value pf
with certainty. Assuming that the asset price p will sooner or later re-
vert to its fundamental value, fundamental traders simply buy (sell) the
asset when its price is below (above) its fundamental value.39 Chartists,
who are either optimistic or pessimistic about the market’s development
in the near future, derive their price expectations from past asset price
movements.
Agents meet randomly, compare the profitability of their strategies,
and may switch to another trading group if they perceive the other

38
Some early computational models belonging to this group are discussed in [243].
39
The fundamental value in Lux’ model is simply a constant. In [268, 67], an exoge-
nous news arrival process is introduced and the log changes of the fundamental
value are assumed to be Gaussian random variables.
84 5 Replicating the Stylized Facts of Financial Markets

strategy to be more promising in the short run. Thus, even funda-

mental traders may be induced to become chartists since the market
price p may stay above or below the fundamental value pf for several
periods.40 The switching behavior within the chartist group between
optimistic and pessimistic attitudes is modeled through mimetic con-
tagion. A positive opinion index x ∈ [−1, +1], i.e., when there are more
optimistic chartists than pessimists, increases the probability that the
latter jump on the bandwagon and become optimists themselves or
vice versa. Switching between these different trader groups is formal-
ized through six dynamic transition probabilities. Obviously, time vary-
ing fractions of optimistic or pessimistic chartists and fundamentalists
effect the aggregate supply and demand for the risky asset in this finan-
cial market. The asset price dynamics are finally determined through
a market maker whose price adjustment rule reacts sluggishly to order
imbalances. Note that this specification implies that the market maker
has to temporarily absorb additional excess demands.
Lux and Marchesi [269] first analyze the population dynamics in
their artificial market. If z denotes the fraction of chartists, the market
possesses the following stationary solutions:
1) x∗ = 0 and z ∗ = 1 with arbitrary p,
2) z ∗ = 0 and p∗ = pf with arbitrary x,
3) x∗ = 0 and p∗ = pf with arbitrary z.
Lux and Marchesi show that if the opinion index x is biased (x = 0) and
if the asset price does not equal its fundamental value, stable states do
not exist. The two absorbing states 1) and 2) at which neither chartists
nor fundamentalists exist are excluded from the numerical simulations
by additional assumptions. The stability of the equilibria in 3), i.e.,
when the market price equals the fundamental value and when the
chartist group as a whole is neither pessimistic nor optimistic, depends
crucially on the fraction of noise traders z in the market. In the vicinity
of a critical point, the artificial market is characterized by permanent
fluctuations of the portion of both chartists and fundamentalists. Tem-
porary increases in volatility occur when the number of chartists is
relatively large. The market behavior, however, tends to stabilize itself

40
That is, arbitrage profits for fundamentalist tends to be realized over the course
of several trading periods. Capital gains and losses, however, accrue immediately.
This asymmetry might explain why myopic fundamentalists might choose not to
use their information about the asset’s fundamental value.
 5.4 Agent-Based Computational Models of Financial Markets 85

through the tendency of agents to become fundamentalists when there

are large deviations between market price and fundamental value.41
Numerical simulations with 500 agents confirmed the theoretical
results in that extended periods of quiet were followed by sudden bursts
of clustered volatility in returns. The generated return distribution is
thus heteroskedastic and leptokurtotic with its tails following a power
law with a tail index between 2 and 4. Similar to empirical data, the
autocorrelations of squared and absolute returns are positive and decay
hyperbolically while the raw returns are almost uncorrelated. Given
the Gaussian news arrival process and the deviations of the return
distributions from normality, Lux and Marchesi [268, 269] and Chen
et al. [67] conclude that market interactions of agents magnify and
transform exogenous noise (news) into fat-tailed returns with clustered
volatility, a phenomenon they labeled the Interacting Agent Hypothesis
(see also section 5.3.4).
Similar microscopic simulation models of chartist-fundamentalist in-
teractions have been proposed in [19, 121, 151, 152, 153]. Analytical
models of chartist-fundamentalist interactions are surveyed in [190].

5.4.4 Many-Strategy Models with Learning

The models discussed so far were, in terms of model structure and

agent rationality, quite simple. None of the models explicitly assumed
utility maximizing agents. Rather, demands were often random and/or
binary, i.e., agents either bought or sold exactly one unit of stock.
Furthermore, the choices that agents faced were very limited in that
they were restricted to buying or selling one unit of stock or following
a chartist or fundamentalist trading strategy.
Many-strategy models, on the other hand, usually abandon many
of these limitations and focus on the question of emergence. They are
interested in which types of trading strategies will appear and survive
in a dynamically changing market environment. The question whether
the learning agents with their constantly co-evolving mix of trading
strategies will give rise to a rational expectations equilibrium is an-
other commonality in this model class. Often, these models assume a
higher level of agent rationality by using agents who derive their opti-
mal demands by maximizing utility functions.

41
In the physics literature, this dynamic behavior is known as on-off intermittency,
i.e., an attracting state becomes temporarily unstable due to a local bifurcation.
Endogenous forces, though, drive the physical system back to its equilibrium
state.
86 5 Replicating the Stylized Facts of Financial Markets

A simple example is the artificial market by Lettau [246]. Although

the market structure is still very simple, the model takes us one step
further with respect to agent rationality. Instead of having only two
or three decision alternatives, Lettau’s agents are confronted with a
continuum of choices of how many shares x to hold of a risky stock
that pays a stochastic dividend d with mean d¯ and variance σd2 . Even
though agents possess a constant absolute risk aversion utility function
with coefficient λ
U (w) = −e−λw , (5.18)
w = x(d¯ − p) being the net payoff, they do not determine an optimal
amount x . Lettau is rather interested in how close adaptive agents
equipped with an evolutionary GA-learning algorithm could come to
the optimal solution
  1 ¯ 
 d¯ − p =
=α
x 2 d−p (5.19)
λσd

of the maximization problem.42 In his computer model, agents are thus

depicted as binary solution strings43 of length L, where each string
ωt = (ω1,t , . . . , ωL,t ) is an instance of αt in the interval [min, max]. The
binary coding of an agent is thus simply
L j−1
j=1 ωj,t 2
αt = min +(max − min) (5.20)
2L−1
(compare also section 4.4.2). At the beginning of the simulation, all bit
positions of an agent are randomly initialized to either zero or one. Each
period t is divided into S subintervals in which agents order stock ac-
cording to the current value of α. At the end of each period, a new gen-
eration of agents is generated. The probability of agent i being copied
into the next generation is positively correlated with its cumulative
utility in t

42
In order to focus on learning behavior, Lettau assumes that the stock price is
exogenously determined and not influenced by the adaptive agents.
43
Note that this specification is a single-population GA where each agent repre-
sents a solution. In multi-population GAs, each agent holds a variety of different
candidates and the GA is actually run inside each agent. Single-population GAs
are sometimes equated with social learning since the genetic operators are ap-
plied across different individuals. Multi-population GAs then refer to individual
learning since no exchange of genetic material between individuals occurs. Vriend
showed that there are different aggregate outcomes depending on whether learn-
ing is modeled as social or individual learning [431].
 5.4 Agent-Based Computational Models of Financial Markets 87
S


Ucum = Ui (ωi,j ). (5.21)
j=1

S determines the sample length over which a rule’s fitness is evaluated.

After copying has taken place, agents are modified by mutation and
crossover.
The length of the rule evaluation period S turns out to be crucial for
the effectiveness of the learning algorithm. For smaller S, agents hold
considerably more of the risky asset than the optimal amount. For large
S, Lettau’s agents come very close to the optimal portfolio weight α̂,
yet the bias to hold more risky stock remains. The intuition behind this
bias is quite simple. For small sample sizes, those agents that took risks
and did well because of favorable realizations of the random dividends
have a higher reproductive probability than conservative agents. As
the rule evaluation period S increases and agents learn more about the
true dividend distribution, former lucky agents will experience some
rare negative events and the fraction of conservative agents tends to
grow. As S → ∞, the bias vanishes.
Lettau’s basic model is interesting in itself because of the details it
reveals about genetic algorithm learning. He then continues to explain
the flows in and out of mutual funds through a slightly modified model
version. Standard financial theory such as the CAPM [288, 289, 389]
cannot explain why the flows into mutual funds are positively correlated
with returns and why investors act more sensitively to negative returns
than to positive ones. These empirical mutual fund flows, however, are
surprisingly well replicated by Lettau’s market when setting S at 1 and
replacing 3 agents in each period with new randomly initialized ones.
Because Lettau’s artificial market used an exogenous price process
which was not influenced by the agent’s buying and selling decisions,
it is not well suited to replicate the aforementioned stylized facts of
real financial markets. Another market with many co-evolving trading
strategies and agent learning was suggested by Chen and Yeh [68]. Their
basic model framework is similar to Grossman and Stiglitz [164], i.e.,
agents have a constant absolute risk aversion utility function which they
maximize subject to a budget constraint. When deriving their optimal
demands, they need to form expectations about next period’s price and
dividend.44 They do so by acquiring specific forecasting models from a
business school where faculty members engage in creating and publish-

44
The SFI-ASM [244] has the same basic structure and will be introduced in more
detail in the next chapter.
88 5 Replicating the Stylized Facts of Financial Markets

ing new forecasting models.45 The creation of new trading strategies

is implemented by means of genetic programming. Even though Chen
and Yeh do not intentionally fine-tune their model parameters for the
purpose of replicating the stylized facts of financial time series, the
endogenous price and return series still turn out to be non-normally
distributed. Besides fat-tailedness, return series are also independently
and identically distributed (iid), which supports the EMH. Chen and
Yeh emphasize that this feature is especially surprising since martin-
gale traders, i.e., those whose set of trading strategies reflect a belief in
the EMH, become extinct in the course of a typical simulation run. In a
newer paper, Chen and Yeh thus interpret the validity of the EMH on
the macro-level as an emergent property since it should not be expected
from the behavior of individual investors [69].46 In addition, they test
the Rational Expectations Hypothesis by examining the series of ag-
gregate forecasting errors. The mean forecasting errors are found to
be only insignificantly different from zero, thus, on the aggregate level,
they do not make systematic errors in the mean. Furthermore, by look-
ing for linear and non-linear patterns in the error series, the null of
iid-nes cannot be rejected for most subperiods. Thus, Chen and Yeh
conclude that the Rational Expectations Hypothesis is another emer-
gent property in their artificial stock market.

45
This learning structure addresses some methodological criticism concerning the
spreading of trading strategies within populations. It is argued that in single-
population GAs/GPs, only a sequence of actions may be observable, but not the
strategies that lead to these actions. Thus, imitation cannot explain the spread-
ing of strategies within the population. Since Chen and Yeh consider the multi-
population GA/GP approach to be an unsatisfactory response to this criticism,
they resolve this methodological issue through a “business school” in which fac-
ulty members are forced to publish their results.
46
Contrary to their older paper [68], the share of martingale believers is not zero,
but with a share of approximately 1%, very small.
 Part II

The Santa Fe Institute Artificial Stock Market

Model Revisited
6
The Original Santa Fe Institute Artificial Stock
Market

Science is a long history of learning how not to fool ourselves.

Richard P. Feynman1

6.1 Introduction

Similar to the models discussed in the previous section, the Santa Fe

Institute Artificial Stock Market (SFI-ASM) is a model with many
trading strategies which are improved over time. The departures from
the neoclassical framework are only minimal. Agents derive their opti-
mal demands by maximizing a utility function and, therefore, behave
quite rationally in the standard economic sense. Only because they are
assumed to be heterogeneous with respect to their expectations, can
they not be modeled as fully rational.
By focusing on the dynamics of learning, the SFI-ASM identifies a
single parameter, i.e., the learning speed of agents, which is able to shift
the model to either a regime that is close to the homogeneous ratio-
nal expectations equilibrium, or to a more complex regime that better
fits the empirical facts. The complex regime emerges for fast learning
rates and is characterized by more complicated price time series and
by substantial levels of technical trading.
After a detailed description of the SFI-ASM’s basic model structure
and its results, it will be shown in this chapter that the emergence
of the complex regime could be an artifact of design assumption. A
closer investigation of the genetic algorithm that updates the trading
rules of agents reveals that the SFI mutation operator causes a sys-
tematic upward bias in the level of set bits in the condition part of
1
Quoted in Cole [77, p. 10].
92 6 The Original Santa Fe Institute Artificial Stock Market

trading rules. Faster learning speeds with more mutations per time
period imply increased levels of technical (and fundamental) trading.
Thus, postulating the emergence of technical trading by only looking
at the aggregate level of technical trading bits may be too premature.

6.2 The Marimon-Sargent Hypothesis and the SFI-ASM

According to Waldrop [433, p. 270], the development of the SFI-ASM

was a direct result of a discussion that took place at the Santa Fe Insti-
tute in March 1989. Ramon Marimon and Thomas Sargent contended
that adaptive agents in an artificial stock market model would quickly
discover the rational expectations equilibrium solution. In spite of a few
random fluctuations up and down, prices would converge to the fun-
damental values of traded stocks as predicted by neoclassical theory.
Marimon and Sargent bolstered their claim by their own research [285]
in which they had adaptive classifier agents assigned to solve Wick-
sell’s triangle.2 Their artificially intelligent agents were always able to
learn the neoclassical solution previously found by Kiyotaki and Wright
[226], i.e., the good with the lowest storage cost emerged as a generally
accepted medium of exchange.3
John Holland and Brian W. Arthur, on the other hand, could not
conceive of adaptive agents being able to find the neoclassical equilib-
rium solution. In their view, an artificial stock market would be far
more complicated than the simple and well-defined problems that had
been proven solvable by adaptive agents. The rational expectations
equilibrium would require that heterogeneous agents infer their own
expectations by an infinite regress of other agents’ expectations, a task
that Holland and Arthur considered to be too difficult to be solved by
realistically modeled artificial agents. The complications arising from
2
Wicksell’s triangle refers to a situation in which three types of agents would like
to consume the goods possessed by one other type of agent, yet they have nothing
to offer that has any consumption value to these agents. In other words, there are
no double coincidences of wants in the economy. Assuming that the reallocation of
consumption goods has to be quid pro quo, a Pareto efficient allocation can never
be achieved as long as agents are not willing to give up their good in exchange
for something they do not value for their own consumption.
3
Sargent’s general view of the use of genetic algorithms in economics becomes
clearer in his book on bounded rationality in macroeconomics [372]. In it, he
mainly uses genetic algorithms as a substitute to justify rational expectations in
macroeconomic general equilibrium models. This approach, however, is criticized
by Moss and Edmonds who see the role of genetic algorithms unduly reduced to
bring about the rational expectations equilibrium [312].
 6.3 An Overview of SFI-ASM Versions 93

the interactions of many interacting heterogeneous agents would lead

to a complex stock market behavior quite different from a homogeneous
rational expectations regime.

6.3 An Overview of SFI-ASM Versions

Since the debate could not be settled without having an actual version
of an artificial stock market, Arthur and Holland decided to program
their own stock market simulation which became known as the Santa
Fe Institute Artificial Stock Market (SFI-ASM). The first version was
operable by the end of 1989 and was described in [331]. The studies
by Arthur et al. [11] and LeBaron et al. [244] were based on a modi-
fied Objective-C version which used a market clearing mechanism in-
stead of an excess demand price adjustment mechanism. It was later
revised by Brandon Weber and Paul Johnson to run with the Swarm li-
braries, a well-known toolkit for agent-based simulations for Objective-
C and Java. This ongoing effort is currently hosted by Paul Johnson
at http://ArtStkMkt.sourceforge.net [204]. The SFI-ASM has inspired
various modelers to do their own research. Joshi et al. [206, 207] slightly
adapted the original Objective-C version to analyze wealth levels. Tay
and Linn [418] extended the original model by using fuzzy logic for
expectation formation. Wilpert [437], who used his own Borland C++
implementation, tested the model with different modifications. He used,
for instance, the generated profits of trading rules instead of forecast
accuracies as a fitness criterion. Gulyás [166] programmed a participa-
tory market model in which real humans were placing market orders
alongside the artificial adaptive agents. They were thus able to study
the effects on actual human decision making on the market behavior
and vice versa.
For this book, the Objective-C version 7.1.2. was ported to Java
(Java 2 SDK, standard edition, version 1.4.0).4 In order to distinguish
between the original SFI-ASM and the current Java-version with which
the simulation results were derived, I will refer to the latter as either
the Java-version of the SFI-ASM, or, when referring to the suggested
modification of the mutation operator, as the modified SFI-ASM. The
Java-version of the SFI-ASM was programmed by using the Repast
(Recursive Porous Agent Simulation Toolkit) library [78].5 Agent-based

4
The source code is available upon request.
5
Repast is being developed at the University of Chicago and is freely avail-
able under the BSD license agreement. The Repast website is located at
94 6 The Original Santa Fe Institute Artificial Stock Market

simulation toolkits such as Swarm and Repast save agent-based model-

ers from building their computer models from scratch. These libraries
provide non-content specific routines such as simulation interfaces or
data output procedures that are designed and thoroughly tested by
professional developers. Social scientists are thus able to concentrate
on their model specific structures and interactions. The apparent stan-
dardization that comes with the use of a commonly accepted agent-
based toolkit also allows for a more efficient exchange of program code.
This enhances the understanding and readability of the by now many
agent-based computer simulations.

6.4 The Basic Structure of the SFI-ASM

The SFI-ASM borrows much of its structure from Grossman and
Stiglitz [164]. It is inhabited by N traders, who are all initially en-
dowed with one unit of risky stock and 20, 000 units of cash. During
each period, traders have to decide how much to invest in risky stock
and how much to keep in cash which yields a risk-free rate of return
rf .
The stock pays a stochastic dividend per period which is generated
by a mean-reverting autoregressive Ornstein-Uhlenbeck process
dt+1 = d¯ + ρ(dt − d)
¯ + ǫt+1 . (6.1)
d¯ denotes the dividend mean, ρ is the speed of mean reversion, and ǫ
refers to stochastic shocks which are normally distributed with mean
zero and variance σǫ2 . Traders are homogeneous with with respect to
their utility function, i.e., they all have the same constant absolute risk
aversion (CARA) expected utility function
U (Wi,t+1 ) = −e−λWi,t+1 , (6.2)
with λ being the degree of risk aversion and Wi,t+1 being agent’s i
expected wealth level in the next period. In determining their optimal
http://repast.sourceforge.net from which the latest distribution can be down-
loaded. While the reprogramming of the SFI-ASM used Repast 1.4, the Repast
version at the time of writing is 3.1. In a recent comparison of free JAVA-
libraries for agent-based simulation in the social sciences [423], Repast has been
found to be the clear winner. Among the several rating criteria were official pro-
gram documentation, statements by developers and users, and experiences and
impressions by the evaluation team. Another review of agent-modeling toolk-
its is in [148]. A good overview of current agent-based simulation toolkits is
maintained by Leigh Tesfatsion at the University of Iowa and can be found at
http://econ.iastate.edu/tesfatsi/acecode.htm.
 6.4 The Basic Structure of the SFI-ASM 95

demand for the risky stock, agents are perfectly myopic in that they
only consider next period’s expected returns. Agents maximize their
expected utility subject to the budget constraint

Wi,t+1 = xi,t (pt+1 + dt+1 ) + (1 + rf )(Wi,t − pt xi,t ), (6.3)

where xi,t is the amount of stock agent i holds in period t. Under the
assumption of normally distributed stock returns, the optimal amount
of stock x
i,t that agents desire to hold is then determined as

Ei,t [pt+1 + dt+1 ] − pt (1 + r)

x
i,t = 2 , (6.4)
λσt,p+d

where Ei,t [pt+1 + dt+1 ] is i’s expectation in t about next period’s real-
ization of the stock price and dividend, and σt,p+d2 the empirically ob-
served variance of the stock’s combined price plus dividend time series.
LeBaron et al. point out that the normality assumption of stock returns
holds in the homogeneous rational expectations equilibrium (hree), but
outside the hree-regime, it is not clear whether stock returns will be
normally distributed [244].
The effective demand of an agent is the difference of his actual
and desired stock holdings. Once agents have determined their effective
demands, they submit them as well as their partial derivatives with
respect to the price to a specialist, who tries to balance the effective
demands by setting a market clearing price. If the specialist is not able
to find a market clearing price in the first place, an iterative process
is started in which new trial prices are announced and agents update
their effective demands and partial derivatives accordingly. If complete
market clearing is not reached within a specified number of trials, one
side of the market will be rationed.

6.4.1 Trading Rules and Expectation Formation

While traders are homogeneous with respect to their utility functions

and degrees of risk aversion, they differ when deriving their expecta-
tions about future prices and dividends Ei,t [pt+1 + dt+1 ]. In a way, they
differ in processing an identical information set. Price and dividend
forecasts are generated by using individual trading rules of the form

if (condition fulfilled), then (derive forecast).

These “condition-forecast” rules are a modified version of the condition-

action classifier system by John Holland [185]. While the latter maps
96 6 The Original Santa Fe Institute Artificial Stock Market

directly from condition into action, the “condition-forecast” rules do so

only indirectly.6 First, a forecast is produced which, by using equation
6.4, then will be converted into an action, i.e., an agent’s bid or offer
for the risky stock.
Each of the j = 1 . . . 100 trading rules that every agent possesses
consists of a condition part, a forecast part (predictor), a numerical
2 , i.e.,
value Φt,i,j for its fitness, and a value for its forecast accuracy νt,i,j

rulei,j = {(condition part); (predictor); fitness, forecast accuracy}

The condition parts are checked against a Boolean market descriptor

Dt which holds current and past price and dividend information. For
example, a particular market state could be that the price of the stock is
greater than n-times its fundamental value, while at the same time, the
25-period moving average of the stock price is greater than the current
price. When a particular predefined condition is met, the corresponding
descriptor bit is set to 1, and otherwise to 0.
A rule’s condition part, on the other hand, is coded as a ternary
string holding either 1 or 0, depending on whether the corresponding
bit in the market descriptor has to be matched or not, or holding # if
the rule ignores that particular descriptor bit.7
An example of how rules are checked against a market descriptor
with seven conditions A – H is shown in table 6.1. Rule 2 and rule 100
are not activated, since in both cases, two conditions are not fulfilled.
On the other hand, rule 1 and rule 99 match the market descriptor
and hence, are activated. As one can see, rule 1 is rather general. Since
it contains numerous #-signs, it is quite insensitive to changes in the
environment and will probably be activated quite often, as it is the
case in this period. All other rules are more detailed and describe more
specific market states.
The bits of a trading rule may be characterized as either technical or
fundamental. Technical bits check only price information, for instance,
whether the stock price has gone up or down during the last periods,
6
The first SFI-ASM version mapped directly from states of the world into actions.
All later versions use the described condition-forecast rules.
7
Technically, the units in the ternary strings should be called trits. A trit is the
smallest unit that can hold three values. However, as is usually done in the litera-
ture, I will refer to them as bits. Alternatively, when using the terminology known
from genetic algorithms, I will refer to these bits as genes. The condition parts
correspond to chromosomes. The bit positions at which genes reside in the chro-
mosome are called gene loci, and the three possible values that a bit (gene) can
take on are called alleles. The cardinality of the allele-alphabet in the SFI-ASM
is three.
 6.4 The Basic Structure of the SFI-ASM 97

Table 6.1. Comparison of an agent’s rule set against a hypothetical Boolean

market descriptor. Rules with matching condition parts are marked as active.
condition A B CD E F G H match ?
market descriptor Dt 0 0 0 1 0 1 0 1 -
rule 1 0 ###### #

rule 2 # 1 # 0 # 1 # # -
. . . . . . . . . .
. . . . . . . . . .
rule 99 0 ## 1 # 1 # 1

rule 100 0 1 ### 0 0 # -

or whether certain moving averages of prices are bigger or smaller than

other moving averages. Fundamental bits, on the other hand, relate the
price of a stock to its fundamental value by using dividend information.
For instance, prices are checked against a stock’s fundamental value by
comparing for each ratio in the brackets whether
 
1 1 3 7 95 9
price x interest rate/dividend > , , , , , 1, (6.5)
4 2 4 8 100 8
is fulfilled. The conditions used by the SFI-ASM versions differ from
those in the current Java SFI-ASM version.8 Increasing the number of
checked trading conditions might have an effect on what Axelrod [13]
calls distributional equivalence, i.e., different statistical properties of
time series, but the relational equivalence should remain unaffected.
From the set of 100 individual trading rules, normally several rules
match the market descriptor. From the set of activated rules, agents
now have to choose one for their forecast production. LeBaron et al.
use the forecast accuracies of activated rules as a tie breaker, i.e., the
best rule is always selected.9 Finally, agent i determines his forecast of
8
To state how many conditions are actually checked in “the original SFI-ASM” is
quite difficult to say. LeBaron et al. [244] document only 12 conditions (including 2
dummy bits), while the Objective-C version 7.1.2 used by Joshi et al. [206], which
served as the blueprint for the Java SFI-ASM, had a total of 61 conditions with
three dummy bits. The reprogrammed Java version checks a total of 64 conditions,
32 fundamental and 32 technical trading bits. The differences in checked trading
conditions between the three SFI-ASM versions are documented and discussed in
appendix 11.2.
9
Instead of always selecting the most accurate trading rule, Joshi et al. [206, 207]
use the roulette wheel mechanism. This selection algorithm assigns to each rule
a wedge on a roulette wheel which is proportional in size to its relative fitness.
Thus, rules with higher fitness values are more likely to be chosen than those with
98 6 The Original Santa Fe Institute Artificial Stock Market

next period’s price and dividend according to the linear equation

Et,i [pt+1 + dt+1 ] = at,i,j (pt + dt ) + bt,i,j , (6.6)

where at,i,j and bt,i,j are real-valued parameters constituting the pre-
dictor part of the chosen trading rule j. Only when no rules match
the market descriptor, parameters a and b are determined as a fitness-
weighted average of all trading rules in his rule set.
One period later, the accuracy of all activated rules is checked by
comparing their predictions E[pt+1 + dt+1 ] with the actual realization
of (pt+1 + dt+1 ). A rule’s forecast accuracy is determined as
 2
2 1 2 1
νt,i,j = 1 − νt−1,i,j + (pt +dt )−[at,i,j (pt−1 +dt−1 )+bt,i,j ] . (6.7)
θ θ
This forecast accuracy is measured as a weighted average of previous
and current squared forecasting errors. The parameter θ determines the
size of the time window that agents take into account when estimating
a rule’s accuracy. As LeBaron et al. have pointed out, the value of θ is
a crucial design question since it strongly affects the speed of accuracy
adjustment and learning in the artificial stock market. If θ = 1, trading
rules would be judged only on last period’s performance, and forecast
accuracy would be strongly prone to noise. As in LeBaron et al., a value
of 75 is chosen for θ.
2
The forecast accuracy νt,i,j is used as a rule’s variance estimate
2
σt,(p+d) , which is used in equation 6.4 to determine the optimal stock
holdings of agents. Furthermore, it is the main determinant of a rule’s
fitness  2 
Φt,i,j = C − νt,i,j + bit cost × specificity , (6.8)
with specificity being the number of conditions in a rule that are not
ignored, with bit cost as an associated cost for each non-ignored con-
dition, and C as a positive constant to ensure positive fitness.10 Non-
zero-bit costs per set trading bit (0 or 1) could be interpreted as the
cost of acquiring and evaluating new information. It could also be seen
as a complexity aversion, since simple rules are favored over more spe-
cific ones. Most importantly, however, LeBaron et al. emphasize that
positive bit costs would ensures that non-# bits in a trading rule serve

low fitness. The specific problems that arise from roulette wheel selection were
already discussed in section 4.4.2 on page 41.
10
The maximum squared forecast error of a trading rule has an arbitrary ceiling of
100, hence, the numerical value for C is set at this value.
 6.4 The Basic Structure of the SFI-ASM 99

a useful purpose, i.e., that they have some informational content.11

Since bit costs bias the resulting bit distribution towards the all-#
rule, LeBaron et al. claim that non-# trading bits will only survive if
they have some predictive value.

6.4.2 Learning and Rule Evolution

So far, agents have been equipped with a static rule set. Feedback learn-
ing in the stock market has taken place by identifying and using the
rules that produced better forecasts than others. The quality and the
speed of this type of learning was strongly dependent on the parameter
θ which determined over how many past periods agents averaged the
forecast accuracy of their trading rules. However, if agents started with
a rule set that contained only bad performing rules, in the absence of
any other learning mechanism, they would not be able to find better
ones.
Agents therefore use an additional genetic algorithm (GA) learn-
ing procedure that allows them to alter their rule set by replacing
poorly performing rules with new, possibly better ones.12 Exploratory
GA learning usually happens on a slower evolutionary time scale than
the feedback learning and examines the search space in a random, yet
not directionless fashion. For each agent, the GA is, on average, asyn-
chronously invoked every K periods and replaces the 20 worst rules of
the rule set. The GA-invocation interval K alters the learning speed of
agents and turns out to be the most crucial model parameter.
New trading rules are created either through mutation (with pre-
dictor mutation probability Π = 0.9) or through crossover (with prob-
ability 1 − Π). Crossover is a sexual genetic operator that needs two
parents to work, both of whom are chosen by tournament selection.13
Even though there are a variety of different crossover operators avail-
able, the SFI-ASM uses exclusively uniform crossover for the condition

11
“The purpose of this bit cost is to make sure that each bit is actually serving
a useful purpose in terms of a forecasting rule” (p. 1497). Emphasis added by
author.
12
See, also, section 4.4.2 on GA learning.
13
For tournament selection in the SFI-ASM, two genetic individuals, i.e., trading
rules, are randomly selected from the gene pool, and the fitter of both is chosen.
Goldberg [154] points to another popular tournament selection algorithm by Wet-
zel. Instead of randomly selecting two candidates, Wetzel-ranking picks the two
candidate solutions by using roulette wheel selection. It is not clear whether sim-
ple tournament or Wetzel-ranking is superior or whether it greatly affects results,
yet Wetzel-ranking should return, on average, fitter parents.
100 6 The Original Santa Fe Institute Artificial Stock Market

parts. Here, an offspring’s bit is chosen with equal probability from the
corresponding bit positions of either parent.

Table 6.2. Example for uniform crossover on the condition part.

parent 1 # 0 1 # # # # 1 1
parent 2 # 0 1 1 1 0 0 0 0
offspring # 0 1 # 1 # 0 1 0

Note that the fraction of set trading bits in an offspring is an un-

weighted average of the two parents’ bit fractions. There is no sys-
tematic influence on the average specificity through the working of the
crossover operator.
As for the real-valued forecasting parameters a and b, one of three
methods is randomly chosen with equal probability. First, both param-
eters are exclusively taken from one randomly selected parent. Second,
each individual parameter is selected from either one of the two parents.
Third, new parameter values are created by determining a weighted av-
erage of the two parents’ values, with 1/σj,p+d2 as the weight for each
parent. The weights are normalized to sum up to one.14
It is apparent that the crossover operator is incapable of introduc-
ing either zero, one, or #-bits to a certain bit position if they are not
yet contained in the parent gene pool for that particular bit position.
Crossover is also unable to generate real-valued parameters for a and
b that are outside the interval created by the minimum and maximum
values in the parent gene pool. These limitations are overcome by mu-
tation, which is an important part for any evolutionary algorithm. It
helps to maintain a diverse population by forcing new solutions into
the population and thus, avoids premature convergence of the search
algorithm. For mutation, one parent is chosen by using tournament se-
lection. First, a genetically identical offspring of that parent is created.
The real-valued parameters of the predictor parts are mutated in one of
three possible ways. With probability 0.2, they are uniformly changed
to a value within the permissible ranges of the parameters which is
[0.7, 1.2] for the a parameter and [−10.0, 19.0] for the b parameter.
With probability 0.2, the current parameter value is uniformly dis-
14
LeBaron et al. pointed out that there is little experience in the GA community
on how to perform real-value crossover. By now, it has become more common and
the early distinction between genetic algorithms and evolutionary strategies has
become fuzzy.
 6.4 The Basic Structure of the SFI-ASM 101

tributed within ± 5% of its current value, and it is left unchanged for

the remaining cases.
The individual bits in the condition parts are mutated with a small
bit mutation probability of π = 0.03. Once a bit in the condition part
has been chosen for mutation, it will be changed according to the fol-
lowing bit transition probabilities:

0 1 #
 
0 0 1/3 2/3
P= 1  1/3 0 2/3  . (6.9)
# 1/3 1/3 1/3
This matrix of transition probabilities specifies that a 0-bit is
changed with a probability of one third to a 1-bit, and with probability
of two thirds to a #-bit. Similarly, a 1-bit changes with one third prob-
ability to 0, and with two thirds to #. Don’t care signs # change with
equal probability of one third to either 1 or 0, or remain unchanged.
LeBaron et al. assert that these transition probabilities would, on av-
erage, maintain the specificity, i.e., the fraction of #’s in a rule.
For crossover, the offspring inherits the average forecast accuracy of
its two parents. For mutation, the offspring’s forecast accuracy is set
at the median forecast error over all rules in the agent’s rule set.

6.4.3 Other Programming Details and Initialization of

Model Parameters

Trading rules in the SFI-ASM contain bit sequences that belong to-
gether. Within such a sequence, certain bit combinations are invalid
and constitute an illogical trading rule.

Table 6.3. Example of an illogical trading rule.

price x interest Rate 1 3 7
dividend > 2 1 98 54 32
4 8
rule 1 ##01#0#

An example of an illogical trading rule is shown in figure 6.3. It

describes a situation in which pt rf /dt is smaller than 7/8, yet at the
same time, that ratio is supposed to be greater than 1. Obviously, rules
of this type will never be matched, but they will get inserted into an
agent’s rule set since the genetic algorithm does not check the trading
rules it creates for logical consistency.
102 6 The Original Santa Fe Institute Artificial Stock Market

The SFI-ASM deals with this problem in two ways: through a large
rule set, and with a generalization procedure. Once a trading rule has
not been matched for more than 4,000 periods, it is generalized by
converting one fourth of its 0 and 1 trading bits to #. The fitness of a
generalized rule is set at the median value.
Agents also face some trading restrictions in that their orders are
constrained to lie within their budget constraint and that they cannot
go short more than five shares. In addition, the stock price is capped
by the specialist at 200 and is bounded from below at 0.01. These
constraints, however, seem to be binding only in the beginning of the
simulation when the randomly initialized trading rules of untrained
agents result in causing the stock price to fluctuate wildly. The same is
true for a constraint that limits the squared forecast error to 100. This
allows us to choose the constant C in equation 6.8, on page 98, such
that fitness values are always positive.
A graphical depiction of the timing sequence of major activities in
the SFI-ASM can be found in appendix 11.1

6.5 The Homogeneous Rational Expectations

Equilibrium

Since agents are assumed homogeneous with respect to their prefer-

ences, the known structure of the dividend process makes it possible
to determine the properties of the homogeneous rational expectations
equilibrium (hree). Since the hre-equilibrium is in the set of possible
outcomes in the SFI-ASM, it will be interesting to see whether agents
will be able to converge to this hree-solution.
A rational expectations equilibrium specialist would determine an
equilibrium price such that the expectations of agents

Ethree (pt+1 + dt+1 ) = ahree (pt + dt ) + bhree (6.10)

are, on average, fulfilled. In the case of a linear rational expectations

equilibrium, we can assume a linear function mapping the current div-
idend into a price
phree
t = f dt + g. (6.11)
We seek the parameters f and g for the hree-specialist and the param-
eters ahree and bhree for the agents forecasts that are compatible with
a hre-equilibrium. Since agents are homogeneous and hold exactly one
unit of the risky stock in this benchmark scenario, the known structure
 6.6 The Marimon-Sargent Hypothesis Refined 103

of the dividend process allows us to determine the hree-parameters for

the specialist as
ρ
f= (6.12)
1 + rf − ρ
and
(1 + f )(1 − ρ)d¯ − λσp+d
2
g= (6.13)
rf
with
2
σp+d = (1 + f )2 σǫ2 (6.14)
being the variance of the combined price plus dividend time series. For
the agents, their hree-forecast parameters compute to

ahree = ρ (6.15)

and  
bhree = (1 − ρ) (1 + f )d¯ + g . (6.16)
The hree-specialist thus sets the hree-price according to the given div-
idend. This price then ensures that all agents want to hold exactly one
unit of stock, given their hree-forecast of next period’s price plus div-
idend. The normal model behavior for heterogeneous agents can now
be assessed by comparing it with the properties of the homogeneous
rational expectations equilibrium.

6.6 The Marimon-Sargent Hypothesis Refined

Having determined the properties of a hre-equilibrium within the SFI-

ASM model structure, it is now possible to refine the Marimon-Sargent
Hypothesis so that adaptive agents would quickly converge to this
benchmark solution. Within the SFI-ASM framework, the Marimon-
Sargent Hypothesis comprises three parts.
First, the hre-equilibrium is characterized by a no-trade situation
since no agent will ever deviate from his optimal position of holding
exactly one unit of the risky stock. Given that the ongoing stochas-
tic GA learning always introduces some noise into the system, minor
trading activities will be inevitable.
Second, equation 6.11 tells us that the statistical properties of the
price process are simply a linear transformation of the known properties
of the stochastic dividend process. In the hre-equilibrium, the price
process should thus be linear, normally distributed, and non-skewed.
104 6 The Original Santa Fe Institute Artificial Stock Market

Finally, equation 6.10 reveals that agents only need to know last pe-
riod’s price and dividend to derive their forecasts about next period’s
price and dividend, given the mean-reverting dividend process. There-
fore, all the information provided by the condition parts of the classifier
system is unnecessary. Given that the use of technical and fundamen-
tal trading bits is punished with associated bit costs, the hree-solution
should be characterized by complete negligence of technical and funda-
mental trading bits.

6.7 Simulation Results of the SFI-ASM

6.7.1 Time Series Behavior

The simulation results obtained by the original SFI-version are doc-

umented in LeBaron et al. [244]. Since the computer model used in
this book was a reprogrammed Java-version, it was necessary to test
whether the two model versions behave similarly for the same param-
eter settings. This was done by running the same statistical tests on
the time series and comparing the results with those published in [244,
p. 1499-1512]. All model parameters were set to the same values re-
ported there.
In order to allow for sufficient learning to occur, price and dividend
time series were recorded between the periods 250,000 and 260,000. The
simulated data of 25 simulation runs for two learning speeds were then
compared by LeBaron et al. to the hree-benchmark case. For the slow
learning case, the GA was invoked by the agents every 1,000 periods,
on average. In the fast learning regime, that parameter was set to 250
periods.
In the hree-mode, the dividend and market price should be a linear
function of their first order lags. Therefore, they are regressed on a lag
and a constant

pt+1 + dt+1 = a(pt + dt ) + b + ǫt , (6.17)

and the estimated residual time series ǫ̂t is analyzed to see whether it
satisfies being iid and N(0,4) distributed. The results are summarized
in table 6.4.
First of all, one notices that for both learning speeds, the new Java-
version of the SFI-ASM produces time series that are close to those of
the original SFI-ASM, but generally a little bit further away from the
hree-benchmark. The standard deviations in the residuals are slightly
 6.7 Simulation Results of the SFI-ASM 105

Table 6.4. Comparison of simulated data of the reprogrammed SFI version

with the reported data from the original Objective-C version. Means over 25
runs. Numbers in parentheses are standard errors estimated using the 25 runs.
Numbers in brackets are the fraction of tests rejecting the no-ARCH or iid-
hypothesis for the ARCH and BDS tests, respectively, at the 95% confidence
level.
Description GA 1,000 GA 250
Java-SFI Obj.-C SFI Java-SFI Obj.-C SFI
Std. Dev. 2.145 2.135 2.225 2.147
(.010) (.008) (.015) (.017)
Excess kurtosis 0.085 0.072 0.285 0.320
(.015) (.012) (.033) (.020)
ρ1 0.031 0.036 0.025 0.007
(.008) (.002) (.012) (.004)
ARCH(1) 3.502 3.159 25.34 36.98
[0.60] [0.44] [1.00] [1.00]
ρ21 0.020 0.017 0.075 0.064
(.002) (.002) (.008) (.004)
BDS 1.41 1.28 4.08 3.11
[0.32] [0.24] [0.92] [0.84]
Excess return 2.92% 2.89% 3.25% 3.06%
(.03%) (.03%) (.08%) (.05%)
Trading volume 0.364 0.355 0.854 0.706
(.025) (.021) (.065) (.047)

higher, thus, indicating a small increase in price variability. Excess kur-

tosis, albeit small, is positive for both the fast and slow learning cases.
Empirical return distributions are usually more fat-tailed and higher-
peaked than the simulated return distributions. The autocorrelation in
the residuals, as shown in the third row, demonstrates that there is lit-
tle linear structure remaining except for the extreme case of updating
the rule set in every period. As LeBaron et al. indicate, any artificial
stock market should exhibit negligible autocorrelations since they are
very low for real markets.
The next row reports the means of the test statistics for the ARCH
test proposed by Engle [110]. The ARCH dependence in the residuals
increases with learning speed and is slightly higher for the Java-version.
The table gives the means of the test statistics, averaged over 25 sim-
ulation runs. The numbers in brackets tell the fraction of runs that
reject the no-Arch hypothesis at the 95% confidence level. In row five,
the first order autocorrelation of the squared residuals is another test
106 6 The Original Santa Fe Institute Artificial Stock Market

for volatility persistence. Again, it increases for faster learning speeds,

but is generally lower for the original model version.
The BDS test in row six is a test for nonlinear dependence devel-
oped by Brock et al. [58]. Its test statistic is asymptotically standard
normally distributed under the null hypothesis of independence.15 One
can notice an increasing amount of nonlinearities for faster exploration
rates, yet again, it is slightly higher for the new Java-version. Trading
volume, which should be zero in the hree-case, increases significantly for
faster learning speeds. This points to a greater degree of heterogeneity
between the agents.
Overall, the conclusion that the learning speed affects the price se-
ries behavior is confirmed by the new model version. The observed
changes in both model versions have the same direction, and the ab-
solute differences between them are relatively small. The tendency of
the new Java-version being further away from the hree-benchmark case
could be a result of the substantial increase in checked trading condi-
tions. The original version checked only 12 conditions, while the new
Java-version increased this to 64 trading conditions.

6.7.2 Forecast Properties

When analyzing whether the agents made use of the trading informa-
tion given in the condition parts of their trading rules, LeBaron et al.
analyzed the fraction of set trading bits, i.e., the fraction of all non-#
bits, averaged over all the rules of all agents. They reported that for
both learning speeds, the equilibrium bit level settles to levels below
5%. More important, though, is in their view the elevated bit level
for the fast learning case, which they define to be at a GA-invocation
interval of 250, with an indication of a continued increase beyond the
250,000th period. One long-run test with 1,000,000 periods yielded that
the bit level eventually stops increasing, but it continued to show large
swings. When focusing only on the 4 moving average trading bits in
the condition parts, they found the increase in the bit level to be even
more pronounced than in the slow learning case.
Figure 6.1 shows that the new Java-SFI version does not exactly
replicate the described bit behavior. At period 250,000, the aggregate
bit level for the fast learning case is slightly below the bit level for the
slow learning case. While this simulation evidence was at first disturb-
ing, extending the simulation horizon revealed that the equilibrium bit
15
There are two free parameters for this test. The distance r is measured as a
fraction of the standard deviation and was set to a value of 0.5, while for the
embedding dimension m, a value of two was chosen.
 6.7 Simulation Results of the SFI-ASM 107

.......
0.12 ....... ............
..
Total fraction of bits set ...
..
.. ....
................ ...
...
........ .. .. ..
..... ....... .....
0.11
...
.......... .... ...
...
....... .... .... ...
.......... ... ... ...
...
........... ...... .... ...
0.1 .......... .......
...... ......
...... ...... ....
..... ........ ...
...
..
...
...
...
...
...
......... ... .. ...
.. ......... ... ........................ ....
0.09
... ........... .... ..................
... ................. ... .................................................................
.... ........... ..... . ..................
. .
................................................................... ... .........................
......................... ................. . .................................................................
... ....................... .............. ...... . .................. .
... ......................................................... ................................................................................ ....... . .
... ........... ................................................ .............. .. .. . ...... .............................................................................................................
0.08 .... ........ ........ ............................................................................................. . . ...................
..... ................ ...... ................................. .................................................................................... ..... .........
.... . . ...... .. . . . . ... ............................ ....... ....... ..................................................................................................................................... .... ...
..... . ............................. ................ ...................
...... .... ................................. .................................
........ ........ ............... .........................
............ ........ ............................ .............................. ...
...................... ... ......... ............................................. .....
0.07
.............. ........ . ... . .... . .
......................................................... . .... ..............
...... .. ... . .. .
...................... .. .. ...................................................
.................................................................. ............................. .
.......................
..................................................... .................................................. . ..
.......................................................................... .... .........................................................
............................................................................. ...........
................................................ ..............................................................................................................................................................
.... .................................................................................. . .

0.06
............................................................................ ..... ..
. .... ....

0 200,000 400,000 600,000

Periods

Fig. 6.1. Total fraction of bits set as a function of learning speed in the
replicated Java version of the SFI-ASM. Data were averaged over 5 separate
runs at different random seeds for a GA-invocation intervals of 1,000 (bottom
at period 750,000), 250, 100, 75, 50, and 25 (top at period 750,000).

levels had not been reached at period 250,000. Even after 750,000 peri-
ods, some of the curves in figure 6.1 did not settle down to their equi-
librium levels. However, the basic statement of the original Objective-C
SFI-ASM can be confirmed. There is an increase in the aggregate bit
level for faster learning speeds.
When trying to identify the reasons why the adjustment processes
towards bit equilibrium obviously take longer than in the original ver-
sion, one immediately thinks of the increase in checked trading con-
ditions. The original version had a meager amount of 12 conditions
(including two dummy bits), while the current Java-version checked a
total of 64 trading conditions.16 It seems logical that when agents have
many more opportunities to combine trading conditions to create elab-
orate trading rules, it will take much longer to test them all. Chances
16
The hypothesis that an increase in trading conditions leads to an increase in
adjustment time was confirmed on two occasions. At first, there were only 57
conditions encoded in the Java-version. The two long integers that represent the
condition parts, however, can hold up to 64 conditions. When making full use of
this available space, a slight increase in adjustment time was observable. When
realizing that the break points in the Objective-C version of Joshi et al. [206]
were not optimally chosen, a reassignment of the break points much closer to the
one’s used by LeBaron et al. resulted in an even higher increase in adjustment
time and a higher equilibrium bit level.
108 6 The Original Santa Fe Institute Artificial Stock Market

are also higher that they end up with a higher fraction of useful trading
rules. This could explain the higher equilibrium levels that are attained
by the new Java-version. There is, however, one caveat at this point.
When starting with different initial bit probabilities, the final equilib-
rium bit levels are different, too. There seems to be a path dependency
hidden in the model structure. The reasons for this path dependency
will become clear in the final chapter.
In summing up the behavior under the two learning speeds, Arthur
et al. [11] label the slow learning case as the rational expectations
regime. The price closely tracks the value predicted by the homogeneous
rational expectations equilibrium, and the trading volume is small. The
low number of set trading bits implies that the information contained in
the condition parts is largely neglected by agents. The fast learning case
is characterized by Arthur et al. as the complex or rich psychological
regime. The price series exhibit larger deviations from the hree-case
such as a larger kurtosis and volatility persistence. Equally important,
though, is the increase in the level of set trading bits. Agents appear
to create more trading rules that can exploit useful information in the
price series.
The difference in bit levels between the two GA-invocation intervals
of 1,000 and 250 shown in figure 6.1 does not seem that impressive in
the new Java implementation. The increases in bit levels for learning
speeds faster than 250 are more apparent. Even though the boundary
between the slow and fast learning regime is not a strict one, I prefer
to speak of a complex regime for learning speeds equal to or smaller
than a GA-invocation interval of 100.

6.8 A Potential Problem: A Biased Mutation Operator

For the creators of the SFI-ASM, the increase in technical trading bits
for faster learning speeds pointed to emergent technical trading. Be-
cause of the cost they had attached to every non-# bit, they conjec-
tured that emerging trading bits must have, on average, some fitness-
based advantages by producing more accurate forecasts. Their intuition
guided them to conclude that the classifier systems enabled agents to
detect short-term trends in the price series upon which they started to
act. At the same time, the price dynamics became more complicated,
yet more realistic, and fast learning regimes were labeled as complex.
The dependence of equilibrium bit levels on learning speed became one
of the main results of the original SFI-ASM and has been replicated by
subsequent studies, e.g., by Joshi et al. [206, 207]), and Wilpert [437].
 6.8 A Potential Problem: A Biased Mutation Operator 109

Arthur et al. asked themselves to what extent the existence of

the complex regime is an artifact of design assumptions in their
model. They found “by varying both the model’s parameters and the
expectational-learning mechanism, that the complex regime and the
qualitative phenomena associated with it are robust. These are not an
artifact of some deficiency in the model ” [11, p. 35].
However, there seems to be a decisive influence of another model
parameter on the simulated outcome that has been neglected so far.
When varying the rate of mutation in the SFI-ASM, one notices a
strict dependence of the equilibrium bit level.

0.125
..
.....
... ...
... ....
...... ...
Total fraction of bits set

. .
..... ....
0.1 .............
........
...
....
.....
Π = 0.90
....... ...........
..... ...... ............
...... .. .......
...... ... .............
...... .... ........................
........................
...... ... ..................
..... . .... .................................
............................
Π = 0.75
...... .... .................................
.....
0.075
...... .......... .......................................
..................................... .
...... ................... .............................................................
...... ......................... .
... .... .
...............................................................................
......... .. .......................... ..................................................
... .... ..............
................................................
..... ... ................................................................ .
....................................................................................
...... .. . . . . ..... . .
...... .. . ........ ........ .. . ..... . .
. . ................................................................................................................. .......
...... ... . ... . .
...... ...
...... ...

0.05
...... ...
...... ...
..... ...
......
... ...
.. ...
Π = 0.50
.....
.........
.......................................... .....
.... ...................................................................................................................................................................................................................................................................... ............................... ...............
.. .. .. . . . . .... . ..... .....................................................................................
.. ..
.. ..
Π = 0.25
... ...
.. ...
.. .... .......... ..............................................................................................................................................................................................................................................................................................................................................................................................
..
... ................................................................................ ...... .
0.025 ..
...........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................

Π = 0.00

0.0
0 200,000 400,000 600,000
Periods

Fig. 6.2. Total fraction of bits set as a function of mutation probability Π in

the Java-version of the SFI-ASM. Data were obtained from a cross-section of
5 separate runs at different random seeds for a GA-invocation interval of 100.

It is apparent from figure 6.2 that the level of bit usage in the
model does not only depend on learning speed, but also on mutation
probability.17 It should thus be crucial to investigate why increasing
17
In figure 6.2, the predictor mutation probability Π was adjusted. A similar effect
could have been achieved by changing the bit mutation probability π and holding
Π constant. Figure 6.2 nicely demonstrates that the GA converges quickly for no
mutation at all. The more agents experiment, the longer it takes them to arrive
their equilibrium bit levels.
110 6 The Original Santa Fe Institute Artificial Stock Market

the mutation probability, and hence, decreasing the rate of crossover,

has a bit-increasing effect on the condition parts. Can we really be sure
that the emergence of the complex regime is not an artifact of design
assumption in the SFI-ASM? Postulating emergent technical trading
due to faster mutation rates would, after all, be hard to justify. A
technical explanation based on model design seems more appropriate
to explain this behavior.
Increasing the rate of mutation in the SFI-ASM model means de-
creasing the rate of crossover. That is, the equilibrium bit level could
also be affected by either the crossover operator or the mutation oper-
ator. Recall, however, that the uniform crossover operator chooses an
offspring’s bit with equal probability from the corresponding bit posi-
tions of either one or the other parent. Hence, the offspring’s fraction of
bits set is an unweighted average of the two parents’ bit fractions, thus,
no systematic influence on the resulting bit level through the working
of the crossover operator can be stated.
The picture is quite different for the mutation operator though. In
order to analyze its influence on the bit dynamics, we have to look
again at the matrix of bit transition probabilities
 
0 1/3 2/3
P =  1/3 0 2/3  . (6.18)
1/3 1/3 1/3

LeBaron et al. have asserted that these transition probabilities would,

on average, maintain the specificity, i.e., the fraction of #’s in a rule.
However, a Markov chain analysis reveals that this statement is not
true. If we denote the vector of probabilities of the three possible states
in period t as pt = {pt0 , pt1 , pt# }, in equilibrium pt = pt P = pt+1
must hold. By repeatedly invoking the mutation operator, the vector
of probabilities
  will converge to its equilibrium distribution of p∗ =
1 1 1
4 , 4 , 2 , i.e., on average, a quarter of all bits will be zero, another
quarter will be one, and the remaining fraction of one half will be the
don’t care sign #. That is, the pure undiluted effect of mutation is to
drive the equilibrium bit level towards its fixed point of one half.18

18
An alternative approach to derive the fixed point of one half is to start by denoting
the initial fraction of bits set before mutation with P ∈ [0, 1]. The non-# bits are
′ 1
mutated to non-# bits with a probability of P|0,1 →1,0 = 3 P , while the probability
′
that a #-bit is mutated to either 1 or 0 equals P|#→0,1 = 32 (1 − P ). Thus, for
any given P , the fraction of bits set after mutation is determined by adding the
′ 1 ′
two probabilities above, i.e., P|0,1,# →1,0 = 3 (2 − P ). Since P|0,1,#→1,0 ∈ [0, 1] is
a continuous function for all P ∈ [0, 1], we know by a fixed point theorem that
 6.8 A Potential Problem: A Biased Mutation Operator 111

This analysis shows that the mutation operator is not neutral to

the initial level of non-# bits. Because the model usually functions well
below this level, the mutation operator introduces an upward bias in
the bit distribution. Consequently, when increasing the learning speed,
the mutation operator is invoked more often per time period, and its
upward bias results in a higher equilibrium bit level.
The influence of the mutation operator on the equilibrium bit level
probably went undiscovered so far because the theoretical level of one
half is usually far from being attained. The key parameter that exerts
a downward influence on the bit level is the fractional rule replacement
through the GA. Usually, only a fraction of trading rules, typically one
fifth, is changed by the GA. Another downward pressure is brought
about by the positive bit costs which causes the GA to favor rules with
low specificity for reproduction.

0.5 p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p pppppppppppp pppppp ppppppppppppppppppppppppppppppp pp ppppp ppp p ppp ppppppppppppppppppppppppppp pppppppppppppppppppppppppppppppppppppppppppppppppppp pppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppp ppppppppppppppppppppppppppppppp ppppp pppppppppppppppppp ppppp ppppp pppppppppppppppppppp ppppppppppppp
Fractions of bits set

p p p ppp ppp
ppp p p p p p
pppp pp p pp p pp pp pp p ppp
p
pp
pp pp
0.25 p p pp p
p pp p p
pp p
pppp
ppppppppppppppppppppp technical bits
p p pp p p p
ppp pp pp pp pp pp fundamental bits
0.1 ppp
p
p ppp
ppp
0
0 50 100 150 200 250
Periods

Fig. 6.3. Fixed point convergence of trading bits in the SFI-ASM for Π = 1.0
and π = 1.0. Further parameter settings were 25 agents, a GA-invocation
interval of 10, no bit costs, and a complete generational replacement of all
trading rules.

′
an equilibrium exists. By repeatedly invoking the mutation operator, P|0,1,# →1,0
converges to its equilibrium value of one half. This result also holds if every bit
in the bitstring is mutated with a probability of less than one. In the SFI-ASM,
this mutation probability π is set at 0.03. Deriving the fraction of non-#  bits
′ 2 2
in the same manner as above yields P|0,1,# →1,0 = 3 π(1 − P ) + P 1 − 3 π . It is
easy to check this formula by setting π = 1, which will yield equation (18), or
′
by setting π = 0, which will yield P|0,1,# →1,0 = P since no mutation will ever be
performed. The equilibrium level equals 1/2, even though convergence occurs a
little bit slower than before.
112 6 The Original Santa Fe Institute Artificial Stock Market

When using a complete generational replacement of trading rules

and setting bit costs to zero, figure 6.3 shows that the equilibrium bit
level quickly converges to its theoretical level of one half. To speed
up convergence, predictor mutation probability Π and bit mutation
probability π are set at their maximum values of 1.0, and the GA-
invocation interval has been set to a very high rate of 25. For the fastest
possible learning rate of GA=1, the convergence to one half would
occur immediately.19 In the long run, the generalization procedure also
introduces a slight downward pressure on the bit level by converting
some of the 0 and 1-# bits in illogical rules to #-bits.
In light of the upwardly biased mutation operator, one has to be cau-
tious about the claim by LeBaron et al. that positive bit costs would
ensure that each surviving trading bit contains useful information. Be-
cause of the constant upward pressure on the bit level, the simple ex-
istence of trading bits does not necessarily mean that they are useful.
LeBaron’s argument would imply that the model could be forced into
a zero-bit solution if bit costs were only high enough. It can be shown,
however, that there will always be some trading bits in the rule sets of
agents, no matter how big the associated bit costs are.

19
Upon closer inspection, however, the mean bit levels seem to hover slightly below
the value of half. This either points to an unidentified parameter that also affects
the equilibrium level, or a minor plus/minus one programming problem when
determining the bit fractions in the model. A careful analysis of the source code,
however, leads me to believe that the latter is highly unlikely.
7
A Suggested Modification to the SFI-ASM

An expert is a man who has made all the mistakes which can
be made in a very narrow field.
Niels Bohr

7.1 Introduction

Along with complex price series behavior, the endogenous appearance

of technical trading bits for faster learning speeds was one of the most
striking results of the SFI-ASM. However, the theoretical and experi-
mental analysis in the previous chapter has shown that the interpreta-
tion of this increase in trading bits as emergent technical trading may
be a design artifact caused by a biased mutation operator. Announc-
ing emergent technical trading simply because the number of trading
bits increases for faster learning is premature when a biased mutation
operator injects them at increasing rates.
This chapter, therefore, suggests a modification to the SFI-ASM by
developing an unbiased mutation operator. The following simulations
show that the results are drastically altered with respect to bit lev-
els. All agents now seem to discover the hree-solution of non-bit usage,
suggesting that the classifier system does not provide any useful in-
formation in terms of improved forecast predictions. Since the finding
of zero-bit usage in equilibrium is completely different from that of
the original SFI-ASM, the remainder of this chapter then focuses on
checking the reliability and stability of the zero-bit solutions. Finally, a
rule consistency check as an additional improvement over the original
SFI-ASM is introduced.
114 7 A Suggested Modification to the SFI-ASM

7.2 An Unbiased Mutation Operator

In order to derive valid conclusions about the bit usage in the model,
one should take care in designing bit-neutral operators and procedures.
Bit-neutral refers to the feature of leaving the fractions of set bits unal-
tered, unless an impact is explicitly desired. The bit decreasing effect of
the bit cost parameter, for instance, is desirable as it is a fitness-based
influence. The bit-increasing effect of the SFI mutation operator, on
the other hand, seems problematic since it is completely technical and
economically uninterpretable.
The suggested alternative bit-neutral mutation operator works with
dynamically adjusting bit transition probabilities. In order to infer
whether technical and fundamental bit usage differs in the stock mar-
ket, this mutation operator works separately for fundamental and tech-
nical trading bits.1 Therefore, it is necessary to distinguish between the
initial fraction of fundamental bits set Ff und. , and the initial fraction
of technical bits set Ftechn. . The transition matrix for the fundamental
bits is then given by
 
0 Ff und. 1 − Ff und.
Pf und. =  Ff und. 0 1 − Ff und.  , (7.1)
1 1
2 F f und. 2 Ff und. 1 − Ff und.

and, similarly, for the technical bits by

 
0 Ftechn. 1 − Ftechn.
Ptechn. =  Ftechn. 0 1 − Ftechn.  . (7.2)
1 1
2 Ftechn. 2 Ftechn. 1 − Ftechn.
t
It is easy to verify that these transition matrices ensure that Ftechn. =
t+1 t t+1
Ftechn. and Ff und. = Ff und. , i.e., the fractions of set bits remain, on
average, unaltered. Unlike the original SFI mutation operator, there
is no built-in attractor towards which the resulting bit distribution
converges. Without an artificial upward bias, surviving trading bits
should only emerge through competition and fitness considerations,
implying that they indeed contain useful information.
While the predictor mutation probability Π and crossover probabil-
ity 1−Π in the stock market model remain unchanged, the effective rate
1
Theoretically, there is no clear distinction between fundamental and technical
trading bits, neither of which have any use in a rational expectations equilib-
rium. One reason to treat them differently is motivated by the SFI-ASM, which
explicitly distinguishes between these two types of trading bits. Note that the
possibility of divergent bit behaviors does not create a bias for it.
 7.3 Simulation Results with the Modified SFI-ASM 115

with which an individual bit in a trading rule is mutated is constantly

adapted such that the expected number of set trading bits before mu-
tations equals the expected number of set trading bits after mutation.
Self-adapting operator rates in a GA have recently been investigated
by Gomez who, however, based the adaption mechanism on the per-
formance achieved by the offspring [157]. If a genetic operator such as
mutation produced fitter offspring than those of another operator such
as crossover, its invocation rate was subsequently increased. Gomez
found such self-adapting operator rates quite efficient for a variety of
problems.2
When analyzing the properties of the new mutation operator, two
points in the possible bit distribution deserve attention. Once the bit
level has reached zero, the updated mutation operator will become in-
active. A state j that may never be left again once it is entered is known
in Markov chain analysis as an absorbing state. Its transition proba-
bility is pjj = 1. At a bit level of one, a 1-bit will always be changed
to zero and vice versa, but never to the don’t care sign #. A state j
at the one-bit level thus has an alternating counterpart, both of which
are repeated after two mutation steps. Therefore, the zero and one bit
levels are limiting distributions in the sense that once agents have ar-
rived at either of them, they will be stuck there forever. However, these
corner solutions are not attracting states in that the bit distribution is
not torn towards either of them. Because of these two limiting distri-
butions, one could conceive of imposing a minimum and maximum bit
level when determining the bit transition matrices, e.g.,
 
Ff und. = min max(Ffmin und. , Ff und. ), F max
f und. (7.3)

for an agent’s fundamental bit level with 0 < Ffmin max

und. < Ff und. < 1.
To accentuate the following results, however, this idea was not imple-
mented.

7.3 Simulation Results with the Modified SFI-ASM

7.3.1 Trading Bit Behavior

The main difference in the model behavior is shown in figure 7.1. No

matter what GA-invocation interval is used, most agents choose to give
2
Another similarity between Gomez’ approach and the mutation operator defined
by 7.1 and 7.2 is that they both work on the level of individual agents, i.e., they
employ the same strategy of decentralized parameter control.
116 7 A Suggested Modification to the SFI-ASM

1.0 p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p
p p p p p pp p p
Fraction of Zero-Bit Agents
p p p p p p p p p p pp p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p
pppp pppp p pp pp ppp p p p p pp p pp ppp pp ppp
0.8 pp p ppppp pp ppppppppp p ppp pp ppppp ppppp p ppp ppppp p p p
pp pp pp pp p p p p p p p p pp ppp p pp pppppp
pp p pp pp pp pp ppp p pppppp p ppp pppppppppppp ppppppppppp p p p p p pppppppp pppppppp pppppppp pppppppp pppppppp
ppppppppppppppppppppppppppppppppppppp p p p pppppp pppppppp
pppppppppppppppp p ppp p ppp p
0.6 p p p p p p p
p p pppppppppp p p p p pp p pp pp p pp p p p p
p
pp p p p p p p p p p p pp ppp
p pp p ppp p
pppp p p p ppp p p p p pp p p pp
p p pp
pp p pp pp ppp p pppppppppppppppppppppppppppppppppppppppppppppp
0.4 ppp pp ppp p p GA= 25
pp pp pp
ppp pp pp p pp p pp pp pp pp pp pp pp pp pp pp pp pp p GA= 100
pppp ppp pp p
pp p ppp
p p
0.2 ppp p
ppp pp p p
pp
ppp p
pp pp ppppp p ppp ppppp ppppp ppppp ppp GA= 250
p p p
pp p pp p p p pp p pp
p pppppppp pppppppp pppppppp ppp GA= 1, 000
pp p pp p p ppp pp
pp p pp p p pp
0 ppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppp pppppppp p pppppppp ppppp ppppppppp ppppp p pppppp pppppppppp pppppppp pppppppp p ppppp p
100 250 500 1,000 5,000 25,000 100,000 500,000
Periods

Fig. 7.1. Fraction of bit-neutral agents which discovered the correct hree
non-bit usage solution (Zero-Bit Agents) recorded for different GA-invocation
intervals and averaged over 10 simulation runs. (25 agents with λ = 0.3, bit
cost = 0.01, and initial bit probability of 0.01.)

up the use of their classifier system and neglect any information pro-
vided by it. Since homogeneous agents in a rational expectations equi-
librium would be characterized by a total neglect of trading information
provided by the classifier system, the Marimon-Sargent Hypothesis is
finally supported with respect to bit behavior. When replacing the orig-
inal mutation operator with an updated operator that has the desirable
property of being bit-neutral, most agents seem to realize that, under
the given dividend process, all they need for their forecast production
is the last period’s price and dividend information. Given the biased
mutation operator in the original SFI-ASM, emergent technical trading
bits thus seem indeed to be a design artifact and not a reflection of on-
going technical trading in the market. Changing the mutation operator
appears to be a relatively small change in the design of the artificial
stock market. It leads, however, to radically different results with re-
spect to bit behavior from the original model. I will henceforth refer to
this version as the modified SFI-ASM and to the agents as bit-neutral
agents.
Bit-neutral agents who sooner or later arrive at the zero-bit solu-
tion do so by using smaller and smaller mutation rates. Since Fogarty
[130] has shown that decreasing mutation rates may improve the per-
 7.3 Simulation Results with the Modified SFI-ASM 117

formance of a genetic algorithm, it is conceivable that the dynamically

adjusting bit transition probabilities of bit-neutral agents assist them
in zeroing in on the optimal hree-solution. However, since a zero-bit
agent is characterized my no mutations at all, a baseline mutation rate
greater than zero is highly recommended. Only if mutation remains an
option for agents are they able to react to structural breaks such as a
change in the dividend process. DeJong has suggested using 1/(number
of bits in the population) as a minimum mutation rate [91].

150,000 pppp∗pp∗pppp
p∗ppppp ∗pp ∗ ∗ppp
pp∗pp pp∗pppp∗pp ∗p ∗p∗pp∗pp∗pppp pppppp
50,000 ∗∗p∗∗∗∗pppppp
Average GA-invocations

p
∗pppp ∗∗p ppppp p p p
ppppp ∗ppp∗ppppp
25,000 pppp ∗pppppppp pp∗pppppppppppppppppp∗ppppppppppppppppppp∗pppppp ρ = 1.00
p p p p pp
pppp
pppp


p
p ∗pppppp
10,000
p


p p p p p
ppp
pp
pp
ppp

ppp

pp

pppppppp
p p

ppp ppp ∗ppppp
p p p p
pppppppppppppppppp
ppppppppppppppppppp
pppppp ρ = 0.99
∗ppp

p pp



pp



p p p pp
pp pppppp ∗pp p p p p p p p p p p p p⋆p p p p p p p p⋆p p ρ = 0.95
5,000 ppp pp p

ppppp p p p ⋆
pp
pp
p

pp
pp ppppp ∗ppppppppp
p
∗pppppp pp pp pp ppp
pp

pp ∗ p pppppppp
2,500 pppppppppp ∗p
ppppppppp pp ppp p ppp
ppppp pp
p

p ⋆
pppp⋆pp⋆p⋆p p ⋆pppp ⋆p ⋆pp

pp p


pppppppp
pppp ∗pp∗ppppppp
p
p pp p p p p p p
p p p p p p p p p pp p p p p p ⋆ppp⋆p ⋆p pp ⋆p ⋆⋆pp ppppp⋆pppp⋆p⋆p⋆ppp p p⋆p p p p p pppp

p ∗ ∗ppppppp
pppp ∗ppp pppppp
p ⋆ p p ppppp
1,000 ⋆p p p
p p ⋆ p pp ⋆ p p
⋆ p p p p ⋆ p p p p p p p
p p p p pp
ppppppp p p
p ⋆ p pp p p
⋆p p ppp ⋆ ppp
pp

p p p p p pp p p p p p p∗ppppppppp
⋆
pp
ppppp ppp ∗ pppppppp∗ppppppp
p p p ⋆ ⋆p p p⋆p p ⋆
p p pppp
pppp ⋆ p p p ⋆ ⋆p p pp p p p p p p p p
500 pppp p p p pppppp p⋆ ppp pp
⋆
p p
ppp
p p p p pp p
ppppppp
pppppp ∗ ppp∗ppppp∗ppppppppp∗ppp
⋆p p p ⋆ p p p pp p p p

ppppppppp p p p p p
⋆p p p p p pp p p⋆
p
⋆ ⋆p p

250
pppp
⋆p p p
p p p p
p p p pppp

p p p


pppppp
pppppppp pppppp p
⋆p p p p p p ⋆p p p p p pp
pppppppp
∗pppppppppppppppp
ppp
⋆pp p⋆p p p pp ∗
p ⋆p p p p⋆p p ⋆ppp p p ⋆p p p p p
100 ⋆p p p⋆p p⋆p p p p⋆p p p p p p⋆p
75
1 5 10 25 50 100 250 500 1,000 2,500 7,500
GA-invocation interval

Fig. 7.2. Number of GA-invocations needed until 80% of bit-neutral agents

have reached the zero-bit level for the fundamental bits, recorded for different
speeds of mean reversion for the dividend process. It was averaged over 25
simulation runs with 30 agents, λ = 0.3, and bit cost = 0.01.

Depending on the characteristics of the dividend process, however,

agents may experience problems in discovering the zero-bit solution.
Figure 7.2 plots the average number of GA-invocations needed until
the fundamental bits of 80% of the agents have reached the zero-bit
level. This was done for several speeds of mean reversion of the div-
idend process.3 First, one notices a pronounced hump in the curves
3
The behavior of technical trading bits is qualitatively similar. However, they
generally seem to arrive earlier at the zero-bit level. These results, however, were
derived with an earlier Java-version of the SFI-ASM which encoded four technical
bits less than fundamental bits, which could account for this difference. Because
of the enormous computational costs—each data point in figure 7.2 was averaged
118 7 A Suggested Modification to the SFI-ASM

for GA-invocation intervals between 3 and approximately 250. These

small invocation intervals seem to hamper the ability of the agents
to discover the long-term randomness of what they first believe to be
regular patterns. More interestingly, though, is that figure 7.2 clearly
demonstrates that the closer the mean-reverting dividend process gets
to a random walk, the more difficult it is for the agents to arrive at
the zero-bit level.4 It implies that pure random asset behavior in real
financial markets is much harder to detect than more predictable as-
set behavior. Investors who keep testing the vast amount of possible
trading rules may simply be too impatient to discard non-profitable
strategies or lack the necessary memory to do so. In real financial mar-
kets, the constant departure of experienced traders and arrival of new
traders might lead to constant out-of-equilibrium behavior of markets.
Given the average lifetime of an investor, figure 7.2 suggests that any
practical learning speed could put us in the region where it is extremely
difficult to discover the randomness of asset prices.

7.3.2 Time Series Properties

A typical neoclassical rational equilibrium solution would not only be

characterized by a total neglect of any additional information contained
in condition bits, it would also satisfy the second part of the Marimon-
Sargent Hypothesis by exhibiting “nice” price series properties. One
should keep in mind that the proposed change to the GA only affects
the condition part and not the real-valued forecast parameters of a
trading rule. Thus, one would expect the two models to produce similar
time series, that is, “well behaved” ones for the slow learning case and
more complicated ones for faster GA-invocation intervals.
This hypothesis was tested by comparing the statistical properties
of the price time series with those of the Java SFI-version from table 6.4
on page 105. Remember that in hree-mode, the dividend and market
price are a linear function of their first order lags, i.e.,

pt+1 + dt+1 = a(pt + dt ) + b + ǫt . (7.4)

aver 25 simulation runs which sometimes lasted several million periods—, the
simulations were not repeated when the model was upgraded to utilize all possible
32 bit locations in the two condition parts.
4
There is, however, no monotone relationship between speed of mean reversion and
required time to find the zero-bit solution. For very small ρ, the dividend process
is very close to a white noise process, and agents again experienced difficulties in
detecting this randomness.
 7.3 Simulation Results with the Modified SFI-ASM 119

The estimated residual time series ǫ̂t are then analyzed to see whether
they satisfy being i.i.d. and N(0,4) distributed. The results are summa-
rized in table 7.1.

Table 7.1. Time series comparison of the modified SFI (M-SFI) and the
replicated Java SFI-ASM with the original mutation operator (J-SFI). Means
over 25 runs. Numbers in parentheses are standard errors estimated using
the 25 runs. Numbers in brackets are the fraction of tests rejecting the no-
ARCH or iid-hypothesis for the ARCH and BDS tests, respectively, at the
95% confidence level.
Description GA 1,000 GA 250 GA 20 GA 1
M-SFI J-SFI M-SFI J-SFI
Std. Dev. 2.084 2.145 2.141 2.225 2.229 3.397
(.009) (.010) (.013) (.015) (.013) (.034)
Excess kurtosis 0.004 0.085 0.001 0.285 0.050 9.046
(.009) (.015) (.001) (.033) (.011) (1.56)
ρ1 0.011 0.031 0.014 0.025 0.029 0.491
(.002) (.008) (.002) (.012) (.001) (.006)
ARCH(1) 2.610 3.502 2.754 25.34 5.722 1,871.9
[0.20] [0.60] [0.40] [1.00] [0.48] [1.00]
ρ21 0.013 0.020 0.015 0.075 0.020 0.425
(.002) (.002) (.004) (.008) (.003) (.017)
BDS 1.06 1.41 1.10 34.08 1.44 38.63
[0.20] [0.32] [0.24] [0.92] [0.28] [1.00]
Excess return 1.52% 2.92% 1.59% 3.25% 1.51% 25.34%
(.02%) (.03%) (.03%) (.08%) (.03%) (3.41%)
Trading volume 0.244 0.364 0.271 0.854 0.876 1.359
(.008) (.025) (.007) (.065) (.009) (.015)

First of all, one notices that the modified SFI-ASM produces time
series that are usually closer to the hree-benchmark than those of the
original SFI-ASM. The standard deviations in the residuals are gen-
erally smaller, thus indicating less price variability. Excess kurtosis is
almost negligible for both the fast and slow learning cases, which does
not line up very well with empirical fat tailed return distributions.
Yet when further enhancing the learning speed, both the increase in
standard deviation and excess kurtosis suggest that the modified SFI-
model shifts into a more complex regime for faster learning rates than
120 7 A Suggested Modification to the SFI-ASM

the original SFI-model suggests.5 The autocorrelation in the residuals,

as shown in the third row, demonstrates that there is little linear struc-
ture remaining except for the extreme case of updating the rule set in
every period. As LeBaron et al. indicate, any artificial stock market
should exhibit negligible autocorrelations since they are very low for
real markets. The large autocorrelation coefficient for GA=1 indicates
that the economic structure of the model might break down at this
speed, i.e., equation (7.4) becomes misspecified.6
The next row reports the means of the test statistics for the ARCH
test proposed by Engle [110]. There is considerably less ARCH depen-
dence in the residuals for the modified SFI-version. It is interesting
to note that even for very small GA-invocation intervals, some test
runs are not able to reject the no-ARCH hypothesis. Only for a GA-
invocation in every period can extreme ARCH-behavior for all test runs
be observed.7 In row five, the first order autocorrelation of the squared
residuals is another test for volatility persistence. Again, it increases
for faster learning speeds but is generally lower than for the SFI-case.
The BDS test in row six is a test for nonlinear dependence developed
in [58]. Its test statistic is asymptotically standard normally distributed
under the null hypothesis of independence.8 One can notice an increas-
ing amount of nonlinearities for faster exploration rates, yet again it is
substantially lower for the modified SFI-version. Trading volume, which
should be zero in the hree-case, increases significantly for faster learn-
ing speeds. This points to a greater degree of heterogeneity between
the agents.
Overall, the original conclusion that the learning speed affects the
price series behavior can still be confirmed after the proposed change.
However, it is also apparent that for identical GA-invocation intervals,
the modified SFI-results are generally closer to the hree-benchmark.

5
While LeBaron et al. have reported the results only for the GA-intervals of 1,000
and 250, two additional learning speeds are included in table (7.1). The statistical
tests were performed for even more GA-intervals, in particular, for 100, 50 ,25,
20, 10, 5, 2, and 1.
6
The autocorrelation coefficient for a GA-interval of 2 with a value of 0.06 (stan-
dard deviation 0.0035) is considerably lower than for an invocation interval of
one. This supports the hypothesis that there is a structural break in the model
when using the fastest possible learning speed.
7
Even for an invocation interval of two, the no-ARCH hypothesis cannot be re-
jected for 16% of the test runs.
8
There are two free parameters for this test. The distance r is measured as a
fraction of the standard deviation and has been set to a value of 0.5, while for
the embedding dimension m, a value of two has been chosen.
 7.4 Robustness of the Zero-Bit Solution 121

Compatible with these results are the findings of Wilpert [437]. He

reports less kurtosis in the residuals and less trading volume when
agents have no access to their classifier system in the first place. In
the modified SFI-ASM, agents endogenously arrive at neglecting the
classifier system and converge to the hree-solution.

7.4 Robustness of the Zero-Bit Solution

7.4.1 Stochastic versus Periodic Dividends and the Classifier

System

Since the finding of zero-bit usage by bit-neutral agents contradicts the

findings of original SFI-ASM, various tests were designed to check the
robustness of this result. First, in order to check the proper working
of the classifier system, the model behavior was tested for classifier
mode and non-classifier mode. For the latter, agents had no access to
condition bits at all, and their trading rules consisted only of prediction
parts with associated fitness information. Classifier and non-classifier
agents were both confronted with a noiseless periodic dividend stream
such as a sine wave or a square wave. Even though the simulated price
series tracked the crude risk-neutral price astoundingly well in the non-
classifier mode, the tracking behavior in the classifier mode was better
for most GA intervals.9 The bit usage of most classifier agents was
characterized by bit positions in their trading rules that were either
almost completely set or completely ignored within a relatively short
amount of time.
This finding was revealed by figure 7.3. When looking at how often
certain bit positions in the whole economy are set, one notices that they
are often multiples of 100, which is the number of trading rules that
agents possess. One can easily see that the fundamental bit no. 32 has
been independently discovered by seven bit-neutral agents, while four
agents found technical bit no. 31 to be useful. Many other bit positions
were heavily used by either one or two agents. In spite of the periodic
9
There are, however, some parameter restrictions that may make it tricky for
agents to recognize the periodicity of the dividend process. First, the wavelength
should be shorter than the parameter τ which controls the time window over which
trading rules are evaluated. It should also be shorter than the GA-invocation in-
terval. Furthermore, the forecast parameters in the prediction parts of trading
rules can only be changed by the GA within very narrow intervals which were
calculated on the basis of a mean-reverting AR(1) dividend process. For a com-
pletely different dividend process, the required parameter values may lie outside
the allowable interval and cannot be reached by the GA.
122 7 A Suggested Modification to the SFI-ASM

Fig. 7.3. Bit Distribution of 25 bit-neutral agents when confronted with

periodic square wave dividends (wavelength = 50, Ga-interval = 250, θ=75).

nature of the dividend process without any added noise, agents became
heterogeneous in that most of them decided to use a different technical
trading rule.
In summing up one can state that the simulation results under pe-
riodic and stochastic dividends show that the classifier system works
quite efficiently. When confronted with periodic dividends, it detects
these patterns, yet when working with stochastic data, it also discovers
the “right” solution of non-bit usage. Even though the mean-reverting
dividend process is able to produce short term trends toward its mean,
these are by no means regular. Therefore, in the long run, the stochas-
tic nature of the dividend process dominates any (random) short term
trends and patterns.

7.4.2 Dependence on Other Parameter Values

The zero-bit solution proved robust under a wide range of parameter

combinations. For instance, changing the selection procedure for rule
usage (select best or roulette wheel selection) did not alter the main
 7.4 Robustness of the Zero-Bit Solution 123

finding of convergence towards the hree zero-bit solution. Increasing

the bit cost increased the convergence speed, especially for “latecomer”
agents who experience difficulties in finding the correct hree-solution.
Varying the crossover and mutation rates revealed some differences
in convergence speed, but did not alter the main result of agents finding
the zero-bit solution. In the crossover only case, some zero-bit agents
emerged rather quickly. Other agents had some bits set in all their
trading rules while the other bits tended to be completely unused. In
the extreme case of mutation only with both mutation probabilities set
to one, trading bits were uniformly distributed across all agents and
their trading rules. Since the mutation only case could be seen as pure
experimentation without making use of previously good solutions, it is
not surprising that agents gave up their classifier system, too.
The zero-bit solution thus turns out to be robust under wide pa-
rameter combinations. The finding that the classifier system in this
artificial stock market does not provide any advantage or useful in-
formation that agents could exploit is also supported by Wilpert who
replaces the fitness evaluation based on forecasting ability with profit
generating ability. He, too, finds that giving up the condition parts of
the trading rules does not have drastic effects on the model’s behavior,
and he questions the usual interpretation of bit usage. He emphasizes
that its importance should not be overestimated.

7.4.3 Generalization or Consistent Trading Rules?

In section 6.4.3 it was mentioned that the GA in the SFI-ASM reg-

ularly produced illogical trading rules that would never be activated.
The solution to this problem was to have a large rule set and to gener-
alize rules that have not been matched for more than 4,000 periods by
converting some of their 0 and 1-bit to #.
Since the original SFI-mutation operator introduced an upward bias
in the bit distribution, it was impossible to assess the degree of down-
ward pressure exerted by the generalization procedure. Even though
the zero-bit solution seemed largely unaffected by changes of the gen-
eralization parameters, i.e., the maximum number of inactive periods
before a rule is generalized and the fraction of bits to be generalized,
it cannot be excluded that it is the generalization procedure that is
responsible for the zero-bit solution.
To exclude that possibility, a rule consistency had been developed
which ensures that only logical trading rules are accepted into an
agent’s rule set.
124 7 A Suggested Modification to the SFI-ASM

Table 7.2. Creating logical trading rules through a consistency check.

price x interest Rate 1 3 7 9 5 3
dividend > 2 4 8 1 8 4 2
rule 1 # # 0 1 # 0 #
rule 2 # # 1 1 # 0 #
←−−−
rule 1 −→ rule 3 # ## 1 # 0 #
−−−→
rule 1 −→ rule 4 # # 0 # ## #

An illogical trading rule is rule 1 in figure 7.2. It describes a situation

in which pt rf /dt is smaller than 7/8, yet at the same time, that ratio is
supposed to be greater than 1. Obviously, rules of this type will never
be matched but will get inserted into an agent’s rule set. Rule 2, on
the other hand, might be considered a corrected version. However, the
third rule demonstrates that the same information can be coded by
using less activated bits.
The consistency check in the modified SFI-ASM is thus designed
not only to create logical trading rules, but also to express corrected
bit sequences with the least amount of non-# bits, i.e., one or two.
However, depending on whether the checking procedure starts analyz-
ing the bit sequence from the lower bits upward or vice versa, different
results will be returned by it. Rule 3 would be returned by the consis-
tency check if it starts analyzing from the higher order bits. Otherwise,
the fourth rule would be returned. The consistency check thus chooses
both variants with equal probability.
Since the consistency check removes bits from illogical trading rules,
one could easily argue that it aggravates a possible downward pressure.
The conjecture of a downward bias through the consistency check, how-
ever, turns out to be wrong. Surprisingly, figure 7.4 shows that turning
the consistency check on increases the level of set trading bits. The
key for understanding this astonishing result lies in the way the bit
fractions are calculated.
Consider the case without the consistency check. If we denote the
number of set fundamental bits with y, the fundamental bit fraction
in a trading rule is determined as y/32, 32 being the number of all
fundamental conditions. With the consistency check, however, the bit
fractions are calculated differently. Recall that 2 bit sequences of 7
bits are checked in the fundamental condition word of 32 bits. Within
each bit sequence, at maximum, 2 bits can be set. The maximum bit
fraction of 100 % would be achieved if all 18 unchecked bits are set, in
addition to the 4 bits in the two checked sequences. Instead of 32 bits,
the denominator for the calculation of bit fractions is then 18 + 4 = 22
 7.4 Robustness of the Zero-Bit Solution 125

Fig. 7.4. Demonstration of the upward pressure on the bit distribution when
switching the consistency check on in mid-simulation.

bits. Using a smaller denominator increases the bit fraction for a given
number of set trading bits. The simple recalculation explains the jump
of the bit fractions the moment the consistency check is turned on.
But remember that these augmented bit fractions for technical and
fundamental bits are then used to determine the dynamically adjusting
bit transition probabilities for the updated mutation operator. Since
the bit-neutral mutation operator applies these elevated bit transition
probabilities on the whole bit string, turning the consistency check on
results in a slight increase in the bit distribution in addition to the one
time effect shown in figure 7.4.
The major result of agents finding the correct hree-solution are usu-
ally replicated with or without the consistency check. It is probably
a useful extension of the SFI-ASM since it allows agents to focus on
creating and testing only logical trading rules. From a computational
perspective, it could even help save computational time since agents do
not need a huge rule set to work with.
8
An Analysis of Wealth Levels

If you’re so smart, why aren’t you rich?

8.1 Introduction

The previous chapter has presented evidence that fundamental and

technical trading bits do not improve the forecast accuracy of agents’
trading rules when confronted with stochastic dividend data. As a re-
sult, agents abandon their classifier system in the long run. Unless a
rule’s forecast accuracy is a wrong proxy for its profitability, the con-
vergence at the zero-bit solution should imply that classifier agents,
i.e., those who use the information provided by their classifier system,
cannot outperform agents who neglect the classifier system altogether.
This is in contradiction to studies by Joshi et al. [206, 207]) who found
that agents with access to technical trading bits acquire more wealth
than fundamental traders.
This chapter first briefly introduces the two studies by Joshi et al.
and presents the explanations given for the observed differences in
wealth levels. However, a sensitivity analysis of wealth levels reveals
that things are more complicated than previously thought. Wealth
differences between different trader types—technical or fundamental
traders, fast or slow learning agents, bit-neutral or unmodified SFI
agents, classifier or non-classifier agents—generally arise. While the
explanation of these wealth differences has been sought by Joshi et al.
through elaborate economic reasoning, the following analysis identifies
the number of active trading rules and the chosen selection mechanism
as main determinants for varying terminal wealth. Additionally, two
benchmark levels are developed with which the absolute wealth levels
128 8 An Analysis of Wealth Levels

attained by traders can be compared. This chapter concludes by stat-

ing that the economic structure of the SFI-ASM makes it unsuitable
to give an economic interpretation to wealth effects.

8.2 Wealth Levels in the SFI-ASM: An Economic(al)

Explanation

Since agents in the modified SFI-ASM endogenously give up the use of

their classifier system, it is strongly implied that it does not provide any
profitable trading information. If technical trading rules were to gener-
ate excess profits, technical trading should flourish rather than vanish.1
The conclusion of a useless classifier system is in direct contradiction to
Joshi et al. [206], who find that agents with access to technical trading
bits acquire more wealth than fundamental agents. The wealth differ-
ences that they report are not negligible and grow over time, as figure
8.1 vividly illustrates.
When trying to explain the source of these wealth differences, Joshi
et al. [206] offer some vague economic rationalization. They argue that
the use of technical trading rules creates a negative externality for
other agents. If a single agent with access to technical trading bits
detects a short term price trend, he might be able to exploit this pattern
without it dissipating. If more and more agents detect the price trend,
the particular technical trading rules are reinforced through positive
feedback, thus making them self-fulfilling prophecies, which can cause
bubbles and crashes. The latter are negative market externalities which
worsen everyone else’s strategies by lowering their forecast accuracy. As
a result, stock prices are driven away from their fundamental values.
This is a loss of market efficiency and Joshi et al. claim that this would
translate into lower average returns and less accumulated wealth.2
1
Strictly speaking, the trading rules are not evaluated according to their wealth
generating ability, but on their forecast accuracy. While it is possible that a more
accurate trading rule is less successful in terms of generated wealth, Arthur et al.
[11] believed this scenario to be highly unlikely. Their intuition was confirmed by
Wilpert [437], who evaluated trading rules according to their generated profits.
His results did not differ much from those of the original SFI-ASM.
2
Similar arguments on the emergence and profitability of technical trading can
be found in [11, 205]. The papers by Arthur et al. [11], Palmer et al. [330], and
LeBaron et al. [244] do not discuss any possible wealth effects. In [331], we find
the rather vague claim that “if a trained agents is extracted from the market
and then reinserted much later, it tends to do rather poorly.” “Doing poorly”
probably means acquiring less wealth. Furthermore, they report that the wealth
distribution evolves into a wide distribution, with some agents being very rich
 8.3 Previous Studies Based on Wealth Levels in the SFI-ASM 129

Fig. 8.1. Wealth levels for the situation when one agent includes technical
rules while all others exclude them. Note that the singular agent using techni-
cal rules accumulates significantly more wealth than those agents using only
fundamental rules almost all through the run, and that this difference grows
over time. Source: [206, p. 11]

8.3 Previous Studies Based on Wealth Levels in the

SFI-ASM

The observed differences in wealth accumulation between fundamental

and technical traders were used by Joshi et al. for two game-theoretic
analyses which will be briefly introduced in the following two sections.

8.3.1 Financial Markets Can Be at Sub-Optimal Equilibria

The learning rate of agents in the SFI-ASM was treated as an exogenous

variable by LeBaron et al. For slow learning rates, the model closely
resembled the homogeneous rational expectations equilibrium, but for
faster learning rates, the model shifted into a more complex regime that
better fits the empirical facts. However, the question of which learning
rate would be chosen by agents if given the opportunity to do so was
not addressed by them.
Joshi et al. [207] close this gap by combining their simulation results
with a game-theoretic analysis. They assume that there are S possible

and some being poor. While the identity of winners and losers changes over time,
the distribution stays rather constant.
130 8 An Analysis of Wealth Levels

learning rates that agents can chose from.3 Given the assumption that
N − 1 traders in the market have adopted a common, but unknown
revision rate K, a single trader has to find his optimal learning speed
K ∗ .4 Since all traders in the market attempt to find their optimal
responses, this situation constitutes a symmetric simultaneous-move
N -person game with a S × S decision matrix in which a single trader’s
dominant strategy would be a dominant strategy for all other agents.
The payoffs in the S ×S decision matrix are determined by fixing the
base revision rate K to one of the S different values, and then recording
the terminal wealth levels that the remaining agent can acquire by
varying his learning rate K. The information in the S × S decision
matrix can be summarized in a reaction function, which assigns to
every K the optimal response K ∗ at which the single trader maximizes
his terminal wealth level.

Fig. 8.2. A trader’s reaction function, indicating his optimal GA-invocation

interval K ∗ , given the baseline GA-invocation interval K used by all other
traders. The symmetric Nash equilibrium is where K = K ∗ = 100. Source:
Joshi, Parker, and Bedau [207, p. 12]

3
These rates are 5, 10, 50, 100, 250, and 5,000.
4
For simplicity, Joshi et al. assume that all other agents chose the same single
base revision rate K. They claim, however, that their results do not change much
when other agents are allowed to follow mixed strategies.
 8.3 Previous Studies Based on Wealth Levels in the SFI-ASM 131

It is apparent from figure 8.2 that an agent’s optimal response

matches the base GA-invocation interval K only at one point, i.e.,
where K = K ∗ = 100. The single trader cannot gain by deviating
from his strategy, thus he keeps updating his rule set at the same fre-
quency as everyone else does in the economy. Since all traders in the
market face an identical situation, no one has an incentive to deviate
from this learning rate either. K = K ∗ = 100 thus constitutes a unique
symmetric Nash equilibrium.
Joshi et al. emphasize that this market equilibrium is sub-optimal
since it clearly falls into the complex regime as described earlier. Price
series are complicated and exhibit a higher variance than the rational
expectations regime, making investments riskier. Furthermore, the con-
dition parts of trading rules contain many set fundamental and techni-
cal trading bits which requires more resources devoted to creating and
evaluating them. Last, but not least, they claim that the wealth levels
at the end of the simulation are lower for the complex regime than for
the rational expectations regime. Everyone in this Nash equilibrium is
worse off than if all agents would pursue more long-term oriented trad-
ing strategies. This is a typical prisoner’s dilemma situation in which
the rational behaviors of individual agents cause the optimal social out-
come, the rational expectations regime, to be missed. Joshi et al. there-
fore conclude that financial markets can operate at such sub-optimal
equilibria in which we have significant levels of technical trading.

8.3.2 Technical Trading as a Prisoner’s Dilemma

In a similar and related study, Joshi et al. [206] vary their focus and
ask whether agents will choose to include technical trading information
in their trading rules. Throughout this study, a constant learning rate
of K = 100 is used. This is done for two reasons: First, at this rate,
all agents may use technical trading information since they are in the
complex regime. Second, they have shown in their 2002 article that this
would be the learning rate chosen by agents.
The framework is slightly modified from the study above in which
agents had to choose their optimal learning rate. This time, a single
agent faces a classic 2 × 2 decision problem of whether to include the
technical trading information or not, given his assumption that all of
the other traders either include them or not.5 Again, this design resem-
bles a multi-person simultaneous-move game. Thus, the only potential
5
All traders have access to fundamental trading information though. The exclusive
use of technical trading information is ruled out as unrealistic.
132 8 An Analysis of Wealth Levels

Nash equilibria are symmetric in the sense that all agents either include
or exclude the technical trading bits.

Table 8.1. “The decision table for an agent contemplating whether to include
technical trading rules to make her market forecasts, when she is uncertain
whether the other traders in the market are doing so. The agent’s payoff in
each of the four situations A–D is her expected final wealth (divided by 104 ,
to make more readable), derived by averaging the results of 45 simulation
runs of each situation. Error bounds are calculated using standard deviations
of each set of 45 simulations.” Source: Joshi, Parker, and Bedau [206, p. 9]

all other agents

include exclude
technical rules technical rules
includes
single agent

A: 113 ± 7 B: 154 ± 7
technical rules

excludes
C: 97 ± 7 D: 137 ± 5
technical rules

The payoffs in table 8.1 are averaged terminal wealth levels.6 It is

obvious from the payoff table, that including technical trading rules in
one’s rule set is a dominant strategy for a single trader since A > C
and B > D. Because we can have only symmetric Nash equilibria, the
Nash equilibrium that will be selected is the one in which all agents
include technical trading information. All other states are unstable.
According to this game-theoretic analysis, it is rational for each
agent to include technical trading information, even though it leads to
an outcome in which all earn less wealth than as if they all agreed on
ignoring technical trading bits. Thus, Joshi et al. [206] conclude that
technical trading is a typical multi-person prisoner’s dilemma and that
financial markets can get locked into such an inefficient equilibrium.

6
For each cell, Joshi, Parker, and Bedau averaged over 45 simulations, each of
which was run for 300,000 periods.
 8.4 Wealth Levels in the SFI-ASM: Alternative Explanations 133

8.4 Wealth Levels in the SFI-ASM: Alternative

Explanations

The following sections, however, will provide other answers to the ques-
tion of what determines the wealth accumulated by agents. Section 8.4.1
analyses the model design and finds that the economic explanations for
absolute wealth levels given by Joshi et al. [206] turn out to be conjec-
tures that are not supported by the SFI-ASM design. This theoretical
analysis is supplemented by some simulation evidence in section 8.4.2.
The question of relative wealth, i.e., why certain trader types become
richer than others, is addressed in section 8.4.3.

8.4.1 Risk-Premium, Taxation, and Two Benchmark Wealth

Levels

When observing differences in absolute wealth levels due to learning

speed or access to certain trading information, a mind trained in eco-
nomics easily spots some potential for optimization. For instance, Joshi
et al. [206] asked themselves whether these wealth differentials could
be helpful in finding an agent’s optimal learning speed. Because of the
specific design of the SFI-ASM, however, wealth levels are somewhat
counterintuitive to interpret and optimization attempts might fail.
Agents in the SFI-ASM increase their wealth in two ways. First,
the cash they own earns them a risk-free interest income and second,
they collect the stochastic dividends on their stock holdings. To avoid a
long-run explosion of wealth levels through compound interest effects,
however, an agent’s wealth is taxed at a rate equal to the exogenous
interest rate. The cash and wealth positions of an agent i are computed
according to

Wi,t = Ci,t−1 + pt xi,t (8.1)

Ci,t = Ci,t−1 + rf Ci,t−1 + xi,t dt (8.2)
Ci,t = Ci,t−1 − rf Wi,t , (8.3)

xi,t being the amount of stock held by agent i. Equation 8.1 denotes the
previous wealth before adjustment, and equation 8.2 takes interest and
dividend earnings into account, while equation 8.3 lowers cash through
tax payments at a tax rate equal to the risk-free interest rate. Equations
8.1 − 8.3 can be summarized as

Ci,t = Ci,t−1 + xi,t (dt − rf pt ) , (8.4)

134 8 An Analysis of Wealth Levels

which is the equation with which the cash positions of traders in the
model are updated. Note that for risk aversion, the term in parentheses
reflects a positive risk premium, and Ci,t > Ci,t−1 holds.
Gulyás et al. [166] point out that in the SFI-ASM, agents usually
increase their wealth more or less independent of their actions. It is,
therefore, possible to determine a benchmark wealth level under the
assumption of inactivity. Agent i would hold onto his one unit of stock,
i.e., xi,t = 1 for all t, and would take the market price as given. In this
case, equation 8.4 would simplify to

Ci,t = Ci,t−1 + (dt − rf pt ). (8.5)

Since wealth is defined as an agent’s cash plus the value of his stock
position, which is xi,t = 1 in this case, the first benchmark wealth level
is calculated as
Wi,t = Ci,t−1 + dt + pt (1 − rf ) (8.6)
and will be referred to as base wealth.
To determine an agent’s base wealth, however, it is necessary to
have an actual simulation run to have stock prices with which stock
positions can be valued. One way to determine an absolute benchmark
wealth level without the need for a simulation is to assume that the
market is run in hree-mode, i.e., not only one, but all agents are inac-
tive, holding onto their one unit of stock, collecting its dividend, and
receiving interest on their cash. An approximation for hree-base wealth
can be obtained by substituting the simulated prices and dividends in
period t by their theoretical averages d and phree .7 Since these are con-
stants, the recursive relationship of equation 8.5 can then be written
as a function of initial cash endowment C0 , i.e.,
 
Wthree ≈ Cthree = C0 + t d − rf phree . (8.7)

Since agents are identical, an agent subscript is not needed. In the long
run, the value of one unit of stock becomes negligible in comparison
to the cash value, hence, the theoretical hree-base wealth level Wthree
in period t is reasonably well approximated by hree-base cash. Again,
risk aversion implies that phree < d/rf , hence, 8.7 states that hree-
base wealth Wthree grows linearly in t and that the only source for

7
The hree-price was derived in section 6.5. It is a linear function of agent’s risk
aversion, the risk-free interest rate, the theoretical dividend mean, speed of mean
reversion of the dividend process, and the variance of the dividend noise process.
 8.4 Wealth Levels in the SFI-ASM: Alternative Explanations 135

wealth accumulation in the SFI-ASM are the risk premiums that agents
collect.8

Table 8.2. Some numerical examples for hree-prices and hree-base wealth
levels at different periods. Other parameter values were ρ = 0.95, rf = 0.1,
d¯ = 10, and C0 = 20, 000.

risk aversion phree (d¯ = 10) Wt=250,000

hree hree
Wt=300,000
λ = 0.0 100.00 20,000 20,000
λ = 0.3 88.00 320,000 380,000
λ = 0.5 80.00 520,000 620,000

It is important to realize at this point that base wealth usually

exceeds hree-base wealth since the long-term average of simulated prices
pt is smaller than the theoretical hree-price average. This reflects an
additional risk premium since the constant learning of agents introduces
some extra noise. Equation 8.7 also implies that an increase in the risk
premium due to higher volatility, reflected in smaller values for pt , will
result in higher wealth levels in the SFI-ASM. It is easy to see that
wealth could be maximized if stock prices would be zero. Contrary to
the statements made by Joshi et al. [206], an efficient market outcome
in the SFI-ASM is thus not characterized by high wealth levels, but
quite the contrary. Considering that simulated prices are usually below
the hree-prices and seldom overshoot, the hree-benchmark equilibrium
constitutes a lower bound for non-hree wealth levels.

8.4.2 Average Stock Holdings and Wealth Levels

The previous section only focused on the problem of aggregate wealth.

It found that when the risk premium in the stock market increases,
all agents are able to acquire more wealth. The question as to where
the wealth differences come from between technical and fundamental
traders or fast and slow learners reported by Joshi et al. [206, 207] was
not addressed. Given that different trader types in a particular simu-
lation face the same risk premium, the question remains why technical
8
Another reason why equation 8.7 is only an estimate is that it considers the
theoretical averages for prices and dividends and not the empirical averages for
a particular simulation run. In the long run, however, these averages converge to
their theoretical means. The exact hree-base wealth in t for an actual simulation
run would be Wthree = C0 +t(d−rf phree )+phree
t , where d and phree now represent
the empirical averages.
136 8 An Analysis of Wealth Levels

traders accumulate different wealth levels than fundamental traders.

Since we have established in chapter 7 that agents endogenously give
up the use of their classifier system, it seems highly unlikely that there
are possible gains from technical trading. It is implied that the patterns
agents might temporarily detect are random and not profitable in the
long run.
When investigating the possible impact of the classifier system on
wealth accumulation, it seems reasonable to introduce another trader
type. Since most bit-neutral agents end up not using their classifier
systems in the long run, it seems obvious to shorten the adjustment
process and to have non-classifier agents to begin with. Non-classifier
agents do not check any environmental information at all and have
trading rules that only consist of prediction parts and fitness informa-
tion. How would these agents compare if they had to compete with
old-fashioned SFI-classifier agents?
Wealth (SFI − Non-Classifier Agents)

125,000 .........
.........
.........
.........
.........
.........
.........
.........
100,000 .........
.........
.........
.........
.........
.........
.........
.........

75,000
.........
.........
.........
.........
.........
.........
.........
.........
.........

50,000
.........
.........
.........
.........
.........
.........
...........
................
......................
25,000 ..............................
...............................
... .................................
.........................................
.............................................. .
................................................. .......................................
.................................................... ..........................................................................................
................................................ ...................................................................... ...........................................................................................................................................
......................................................................................................................................................................... . ........
....................................
................................................ .................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................
0
.. .. . . .
.................................... .
.........................................................................................................................................................................................................................................................................................................................................
........................................... ................................................... .............................. .....................
................................... .................................................
...................................... .........................................................................................................

5 10 25 50 75 100 250 500 1,000

......................................................................
.........................................................................
...............................................
.......................................
.......................
............
-25,000 GA-invocation interval

Fig. 8.3. Wealth differences between SFI and non-classifier agents, both
trader types using select best for rule selection, recorded for different GA-
invocation intervals and averaged over 25 simulation runs. Positive values in-
dicate that SFI agents outperform the non-classifier agents. The shaded area
indicates mean wealth difference ± one standard deviation.

It was my initial hypothesis was that non-classifier agents should

never outperform classifier agents since only the latter were equipped
with an additional tool to analyze and exploit potentially useful in-
 8.4 Wealth Levels in the SFI-ASM: Alternative Explanations 137

formation.9 Figure 8.3, however, tells a different story. While for most
GA-invocation intervals the classifier agents acquire more wealth, there
is a significant dip in the curve when non-classifier agents do better
than their SFI-counterparts. Better performing non-classifier agents,
however, are a clear indicator that reasons other than pattern recog-
nition and exploitation are responsible for differences in accumulated
wealth levels.10
Since the risk premium is the only source for wealth accumulation in
the model, one obvious conclusion is that wealthier trader groups hold,
on average, more risky stocks. Figure 8.4 verifies this hypothesis by
plotting the acquired wealth levels of individual SFI- and non-classifier
traders as a function of their average stock holdings over their entire
lifetime.11
Besides the highly linear relationship between average stock holdings
and final wealth, figure 8.4 also shows that the non-classifier agents, as
a group, hold more stock than SFI agents. When regressing individual
wealth Wi on average stock holdings xi ,

Wi = a0 + a1 xi , (8.8)

the regression coefficients as shown in table 8.3 are highly significant.

Table 8.3. Regression results for the linear regression Wi = a0 + a1 xi

regr. coefficient std. error t-value coeff. of determination
a0 = 85, 380 7,400 11.5 0.9506
a1 = 223, 342 7,350 33.4

9
That hypothesis was posited before I actually established that risk premiums are
the only source of possible wealth accumulation in the SFI-ASM.
10
Initially, I doubted the validity of the simulation results in [206]. In [105], I re-
ported that wealth levels of classifier and non-classifier agents rise equally after
some initial divergence during the warm-up phase, which is in contradiction to
figure 8.1 with diverging wealth slopes. I then noticed, however, that while the
absolute wealth differences were tiny, the classifier agents did slightly better in
23 out of 25 simulation runs. A paired two sided t-test then revealed that these
minuscule wealth differences were significant, sometimes even at the 1% percent
level. It was this outperformance of the classifier agents that prompted me to look
for alternative explanations as to how wealth differences in the SFI-ASM come
about.
11
The specific parameter combination was chosen in order to clearly visualize that
the two trader groups hold, on average, different amounts of stock.
138 8 An Analysis of Wealth Levels

375,000 × – Wealth of SFI-agents

• – Wealth of non-classifier agents •
350,000 • •
• • •
• ××
••• •
325,000 ••
• •×
Wealth

• • ו• • ×
× ×• × •
300,000 •×
× ×× ×
××
275,000 × ××
×
×× ××

250,000 × ×
0.7 0.8 0.9 1.0 1.1 1.2
average stock holdings of an individual trader

Fig. 8.4. Final wealth levels of SFI- and non-classifier agents as a function
of average stock holdings (period 250,000; GA-interval = 250; risk aversion
=0.3; 10 trading rules; 25 SFI- and 25 non-classifier agents).

The coefficient of determination indicates that more than 95% of

wealth variations are explained through average stock holdings. While
this seems like a reasonably high power of explanation, the remaining
five percent of unexplained wealth differences could be, in principle,
attributed to profitable pattern recognition and exploitation, perhaps
of the “buy low and sell high” type.
However, equations 8.4 and 8.7 lead me to conclude that the only
source of wealth accumulation in the SFI-ASM framework is the risk
premium that agents collect. As long as stock carries a positive risk
premium at all, selling stock inevitably reduces the wealth that an
agent could otherwise acquire. This seems counterintuitive since my-
opic CARA-agents maximize their expected utility, which is a mono-
tone function of expected wealth in the next period. Even if they expect
a sharp drop in stock value, they should never sell. The explanation for
desired stock sales as a result of utility maximization is that agents
maximize their pre-tax wealth, neglecting that shifting wealth from
stock to cash will hurt them through subsequent taxation. The remain-
ing five percent of unexplained wealth variation thus, most likely reflect
differences in taxation over the lifetime of different agents. Even if the
average stock holdings of two agents happen to be identical in a given
period, they probably were not identical in all the periods before. This
 8.4 Wealth Levels in the SFI-ASM: Alternative Explanations 139

led to different tax payments which, when added up, may explain the
difference in final wealth.

8.4.3 Activated Rules and Rule Selection

A first step in identifying possible reasons why distinct trader groups

consistently hold more or less risky stock is to look at how they possi-
bly differ from each other in their properties. A plausible explanation is
the number of activated trading rules that systematically varies for dif-
ferent trader types. For instance, technical agents who check technical
conditions in addition to fundamental ones end up with less activated
rules than fundamental agents. Fast learning SFI agents invoke their
biased mutation operator more often per period than slow learning
SFI agents and thus have more specific and illogical rules. The size
of the activated rule set for SFI agents is generally smaller than for
non-classifier agents since for the latter, it is largely independent of the
GA-invocation interval and usually coincides with the total number of
trading rules they possess.12
It is thus, my hypothesis that it is the number of activated rules that
is mainly responsible for wealth differences between different trader
groups. Since the size of the set of activated trading rules depends on
the number of trading rules that agents possess, it seems reasonable to
compare agents with varying numbers of trading rules. In addition, one
also notices an important difference between the SFI-ASM version used
by LeBaron et al. and those used by Joshi et al. [206, 207]. In LeBaron
et al., always the best, i.e., the rule with the lowest variance estimate
from equation 6.7, is selected. Joshi et al., however, use roulette wheel
selection in which a rule is randomly chosen from the active rule set
with a probability proportional to its fitness value.
To test whether the number of activated trading rules may have an
influence on wealth accumulation of different trader types, wealth dif-
ferences were recorded for different rule set sizes under both selection
procedures. Figures 8.5 for select best and 8.6 for the roulette wheel
12
There is one caveat at this point. For both agent types, trading rules that had
been activated less than “mincount” times since their creation are not included in
the active rule set either. The mincount-parameter was set at a value of five and
thus, tends to decrease the number of activated rules only for very fast learning
speeds. Even though non-classifier agents have all their trading rules activated by
definition, a rule that is younger than mincount-periods will not be considered for
rule selection. For SFI agents, on the other hand, the requirement of minimum
activation may take longer to fulfill. The higher the learning rate of SFI agents,
the longer it takes, on average, for rules to become eligible for use.
140 8 An Analysis of Wealth Levels

15000

5000
Wealth Difference

-5000

-15000

-25000

-35000

100 Rules
-45000 75 Rules
5 50 Rules
10
25 25 Rules
50
75 10 Rules
100
250
GA-interval 500 5 Rules
750
1000

Fig. 8.5. Wealth differences between SFI and non-classifier agents for select
best as a function of learning speed and number of trading rules they possess.
Data were averaged over 25 simulation runs. Note that non-classifier agents
outperform the SFI agents for most parameter combinations (negative wealth
differences) and that a peak in the upper left corner reaching as high as 300,000
was truncated for better visibility.

mechanism thus give a more complete impression of wealth behavior

than figure 8.3 does. First, one notices that wealth differences between
SFI- and non-classifier agents do not obey any simple relationship when
changing either the learning speed or the rule set size. Both directions
have pronounced “wealth valleys”, thus indicating at least two or more
overlapping factors that influence an agent’s wealth accumulation. Sec-
ondly, while the general shape in both wealth graphs seems similar,
there are some noticeable differences such as the pronounced peak in
the upper left corner for select best.

Instead of focusing on the wealth differences between competing

SFI- and non-classifier agents, it might be more insightful to investigate
 8.4 Wealth Levels in the SFI-ASM: Alternative Explanations 141

5000

-5000

-10000
Wealth Difference

-15000

-20000

-25000

-30000

-35000

-40000

-45000 100 Rules

75 Rules
5
10 50 Rules
25
50 25 Rules
75
100 10 Rules
250
GA-interval 500 5 Rules
750
1000

Fig. 8.6. Wealth differences between SFI and non-classifier agents for select
roulette as a function of learning speed and number of trading rules they
possess. Data were averaged over 25 simulation runs. Note that non-classifier
agents outperform the SFI agents for most parameter combinations (negative
wealth differences).

the absolute wealth levels of one agent type alone. They are shown in
the top sections of figure 8.7 for SFI agents and 8.8 for non-classifier
agents. It is interesting to note that subtracting the absolute wealth
levels of non-classifier agents from those of SFI agents does not yield the
same wealth differences as depicted in figures 8.5 and 8.6. Obviously, the
co-evolution of two competing agent types creates a different economic
environment in terms of aggregate price level and risk premium than
just having one agent type in a simulation.

The middle and bottom sections in figures 8.7 and 8.8 point to
another reason why wealth levels may vary for different trader types
and learning speeds. Even when told to use select best or roulette wheel
selection for forecast production, SFI agents may not be able to do so
142 8 An Analysis of Wealth Levels

450000
475000

425000

Wealth
400000

375000
350000
100 Rules 325000
100 Rules 300000 75 Rules

1000
750
75 Rules 50 Rules

1000

500
750

250
50 Rules 25 Rules

500

100
75
250
25 Rules 10 Rules

50
100

25
75

5 Rules

10
10 Rules
50

5
25
10

5 Rules
5

2500

2000
2500

Active Rules
1500
2000
1500
1000
1000
500 500
100 Rules 0
75 Rules 100 Rules 0
1000

75 Rules
750

1000
750
50 Rules 50 Rules
500

500
250
250

25 Rules

100
25 Rules
100

75
10 Rules
50
75

25

10 Rules 5 Rules
50

10
25

5 Rules
10
5

8000000 8000000

Select Average
6000000 6000000

4000000
4000000
2000000
2000000
100 Rules 0
100 Rules 0 75 Rules
1000

75 Rules
750
1000

50 Rules
500
750

50 Rules
250
500

25 Rules
100
250

25 Rules
75
100

10 Rules
50
75

25

10 Rules
50

5 Rules
10
25

GA-interval
5

5 Rules
10

GA-interval
5

Fig. 8.7. Top: Final wealth levels for SFI agents with select best (left) and
roulette wheel selection (right). Middle: Number of activated rules. Bottom:
Number of “select averages” during simulation. Data were averaged over 10
simulation runs.

and employ certain fall-back methods. By increasing the learning speed

for SFI agents or by reducing the number of rules they possess, the size
of their activated rule set shrinks as shown in the middle section in
figure 8.7. When there are no active rules at all, SFI agents derive
their forecast parameters as a fitness-weighted average of all the rules
 8.4 Wealth Levels in the SFI-ASM: Alternative Explanations 143

425000
475000

Wealth
375000 425000
375000
325000 325000
100 Rules 100 Rules 275000
275000
75 Rules

1000
75 Rules

1000

750
50 Rules

750

500
50 Rules

500

250
25 Rules
250

100
25 Rules
100

75
10 Rules
75

50
10 Rules
50

25
5 Rules
25

10
5 Rules
10

5
5

3000 3000
2500 2500

Active Rules
2000 2000
1500 1500
1000 1000
500 500
100 Rules 100 Rules 0
0
75 Rules 75 Rules

1000
1000

750
750

50 Rules

500
50 Rules
500

250
250

25 Rules

100
25 Rules
100

75
75

10 Rules
50
10 Rules
50

25
25

5 Rules
10

5 Rules
10

5
5

8000000 8000000

Select Average
6000000 6000000

4000000 4000000

2000000 2000000

100 Rules 0 100 Rules 0

75 Rules 75 Rules
1000
1000

750

50 Rules 50 Rules
750

500
500

250

25 Rules
250

25 Rules
100
100

75

10 Rules
75

10 Rules
50
50

25

5 Rules
25

10

5 Rules GA-interval
10

GA-interval
5
5

Fig. 8.8. Top: Final wealth levels for non-classifier agents with select best
(left) and roulette wheel selection (right). Middle: Number of activated rules.
Bottom: Number of “select averages” during simulation. Data were averaged
over 10 simulation runs. Note that the wealth peak in the upper left corner for
select best, reaching as high as 1.7 million, was truncated for better visibility.

that have been activated at least mincount-times in their lifetime. If

none of the rules have been activated at least mincount times, SFI
agents resort to using the global average of past prices and dividends
as their forecast. Each of the intended or fall-back selection methods,
however, has an influence on the quality of the forecast production and
144 8 An Analysis of Wealth Levels

thus influences the model behavior. The relative frequencies with which
those methods are used depend on the the parameterization, especially
on learning speed, and will thus lead to different model behaviors. I
subsume both fall-back mechanism in the SFI-ASM as select average.
The bottom section in figure 8.7 shows how SFI agents increasingly
resort to select average when the number of activated rules becomes
smaller.
Non-classifier agents, on the other hand, are not plagued by the
problem of a diminishing active rule set. Their rules are always acti-
vated (middle section of figure 8.8) and hence, there is no need for them
to resort to any of the fall-back methods (bottom section). Even though
non-classifier agents indeed stick to the intended selection mechanism,
we still see a wealth peak for GA-intervals around 50 and a valley at
about 500 for roulette wheel selection. These remaining wealth differ-
ences most likely reflect the working of the GA which changes the fore-
cast parameters, yet other still unidentified influences are possible, too.
Separating and attributing price and wealth effects to each individual
factor seems impossible, though. Changing one factor affects the price
series, which in turn will affect the GA and the forecast parameters,
which in turn will affect the triggering change. Because of these inter-
dependencies, an exhaustive explanation of price and wealth behavior
cannot be given.
Identifying all causes and interdependencies of wealth dynamics,
however, is not necessary at this point and would not justify the com-
putational expenses. All that is necessary is to show that some reasons
other than pattern recognition and exploitation influence the wealth
distribution to a much greater extent than technical pattern exploita-
tion ever could. Even though the game-theoretic analyses by Joshi et
al. themselves are correct, they are based on wealth levels that should
not be given any economic meaning.
Up to this point, we have not addressed whether one of the two
selection mechanisms, select best or select roulette, is better suited.
However, section 4.4.2 has already discussed the problem of scaling
invariance for simple roulette wheel selection. Remember that scaling
invariance means that adding an offset to all fitness values tends to
equalize the selection probabilities in roulette wheel selection. In order
to avoid negative fitness for trading rules, a constant C with a value
of 100 is added to each raw fitness.13 This effectively leads to almost
uniform selection probabilities. The fitness information is thus widely
neglected, especially in large active rule sets. Hence, we can safely as-
13
See also equation 6.8 on page 98.
 8.5 A Verdict on Wealth Analysis in the SFI-ASM 145

sume that unscaled roulette wheel selection as used by Joshi et al. is

inferior to the select best mechanism used by LeBaron et al.

8.5 A Verdict on Wealth Analysis in the SFI-ASM

This chapter has clearly demonstrated that the only source for wealth
accumulation in the SFI-ASM framework is the aggregate risk pre-
mium. Since simulated prices are usually below the hree-benchmark,
wealth generated in the hree-regime can be said to constitute a lower
bound for wealth in non-hree simulations. In a way, the smartest agents
who are able to coordinate on the correct hree-solution make the least
amount of money. It was furthermore shown that an agent’s wealth is
highly affected by the size of his activated rule set and the selection
procedure that he uses to choose a trading rule to act upon. Last, but
not least, agents in the SFI-ASM are modeled as myopic by having
a constant absolute risk aversion utility function. Since the CARA-
assumption implies that an agent’s wealth level has no impact on his
demand function, wealthier traders do not have a greater impact on
market prices than poorer traders [242]. Because of all these issues, the
SFI-ASM is highly unsuitable to address issues of wealth accumulation.
The game-theoretic analyses by Joshi et al. thus rest on simulation re-
sults that should not be given any economic meaning.
It was also noted before that the trading rules in the SFI-ASM are
not evaluated according to their wealth generating capabilities, but
on their forecast accuracy. Wilpert [437] based the fitness evaluation
of trading rules according to the wealth they would have generated if
used by agents.14 Equation 6.7 on page 98 as the main determinant of
rule fitness (equation 6.8) would have to be adapted to

1 1
fW,t = 1 − fW,t−1 + [pt + dt − (1 + rf )pt−1 ] x
t−1 , (8.9)
θ θ
with fW,t being a raw fitness values based on wealth, and x t−1 the op-
timal demand of last period’s portfolio decision. The term in brackets
multiplied by that optimal demand would denote an excess profit. It
is now a legitimate question to ask whether this type of rule evalua-
tion is still justified since we just have derived that wealth levels in the
SFI-ASM are economically not meaningful. There is, however, noth-
ing wrong with this profit based rule evaluation. Agents derive their
optimal demands by neglecting any subsequent wealth taxation. Since
14
This type of fitness evaluation was first suggested by Brock and Hommes [54].
146 8 An Analysis of Wealth Levels

the profitability of trading rules, too, is based on their excess profit

generation before taxation, it would be completely justified to use this
fitness evaluation.
9
Selection, Genetic Drift, and Technical Trading

I have yet to see any problem, however complicated, which,

when you look at it in the right way, did not become more
complicated.
Poul W. Anderson

9.1 Introduction

A mutation operator modification suggested in chapter 7 resulted in

drastically different model behavior. Instead of discovering profitable
technical trading possibilities, most agents voluntarily gave up the
use of their classifier system. On the one hand, this model behavior
prompted me to analyze the implications for wealth accumulation by
agents. On the other hand, the model behavior was so radically different
from the original results that I kept designing tests and procedures that
could bolster my argument of an inherent uselessness of the classifier
system. In the course of these tests, the robustness of the zero-bit so-
lution proved to be very robust. In fact, it still emerged when I did not
expect it to under the given parameter combinations. Before presenting
this simulation evidence in section 9.3, I will discuss a misconception
about the proper way of detecting technical trading in the SFI-ASM.
The main part of this chapter will then identify and discuss the zero-bit
solution as a side effect of genetic drift. In population genetics and in
evolutionary programming, genetic drift is known to affect the genetic
variation in finite population sizes even in the absence of any selec-
tion pressure. By transferring the longstanding selectionist-neutralist
debate from evolutionary biology to the field of genetic algorithms, I
will separate the effects caused by genetic drift and those caused by
148 9 Selection, Genetic Drift, and Technical Trading

selective forces in the two SFI-ASM versions. As a result, I will deter-

mine that the original mutation operator is effectively combatting the
influences of genetic drift while the modified operator is subject to ge-
netic drift itself. By using different tests than a superficial look at the
aggregate bit level, I will finally resolve the question whether agents in
the SFI-ASM use technical trading information or not.

9.2 Technical Trading and the Aggregate Bit Level

The first signs that a superficial look at the aggregate bit level is in-
appropriate in judging the emergence of technical trading appeared in
figure 7.3 on page 122. It was obvious that certain bits within the econ-
omy were often set in multiples of 100, which is the number of trading
rules agents possess. This implied that some bit positions in the rule
sets of bit-neutral agents were completely filled with 0- or 1-bits while
other bit positions seemed to be completely neglected. While figure 7.3
was derived for a noiseless periodic dividend process, the bit behavior
should have been similar for stochastic dividend processes if techni-
cal trading had emerged. Under the standard AR(1) dividend process,
the number of instances of bit fixations at the 100% level almost com-
pletely disappeared for bit-neutral agents. The few exceptions were seen
as temporary lock-ins at a suboptimal solution since in the long run,
they also ended up with the zero-bit solution.
The simulation evidence of certain bit positions in an agent’s rule be-
coming completely filled with non-# bits is consistent with the expected
long run asymptotic behavior of genetic algorithms. It was shown by
Holland that once an allele or a schema has been found to be useful,
it will spread through the whole population [182]. In the long run, all
chromosomes in a population will become identical. Once an agent has
found a good trading rule, one would expect the complete rule set to
be characterized by bit fixations at either the 0 or 100% level.1
Therefore, the remark by LeBaron et al. that bit costs were intended
to ensure that each bit is actually useful in improving the forecast
ability of a trading rule takes on a different meaning [244, p. 1497].
For a long time I, had interpreted that remark literally, i.e., I simply
assumed that it meant one single bit in one particular trading rule.
It now becomes clear that “each bit” only makes sense if it refers to
1
GAs that employ niching methods [155] such as fitness sharing and crowding are
capable of locating and maintaining several solutions within a single population.
Mahfoud [270] derives lower bounds for the population size to maintain a desired
number of niches. An overview of niching methods can be found in [272].
 9.2 Technical Trading and the Aggregate Bit Level 149

a particular bit position in the entire rule set of an agent.2 In order

to infer whether technical trading has emerged or not, one should not
look at the general bit level. The proper way of assessing the existence
of technical trading in the SFI-ASM is to check whether there are bit
positions in agents’ rule sets that are completely filled with non-# bits.
Since John Holland, the creator of genetic algorithms, was one of the
founding fathers of the SFI-ASM, it is hard to believe that the SFI-ASM
architects were not aware of this. It is more likely that they did not
take the implications of bit convergence fully into account. Nowhere in
their series of papers that describe the SFI-ASM [331, 11, 330, 244] do
they refer to bit positions and bit fixations at the 100% level. Instead,
they always focused on the aggregate bit level, averaged over all trading
rules of all agents. Only vaguely can one sometimes infer an interest
in the use of individual bits. In Arthur et al., for instance, one finds
the statement that “it can be said that technical trading has emerged
if bits 7–10 become set significantly more often, statistically, than the
control bits” [11, p. 27]. First, I suppose that this statement was still
intended to apply to the aggregate bit level. To detect a supposedly
useful trading condition on the level of an individual trader, one does
not need elaborate statistical tests since it is easy to spot a bit fixation
at the 100% level, especially when all other trading conditions should
have very low bit frequencies. More importantly, though, is that the
significance test as suggested above cannot be applied in the original
model. Remember that the control bits were, by definition, always set
to zero or one. It is, therefore, impossible for any trading condition to
be set more often. Another indication of an interest in disaggregated bit
information can be found on page 1507 of the article by LeBaron et al.
[244]. There, they narrow down their focus to the use of one particular
bit position, yet it is still averaged over all agents in the economy, and
the question of possible bit fixations was not addressed.
Since the aggregate bit level may get artificially inflated at higher
learning speeds through the mutation operator, it is unsuitable to in-
fer the degree of technical trading in the model as long as we do not
know the exact impact that mutation has on it. There may be a high
correlation between the aggregate bit usage and bit fixations, but we
cannot simply assume this correlation, especially when knowing about
the upward bias of the mutation operator. In order to measure the de-

2
Depending on the interpretation, the “each bit” formulation in their papers may
not be wrong, but I would call it at least unfortunate. It may have caused consid-
erable confusion and misinterpretation in the area of genetic algorithms among
novices, such as I was when I started working with the SFI-ASM.
150 9 Selection, Genetic Drift, and Technical Trading

gree of technical trading in the model, one has to investigate whether

there are bit fixations. To assess whether there is emergent technical
trading at higher learning speeds, one has to test whether the number
of bit fixations increases for faster learning speeds.

9.3 The Zero-Bit Solution: Some Disturbing Evidence

When looking at figure 7.1 on page 116, one notices that there are
some latecomer agents who experience difficulties in finding the correct
hree-solution. They appear to be temporarily locked into a suboptimal
solution, i.e., one or two specific trading conditions are set to either
zero or one in all of their trading rules. Since these bit fixations cannot
be changed by the the crossover operator, they are effectively fixed.
Only mutations and generalizations are able to change several of these
bits such that the #-bits can finally take over. It is obvious that this
may take a long time to happen. Some long run testing with bit costs
greater than or equal to 0.01, however, yielded that the latecomers
are eventually able to arrive at the zero-bit solution, too. The question
arises, though, as to why some of those bit fixations arose to begin with.
Was it a random mistake on the agents’ part, or did they temporarily
discover some useful trading information?
Another peculiarity that figure 7.1 reveals is the apparent speed
with which the majority of agents is able to find the zero-bit solution.
No matter what GA-invocation interval is used, the first agents to dis-
card their classifier system need only between 10 to 20 GA-invocations.
Sixty to seventy percent of the other agents follow shortly after. This,
however, seems quite fast for a GA-learning algorithm to discover the
optimum solution. It is only later on that the discovery process slows
down.3
The simulation results with the modified SFI-ASM have so far un-
equivocally shown that the zero-bit solution is robust. My growing con-
viction of its validity, however, was shaken when I ran the model under
no bit cost. If small efficiency gains from using condition bits had been
overcompensated by the associated bit costs, agents would have been
forced into the zero-bit solution. Thus, the model behavior was inves-
tigated under zero bit cost. To exclude that the bit decreasing effect
3
I should add at this point that figure 7.1 was derived with the consistency check
switched on. Since the rule set then contains only logical trading rules, the num-
ber of rules were reduced to 50. The reason why this might have an impact on
the speed with which agents find the zero-bit solution will become clear in the
following sections.
 9.3 The Zero-Bit Solution: Some Disturbing Evidence 151

Fig. 9.1. Aggregate bit level in the modified SFI-ASM for zero-bit costs,
no generalization, 25 bit-neutral agents with 100 trading rules, an initial bit
probability of 0.05, a crossover probability of 0.1, and a GA-invocation interval
of 100. In spite of zero-bit costs, 4 agents arrived at the zero bit level for
fundamental bits and 5 agents for technical bits within less than 17,000 peri-
ods.

of the generalization procedure was responsible for the appearance of

zero-bit agents, generalization was deactivated.
For no bit cost at all I intuitively expected a global bit fraction fluc-
tuating around its initial level and no zero-bit agents. It was surprising
when many, albeit less agents, still ended up in the zero-bit state. In
figure 9.1, a somewhat stable aggregate bit level materialized only af-
ter four agents had given up all their fundamental trading bits and five
other agents had given up their technical trading bits. The continuing
existence of the zero-bit solution under no bit cost may point towards
a systematic influence that exerts a downward pressure on bit usage.
The hypothesis of a stationary bit level and no zero-bit agents for
zero-bit costs, on the other hand, was supported for an initial bit proba-
bility greater than a certain threshold level and for an inactive general-
ization procedure. In the simulation run shown in figure 9.2, no zero-bit
152 9 Selection, Genetic Drift, and Technical Trading

Fig. 9.2. Aggregate bit level in the modified SFI-ASM for zero-bit costs,
no generalization, 25 bit-neutral agents with 100 trading rules, an initial bit
probability of 0.15, a crossover probability of 0.1, and a GA-invocation interval
of 100. Under this parameterization, no zero-bit agents appeared within the
first 30,000 periods.

agents appeared within the first 30,000 periods and the aggregate bit
level seemed to have reached its long run equilibrium. Note that the
only difference in the parameterization between figures 9.1 and 9.2 is
the initial bit level with which trading rules were initialized.

9.4 Random Genetic Drift in Genetic Algorithms

Because of the emergence of zero-bit agents under no bit costs, the

zero-bit solution turns out to be more robust than one would normally
expect. It became increasingly clear to me that zero-bit agents most
likely were a result of a phenomenon known as genetic drift. In pop-
ulation genetics and in evolutionary programming, this concept refers
to the loss of genetic diversity due to accumulated stochastic selec-
 9.4 Random Genetic Drift in Genetic Algorithms 153

tion errors in finite populations, even in the absence of any selection

pressure.4
After I had realized that useful trading bits should spread through
the whole population of trading rules, I had assumed, in line with Hol-
land’s theoretical results on genetic algorithms, that this can only be
the result of natural selection. Holland’s results, however, were derived
under some simplifying assumptions, such as an infinite population size,
an accurate mapping of the utility values of a solution into a fitness
function, and the absence of gene interaction in a chromosome (epis-
tasis) [182].
In their short review of the history of evolutionary computation, De-
Jong et al. mentioned that the initial lack of computational resources
led GA-researchers to do very few simulation runs with populations
consisting of generally less than 20 individuals [93]. Those first ex-
periments often produced results that deviated from the theoretically
expected behavior. It soon became clear that those deviations were the
consequence of genetic drift within finite populations.
Goldberg and Segrest [156] distinguished two types of stochastic er-
rors in GAs with a finite population size which may prevent a GA from
converging to the optimal solution. Sampling errors may occur because
the population is not representative in that it does not contain a highly
fit schema that can propagate. The stochastic errors in selection, how-
ever, are more important and were given a special name—genetic drift.
Because a new population is generated by selecting offspring from a
finite sized parent generation, allele frequencies in this population are
subject to random selection errors. With infinite population sizes, these
errors will cause only negligible changes, yet in small populations, they
will be more than noticeable. Even under no selection pressure, i.e.,
with a constant fitness function, the accumulation of these stochastic
errors will cause certain alleles to become more predominant, and the
members of a genetic population will eventually converge to a single in-
stance in the solution space. Beasley et al. [27] describe the mechanism
of genetic drift as a ratchet effect. That is, once an allele has spread to
all members of a finite population, it is effectively fixed since crossover
cannot change it anymore. This ratchet effect will eventually cause each
individual gene in the population to become fixed. The effect of genetic

4
Genetic drift is also sometimes characterized as random fluctuations in finite
populations, “evolutionary forgetting”, or allele loss. It is occasionally referred to
as the Sewall Wright effect, named after one of the founders of population genetics.
Genetic drift in the context of population genetics is discussed, for instance, in
[414, 170, 174].
154 9 Selection, Genetic Drift, and Technical Trading

drift thus, can lead to a premature convergence of the GA before the

optimal solution has been found.
Goldberg and Segrest analyzed genetic drift in a GA acting with the
operators of reproduction and mutation on a one-locus binary struc-
ture. By using a Markov chain analysis, they calculated the expected
time of first passage to various convergence levels. They found that
increasing the mutation rates increased the expected time until a par-
ticular convergence level was reached for the first time. This implies
that an increase in mutation rates reduces the impact of genetic drift.
For mutation rates that are too high, on the other hand, the search be-
comes effectively random since the gradient information in the fitness
function is not exploited, as Beasley et al. point out.
Unlike the calculations in terms of convergence time, approaches
to quantify the effect of genetic drift based on population variances
allow exact analytical expressions. Beyer [31] calculates genetic drift
in unrestricted and binary search spaces. Let us assume that there is
a parent population of µ single-locus individuals whose gene values
are bi ∈ 0, 1 for i = 1, . . . , µ for the binary case. Beyer defines the
population variance σP2 [b] as
µ µ
1
1

σP2 [b] := (bi − b̄)2 , b̄ = bi , (9.1)
µ µ
i=1 i=1

with b̄ being the average gene value. The offspring generation is created
through random sampling, i.e., one offspring is obtained by randomly
choosing one of the µ parents, a procedure that is repeated µ times.
Beyer then derives the expected population variance of the offspring
generation as
2 µ−1 2
σO = σP . (9.2)
µ
Equation 9.2 demonstrates that random selection in finite popula-
tions reduces the expected population variances. A similar calcula-
tion of genetic drift based on population variance was done by Rogers
and Prügel-Bennet who then applied their results to various selection
schemes such as steady generational replacement, state selection, and
generation gap methods [358].

9.5 The Neutralist–Selectionist Controversy

Even though genetic drift is often defined as the phenomenon of allelic

convergence under no selection pressure, it is important to realize that
 9.5 The Neutralist–Selectionist Controversy 155

it is still at work when there is a fitness function. This immediately

leads to the question of whether one is able to separate the effects of
“natural” selection from those of genetic drift, and which of the two
evolutionary forces are more important in the evolutionary process.
In fact, Harrison et al. point out that the question about the rela-
tive importance of genetic drift versus natural selection in determining
evolutionary change was and still remains one of the most controversial
issues in population genetics [170]. The so-called neutralist-selectionist
debate arose between Sewall Wright and Ronald A. Fisher in the 1930’s
and 1940’s when they started to investigate the fixation probabilities of
newly mutant alleles in a given population. It has been rekindled with
the ascent of the neutral and near-neutral theories of molecular evolu-
tion by Kimura [222, 223] and Ohta [326, 327] about the neutrality or
near-neutrality of genes.
The following description of the Wright-Fisher model is intended to
highlight the differences between these two different schools of thought
in population genetics. The Wright-Fisher model will also form the
basis for our following analysis of whether selection or random ge-
netic drift is more important in the two versions of the SFI-ASM. The
Wright-Fisher model assumes a population of N haploid organisms5
who create an offspring generation of N individuals through random
mating. Let us consider the case for two alleles A and a with corre-
sponding allele frequencies p and q. Assume that there are i copies of
A at a certain gene position, i.e., p = i/A and q = 1 − i/A. A’s relative
fitness w is defined as its absolute fitness, often the number of expected
offspring, divided by the average fitness in the population. The selec-
tion coefficient favoring A is denoted by s = w − 1 and can take any
value in the interval [−1, ∞).6 We are interested in the fixation prob-
ability U (p) of this allele for varying degrees of selection pressure and
differing population sizes. There is, however, no closed-form solution to
this problem in discrete time, and enumerative algorithms work only
for very small population sizes. A diffusion approximation in a contin-
uous time formulation was developed by Kimura [220, 221],7 and A’s
probability of fixation is determined as
5
A haploid organism contains only one copy of each chromosome. Diploid organ-
isms, i.e., most sexually reproducing species, contain two copies of each chromo-
some, one from each parent. The genetic algorithm in the SFI-ASM, as most GAs,
is a haploid GA-representation.
6
The selection coefficient could also be defined as being against A. In this case,
s = 1 − w would be in the interval (−∞, 1].
7
A relatively easy to follow derivation of this result can be found in [224] and [126].
Since the diploid case, i.e., each individual owning two sets of chromosomes,
156 9 Selection, Genetic Drift, and Technical Trading

1 − e−2N sp
U (p) ≈ . (9.3)
1 − e−2N s

1.0 ....
......
.............................................................................................................................................................................................................................................................................................
..............
....................
.. ..................
...... ...... .....
... ........... ...... ........ .....
... ......... ...... .... .. .. ..
..... ........ ...... ..... ... .. ..
0.9 50 ....
......
.... ...........
. .. ......... ......... ..... .... .....
...... ........ ......
. . .. ..
Fixation Probability U (p)
.. ...... ..... ... ...
... ..... ......
.
..... ....... ... ...
... .... ...... .... . ... ..
..... ..... ..... .......
0.8 ..
...
...
.....
....
....
..........
.....
.....
.
. ........
.....
.
.....
.. ....
.
....
.
....
..
.. .
. ... .....
.. ..
... .....
.... ..... ..... .... ...
... .... .... .... .... ... ..
... .... ..... ..... .... ...
0.7 .....
.... 5 .. ...
....
.....
..
. .....
... .....

.....
.........
........
....
...
.....
.
.....
.
...
.
...
...
.
.
.
.
..
...
..
. ... ..... .... ..... .. ...
... ... ..... ..... ..... ... ...
... .. ..... .... ....
0.6 ...
.....
...
.
.
...
...
1 ....
........
.....
.....
.........
.....
....
........
....
.
.
...
...
... .
.
.
...
...
...
.. ..... .... .... ...
... ... ..... ..... .... ... ...
... .. ..... .... .... ...
0.5 ..
...
.... ..
...
.
...
...
....
.. .
.....
.....

....
.....
........
.....

.....
.....
........
....
...
..
.
...
.. ...
.
.
..
...
... .... ..... ..... ...
... .... .... ...
... ... .... ..... .... .. ...
0.4 .
...
..... ..
...
.
....
.....
.....
......
.
....

....
.
.....
........
.
.....
.....
.........
.....
−1 ...
....
.
... ..
.
.
..
...
... ... ..... ..... ..... .. ...
... ... .... .... ..... ... ...
... .... ..... ..... ...
0.3 .....
..
.
.
...
...
.
....
....
.......
.
....
.....
........
....
.....
. ..
. .....
−5 ....
....
.... .
..
...
...
.
... ... .... ..... ..... .... ...
... ... .... .... ..... .... ...
... ... .... ..... ..... ....
0.2 ..... ....
.. ...
... .. ....
.
...
.....
.
.... ........
.... ..
......

......
.....
.... . ........

.....
....
.....
..
......
..
...
.
....

.. ... ... ....

..
...... ..... ...
... ... .... ..... ...... ...... ...
... .....
0.1
..... ...... ...
..... .... .... ........ ..........
.. ... .... ...... ....... ........
.......
.. ......
−50 ...
.
.
..
...... ... ...... ......... .......... ...
...... ................. ............ ...
.......................... ...............
...................... ....
0 .........................................................................................................................................................................................................................................................................................

0.0 .25 .5 .75 1.0

Initial Allele Frequency p

Fig. 9.3. Probabilities of fixation of an allele for various values of 2N s and p.

The values of 2N s are shown adjacent to the curves, except for the diagonal
with no selection pressure (2N s = 0).

Figure 9.3 plots the fixation probability U (p) as a function of p for

various values of 2N s. Let us first consider the case under no selection
pressure. For s = 0, both the numerator and denominator of equation
9.3 are zero, but by using L’Hôpital’s theorem, we obtain that the
probability of fixation for an allele equals its initial allele frequency p.
For positive selection coefficients, the fixation probabilities of A exceed
its initial allele frequencies. It is, however, obvious from equation 9.3
that the selection coefficient s, alone, is not sufficient to decide about
the relative importance of natural selection versus genetic drift. It has
become commonplace in evolutionary biology to regard
1
s> (9.4)
2N
as the threshold at which natural selection begins to have a significant
impact on allele fixation. Otherwise, the alleles are either called nearly
is more common in nature, one often finds the diploid formula U (p) ≈ (1 −
e−4N sp )/(1 − e−4N s ).
 9.6 Fitness Driven Selection or Genetic Drift? 157

neutral or effectively neutral, i.e., their frequencies are determined more

by genetic drift [80].
While Fisher explicitly assumed that the population size (up to
1012 individuals) is extremely large, Wright thought more of local pop-
ulations and explicitly considered effective population sizes.8 For ex-
tremely large N as Fisher had assumed, very small selective forces are
sufficient, given enough time, to cause large adaptive changes. In con-
trast to this Darwinian or Fisherian gradualism as it is sometimes
called [432], Wright espoused a much larger role of genetic drift in
evolutionary adaptation due to smaller population sizes. He rejected
Fisher’s idea that either selection or genetic drift had to alone be re-
sponsible for the observed fluctuations in gene frequencies. Instead, he
accepted that several evolutionary forces—mutation, random genetic
drift, selection, and migration—act simultaneously. As a result of this
controversy, random genetic drift was dubbed the Sewall Wright Effect
by Fisher. The modern version of this debate, the neutral and nearly
neutral theories of molecular evolution, are mainly concerned with the
percentage of alleles that are neutral or near-neutral, i.e., with zero or
only negligible effects on an individual’s fitness. Similar to Wright, the
difference in the traditional selection theory lies in the magnitude of
the N s coefficient [327].

9.6 Fitness Driven Selection or Genetic Drift?

9.6.1 Selection or Genetic Drift in the Modified SFI-ASM?

The idea that the fixation of trading bits at the 100% level in the SFI-
ASM necessarily reflects their advantageousness is obviously equivalent
to the position held by the selectionist faction. The view that vanishing
trading bits are likewise a result of selection would equally be favored
by the selectionists.
The emergence of zero-bit agents under no bit costs, however, clearly
points to genetic drift as a major evolutionary force in the modified SFI-
ASM. Since genetic drift has not been taken into account yet, the sim-

8
The N in equation 9.3 actually refers to effective population size. In order to
determine effective population sizes, evolutionary biologists make adjustments
for unequal sex ratios, strong sexual selection, or fluctuating population sizes
[80, 174]. In genetic algorithms, such adjustments are usually unnecessary. If,
however, the SFI-ASM founders would have opted to prevent illogical trading
rules from reproducing, the effective population size would have been lower than
the number of trading rules.
158 9 Selection, Genetic Drift, and Technical Trading

ulation results presented so far are in need of reinterpretation. Judge-

ment about whether there is technical trading in the SFI-ASM or not
has to be postponed.
The Zero-Bit Solution and Initial Bit Probability
In hindsight, the choice of initial parameter values made bit-neutral
SFI agents particulary vulnerable to getting locked into the zero-bit
solution through genetic drift. The trading rules were initialized with a
bit probability of only 0.05. Since 95 % of all trading bits were already
don’t care signs #, random genetic drift was much more likely to fixate
the #-bits than to propagate the 0 and 1-alleles through the population.
When switching off mutation and setting bit costs to zero, equation 9.3
tells us that, on average, 95% of all bit positions would fixate at the
zero-bit level. The remaining five percent will eventually fixate at the
100% level.9
This also explains why a simple increase in the initial bit level of
trading rules drastically reduced the emergence of zero-bit agents in fig-
ure 9.2 on page 152. Even though no zero-bit agent appeared within the
first 30,000 periods in that particular simulation, additional tests have
shown that further increasing the initial bit probability cannot prevent
zero-bit agents from eventually showing up in the long run. Zero-bit
agents are, in a way, inevitable. In cases with high initial bit probabili-
ties, the generalization procedure first lowers the bit level, which paves
the way for genetic drift to work its way toward the zero-bit solution.
Differentiation of Trading Rule Sets
It is also known from population genetics that random genetic drift
tends to create subpopulation differentiation [80]. The analogy between
an agent’s rule set and the subpopulation concept in population genet-
ics is hardly contentious. All rule sets were initialized randomly and
thus, had the same stochastic properties such as allelic frequencies. Yet,
after they started to independently evolve, random genetic drift led to
an increase in the among-population variance (differentiation), and to a
decrease in the within-population variation.10 When discussing the bit
distribution under periodic dividends in figure 7.3 on page 122, it was
pointed out that each agent converged to a single trading rule that he
was using, yet all agents seemed to have discovered a different techni-
9
The Wright-Fisher model in the previous section did not take random mutations
into account and assumed random mating. Mutation and the fact that the two
SFI-ASM versions choose parent rules through tournament selection will have an
impact on the exact fixation probabilities.
10
The decrease in the within-population variance was already captured in equation
9.2.
 9.6 Fitness Driven Selection or Genetic Drift? 159

cal trading rule. In simulation runs with many traders, subpopulation

differentiation may prevent one from detecting bit fixations when only
looking at the bit usage across all agents.
Latecomer bit-neutral agents
Let us assume for a while that the majority of bit-neutral agents who
arrive at the zero-bit level rather quickly do so through genetic drift
and not by fitness driven selection. The latecomer agents then suddenly
become of particular interest to us.

Table 9.1. Typical rule set of a bit-neutral “latecomer” agent. The rule set
contains 100 trading rules, each with 64 trading bits. The last row shows the
sum of all non-# bits.

1 2 3 4 5 6 7 · · · · · · 58 59 60 61 62 63 64
1 0 # 0 # # # # ······ # 1 # # # # #
2 # # 0 # # # # ······ # 1 # # # # #
3 # # 0 # # # # ······ # 1 # # # # #
.. .. ..
. . .
98 # # 0 # # # # · · · · · · # 1 # # # # #
99 # # 0 # # # # · · · · · · # # # # 1 # #
100 # # 1 # # # # · · · · · · # 1 # # # # #

1 0 100 0 0 1 0 0 99 0 0 1 0 0

The fixation of certain bit positions in the trading rules of latecomers

could be either a mistake on their part, i.e., a result of genetic drift, or
indeed reflect fitness driven selection. A typical rule set of a latecomer
agent is shown in figure 9.1. Long run simulations have shown though
that those agents will sooner or later become zero-bit agents themselves.
The reversal of bit fixations, however, is not necessarily a sign of fitness-
driven forces finally gaining ground. This interpretation is likely, but
again, equation 9.3 tells us that there is a small, but positive probability
of this happening, even under no selection pressure.

9.6.2 Selection or Genetic Drift in the Original SFI-ASM?

Having inferred from the zero-bit solution under no bit costs that ran-
dom genetic drift plays a major role in the modified SFI-ASM, does
its absence in the SFI-ASM indicate the dominance of selective forces?
Not necessarily. Given the upward biased mutation operator, a zero-bit
solution in the SFI-ASM is impossible to begin with.
160 9 Selection, Genetic Drift, and Technical Trading

Table 9.2. Typical rule set of a SFI agent. The rule set contains 100 trading
rules, each with 64 trading bits. The last row shows the sum of all non-# bits.

1 2 3 4 5 6 7 · · · · · · 58 59 60 61 62 63 64
1 # 0 # 0 # 1 # ······ # # 1 # 0 # #
2 # # # 1 # 1 # ······ # # 1 # 0 # #
3 # # # 0 # 1 # ······ # # 1 1 0 # #
.. .. ..
. . .
98 # # # # # 1 # · · · · · · # # # # # 1 #
99 # 0 # 0 # 1 # · · · · · · # # 1 1 0 # #
100 # # # # # 1 # · · · · · · # # # # 0 1 #

1 4 0 96 0 100 0 0 1 93 5 99 5 0

A typical rule set of SFI agents shown in table 9.2 reveals that there
are indeed bit positions that have fixated. Instead of providing data on
the aggregate bit level in the economy, this could have been a much
stronger argument for emergent technical trading. After all, those bits
close to the 100% level have emerged from low initial bit frequencies.
But again, it is possible that those bit positions became fixed through
random genetic drift and not through selective forces. If we compare the
two rule sets shown in tables 9.1 and 9.2, one notices several things.
First, compared to bit-neutral latecomers, SFI agents have more bit
positions that are close to zero, but not exactly zero. Second, some bit
positions in SFI-rule sets tend to be only “nearly” fixated, i.e., they
are also close to, but not exactly at the 100% level. In the rule sets
of bit-neutral agents, the fixation is usually complete. Both effects are
probably an effect of the SFI-mutation operator which is more likely,
i.e., with probabilities of two thirds, to introduce a 0- or a 1-bit when
it finds a #-bit at a particular position, or to change non-# bits to the
don’t care symbol.
It is also possible to offer an explanation for the apparent path
dependency of equilibrium bit levels. In section 6.7.2, it was pointed
out that the equilibrium bit frequencies differ when the trading rules
are initialized with different bit probabilities. Equation 9.3 showed that
the fixation probability of trading bits increases for higher initial allele
frequencies. Starting at higher initial bit probabilities, thus, will lead to
more bit fixations and a higher aggregate bit level. It does not matter
at this point whether bit fixation is a result of genetic drift or selection,
but once it has occurred, it is hard to reverse.
 9.6 Fitness Driven Selection or Genetic Drift? 161

9.6.3 Genetic Drift, Fitness Gradient, and Population Size

How much selective pressure is necessary to ensure that the two model
versions of the SFI-ASM are working in the regime where selection
dominates genetic drift? Remember that section 9.5 established that
the condition s > 1/(2N ) has to be satisfied for selection to be primarily
responsible for allele frequencies. Certainly, this is more a rule of thumb
than a strict border separating the two regimes, but it will help us get
an idea of how much a trading bit has to improve a rule’s forecast
accuracy in order to be selected.
An individual’s i relative fitness wi was defined as its absolute fit-
ness fi over mean population fitness f¯, and the selection coefficient
in favor of individual i as s = w − 1 with s ∈ [−1, ∞). Inspection of
typical trading rule sets shows the mean fitness values of trading rules
to be mostly in the interval 92–94. Since we are only interested in an
estimate of the magnitude of the fitness differential, the exact value
does not matter that much. If we then assume that mean fitness is
93.0, by how much does a rule’s fitness value have to be improved by
an additional trading bit such that this bit is likely to become fixated
through selection? Substituting s with w − 1 and rearranging equation
9.4 yields

¯ 1
fi > f 1 + (9.5)
2N

1
> 93.000 1 +
200
> 93.465.

Each trading bit in a rule should, thus, improve a rule’s forecast

accuracy not only by the associated bit costs of 0.005, but by an ad-
ditional offset of about 0.5 which is necessary to counter the effect of
genetic drift.11
However, having no idea whether, and if so, by how much, a single
trading bit is able to improve a rule’s forecast accuracy, this calculation
does not help us in deciding whether the gradient information contained
in fitness functions is sufficient for selection to dominate genetic drift.
It does illustrate, however, that a trading bit offsetting only slightly
more than its associated bit costs is not guaranteed for being selected
and for spreading through the whole rule set. If, however, we would
11
There is, however, the possibility of epistasis, that is, a certain trading bit is only
useful in conjunction with another trading bit. In the derivation of 9.3, epistasis
has not been accounted for.
162 9 Selection, Genetic Drift, and Technical Trading

know by how much a certain allele would contribute to the fitness of

a chromosome, equation 9.5 could be solved for the minimum effective
population size N for a given GA-problem.

9.7 The Effect of Mutation on Genetic Drift

Since we do not know the possible magnitude of the fitness contribution

of useful trading bits, it is also possible that the fixated bit positions
in both model versions reflect genetic drift and not selection. Dynamic
stability or instability of fixated bit position can also be both a sign of
genetic drift or selection. If a gene is found to be useful, one could argue,
its alleles should remain fixed over time. On the other hand, the SFI-
ASM was developed to show that the stock market is in a constantly
co-evolving dynamic environment. Fixation reversal thus could indicate
changes in the economic environment.
We can do nothing other than concede at this moment that we are
in the midst of the neutralist-selectionist debate. How can we judge
whether the fixation of some trading bits reflects technical trading or
not? How can we separate whether the effects on the bit level have
mainly been caused by genetic drift or by fitness based selection?
In evolutionary biology, this issue is often resolved by taking the
neutral theory as a “null model” and testing whether observed differ-
ences in DNA sequences can be explained by genetic drift or not.12
In computer simulations with genetic algorithms, we do not need to
resort to the theoretical predictions by the neutral model since we are
able to simply test our genetic algorithm under no selection pressure.
The observed behavior can then serve as the null model with which the
full-blown model behavior is compared.

9.7.1 Genetic Drift, Mutation, and Crossover Only in SFI

Agents

To obtain the “null behavior” of the classifier system under no selection

pressure, the model was run in GA-only mode. In this mode, the only
activity of agents was to alter their trading rules by invoking the GA.
In addition to disabling all other model routines, certain adjustments
12
Evolutionary biologists resort to this type of test when they already have an
estimate of how much time has passed since two lineages have diverged. They
then compare whether the observed genetic differences may be explained by the
neutral model of genetic drift or not, i.e., whether selection has also had an impact
on the genetic sequences.
 9.7 The Effect of Mutation on Genetic Drift 163

to the GA had to be made. Since there are no “bad” trading rules

under a constant fitness function, newly created rules had to randomly
replace older trading rules in the rule set. The GA-parameters were set
at the same values that were used in realistic simulations runs.

1400

1200
number of instances

1000

800

600

400

200

0
0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1
bit frequency

Fig. 9.4. Number of bit positions that have a certain bit frequency within an
agent’s rule set. Data were obtained at period 250,000 from a single GA-only
simulation run with 1,000 SFI agents, each endowed with 100 trading rules
(Π = 0.9, π = 0.03, bit costs = 0, GA-interval = 1,000).

Figure 9.4 shows the bit frequency of bit positions in the rule sets of
SFI agents. A bit frequency of zero means that a certain bit position in
an agent’s rule set has no set trading bits at all. A bit frequency of one
refers to the situation when all 100 trading bits are set to either zero or
one. In a simulation run with 1,000 SFI agents who all possessed 100
trading rules consisting of 64 trading bits, a total of 64,000 bit positions
could be investigated at once. Figure 9.4 shows that a bit frequency of
zero is the most common one, even though it does not fit very well into
the overall shape of the remaining histogram. Apart from this outlier,
the bit frequencies between 0.15 to 0.35 are the most common ones.
It is also worth noticing that a few bit positions with a bit frequency
of one have emerged even though the initial bit probability was set at
0.05.
164 9 Selection, Genetic Drift, and Technical Trading
1200

1000
number of instances

800

600

400

200

0
0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
bit frequencies

Fig. 9.5. Number of bit positions that have a certain bit frequency within an
agent’s rule set. Data were obtained at period 250,000 from a single GA-only
simulation run with 1,000 SFI agents, each endowed with 100 trading rules
(Π = 0.9, π = 0.03, bit costs = 0, GA-interval = 50).

Figure 9.5 was obtained by increasing the GA-invocation interval

to 50. One notices that the mode of the distribution has shifted to the
right, that the bit frequency of zero has become less common, and that
the number of bit fixations at the 100 % level has increased. Since SFI
agents do nothing other than randomly altering the condition bits of
their trading rules, increasing the “learning speeds” simply speeds up
convergence to the equilibrium bit frequency distribution. Convergence
would be fastest if agents would call the GA in every period. The shift-
ing shape of the bit frequency distribution towards the center suggests
that the equilibrium bit distribution is centered around one half. The
remaining peak at a bit frequency of zero could be caused through the
crossover operator. Its decrease for faster GA-rates makes it more likely
though that the bit distribution has not yet reached its equilibrium. I
suppose that in the limit, the number of instances with a bit frequency
of zero equals the number of bit fixations at the 100% level.
 9.7 The Effect of Mutation on Genetic Drift 165
50.000

45.000

40.000
number of instances

35.000

30.000

25.000

20.000

15.000

10.000

5.000

0
0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1
bit frequencies

Fig. 9.6. Number of bit positions that have a certain bit frequency within
an agent’s rule set when bit costs are positive. Data were obtained at period
250,000 from a single GA-only simulation run with 1,000 SFI agents, each
endowed with 100 trading rules (Π = 0.9, π = 0.03, bit costs = 0.005, GA-
interval = 5).

Figure 9.6 shows a dramatic change in equilibrium bit frequencies

when bit costs are included.13 Most of the bit positions are completely
filled with #-signs. It is remarkable though that genetic drift has re-
sulted in 11 bit fixations at the 100% level and 3 instances at the
99%-level.
In the presence of bit costs, the proper null-model with which the
bit distribution of real SFI agents should be compared is figure 9.6.
The significance of the equilibrium bit distribution under no selection
pressure at all will become clear in section 9.7.3.

13
Under positive bit costs, the 20 worst trading rules were replaced. With no other
fitness influence than bit costs, the exact value does not matter, i.e., the bit
distributions for different bit costs would look identical.
166 9 Selection, Genetic Drift, and Technical Trading

9.7.2 Genetic Drift, Mutation, and Crossover Only in

Bit-Neutral Agents

For bit-neutral agents, the bit distribution under no selection pressure

is shown in figure 9.7. Qualitatively, it resembles the bimodal bit dis-
tribution of ordinary SFI agents in the presence of bit costs.

60.000

50.000
number of instances

40.000

30.000

20.000

10.000

0
0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1
bit frequencies

Fig. 9.7. Number of bit positions that have a certain bit frequency in the
rule sets of all bit-neutral agents when bit costs are zero. Data were obtained
at period 250,000 from a single GA-only simulation run with 1,000 bit-neutral
SFI agents, each endowed with 100 trading rules (Π = 0.9, π = 0.03, bit costs
= 0, GA-interval = 5).

There are 3,403 out of 64,000 possible bit fixations at the 100%
level, that is, 5.3% of trading bits became fixated. Given an initial bit
probability of 0.05, this is remarkably close to the theoretical prediction
of 5% by equation 9.3. The bit equilibrium distribution in the presence
of bit costs is trivial—there are no remaining bits at all.

9.7.3 An Equilibrium Analysis of Genetic Drift and

Mutation

The previous sections have shown that replacing the original mutation
operator with an unbiased operator will result in drastically different
 9.7 The Effect of Mutation on Genetic Drift 167

bit equilibrium distributions under no selection pressure. These equi-

librium distributions are the result of the interaction of genetic drift,
mutation, and crossover. Since the equilibrium properties of mutation
and genetic drift have already been studied in population genetics ,
we are able to investigate the two mutation operators in more detail.
Consider a neutral model from population genetics with reversible mu-
tation, i.e., an allele A mutates to a at rate µ and a alleles mutate to
A with probability u. If we denote the allele frequency of A at time t
with pt , A’s frequency in t + 1 is then given as

pt+1 = (1 − pt )u − pt µ. (9.6)

By using a diffusion approximation approach, the final equilibrium den-

sity for the haploid case can be shown to be

φ(p) = Cp2N u−1 (1 − p)2N µ−1 . (9.7)

A derivation for the diploid model can be found in Felsenstein [126,

chapter VII.9] and Bustamante [60]. Because equation 9.7 is the density
of a Beta distribution with parameters 2N u and 2N µ, the constant C
is
Γ (2N u + 2N µ)
C= . (9.8)
Γ (2N u)Γ (2N µ)
Figure 9.8 shows some probability density plots of allele frequencies
when mutation rates from A to a and vice versa are symmetrical. For
values of 2N u = 2N µ > 1, the Beta distribution is unimodal and the
probability mass of allele frequencies is centered around its theoretical
mean of half.14 For large values of 2N u = 2N µ ≫ 1, allele frequencies
are tightly clustered around the deterministic equilibrium, because once
genetic drift moves the allele frequencies away from this equilibrium,
mutation pushes it right back. Mutation thus dominates genetic drift.
For small values of 2N u = 2N µ < 1, the Beta distribution is U-
shaped with most of its probability mass concentrated near the ab-
sorbing states at p = 0 and p = 1. In this case, genetic drift dominates
mutation and has pushed the equilibrium allele frequencies to the ab-
sorbing states with little resistance from mutation. There are some
occasional bit reversals from one tail of the curve to the other when

14
The mean of a Beta distribution with parameters α and β is α/(α + β). The mode
of a Beta distribution (α − 1)/(α + β − 2) is only defined for α > 1 and β > 1.
168 9 Selection, Genetic Drift, and Technical Trading

10.0 ..
......
... ...
... ...
..... ..... 80
.. ..
... .....
... ...
... ...
.... ...
... ...
7.5 ..... ...

Probability Density
...
.... ...
.... ...
... ...
...
.... ...
0.1
..
. ...
... . ... .
... .
.
. ... ...
... ..... ... ...
... . ... ...
. ...
5 ...
... . ... .
.
... . ... ..
... ... ... ...
... ... ... ...
... ... ... ..
... ..
.
. .
. ...
.
.
... ...
... ... . ..
...
... ... .................. .... ...
... ... ...... .......
...
... ...... ......
...
... . .
......... .......
. .. .
2.5
. ... ..... .
... .... .. .
...
10
... .... ... ... .....
.... .. ... ..... ...
... ... .... ... .... ...
... .... ....
...
... .
.....
. ...
.
.
...
. ....
. .
.
. . ... .... .
... .... ... .... ...
... .... ...
... .... ...
1
. .... . .
.....................................................................................................................................................................................................................................................................................
..... .
.....
....... ..
.
.
.... ....
.
...
. ..... . ........
............ .. ... ..... ..
................................................. .. . ............
....................................................................................................................................................................
0.0
...
......................................................................................................... ...............................................................................................................

0.0 .25 .5 .75 1.0

Allele Frequency p

Fig. 9.8. Equilibrium distribution of allele frequencies under genetic drift

and mutation for symmetric mutation ratios 2N u = 2N µ. The values for 2N u
are shown next to the curves.

a new mutation succeeds in spreading through the population.15 The

uniform distribution at 2N u = 2N µ = 1 is the dividing line where
mutation and genetic drift balance each other out exactly.
It is now straightforward to apply equation 9.7 to the SFI-mutation
operator and to interpret the resulting equilibrium density function of
allele frequencies. Technically, our “genes” in the SFI-ASM come in
three alleles, 0-, 1-, or #-bits. We can, however, subsume the 0- and
1-bits into one “super allele” A with the meaning “bit is set”. From
the bit transition matrix 6.9 on page 101, we see that the probability
of A being mutated to # is two thirds. Similarly, the mutation rate
of #-bits to zero or one is also two thirds. Since the mean of a Beta
distribution with parameters α and β is α/(α + β), mutation presses
the gene frequencies towards an equilibrium of one half. Note that
this is the same fixed point that we have determined with our Markov
chain analysis of the SFI-mutation operator. In order to determine the
effective mutation rates of alleles in the SFI-ASM, one has to consider
the probability with which individual bits are chosen for selection, i.e.,
15
Technically, the Beta distribution for 2N u = 2N µ < 1 is not defined at its tails
for p = 0 and p = 1, but we are only approximating a discrete histogram with a
continuous density function. The diffusion approximation never predicts an allele
frequency of exactly zero or one. Practically, there will be numerous genes with
either none or all of their alleles set at a common value.
 9.7 The Effect of Mutation on Genetic Drift 169

we have to take predictor mutation probability Π and bit mutation

probability π into account. The effective mutation rates u and µ are
then determined as
2 2
u = µ = Ππ = × 0.9 × 0.03 = 0.018. (9.9)
3 3
With our population of 100 haploid trading rules, the numerical
values for 2N u = 2N µ equal 3.6, and the resulting probability density
function of allele frequencies is shown in figure 9.9. Since 2N u and
2nµ are greater than one in the SFI-ASM, the equilibrium density of
allele frequencies is clustered around the deterministic equilibrium of
one half. The unimodal probability density of allele frequencies shows
that the original SFI-mutation operator dominates genetic drift. Now
it becomes clear that the two empirical bit distributions in figures 9.4
and 9.5 are on their way to their final equilibrium distribution shown
in figure 9.9.

2.0
......................
....... .....
..... .....
..... .....
..... ....
...
1.75
...
.... ...
.. ...
Probability Density

.. ...
..... ...
... ...
1.5
. ...
.... ...
.... ...
...
... ...
1.25
.
...
.
...
...
...
. ...
... ...

1.0
. ...
...
. ...
.... ...
...
.... ...

0.75 ... ...

. ...
.... ...
...
.... ...
...
0.5
. ...
... ...
. ...
.... ...
... ...

0.25
.. ...
.
... ...
.
.... .....
...
. .....
........
. .....
......
0
...... ...
...............................................................................................................................................................................................................................................................................

0.0 .25 .5 .75 1.0

Allele Frequency p

Fig. 9.9. Equilibrium distribution of allele frequencies under genetic drift and
mutation with the original SFI-mutation operator.

The effective mutation rates u and µ for the bit-neutral agents are
dynamically self-adjusting. We can, however, generate a probability
density plot for equilibrium allele frequencies if we keep the bit tran-
sition probabilities from the bit transition matrix fixed at their values
for the initial bit frequency of 0.05. The effective mutation rates from
A to #-bits and vice versa are not symmetrical anymore and equal
u = 0.95Ππ = 0.95 × 0.9 × 0.03 = 0.02565 (9.10)
µ = 0.05Ππ = 0.05 × 0.9 × 0.03 = 0.00135. (9.11)
170 9 Selection, Genetic Drift, and Technical Trading

The parameters for the Beta distribution are then calculated as

2N u = 200 × 0.02565 = 5.13 (9.12)
2N µ = 200 × 0.00135 = 0.27. (9.13)

1.25 ...
...
...
...

1.00
...
Probability Density

...
...
...
...
...
...
...
0.75 ...
...
...
...
...
...
...
0.50
...
...
...
...
...
...
...

0.25
...
...
....
....
......
.......
..........
.................
.................................
.....................................................................................
0.00
.............................................................................................................................

0 0.02 0.04 0.06 0.08 0.1

Allele Frequency p

Fig. 9.10. Equilibrium distribution of allele frequencies under genetic drift

and mutation with the modified mutation operator when population size is
100 trading rules and effective mutation rates are kept fixed at the initial
bit probability. Note that the probability density values for allele frequencies
> 0.1 have been cut off for better visibility.

The corresponding density plot of equilibrium allele frequencies is

shown in figure 9.10. It clearly shows that most of the probability mass
is torn towards the tail with p = 0. With only 100 trading rules and the
newly suggested mutation operator, genetic drift dominates mutation.
The simulation evidence of bit fixation at the zero-bit level is, thus,
evidently a result of genetic drift and not of selective forces as previously
thought.
Equation 9.7 also allows us to determine a critical effective popu-
lation size for any given mutation operator. For symmetrical mutation
rates, the dividing line at which mutation and genetic drift balance each
other out is where 2N u = 2N µ = 1. With these parameters, the Beta
distribution turns into a uniform distribution. For unequal mutation
rates, the dividing line between genetic drift and mutation is crossed
when the smaller of the two parameters exceeds 1.0.
With the new mutation operator, the condition
2N µ ≥ 1.0
N ≥ 371
 9.7 The Effect of Mutation on Genetic Drift 171

0.04 ...
.....................
.... .......
......
... .......
.......
..... ......
.......
....

Probability Density
.......
.... ......
.......

0.03
... .......
.......
..... ......
.......
.... .......
... .......
.......
..... .......
.......
.... ......
......
.... .......
0.02 ...
.....
.......
.......
.......
.......
........
.... ........
.... .........
.........
... .........
..........
..... ...........

0.01 .... ...........

............
.... .............
..............
... ...............
..
.....
....
...
.....
0.00 .

0 0.02 0.04 0.06 0.08 0.1

Allele Frequency p

Fig. 9.11. Equilibrium distribution of allele frequencies under genetic drift

and mutation with the modified mutation operator when population size is
400 trading rules and effective mutation rates are kept fixed at the initial
bit probability. Note that the probability density values for allele frequencies
> 0.1 have been cut off for better visibility.

has to be satisfied for mutation to dominate genetic drift.16 Figure 9.11

shows a unimodal probability density function of equilibrium allele fre-
quencies when agents possess 400 trading rules. Note that even though
the continuous approximation of the density function starts at zero,
the first class in the discrete histogram is very likely filled with many
instances of allele frequencies of zero. It should be mentioned at this
point that the prior intuition about reducing the rule set size for an
activated rule consistency check was not justified. It just aggravated
the problem of genetic drift and resulted in an even earlier convergence
for most bit-neutral agents at the zero-bit solution.

9.7.4 A Final Assessment of the Two Mutation Operators

The previous analysis has shown that the original mutation operator
is more efficient in combating genetic drift than the new bit-neutral
operator. Its effective mutation rates for individual bits are higher and
it is more disruptive. The difference in disruptiveness between the two
operators is highest when the bit level has reached zero. For the bit-
neutral mutation operator, the zero-bit level is truly an absorbing state.
It becomes inactive and the genetic algorithm converges prematurely.

16
The critical population size for the original mutation operator turns out to be
only 28 trading rules.
172 9 Selection, Genetic Drift, and Technical Trading

Even though the built-in upward bias of the original mutation op-
erator may be questionable at first sight, a deeper analysis reveals that
it is superior to the bit-neutral design. With the benefit of hindsight,
the new mutation operator did not effectively combat genetic drift; it
even became subject to genetic drift itself. Since the new operator was
designed to keep the bit fraction before and after mutation, on aver-
age, unaltered, the bit level in an agent’s rule set can successively be
changed in one direction several periods in a row by pure chance.17
Eventually, one of the two absorbing states will be reached. With low
initial bit probabilities, the absorbing state at the zero-bit level is the
more likely one.

9.8 Detection of Emergence of Technical Trading

As it turns out, the problem with the original SFI-ASM is not in model
design, but in the interpretation of certain simulation results. Given the
original mutation operator’s fixed point of one half, an isolated look at
the aggregate bit level to detect the existence of technical trading is not
sufficient. Without prior knowledge of the degree to which the increase
in learning speed affects the aggregate bit level through the biased
mutation operator, additional tests need to be done to establish the
existence of technical trading in the model.

9.8.1 Predictability in the Price Series

Instead of providing hard and unquestionable evidence, LeBaron et al.

often present arguments that do not establish the existence of techni-
cal trading in the model beyond any doubt. For instance, they showed
that the simulated price series contain useful technical information that
could be successfully exploited. LeBaron et al. regressed a simulated
price series on lagged prices and dividends and added an extra ex-
planatory variable, e.g.,

pt+1 + dt+1 = a + b(pt + dt ) + cIt,M A500 + ǫt+1 . (9.14)

This extra variable was either an indicator showing whether the
price is above or below a 5-period or a 500-period moving average, or
an indicator using the price dividend ratio. The results reported by
LeBaron et al. are reproduced in table 9.3.
17
Under no selection pressure and mutation only, the bit level in an agent’s rule set
would follow a random walk.
 9.8 Detection of Emergence of Technical Trading 173

Table 9.3. Predictability in the simulated price time series as documented in

[244, p. 1503].
Forecasting regressions

Description Fast learning Slow learning

MA(5) 0.009 !0.008

(0.013) (0.007)
MA(500) 0.074 !0.025
(0.014) (0.015)
rP/D'3/4 !0.443 0.050
(0.104) (0.093)

Note: Means over 25 runs. Numbers in parenthesis are standard errors estimated using the 25 runs.

LeBaron et al. find no extra predictability coming from the 5-period

MA for both the slow and fast learning case. While not being significant
for the slow learning case, the coefficients for the 500-period MA and
the price dividend ratio are, on the other hand, statistically significant
for the fast learning case. This points to some technical predictability
in the price series that agents could exploit by using their classifier
system.
While this regression demonstrates that there is indeed predictabil-
ity remaining in the simulated price series, it does not show whether
SFI agents are able to discover this predictability with their classifier
system. Neither the increase in the aggregate bit level nor the sign of
remaining predictability are convincing evidence whether agents really
use specific trading rules more often than general ones.

9.8.2 Trading Bits and Fitness Values

To really establish whether set trading bits reflect technical trading or

not, one could utilize the fitness information attached to each trad-
ing rule. If they indeed contain useful forecasting information, specific
trading rules should then be more accurate and possess higher fitness
values.
In order to test this hypothesis, 10 simulations were run over 500,000
periods in which 25 ordinary SFI agents were competing against 25 non-
classifier agents. As learning speeds, GA-invocation intervals of 100 and
1,000 were chosen. There are a variety of significance tests possible with
this setup. First, I tested the mean fitness values of all trading rules
over all simulation runs. The mean fitness values of SFI agents in one
simulation run were adjusted for bit costs, i.e., the values reported in
174 9 Selection, Genetic Drift, and Technical Trading

table 9.4 show the pure forecast accuracy of trading rules, independent
of their specificity.

Table 9.4. Comparison of bit cost adjusted mean fitness values for SFI- and
non-classifier agents. for. Fitness values were averaged over all rules of one
agent type within one simulation run.

Run GA-1,000 GA-100

Mean Fitness Mean Fitness Mean Fitness Mean Fitness
Adjusted for Bit over All Adjusted for Bit over All
Costs for SFI Non-Classifier Costs for SFI Non-Classifier
Agents Agents Agents Agents
1 92.672 93.449 93.767 93.771
2 92.912 92.442 94.612 93.975
3 92.737 92.732 94.094 94.506
4 91.123 91.159 92.824 92.432
5 92.774 93.121 93.531 93.613
6 91.940 92.617 94.365 93.654
7 90.666 91.263 93.330 93.340
8 93.280 92.404 94.226 94.001
9 92.672 92.370 92.888 92.738
10 91.977 91.923 93.740 93.411
mean 92.275 92.348 93.738 93.544
std.-dev. 0.837 0.731 0.605 0.609
t-stat. -0.430 1.792

In the slow learning case, non-classifier agents developed rule sets

that have a slightly higher mean fitness than those of their SFI counter-
parts. For the fast learning case, that relationship was reversed though.
Even though not yet significant at the 5%-level, SFI agents seem to gen-
erate trading rules that indeed produce more accurate forecasts. The
average specificity of their trading rules is about 4 trading bits per rule,
with a small increase in specificity for the fast learning case.
While an analysis of the mean fitness values in the economy showed
no evidence of advantageousness for technical trading in the slow learn-
ing regime and weak evidence for it in the fast learning case, the pic-
ture looks quite different if we only focus on the best trading rules. As
a maximum fitness value for each simulation run in table 9.5, the bit
cost-adjusted fitness value of the best trading rule in the whole econ-
omy per agent type was used. The use of only the best trading rules
makes sense since agents use the best of all activated trading rules for
 9.8 Detection of Emergence of Technical Trading 175

Table 9.5. Comparison of bit cost adjusted maximum fitness values for SFI-
and non-classifier agents. For each simulation run, the maximum fitness value
of all rules over all agents within the economy was used.

Run GA-1,000 GA-100

Max Fitness Max Fitness
Max Fitness of Max Fitness of
Adjusted for Bit Adjusted for Bit
Non-Classifier Non-Classifier
Costs for SFI Costs for SFI
Agents Agents
Agents Agents
1 97.390 97.113 96.474 96.258
2 96.881 96.462 97.181 96.977
3 96.827 96.467 96.759 96.594
4 96.812 95.012 95.544 95.112
5 96.875 96.470 96.325 96.031
6 96.646 96.346 96.933 96.401
7 96.085 94.939 96.248 95.418
8 96.729 96.403 97.099 96.528
9 96.998 96.811 95.998 95.272
10 96.393 95.896 96.309 95.899
mean 96.664 96.192 96.487 96.049
std.-dev. 0.459 0.713 0.513 0.620
t-stat. 5.16 6.11

their forecast production. When using only the best trading rule per
simulation run, one suddenly realizes that SFI agents are able to pro-
duce significantly better trading rules than non-classifier agents. Both
test statistics from a paired t-test are highly significant for both the
one- and the two-sided versions of the test. One also notices that in
the slow learning regime, the best trading rules produce slightly better
forecasts than in the fast learning case.18
The analysis of fitness information finally proves beyond a doubt
that there is indeed technical trading in the model. A superficial look at
the aggregate bit level or a reference to remaining predictability in the
price series does not have the same power of persuasion than the fitness
information stored for each trading rule. If the authors of the original
SFI-ASM would have performed such a fitness analysis to begin with,
no questions would have arisen about whether the continued existence
of trading bits reflects technical trading or not.
18
For the best trading rules of SFI agents, the t-statistics for an unpaired t-test
with unknown and unequal variances is 0.81. The t-value is only 0.48 for the best
trading rules of non-classifier agents, hence, in both cases, the forecasting quality
in the slow learning regime is better, but not significantly better.
176 9 Selection, Genetic Drift, and Technical Trading

9.8.3 Equilibrium Bit Frequencies and Bit Fixations

While the previous section proves that there is technical trading in

the model, it does not yet prove whether there is emergent technical
trading for faster learning speeds.19 The difference lies in the degree of
technical trading at varying learning speeds. Is there really an increase
in the amount of technical trading when agents learn faster, or does
the augmented bit level reflect a rise in the background level of bit
frequencies because of the biased mutation operator?
A definite answer to this question can be given when analyzing
the bit frequencies at different GA-invocation intervals. Figure 9.6 on
page 165 provides an approximation to the null model of equilibrium
bit frequencies with which the bit frequencies of SFI agents in normal
simulation runs should be compared. Remember that 9.6 was derived
under bit costs of 0.005, but with no selection pressure stemming from
forecast evaluation. Any differences in bit frequencies should then re-
flect the influence that the fitness evaluation of forecasts exerts on the
bit dynamics.
Figure 9.12 shows the distribution of bit frequencies after 250,000
periods for 30 slow learning SFI agents. First, one notices that there
about 10,000 fewer occurrences with a bit frequency of zero. The ab-
solute occurrences, however, cannot be compared since we analyzed
64,000 bit positions at once in figure 9.6 (1,000 agents x 64 condi-
tions), and only 48,000 bit positions in figure 9.12 (25 runs x 30 agents
x 64 conditions). If we compare the relative commonness of zero-bit
fixations, we have a drop from 0.735 down to 0.660. This reduction
certainly implies an increase in bit frequencies greater than zero. Of
main interest, however, is the change in complete or near bit fixations
at the 100% level. While the null-model has only 0.02% possible bit
fixations, 3.35% of all possible bit positions are completely filled with
non-# bits under real simulation conditions. The increase compared to
the null model is another irrefutable test that proves the existence of
technical trading in the model.
The distribution of bit frequencies of SFI agents in the fast learn-
ing regime is shown in figure 9.13. To test whether there is indeed an
increase in the degree of technical trading, we need to check whether
there are more bit fixations at the 100% level compared to the slow

19
The increased fitness differential between the best rules of SFI and non-classifier
agents in the fast learning case suggests that there is a potential for higher degrees
of technical trading. But again, the fitness values alone do not allow a statement
about levels of technical trading at both learning speeds.
 9.9 An Evolutionary Perspective on Technical Trading 177
32500
30000
27500
25000
number of instances

22500
20000
17500
15000
12500
10000
7500
5000
2500
0
0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1
bit frequency

Fig. 9.12. Number of bit positions that have a certain bit frequency in the
rule sets of slow learning SFI agents. Data were obtained at period 250,000
from 25 simulation runs with 30 SFI agents, each endowed with 100 trading
rules (Π = 0.9, π = 0.03, bit costs = 0.005, GA-interval = 1,000).

learning case.20 This is definitely the case, with 4.46% of all possible
bit positions fixed. When checking the significance of the increase from
3.35% to 4.46% with an unpaired t-test with unknown and unequal
variances, the t-value turns out to be 8.9. Since this is significant at
even the highest levels of confidence, we can finally establish the emer-
gence of technical trading for faster learning speeds beyond a doubt.

9.9 An Evolutionary Perspective on Technical Trading

In section 9.8.2, we saw that the average specificity of successful trading

rules is about 4 bits per rule, with a small increase in specificity for
the fast learning case. Remember that we determined the necessary
20
Even though we know from figure 6.1 on page 107 that the final equilibrium
distributions of bit frequencies have not yet been reached, we also know that we
can already test for differences in bit fixations. At period 250,000, the aggregate
bit levels for the two GA-invocation intervals have already reached their relative
position. The enormous requirements in terms of computational time led me to
shorten the time horizon. Figure 6.1 suggests that in the final bit equilibrium,
the difference between bit fixations will be even more pronounced than at period
250,000.
178 9 Selection, Genetic Drift, and Technical Trading
35000
32500
30000
27500
number of instances

25000
22500
20000
17500
15000
12500
10000
7500
5000
2500
0
0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1
bit frequency

Fig. 9.13. Number of bit positions that have a certain bit frequency in the
rule sets of fast learning SFI agents. Data were obtained at period 250,000
from 25 simulation runs with 30 SFI agents, each endowed with 100 trading
rules (Π = 0.9, π = 0.03, bit costs = 0.005, GA-interval = 100).

fitness increase per single trading bit to be around 0.5 in order for
selection to offset the effect of genetic drift (section 9.6.3 on page 161).
With an average specificity of four, the fitness increases of only 0.2 for
mean fitness and of about 0.5 for maximum fitness values in the fast
learning case suggest that selection cannot guarantee that favorable
trading bits will spread through the population. The observed effect of
subpopulation differentiation, i.e., different agents discovering different
trading rules, is most likely a result of insufficient selection pressure
in favor of successful trading bits. With fitness gradients too small to
counter genetic drift, it is hard to say which trading bits will be selected
since bit fixation remains a noise-prone process.
While we have developed this argument within a model that is char-
acterized by successful technical trading rules, the existence of technical
trading strategies in real financial markets remains a puzzle for many
adherents of the EMH. However, they can now use the same argument
to explain that existence even in the presence of efficient markets.
Those who voiced the earliest evolutionary arguments in favor of
emergent rationality or efficient markets seemed to have come from the
selectionist camp. They implicitly assumed that selection is the only
evolutionary force at work and, “if things have had time to hammer
 9.9 An Evolutionary Perspective on Technical Trading 179

logic into men” [382, p. 80], a rational and/or efficient outcome will be
inevitable. An evolutionary perspective that extends its scope beyond
selection is not that certain anymore.
In the light of the above argument, the sheer amount of differ-
ent technical trading strategies in real financial markets suggests two
things. First, their possible profits are at best relatively small. Only
highly profitable and easily detectable trading strategies could, in prin-
ciple, spread through the whole population, which would then lead to
their own demise. Second, to the extent that these trading strategies
are unprofitable (or less profitable than the market portfolio), the losses
they produce are equally too small to quickly eradicate them.21 With
only small selective forces in favor of or against most trading strategies,
“financial evolution” cannot guarantee that only the fittest strategy will
survive.
Under small selection pressures in either direction, random evolu-
tionary forces other than selection might prevail and maintain a suffi-
ciently diverse population of trading strategies over extended periods
of time. Keeping in mind that a moderate increase in checked trading
conditions in the SFI-ASM drastically prolonged the adjustment pe-
riod towards equilibrium, the “evolutionary time scale” in real financial
markets may be too long to observe evolutionary change in real time.22
The existence of technical trading strategies does not necessarily prove
their profitability. It does not disprove the validity of the EMH either.
Technical trading rules may simply be a sign of insufficient selection
pressures against them.

21
Positive profits below the returns of the market portfolio do not strike one as a
particulary strong selective force against a particularly trading strategy.
22
A similar concern was voiced in [242].
10
Summary and Future Research

“... the outcome of any serious research can only be to make

two questions grow where only one grew before.”
Thorstein Veblen

It was argued in the beginning of this book that the current surge in
agent-based simulation models in economics cannot be explained by the
sudden increase in computational resources alone. Thus, the first part
discussed agent-based simulations in the light of the limitations of cur-
rent research methodologies and practices in the economics field. It was
highlighted that agent-based simulation approaches are highly suited
to address heterogeneous agents problems. Giving up the assumption
of agent homogeneity necessarily leads to an abandoning of the rational
expectations paradigm. In addressing what degree of rationality should
be chosen instead, chapter 3 argued in favor of the principle of minimal
rationality [70]. When considering cognitive resources as scarce, agents
should be modeled as economizing on reasoning processes and employ-
ing only as much reasoning power as necessary to perform the task they
are facing. It was shown at the end of chapter 3 that the trigger for
the rational expectations revolution, the famous Phillips-curve debate,
closely followed the principle of minimal rationality.
After discussing some stylized facts of financial markets in the light
of the EMH, chapter 5 introduced several agent-based simulations.
Through their attempt to replicate certain stylized facts, additional
insight can be gained as to which assumptions are essential for the
emergence of these stylized facts. The selection and sequencing of pre-
sentation of agent-based simulations was guided by the principle of
minimal rationality, discussing first the zero-intelligence approach by
182 10 Summary and Future Research

Gode and Sunder [150], followed by models with random communi-

cation structures which are, too, populated by agents with random
demands. Models of chartist-fundamentalist interactions and many-
strategy models with learning were the last to be discussed in chapter
5.
Within the hierarchy of agent rationality, the agents of the SFI-
ASM, the main focus of this book, ranked at the top. Only because
they were modeled as being heterogeneous in analyzing a common in-
formation set, the assumption of rational expectations had to be re-
placed by that of boundedly rational agents that use adaptive learning
procedures. Since these agents derive their demands through optimiza-
tion, the deviations from a typical neoclassical framework were only
minimal.
The main part of this book analyzed a Java-replication of the orig-
inal SFI-ASM. The replication of existing simulation models is, as Ax-
elrod points out, an important, but often neglected step in the field
of agent-based simulations [13]. Recreating a simulation model from
scratch and confirming the basic results is much more powerful than
simply utilizing an existing source code. Unlike many other models,
the SFI-ASM, as one of the first and most famous artificial stock mar-
ket simulations, has been extensively tested and it exists on several
different programming platforms.
During the normal process of discovering and eliminating my own
implementation errors, I stumbled upon an interesting model behavior.
The unexpected and previously undocumented increase in the aggre-
gate bit-level for higher mutation rates prompted me to analyze the
GA in more detail in chapter 6. It was shown that the mutation oper-
ator had a fixed point of one half, i.e., it was upwardly-biased under
normal model parameterization. This upward bias caused me to ques-
tion whether the increase in the aggregate bit level for faster learning
speeds is really a reflection of emergent technical trading. Since, at
faster learning speeds, the mutation operator is called more often per
time interval and injects new trading bits, the increase in the aggregate
bit level could also be a design artifact.
An alternative mutation operator was suggested in chapter 7. This
mutation operator has the property of being bit-neutral, i.e., the num-
ber of trading bits before and after mutation was left, on average, un-
altered. The updated model behavior then supported the Marimon-
Sargent Hypothesis with respect to bit usage, namely that adaptive
boundedly rational agents in an artificial stock market would discover
the homogeneous rational expectations equilibrium. Independent of
 10 Summary and Future Research 183

learning speed, all agents discovered the zero-bit solution which tem-
porarily suggested two things. First, the classifier system does not pro-
vide any informational advantage that agents could exploit. Second, the
emergence of technical trading in the original SFI-ASM is an artifact
of design assumption.
The tentative result of a useless classifier system stood in contra-
diction to two studies by Joshi, Parker, and Bedau [206, 207]. Based
on the terminal wealth levels acquired by agents in the SFI-ASM, they
performed two game-theoretic analyses. In the first study, they asked
whether it would be optimal for agents to include technical trading
rules or not. They concluded that technical trading constitutes a typi-
cal prisoner’s dilemma. They argued that it is rational for each agent to
use technical trading, yet at the same time, they found the aggregate
market outcome to be less efficient than when agents had no access
to technical trading rules. In the second study, they endogenized the
learning speed of agents and found the optimal learning rate to be at
a GA-invocation interval of 100. This clearly falls into the less efficient
complex regime, and Joshi, Parker, and Bedau inferred that financial
markets can operate at sub-optimal equilibria [207]. While the game-
theoretic analyses themselves are correct, chapter 8 deduced that the
SFI-ASM is highly unsuitable for addressing wealth issues. Because of
taxation in the model, a programming trick to avoid a long-run ex-
plosion of wealth, a wedge between long-term wealth and maximized
short-term wealth was introduced. It was analytically shown that the
only source of wealth accumulation in the SFI-ASM is the aggregate
risk premium that agents collect. Wealth contributions of the kind “buy
low and sell high” are impossible. A series of simulations then demon-
strated that wealth differences between different trader types generally
arise for various reasons, none of which has any economic meaning.
In the final chapter, some simulation evidence was presented that
showed the zero-bit solution to be too robust, i.e., it even emerged un-
der parameterizations under which it should not have emerged. As a
reason for this behavior, the effect of genetic drift was identified. Sub-
sequently, the two mutation operators were analyzed in light of genetic
drift. By transferring the Wright-Fisher model from population genetics
to the field of genetic algorithms, conditions were derived under which
selection dominates genetic drift. In a second step, a so-called neutral
model from population genetics was used to analyze under which con-
ditions mutation dominates genetic drift. While the original mutation
operator proved to dominate genetic drift, the same could not be said
for the suggested bit-neutral operator. In addition to being unable to
184 10 Summary and Future Research

effectively combat genetic drift, it even became subject to genetic drift

itself.
Even though the original mutation operator looked suspicious at
first, it turned out that there was no problem in model design as previ-
ously thought. With the benefit of hindsight, however, it became clear
to me that there was a problem in the interpretation of the original
simulation results. The aggregate bit level, which was used to bolster
the argument of emergent technical trading, is highly unsuitable for
advancing such claims, especially in the light of an upwardly-biased
mutation operator.1 The proper way to test for technical trading is to
utilize the attached fitness information for each trading rule. To test for
emergent technical trading, i.e., whether the level of technical trading
is higher at faster learning speeds, one has to analyze whether the num-
ber of bit fixations at the 100%-level has significantly increased. Both
tests unequivocally showed in the end that there is indeed emergent
technical trading in the SFI-ASM.
The final disapproval of the bit-neutral mutation operator made it
necessary to reconsider some of the claims and results that were pre-
sented in the course of this book. Even though the initial reason for
reconsidering two game-theoretic analyses by Joshi, Parker, and Bedau
turned out to be ungrounded, the finding that accumulated wealth lev-
els should not be given any economic meaning in the SFI-ASM remains
unaffected.
Therefore, future research again needs to tackle the problem of an
optimal learning speed of agents. Since wealth levels are not appropri-
ate in addressing this issue, one either has to find another venue for
doing this, or one has to alter the economic structure such that accu-
mulated wealth becomes economically meaningful. A related issue is
the question whether there is a maximum learning speed beyond which
the model structure would break down. A GA-invocation interval of
one has been shown to be such a case.
Figure 7.2 on page 117 showed that bit-neutral agents took longer to
arrive at the zero-bit solution the closer the stochastic dividend process
came to a random walk. This increase in time suggests that there are
1
It was the aggregate bit level argument that caused me to suggest the bit-neutral
mutation operator. My own statements in favor of the bit-neutral mutation op-
erator were, too, based on the aggregate bit level. The approval of my arguments
from the scientific economics community at various conferences, the acceptance
of an article to be published in a leading economics journal, and the absence of
a well-founded rebuttal by the original authors show me that there was, indeed,
substantial room for misinterpretation, and that the dynamics of GA-behavior
were not fully understood by the groups involved.
 10 Summary and Future Research 185

more possibilities for emerging patterns to occur at near random walk

behavior. It is, therefore, my hypothesis that a similar effect should
be prevalent with the original mutation operator, i.e., the closer the
dividend is to a random walk, the higher the degree of technical trading
in the model. Having now pointed to a proper way of testing the degree
of technical trading in the model, this hypothesis could be easily tested
in further research.
An important task that remains to be done is the calibration of the
model to real empirical data. This includes refining the concept of a
trading period and possibly allowing for intra-period trading. Replacing
the CARA-assumption with constant relative risk aversion would allow
for wealth to impact the optimal demands.2 Instead of using myopic
agents, agents could be modeled as having fully intertemporal prefer-
ences. While classifier systems lend themselves to an easy interpreta-
tion and analysis of the learning dynamics, LeBaron points out that
they remain a controversial modeling tool in economics [242]. Other
approaches such as using neural networks for forecast production seem
to be becoming more popular [240, 241]. It is however, likely, that in
addressing some or all of these issues, the economic structure of the
model has to be altered in a way that one should no longer speak of
“the SFI-ASM”.
Once the various approaches of designing agent-based artificial mar-
kets result in an adoption of commonly accepted principles or building
blocks, it is likely that agent-based financial markets will be able to
address practical problems. Farmer predicts that in several years, they
will be used for investment applications [120] while LeBaron conceives
of a role for agent-based models in evaluating the stability and efficiency
of different trading mechanisms [239]. Until then, however, much work
remains to be done.

2
CRRA preferences are used, for instance, in [247, 241].
11
Appendix

11.1 Timing in the Stock Market

The general sequence of activities in the SFI-ASM is summarized and

represented in figure 11.1. At the beginning of period t, a new dividend
dt is announced. For each agent i = 1, . . . , N , the GA is invoked with
probability 1/K to change an agent’s rule set ℑi . Afterwards, agents de-
termine their active rule set ωi ⊆ ℑi by comparing the condition parts
of each trading rule j = 1, . . . , 100 with the binary market descrip-
tor. From that active rule set, agents select the rule with the highest
forecast accuracy for their forecast production. The price formation
process is then initiated by the specialist who announces a trial price
ptrial equal to last period’s price. Based on this trial price, agents form
their expectation about next period’s price and dividend, determine
their optimal demand for the risky stock, and then submit their offers
and bids to the specialist. If the bids and offers cannot be matched, the
specialist determines a trial price and the whole process starts all over.
This iterative process ends when the offers are balanced by the sub-
mitted bids or after 10 unsuccessful trial rounds. In the latter case, one
side of the market will be proportionally rationed. The last trial price
is announced to be the stock price for period t and all trades between
agents are executed at that price. Since at this time, the dividend and
the stock price in period t are known to the agents, they are now able
to update the forecast performance of all rules which were activated
in the last period and actually produced a forecast about this period’s
price and dividend. After completion, a new period starts when the
next stock dividend is revealed.
Artificial Stock Market
Fig. 11.1. Timing sequence of major activities in the Santa Fe Institute

188
∀ i: with prob 1 K : invoke GA on ℑi ∀ i: with prob 1 K : invoke GA on ℑi

{ } { }

11 Appendix
∀ i: choose j ∈ ω i ,t = k ∈ ℑi : cond. part k = market descr. ∀ i: choose j ∈ ω i ,t +1 = k ∈ ℑi : cond. part k = market descr.
t t+1

specialist: ptrial = pt −1 specialist: ptrial = pt

[ ]
∀ i: Ei ,t pt +1 + d t +1 = ai , j ( ptrial + d t ) + bi , j ∀ i: Ei ,t +1[ pt +2 + dt +2 ] = ai , j ( pt +1 + dt +1 ) + bi , j

spec.: ptrial = f (∑ N
i =1
xi ,t − N ) spec.: ptrial = f (∑ N
i =1
xi ,t +1 − N )
Ei ,t [ pt +1 + dt +1 ] − ptrial (1 + r ) Ei ,t +1[ pt +2 + d t +2 ] − ptrial (1 + r )
∀ i: xi ,t = ∀ i: xi ,t +1 =
λσ t2,i , j , p +d λσ t2+1,i , j , p +d

N N
spec.: ∑ x i ,t = N? NO spec.: ∑ x i ,t +1 = N? NO
i =1 i =1

YES YES

pt = ptrial ∀ i: ∀ j ∈ ω i ,t −1: update pt +1 = ptrial ∀ i: ∀ j ∈ ω i ,t : update t

dt .
d t +1 .

. fitness . fitness

period t period t+1

 11.2 Fundamental and Technical Trading Bits 189

11.2 Fundamental and Technical Trading Bits

This section documents the differences in condition bits used by several

SFI-ASM versions. Furthermore, it explains why certain changes are
made to the Java version of the SFI-ASM.

Table 11.1. Condition bits documented in [11, p. 27] and [244, p. 1494].

Bit Condition

1 Price * interest/dividend '1/4

2 Price * interest/dividend '1/2
3 Price * interest/dividend '3/4
4 Price * interest/dividend '7/8
5 Price * interest/dividend '1
6 Price * interest/dividend '9/8
7 Price '5-period MA
8 Price '10-period MA
9 Price '100-period MA
10 Price '500-period MA
11 On: 1
12 O!: 0

Arthur et al. [11] and LeBaron et al. [244] report only 12 condi-
tions for the SFI-ASM. Bits 1 to 6 are the fundamental bits. Note that
even though these are 6 individual bits, they can only code one piece
of fundamental information since they logically belong together. The
rule consistency check introduced in section 7.4.3 compresses this bit
sequence to a maximum of 2 bits, and a neural network trading rule
would only use one real valued number. Bits 7 to 10 are the technical
trading bits. LeBaron et al. remark that removing 3 out of the 4 tech-
nical bits can have a big impact, but removing only one will not change
things. They do not, however, mention what would happen if the list
of trading bits were substantially extended.
The last two bits convey no useful market information, but Arthur
et al. explain that they were intended as experimental controls to check
to which degree agents act upon useless information. They also assert
that these two bits would allow them to detect emergence of technical
trading “if bits 7–10 become set significantly more often, statistically,
than the control bits” [11, p. 27]. Emergence of technical trading in
the SFI-ASM has surely been declared without ever having done such
a test. Since the two control bits are always set, i.e., they are never
#, the “benchmark” of control bits is always 100 %. It is, therefore,
190 11 Appendix

impossible for any trading bit to be set significantly more often. Since
these bits are not mentioned anywhere in the discussions of simulation
results, they are not implemented in the Java version of the SFI-ASM.

Table 11.2. First part of the condition bits in the SFI-ASM Objective-C
version 7.1.2 used by Joshi et al. (1998 and 2002).

Bit Condition
1 dummy bit – always on
2 dummy bit – always off
3 dummy bit – random on or off
4 dividend went up this period
5 dividend went up one period ago
6 dividend went up two periods ago
7 dividend went up three periods ago
8 dividend went up four periods ago
9 5-period moving average of dividend went up
10 20-period moving average of dividend went up
11 100-period moving average of dividend went up
12 500-period moving average of dividend went up
13 dividend > 5 period moving average
14 dividend > 20 period moving average
15 dividend > 100 period moving average
16 dividend > 500 period moving average
17 dividend: 5-period moving average > 20-period moving average
18 dividend: 5-period moving average > 100-period moving average
19 dividend: 5-period moving average > 500-period moving average
20 dividend: 20-period moving average > 100-period moving average
21 dividend: 20-period moving average > 500-period moving average
22 dividend: 100-period moving average > 500-period moving average
23 dividend/mean-dividend > 1/4
24 dividend/mean-dividend > 1/2
25 dividend/mean-dividend > 3/4
26 dividend/mean-dividend > 7/8
27 dividend/mean-dividend > 1
28 dividend/mean-dividend > 9/8
29 dividend/mean-dividend > 5/4
30 dividend/mean-dividend > 3/2
31 dividend/mean-dividend > 2
32 dividend/mean-dividend > 4

The Objective-C version 7.1.2 used by Joshi et al. [206], which

served as the blueprint for the current Java implementation of the
 11.2 Fundamental and Technical Trading Bits 191

FI-ASM, had a total of 61 conditions. These conditions are docu-

mented in tables 11.2 and 11.3. Besides checking many more condi-
tion bits than in the model version by LeBaron et al., the ratios for
the price*interest/dividend (prf /d) bit sequence were also changed by
Joshi et al.

Table 11.3. Second part of the condition bits in the SFI-ASM Objective-C
version 7.1.2 used by Joshi et al. (1998 and 2002).

Bit Condition
33 price*interest/dividend > 1/4
34 price*interest/dividend > 1/2
35 price*interest/dividend > 3/4
36 price*interest/dividend > 7/8
37 price*interest/dividend > 1
38 price*interest/dividend > 9/8
39 price*interest/dividend > 5/4
40 price*interest/dividend > 3/2
41 price*interest/dividend > 2
42 price*interest/dividend > 4
43 price went up this period
44 price went up one period ago
45 price went up two periods ago
46 price went up three periods ago
47 price went up four periods ago
48 5-period moving average of price went up
49 20-period moving average of price went up
50 100-period moving average of price went up
51 500-period moving average of price went up
52 price > 5-period moving average
53 price > 20-period moving average
54 price > 100-period moving average
55 price > 500-period moving average
56 price: 5-period moving average > 20-period moving average
57 price: 5-period moving average > 100-period moving average
58 price: 5-period moving average > 500-period moving average
59 price: 20-period moving average > 100-period moving average
60 price: 20-period moving average > 500-period moving average
61 price: 100-period moving average > 500-period moving average

The cutoff points for the prf /d-ratios had been chosen by LeBaron
et al. to have good coverage of the range of relevant ratios. Note that the
192 11 Appendix

cutoff points around one are non-symmetrical, i.e., it is very unlikely

for the stock to be highly overvalued. LeBaron et al. reported that the
extreme events at 1/4 and 9/8 were visited with probabilities of less
than 0.01. Hence, the bits 39 − 42 used in Joshi et al. can be safely
neglected. Since the dividend process is symmetrical around its mean,
¯
the d/d-ratios are also symmetrical. Applying the same principle of
good coverage to the dividend/mean-dividend ratios (d/d), ¯ however,
¯
led me to propose cutoff points for the d/d-ratios that are much closer
than those used by Joshi et al. The condition bits for the Java version
of the SFI-ASM are shown in table 11.4.
The reprogrammed SFI-ASM differs in the number of checked condi-
tions. All conditions labeled as fundamental are coded in a long integer
which consists of 8 bytes (64 bits). Since a trading bit can take on three
allele values (0, 1, and #), one encoded trading condition requires 2
bytes. Therefore, a long integer can hold 32 conditions. One long inte-
ger is used to hold all 32 fundamental conditions; a second holds the
32 technical conditions.
 11.2 Fundamental and Technical Trading Bits 193

Table 11.4. Fundamental and technical condition bits in the current Java
version of the SFI-ASM.
Bit Fundamental conditions Technical conditions
0 dividend / dividend-mean > 0.6 price / price-mean > 0.25
1 dividend / dividend-mean > 0.8 price / price-mean > 0.5
2 dividend / dividend-mean > 0.9 price / price-mean > 0.75
3 dividend / dividend-mean > 1.0 price / price-mean > 0.875
4 dividend / dividend-mean > 1.1 price / price-mean > 1.0
5 dividend / dividend-mean > 1.12 price / price-mean > 1.125
6 dividend / dividend-mean > 1.4 price / price-mean > 1.25
7 price * interest / dividend > 0.25 price went up this period
8 price * interest / dividend > 0.5 price went up one period ago
9 price * interest / dividend > 0.75 price went up two periods ago
10 price * interest / dividend > 0.875 price went up three periods ago
11 price * interest / dividend > 0.95 price went up four periods ago
12 price * interest / dividend > 1.0 5-period price-MA went up
13 price * interest / dividend > 1.125 10-period price-MA went up
14 dividend went up this period 20-period price-MA went up
15 dividend went up one period ago 100-period price-MA went up
16 dividend went up two periods ago 500-period price-MA went up
17 dividend went up three period ago price > 5-period price-MA
18 5-period dividend MA went up price > 10-period price-MA
19 10-period dividend MA went up price > 20-period price-MA
20 100-period dividend MA went up price > 100-period price-MA
21 500-period dividend MA went up price > 500-period price-MA
22 dividend > 5-period div.-MA price: 5-period MA > 10-period
23 dividend > 10-period div.-MA price: 5-period MA > 20-period
24 dividend > 100-period div.-MA price: 5-period MA > 100-period
25 dividend > 500-period div.-MA price: 5-period MA > 500-period
26 div.: 5-period MA > 10-period MA price: 10-period MA > 20-period
27 div.: 5-period MA > 100-period MA price: 10-period MA > 100-period
28 div.: 5-period MA > 500-period MA price: 10-period MA > 500-period
29 div.: 10-period MA > 100-period MA price: 20-period MA > 100-period
30 div.: 10-period MA > 100-period MA price: 20-period MA > 500-period
31 div.: 100-period MA > 500-period MA price: 100-period MA > 500-period
References

[1] C. Ahn, E.-B. Lee, and E.-H. Suh. Noise trading in the Korean
foreign exchange market: Some questionnaire evidence. Bank of
Korea Economic Papers, 5(2):133–155, 2002.
[2] V. Akgiray, G. G. Booth, and O. Loistl. Stable laws are in-
appropriate for describing German stock returns. Allgemeines
Statistisches Archiv, 73:115–121, 1989.
[3] A. A. Alchian. Uncertainty, evolution, and economic theory. Jour-
nal of Political Economy, 58(3):211–222, 1950.
[4] M. Allais. Le comportement de l’homme rationnel devant le
risque: Critique des postulats et axiomes de l’ecole americaine.
Econometrica, 21(4):503–546, 1953.
[5] F. Allen and R. Karjalainen. Using genetic algorithms to
find technical trading rules. Journal of Financial Economics,
51(2):245–271, 1999.
[6] F. Allen, S. Morris, and H. S. Shin. Beauty contests, bubbles
and iterated expectations in asset markets. Cowles Foundation
Discussion Paper 1406, 1406, March 2003.
[7] H. L. Allen and M. P. Taylor. Charts, noise, and fundamen-
tals in the London foreign exchange market. Economic Journal,
100(400):49–59, 1990.
[8] H. M. Amman, D. A. Kendrick, and J. Rust, editors. Handbook
of Computational Economics, volume 1. Elsevier, Amsterdam,
New York, 1996.
[9] J. Arifovic. Genetic algorithm learning and the cobweb model.
Journal of Economic Dynamics and Control, 18(1):3–28, 1994.
[10] B. W. Arthur, S. Durlauf, and D. Lane, editors. The Economy
as an Evolving Complex System II. SFI Studies in the Sciences
of Complexity, vol. 27. Addison-Wesley, Reading, MA, 1997.
196 References

[11] W. B. Arthur, J. H. Holland, B. LeBaron, R. Palmer, and

P. Tayler. Asset pricing under endogenous expectations in an ar-
tificial stock market. In W. B. Arthur, S. N. Durlauf, and D. A.
Lane, editors, The Economy as an Evolving Complex System II,
volume 27 of SFI Studies in the Sciences of Complexity, pages
15–44. Santa Fe Institute, Addison-Wesley, 1997.
[12] R. J. Aumann. Agreeing to disagree. Annals of Statistics,
4(6):1236–1239, 1976.
[13] R. Axelrod. Advancing the art of simulation in the social sciences:
Obtaining, analyzing, and sharing results of computer models.
Complexity, 3(2):16–22, 1997.
[14] C. Azariadis and R. Guesnerie. Sunspots and cycles. Review of
Economic Studies, 53(5):725–737, 1986.
[15] L. Bachelier. Théorie de la spéculation. Annales Scientifiques de
lÉcole Normale Supérieure, 17:21–86, 1900.
[16] T. Bäck. Evolutionary Algorithms in Theory and Practice. Oxford
University Press, New York, 1996.
[17] T. Bäck. Mutation: Binary strings. In T. Bäck, D. B. Fogel, and
Z. Michalewicz, editors, Handbook of Evolutionary Computation,
chapter C3.2.1. IOP Publishing Ltd and Oxford University Press,
Oxford, UK, 1997.
[18] R. T. Baillie, T. Bollerslev, and H. O. Mikkelsen. Fractionally
integrated generalized autoregressive conditional heteroskedastic-
ity. Journal of Econometrics, 74(1):3–30, 1996.
[19] P. Bak, M. Paczuski, and M. Shubik. Price variations in a stock
market with many agents. Physica A, 246:430–453, 1997.
[20] P. Bak, C. Tang, and K. Wiesenfeld. Self-organized criticality.
Physical Review A, 38:364–373, 1988.
[21] R. Ball. Anomalies in relationships between securities’ yields and
yield-surrogates. Journal of Financial Economics, 6(2/3):103–
126, 1978.
[22] S. C. Bankes. Agent-based modeling: A revolution? Proceedings
of the National Academy of Science of the USA, 99:7199–7200,
2002.
[23] Y. Bar-Yam. Dynamics of Complex Systems. Addison-Wesley,
Reading, MA, 1997.
[24] N. Barberis, A. Shleifer, and R. Vishny. A model of investor
sentiment. Journal of Financial Economics, 49(3):307–343, 1998.
[25] R. J. Barro. Rational expectations and macroeconomics in 1984.
American Economic Review, Papers and Proceedings, 74(2):179–
182, May 1984.
 References 197

[26] T. A. Bass. The Predictors. Henry Holt, New York, 1999.

[27] D. Beasley, D. R. Bull, and R. R. Martin. An overview of Ge-
netic Algorithms: Part 1, Fundamentals. University Computing,
15(2):58–69, 1993.
[28] G. S. Becker. Irrational behavior and economic theory. Journal
of Political Economy, 70(1):1–13, February 1962.
[29] J. Benhabib and R. E. A. Farmer. Indeterminacy and sunspots
in macroeconomics. In J. B. Taylor and M. Woodford, editors,
Handbook of Macroeconomics, volume 1a, chapter 6, pages 387–
448. North-Holland, Amsterdam, New York, 1999.
[30] A. K. Bera and C. M. Jarque. Efficient tests for normality, ho-
moscedasticity and serial independence of regression residuals:
Monte Carlo evidence. Economics Letters, 7(4):313–318, 1981.
[31] H.-G. Beyer. On the dynamics of EAs without selection. In
W. Banzhaf and C. Reeves, editors, Foundations of Genetic Al-
gorithms, volume 5, pages 5–26. Morgan Kaufmann, 1999.
[32] H. S. Bhamra. Imitation in financial markets. International Jour-
nal of Theoretical and Applied Finance, 3(3):473–478, 2000.
[33] B. R. Biais and P. Bossaerts. Asset prices and trading volume
in a beauty contest. Review of Economic Studies, 65(2):307–340,
1998.
[34] B. R. Biais and R. Shadur. Darwinian selection does not eliminate
irrational traders. European Economic Review, 44(3):469–490,
2000.
[35] S. Bikhchandani and S. Sharma. Herd behavior in financial mar-
kets: A review. IMF Working Paper, 00/48, 2000.
[36] F. Black. Noise. Journal of Finance, 41(3):529–543, 1986.
[37] F. Black and M. Scholes. The pricing of options and corporate
liabilities. Journal of Political Economy, 81(3):637–654, 1973.
[38] T. Blickle. Tournament selection. In T. Bäck, D. B. Fogel, and
Z. Michalewicz, editors, Handbook of Evolutionary Computation,
chapter C 2.3. IOP Publishing Ltd and Oxford University Press,
Oxford, UK, 1997.
[39] T. Blickle and L. Thiele. A comparison of selection schemes used
in genetic algorithms. TIK Report 11, Version 2, Computer Engi-
neering and Communication Networks Lab TIK, Zurich, Decem-
ber 1995.
[40] L. E. Blume and D. Easley. Learning to be rational. Journal of
Economic Theory, 26(2):340–351, 1982.
[41] L. E. Blume and D. Easley. Evolution and market behavior.
Journal of Economic Theory, 58(1):9–40, 1992.
198 References

[42] L. E. Blume and D. Easley. Evolution and rationality in compete-

tive markets. In A. Kirman and M. Salmon, editors, Learning and
Rationality in Economics, chapter 11, pages 324–342. Blackwell,
Oxford, Cambridge, 1995.
[43] L. E. Blume and D. Easley. Optimality and natural selection in
markets. Journal of Economic Theory, 107(1):95–135, 2002.
[44] N. Boitout and L. Ureche-Rangau. Towards a multifractal
paradigm of stochastic volatility? International Journal of The-
oretical and Applied Finance, 7(7):823–851, 2004.
[45] T. Bollerslev. Generalized autoregressive conditional hetero-
scedasticity. Journal of Econometrics, 31(3):307–327, 1986.
[46] T. Bollerslev, R. Y. Chou, and K. F. Kroner. ARCH modeling in
finance: A review of the theory and empirical evidence. Journal
of Econometrics, 52(1):5–59, 1992.
[47] L. B. Booker, D. B. Fogel, D. Whitley, and P. J. Angeline. Recom-
bination. In T. Bäck, D. B. Fogel, and Z. Michalewicz, editors,
Handbook of Evolutionary Computation, chapter C 3.3. IOP Pub-
lishing Ltd and Oxford University Press, Oxford, UK, 1997.
[48] T. Börgers. On the relevance of learning and evolution to eco-
nomic theory. Economic Journal, 106(438):1374–1385, 1996.
[49] K. E. Boulding. Evolutionary Economics. Sage, Beverly Hills and
London, 1981.
[50] G. H. Bower and E. R. Hilgard. Theories of Learning. Prentice
Hall, Englewood Cliffs, N.J., 1981.
[51] T. Brenner. Can evolutionary algorithms describe learning pro-
cesses? Journal of Evolutionary Economics, 8(3):271–283, 1998.
[52] T. Brenner. Modelling Learning in Economics. Edward Elgar,
Cheltenham, Northampton, 1999.
[53] T. Brenner. Agent learning representation: Advice in modelling
economic learning. Papers on Economics & Evolution 0416, Max
Planck Institute for Research into Economic Systems, Evolution-
ary Economics Group, 2004.
[54] W. A. Brock and C. H. Hommes. A rational route to randomness.
Econometrica, 65(5):1059–1095, September 1997.
[55] W. A. Brock and C. H. Hommes. Heterogeneous beliefs and routes
to chaos in a simple asset pricing model. Journal of Economic
Dynamics and Control, 22(8-9):1235–1274, 1998.
[56] W. A. Brock, J. Lakonishok, and B. LeBaron. Simple techni-
cal trading rules and the stochastic properties of stock returns.
Journal of Finance, 47(5):1731–1764, 1992.
 References 199

[57] W. A. Brock and B. LeBaron. A dynamic structural model for

stock return volatility and trading volume. Review of Economics
and Statistics, 78(1):94–110, 1996.
[58] W. A. Brock, J. A. Scheinkmann, W. D. Dechert, and B. LeBaron.
A test for independence based on the correlation dimension.
Econometric Reviews, 15(3):197–235, 1996.
[59] H. S. Buscher. Angewandte Zeitreihenanalyse. In M. Schröder,
editor, Finanzmarkt-Ökonometrie: Basistechniken, Fortgeschrit-
tene Verfahren, Prognosemodelle, chapter III, pages 131–212.
Schäffer-Poeschel, Stuttgart, 2002.
[60] C. D. Bustamante. Population genetics of molecular evolution.
In R. Nielsen, editor, Statistical Methods in Molecular Evolution,
Statistics for Biology and Health, chapter 4. Springer, New York,
2005.
[61] R. J. Caballero. A fallacy of composition. American Economic
Review, 82(5):1279–1292, 1992.
[62] P. Cagan. The dynamics of hyperinflation. In M. Friedman,
editor, Studies in the Quantity Theory of Money, pages 25–117.
University of Chicago Press, Chicago, London, 1956.
[63] L. Calvet and A. Fisher. Multifractality in asset returns: Theory
and evidence. Review of Economics and Statistics, 84(3):381–406,
2002.
[64] C. Camerer. Individual decision making. In J. H. Kagel and
A. E. Roth, editors, The Handbook of Experimental Economics,
chapter 8, pages 587–703. Princeton University Press, Princeton,
N.J., 1995.
[65] J. Y. Campbell, A. W. Lo, and A. C. MacKinlay. The Economet-
rics of Financial Markets. Princeton University Press, Princeton,
NJ, 1997.
[66] I. Chang, D. Stauffer, and R. B. Pandey. Asymmetries, corre-
lations and fat tails in percolation market model. International
Journal of Theoretical and Applied Finance, 5(6):585–597, 2002.
[67] S.-H. Chen, T. Lux, and M. Marchesi. Testing for non-linear
structure in an artificial financial market. Journal of Economic
Behavior and Organization, 46(3):327–342, 2001.
[68] S.-H. Chen and C.-H. Yeh. Evolving traders and the business
school with genetic programming: A new architecture of the
agent-based artificial stock market. Journal of Economic Dy-
namics and Control, 25(3-4):363–393, 2001.
[69] S.-H. Chen and C.-H. Yeh. On the emergent properties of artifi-
cial stock markets: the efficient market hypothesis and the ratio-
200 References

nal expectations hypothesis. Journal of Economic Behavior and

Organization, 49(2):217–239, 2002.
[70] C. Cherniak. Minimal Rationality. MIT Press, Cambridge, MA,
1986.
[71] Y.-W. Cheung and M. D. Chinn. Macroeconomic implications
of the beliefs and behavior of foreign exchange traders. NBER
working paper 7417, 7417, 1999.
[72] Y.-W. Cheung, M. D. Chinn, and I. W. Marsh. How do UK-based
foreign exchange dealers think their market operates? CEPR
Discussion Paper 2230, 2230, 1999.
[73] C. Chiarella, R. Dieci, and L. Gardini. Speculative behaviour
and complex asset price dynamics: A global analysis. Journal of
Economic Behavior and Organization, 49(2):173–197, 2002.
[74] P. K. Clark. A subordinated stochastic process model with fi-
nite variance for speculative prices. Econometrica, 41(1):135–156,
1973.
[75] D. Cliff and J. Bruten. More than zero intelligence needed for
continuous double-auction trading. Technical Report HPL-97-
157, Hewlett-Packard Laboratories, Bristol, UK, December 1997.
[76] D. Cliff and J. Bruten. Zero is not enough: On the lower limit of
agent intelligence for continuous double auction markets. Tech-
nical Report HPL-97-141, Hewlett-Packard Laboratories, Bristol,
UK, 1997.
[77] K. C. Cole. The Universe and the Teacup: The Mathematics of
Truth and Beauty. Harcourt Brace, New York, 1998.
[78] N. Collier, T. Howe, and M. North. Onward and upward: The
transition to Repast 2.0. In Proceedings of the First Annual
North American Association for Computational Social and Or-
ganizational Science Conference, Electronic Proceedings, Pitts-
burgh, PA, 2003.
[79] J. Conlisk. Why bounded rationality? Journal of Economic Lit-
erature, 34(2):669–700, June 1996.
[80] J. K. Conner and D. L. Hartl. A Primer of Ecological Genetics.
Sinauer Associates, Sunderland, MA, 2004.
[81] R. Cont and J.-P. Bouchaud. Herd behavior and aggregate fluctu-
ations in financial markets. Macroeconomic Dynamics, 4(2):170–
196, 2000.
[82] P. H. Cootner, editor. The Random Character of Stock Market
Prices. MIT Press, Cambridge, MA, 1964.
 References 201

[83] T. E. Copeland. A model of asset trading under the assumption

of sequential information arrival. Journal of Finance, 31(4):1149–
1168, 1976.
[84] K. Cuthbertson. Quantitative Financial Economics: Stocks,
Bonds, and Foreign Exchange. Wiley, Chichester, UK, 1996.
[85] D. M. Cutler, J. M. Poterba, and L. H. Summers. Speculative
dynamics and the role of feedback traders. American Economic
Review, Papers and Proceedings, 80(2):63–68, 1990.
[86] R. B. D’Agostino and M. A. Stephens. Goodness-of-fit Tech-
niques. Marcel Dekker, New York, 1986.
[87] F. J. Damerau and B. B. Mandelbrot. Tests of the degree of word
clustering in samples of written english. Linguistics, 102:58–75,
1973.
[88] K. Daniel, D. Hirshleifer, and A. Subrahmanyam. Investor psy-
chology and security market under- and overreactions. Journal
of Finance, 53(6):1839–1885, 1998.
[89] H. Dawid. Adaptive Learning by Genetic Algorithms. Springer,
Berlin et al., 2nd, revised and enlarged edition, 1999.
[90] R. Dawkins. The Selfish Gene. Oxford University Press, Oxford,
UK, 1976.
[91] K. De Jong. Adaptive system design: A genetic approach. IEEE
Transactions on Systems, Man and Cybernetics, 10(9):566–574,
Sept. 1980.
[92] K. De Jong. Genetic algorithms are not function optimizers. In
L. D. Whitley, editor, Foundations of Genetic Algorithms, pages
5–17. Morgan Kaufmann, San Mateo, CA, 1993.
[93] K. De Jong, D. B. Fogel, and H.-P. Schwefel. A history
of evolutionary computation. In T. Bäck, D. B. Fogel, and
Z. Michalewicz, editors, Handbook of Evolutionary Computation,
chapter A2.3. IOP Publishing Ltd and Oxford University Press,
Oxford, UK, 1997.
[94] K. Deb. Encoding and decoding functions. In T. Bäck, D. B. Fo-
gel, and Z. Michalewicz, editors, Handbook of Evolutionary Com-
putation, chapter C4.2. IOP Publishing Ltd and Oxford Univer-
sity Press, Oxford, UK, 1997.
[95] W. F. M. DeBondt and R. Thaler. Does the stock market over-
react? Journal of Finance, 40(3):793–805, 1985.
[96] J. B. DeLong, A. Shleifer, L. H. Summers, and R. J. Waldmann.
Noise trader risk in financial markets. Journal of Political Econ-
omy, 98(4):703–738, August 1990.
202 References

[97] J. B. DeLong, A. Shleifer, L. H. Summers, and R. J. Waldmann.

Positive feedback investment strategies and destabilizing rational
speculation. Journal of Finance, 45(2):379–395, June 1990.
[98] J. B. DeLong, A. Shleifer, L. H. Summers, and R. J. Waldmann.
The survival of noise traders in financial markets. Journal of
Business, 64(1):1–19, 1991.
[99] A. Devenow and I. Welch. Rational herding in financial eco-
nomics. European Economic Review, 40(3-5):603–615, 1996.
[100] E. Dimson and M. Mussavian. A brief history of market efficiency.
European Financial Management, 4(1):91–103, 1998.
[101] K. Dopfer, editor. Evolutionary Economics: Program and Scope.
Kluwer, Boston, 2001.
[102] G. Dosi and R. R. Nelson. An introduction to evolutionary theo-
ries in economics. Journal of Evolutionary Economics, 4(3):153–
172, 1994.
[103] J. Duffy. Agent-based models and human subject experiments.
In K. L. Judd and L. Tesfatsion, editors, Handbook of Computa-
tional Economics, volume 2, chapter 4. North-Holland, Amster-
dam, 2006.
[104] F. Y. Edgeworth. Mathematical Psychics: An Essay on the Ap-
plication of Mathematics to the Moral Sciences. Augustus M.
Kelley, New York, 1881.
[105] N. Ehrentreich. The Santa Fe Artificial Stock Market re-
examined: Suggested corrections. Betriebswirtschaftliche Diskus-
sionsbeiträge 45/02, Martin Luther University Halle-Wittenberg,
2002.
[106] N. Ehrentreich. Technical trading in the Santa Fe Institute Arti-
ficial Stock Market revisited. Journal of Economic Behavior and
Organization, 61(4):599–616, 2006.
[107] R. N. Elliott. The Wave Principle. Elliott, New York, 1938.
[108] R. N. Elliott. Nature’s Law—The Secret of the Universe. Elliott,
New York, 1946.
[109] D. Ellsberg. Risk, ambiguity, and the Savage axioms. Quarterly
Journal of Economics, 75(4):643–669, 1961.
[110] R. F. Engle. Autoregressive conditional heteroskedasticity with
estimates of the variance of United Kingdom inflation. Econo-
metrica, 50(4):987–1008, 1982.
[111] T. W. Epps and M. L. Epps. The stochastic dependence of secu-
rity price changes and transaction volumes: Implications for the
mixture-of-distributions hypothesis. Econometrica, 44(2):305–
321, 1976.
 References 203

[112] L. J. Eshelman. Genetic algorithms. In T. Bäck, D. B. Fogel, and

Z. Michalewicz, editors, Handbook of Evolutionary Computation,
chapter B1.2. IOP Publishing Ltd and Oxford University Press,
Oxford, UK, 1997.
[113] E. F. Fama. Mandelbrot and the stable Paretian hypothesis.
Journal of Business, 36(4):420–429, 1963.
[114] E. F. Fama. The behavior of stock-market prices. Journal of
Business, 38(1):34–105, 1965.
[115] E. F. Fama. Efficient capital markets: A review of theory and
empirical work. Journal of Finance, 25(2):383–417, 1970.
[116] E. F. Fama. Efficient capital markets II. Journal of Finance,
46(5):1575–1617, 1991.
[117] E. F. Fama, L. Fisher, M. C. Jensen, and R. Roll. The adjust-
ment of stock prices to new information. International Economic
Review, 10(1):1–21, 1969.
[118] E. F. Fama and K. R. French. Dividend yields and expected stock
returns. Journal of Financial Economics, 22(1):3–25, 1988.
[119] J. D. Farmer. Physicists attempt to scale the ivory towers of
finance. Computing in Science & Engineering (IEEE), 1(6):26–
39, 1999.
[120] J. D. Farmer. Toward agent-based models for investment. In
Developments in Quantitative Investment Models, pages 61–71,
AIMR
http://www.santafe.edu/˜jdf/papers/aimr.pdf, 2001.
[121] J. D. Farmer and S. Joshi. The price dynamics of common trad-
ing strategies. Journal of Economic Behavior and Organization,
49(2):149–171, 2002.
[122] J. D. Farmer and A. W. Lo. Frontiers of finance: Evolution and
efficient markets. Proceedings of the National Academy of Science
of the USA, 96:9991–9992, 1999.
[123] J. D. Farmer, P. Patelli, and lija I. Zovko. The predictive power of
zero intelligence in financial markets. Proceedings of the National
Academy of Science of the USA, 102(6):2254–2259, 2005.
[124] M. J. Farrell. Profitable speculation. Economica, 33(130):183–
193, 1966.
[125] E. Fehr and J.-R. Tyran. Individual irrationality and aggregate
outcomes. Journal of Economic Perspectives, 19(4):43–66, 2005.
[126] J. Felsenstein. Theoretical Evolutionary Genetics – Genome
562. http://evolution.gs.washington.edu/pgbook/pgbook.html,
March 2005.
204 References

[127] F. Fernandez-Rodriguez, C. Gonzalez-Martel, and S. Sosvilla-

Rivero. On the profitability of technical trading rules based on
artificial neural networks: Evidence from the Madrid stock mar-
ket. Economics Letters, 69(1):89–94, 2000.
[128] R. A. Fisher. The Genetical Theory of Natural Selection. Claren-
don Press, Oxford, 1930.
[129] S. Focardi, S. Cincotti, and M. Marchesi. Self-organization and
market crashes. Journal of Economic Behavior and Organization,
49(2):241–267, 2002.
[130] T. C. Fogarty. Varying the probability of mutation in the genetic
algorithm. In J. D. Schaffer, editor, Proc. 3rd Int. Conf. on Ge-
netic Algorithms, pages 104–109. Morgan Kaufmann, San Mateo,
CA, 1989.
[131] D. B. Fogel. Other selection methods. In T. Bäck, D. B. Fogel, and
Z. Michalewicz, editors, Handbook of Evolutionary Computation,
chapter C 2.6. IOP Publishing Ltd and Oxford University Press,
Oxford, UK, 1997.
[132] C. Forner and J. Marhuenda. Contrarian and momentum strate-
gies in the Spanish stock market. European Financial Manage-
ment, 9(1):67–88, 2003.
[133] J. Forrester. Industrial Dynamics. MIT Press, Cambridge, MA,
1961.
[134] J. Forrester. Urban Dynamics. MIT Press, Cambridge, MA, 1969.
[135] J. Forrester. World Dynamics. Wright-Allen Press, Cambridge,
MA, 1971.
[136] F. D. Foster and S. Viswanathan. Variations in trading volume,
return volatility and trading costs: Evidence on recent price for-
mation models. Journal of Finance, 48(1):187–211, 1993.
[137] C. Fourgeaud, C. Gourieroux, and J. Pradel. Learning proce-
dures and convergence to rationality. Econometrica, 54(4):845–
868, July 1986.
[138] J. A. Frankel and K. A. Froot. Chartists, fundamentalists, and
trading in the foreign exchange market. American Economic Re-
view, Papers and Proceedings, 80(2):181–185, 1990.
[139] D. Friedman. Monty hall’s three doors: Construction and de-
construction of a choice anomaly. American Economic Review,
88(4):933–946, 1998.
[140] M. Friedman. The methodology of positive economics. In Essays
in Positive Economics, pages 3–43. University of Chicago Press,
Chicago, 1953.
 References 205

[141] M. Friedman. The role of monetary policy. American Economic

Review, 58(1):1–17, 1968.
[142] M. Friedman. Nobel lecture: Inflation and unemployment. Jour-
nal of Political Economy, 85(3):451–472, 1977.
[143] A. R. Gallant, P. E. Rossi, and G. Tauchen. Stock prices and
volume. Review of Financial Studies, 5(2):199–242, 1992.
[144] M. Gardner. Mathematical games: The curious properties of the
Gray code and how it can be used to solve puzzles. Scientific
American, 227(2):106–109, August 1972.
[145] A. Gaunersdorfer. Endogenous fluctuations in a simple asset pric-
ing model with heterogeneous agents. Journal of Economic Dy-
namics and Control, 24(5-7):799–831, 2000.
[146] M. Gell-Mann. What is complexity? Complexity, 1(1), 1995.
[147] R. Gençay, F. Selçuk, and A. Ulugülyaǧci. High volatility, thick
tails, and extreme value theory in value-at-risk estimation. In-
surance: Mathematics and Economics, 33(2):337–356, 2003.
[148] N. Gilbert and S. Bankes. Platforms and methods for agent-based
modeling. Proceedings of the National Academy of Science of the
USA, 99(3):7197–7198, 2002.
[149] J. Gleick. Chaos: Making a New Science. Penguin, New York,
1987.
[150] D. K. Gode and S. Sunder. Allocative efficiency of markets with
zero-intelligence traders: Market as a partial substitute for indi-
vidual irrationality. Journal of Political Economy, 101(1):119–
137, Feb. 1993.
[151] D. Goldbaum. Cycles of market stability and instability due to
endogenous use of technical trading rules. In Y. S. Abu-Mostafa,
B. LeBaron, A. W. Lo, and A. S. Weigand, editors, Computational
Finance 1999. MIT Press, New York and Cambridge, MA, 2000.
[152] D. Goldbaum. Profitable technical trading rules ans a source of
price instability. Quantitative Finance, 3(3):220–229, 2003.
[153] D. Goldbaum. Market efficiency and learning in an endogenously
unstable environment. Journal of Economic Dynamics and Con-
trol, 29(5):953–978, 2005.
[154] D. E. Goldberg. Genetic Algorithms in Search, Optimization, and
Machine Learning. Addison-Wesley, Reading, MA, 1989.
[155] D. E. Goldberg and J. Richardson. Genetic algorithms with shar-
ing for multimodal function optimization. In J. J. Grefenstette,
editor, Genetic Algorithms and their Applications: Proceedings of
the Second International Conference on Genetic Algorithms Juy
206 References

28-31, 1987 MIT Cambridge, pages 41–49. Lawrence Erlbaum

Associates, Hillsdale, NJ, 1987.
[156] D. E. Goldberg and P. Segrest. Finite Markov chain analysis of
genetic algorithms. In J. J. Grefenstette, editor, Genetic Algo-
rithms and their Applications: Proceedings of the Second Interna-
tional Conference on Genetic Algorithms Juy 28-31, 1987 MIT
Cambridge, pages 1–8. Lawrence Erlbaum Associates, Hillsdale,
NJ, 1987.
[157] J. Gomez. Self adaptation of operator rates in evolutionary al-
gorithms. In K. Deb, editor, Genetic and Evolutionary Com-
putation - GECCO 2004 Genetic and Evolutionary Conference
Seattle, WA, USA, June 26-30, 2004, Proceedings Part I, Lec-
ture Notes in Computer Science 3102, pages 1162–1173. Springer,
2004.
[158] P. Gopikrishnan, V. Plerou, L. A. N. Amaral, M. Meyer, , and
H. E. Stanley. Scaling of the distribution of fluctuations of finan-
cial market indices. Physical Review E, 60(5):5305–5316, 1999.
[159] P. Gopikrishnan, V. Plerou, Y. Liu, L. A. N. Amaral, X. Gabaix,
and H. E. Stanley. Scaling and correlation in financial time series.
Physica A, 287:362–373, 2000.
[160] W. M. Gorman. Community preference fields. Econometrica,
21(1):63–80, 1953.
[161] H. A. J. Green. Aggregation in Economic Analysis: An Introduc-
tory Survey. Princeton University Press, Princeton, N.J., 1964.
[162] J. Grefenstette. Proportional selection and sampling algorithms.
In T. Bäck, D. B. Fogel, and Z. Michalewicz, editors, Handbook
of Evolutionary Computation, chapter C2.2. IOP Publishing Ltd
and Oxford University Press, Oxford, UK, 1997.
[163] J. Grefenstette. Rank-based selection. In T. Bäck, D. B. Fogel,
and Z. Michalewicz, editors, Handbook of Evolutionary Compu-
tation, chapter C 2.4. IOP Publishing Ltd and Oxford University
Press, Oxford, UK, 1997.
[164] S. Grossman and J. E. Stiglitz. On the impossibility of informa-
tionally efficient markets. American Economic Review, 70(3):393–
408, 1980.
[165] D. M. Guillaume, M. M. Dacorogna, R. R. Davé, U. A. Müller,
R. B. Olsen, and O. V. Pictet. From the bird’s eye to the mi-
croscope: A survey of new stylized facts of the intra-daily foreign
exchange markets. Finance and Stochastics, 1(2):95–129, 1997.
[166] L. Gulyás, B. Adamcsek, and A. Kiss. An early agent-based stock
market: Replication and participation. In Proceedings 4th Soft-
 References 207

Computing for Economics and Finance Meeting, May 29, 2003,

University Ca’ Foscari, Venice, Italy, 2003.
[167] P. J. B. Hancock. A comparison of selection mechanisms. In
T. Bäck, D. B. Fogel, and Z. Michalewicz, editors, Handbook of
Evolutionary Computation, chapter C 2.8. IOP Publishing Ltd
and Oxford University Press, Oxford, UK, 1997.
[168] P. R. Hansen and A. Lunde. A forecast comparison of volatility
models: Does anything beat a GARCH(1,1)? Journal of Applied
Econometrics (forthcoming), 2005.
[169] P. R. Hansen, A. Lunde, and J. M. Nason. Choosing the best
volatility models: The model confidence set approach. Oxford
Bulletin of Economics and Statistics, Supplement, 65(0):839–861,
2003.
[170] G. A. Harrison, J. M. Tanner, D. R. Pilbeam, and P. T. Baker.
Human Biology: An Introduction to Human Evolution, Variation,
Growth, and Adaptability. Oxford University Press, Oxford, UK,
3 edition, 1988.
[171] R. F. Harrod. Price and cost in entrepreneurs’ policy. Oxford
Economic Papers, 2(1):1–11, 1939.
[172] J. E. Hartley. The Representative Agent in Macroeconomics.
Routledge, London, New York, 1997.
[173] F. G. Heath. Origins of the binary code. Scientific American,
227(2):76–83, August 1972.
[174] P. W. Hedrick. Genetics of Populations. Jones and Bartlett,
Sudbury, MA, 2 edition, 1999.
[175] M. Heinemann. Adaptive learning of rational expectations using
neural networks. Journal of Economic Dynamics and Control,
24(5-7):1007–1026, 2000.
[176] M. F. Hellwig. Zur Informationseffizienz des Kapitalmarktes.
Zeitschrift für Wirtschafts- und Sozialwissenschaften, 102(1):1–
27, 1982.
[177] B. Hergenhahn. An Introduction to Theories of Learning. Pren-
tice Hall, Englewood Cliffs, N.J., 3 edition, 1988.
[178] I. N. Herstein and J. Milnor. An axiomatic approach to measur-
able utility. Econometrica, 21(2):291–297, 1953.
[179] J. Hirshleifer. The expanding domain of economics. American
Economic Review, 75(6):53–68, 1985.
[180] G. M. Hodgson. Economics and Evolution - Bringing Life Back
into Economics. Economics, Cognition, and Society. University
of Michigan Press, Ann Arbor, 1993.
208 References

[181] G. M. Hodgson. Darwinism in economics: From analogy to on-

tology. Journal of Evolutionary Economics, 12(3):259–281, 2002.
[182] J. H. Holland. Adaptation in Natural and Artificial Systems: An
Introductory Analysis with Applications to Biology, Control, and
Artificial Intelligence. University of Michigan Press, Ann Arbor,
MI, 1975.
[183] J. H. Holland. The global economy as an adaptive process. In
P. W. Anderson, K. J. Arrow, and D. Pines, editors, The Economy
as an Evolving Complex System, volume 5 of Santa Fe Institute
studies in the sciences of complexity, pages 117–124. Westview,
Santa Fe, N.M., 1988.
[184] J. H. Holland. Adaptation in Natural and Artificial Systems: An
Introductory Analysis with Applications to Biology, Control, and
Artificial Intelligence. MIT Press / Bradford Books edition, Cam-
bridge, MA, 1992.
[185] J. H. Holland, K. J. Holyoak, R. E. Nisbett, and P. R. Thagard.
Induction: Processes of Inference, Learning, and Discovery. MIT
Press, Cambridge, MA, 1986.
[186] J. H. Holland and J. H. Miller. Artificial adaptive agents in eco-
nomic theory. American Economic Review, Papers and Proceed-
ings, 81(2):365–370, 1991.
[187] J. H. Holland and J. S. Reitman. Cognitive systems based on
adaptive algorithms. In D. A. Waterman and F. Hayes-Roth,
editors, Pattern-Directed Inference Systems, pages 313–329. Aca-
demic Press, New York, 1978.
[188] R. B. Hollstien. Artificial Genetic Adaptation in Computer Con-
trol Systems. Phd thesis, University of Michigan, 1971.
[189] C. H. Hommes. Financial markets as nonlinear adaptive evolu-
tionary systems. Quantitative Finance, 1(1):149–167, 2001.
[190] C. H. Hommes. Heterogeneous agent models in economics and
finance. In K. L. Judd and L. Tesfatsion, editors, Handbook of
Computational Economics, volume 2. Elsevier, Amsterdam, 2006.
[191] C. H. Hommes and G. Sorger. Consistent expectations equilibria.
Macroeconomic Dynamics, 2(3):287–321, 1998.
[192] J. Hong and J. Stein. A unified theory of underreaction, mo-
mentum trading, and overreaction in asset markets. Journal of
Finance, 54(6):2143–2184, 1999.
[193] J. Horgan. From complexity to perplexity. Scientific American,
272(6):104–110, 1995.
 References 209

[194] G. Iori. A threshold model for stock return volatility and trad-
ing volume. International Journal of Theoretical and Applied
Finance, 3(3):467–472, 2000.
[195] G. Iori. A microsimulation of traders activity in the stock market:
The role of heterogeneity, agents’ interactions and trade frictions.
Journal of Economic Behavior and Organization, 49(2):269–285,
2002.
[196] J. F. Jaffe. Special information and insider trading. Journal of
Business, 47(410):410–428, 1974.
[197] C. M. Jarque and A. K. Bera. Efficient tests for normality, het-
eroscedasticity and serial independence of regression residuals.
Economics Letters, 6(3):255–259, 1980.
[198] T. Jasic and D. Wood. The profitability of daily stock market
indices trades based on neural network predictions: Case study
for the S&P 500, the DAX, the TOPIX and the FTSE in the
period 1965-1999. Applied Financial Economics, 14(4):285–297,
2004.
[199] N. Jegadeesh. Foundations of technical analysis: Computational
algorithms, statistical inference, and empirical implementation:
Discussion. Journal of Finance, 55(4):1765–1770, 2000.
[200] N. Jegadeesh and S. Titman. Returns to buying winners and
selling losers: Implications for stock market efficiency. Journal of
Finance, 48(1):65–91, 1993.
[201] M. C. Jensen. Some anomalous evidence regarding market effi-
ciency. Journal of Financial Economics, 6(2-3):95–101, 1978.
[202] M. C. Jensen and G. A. Benington. Random walks and tech-
nical theories: Some additional evidence. Journal of Finance,
25(2):469–482, 1970.
[203] M. C. Jerison. Aggregation and pairwise aggregation of demand
when the distribution of income is fixed. Journal of Economic
Theory, 33(1):1–33, June 1984.
[204] P. E. Johnson. Agent-based modeling: What I learned from
the artificial stock market. Social Science Computer Review,
20(2):174–186, 2002.
[205] S. Joshi and M. A. Bedau. An explanation of generic behavior in
an evolving financial market. Santa Fe Institute Working Paper
98-12-114, 1998.
[206] S. Joshi, J. Parker, and M. A. Bedau. Technical trading creates
a prisoner’s dilemma: results from an agent-based model. Santa
Fe Institute Working Paper 98-12-115, 1998.
210 References

[207] S. Joshi, J. Parker, and M. A. Bedau. Financial markets can be

at sub-optimal equilibria. Computational Economics, 19(1):5–23,
2002.
[208] K. L. Judd. Computational economics and economic theory: Sub-
stitutes or complements? Journal of Economic Dynamics and
Control, 21(6):907–942, 1997.
[209] K. L. Judd and L. Tesfatsion, editors. Handbook of Computational
Economics, volume 2. Elsevier, Amsterdam, 2006.
[210] J. Jung and R. J. Shiller. Samuelson’s dictum and the stock
market. Economic Inquiry, 43(2):221–228, 2005.
[211] M. A. Kaboudan. Genetic programming prediction of stock
prices. Computational Economics, 16(3):207–236, 2000.
[212] D. Kahneman and A. Tversky. Prospect theory: An analysis of
decision under risk. Econometrica, 47(2):263–291, 1979.
[213] J. M. Karpoff. A theory of trading volume. Journal of Finance,
41(5):1069–1087, 1986.
[214] J. M. Karpoff. The relation between price changes and trading
volume: A survey. Journal of Financial and Quantitative Analy-
sis, 22(1):109–126, 1987.
[215] S. Kauffman. At Home in the Universe: The Search for Laws of
Self-Organization and Complexity. Oxford University Press, New
York, Oxford, 1995.
[216] M. Kendall. The analysis of economic time series, part i: Prices.
Journal of the Royal Statistical Society, Series A, 96:11–25, 1953.
[217] D. A. Kendrick. Sectoral economics. In H. M. Amman, D. A.
Kendrick, and J. Rust, editors, Handbook of Computational Eco-
nomics, chapter 6, pages 295–332. Elsevier, Amsterdam, 1996.
[218] J. M. Keynes. The General Theory of Employment, Interest,
and Money. MacMillan, New York, reprinted 1964 San Diego,
Harcourt Brace, 1936.
[219] G. A. Kimble. The definitions of learning and some useful dis-
tinctions. In G. A. Kimble, editor, Foundations of Condition-
ing and Learning, The Century Psychology Series, pages 82–99.
Appleton-Century-Crofts, New York, 1967.
[220] M. Kimura. Solution of a process of random genetic drift with a
continuous model. Proceedings of the National Academy of Sci-
ence of the USA, 41(3):144–150, March 1955.
[221] M. Kimura. On the probability of fixation of mutant genes in a
population. Genetics, 47(6):713–719, 1962.
[222] M. Kimura. Evolutionary rate at the molecular level. Nature,
217:624–626, 1968.
 References 211

[223] M. Kimura. The Neutral Theory of Molecular Evolution. Cam-

bridge University Press, Cambridge, UK, 1983.
[224] M. Kimura and T. Ohta. Theoretical Aspects of Population Ge-
netics. Princeton University Press, Princeton, NJ, 1971.
[225] A. P. Kirman. Whom or what does the representative individ-
ual represent? Journal of Economic Perspectives, 6(2):117–136,
Spring 1992.
[226] N. Kiyotaki and R. Wright. On money as a medium of exchange.
Journal of Political Economy, 97(4):927–954, 1989.
[227] T. S. Kuhn. The Structure of Scientific Revolutions. University
of Chicago Press, Chicago, London, 3 edition, 1996.
[228] K. Y. Kwon and R. J. Kish. A comparative study of techni-
cal trading strategies and return predictability: An extension of
Brock, Lakonishok, and LeBaron (1992) using NYSE and NAS-
DAQ indices. Quarterly Review of Economics and Finance,
42(3):611–631, 2002.
[229] K. Y. Kwon and R. J. Kish. Technical trading strategies
and return predictability: NYSE. Applied Financial Economics,
12(9):639–653, 2002.
[230] A. S. Kyle. Continuous auctions and insider trading. Economet-
rica, 53(6):1315–1336, 1985.
[231] R. N. Langlois. Bounded rationality and behavioralism: A clar-
ification and critique. Journal of Institutional and Theoretical
Economics, 146(4):691–695, 1990.
[232] R. N. Langlois. Rule-following, expertise, and rationality: a
new behavioral economics? In K. Dennis, editor, Rationality in
Economics: Alternative Perspectives, pages 57–80. Kluwer, Dor-
drecht, 1998.
[233] R. N. Langlois and L. Csontos. Optimization, rule following,
and the methodology of situational analysis. In B. Gustafsson,
C. Knudsen, and U. Mäki, editors, Rationality, Institutions, and
Economic Methodology. Routledge, Nondon, New York, 1993.
[234] J. S. Lansing, J. N. Kremer, and B. B. Smuts. System-
dependent selection, ecological feedback and the emergence of
functional structure in ecosystems. Journal of Theoretical Biol-
ogy, 192(3):377–391, 1998.
[235] T. Lawson. Economics and Reality. Routledge, London, New
York, 1997.
[236] B. LeBaron. Building financial markets with artificial agents:
Desired goals, and present techniques. In G. Karakoulas, editor,
Computational Markets. MIT Press, 1999.
212 References

[237] B. LeBaron. Technical trading rule profitability and foreign

exchange intervention. Journal of International Economics,
49(1):125–143, 1999.
[238] B. LeBaron. Agent-based computational finance: Suggested read-
ings and early research. Journal of Economic Dynamics and Con-
trol, 24(5-7):679–702, 2000.
[239] B. LeBaron. A builder’s guide to agent based financial markets.
Quantitative Finance, 1(2):254–261, March 2001.
[240] B. LeBaron. Empirical regularities from interacting long and
short memory investors in an agent based stock market. IEEE
Transactions on Evolutionary Computation, 5(5):442–455, 2001.
[241] B. LeBaron. Evolution and time horizons in an agent-based stock
market. Macroeconomic Dynamics, 5(2):225–254, 2001.
[242] B. LeBaron. Building the Santa Fe Artificial Stock Market. work-
ing paper, Brandeis University, 2002.
[243] B. LeBaron. Agent-based computational finance. In K. L. Judd
and L. Tesfatsion, editors, The Handbook of Computational Eco-
nomics, volume 2, chapter 9. Elsevier, Amsterdam, 2006.
[244] B. LeBaron, W. B. Arthur, and R. Palmer. Time series properties
of an artificial stock market. Journal of Economic Dynamics and
Control, 23(9-10):1487–1516, 1999.
[245] S. F. LeRoy. Efficient capital markets and martingales. Journal
of Economic Literature, 27(4):1583–1621, 1989.
[246] M. Lettau. Explaining the facts with adaptive agents: The case of
mutual fund flows. Journal of Economic Dynamics and Control,
21(7):1117–1147, 1997.
[247] M. Levy, H. Levy, and S. Solomon. A microscopic model of the
stock market: Cycles, booms, and crashes. Economics Letters,
45(1):103–111, 1994.
[248] M. Levy, H. Levy, and S. Solomon. Microscopic Simulation of Fi-
nancial Markets: From Investor Behavior to Market Phenomena.
Academic Press, San Diego et al., 2000.
[249] A. Lewbel. Exact aggregation and a representative consumer.
Quarterly Journal of Economics, 104(3):621–633, 1989.
[250] R. C. Lewontin. The Genetic Basis of Evolutionary Change.
Columbia University Press, New York, 1974.
[251] W. K. Li, S. Ling, and M. McAleer. Recent theoretical results
for time series models with GARCH errors. Journal of Economic
Surveys, 16(3):245–269, 2002.
 References 213

[252] J. Lintner. The valuation of risk assets and the selection of risky
investments in stock portfolios and capital budgets. The Review
of Economics and Statistics, 47(1):13–37, 1965.
[253] B. L. Lipman. How to decide how to decide how to ... : Modeling
bounded rationality. Econometrica, 59(4):1105–1125, 1991.
[254] R. G. Lipsey. The relation between unemployment and the rate
of change of money wages in the United Kingdom 1862-1957: A
further analysis. Economica, 27(105):1–31, 1960.
[255] A. W. Lo. The Adaptive Markets Hypothesis: Market efficiency
from an evolutionary perspective. Journal of Portfolio Manage-
ment, 30:15–29, 2004.
[256] A. W. Lo, H. Mamaysky, and J. Wang. Foundations of techni-
cal analysis: Computational algorithms, statistical inference, and
empirical implementation. Journal of Finance, 55(4):1705–1765,
2000.
[257] I. N. Lobato and C. Velasco. Long memory in stock-market
trading volume. Journal of Business and Economic Statistics,
18(4):410–426, 2000.
[258] G. Loomes and C. Taylor. Non-transitive preferences over gains
and losses. Economic Journal, 102(411):357–365, 1992.
[259] M. Loretan and P. C. B. Phillips. Testing the covariance station-
arity of heavy-tailed time series. Journal of Empirical Finance,
1(2):211–248, 1994.
[260] J. Lucas, Robert E. Expectations and the neutrality of money.
Journal of Economic Theory, 4(2):103–124, April 1972.
[261] J. Lucas, Robert E. Econometric policy evaluation: A critique. In
K. Brunner and A. H. Meltzer, editors, The Phillips Curve and
Labor Markets, volume 1 of Carnegie-Rochester Conference Series
on Public Policy, pages 19–46. North-Holland, Amsterdam, 1976.
[262] J. Lucas, Robert E. Asset prices in an exchange economy. Econo-
metrica, 46(6):1429–1445, 1978.
[263] Y.-H. Lui and D. Mole. The use of fundamental and technical
analyses by foreign exchange dealers: Honk Kong evidence. In-
ternational Journal of Finance and Economics, 17(3):535–545,
1998.
[264] T. Lux. Herd behaviour, bubbles, and crashes. Economic Journal,
105(431):881–896, 1995.
[265] T. Lux. The stable Paretian hypothesis and the frequency of
large returns: An examination of major German stocks. Applied
Financial Economics, 6(6):463–475, 1996.
214 References

[266] T. Lux. The socio-economic dynamics of speculative markets:

Interacting agents, chaos, and the fat tails of return distributions.
Journal of Economic Behavior and Organization, 33(2):143–165,
1998.
[267] T. Lux and M. Ausloos. Market fluctuations I: Scaling, multi-
scaling and their possible origins. In A. Bunde, J. Kropp, and
H.-J. Schellnhuber, editors, The Science of Desaster: Climate
Disruptions, Heart Attacks, and Market Crashes, pages 373–409.
Springer, 2002.
[268] T. Lux and M. Marchesi. Scaling and criticality in a stochastic
multi-agent model of a financial market. Nature, 397:498–500,
1999.
[269] T. Lux and M. Marchesi. Volatility clustering in financial mar-
kets: a microsimulationof interacting agents. International Jour-
nal of Theoretical and Applied Finance, 3(4):675–702, 2000.
[270] S. W. Mahfoud. Population size and genetic drift in fitness shar-
ing. In L. D. Whitley and M. D. Vose, editors, Foundations of
Genetic Algorithms 3, volume 3, pages 185–223. Morgan Kauf-
mann, San Francisco, CA, 1995.
[271] S. W. Mahfoud. Boltzmann selection. In T. Bäck, D. B. Fogel, and
Z. Michalewicz, editors, Handbook of Evolutionary Computation,
chapter C 2.5. IOP Publishing Ltd and Oxford University Press,
Oxford, UK, 1997.
[272] S. W. Mahfoud. Niching methods. In T. Bäck, D. B. Fogel, and
Z. Michalewicz, editors, Handbook of Evolutionary Computation,
chapter C6.1. IOP Publishing Ltd and Oxford University Press,
Oxford, UK, 1997.
[273] U. Mäki. ‘The Methodology of Positive Economics’ (1953) does
not give us the methodology of positive economics. Journal of
Economic Methodology, 10(4):495–505, 2003.
[274] B. G. Malkiel. A Random Walk Down Wall Street : Including a
Life-Cycle Guide to Personal Investing. Norton, New York, 1999.
[275] B. G. Malkiel. The efficient market hypothesis and its critics.
Journal of Economic Perspectives, 17(1):59–82, 2003.
[276] B. B. Mandelbrot. The variation of certain speculative prices.
Journal of Business, 36(4):394–419, 1963.
[277] B. B. Mandelbrot. Forecasts of future prices, unbiased markets,
and “martingale” models. Journal of Business, 39(1, Part 2: Sup-
plement on Security Pricing):242–255, 1966.
 References 215

[278] B. B. Mandelbrot. Scaling in financial prices: 3: Cartoon Brown-

ian motions in multifractal time. Quantitative Finance, 1(4):427–
440, 2001.
[279] B. B. Mandelbrot, A. Fisher, and L. Calvet. A multifractal model
of asset returns. Cowles Foundation Discussion Paper #1164,
1997.
[280] B. B. Mandelbrot and R. L. Hudson. The (Mis)behavior of Mar-
kets: A Fractal View of Risk, Ruin, and Reward. Basic Books,
New York, 2004.
[281] B. B. Mandelbrot and J. R. Wallis. Noah, Joseph, and operational
hydrology. Water Resources Research, 4(3):909–918, 1968.
[282] R. N. Mantegna and H. E. Stanley. Scaling behavior in the be-
havior of an economic index. Nature, 376:46–49, 1995.
[283] R. N. Mantegna and H. E. Stanley. An Introduction to Econo-
physics: Correlations and Complexity in Finance. Cambridge
University Press, Cambridge, UK, 2000.
[284] R. Marimon. Adaptive learning, evolutionary dynamics, and equi-
librium selection in games. European Economic Review, 37(2-
3):603–611, 1993.
[285] R. Marimon, E. McGrattan, and T. Sargent. Money as a medium
of exchange in an economy with artificially intelligent agents.
Journal of Economic Dynamics and Control, 14(2):329–373, 1990.
[286] S. Markose. Computability and evolutionary complexity: Mar-
kets as Complex Adaptive Systems (CAS). Economic Journal,
forthcoming, 2005.
[287] S. Markose, E. Tsang, and S. Martinez-Jaramillo. The Red Queen
principle and the emergence of efficient financial markets: An
agent based approach. In T. Lux, editor, Proceedings of the
8th Annual Workshop on Economics with Heterogeneous Inter-
acting Agents (Kiel, Germany, May 29-31, 2003), Heidelberg,
New York, 2004.
[288] H. M. Markowitz. Portfolio selection. Journal of Finance,
7(1):77–91, March 1952.
[289] H. M. Markowitz. Portfolio Selection: Efficient Diversification of
Investments. Wiley, New York, 1959.
[290] A. Marshall. Principles of Economics. Macmillan & Co., Ltd.,
London, 1890.
[291] T. Mayer. Truth versus Precision in Economics. Edward Elgar,
Aldershot, 1993.
[292] D. McFadden. Rationality for economists? Journal of Risk and
Uncertainty, 19(1-3):73–105, December 1999.
216 References

[293] G. McQueen and K. P. Vorkink. Whence GARCH? A preference-

based explanation for conditional volatility. Review of Financial
Studies, 17(4):915–949, 2004.
[294] D. H. Meadows, D. L. Meadows, and J. Randers. Beyond the
Limits. Chelsea Green, Post Mills, VT, 1992.
[295] D. H. Meadows, D. L. Meadows, J. Randers, and W. W. Behrens
III. The Limits to Growth. Universe Books, New York, 1972.
[296] D. H. Meadows, D. L. Meadows, J. Randers, and W. W. Behrens
III. The Dynamics of Growth in a Finite World. Wright-Allen
Press, Cambridge, MA, 1974.
[297] D. H. Meadows, J. Randers, and D. L. Meadows. Limits to
Growth: The 30-Year Update. Chelsea Green, White River Jct.,
VT, 2004.
[298] L. Menkhoff. Examining the use of technical currency analysis.
International Journal of Finance and Economics, 2(4):307–318,
1997.
[299] L. Menkhoff. The noise trading approach–questionnaire evidence
from foreign exchange. Journal of International Money and Fi-
nance, 17(3):547–564, 1998.
[300] M. Metghalchi and Y. H. Chang. Profitable technical trading
rules for the Italian stock market. RISEC – International Review
of Economics and Business, 50(4):433–450, 2003.
[301] D. Michie, editor. Expert Systems in the Micro-electronic Age.
Edinburgh University Press, Edinburgh, 1979.
[302] T. Mikosch and C. Stărică. Is it really long memory we see in fi-
nancial returns? In P. Embrechts, editor, Extremes and Integrated
Risk Management, chapter 12, pages 149–168. Risk Books, Lon-
don, 2000.
[303] T. Mikosch and C. Stărică. Long range dependence effects and
ARCH modelling. In P. Doukhan, G. Oppenheim, and M. Taqqu,
editors, Theory and Applications of Long Range Dependence,
pages 439–460. Birkhauser, Boston, 2003.
[304] T. Mikosch and C. Stărică. Changes of structure in financial time
series and the GARCH model. REVSTAT Statistical Journal,
2(1):41–73, 2004.
[305] T. Mikosch and C. Stărică. Non-stationarities in financial time se-
ries, the long range dependence and the IGARCH effects. Review
of Economics and Statistics, 86(1):378–390, 2004.
[306] P. Milgrom and N. Stokey. Information, trade, and common
knowledge. Journal of Economic Theory, 26(1):17–27, 1982.
 References 217

[307] P. E. Mirowski. Review of: Fractals and scaling in finance: Dis-

continuity, concentration, risk. Journal of Economic Literature,
39(2):585–587, 2001.
[308] M. Mitchell. An Introduction to Genetic Algorithms. MIT Press,
Cambridge, MA, 1996.
[309] S. K. Mitra. Profiting from technical analysis in Indian stock
market. Finance India, 16(1):109–120, 2002.
[310] P. Mongin. The marginalist controversy. In J. Davis, W. Hands,
and U. Maki, editors, Handbook of Economic Methodology, pages
558–562. Edward Elgar, London, 1997.
[311] R. D. Montassér. Die Rolle des Volumens bei der Aktienkurs-
prognose unter besonderer Berücksichtigung der AVAS-Trans-
formation. Gabler Edition Wissenschaft: Hallesche Schriften zur
Betriebswirtschaft 12. Deutscher Universitätsverlag, Wiesbaden,
2003.
[312] S. Moss and B. Edmonds. Modeling economic learning as mod-
eling. Cybernetics and Systems, 29(3):215–247, 1998.
[313] J. Mossin. Equilibrium in a capital asset market. Econometrica,
34(4):768–783, 1966.
[314] R. F. Mulligan. Fractal analysis of highly volatile markets: An ap-
plication to technology equities. Quarterly Review of Economics
and Finance, 44(1):155–179, 2004.
[315] J. F. Muth. Rational expectations and the theory of price move-
ments. Econometrica, 29(3):315–335, 1961.
[316] A. Nagurney. Parallel computation. In H. M. Amman, D. A.
Kendrick, and J. Rust, editors, Handbook in Computational Eco-
nomics, chapter 7, pages 335–404. Elsevier, Amsterdam, 1996.
[317] C. J. Neely. The temporal pattern of trading rule returns and ex-
change rate intervention: Intervention does not generate technical
trading profits. Journal of International Economics, 58(1):211–
232, 2002.
[318] D. B. Nelson. Conditional heteroskedasticity in asset returns: A
new approach. Econometrica, 59(2):347–370, 1991.
[319] R. R. Nelson and S. G. Winter. An Evolutionary Theory of Eco-
nomic Change. MIT Press, Cambridge, Mass., 1982.
[320] S. A. Nelson. The ABC of Stock Speculation. Fraser Publ. Co.,
New York, 1903.
[321] M. Nerlove. Adaptive expectations and cobweb phenomena.
Quarterly Journal of Economics, 72(2):227–240, 1958.
[322] G. Nicolis and I. Prigogine. Exploring Complexity : An Introduc-
tion. Freeman, New York, 1989.
218 References

[323] W. D. Nordhaus. World dynamics: Measurement without data.

Economic Journal, 83(332), 1973.
[324] W. D. Nordhaus. Lethal model 2: The limits to growth revisited.
Brookings Paper on Economic Activity, 2, 1992.
[325] T. Oberlechner. Importance of technical and fundamental analy-
sis in the European foreign exchange market. International Jour-
nal of Finance and Economics, 6(1):81–93, 2001.
[326] T. Ohta. Slightly deleterious mutant substitutions in evolution.
Nature, 246:96–98, 1973.
[327] T. Ohta. Near-neutrality in evolution of genes and gene regula-
tion. Proceedings of the National Academy of Science of the USA,
99(25):16134–16137, December 2002.
[328] A. Pagan. The econometrics of financial markets. Journal of
Empirical Finance, 3(1):15–102, 1996.
[329] S. E. Page. Computational models from A to Z. Complexity,
5(1):35–41, 1999.
[330] R. G. Palmer, W. B. Arthur, J. H. Holland, and B. LeBaron. An
artificial stock market. Artificial Life and Robotics, 3(1):27–31,
1999.
[331] R. G. Palmer, W. B. Arthur, B. LeBaron, and P. Tayler. Artifi-
cial economic life: A simple model of a stockmarket. Physica D,
75:264–274, 1994.
[332] R. Pancs and N. J. Vriend. Schelling’s spatial proximity model of
segregation revisited. Queen Mary University of London, Dept.
of Economics Working Paper 487, 2003.
[333] K. Pearson. The problem of the random walk. Nature, 72:294,
July 1905.
[334] M. H. Pesaran. The Limits to Rational Expectations. Basil Black-
well, Oxford, New York, 1987.
[335] E. E. Peters. Chaos and Order in the Capital Markets: A New
View of Cycles, Prices, and Market Volatility. Wiley, New York,
1991.
[336] E. E. Peters. Fractal market Analysis: Applying Chaos Theory to
Investment and Economics. Wiley, New York, 1994.
[337] J. Pheby. Methodology and Economics: A Critical Introduction.
MacMillan Press, Houndsmill et al., 1988.
[338] E. S. Phelps. Phillips curves, expectations of inflation, and opti-
mal unemployment over time. Economica, 34(135):254–281, 1967.
[339] A. W. Phillips. The relation between unemployment and the
rate of change of money wages in the United Kingdom 1861-1957.
Economica, 25(100):283–299, 1958.
 References 219

[340] K. R. Popper. The Open Society and Its Enemies, volume 2.

Princeton University Press, Princeton, 2 edition, 1966.
[341] K. R. Popper. Objektive Erkenntnis: Ein evolutionärer Entwurf.
Hoffman and Campe, Hamburg, 3 edition, 1995.
[342] R. R. Prechter. The Major Works of R.N. Elliott. New Classics
Library, New York, 1980.
[343] W. H. Press, S. A. Teukolski, W. T. Vetterling, and B. P. Flan-
nery. Numerical Recipes in C. The Art of Scientific Computing.
Cambridge University Press, Cambridge, UK, 2 edition, 1992.
[344] M. Raberto, S. Cincotti, S. M. Focardi, and M. Marchesi. Agent-
based simulation of a financial market. Physica A, 299:319–327,
2001.
[345] M. Raberto, S. Cincotti, S. M. Focardi, and M. Marchesi. Traders‘
long-run wealth in an artificial financial market. Computational
Economics, 22(2-3):255–272, 2003.
[346] S. Rachev and S. Mittnik. Stable Paretian Models in Finance.
Wiley, Chichester, UK, 2000.
[347] N. J. Radcliffe. Schema processing. In T. Bäck, D. B. Fogel, and
Z. Michalewicz, editors, Handbook of Evolutionary Computation,
chapter B2.5. IOP Publishing Ltd and Oxford University Press,
Oxford, UK, 1997.
[348] M. Ratner and R.-P. C. Leal. Tests of technical trading strate-
gies in the emerging equity markets of Latin America and Asia.
Journal of Banking and Finance, 23(12):1887–1905, 1999.
[349] J. W. S. Rayleigh. The problem of the random walk. Nature,
72:318, August 1905.
[350] M. J. Ready. Profits from technical trading rules. Financial
Management, 31(3):43–61, 2002.
[351] I. Rechenberg. Evolutionsstrategie: Optimierung technischer Sys-
teme nach Prinzipien der technischen Evolution. Frommann-
Holzboog, Stuttgart, 1973.
[352] I. Rechenberg. Evolutionsstrategie ’94. Frommann-Holzboog,
Stuttgart, 1994.
[353] M. W. Reder. Remarks on the ’The methodology of positive
economics’. Journal of Economic Methodology, 10(4):527–530,
2003.
[354] T. Riechmann. Learning and behavioral stability: An economic
interpretation of genetic algithms. Journal of Evolutionary Eco-
nomics, 9(2):225–242, 1999.
220 References

[355] T. Riechmann. Genetic algorithm learning and evolutionary

games. Journal of Economic Dynamics and Control, 25(6-
7):1019–1037, 2001.
[356] T. Riechmann. Learning in Economics: Analysis and Application
of Genetic Algorithms. Physica, Heidelberg, 2001.
[357] A. J. Robson. The evolution of rationality and the Red Queen.
Journal of Economic Theory, 111(1):1–22, 2003.
[358] A. Rogers and A. Prügel-Bennett. Genetic drift in genetic al-
gorithm selection schemes. IEEE Transactions on Evolutionary
Computation, 3(4):298–303, November 1999.
[359] R. Roll. What every ceo should know about scientific progress
in economics: What is known and what remains to be resolved.
Financial Management, 23(3):69–75, 1994.
[360] S. A. Ross. Finance. In J. Eatwell, M. Milgate, and P. Newman,
editors, The New Palgrave: A Dictionary of Economics. W. W.
Norton, New York, London, 1987.
[361] J. B. Rosser Jr. On the complexities of complex economic dy-
namics. Journal of Economic Perspectives, 13(4):169–192, Fall
1999.
[362] M. Rubinstein. Rational markets: Yes or no? The affirmative
case. Financial Analysts Journal, 57(3):15–29, May-Jun 2001.
[363] P. Saacke. Technical analysis and the effectiveness of central
bank intervention. International Journal of Money and Finance,
21(4):459–479, 2002.
[364] L. Samuelson. Evolutionary Games and Equilibirum Selection.
MIT Press Series on Ecvonomic Learning and Social Evolution.
MIT Press, Cambridge, MA, 1997.
[365] P. A. Samuelson. Proof that properly anticipated prices fluctuate
randomly. Industrial Management Review, 6(1):41–50, 1965.
[366] P. A. Samuelson. Summing up on business cycles: Opening ad-
dress. In J. C. Fuhrer and S. Schuh, editors, Beyond Shocks:
What Causes Business Cycles? Federal Reserve Bank of Boston,
Boston, 1998.
[367] P. A. Samuelson and R. M. Solow. The problem of achiev-
ing and maintaining a stable price level: Analytical aspects of
anti-inflation policy. American Economic Review, 50(2):177–194,
1960.
[368] A. M. Santomero and J. J. Seater. The inflation-unemployment
trade-off: A critique of the literature. Journal of Economic Lit-
erature, 16(2):499–544, June 1978.
 References 221

[369] A. Sapusek. Informationseffizienz auf Kapitalmärkten: Konzepte

und empirische Ergebnisse. Gabler, Wiesbaden, 1998.
[370] T. J. Sargent. Macroeconomic Theory. Academic Press, New
York, 1979.
[371] T. J. Sargent. Beyond demand and supply curves in macroe-
conomics. American Economic Review, Papers and Proceedings,
72(2):382–389, May 1982.
[372] T. J. Sargent. Bounded Rationality in Macroeconomics. Claren-
don Press, Oxford, 1993.
[373] J. Sarma and K. D. Jong. Generation gap methods. In T. Bäck,
D. B. Fogel, and Z. Michalewicz, editors, Handbook of Evolution-
ary Computation, chapter C2.7. IOP Publishing Ltd and Oxford
University Press, Oxford, UK, 1997.
[374] L. J. Savage. The Foundations of Statistics. Wiley, New York,
1954.
[375] T. C. Schelling. Models of segregation. American Economic Re-
view, Papers and Proceedings, 59(2):488–493, 1969.
[376] T. C. Schelling. Micromotives and Macrobehavior. W.W. Norton
& Co., New York, London, 1978.
[377] K. H. Schlag. Why imitate, and if so, how? a boundedly rational
approach to multi-armed bandits. Journal of Economic Theory,
78(1):130–156, 1998.
[378] K. H. Schlag. Which one should I imitate? Journal of Mathemat-
ical Economics, 31(4):493–522, 1999.
[379] J. Schmidt. Die Grenzen der Rational Choice Theorie: Eine
kritische theoretische und empirische Studie. Leske + Budrich,
Opladen, 2000.
[380] R. Schmidt and S. Wulff. Zur Entdeckung von Insider-Aktivitäten
am deutschen Aktienmarkt. Zeitschrift für Bankrecht und Bank-
wirtschaft, 5(1):57–68, 1993.
[381] W. Schoutens. Lévy Processes in Finance: Pricing Financial
Derivatives. Wiley, New York, 2003.
[382] J. A. Schumpeter. The Theory of Economic Development. Har-
vard University Press, 1934.
[383] P. Schuster and K. Sigmund. Replicator dynamics. Journal of
Theoretical Biology, 100(3):533–538, 1983.
[384] F. Schweitzer. Modelling migration and economic agglomeration
with active Brownian particles. Advances in Complex Systems,
1(1):11–37, 1998.
222 References

[385] F. Schweitzer. Brownian Agents and Active Particles: Collective

Dynamics in the Natural and Social Sciences. Springer Series in
Synergetics. Springer, Berlin, 2003.
[386] F. Schweitzer and J. A. Holyst. Modelling collective opinion for-
mation by means of active Brownian particles. European Physical
Journal B, 15(4):723–732, 2000.
[387] S. Sehgal and A. Garhyan. Abnormal returns using technical
analysis: The Indian experience. Finance India, 16(1):181–203,
2002.
[388] R. Selten. What is bounded rationality? In G. Gigerenzer and
R. Selten, editors, Bounded Rationality: The Adaptive Toolbox,
chapter 2, pages 13–36. MIT Press, Cambridge, MA, 2001.
[389] W. F. Sharpe. Capital asset prices: A theory of equilibrium under
conditions of risk. Journal of Finance, 19(3):425–442, 1964.
[390] R. J. Shiller. Investor behavior in the October 1987 stock market
crash: Survey evidence. NBER working paper 2446, 1987.
[391] R. J. Shiller. From efficient market theory to behavioral finance.
Cowles Foundation Discussion Paper 1385, 2002.
[392] A. Shleifer. Inefficient Markets : An Introduction to Behavioral
Finance. Oxford University Press, Oxford, New York, 2000.
[393] A. Shleifer and L. H. Summers. The noise trader approach to
finance. Journal of Economic Perspectives, 4(2):19–33, 1990.
[394] H. A. Simon. Models of Man. Wiley, New York, 1957.
[395] H. A. Simon. From substantive to procedural rationality. In S. J.
Latsis, editor, Method and Appraisal in Economics, pages 129–
148. Cambridge University Press, Cambridge, New York, 1976.
[396] H. A. Simon. Rationality as a process and as product of thought.
American Economic Review, 68(2):1–16, 1978.
[397] H. A. Simon. Rational decision making in business organizations.
American Economic Review, 69(4):493–513, 1979.
[398] H. A. Simon. Complex systems: The interplay of organizations
and markets in contemporary society. Computational and Math-
ematical Organization Theory, 7(2):79–85, 2001.
[399] B. F. Skinner. Are theories of learning necessary? Psychological
Review, 57(4):193–216, 1950.
[400] T. Slembeck. Learning in economics: Where do we stand? a
behavioral view on learning in theory, practice and experiments.
University of St. Gallen, VWA Discussion Paper No. 9907, 1999.
[401] T. Slembeck. Evolution und Lernen im Vergleich: Individuelles
Lernen als Voraussetzung für Selbsttransformation und die Be-
deutung von Lernbedingungen. Paper delivered at the Annual
 References 223

Meeting of the Committee for Evolutionary Economics of the

Verein für Socialpolitik, Schloss Reisensburg, July 2000, 2000.
[402] R. E. Smith. Learning classifier systems. In T. Bäck, D. B. Fogel,
and Z. Michalewicz, editors, Handbook of Evolutionary Computa-
tion, chapter B1.5.2. IOP Publishing Ltd and Oxford University
Press, Oxford, UK, 1997.
[403] S. F. Smith. A learning system based on genetic adaptive algo-
rithms. Phd thesis, University of Pittsburgh, Pittsburgh, PA,
1980.
[404] V. L. Smith. An experimental study of competitive market be-
havior. Journal of Political Economy, 70(3):111–137, 1962.
[405] J. Sobel. Economists’ model of learning. Journal of Economic
Theory, 94(2):241–261, 2000.
[406] L. Sögner and H. Mitlöhner. Consistent expectations equilibria
and learning in a stock market. Journal of Economic Dynamics
and Control, 26(2):171–185, 2002.
[407] D. Stauffer, P. M. C. de Oliveira, and A. T. Bernardes. Monte
Carlo simulation of volatility clustering in market model with
herding. International Journal of Theoretical and Applied Fi-
nance, 2(1):83–94, 1999.
[408] D. Stauffer and T. J. P. P. Penna. Crossover in the Cont-
Bouchaud percolation model for market fluctuations. Physica
A, 256:284–290, 1998.
[409] D. Stauffer and D. Sornette. Self-organized percolation model for
stock market fluctuations,. Physica A, 271:496–506, 1999.
[410] M. Steiner and H.-G. Wittkemper. Portfolio optimization with a
neural network implementation of the coherent market hypoth-
esis. European Journal of Operational Research, 100(1):27–40,
1997.
[411] J. E. Stiglitz. The allocation role of the stock market. Journal of
Finance, 36(2):235–251, 1981.
[412] T. M. Stoker. Exact aggregation and generalized slutsky condi-
tions. Journal of Economic Theory, 33:368–377, 1984.
[413] S. Sunder. Markets as artifacts: Aggregate efficiency from zero-
intelligence traders. Yale ICF Working Paper No. 02-16, May
2003.
[414] D. T. Suzuki, A. J. F. Griffiths, J. H. Miller, and R. C. Lewon-
tin. An Introduction to Genetic Analysis. Freeman, New York, 4
edition, 1989.
224 References

[415] G. Syswerda. Uniform crossover in genetic algorithms. In J. D.

Schaffer, editor, Proc. 3rd Int. Conf. on Genetic Algorithms,
pages 2–9. Morgan Kaufmann, San Mateo, CA, 1989.
[416] M. S. Taqqu. Bachelier and his times: A conversation with
Bernard Bru. Finance and Stochastics, 5(1):3–32, 2001.
[417] G. E. Tauchen and M. Pitts. The price variability-volume rela-
tionship on speculative markets. Econometrica, 51(2):485–505,
1983.
[418] N. S. P. Tay and S. C. Linn. Fuzzy inductive reasoning, expec-
tation formation and the behavior of security prices. Journal of
Economic Dynamics and Control, 25(3-4):321–361, 2001.
[419] M. P. Taylor and H. L. Allen. The use of technical analysis in the
foreign exchange market. International Journal of Money and
Finance, 11(3):304–314, 1992.
[420] L. Tesfatsion. Economic agents and markets as emergent phe-
nomena. Proceedings of the National Academy of Science of the
USA, 99:7199–7200, 2002.
[421] L. S. Tesfatsion. Agent-based computational economics: Mod-
elling economies as complex adaptive systems. Information Sci-
ences, 149(4):263–269, 2003.
[422] J. Tirole. On the possibility of speculation under rational expec-
tations. Econometrica, 50(5):1163–1182, 1982.
[423] R. Tobias and C. Hofman. Evaluation of free java-libraries for
social-scientific agent based simulation. Journal of Artificial So-
cieties and Social Simulation, 7(1), 2004.
[424] H. Tong and K. S. Lim. Threshold autoregression, limit cycles
and cyclical data. Journal of the Royal Statistical Society, Series
B (Methodological), 42(3):245–292, 1980.
[425] R. M. Townsend. Forecasting the forecasts of others. Journal of
Political Economy, 91(4):546–588, 1983.
[426] T. Vaga. The coherent market hypothesis. Financial Analysts
Journal, 46(6):36–49, 1990.
[427] L. van Valen. A new evolutionary law. Evolutionary Theory,
1(1):1–30, 1973.
[428] H. R. Varian. Differences of opinion in financial markets. In C. C.
Stone, editor, Financial Risk: Theory, Evidence and Implications,
Proceedings of the 11th Annual Economic Policy Conference of
the Federal Reserve Bank of St. Louis, pages 3–37, Boston, 1988.
Kluwer.
[429] T. Veblen. Why is economics not an evolutionary science? Quar-
terly Journal of Economics, 12(4):373–397, 1898.
 References 225

[430] J. von Neumann and O. Morgenstern. Theory of Games and

Economic Behavior. Princeton University Press, Princeton, 1944.
[431] N. J. Vriend. An illustration of the essential difference between
individual and social learning, and its consequences for compu-
tational analyses. Journal of Economic Dynamics and Control,
24(1):1–19, 2000.
[432] M. Wade. Evolutionary genetics. In E. N. Zalta, editor, The
Stanford Encyclopedia of Philosophy,
http://plato.stanford.edu/entries/evolutionary-genetics/, 2005.
[433] M. M. Waldrop. Complexity: The Emerging Science at the Edge
of Order and Chaos. Touchstone, New York, 1992.
[434] D. A. Waterman. A Guide to Expert Systems. Addison-Wesley,
Reading, MA, 1986.
[435] J. W. Weibull. Evolutionary Game Theory. MIT Press, Cam-
bridge, London, 1996.
[436] U. Wilensky. Netlogo segregation model. Center for Connected
Learning and Computer-Based Modeling, Northwestern Univer-
sity, Evanston, IL., 1998.
[437] M. Wilpert. Künstliche Aktienmarktmodelle auf Basis von
Classifier-Systems. Knapp, Frankfurt/M., 2004.
[438] S. Winter. Economic ’natural selection’ and the theory of the
firm. Yale Economic Essays, 4:225–272, 1964.
[439] U. Witt, editor. Evolutionary Economics, volume 25 of Interna-
tional Library of Critical Writings in Economics. Edward Elgar,
Aldershot, 1993.
[440] U. Witt. What evolutionary economics is all about? In U. Witt,
editor, Evolutionary Economics, volume 25 of International Li-
brary of Critical Writings in Economics, pages xiii–xxvii. Edward
Elgar, Aldershot, 1993.
[441] W. K. Wong, M. Manzur, and B. K. Chew. How rewarding is
technical analysis? Evidence from Singapore stock market. Ap-
plied Financial Economics, 13(7):543–551, 2003.
[442] S. Wright. The roles of mutation, inbreeding, crossbreeding and
selection in evolution. In Proc. 6th Int. Congress on Genetics,
volume 1, pages 356–366, 1932.
[443] J. Yao, C. L. Tan, and H. L. Poh. Neural networks for technical
analysis: A study on KLCI. International Journal of Theoretical
and Applied Finance, 2(2):221–241, 1999.
[444] C. E. Zeeman. On the unstable behavior of the stock exchanges.
Journal of Mathematical Economics, 1(1):39–44, 1974.
Index

Adaptive Market Hypothesis 72–74 classifier system 29, 36, 46–49, 95,
aggregation 10–11 105, 110, 115, 118, 123–125, 129,
allele 36, 44, 96, 150, 155, 157–159, 130, 137, 149, 165, 175, 185
164 Coherent Market Hypothesis 70–71
allele frequency 155, 162 complex adaptive systems 6, 14–16
AR(1)-process 94, 123, 150 conditional heteroskedasticity
ARCH models autoregressive 62
ARCH(q) model 62 consistent expectations equilibria
Exponential ARCH 63 23
Fractionally Integrated ARCH crossover 32, 33, 37, 43–45, 49, 86,
63 99–101, 110, 116, 125, 152, 155,
Generalized ARCH(p,q) 62 167, 169
artificial neural network 30, 36, 71, crowding methods 150
187 CRRA see risk aversion, constant
as-if argument 19–21 relative
aspiration level 22, 26 cybernetics 15

Darwinian gradualism 159

BDS test 107, 122
Darwinism 33
beauty contest 12
deduction 3, 12, 17
behaviorism 29
demarcation criterion 65
dictum for the stock market 53
canonical genetic algorithm 37 diploid 36, 157, 169
CARA see risk aversion, constant Dow Theory 65
absolute
cardinality 36, 45, 46, 48, 96 economic principle 26
catastrophe theory 15 econophysics 58, 79, 80
chaos theory 15, 16, 70 Efficient Market Hypothesis 4, 20,
characteristic exponent 59, 70 51–56, 65, 67, 69, 71, 72, 74, 87,
chromosome 36, 39, 43, 96, 150, 155, 181, 183
157, 164 elitism 33
228 Index

Elliott-wave theory 65 induction 3, 17

emergence 6, 16, 73, 85, 110, 130, Interacting Agent Hypothesis
150, 174 74–75, 82, 85
epistasis 155, 163 Ising model 71, 81–82
evolutionary stable strategy 23
extreme value theory 60 Joseph-effect 60

fallacy of composition 10–11 kurtosis 55–59

fat tails 55, 60, 64, 70, 75, 80
law of iterated expectations 11
fictitious play 31
learning
Fisher-Eigen equation see selection-
Bayesian 30–32
mutation equation
cognitive 30
Fisherian gradualism 159
definitions 29–31
fitness
least square 30
definition of 33
reinforcement 30–31
function 33, 35, 40, 98, 147, 155,
Lotka-Volterra equation 34
157, 164, 165
Lucas-critique 7–9
gradient 163, 178
landscape 40–41 marginalism controversy 19, 34, 35
scaling 40, 42 Marimon-Sargent Hypothesis
sharing 150 92–93, 105, 118, 120, 184
fractal distribution 59 Markov chain analysis 4, 112, 117,
Fractal Market Hypothesis 69–70 156, 171
fractals 70 martingale 11, 53–55, 64, 74, 87
Fundamental Theorem of Natural mimetic contagion 69, 83
Selection 34, 45 mixture of distributions hypothesis
58, 65
generalization 103, 113, 125–127, molecular evolution 157, 159
152–154, 160 Monty Hall’s three-door anomaly
genetic algorithm 29, 32, 36–46, 91, 32
92, 96, 100, 103, 119, 149–151, Multifractal Model of Asset Returns
154–156, 159, 164, 174, 185 63
genetic drift 40, 149, 154–173 multifractal time 63
genetic operator 33 mutation 4, 33, 35, 40, 44–45, 49,
genetic programming 31, 67, 87 86, 91, 93, 99–101, 110–113,
Genoa artificial stock market 82 116–119, 164–173
genome 36
genotype 32, 33, 36 natural rate of unemployment 27
neighborhood segregation 6
haploid 36, 157, 169, 171 neural network see artificial neural
herding behavior 20, 64, 72, 80 network
Heterogeneous Market Hypothesis neutralist-selectionist controversy
75 149, 156–159, 164
heteroskedastic 84 New Institutional Economics 23
Hurst exponent 70 niching methods 150
 Index 229

no-trade theorem 10, 64 replicators 34

Noah-effect 60 representative agent 7–11, 64
rescaled range analysis 70
Occam’s razor 25–26 risk aversion
Ornstein-Uhlenbeck process 94 constant absolute 85, 87, 94, 140,
147, 187
Paretian constant relative 187
distribution 55, 59
hypothesis 58–60 Santa Fe
path dependency 36, 109, 162 Institute 3, 5, 14
percolation model 79–81 perspective of complexity 15
Phillips curve 26–28, 183 satisficing 22
polyploid 36 scaling function 40
population genetics 33, 36, 39, 149, scaling invariance 41–42, 146
154, 157, 160, 169, 185 schema 45–46, 150, 155
power law 58, 80, 84 schema theorem 45–46
predator-prey equation see selection
Lotka-Voltera equation Boltzmann 42
premature convergence 44, 100, fitness proportionate 41
156, 174 natural 19, 21, 34, 72, 155,
Prime Directive 24–25 157–159
operator 33, 41
random communication structure ranking-based 42
79–82, 184 roulette wheel 41, 42, 97, 99, 124,
random genetic drift see genetic 141, 143, 146
drift tournament 42, 99, 100, 160
random walk 53–55, 69–71, 74, 120, selection-mutation equation 35
174, 187 self-affine 70
rational choice theory 22 self-organized criticality 16, 81
Rational Expectations Hypothesis self-similarity 70
12–14, 20, 23, 26, 88 Sewall Wright effect 155, 159
rationality shifting balance theory 40
axioms 21, 22 sigma scaling 42
bounded 22, 23 situational analysis 22–23, 26
economic 22 skewness 55, 57–59
emergent 20 social Darwinism 34
minimal 24–28, 76, 183, 184 stochastic universal sampling 41
perception 21 strategic substitutability 20
preference 21 subpopulation differentiation 160,
procedural 22 181
recombination see crossover sunspots 12
Red Queen principle 74
replicator TAR model see threshold autore-
dynamics 29, 31, 34–36, 45 gressive model
equations 34 threshold autoregressive model 61
230 Index

trading volume 48, 64–65, 81, 107, dependence 75

109, 122 volatility persistence see volatility
trend persistence 70 clustering

volatility Wetzel-ranking 99
clustering 55, 60–64, 74, 75, 82, Wright-Fisher model 157–160, 185
84, 107, 109, 122
conditional 61, 62 zero-intelligence traders 76–79, 184
Lecture Notes in Economics
and Mathematical Systems
For information about Vols. 1–512
please contact your bookseller or Springer-Verlag

Vol. 513: K. Marti, Stochastic Optimization Techniques. Vol. 534: M. Runkel, Environmental and Resource Policy
VIII, 364 pages. 2002. for Consumer Durables. X, 197 pages. 2004.
Vol. 514: S. Wang, Y. Xia, Portfolio and Asset Pricing. XII, Vol. 535: X. Gandibleux, M. Sevaux, K. Sörensen, V. T’kindt
200 pages. 2002. (Eds.), Metaheuristics for Multiobjective Optimisation.
IX, 249 pages. 2004.
Vol. 515: G. Heisig, Planning Stability in Material
Requirements Planning System. XII, 264 pages. 2002. Vol. 536: R. Brüggemann, Model Reduction Methods for
Vector Autoregressive Processes. X, 218 pages. 2004.
Vol. 516: B. Schmid, Pricing Credit Linked Financial
Instruments. X, 246 pages. 2002. Vol. 537: A. Esser, Pricing in (In)Complete Markets. XI,
122 pages, 2004.
Vol. 517: H. I. Meinhardt, Cooperative Decision Making
in Common Pool Situations. VIII, 205 pages. 2002. Vol. 538: S. Kokot, The Econometrics of Sequential Trade
Models. XI, 193 pages. 2004.
Vol. 518: S. Napel, Bilateral Bargaining. VIII, 188 pages.
Vol. 539: N. Hautsch, Modelling Irregularly Spaced
2002.
Financial Data. XII, 291 pages. 2004.
Vol. 519: A. Klose, G. Speranza, L. N. Van Wassenhove
Vol. 540: H. Kraft, Optimal Portfolios with Stochastic
(Eds.), Quantitative Approaches to Distribution Logistics
Interest Rates and Defaultable Assets. X, 173 pages. 2004.
and Supply Chain Management. XIII, 421 pages. 2002.
Vol. 541: G.-y. Chen, X. Huang, X. Yang, Vector
Vol. 520: B. Glaser, Efficiency versus Sustainability in
Optimization. X, 306 pages. 2005.
Dynamic Decision Making. IX, 252 pages. 2002.
Vol. 542: J. Lingens, Union Wage Bargaining and
Vol. 521: R. Cowan, N. Jonard (Eds.), Heterogenous Economic Growth. XIII, 199 pages. 2004.
Agents, Interactions and Economic Performance. XIV,
339 pages. 2003. Vol. 543: C. Benkert, Default Risk in Bond and Credit
Derivatives Markets. IX, 135 pages. 2004.
Vol. 522: C. Neff, Corporate Finance, Innovation, and
Vol. 544: B. Fleischmann, A. Klose, Distribution Logistics.
Strategic Competition. IX, 218 pages. 2003.
X, 284 pages. 2004.
Vol. 523: W.-B. Zhang, A Theory of Interregional
Vol. 545: R. Hafner, Stochastic Implied Volatility. XI, 229
Dynamics. XI, 231 pages. 2003.
pages. 2004.
Vol. 524: M. Frölich, Programme Evaluation and Vol. 546: D. Quadt, Lot-Sizing and Scheduling for Flexible
Treatment Choise. VIII, 191 pages. 2003. Flow Lines. XVIII, 227 pages. 2004.
Vol. 525: S. Spinler, Capacity Reservation for Capital- Vol. 547: M. Wildi, Signal Extraction. XI, 279 pages. 2005.
Intensive Technologies. XVI, 139 pages. 2003.
Vol. 548: D. Kuhn, Generalized Bounds for Convex
Vol. 526: C. F. Daganzo, A Theory of Supply Chains. VIII, Multistage Stochastic Programs. XI, 190 pages. 2005.
123 pages. 2003.
Vol. 549: G. N. Krieg, Kanban-Controlled Manufacturing
Vol. 527: C. E. Metz, Information Dissemination in Systems. IX, 236 pages. 2005.
Currency Crises. XI, 231 pages. 2003.
Vol. 550: T. Lux, S. Reitz, E. Samanidou, Nonlinear
Vol. 528: R. Stolletz, Performance Analysis and Dynamics and Heterogeneous Interacting Agents. XIII,
Optimization of Inbound Call Centers. X, 219 pages. 2003. 327 pages. 2005.
Vol. 529: W. Krabs, S. W. Pickl, Analysis, Controllability Vol. 551: J. Leskow, M. Puchet Anyul, L. F. Punzo, New
and Optimization of Time-Discrete Systems and Tools of Economic Dynamics. XIX, 392 pages. 2005.
Dynamical Games. XII, 187 pages. 2003. Vol. 552: C. Suerie, Time Continuity in Discrete Time
Vol. 530: R. Wapler, Unemployment, Market Structure Models. XVIII, 229 pages. 2005.
and Growth. XXVII, 207 pages. 2003. Vol. 553: B. Mönch, Strategic Trading in Illiquid Markets.
Vol. 531: M. Gallegati, A. Kirman, M. Marsili (Eds.), The XIII, 116 pages. 2005.
Complex Dynamics of Economic Interaction. XV, 402 Vol. 554: R. Foellmi, Consumption Structure and Macro-
pages, 2004. economics. IX, 152 pages. 2005.
Vol. 532: K. Marti, Y. Ermoliev, G. Pflug (Eds.), Dynamic Vol. 555: J. Wenzelburger, Learning in Economic Systems
Stochastic Optimization. VIII, 336 pages. 2004. with Expectations Feedback (planned) 2005.
Vol. 533: G. Dudek, Collaborative Planning in Supply Vol. 556: R. Branzei, D. Dimitrov, S. Tijs, Models in
Chains. X, 234 pages. 2004. Cooperative Game Theory. VIII, 135 pages. 2005.
Vol. 557: S. Barbaro, Equity and Efficiency Considerations Vol. 583: I.V. Konnov, D.T. Luc, A.M. Rubinov† (Eds.),
of Public Higer Education. XII, 128 pages. 2005. Generalized Convexity and Related Topics. IX, 469 pages,
Vol. 558: M. Faliva, M. G. Zoia, Topics in Dynamic Model 2006.
Analysis. X, 144 pages. 2005. Vol. 584: C. Bruun, Adances in Artificial Economics: The
Vol. 559: M. Schulmerich, Real Options Valuation. XVI, Economy as a Complex Dynamic System. XVI, 296 pages,
357 pages. 2005. 2006.

Vol. 560: A. von Schemde, Index and Stability in Bimatrix Vol. 585: R. Pope, J. Leitner, U. Leopold-Wildburger, The
Games. X, 151 pages. 2005. Knowledge Ahead Approach to Risk. XVI, 218 pages, 2007
(planned).
Vol. 561: H. Bobzin, Principles of Network Economics.
XX, 390 pages. 2006. Vol. 586: B.Lebreton, Strategic Closed-Loop Supply Chain
Management. X, 150 pages, 2007 (planned).
Vol. 562: T. Langenberg, Standardization and
Expectations. IX, 132 pages. 2006. Vol. 587: P. N. Baecker, Real Options and Intellectual
Property: Capital Budgeting Under Imperfect Patent
Vol. 563: A. Seeger (Ed.), Recent Advances in Protection. X, 276 pages , 2007.
Optimization. XI, 455 pages. 2006.
Vol. 588: D. Grundel, R. Murphey, P. Panos , O. Prokopyev
Vol. 564: P. Mathieu, B. Beaufils, O. Brandouy (Eds.), (Eds.), Cooperative Systems: Control and Optimization.
Artificial Economics. XIII, 237 pages. 2005. IX, 401 pages , 2007.
Vol. 565: W. Lemke, Term Structure Modeling and Vol. 589: M. Schwind, Dynamic Pricing and Automated
Estimation in a State Space Framework. IX, 224 pages. Resource Allocation for Information Services:
2006. Reinforcement Learning and Combinatorial Auctions.
Vol. 566: M. Genser, A Structural Framework for the XII, 293 pages , 2007.
Pricing of Corporate Securities. XIX, 176 pages. 2006. Vol. 590: S. H. Oda, Developments on Experimental
Vol. 567: A. Namatame, T. Kaizouji, Y. Aruga (Eds.), The Economics: New Approaches to Solving Real-World
Complex Networks of Economic Interactions. XI, 343 Problems. XVI, 262 pages, 2007.
pages. 2006. Vol. 591: M. Lehmann-Waffenschmidt, Economic
Vol. 568: M. Caliendo, Microeconometric Evaluation of Evolution and Equilibrium: Bridging the Gap. VIII, 272
Labour Market Policies. XVII, 258 pages. 2006. pages, 2007.
Vol. 569: L. Neubecker, Strategic Competition in Vol. 592: A. C.-L. Chian, Complex Systems Approach to
Oligopolies with Fluctuating Demand. IX, 233 pages. Economic Dynamics. X, 95 pages, 2007.
2006. Vol. 593: J. Rubart, The Employment Effects of
Vol. 570: J. Woo, The Political Economy of Fiscal Policy. Technological Change: Heterogenous Labor, Wage
X, 169 pages. 2006. Inequality and Unemployment. XII, 209 pages, 2007
Vol. 571: T. Herwig, Market-Conform Valuation of Vol. 594: R. Hübner, Strategic Supply Chain Management
Options. VIII, 104 pages. 2006. in Process Industries: An Application to Specialty
Vol. 572: M. F. Jäkel, Pensionomics. XII, 316 pages. 2006 Chemicals Production Network Design. XII, 243 pages,
2007
Vol. 573: J. Emami Namini, International Trade and
Multinational Activity. X, 159 pages, 2006. Vol. 595: H. Gimpel, Preferences in Negotiations: The
Attachment Effect. XIV, 268 pages, 2007
Vol. 574: R. Kleber, Dynamic Inventory Management in
Reverse Logistics. XII, 181 pages, 2006. Vol. 596: M. Müller-Bungart, Revenue Management
with Flexible Products: Models and Methods for the
Vol. 575: R. Hellermann, Capacity Options for Revenue Broadcasting Industry. XXI, 297 pages, 2007
Management. XV, 199 pages, 2006.
Vol. 597: C. Barz, Risk-Averse Capacity Control in
Vol. 576: J. Zajac, Economics Dynamics, Information and Revenue Management. XIV, 163 pages, 2007
Equilibnum. X, 284 pages, 2006.
Vol. 598: A. Ule, Partner Choice and Cooperation in
Vol. 577: K. Rudolph, Bargaining Power Effects in Networks: Theory and Experimental Evidence. Approx.
Financial Contracting. XVIII, 330 pages, 2006. 200 pages, 2007
Vol. 578: J. Kühn, Optimal Risk-Return Trade-Offs of Vol. 599: A. Consiglio, Artificial Markets Modeling:
Commercial Banks. IX, 149 pages, 2006. Methods and Applications. XV, 277 pages, 2007
Vol. 579: D. Sondermann, Introduction to Stochastic Vol. 600: M. Hickman, P. Mirchandani, S. Voss
Calculus for Finance. X, 136 pages, 2006. (Eds.): Computer-Aided Scheduling of Public Transport.
Vol. 580: S. Seifert, Posted Price Offers in Internet Auction Approx. 424 pages, 2007
Markets. IX, 186 pages, 2006. Vol. 601: D. Radulescu: CGE Models and Capital Income
Vol. 581: K. Marti; Y. Ermoliev; M. Makowsk; G. Pflug Tax Reforms: The Case of a Dual Income Tax for Germany.
(Eds.), Coping with Uncertainty. XIII, 330 pages, 2006. XVI, 168 pages, 2007
Vol. 582: J. Andritzky, Sovereign Default Risks Valuation: Vol. 602: N. Ehrentreich: Agent-Based Modeling: The
Implications of Debt Crises and Bond Restructurings. Santa Fe Institute Artificial Stock Market Model
VIII, 251 pages, 2006. Revisited. XVI, 225 pages, 2008