528 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 17, NO.

4, AUGUST 2013

Evolutionary Foundation of Bounded

Rationality in a Financial Market
Kanta Kinoshita, Kyoko Suzuki, and Tetsuya Shimokawa

Abstract—This paper provides a foundation for decision mak- I. Introduction

ing with bounded rationality in economic entities from the
viewpoint of evolutionary theory. To this end, first, we conducted
an investment test with participants to extract a behavioral
learning model for activities with bounded rationality. We found
R ECENT developments in behavioral economics and neu-
roeconomics have forced a re-evaluation of the con-
ventional concept of rationality used in economics (see, for
that the decision-making model obtained from this behavioral example, [1]–[3]).1
science approach has characteristics that are frequently seen in The traditional economics model assumes that economic
the results of observations of instances of bounded rationality.
Furthermore, the model presents some well-known biases in entities are involved in rational decision making; however,
decision making, such as profit-and-loss asymmetry in risk developments in the areas of behavioral economics and neu-
avoidance, reference point dependence, and the asset effect. roeconomics have led to advancements in the modeling and
Next, using agent-based simulations, we examined whether our identification of bounded rationality in humans by focusing on
behavioral-learning model for activities had the capacity to some of the biases in decision making. Further clarification
become a stable strategy in a market environment where selection
pressure exists. When, in response to maximum loss, a drawdown of these economic behaviors is expected through analyses of
is set as an evaluation criterion for selection, the results of our studies in these areas. This paper will conduct an examination,
simulations imply the following: 1) our decision-making model from an evolutionary perspective, on whether the typical biases
with bounded rationality has the capacity to become a stable in decision making (or bounded rationality) reported in the
evolutionary strategy and 2) entities with bounded rationality studies related to these areas might be the result of economic
can survive in a competitive market. These results are antithetical
to the evolutionary explanations used as a basis for rationality activities in the marketplace.
in traditional economics, and they indicate the possibility that According to the famous statement by Friedman [8], which
many well-known biases in decision making can be derived has been frequently used as a basis for the validity of the
evolutionarily from a single criterion. rationality of economic entities (agents), if an economic entity
Index Terms—Agent-based simulation, artificial market, is not rational, it will eventually lose out to the competition and
evolution, experimental economies, sequential investment task. withdraw from the market, which will lead to the remainder of
the market behaving rationally. According to this Lamarckian
concept, entities that follow rational decision making exist in
Manuscript received September 20, 2009; revised July 14, 2011, October 21, the market as a result of the selection of economic entities by
2011, January 31, 2012, and April 4, 2012; accepted June 23, 2012. Date of the market.
publication July 13, 2012; date of current version July 25, 2013. This work
was supported by JSPS Grant-in-Aid for Scientific Research, under Grant
This paper employs an evolutionary perspective and inves-
21530313. tigates some of the reasons and probable lines for justifying
K. Kinoshita is with Rakuten, Inc., Tokyo 140-0002, Japan (e-mail: bounded rationality.2 Specifically, simulations of a market with
kanta [email protected]).
K. Suzuki is with the Graduate School of Economics, Tokyo University of
evolutionary pressures are conducted to show that an agent
Science, Tokyo 113-0033, Japan (e-mail: [email protected]). with observed biases in decision making survives for a long
T. Shimokawa is with the School of Management, Tokyo University of term under market conditions that are close to reality, and that
Science, Tokyo 346-8512, Japan (e-mail: [email protected]).
This paper has supplementary downloadable multimedia material available
many of the biases in decision making that have been observed
at http://ieeexplore.ieee.org provided by the author. The files contain Tables in the past can be derived from a single evaluation criterion for
V–VII and Appendixes. This material is 372 KB in size. entry into, and withdrawal from, the market. In other words,
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
this paper emphasizes that many of the observed biases in
Digital Object Identifier 10.1109/TEVC.2012.2208465 decision making have certain evolutionary rationalities. This
1 For example, Benartzi and Thaler [4] and Barberis et al. [5] proposed equi-
librium models based on prospect theory and clarified that the loss aversion
feature was effective in explaining certain stylized facts of a financial return, 2 The idea of bounded rationality using evolutionary computation was
such as the equity premium puzzle and the volatility puzzle. Shimokawa supported by Tsang [9]. It is possible to list several works of the literature in
et al. [6] showed that the price process, which is endogenously generated which subjects with bounded rationality survive by presuming some aspects of
through their model with loss-averse traders, not only has consistencies with market imperfection, such as asymmetric information, limitations on trading,
the equity premium puzzle and the volatility puzzle but also has great kurtosis, or myopic decision making. However, the uniqueness of this analysis lies
asymmetry of return distribution, autocorrelation of return volatility, and cross- in its examination of whether decision-making biases obtained from actual
correlation between return volatility and trading volume. We can point out this experiments using behavioral economics methods can be justified based on
kind of rethinking and attempt to surpass traditional theory in various fields the market mechanisms instead of merely showing the possible presence of
of finance (please see the studies summarized by Barberis and Thaler [7]). such biases.
1089-778X/$31.00 
c 2012 IEEE
KINOSHITA et al.: EVOLUTIONARY FOUNDATION OF BOUNDED RATIONALITY IN A FINANCIAL MARKET 529

justify whether it is valid based on the market selection

viewpoints.4
A standard equilibrium model in financial theory by Gross-
man and Stiglitz [29] was used as the market model on the
assumption of the existence of rational entities presumed in
economics, and of entities with bounded rationality whose
decision-making models are based on the behavioral-learning
theory observed in the sequential investment task above.5 We
used an equilibrium model that has typically been adopted in
the field of agent-based simulation [30]; in addition, we used
evolutionary selection pressures that have been a standard for
genetic algorithms.
As described above, the results of this simulation affirm
Fig. 1. Objectives of the agent-based simulation. decision making with bounded rationality and differ from
results obtained by conventional agent-based simulation
result contrasts with the basic concept of rationality established studies. For example, Chen et al. [24] showed that the
by Friedman [8] and relativizes the assumption of rationality market becomes effective by adopting the mean absolute
conventionally used in economics. prediction error (MAPE) as the fitness function in an artificial
First, we conducted a sequential investment task test (with market analysis while using genetic programming. The reason
159 examinees between the ages of 18–32 years) [10]–[12] our result differs from the conventional result is that our
in order to describe the bounded rationality of financial simulation accounts for a drawdown, which has become
markets.3 Using these data as a basis, we performed a sta- analogous with the standards for entry into, and withdrawal
tistical calculation to create a behavioral-learning model from from, the actual market. This research contributes to financial
the responses of the examinees. The behavioral-learning model theory by showing that some of the well-known biases in
was created by developing reinforcement learning, reflecting decision making have certain rationalities that are evident
the profit-and-loss asymmetry established by Bereby-Meyer from an evolutionary perspective under this criterion. This
and Erey [17]. Our model does not identify the influence of paper bridges two important research fields: behavioral
fictive error (addressed by Camerer and Ho [18] and Lohrenz economics and agent-based computational economics.
et al. [10]), although it can be considered a typical learning The remainder of this paper is composed as follows. Sec-
model. tion II specifies the features of bounded rationality using an
In fact, this model is consistent with many of the typical experimental approach. We conducted a behavioral investment
biases observed in decision making and has the following experiment with 159 subjects to identify their decision rule
characteristics: or biases. Section III shows the robustness of the observed
1) adjustable reference point (addressed by Erev and Roth parameters of the bounded rational model in the previous
[15] and Bereby-Meyer and Erev [17]); experiment (Section II) from the view of the agent-based
2) asymmetry of reinforcement learning (related to ev- market simulation with evolutionary pressure. Section IV
idence from neuroscience, e.g., [19] and [20], and indicates the consistency of the agent-based market simulation
prospect theory); with evolutionary pressure with the stylized facts of a financial
3) disposition effect (addressed by Shefrin and Statman time series and the forecastability of returns. Section V reveals
[21] and Odean [22]); market situations where Friedman’s assumption of market ra-
4) market events: deviation, trend reversal, and volatility tionality is broken by using the agent-based market simulation.
change (addressed by Lo and Repin [23]). Section VI discusses the validity of the evolutionary approach
Next, we conducted an agent-based market simulation in a market setting and the limitations of our analysis. The
[6], [24]–[28]. Our objective is to show whether a human final section provides conclusions.
decision-making process derived from the most typical be-
havioral finance methods or models can be justified based II. Modeling Bounded Rationality
on evolutionary viewpoints, such as market selection (Fig. 1).
A. Experiment
Therefore, the research conducted performs experiments first,
and then specifies the decision-making model using meth- First, we conducted a behavioral investment experiment
ods of behavioral finance. The evolutionary simulation is to identify the biases that people have in a decision-making
adopted not to derive the decision-making model but to activity in order to describe decision making with bounded ra-
3 A great deal of research has been done on behavioral-learning theory, 4 The reader should keep in mind that the purpose of performing the

to analyze how an entity selects its behavior on the basis of information evolutionary simulations was not to derive decision-making models, but to
obtained from the environment (see [13] for a good survey on this topic). The evaluate whether they are valid based on the perspectives of the market
reinforcement learning models created by Sutton and Barto [14], Erev and selection process. Therefore, we do not use the evolutionary algorithm to
Roth [15], [16], and Bereby-Meyer [17], along with the experience-weighted select decision-making models.
attraction learning models by Camerer and Ho [18] (which are a hybrid of 5 A behavioral bias is not akin to bounded rationality. However, we believe
reinforcement learning and hypothetical behavior), have their own merits and that facts observed as a behavioral bias can provide us with important
weaknesses. guidelines for modeling bounded rationality in decision making.
530 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 17, NO. 4, AUGUST 2013

Fig. 2. Price history, current investment rate, time remaining for the next price update, and the remaining number of price updates are displayed in the
middle of the screen. When the simulation is started, the price data for 225 periods is displayed, and then the subject makes a decision. At the bottom right of
the screen, the current asset price and the gain up until now are displayed. In this figure, the subject has assets worth 1 025 070 yen, has just won 25 070 yen,
and has 70% of his assets invested in the market. In order to change the investment rate, the subject either pushes the up or down arrow keys on the keyboard
(the up key for increasing the rate and the down key for decreasing it), or clicks the “investment rate” button on the bottom left of the screen display. The
slider bar at the bottom middle of the screen indicates the current investment rate. The investment rate moves in increments of 10%.

tionality. We conducted sequential investment task experiments invested in stocks, and the remaining assets were invested in
by referring to [10]. Many studies have been conducted to deposits. We set the deposit rate at 0%.6 The stock allocation
describe bounded rationality according to behavioral-learning changed depending on fluctuations in stock prices, which were
theory [13]. This paper is also based on the standard model automatically updated every 10 s for 75 iterations. Fig. 2
in this field. The details of the experiment are as follows. shows a typical screen display. When we ran the simulator,
The study used 159 subjects (95 males and 64 females, the price data for 225 periods were displayed, after which
aged 18–32) in experimental sessions conducted at the Tokyo the subject made a decision. In Fig. 2, the subject has assets
University of Science, Tokyo, Japan. All subjects were amateur worth 1 025 070 yen, has just gained 25 070 yen, and has 70%
investors with limited trading experience; professional traders of his assets invested in the market. Each subject performed
were not included. Specifically, 24 subjects had investment this experiment twice with different price data each time. In
experience (19 males, 5 females). Thirteen subjects had invest- our experiment, there were ten price data points that constitute
ment experience of one year or less, 10 had between one and the actual price history. The summary statistics for the price
two years of experience, and 1 had over two years but less than data, method used to assign the price data to the subjects, and
five years. The investment amounts for experienced partici- other relevant information are summarized in Appendix A.
pants averaged about 825 000 yen, and the highest amount was
3 500 000 yen. Each experimental session lasted approximately B. Model
90 min, including instruction time. As a reward, subjects were The most suitable learning model related to the subjects’
offered 1500 yen (about 12 U.S. dollars according to the investment behavior was selected from 22 candidate models
exchange rate at the time of the experiment) as a payment for constructed using six factors considered to have a crucial
participating, and this was adjusted ±1500 yen in accordance impact on decision making under uncertainty using the Akaike
with earnings. Each subject was assigned a computer terminal. information criterion (AIC) and the Bayesian information
Our experiment is similar to that of Lohrenz et al. [10]; criterion (BIC)7 (Fig. 3). The following studies on behavioral
however, there are some differences between their design and finance provide validation and interpretations of characteristic
ours that are described in Appendix A. Our experimental features of the candidate models:
setting is as follows. All traders started out with 1 000 000 1) rational expectation (adopted in traditional financial the-
yen. The traders allocated their funds between stocks and ory);
risk-free assets (i.e., deposits). They chose their investments 2) adjustable reference point (addressed by Erev and Roth
for stocks or the allocation rate of their assets, which they [16] and Bereby-Meyer and Erev [17]);
could change throughout the experiment. A certain amount of 3) fictive error or fictitious investment (addressed by
money was calculated according to their investment rate and Camerer and Ho [18] and Lohrenz et al. [10]);
7 The statistical significance of the difference between the AIC values was
6 Stock price data supplied in the experiment were selected from data determined using the Kishino–Hasegawa (KH) test [31]–[33]. The bootstrap
observed over the past six years in the Tokyo Stock Exchange. A deposit method was used to calculate the statistics of the KH test. The hypothesis
rate of 0% was used because the Japanese market has had a zero interest rate that the AIC value of our model is the same as in other models was rejected
since 2001. with a 99% confidence level.
KINOSHITA et al.: EVOLUTIONARY FOUNDATION OF BOUNDED RATIONALITY IN A FINANCIAL MARKET 531

Fig. 3. Candidate models.

4) asymmetry of reinforcement learning (related to the The other 21 models are introduced in Appendix B. The model
evidence from neuroscience in [19] and [20], and was selected using AIC and BIC. The investment rate at period
prospect theory); t is denoted by i (t), the market price by p (t), market return
5) disposition effect (addressed by Shefrin and Statman by r (t), and invested assets by S (t)
[21] and Odean [22]);
6) market events such as deviation, trend reversal, and i(t) = φ + ρ · i(t − 1) + δ± · (r(t − 1) − r.p.(t − 1)) · S(t − 1)
volatility change (addressed by Lo and Repin [23]). +ω± · unrealized gain(t − 1) + ε (1)
This paper arranged as many models as possible, all of which
are considered to be in the field of behavioral economics, where the reference point, r.p.(t), is defined as
to specify the most consistent decision-making model with r.p.(t) = (1 − γ ± ) · r.p.(t − 1) + γ ± · r(t − 1) (2)
actual observation data. Therefore, each model should be  +
δ , if r(t − 1) ≥ r.p.(t − 1)
justified based upon the studies conducted on behavioral δ± = (3)
δ− , if r(t − 1) < r.p.(t − 1)
economics to date.  +
We calibrated the parameter values of the candidate models ± ω , if unrealized gain(t − 1) ≥ 0
ω = (4)
using the maximum-likelihood method with the assumption ω− , if unrealized gain(t − 1) < 0
 +
of a prediction residual having normal distribution. In this ± γ , if r(t − 1) ≥ 0
γ = (5)
subsection, we will only examine the model with the best fit. γ − , if r(t − 1) < 0
532 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 17, NO. 4, AUGUST 2013

TABLE I TABLE II
Values of the Parameters in the Most Suitable Model Estimated Value and the Estimated Error (in Parentheses) for
Each Parameter Value, Which Is Calculated Using the
φ ρ δ+ δ−
0.1298 0.6687 1.3862 2.9567 Bootstrap Method With 1000 Times Resampling
ω+ ω− γ+ γ− φ ρ δ+ δ−
−0.0994 −1.4109 0.0822 0.0097 Mean 0.1338 0.665 1.5941 2.8808
We can observe the following properties: 1) the adjustment of the reference Standard deviation (0.0153) (0.0324) (0.5218) (0.6114)
point to the positive return is faster and later than the negative return (γ + > ω+ ω− γ+ γ−
γ − ); 2) the reinforcement learning from the excess return, defined as the Mean −0.2461 −1.3016 0.1279 0.0261
previous return minus the reference point of return, is highly sensitive to Standard deviation (0.2905) (0.2746) (0.0609) (0.0641)
the excess loss (δ+ < δ− ); and 3) the disposition effect is relatively strong
and has gain-loss asymmetry (|ω+ | < |ω− |). Statistically, γ − and ω+ cannot be distinguished from 0.

and the unrealized gain rate at period t is calculated as

  sample paths. Table II provides the estimated value and esti-
p(t) − purchase price(t) i(t) · asset(t) mated error (in parentheses) for each parameter value, which
unrealized gain(t) =
purchase price(t) asset(0) is calculated using 1000 bootstrap resampling iterations. The
(6) statistical significance of each parameter estimation can also
be confirmed using this estimation. These tables suggest that
where the term asset(0) is used as the normalized factor and
the features of the decision-making model show the following
the purchase price is composed of the average price.8
characteristics: the adjustable reference point, the asymmetry
In an intuitive sense, the model is a typical reinforcement
of reinforcement learning, and the disposition effect.
learning model that additionally takes the asymmetry of a
reference point and the disposition effect into consideration. In 1) The autocorrelation of decision making can be con-
(1), the second term is an autocorrelation term for investment firmed by the high value of ρ.
and the third term is a reinforcement learning term. With 2) Reinforcement learning from the excess return, defined
the reinforcement learning term, the model learns that if the as the previous return minus the reference point of
return on investment (ROI) received in the current period is return, is highly sensitive to the excess loss (δ+ < δ− ).
greater than the reference ROI, then the investment rate for 3) The disposition effect is relatively strong in the loss
the coming period increases. In decision making, the rate of domain and relatively small in the gain domain. There
return that serves as a reference point for profit and loss is is gain-loss asymmetry (|ω+ | < |ω− |).
determined by weighted moving averages according to (2). 4) The adjustment of the reference point to the positive
The fourth term represents the disposition effect; that is, the return is faster and later than the adjustment to the
investment rate concerning investment assets decreases with negative return (γ + > γ − ). Statistically, γ − and ω+
greater “unrealized” profit (indicating that profits are locked cannot be distinguished from 0.
in). Conversely, investment tends to continue with greater Some of these characteristics are consistent with the results of
“unrealized” loss (indicating the inability to cut losses). Given the psychological experiments on traders carried out by using
(6), “unrealized” profit and loss are calculated as the difference Kahneman and Tversky’s [1] value function in prospect theory
between the purchase price and the current price. In addition, and by Lo and Repin [23]. Furthermore, these characteristics
the superscript “'” for individual parameters represents asym- explain the over-reaction and under-reaction often reported in
metrical learning in the areas of profit and loss. As Bereby- price formation in the market [3], [36]–[40], a phenomenon
Meyer and Erev [17] noted, the asymmetry of such learning referred to as the disposition effect [41], [42]. Concerning
substantially increases a model’s descriptiveness.9 the characteristics of market return sequences, the following
features have been confirmed in the relationship between
C. Characteristic Features of the Model financial stylized facts and the decision biases of the bounded
Table I shows the calibrated parameter values of the model rational traders:
generated by applying the maximum likelihood method to ten 1) the return does not follow a normal distribution and has
8 In order to confirm the robustness of the model, we conducted the
great kurtosis;
overfitting test and a χ2 test. We conducted the over-fitting test as follows: 2) there exists an autocorrelation of the volatility of return
1) we separated the samples into two groups: in-sample and out-of-sample, distribution;
and then we used the in-sample data to calibrate the parameter values with 3) the forecastability of returns exists.
maximum log likelihood; and 2) we compared the mean square deviation
of the expected values with that of the actual values for investment behavior Section IV shows these relationships.
within the out-of-sample range of each model. As a result, our model achieved
the highest prediction accuracy among all models. We also conducted a χ2 III. Evolutionary Foundation
test, the result of which verified our conclusion that the model we have chosen
best fits the investment behaviors of the examinees. A. Evolutionary Artificial Market Simulation
9 We can point out some attempts to refine learning models from the
viewpoint of pattern recognition. For example, Shimokawa et al. [12] intended We conducted an agent-based simulation using an artificial
to reform the ordinal reinforcement learning model in order to take into market with selection pressures to discover the conditions
account nonlinear decision making by using a Bayesian three-layer percep- capable of introducing an evolutionary basis into the bounded
tron. Kampouridis et al. added evolutionary decision-making to the market
fraction hypothesis [34] and the dinosaur hypothesis [35]. They implemented rationality model (obtained from the observation above) and
reinforcement learning with genetic programming. to determine whether the model can become a stable market
KINOSHITA et al.: EVOLUTIONARY FOUNDATION OF BOUNDED RATIONALITY IN A FINANCIAL MARKET 533

strategy. Evolutionary approaches to economic and financial Step 3) evaluation of agents according to transaction re-
modeling encourage a re-examination of “rationality,” as it sults in the market;
is assumed in traditional economic and financial models; Step 4) manipulation according to evolutionary theory (se-
furthermore, they have helped to describe “the learning lection, crossover, and mutation);
process of actors in a market.” For example, Chen and Yeh Step 5) repetition of Steps 2–4 G times (here, 1000 gener-
[43], [44] used genetic programming to study traders with ations).
rational expectations and the attainment of market efficiency. The details of these steps are explained as follows.
They also modeled how market participants were retrained Step 1: This market has 100 traders in every generation. The
using strategies from a strategic pool in what they termed traders are divided into rational traders and bounded rational
a “business school” and proposed a more realistic learning traders at a ratio of 3:7 (this condition is eased later). The
process [24]. In addition, Markose et al. [45], Robson [46], chromosomes of the bounded rational traders are set to the
[47], and Jaramillo and Tsang [25] indicated that the “Red parameter values of the model (φ, ρ, δ+ , δ− , ω+ , ω− , γ + , γ − ).
Queen principle,” which states that one must continue learning To examine whether the model obtained from the experiment is
in order to retain one’s current position in a market, is the valid, the initial parameters are given in a normal distribution,
result of coevolution in the market.10 with a mean calibrated from the experiment results shown in
The design of the artificial market was in accordance with Table I and a variance of 1.
a financial market model [6], [26], [28] with typical hetero- Step 2: Trader agents in each generation conduct transac-
geneous trader agents, as in [29]. We assumed that rational tions in the artificial market according to their own decision-
traders and bounded rational traders exist in this artificial making model for a certain period (here, 50 periods11 ), and the
financial market and that the bounded rational traders make prosperity of the next generation is evaluated on the basis of
their decisions according to the behavioral-learning model their transaction results.12 The details of the artificial market,
observed in the previous section. The modeling of the rational the demand function for each type of trader, and the pricing
traders in our study is the same as in [29], which is typically model are described as follows.
used in traditional economic models. Specifically, the current a) Design of the artificial market: The basic economic
artificial market has the following features. structure of this market simulation draws on the tra-
1) Preference: Rational agents feature a constant absolute ditional rational expectation model in which heteroge-
risk aversion (CARA) utility function and determine neous agents have different information with CARA
an investment strategy for each period in order to utility functions [29].
maximize their investments over the long term. In con- We consider a finite horizon (N-period) model. Two
trast, bounded rational agents feature a decision-making assets (risky assets and risk-free assets) exist in this
scheme (1)–(6), as was actually observed (and described economy. For simplicity, the dividend rate and the risk-
in Section II-B). Parameters for this scheme are learned free rate rf are assumed to be 0. The fundamental value
coevolutionally. of the risky assets, denoted by v, is settled at period
2) Price determination: Market prices for each period 0 and cleared with the value in period N + 1. The
are determined by the supply and demand of market value v is determined according to a normal distribution
participants meeting (the market-clearing condition). A N(v̄, 1/τv ). All the traders cannot directly observe value
Walrasian price adjustment process is not assumed ex- v; they infer the fundamental value from the available
ogenously. information, past prices, and private signals at each
3) Information representation: In accordance with [29], period. Following the tradition of heterogeneous agent
information concerning the true value of assets is pro- models, our model is assumed to have rational traders
vided along with noise to rational agents. This informa- (informed traders) and bounded rational traders who use
tion is communicated to market participants via market our learning model (1) to determine investments.
prices. In each period, some agents are selected from trader
4) Social learning: Individual parameters for the decision- groups to participate in the trade market. To simplify the
making scheme (1) are learned using a genetic algorithm analysis, we suppose that the number of market partici-
that causes the bounded rational agents to attempt to pants in both trader groups is constant in every period.
reduce their own losses. This learning affects market We denote the number of rational and irrational traders
price formation and consequently affects the learning of as P R and P B , respectively. Each trader is assumed to
other actors. Such social learning occurs coevolutionally. 11 The evaluation period (50 periods) is determined ad hoc. There is no

We conducted simulations for markets with selection under corroboration for this criterion based on observations, but the effects of
relaxing this assumption are shown in Fig. 12. Generally, decision-making
the following procedures: bias is more apparent when the criteria for evaluation are more short-term
Step 1) initialization; (i.e., investors make decisions more myopically).
12 Here, the market price is determined endogenously. This means that
Step 2) transaction in the market for a certain period;
the strategies of all the traders have mutual effects on their performances.
Coevolution represents the process wherein the strategies of each agent have
10 Our study does not aim to re-evaluate rationality, but instead aims to an influence on the others; consequently, there is a change in their decisions
examine whether decision-making biases obtained from actual experiments, on issues such as fitness and the adaptive landscape. Kauffman [48] pointed
using behavioral economics methods, can be justified from market mechanism out that coevolution applies to economic literature. Our simulation is also a
perspectives. process of coevolution.
534 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 17, NO. 4, AUGUST 2013

be selected and to have participated in trading only once c) Modeling of the bounded rational traders: The bounded
in each period, although he or she can observe the past rational traders are assumed to have observable de-
price sequence of the risky assets, and hold the traded cision biases in our experiment, i.e., their investment
assets for N + 1 periods. This setting is similar to the decisions exhibit tendencies such as adjusted reference
short-term investment model in [49], in which investors points, asymmetric response, and disposition effects.
myopically make decisions, although they have a long The bounded rational traders are assumed to make
memory. The timing of the participation in the trade decisions according to our model given in Section II-B.
market is exogenously given; a trader’s strategic decision Simply put, the bounded rational traders are assumed to
about the timing of participation is not considered in this be homogeneous and their holding assets are normalized
paper. Let the rational traders have demand functions as 1. Given that the investment rate is i(t), the bounded
corresponding to their limit prices. The price in each rational traders’ demand is
period is determined such that demand and supply can
be matched. xiB (t) = i(t) (10)
b) Modeling of the rational traders: In any period t, the
rational traders predict the values of their assets using where i(t), which is the behavior model obtained in
private signals and the past price sequence P t−1 = Section II-B, is defined as
(p(0), p(1), ..., p(t − 1)), and then they determine their
i(t) = φ + ρ · i(t − 1) + δ± · (r(t − 1)
demand so as to maximize expected utility, given a
limit order price. Traders offer a demand function corre- −r.p.(t − 1)) · S(t − 1)
sponding to each limit price. The rational traders receive +ω± · unrealized gain(t − 1) + ε (11)
different signals. The private signals that trader i receives
in period t are expressed as si (t) = v +
s (t), where where r.p.(t), which is the reference point, is defined as

s (t) represents the white noise of a signal following the
normal distribution N(0, 1/τ
). When new information is r.p.(t) = (1 − γ ± ) · r.p.(t − 1) + γ ± · r(t − 1). (12)
gained, the traders update their predictions with respect
The unrealized gain rate at period t is calculated as
to the asset value v using Bayes’s rule. Since it is
assumed that the investors trade in the market only once, unrealized gain(t) =
their decision becomes similar to a static optimization  
p(t) − purchase price(t) i(t) · asset(t)
problem. More precisely, as shown below, the expected . (13)
value of the exponential utility function is maximized purchase(t) asset(0)
under the usable information in the current period. The Since the rational traders are informationally superior to
initial wealth possessed at t by a rational trader is the bounded rational traders and our model is supposed
represented as Wi (t), the value of the risk-free assets to be common knowledge, the bounded rational traders’
held at t is represented as Si (t), and the value of the decision does not create informational problems for the
risky assets held at t is represented as xiR (t). The rational rational traders’ decision problems.
trader’s optimization problem is d) Endogenous determination of the equilibrium price by

Max E{− exp[−ηWi (t)] si (t), P t−1 } the market: The equilibrium price is determined at
R
xi (t) (7) the point where the supply of assets matches demand.
s.t. Wi (t) = Si (t) + P 0 (t)xi (t) Hence, the equilibrium price is determined in period t
where P 0 (t) is the limit price given at the time of so as to satisfy
decision making and η is the degree of risk aversion.
 
R B
P P
Under the assumption that the risk-free rate is 0, the xiR (t)(p(t)) + xiB (t) = 0. (14)
term-end assets are given as i i

Wi (t) = (1 + rf )N−t Si (t) + vxi (t) Proposition:

= Si (t) + vxi (t) (8) Under these circumstances, the equilibrium price is
determined as follows:
since risky assets should be liquidated by the fundamen-
tal value v. To solve this problem, we first rewrite the 
t−1
τε s̃(t) + τv v̄ + τε P R s̃(j)  
trader’s demand xiR (t) given a limit price P 0 (t) at t as j=1 PB i(t)
 p(t) = + η (15)
E{v si (t), P t−1 } − P 0 (t) τt PR τt
xiR (t) =  (9)
ηVar{v si (t), P t−1 }

PR
 where τt = τ
+ τv + (t − 1)P R τ
and s̃(t) = s(t)
PR
.
where E{v si (t), P t−1 } represents the conditional
 expec- i
Please see Appendix C or refer to [6] and [11] for a
tation of v by trader i in period t, and Var{v si (t), P t−1 }
verification of this equilibrium price formula.
represents the conditional variance.13
In the following simulations, the parameter values v̄ =
13 We referenced the introduction of [50] to derive the demand function. 100, τv = 0.01, τ
= 0.004, and η = 2 are adopted.
KINOSHITA et al.: EVOLUTIONARY FOUNDATION OF BOUNDED RATIONALITY IN A FINANCIAL MARKET 535

Fig. 4. Main result: histogram of the parameter values of the agents in the last generation.

Step 3: As a result of the market transactions described This selection strategy implies that the trader whose drawdown
above, the value of the assets possessed by each trader is is very large should leave the market and it provides a
endogenously determined. natural analogy for the selection pressure in the real market.
To characterize each bounded rational agent, the chromo- Professional traders would be selected using this kind of
some randomly assigned in Step 1, which is composed of evaluation in the real world. The crossover rate is set to 0.15
the parameter values of the investment function, is used. To and the mutation rate is set to 0.01.
assess a trader’s fitness, the performance of the trader in the
artificial market defined in Step 2 during the evaluation period B. Results of the Simulation
is adopted. We use a drawdown as a function for selection. 1) Primary Result: We conducted a simulation of ar-
Here, the drawdown implies the minimum amount of assets tificial markets according to the parameter values in Table
held (i.e., the maximum amount of loss) within the evaluation III. Fig. 4 shows the distribution of each chromosome of
period. This evaluation standard forces a trader to withdraw the bounded rational traders in the last generation of the
from the market once he or she makes a huge loss, despite the simulations. The horizontal axis indicates the parameter values
large profits he or she may have made in the past. For example, and the vertical axis shows the corresponding number of agents
general traders are forced to withdraw from the market under in each histogram; the total number of agents was 14 000
their own credit restrictions if their loss exceeds a certain (100 × 0.7 agents × 200 simulations).
amount, while professional traders usually have limits on the Table IV shows the average and deviation of each histogram.
loss allowable over a certain period of time. The drawdown By comparing the results of actual trading experiments in Sec-
has become analogous with the standard for entry into and tion II-B with these results, it is clear that all the characteristics
withdrawal from the actual market.14 Moreover, the drawdown match. Therefore, the decision-making model of (1) can be
represents a situation that our evolutionary analogy defines as a represented as a stable strategy endogenously derived from
fatal situation. The use of this kind of evaluation condition is a the market simulations, and the bounded rationality obtained
feature of this paper. As we demonstrate later, this evaluation from the experiments can be said to have certain kinds of
method can define many of the biases in decision making, evolutionary rationality.
explaining them as results of evolutionary selection. This is because the use of a drawdown in the evaluation
Step 4: Following the evaluation, the next generation rules promotes the survival of traders with decision-making
(offspring) is generated by applying three operators: selection, biases. Such biases are represented by the following three
two-point crossover, and mutation. characteristics of (1):
We adopt an elite selection strategy, in which genes ranked 1) δ+ < δ− ;
in the top 70% (we refer to this factor as the “eselection 2) |ω+ | < |ω− |;
rate”) of the evaluation are selected for the next generation. 3) γ + > γ − .
14 De Long et al. [51] and Shleifer and Vishny [52] stated that more realistic
First, δ+ < δ− indicates the tendency to reduce investment to
evaluation standards than those set above can block arbitrage and consequently avoid risk when the price falls below a certain reference point.
affect the distortions of the market price. More specifically, the first characteristic shows the existence
536 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 17, NO. 4, AUGUST 2013

Fig. 5. Independence on the initial distribution. These figures show the relation between the number of generations and the parameter values. The upper row
represents the mean estimate values of parameters, and the lower row shows the variance estimate values of parameters.

TABLE III
Parameters Used in the Above Simulation

Name of the Parameter Explanation Value

S Number of simulations 200
G Number of generations 1000
N Number of agents 100
P B /N Ratio of the bounded rational traders 0.7
T Evaluation period 50
Probs Selection probability (rate) 0.7
Probc Crossover probability (rate) 0.15
Probm Mutation probability (rate) 0.01
Asset(1) Initial asset value 1
v̄ Average of the true fundamental values of the risky assets 100
τv Accuracy of the true fundamental values of the risky assets 0.01
τ
Accuracy of private signal noise 0.004
η Degree of risk aversion 2
φ, ρ, δ+ , δ− , ω+ , ω− , γ + , γ − Initial parameter values Normal distribution of Distribution 1 whose
average is set to the values in Table I

TABLE IV
Results of the Simulations: The Average and the Distributions of the Parameter Values

φ ρ δ+ δ− ω+ ω− γ+ γ−
Mean 0.1774 0.6913 1.3630 2.9963 −0.1416 −1.4529 0.0806 0.0645
Standard deviation 0.9432 0.9529 0.9680 0.9638 0.8766 0.9459 0.9382 1.1038

of asymmetry in the risk aversion measurements in profits falling prices. Tversky and Kahneman [1] reported that
and losses, on the basis of the reference point in addition subjectivity to the value of the gain was 2–2.5 times greater
to the dependence on the reference point. The reason these than subjectivity to the value of the loss. Furthermore, in
characteristics are derived under the evaluation criterion using our simulation, the sensitivity (which is represented by the
a drawdown is because the biases in decision making help parameter delta) to a loss is approximately 2.2 times larger
decrease the maximum loss through excessive risk avoidance than the sensitivity to a gain. As shown in δ+ < δ− in
behavior in the area where loss is generated. (1), the quick adjustment of reference points in a period of
Next, γ + > γ − , which is the reference point factor, rising prices immediately reduces the influence of the price
represents the tendency for the quick adjustment of reference increase on the investment ratio, while the slow adjustment
points in a period of rising prices. Furthermore, it represents of reference points in a period of falling prices results in the
the slow adjustment of reference points in a period of continued influence of the price decline on the investment
KINOSHITA et al.: EVOLUTIONARY FOUNDATION OF BOUNDED RATIONALITY IN A FINANCIAL MARKET 537

Fig. 6. Market price sequence of the benchmark case.

ratio for a relatively long period of time. These characteristics number of generations and the parameter values. The vertical
promote a careful attitude toward a price decline, which axes of the graphs on the top represent the average parameter
leads to a smaller drawdown. In other words, when prices values mentioned in the captions, while the vertical axes of
are below the reference point, a number of risk-avoidance the graphs on the bottom represent the standard deviation of
behaviors set in, as seen in the characteristics of δ+ < δ− . the parameters mentioned in the captions. The horizontal axes
Finally, |ω+ | < |ω− | represents the tendency to avoid risks indicate the number of generations.15
when there are latent profits (unrealized returns), and the The results indicate that the average parameter values are
tendency to prefer risk when there are latent losses (unrealized constant during the simulations and that variance values other
losses). In other words, these are disposition effects that than the γ − values converge to the mean estimate value.
promote the disposal of assets with hidden profits (unrealized This is consistent with our experiment, where the γ − values
gains), since the disposal of risky assets with latent losses could not be distinguished from zero. These results confirm
cannot be implemented. The reason why this tendency is that the characteristics obtained from our experiments can
consistent with the drawdown is that a trader who has obtained provide a stable solution. As mentioned above, it is clear
latent profits at an early stage tries to suppress the drawdown that the characteristics of the investment behaviors obtained
by assuming risk-avoidance behaviors in the later stages in from our experiments can be used to develop stable strategies
order to survive; furthermore, a trader who has suffered for the adaptive landscape in the market environment we
latent losses and who intends to launch a come-from-behind assumed.
bid for victory with one last effort has a better chance of
survival. For these reasons, our decision-making model (1) is
expected to emerge from our market simulations as a stable, IV. Statistical Features of the Equilibrium
endogenously derived solution. This implies that evolutionary Price Sequence
simulation with a drawdown as an evaluation criterion can
A. Financial Stylized Facts
explain dependence on reference points, asymmetry in the
measure of risk aversion, and asset effects, which are the The price sequence of our model has the following features.
representative biases in prospect theory. 1) When the rational traders dominate the market: The
2) Robustness Test: In evolutionary simulations, such as first term of the market price (15), which represents a
genetic algorithm (GA), the dependence of the results on the weighted average of information gain, converges with
initial values is always a matter of concern. In this section, we the fundamental value v and the second term, which
conduct two types of examinations to check the robustness represents the effect of the bounded rational agents to
of the results in Section III-B1. We increase the number the price, goes to 0 as information is accumulated or
of generations and observe the changes in order to further precision rises. Thus, the market price converges to the
examine dependence on the initial distribution. If the results
15 The parameters are determined based on the mean value of a stationary
depend on the initial values, a transition to other reasonable
distribution. To confirm the robustness of the main results, simulations with
and more stable solutions should be observed as the number of 1000 generations are performed and the Geweke’s convergence test [53] is
generations increase. Fig. 5 shows the relationship between the applied to the latter 500 generations to identify the convergence.
538 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 17, NO. 4, AUGUST 2013

Fig. 7. Market return sequence of the benchmark case.

Fig. 8. Autocorrelation function on volatility of returns: the horizontal axis represents lag order and the lines at plus and minus 0.09 represent 5% confidence
interval with the presupposition that the process of returns follows the MA process of lag order.

fundamental value, and if noise traders are considered, that the process of returns follows the MA process of
the price sequence completely becomes a random walk. lag order. The autocorrelation of volatility of return
2) When the bounded rational traders survive (or under seems to continue for the long term. By using Ljung
the convergence process of case 1): In this case, the and Box’s Q-statics test, we confirmed the existence
decision biases of the bounded rational traders have of autocorrelation of volatility with first, fourth, eighth,
some influence on market price formation. Figs. 6 and and 12th lags, respectively (5% significance level). The
7 represent sample sequences of price and return in the autocorrelation of volatility is consistent with previously
benchmark case, respectively (the parameter values of observed facts [54], [55]. Model selection using AIC
investment are defined in Table IV, and other parameter shows that the GARCH model where an autocorrelation
values are defined in Table III). Table V16 shows the of the volatility of return distribution is allowed, is
moments of return distribution. Obviously, the return most suitable for describing the return sequence. Of
does not follow a normal distribution and has great the ARMA, GARCH, and EGARCH models [56], [57],
kurtosis. This is because the boundedly rational decision GARCH(1, 1) is the most suitable for our purposes.
defined by (1) has strong autocorrelation, which causes
traders to overshoot the market price. Fig. 8 represents B. Forecastability of Returns
the autocorrelation function of return volatility. The The features of the bounded rationality of market partic-
vertical axis represents lag order and a solid line repre- ipants can enable the forecasting of market returns that are
sents a 5% confidence interval with the presupposition frequently observed in empirical analysis.
16 The reader can find Tables V–VII in a supplementary file, which is Fig. 9 shows the average response (1000 simulations) of
available for download through the IEEExplore website. market price to fundamental shocks and displays changes
KINOSHITA et al.: EVOLUTIONARY FOUNDATION OF BOUNDED RATIONALITY IN A FINANCIAL MARKET 539

Fig. 9. Responses of the market price and changes in each trader’s investment rates to the shocks in which the fundamental value changes by 10% at 30
periods.

in each trader’s investment rates, which are generated from The lower row of each column displays the changes in
the above price formation and learning model, to the shocks investment rates to the exogenous fundamental shock. The
wherein the fundamental value changes by 10% over 30 solid line represents the reaction of the bounded rational
periods. The left column of the figure represents a positive investor and the dotted line represents the reaction of the
10% shock (good news) and the right column represents a rational investor. We can observe that the rational investor
negative 10% shock (bad news). changes his or her investment rate to absorb the bounded
The upper row of each column represents the responses of rational behavior of the bounded rational traders. For example,
the market price. The broken line represents the case where if a negative shock causes bounded rational traders to panic
only the rational investor exists, i.e., the population ratio and remain extremely passive, the reduction of investment by
between rational investors and bounded rational investors is the bounded rational traders indicates an overreaction to the
1:0. It is observed that the market price gradually converges market price. Thus, rational traders increase their investment
to the new fundamental value as the rational investor’s private rate as a result.
information is accumulated and the information is transmitted The predictability of security prices has been one of
among investors through the price. The dotted line represents the most important subjects in financial research. In recent
a case where the rational traders and bounded rational traders decades, experimental studies in this field have reported the
in our learning model exist at a ratio of 0.75:0.25. In this existence of the predictability of stock prices and denied the
case, an overreaction of price (compared with the dotted line) efficient market hypothesis. There are two types of predictabil-
occurs in the short term and an underreaction is observed ities: contrarian, which is based on the investor overreaction
in the medium and long terms. This underreaction depends hypothesis, and momentum, which is based on underreaction
on the adjustment of the reference point and the disposition [36]–[40].
effect in the learning model. The solid line provides price In our learning model, overreaction is caused in the rel-
changes when the population ratio is 0.3:0.7 (as in Table III). ative short term and underreaction is observed in the mid
Consequently, it can be observed that the abovementioned to long term. Moreover, overreaction is remarkable when a
tendency becomes much stronger. Moreover, a fluctuation is price falls, as is underreaction when the price rises, owing
caused. The overreaction is especially remarkable in the case to the asymmetrical feature of reinforcement learning (i.e.,
of a negative shock. investment rate is more sensitive to a loss than to a gain). In
In the long run, the market price settles to a fundamental addition, the asymmetrical adjustment of the reference point
value in each case. This is because rational investors gradually (i.e., adjustment of a reference point is much faster for a
adjust the distortion caused by the decision making of the gain than for a loss) creates crucial decision-making biases
bounded rational investor. in our model. Thus, it can be expected that stock prices
540 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 17, NO. 4, AUGUST 2013

Fig. 10. Changes in the proportion of rational traders and bounded rational traders during simulations.

Fig. 11. Decreasing informational superiority of the rational traders increases the bounded rational traders. The vertical axis shows the survival probability
of the bounded rational traders, and the horizontal axis represents the precision of the private signal of the rational trader.

will increase comparatively calmly during rising periods and an extremely favourable setting for rational traders. This is
decrease rapidly and considerably during falling periods. This because the rational traders have information about the true
also means that the momentum strategy is effective when the fundamental values of risky assets, and these values have a
price is rising, and the contrarian strategy is effective when great influence on pricing rules. Therefore, if the bounded
the price is falling.17 rational traders can survive in this market environment, it
would prompt a serious questioning of the basis of the ratio-
V. Emergence of an Irrational Market nality argument established by Friedman [8] This subsection
specifies which situations allow the emergence of an irrational
We examined the survival possibility of bounded rational market and investigates whether Friedman’s assertion holds.
traders by changing the proportion of rational traders and The vertical axes of the two graphs in Fig. 10 indicate the
bounded rational traders in the market. Considering that selec- number of agents and the horizontal axes show the number
tion pressure applies to rational traders, our market provides of generations. The straight line represents bounded rational
17 Grinblatt and Han [41] and Frazzini [42] examined whether or not the traders, while the dotted line represents rational traders. The
tendency of some investors to hold on to their losing stocks, driven by prospect graph on the left depicts the case where rational traders control
theory and mental accounting, creates a spread between a stock’s fundamental the market, while the graph on the right represents the case
value and its equilibrium price, as well as price underreaction to information.
Grinblatt and Han [41] used Fama-MacBeth-type cross-section regression to
where bounded rational traders dominate the market. While
illustrate the possibility that a variable proxying for aggregate unrealized most simulations are akin to those shown in the graph on the
capital gains is the key to generate the profitability of a momentum strategy. left, the cases shown on the right are more similar to reality.
Frazzini [42] also tested whether a disposition effect induces the underreaction
to news and suggested that it may do so, leading to return predictability and
Table VI summarizes the survival probabilities of bounded
a postannouncement price drift (momentum effect). rational agents under three conditions. The table indicates that
KINOSHITA et al.: EVOLUTIONARY FOUNDATION OF BOUNDED RATIONALITY IN A FINANCIAL MARKET 541

Fig. 12. Myopic assessment increases the bounded rational traders. The vertical axis shows the survival probability of the bounded rational traders, and the
horizontal axis represents the evaluation periods.

as the informational advantage of the rational agents is eased, probability of survival as the selection probability increases
as the evaluation period is shortened, and as the selection (since the selection function becomes progressively smaller).
becomes loose, the survival probability of the bounded rational This result may imply that the market becomes more rational
traders increases. with increased competition.
Standard economic theory assumes that the market is strictly The reader may observe that the survival rate of the bounded
rational and that irrational entities immediately withdraw from rational traders is relatively low. However, the low ratio of the
the market. However, our simulation results clearly contradict bounded rational traders has a sufficiently large effect on mar-
this hypothesis. We believe that an agent-based simulation in ket price formation. Table VII shows the relationship between
the market is an effective approach to determine which market the ratio of the bounded rational agents and the distortions of
situation allows bounded rational agents to survive. the market returns, which are endogenously generated by the
Our setting is extremely favourable for rational traders, as market simulation. We can see that the market return distribu-
they have advance information about the true fundamental tion deviates sufficiently from the normal distribution, which
value of assets. Fig. 11 shows the reduction of the information emerges when there is no bounded rational agent; this occurs
advantage and the survival probability of bounded rational even if the ratio of the bounded rational agent is under 5%.
traders in relation to the precision of the private signal, τε .
The figure indicates that the survival probability of bounded
rational traders rises greatly as the information superiority of VI. Discussion and Limitations
the rational traders decreases. A. Evolutionary Approach in a Market Setting
The vertical axis of Fig. 12 shows the survival probabilities The reader may question the scientific importance of justify-
of the bounded rational traders when the evaluation period is ing bounded rationality and decision biases in an evolutionary
changed. The horizontal axis represents the evaluation period. sense in a market setting. Such a question is relevant when an
Note that the data includes nine types of selection probabilities evolutionary approach is used to validate economic modeling.
and the initial proportion in order to remove the influence of In the field of economic research, the theory of the evo-
other parameter values.18 The results show that the bounded lutionary approach has been adopted as one of the grounds
rational traders have a progressively better chance for survival from which man’s decision-making rules and market structure
as the evaluation period shortens. This implies that the short- are generated. This paper also adopts this approach on the
term pursuit of profits in the present helps a large number of grounds of bounded rationality and the fact that Friedman
bounded rational traders survive in the market. used the evolutionary approach to justify market rationality.
The vertical axis of Fig. 13 represents the ratio of the In other words, we adopted the approach to determine which
bounded rational traders who survived the simulations in decision-making tendencies of traders can survive in the mar-
relation to change in the selection rate, while the horizontal ket, which market conditions strengthen the survival capability
axis shows the selection probability. This figure indicates of bounded rationality, and how individual actors develop
that bounded rational traders have an increasingly higher decision-making rules through adaptive learning (thinking).
18 For example, when the evaluation period for the given data is 10, the
Of course, this evolutionary simulation in a market setting is
data includes the simulation results for cases where the selection probability not the same as the evolutionary process in a biological sense;
lies between 0.1 and 0.9 (stepwise count of 0.1). Furthermore, the simulation however, we believe that this simulation is an effective analogy
results for cases where the initial proportions of bounded rational traders lie to analyse decision-making rules or biases in the market.
between 0.1 and 0.9 (stepwise count of 0.1) are also included. In other words,
the data at each point represents the result of 16 200 simulations (nine types of Although almost all existing research specifies the bounded
selection probabilities × nine types of initial proportion × 200 simulations). rationality of economic decisions from an experimental
542 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 17, NO. 4, AUGUST 2013

Fig. 13. Loose selection increases the bounded rational traders. The vertical axis shows the survival probability of the bounded rational traders, and the
horizontal axis represents the selection rate.

approach, it is also necessary to validate decision-making the prices in the experiments should be given endogenously,
rules from the viewpoint of survivability through coevolution and they should take the suitable degree of the aggregate
in a market. In this sense, evolutionary simulation using agent- effects into consideration. Therefore, experimental designs in-
based modeling can play an important role in decision-making volving the double auction should be considered in the future.
research. However, the effects are relatively weak in large markets
There may also be questions about the evolutionary such as financial markets and therefore tend, instead, to be
criteria (“the selection criterion” and “mutation”) used in this too strong with only less than ten participants. It is anticipated
manuscript, as they obviously do not accurately correspond that experiments involving a few people would create strategic
to evolutionary processes in actual markets. Having said that, conditions that make it difficult to estimate a suitable decision-
those criteria are the most acceptable approximation of an making model. An attempt to treat this aggregation effect
adaptive learning process using a traditional genetic algorithm endogenously in our analysis requires a large-scale experiment
process. (estimating simulations with 100 agents) and a large sum of
experiment costs. With such foreseeable limitations, it was felt
B. Generalization of Our Experimental Result
that adopting competitive markets where the prices are given
In our analysis, we do not pay much attention to the exogenously would be more appropriate to approximate the
subjects’ investment experience. Because of this limitation in weak aggregation effects rather than performing the small-
subject selection, we have to consider the generalizability of scale double auction.19
our analysis result for the decision-making model. We believe
that our results can be generalized in cases where the levels of
investment experience differ. We have come to this conclusion D. Conditions in the Context of Real Markets
because, first, the same decision-making model is selected Further discussions are necessary to explain the conditions,
for both subject groups, even when the groups are divided in the context of real markets, where bounded rational agents
according to investment experience. Second, decision-making can gain a sizable population. For instance, Fig. 11 indicates
models in other studies, that comprise a wider range of subject that the ratio of the bounded rational agents can increase up to
groups have shown similar tendencies as those in our research. 22%. This occurs when the accuracy of the information that
For example, the reinforcement learning model including the rational agents obtain declines significantly (this means the
asymmetry was employed by Bereby-Meyer and Erev [17]. information accuracy decreases to one-tenth of the benchmark:
Moreover, Weber and Camerer [58] indicated that the disposi- τε = 0.004 to τε = 0.0004). Realistically, this may be a difficult
tion effect was robustly observed among professional traders. condition. Additional discussions are essential to evaluate the
However, the relationship between investment experience validity of cases where the information accuracy declines
and the approach to decision making is important and requires dramatically as such.
further examination [23]. We would like to address this This paper anticipates cases where rare events such as a
problem in a future study. market crash occur. In such circumstances, we assume that
C. Price Formation obtaining adequate information would generally be difficult
In our analysis, the price sequence is given exogenously 19 Suzuki et al. [59] conducted a market experiment using endogenous
for decision-making experiments using actual stock prices, prices. Here, the price information method is elected using the double auction.
whereas it is decided endogenously in the evolutionary simula- However, this is a small-scale case involving only four market participants.
The result of this experiment has also confirmed that “the asymmetric response
tion. In this regard, readers should keep in mind that the exper- to profit and loss,” “reference point dependency,” and “the disposition effect”
iments and simulation lack consistency. Certainly and ideally, have significant effects on investment behaviors.
KINOSHITA et al.: EVOLUTIONARY FOUNDATION OF BOUNDED RATIONALITY IN A FINANCIAL MARKET 543

and that defining subjective probability rationally would be- For a long time, researchers have been trying to explain the
come challenging. Moreover, if a lack of market liquidity main concept of finance theory from an evolutionary view-
and myopic decision making emerge within markets, the point, and several studies employing evolutionary approaches
bounded rational agents are expected to obtain sufficient sizes have contributed to financial research [61]. These works re-
coupled with such conditions. However, further discussions examine the validity of assumptions that were considered
are required to evaluate the validity of these parameters in to be natural in traditional financial models, and they have
comparison to actual markets. succeeded in replicating various financial facts. Our research
follows this tradition.
VII. Conclusion The reader can find Tables V–VII, and Appendixes in a
supplementary file. The supplementary file is available for
Instinctual human traits, including decision biases, can often
download through the IEEExplore website.
be explained from the viewpoint of evolution. For example,
LeDoux [60] provided the following evolutionary basis for the
activity of the amygdala and our emotional reaction to fear. References
The system is not, strictly speaking, a system that [1] A. Tversky and D. Kahneman, “Advances in prospect theory: Cumulative
representation of uncertainty,” J. Risk Uncertainty, vol. 5, no. 4, pp. 297–
results in the experience of fear. It is a system 323, 1992.
that detects danger and produces responses that [2] R. Thaler, “Mental accounting matters,” J. Behavioral Decision Making,
maximize the probability of surviving a dangerous vol. 12, no. 3, pp. 183–206, 1999.
[3] C. Camerer, J. Kagel, and A. Roth, The Handbook of Experimental
situation in the most beneficial way. It is, in other Economics. Princeton, NJ: Princeton University Press, 1995.
words, a system of defensive behavior. As noted [4] S. Benartzi and R. Thaler, “Myopic loss aversion and the equity premium
above, I believe that emotional behaviors, such as puzzle,” Quart. J. Econ., vol. 110, no. 1, pp. 73–92, Feb. 1995.
[5] N. Barberis, M. Huang, and T. Santos, “Prospect theory and asset prices,”
defensive green behaviors, evolved independent of Quart. J. Econ., vol. 116, pp. 1–53, Feb. 2001.
conscious feeling, and that we should not assume [6] T. Shimokawa, K. Suzuki, and T. Misawa, “An agent-based approach
that animals other than humans feel afraid when to financial stylized facts,” Physica A Statist. Mech. Its Applicat., vol.
379, no. 1, pp. 207–225, 2007.
they are in danger. We should, in other words, take [7] N. Barberis and R. Thaler, “A survey of behavioral finance,” in Hand-
defensive behaviors at face value: they represent book of the Economics of Finance, G. Constantinides, M. Harris, and
the operation of brain systems that have been pro- R. M. Stulz, Eds. Amsterdam, The Netherlands: Elsevier, 2003, ch. 18,
pp. 1053–1128.
grammed by evolution to deal with danger in routine [8] M. Friedman, Essays in Positive Economics. Chicago, IL: Univ. Chicago
ways [60, Section 5, p. 128]. Press, 1953.
This paper provided an evolutionary foundation for the [9] E. Tsang, “Computational intelligence determines effective rationality,”
Int. J. Autom. Comput., vol. 5, no. 1, pp. 63–66, 2008.
decision biases observed in financial markets. First, we con- [10] T. Lohrenz, K. McCabe, C. Camerer, and P. Montague, “Neural signature
ducted experiments and described boundedly rational trade of fictive learning signals in a sequential investment task,” Proc. Natl.
behaviors to provide an evolutionary foundation for the subject Acad. Sci., vol. 104, no. 22, p. 9493, 2007.
[11] T. Shimokawa, K. Suzuki, T. Misawa, and K. Miyagawa, “Predictability
of bounded rationality. The model obtained in this paper of investment behavior from brain information measured by functional
explained issues such as the adjustment of reference points, near-infrared spectroscopy: A Bayesian neural network model,” Neuro-
asymmetry in the profit-and-loss response, and the disposition science, vol. 161, no. 2, pp. 347–358, 2009.
[12] T. Shimokawa, K. Suzuki, and T. Misawa, “Argumented reinforcement
effect; these were consistent with the reports on bounded learning and investment decision making,” Neurocomputing, vol. 72, nos.
rationality that were presented by previous studies. 16–18, pp. 3447–3461, 2009.
Next, we conducted an evolutionary simulation using an [13] C. Camerer and R. Foundation, Behavioral Game Theory: Experiments
in Strategic Interaction. Princeton, NJ: Princeton Univ. Press, 2003.
artificial market model with selection pressures in order to [14] R. Sutton and A. Barto, Reinforcement Learning: An Introduction.
check whether the observed bounded rationality could have Cambridge, MA: MIT Press, 1998.
an evolutionary basis. The results confirmed the following. [15] I. Erev and A. Roth, “On the need for low rationality cognitive game
theory: Reinforcement learning in experimental games with unique
1) The decision biases in investment behaviors obtained mixed strategy equilibria,” in Mimeo. Pittsburgh, PA: Univ. Pittsburgh,
from our experiments can be made into stable strategies 1996.
[16] I. Erev and A. Roth, “Predicting how people play games: Reinforcement
for the adaptive landscape of the market environment learning in experimental games with unique, mixed strategy equilibria,”
presented in this paper. In other words, these decision Amer. Econ. Rev., vol. 88, no. 4, pp. 848–881, 1998.
biases were evolutionarily rational. [17] Y. Bereby-Meyer and I. Erev, “On learning to become a successful loser:
A comparison of alternative abstractions of learning processes in the loss
2) The observed decision biases were consistent with well- domain,” J. Math. Psychol., vol. 42, nos. 2–3, pp. 266–286, 1998.
known biases such as dependence on reference points, [18] C. Camerer and T. Ho, “Experience-weighted attraction learning in
asymmetry in the measurement of risk aversion, and the games: A unifying approach,” Econometrica, vol. 67, no. 4, pp. 827–
874, 1999.
asset effect. Furthermore, these biases can be derived [19] H. Breiter, I. Aharon, D. Kahneman, A. Dale, and P. Shizgal, “Functional
using a drawdown as the single evaluation criterion from imaging of neural responses to expectancy and experience of monetary
an evolutionary viewpoint. This drawdown is analogous gains and losses,” Neuron, vol. 30, no. 2, pp. 619–639, 2001.
[20] B. Knutson and R. Peterson, “Neurally reconstructing expected utility,”
to the evaluation standards for entry and withdrawal in Games Econ. Behavior, vol. 52, no. 2, pp. 305–315, 2005.
the actual market. [21] H. Shefrin and M. Statman, “The disposition to sell winners too early
3) These results were reinforced by the informational in- and ride losers too long: Theory and evidence,” J. Finance, vol. 40,
no. 3, pp. 777–790, Jul. 1985.
feriority of the rational agent, the myopic evaluation of [22] T. Odean, “Volume, volatility, price, and profit when all traders are above
traders, and loose selection. average,” J. Finance, vol. 53, no. 6, pp. 1887–1934, Dec. 1998.
544 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 17, NO. 4, AUGUST 2013

[23] A. Lo and D. Repin, “The psychophysiology of real-time financial risk [49] X. Vives, “The speed of information revelation in a financial market
processing,” J. Cognitive Neurosci., vol. 14, no. 3, pp. 323–339, 2002. mechanism,” J. Econ. Theory, vol. 67, no. 1, pp. 178–204, 1995.
[24] S. Chen and C. Yeh, “Evolving traders and the business school with [50] C. Chamley, Rational Herds: Economic Models of Social Learning.
genetic programming: A new architecture of the agent-based artificial Cambridge, U.K.: Cambridge Univ. Press, 2003.
stock market,” J. Econ. Dynamics Control, vol. 25, nos. 3–4, pp. 363– [51] J. De Long, A. Shleifer, L. Summers, and R. Waldmann, “Noise trader
393, 2001. risk in financial markets,” J. Political Econ., vol. 98, no. 4, pp. 703–738,
[25] S. Martinez-Jaramillo and E. Tsang, “An heterogeneous, endogenous 1990.
and coevolutionary GP-based financial market,” IEEE Trans. Evol. [52] A. Shleifer and R. Vishny, “The limits of arbitrage,” J. Finance, vol. 52,
Comput., vol. 13, no. 1, pp. 33–55, Feb. 2009. no. 1, pp. 35–55, 1997.
[26] B. LeBaron, “Agent-based computational finance,” in Handbook of [53] J. Geweke, “Evaluating the accuracy of sampling-based approaches to
Computational Economics, vol. 2. Amsterdam, The Netherlands: the calculation of posterior moments,” in Bayesian Statistics, vol. 4.
Elsevier/North-Holland, 2006, pp. 1187–1233. Oxford, U.K.: Oxford Univ. Press, 1992, pp. 169–193.
[27] C. Hommes, “Heterogeneous agent models in economics and finance,” [54] Z. Ding and C. Granger, “Modeling volatility persistence of speculative
in Handbook of Computational Economics, vol. 2. Amsterdam, The returns: A new approach,” J. Econometrics, vol. 73, no. 1, pp. 185–215,
Netherlands: Elsevier/North-Holland, 2006, pp. 1109–1186. 1996.
[28] K. Suzuki, T. Shimokawa, and T. Misawa, “Agent-based approach to [55] S. Taylor, “Modeling stochastic volatility: A review and comparative
option pricing anomalies,” IEEE Trans. Evol. Comput., vol. 13, no. 5, study,” Math. Finance, vol. 4, no. 2, pp. 183–204, 1994.
pp. 959–972, Oct. 2009. [56] R. F. Engle, “Autoregressive conditional heteroscedasticity with esti-
[29] S. Grossman and J. Stiglitz, “On the impossibility of informationally mates of variance of united kingdom inflation,” Econometrica, vol. 50,
efficient markets,” Amer. Econ. Rev., vol. 70, no. 3, pp. 393–408, Jun. no. 4, pp. 987–1008, 1982.
1980. [57] T. Bollerslev, “Generalized autorregressive conditional heteroskedastic-
[30] B. LeBaron, W. Arthur, and R. Palmer, “Time series properties of an ity,” J. Econometrics, vol. 31, no. 3, pp. 307–327, 1986.
artificial stock market,” J. Econ. Dyn. Control, vol. 23, nos. 9–10, pp. [58] M. Weber and C. Camerer, “The disposition effect in securities trading:
1487–1516, 1999. An experimental analysis,” J. Econ. Behavior Organization, vol. 33,
[31] H. Linhart, “A test whether two AICs differ significantly,” South Afr. no. 2, pp. 167–184, 1998.
Stat. J., vol. 22, no. 1, pp. 153–161, 1988. [59] K. Suzuki, K. Kinoshita, K. Miyagawa, S. Shiomi, T. Misawa, and
[32] H. Kishino and M. Hasegawa, “Evaluation of the maximum likelihood T. Shimokawa, “An investment behavior analysis using by brain com-
estimate of the evolutionary tree topologies from DNA sequence data, puter interface,” Trans. Japan. Soc. Artif. Intell., vol. 25, no. 1, pp.
and the branching order in hominoidea,” J. Molecular Evol., vol. 29, 183–195, 2010.
no. 2, pp. 170–179, 1989. [60] J. LeDoux, The Emotional Brain. New York: Simon and Schuster, 1996.
[33] H. Shimodaira, “Assessing the error probability of the model selection [61] E. Tsang and S. Martinez-Jaramillo, “Computational finance,” IEEE
test,” Ann. Instit. Statist. Math., vol. 49, no. 3, pp. 395–410, 1997. Comput. Intell. Soc. Newsletter, pp. 3–8, Aug. 2004.
[34] M. Kampouridis, S. Chen, and E. Tsang, “The market fraction hypoth-
esis under different genetic programming algorithms,” in Information
Systems for Global Financial Markets: Emerging Developments and
Effects. Pennsylvania: GIG Global, 2011, pp. 37–54.
[35] M. Kampouridis, S. Chen, and E. Tsang, “Testing the dinosaur hypoth- Kanta Kinoshita received the Masters degree in
esis under empirical datasets,” in Parallel Problem Solving From Nature economics from the Graduate School of Economics,
XI, LNCS 6239. 2010, pp. 199–208. Waseda University, Tokyo, Japan.
[36] W. DeBondt and R. Thaler, “Does the stock market overreact,” J. He is currently engaged in developing large-scale
Finance, vol. 40, no. 3, pp. 793–805, 1985. databases and in analyzing consumer behavior with
[37] J. Lakonishok, A. Shleifer, and R. Vishny, “Contrarian investment, Rakuten, Inc., Tokyo. His current research interests
extrapolation, and risk,” J. Finance, vol. 49, no. 5, pp. 1541–1578, 1994. include pattern recognition, multiagent systems, ex-
[38] E. Fama and K. French, “The cross-section of expected stock returns,” perimental economics, and neuroeconomics.
J. Finance, vol. 47, no. 2, pp. 427–465, 1992.
[39] N. Jegadeesh and S. Titman, “Returns to buying winners and selling
losers: Implications for stock market efficiency,” J. Finance, vol. 48,
no. 1, pp. 65–91, 1993.
[40] A. W. Lo and A. C. MacKinlay, “Stock market prices do not follow
random walks: Evidence from a simple specification test,” Rev. Financial Kyoko Suzuki received the B.E. and M.E. degrees
Stud., vol. 1, no. 1, pp. 41–66, 1988. from the Tokyo University of Science, Tokyo, Japan,
[41] M. Grinblatt and B. Han, “Prospect theory, mental accounting, and in 2005 and 2007, respectively, where she is cur-
momentum,” J. Financial Econ., vol. 78, no. 2, pp. 311–339, 2005. rently pursuing the Ph.D. degree in economics with
[42] A. Frazzini, “The disposition effect and under-reaction to news,” J. the Graduate School of Economics.
Finance, vol. 61, no. 4, pp. 2012–2046, 2006. She is currently a Research Fellow (DC) with
[43] S. Chen and C. Yeh, “On the emergent properties of artificial stock the Japan Society for the Promotion of Science.
markets: The efficient market hypothesis and the rational expectations Her current research interests include finance and
hypothesis* 1,” J. Econ. Behavior Organization, vol. 49, no. 2, pp. 217– multiagent systems.
239, 2002.
[44] C. Yeh and S. Chen, “Market diversity and market efficiency: The
approach based on genetic programming,” Artif. Intell. Simul. Behaviour
J., vol. 1, no. 1, pp. 147–167, 2001. Tetsuya Shimokawa received the Ph.D. degree in
[45] S. Markose, E. Tsang, and S. Jaramillo, “The red queen principle and economics from the University of Tokyo, Tokyo,
the emergence of efficient financial markets: An agent based approach,” Japan.
in Nonlinear Dynamics and Heterogeneous Interacting Agents (Lecture He is currently an Associate Professor with the
Notes in Economics and Mathematical Systems), vol. 550, part V, T. School of Management, Tokyo University of Sci-
Lux, S. Reitz, and E. Samanidou, Eds. Berlin/Heidelberg, Germany: ence, Tokyo. He was a Research Fellow (DC) of
Springer, pp. 287–303, 2005. the Japan Society for the Promotion of Science in
[46] A. Robson, “The evolution of rationality and the Red Queen* 1,” J. 1997 and of the Research on the Japanese Econ-
Econ. Theory, vol. 111, no. 1, pp. 1–22, 2003. omy University of Tokyo in 1999. He was also a
[47] A. Robson, “Complex evolutionary systems and the Red Queen*,” Econ. Research Fellow (PD) of the Japan Society for the
J., vol. 115, no. 504, pp. F211–F224, 2005. Promotion of Science in 2000. His current research
[48] S. Kauffman, At Home in the Universe: The Search for Laws of Self- interests include financial equilibrium theory, financial time-series analysis,
Organization and Complexity. Oxford, U.K.: Oxford Univ. Press, 1995. and behavioral finance.