Market Efficiencies and Drift: A Computational Model

John Dickhaut
Emeritus, University of Minnesota
Professor of Accounting and Economics, Chapman University
[email protected]

Baohua Xin
Assistant Professor
University of Toronto
[email protected]

We have benefited from discussions with Sudipta Basu, Joyce Berg, Peter Bossaerts, Amy Choy,

Sheri Dickhaut, Murray Frank, Frank Gigler, Steve Gjerstad, Kareem Jamal, Bruce Johnson,

Chandra Kanodia, Jack Kareken, Steven Kachelmeier (editor), John OBrien, Dave Porter,

Madhav Rajan (editor), Tom Rietz, Jack Stecher, Shyam Sunder, Greg Waymire, Rick Young,

two anonymous referees, and workshop participants at Chapman University (Economic Science

Institute), Emory University, Purdue University, the University of Iowa, the University of

Minnesota, the University of Pittsburgh, the University of Wisconsin at Madison, Norwegian

School of Economics, the Economic Science Association, the Southern Economic Association,

and EIASM Managerial Accounting Conference.

1
 Market Efficiencies and Drift: A Computational Model

John Dickhaut

University of Minnesota and Chapman University

Baohua Xin

University of Toronto

ABSTRACT: Accounting and finance researchers show semi-strong form efficiency or lack

thereof by using sequences of prices from Center for Research in Security Prices (CRSP) and

Compustat data for which there is no model of how these prices arise from individual decisions.

One needs a setting in which prices (including bids and asks) as well as information about

individuals making the choices are both available. To begin to bridge the gap between theory and

data, we extend work done by experimental economists on the double auction and model price

formation that is or is not semi-strong efficient. Agents in the model uncover prices in a manner

consistent with Hayeks notion of price discovery (Hayek 1948).

Keywords: market efficiency; drift; computational model; agent-based simulation.

Data Availability: Data are available on request.

2
 What individual has chosen prices? In the formal theory, at least, no one. They are

determined on (not by) social institutions known as markets, which equate supply

and demand. . . . The failure to give an individualistic explanation of price

formation has proved to be surprisingly hard to cure.

Arrow (1994, 4)

I. INTRODUCTION

For 40 years, accountants using archival data have sought an understanding of the

relationship between stock price behavior and information (and, in particular, accounting

information). The issue described by Arrow (1994) is always apparent in this research, namely,

that there is no individualistic explanation of how prices in archival data are formed.

Accountants show semi-strong form efficiency or lack thereof by using sequences of prices from

Center for Research in Security Prices (CRSP) and Compustat data, but without a model of how

these prices arise from individual decisions. The fact that such a puzzle has been unsettled for so

long suggests that it will not be resolved completely overnight. One needs a setting in which

prices (as well as bids and asks) and information about individuals making the choices are

available. To begin to bridge the gap between theory and data, we extend work done by

experimental economists on the double auction to model price formation that is/is not semi-

strong efficient.

A related motivation stems from the recent financial collapse; it has produced challenges

to the viability of economic thought underlying policy selection (and, in particular, the efficient

market hypothesis). Given Alan Greenspans admission of overreliance on the efficient market

3
hypothesis, it seems that, showing when such a hypothesis holds, could contribute to a better

understanding of the forces underlying economic disorder.

We use a general computational model of price formation with risky assets to give

insight into the type of structure that can support many of the (at times, apparently conflicting)

results. In principle, this model applies to any setting with a finite set of actors who behave

according to any well-known risk-preference structure (expected utility, prospect theory, rank-

dependent utility theory), and is void of strong (rational expectations) informational knowledge

on the part of actors in the economies. The model is general, yet it is computational in that it is

not of a closed form, and yields testable propositions when parameters are specified explicitly. In

this sense, it offers more than just experimental and/or agent-based simulation results; rather, it

offers an entire framework to examine theoretical predictions across a wide set of parametric

specifications and contexts.

Background Literature

Experimental economics has spawned an extensive literature, one aim of which is to

understand how individuals behave in double auctions. These auctions have been objects of

study because they bear resemblance to real-world trading institutions such as the New York

Stock Exchange (NYSE) and the Chicago Mercantile Exchange. Fundamental research on

double auction markets was conducted in a series of experiments summarized by Plott (1982)

and Smith (1982), which demonstrated that a competitive equilibrium (prices, quantities, and

consumer surplus) can be achieved.

Results from double auction experiments did not come without an accompanying puzzle:

why did the prices arrive where they did? After all, market participants had only limited

4
information; they did not know the private valuations of other traders, nor the equilibrium price,

but only the rules of the game and their own valuations. The work of Easley and Ledyard (1993),

Cason and Friedman (1993, 1996), and Wilson (1987) provide theoretical and empirical insight

into how this convergence might occur. Gode and Sunder (1993) provide a dramatic example of

how a very simple computer representation of each individual agent (a zero-intelligence agent)

allowed such agents to arrive at the competitive equilibrium consumer surplus. In their agent-

based simulation, agents were not concerned with any historical information about previous

behavior in the market, so that prices did not necessarily inform them what the next observable

price would be. Gjerstad and Dickhaut (1998) explore how historical information in the markets

(bids, asks, and trades) might be used by computerized agents to form beliefs in a way that could

produce price paths like those observed in classic double auction experiments. This latter

approach of asking how modeling individual agents could lead toward understanding

observables such as price paths in markets is the one adopted here. This approach can be used to

formulate theory, from which new experiments can be designed (Duffy 2004).

In related literature, Grossman (1976, 1978) and Grossman and Stiglitz (1980) show that

equilibrium exists only if there is noise in the price system that prevents traders from inferring

perfectly the information from prices. For example, there could be shocks in the supply of the

risky asset. They focus on the equilibrium (or disequilibrium) analysis: all traders move

simultaneously, and price incorporates all available information instantaneously. There are

several significant differences in our approach. First, we model the dynamics of price/bid/ask

formation, instead of focusing on the equilibrium results. Second, we use the double auction,

instead of the Walrasian auctioneer, because the latter is far removed from the auction structures

that are studied in both laboratory and archival studies. Finally, the traders in our model are

5
heterogeneous (in terms of preferences), and there is no shock in the supply of the risky asset. In

the Grossman (1976, 1978) and Grossman and Stiglitz (1980) models, there is no trade if

everyone knows the revealed information when it is revealed, whereas in our model, risk sharing

will always allow trade to take place.

Marcet and Sargent (1989a, 1989b) study a class of linear stochastic models in which the

actual law of motion depends on the agents perceived law of motion. They show the

convergence of recursive least squares learning schemes to a rational expectations equilibrium.

Marcet and Sargent use a description of how beliefs about the relationship between choices and

outcomes arise from repeated processing. Through repeated processing of information and

allocations based on that processing, the predicted law of motion becomes the true law of

motion. In our model, learning occurs directly. This learning involves examining prices, bids,

and asks, and building conditional beliefs that a particular bid or ask will be taken based on a

particular mechanism.

Market agents in Gjerstad and Dickhaut (1998) trade in riskless assets (i.e., assets with a

constant payout), yet many archival market studies based on prices from the NYSE are

concerned with how risky assets are priced, particularly in response to the arrival of information.

Fama (1970) distinguishes between semi-strong and strong form efficiency. Under semi-strong

(strong) form efficiency, profitable trades cannot be achieved by individual agents after the

release of public (private) information. Such efficiency is often a natural consequence of the

behavior of agents whose expectations are rational. When observed in archival data, the returns

after information is released are flat, and in numerous instances, it is impossible to build

advantageous trading algorithms on such prices. Figure 1A, taken from Fama et al. (1969;

hereinafter referred to as FFJR), shows the positive change in cumulative residuals (excess

6
returns) for a portfolio of firms up to the point of announcement of a stock split (date 0) with a

nearly flat curve thereafter. A stock split is public information such that an interpretation of the

residuals before date 0 indicates that their change reflects private information coming into the

market, while the behavior of the residuals after time 0 indicates the absence of advantageous

trades, thereby implying semi-strong form efficiency.

(insert Figure 1 about here)

We show, by adapting the Gjerstad and Dickhaut (1998) trading algorithm to trade in a

risky asset, how it is possible to produce prices that are semi-strong efficient but not strong

efficient.

Figure 1B, taken from Ball and Brown (1968; hereinafter referred to as BB), shows the

positive (negative) abnormal performance index up to the point of earnings announcements (date

0) for various portfolios of firms and relatively flat curves thereafter. An earnings announcement

is public information, and an interpretation of the residuals before date 0 indicates that their

change reflects private information coming into the market, while the behavior of the residuals

after time 0 is consistent with a notion of a mild drift.

Point 0 is when public information in the form of the earnings announcements arrives for

a collection of firms. Each line represents different cumulative average residuals for a collection

of forecast/earnings combinations. As is evident in Figure 1B, the drift after announcement is not

independent of the type of information generated. This particular study does not have an explicit

theory associated with it that would explain such a drift. We are able to show that this type of

drift can be well approximated by the interaction of agents in our model of price formation.

7
A Brief Description of the Dynamic Process

We describe the economy at an arbitrary point in time, t, in the life of the risky asset. At

that point in time, there are a finite number of traders, n. Each trader has an endowment of the

risky and a riskless asset. The traders risk attitudes are all different. Traders may also differ in

their information: some traders may be better informed than others. The traders are not assumed

to have any direct knowledge of the other types of traders in the environment at any particular

point of time.

We model how a sequence of trades comes about in a double auction in which agents

post bids and asks. Such bids and asks can start at any amount (here we assume that agents

confine their amounts between the high and low payoffs). The critical feature of the auction is

that bids and asks are subject to an improvement rule; that is, if there is an outstanding bid and

ask at time t, agents may only submit a t + 1 bid that is higher or a t + 1 ask that is lower than the

outstanding bid or ask. The auction processes only one possible improvement at a time, and a

legitimate bid/ask submitted by the first agent will become the new outstanding bid or ask at t +

1. A trade occurs when a bid equals the outstanding ask, or vice versa. If a trade does not occur,

all agents must decide what to do at t + 1 in an attempt to be the outstanding bid/ask at t + 2.

At any point in time, besides having some information on the payoff of the risky asset, an

agent also sees the history of previous bids and asks and how many bids and asks were or were

not taken for each ask. There are two types of information processing in which a market

participant engages. First, he or she analyzes the implications of his or her own information on

the value of a traded asset. Second, he or she analyzes the implications of publicly available

prices, bids, and asks to determine his or her action choice in the market. That action can be

8
either making a bid or ask or taking an outstanding bid or ask based on his or her assessment of

value. Because agents can have different risk attitudes as well as information, agents can have

different valuations of the risky asset. When there are different valuations, there is a potential for

an advantageous trade between two agents if the agreed-upon price is between the valuations. An

agent need not bid/ask his or her valuation; however, if there is an opportunity to trade at an

advantageous price (i.e., buy at a price less than valuation or sell at a price greater than

valuation), the agent will try to take advantage of the opportunity.

When each agent has decided precisely what to do, there will be an expected profit from

a potential trade. Traders with higher expected profit from trade are more anxious to act faster. 1

However, another trader may get to the market first. The chance that any particular trader takes

his or her action first is a monotonic function of the ratio between his or her expected profit and

the sum of all traders expected profits. Once an action is taken by a trader, the market

incorporates that action by either executing a trade or posting a new ask or bid. Then the same

process is repeated.

II. DESCRIPTION OF THE MODEL

Figure 2 describes the four basic steps of the model. In step 1, each agent determines his

or her value. In step 2, the agent uses market information to assess the likelihood of being able to

trade. In step 3, the agent takes an optimal trading action, and step 4 describes the stochastic

arrival of agents to the market. We examine each step in detail.

(insert Figure 2 about here)

1
We use agent (agents), trader (traders) and market participant (participants) interchangeably throughout the paper.

9
Step 1: Valuation

Each trader comes to the market with a specific risk attitude. Not all traders have the

same risk attitude. To make this assumption specific, we describe a risk attitude via a risk

parameter later on. However, there is no conceptual restriction on how risk is incorporated. A

trader uses the information he or she has to arrive at his or her assessment of the reservation

value of the traded asset (i.e., the lowest market price at which the trader would sell and the

highest market price at which the trader would buy). The method for making this assessment

could involve any type of decision-making model such as expected utility theory, prospect

theory, or rank-dependent utility theory.2 The actor may even choose to include previous market

price in the formulation. So long as certainty equivalents can be stated for each probability

payoff pair (and wealth combination), the theory works the same way, although the

implementation becomes more challenging. Because traders have different risk attitudes, trade is

likely to occur.

A market trader may or may not have personal private information. A traders private

information could, in principle, derive from word of mouth or personal observation. Information

is not purchased, but rather is simply endowed to traders who are at the right place at the right

time. Informed traders do not know how many other traders are informed. All informed traders

share the same beliefs about the dollar outcome from the asset when computing their estimates of

2
To implement the structure, we will assume that this assessment is made by computing expected utility theory
given available information and then deriving the related certainty equivalent. We derive here the implications of
having ten agents with different risk coefficients who also have the same underlying constant absolute risk
aversion (CARA) utility function. The approach can be extended to a general two-state model in which we assume
two payoffs and a reservation value. Note that any structure that has a reservation value for the gamble will also
work. For simplicity, we employ a model that makes the calculation of the reservation value a direct function of
payoffs and risk characterization, although there is no reason to make the model this specific. It is always assumed
that the reservation value will be between the high and low payoffs. It may be that the agent has two reservation
values, one related to being a seller and the other related to being a buyer. As long as the agent is still interested in
the difference between the reservation values and the price received in the market, the structure can work out in the
same manner.

10
value. This assumption is implemented by incorporating Bayesian revision relative to the arrival

of information.3 Uninformed traders will have processes for generating beliefs that include all

available public information or some portion of it. Uninformed market participants do not know

when informed market participants receive information. Because prices, bids, and asks arise

from both new information to informed market participants and risk-sharing considerations,

uninformed traders are unable to disentangle these effects, thereby basing their assessments of

value only on released public information. Accordingly, uninformed market participants are

assumed to use Bayesian revision relative only to the arrival of public information.

In step 1, agents are just determining their value before they take an action in the market.

We assume that this activity can be done with some degree of precision because experts may be

consulted, and there can be extensive representations of whatever information there is. In steps 2

and 3, the agent decides what action to take in the market (i.e., an appropriate bid, an appropriate

ask, or a take).

Step 2: Agent Computations of Likelihoods

An agent assesses the conditional probabilities of bids and offers being accepted based on

history. In step 2, the agent gathers additional inputs for deciding what action to take in the

market (bid, ask, or take of an available bid or ask outstanding). The agent uses the certainty

equivalent of his or her computed value from the previous step as well as the available market

information (prices, bids, and asks). At any point in time t, there is a set of messages. Each of

these messages is sent in the form of a bid, ask, taken bid, or taken ask by a specific agent. The

3
Note that pure Bayesian is not absolutely necessary for the general structure of the model. An agent could
differentially underadjust or overadjust. This characteristic, of course, would need to be built specifically into the
adjustment process of the model.

11
message at time t reflects an action by one of the agents in the economy that maximizes that

agents expected surplus.4

The additional information the agent computes in this step is the likelihood that a

particular bid or ask will be taken by someone else in the market. In the absence of differential

information, traders still do not know what types are in the market, and so engage in trades based

only on available bid, ask, and trade information, identifying pseudo-arbitrage opportunities

based on that information and private knowledge of risk. This preference process is stimulated

by the ability to re-trade. While pseudo-arbitraging is taking place, differential information can

arrive to traders. Traders with new information revalue; people without do not revalue. Traders

do not know anyone else received new information and what that information might be. They

infer from price the advantageous trade relative to that price. We assume that agents have limited

memory.

While a trader knows his or her value of the asset, he or she knows nothing about the risk

distribution of other trader types, nor how the most recent market observations reflect the

consequence of differential information or risk sharing. Thus every trader bases his or her

assessment on the likelihood of a particular bid (ask) being accepted on recent observed

information in the market about whether that bid (ask) will be successful or unsuccessful. We

assume, because the trader simply has such limited information and no basis for assigning priors,

that the only information on which the trader can base his or her action regarding a bid or ask is

the recent prices, bids, and asks. He or she then resorts to a simple counting system (which can

be subconscious). For example, the likelihood that a specific dollar bid B will be taken is based

on (1) the number of times this specific dollar bid has been successfully taken, #BT; (2) the

4
The timing of messages is independent of the type of risk preferences the agent has for representing the basic
gamble as well as the belief revision policy.

12
number of times an ask below this bid has been made, #(A < B); and (3) the number of times a

bid above this bid has not been taken, # ( BNT B) . The likelihood is then computed using the

following formulation:

# BT # ( A B)
.
# BT # ( A B) # ( BNT B)

A similar counting algorithm applies to the calculation of the likelihood of being taken

for a specific dollar ask A:

# AT # ( B A)
,
# AT # ( B A) # ( ANT A)

where #AT is the number of times this specific dollar ask has been successfully taken; #(B > A) is

the number of times a bid above this ask has been made; and # ( ANT A) is the number of

times an ask below this ask has not been taken.

Why does a trader form such a belief? Given the paucity of information in the

environment, we suggest that the traders resort to primitive mechanisms that have evolved

through millions of years. In referring to primitive mechanisms, we refer not only to human

evidence, but also to evidence that has been more precisely gathered with animals other than

humans. There are now findings that the brain has the ability to represent numbers as if taken

from a line segment, and that this ability has existed somewhat indefinitely. For example, rats are

capable of distinguishing the numbers 1, 2, 3, and so on, based on whether a reward is given

when a rat presses a lever exactly one, two, or three times. Monkeys can also do simple auxiliary

mathematical operations such as addition. This behavior occurs conjointly with activation in the

posterior parietal sulcus and dorsolateral prefrontal cortex of the monkey brain. Humans have a

very extensive system for representing numbers inexactly in the horizontal intraparietal sulcus

and have mechanisms for adding, subtracting, multiplying, and dividing. These operations can be

13
carried on by the brain regardless of whether an individual is conscious of performing these

operations. Recently, Fiorillo et al. (2003) found evidence that specific cells of the monkey can

be tuned to different conditional probabilities of reward that exist in the environment. Thus we

are not assuming conscious mechanisms in the mind of the trader, but rather that traders have a

feel for what is the best bid or ask, even though they may not be aware how such feelings arise.

Gigerenzer and Ulrich (1995) trained experimental participants to behave in a Bayesian fashion

based on variations in the types of counting systems used here, although they assume more

explicit information.

An alternative way to think about this is to equate the individuals in the economy to

econometricians in the sense of Sargent (1993, 34). The critical point we are trying to make

here is that even without being aware of any explicit statistical expertise, our agents have ways

of examining conditional probabilities, which would be at the heart of any econometric

formulation.

Step 3: Determination of Optimal Action

In step 3, the agent determines the expected monetary profit for each action he or she

might take.5 The expected profit for particular agent i, with a valuation Vi , for bid B, will

be P( B) (Vi B) , where for an observed B,

# BT # ( A B)
P( B) ,
# BT # ( A B) # ( BNT B)

while for previously unobserved Bs, the calculation for P(B) is determined by interpolation. The

resulting P(B) is nondecreasing in B.

5
We omit here the construction that would be necessary if the agent had a different certainty equivalent for buying
and selling.

14
 Similarly, the expected profits for particular agent i, with a valuation Vi , for bid A, will

be P( A) ( A Vi ) , where for an observed A,

# AT # ( B A)
P( A) ,
# AT # ( B A ) # ( ANT A)

while for previously unobserved As, the calculation for P(A) is determined by interpolation. The

resulting P(A) is nonincreasing in A.

Choice calculation proceeds as follows: the agent behaves as if doing four calculations

Max P( B) (Vi B) ,
B

Max P( A) ( A Vi ) ,
A

Vi OB ,

OA Vi ,

where OB (OA) is the existing outstanding bid (ask). Determining the maximum profit from

these calculations leads to the selection of an action that will be a particular B, A, OB, or OA,

where the maximum profit is greater than 0. Each subject can make a bid, make an ask, or take

an outstanding bid or ask. Thus he or she may be a buyer or a seller.

Step 4: Whose Bid/Ask Is Posted FirstA Decentralized Process

To summarize, through the first three steps, each trader arrives at what he or she would

do if the trader could act first in the auction; that is, he or she knows whether to (1) make a bid,

(2) make an ask, (3) accept an outstanding bid, or (4) accept an outstanding ask. A player will

only take one of these four actions. The expected profit for each action is computed in choice

calculation.

15
 The auction mechanism acknowledges only one action of only one player. Once this

action is incorporated into the message structure of the auction, all players are notified, and the

bidding process begins again. What is important is the determinate of who moves in the auction

and that we are not assuming the mechanism makes the decision of who moves first.

It is crucial that the auction be decentralized if prices are a consequence of individual

choice. This means that we need a behavioral theory of how such decentralization takes place.

No single person (or mechanism) knows all profits or valuations of each actor. We assume that

on average, the more profit a single player makes, the more likely he or she will enter the

auction; that is, higher profits to a player induce that player to act faster. Interpreting this

proposition strictly would imply that the person with the highest expected profit would always

move first. However, we assume some noise in the response, which could be attributed to the

circumstances of individual players. This assumption is captured in the following way: if Si is the

expected profit of the optimal action for the ith player, then the likelihood that player i will move

first is

si
.
s1 si sn
Any monotonic function defined on the Si will result in a similar construction that will generate

noise in trades. The important point here is that if traders move to get in the auction based on

their profits, then the entire process is decentralized.6

The assumption that speed of reaction depends on payoffs is well established in many

studies. A general theory of reaction times is that an alternative with higher profits results in

arriving at a decision faster when compared with a zero-profit alternative (doing nothing). Such

6
Each subject can make a bid, make an ask, or take an outstanding bid or ask. Thus he or she may be a buyer or a
seller. Agents learn the consequences of their actions and remember them for four trades.

16
an approach can be incorporated into an established theory of reaction time. For example,

Dickhaut et al. (2008) use Ratcliffs (1978) theory to show that the further away a choice is from

zero profits as compared to an alternative, the faster is the reaction time. Furthermore, the notion

that higher rewards stimulate faster reaction times has strong support with respect to primitive

brain development. Tremblay and Shultz (1999) find that specific neurons in the orbital frontal

cortex respond not only when a reward is received, but respond faster the higher the reward. The

latter evidence suggests that the speed of response can be determined without conscious

awareness.

Our mathematical formulation of the theory is in the appendix.

III. RESULTS FROM COMPUTATIONAL ECONOMIES

Agent-Based Computational Models

Our results are computer computations that result from studying the behavior of agents

through repeated interactions in our markets. Each agent acts separately and independently

(except through market information) of other agents. This approach goes by the name agent-

based modeling and is a particular form of simulation that has grown increasingly popular for

studying emergent phenomena, that is, phenomena that, in the aggregate, arise from the

interaction of individuals who have no understanding of their impact on aggregate phenomena

(Bonabeau 2002). Thus we are showing how the market prices that arise when different market

phenomena are observed, are a consequence of agents with the same underlying characteristics.

Only aspects of the market setting have changed. The program was written in Mathematica and

is available on request from the authors.

17
Characteristics of All Economies

In the computational economies discussed in this section, we follow the general

hypothesis arising from experimental economics that properties of large-scale economies can be

approximated with a small number of agents in the double auction (Smith 1962). For our

graphical results, we always use either 10 agents or 20 agents. Each agent is endowed with the

same number of risky and riskless assets. In the specific economies, agents have different risk

coefficients, but across economies, the collection of risk coefficients is held constant.

Possible states and signals are always the same, and priors are held the same on all states

in the first two economies. There are two possible public information signals in economies 1 and

2, and the realization of public information is the same across economies. In our economies, the

public signal takes on a value of 40. Furthermore, the conditional probabilities of public signal

given the state are always the same. These features of the economies are characterized in Table

1, panels A and B.

(insert Table 1 here)

Computational Economy 1: Strong versus Semi-strong Form Market Efficiency

(Approximating FFJR)

The first study shows that, with public information, semi-strong form efficiency arises

without strong form efficiency. There are 1,000 points of history that are potential trading dates.

At any of these points in history, an outstanding bid and/or ask as well as the previous history

exists. The trader may have private or public information at any point. Different traders receive

the same piece of private information at time 100 (two agents get information); at time 200 (three

18
different agents get information); and at time 300 (three more agents get information). The exact

same piece of information is made public at time 400.

Each agent can determine a different certainty equivalent relative to his or her

information and risk attitude (see Step 1: Valuation). When informed agents receive a signal,

they update their probabilities of the state outcome of 40 or 80 dollars accordingly. For the

economies studied here, the private signal takes on a value of 40, and the appropriate conditional

probabilities of private signal given the state are represented in Table 1, panel C. The

information structure, timeline, and belief updating are summarized in scenario 1 of the model in

the appendix.

A market agent would like to sell above/buy below his or her certainty equivalent. If a

trader were at point in time 200, he or she would see all the information portrayed in Figure 3

prior to point 200. A trader also knows which bids and asks were successful in leading to an

exchange, so he or she can see for a particular bid just how frequently it was chosen. The trader

knows nothing about other traders in the market such as whether, at any point in time, other

traders have received private information. The limited trader information about whether

particular bids or asks were taken gives the trader some insight (although the lack of specificity

of the problem makes this insight non-Bayesian) and an ability to calculate relative frequencies

in an as-if manner.7

Each trader is guided by maximum profit in the trading market relative to the certainty

equivalent. In period 0 to 100, all traders, though not consciously aware, have the same identical

information for determining their values. In this case, priors are 0.2 and 0.8 on the states 40 and

80 respectively, which are final payoffs. Note that prices need not be constant because trades at a

7
The exact description of traders in this economy is given in the appendix. While we used a specific form of the
utility function here, the theory is not at all constrained to this function, or to expected utility maximization, for
determining the certainty equivalent of a share.

19
new price for one trader can change the expected gains to trade for an action (bid, ask, or take)

of the other traders.

(insert Figure 3 about here)

Computational Economy 1: Bid/Ask Results

In Figure 3A, the gray points are simulated asks and the black points are simulated bids.

There is a tendency toward scalloping (i.e., a downward string of asks or an upward string of

bids) in the data, particularly in the early stages. This scalloping most often represents a single

agent who tries to make an advantageous trade and then is disciplined by the market in the sense

that the market will not jump at the over- (under-) priced ask (bid). A participant represented by

a scallop has a much higher expected profit from his or her action than other traders have from

their actions, leading that participant to try to take advantage of the profit opportunities. The

improvement rule forces a potential seller (or buyer) to reduce asks (or increase bids) until a

trade is made. A particular seller keeps lowering his or her ask until his or her profits are more in

line with those of other traders in the market. As more traders enter, we see competition in

bidding behavior. After a trade, the outstanding bid and ask are reset, and we can see scalloping

again. Bids and asks become more and more concentrated because the agents most frequently

tend to use asks and bids that have had some degree of success in generating exchanges in the

past. In this setting, sellers appear to undercut each other, but in a formal sense, they are not

strategic insofar as their actions do not anticipate other traders responses. Sellers simply identify

optimal asking prices based on the information that is available in the market, that is, bids, asks,

20
and takes. This competition is not arising from considering what other agents might be thinking,

but rather, it is a form of arbitrage relative to market bids and asks.

In this scenario, the market is always discovering information in a Hayekian sense

(Hayek 1948); that is, an order is emerging in the market in a way that none of the participants

anticipate, such that the distribution of bids and asks is an unintended consequence of all the

actors in the economy.

Computational Economy 1: Pricing Results

Figure 3B displays the time series for price in this economy. Trades are represented by

points on the graph. Trades in the interval 0 to 100 reflect gains to risk sharing. At point in time

100, the private signal begins to be absorbed in market price, and we see price gradually shifting

downward as more and more traders learn the private message. At point in time 400, public

information is released and prices settle down to a steady state, but prices continue to vary

because any trade can affect the profit opportunities of others. Up until time 400, it is possible to

see private information leaking into prices. During this time, there are always two sources of

uncertainty when trading: the potential existence of private information and uncertainty about

whether a particular action will lead to a consummated transaction.

We interpret these results as being consistent with semi-strong form efficiency but not

strong form efficiency. The graph has a qualitative similarity with the earlier graph for FFJR. To

get a better feel for this result, we bootstrapped this procedure 100 times, yielding a distribution

of regression coefficients of prices on time. We compare the resulting diagrams, together with

the BB computational economy described in the next section.

21
Computational Economy 2 (Approximating BB)

To approximate the BB results, we keep the scenario the same as the FFJR case but

assume that private and public information are not equivalent. The public signal is correlated, but

not perfectly, with the private signal. There are, again, 1,000 points of history that are potential

trading dates. At any of these points, an outstanding bid, ask, or exchange may occur. Private and

public information arrive identically to economy 1. The bid/ask data for the correlated setting

have the same general properties as the complete revelation setting, although offers are

disregarded a bit more as time increases after the public information announcement.

The information structure, timeline, and belief updating are summarized in scenario 2 in

the model in the appendix. For this economy (and economy 3, later), the appropriate additional

conditional probabilities are represented in Table 1, panel D.

(insert Figure 4 about here)

The double auction price data are captured in Figure 4A. There is an apparent drift in

these data. As in computational economy 1, we also bootstrapped this procedure 100 times,

yielding a distribution of regression coefficients of prices on time. The distributions of

coefficients of time series for economies 1 and 2 are displayed in Figure 5.

(insert Figure 5 about here)

The top graphs show the coefficients corresponding to the private information times (100

to 399) and public information times (400 to 1,000) of the FFJR agent-based simulation

22
(economy 1). Note that the distribution of coefficients is significantly shifted to the left in the

private information case and is consistent with the period 100 to 399 in the single economy of

Figure 3. The coefficients in FFJR/public correspond to 100 runs of periods 400 to 1,000. The

coefficients, while slightly shifted to the left, are indistinguishable from 0. The bottom graphs

show the BB private and public information (economy 2). Since both private information graphs

for FFJR and BB are drawn on the same scale, it is apparent that prior to public information, the

distribution of coefficients is virtually identical. On the other hand, after release of public

information, the distribution of coefficients under BB is further skewed to the left and is

significantly different from that under FFJR (with p < 0.001 using a Kolmogorov-Smirnov test).

To examine for similarities between our agent-based simulations and experimental work,

in which data can be used to indicate differential information under asymmetric versus public

state uncertainty, we also compared an estimate of the autocorrelation process using generalized

autoregressive conditional heteroskedasticity (GARCH; Bollerslev 1986) techniques. By

expanding the number of agents in the economy to 50, we can show that when asymmetric

information is present, prices follow a GARCH (2,1) process. Concurrently, when all

information is public, a GARCH process cannot be detected. These results are consistent with

those documented in experimental data (e.g., Bruguier et al. 2008).

We also compared the allocative efficiency of prices in a setting with public information

to a setting with private information. When economies are constrained to have the same length

(1,000 points of time) as well as the same endowments and risk characteristics as in our BB and

FFJR examples, allocative efficiency in the private setting is 55 percent, while in the public state

information setting, the allocative efficiency is 99.5 percent. For a discussion of this

computation, see Plott and Sunder (1982).

23
Computational Economy 3: Volatility

The literature contains numerous results regarding volatility. We focus on a volatility

result for which there is theoretical basis. Veronesi (1999) proposes that volatility in response to

information regarding a particular event is conditional on the priors of that event. In particular,

given the same information event, the market with the higher prior probability will have more

volatility in the market price when the same information is revealed. In the Veronesi model,

prices move from one equilibrium level to another. There is no mechanism by which individual

agents formally enter into the price setting process through making bids or asks. In our model, it

is possible to produce Veronesis volatility result when traders are actively engaged in trading in

an auction; we can then show how differences in measured volatility arise as a consequence of

differences in prior probabilities.

Figure 4B portrays the outcome of applying our price formation mechanism under

Veronesis assumptions. In one economy, agents begin with high priors (0.8) of the high state

outcome (80), while in the other economy, agents begin with low priors (0.2) of the high state

occurring. The remaining aspects of these two economies are identical and reflected in Table 1,

panels A and B. We assume that a public low signal (40) is received at trading point 50 in each

setting. The interior graph plots price on the vertical axis and history of trades on the horizontal

axis when all agents have a prior of 0.2 and receive a public signal of 40. When everyone

receives identical information, the price adjustment is very quick, even though individual agents

are trading only on their private accounts and do not assess a distribution of future prices. Note

that there is some variability after the adjustment occurs. In the interior graph, the ticks on the

vertical axis increase at a rate of 0.2, which suggests that volatility is low. If we measure

24
volatility as the standard deviation of price after the release of the information, it is 0.036. On the

other hand, the description of a market reaction when there are high priors is portrayed in the

outer graph, which shows that agents trade on their own accounts and prices adjust quickly. Once

again, there is no anticipation by agents of what other agents will do, and in this sense, they do

not behave strategically. The range of the outer graph is 40 to 50, which seems to suggest that the

volatility for prices after receiving information is greater than when the priors are 0.2. Volatility

as measured by standard deviation of prices after the information release is now 0.155, so the

Veronesi (1999) claim is reflected in these data.

Robustness

There are a number of ways to examine the robustness of our price formation dynamic.

One is to simply replicate the same economy for a large number of trials. We used this approach

for describing the initial studies of semi-strong form efficiency and failure of semi-strong form

efficiency, finding consistent results with a reasonable tolerance level (see Figure 5). In addition,

we varied the specifications of the utility functions, the endowments of agents, and the initial

probabilities of agents with consistent findings across all cases.

We consider the strongest strength of the model the fact that, not only the results on

efficiency, but in addition, other fundamental results emerge, including the information

transmission in forecasts, volatility of prices in settings with and without asymmetric

information, bid-ask spreads in relations to asymmetric information, and the possibility of drift

processes induced by non-standard belief revision. This extended set of results, along with the

discussion of other auction mechanisms (e.g., dealers) that could yield similar conclusions, are in

25
a companion piece Learning about field behavior using a laboratory based model (Dickhaut

and Xin 2009).

IV. CONCLUSIONS

This study builds a model of price formation and shows that the model leads to a variety

of phenomena, including semi-strong from market efficiency, semi-strong market inefficiency,

and predictable volatility. Rather than specify just prices, the model specifies the generation of

all messages as they reach the market. Our model yields the ability to test parametric predictions

of the theory in a variety of experimental settings in which tastes can be estimated or induced.

OBriens (1992) and Dickhaut et al.s (1993) results from inducing preferences in auctions

suggest the plausibility of tracking the predictions of such a model. OBrien (1992) does not

explicitly attempt to specify the evolution of all messages in the market. Recently, Gjerstad

(2006), in a non-risky setting, has been able to explicitly test the predictions of this price

formation process when individuals values are certain and there is no asymmetric state

information.8

To elaborate the test of our model, suppose one takes the model predictions in

computational economy 1. That economy yields bid/ask spreads, prices, volatility, and a lagged

correlational structure. These predictions can be tested with an experiment that induces the risk

preferences used in that computational economy, randomly releases private information, and has

a public information release in a double auction (e.g., using the methodology of Plott and Sunder

[1982]).

8
There could, of course, be other models. The model of Bosch-Domenech and Sunder (2000) is a model of price
formation that does not achieve the specifics of the current setup, but conceivably, a variant of this approach might
be made to create the pricing information described here.

26
 Beyond predicting in the laboratory, we explore fundamental findings from empirical-

archival accounting and finance. Notions of weak and strong form market efficiency have been

replicated extensively in archival settings (Fama 1970, 1991). Our contribution is to provide a

theory derived from laboratory experimentation that can explain such results and provide

conditions under which they might occur. Unlike standard accounting and finance theory, we

incorporate the choices of individual traders involving how to submit bids and asks and take

outstanding bids and asks.

27
 APPENDIX

THE MODEL

Environment at Time t

Preferences and Valuations

The model can be applied to an arbitrarily large finite set of states and signals. For

concreteness, we focus on a world with two states and two possible signals each time

information is released. We assume an economy with n agents, all of whom will consume at a

future time T. There are two types of assets: riskless and risky. At T, the risky asset pays off one

of two amounts of the riskless asset, xL , xH , where x H x L . At any time t, agent i holds a

belief PH ,i ,t (0 PH ,i ,t 1) regarding the probability of the high outcome. We assume that each

agent i has a constant absolute risk averse utility function defined relative to the holdings of the

riskless asset at time T9:

Ui (wi ) ei wi ,

where wi denotes is wealth at T and 0 i is agent is risk parameter. Given i , PH ,i ,t , and

{x L , x H } , the expected utility of the risky asset to i at t will be

EUi (i , PH ,i ,t ,{xL , xH }) (1 PH ,i ,t ) (ei xL ) PH ,i ,t (ei xH ) .

The valuation of the gamble in units of the riskless asset is that value of ci ,t that solves

i ci ,t
e (1 PH ,i ,t ) (ei xL ) PH ,i ,t (ei xH ) ,

where ci ,t represents that amount at which agent i is just willing to buy or sell at time t.

9
Any structure that can define certainty equivalents for each probability payoff pair (and endowment) will also work
here (e.g., prospect theory).

28
 In the preceding setting, the agents may or may not have homogeneous beliefs. Later, we

consider settings in which the market might not fully incorporate information. Furthermore, t is

meant to be any arbitrary t less than T, thus t might be the initial time (t = 0), or it may represent

the time at which private or public information is received, or the time of a forecast.

Information Structures

We focus on two prototypical information scenarios. In both scenarios, all agents start out

with the same prior beliefs. Privileged agents receive private information, and later public

information is released. These scenarios are captured by Figure A1.

(insert Figure A1 about here)

In each setting, there are initial prior beliefs for the low and high states represented by

P
S L .
PH

At time tprivate, selected agents are informed of a private signal, which can take on the

value L or H. The likelihoods of the private signal given low and high states are given

PL|L PH |L
SM 1 ,
PL|H PH |H

where Pi| j is the conditional probability of message i given state j.

The scenarios differ at time tpublic. In scenario 1, the public information is identical to the

private, and so the likelihood of the public signal given the private signal is simply equivalent to

the private signal given the state. In scenario 2, the public signal is a separate random draw, and

29
thus the public information is only partially correlated with the private information. An

appropriate probabilistic representation of these facts is

PL|L, L PL|H , L PL|L , H PL|H , H

SM 2M 1L SM 2M 1H ,
PH |L, L PH |H , L PH |L , H PH |H , H

where Pi|j,k represents the probability that the public signal is i given that the private signal is j

and the state is k.

In this construction, we assume that agents, after seeing information, perform Bayesian

revisions to arrive at revised probabilities. Thus, in scenario 1, an agent would calculate the

probability of the high state given low private and public messages as

SM 12,1 S2 / SM 12,1 S2 SM 11,1 S1 .

On the other hand, in scenario 2, the agent calculates the probability of the high state

given low private and public messages based on the dependency between the messages as

(SM 2M 1H )1,1 SM 12,1 S2 / [(SM 2M 1H )1,1 SM 12,1 S2 (SM 2M 1L )1,1 SM11,1 S1 ] .

These information structures are reflected in Table A1.

(insert Table A1 about here)

Institution: The Double Auction

In the double auction, potential sellers post asks and potential buyers post bids. The

message space defines a set of allowable messages for each agent. We consider the double

auction with a bid/ask spread reduction rule (defined later). In effect, this produces restrictions

on any agents messages as a function of previous messages from all agents. The double auction

imposes no restrictions on the sequencing of messages: any agent can send a message at any time

30
during the trading period. Allocation of units is by mutual consent between any buyer and seller.

If a sellers ask is acceptable to a buyer, then a transaction is completed when the buyer takes

(accepts) the sellers ask. Similarly, a buyers bid may be accepted by a seller.

Definition 1: Asks. An ask a is an amount that a potential seller i is willing to accept from

any buyer as payment for a unit of the commodity being traded. To submit an ask of a,

seller i sends the message (i, 0, a).

Definition 2: Bids. A bid b by a potential buyer j is an amount that j is willing to pay to

any seller for a unit. Buyer j submits this bid by sending the message (0, j, b).

Definition 3: Spread Reduction Rule. The lowest ask in the market at any time is called

the outstanding ask and is denoted oa. At any time, agents must place asks a satisfying a

< oa. The highest bid is called the outstanding bid, denoted ob. If agents place a bid, it

must be above the outstanding bid. The outstanding ask oa and outstanding bid ob define

the bid/ask spread [ob, oa]. In markets with a spread reduction rule, all bids and asks fall

within the bid/ask spread.

Definition 4: Acceptance. If a potential seller i sends the message (i, 0, a), and holds the

outstanding ask oa = a, then a take of oa by a buyer j is an agreement by j to purchase a

unit from seller i at the transaction price p = oa. Buyer j accepts the outstanding ask oa

by sending the message (0, j, b), where b = oa. Similarly, if the outstanding bid ob is held

by buyer j, then a take of ob by seller i is an agreement by i to sell a unit to buyer j at

the transaction price p = ob.

Definition 5: Trades. A trade is represented by (i, j, oa) if j takes the ask of i or (i, j, ob)

if i takes the bid of j.

31
Observed Histories

This section demonstrates a method of describing data to arrive at what needs to get into

the calculation of probabilities. A record of bids, asks, and takes constitutes a history of

transactions. For example, consider the following history:

H2 = {h1, h2} = {(3, 0, 45.00), (3, 1, 45.00)}.

This history captures the fact that agent 3 submitted an ask of 45.00 and agent 1 accepted that

ask.

Definition 6: Histories. After n messages have been sent, there is a history Hn of n

ordered triples. For any message mn+1 = (mn+1,1, mn+1,2, mn+1,3) that is sent, one of six cases

will hold:

Invalid ask or bid. A message mn+1 = (i, 0, a) is not valid if a oa . An invalid

ask will not be included in the history. In effect, the institution ignores messages

that violate the spread reduction rule. Similarly, a message mn+1 = (0, j, b) is not

valid if b ob .

No ask outstanding. If no ask has been made since the last transaction, then there

is no outstanding ask, and any ask a is valid. If, in addition, mn+1,3 > ob, then hn+1

= mn+1.

No bid outstanding. Similarly, if no bid has been made since the last transaction,

then there is no outstanding bid, and any bid b is valid. If mn+1,3 < oa, then hn+1 =

mn+1.

Acceptance of ob. If mn1,1 0 and mn+1,3 = ob, then seller mn+1,1 is making an

offer at ob, so that mn+1 is an acceptance of ob. The buyers identity is found by

32
 looking back in Hn and finding the last hk with hk , 2 0, that is, k*= max{k :

hk , 2 0} . Then (hn+1,1, hn+1,2, hn+1,3) = (mn+1,1, hk*,2, ob).

Acceptance of oa. If mn1, 2 0 and mn+1,3 = oa, then buyer mn+1,2 is making a bid

at oa, so the mn+1 is an acceptance of oa. The sellers identity is found by looking

back in Hn and finding k * max {k : hk ,1 0} . Then (hn+1,1, hn+1,2, hn+1,3) = (hk*,1,

mn+1,2, oa).

Improving ask or bid. If mn1,3 (ob, oa) , then mn+1 is either an improving ask or

an improving bid, and hn+1 = mn+1.

Behavior

At any point in time, agent i has a valuation for the risky asset ci,t, which represents the

price at which the agent would either buy or sell.

Definition 7: Trading Profits. Given a trade (i, j, hn+1,3), in which i sells j a unit of the

risky asset with given valuations cj,t and ci,t, then the trading profits to the seller and buyer

are (hn+1,3 c i, t) and (cj, t hn+1,3), respectively.

After observing the most recent message mn, each agent decides whether to submit a bid

or ask and the amount of the bid or ask. Every agent i assumes that is behavior will not affect

the likelihood of a particular bid or ask resulting in a trade. Furthermore, the agent assumes that

the likelihood of a particular bid or ask resulting in a trade can be derived from information that

the agent has observed and remembers from history.

We first provide the intuition of how the optimal bid and/or ask is determined by an

agent. Suppose agent i is considering asking a. Then the profit if the ask is accepted is (a ci,t).

The question for the agent, then, is, how likely is it that a will be accepted? In past trading
33
behavior, there may be instances in which a seller offered to sell at a and it was (not) accepted.

And there is additional information in the environment. First, an ask above a that was accepted is

evidence that an ask at a will be accepted. Also, a bid above a that is made previously is

evidence that ask a will be accepted. A similar intuition describes the history of past bids.

Implementation

Definition 8: Remembered History. For H n n, we make the following definitions.

Trade function. For a vector Hn, define a function T: by setting

Tk ( H n ) I {hk ,1hk , 2 0} ( H k ). Then each component Tk of T indicates whether a trade

occurred in the kth element of history.

Number of trades. Let x = (x1, x2 ,..., xn). For each n, define Sn: {0,1}n N by

n
setting Sn(x) = k 1
x k . Then Sn(T(Hn)) is the number of trades resulting from the

first n messages.

Remembered history. Let L be the memory length of a given agent. For fixed n

and Hn, to simplify notation, let S = Sn(T(Hn)). Let n be the position of trades (S
~
L) for S > L, and let n = 0 if S L . Define H n( L ) {hn '1, hn '2 ,..., hn } .

Deletion of oa and ob from history. Let n = max {k:Tk(Hn)) = 1}; that is, n is the

index of the most recent trade. Let n * be the index of the lowest (most recent) ask

in the vector {hn+1, hn+2, ..., hn}. Let n* be the index of the highest bid in the

vector {hn+1,. ..., hn}. Note that if Tn(Hn) = 0, hn,3 is either the outstanding ask or

the outstanding bid as a consequence of the spread reduction rule, and if Tn(Hn) =

1, then there is no outstanding bid and no outstanding ask. If Tn(Hn) = 0 and hn,1 =

34
 0, then hn,3 is the outstanding bid. If hk ,1 0 for some k {n"1,..., n 1} , then

n * {} , and we define H n( L ) by H n( L ) {hn '1 , hn '2 ,.hn* 1 , hn* 1 ,.., hn1} ; that is,

~
H n( L ) is H n( L ) with hn* removed. The other case, in which hn,3 is the outstanding

ask, is treated similarly, because it is not known at time n if the outstanding ask or

bid will be accepted. Then H n( L ) is the history remembered by agents with

memory length L, who observe the history Hn.

Sets of asks and bids. Let Dn( L ) be the set of all asks and bids that have been made

in H n( L ) , i.e., Dn( L ) k{n '1,...,n}\{n ,n* }{hk ,3}.

*

Definition 9: Ask Frequencies. For each d Dn( L ) , let A(d) be the total number of asks

that have been made at d, and let TA(d) be the total number of these that have been

accepted. Let RA(d) = A(d) TA(d) be the rejected asks at d.

For A(d), the counting procedure is as follows. For each k {n + 1,..., n}\{ n * , n* }, if

hk,3 = d, hk ,1 0 , and hk,2 = 0, then A(d) is incremented by 1. If hk,3 = d and Tk (hk) =1, then hk is

either a taken ask or a taken bid. To determine which is the case, find

m* min{m 1 : hk m,3 hk ,m }. If hk m* ,1 0 , then A(d) and TA(d) are incremented by 1. The

rejected asks at d are given by RA(d) = A(d) TA(d).

Definition 10: Bid Frequencies. For each d Dn( L ) , let B(d) be the total number of bids

that have been made at d, and let TB(d) be the total number of these that have been

accepted. Let RB(d) = B(d) TB(d) be the rejected asks at d. The interpretations and

counting procedures for B(d), TB(d), and RB(d) are analogous to those described in

definition 9 for asks.

35
 Note that in what follows, the sets of asks and bids are frequently denoted D, with the

subscripts and superscripts omitted. After n messages have been sent, the relevant set of asks and

bids is Dn( L ) , and the relevant history is H n( L ) .

Beliefs

Definition 11: Beliefs an Ask a Will Be Taken. For each potential ask a D , define

p (a) ( d a TA(d ) d a B(d )) / ( d a TA(d ) d a B(d ) d a RA(d )) . Then

p (a) is an agents belief that an ask a will be taken by a buyer. We assume that agents

always believe that an ask at a = 0.00 will be accepted with certainty and that there is

some valuation M > 0 such that p (M ) 0.

Definition 12: Beliefs a Bid b Will Be Taken. For each potential bid b D , define

q (b) (d b TB(d ) d b A(d )) /( d b TB(d ) d b A(d ) d b RB(d )) . Then q (b)

is an agents belief that a bid b will be taken by a seller. We assume that agents always

believe that a bid at b = 0.00 will be rejected with certainty and that there is some

valuation M > 0 such that q (M ) 1.

Spread Reduction Rule and Beliefs

The spread reduction rule has the effect of making the probability of a take for an ask

a oa equal to 0 (where oa is the outstanding offer from definition 4). We denote this

p (a) , where ~
modification of p (a) by ~ p(a) p (a) if a < oa and ~
p (a) 0 if a oa . Similarly,

q~(b) 0 for all b ob. These facts are incorporated into agents beliefs in the following

definition.

36
 Definition 13: Spread Reduction Rule. Let ~
p(a) p (a) I [0,oa] (a) for each a D ; that is,

~
p(a) p (a) if a < oa and ~
p (a) 0 if a oa . For all b D , let q~(b) q (b) I [ ob,M ] (b) .

p (a) is nonincreasing and the function q~(b) is nondecreasing.

Note that the function ~

Expected Surplus Maximization

At time t, an agent has a certainty equivalent ci,t. Should the agent make the ask a where a

is taken from the set of previous bids and asks (absent ob and oa), the agents expected surplus

will be

E[i ,t (a, ci ,t )] (a ci ,t ) p (a) . (A1)

The maximum expected gain over all such previous asks is given by

Maxa(ob,oa) E[ i ,t (a, ci ,t )] . (A2)

Let aj be the jth member of bids and asks in the interval that solves (A2).

Using linear interpolation, the agent assesses probability functions p between aj1 and aj

and aj+1 and solves

Maxa( a j 1 ,a j 1 ) E[i ,t (a, ci ,t )] , (A3)

where

i ,t (a, ci ,t )] (a ci ,t )
E[ p( a) (A4)

to find an a*.

On the other hand, the agent may choose to bid. Should the agent offer the bid b, the

agents expected surplus will be

E[i ,t (b, ci ,t )] (ci ,t b) q (b) . (A5)

The maximum expected gain over all bids is given by

37
 Maxb( ob,oa) E[ i ,t (b, ci ,t )] . (A6)

Let bj be the jth member of bids and asks in the interval that solves (A6).

Using interpolation, the agent assesses probability functions q between bj1 and bj and

bj+1 and solves

Maxb(b j 1 ,b j 1 ) E[i ,t (b, ci ,t )] , (A7)

where

i ,t (b, ci ,t )] (ci ,t b) q(b)

E[ (A8)

to find a b*.

Since (A3) and (A7) can result in negative surplus, the agent may choose to do no action

at all. Therefore the overall maximized surplus for agent i at time t is given by

S i ,t Max{E[i ,t (a * , ci ,t )], E[i ,t (b* , ci ,t )],0} . (A9)

Timing of Messages

At any point in time t, there is a set of messages {h1, h2, ..., ht1}. Each of these messages

was sent in the form of a bid, ask, taken bid, or taken ask by a specific agent. The message at

time t will reflect an action by one of the agents in the economy, and that action will maximize

that agents expected surplus. How fast an agent moves to take an action is assumed to be a

stochastic function of the amount of expected surplus that an agent can achieve. Let i ,t represent

the time at which agent i sends a message. Then

Si , t
P(i ,t j ,t , j i ) .
j 1 S j ,t
n

38
This probability then determines the likelihood an agents bid will be the one that appears in the

history of the double auction.

39
 REFERENCES

Arrow, K. J. 1994. Methodological individualism and social knowledge. American Economic

Review 84 (May): 19.

Ball, R., and P. Brown. 1968. An empirical evaluation of accounting income numbers. Journal of

Accounting Research 6 (Autumn): 159178.

Bollerslev, T. 1986. Generalized autoregressive conditional heteroskedasticity. Journal of

Econometrics 31: 307327.

Bonabeau, E. 2002. Agent-based modeling: Methods and techniques for simulating human

systems. Proceedings of National Academy of Sciences 99 (3): 72807287.

Bosch-Domenech, A., and S. Sunder. 2000. Tracking the invisible hand: Convergence of double

auctions to competitive equilibrium. Computational Economics 16 (3): 257293.

Bruguier, A., S. Quartz, and P. Bossaerts. 2008. Exploring the nature of trading intuition.

Working paper, California Institute of Technology.

Cason, T. N., and D. Friedman. 1993. An empirical analysis of price formation in double auction

markets. In The Double Auction Market: Institutions, Theories, and Evidence, edited by

D. Friedman and J. Rust. Reading, MA: Addison-Wesley.

, and . 1996. Price formation in double auction markets. Journal of Economic

Dynamic Control 20 (August): 13071337.

Dickhaut, J., A. Martin, D. Senkow, and G. Sevcik. 1993. The price/quantity sealed bid/offer

auction with pro-rata rationing: Experimental evidence. The Economic Journal 103

(May): 547569.

, O. Lungu, V. Smith, B. Xin, and A. Rustichini. 2008. A neuronal theory of human

economic choice. Working paper, University of Minnesota.

40
, and B. Xin. 2009. Learning about field behavior using a laboratory based model.

Working paper, Chapman University and University of Toronto.

Duffy, J. 2004. Agent based models and human subject experiments. Working paper, University

of Pittsburgh.

Easley, D., and J. Ledyard. 1993. Theories of price formation and exchange in double oral

auctions. In The Double Auction Market: Institutions, Theories, and Evidence, edited by

D. Friedman and J. Rust. Reading, MA: Addison-Wesley.

Fama, E. F. 1970. Efficient capital markets: A review of theory and empirical work. The Journal

of Finance 25 (May): 383417.

. 1991. Efficient capital markets: II. The Journal of Finance 46 (December): 15751617.

, L. Fisher, M. Jensen, and R. Roll. 1969. The adjustment of stock prices to new

information. International Economic Review 10 (February): 121.

Fiorillo, C. D., N. Philippe, P. Tobler, and W. Schultz. 2003. Discrete coding of reward

probability and uncertainty by dopamine neurons. Science 299: 18981902.

Gigerenzer, G., and H. Ulrich. 1995. How to improve Bayesian reasoning without instruction:

Frequency formats. Psychological Review 102: 684704.

Gjerstad, S. 2006. The competitive market paradox. Journal of Economic Dynamics and Control

31 (May): 17531780.

, and J. Dickhaut. 1998. Price formation in double auctions. Games and Economic

Behavior 22 (January): 129.

Gode, D., and S. Sunder. 1993. Allocative efficiency of markets with zero intelligence traders:

Market as a partial substitute for individual rationality. Journal of Political Economy 101

(February): 119137.

41
Grossman, S. 1976. On the efficiency of competitive stock markets where trades have diverse

information. Journal of Finance 31 (May): 573585.

. 1978. Further results on the informational efficiency of competitive stock markets.

Journal of Economic Theory 18: 81101.

, and J. Stiglitz. 1980. On the impossibility of informationally efficient markets.

American Economic Review 70 (June): 393408.

Hayek, F. A. 1948. Use of Knowledge in Society. Reprinted in The Essence of Hayek, edited by

C. Nishiyam and K. R. Leube. Stanford, CA: Hoover Institution.

Marcet, A., and T. Sargent. 1989a. Convergence of least squares learning mechanisms in self-

referential liner stochastic models. Journal of Economic Theory 48: 337368.

, and . 1989b. Convergence of least-squares learning in environments with hidden

state variables and private information. Journal of Political Economy 97 (December):

13061322.

OBrien, J. 1992. Experimental stock markets with controlled risk preferences. Journal of

Accounting, Auditing and Finance 7 (Spring): 117134.

Plott, C. 1982. Industrial organization theory and experimental economics. Journal of Economic

Literature 20 (December): 14851527.

, and S. Sunder. 1982. Efficiency of experimental security markets with insider

information: An application of rational expectations models. Journal of Political

Economy 90 (August): 663698.

Ratcliff, R. 1978. A theory of memory retrieval. Psychological Review 85 (2): 59-108.

Sargent, T. 1993. Bounded Rationality in Macroeconomics. New York: Oxford University Press.

42
Smith, V. L. 1962. An experimental study of competitive market behavior. Journal of Political

Economy 70 (April): 111137.

. 1982. Microeconomic systems as an experimental science. American Economic Review

72 (December): 923955.

Tremblay, L., and W. Schultz. 1999. Relative reward preference in primate orbitofrontal cortex.

Nature 398 (April): 704708.

Veronesi, P. 1999. Stock market overreaction to bad news in good times: A rational expectations

equilibrium model. The Review of Financial Studies 12 (Winter): 9751007.

Wilson, R. B. 1987. On equilibria of bid-ask markets. In Arrow and the Ascent of Modern

Economic Theory, edited by G. W. Feiwel. New York: New York University Press.

43
 TABLE 1
Distributions of States, Signals, and Forecasts

A. State Characteristics
Possible states 40 80
Prior probabilities of states 0.2 0.8

B. Conditional Probabilities of Public Signal

Given States
State = 40 State = 80
Public Signal = 40 0.8 0.2
Public Signal = 80 0.2 0.8

C. Conditional Probabilities of Private Signal

Given States
State = 40 State = 80
Private Signal = 40 0.8 0.2
Private Signal = 80 0.2 0.8

D. Conditional Probabilities of Public Signal

Given State and Private Signal
State = 40
Private Signal = 40 Private Signal = 80
Public Signal = 40 0.9 0.7
Public Signal = 80 0.1 0.3
State = 80
Private Signal = 40 Private Signal = 80
Public Signal = 40 0.3 0.1
Public Signal = 80 0.7 0.9

44
 TABLE A1
Information Structures in the Computational Economies

Scenario 1
Initial beliefs, Private information, Public information fully redundant with
P PL|L PH |L private information,
S L SM 1
PL|L PH |L
PH PL|H PH |H SM 2M 1 SM 1
P P
L| H H | H
Scenario 2
Initial beliefs, Private information, Public information partially correlated with
P PL|L PH |L private information,
S L SM 1
PL|L, L PL|H , L
PH PL|H PH |H SM 2M 1L ,
P P
H | L , L H | H , L
PL|L , H PL|H , H
SM 2M 1H

PH |L , H PH |H , H
Pi| j is the conditional probability of message i given state j.
Pi|j,k is the probability that the public signal is i given that the private signal is j and the state
is k.

45
 FIGURE 1

Market Efficiency and Drift

A. Semistrong Form Market Efficiency (from FFJR 1969)

B. Post Announcement Drift (from Ball and Brown 1968)

46
 FIGURE 2

The Model

Agent 1 Agent 2 ,.,,.,.,. Agent .i-1 Agent i Agent i+1 .,..,.,, Agent n-1 Agent n

New private Determines Determines New private

information value based Determines information
value based
arrives on r1 on ri Step 1 value based
arrives
on rn

Assesses Assesses Assesses

conditiona l conditiona l conditiona l
No new No new
probabilities of probabilities of probabilities of
information information
bids and offers bids and offers
being accepted
Step 2 bids and offers
being accepted
being accepted
based on history based on history based on history

Determines action Determines action Determines action

( bid, ask or take) ( bid, ask or take) ( bid, ask or take)
that yields ones that yields is Step 3 that yields ns
optimal profit,S1. optimal profit, Si. optimal profit, , Sn

Attempts to Attempts to
place order in place order in
the market. the market.
The chances the ith player is the first
to the market are is profits divided by the sum of
total profits of all players.
Si
S1+ . . + Si + . . + Sn
Step 4
This new order becomes either the outstanding
bid/ask or a take of the existing outstanding bid/ask

If market still open If market still open

Market closed
 FIGURE 3

Messages from Computational Economy 1

A. Double Auction Messages (Bids and Asks)

Double Auction Me ssage s

B i dAsk

54

52

50

48

46 Asks

44

42

B ids
40 Hi story
0 200 400 600 800 1000

B. Pricing Messages

Public Inform ation Efficie ncy

Pri ce
47

46

45

44

43

42

41

40 Hi story
0 200 400 600 800 1000
 FIGURE 4

Messages from Computational Economies 2 and 3

A. Price Messages from Computational Economy 2

Public Inform ation Drift

Pri ce
47

46

45

44

43

42

41

40 H i story
0 200 400 600 800 1000

B. Price Messages Showing Volatility from Economy 3

Price Volatility and Priors

50
Low
Information 41.6
Release 41.4 Low Information Release
48 41.2
41.

Volatility at Low .2 P rior .036

40.8 P ost Announcement
46
40.6
40.4
40.2
44
40.
50 100 150 200

42

Volatility at High .8 Prior .155

Post Announcement

Trade s
50 100 150 200

49
 FIGURE 5

Distributions of the Regression Coefficients of Prices against Time

N FFJR Pri vate N FFJR Pu blic

50
25
40 20
30 15

20 10

10 5

-0.04 -0.03 -0.02 -0.01 0 -0.002 -0.001 0

N B B Private N B B Pu blic
50 20
40
15
30
10
20

10 5

-0.04 -0.03 -0.02 -0.01 0 -0.002 -0.001 0

50
 FIGURE A1

Timeline of Information Scenarios

51