Expectation formation in an experimental

foreign exchange market

Johannes Leitner1 , and Robert Schmidt2
1
Department of Statistics and Operations Research, University of Graz
Universitätsstrasse 15/E3, 8010 Graz, Austria
e-mail: [email protected]

2
Department of Economics, University of Würzbrug
Sanderring 2, D-97070 Würzburg, Germany
e-mail: [email protected]

Abstract
Participants of an experimental foreign exchange market forecast an ex-
change rate with an unknown price reaction function. Aggregate demand
is derived from their own forecasts and random shocks. Our experimental
results indicate that the expectations of the subjects tend to be coordinated
on a common prediction strategy. This strategy is best described as a trend-
extrapolative, destabilizing expectation formation scheme. Deviations from
common expectations are mainly caused by random shocks, which can be
ascribed to the similarity of the subjects’ behavior within and between the
different markets. The findings can be explained using insights of behavioral
economics.

Key words: Foreign exchange market, experimental economics, behavioral

finance, expectation formation.

1 Introduction
According to economic theory, exchange rates are basically determined
by rational expectations about future fundamentals. However, the empiri-
cal support for such models is rather poor. The empirical failure of economic
exchange rate models is often explained either by the irrational behavior or
excessive speculative trading activities of at least some market participants.
Keynes (1936) was one of the first who analyzed the impact of a speculative
market environment on the individual expectation formation. He argued
that, in speculative markets, in addition to fundamental considerations,
non-fundamental factors become more important. In particular, the expec-
tations of other market participants concerning the future exchange rate
become relevant. According to Keynes (1936), participants in speculative
markets are mostly concerned with the anticipation of the expectations of
the other market participants.1 Keynes (1936) calls this information of the
third degree. A reasonable behavior in such decision-making situations is to
be geared to existing market conventions. The present study is concerned
with the experimental investigation of expectation formation in the context
of foreign exchange markets. In particular, we are interested in the indi-
vidual behavior in a market environment that is characterized by feedback
loops.
Related research on expectation formation can be classified into forecasting
experiments and experimental markets. In forecasting experiments partici-
pants judgmentally predict future values of a time series based on its past
realizations. The behavior of the participants is analyzed by testing hy-
potheses of expectation formation. In the majority of the studies, the time
series are generated by linear autoregressive processes (Hey (1994)) or they
are pure random walks (Dwyer et al. (1993), Beckman and Downs (1997)).
For instance it was shown that subjects are able to behave rationally when
forecasting random walks but fail to do so with more complex time series.
In an experiment by Becker et. al (2007) subjects were given several indi-
cators, i.e. time series with a lead of one period, for the forecast of another
time series. The authors presented a heuristic that models average fore-
casting behavior of the subjects much better than the rational expectations
hypothesis.
Only a few studies have been mentioned, but these generally have in com-
mon that they do not account for expectation feedback as the applied time
series are fixed in advance.
Although there is numerous literature on experimental markets (see Sun-
der (1995) for a review, Williams (1987) and Smith et al. (1988)) only
a few of the studies are focused on foreign exchange markets. Arifovic
(1996) explored the celebrated result that the exchange nominal rate is not
determined in an overlapping generations model. He found weak support
for the theory in an experimental market. Noussair et al. (1997) tested
a two-country model with a real side and two cash-in-advance constraints.
These authors also reported mixed evidence for simple elements of exchange
rate theory. Fisher and Kelly (2000) studied essentially identical assets in
a non-stationary environment and showed that cross-asset asset arbitrage
held, even though every asset had a significant bubble. This result was
interpreted as support for simple exchange rate theories. Even though sub-
jects did not perform backward induction, they understood that asset prices
were related, and the bubbles for the different assets were almost perfectly
correlated.
Although all of these experiments include a dynamic feedback component,
the pure analysis of expectations hypotheses is not possible. The observed
market behavior also includes other behavioral features that may be due to
e.g. trading activity. In our experiment, we explicitly consider expectation
1 Note that both reasons are closely interrelated: holding excessive speculation respon-

sible for deviations of the exchange rate from its fundamental level implicitly assumes
that speculation is not based on fundamental, and thus rational, considerations.
feedback. Individuals’ expectations directly influence the actual realization
of the exchange rate, i.e. the time series they forecast. This is an impor-
tant feature of the experimental design in order to accurately reproduce the
decision-making situation in financial markets. We focus on the expecta-
tions of subjects to exclude undesirable effects, e.g. trading decisions. There
are only a few studies with similar characteristics. Gerber et al. (2002) in-
vestigate forecasting behavior in a beauty contest experiment. According to
their experimental setting, participants are only able to sell and buy an as-
set whose price is determined by aggregated orders and a noise component.
Although the framework is quite similar to our experimental setting, it is
rather abstract and expectations are not explicitly analyzed. We closely re-
late our experimental setting to Hommes et al. (2002, 2005) and Heemeijer
et al. (2004) who explore expectation formation in an asset market with
expectations feedback. In their experiments, participants forecast the fu-
ture price of an asset which is given as a function of the average forecasts
of all subjects, a noise component and (in a few treatments) the forecasts
of some fundamentalist computer traders who always forecast the funda-
mental value of the asset. Hommes et al. (2002, 2005) report that in many
of the experiments extreme speculative bubbles arise, although all partic-
ipants can easily compute the fundamental value under perfect knowledge
of dividends and interest rates. Their results can be ascribed to the specific
price function of the asset that requires participants to make forecats two
periods ahead. This favors tendencies to extrapolate trends. This is also
the case when the price is forced into the direction of the fundamental value
by artifical traders. Heemeijer et al. (2004) basically only change the price
function of this experiment by collecting one period ahead forecasts. The
participants still tend to create bubbles in their forecasts but these are not
as extreme as in the Hommes et al. experiments.
We apply the feedback design to a simplified cooperative exchange market.
Hommes et al. and Heemeijer et al. find that individuals within a market
coordinate on a common prediction strategy such as naïve, adaptive or au-
toregressive expectations. While these authors notice strong coordination
within the markets, we notice that the market developments between the
groups are also extremely similar. The similarity is caused by the character-
istics of the random component in the price reaction function and implies
coordination on a higher level.
In the next section we discuss the exchange rate behavior and explore the
impact of a speculative market environment on the expectation formation
in more detail. We emphasize the relevance of psychological factors and
the influence of uncertainty. Section 3 is concerned with the experimental
setup. In section 4 the experimental investigation of expectation formation
in a foreign exchange market context in investigated. We first analyze the
aggregated market behavior and afterwards examine the individual behav-
ior in experimental foreign exchange markets. In section 5 we discuss our
findings.
2 Functioning of Speculative Asset Markets
Excessive speculation is often held responsible for the poor performance of
fundamental-oriented exchange rate models. According to surveys of foreign
exchange traders, the excessive speculation of market participants hinders
the determination of exchange rates by macroeconomic fundamentals so
that, in speculative markets, the importance of macroeconomic fundamen-
tals obviously drops (see e.g. Cheung and Chinn, 2001).
How does speculation affect the behavior of foreign exchange traders? Com-
mon definitions of speculation illustrate that speculation usually includes
two central attributes: i) speculative transactions always rest upon expecta-
tions concerning the future development of the corresponding market value,
and ii) as the future development of the market value cannot be predicted
for certain, speculative transactions are always accompanied by a high de-
gree of uncertainty so that speculation is always risky. An analysis of both
attributes of speculation will help to understand why excessive speculation
may cause deviations from fundamental levels.

2.1 Behavioral Economics and Expectations

The first attribute of speculation is concerned with the nature of expec-
tations. The impact of speculation on exchange rates depends primarily
on the way market participants form their expectations. Traditionally, the
economic literature assumes that expectations about future exchange rates
are formed rationally, i.e. speculators base their speculative engagements
on the fundamental value of currencies and maximize their wealth. This
point of view goes back to the influential contributions of Friedman (1953),
who suggests that rationally acting market participants ensure that prices
correspond to fundamental values. Within this view, it is impossible for
"irrational" speculators to influence exchange rates substantially as they
would suffer permanent losses. In recent times, more and more economists
critically scrutinize the traditional view of the impact of speculation. On the
one hand, there exists a strong theoretical doubt that rational speculators
always stabilize exchange rates. This criticism is mainly attributed to the
limits of arbitrage induced by the institutional conditions of speculators in
foreign exchange markets (see Shleifer, 2000). On the other hand, recent re-
search raises serious doubts about the rationality assumptions in economics.
According to the Behavioral Economics approach, actual human behavior
deviates systematically from the predictions of economic rationality due
to cognitive limitations. The new research field of Behavioral Economics
explicitly considers these limitations and tries to develop a more realistic
view of human behavior in the context of economics. Thereby, behavioral
economists often refer to the work of behavioral decision theorists who in-
troduced new concepts under the general heading of Bounded Rationality.
The concept of Bounded Rationality is mainly associated with the work
of Herbert Simon (see Simon, 1955). According to Simon, Bounded Ra-
tionality consists of two interlocking components: first, the limitations of
the human mind and, second, the structure of the environments in which
the mind operates. The first component requires that models of human
judgment and decision-making deal with the actual functionality of human
mind. In this context, Bounded Rationality is used to indicate a reasonable
(procedurally rational) decision-making behavior that takes into account the
cognitive limitations of the decision-maker (Simon, 1997). The second com-
ponent of Bounded Rationality emphasizes the importance of considering
the environmental structure because it can explain when and why a specific
behavior performs well (Gigerenzer and Todd, 1999a). Against the back-
ground of the limited cognitive resources, psychologists have investigated
the processes by which decision-makers reach their conclusions. Psychol-
ogists argue that human beings tend to use simple heuristics in complex
decision-making situations. A simple heuristic can be characterized as a
simple rule of thumb which allows quick and efficient decisions even under
a high degree of uncertainty (see Fiedler and Bless, 2001). As Gigerenzer
and Todd (1999b) have shown, there are many simple heuristics that pro-
vide a good compromise between economic rationality and an efficient use
of scarce human cognitive resources.

2.2 The Role of Uncertainty in Foreign Exchange Mar-

kets
The second attribute of speculation is concerned with the risk that is
associated with speculative activities. The risk is due to the uncertainty
speculators are faced with, whereby uncertainty arises from two different
sources: First, speculators are uncertain about the future development of
macroeconomic fundamentals, and, second, speculators are uncertain about
the evaluation of other market participants concerning the relevant exchange
rate. Thus, it is useful to distinguish between two different kinds of uncer-
tainty in this context. The first kind of uncertainty can be characterized
as exogenous uncertainty as it is more or less independent of the action
of speculators (see Muthoo, 1999). The second kind of uncertainty can be
characterized as endogenous uncertainty. It is attributed to the impact of
actions chosen by some other market participants. In this case, the state
of the nature is determined strategically by some other traders because the
state of the nature is a strategy choice of some other traders (Muthoo,
1999). As the characteristic of this uncertainty depends crucially on the
actual behavior of market participants it can also be denoted as behavioral
uncertainty. The existence and prevalence of behavioral uncertainty is due
to the capacity of individuals to adapt and react to one another in a non-
negligible manner (Pesaran, 1988). The extent of behavioral uncertainty is
closely related to the degree to which individuals may be able to influence
the actions of others by their own actions, and also to which extent they
are themselves influenced by the actions of others. Obviously, the influ-
ence of behavioral uncertainty on the individual’s decision-making behavior
increases with the degree of speculation in foreign exchange markets and
appears to be of particular importance. A very prominent and illustrative
description of the individual decision-making behavior in an environment
that is primarily characterized by behavioral uncertainty was first given by
Keynes (1936). He compared the individual decision-making situation in
an asset market with the decision-making situation in a newspaper beauty
contest. The objective of the beauty contest is to guess which picture will
get the most votes. Thereby, the appropriate behavior for the participants
is not to choose the picture they favor, but to elect the pictures favored by
most other participants. Consequently, they have to estimate "what average
opinion expects the average opinion to be" (Keynes, 1936, p. 156). Keynes
describes this purpose as "information of the third degree". The implication
of Keynes’ parable is that an understanding of financial markets requires
not only an understanding of market participants’ evaluation of assets’ fu-
ture returns, but also an understanding of market participants’ evaluation
of other market participants’ evaluation and higher order evaluations (Allen
et al., 2003).
From a theoretical point of view, the solution of the Beauty Contest Prob-
lem can be found in the concept of "focal points". Focal points are cues
that induce people to behave in a similar manner. The focal solution of a
decision-making problem under behavioral uncertainty emerges because a
group of people has come to believe that the members of this group will
behave consistently with this equilibrium, but which solution is a priori un-
known and depends on the co-ordination problem and the decision-making
environment (see Young, 1996). Arrow (1986) notes that in this context the
solution of such decision-making problems asks for a rationality that refers
to social phenomena. As soon as one particular solution is known to be
focal it becomes reasonable or "rational" for each decision-maker to expect
that all others will decide consistent with this solution and to act on this
expectation. An important characteristic of focal points is that nobody has
an incentive to change once a common expectation has been established.
Consequently, the expectations, which rely on existing focal points, feature
a high degree of persistence (see Duncan and Isaac, 2000). Closely related
to the concept of focal points is the concept of convention. A convention is
typically defined as "a pattern of behavior that is customary, expected and
self-enforcing" (Young, 1996, p. 105). The main characteristic of convention
is that everyone conforms to it, everyone expects others to conform to it, and
everyone has good reason to conform because conforming is in each person’s
best interest when everyone else plans to conform (Young, 1996). The main
feature of a convention is that, out of a host of conceivable choices, only
one is actually used. Hence, conventions solve problems of indeterminacy in
interactions that have multiple equilibria. In general, one may discern two
ways in which conventions become established: first, by authority, and sec-
ond, by the gradual accretion of precedent. At first glance, in the context of
asset markets, the second process by which conventions may emerge seems
to be of particular interest. The idea is that, in a repeated co-ordination
game, one particular way of interaction emerges successively as a superior
way of resolving the game. Thus, it reaches a greater degree of prominence,
which in turn entails that more people notice it, which leads to more people
using it, and so forth. Consequently, a positive feedback loop is created.

2.3 Consequences for Speculative Asset Markets

The foregoing subsections denote two interrelated effects which are, in
our opinion, inevitably linked with speculative trading in foreign exchange
markets. First, there is a high degree of endogenous, behavioral uncertainty.
This may force market participants to base their own trading activities on
the average assessment of the market. In the context of foreign exchange
markets, it seems likely that the basic attitude of market participants is
rather stable, i.e. market participants judge a certain currency for a certain
time as strong regardless of the fundamental development. This rating of
market participants consequently results in a strong trend behavior of ex-
change rates. In our view, such conventions about the actual strength of
a currency help market participants to orientate in an environment that is
mainly characterized by behavioral uncertainty.
Second, due to cognitive limitations market participants are obliged to
use simple heuristics in their very complex decision-making situation. As
Gigerenzer and Todd (1999b) argue, people tend to make use of such sim-
ple heuristics that provide accurate solutions for a complex decision-making
situation due to their adaptiveness. Thus, an important characteristic of a
reasonably simple heuristic is that it considers the environmental structure.
In the context of foreign exchange markets, trends seem to characterize the
environmental structure very well. Consequently, we assume the existence
of simple trend heuristics in foreign exchange markets. Such a heuristic
would have the advantage of saving cognitive costs by leading to a reason-
able performance at the same time.

3 Experimental design
The experimental design of the foreign exchange market focuses on the
exploration of the expectation formation of the subjects. Consequently, the
only task of the subjects is to forecast the exchange rate development. No
trading activity is involved.
In order to make it clear to the subjects why their forecasts have an influ-
ence on the development of the exchange rate, the experiment is described
as follows: The subjects are trading floor economists in different leading
European banks. They watch the Euro/US Dollar exchange rate and brief
the currency dealers of the bank at the beginning of each period about the
expected development of the exchange rate. The participants are told that
the trading decisions of the dealers and consequently the demand for Euros
and US Dollars are exclusively based on their forecasts. Thus, profits from
foreign exchange trading activities realized by the bank are strongly related
to the quality of the individual forecasts, and therefore the payoffs of the
subjects in the experiment are inversely proportional to their forecast er-
rors.
Besides this information about their task, the subjects are given background
information about the exchange rate market. They know that the exchange
rate is influenced by the forecasts of the other participants in the experi-
ment and slightly influenced by the demand of private investors. However,
the subjects do not know the exact market equilibrium equation. They also
know values of domestic and foreign interest rates and expected inflation
rates, which give them the possibility of calculating the fundamental values
of the Euro/US Dollar exchange rate.
The price reaction function generating the experimental exchange rate St is
given by

St+1 = St + ∆Etavg St+1 + ξt+1 (1)

avg 1
Pn i
where ∆Et St+1 = n i=1 Et St+1 − St denotes the average forecasted
change of n individual subjects i (i = 1, . . . , n) made in period t for the
exchange rate in the period t + 1. ξt+1 is an independently normally dis-
tributed error term with ξt+1 ∼ N (0, 1) and represents random demand and
supply shocks from private investors. The initial value of the exchange rate
S1 equals 50. The fundamental value S f is given by the standard Fisher
relation
f
St+1 = St + St (rt − rt∗ ) = St + St (πt − πt∗ ) (2)
where r (r∗ ) denotes the domestic (foreign) interest rate and π (π ∗ ) denotes
the expected domestic (foreign) inflation rate. The values of these variables
were chosen such that a rather stable fundamental value of about 50 is
ensured. The values of the four variables were presented to the subjects in
each period.
In each period t the task of the subjects is to forecast the value of the
exchange rate in period t + 1. The information set of subject i consists of
the realizations of the exchange rate St , St−1 , . . . , S1 and his/her own past
i i
forecasts Et−1 St , Et−2 St−1 , . . . , E1i S2 . After all participants have completed
their forecast for period t + 1, the actual exchange rate of period t + 1 is
presented on the screen and subjects are asked for the next forecast. This
procedure is repeated for 49 periods. All past values of the exchange rate
and the participant’s own forecasts are graphed in different colors and are
additionally presented in a tabular format. Furthermore, the information
about the expected domestic and foreign inflation rates and interest rates
are presented on the screen. The subjects are told that their forecasts have
to be between 0 and 100.
The participants were recruited from an undergraduate course of business
administration at the Karl Franzens University Graz. They did not have
any special education in financial markets and they had never participated
in a similar experiment before. They were not offered any course credit for
their participation. The reward scheme of the subjects is only related to
their performance in the experiment. The payoff pit of subject i in period t
is given by
pit = max{30 − 10|St − Et−1
i
St |; 0} (3)
where the unit of pt is cents. For a perfect forecast, the subjects are remu-
nerated 30 Cents, for a deviation or more than 3 units, the payment equals
0. The average payoffs were 6.94 Euros for an average duration of about 45
minutes.
Altogether, 36 students participated in the experiment. The parameter n
was set to 6 in all experiments, i.e. each market consists of six subjects.
The random shock sequence ξt was identical in each of the six markets.

4 Experimental results
Our analysis of the experimental results is divided into two parts. We
first examine the aggregated behavior of experimental exchange rates. In
this context, we are particularly interested in the observable exchange rate
developments. Furthermore, our focus is concerned with the efficiency of
experimental exchange rate markets. In the second subsection, we analyze
the individual behavior of the participants in the experiment in more detail.
We are interested in the rationality of experimental expectations and the
nature of those expectations.

4.1 Aggregated behavior in experimental foreign ex-

change markets
4.1.1 Similarity of experimental exchange rates
Figure 1 shows the realized exchange rates of the six experimental groups
and the development of the fundamental value. The most striking feature
of the results is the obvious similarity of the group behavior. The exchange
rates start with an upswing, fall in period 5 and stabilize around periods
12 to 21. The exchange rates continue their downtrend until period 30 and
turn upwards for the final 20 periods. This applies for almost all groups,
only group 2 is inconsistent with this pattern. This is due to several ex-
treme forecasts of one subject in the periods 7, 14, 31 and 33. This can
be seen from the charts of individual forecasts in Figure 1. To underline
these findings, the correlation of the exchange rates are presented in Ta-
ble 1. The correlation coefficients are highly significant and indicate very
strong associations between the groups. The correlations with group 2 are
comparatively weaker, but still significant and at a minimum level of 0.614.
70

t
u
t
u
t r ru
u t
60 r
r ut
r ut u
t
t r
u r r r
r ru
r u
t
t t
u
r
r
t
u t
u
t
ru t
u
rr u t
t u ut u u u
t t t u
t
50 u t
t u u
t
t t
u u u
t
r
t
u
r r
ru
t r
r u
t
t
u u
t
r
r rr r r r r r r t
u t
u
40 t u
u t ru r r u
t t
u t
u
t
u
r t ut t
u u r
t u
t t
u
t
u r t
u
t
u t
u r
t rr rr
u u
t t t
u
u
t t
u t u
t u
u u r r
t r
r r
u
t r
t
u r r t
u r
r u
t t
u u t
t u t
u r r
t
u r u t
t r
30 r u
r
r
t
u
t
u r u
t
r
t r
u r t
u t
u
r r
t
u
t
u u
t
r r r r r t
u r
t
u r r
r r r t
u r
20 r t
u t
u
r r r t r r r
u
Group 1 t r r
u t r
ru
r r
Group 2 t
u u t r r
t u r
r r t
u
t u
u t t
u
r r Group 3 t u
u t u
t

10 r r Group 4
t
u tGroup 5
u
t
u tGroup 6
u

Fundamental value
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
100 100

90 Group 1 90 Group 2

80 80

70 70

60 60

50 50

40 40

30 30

20 20

10 10

0 0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
100 100

90 Group 3 90 Group 4

80 80

70 70

60 60

50 50

40 40

30 30

20 20

10 10

0 0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
100 100

90 Group 5 90 Group 6

80 80

70 70

60 60

50 50

40 40

30 30

20 20

10 10

0 0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49

Figure 1: Realized exchange rates and individual expectations

 The reason for this astonishing result can be found in the structure of
the random shocks ξ of equation (1). We calculate the average changes of
the predictions ∆Etavg St+1 = Etavg St+1 − Et−1
avg
St and compare them to the
random shocks ξt . These values are highly correlated at a 99% level of sig-
nificance. The corresponding correlation coefficients are reported in Table
1. Again, only group 2 shows a weaker association.

Group 1 2 3 4 5 6
2 0.614** - - - - -
3 0.846** 0.823** - - - -
4 0.953** 0.739** 0.952** - - -
5 0.891** 0.803** 0.992** 0.977** - -
6 0.901** 0.777** 0.968** 0.955** 0.968** -
Shock 0.881** 0.653** 0.962** 0.937** 0.969** 0.900**
** denotes that correlation is significant at the 0.01 level (two tailed)

Table 1: Correlation between aggregated group forecasts and correlation

between forecasted changes and random shocks

The development of the exchange rate is obviously driven by the random

shocks to a high degree. The occurrence of local extrema and longer trends
can be ascribed to the number of random shocks. The periods 12 to 21 are
representative for this phenomenon. The random shocks are very small in
five periods successively. They amount to absolute values of about 0.5. In
periods 21 and 22, random shocks of -3.33 and -2.06 occur, respectively, and
start a downtrend that lasts for about ten periods. Between the periods 21
and 31 the random shocks are negative in nine cases, and a large positive
shock stops the trend. The exchange rates have many characteristic local
minima and maxima in common due to the same random shocks in all ex-
periments. This is especially noticeable in the final periods, e.g. periods
43, 45 and 46. The changes of average forecasts correspond to the sign of
the shocks in almost all periods. Only in group 2 can larger deviations be
observed, due to the reasons discussed above.
It can be concluded that the subjects extrapolate the trends that are mainly
caused by random shocks. The influence of the shocks is remarkably stable
for all experiments. Furthermore, our findings of mainly shock-driven ag-
gregated forecasts imply that the fundamental value of the exchange rate
does not have a noticeable influence on the realized prices in all groups.

4.1.2 Efficiency of experimental foreign exchange markets

A central building block of traditional exchange rate models is the as-
sumption that exchange rates are determined in efficient markets. A testable
implication of the efficient market hypothesis is that excess returns xt =
ln(St ) − ln(St−1 ) should be serially uncorrelated with any excess return in
the past or future.
To analyze the autocorrelation of excess returns we carry out a variance
ratio test proposed by Lo and MacKinlay (1988). The variance ratio test
exploits the fact that the variance of the increments in a random walk time
series must be linear in the sampling interval. That is, if a time series follows
a random walk process, the variance of its q-differences would be q times
the variance of its first difference:

var(St − St−q ) = q · var(St − St−1 ) (4)

Therefore, the relevant null hypothesis of the variance ratio test implies that
the following ratio is equal to one:

σ 2 (q)
VR= = 1, (5)
σ2
where σ 2 (q) is 1/q times the variance of the q-th period returns and σ 2 is
the variance of the one-period returns. With regard to the autocorrelation
structure of the time series, Cochrane (1988) shows that the variance ratio,
V R(q), can be approximated by the following expression:
q−1
X q−j
V R(q) ∼
=1+2 = ρbj , (6)
j=1
q

where ρbj denotes the q-th order autocorrelation coefficient estimator of the
first differences of Xt (see Cochrane, 1988). Equation (6) provides a simple
interpretation for the variance ratios computed with an aggregation value q:
They are approximately linear combinations of the first q−1 autocorrelation
coefficient estimators of the first differences with arithmetically declining
weights (see Lo and MacKinlay, 1999). Thus, variance ratios larger than
unity indicate the presence of positive serial correlation in the series, which
is consistent with trend behavior in exchange rate series and, in contrast,
variance ratios smaller than unity suggest the presence of negative serial
correlation, which is consistent with a mean reverting behavior in exchange
rate series.
The results for the variance ratio test are summarized in Table 2. Almost
all V R(q)’s are larger than 1, indicating a tendency for trend behavior in
the exchange rate. Only the V R(3) for the "shock price" is well below 1.
However, the V R(q)’s for group 1 and group 2 are on the whole not statis-
tically significant, so we have to conclude that those experimental exchange
rates do not exhibit an autocorrelation pattern which can be exploited for
abnormal trading profits. In contrast, the results for the V R(q)’s of the ex-
perimental exchange rates in group 3 to 6 show distinct evidence for positive
autocorrelation in the excess returns. All V R(q)’s are statistically signifi-
cant above one and thus indicate a strong trend behavior in those exchange
rates. Skeptics might suspect that the positive serial correlation in the ex-
cess returns is due to the chosen shock sequences. We therefore also analyze
the autocorrelation of the exchange rate, which is only determined by the
shock sequence. The results indicate that this exchange rate exhibits no
tendency for significant autocorrelation patterns in excess returns. Thus,
the positive autocorrelation in most of the experiments is probably caused
by the behavior of the participants in the experiment.

Group 1 2 3 4 5 6 Shock
VR(2) 1.154 1.060 1.314 1.493 1.481 1.406 1.005
T-stat 1.077 0.417 2.199 3.452 3.365 2.844 0.033
(hom.) (0.282) (0.677) (0.028) (0.001) (0.001) (0.005) (0.974)
T-stat 1.342 0.414 2.046 3.112 3.014 2.760 0.038
(het.) (0.180) (0.679) (0.041) (0.002) (0.003) (0.006) (0.970)
VR(3) 1.099 1.201 1.496 1.765 1.829 1.550 0.942
T-stat 0.459 0.945 2.328 3.592 3.891 2.583 -0.273
(hom.) (0.646) (0.345) (0.020) (0.000) (0.000) (0.010) (0.785)
T-stat 0.531 0.888 2.171 3.288 3.483 2.594 -0.302
(het.) (0.595) (0.374) (0.030) (0.001) (0.001) (0.010) (0.763)
VR(4) 1.213 1.429 1.760 2.058 2.211 1.691 1.123
T-stat 0.797 1.606 2.843 3.960 4.530 2.585 0.465
(hom.) (0.426) (0.108) (0.005) (0.000) (0.000) (0.010) (0.642)
T-stat 0.881 1.486 2.643 3.671 4.077 2.661 0.491
(het.) (0.378) (0.137) (0.008) (0.000) (0.000) (0.008) (0.623)
VR(5) 1.290 1.561 1.874 2.289 2.490 1.679 1.137
T-stat 0.925 1.793 2.791 4.119 4.759 2.170 0.439
(hom.) (0.355) (0.073) (0.005) (0.000) (0.000) (0.030) (0.661)
T-stat 1.009 1.674 2.602 3.855 4.325 2.275 0.454
(het.) (0.313) (0.094) (0.009) (0.000) (0.000) (0.023) (0.650)
Notes: p-values are given in parenthesis.
hom. denotes the standard normal test statistic under homoscedasiticity
het. denotes the heteroscedasticity-consistent test statistic

Table 2: Variance ratio test

Overall, we have to reject the hypothesis of efficient markets in the con-

text of our experimental analysis in 4 of 6 cases. The rejection of the efficient
market hypothesis is always due to positive autocorrelation in excess returns
suggesting trend behavior. Furthermore, the rejection of the efficient mar-
ket hypothesis cannot be linked to the shock sequence.

4.2 Individual Behavior in Experimental Foreign Ex-

change Markets
4.2.1 Rationality of Individual Expectations
For asset pricing models, the rational expectations hypothesis (REH) is
of crucial importance. The hypothesis states that agents’ subjective expec-
tations with respect to a variable (e.g. exchange rates) are equal to the
mathematical expectations conditional on an information set containing all
publicly available information. The conditional mathematical expectation
value is thereby based on the true probability distribution of the variable.
Within asset pricing models the conditional mathematical expectation value
is derived from the ’true’ economic model and it is assumed that the market
participants behave as if they form their subjective expectations as the con-
ditional mathematical expectations, so that subjective and objective expec-
tations coincide. A major drawback of the rational expectations hypothesis
is that it makes implausible demands on the human information processing
capabilities.
The REH first implies that forecast errors of rational subjects (et ) condi-
tioned on the available information set (Ωt ) should be purely random,

et+1 = St+1 − E(St+1 |Ωt ), with et+1 ∼ (0, σ 2 ) (7)

where S denotes the nominal spot exchange rate and E is the rational ex-
pectations operator. Thus, the unbiasedness hypothesis implies that under
REH forecasts errors are expected to be zero, i.e. they fluctuate randomly
so that ex post no systematic deviations of the actual spot rate from the ex-
pected rate should be observed. The unbiasedness hypothesis can be tested
econometrically by regressing the actual change in the spot exchange rate on
the expected change according to the professional forecasts. Thus, the null
hypothesis of unbiasedness implies that it is possible to decompose st+1 − st
as

st+1 − st = α + β(Et st+1 − st ) + ǫt+1 (8)

where s is the logarithm of the nominal spot exchange rate, α = 0, β = 1
and ǫt+1 has a mean of zero and is uncorrelated with Et St+1 − St .
A second implication of the rational expectation hypothesis is that forecast
errors of rational subjects are serially uncorrelated. This condition can
directly be tested by estimating

et = α + β1 et−1 + β2 et−2 + . . . + βn et−n + ǫt (9)

The hypothesis of serially uncorrelated forecast errors implies that α = β1 =
β2 = . . . = βn = 0. Furthermore, the rational expectation hypothesis im-
plies that rational subjects generate their forecasts by using all available
information efficiently. This implication is often called the orthogonality
condition. According to the orthogonality hypothesis, rational forecasts in-
corporate all available information, so that their predictive power cannot
be improved by the inclusion of any variable that is known at the time of
expectation formation. Consequently, forecast errors must be uncorrelated
with any variable in the available information set. The orthogonality hy-
pothesis can be tested by regressing the ex post forecast errors against some
known information available when market participants form their forecasts,

st+1 − Et st+1 = α + βXt + ǫt+1 (10)

where Xt is a set of information known at time t and the orthogonality
hypothesis holds if α = 0 and β = 0. In our regression approach, the
information set Xt contains lagged exchange rates, so that the regression
equation is given as

st+1 − Et st+1 = α + β1 st + β2 st−1 + . . . + βn st−n−1 + ǫt+1 (11)

Table 3 summarizes the results of the tests for rational expectations. The re-
sults reveal that, although forecasts anticipate the future direction of change
rather well (almost all β coefficients are larger than zero), the hypothesis of
unbiased forecast is rejected for most of the participants in the experiment.
Furthermore, the participants tend to make only inefficient use of past infor-
mation. Thus, the orthogonality hypothesis must be rejected in most cases.
We reach a similar conclusion for the hypothesis of uncorrelated forecast er-
rors. Nearly all participants form expectations concerning future exchange
rates in a way that the forecast errors are serially correlated. Consequently,
we have to reject the hypothesis of serially uncorrelated forecast errors.

Efficient Uncorrelated
No. of Unbiased
Group β>0 use of forecast
Subjects forecasts
information errors
1 6 1 6 2 1
2 6 1 5 1 4
3 6 3 5 0 0
4 6 2 6 1 2
5 6 3 6 0 3
6 6 0 5 1 1
Note: the figures are based on the results of the corresponding F-tests; the
considered significance level is 10%.

Table 3: Tests for Rational Expectations

4.2.2 Individual expectation formation mechanisms

Our analysis of how participants in the experiment form their expecta-
tions concerning the future exchange rate is based on adaptive, extrapolative
and regressive expectation mechanisms. The adaptive expectations hypoth-
esis describes the change of the prediction as an adjustment depending on
the error between the actually observed exchange rate St and the last fore-
cast Et−1 St . For this reason it is also referred to as the error-learning
model:

Et St+1 − Et−1 St = α + β(St − Et−1 St ) + ǫt (12)

The adaptive expectation hypothesis requires that 0 ≤ β ≤ 1. In this case,
expectations are judged to be stabilizing in the sense that an unexpected
appreciation of the exchange rate leads to an expected depreciation of the
exchange rate. In contrast, for β > 1 those expectations are characterized
as destabilizing since an unanticipated appreciation leads to a continued
expected appreciation of the exchange rate. The case β = 1 represents
static expectations.
According to extrapolative expectations, the forecasts are affected solely by
past realizations:

Et St+1 − St = α + β(St − St−1 ) + ǫt (13)

Crucial for the interpretation of the results is the sign of the coefficient β.
If β < 0, expectations are stabilizing in the sense that a recent movement
in the exchange rate gives rise to the expectation of a reserve change in
the future. For β > 0, market participants expect that current exchange
rate movements will recur in the future. This phenomenon is called band-
wagon expectations. For β = 0, individuals have static expectations. In the
regressive model, the exchange rate is expected to develop towards a refer-
ence level, for instance a constant or a fundamental value. We test whether
the subjects adjust their forecasts according to changes of the fundamental
value Stf by estimating the equation:

Et St+1 − St = α + β(Stf − St−1 ) + ǫt (14)

Here, β > 0 indicates that individuals expect the exchange rate to move
towards the reference level. In contrast, β < 0 corresponds to an expected
deviation of the exchange rate from its reference level. Again, β = 0 alludes
to static expectations.
Table 4 presents the summarized results of the tests for various expectation
formation mechanisms. Our results indicate that for group 1 and group
2 the majority of individuals show static expectations. However, some in-
dividuals depart significantly from the static expectation formation mech-
anism, especially in group one in which at least three participants form
trend-extrapolative expectations. For group 3 to 6 we can conclude that
trend-extrapolative, destabilizing expectations play a decisive role. In all of
these groups, most participants generate expectations consistent with the
extrapolative expectation model. As most β coefficients are larger than
zero, those expectations can be characterized as bandwagon expectations.
Conforming to our findings for the extrapolative expectation scheme, the
results for the adaptive expectations indicate that individuals either possess
static expectations or form destabilizing expectations in the sense that they
expect a continued appreciation of the exchange rate due to an unantici-
pated appreciation.
The results for the regressive expectation scheme demonstrate that the fun-
damental value of the exchange rate has only a minor influence on the
expectation formation of individuals. Thus, the regressive expectation hy-
pothesis has to be rejected for the majority of the subjects.2

Group Extrapolative Adaptive Regressive

b=0 b>0 b<0 b<1 b=1 b>1 b=0 b>0 b<0
1 2 3 1 1 4 1 4 2 0
2 5 1 0 0 5 1 3 3 0
3 0 6 0 0 4 2 6 0 0
4 0 6 0 0 2 4 5 1 0
5 0 6 0 0 0 6 6 0 0
6 1 5 0 0 2 4 6 0 0

Table 4: Individual expectation formation mechanisms

Overall, the results for the individual expectation formation correspond

with the aggregated behavior of the market participants discussed above.
The subjects extrapolate the trends initiated by random shocks and there-
fore show a high degree of similarity between the groups. The fundamental
value is obviously not considered in the expectation formation process of the
market participants. In our opinion, the previous results inevitably raise the
question of coordinated expectations as suggested by Keynes (1936).

4.3 Keynes Beauty Contest - Coordinated Expecta-

tions?
Keynes’ (1936) metaphor of the beauty contest suggests that market par-
ticipants tend to coordinate their expectations concerning future asset prices
in speculative markets. As Figure 1 shows, participants within one group
tend to coordinate their expectations. We test for coordinated expectations
by estimating

Eti st+1 − Etavg st+1 = ci + ǫt+1 (15)

where Eti st+1
are
Pnthe individual (log) expectations for st+1 and Etavg st+1
is defined as n1 i=1 Eti st+1 . Under the null hypothesis of nonsystematic
deviations in expectations which consequently corresponds to homogeneous
expectations, it is expected that the mean ci should be zero. Estimat-
ing equation (15) using the corresponding expectations of participants in
2 We furthermore estimated a more general framework for the analysis of expectation

formation mechanisms as was suggested by MacDonald (2000). The three concepts are
nested in one regression of the form: Et st+1 − st = α + β1 (st − st−1 ) + β2 (Et−1 st − st ) +
β3 (sft − st ) + ǫt+1 . The results obtained from these tests were similar, but we chose the
isolated tests as some subjects did not have any significant coefficients and could not be
classified.
groups 1 to 6 yields to 0/1/0/1/0/1 subjects with heterogeneous expecta-
tions. These results indicate that the expectations of participants in each
group do not deviate systematically from the average group expectations.
Thus, the results of testing for heterogeneous expectations provides further
evidence for homogenous, coordinated expectations in each group.

5 Discussion of the results and conclusion

Our experiment deals with the expectation formation of subjects. We
are particularly interested in whether the subjects tend to coordinate their
expectations as suggested by Keynes (1936) and whether they base their
expectations on trends.
A central result of our experiment is that participants tend to coordinate
their expectations on a common prediction strategy. Merely a few partic-
ipants deviate significantly from the average expectations. Thus, it is rea-
sonable to conclude that Keynes‘ description of how individuals form their
expectations in asset markets is quite accurate. In this context, Lawson
(1985) states that the Keynesian "conventional judgment" is not irrational
as it allows for an efficient usage of scarce cognitive resources.
Akerlof (2002) characterizes Keynes as the progenitor of the modern Behav-
ioral Economics view of asset markets. This brings us directly to our second
proposition. According to Behavioral Economics, market participants use
simple heuristics instead of calculating exact, rational solutions. We suggest
that it is reasonable for market participants to use a simple trend heuris-
tic. This suggestion is also approved in our experimental setting. For most
of the participants in the experiment we find that a trend-extrapolative,
destabilizing expectation formation mechanism is an accurate description
of their prediction strategy. This is especially true for the groups that show
a high tendency for coordinated expectations.
Most of the realized prices in the experiments by Hommes et al. show strong
oscillations, of which some were persistent over the whole duration of the ex-
periment and some were converging to the fundamental value. Asset prices
were fairly stable only in a few groups and thus more consistent with our
results. While Hommes et al. observe coordination within the groups but
find clear differences between the groups, our results of six different market
exchange rates show very similar behaviors of market participants. This
finding can be ascribed to the random shocks that mainly cause deviations
from coordination results. The subjects do not know whether the trend is
generated by a random shock, other market participants or robot traders
who forecast the fundamental value. The subjects also do not need to know.
They simply apply the heuristic of trend following. We relate our findings
to the concept of focal points. It is suggested that subjects do not have an
incentive to deviate from a forecast once a common expectation has been
established. This is exactly what we observe.
Our results provide a plausible explanation for the behavior of market par-
ticipants in speculative environments such as exchange markets.

6 References
Akerlof, G.A. (2002). Behavioral Macroeconomics and Macroeconomic
Behavior. American Economic Review, 92, 411-433.
Allen, F., Morris, S. & Shin, H.S. (2003). Beauty Contests, Bubbles and It-
erated Expectations in Asset Markets. Cowles Foundation Discussion
Paper No. 1406, Yale University.
Arifovic, J. (1996). The Behavior of the Exchange Rate in the Genetic
Algorithm Experimental Economies. Journal of Political Economy,
104, 510-541.
Arrow, K.J. (1986). Rationality of Self and Others in an Economic System.
In R. Hogarth & M. Reder (Eds.), Rational Choice: The Contrast
Between Economics and Psychology. Chicago, 200-216.
Becker, O., Leitner J. & Leopold-Wildburger, U. (2007). Heuristic mod-
eling of expectation formation in a complex experimental information
environment. European Journal of Operational Research, 176, 975-
985.
Beckman, S.R. & Downs, D.H. (1997). Forecasters as Imperfect Infor-
mation Processors: Experimental and Survey Evidence. Journal of
Economic Behavior & Organization, 32, 89-100.
Cheung, Y.-W. & Chinn, M.D. (2001). Currency Traders and Exchange
Rate Dynamics: A Survey of the US Market. Journal of International
Money and Finance, 20, 439-471.
Cochrane, J.H. (1988). How Big Is the Random walk in GNP? Journal of
Political Economy, 96, 893-920.
Duncan, J. & Isaac, R.M. (2000). Asset Markets: How They Are Af-
fected by Tournament Incentives for Individuals. American Economic
Review, 90, 995-1004.
Dwyer, G. P., Williams, A.W., Battalio, R.C. & Mason, T.I. (1993). Tests
of rational expectations in a stark setting. The Economic Journal,
103, 586-601.
Elliott, G. & Ito, T. (1999). Heterogeneous Expectations and Tests of
Efficiency in the Yen/Dollar Forward Exchange Rate Market. Journal
of Monetary Economics, 43, 435-456.
Fiedler, K. & Bless, H. (2001). Social Cognition. In M. Hewstone & W.
Stroebe (Eds.), Introduction to Social Psychology. Oxford, 115-149.
Fisher, E.O’N. & Kelly, F.S. (2000). Experimental Foreign Exchange Mar-
kets. Pacific Economic Review, 5, 365-387.
Friedman, M. (1953). The Case for Flexible Exchange Rates. In: M.
Friedman (Ed.), Essays in Positive Economics, Chicago, 157-203.
Gerber, A., Hens, T. & Vogt, B. (2002). Rational Investor Sentiment.
Working Paper No. 126, Institute for Empirical Research in Eco-
nomics, University of Zurich.
Gigerenzer, G. & Todd, P.M. (1999a). Fast and Frugal Heuristics. In G.
Gigerenzer & P. Todd (Eds.), Simple Heuristics That Make Us Smart,
Oxford, 3-34.
Gigerenzer, G. & Todd, P.M. (1999b). Simple Heuristics That Make Us
Smart, Oxford.
Heemeijer, P., Hommes, C., Tuinstra J., & Sonnemans, J. (2004). Form-
ing Price Expectations in Positive and Negative Feedback Systems.
CeNDEF Working Paper 04-15, Center for Nonlinear Dynamics in
Economics and Finance, University of Amsterdam.
Hey, J.D. (1994). Expectations Formation: Rational or Adaptive or ...?
Journal of Economic Behavior & Organization, 25, 329-349.
Hommes, C.H., Sonnemans, J., Tuinstra, J. & van de Velden, H. (2002).
Expectations and Bubbles in Asset Pricing Experiments. CeNDEF
Working Paper Series, WP-02-05, Center for Nonlinear Dynamics in
Economics and Finance, University of Amsterdam.
Hommes, C.H., Sonnemans, J., Tuinstra, J. & van de Velden, H. (2005).
Coordination of Expectations in Asset Pricing Experiments, Review
of Financial Studies, 18, 955-980.
Keynes, J.M. (1936). The General Theory of Employment, Interest, and
Money. London.
Lawson, T. (1985). Uncertainty and Economic Analysis. Economic Jour-
nal, 95, 909-927.
Lo, A.W. & MacKinlay, A.C. (1988). Stock Market Prices Do Not Fol-
low Random Walks: Evidence from a Simple Specification Test. The
Review of Financial Studies, 1, 41-66.
Lo, A.W. & MacKinlay, A.C. (1999). A Non-Random Walk Down Wall
Street. New Jersey.
MacDonald, R. (2000). Expectations Formation and Risk in Three Fi-
nancial Markets: Surveying What Surveys Say. In F. Gardes and
G. Prat (Eds.), Price Expectations in Goods and Financial Markets:
New Developments in Theory and Empirical Research. Cheltenham,
172-212.
Muthoo, A. (1999). Rationality in the Face of Uncertainty. In P. Earl
& S. Kemp (Eds.), The Elgar Companion to Consumer Research and
Economic Psychology. Cheltenham, 487-495.
Noussair, C.N., Plott, C.R. & Riezman, R.G. (1997). The Principle of Ex-
change Rate Determination in an International Finance Experiment.
Journal of Political Economy, 105, 822-861.
Pesaran, M.H. (1988). The Limits to Rational Expectations, New York.
Shleifer, A. (2000). Inefficient Markets - An Introduction to Behavioral
Finance. Oxford.
Simon, H.A. (1955). A Behavioral Model of Rational Choice. Quarterly
Journal of Economics, 69, 99-118.
Simon, H.A. (1997). Models of Bounded Rationality, Vol. 3 Empirically
Grounded Economic Reason. Cambridge.
Smith, V.L., Suchanek, G.L. & Williams A.W. (1988). Bubbles, Crashes
and Endogenous Expectations in Experimental Spot Asset Markets.
Econometrica, 56, 1119-1151.
Sunder, S. (1995). Experimental Asset Markets: A Survey. In A. Roth &
J. Kagel (Eds.) The Handbook of Experimental Economics. Princeton,
NJ: Princeton University Press, 445-500.
Williams, A.W. (1987). The Formation of Price Forecasts in Experimental
Markets. Journal of Money, Credit and Banking, 19, 1-8.
Young, H.P. (1996). The Economics of Convention. Journal of Economic
Perspectives, 10, 105-122.