Teaching Nash Equilibrium and

Dominance: A Classroom Experiment

on the Beauty Contest
Virtudes Alba-Fernández, Pablo Brañas-Garza,
Francisca Jiménez-Jiménez, and
Javier Rodero-Cosano

Abstract: The authors’ aim in this article was to show how the use of classroom
experiments may be a good pedagogical tool to teach the Nash equilibrium (NE)
concept. The basic game is a version of the beauty contest game (BCG), a simple
guessing game in which repetition lets students react to other players’ choices and
converge iteratively to the equilibrium solution. The authors perform this experi-
ment with undergraduate students with no previous training in game theory. After
four rounds, they observe a clear decreasing tendency in the average submitted
number in all groups. Thus, the findings show that by playing a repeated BCG,
students quickly learn how to reach the NE solution.
Key words: beauty contest game, classroom experiments, Nash equilibrium,
teaching
JEL codes: A22, C99, D83

Generally, the teaching of game theory begins with the study of strategic inter-
actions among players considering only pure strategies. Once players understand
that each player’s payoffs depend not only on their own actions but also on those
of other players, the next step is to explain the process of elimination of domi-
nated strategies. To do this, only two basic assumptions are required: rationality
(maximization) and common knowledge (of rationality). Both concepts allow one
to solve (dominance solvable) games and also predict some particular behavior.
However, some games are not dominance solvable even if they have a Nash equi-
librium (NE). In this case, the NE requires an additional hypothesis: common

Virtudes Alba-Fernández is a professor of statistics, University of Jaén and researcher of centrA,

Spain. Pablo Brañas-Garza is a professor of economics, University of Grenada, Spain, and
researcher of centrA, Spain. Francisca Jiménez-Jiménez is a teaching assistant of economics,
University of Jaén and researcher of centrA, Spain, and Javier Rodero-Cosano is a professor of
economics, University of Mélaga, Spain (e-mail: [email protected]). The authors thank Rosemarie
Nagel, Curro Martínez Mora, Quique Fatás, and the participants at the Economic Science Association
Meeting in Tucson (2003), especially David Reiley. They thank three anonymous referees for valuable
comments. They thank Samuel Hernández for his contribution to the data gathering and processing
and Martha Gaustad for the language revision. Pablo Brañas acknowledges the hospitality of IESA-
CSIC during this research. They gratefully acknowledge the financial support provided by the
University of Jaén R&D program (# 20210/148). Copyright © 2006 Heldref Publications

Summer 2006 305

knowledge of players’ beliefs about their rivals’ actions. Hence, the best response
becomes the suitable mechanism for reaching the equilibrium solution. Best
response is the best strategy that players may choose given the strategies chosen
by their rivals. The NE is reached when all agents play (some of) their best
response. Thus, the NE is self-enforcing because no player has incentives to devi-
ate from it. Clearly, the use of the NE eliminates circular reasoning (player 1
thinks that player 2 thinks that player 1 thinks…) in games with at least one pure
NE, not in a game with a unique mixed NE.
We propose a pedagogical tool for simplifying the teaching of the NE: suc-
cessive repetitions of a dominance solvable game, the beauty contest game
(BCG) introduced by Moulin (1986). This game is very useful for showing, in
an intuitive manner, the two procedures described above. The game is played
four times in the form of a classroom experiment, prior to teaching a game
theory class.
Our findings clearly illustrate that students inexperienced in game theory apply
both iterated elimination of dominated strategies and best-reply behavior in each
round to the previous choices. On average, we observed a recursive approxima-
tion to the Nash prediction, which we call the learning effect. Therefore, by
repeating this static game, we ensure that students learn how to solve it in some
way.

THEORETICAL BACKGROUND

Since the introduction of the concept by Nash (1951), the NE has not only
become the standard tool for the economic scientific community but also the basis
for the systematic teaching of the discipline.
In game theory, it is common practice to assume that individuals magically
choose a set of actions such that all the (infinitively recursive) predictions come
true. To reach this hard-to-believe solution, theorists can follow one of the two
original Nash interpretations: either agents are perfectly rational or an evolution-
ary equilibrium exists.1 The first interpretation relies on the assumption that all
agents are able to compute the game equilibria and reach one of them. In contrast,
the second interpretation presupposes the existence of a large population of ordi-
nary people that play the game in an evolutionary framework: that is, they pick
up one strategy randomly, if the outcome is good, they repeat it again, or if the
strategy is bad, it will disappear. After some time, the game will converge to one
of the NE. The BCG is a good example of the underlying convergence to both
interpretations. From a rational point of view, it is not credible that the subjects
will solve this problem (few people can at the first try!), but if it is repeated, the
solution will converge to equilibrium most of the time.
The beauty contest game is a simple guessing game that makes it easy to eval-
uate individuals’ level of reasoning. The basic BCG is as follows: A certain num-
ber of subjects are invited to play a game in which all of them must
simultaneously choose a real number from an interval (generally between 0 and
100). The winner is the player who chooses the number that is closest to p times
the mean of all the numbers chosen, where 1  p  0. The winner receives a
306 JOURNAL OF ECONOMIC EDUCATION
prize, the losers get nothing. Under these rules, the unique NE is 0 for all
players.2 Other parameter configurations may produce different or even multiple
equilibria (Bosch-Domènech et al. 2002).
As Stahl (1996) pointed out, the distribution of chosen numbers lets us analyze
the depth of reasoning of the agents. In this case, level 1 includes people who
expect that the other players behave randomly so they choose p * mean (being
mean  50 if the choice distribution is uniform); level 2 contains people expect-
ing that the other’s depth of reasoning is level 1, and, choosing then p2 * mean, …;
generalizing, at level K are people who choose pK * mean because they believe
that the other people are at level K  1. If K is big, pK * mean  0, so if the
process is repeated ad infinitum (K  ), the theoretical solution 0, is reached,
that is, the highest level of reasoning (Figure 1 represents this process with
p  2/3). Random answers are called level 0 of reasoning. A level of reasoning
higher than 3 is rare in BCG experiments (Bosch Domènech et al. 2002),
although this iterated best replay behavior is quite commonly found (Ho,
Camerer, and Weigelt 1998).
In Figure 2, taken from Ho, Camerer, and Weigelt (1998, 951), we show the
convergence to the 0 theoretical solution from a dominance iterated point of view.3
Given p  2/3, any number chosen between 66.6 and 100 is dominated by 66.6 
100 * 2/3. Hence, the interval [66.6–100] corresponds to irrational behavior, which
we call R(0). Rational individuals will always choose a number in the [0–66.6]
interval. Applying the same reasoning, R(1) players will choose a number below
66.6 but above 44.4  2/3 * 66.6, whereas R(2) players will choose a number in
the interval [29.6–44.4]. Following this iterated reasoning level process ad infini-
tum, one reaches the unique Nash equilibrium (0, with R()). Thus, this game is
called dominance solvable. Camerer (2003, ch. 5) described some other games that
also tend to unravel and would be fun to teach in class. These include the patent
race game, in which iteration eliminates some strategies but not others, centipede
games, price competition, or traveler’s dilemma, in which the dominance reduces

FIGURE 1. An example of different reasoning levels.

FIGURE 2. Iterated reasoning of individuals by eliminating dominated strategies.

Summer 2006 307

the collective payoffs to players, and the dirty faces game, in which an increase in
the number of steps of iterated deletion increases payoffs.
Nevertheless, this game makes a much more important point in practice: If a
subject plays the NE (0), in the majority of the cases, he or she will not win. Note
that the key to this game is to take one more step of deletion of dominated strate-
gies than the other players (but not too many).4 Hence, the NE is useful for know-
ing where the adaptive process leads, but it is not a useful normative theory as a
basis for advice.5
The original idea behind the BCG was first put forward by Keynes (1936, 155)
to show that clever investors have to “anticipate the basis of conventional valua-
tion a few months hence, rather than…over a long term of years,” if they are to
act in the stock market before other investors do.
The unique equilibrium of the formal game model is 0 for p  1 and is
obtained by iterated elimination of weakly dominated strategies.6 After this basic
framework was laid down, some experimental researchers began to investigate
BCG or p-beauty (Ho, Camerer, and Weigelt 1998). The first experimental
study can be found in Nagel (1994, 1995).7 Other studies have been done by
Bosch-Domènech and Nagel (1997a, 1997b); Bosch-Domènech et al. (2002) and
Duffy and Nagel (1997). Nagel (1998) provided a survey of the literature.
A wide variety of BCG experimental designs can be found.8 In these, some-
times participants are students, at other times they are professors or newspaper
readers. Some designs are one-shot, others are repeated, with communication ver-
sus no communication or laboratory versus field experiments.
Generally, BCG is run with individual subjects. Kocher and Sutter (2001),
however, compared individual versus group behavior in this type of game. They
found that although groups did not apply deeper levels of reasoning, they did, in
fact, learn faster.
Repetition permits individuals to learn dynamically from other people’s
expected behavior. Our experiment was very similar to one in Ho, Camerer, and
Weigelt (1998). In their design, individuals were given information about pre-
vious period choices and, therefore, the learning process was based on an
evolutive game. In contrast, Weber (2003b) argued that learning with the BCG
can happen even without feedback. He found that there was convergence toward
the NE in the game, even when subjects did not receive any information
between periods.
Over the past years, research on the BCG has returned to Keynes’ (1936) orig-
inal idea. Hirota and Sunder (2002) experimentally explored whether price bub-
bles in security markets are generated by a beauty-contest mechanism: If
dividends are paid beyond investors’ personal investment horizons, investors must
create beliefs about others’ beliefs (second-order beliefs) that depend on third-
order beliefs, which in turn depend on fourth-order beliefs, and so on. They con-
cluded that when the realization of the dividend is distant and well beyond
investors’ investment horizon, investors found it difficult to induce the funda-
mental value of securities from the future to the present (backward induction).
This difficulty gives rise to price bubbles because, in this case, investors adjust
their expectations on the basis of observed prices (forward induction).
308 JOURNAL OF ECONOMIC EDUCATION
 Up to now, the BCG has primarily been used to study the depth of individuals’
level of reasoning. However, we propose a new application of this framework:
classroom experiments with a pedagogical aim.9

THE EXPERIMENT

The experiment involves using a repeated version of the beauty contest game
with several groups of students to teach the NE; specifically, the iterated elimi-
nation of dominated strategies. Given that the NE is a general topic in both inter-
mediate microeconomics and industrial organization (and, obviously, in game
theory), we expect this pedagogical tool to be of great help to educators.
Three fundamental issues must be taken into account when running the
experiment.
1. The group should be large enough to reduce the effect of any one person’s
guess on the average. On the other hand, a very large group may be too
cumbersome for the instructor.10 Nagel (1995) used 12 and 17 subjects in
each group; Ho, Camerer, and Weigelt (1998) reduced the group size to 7
subjects and 3 subjects (like Kocher and Sutter 2001). Following these
guides, we used small groups (5–6 subjects) and large groups (10–11 sub-
jects).11 Each group played its own game in which each subject played a
repeated BCG against the other members of the group.
2. To motivate students, some type of reward should be given. In our class-
room experiment, the winner of each round was awarded 0.25 extra-credit
points applicable toward the final exam. In the case of several winners in the
same round within the same group, the prize was split among them.
3. Another important issue is the number of rounds that experimental subjects
must play in the BCG. Because our aim was for students to learn the NE
through the iterated elimination of dominated strategies, several rounds were
needed to observe this learning effect. Kocher and Sutter (2001) proved that
four rounds are enough to approximate the theoretical prediction.
We performed our experiment in an intermediate microeconomics theory
course during the spring semester of the 2002–2003 academic year with three dif-
ferent classroom groups: business majors (morning and afternoon groups, B–1
and B–2) and business  law majors (one single group, B  L). All groups had
an identical individual answer handout, instructor, and grading system (ranging
from 0 to 10). This subject is a requisite in both majors.
The experiment lasts about one hour and the following post-experimental ses-
sion half an hour (it is worth running both sessions consecutively, approximately
1.5 hours).

Procedures

1. The instructor defined the size of each group. However, it is not important
if one group differs slightly in size from the other groups. Nevertheless, for
purposes of comparison, it is desirable to have similarly sized groups.
Summer 2006 309
 2. After determining group size, the instructor randomly selected one monitor
for each group from the student pool. Each monitor was awarded 0.25
points. The monitor’s task was to help the instructor to record and monitor
the experiment in his or her group. Each monitor was given a calculator and
a monitoring sheet correlatively numbered (see appendix). The monitor kept
track of the individual choices in the group and calculated the mean and the
winning number. The monitoring sheets were crucial to explaining the
results after the experimental session as these findings would easily illus-
trate basic concepts of game theory, namely the iterated deletion of domi-
nated strategies.
3. The third task consisted of creating groups, that is, students were randomly
assigned to different groups.
4. To make the monitor’s tasks easier, the instructor sat each group in a single
row (or column).
5. The instructor then gave the monitors their instructions and explained the
procedures to them. When this was done, the monitors handed out the
instruction sheets and the individual answer sheets to the experimental sub-
jects (see appendix).
6. The instructor explained the instructions out loud. It is important to avoid
numerical examples as the generation of focal points was immediate in this
kind of game. Any questions were answered aloud. Experimental subjects
were told that speaking was absolutely forbidden.12 When everyone fully
understood the rules, the experiment began.
7. Round 1: Experimental subjects filled in their answer sheet. The monitor
then collected all guessed numbers and calculated the mean and 2/3 * mean.
By comparing this value to those reported by the experimental subjects, the
monitor determined the winner. The group was told the mean, 2/3 * mean,
and the winner’s guess (but not who the winner is). Nagel (1999) also gave
the students the complete set of guesses.13
8. Round 2: Following round 1, the students were informed that they would play
another round, independent of the previous one. The procedure for round 2 was
identical to round 1. Groups remained the same and did not communicate.
9. Rounds 3 and 4 were the same as round 2.

POSTEXPERIMENTAL DISCUSSION

Immediately following the experimental session, the postexperimental discus-

sion was begun. In this second part of the experiment, we introduced the theoret-
ical background of the NE. This step involved three topics:
1. We present the iterated elimination of dominated strategies underlying the
BCG by means of Figure 2. By explaining this process, students can under-
stand that if rationality is common knowledge, nobody will choose a num-
ber within the interval [66.6–100] because this subset of numbers is
dominated by [0–66.6]. Following this reasoning ad infinitum—students
fully capture this idea.
310 JOURNAL OF ECONOMIC EDUCATION
 At this moment, it is a good idea to show students the first round results
of the groups involved in the session. Alternatively, a simple histogram of
the pooled data may be used to cover up any embarrassing answers (for
more details, see the Results section). Students may compare their own cho-
sen numbers with those reported in Figure 2. Moreover, it could be fun and
interesting to show other experimental results to offer students the opportu-
nity to compare themselves with others (for example, high school students,
80-year-olds, corporate CEOs, or even game theorists). Nagel (1999) and
Camerer, Ho, and Chong (2003b) reported a high number of BCG experi-
ments with several contexts and subject pools.14
2. Once common knowledge rationality was clearly defined, it was time to explain
best response behavior. Students observed that their own performance depended
on their beliefs about other players’ actions. Furthermore, they understood that
other players’ behaviors also depended on their beliefs about others and so on.
As an example, we presented some preliminary results of their own
responses. To do this, some monitoring sheets (see appendix Table A1) were
used to illustrate the average behavior of the groups. Because the average
number chosen decreased through successive rounds, students realized that
game theory approximately predicted their performance.
3. We now had all the necessary ingredients to “cook up” the NE concept. Let
–x be player i’s forecast of the mean of all numbers submitted. Player i chose
xi with the objective of minxi[0,100] |xi – px–|. All players behaved identically.
Common knowledge of beliefs, together with mutual best response, implied
that in any equilibrium, x  px. Hence, the unique equilibrium must be that
x  0, the number that is equal to p times itself.
Because students had experienced the recursive dynamics of the NE, we
explained what the abstract concept implied and how this could be recon-
ciled with the real world. In the theoretical model, we supposed that every-
thing was instantaneously adjusted. The BCG had to be explained as if it
were a slow motion picture: Repetition lets individuals reach the theoretical
solution step by step. Perfectly rational agents do not need this, but people
(even economists) are not fully rational. This is the second remarkable
point: Repetition could be a good substitute for rationality. In fact, the mod-
ern view in game theory is that equilibrium arises from adaptation, evolu-
tion, communication, or imitation. Because these processes take time,
equilibration should not occur instantly in one-shot games.
Nagel (1998), Stahl (1996), and, most recently, Camerer, Ho, and Chong
(2003b) presented very precise formal models of limited rationality that
accurately described the experimental results of these games. Specifically,
Camerer, Ho, and Chong (2003a) developed a nonequilibrium model for
one-shot games with decision rules for players doing different steps of
thinking.15 Their prediction for dominance solvable games will converge to
the NE as a limit result.

Finally, as an anecdote, we mentioned A Beautiful Mind, the movie about John

Nash’s life. The film tries to explain the NE to ordinary people using an allegorical
Summer 2006 311
sequence. In this scene, John is in a bar with some friends when a group of girls
walks in. If all the guys try to pick up only the prettiest girl, all the girls would
leave, and all the guys would be out of luck.16 The BCG can be seen as a similar
symbolic game: Individuals have to compete to outguess the rest of the partici-
pants, but without cooperation, costly errors can occur. Knowing game theory
concepts could reduce these costs.17

RESULTS

The experimental BCG was played in three different sessions, using samples
B–1, B–2, and B  L. Table 1 shows the distribution of subjects participating
in the experimental sessions by group. Our entire sample consisted of 139 sub-
jects assigned to 6 groups per session (5 in B  L). All groups played four
rounds of the BCG.
Although our interest was specifically in the process of learning the NE, we
considered it meaningful to show first-period behavior. In this way, we were able
to relate our experimental results to the theoretical reasoning levels shown in
Figure 2. The relative distribution of first-round iterated steps is shown in
Figures 3–5. As seen, the behavior of subjects in sessions B–1 and session B 
L were quite similar, whereas the subjects in session B–2 seemed to have a higher
level of reasoning than the other two.18
Table 2 shows the aggregate level of the average chosen number and the stan-
dard deviation across sessions and rounds. The mean of the chosen numbers
decreased throughout successive rounds in all cases.19 However, we observed that
the path to convergence was not identical. In Figure 6, we illustrate each trend in
the experimental session.
At first sight, sessions B–1 and B  L behaved similarly, but the trend in B–2
showed a flatter slope. As our main interest was to study the learning effect, we
analyzed the average variation more accurately round after round. We defined
speed of convergence as: ( i  i1)/ i1 where i is the average guessed num-
ber in round i (i  1, …, 4) (shown in Table 3).

TABLE 1. Distribution of Group Size, Number of Subjects per Group and

Session

Session
Group B1 B2 BL

1 10 6 10
2 11 6 8
3 9 6 9
4 9 5 8
5 10 6 10
6 10 6 —
Session pop. 59 35 45

312 JOURNAL OF ECONOMIC EDUCATION

 FIGURE 3. Relative frequencies in the first round group B  1.

FIGURE 4. Relative frequencies in the first round group B  2.

Summer 2006 313

 FIGURE 5. Relative frequencies in the first round group B  L.

TABLE 2. Average Number and Standard Deviation per Round and Session

Session
B1 B2 BL
Round Average S.D. Average S.D. Average S.D.

1 38.885 19.348 20.782 17.611 37.067 18.891

2 32.087 14.235 17.060 12.362 29.008 11.554
3 22.660 16.750 13.042 7.318 22.712 11.577
4 14.419 13.360 9.748 7.503 19.142 13.944

Note that all the sessions shared a similar initial speed of convergence close to
20 percent. However, this speed varied dramatically within and between groups
from round 2 to 3 and from round 3 to 4. Thus, the learning effect was clearly
observable in all rounds and sessions.
After this descriptive approximation, we statistically analyzed the results
obtained. To be more specific, we were interested in testing if students were
induced to modify their behavior toward the NE by repeatedly playing a BCG
a set number of times. Therefore, we expected the average number in a round
to be different from the following round and, furthermore, that the difference
between one round and the following round would be greater than 0, that is, that
it would decrease. To verify this, we formulated the following null hypothesis:
314 JOURNAL OF ECONOMIC EDUCATION
 FIGURE 6. Evolution of session averages.

TABLE 3. Speed per round and session

Session
Speed of Convergence for round B –1 B–2 B L

( 2  1)/ 1 0.17 0.18 0.22

( 3  2)/ 2 0.29 0.24 0.22
( 4  3)/ 3 0.36 0.25 0.16

H0 : ij  kj  0, i  k and i, k  1, …, 4
and alternative hypothesis:
H1 : ij  kj  0, i  k and i, k  1, …, 4,
where i, k represent two consecutive rounds, and j represents the session number
(1 for B–1, 2 for B–2, and 3 for B  L).
The statistical analysis is developed in depth in the appendix. Briefly, we
checked if each population followed the normal distribution. If one of them did
not follow the normal distribution, we used a parametric and a nonparametric test
to contrast our null hypothesis. Our main results showed that, in sessions B–1 and
B  L, the learning effect was observed in the entire experiment, and, in session
B–2, the learning effect was observed in the last two rounds.
To sum up, these results indicate that this is a powerful tool to aid students in
reaching the NE.

VARIATIONS OF THE EXPERIMENT

Multiple possible variations of the BCG could be used as a classroom experi-

ment. We describe some of the most important of these here.
Summer 2006 315
The Two-Person BCG

This version of the game is interesting because, unlike the general BCG in
which there are more than two players, in a two-person game, the player who
picks the lowest number wins (if p  1). Thus, the subject who plays the NE strat-
egy (0) always wins. Therefore, 0 is not only a weakly dominant strategy but also
the best response for all the choices of the other person. For this reason, this vari-
ant is a good example of the attractive power of the equilibrium point. As
Camerer, Ho, and Chong (2003a) pointed out, this special game can be solved by
only one step of weak dominance. In their model, all players using one or more
thinking steps will choose 0. However, it is important to recall that the most
important point of the experiment is to show students the strength of the equilib-
rium concept, even for complicated games.
However, Grosskopf and Nagel (2001) showed that usually subjects do not
choose the 0 in the one-shot game and then they do not reach the equilibrium imme-
diately. Even under full information, students need some repetitions to realize that
0 is the NE. They concluded that “convergence toward equilibrium is driven by imi-
tation and adaptation rather than self-initiated rational reasoning” (2001, 1).
Nonetheless, an interesting option could be to perform both the n  2 and
n  2 experiments in the same class. Some groups could play the first version,
while the others play the second one. It would then be interesting to compare both
versions in the post-experimental session and show students that the NE some-
times fails to describe the reality if the convergence process is too long (Grosskopf
and Nagel 2001).

The Entire-Class Experiment

If the class is not large enough, an interesting alternative is to perform the exper-
iment using the whole class as a single group. The main advantage to this approach
is that no student needs to sit out as a monitor. Also, the individual effect on the
population average would decrease slightly. In contrast, if the group is too large,
the process can become a bit cumbersome for the teacher. The most straightfor-
ward way to run this BCG is for the instructor (or any ad hoc volunteer) to write
the submitted numbers on the board while another volunteer calculates the mean
and p * mean. Note that all guessed numbers should be collected beforehand.

Web-based Experiments

In his Web page at the University of Virginia,20 Charles Holt provides some excel-
lent internet software for playing the BCG (among many other games). Not only can
students play the BCG on this Web site, but they can also check their own results and
performance online as well as compare their outcomes with other previous experi-
ments. However, running the experiment in this way requires having a computer room
with Internet access and banning its use in the classroom. Except for this minor prob-
lem, Veconlab is an amazing tool for teaching. Another point to take into considera-
tion, in our case, is the fact that Veconlab is available only in English at this time.
316 JOURNAL OF ECONOMIC EDUCATION
Other Variations

There are several additional variations on the game. A good idea would be to
run some of these simultaneously with different groups in the same class to com-
pare the different outcomes arising from different rules.
• When p  1, Nagel (1998) showed that this variation of the game was not
trivial. She considered two cases: (a) under p  1 and the interval [0–100],
there are two equilibria: everybody chooses either 0 or 100. The latter is the
perfect equilibrium.21 (b) For p  1 and the interval [100–200], the only
equilibrium is 200; the upper bound of the interval. The iterated deletion
process starts from the lower bound 100p, 100p2, and so on.22
• Using the median instead of the mean, for example, can help to illustrate to
students that the game does not change and the convergence toward the equi-
librium is basically the same under different conditions (Nagel et al. 1999).
• By using large and small groups, identical results may be found in the one-
shot game, but faster convergence to the equilibrium occurs in larger
groups. Duffy and Nagel (1997) studied the influence of a single player on
aggregated performance. They use three treatments changing the order sta-
tistic from mean to maximum or median for p  1/2, so the extreme choices
had very different weights.
• Giving students different feedback and when comparing feedback-free
BCG with normal BCG, we observed that convergence was slower when
subjects were not informed about the numbers chosen by their partners. The
lack of information constrained subjects’ learning to their own experience
(Weber 2003a, 2003b). In our case, we gave students very little information
(only aggregated data). Given that one of the goals of the experiment was to
show students the attractive power of the equilibrium, we preferred a low
information variant.23

FINAL REMARKS

We have formulated an interesting classroom experiment: a repeated version of

the beauty contest game. Our experience is that this sort of game is a good way
to introduce the complexity of the equilibrium concept. Not only does it permit
students to see the mechanism in action but also to comprehend the difference
between a theoretical abstraction and real-world dynamics toward equilibrium.
The BCG is one of those tricky games worth playing. Although it is plainly sim-
ple in its formulation and the solution is always obvious ex post, students realize
the difficulties of outsmarting other people while having a good time doing it.
The classroom experiment needs some careful preparation to make conve-
niently mixed groups. However, once everything is ready to go and the monitors
are well trained, rounds should run quite smoothly and increasingly fast.
Although for our purposes, 4 rounds are enough, it would be easy to play as many
as 10 rounds, if necessary. Continuous repetitions would be very rewarding for
students, as this would let most of them reach the theoretical solution on their
own before the postexperimental session.
Summer 2006 317
 Once the BCG is done and the postexperimental session is concluded, it is pos-
sible to run other experiments so students realize how game theory training helps
them to solve interesting (and sometimes lucrative) puzzles.

NOTES
1. Because of space constraints Nash used only the rationality explanation in the published version.
2. This is not as obvious as it looks, Grosskopf and Nagel (2001) tried a one-shot BCG with game
theorists. The answers were very different from 0 but better than the usual ones!
3. See Rapoport and Amaldoss (2000) for an experimental study of iterated elimination of (strongly)
dominated strategies.
4. Obviously, this process of reasoning involves knowing who you are playing with. In this case, an
individual who knows the 0 solution could also guess that most people will not achieve the 0 solu-
tion. Thus, clever individuals will be able to link their answer to their estimation of the average
rationality level. We call this rationality level  plus (Keynes’ “clever investor”). Alba-Fernández
et al. (2004) reported some examples of  plus behavior arising from internet BCG experiments.
Bosch-Domènech et al. (2002, 1693) found strong evidence of this behavior, reporting some com-
ments from students who followed this strategy. Grosskopf and Nagel (2001) argued that most
individuals think that other people are not fully rational, thus explaining why the equilibrium is
not reached immediately.
5. This is one of the main advantages of using the BCG for teaching purposes. Nagel (1999), using
the same graph shown in Figure 2, illustrated how to explain iterated elimination of dominated
strategies in a classroom setting.
6. It can also be obtained by using a best-response argument (Nagel 1999).
7. Nagel’s (1994, 1995) main purpose was to contrast an iterated best-reply dominance model.
8. See Camerer, Ho, and Chong (2003b) for a list of 24 different BCG experiments.
9. See Nagel’s 1999 study in which she described this application.
10. Ho, Camerer, and Weigelt (1998) showed that smaller groups (three subjects in their research)
need more time to converge.
11. It is difficult to have identically sized groups in a class. Usually, the instructor does not have the
exact number of students required.
12. Subjects were required to keep a maximum level of confidentiality for their own sake. If any sub-
jects were to know their rival’s guess, they used the best reply rule.
13. To simplify the process, we did not give our students the entire set of numbers. Nonetheless, the
type of information provided in our experiment affected the convergence speed: Weber (2003a,
2003b) analyzed the effect of different feedback conditions on learning. He found that conver-
gence toward the equilibrium occured under all circumstances, although it was faster with feed-
back. See also the Variations section to follow.
14. In general, these studies reflect a substantial regularity across very different groups, except some
highly trained ones.
15. The link between thinking steps and the iterated deletion of dominated strategies lies in the fact
that players who are not thinking strategically randomize, whereas players doing k steps of think-
ing accurately predict what lower-level players do and best respond accordingly.
16. This “equilibrium” (all the men including Nash, ignore the pretty blonde girl), in fact, is not an
equilibrium at all because any man in the group who is willing to score her could do it given the
strategies of the others, as Anderson and Engers (2002, 2) point out.
17. It is important to remind students about the importance of their intellectual achievement: John
Nash won the Nobel Prize thanks to this simple discovery. See Varian (2002) for a nice explana-
tion about Nash’s brilliant and beautiful contribution. In contrast to what viewers of the film
might think, it is clearly not a new strategy to pick up girls. Myerson (1999) gave an interesting
historical perspective to this important innovation.
18. Camerer, Ho, and Chong (2003a) estimated the average number of thinking steps at 1.61 across
24 one-shot beauty contest games. They found quite low estimates (0  1) when p  1 and high
estimates (3  5) in games in which the equilibrium was within the interval.
19. Although the initial average observed in our experiment was similar to that of Kocher and Sutter
(2001), our final average was very different! After running four rounds as well, their final aver-
ages were around 7, whereas ours were clearly higher (from 9.7 to 19.1).
20. Found at http://www.people.virginia.edu/~cah2k/.
21. Nevertheless, there are no dominated strategies in this version and, thus, no process of iterated
elimination of dominated strategies leading to equilibrium.

318 JOURNAL OF ECONOMIC EDUCATION

22. In this case, the number of eliminations is finite. That is the reason why these games are called
finite threshold games. See also Ho, Camerer, and Weigelt (1998).
23. As Weber (2003b) pointed out, even if you do not give any feedback, convergence will be
achieved, although the pace will be slower. It will take more time to make the calculation,
although the rounds will be faster. It is up to the instructor to choose the desired mix of informa-
tion and convergence speed.

REFERENCES
Alba-Fernández, V., P. Brañas-Garza, F. Jiménez-Jiménez, and J. Rodero-Cosano. 2004. Communication,
coordination and competition in the beauty contest game: Eleven classroom experiments. Central
Europe Journal of Operations Research 12 (4): 369–87.
Anderson, S. P., and M. Engers. 2002. A beautyful blonde: A Nash coordination game. Working Paper
359. University of Virginia, Charlottesville.
Bosch-Domènech, A., and R. Nagel. 1997a. El juego de adivinar el número x: Una explicación y la
proclamación del vencedor. Expansión June 16:42–43.
———. 1997b. Guess the number: Comparing the FT’s and expansión results. Financial Times June
30:14.
Bosch-Domènech, A., J. García Moltalvo, R. Nagel, and A. Satorra. 2002. One, two, (three), infinity, …:
Newspaper and lab beauty-contest experiments. American Economic Review 92 (5): 1687–1701.
Camerer, C. F. 2003. Behavioral game theory: Experiments in strategic interaction. Princeton, NJ:
Princeton University Press.
Camerer, C. F., T. H. Ho, and J.K. Chong. 2003a. A congnitive hierarchy theory of one-shot games
and experimental analysis. Working paper, Caltech, http://www.hss.caltech.edu/camerer.
———. 2003b. Models of thinking, learning, and teaching in games. American Economic Review 93
(2): 192–95.
Duffy, J., and R. Nagel. 1997. On the robustness of behavior in experimental beauty contest games.
Economic Journal 107: 1684–1700.
Grosskopf, B., and R. Nagel. 2001. Rational reasoning or adaptative behavior? Evidence from two-
persons beauty contest games. Harvard NOM Research Paper 01-09. Harvard Business School,
Cambridge, MA.
Hirota, S., and S. Sunder. 2002. Stock market as a “beauty contest” investor beliefs and price bubbles
sans dividend anchors. Working Paper 02-42, Yale International Center for Finance (ICF), New
Haven, CT.
Ho, T.H., C. F. Camerer, and K. Weigelt. 1998. Iterated dominance and iterated best response in exper-
imental “p-beauty contests.” American Economic Review 88 (4): 947–69.
Keynes, J. M. 1936. The general theory of interest, employment and money. London: MacMillan.
Kocher, M. G., and M. Sutter. 2001. The “decision maker” matters: Individual versus group behavior
in experimental “beauty-contest” games. Working paper, University of Innsbruck, Austria.
Moulin, H. 1986. Game theory for social sciences. New York: New York Press.
Myerson, R. B. 1999. Nash equilibrium and the history of economic theory. Unpublished Manuscript.
Nagel, R. 1994. Reasoning and learning in guessing games and ultimatum games with incomplete
information: An experimental study. PhD diss., University of Bonn, Germany.
———. 1995. Unraveling in guessing games: An experimental study. American Economic Review 85
(5): 1313–326.
———. 1998. A survey on experimental “beauty-contest games”: Bounded rationality and learning.
In Games and human behavior, essays in honor of Amnon Rapoport, eds., D. Budescu, I. Erev, R.
Zwicke. 105–42. New Jersey: Lawrence Erlbaum.
———. 1999. A Keynesian beauty contest in the classroom. Classroom Expernomics 8 (Fall).
http://www.marietta.edu/ delemeeg/expernom/f99.html#nagel.
Nagel, R., A. Bosch-Domènech, A. Satorra, and J. Garcia-Montalvo. 1999. One, two, (three),
infinity: Newspaper and lab beauty-contest experiments. Working Paper 438, UPF, Barcelona,
Spain.
Nash, J. F. 1951. Non-cooperative games. Annals of Mathematics 54 (2): 286–95.
Rapoport, A., and W. Amaldoss. 2000. Mixed strategies and iterative elimination of strongly domi-
nated strategies: An experimental investigation of states of knowledge. Journal of Economic
Behavior & Organization 42 (4): 483–521.
Stahl, D. O. 1996. Boundedly rational rule learning in a guessing game. Games and Economic
Behavior 16 (2): 303–30.
Varian, H. R. 2002. What, exactly, was on John Nash’s beautiful mind? Economic Scene. New York
Times, April 11.

Summer 2006 319

Weber, R. A. 2003a. Learning and transfer of learning with no feedback: An experimental test across
games. Working paper, Carnegie Mellon, University, Pittsburgh, PA.
———. 2003b. “Learning” with no feedback in a competitive guessing game. Games and Economic
Behavior 44 (1): 134–44.

APPENDIX

Experimental Instruments

Instructions
You are going to participate in a microeconomic experiment. You have been randomly
assigned to a 6–10 person group. In this game you will have to make decisions repeatedly
in four rounds. These people will be your partners throughout the four stages. Moreover,
your group has been assigned a monitor that will oversee the procedure.
The rules of the game in each period are as follows: you should choose a (integer or dec-
imal) number in the interval [0–100]. You are allowed to choose 0 and 100. Once the mon-
itor has collected your group’s choices, the winner will be determined. The winning
number is the number closest to 2/3 of the average of all the numbers chosen by your
group:
n

2 a xi
i1
x . .
3 n

The winner of each round will receive a prize of 0.25 extra points valid toward the
Micro II final exam. If two or more people are equally close to x, the prize will be split
equally among them. A person so lucky as to guess the right number all four times will get
a whole point!

TABLE A1. Monitoring Sheet

Monitor:_________ Group:_________ Session:_________

Subject Full name Round 1 Round 2 Round 3 Round 4

1
2
3
4
5
6
7
8
9
10
Average
2/3 of average

320 JOURNAL OF ECONOMIC EDUCATION

 TABLE A2. Individual Answer Sheet

Code: ______________ Group: ________________

First round
Group average: ______________
Your choice: 2/3 of the average ______________
Winning number: ______________
Second round
Group average: ______________
Your choice: 2/3 of the average ______________
Winning number: ______________
Third round
Group average: ______________
Your choice: 2/3 of the average ______________
Winning number: ______________
Fourth round
Group average: ______________
Your choice: 2/3 of the average ______________
Winning number: ______________

TABLE A3. Kolmogorov-Smirnof Goodness-of-Fit Results

Session
B–1 B–2 BL
Round z p z p z p

1 1.116 0.165 0.937 0.344 1.135 0.152

2 1.063 0.208 0.713 0.690 0.681 0.742
3 1.579 0.014* 0.384 0.998 0.792 0.556
4 1.285 0.074 0.709 0.696 0.867 0.440

* Significant at 0:05 Type I error level.

First you must write down the group and code you were given at the top of the attached
sheet. Then in each round, you must select your number. When everyone has finished, the
monitor will collect your responses. Afterwards, you will be given back your sheet with
some additional information: the average of your group, 2/3 of this average, and the win-
ning number. This procedure will be repeated four times.
If you have any questions, please raise your hand and the instructor will come to you.
You are not allowed to speak during the experiment.

Statistical Analysis
Let Xij be the number chosen by a student at round i in session j, where i  1,…, 4 and
j  1, 2, 3. Let Fij(X) be the distribution function associated with each variable Xij.
First, verify if each Fij(X) follows a normal distribution with mean ij and standard devi-
ation
ij. To do so, use the Kolmogorov-Smirnof goodness-of-fit test. The results are
shown in Table A3.
Summer 2006 321
 TABLE A4. Parametric Test for Mean Equality

Session
B–1 B–2 BL
Round
Comparison t p t p t p

1 vs. 2 2.315 0.012 1.252 0.109 2.951 0.002

2 vs. 3 — — 2.107 0.021 3.148 0.001
3 vs. 4 — — 2.859 0.003 1.758 0.042
2 vs. 4 7.586 0 — — — —

TABLE A5. Wilcoxon Nonparametric Test for Mean Equality

Session
B1 B2 BL
Round
Comparison z p z p z p

1 vs. 2 2.378 0.008 1.077 0.1414 2.467 0.006

2 vs. 3 4.612 0 2.001 0.022 2.871 0.002
3 vs. 4 5.225 0 3.459 0 3.45 0

As seen, individual choices adjust to a normal distribution except in the third round of
the first session (B–1) of the game.
Using a parametric test for the equality of means between paired rounds, the results are
given in Table A4.
Observe that it is not possible to apply this test to the third round of the first session
(B–1) because it does not follow a normal distribution.
From the p values shown in Table A4 one can say that the difference between succes-
sive average choices is statistically positive in all cells except for the first two rounds in
the second session (B–2). In this last case, the null hypothesis of mean equality is accepted.
To include the third round of the first session in the statistical analysis, apply the
Wilcoxon nonparametric test using the same hypothesis as above. Results are summarized
in Table A5.
Given these p values, only the means of the first two rounds in the second session are
equal, and the null hypothesis is rejected for the remainder of cases.

322 JOURNAL OF ECONOMIC EDUCATION