09-045

Research Group: Finance in Toulouse May, 2009

Rational and Irrational Bubbles:

An Experiment

SOPHIE MOINAS AND SÉBASTIEN POUGET

 Rational and Irrational Bubbles: An Experiment1

Sophie Moinas
University of Toulouse (IAE and Toulouse School of Economics)
Place Anatole France, 31000 Toulouse, France
[email protected]

and

Sebastien Pouget
University of Toulouse (IAE and Toulouse School of Economics)
Place Anatole France, 31000 Toulouse, France
[email protected]

MAY, 2009 (PRELIMINARY)

1
We are grateful to Bruno Biais, Christophe Bisière, Xavier Gabaix, Jean Tirole, Reinhard Selten,
Paul Woolley, and especially Thomas Mariotti, as well as to seminar participants in Bonn University,
Luxembourg University, Lyon University (GATE), Toulouse University, for helpful comments. This re-
search was conducted within and supported by the Paul Woolley Research Initiative on Capital Market
Dysfonctionalities at IDEI-R, Toulouse.
 Abstract

This paper proposes a theory of rational bubbles in an economy with finite trading opportunities.
Bubbles arise because agents are never sure to be last in the market sequence. This theory is
used to design an experimental setting in which bubbles can be made rational or irrational
by varying one parameter. This complements the experimental literature on irrational bubbles
initiated by Smith, Suchanek and Williams (1988). Our experimental results suggest that it is
pretty difficult to coordinate on rational bubbles even in an environment where irrational bubbles
flourish. Maximum likelihood estimations show that these results can be reconciled within the
context of Camerer, Ho, and Chong (2004)’s cognitive hierarchy model, and Mc Kelvey and
Palfrey (1995)’s quantal response equilibrium.

Keywords: Experiment, rational bubbles, irrational bubbles, cognitive hierarchy model, quantal
response equilibrium.
1 Introduction

This paper presents an experimental investigation of speculative behavior in asset markets where
both rational and irrational speculation can arise. Recent economic developments suggest that
financial markets experienced various periods of bubbles and crashes. The dot com mania, and
the subprime mortgage frenzy are frequently interpreted as evidence that asset prices on financial
markets reach levels well above their fundamental value.1 Likewise, Dutch Tulip, South Sea, and
Mississippi are names often associated with the term bubble to refer to more ancient episodes
of price run ups followed by crashes. However, to the extent that fundamental values cannot be
directly observed in the field, it is very difficult to empirically demonstrate that these episodes
actually correspond to mispricings.

To overcome this difficulty and study bubble phenomena, economists have relied on the ex-
perimental methodology: in the laboratory, fundamental values are induced by the researchers
who can then compare them to asset prices. Starting with Smith, Suchanek and Williams (1988),
many researchers document the existence of irrational bubbles in experimental financial markets.
These bubbles are irrational in the sense that they would be ruled out by backward induction.
The design created by Smith, Suchanek and Williams (1988) features a double auction market
for an asset that pays random dividends in several successive periods. The subsequent litera-
ture shows that irrational bubbles also tend to arise in call markets (Van Boening, Williams,
LaMaster, 1993), with a constant fundamental value (Noussair, Robin, Ruffieux, 2001), with
lottery-like assets (Ackert, Charupat, Deaves, and Kluger, 2006), and tend to disappear when
some traders are experienced (Dufwenberg, Lindqvist, and Moore, 2005), when there are futures
markets (Porter and Smith, 1995) and no short-selling restrictions (Ackert, Charupat, Church,
and Deaves, 2005). Lei, Noussair and Plott (2001) modify the Smith, Suchanek and Williams
(1988) ’s framework to show that, even when they cannot resell and realize capital gains, some
participants still buy the asset at a price which exceeds the sum of the expected dividends to be
distributed, a behavior consistent with risk-loving preferences or judgmental errors.

The objective of the present paper is to propose an experimental design in which rational and
as well as irrational bubbles can be studied. To do so, we develop a theory of bubbles with a
finite number of trading opportunities. This theory complements the analysis of Allen, Morris,
1
We define the fundamental value of an asset as the price at which agents would be ready to buy the asset given
that they cannot resell it later. See Camerer (1989) and Brunnermeier (2008) for surveys on bubbles.

1
and Postlewaite (1993) in that it is simpler and features an economy where the bubble is common
knowledge. This theory also complements the literature on rational bubbles that considers infinite
trading opportunitites via infinite time models (see, for example, Tirole (1985) for deterministic
bubbles, and Blanchard (1979) and Weil (1987) for stochastic bubbles), via continuous trading
models (see Allen and Gorton (1993)), or via clock games (see Abreu and Brunnermeir (2004)).
We use our theoretical analysis in order to design a new experimental setting where rational
bubbles can be studied. By appropriately modifying one parameter of the experimental design
(in our case, by imposing a price cap), it is also possible to study irrational bubbles. Having
such a unified framework is useful because it allows us to compare the formation of irrational and
rational bubbles. It also enables us to study how our results on rational bubbles compare to the
previous literature on irrational bubbles. Finally, our paper analyzes individual behavior (and not
only market-level data) to better understand the sources of speculative behavior. This analysis
is based on two models of bounded rationality: the cognitive hierarchy model developped by
Camerer, Ho and Chong (2004), and the quantal response equilibrium of Mc Kelvey and Palfrey
(1995). Studying individual behavior enables to better identify and study three different types
of speculative behavior that might play a role in bubbles, namely irrational speculation (due to
mistakes or erroneneous beliefs), speculation on others’ irrationality (betting on the fact that
others are going to do mistakes), and speculation on others’ rationality (betting on the fact that
others are going to act rationally).

As noted above, our theory extends the analysis of rational bubbles to show that these can
arise even with a finite number of trading opportunities. We consider a model where the market
proceeds sequentially. The basic element that allows for the possibility of a rational bubble in
this setting is that market participants are never sure to be last in the market sequence. This is
achieved by including in the economy adequate uncertainty on participants’ position in the market
and on price paths.This idea is in the spirit of Allen and Gorton (1993). We complement their
analysis by showing that limited liability and continuous trading are not required for a bubble
to be sustainable at equilibrium. We then show that bubbles can be sustained as the outcome
of a Bayesian Nash equilibrium.2 Indeed, conditionally on observing a positive price at which
they can buy the asset (which is assumed to have a zero fundamental value), each participant
in the market sequence is better off entering the bubble. Thus, even in this zero-sum game,
2
Our model of financial market can be viewed as a generalization of a constant-sum centipede game in which
players do not know at which node of the game they play.

2
risk-neutral (or even risk averse) agents would be willing to participate. However, this situation
does not create ex-ante surplus because the unconditional expected payoffs cannot be computed
(the unconditional expectation of price levels as well as of profits and losses are infinite).

Our experimental design uses this theoretical analysis in order to empirically study bubbles.
The previous theoretical environment cannot be implemented as it is in the laboratory because
players are exposed to potentially infinite losses. In order to offer limited liability to our subjects,
we thus devise a delegated portfolio management situation that is closely linked to the theoretical
benchmark. We endow subjects one monetary unit that they can invest (and thus potentially
loose). In the paper, we refer to the subjects as portfolio managers or managers. If the amount
needed to buy the asset is higher than their endowment, managers receive capital from an outside
financier, and profits and losses are shared in proportion of the initial stakes. We check that there
exists a Bayesian Nash equilibrium of this game in which it is perfectly rational for managers to
enter a bubble and for financiers to provide capital.3 If potential prices are capped, we are back to
a situation similar to the one analyzed by the previous experimental literature in which bubbles
cannot arise at equilibrium. This enables us to compare our results with the previous findings.
It also allows us to study the impact of bounded rationality on bubble formation. Indeed, the
higher the cap is, the higher is the number of iterated steps of reasoning that is required to reach
equilibrium. By varying the level of the cap on prices, we thus vary the level of sophistication
that is needed for the subjects to realize that it is not in their interest to enter into the bubble.

Our experimental protocol is as follows. We consider a setting with three market participants.
1
Ex-ante, each subject has the same probability 3 to be first, second or third in the sequence.
Prices experience a tenfold increase at each trading date. The price proposed to the first trader
in the market is a power of 10. This power is randomly determined according to a geometric
distribution of parameter 12 . We run the experiment under different conditions. In one session,
there is no cap on prices and there thus exists a bubble equilibrium. In other sessions, there is a
cap K on the first price with K equals to 10,000, 100, or 1.

Our experimental results are as follows. First, we show that, when there is no price cap,
bubbles are not observed for sure. There is a great deal of uncertainty as whether participants
coordinate on the bubble equilibrium. Second, when there is a cap on the first price, we observe
3
In our experiment, the price path is not left at the discretion of market participants. We thus do not have
pricing results per se but instead focus on the decision to speculate by entering bubbles.

3
a lot of bubbles: the likelihood of a bubble is not significantly different whether there is a cap or
not on the first price. This result shows that in our simple design, we are able to replicate the
results found in the previous experimental literature. We complement this literature by showing,
through a regression analysis, that the propensity for a subject to enter a bubble is positively
related to the conditional probability to be first in the market sequence (from a participant’s
perspective), and to the number of reasoning steps needed to realize that subsequent participants
can realize they are last. The propensity to enter a bubble is negatively related to the conditional
probability to be last in the market sequence, and to risk aversion. These results suggest that
subjects’ decision to enter bubbles, if they are not in line with the level of rationality required to
achieve Nash equilibrium, demonstrate some level of rationality.

To better understand subjects’ behavior, we take two models of bounded rationality to the
data: Camerer, Ho and Chong (2004)’s cognitive hierarchy (CH) model, and Mc Kelvey and
Palfrey (1995)’s quantal response (QR) equilibrium. The CH model states that agents best-
respond to mutually inconsistent beliefs. In particular, the model features players with different
levels of sophistication: level-0 agents play randomly; level-1 agents believe that other players are
level-0 and best-respond; type-2 agents believe that other players are level-1 or level-0, and best-
respond,... The model considers that levels of sophistication are distributed according to a poisson
distribution. The only free parameter in this model is thus the average level of sophistication.
Maximum likelihood estimation indicates an average level of sophistication equal to 0.67 which
is in line with the estimations of Camerer, Ho and Chong (2004) on other types of games.

The QR equilibrium also takes into account the fact that players make mistake but retains
beliefs’ consistency. Players’ expectations are rational in the sense that they take into account
the fact that others make mistakes. The only free parameter is the responsiveness of players’
choice to the expected payoffs of the various actions: if it is 0, players choose uniformly among
the available actions. If the responsiveness is infinite, players best-respond. Otherwise, players
choose their actions stochastically with high expected payoff actions being more likely than low
ones. Maximum likelihood estimation indicates an average responsiveness of 2.54 in line with
previous estimations of Mc Kelvey and Palfrey (1995) on other types of games.

The CH model and the QR equilibrium both fit the data pretty well compared to the Nash
equilbrium. These theories further appear to offer a good description of behavior both in rational
and irrational bubbles. This suggests that it might be useful for market practitionners to include

4
limited sophitistication and noisy responses in their analysis when thinking about the causes and
consequences of bubbles in financial markets, and when trying to fight against these bubbles. Our
analysis suggests that insitutional features that reduce the required level of iterated reasoning
necessary to find bubbles unprofitable might be useful in mitigating the development of such
bubbles. Price caps (potentially temporary) can in this respect be useful to prevent the occurrence
of bubbles.

Our experimental analysis is related to Brunnermeier and Morgan (2006) who study clock
games both from a theoretical and an experimental standpoint. These clock games can indeed
be viewed as metaphors of “bubble fighting” by speculators, gradually and privately informed
of the fact that an asset is overvalued. Speculators do not know if others are already aware of
the bubble. They have to decide when to sell the asset knowing that such a move is profitable
only if enough speculators have also decided to sell. Their experimental investigation and ours
share two common features. First, the potential payoffs are exogenously fixed, that is, there is
a predetermined price path. Second, there is a lack of common knowledge over a fundamental
variable of the environment. In Brunnermeier and Morgan (2006), the existence of a bubble is
not common knowledge. In our setting, the existence of the bubble is common knowledge but
traders’ position in the market sequence is not. One difference between our approach and theirs
is the time dimension. The theoretical results tested by Brunnermeier and Morgan (2006) depend
on the existence of an infinite time horizon. They implement this feature in the laboratory by
randomly determining the end of the session. On the contrary, we design an economic setting in
which there could be bubbles in finite time with finite trading opportunities, even if traders act
rationally.

The rest of the paper is organized as follows. Next section presents a model in which bubbles
arise at equilibrium in an economy with finite trading opportunities. Section 3 details the exper-
imental design. Results are in Section 4. Section 5 concludes and provides potential extensions.

2 A theory of rational bubbles with finite trading opportunities

The objective of this section is to show that bubbles can emerge in a financial market with
perfectly rational traders and finite trading opportunities. Consider a financial market in which
trading proceeds sequentially. There are T agents, referred to as traders. Traders’ position in the

5
market sequence is random with each potential ordering being equally likely. Traders can trade
an asset that is issued by agent 0, referred to as the issuer.4 The first trader in the sequence
is offered to buy the asset at a price P1 . If he does so, he proposes to resell at price P2 to the
second trader. More generally, the i-th trader in the sequence, i ∈ {1, ..., T − 1}, is offered to buy
the asset at price Pi and resell at price Pi+1 to the i + 1-th trader. Traders take the price path
as given, with Pi > 0 for i ∈ {1, ..., T }. Finally, the last trader in the sequence is offered to buy
the asset at price PT but is unable to resell. If the i-th trader buys the asset and is able to resell
it, his payoff is Pi+1 − Pi . If he is unable to resell the asset, his payoff is −Pi . For simplicity,
we consider that if a trader refuses to buy the asset, the market process stops. Without loss of
generality, we asssume that the asset is worth zero. Traders are risk neutral and have an initial
wealth denoted by Wt , t ∈ {1, ..., T }.5 As a benchmark, consider the case in which traders have
perfect information, that is, each trader t knows his position in the sequence and this is common
knowledge. In this perfect information benchmark, it is straightforward to show that no trader
will accept to buy the asset except at a price of 0 which corresponds to the fundamental value
of the asset. Indeed, the last trader in the queue, if he buys, ends up with WT − PT which is
lower than WT . Since he knows that he is the last trader in the queue, he prefers not to trade.
By backward induction, this translates into a no-bubble equilibrium. This result is summarized
in the next proposition.

Proposition 1 When traders know their position in the market sequence, the unique perfect
Nash equilibrium involves no trade.

Let’s now consider what happens when traders do not initially know their position in the
maket sequence, and this is common knowledge. We model this situation as a Bayesian game.
The set of players is {1, ..., T }. The set of states of the world is Ω which includes the T ! potential
orderings. ω refers to a particular ordering. The set of actions is identical for each player t and
is denoted by A = {B, ∅} in which B stands for buy and ∅ for refusal to buy. The set of signals
that may be observed by player t is the set of potential prices denoted by P . Denote by ωit ⊂ Ω
the set of orderings in which trader t’s position in the market sequence is i. The signal function
of player t is τ (t) : ωit → Pi , in which Pi refers to the price that is proposed to the i-th trader
4
The potential bubbles that may arise in our environment can be interpreted as Ponzy schemes, and the issuer
of the asset as the scheme organizer.
5
In our model, traders might end up with negative wealth.

6
in the market sequence. The price path Pi is defined as follows. The price P1 proposed to the
first trader in the sequence is random and is distributed according to the probability distribution
g (.) on P .6 Other prices are determined as Pi+1 = fi (Pi ), with fi (.) : P → P being a strictly
increasing function that controls for the explosiveness of the price path. A strategy for player t is
a mapping St : P → A in which St (Pt ) indicates what action is chosen by player t after observing
a price Pt . We denote by S−t (Pt ) the actions chosen by players other than t when this player t
has observed Pt . Player t indeed understands that the next player in the market sequence will
observe ft (Pt ), and that he choses St+1 (ft (Pt )). Using the signal function, players may learn
about their position in the market sequence. A strategy profile {S1∗ , ..., ST∗ } is a Bayesian Nash
equilibrium if the following individual rationality (IR) conditions are satisfied:
£ ¡ ¢ ¤ £ ¡ ¢ ¤
E π St∗ (Pt ) , S−t
∗ ∗
(Pt ) |Pt ≥ E π St (Pt ) , S−t (Pt ) |Pt , ∀t ∈ {1, ..., T } , and ∀Pt ∈ P.
¡ ¢
∗ (P ) represents the payoff received by player t given that he chooses action S (P )
π St (Pt ) , S−t t t t
∗ (P ).
and that other players choose actions S−t t

We now study under what conditions there exists a bubble equilibrium {S1∗ = B, ..., ST∗ = B}.
The crucial parameter a player t has to worry about in order to decide whether to enter a bubble
¡ ¢
is the conditional probability to be last in the market sequence, P ω ∈ ωTt |Pt . The IR condition
can be rewritten as:

¡ £ ¤¢ £ ¤
1 − P ω ∈ ωTt |Pt ×(Wt + ft (Pt ) − Pt )+P ω ∈ ωTt |Pt ×(Wt − Pt ) ≥ Wt , ∀t ∈ {1, ..., T } , and ∀Pt ∈ P

¡ £ ¤¢
⇔ 1 − P ω ∈ ωTt |Pt × ft (Pt ) ≥ Pt , ∀t ∈ {1, ..., T } , and ∀Pt ∈ P.

£ ¤
If P ω ∈ ωTt |Pt = 1 for some Pt , the IR condition is not satisfied and the bubble equilibrium
does not exist. This is for example the case when the support of the distribution g (.) is bounded
above by a threshold K. Indeed, a trader upon observing Pt = f1 ◦ ... ◦ fT −1 (K) knows that
he is last and refuses to trade. Backward induction then prevents the existence of the bubble
equilibrium. The IR function is also not satisfied if the signal function τ (t) is injective. Indeed,
by inverting the signal function, players, including the one who is last in the sequence, learn what
their position is. These results are summarized in the following proposition.
6
One can consider that this first price P1 is chosen by Nature or by the issuer according to a mixed strategy
characterized by g (.).

7
Proposition 2 The no-bubble equilibrium is the unique Bayesian nash equilibrium if i) the signal
function is injective, ii) the first price is randomly distributed on a support that is bounded above,
iii) the price path is not explosive enough, or iv) the probability to be last in the market sequence
is too high.

We now propose an environment where the IR condition derived above is satisfied. Consider
that the set of potential prices is defined as P = {mn for m > 1 and n ∈ N}, that is, prices are
powers of constant m > 1. Also, assume that g (P1 = mn ) = (1 − q) q n , that is, the power n
follows a geometric distribution of parameter q ∈ (0, 1). Finally, we set fi (Pi ) = f (Pi ) = m × Pi .
Given these assumtions, upon being proposed to buy at a price of m2 , a player can be first in
the market sequence with probability 18 , second with probability 41 , and third with probability 12 .
Also, if there are T players on the market, the probability that a player t is last in the sequence,
£ ¤
conditional on the price Pt that he is proposed, is computed by Bayes’ rule: P ω ∈ ωTt |Pt = mn =
P[Pt =mn |ω∈ωT ] [ T]
t ×P ω∈ω t
(1−q)q n−(T −1) × T1 1−q £ t
¤
n =0
P[Pt =mn ] = Pj=n j 1 = 1−q T if n ≥ T − 1, and P ω ∈ ωT |Pt = m
j=n−(T −1)
(1−q)q × T
if n < T − 1. Under our assumptions, Bayes’ rule implies that i) the conditional probability to
be last in the market sequence is either 0 if the proposed price is smaller than mT −1 , and ii) if
the proposed price is equal to or higher than mT −1 , this conditional probability does not depend
on the level of the price that is proposed to the players.7 The IR condition can be rewritten:

µ ¶
q − qT
× m ≥ 1, ∀t ∈ {1, ..., T } , and ∀Pt ∈ P .
1 − qT

This condition is less restrictive when there are more traders present on the market. This
1−q T
condition is equivalent to m ≥ q−q T
, ∀t ∈ {1, ..., T }, and ∀Pt ∈ P .

There thus exists an infinity of price paths that sustain the existence of a bubble equilibrium.
Obviously, there always exists a no-bubble equilibrium.8 Indeed, if players anticipate that other
7
We implicitly assume here that players cannot observe if transactions occured before they trade. However, we
do not need such a strong assumption. For example, if each transaction was publicly announced with a probability
strictly smaller than a threshold, our results would still hold. This threshold should be small enough such that the
likelihood of being last in the sequence does not get too high.
8
When there exists a bubble-equilibrium in pure strategies, there can also exist mixed-strategies equilibria
in which traders enter the bubble with a positive probability that is lower than 1. We have characterized these
equilibria for the two-player case. They involve peculiar evolutions of the probability to enter the bubble depending
on the price level that is observed.

8
players do not enter the bubble, then they are better off refusing to trade. These results are
summarized in the next proposition.

Proposition 3 If i) the T traders are equally likely to be last in the market sequence, ii) the
price P1 proposed to the first trader in the sequence is randomly chosen in powers of m according
to a geometric distribution, and iii) Pt = m × Pt−1 , ∀t ∈ {2, ..., T }, there exists a bubble Bayesian
1−q T
Nash equilibrium if and only if m ≥ q−q T
, ∀t ∈ {1, ..., T }, and ∀Pt ∈ P . There always exist a
no-bubble equilibrium.

Our results hold even if one introduces randomness in the underlying asset payoff, and (po-
tentially random) payments at interim dates. In Appendix A, we show that our results can still
hold if traders are risk averse. One could be tempted to interpret our results as an inverse-
Hirshleifer effect: going from perfect to imperfect information seem to imply a creation of gains
from trade in our setting even with risk-neutral agents. However, note that it is not possible to
compute the ex-ante welfare created by the game of imperfect information. Indeed, this would
require computing the unconditional expectation of the price P1 that is proposed to the first
trader in the market sequence. Given our assumptions, this expectation is given by the expres-
P (1−q)[1−(qm)x+1 ]
sion E [P1 ] = limx→+∞ n=x
n=0 1−qm . This sum converges if and only if qm ≤ 1. This
1−q T
inequality is in conflict with the IR condition according to which m ≥ q−q T
> 1. This implies
that the only games in which the ex-ante welfare is well-defined are the games where only the
no-bubble equilibrium exists. This makes it hard to conclude that the imperfect information
game is actually creating welfare even if interim (that is, knowing the proposed price), all traders
are strictly better off entering the bubble if they anticipate that other traders are also going to
do so.

The next section will use these theoretical results in order to design an experimental setting
where the existence of a bubble equilibrium depends on an institutional parameter of the economy.
This allows us to study bubble formation in the laboratory in a context where bubbles can be
individually rational for all the participants.

9
3 Experimental design

This section proposes a simple experimental design in which bubbles can arise at equilibrium.
Based on the previous theoretical analysis, this design features a sequential market for an asset
whose fundamental value is 0. There are three traders on the market. 9 Trading proceeds
sequentially. Each trader is assigned a position in the market sequence and can be first, second
or third with the same probability 31 . Traders are not told their position in the market sequence
but get some information when observing the price at which they can buy the asset. Prices
are exogenously given and are powers of 10. For simplicity, in this experimental design, we do
not include the issuer of the asset. The first trader is offered to buy at a price P1 = 10n . In
1
the baseline experiment, the power n follows a geometric distribution of parameter 2, that is
1 j+1
P (n = j) = 2 , with j ∈ N The geometric distribution is useful from an experimental point
of view because it is simple to explain and implies that the probability to be last in the market
sequence conditional on the proposed price is constant. This probability is equal to 0 if the
4
proposed price is 1 or 10, and is equal to 7 otherwise.10 If he decides to buy the asset, trader i
in the market sequence proposes the asset to the next trader at a price Pi+1 = 10Pi . In order
to avoid participants from discovering their position in the market sequence by hearing other
subjects making their choices, we propose simultaneously to the first, second, and third traders
to buy the asset. Once P1 has been randomly determined, the first, second and third traders are
simultaneously offered prices of P1 , P2 , and P3 , respectively.11

If we stopped here, participants’ net payoffs (that is, their gains and losses relative to their
initial wealth) would be as illustrated in Panel A of Figure 1. Payoffs depend on the various
traders’ decisions. For example, if the first and the second traders buy while the third one refuses
to buy, payoffs are P2 − P1 , −P2 , and 0, respectively. Except for the case in which the first trader
refuses to buy (so that the bubble does not start), each potential market outcome of the game
9
We could have designed an experiment with only two traders per market. However, this would have required
higher payments for bubbles to be rational. Indeed, the conditional probability to be last would be higher. We
could also have chosen to include more than three traders per market. We decided not to do so in order to have
a high number of markets. This is useful from an econometric perspective since markets constitute statistically
independent observations.
10
The probabilities to be first, second or third conditional on the prices, which are computed using Bayes’ rule,
are given to the participants in the Instructions.
11
When a trader does not accept to buy the asset, subequent traders end up with their initial wealth whatever
their decisions.

10
translates into a loss for one of the market participants (the last one in the market sequence).
12 Since the first price is distributed on an unbounded support, this loss can be very large. This
feature is unappealing because experimental subjects cannot be asked to pay large amounts of
money. We thus introduce limited liability in a way that does not affect subjects incentive to
enter into bubbles. To do so, we rely on a delegated portfolio situation that we refer to as the
manager/financier game (as opposed to the trader game that we considered above). Each trader
in the previous game is replaced by an asset manager who is endowed with an initial wealth of 1
euro. If additional capital is required in order to buy the asset at price Pt , this additional capital
(that is, Pt − 1) is provided by an outside financier. We assume that each manager is financed
by a different financier. Net payoffs (potential gains and losses) are then divided between the
1
manager and the financier according to their share in the initial capital: Pt for the manager, and
Pt −1
Pt for the financier. Consequently, if the manager decides to buy the asset at price Pt , his loss
in case the next trader refuses to trade is:

1
× (−Pt ) = −1.
Pt

The manager has invested 1 euro (along with the Pt − 1 euros of the financier) in order to
buy an asset at a price of Pt but he is unable to resell the asset which has a liquidation value of
0. He ends up with a final wealth of 0 since he has lost 1 euro due to the fact that the bubble
bursted after he entered. Likewise, if the manager decides to buy the asset at price Pt , his gain
in case the next trader accepts to buy is:

1 1
× (Pt+1 − Pt ) = × (10Pt − Pt ) = 9.
Pt Pt

1
When the manager is able to resell the asset, he gets Pt percent of the proceed Pt+1 = 10Pt
and thus ends up with a final wealth of 10 which corresponds to 1 euro of initial wealth plus a
gain of 9 euros. Overall, whatever the price at which the manager buys, he can loose 1 euro or
win 9 euros. If he anticipates that other managers buy the asset, it is in a manager’s best interest
to also buy the asset if the following individual rationality (IRm) condition is satisfied to buy the
asset if and only if:
3 4
Um (Wm + 10) + U (Wm ) ≥ U (Wm + 1) ,
7 7
12
At each outcome of the game (except if the first trader refuses to buy), the total payoff is equal to −P 1. This
aggregate loss corresponds to the gain of the issuer of the asset (who is not part of the experiment).

11
 where Um (.) is a manager’s utility function and Wm his initial wealth. For a bubble equi-
librium to exist, the IRm condition has to be satisfied for all managers on the market. It is
straightforward to show that there exists functions Um (.) for which the IRm condition hold for
all Wm .

This IRm condition in the manager/financier game echoes the IR condition that prevails
for the trader game presented in the previous section (see appendix A, for an analysis of the
trader game with risk averse agents). The strategic incentives that agents face in both games are
similar except that in the trader game potential gains or losses are potentially infinite. The fact
that we have finite gains and losses for the manager/financier game is extremely.useful from an
experimentl point of view since it enables the experimenter to assign limited liability to subjects.

In order to show that the manager/financier game is meaningful, we now check that the
financier has an interest in providing capital to the manager. The individual rationality of the
outside financier (IRf) is written as:

3 4
Uf (Wf + ft (Pt ) − Pt − 10) + Uf (Wt − Pt + 1) ≥ Uf (Wt ) , ∀t ∈ {1, ..., T } , and ∀Pt ∈ P ,
7 7

where Uf (.) is a financier’s utility function and Wf his initial wealth. For a bubble equilibrium
to exist, the IRf condition has to be satisfied for all financiers on the market. It is straightforward
to show that there exist functions Uf (.) for which the IRf condition holds for all Wf . In order
to limit subjects’ potential losses in our experiment, the outside financiers are not part of the
experiment. The timing of the game in which subjects are involved is illustrated in Panel B of
Figure 1 (the payoffs of the financiers who are not part of the experiment are also indicated for
completeness). Notice that the sum of managers’ and financiers’ payoffs in the manager/financier
game equals the payoffs of the traders in the trader game.

The timing of the game as it was presented to subjects in our experiment is given in Figure
2.13 In this game, subjects face strategic incentives that are identical to the one faced by a
trader who has to decide to enter into a bubble. In order to study how formation of bubbles is
influenced by their rationality, we also focus on an experimental design where there is a cap K
on the first price. This will allow us to relate our work to the previous experimental literature
13
This timing does not correspond to the extensive form of the game. Indeed, it leaves aside the issue of which
player is first, second, or third. The extensive form game is provided in Appendix B for the two-player case.

12
on bubbles initiated by Smith et al. (1988) that focuses on irrational bubbles. Indeed, as shown
in the previous section, in this design, bubbles are irrational in the sense that they would be
ruled out by backward induction. The cap on the first price translates into a cap on the highest
potential price in the experiment. Upon being proposed this price, a subject should understand
that he is last in the market sequence and, consequently, should refuse to buy. Anticipating this
refusal, subjects who are proposed lower prices should also refuse to buy. At equilibrium, the
bubble never starts. However, given the experimenatal results on centipede games (see Mc Kelvey
and Palfrey (1992) or Kawagoe and Takizawa (2009), for example), we expect that this will not
happen in our experiment (because of altruism or failures in backward induction). Varying the
level of the cap K then offers potentially interesting comparative statics because it controls the
number of iterated steps of reasoning that are needed in order to reach the Nash equilibrium.
When the proposed price is P = 100K, a subject knows that he is last and there is no iterated
step of reasoning. When the proposed price is P = 10K, a subject knows that he is not first in the
market sequence (he can be second or third). At equilibrium, he has to anticipate that the next
trader in the market sequence (if any) would not accept to buy the asset. One step of iterated
reasoning is thus needed to derive the equilibrium strategy. More generally, when the proposed
¡ ¢
price is 1 ≤ P ≤ 100K, the required number of iterated steps of reasoning is log10 100K
P . In
order to study whether this required number of iterated steps of reasoning could affect bubble
formation, we have chosen to experimentally study cases in which K equals 1, 100, and 10,000.

The experimental protocol runs as follows. Our experiment includes a total of 93 subjects.
Subjects are in the last year of the Bachelor in Business Administration at the University of
Toulouse. We have four sessions with 21 to 24 subjects per session. Each subject participates
in only one session. Each session includes only one replication of the trading game.14 Subjects’
risk aversion is measured thanks a procedure which is inspired from Laury and Holt (2002). We
adjust their questionnaire in order to match the set of possible decisions to the decisions subjects
may actually face in our experiment.15 The experiment is run with paper and pencil. Subjects
are given the conditional probabilities to be first, second, and third given the price they are
proposed. The instructions for the case where K = 10, 000 are in Appendix C.
14
Because we only consider a single replication of the experiment, we thus cannot discuss learning issues that
are left for future research.
15
The questionnaire is composed of 14 decisions. For each decision i, subjects may choose between the riskless
option A, which is to receive 1 euro for sure, or the risky option B, which is to receive 10 euros with probability
i 14−i
14
, or 0 euro with probability 14
.

13
 Our experimental design is summarized in Table 1.

TABLE 1

Session Number of Subjects cap on initial price, K Equilibrium

1 24 1 no-bubble
2 21 100 no-bubble
3 24 10,000 no-bubble
4 24 ∞ no-bubble or bubble

4 Results

4.1 Market behavior

We first start by analyzing overall market behavior. At this aggregate level, we can measure
the frequency as well as the magnitude of bubbles. The frequency of bubbles is defined as the
proportion of replications where the first trader accepted to buy the asset. The magnitude of
bubbles is referred to as large if all three traders accepted to buy the asset, medium if the first
two traders accepted, and small if only the first trader accepted. Figure 3 presents the results
per session.

Figure 3 shows that there are bubbles in an environment where backward induction is sup-
posed to shut down speculation, namely when there exists a price cap. This is in line with the
previous experimental litterature cited in the introduction and contradicts our Proposition 2 (ii).
Figure 3 helps us identifying three types of speculation. First, we observe large bubbles even in
situations where the existence of a cap enables some subjects to perfectly infer that they are last
in the market sequence. A possible explanation is related to bounded rationality.16 It is possible
that some traders make mistakes and buy when they are last in the market sequence. Second,
this observation gives rise to a second motive for speculation: if initial traders in the sequence
anticipate that some participants will behave as described above, it may become optimal for them
to buy the asset. In this case, participants rationally speculate on others’ irrationality (betting on
the fact that others are going to do mistakes). Third, we observe bubbles when there is no price
16
An alternative explanation could be related to social preferences. However, extreme altruism would be required
in order for a subject to be willing to loose in order to let other subjects gain. We thus do not focus on this
interpretation.

14
cap, that is, when there exists a bubble equilibrium. This suggests that participants speculate
on others’ rationality (betting on the fact that others are going to act rationally).

Figure 3 also shows that bubbles are not more likely when there is no price cap. This indicates
that traders fail to perfectly coordinate on the bubble equilibrium. Proposition 3 indicates that
this could be due to strategic uncertainty because there always exists a no-bubble equilibrium,
or to risk aversion because if traders are sufficiently risk averse, it is not beneficial to enter the
bubble. An additional interpretation is that the possible existence of irrational traders, who may
not buy when it would be rational, increases the risk of entering the bubble for rational traders
and may prevent them from doing so.

4.2 Individual behavior

To gain more insights on bubbles formation, we now study individual decisions to buy the over-
valued asset. Figure 4 plots our entire data set. For each session, each bar represents the number
of times a given price has been proposed. Within each bar, the dark grey part corresponds to buy
decisions while the white part corresponds to refusals to trade. Figure 4 complements the results
found in the previous subsection. Let’s first focus on the case where there is no cap on the first
price. In this treatment, participants who are proposed prices of 1 or 10 buy the asset while those
who are proposed higher prices tend to refuse to trade quite often. This pattern is consistent
with Nash equilibrium if one takes into account subjects’ risk aversion: when they are proposed
a price of 1 or 10, subjects are sure not to be last, whereas when they are proposed higher prices,
they have 4 chances out of 7 to be last. Their level of risk aversion may thus become binding.
When prices are above 100, if participants coordinate on the same equilibrium, their decisions
should be the same for all price levels. We cannot reject this equality using a Wilcoxon rank
sum test that compares the proportion of buy decisions after observing prices of 100, 1,000, and
10,000 to this proportion after observing higher prices.

The high probabilities to buy in the sessions where there is a price cap are inconsistent with
Nash equilibrium. However, we keep observing the same pattern as in the situation where there
is no cap on the initial price. In the former cases though, higher prices are informative on a
trader being last in the sequence. To investigate this result, Figure 5 reports the probability to
buy, conditional on participants’ inference on their position. On the one hand, the probability to
buy decreases with the likelihood that a trader is last in the market sequence. Traders buy very

15
often when they are sure to be first and sure not to be last, while they buy half of the time when
they cannot infer their position. This is consistent both with the fact that they face a more risky
decision and with the fact that traders are reluctant to coordinate on the bubble equilibrium.
Also, this indicates that there are some elements of rationality in subjects’ decisions. On the
other hand, this result holds whether bubbles are irrational or not. In particular, the propensity
of subjects to enter a bubble is extremely large when they know that they are first or second,
even in a situation in which there exists a price cap. This result contradicts the predictions of
Nash equilibrium, so that rationality does not appear to be perfect.

These results suggest that the number of steps of iterated reasoning which are needed to
derive the equilibrium strategy may help explaining our descriptive statistics, in situations where
there exists no bubble equilibrium. We therefore run a probit regression that tries and explains
the propensity to buy the overvalued asset as a function of several variables: the number of steps
of iterated reasoning needed to derive the equilibrium strategy, the coefficient of risk aversion, the
probability to be first conditional on the observed price, and the probability to be last conditional
on the observed price. Conditional probabilities are computed following Bayes’ rule17 . A probit
regression is adequate because we try to explain a dummy variable: either the subject chooses to
buy the asset (a decision coded as 1) or not (coded as 0).

The results are in Table 2 (presented at the end of the paper) and show that, apart from the
level of risk aversion, all the explanatory variables have statistically significant coefficients. The
sign of these coefficients is intuitive: a subject’s propensity to buy the overvalued asset increases
with his probability to be first in the market sequence and with the number of steps of iterated
reasoning needed to derive the equilibrium strategy, and decreases with his probability to be
last, and with his risk aversion. The risk aversion coefficient is not significant due to a spurrious
correlation with the probability to be third. This correlation is spurrious since subjects have
been randomly assigned a position in the experiment, independently from their answers to the
questionnaire aiming at measuring their risk aversion coefficient. We therefore report in Table
2 the results of the regression excluding the conditional probability to be last. Risk aversion is
now significantly and negatively related to the propensity to enter into the bubble.
17
Remember that these probabilities were given to subjects during the experiment.

16
4.3 Fitting models of bounded rationality

Our results so far suggest that some players may have bounded rationality. The fact that some
agents either play randomly or make errors may induce rational players to enter a bubble even if
they would not do so in presence of rational traders only. Likewise, these agents with bounded
may induce rational players not to enter bubbles even if they would do so in presence of rational
traders only. To account for these phenomenon, we extend our analysis by estimating models
that explictly incorporate bounded rationality: the cognitive hierarchy model and the quantal
response equilibrium.

In cognitive hierarchy (CH) theories, traders best-respond to mutually inconsistent beliefs.

Traders differ in their level of sophistication s, and each player believes he understands the game
better than the other players. Specifically, level-0 traders play randomly, level-1 traders believe
that other traders are level-0, and level-s traders believe that other traders are a mixture of s − 1,
s − 2, , 0, and best-respond to this belief. Cognitive hierarchy models posit decision rules for
players doing s steps of thinking, that reflect this iterated process of strategic thinking. As in
Camerer, Ho and Chong (2004), we assume that traders’ types are distributed according to a
Poisson distribution. let τ denote the average level of sophistication. For each parameter τ and
each level of price P , we find the best-response of a risk-neutral level-1 trader who considers that
the next trader is a level-0 player observing P × 10. After one iteration, we find the best-response
of a risk-neutral level-2 player who considers that the next trader, observing P × 10, is a level-0
³ ´
)×τ 0 exp(−τ )×τ 0
player with probability exp(−τ0! , and a level-1 player with probability 1 − 0! . We
use a similar iterative process to find the best-response of a level-s player who considers that
exp(−τ )×τ j
the next trader, observing P × 10, is a level-j player with probability j! for j < s − 1,
and a level-(s − 1) player with the complementary probability. This process is fully discribed
in Appendix D for the case where the price cap is K = 1. This enables us to determine the
likelihood function under the assumption that subject are risk-neutral. We then estimate the
parameter τ by maximum likelihood in each session and for the entire dataset.

Results are reported in Table 3 (presented at the end of the paper). The fit of the CH-
Poisson model is compared with the fit of Nash equilibrium, by session and overall. The first two
lines describe our data, namely, the group size, and the observed average probability to buy. The
middle three lines show the predictions of the Nash equilibrium under a similar assumption of risk-

17
neutrality, the log-likelihood of this model, and its mean squared deviation.18 The mean choices
are generally far off from the Nash equilibrium; the probability to buy is too low when there exists
a bubble-equilibrium, and too high when it does not exist. The five next lines report the estimate
of the parameter τ , the predictions of the cognitive hierarchy model for this value of the estimate,
and the corresponding log-likelihood and mean squared deviation. We further compute the 90
percent confidence interval for τ estimated from a randomized resampling (bootstrap) procedure
using 10,000 simulations. We estimate an overall average level of sophistication of 0.67, which
is consistent with the estimates reported in Camerer, Ho and Chong (2004). This suggests a
high proportion of level-0 players, that is, almost 50%. Interestingly, what drives this result is
not really the fact that traders enter too much into bubbles when they should not. Indeed since
there is only around 10% of subjects who buy when they are last in the market sequence and
could understand that. This would suggest a proportion of level-0 players equal to 20%. What
explains the high estimated proportion of level-0 players is rather the fact that subjects do not
buy as much as expected by the cognitive hierarchy model (with a higher average sophistication
level) when there is no cap on the initial price or when the cap is large. The reason why subjects
do not buy as much as expected by the risk-neutral cognitive hierarchy model can be related to
risk aversion. This suggests that the average level of sophistication may be underestimated due
to the risk-neutrality assumption.

The Poisson CH model retains best-response (except for level-0 players), but it weakens
equilibrium (that is, belief-choice consistency). McKelvey and Palfrey (1995) propose an alterna-
tive approach, which retains equilibrium but weakens best-response. In their Quantal Response
Equilibrium, players may make mistakes. However, the likelihood of these errors depends on the
impact of such errors on players’ expected utility. More specifically, the following logit specifica-
tion of the error structure is assumed, so that, if the buy decision yields an expected profit of uB
while the no buy decision yields an expected profit of u∅ , the probability to buy writes:

eλuB
Pr (B) =
eλuB + eλu∅

This enables us to determine the likelihood function under the assumption that subject are risk-
neutral (Appendix D below describes how this probability to buy is computed conditionally on
18
We consider that traders coordinate on the bubble equilibrium when there is no cap on the inital price. The
no-bubble Nash equilibrium has a lower log-likelihood. In order to compute these likelihoods, we assume that
players choose non-equilibrium strategies with a probability of 0.0001.

18
the proposed price for the case where the price cap is K = 1). We then estimate the parameter λ,
which we refer to as responsiveness to expected payoffs, by maximum likelihood in each session
and for the entire dataset. Responsiveness is inversely related to the level of errors made by
subjects. The results are reported in the last five lines of Table 3. We estimate an overall average
responsiveness of 2.54, which is consistent with the results of McKelvey and Palfrey (1995). The
QRE seems to fit better our data than the cognitive hierarchy model, especially when there is a
cap on the initial price. This result seems to contradict those of Kawagoe and Takizawa (2008),
who compare the goodness of fit of both models in laboratory experiments of the centipede game.
To further investigate this issue, we compare in Figure 7 the probability to buy conditional on
the proposed price, in the CH model and in the QRE, with our observations. What the QRE
seems to better capture in our data is the drop in the probability to buy for prices larger than
P = 100. In the CH model, this pattern either does not characterize the expected outcome (see
the case in which there is no cap on the initial price), or captures it less intensively and with a lag
(see the cases in which there is a cap at K = 100 or K = 10, 000). In the QRE however, agents’
mistakes have the feature that costlier (in terms of expected payoff) mistakes are less likely. This
model is thus able to capture the drop in players’ expected utility of buying related to the fact
that they are proposed a price P ≥ 100 in which case the conditional probability to be third is
4
greater than or equal to 7 instead of a price of 1 or 10 in which case their conditional probability
to be third is zero. This feature is present in our design and is different from the centipede games
analysed by Kawagoe and Takizawa (2008).

5 Conclusion

This paper explores speculative behavior in a laboratory experiment. The objective is to be

able to better understand the various types of speculation that may be observed during bubble
episodes: irrational speculation (due to mistakes or erroneneous beliefs), rational speculation
on others’ irrationality (betting on the fact that others are going to do mistakes), and rational
speculation on others’ rationality (betting on the fact that others are going to act rationally). In
order to design an experimental environment (that necessarily induces a finite number of trading
opportunities) in which bubbles can occur at equilibrium, a theory is developped that extends
the insights of Tirole (1982), Allen and Gorton (1993), and Abreu and Brunnermeieir (2003).
The idea is to model financial markets as a game in which agents trade an asset sequentially.

19
When agents do not know at which position they are in the market sequence, it can be in their
interest to enter the bubble. We design an experimental setting based on this insight. Our design
includes several treatments that defer by only one parameter, namely the level of a cap on prices.
When there is no cap (or an infinite cap), there exists a bubble equilibrium. When there is a cap,
there is only a no-bubble equilibrium. However, the higher the cap is, the higher is the number
of iterated steps of reasoning that is needed to reach equilibrium.

Our results show that bubbles are frequently observed whether there is a price cap or not. This
relates to the previous literature on bubbles initiated by Smith, Suchanek and Williams (1988)
that shows that bubbles arise even when theory predicts they are irrational. We complement
this literature by showing that, when bubbles can be expected in theory, they do not always
materialize. We also show that the decision to speculate and enter into a bubble is positively
related to the likelihood to be first in the market sequence, and to the number of iterated steps
of reasoning required to rule out bubbles. This decision is negatively related to risk aversion
and to the likelihood to be last in the market sequence. We reconcile these results thanks to
Camerer, Ho, and Chong (2004)’s cognitive hierarchy (CH) model. In this model, players have
different levels of sophistication and best respond to lower-level players’ behavior. The average
sophistication level estimated in our data through maximum likelihood is in line with the ones
estimated by Camerer, Ho, and Chong (2004). Our results are also in line with the Quantal
Response Equilibrium developped by McKelvey and Palfrey (1995). In this model, players make
mistakes but costlier mistakes are more likely. Overall, both the cognitive hierarchy model and
the quantal response equilibrium capture two of our main experimental observations: we observe
bubbles when there is a price cap and we do not observe more bubbles when there is no price
cap. This is because both models enable one to model the following two features. First, the
presence of irrational traders, who buy when it is not beneficial, may create a rationale motive
for speculation, even if speculation would be ruled out by backward induction if all traders were
rational. Second, the existence of irrational traders, who do not buy when it would be beneficial,
decrease the expected payoffs of entering the bubble for rational traders, so that buying may
become suboptimal.

The experimental setting proposed in the present paper opens several avenues of research.
It may indeed be used to study the manipulability and controllability of speculative markets.
In particular, it could be interesting to study whether the occurrence of bubbles (rational and

20
irrational) vary with the number of traders, the introduction of risk in the underlying asset payoff,
the level of transparency (one could proxy for transparency by setting a non-null probability that
a trade is publicly announced). It would also be fruitful to study what is the impact of learning
on bubble formation. Indeed, one can expect that irrational bubbles would be less likely when
traders are experienced, whereas rational bubbles would be more frequent. Finally, it would be
interesting to extend this setting to cases in which the price path and the timing are left at the
discretion of traders. This would allow testing whether traders are able to coordinate on a price
path and a timing that sustains rational bubbles.

21
6 Appendix

Appendix A: Bubble equilibrium with risk aversion

Consider the environment in which a bubble-equilibrium exists when players are risk-neutral.
We now show that a bubble equilibrium can still exist if players are risk averse. The envi-
ronment is as follows. There are T players. The set of potential prices is defined as P =
{Pn = mn for m > 0 and n ∈ N}. P1 is randomly determined following a geometric distribution:
g (P1 = mn ) = (1 − q) q n with q ∈ (0, 1). Finally, the price path is defined as Pi+1 = m × Pi
for i ∈ [1, ..., T − 1]. For simplicity, we assume that utility functions are piecewise linear with
a kink at agents’ initial wealth , that is palyer t’s utility function is: Ut (x) = x1x≤Wt +
[Wt + (1 − γt ) (x − Wt )] 1x>Wt ,.where γt ∈ [0, 1] is a measure of player t’s risk aversion. The
IR condition is now written as:

 ¡ £ ¤¢ 
1 − P ω = ωTt |Pt × Ut (Wt + ft (Pt ) − Pt )
 £ ¤  ≥ Ut (Wt ) , ∀t ∈ {1, ..., T } , and ∀Pt ∈ P
t
+P ω = ωT |Pt × Ut (Wt − Pt )

 ¡ £ ¤¢ 
1 − P ω = ωTt |Pt × [Wt + (1 − γt ) (ft (Pt ) − Pt )]
⇔ £ ¤  ≥ Wt , ∀t ∈ {1, ..., T } , and ∀Pt ∈ P.
+P ω = ωTt |Pt × (Wt − Pt )

£ ¤
P ω = ωTt |Pt × Pt
⇔ γt ≤ 1 − ¡ £ ¤¢ , ∀t ∈ {1, ..., T } , and ∀Pt ∈ P.
1 − P ω = ωTt |Pt (ft (Pt ) − Pt )

(1 − q)
⇔ γt ≤ 1 − , ∀t ∈ {1, ..., T } .
(q − q T ) (m − 1)

This inequality indicates that, if players are not too risk averse, there exists a bubble equi-
librium. It also shows that, when m gets larger, the range of risk aversion for which a bubble
equilibrium exists is larger.

22
Appendix B: Extensive form of the game with two players
At each node, Nature (N), player i or player –i choose an action. (x;y) represents the payoffs; x for
player i, and y for player –i. Dotted lines relate nodes that are observationally equivalent. b refers to
the buy decision, nb to the refusal decision.
-i b (10;0)
b
i (0;1)
nb
1/2 -i b (1;1)
N nb
(1;1)
n=0 nb
1/2 -i b (0;10)
i b
nb (1;1)
1/2
-i b (1;0)
nb
nb (1;1)

-i (10;0)
i (0;1)
1/2 -i (1;1)
N N
1/4 (1;1)
n=1
1/2 -i (0;10)
i (1;1)
-i (1;0)
1/8
(1;1)

-i (10;0)
i (0;1)
1/2
1/16 -i (1;1)
N (1;1)
n=2
1/2 -i (0;10)
i (1;1)
-i (1;0)
… (1;1) 23
 Appendix C: Instructions for the case where K = 10, 000

Welcome to this market game. Please read carefully the following instructions. They are
identical for all participants. Please do not communicate with the other participants, stay quiet,
and turn off your mobile phone during the game. If you have questions, please raise your hand.
An instructor will come and answer.

As an appreciation for your presence today, you receive a participation fee of 5 euros. In
addition to this amount, you can earn money during the game. The game will last approximately
half an hour, including the reading of the instructions.

Exchange process

To play this game, we form groups of three players. Each player is endowed with one euro
which can be used to buy an asset. Your task during the game is thus to choose whether you want
to buy or not the asset. This asset does not generate any dividend. If the asset price exceeds one
euro, you can still buy the asset. We indeed consider that a financial partner (who is not part
of the game) provides you with the additional capital and shares profits with you according to
the respective capital invested. The market proceeds sequentially. The first player is proposed to
buy at a price P1 . If he buys, he proposes to sell the asset to the second player at a price which
is ten times higher, P2 = 10 × P1 . If the second player accpets to buy, the first player ends up
the game with 10 euros. The second player then proposes to sell the asset to the third trader at
a price P3 = 10 × P2 = 100 × P1 . If the third player buys the asset, the second player ends up
the game with 10 euros. The third player does not find anybody to whom he can sell the asset.
Since this asset does not generate any dividen, he ends up the game with 0 euro. This game is
summarized in the following figure19 .

At the beginning of the game, players do not know their position in the market sequence.
Positions are randomly determined with one chance out of three for each player to be first, second
or third.

Proposed prices

The price P1 that is proposed to the first player is random. This price is a power of 10 and
is determined as follows:
19
See Figure 2 in the present paper.

24
 Price P1 Probability that this price is realized
1 1/2 (50%)
10 1/4 (25%)
100 1/8 (12.5%)
1,000 1/16 (6.3%)
10,000 1/16 (6.3%)

Players decisions are made simultaneously and privately. For example, if the first price P1 = 1
has been drawn, the prices that are simultaneously proposed to the three players are: P1 = 1 for
the first player, P2 = 10 for the second player, and P3 = 100 for the third player. Identically,
if the first price P1 = 10, 000 has been drawn, the prices that are simultaneously proposed to
the three players are: P1 = 10, 000 for the first player, P2 = 100, 000 for the second player, and
P3 = 1, 000, 000 for the third player.

The prices that you are been proposed can give you the following information regarding your
position in the market sequence:

- if you are proposed to buy at a price of 1, you are sure to be first in the market sequence;

- if you are proposed to buy at a price of 10, you have one chance out of three to be first and
two chances out of three to be second in the market sequence. You are sure not to be last;

- if you are proposed to buy at a price of 100 or 1,000, you have one chance out of seven to be
first, two chances out of seven to be second, and four chances out seven to be last in the market
sequence;

- if you are proposed to buy at a price of 10,000, you have one chance out of four to be first,
one chance out of four to be second, and two chances out four to be last in the market sequence.

- if you are proposed to buy at a price of 100,000, you have one chance out of two to be
second, and one chance out of two to be third. In this case, you you are sure not to be first in
the market sequence.

- if you are proposed to buy at a price of 1,000,000, you are sure to be last in the market
sequence.

In order to preserve anonymity, a number will be assigned to each player. Once decision will
be made, we will tell you (anonymously) the group to which you belong, your position in the

25
market sequence, if you are proposed to buy, and your final gain.

Do you have any question?

26
 Appendix D: Solving and estimating models of bounded rationality
Consider the environment in which the cap on the initial price is equal to K = 1. We derive
the conditional probabilities to buy for risk-neutral traders observing prices of P = 1, P = 10
and P = 100 respectively, in the Poisson-cognitive hierarchy model of Camerer et al. (2004),
and in the logit-quantal response equilibrium model of Mc Kelvey and Palfrey (1995). These
probabilities are then used to estimate the deep parameter of each model thanks to maximum
likelihood technics.

Cognitive hierarchy model

Consider first the case of a trader observing a price P = 100. This trader perfectly infers from this
observation that he is third in the sequence. Consequently, in the CH model, he only buys if he is a
τ 0 ×exp(−τ )
level-0 player. Given that there is a fraction f (0) = of such traders in the population,
0!
³ ´
and given that these traders buy randomly with probability Pr B|P = 100, S̃ = 0 = 12 , the
probability to observe a trader buying at a price of P = 100 is:

1
Pr (B|P = 100) = exp (−τ )
2

Consider now the case of a trader observing a price P = 10. This trader perfectly infers from
this observation that he is second in the sequence.
³ ´
- If he is a level-0 player, he buys with probability Pr B|P = 10, S̃ = 0 = 21 .

- If he is a level-s player with s ≥ 1, he thinks that the next player observing the price
P3 = 10 × P2 is a level-0 with probability f (0) = exp (−τ ), a level-1 with probability f (1) =
P
τ ×exp (−τ ),..., and a level-s−1 player with the truncated probability 1− s−2
i=0 f (i).Consequently,

his expected profit if he buys writes:

f (0) 1
EΠs≥1 (B|P = 10) = Ps−1 × × 10.
i=0 f (i)
2
Ps−1 τi
He is strictly better off buying if and only if i=0 i! < 5. This condition depends on s and on τ .
Consider for instance the case of a level-1 player. He thinks that the next player observing the
price P3 = 10 × P2 is a level-0 player with probability 1, and that such a trader would buy with
probability 12 . Consequently, his expected profit if he buys is EΠs=1 (B|P = 10) = 21 ×10 which is
strictly larger than his profit if he does not buy, that is, 1. So level-1 players buy with probability
³ ´
Pr B|P = 10, S̃ = 1 = 1. But if he is a level-2 player, he thinks that the next player is a level-0

27
with probability f (0) and a level-1 with the complementary probability 1 − f (0).Consequently,
f (0) 1
his expected profit if he buys decreases to EΠs=2 (B|P = 10) = P1 × 2 × 10. This is lower
i=0 f (i)

that the expected payoff of a level-1 trader since a level-2 perceives that there is a lower proportion
of level-0 players who would buy when observing P = 100. Consequently, a level-2 player would
only enter for lower levels of τ than level-1 players, namely for τ < 4.
Ps−1 τi Ps−1 τi
Now, i=0 i! is increasing in s and lims→∞ i=0 i! = eτ . If τ ≤ ln (5), then the inequality
Ps−1 τi
i=0 i! < 5 holds for all s: the proportion of level-0 players who buy after observing a price
P3 = 100 is sufficiently high to induce all higher level players to buy when they observe P2 = 10.
Ps∗1 −1 τ i
If τ > ln (5) however, there exists a s∗1 ≥ 1 such that s∗1 -level players buy, that is, i=0 i! < 5,

but not level-s∗1 + 1 players who have a more accurate perception of the distribution of lower-level
Ps∗1 −1 τ i
types, that is, i=0 i! ≥ 5.

Finally, given the distribution of players’ types, the probability to buy conditionnal on the
price being P = 10 writes:
" ∞ µ ¶#
1 X τs
Pr (B|P = 10) = exp (−τ ) × + × 1Ps−2 τ i <5
2 s! i=0 i!
s=1

To compute this probability as a function of τ , we use the following process.

i) If the condition τ ≤ ln (5) is satisfied, then

1
For τ ≤ ln (5) : Pr (B|P = 10) = 1 − × exp (−τ )
2

ii) Otherwise, we define s∗1 and:

 ∗ 
s1
X τ s 1
For τ > ln (5) : Pr (B|P = 10) = exp (−τ ) ×  − ,
s! 2
s=0
s∗1 −1 s∗1
X τi X τi
with s∗1 ∈ ℵ∗ such that: ≤ 5 and > 5.
i! i!
i=0 i=0

Consider finally the case of a trader observing a price P = 1. This trader perfectly infers from
this observation that he is first in the sequence.
³ ´
- If he is a level-0 player, he buys with probability Pr B|P = 1, S̃ = 0 = 12 .

- If he is a level-s player with s ≥ 1, he thinks that the next player observing the price
P2 = 10 × P1 is a mixture of level-0, level-1 ... level-s − 1 players. Now, depending on the value

28
of τ , although level-i players would not buy at date 3 for i > 0, we have shown above that they
may be willing to buy at date 2 if they anticipate that they would be able to resell to third level-0
players sufficiently frequently. The incentive to buy of a sophisticated player is thus higher after
observing 1 than after observing 10. His expected profit writes:
 Ps−1 ³ ´ 
f (0) 1 i=1 Pr B|P = 10, S̃ = i
EΠs≥1 (B|P = 1) =  Ps−1 × + Ps−1 × 1s>1  × 10.
i=0 f (i) 2 i=0 f (i)

Again, this expected profit depends on the value of τ . Straightforward manipulations lead to the
following condition for a level-s player to buy when observing P = 1:
s−1
à à !!
X τj
4+ × 10 × 1 P 5 >1 − 1 × 1s>1 > 0
j! j−1 τ i
j=1 i=0 i!

³ P ´
τj
Now, if the condition τ ≤ ln (5) is satisfied, then the inequality becomes 4 + 9 × s−1
j=1 j! ×
1s>1 > 0, which always holds. If τ > ln (5), s∗1 , defined above, is such that level-s∗1 + 1 players
would not buy when observing P2 = 10. But for s > s∗1 , we can rewrite the first element as
Ps∗1 −1 τ j Ps−1 τ j
4 + 9 j=1 j! − j=s∗ j! , which is decreasing in s. Given that
1

s∗1 −1
s−1 j
∗
s1 −1 j
X τj X τ X τ
lim 4 + 9 − = 5 + 10 − exp (τ ) ,
s→∞ j! ∗
j! j!
j=1 j=s1 j=1

depending on the value of τ , there may exist a s∗2 > s∗1 such that the condition () is not satisfied.

Finally, given the distribution of players, the probability to buy conditionnal on the price
being P = 1 writes:
  
1 ∞  s
X 
 τ    
Pr (B|P = 1) = exp (−τ ) ×  + τ +  ×1 
2  k! 4+
Ps−1 τ j
s=2  ×10×1 −1>0

j=1 j! 5 >1
Pj τi
i=0 i!

To compute this probability as a function of τ , we use the following process.

i) If the condition τ ≤ ln (5) is satisfied, then

1
For τ ≤ ln (5) : Pr (B|P = 1) = 1 − × exp (−τ )
2

ii) Otherwise, we define s∗1 as above.

ii-a) If the following condition is satisfied:

29
 s∗1 −1
X τj
5 + 10 − exp (τ ) ≥ 0
j!
j=1

then:

s∗1 −1
X τj 1
For τ such that 5 + 10 − exp (τ ) ≥ 0 : Pr (B|P = 1) = 1 − × exp (−τ )
j! 2
j=1

ii-b) Otherwise, we defined s∗2 and:

 ∗ 
s∗1 −1 s2
X τj X τ s 1
For τ such that 5 + 10 − exp (τ ) < 0 : Pr (B|P = 1) = exp (−τ ) ×  − 
j! s! 2
j=1 s=0

with s∗2 ∗
∈ ℵ such that:
k1∗ k2∗ −1 k1 ∗ ∗
k2
X τj X τj X τj X τj
4+9× − ≥ 0 and 4 + 9 × − <0
j! ∗
j! j! ∗
j!
j=1 j=k1 +1 j=1 j=k1 +1

Quantal response equilibrium

Let ui,B be the expected payoff of a risk-neutral player if he buys after observing P = mi−1 ,
and ui,∅ his expected payoff if he does not buy. In the quantal response model, the probability
with which the trader buys conditional on observing P = Pi writes:

eλui,B
Pr (B|P = Pi ) =
eλui,B + eλui,∅

Consider first the case of a trader observing a price P = 100. This trader perfectly infers from
this observation that he is third in the sequence. Consequently, his expected payoffs for buying
and not buying respectively write:

u3,B = 0

u3,∅ = 1

The probability to buy is therefore:

1
Pr (B|P = 100) = .
1 + eλ

30
Consider now the case of a trader observing a price P = 10. This trader perfectly infers from
this observation that he is second in the sequence. To compute his expected payoff from buying,
he anticipates that the probability to buy of the third trader is not equal to zero. His expected
payoffs for buying and not buying respectively write:

u2,B = Pr (B|P = 100) × 10

u2,∅ = 1

The probability to buy is therefore:

10λ
e 1+eλ 1
Pr (B|P = 10) = 10λ = ³
10
´.
λ× 1−
e 1+eλ + eλ 1+e 1+eλ

Consider finally the case of a trader observing a price P = 1. This trader perfectly infers from
this observation that he is first in the sequence. To compute his expected payoff from buying, he
anticipates that the probability to buy of the second trader is not equal to zero. His expected
payoffs for buying and not buying respectively write:

u1,B = Pr (B|P = 10) × 10

u1,∅ = 1

The probability to buy is therefore:

10λ
e 1+eλ
e10λ 10λ
e 1+eλ +eλ 1
Pr (B|P = 1) = 10λ =  .
1+eλ
10λ e  µ 10

10λ λ1− ¶
1+eλ λ 1− 10
e e +eλ + eλ 1+e 1+e 1+eλ

31
7 References

References

Abreu D., and M. K. Brunnermeier, 2003, Bubbles and Crashes, Econometrica, 71(1), 173-204.

Ackert L. F., N. Charupat, B. K. Church, and R. Deaves, 2005, Bubbles in Experimental Asset
Markets: Irrational Exuberance No More, mimeo.

Ackert L. F., N. Charupat, R. Deaves, and B. Kluger, 2006, The Origins of Bubbles in Laboratory
Asset Markets, mimeo.

Allen, F., and G. Gorton, 1993, Churning Bubbles, Review of Economic Studies, 60, 813-836.

Allen, F., S. Morris, and A. Postlewaite, 1993, Finite Bubbles with Short-Sale Constraints and
Asymmetric Information, Journal of Economic Theory, 61, 206-229.

Blanchard O., 1979, Speculative Bubbles, Crashes, and Rational Expectations, Economic Letters,
3, 387-389.

Brunnermeier M. K., 2008, Bubbles, Entry in The New Palgrave Dictionary of Economics, edited
by Steven Durlauf and Lawrence Blume, 2nd edition, Forthcoming.

Brunnermeier M. K., and J. Morgan, 2006, Clock Games: Theory and Experiments, Competition
Policy Center, Paper CPC07-072.

Camerer C., 1989, Bubbles and Fads in Asset Prices: A Review of Theory and Evidence, Journal
of Economic Surveys, vol. 3, 3-41.

Camerer C., T.-H. Ho, and J.-K. Chong, 2004, A cognitive hierarchy model of one-shot games,
Quarterly Journal of Economics, vol. 119, 861-898.

Dufwenberg M., T. Lindqvist, and E. Moore, 2005, Bubbles and Experience: An Experiment,
American Economic Review, vol. 95, 1731-1737.

Holt, C. A., and S. K. Laury, 2002, Risk Aversion and Incentive Effects, American Economic
Review, 92, 1644-1655

Kawagoe, T. and H. Takizawa, 2009, Level-K Analysis of Experimental Centipede Games, mimeo.

32
Lei V., C. Noussair and C. Plott (2001), Nonspeculative Bubbles in Experimental Asset Markets:
Lack of Common Knowledge of Rationality vs. Actual Irrationality, Econometrica, 69(4), 831.859.

Mc Kelvey, R. D., and T. R. Palfrey, 1992, An Experimental Study of the Centipede Game,
Econometrica, vol. 60, 803-836.

McKelvey, R. D. and T. R. Palfrey, 1995, Quantal Response Equilibrium for Normal Form Games,
Games and Economic Behavior, 10, 6-38.

Noussair, C., S. Robin, and B. Ruffieux, 2001, Price Bubbles in Laboratory Asset Markets with
Constant Fundamental Values, Experimental Economics, 4, 87-105.

Porter, D. P & Smith, V. L, 1995, Futures Contracting and Dividend Uncertainty in Experimental
Asset”, University of Chicago Press, vol. 68, 509-41.

Smith V., G. Suchanek, and A. Williams, Bubbles, Crashes, and Endogenous Expectations in
Experimental Spot Asset Markets, Econometrica, 1988, 56, 1119-1151.

Tirole J., On the possibility of speculation under rational expectations, Econometrica, 1982, 50,
1163-1181.

Tirole J., Asset Bubbles and Overlapping Generations, Econometrica, 1985, 53, 1499-1528.

Van Boening, M.V., Williams, A.W., and LaMaster S. 1993. Price bubbles and crashes in exper-
imental call. markets. Economic Letters, 41, 179-185.

Weil P., 1987, Confidence and the Real Value of Money in an Overlapping Generations Economy,
Quarterly Journal of Economics, 102, 1-22.

33
 Figure 1 – Panel A: Timing of the market in the trader game
Market proceeds sequentially. Traders are equally likely to be first, second or third. The first, second and
third traders are offered to buy at prices P1, P2, and P3 respectively. This figure displays traders’ net payoff,
that is their gains or losses relative to their initial wealth. Traders can end up loosing very large amount of
money.

P1 Buy
P2 Buy
P3 Buy
(P2-P1,P3-P2,-P3)

Buy Buy Buy

(0,0,0) (-P1,0,0) (P2-P1,-P2,0)

Figure 1 – Panel B: Timing of the market in the manager/financier game

Market proceeds sequentially. Portfolio managers can invest one monetary unit along with additional capital
if needed coming from an outside financier. Managers are equally likely to be first, second or third. The first,
second and third managers are offered to buy at prices P1, P2, and P3 respectively. This figure displays
managers’ net payoff, that is their gains or losses relative to their initial wealth. Financiers’ net payoffs are
indicated for completeness but these agents are not part of the experiment.

P1 Buy
P2 Buy
P3 Buy
(9,9,-1)
(P2-P1+9,P3-P2+9,-P3+1)
Buy Buy Buy

managers (0,0,0) (-1,0,0) (9,-1,0)

financiers (0,0,0) (-P1+1,0,0) (P2-P1+9,-P2+1,0)

Figure 2: Timing of the market and payoffs in the experiment

Market proceeds sequentially. Subjects are endowed with one monetary unit that they can invest in a
financial asset. If they do so, they receive ten monetary units if the next subject also decides to invest (the
additional capital that is potentially needed to buy the asset is provided by outside financiers that are not part
of the experiment). Subjects are equally likely to be first, second or third. The first, second and third subjects
are offered to buy at prices P1, P2, and P3 respectively. This figure displays subjects’ payoff (if they refuse to
buy they keep one monetary unit).

P1 Buy
P2 Buy
P3 Buy
(10,10,0)

Buy Buy Buy

(1,1,1) (0,1,1) (10,0,1)

34
 Figure 3: Probability to observe bubbles, depending on the cap on the
initial price
80,00%

70,00%

60,00%

50,00%

40,00%

30,00%

20,00%

10,00%

0,00%
No cap Cap K=10,000 Cap K=100 Cap K=1 all

no bubble small bubbles medium bubbles large bubbles

Figure 4

Buy decision, depending on the initial price # no buy Buy decision, depending on the initial price # no buy
No cap on the initial price #buy Cap on the initial price K=10,000 #buy

8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
0 0
1 10 100 1 000 10 000 100 000 1 000 000 10 000 1 10 100 1 000 10 000 100 000 1 000 000 10 000 000
000

Buy decision, depending on the initial price # no buy Buy decision, depending on the initial price # no buy
Cap on the initial price K=100 #buy Cap on the initial price K=1 #buy

8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
0 0
1 10 100 1 000 10 000 100 000 1 000 000 10 000 000 1 10 100 1 000 10 000 100 000 1 000 000 10 000 000

35
 Figure 5: Probability to buy conditional on subjects' inferences

100%

90%

80%

70%

60%

50%

40%

30%

20%

10%

0%
FIRST NOT T HIRD NOT FIRST T HIRD NO INFERENCE

Figure 6

Probability of Buy decision, depending on the initial price Probability of Buy decision, depending on the initial price
No cap Cap K=10,000

100% 100%
90% 90%
80% 80%
70% 70%
60% 60%
50% 50%
40% 40%
30% 30%
20% 20%
10% 10%
0% 0%
1

1
10

10
0

0
0

0
0

0
10

10
00

00

00

00

00

00
00

00

00

00
1

10

10
0

00

0
10

00

10

00

00
10

10

10

CH: Tau = 0.41 LogL= -15.28 CH: Tau = 1.37 LogL= -10.36
QR: Lambda = 0.67 LogL= -5.44 Data CH QR QR: Lambda = 2.69 LogL= -7.99 Data CH QR

Probability of Buy decision, depending on the initial price Probability of Buy decision, depending on the initial price
Cap K=100 Cap K=1

100% 100%
90% 90%
80% 80%
70% 70%
60% 60%
50% 50%
40% 40%
30% 30%
20% 20%
10% 10%
0% 0%
1

10

00

00
1

10

10

00

00

00
10

00

00

00

00

00

00

00
1

10

0
1

10

10

00

00
10

00

00

10
1

10

CH: Tau = 0.27 LogL= -13.96 CH: Tau = 2.65 LogL= -9.26
QR: Lambda = 2.49 LogL= -7.53 Data CH QR QR: Lambda = 1.93 LogL= -8.31 Data CH QR

36
Table 2 Logit Regression on the Buy Decision

Coefficient Statistic p-value Coefficient Statistic p-value

Number of steps of iterated reasoning from maximal price 0.47 2.38 0.02 0.40 2.33 0.02
Risk aversion -0.82 -1.48 0.14 -1.40 -3.02 0.00
Conditional probability to be first 1.96 1.54 0.12 2.54 2.24 0.03
Conditional probability to be third -1.23 -2.12 0.03

Log likelihood -22.68 -25.28

Number of observation 63 63

Table 3 Comparison of fits of Nash, Cognitive Hierarchy and Quantal Response Equilibrium

Session
No Cap Cap K=10,000 Cap K=100 Cap K=1 All
Data
Sample size 24 24 21 24 93
Av. probability buy 67% 83% 48% 58% 65%

Nash Equilibrium
Av. probability buy 100% 0% 0% 0% 26%
Log L -73,68 -184,21 -92,10 -128,95 -478,94
Mean Squared Deviation 0,11 0,74 0,36 0,47 0,42

Cognitive Hierarchy
Tau 0,41 1,37 0,27 2,65 0,67
Av. probability buy 67% 86% 58% 56% 69%
Log L -15,28 -10,36 -13,96 -9,25 -53,04
Mean Squared Deviation 0,000 0,042 0,116 0,004 0,054
90% CI [0 - 1.10] [0.67 - 3.91] [0 - 1.79] [0.87 - 4] [0.48 - 1.12]

Quantal Response
Lambda 0,67 2,69 2,49 1,93 2,54
Av. probability buy 55% 79% 48% 58% 58%
Log L -5,44 -7,99 -7,53 -8,31 -25,90
Mean Squared Deviation 0,016 0,033 0,001 0,000 0,022
90% CI [0.56 - 2.52] [1.77 - 2.86] [0.54 - 2.83] [0.46 - 1.16] [0.51 - 2.74]

37