Information, Reputation and Imitative Choice. A Simple Bayesian Model

Full text

EJESS – 19/2006. Beliefs, Norms and Markets, pages 191 to 212
Information, reputation and imitative choice
A simple Bayesian model
Sylvain MARSAT
*
Faculté de Sciences Economiques et de Gestion
41, Boulevard François Mitterrand
63002 Clermont Ferrand Cedex
and
Centre de Recherches en Sciences de Gestion (CRSG)
Sylvain.marsat@u-clermont1.fr
Submitted November 2005 – Revised October 2006
ABSTRACT
. How does an agent choose an investment when his private information and the
behaviour of other preceding actors are opposed? What are the factors encouraging him to act
independently of the behaviour of others, or, on the contrary, to imitate them? We propose an
extension of the Bayesian model of Scharfstein and Stein (1990), in order to introduce the
informational aspect developed by Bikhcandani, Hirshleifer and Welch (1992) and Orléan
(1989, 1990, 1992). Agent B can be induced to imitate an agent or a group of agents A
preceding him (i) because the information held by A is more reliable than his own information,
(ii) because agent B a priori relies more on agent A than on its own abilities or (iii) because he
doesn’t want to deviate from A, in order to preserve his reputation. We thus seek to synthesize
the informational and reputational approaches in order to better understand their respective
importance and relationships in imitative investment choices.
KEYWORDS
: Herding Behaviour, Bayesian Modelling, Decision Making, Investment Choice,
Reputation, Informational Cascade
DOI:10.3166/EJESS.19.191-212 © 2006 Lavoisier, Paris
*
Université d’Auvergne – Clermont Ferrand 1, Faculté de Sciences Economiques et de Gestion, 41
Boulevard François Mitterrand, 63 002 Clermont-Ferrand Cedex, Centre de Recherches en Sciences de
Gestion (CRSG), E-Mail : sylvain.marsat@ecogestion.u-clermont1.fr. I am grateful for helpful comments
and advices from Lorraine April, Solange Gannat, André Orléan, Jean-Pierre Védrine, Benjamin
Williams, the participants at the ACSEG 12th Annual Meeting in Aix-en-Provence, and two anonymous
referees who helped me to improve the paper with their numerous suggestions. All remaining errors are
my own.

Introduction
« Most people are other people. Their thoughts are
someone else's opinions, their lives
a mimicry,their passions a quotation. »
Oscar Wilde (1854 - 1900), De Profundis, 1905
From teenagers’ fashions to financial investments, sports in vogue to medical practices,
social influence has a significant impact in the majority of human activities (Shiller
[1984]). In a great number of their decisions, individuals take the choices and opinions
of others into account. For the social psychologist Bandura (1986), observation of
others is an essential element in the human learning process.
According to Moschetto (1998), the emergence of the concept of imitation in economy
and management was not an easy thing. Indeed, the economic agent in the classical
theory is rational, independent of the other individuals, as Katona (1953) mentions it.
Mimicry is often attached, probably because of the ethologic origin of the concept, with
gregarious behaviours, inevitably animal and irrational.
However, based on a keynesian approach, Orléan (1989) shows that rationality and
imitation are not inevitably antinomic. Indeed, if an agent does not know anything
about the outcome of a decision, he may find it beneficial to imitate another agent who
is at least as informed as himself. In a more general way, the behaviour of the other
agents partly reveals information that they hold on the situation and, within this
framework, imitation makes possible to optimize a decision in a fully rational manner.
Two major mainstreams used a Bayesian approach to explain imitation in investment
decisions. Seminal work from Bikhchandani, Hirshleifer and Welch (hereafter referred
to as BHW, 1992) models a mimetic chain based on the use of others’ behaviour,
revealing their own information. In this framework, information signals have the same
precision among individuals. If only two individuals made the same decision, the third
one, if he acts in a rational Bayesian way, must follow the others whatever his private
information. An informational cascade is engaged and every agent who arrives then
rationally follows.

Scharfstein and Stein (1990) propose an alternative approach by linking imitation and
reputation (Holmström [1982]). The agent does not seek to imitate in order to profit
from the information transmitted by other agents’ behaviours, but because if he deviates
from the preceding actors, he will be judged by the principal as being not qualified
enough in the task assigned to him. In fact, it is assumed that good informed managers
act in the same way in response to “good identical information”. As the informative
signals of two good informed people are correlated, the agent seeks through imitation to
be considered as a good informed agent, and thus to preserve his reputation.
This paper jointly analyzes some of the various motivations which induce an agent to
imitate another one, and their relationships. We propose an extension of the model of
Scharfstein and Stein (1990), in order to introduce the informational aspect approached
in particular by Bikhchandani, Hirshleifer and Welch (1992) and Orléan (1989, 1990,
and 1992). After a presentation of the stakes and assumptions of this simple model (1),
the impact of the reliability of information is analysed (2) as well as a priori confidence
(3), and reputation (4).
1 Assumptions
1.1 Objectives and settings
The aim of this paper is to develop a clearer understanding of the choice of information
for an agent, B, thanks to a Bayesian model. We seek to identify some of the forces that
can lead to imitation, individuals neglecting their personal information to join the group.
This analysis applies generally to any decision of investment including the framework
of the decisions of portfolio managers on financial markets
1
.
Scharfstein and Stein (1990) rely on reputation to explain imitative behaviour. This
work is analysed in detail by the famous comment of Ottaviani and Sorensen (2000),
who develop it and show how this model can be close to the BHW (1990) cascade
model. The main difference between the two models is in the correlation of signals used
by Schafstein and Stein (1992), which can be released partially (Graham [1999]) or
totally (Ottaviani and Sorensen [2000]).
Most of the theoretical models postulate signal independence in their settings. This
allows multi agent settings’ convenient analysis. Chamley (2004, p.228) rejects the idea
of signal correlation, which he considers as “irrelevant and confuses the issue”,
1
As Scharfstein and Stein do (1990, p.477), considering a perfectly elastic supply.

postulating that the decision model of each agent is perfectly known by the principal.
However, if not needed for herding, this correlation can be interesting insofar as it
introduces another payoff for the agent, rewarding agents acting like previous ones.
Scharfstein and Stein (1990) show that in uncertainty, principals are based to update
their beliefs on two pieces of evidence: (1) whether the investment was profitable and
(2) whether the agent’s behaviour was similar to other agents. Since systematically
unpredictable components affect the profitability, competent agents can receive
misleading signals. The fact that all made the same choice can be viewed as a proof of
ability. Reputation is then considered to be higher if agents are acting together, which is
another constraint and payoff for an agent.
In the following formalization, we consider a principal, delegating his portfolio, with
bounded rationality, since he does take into account the fact that the agent could act
strategically and fool him. This assumption could seem too simplistic but is consistent
with clients who have a clear preference for relative performance, without interest on
the possible effects of this evaluation on the manager, or accepting them. This effect has
been highlight by Maug and Naik (1996): principals often “prefer the insurance
possibilities of a relative performance contract to the higher returns available if fund
managers do not herd “.
We present a simple binary signal and dichotomous choice model based on Scharfstein
and Stein (1990). The independence signal setting will be used to present basic
informational conclusions (sections 2 and 3), adding the a priori ability of agents.
Hence, section 4 will introduce the judgement of a principal. If this principal has the
same information set, the decision of the agent will not be affected at all. An asymmetry
of beliefs is then postulated to introduce the relative performance bias: the principal will
estimate the ability of the agent on both the probability of making a good decision and
the similarity of behaviour with the preceding agent.
1.2 Theoretical assumptions
In order to introduce the concept of reputation, we rely on the work of Scharfstein and
Stein (1990) and Ottaviani and Sorensen (2000). We consider two states of the world,
{R=P} or {R=N}. {R=P} corresponds to the positive outcome of an investment, and
{R=N} corresponds to the negative outcome of the same value
2
. We consider the state
2
Agent’s utility is considered to be symmetric, assuming x
l
+ x
h
= 0 in Scharfstein et Stein (1990)
framework, as it is assumed in Birkhchandani, Hirshleifer et Welch (1992).

of the world R has a prior probability of ω to be positive, and (1 - ω) to be negative. In
order to simplify the model
3
, we set ω = ½. The price is exogenously fixed.
In this very simple framework, two agents A and B act sequentially. Agent A can be
regarded as one or more agents
4
. In order to make a decision, agent B receives a private
signal which can take value in {-; +}. A {+} signal is a buying signal, i.e. indicates that
profitability is positive. Conversely, a {-} signal indicates to the agent that he must sell
to avoid a negative profitability. On the stock market, this signal comes from his own
fundamental interpretation of the value of the stock and we therefore call it the
fundamental signal, Sf. Agents A and B have to optimize their portfolio in executing a
binary choice: to buy if they estimate that the future profitability of the action will be
positive, and to sell otherwise.
The interpretation of this signal is complex insofar as there are two types of agents on
the market. The first type corresponds to rational, informed agents called “smart”.
Conversely, “dumb” is the category of the agents acting in an irrational way
5
. Agent B,
just like agent A, is unaware of the category in which he belongs, “smart” or “dumb”,
whose prior probabilities are:
P(A=smart) = P(B=smart) = θ
P(A=dumb )=P(B=dumb)= (1- θ)
If he belongs in the “smart” category, the fundamental signal received by agent B is
informative, which means that the signal {Sf=+} has more probability to occur in the
{R=P} state than in the {R=N}state:
P(Sf=+/R=P, B=smart) = p = P(Sf=-/R=N, B=smart)
P(Sf=+/R=N, B=smart) = 1-p = P(Sf=-/R=P, B=smart)
With p > ½
3
This assumption “eliminates any incentive for manager 1 to signal his ability by deviating from the
efficient outcome” (Avery and Chevalier [1999]). Agent A knowing that ω is close to 0 is incited to sell
to be seen as an informed agent. With ω= ½, agent A has no a priori incentive to sell or buy, and uses
only her private information signal as a cutoff.
4
If agent A is a group of agents, we consider this group to be globally perceived by agent B, and have
only aggregated characteristics.
5
They can be qualified of “noise traders” by the literature see for instance De Long et al. (1990). Noise
traders can however also be motivated by liquidity constraints, which is not the case in this framework.

On the contrary, if agent B belongs to the category “dumb”, which occurs with a
probability (1- θ), he receives completely uninformative fundamental signals and has as
much probability of receiving the signal {Sf=+} than the signal {Sf =-}.
P(Sf=+/R=P, B=dumb) = P(Sf=+/R=N, B=dumb) = ½
We assume that agent B receives also a signal called “Sm”, the signal of agent A, who
acted before him. This “mimetic” signal thus corresponds for agent B to the action of
the preceding agent. The reliability q of the mimetic signal depends on the category of
agent A. If he is “smart”, this signal is informative with a reliability q:
P(Sm=+/R=P, A=smart) = q = P(Sm=-/R=N, A=smart)
P(Sm=+/R=N, A=smart) = 1-q = P(Sm=-/R=P, A=smart)
With q > ½
On the other hand, an agent A “dumb” does not transmit information through his
decisions because he also receives uninformative signals:
P(Sm=+/R=P, A=dumb) = P(Sm=+/R=N, A=dumb) = ½
2 Simple informational Bayesian choice
2.1 Choice of agent B with a private signal
Let us suppose that agent B takes his decision starting from only one signal of
information: his private, fundamental signal. One can easily calculate the probabilities
of realization of the events according to the received signal with P(R=P)= ½ and
P(Sf=+)= ½
6
:
P(R=P/Sf=+)= )/,()/,(
+
=
=
=
+
+
=
=
=
SfdumbBPRPSfsmartBPRP =
)().,/()().,/( dumbBPdumbBPRSfPsmartBPsmartBPRSfP
=
=
=
+
=
+
=
=
=
+
=
(2.1.1)
P(R=P/Sf=+)= ½ + θ (p- ½) (2.1.2)
In the same way,
P(R=P/Sf=-)= ½ + θ (½ - p) (2.1.3)
P(R=N/Sf=-)=½ + θ (p- ½) (2.1.4)
P(R=N/Sf=+)=½ + θ (½ - p) (2.1.5)
6
See proof in appendix.

The choice of agent B depends on two parameters: θ and p. These two variables enable
him to increase his probability of making a sound decision, compared to a random
choice, of probability ½. The reliability of the signal makes it possible to improve the
choice, insofar as this signal is informative
7
. The confidence of agent B in his ability,
i.e. the probability that the actor considers himself as “smart” plays a multiplying role of
this reliability of the signal.
Following Scharfstein and Stein (1990), we assume that the investment is attractive if a
positive signal was received and on the other hand that it is not attractive if the signal is
negative. Insofar the fundamental signal is informative, p
]
[
1;5.0∈
then if θ>0
8
:
P(R=P/Sf=+)= P(R=N/Sf=-)> ½
P(R=P/Sf=-)= P(R=N/Sf=+)< ½
According to this result, agent B follows his fundamental signal whatever it is, since the
state of the world {R=P} (respectively {R=N}) has a greater probability of being
realized if the received fundamental signal is positive: {Sf=+} (resp. negative: {Sf =-
}).
2.2 Choice of agent B on private signal and information on A’s
behaviour
Let us suppose a sequential investment. In a first period, agent A decides to invest or
not. During the second period, agent B chooses, with his own fundamental signal and
the behaviour of agent A. The signal transmitted by the action of A will constitute for
agent B a “mimetic” signal, Sm, with reliability q and as the first mover, his action
reveals his signal. Sf and Sm are considered as independent.
It is impossible to distinguish an imitative behaviour if two information signals induce
the same choice. The description of an informational choice thus requires a divergence
of the mimetic and fundamental behaviours. We will study the situations for which
agent B must deal with two contradictory signals, which correspond to the events
{Sf=+, Sm=-} and {Sf=-, Sm=+}. From these two information signals, B must make a
binary choice: to follow his fundamental signal, or the mimetic signal, according to the
buying and selling signals received. If the second agent only acts on the basis of the two
7
i.e. when p > ½. Indeed, if p = ½, the signal does not inform more the agent B which is facing a
random choice because P(R=P/Sf=+)= P(R=P/Sf=-)= ½.
8
If θ =0, agent B does not believe in his own ability and the choice is no more that a random one. If p= ½
the choice amounts to a random choice too.

independent signals which are transmitted to him, one can calculate the probability of
the event {R=P}, given the signals {Sf=+, Sm = -}
9
:
P(R=P/Sf=+, Sm = -) =
)/()./()/()./( )/()./(
NRSmPNRSfPPRSmPPRSfP
PRSmPPRSfP =−==+=+=−==+=
=
−
=
=
+
=
(2.2.1)
And thus
10
:
P (R=P/Sf=+, Sm=-) = )]
2
1
(
2
1
)].[
2
1
(
2
1
[()]
2
1
(
2
1
)].[
2
1
(
2
1
[
)]
2
1
(
2
1
)].[
2
1
(
2
1
[
−+−++−+−+
−+−+
qpqp
qp
θθθθ
θθ
(2.2.2)
The rule of decision making according to the information signals is as follows
11
:
- if P (R=P/Sf=+, Sm = -) > ½, the probability that profitability is positive is
higher than ½, agent B buys;
- if P (R=P/Sf=+, Sm = -) < ½, the probability that profitability is positive is
lower than ½, agent B sells.
As in agent B’s decision having only the fundamental signal, θ plays an amplifying role
of the reliability of information
12
. In this configuration nevertheless, this amplification
takes place in the same way for the fundamental signal and the mimetic one, which
cancels its effect. In other words, the probability for agent A of being “smart” is equal
to that of the agent B to be “smart”. Under these conditions, parameters p and q, the
reliability of the signals received by the agents, make the difference. Agent B chooses
the decision whose reliability is higher: he buys if p>q, sell if p<q. If p=q, he chooses
one or the other with probability ½.
9
We now consider the {Sf=+, Sm = -} signals. These results can be extended to {Sf=-, Sm = +} given the
symmetry of assumptions.
10
See proof in appendix.
11
P(R=P/Sf=+, Sm = -) can be seen as a relative a posteriori confidence between A and B
12
If θ =1, the reliability of the signal p is equal to the posterior probability P(Sf = +/R=P) = P(Sf = -
/R=N). If p=q, this model converges with that of Birkhchandani, Hirshleifer and Welch (1992). Within
our framework of analysis, the reliability of mimetic signal is not a direct function of the fundamental
signal (q=f(p)), as it is the case in the mimetic sequence suggested by these authors, who postulate that q
increases with the number of actors who chose the same action in such a way that as soon as two
succeeding people acted in concert, q is systematically higher than p, therefore which it is
informationnally rational to imitate. We do not study the mimetic mechanism of sequence and we restrict
with a model with two actors, in the line of Scharfstein and Stein (1990).

0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,5
0,55
0,6
0,65
0,7
0,75
0,8
0,85
0,9
0,95
1
p
P(R=P/Sf=+,Sm=
-)
Figure 1. Variation of P(R=P/Sf=+, Sm=-)
Parameters: q=0.6; θB=θA=½
As an example, let q= 0.6 and θ = ½. In this special case, P(R=P/Sf=+, Sm = -) vary
according to p on figure 1. This probability is lower than ½ when p<0.6, and that on the
contrary, when p>0.6, is higher than ½. Agent B which has a mimetic signal of a
reliability of 0.6 will thus imitate A when the reliability of the mimetic signal q is higher
than the reliability of the fundamental signal, p.
Let us suppose nevertheless that the actor receives a highly reliable fundamental signal,
but that his preferences encourage it to act according to the mimetic signal. The
experiment of Asch (1951) presents a striking illustration. An individual subjected to
these tests of length recognition has a negligible probability to be mistaken. However, in
more than one third of the cases, he chooses to act as the group. Asch mentions that
some subjects hesitate, lack confidence in themselves and thus feel a strong tendency to
join in the majority. This dimension
13
can be integrated by differentiating the capacities
13
Of course, overconfidence is a well documented bias in the behavioural literature. Asch’s experiment is
invocated to motivate that people have a relative-low confidence in their ability. An agent can be
overconfident (θ
B”overconfident”
) and however less confident in his ability than in agent A (θ
B”overconfident”
< θ
A
).

of actors A and B. Beyond the reliability of the signal, the a priori confidence of the
ability of actors can explain some herding behaviours.
3 Informational influence and a priori confidence in
the source
3.1 Distinction of a priori confidence and consequences
Let us suppose now that the ability of the two actors have different probabilities for
agent B. As Avery and Chevalier (1999), we assume that agents have information about
their ability
14
. There are therefore two distinct prior probabilities: the a priori
probability of ability of agent B himself and of the first agent A
15
.
P(A=smart) = θ
A
; P(A=dumb)= (1- θ
A
)
And P(B=smart) = θ
B
; P(B=dumb)= (1- θ
B
)
Agent B does not know to which category he belongs, but estimates the probabilities of
agent A and himself to be “smart”. The choice does not then depend any more on the
only reliability of information signals, but also on the a priori confidence the actor
grants to the agents. In this case, the equation (2.2.2) becomes:
P (R=P/Sf=+, Sm=-) = )]
2
1
(
2
1
)].[
2
1
(
2
1
[()]
2
1
(
2
1
)].[
2
1
(
2
1
[
)]
2
1
(
2
1
)].[
2
1
(
2
1
[
−+−++−+−+
−+−+
qpqp
qp
ABAB
AB
θθθθ
θθ
(3.1.1)
The decision rule remains unchanged: agent B invests when the probability of a
profitability knowing the two signals is positive, otherwise he sells. In this new
configuration, the prior probabilities play a differentiated role on p and q. When θ
A
=
θ
B
, one goes back to the preceding case of an informational Bayesian choice only based
on the reliability of information. If θ
A
> θ
B
, agent B grants a greater a priori confidence
in the ability of agent A than in his own ability and this induces an amplification of the
mimetic signal q more significant than that of the fundamental signal: agent B
propensity to imitate is thus more important. An individual who does not feel capable,
and lacks confidence in his abilities, will have a low θ
B
. Consequently, his future
Overconfidence can explain why people are experimentally less prone to follow the herd (with
θ
B”overconfident”
> θ
B
,) which is consistent with Kramer, Nörth and Weber (2006).
14
Differing from Avery and Chevalier (1999), this information is not private since these probabilities are
assumed to be common knowledge.
15
Graham (1999, p.241) uses the term “initial reputation” to define this probability.

decision will be affected by minimizing the impact of the reliability of individual
information. Otherwise, if θ
A
< θ
B
, agent B relies more on his own ability and he will
attach a priori more importance to the reliability of fundamental signal, p. His
propensity to imitate is thus less significant.
Let us once again take an example with q=0.6 and θ
A
= 0.7 while varying a priori
confidence of agent B in his ability. As shown in figure 2, the more θ
B
is weak, i.e. the
less the agent believes in his own capacities, the more p must be significant so that he
can adopt a fundamentalist behaviour. If θ
B
is very weak (for example θ
B
= 0.1), the
agent B imitates whatever the reliability of the private information is. On the other
hand, if agent B relies a lot on his capacities (θ
B
= 0.9), then a less reliable signal
(p<0.6) can nevertheless enable him to adopt a fundamentalist behaviour.
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
0,5
0,6
0,7
0,8
0,9
1
p
P(R=P/Sf=+, Sm=-)
θ
B
= 0,1
θ
B
= 0,3
θ
B
= 0,5
θ
B
= 0,7
θ
B
=0,9
Figure 2. Variation of P(R=P/Sf=+, Sm=-)
Parameters: q=0.6; θ
A
= 0.7
3.2 Reliability of information signal and credibility of the source

The interest of this concept of credibility is to distinguish for the actor (i) the
information objectively received and his interpretation (p and q), of (ii) the beliefs or the
preferences on their source, which also plays a role, whatever this information signal is.
One could thus differentiate in the precision of the signal two different components:
- the acquisition of information and interpretation in buying and selling signals of
a more or less great reliability (p and q);
- its weighting according to personal criteria with the agent, relating to the ability
of the agents (θ
A
and θ
B
).
Most of the models stemming from BHW (1992) elude this difference, assuming that all
agents receive signals of the same probability (p=q) and that all have a perfect
confidence in their ability (θ
A
= θ
B
=1). Differentiating these two parameters can be
interesting since in an informational efficient market with no informational asymmetry
(p=q), the difference of ability can be a factor explaining herding behaviour for a
second agent, less confident in his a priori ability. This is consistent with several
empirical studies, such as Lamont (1995), Hong and al. (2000), or Chevalier and Ellison
(1999) who find that younger decision-makers herd more than their older counterparts,
more experienced and therefore having more prior probability of being “smart”
16
. An
interesting experiment of Cote and Sanders (1997) propose the same information to
evaluate a firm: subjects who were less confident in their ability were proved to be more
influenced by consensus. Signal precision seems then to have an endogenous
component, depending on the subject and not the only information, captured by this
prior confidence.
4 Reputational aspects
Asch (1951) notes, in order to explain the conformism observed during his experiment,
the assumption according to which subjects proceed in a “distortion of the action”.
Even if they think they are right, they are afraid that their “true judgement” isolate them,
and to be badly perceived by the group. They do not give up their judgement which they
believe true, but modify the action and act in order to conform to others’ judgement.
Thus, they do not adopt the judgement of the group -which would be a “distortion of
judgement”- but only its action -from there this concept of “distortion of action”-.
We consider an agent B sensitive to the evaluation of his behaviour by the principal.
The agent thus worries about the signal transmitted by his behaviour. If this signal is
16
i.e. with E(θ
youngerDMSr
) < E(θ
olderDMSr
), DMS for decision makers.

different to agent A’s choice, he may have problems with his reputation. As Keynes
(1936, p.158) mentions it, “worldly wisdom teaches that it is better for the reputation to
fail conventionally than to succeed unconventionally”.
In the absolute, agents should only seek to maximize their profit expectation concerning
the investment, and thus to use information indicating that to invest has a positive
expectation, as we saw previously. However, two smart managers could receive two
misleading signals because of systematically unpredictable components. Therefore,
similarity in A and B behaviours is considered as an important piece of information for
the principal, estimating the ability of agent B.
Scharfstein and Stein (1990) decide to allow external observers
17
, customers, employers
or peers, to update their beliefs on agents’ abilities.
θ
ˆ
is their revised estimate of the
probability that a manager is “smart”. Taking into account the competition between
managers, the level of wages is regarded as directly related: a manager perceived as
being more qualified will have higher wages. Insofar as their wages increase in a linear
way, the managers are incited to maximize the expected value of
θ
ˆ
, rather than to
invest in an efficient way.
We postulate that the principal has a bias towards relative performance but does not take
into account that the agent may act strategically. The principal either (i) delegated the
decision and do not want to spend time or resources on it, or (ii) is conscious of the
possible strategic action of agent B but clearly prefers a relative performance than a
higher return, implying a deviant choice, as pinpointed by Maug and Naik (1996). This
is consistent with the study of Arnwald (2001) who shows this relative performance
concern in a survey among German portfolio managers: their first objective is to beat
the benchmark, their first risk, under-performance. Likewise, evaluation criteria are
relative to benchmarks and results of comparable funds, by far more than absolute
return or risk-adjusted measures.
4.1 Correlation of signals
Acting like others can be considered for the principal as a proof of ability, and relative
performance insurance, since investment decision contains systematically unpredictable
components. Introducing this aspect in the model, an asymmetry of beliefs between
17
We will now refer to the principal, estimating the ability of the agent B who has access to the same set
of information than agent B but assumes the correlation of signals between two smart agents.

agent B and the principal is needed
18
. In this agency relationship, the principal updates
the ability of agent B according to the probability of signals, and assuming they are
correlated if both are “smart”. Agent B, who believes in an independent signal setting,
knows the principal estimation function as well as this signal correlation, and will try to
optimize his reputation.
Correlation
19
of signals between “smart” agents has been introduced by Scharfstein and
Stein (1990). Two agents are assumed to observe exactly the same signal if they are
“smart”. “Smart” agents receive perfectly correlated signals since they are assumed to
share the same vision of part of the “truth”, whereas the signals of “dumb” agents are
not correlated at all, since they observe a noise which has nothing to do with
fundamentals
20
. In other words, if two agents receive different signals, at least one of
them must be dumb. Then, the probability does not follow an independent draw relating
to the initial distributions from the principal standpoint.
Since for the principal the signals of “smart” agents are perfectly correlated, agent B
tends to make the same decisions as A, indicating they have received the same
information, and thus belong to the “smart” category. If the choice of agent B deviates
from that of agent A, this means that the signal Sf received is different from Sm,
therefore that his signal is different from that of the group. In this case, either agent B is
“dumb”, or agent A is “dumb”, or both are “dumb”. On the other hand, if agent B does
not use his information and follows agent A, the principal estimation function will judge
that the two agents have a strong probability of being “smart” insofar as, apparently, the
two agents seem to act according to same information
21
.
4.2 Estimation fonction
θ
ˆ
Let us examine now when the second manager, B, seeks to maximize the expected value
θ
ˆ
B
in the event {Sf=+,Sm=-,R=P}.
θ
ˆ
B
(Sf=+,Sm=-, R=P) is the revision of the
estimate of the principal i.e. the posterior probability that the manager B is “smart”
18
If the principal perfectly knows the agent problem and constraints, agent B has no stronger incentive to
follow agent A and he his placed in a simple informative decision as in the previous sections. Without
signal correlation, the estimation function is not biased and correctly estimates the choice of the agent
according to an efficient Bayesian choice.
19
Here, only perfect correlation is analysed. This could be extended to partially correlated signals (see
Graham [1999])
20
Thus, the probability that the two agents observe {Sf=+, Sm=+ } when the state of the world is {R=P}
is q.
21
Scharfstein and Stein (1990) assume that two agents making the same error « share the blame », i.e.
that if both made a bad investment, it was because of an unpredictable factor and not because of their
ability.

given these events, assuming the correlation of signals. According to the definition of
the estimation function, the probability agent B to be “smart” with {Sf=+, Sm = -}
signal set is
22
:
θ
ˆ
B
(Sf=+, Sm=-, R=P) = )1)(1()1()1(2)1(2 )1(2
ABABAB
AB
qp
p
θθθθθθ
θ
θ
−−+−−+−
−
(4.2.1)
In the same way, the probability that the agent is “smart” if {R=N} is:
θ
ˆ
B
(Sf=+, Sm=-, R=N) = )1)(1()1(2)1)(1(2 )1)(1(2
ABABAB
AB
qp
p
θθθθθθ
θ
θ
−−+−+−−
−
−
(4.2.2)
And the a posteriori expectation of reputation while acting according to the signals he
acquires is :
Θ
B
(+,-)=
θ
ˆ
B
(Sf=+, Sm=-, R=P) P(R=P/ Sf=+, Sm=-)
+
θ
ˆ
B
(Sf=+, Sm=-, R=N) P(R=N/ Sf=+, Sm=-) (4.2.3)
This expectation can easily be calculated thanks to equations (3.1.1), (4.2.1) and
(4.2.2)
23
.
4.3 Agent B hides his private information
If agent B decides to imitate A, he transmits to the market a signal on his ability. He
shows that his private or fundamental signal is coherent with the signal transmitted by
the agent or the group of agents A. Thus let us suppose that the individual chooses to
imitate, i.e. that he deliberately chooses to transmit a signal different from his private
information. This signal transmitted thus means that the private signal of information is
coherent with the mimetic signal, which is not the case since both signals are opposed.
In this case, one obtains the following probabilities of his estimation by the principal:
θ
ˆ
B
(Sf=-, Sm=-, R=P) = )1)(1()1)(1(2)1(4)1)(1(2 )1(4)1)(1(2
BABABAAB
BAAB
qqp
qp
θθθθθθθθ
θ
θ
θ
θ
−−+−−+−+−−
−
+
−
−
(4.3.1)
22
On numerator is the probability of agent B being smart with a {Sf=+} signal. Since signals are
correlated, A and B being smart is not possible with different signals. On the denominator, all the possible
cases: B smart and A dumb, A smart and B dumb, and both dumb, according to {Sf=+} and {Sm=-}
signals.
23
See the detailed equation in appendix.

and
θ
ˆ
B
(Sf=-, Sm=-, R=N) = )1)(1()1(2)1(24 )1(24
BAABABBA
ABBA
qpq
pq
θθθθθθθθ
θ
θ
θ
θ
−−+−+−+
−
+
(4.3.2)
With equations (3.1.1), (4.3.1) and (4.3.2) the expectation of a posteriori reputation is:
Θ
B
(-,-)=
θ
ˆ
B
(Sf=-, Sm=-, R=P) P(R=P/ Sf=+, Sm=-)
+
θ
ˆ
B
(Sf=-, Sm=-, R=N) P(R=N/ Sf=+, Sm=-)
(4.3.3)
4.4 Optimal choice with reputation
According to the simplifications of their model, the choice of signal is not informational
any more
24
and Scharfstein and Stein (1990) show that an optimal behaviour is
systematically mimetic. In this framework, herding behaviour depends on the
parameters p, q, θ
A
and θ
B
. An optimal decision for agent B is then to have the highest
expectation of a posteriori reputation:
- if Θ
B
(-,-)<Θ
B
(+,-), agent B acts according to his private signal;
- if Θ
B
(-,-)>Θ
B
(+,-), agent B hides his private signal and follows agent A.
When agent B hides his private information not to appear inadequate towards others
actors, because of the correlation of the signals, he has the possibility to increase the
probability of being perceived like “smart”. In fact, he has by far a more significant
propensity to hide his private information compared to an informational choice as we
saw in sections 2 to 4.
By adding the dimension of reputation, the conditions by which the agent B acts in a
fundamental, non imitative way are much more restrictive: he must receive a
fundamental signal of a great reliability p, have a great confidence in this signal θ
B
but
also a weak confidence θ
A
in the ability of agent A, of which the reliability of the signal
q must be low (close to 0.5). Let us take once more the example where q is 0.6 and θ
A
=θ
B
= ½. For agent B, which is his estimation by the principal when he shows or when
he hides his private information?
24
i.e. P(R=P/Sf=+, Sm = -) = ½.

If θ
A
= θ
B
=½, P(R=P/Sf=+, Sm=-) is higher than ½ when p exceeds q, i.e. when p>0.6.
By taking into account his reputation, agent B puts aside this reliability of information
to stick to the maximization of his perceived ability by the principal. One can see that
this constraint of maximization means that he must systematically hide his information
(Θ
B
(-,-)>Θ
B
(+,-)) whatever the reliability of his private information may be.
Figure 3. Variation of P(R=P/Sf=+, Sm=-), Θ
B
(-,-) and Θ
B
(+,-))
Parameters: q=0.6; θ
A
= θ
B
=0.5
What does occur if one modifies the values of θ
A
and θ
B
? Let us take for example a
higher a priori confidence of agent B in his ability than those of agent A, so that θ
A
=0.3
and θ
B
=0.7 (see figure 4 below).
The modification of a priori confidence changes estimation fonctions Θ
B
(-,-) and Θ
B
(+,-). Because a priori confidence in the ability of A is less significant, and confidence
in his own capacities greater, agent B can be found in situations in which it is not
optimal to imitate. These situations are nevertheless very restrictive. Indeed, the agent is
rationally induced to act according to his signal (Θ
B
(+,-)>Θ
B
(-,-)) only when this

signal is highly reliable and almost equal to one. His fear of being regarded as “dumb”
if he uses his own information leads him to a very great prudence. The concept of
reputation developed by Scharfstein and Stein (1990) thus adds to the simple quality of
information a dimension of the pressure of the group, outside informational
maximization. Agent B may find it beneficial to change his decision, in order to
transmit, by his action, a signal which identifies him by the principal as being a “smart”
agent.
Figure 4. Variation of P(R=P/Sf=+, Sm=-), Θ
B
(-,-) and Θ
B
(+,-))
Parameters: q=0.6; θ
A
=0.3; θ
B
=0.7
If agent B acts isolated, it is necessary for him to have a lot of confidence in his ability,
a highly reliable private information, and at the same time a low reliability and
credibility in agent A. Otherwise, if he has uncertainties about his private information,
agent B seeks to reduce its idiosyncratic risk: he imitates, for fear of feeling set aside.
With the requirement of reputation maximization, agent B thus has a tendency to focus
his attention more on the action of agent A and less on his private signal compared to an
informationally efficient decision.
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
0,5
0,6
0,7
0,8
0,9
1
p
P(R=P /Sf=+,Sm=-)
Θ
B
(+,-)
Θ
B
(-,-)

Conclusion
Through a principal-agent asymmetry of beliefs, this simple model allows the analysis
of the decision of an agent B facing two kinds of payoffs: an informational one and a
reputational one. As in traditional informational models
25
, signal precision is an
important component in the decision. This model introduces a priori confidence in
ability. Low ability introduces noise and affects the precision of signal, which is
consistent with some empirical works
26
showing that more experienced agents are less
prone to herd.
The principal’s signal correlation belief modifies the pure informational decision of
agent B, since the estimation function of ability rewards taking the same behaviour than
agent A. Aware of this agency problem, agent B then tries to fool the principal in acting
like the previous agent. Reputation is analysed as a constraint fostering imitation,
already present for informational reasons.
This framework allows informational and reputational factors to be studied together.
The most important results are the view of imitation in a multiple factor analysis,
building a bridge between informational and reputational works. It is particularly
interesting to analyze the impact of traditional “informational” parameters like signal
reliability and prior ability on reputation. An agent B with high reliability signal and a
high a priori confidence will not herd neither for informational reasons nor for
reputational ones, since the principal’s estimation is then based more on these
parameters than on the similarity of behaviour between A and B. This result was not
possible in Scharfstein and Stein (1990) framework in which signals were not
informative.
This simple model proposes a few concrete hypotheses on imitation. A natural further
task would be to develop it with partial correlation, a dynamic approach and the
introduction of a price mechanism. Besides, testing the theoretical hypotheses through
an experimental setting, and finding empirical evidence of the different factors involved
in herding behaviour in investment decisions seems to be a fertile area for future
research.
25
These models tem from BHW (1992).
26
e.g. Chevalier and Ellison (1999).

References
ARNWALD, T., 2001, “Investment Behaviour of German Equity Fund Managers: An Exploratory
Analysis of Survey Data”, Discussion paper 08/01, Economic Research Centre of the Deutsche
Bundesbank
ASCH S. E., 1951, “Effects of group pressure upon the modification and distortion of judgements”, in
Guetnow, H. (ed.) Group, leadership and men, Pittsburg, Carnegie Press, pp.177-190
AVERY, C., CHEVALIER J., 1999, “Herding over the Career”, Economics Letters, n° 63, p.327–33
BANDURA, A, 1986, L’apprentissage social, Bruxelles, Belgique : Pierre Mardaga Eds.
BIKHCHANDANI, S., HIRSHLEIFER, D., WELCH, I., 1992, “A Theory of Fads, Custom, and Cultural
Change as Informational Cascades”, Journal of Political Economy, vol. 100, n°5, pp.992-1023
CHAMLEY, C., 2004, Rational Herds: Economic Models of Social Learning, Cambridge Press
University, Cambridge (UK), 402p.
CHEVALIER, J., ELLISON, G., 1999, “Career concerns of mutual fund managers”, The Quarterly
Journal of Economics, Vol.114, n°2, pp. 389-432
COTE, J., SANDERS, D., “Herding Behavior: Explanation and Implications”, Behaviorial Research in
Accounting, vol.9, pp.20-45
DELONG, J., SHLEIFER, A., SUMMERS, L., WALDMAN, R., 1990b, “Noise Trading Risk in
Financial Markets”, Journal of Political Economy, vol.98, n°4, pp.703-738
GRAHAM, J., 1999, “Herding among Investment Newsletters: Theory and Evidence”, Journal of
Finance, vol.54, n°1, pp.237-268
HOLMSTRÖM, B., 1982, “Managerial Incentive Problems: a Dynamic Perspective”, in Essays in
Economics and Management in Honor of Lars Wahlberck, Helsinki: Swedish School of Economics
HONG, H., KUBIK, J., SOLOMON, A., 2000, “Security Analysts Career Concerns and the Herding of
Earnings Forecasts”, Rand Journal of Economics, vol.31, pp.121-44
JENSEN M.C., MECKLING W.H., 1976, “Theory of the Firm: Managerial Behavior, Agency Costs and
Ownership Structure”, Journal of Financial Economics, vol.3, pp.305-360
KATONA, G., 1953, “Rational Behavior and Economic Behavior”, Psychological Review, vol.60,
pp.307-318
KEYNES, J. M., 1936, The General Theory of Employment, Interest and Money, London, Macmillan
KRAMER, C., NÖRTH, M., WEBER M., 2006, “Information Aggregation with Costly Information and
Random Ordering: Experimental Evidence”, Journal of Economic Behavior & Organization, vol. 59, n°3,
p. 423-432
LAMONT, O., 1995, “Macroeconomic forecasters and microeconomic forecasts”, NBER, Working paper
5284
MOSCHETTO, B.-L., 1998, Mimétisme et marchés financiers, Economica, Paris, 249p.
MAUG, E., NAIK, N., 1996, “Herding and Delegated Portfolio Management”, London Business School
mimeo.

ORLEAN, A., 1992, « Contagion des opinions et fonctionnement des marchés financiers », Revue
Economique, vol.43, n°4, pp.685-697
ORLEAN, A., 1990, « Le rôle des influences interpersonnelles dans la détermination des cours boursiers
», Revue Economique, vol.41, n°5, pp.839-868
ORLEAN, A., 1989, « Comportement mimétique et diversité des opinions sur les marchés financiers », in
BOURGUINAT, Henri, ARTUS, Patrick, Théorie économique et crises des marchés financiers,
Economica, Paris, 208p., pp.45-65
OTTAVIANI, M., SORENSEN, P., 2000, “Herd Behavior and Investment: Comment”, American
Economic Review, vol.90, pp.695-704
SCHARFSTEIN, D., STEIN, J., 1990, “Herd Behavior and Investment”, American Economic Review,
vol. 80, n°3, pp.465-479
SHILLER, R., 1984, “Stock Prices and Social Dynamics”, Brookings Papers on Economic Activity, vol.2,
pp.457-510
Appendix
A. List of notation
Notation Meaning
R State of the world
P Profit
N Loss
ω P(R=P), a priori probability of a profit (set to ½)
Sm Signal from the choice of agent(s) A
Sf Private signal
+ Buying signal
- Selling signal
p P(Sf=+ /R=P, B=smart)
Reliability of Sf, when B is « smart »
q P(Sm=+ /R=P, A=smart)
Reliability of Sf, when A is « smart »
θ
A
P(A=smart)
ex ante probability that agent A is “smart”
θ
B
P(B=smart)
ex ante probability that agent B is “smart”
θ
ˆ
B
a posteriori probability that agent B is “smart”
Θ
B
a posteriori expectation of reputation of l’agent B
B. Proof that P(Sf=+)= ½
P(Sf=+) = P(Sf=+/B=smart).P(B=smart)+P(Sf=+/B=dumb).P(B=dumb)
= P(Sf=+/B=smart, R=P).P(B=smart).P(R=P)+ P(Sf=+/B=dumb, R=P).P(B=dumb).P(R=P)
+ P(Sf=+/B=smart, R=N).P(B=smart).P(R=N)+ P(Sf=+/B=dumb, R=N).P(B=dumb).P(R=N)
And from the hypotheses:

P(Sf=+) = p. θ. ½ + ¼.(1- θ)+ ½.(1-p). θ+ ¼.(1- θ)
Then: P(Sf=+) = ½, Q.E.D.
C. Equation (2.2.2) - Proof
P(R=P/Sf=+, Sm = -) =
)/()./()/()./( )/()./(
NRSmPNRSfPPRSmPPRSfP
PRSmPPRSfP =−==+=+=−==+=
=
−
=
=
+
=
(2.2.1)
With:
P(Sf =+/R=P)= P(Sf=+ /R=P, B=smart) . P(B=smart) + P(Sf=+/R=P, B=dumb). P(B=dumb)
= ½ + θ (p- ½)= P(Sf=-/R=N)
P(Sf=- /R=P) =P(Sf=- /R=P, B=smart) . P(B=smart) + P(Sf=- /R=P,B=dumb). P(B=dumb)
= ½ + θ (½ - p)= P(Sf=+/R=N)
And in a similar manner:
P(Sm=+ /R=P) = ½ + θ (q-½ ) = P(Sm=- /R=N)
P(Sm=- /R=P) = ½ + θ (½ -q) = P(Sm=+ /R=N)
Thus:
P (R=P/Sf=+, Sm=-) =
)]
2
1
(
2
1
)].[
2
1
(
2
1
[()]
2
1
(
2
1
)].[
2
1
(
2
1
[
)]
2
1
(
2
1
)].[
2
1
(
2
1
[
−+−++−+−+
−+−+
qpqp
qp
θθθθ
θθ
(2.2.2)
Q.E.D.
D. Full report of equation (4.2.3)
Θ
B
(+,-)=
θ
ˆ B
(Sf=+, Sm=-, R=P) P(R=P/ Sf=+, Sm=-)
+
θ
ˆ B
(Sf=+, Sm=-, R=N) P(R=N/ Sf=+, Sm=-) (4.2.3)
Given equations: (3.1.1) (4.2.1) (4.2.2), then (4.2.3):
Θ
B
(+,-)=
)1)(1()1()1(2)1(2
)1(2
)]
2
1
(
2
1
)].[
2
1
(
2
1
[()]
2
1
(
2
1
)].[
2
1
(
2
1
[
)]
2
1
(
2
1
)].[
2
1
(
2
1
[
ABABAB
AB
ABAB
AB
qp
p
qpqp
qp
θθθθθθ
θθ
θθθθ
θθ
−−+−−+−
−
⋅
−+−++−+−+
−+−+
( )
)1)(1()1(2)1)(1(2
)1)(1(2
)]
2
1
(
2
1
)].[
2
1
(
2
1
[()]
2
1
(
2
1
)].[
2
1
(
2
1
[
)]
2
1
(
2
1
)].[
2
1
(
2
1
[
1
ABABAB
AB
ABAB
AB
qp
p
qpqp
qp
θθθθθθ
θθ
θθθθ
θθ
−−+−+−−
−−
⋅
−+−++−+−+
−+−+
−+