American Economic Journal: Macroeconomics 2012, 4(1): 1–21

http://dx.doi.org/10.1257/mac.4.1.1

Contagious Adverse Selection†

By Stephen Morris and Hyun Song Shin*

We illustrate the corrosive effect of even small amounts of adverse

selection in an asset market and show how it can lead to the total
breakdown of trade. The problem is the failure of “market confi-
dence,” defined as approximate common knowledge of an upper
bound on expected losses. Small probability events can unravel mar-
ket confidence. We discuss the role of contagious adverse selection
and the problem of “toxic assets” in the recent financial crisis. (JEL
D82, G01, G12, G14)

S ome participants in financial markets have either private information or better

expertise in evaluating new financial instruments and markets. This gives rise to
adverse selection. Financial markets may nonetheless operate smoothly when there
are sufficient gains from trade among traders without private information or exper-
tise. In this case, expert traders are able to extract some rent from their expertise,
but there are enough gains from trade to fund these rents and yet allow mutually
beneficial trade.
However, as illustrated vividly by the recent financial crisis, financial markets
appear to exhibit fragility where shocks to the economy can suddenly lead to abrupt
breakdowns in markets. One element of such market breakdowns to be explored
in our paper is the amplifying effect of strategic complementarities in participa-
tion decisions and the inefficient coordination among ordinary uninformed traders’
participation decisions. If I am expecting other traders to run, then I expect that my
trade will be with an expert trader who will take advantage of me. As an ordinary
uninformed trader, my incentive to participate in the market is greater when other
ordinary uninformed traders participate more.
An important insight from game theory is that in coordination problems, what
matters is not so much what each agent knows about the returns to alternative
actions, but rather what is common knowledge (or approximate common knowl-
edge) between them. It is their shared understanding that matters. This insight fits
well with much commentary on financial markets suggesting that “market confi-
dence” is critical. We can give an interpretation of “market confidence” as embody-
ing such shared understanding. It is not enough that each market participant believes

* Morris: Princeton University, Department of Economics, Princeton, NJ 08544 (e-mail: smorris@princeton.

edu); Shin: Princeton University, Department of Economics, Princeton, NJ 08544 (e-mail: [email protected]).
We are grateful to Christian Hellwig for his discussion of this paper at the University of Michigan November 2009
conference on “The Financial Crisis: Lessons for Economic Theory.” We are also grateful for financial support from
National Science Foundation grant #SES 064 8806 and for comments from two referees, Markus Brunnermeier, and
other participants at the Princeton department-wide seminar in October 2008.
†
To comment on this article in the online discussion forum, or to view additional materials, visit the article page
at http://dx.doi.org/10.1257/mac.4.1.1.

1
2 American Economic Journal: MAcroeconomicsjanuary 2012

that the fundamentals are sound. Market confidence requires that the fundamental
soundness is commonly understood among market participants.
Our purpose in this paper is to develop a link between adverse selection in trading
environments with game theoretic insights about coordination and common knowl-
edge. Shortly, we will describe how our approach relates to the (large) literature on
trade under adverse selection. Our main contribution is to highlight the importance
of small departures from common knowledge, and what kinds of departures matter
for the breakdown of trade.
We consider the following classic trading problem. It is common knowledge
among two groups of traders—potential buyers and potential sellers—that an asset
is worth 2c more to potential buyers than to potential sellers at every state of the
world, where c > 0 is a known constant. Thus, it is common knowledge that the
ex post gains from trade are 2c. Ex ante welfare is thus equal to 2c times the prob-
ability of trade. From a welfare perspective, the traders should always trade with
each other.
However, suppose that a trader attaches a small probability δ to the possibility
that his partner knows something that creates a large benefit M to the partner at his
expense. The “loss ratio” is the ratio of the expected losses of an uninformed agent
and his known gains from a split the difference trade. In the benchmark case where
there is common knowledge about the expected loss ratio, trade takes place if the
loss ratio is less than one.
However, the more intriguing case is when there is incomplete information about
the loss ratio. In other words, each agent is unsure exactly what his partner’s per-
ception of expected losses are. In such circumstances, the adverse selection can be
shown to have a much more corrosive effect where the fear of asymmetric infor-
mation reverberates throughout the information structure and gets amplified in the
process. It is possible that even when the ex ante probability of adverse selection is
small, there can be a catastrophic breakdown in trade. Essentially, the incomplete
information leads to an unraveling result in a coordination game among differen-
tially informed traders. Each uninformed trader would like to trade if the trading
partner is also an uninformed trader. Otherwise, the trading partner is likely to be an
informed party who will take advantage of the uninformed trader.
Drawing on the insights from the earlier literature on common knowledge, we
can characterize the threshold condition for the sustainability of trade. At the core
of our construction is the self-referential nature of “market confidence.” Loosely,
market confidence rests on approximate common knowledge of mutually beneficial
trade. The exact threshold depends on the loss ratio faced by the traders. The higher
the loss ratio, the more rigorous the notion of common knowledge that will sustain
trade must be. Stated more precisely, a trader has market confidence if he expects the
proportion of regular traders who (themselves) have market confidence exceeds his
loss ratio. Notice how this definition is self-referential, and thus, implicitly, incor-
porates a notion of approximate common knowledge. We show that only traders
with market confidence participate in the market. To the extent that approximate
common knowledge can be sensitive to the interaction of the payoff fundamentals
with the information structure, it is possible that even small changes in the underly-
ing parameters of the problem can lead to abrupt breakdowns of market confidence,
Vol. 4 No. 1 morris and shin: Contagious Adverse Selection 3

and, hence, of trade. Exploring the subtleties of how market confidence depends
on the parameters gives us a great deal of insight into the underlying economics of
trade under adverse selection. In the penultimate section of our paper, we draw on
the insights from our framework to revisit the breakdown in the market for securities
based on subprime mortgages in the United States.
The outline of our paper is as follows. Before presenting our formal framework,
we begin with a brief review of the literature on trade with adverse selection, with
an emphasis on how our results can be related to the insights gained from the exist-
ing literature. The formal framework is then presented in several stages. We start by
stating the fundamentals of our trading environment and posit a trading institution
where trade takes place if and only if both traders say “yes” to a proposed trade at a
price that splits the difference. We then introduce adverse selection, and introduce
the idea that the severity of the adverse selection can be a subject of incomplete
information among the traders.
The core of our paper is the characterization of market confidence in terms of
approximate common belief and the demonstration that our notion of market con-
fidence is the right one when considering the occurrence of trade in equilibrium.
Having introduced our key concepts in the initially stark setting, we follow up by
showing that the insights from the simplified setting can be embedded in more gen-
eral settings, and that the intuitions from the common knowledge literature can help
our understanding in these more general settings. As we have flagged already, we
conclude with a brief discussion of how our results can help to shed light on aspects
of the subprime crisis. Before we embark on the main body of our paper, we begin
with a brief survey of how our discussion links with the existing literature on trade
with adverse selection.

Related Literature.—In Akerlof (1970) and the classical adverse selection mod-
els that followed, there is market unraveling with equally informed traders on both
sides of the market. We are concerned with situations where, on both sides of the
market, some traders are informed and some are not informed. This then translates
into a coordination problem among uninformed traders. This coordination problem
among uninformed traders plays an explicit or implicit role in a wide variety of
finance models.
Bhattacharya and Spiegel (1991) consider a competitive model of a market for
a risky asset when there are gains from trade from risk sharing among a pool of
uninformed traders, but there is a single informed trader. The price does not fully
reveal the informed trader’s information because of idiosyncratic motives for trade.
Bhattacharya and Spiegel (1991) identify conditions under which there is a market
breakdown (i.e., no trade). While the analysis is competitive rather than explicitly
strategic, there are strategic complementarities in the sense that lower participa-
tion of uninformed traders in the market reduces the incentive of other uninformed
traders to participate. This framework has been used to address questions, such
as when new securities can shut down markets (Bhattacharya, Reny, and Spiegel
1995) and when public disclosure rules can mitigate the adverse selection prob-
lems (Spiegel and Subrahmanyam 2000). Pagano (1989) and Dow (2004) high-
light the coordination problem among uninformed t­raders and draw implications
4 American Economic Journal: MAcroeconomicsjanuary 2012

for ­designing financial institutions. In these models, the coordination problem is

among identical uninformed traders, so there is no role for considering what is or is
not common knowledge among them. A distinctive feature of our analysis is that we
examine the coordination among traders who, while all uninformed, have different
beliefs about the amount of adverse selection they face.
A large literature in game theory examines the importance of what is or is not
common knowledge or approximate common knowledge in coordination games.
Rubinstein (1989), Monderer and Samet (1989) and Carlsson and van Damme
(1993) are key early contributions. The insight from this literature is that coordina-
tion—i.e., taking an action which is only optimal if others do so—requires approxi-
mate common knowledge, in a sense to be made precise below. The methodological
contribution of this paper is providing a tractable framework in which adverse selec-
tion generates a coordination problem among uninformed ­traders and then showing
how the relevant common knowledge requirement for coordination can be expressed
in terms of the underlying adverse selection problem.
We consider an environment in which it is common knowledge that an object
is worth more to buyers than to sellers, but there is a lack of common knowledge
about a common value component. This environment is a classic one in the mecha-
nism design literature, with Myerson (1985) being an early reference on optimal
mechanisms in this setting. We analyze what happens in simple but realistic trad-
ing mechanisms but with arbitrary beliefs and higher order beliefs of traders. Dang
(2008, 2009) also examines what happens in this environment in simple trading
mechanisms (ultimatum offer bargaining and double auction, respectively) with a
simple information structure but endogenous information acquisition that gives rise
to an endogenous lemons problem.
A number of papers have examined channels by which adverse selection may cause
a market freeze. Adverse selection interacts with funding constraints in Kiyotaki and
Moore (1997). In Moore (2010), adverse selection breaks down markets in financial
assets being re-traded through time, with shocks spreading contagiously through time
instead of (as in this paper) across information sets in static setting. Kurlat (2009)
shows how learning can exacerbate endogenous adverse selection in such settings.
In Bolton, Santos, and Scheinkman (2009, 2011), decisions of firms whether to sell
early or late in response to a liquidity shock create coordination problems because of
endogenous adverse selection. Kirabaeva (2010) introduces adverse selection into a
Diamond and Dybvig (1983) banking model.
The recent papers of Pagano and Volpin (2008) and Dang, Gorton, and Holmstrom
(2009) emphasize the role of information in preventing trade. These papers high-
light the payoff structure of debt as a financial claim that minimizes the information
sensitivity of asset returns. In such a setting, an aggregate shock may have a dispro-
portionate impact because it increases the information sensitivity of asset returns
and, thus, triggers adverse selection. This paper shares with those two papers the
perspective that the arrival of information (not lack of “transparency”) may trig-
ger market collapse. However, our focus is on modeling the commonality of infor-
mation in an abstract trading model, while Dang, Gorton, and Holmstrom (2009)
endogenize the choice of financial instrument and have endogenous information
acquisition in a simpler informational environment.
Vol. 4 No. 1 morris and shin: Contagious Adverse Selection 5

Table 1—Value of the Asset to Agents

State Probability Value to sellers Value to buyers

_ _
Lemon δ ​ − M − c
v​ ​ − M + c
v​
_ _
Normal 1 − 2δ ​ −c
v​ ​ +c
v​
_ _
Peach δ ​ +M−c
v​ ​ + M + c
v​

The notion of market confidence as approximate common knowledge ties in with

a broader set of arguments on the importance of institutions that ensure common
understanding and common knowledge of the important fundamentals. We argued
elsewhere (Morris and Shin 2007a) that there are important trade-offs between pro-
viding accuracy (individually correct beliefs) and commonality (approximate com-
mon knowledge of beliefs) for many problems in economics. Holmstrom (2009) has
argued that this trade-off is particularly important in thinking about the regulatory
reforms on transparency and the coarse nature of credit ratings. The coarse nature
of credit ratings may have a rationale in terms of promoting common understand-
ing at the expense of a finer grid for the fundamentals. In Morris and Shin (2007a),
we argue that coarse accounting standards can be seen as an institution that could
potentially provide the commonality.
Our objective in this paper is to highlight the importance of commonality of infor-
mation, and how the lack of such commonality can lead to suboptimal outcomes. In
order to emphasize our key theme, the model is deliberately stark, but we believe
that our results shed much light on the broader problem of common knowledge and
trade, as well as providing a spur for a more systematic investigation of the role of
adverse selection in exacerbating the current financial crisis.

I. The Setting

There are N potential buyers and N potential sellers of an asset. Sellers each own
one unit of the asset with private value v − c, while buyers’ valuations are v + c. It
would be efficient for all sellers to transfer their object to the buyers, who each have
unit demand, say, at a price of v, which splits the gains from trade.
However, there is uncertainty about the common value component v. Its expected
_
​ , and it is equal to its expectation in what we’ll refer to as the normal state
value is v​
with probability 1 − 2δ. However, with probability δ, the asset is a “peach” and
_ _
v = ​v​ + M, and, with probability δ, the asset is a “lemon” with v = ​v​ − M. Thus,
we have the distribution of the value of the asset to the agents as given by Table 1.
There is adverse selection, with proportion q of both buyers and sellers informed
of the true state, and proportion 1 − q uninformed. We will implicitly assume
throughout our analysis that qN, the number of informed agents, is an integer.
We have in mind the archetypal example where the asset in question is an asset-
backed security backed by subprime mortgages, but where the quality of the mort-
gage pool depends sensitively on the region and date of their origination. The two
groups of traders are equally well-informed most of the time, but we allow the pos-
sibility that one or other of the traders is better informed than his trading partner,
and that the information could be positive or negative. The highly skewed payoffs
6 American Economic Journal: MAcroeconomicsjanuary 2012

associated with subprime CDOs motivates payoffs in the trading game, where for
most of the time there are (small) gains from trade, but for a few states of the world,
there are large payoff consequences of trade. See Coval, Jurek, and Stafford (2009)
for an introduction to the economics of structured finance.

A. Loss Ratio

Throughout the paper we will fix the gains from trade c as a constant. A conve-
nient feature of our basic setup is that ex ante welfare (the sum of the agents’ ex ante
expected utility from trade) is simply 2c times the probability of trade:

ex ante welfare = 2c × probability of trade.

Buyers and sellers are randomly matched with each other. Each buyer and each
_
seller are asked if they would like to trade at price ​v​. If both say yes, they trade. If
either says no, they do not. Clearly, informed buyers will buy if and only if the asset is
normal or a peach, while informed sellers will sell if and only if the asset is normal or
a lemon. We first consider the case where there is common knowledge of the structure
of the model and ask if there is an equilibrium where uninformed agents participate in
the market and trade.
If faced with an uninformed agent on the other side of the market (a probability
1 − q event), an uninformed trader has an expected gain from the trade of c. Thus,
we have the expression (1 − q) c for an agent’s expected gains from trade with an
uninformed agent.
If faced with an informed agent on the other side of the market (a probability q
event), an uninformed agent still hopes to get a gain from the trade of c; but with prob-
ability δ, the informed agent will not participate in trade, and the agent will lose M + c.
Thus, expected losses from trade with an informed agent are q(δ(M + c) − c). An
important parameter for us will be the agents’ loss ratio defined as

ψ = __
expected losses
​     ​
  
expected gains
q(δ(M + c) − c)
= __
  
​      ​
(1 − q) c

≈ _
qδM
​  c ​,

where the approximation is good when q is small (i.e., the incidence of informed
traders is low) and δM is much larger than c—that is, when the possible loss to hold-
ing a defective security is much larger than the possible underlying gains from trade.
Arguably, these are precisely the attributes that were involved in the most toxic of
the mortgage-backed securities, such as CDOs and CDO-squareds.
The loss ratio ψ plays a pivotal role in our paper. Consider the decision problems
of the informed and uninformed traders. Informed traders have a dominant strategy.
Informed buyers will offer to buy if and only if the asset is normal or a peach, while
informed sellers will offer to sell if and only if the asset is normal or a lemon.
Vol. 4 No. 1 morris and shin: Contagious Adverse Selection 7

Table 2—Gains from Trade

Outcome Probability Payoff

Lemon δ (q + (1 − q) p) (− M + c)
Normal 1 − 2δ (q + (1 − q) p) c
Peach δ (1 − q) p(M + c)

Consider the reasoning of an uninformed buyer. For uninformed buyers, their

optimal strategy will be determined by their beliefs about the outcome and the pro-
portion of uninformed agents who trade (denoted by p). Gains from trade condi-
tional on the outcome are depicted in Table 2.
For the buyer, the unfavorable outcome is when the good is a lemon, which hap-
pens with probability δ. Trade happens with the insider with probability q and with
an uninformed trader with probability (1 − q) p. Conditional on trade taking place,
the payoff is − M + c. Hence, the expected payoff for the buyer when the good is
a lemon is given by the top right-hand cell of Table 2. The other cells in the table
are derived through analogous reasoning. For the uninformed trader, the ex ante
expected payoff when offering to trade is then given by

(1) δ(q + (1 − q) p)(− M + c) + (1 − 2δ)(q + (1 − q) p) c

+ δ(1 − q) p(M + c)

= (1 − q) cp − δq(M + c) + qc

( q(δ(M + c) − c)
= (1 − q) c p − ​ __
     
(1 − q) c
 ​ )
(2) = (1 − q) c(p − ψ) .

From (2), we see the important role played by the loss ratio ψ. The expected gain
from trade is positive or negative depending on whether the loss ratio is smaller or
larger than the probability p that the uninformed traders participate in the market. If
the loss ratio ψ is less than or equal to one, and if other uninformed traders offer to
trade (so p equals 1), then (2) is positive. Hence, the best reply is to offer to trade,
meaning that everyone offering to trade is sustainable as an equilibrium. However, if
the loss ratio ψ is strictly greater than one, the payoff to trading is guaranteed to be
negative irrespective of the actions of others. Hence, the only equilibrium is the no
trade equilibrium where all uninformed traders refuse to trade. We summarize our
finding with the following preliminary result.

Lemma 1: Suppose ψ is common knowledge and ψ ≤ 1. Then there is an equi-

librium where trade always takes place. If ψ is common knowledge and ψ > 1, the
only equilibrium is the no trade equilibrium.
8 American Economic Journal: MAcroeconomicsjanuary 2012

In this sense, when ψ is common knowledge, the possibility of trade hinges on

whether the loss ratio is greater than or less than one. However, the introduction of
incomplete information can drastically reduce the possibility of trade. Exactly how
incomplete information reduces the possibility of trade is the topic of our analysis.
We will start by reporting an example showing how lack of common knowledge of
the loss ratio can have a large impact and prevent trade even when ψ is much below
1. We will then explain the logic of this sensitivity in words before reporting formal
analysis in the next section.

B. Incomplete Information

We turn to the case where ψ is not common knowledge. Suppose, instead, that
M is a random variable with possible support on (c, ∞), and each informed and
uninformed agent has different information about M, and hence ψ. We begin with
an example in the spirit of Rubinstein’s (1989) e-mail game. Suppose there is a set
of possible states

Ω = {1, 2, … , 2K + 1}.

There is a uniform prior over Ω. The loss ratio ψ now depends on the realized state
in Ω, and takes two values: high or low. The loss ratio at state k is denoted by ​ψ​k​,
and is given by

​ψ​k​ =  ​​ψH​ ​​ ​

if k = 1,​
​​ψL​​ if
​ k > 1​
{
where 1/2 < ​ψL​​ < 1 and ​ψH​ ​ > 3∕2. The uninformed sellers all have information about
the state described by the following partition:

{{1}, {2, 3}, … , {2K − 2, 2K − 1}, {2K, 2K + 1}},

while the uninformed buyers’ information partition is

{{1, 2}, {3, 4}, … , {2K − 1, 2K}, {2K + 1}} .

In this way, all traders have very good information on the fundamental value of the
loss ratio ψ. Indeed, all types—except buyers with information set {1, 2}—know the
true value of ψ perfectly. Moreover, for all states other than state 1, and thus with
high ex ante probability, the loss ratio ψ is strictly below 1, which is the precondition
for mutually beneficial trade. However, the incomplete information drastically cur-
tails the possibility of trade so that the only equilibrium is the no trade equilibrium,
as we will now show.
Vol. 4 No. 1 morris and shin: Contagious Adverse Selection 9

For an uninformed seller with information {1}, the dominant action is to refuse
to trade, since the loss ratio is known to be greater than 1. For an uninformed buyer
with information {1, 2}, the loss ratio is uncertain, but the average loss ratio is

​ _
1 ​ ​ψ​​ + ​ _
1 ​ ​ψ​ ​ > 1,
2 L 2 H
so that it is optimal not to participate (even if he expects uninformed sellers to
participate).
Now consider the seller with information {2, 3}. This trader knows that the loss
​ L​​ < 1. From (2), the expected payoff to participating in trade is
ratio is ψ

(1 − q) c(p − ​ψL​​),

where p is the probability of meeting an uninformed buyer. However, the buyer

with information {1, 2} is not participating in the market, and so the probability of
meeting an uninformed buyer is 0.5 if the buyers with information {3, 4} are in the
market, and is zero if these buyers are not in the market. In any case, the probability
of meeting an uninformed buyer is at most 0.5. This means that ψ ​ ​L​ > p for sure, so
that the gain from trade is negative. Thus, any seller with information {2, 3} refuses
to participate in the market.
With all {2, 3} sellers out of the market, the {3, 4} buyers refuse to participate.
With the {3, 4} buyers out of the market, the {4, 5} sellers refuse to participate, and
so on. By iteration, all buyers and all sellers refuse to participate. Hence, the only
equilibrium is the no-trade equilibrium.
The example shows that even if the loss ratio is strictly below 1 for almost all
states, a small “contamination” of a high loss ratio in one state can undermine mar-
ket confidence through contagious adverse selection where the uninformed traders
pull out of the market, so that the only equilibrium outcome is for there to be no
trade everywhere.
This is a result that echoes the result in Rubinstein’s (1989) e-mail game.
Morris, Rob, and Shin (1995) discuss the underlying game-theoretic structure of
Rubinstein’s e-mail game, and how the possibility of coordination depends on the
interaction of the payoff function and degree of common knowledge derived from
the individual information structures.
What is the general logic of this example? If there is uncertainty about M (and
thus ψ) a trader will participate in the market if and only if his expectation of the
proportion of others participating (given by p) exceeds his expected loss ratio. If
we assume symmetry between sellers and buyers in their beliefs and higher order
beliefs, a trader will participate in the market if and only if all the following state-
ments hold:

1) His expectation of the loss ratio is less than 1.

2) (1) is true, and his expectation of the proportion of agents on the other side
of the market for whom (1) is true is greater than his expectation of the loss
ratio.
10 American Economic Journal: MAcroeconomicsjanuary 2012

3) (2) is true, and his expectation of the proportion of agents on the other side
of the market for whom (2) is true is greater than his expectation of the loss
ratio.

4) And so on...

While stated informally, this turns out to be an exact characterization of when

trade is possible in equilibrium. To see this, let us first give an informal argument on
the necessity and sufficiency of the above list of statements for a trader to participate
in the market. We will follow in the next section with a more formal argument asso-
ciated with the notion of market confidence.
First, consider the “only if  ” part of the proof. We need to show that if any one of
the above statements is false, then the trader will not participate. If (1) is false, then the
decision to participate is dominated by the decision not to participate. If (2) is false,
then the decision to participate is not the best reply to any strategy profile of other
traders that is undominated. In other words, if (2) is false, the decision to participate
is knocked out at the second round of the iterated deletion of strictly dominated strate-
gies. We can then proceed by induction. If (3) is false, then the decision to participate
is eliminated in the third round of the iterated deletion of strictly dominated strategies.
In this way, if any one of the above list of statements is false, then the decision to par-
ticipate does not survive iterated dominance, and, hence, is not part of any equilibrium.
In the example above, we see the iterated dominance argument in action. The
seller with information {1} and buyer with information {1, 2} refuse to trade as a
dominant action. Thus, for the seller with information {2, 3}, participating in the
market is eliminated in the second round of deletion in the iterative deletion of
dominated strategies. In this way, for any trader, one of the statements in the above
list turns out to be false.
The “if  ” part of the proof rests on a convergence argument, where the list of state-
ments above eventually picks out a set of traders for whom all the statements are
true, and the decision to participate in the market is a best reply to all others in the
set participating. As we will prove formally in the next section, the infinite list of
statements above is equivalent to the following self-referential “fixed point,” state-
ment. Say that an agent has market confidence if and only if his expectation of the
proportion of agents with market confidence is greater than his expected loss ratio.
An agent trades if and only if he has market confidence.

II. Market Confidence

We will now formalize the self-referential idea of having “confidence in the mar-
ket”—namely, that an agent has confidence in the market if his expectation of the
proportion of traders with confidence in the market is greater than his expected loss
ratio. We do this in two steps. First, we introduce a formal language to talk about
an agent’s beliefs, an agent’s expectation of proportions of other agents holding
certain beliefs, and so on, reporting some important properties of such higher order
expectations of proportions. Second, we show how the properties of beliefs we have
introduced characterize equilibria in the trading game.
Vol. 4 No. 1 morris and shin: Contagious Adverse Selection 11

Suppose that we have a collection of N agents. Each agent has a set of possible
types T ​ i​​is a measurable set. We write ​πi​​: ​Ti​​  → Δ(​T−i
​ ​i​  , where T ​ ​) for agent i’s beliefs
about other agents’ types, where Δ(​T− ​ i​) is the notation for the set of all probabilities
distributions for types of all agents other than i.
We will be interested in rectangular events on the type space. A rectangular event
E is the n-tuple

E = (​E1​​, ​E2​​, ⋯, ​EN​ ​),

where each ​ E​i​  ⊆ ​Ti​​. Now let ψ = (​ψ1​​, ​ψ2​ ​,⋯, ​ψN​ ​), where ​ ψ​i​: ​T​i​  → ℝ is the
mapping that associates each type with a number for that type. The n-tuple ψ is
the profile of such functions—one for each agent. Later on we will associate these
functions with variable loss ratios, but for now they will simply be state dependent
threshold beliefs that we want to introduce a language to reason about.
For a given rectangular event E, say that agent i “​ψ​i​  -believes E” if his expectation
of the proportion of agents, whose type t​j​​ belongs in E ​ ​j​, is at least as large as ψ ​ i​​. We
use the notation ​B​  ​ψi​  i​ ​​  (E) to denote the set of i’s types that ​ψ​i​  -believe E. In other words,

​B​  ​ψi​  i​ ​​   (E) = { |∫

 ​t​i​  ∈ Ti  ​ ​ ​ ​  _
​t​−i​ N
1  ​​  # {  j | ​t​​  ∈ ​E​ ​) dπ (​t​​) ≧ ​
j j i ψi​​  . }
Also write

​B​  ψ​(E) = (​B​  ​ψ1​  1​ ​​(E ), ​B​  ​ψ2​  2​ ​​(E ),⋯, ​B​  ​ψNN​​  ​​(E )).

Thus, if the n-tuple of types t belong in ​B​  ψ​(E), then every agent ​ψ​i​-believes E. For
any rectangular event E, we see that B​ ​  ψ​(E ) is also a rectangular event, since B ​ ​  ​ψi​  i​ ​​(E )
is a subset of agent i’s types. Also, note that the ψ -belief operator B ​ ​  ​( ⋅ ) has the
ψ

closure property under logical implication—namely

(3) E ⊆ E′ ⇒ ​B​  ψ​(E ) ⊆ ​B​  ψ​(E′ ).

In other words, if a profile of types believe E and E implies E′, then they believe E′.
The iterated event ​B​  ψ​(B
​ ​  ψ​(E )) is defined as

​ ​  ψ​(E )) = (​B​  ​ψ1​  1​ ​​  (​B​  ψ​(E  )), ​B​  ​ψ2​  2​ ​​  (​B​  ψ​(E )), ⋯, ​B​  ​ψNN​​  ​​(​B​  ψ​(E ))).
​B​  ψ​(B

We use the notation ​(​Bψ​ ​)​  ​(E ) to denote the n-fold iteration of ​B​  ψ​( ⋅ ) on the event E.
n

​ ​  ψ​(E ) as
Define C

(4) ​C​  ψ​(E ) = ​B​  ψ​(E ) ∩ ​B​  ψ​(​B​  ψ​(E )) ∩ ​B​  ψ​(​(​Bψ​ ​)​  2​(E  )) ∩ ⋯

Say that there is common ψ-belief of rectangular event E among the profile of agent
types t if

(5) t ∈ ​C​  ψ​(E ).

12 American Economic Journal: MAcroeconomicsjanuary 2012

Say that rectangular event E is ψ-evident if

(6) E = ​C​  ψ​(E).

In other words, an event E is ψ-evident when it is a fixed point of the common

​ ​  ψ​(⋅). The definition of common ψ-belief and ψ-evident events are
ψ-belief operator C
extensions of the definitions introduced by Monderer and Samet (1989), who in turn
generalized Aumann’s (1976) notion of common knowledge. We have the following
characterization of common ψ-belief.

Proposition 2: Rectangular event E is common ψ-belief at t if and only if there

exists a ψ-evident event E′ such that t ∈ E′ ⊆ E.

PROOF:
For the “if  ” direction, note that since E′ is ψ-evident, we have E′ ⊆ ​B​  ψ​(E′ )
⊆ ​B​  ψ(​ ​B​  ψ​(E′  )) ⊆ ⋯. From (3), we then have E′ ⊆ E ⊆ ​B​  ψ​(E) ⊆ ​B​  ψ​(B ​ ​  ψ​(E ))
⊆ ⋯. Hence, E is common ψ -belief at t. For the “only if” direction, if E is com-
mon ψ-belief at t, then ​C​  ψ​(E ) = ​B​  ψ(​ C​ ​  ψ​(E )), so that ​C​  ψ​(E ) is ψ -evident.
The equivalence of the fixed-point definition and the iterative definition of com-
mon ψ-belief is in the spirit of the characterization of common knowledge by
Aumann (1976), of common p-belief by Monderer and Samet (1989), and of state
dependent p-belief in Morris and Shin (2007b). However, there is an added ingredi-
ent to our definition of common ψ-belief, which links with our earlier discussion of
the possibility of trade.
There is tight connection between the notion of common ψ-belief and equilib-
rium in the trading game. To make the connection, suppose that there are N buyers
and N sellers in the trading game that we discussed earlier. Each buyer may have dif-
ferent beliefs. But make a simplifying symmetry assumption. For each buyer, there
is a seller who has identical beliefs about the loss ratio. Now note that rectangular
events can be interpreted as strategy profiles in a binary action game. For rectan-
gular event E = (​E1​​, ​E2​​,⋯, ​EN​ ​), we interpret type ​ti​​∈ Ei as a type of trader i who
participates in the market. Thus, we may interpret E as a full list of all types of all
traders who participate in the market. Now, consider the belief operator ​B​  ψ​(⋅), where
ψ = (​ψ​1​, ​ψ2​ ​, ⋯ , ​ψN​ ​) is the profile of the expected loss ratios of the traders. Then
the rectangular event ​B​  ψ​(E ) has the following interpretation. When all types in E
participate in the market, then ​B​  ψ​(E ) consists of all types ​ti​​of all traders i who put
probability at least ​ψi​​ to meeting an uninformed trader who is participating in the
market. Hence, ​B​  ψ​(E ) consists of all types ​ti​​ of all traders i whose best reply is to
participate in the market when all types in E participate in the market. We thus have:

Proposition 3: There is an equilibrium in which all types in E participate in the

market if and only if E is ψ-evident.

PROOF:
The proof follows straightforwardly from the fact that E being ψ-evident means E
= ​B​  ψ​(E), so that strategy profile E is a fixed point of the best reply mapping ​B​  ψ​( ⋅ ).
Vol. 4 No. 1 morris and shin: Contagious Adverse Selection 13

Propositions 2 and 3 show the usefulness of our definition of common belief. The
belief operator B​ ​  ψ​( ⋅ ) has two faces. On the one hand, it has the interpretation as a
belief operator. However, it also has the interpretation as the best reply mapping. It
is this dual face of the belief operator that allows us to define the notion of market
confidence that links beliefs directly with equilibrium of the trading game.
Say that trader i has confidence in the market if

t​i​  ∈ ​B​  ​ψi​  i​​​  (T ) ∩ ​B​  ​ψi​  i​​​  (​B​  ψ​  (T )) ∩ ​B​  ​ψi​  i​​​  ((​B​  ψ​ ​)2​​ (T )) ∩ ⋯,
(7) ​

where T is the set of all types. Finally, say that there is market confidence if all trad-
ers have confidence in the market. In other words, there is market confidence if

(8) t ∈ ​C​  ψ​  (T  ).

When there is market confidence, the following list of statements are all true:

1) The loss ratios of all traders is less than 1 (since ​B​  ψ​(T  ) is nonempty).

2) For every agent i, his expectation of the proportion of other agents whose loss
ratios are less than 1 is at least ​ψ​i​.

3) For every agent i, his expectation of the proportion of agents for whom (1)
and (2) are true is at least ψ
​ ​i​.

4) And so on...

Since T is the set of all types, it has the interpretation of the strategy profile where
every uninformed trader participates in the market. Hence, ​C​  ψ​(T  ) has the interpreta-
tion of the “largest” equilibrium in the trading game which maximizes market par-
ticipation of all traders. Therefore, we have the following succint characterization of
equilibrium with trade.

Proposition 4: There is an equilibrium in which type profile t participates in the

market if and only if there is market confidence at t.

This proposition links the iterative definition of common belief (which is also
the iteration of the best reply mapping) with the self-referential, fixed point defi-
nition of market confidence. It enables us to diagnose the failure of the traders to
exploit opportunities for mutually beneficial trade in terms of a failure of market
confidence. The mere fact that mutually beneficial trade exists is not enough for
trade to take place. Nor is it enough for all traders to believe that mutually benefi-
cial trade is possible. Instead, we need a more stringent condition that ensures that
there is sufficient common understanding of the existence of mutually beneficial
trade. It is this common understanding that is captured by our definition of market
confidence.
14 American Economic Journal: MAcroeconomicsjanuary 2012

A Global Games Example.—We can illustrate our notion of market confidence,

why it characterizes trade and why it is surprisingly hard to achieve, with a global
game example (Carlsson and van Damme 1993; Morris and Shin 1998, 2003). We
emphasize that Proposition 4 holds more generally than the global games informa-
tion structure, but use this example to highlight why market confidence may be hard
to achieve.
Suppose that a variable θ is distributed on the real line. There are N buyers and N
sellers. Buyer i and seller i both observe a signal t​ i​​ = θ + ​εi​​, where ​εi​​is independently
distributed normally with mean 0 and standard deviation σ. Thus, as above, we assume
that for every buyer, there is a seller with the same signal about θ: this simplifies the
exposition and, because we will focus on the case of large N, does not effect the quali-
tative conclusions. Also for simplicity, we will assume that θ is distributed uniformly
on the real line, but the argument will go through for any smooth prior if σ is small
enough. Suppose that the loss ratio takes a high value ψ ​ ​H​ > 3∕2 if θ is less than 0, but
takes a low value ​ψ​L​∈ (1/2, 1)if θ is greater than 0. Each trader i’s expected loss ratio is
close to ​ψL​​for positive ​ti​​and close to ​ψH​ ​for negative ​ti​​. Specifically,

​ψ​i​(​t​i​) = Pr
​  ​(θ < 0 | ​ti​​) ​ψH​ ​ + ​ 
Pr​(θ > 0 | ​ti​​) ​ψL​ ​

= (1 − Φ(​_σ​t​​ ​)) ​ψ​ ​ + Φ(​_σ​t​​ ​) ​ψ​ ​
i
H
i
L

≥ ψ
​ ​L​.

This example now fits the framework introduced at the beginning of this section.
In particular, each trader i will have a multinomial normal distribution about the
signals of other N − 1 traders.
We will argue that no matter how accurate the traders’ signals are (i.e., no matter
how small σ is as long as it is strictly positive), and no matter how large the state
θ, and thus how confident traders are that the loss ratio is less than 1, there is never
nontrivial common ψ-belief, as defined above, for sufficiently large N, and thus
there is never market confidence, and thus (by Proposition 4) there is no trade. To
show that common ψ-belief does not arise, it is enough, by Proposition 2, to show
that no rectangular event is ψ-evident. To see why, consider the event that each
trader observes a signal above some high threshold ​t​  *​, i.e.,

(9) ​ ​1​ × ⋯ × ​

E = E EN​ ​

= {t ∈ ​ℝ​N​  | ​ti​​  ≥ ​t​*​for each i }.

When is this event ψ-evident? A trader observing signal ​ti​​ thinks that θ is
n­ ormally distributed with mean ​ti​​and standard deviation σ, and that any other trader
j’s signal—equal_to t​j​​ = θ + ​εj​​—is normally distributed with mean t​i​​ and stan-
dard deviation ​√2 ​σ. Thus, a trader observing signal ​t​i​ ≥ ​t*​​will assign ­probability
1−Φ _
​t*​​ − ​
(​​ 2 ​σ )ti​​
_ ​ to any other trader observing a signal above t​* ​​. Thus, his expected
√
Vol. 4 No. 1 morris and shin: Contagious Adverse Selection 15

proportion of traders with signals above t​*​​ is

​ _ N − ​ 
1  ​  + ​ _
N N
t*​​ − ​t​i​
1 1 − Φ _ _  ​ 
​ 2 ​σ
√ ( (​​ )).
For this to exceed ​ψi​​(​t​i​), it must exceed ​ψL​​ > 1/2. Thus, we must have

(​​
t​*​ − ​ti​​
1 − Φ _
​ 2 ​σ
√
_
_  ​ > ​ 
)
N ​ψ​L​ − 1
N − 1
 ​,

and so
_
​ ​ _ ​ (1 − ​ψL​​) .
​t​i​ > ​t*​​ − ​√2 ​  σ ​Φ−1
N
N − 1 ((​ ) )
​ ​  ψ​({t ∈ ​ℝ​  N​  | ​ti​​  ≥ ​t*​ ​for each i }) is a subset of the event:
Thus, the event B

{t ∈ ​ℝ​ ​  | ​t​​  ≥ ​t​ ​− ​ 2 ​  σ ​Φ​ ​((​_

N
i
* √
_
−1
 ​ (1 − ​ψ​​)) for each i }.
N − 1)
N
L

Since ​ψ​L​ > 1/2, for sufficiently large N, _ (​

N − 1 ​ (1 − ​
N
)
ψL​​) < 1/2, and thus

_
((​ )
​ ​ _ ​ (1 − ​ψL​​)
− ​√2 ​  σ ​Φ−1
N
N − 1 ) > 0.
Thus, any rectangular event E of the form (9) is not ψ-evident for any ​t*​​, since there
is always some type ​t​i​ > ​t*​​included in the event whose expected proportion of trad-
ers observing signals above ​t*​​is less than ​ψL​​. But this in turn implies that no event
that includes an event of the form (9) is ψ-evident, and thus no nontrivial event is
ψ-evident. There is never market confidence, and thus there is never trade.
The intuition for this result is as follows. The loss ratio never falls below one-
half. While this would be enough to sustain trade if the loss ratio were common
knowledge, with noisy signals, a marginal uninformed trader—who is just willing
to trade—will anticipate that about half the traders on the other side of the market
will have observed lower signals and will not be trading. With half the uninformed
traders not trading and a loss ratio more than one-half, trade is not optimal for that
marginal trader—a contradiction.

III. General Asset Returns and Trading Games

We now turn to how our result applies to more general contexts with general asset
payoff distributions and trading games.
16 American Economic Journal: MAcroeconomicsjanuary 2012

A. Asset Returns

Consider the model as before but assume that the common value component
_
of the asset v = ​v​ + η, where η is smoothly distributed according to symmet-
ric density f​​θ​  ( ⋅ ), so θ parameterizes asset returns. Informed sellers will trade
_ _
only if v − c = ​v​ + η − c < ​v​, i.e., if η < c. Informed buyers will trade only if
_ _
v + c = ​v​ + η + c > ​v​, i.e., if η > − c. Thus, a key parameter will be the prob-
ability under the distribution θ that returns in the tails will be more than c from
the mean

δ(θ) = P​r​θ​(η ≥ c)

∫
∞

= ​ ​ θ​(η) dη
​  ​​f​
η=c

= 1 − ​F​θ​(c).

Another key parameter will be the expected deviation of the common value of the
asset from its mean if returns are in one of the tails:

M(θ) = ​Eθ​​  (η | η ≥ c)

∫
∞

= ​ _1  ​ ​ ​ ​η​  ​fθ​​(η) dη.

δ(θ) η=c

These will be the only parameters of returns that will matter in our trading game. In
particular, for each distribution θ, there is a corresponding loss ratio defined as above:

__ q (δ(θ) (M(θ) + c) − c)
   ​ = ​ __    
expected losses
ψ(θ) = ​ 
       ​  .
expected gains (1 − q) c

Maintaining symmetry between beliefs and higher order beliefs of sellers and buy-
ers, the analysis of Section I goes through exactly as before, where an agent’s beliefs
and higher order beliefs about θ determine his beliefs and higher order beliefs about
his and other agents’ loss ratios.
Our ability to derive an exact characterization of when trade occurs relies on a
couple of features of the model. We assume that traders are informed or uninformed
(there is nothing in between). This is crucial to the analysis, since it means that
adverse selection translates into a pure coordination problem, and agents’ assess-
ment of the loss ratio is not correlated with their assessment of the proportion of
agents on the other side of the market trading. A more realistic modeling would
allow for intermediate types. But as we explained in the introduction, we wanted to
focus on the coordination element of adverse selection.
Vol. 4 No. 1 morris and shin: Contagious Adverse Selection 17

B. Losing Market Confidence

How is market confidence lost? Suppose that initially there was common knowl-
edge or approximate common knowledge of a loss ratio greater than one-half but
less than one. There could be an equilibrium where everyone trades and all ex ante
gains from trade are realized. Now suppose that there is a shock. Perhaps the loss
ratio increases or perhaps it decreases, but crucially there is no longer approximate
common knowledge of the loss ratio. Then a crisis will hit. It is the loss of approxi-
mate common knowledge of the loss ratio (which we label “market confidence”)
and not changes in the loss ratio that generate the crisis.
We already illustrated in Section IIA how the noisy private signals of global
games can formalize the idea that agents have very accurate information without
attaining common knowledge or approximate common knowledge. Here, we gen-
eralize this point—appealing to arguments used in the global games literature—to
show how the critical loss ratio sustaining trade is reduced to one-half from one if
there is noisy private information.
In Section IIA, we discussed the case where the common value component of
the asset’s payoffs was parameterized by θ and the loss ratio corresponding to θ
was written as ψ(θ). Let g (⋅) be a prior probability density on θ and suppose now
that ψ(θ) were decreasing in θ, so higher states θ correspond to a lower loss ratio.
Suppose each agent (buyers and sellers) observes a signal t​i​​ = θ + σ ​ε​i​, where σ is
a parameter measuring the size of noise and ε​ ​i​is a noise term distributed according
to density f (⋅) in the population. This information structure describes a type space
as in Section II, parameterized by σ, where each T ​ i​​ = ℝ:

​t​j​  − θ
​π​  σ __ θ j=1 j (​ )
​∫ ​  ​ ​∏ N ​  ​​ f​ ​​​  _​ g(θ) dθ
σ
i​  ​  (​t​−i​  | ​ti​​) = ​ 
       
​ ​ g(θ) dθ
​ ​  ​ ​f​​​  (​_σ )
t​​  − θ
​
∫
j
θ j

and

​∫θ​  ​  ψ​(θ) ​fj​ ​  _
__
(​t​​​  − θ )
j
σ ​ g(θ) dθ
​ψ​  σi​  ​  (​t​i​) = ​ 
       
​
(​ )
tj​​  − θ
​.
∫ _
​ θ​  ​ ​fj​​​  σ ​ g(θ) dθ

Now let θ​ *​​be the point where the loss ratio is reduced to half. Generalizing the argu-
ment in Section II and adapting arguments in the global games literature (see, e.g.,
Morris and Shin 2003, section 4.2), one can show that, as σ → 0, agent i has mar-
ket confidence (i.e., ​ψ​σ​-belief) if and only if ​ti​​ ≥ ​θ*​​. Thus, with small noisy signals,
there is market confidence only when the loss ratio is half as much as when there is
common knowledge of the loss ratio. Such a reduction in the threshold level of the
loss ratio could result in a drastic curtailment of trading, as we saw in the example
earlier in the spirit of Rubinstein’s (1989) e-mail game.
18 American Economic Journal: MAcroeconomicsjanuary 2012

C. General Trading Mechanisms

We have assumed that agents made a yes or no decision whether to trade at some
suggested price. This seems like a realistic assumption in the setting of over-the-
counter (OTC) markets where mortgage-backed securities are traded. However,
there are arguments that suggest that our results will be robust to generalized trad-
ing mechanisms.
To see this, consider a double auction setting where each trader proposes a price
to trade and trade takes place at the average of the proposed prices only if the sell
price is less than the buy price. Consider the following strategy profile in the double
_
auction. All agents propose trade at price ​v​ if they have market confidence. If they
do not have market confidence, informed sellers propose trade at v + c, informed
_
buyers propose trade at v − c, uninformed buyers propose trade at v​ ​ − M + c, and
_
uninformed sellers propose trade at v​ ​ + M − c.
To verify that this is an equilibrium, observe that an uninformed agent will never
get any surplus if his trading partner does not have market confidence, and he will
maximize his surplus if his opponent has market confidence if he proposes trade
_
​ . Thus, under this strategy profile, he has essentially the same payoffs as
at price v​
in the simple trading game. Now consider a seller with a lemon. His expected gain
_
from proposing price v​ ​ is (c + M) times the probability he attaches to having an
uninformed partner who believes in market confidence. His expected gain from pro-
_
posing price ​v​ − M + c is 2c(1 − q). So his strategy is optimal if there is a lower
bound on the probability he attaches to his uninformed partner believing in market
confidence and M is sufficiently large relative to c.

IV. Lessons From the Crisis

Our model is stylized but informed by the recent financial crisis, and it addresses
some puzzling questions that are raised by the financial crisis of 2007–2009. The
problem of “toxic assets” first hit the headlines when the subprime crisis heralded
the beginning of the global financial crisis in August 2007. The market for certain
asset-backed securities, especially those backed by subprime residential mortgages,
was the first to suffer extreme illiquidity, as trading slowed to a trickle and market-
clearing prices became virtually impossible to establish.1 The opaqueness of the
asset-backed securities market and the attendant potential for adverse selection has
frequently been blamed for the sudden drying up of liquidity. Yet, there is a puzzle at
the heart of the crisis. Uncertainty about the true value of an asset should not invari-
ably lead to the breakdown of trade. The stock market is a good illustration of how
financial markets are normally well adapted to aggregating the diverse information
of traders and arriving at a market-clearing price.
Our framework gives a possible avenue to resolving this puzzle. The starting
point of our analysis was adverse selection resulting from information asymmetries
on the true value of the asset. For asset-backed securities, the heterogeneity of the

1
See Gorton (2010) for a detailed description of the subprime securitization process and the initial phase of the
crisis. See also Adrian and Shin (2010), Greenlaw et al. (2008), and Brunnermeier (2009).
Vol. 4 No. 1 morris and shin: Contagious Adverse Selection 19

underlying loan pools that back the securities gives ample scope for greater exper-
tise and information in ascertaining the fundamental value of the securities. When
overall economic fundamentals are strong, such asymmetric information need not
matter for the value of the particular asset-backed security, since such securities
are debt claims that are insensitive to the value of the underlying claims, as noted
by Gorton and Pennacchi (1990). However, when a shock impacts the economy
(such as reversal of the housing market that ultimately underpins the value of the
security), then the true value of the debt security becomes more sensitive to private
information and the asymmetric information begins to exert an influence in the trad-
ing decisions.
Moreover, the new breed of asset-backed securities, such as collateralized
debt obligations (CDOs) written on subprime mortgages, have skewed payoffs in
which they retain their value close to face value in most states of nature, but suf-
fer catastrophic loss in extremely bad states (see Coval, Jurek, and Stafford 2009).
Theoretically, the skewness of payoffs that combines small gains from trade in nor-
mal states with large losses in bad states is captured by the parameter M. Recall that
the loss ratio ψ can be approximated as

ψ ≈ ​ _
qδM
(10) c ​,

where the approximation holds good when q is small (i.e., the incidence of informed
traders is low) and δM is much larger than c—that is, when the possible loss to hold-
ing a defective security is much larger than the possible underlying gains from trade.
M being large is arguably a good description of the most toxic of the mortgage-
backed securities such as CDOs and CDO-squareds.
In addition, note that our results do not rely on ψ going up uniformly across all
types of traders in the market. The breakdown of market confidence can result from
the ψ loss ratios going up for a small number of traders, and it is these traders leav-
ing the market that sets off the vicious circle of greater illiquidity inducing more
traders to leave the market.

V. Concluding Remarks

Our result on the importance of approximate common knowledge in enabling

mutually beneficial trade reiterates the importance of shared understanding as in
many other areas of economic life. Arguably, credit ratings and accounting num-
bers also derive part of their importance from common understanding. Holmstrom
(2009) argues that the coarse nature of credit ratings serve this important purpose,
and that misguided attempts to enhance “transparency” by making finer distinc-
tions may undermine this useful purpose. Elsewhere (Morris and Shin 2007a), we
have argued that accounting numbers also serve the important role of generating
shared understanding. There are inevitable tradeoffs. The imperative for common
understanding can sometimes detract from the precision of accounting numbers.
Common understanding is predicated on the lowest common denominator—the
coarsest shared framework among a set of disparate individuals. So, the coarser the
20 American Economic Journal: MAcroeconomicsjanuary 2012

information, the greater the chance that the information can be understood by all.
However, coarse information is also imprecise information. When communication
is based on the coarsest individual information, there will be many individuals who
are capable of handling more finely nuanced and complex usage. Hence there may
be welfare losses when the opportunity to utilize the greater sophistication is for-
gone in favor of simplicity.
When common understanding is important, it is possible that greater precision
of information can be detrimental to welfare if the greater precision comes at the
expense of greater fragmentation, or if the greater precision of information leads to
an exacerbation of externalities in the use of information that detracts from overall
welfare. Accountants make the distinction between disclosure of information (e.g.,
reporting of numbers in a footnote) and recognition (e.g., inclusion in profit and
loss statement) and observe that the latter has a larger empirical impact than the for-
mer (Barth, Clinch, and Shibano 2003; Espahbodi et al. 2002). The greater impact
of recognized numbers presumably reflects greater common understanding of that
information.
In this paper, we have seen the interaction between asymmetric information and
the coordination motive generated by that asymmetric information. Our theme has
been the corrosive effect of even small amounts of adverse selection in an asset
market and how it can lead to the total breakdown of trade. In our model, there is
common knowledge that an asset is worth strictly more to the buyer than the seller
at every state of the world, and yet there can be a breakdown of trade. The problem
is the failure of common understanding, and in particular what we have termed
“­market confidence,” defined as approximate common knowledge of an upper
bound on expected losses.

References

Adrian, Tobias, and Hyun Song Shin. 2010. “Liquidity and Leverage.” Journal of Financial Interme-
diation, 19(3): 418–37.
Akerlof, George A. 1970. “The Market for ‘Lemons’: Quality Uncertainty and the Market Mecha-
nism.” Quarterly Journal of Economics, 84(3): 488–500.
Aumann, Robert J. 1976. “Agreeing to Disagree.” Annals of Statistics, 4(6): 1236–39.
Barth, Mary E., Greg Clinch, and Toshi Shibano. 2003. “Market Effects of Recognition and Disclo-
sure.” Journal of Accounting Research, 41(4): 581–609.
Bhattacharya, Utpal, Philip J. Reny, and Matthew Spiegel. 1995. “Destructive Interference in an
Imperfectly Competitive Multi-security Market.” Journal of Economic Theory, 65(1): 136–70.
Bhattacharya, Utpal, and Matthew Spiegel. 1991. “Insiders, Outsiders, and Market Breakdowns.”
Review of Financial Studies, 4(2): 255–82.
Bolton, Patrick, Tano Santos, and Jose A. Scheinkman. 2009. “Market and Public Liquidity.” American
Economic Review, 99(2): 594–99.
Bolton, Patrick, Tano Santos, and Jose A. Scheinkman. 2011. “Outside and Inside Liquidity.” Quar-
terly Journal of Economics, 126(1): 259–321.
Brunnermeier, Markus K. 2009. “Deciphering the Liquidity and Credit Crunch 2007–2008.” Journal
of Economic Perspectives, 23(1): 77–100.
Carlsson, Hans, and Eric van Damme. 1993. “Global Games and Equilibrium Selection.” Economet-
rica, 61(5): 989–1018.
Coval, Joshua, Jakub Jurek, and Erik Stafford. 2009. “The Economics of Structured Finance.” Jour-
nal of Economic Perspectives, 23(1): 3–25.
Dang, Tri Vi. 2008. “Bargaining with Endogenous Information.” Journal of Economic Theory, 140(1):
339–54.
Vol. 4 No. 1 morris and shin: Contagious Adverse Selection 21

Dang, Tri Vi. 2009. “Information Acquisition, Noise Trading and Speculation in Double Auction Mar-
kets.” Unpublished. http://www.vwl.uni-mannheim.de/gk/_tri-vi/Paper_DA.pdf.
Dang, Tri Vi, Gary Gorton, and Bengt Holmstrom. 2009. “Opacity and the Optimality of Debt for
Liquidity Provision.” Unpublished. http://www.sloanweb.mit.edu/finance/pdf/Gary%20Gorton.pdf.
Diamond, Douglas W., and Philip H. Dybvig. 1983. “Bank Runs, Deposit Insurance, and Liquidity.”
Journal of Political Economy, 91(3): 401–19.
Dow, James. 2004. “Is Liquidity Self-Fulfilling?” Journal of Business, 77(4): 895–908.
Espahbodi, Hassan, Pouran Espahbodi, Zabihollah Rezaee, and Hassan Tehranian. 2002. “Stock
Price Reaction and Value Relevance of Recognition versus Disclosure: The Case of Stock-Based
Compensation.” Journal of Accounting and Economics, 33(3): 343–73.
Gorton, Gary B. 2010. Slapped by the Invisible Hand: The Panic of 2007. New York: Oxford Univer-
sity Press.
Gorton, Gary, and George Pennacchi. 1990. “Financial Intermediaries and Liquidity Creation.” Jour-
nal of Finance, 45(1): 49–71.
Greenlaw, David, Jan Hatzius, Anil K. Kashyap, and Hyun Song Shin. 2008. “Leveraged Losses: Les-
sons from the Mortgage Market Meltdown.” US Monetary Forum Report No. 2. http://research.
chicagogsb.edu/igm/events/docs/USMPF_final.pdf.
Holmstrom, Bengt. 2009. Comment on “The Credit Rating Crisis,” NBER Macroeconomics Annual
2009. Vol 24 (2010), edited by Daron Acemoglu, Kenneth Rogoff, and Michael Woodford, pages
215–22. University of Chicago Press. Discussion at NBER Macro Annual conference, April.
Kirabaeva, Koralai. 2010. “Adverse Selection, Liquidity and Market Breakdown.” Unpublished.
Kiyotaki, Nobuhiro, and John Moore. 1997. “Credit Cycles.” Journal of Political Economy, 105(2):
211–48.
Kurlat, Pablo. 2009. “Lemons, Market Shutdowns and Learning.” Unpublished.
Monderer, Dov, and Dov Samet. 1989. “Approximating Common Knowledge with Common Beliefs.”
Games and Economic Behavior, 1(2): 170–90.
Moore, John H. 2010. “Contagious Illiquidity.” Unpublished. http://www.ed.ac.ak/polopoly_
fs/1.68984!fileManager/JHM_WC100922.pdf.
Morris, Stephen, and Hyun Song Shin. 1998. “Unique Equilibrium in a Model of Self-Fulfilling Cur-
rency Attacks.” American Economic Review, 88(3): 587–97.
Morris, Stephen, and Hyun Song Shin. 2003. “Global Games: Theory and Applications.” In Advances
in Economics and Econometrics: Theory and Applications, Eighth World Congress, Vol. 1, ed.
Mathias Dewatripont, Lars Peter Hansen, and Stephen J. Turnovsky, 56–114. Cambridge, UK:
Cambridge University Press.
Morris, Stephen, and Hyun Song Shin. 2007a. “Optimal Communication.” Journal of the European
Economic Association, 5(2–3): 594–602.
Morris, Stephen, and Hyun Song Shin. 2007b. “Common Belief Foundations of Global Games.”
Unpublished.
Morris, Stephen, Rafael Rob, and Hyun Song Shin. 1995. “p-Dominance and Belief Potential.” Econo-
metrica, 63(1): 145–57.
Myerson, Roger B. 1985. “Analysis of Two Bargaining Problems with Incomplete Information.” In
Game-Theoretic Models of Bargaining, ed. Alvin E. Roth, 115–47. New York: Cambridge Univer-
sity Press.
Pagano, Marco. 1989. “Endogenous Market Thinness and Stock Price Volatility.” Review of Economic
Studies, 56(2): 269–87.
Pagano, Marco, and Paolo Volpin. 2008. “Securitization, Transparency, and Liquidity.” CEPR Discus-
sion Paper 7105.
Rubinstein, Ariel. 1989. “The Electronic Mail Game: Strategic Behavior under ‘Almost Common
Knowledge’.” American Economic Review, 79(3): 385–91.
Spiegel, Matthew, and Avanidhar Subrahmanyam. 2000. “Asymmetric Information and News Disclo-
sure Rules.” Journal of Financial Intermediation, 9(4): 363–403.
This article has been cited by:

1. József Sákovics,, Jakub Steiner. 2012. Who Matters in Coordination Problems?. American
Economic Review 102:7, 3439-3461. [Abstract] [View PDF article] [PDF with links]