FM Lecture Series

Game Theory in Finance

Anjan V. Thakor
An, an Thakor is the INB National Bank Professor of Finance at Indiana University

It is generally believed that game theory began with

the publication of von Neumann and Morgenstern's
book, The Theory of Games and Economic Behaviour,
in 1944. However, until the 1970s, game theory was not
even an integral part of mainstream economics, let
alone finance. It is not a coincidence that game theory
became more widely adopted in economics in the
1970s, which is the time that Akerlocs [1] path-breaking work on adverse selection marked the beginning of
enormous interest in the economics of information.
Economists began to realize that limitations on the
information possessed by individuals were important
in understanding economic behavior, in part because
such limitations induced people to alter their behavior.
The resulting strategic interactions had potentially
profound implications.
During the 1960s and the early part of the 1970s,
finance was preoccupied with the notion that markets
are efficient and that the details of institutional design
are relatively unimportant in understanding the

The purpose of this paper is to provide an overview

of game theory, particularly as it relates to finance. My

objective is to introduce the subject, so I will be illustrative rather than rigorous and complete. Since it would
be impossible for me to review in this paper all of the
potential applications of game theory in finance, I will
focus on a subset of applications of game theory and
offer a few observations on possible trends. For much
of this paper, I'll work through a simple example to
illustrate how the game-theoretic treatment of a
"standard" asymmetric information problem differs
from the approaches that were popular in finance prior
to the surge of recent game-theoretic models.
This paper is an outgrowth of a talk prepared as part of a tutorial for
the FMA meeting in Orlando, Florida, October 1990. I gratefully
acknowledge helpful comments on an earlier draft by James Ang (the
editor), Mark Bagnoli, Arnoud Boot, Jess Beltz, Michael Brennan,
Sugato Chakravarty, Dan Indro, George Kanatas, Rene Stulz, and
Patty Wilson. They are, of course, absolved of responsibility for any
remaining infelicities.

71

72

FINANCIAL MANAGEMENT/SPRING 1991

functioning of markets. Thus, game theory and information economics did not have much appeal to
finance. However, the work of Leland and Pyle [38],
Ross [52], and Bhattacharya [7] brought information
economics into the arena of finance research. These
early models were attempts to break out of the domain
of irrelevance results provided by the Modigliani and
Miller [43, 44] work on capital structure and dividends.
These papers were followed by numerous papers in the
early 1980s which explained a variety of price reactions,
institutional details, and contract features. This
theoretical work dovetailed nicely with the voluminous
empirical literature on event studies which suggested
that corporate insiders such as managers have
proprietary information not reflected in prices, and
that prices do react to certain actions undertaken by
them.
This early body of work in finance involving information economics did not have very much to do with
game theory. The models belonged to a class I shall
refer to as "Signalling Models with the Uninformed
Moving First" (SMUF). I'll explain later what I mean
by this. However, these models had several limitations
which I'll discuss a little later. As we approached the
end of the last decade, finance theorists began to turn
increasingly to game-theoretic signalling models to
avoid some of the nettlesome problems in SMUF. I
shall refer to these game-theoretic models as "Signalling Models with the Informed Moving First" (SMIF).
These models now occupy center stage in research
involving signalling issues in finance.
What then is game theory? Simply put, it is a study
of situations in which we make some primitive assumptions about the agents involved in a particular interaction (or game) and then try to figure out what happens
when each agent acts to maximize his own expected
utility subject to the constraints imposed by his information (and beliefs), endowments, and production
function. The agents realize that their actions affect
each other. These agents may play this game cooperatively (i.e., they can collude to implement some outcome that jointly maximizes their welfare) or noncooperatively (i.e., each agent selfishly maximizes).
Gaming behavior, which game theory attempts to
study, is all around us. I am reminded of a movie I saw
on television some years back. It was set in the cold-war
years and dealt with the possibility of a nuclear war
between the U.S. and the Soviet Union. In the movie,
the U.S. is at war with the Soviet Union and each
contemplates the use of nuclear weapons. Each side
.

recognizes that if it launches an all-out nuclear strike

first, it can obliterate the other side with limited
damage to itself from retaliation. If both strike simultaneously, both are obliterated. Of course, if neither
side strikes, both can emerge unscathed with a peaceful
resolution. This is a classic example of a noncooperative game. We can represent the payoffs in the following matrix, called the strategic form of the game.
Soviet Union Strategy
Do Not Strike
Unless Struck
Strike First
First
U.S. Strategy

Strike First
Do Not Strike
Unless Struck
First

(,c)

(-1-,w)

L)

(G,G,)

co,

Here the first number in each payoff pair is the payoff

to the U.S. from its chosen strategy and the second
number is the payoff to the Soviet Union. L, Gu and Gs
are positive numbers. It is assumed that if either
country strikes first, the other will retaliate and the loss
to the first striker is L, i.e., its payoff is -L. If neither
country strikes, the U.S. enjoys a utility of Gu and the
Soviet Union enjoys a utility of Gs. What will happen?
In the movie, both countries push the button simultaneously and there is total destruction. As it turns out,
this is a Nash equilibrium in this game. A Nash equilibrium is a set of strategies for ( n > 1 ) agents involved in a noncooperative game. These equilibrium
strategies have the property that, holding fixed the
equilibrium strategies of all the other agents, no agent
can do better than to choose his own equilibrium
strategy. In our war game, suppose the U.S. assumes
that the Soviets will strike first. Then they cannot do
strictly better by deviating from their own equilibrium
strategy of striking simultaneously. Of course, the
Soviets can go through the same reasoning and thus
conclude that they cannot do better than choosing a
preemptive nuclear strike. Hence, one Nash equilibrium in this game is for both to strike simultaneously
and annihilate each other as in the movie. Each country
gets a payoff of 00. Note that this is a Nash equilibrium because, holding each country's equilibrium
strategy fixed, neither can do better for itself by not
striking.
Now, you might object and say, "Isn't it better for
neither side to strike, since then each is better off than

THAKOR/GAMETHEORYINFINANCE

if either side had struck?" True! Indeed, if each side

were to believe that the other side will not strike first,
then each is strictly better off not striking first. In that
case, the Nash equilibrium is for neither side to strike.
Thus, in this game there are two Nash equilibria in
pure strategies (each side's equilibrium strategy is to
choose a particular move with probability one). A
reasonable question to ask is: which Nash equilibrium
will obtain? The answer really depends on what the
agents involved in the game believe. If the Americans
and the Russians trust each other, we get the nice Nash
equilibrium involving a peaceful resolution. If they
don't, we get nuclear holocaust.
Multiple Nash equilibria are common. For example,
in the Diamond and Dybvig [22] bank runs model,
there are two Nash equilibria, one in which there is no
run on banks and one in which there is a run. They focus
on the "bad" Nash equilibrium and show that
governmental intervention with federal deposit insurance can help eliminate the bad Nash equilibrium.
As this discussion indicates, the information that
people have and the beliefs that they form are important ingredients in games. Indeed, three important elements in game-theoretic models are: information and
beliefs (which are often linked), rationality, and
strategic behavior. Much of game theory is built on the
assumption that individuals are rational, although
rationality need not be pervasive, as we will see later.
Often the mere suspicion that some player may be
irrational has far-reaching consequences. I'll argue
later that explicit attention to beliefs in SMIF and the
relative suppression of the role of beliefs in SMUF may
be the major factor in the potentially different implications yielded by these models in finance.
Thus far I've discussed game theory in very general
terms. The rest of this paper will be devoted to specifics.
I'll focus my attention largely on noncooperative game
theory, and then remark briefly on cooperative game
theory. In both these strands of game theory, individual
agents are assumed, at least for the most part, to be
Bayesian rational, i.e., they revise their beliefs according to Bayes' rule whenever possible. I will then have a
little bit to say about a new branch of game theory that
has not been used much in finance, but that might have
interesting applications. It is the concept of evolutionary stable strategies (ESS). This branch of game theory
allows agents to behave in a non-Bayesian fashion and
attempts to understand the economic world as one in
which strategies, institutions, markets, and even

73

preferences are those that survive an evolutionary

process of "fitness maximization."
My discussion of noncooperative game theory will
revolve around an example. I have taken a simple financial signalling model and computed the equilibrium we
would get under SMUF. I then view it as a model
belonging to the SMIF class and point out differences.
I'll follow this with a discussion of selected applications. This discussion is not intended to be an exhaustive review. Finally, I'll conclude with a discussion of
cooperative game theory and ESS.

I. Noncooperative Game Theory: Why It

Matters Who Moves First
A. The Sequence of Moves
Because the early signalling models of Spence [60]
and others were not game-theoretic models, they were
not really concerned with who moved first in asymmetric information settings. We know now, however,
that the sequencing of moves can matter. In most asymmetric information models, we can visualize two parties: the informed and the uninformed. The informed
party has some private information that the uninformed is trying to infer through some action or set of
actions taken by the informed. In Spence's model this
action (or signal) is education, in Ross' [52] model it is
debt, and in Bhattacharya's [7] model it is dividends.
In SMUF, the often-implicit assumption is that the
uninformed party moves first, in the sense that it
specifies a menu of its own responses to various possible values of the signal that the informed agent could
emit. Thus, the informed agent knows what the uninformed agent will do in response to each signal. For
example, in the Ross [52] model of capital structure,
the firm's manager (the informed agent) knows exactly
how the capital market will price his firm if he selects
a particular debt level. We can view that model as one
in which the capital market offers the manager a menu
of choices, where each choice is a combination of a debt
level and an accompanying market value of the firm.
The manager then picks the combination that maximizes his own payoff function. Even though the market
does not know the manager's private information
about his firm's intrinsic worth, it knows the possible
values the firm can have. Thus, it can design the menu
of combinations it offers the manager in such a way
that, regardless of the manager's private information,
he chooses a debt level that correctly reveals his firm's

74

FINANCIAL MANAGEMENT/SPRING 1991

value to the market. This is the familiar incentive compatibility) condition in financial signalling models and
we will see shortly how it works.
A key feature of these models is that the uninformed
agent is assumed to precommit to the menu. That is,
the uninformed agent cannot change its mind about its
own response after observing the signal emitted by the
informed agent. This can often be a problem. To see
why, consider the Bhattacharya [7] dividend signalling
model.2 In that model, the firm's manager announces
a dividend that signals his firm s value. This dividend is
promised as a perpetuity since the manager's information is presumed to not change through time. The
signalling equilibrium is one in which the higher the
value of the firm (privately known by the manager), the
higher is the promised dividend. What ensures incentive compatibility in the model is that the firm is forced
to use external financing to make up the shortfall whenever its realized cash flow from operations is less than
the promised dividend, and external financing is more
costly than retained earnings. This delivers the intuitive result that you will promise a high dividend only if
your expectation of future cash flows is high enough.
That is, this cost structure serves to ensure that an
intrinsically lower valued firm will not be tempted to
mimic its higher valued counterpart by also promising
a high dividend. In this model, however, once the
manager makes his dividend announcement, all of the
informational asymmetry is resolved as the market
knows as much as the manager. Why then should we
subject this poor manager to the tyranny of having to
pay this dividend every period, possibly at the cost of
having to resort to distress financing to make up cash
flow shortfalls? The answer is that if we don't, the
manager will not signal truthfully in the first place. The
difficulty with this is that we are asking agents to abide
ex post by strategies that were determined to be ex ante

dividends. Everybody would know the firm's true value,

the market price would be correct, and no more valuedissipating dividends would need to be paid.3
Another way of viewing the precommitment assumption in SMUF is that the Nash equilibrium4 is an
equilibrium for the overall game, but it is not an equilibrium over every properly defined subgame within
this overall game. Strictly speaking, one cannot make
this claim within the context of Bhattacharya's dividend
signalling model because the firm's only strategy is a
single dividend level promised in perpetuity. But a
natural redefinition of that game would be to make it
truly dynamic and allow the firm to choose a possibly
different dividend level in each period. In that case, a
subgame would be the game that would ensue after the
first dividend is paid, for example. Since the market
knows as much as the firm's manager now, it is not a
Nash equilibrium for the manager to pay dividends
henceforth. That is, in this subgame, if we assume that
investors in the market are Bayesian rational, they
should price the firm at its true value, even if the
manager discontinues paying dividends. Holding fixed
this market response, it is not a Nash equilibrium for
the manager to pay any more dividends.
We have now implicitly extended our notion of what
constitutes an equilibrium. We would like the overall
Nash equilibrium strategies to also be Nash equilibrium strategies over properly defined5 components
(or subgames) of the overall game. When a particular
Nash equilibrium satisfies this requirement, it is called
a subgame perfect Nash equilibrium. Many of the equilibria characterized by models in the SMUF class are
not subgame perfect.
Another difficulty with many of the models in the
SMUF class is that they implicitly study games involving strategic interaction among agents, but they don't

'

As mentioned earlier, this will not "work" if the manager anticipates

this, because then he will misrepresent. Because of its infinite horizon
structure, Bhattacharya's [7] model has two types of precommitment:
the usual precommitment to a specific (price) response by the uninformed agent (the market) and precommitment by the informed
agent (the manager) to keep paying the same dividend perpetually.
4
Not all models of the SMUF class study Nash equilibria.
5
This is the notion of "proper subgames" as in Kreps and Wilson [36].
Loosely speaking, a proper subgame is a component of the overall
game that inherits all of the "essential structure" of the overall game.
3

'By incentive compatibility we mean that no agent has an incentive

to misrepresent, regardless of his private information. The deployment of this condition can be justified on the basis of the revelation
principle (see, for example, Myerson [46]). This principle says that if
there exists a Nash equilibrium in a general reporting game that does
not necessarily impose truth-telling constraints, then the same equilibrium outcome can be replicated in a simpler reporting game with
truth-telling constraints.
2
My comments apply to all SMUF. I do not mean to single out
Bhattacharya [7]. In this, as well as many other dissipative signalling
models, signalling costs are shared by the sender (the manager in this
case) as well as the receiver (the investors) since taxes, transactions
costs, etc., reduce the intrinsic value of investors' holdings.

efficient but are not necessarily ex post efficient. This

is often difficult to ensure. In our dividend signalling
example, the firm's value would be higher if the
manager paid one dividend and then stopped paying

THAKOR/GAME THEORY IN FINANCE

explicitly specify a formal structure for the game, complete with a precisely defined sequence of moves, information sets at various stages of the game, and beliefs.6
This is a problem mainly because Nash equilibrium
requires that we also specify what happens off the
equilibrium path. That is, what if an agent does something he was not supposed to do in equilibrium? Due
to the lack of an extensive form for the game, models
in the SMUF class usually specify a set of ad hoc
restrictions on reactions to out-of-equilibrium moves.
Whether these restrictions would hold up if we could
endow the game with the appropriate extensive form
remains an open question.
This problem, in turn, leads to another potential
difficulty. Quite often, SMUF do not have any Nash
equilibria. Rothschild and Stiglitz [53], in their famous
paper on insurance markets, showed that a Nash equilibrium may fail to exist when the (uninformed) insurance company moves first to specify a set of insurance contracts and the privately informed individual
buying the insurance moves next, even in the case when
the informed agent can be one of only two types. The
reason for this nonexistence will become clear later on
when I analyze a specific model. Later, Riley [51] established the even more powerful result that, under
plausible conditions on the cross-sectional distribution
of types, a Nash equilibrium never exists in the SMUF
class with a continuum of possible types of informed
agents.? This problem of nonexistence of a Nash equilibrium led Riley [51] to define a non-Nash equilibrium
concept called the reactive equilibrium (henceforth
referred to as the Riley Reactive Equilibrium or RRE
for short). Riley provided sufficient conditions for the
existence of an RRE with a continuum of types.8
The picture changes colors when we turn to SMIF.
Instead of assuming precommitment by the uninformed, we now assume that the informed agent moves
first and emits the signal, based on his private information. The uninformed agents rationally interpret this
)This formal structure, with all of the dynamic features intact, is called
the extensive form of the game. See Engers and Fernandez [24] for
an attempt to provide a formal game-theoretic structure to some of
the models in the SMUF class.
7
By a Nash equilibrium here I mean the Rothschild and Stiglitz [53]
version rather than the incompletely specified Nash equilibrium of
Spence [64
8
Earlier, Wilson [66] proposed his "anticipatory equilibrium" as
another non-Nash equilibrium concept. But Riley showed that,
with a continuum of informed agents, an anticipatory equilibrium
also does not exist.
(

75

signal and revise their priors about the informed

agent's type, using Bayes' rule whenever possible. What
do I mean by whenever possible? Well, suppose there
are n possible types of the informed agent, and the
uninformed agents know that there are n possible
types, but they cannot determine precisely a particular
informed agent's type. Suppose that the common belief
is that there will be a separating signalling equilibrium,
with each type of informed agent emitting a distinct
signal. Then there are n possible equilibrium values of
the signal. So if an informed agent emits one of these
n signals, the uninformed can use Bayes' rule to update
their priors. But what if there is a defection from the
conjectured equilibrium, with the informed agent
choosing a signal that is outside the set of n equilibrium
signals? The uninformed agents cannot use Bayes' rule
now since the prior probability of the signal was zero.
How the uninformed agents are supposed to formulate
their beliefs in such a setting has occupied the attention
of game theorists for quite some time now, and I'll
return to this issue later.
This leads me to an interesting digression. In statistical decision theory, zero probability events are inconsequential because, by definition, the likelihood of
their occurrence is zero, so that they don't affect expected utility. But in game theory, we can't ignore such
events. The reason is that what has zero probability is
endogenous to the equilibrium we are focusing on. A
zero probability event in one equilibrium may have
positive probability in another. Moreover, whether an
outcome is an equilibrium depends on the zero probability events, not just the positive probability events.
Returning to SMIF, one major advantage of these
models is that there is no precommitment requirement
of the type encountered with SMUF. Another advantage is that everything is made precise: the sequence
of moves, information sets, beliefs, etc. This eliminates
much of the ad hoc structure inherent in SMUF. Finally, there is usually no paucity of equilibria. Indeed, we
now have an embarrassment of riches, as there is often
a large number of Nash equilibria, and the modeler may
be required to choose between them. Although a final
answer has yet to be provided, game theory has
provided a way to think about this choice. This is an
issue I turn to next. Rather than go through the literature on refinements of Nash equilibrium which is
quite extensive and growing I'll do a simple example
familiar to finance researchers and show how we can
compute the various equilibria.

76

FINANCIAL MANAGEMENT/SPRING 1991

Before getting to that, I would like to make a simple

point. In any signalling game, we have to deal with
defections from the equilibrium, i.e., agents possibly
doing things they are not supposed to in equilibrium.
Who can defect depends on how the moves are sequenced in the game. In signalling games with welldefined first and second movers, it is always the first
mover who has the opportunity to defect. Thus, in
SMUF it is the uninformed agent who may defect by
offering the informed agent a contract that is outside
the equilibrium set of contracts. In SMIF, it is the
informed agent who may defect by emitting a signal that
lies outside the set of equilibrium signals.
B. A Debt Signalling Example: SMUF, SMIF
and Refinements

Consider a single period economy in which all

agents are risk neutral. At t = 0 (the beginning of the
period), firms announce debt levels and at t = 1, the
debt is repaid out of cash flows if possible. There are
three types of firms in the economy: the good, the bad
and the ugly. The density function of the end-of-period
cash flow for the good firm is fg(x), for the bad firm is
fb(x) and for the ugly firm is fax). The support offi(x)
is ( 00,co )ViEl &b,u }.Also,f(x)>OVxE(oo,co
and Vi E { gb,u }. The riskless rate of interest is zero
for simplicity. The utility function of the firm's
manager is defined over the market value of the firm at
t=0 (the manager's wage is all, where a is a positive,
finite, real valued constant and Vis the market value of
the firm at t=0) and whether or not the firm goes
bankrupt at t= 1. Bankruptcy is possible only if the firm
is levered. In the event of bankruptcy, the manager is
assessed a personal (nonpecuniary) penalty of
E( 0,00 ), which is a constant and does not directly
affect firm value. This penalty is purely dissipative, i.e.,
irdoes not accrue to the investors. LetFi be the cumulative distribution function (CDF) corresponding to the
density functionfi. The CDFs of the three types of firms
are ordered by the following first-order stochastic
dominance (FOSD) relationship: g FOSD b FOSD u.
Investors' (common) prior belief is that a fraction 0g of
all firms are good, a fraction 64, are bad and a fraction
= 1 Og Oil are ugly. However, although each
firm's manager knows his own firm's type, investors are
unable to a priori distinguish one type of firm from
another. The manager makes decisions to maximize his
own expected utility. There are no taxes or bankruptcy
costs. We want to analyze this as a signalling game in
which the manager's choice of debt, D (chosen from
some compact interval of real numbers), is a signal of
his private information.9
)

1. Analysis of the Game as SMUF

We can write the expected utility of a manager

choosing debt level Di and managing a firm of type i as
EU(j I i) = aV(Dj) 7r,Fi(Di)

(1)

where we recognize that the firm's current market

value can be based only on the observed debt signal. I
will derive the unique RRE in this game. In our context
the RRE is a set of three distinct signals (debt levels)
such that: (i) each firm is correctly priced (the expected
return of investors who purchase each firm is zero) and
(ii) there does not exist any alternative scheme (a pair
consisting of a signal and an associated firm value) that
could be offered to firms by any coalition of investors,
say C1, such that C1 could earn a nonnegative expected
return on this scheme even after another coalition, say
C2, reacts to the initial defection with yet another
out-of-equilibrium scheme. The key is that C2 should
see a positive expected return in reacting to C1, and no
further reactions can impose strict losses (negative
expected return) on C2.
The RRE for this game is as follows. The worst type
of firm, which is the ugly firm, sets its debt level
D, = 0. The bad firm signals with a debt level Db > 0
which is set to satisfy
WOO aF,ADO = a110)

(2)

where the left hand side (LHS) is the expected utility

of the ugly firm's manager from mimicking the bad
firm's manager and the right hand side (RHS) is the
expected utility of the ugly firm's manager from signalling truthfully. Note that competitive capital market
pricing requires that, consistent with the conjectured
RRE,

vov=".---L.fb(x)dx

(3)

and
Those familiar with Ross [52] will recognize that this is essentially
the Ross model.
9

THAKOR/GAME 'THEORY IN FINANCE

v(0)./i.=Lx.f. (x) dx

77

(4)

wherepb and pu are the mean cash flows of the bad and
ugly firms, respectively.
Substituting (3) and (4) in (2) and solving for Db gives
(aPb POPO

Db

(5 )

Similarly, the good firm signals with a debt level

Dg > 0 which satisfies
aV(Dg) aFb(Dg) = aV(Db)aFb(Db)

(6)

where the LHS is the expected utility of the bad firm

from mimicking the good firm and the RHS is the
expected utility of the bad firm from signalling truthfully. Again, competitive capital market pricing requires that, consistent with the conjectured RRE,
V(D g) = ,ug = x (0

(7)

Substituting (3) and (7) in (6) and solving for Dg gives

us
Dg = Fb 1 (faiug

+ 7rFb(Db) fir)

(8)

where Db is given by (5). It is easy to verify that the ugly

firm will not mimic the good firm and the good firm will
not mimic either of its lower-valued counterparts.
Also, the bad firm will not mimic either the ugly or the
good firm. Moreover, Dg > Db > 0 .
It is appropriate to view this equilibrium as the
outcome of a game in which the uninformed capital
market is a single player that offers the following menu
of choices to firms: (1) choose a debt level of zero and
we will price your firm atpu, (ii) choose a debt level of
Db and we will price your firm at/Ub, and (iii) choose a
debt level of Dg and we will price your firm at pg. The
firm's manager then proceeds to choose a particular
debt level and each firm is priced in equilibrium as if
investors had the same information as managers.
Preconunitment from the market to adhere to this
pricing strategy is almost as crucial here, however, as it
is in the dividend signalling model of Bhattacharya [7]
that we discussed earlier. Recall that in that case the
market was precommitted to a pricing response conditional on the manager paying a particular dividend
perpetually; after the first period dividend, however, all
informational asymmetry is resolved and there is little

reason for the manager to keep paying the dividend,

except for the necessity of the precommitment to be
adhered to for ex ante incentive compatibility reasons.
We can view the nature of precommitment in our debt
signalling model in a similar light. Once the firms have
chosen their respective debt levels, suppose the
manager of the firm that issued debt Ds were to announce to the market; "You know that I have chosen
Dg, so my firm is good. There is no reason now for you
to force me to be levered this way until the end.
Could I not reduce my leverage by repurchasing some
of my debt with an equity issue?" Clearly, in an ex post
sense, investors are no worse off if the manager did
this, and the manager is strictly better off. But
precommitment in this game requires that the
market's price response ofpg be rigidly conditional on
the manager not being able to do this, or else ex ante
incentive compatibility is destroyed. In other words,
this equilibrium hinges on there being precommitment
to persist with an arrangement that is ex post efficient
to dissolve.10 Models in the SMUF class lack the
appropriate structure to deal satisfactorily with this
complication.
Another point to note is that the equilibrium we
have characterized is an RRE under our assumptions.
However, suppose Og is very close to one. Then this
equilibrium will not be a Nash equilibrium (in the
Rothschild and Stiglitz [53] sense). In our context, a
Rothschild and Stiglitz Nash equilibrium is a set of
three distinct signals such that: (i) each firm is correctly
priced in the equilibrium, and (ii) there does not exist
an alternative scheme (a pair of debt signal and associated firm value outside the equilibrium set) such
that the coalition of investors offering it could earn
strictly positive profit and there is at least one type of
firm that would prefer this scheme to its equilibrium
allocation. Intuitively, the reason why a Nash equilibrium fails to exist when 11g is close to I is that a
group of uninformed investors can successfully
defect from the equilibrium by offering a contract
that prices all firms at a pooling price of
=08Ug0tPb + [1 eg 0b]Pis and a debt level of
zero. If Og is sufficiently close to 1, it can be shown that
wOne could argue that the RRE could cope with a possible debt
repurchase decision by the manager by redefining the game to treat
the initial debt issue and the subsequent debt repurchase as joint
signals. However, we are still left with the problem of precommitment
to the repurchase fraction. This precludes arrangements that are ex
post more efficient than these stipulated in the equilibrium.

78

all firms will prefer this defection to their equilibrium

allocations. For example, the manager of the good firm
finds that the gain from signalling via debt, a[llgis insufficient to compensate him fora; the bad and the
ugly firms are, of course, happy to be subsidized. In this
case, no Nash equilibrium exists, since a pooling outcome in such games is never a Nash equilibrium (see
Rothschild and Stiglitz [53]).11
The RRE characterized here has predictions that
are generally in accord with some of the recent empirical evidence, namely that there is a positive association
between debt and firm value.12
2. Analysis of the Game as SMIF

Now imagine that the informed manager is moving

first, choosing his debt signal, and the a priori uninformed capital market moves next by setting a market
value for the firm. An extensive form for this game is
shown in Exhibit 1. This extensive form is drawn as if
the firm had only four debt levels to choose from the
equilibrium levels 0, Dg and Dg (corresponding
respectively to the debt choices of the ugly, bad and
good firms in a perfectly separating equilibrium) and
some other arbitrary out-of-equilibrium debt level D.
This is done merely for simplicity. In principle, the firm
can choose from a continuum of debt levels. Moreover,
for simplicity I also assume that the market can choose
from only four valuation responses. In this extensive
form, a vertical dotted line indicates that the market is
unable to tell which type of firm made that move. Each
circle (clear or full) is a node, and the box above the
node indicates whose turn it is to move. The full black
circles at the extreme left and extreme right of the
figure are payoff nodes the pair (Si,S2) indicates that
Si is the manager's expected utility and S2 is investors'
expected return. (For more details on extensive forms,
see Kreps and Wilson [36]).
Dasgupta and Maskin [20, 21] show that the nonexistence of Nash
equilibrium is due to the restriction to pure strategies and that
randomized strategies can often restore existence, thereby verifying
an earlier conjecture by Ross [52].
12
For example, Masulis [39, 40] reports that exchange offers that
result in increases (decreases) in leverage are associated with positive
(negative) abnormal common stock returns. More recently, Cornett
and Travlos [19] specifically investigate the information effect caused
by a firm's change in capital structure via debt-for-equity and equityfor-debt exchange offers. Their evidence indicates that debt-for-equity swaps lead to abnormal stock price increases, while equity-for-debt
swaps lead to abnormal stock price decreases.
I I

Suppose, for the moment, that DA = Dg = D*>0.

I'll now show that a completely pooling outcome in
which each firm chooses an equilibrium debt level

FINANCIAL MANAGEMENT/SPRING 1991

D*>0 satisfying
a~Ctu>

art orFu(D*)> 0

(9)

is a Nash equilibrium where 17 is the pooling mean

defined earlier. To see why this is a Nash equilibrium,
suppose a manager defects by choosing D # D*. This is
an out-of-equilibrium move and the component of the
game inside the dotted area represents the off-theequilibrium path part of this game. How should the
capital market react to this event which was supposed
to have zero probability? Although the Nash equilibrium concept stipulates how the market should react
to each event (including those that have zero probability), it places no restrictions on reactions to out-ofequilibrium moves, other than that such reactions help
to sustain the equilibrium. So suppose we define a
number K > 0 satisfying
apulK <

2rF.(D*)

and assume that whenever a manager chooses a debt

level r) # D* (where b.- could be zero), his firm will be
priced at ktd.K. That is, this response by the market
is a part of the game that is off-the-equilibrium path and
constitutes a threat used to support the Nash equilibrium. It is straightforward to verify that no manager
will choose to defect from our conjectured Nash equilibrium.
This Nash equilibrium does not make much sense,
however. All firms are being pooled at a common price
and yet their managers are required to issue a positive
amount of debt which is personally costly to them. The
reason why this is a Nash equilibrium is that the equilibrium concept permits too much latitude in specifying reactions to out-of-equilibrium moves. This is
where refinements of Nash equilibrium come in. The
earliest of these refinements was subgame perfection
which I referred to earlier. Unfortunately, subgame
perfection admits too many unreasonable equilibria
under asymmetric information, even though it prunes
away some Nash equilibria. That is, there are many
subgame perfect Nash equilibria that are unreasonable. A class of situations in which this happens
is when the only proper subgame is the overall game

THAKOR/GAME THEORY IN FINANCE

79

Exhibit 1. Illustrative Extensive Form for Debt Signalling Game

I1

'

.-.
.

($1, s2 )

IMArketi

cIs i ' s2)41''- - . . . % .

path of game

_
)
D
Off-equilibzium D

D
Q

itself. In this case it is

obvious that every Nash
equilibrium is subgame

perfect.
A further refinement was
proposed by Kreps and Wilson
[36]. It is called sequential

equilibrium
and it puts
restrictions
on strategies
and beliefs.
In a sequential
equilibrium,
we must not
only specify
strategies
and beliefs
of players at
every node
of the game
that can be
reached in
equilibrium,
but also at
nodes that
cannot be
reached in
equilibrium.
That is, a
sequential
equilibrium
is a Nash
equilibrium
with
the
added
requirements that, at any point in the game, the
strategies of players are Nash, given the strategies and
beliefs of the other players, and the beliefs at each
information set in the game are formulated
according to Bayes' rule

80

wherever possible.13 Simply put, in the context of our

example, sequential equilibrium imposes two restrictions that Nash equilibrium does not. First, over any
(properly defined) subgame of our overall game,
strategies of all players should be Nash.14 Second, as in
the case of Nash equilibrium, beliefs must be revised in
accordance with Bayes' rule whenever an equilibrium
move is observed, but when an out-of-equilibrium
move is observed, there must be some belief on the part
of the uninformed that makes their reaction to this
out-of-equilibrium move rational on the basis of that
belief. In game-theoretic terminology, the reaction of
the uninformed to even an out-of-equilibrium move
should be a best response given some belief. A key
advantage of sequential equilibrium over subgame perfection is that, unlike subgame perfection, the sequential equilibrium concept allows us to eliminate some
Nash equilibria even in games where the only proper
subgame is the overall game itself.
We can see now that our Nash equilibrium is not
necessarily sequential. There is no belief on the part of
investors that would justify a (competitive) market
value of ,/K as a best response for K > 1. In other
words, the threat contained in a response of /
4/K for K > 1 is not credible. Intuitively, the reason
for this is that K > 1 means that the firm choosing an
out-of-equilibrium move is priced at less than its lowest
possible value. That is, whenever the K satisfying (9)
exceeds 1, the Nash equilibrium described earlier is not
sequentia1.15 However, a pooling equilibrium in which
the strategy of each firm is to issue zero debt with
probability one, and all firms are priced at /7 if they
issue zero debt and at /su otherwise is a sequential
equilibrium under the following structure of beliefs:
Pr(firm is of type i I firm issues zero debt) = 0i, the
prior probability that firm is of type i for all i E{g,b,u},
and Pr(firrn is of type i I firm issues positive debt)

FINANCIAL MANAGEMENT/SPRING 1991

= 1 ifi=u to . The sequential equilibrium strategy of all

firms is to issue zero debt with probability one regardless of type. Note that if a firm issues zero debt, the
Bayesian posterior of investors about this firm's type
coincides with the prior belief. Moreover, conditional
on the out-of-equilibrium belief that any deviating firm
is ugly, the market's reaction is a best response. Given
this response, no firm defects from the equilibrium.
Note that what I have really proved is that this is a
Perfect Bayesian Nash equilibrium, since I have not
checked that the equilibrium strategies and beliefs are
the limit of some sequence of (Bayesian) rational
beliefs and strategies. It is, however, straightforward to
check that this third requirement is satisfied here.16
As is transparent, this no-signalling equilibrium
rests on a somewhat objectionable specification of outof-equilibrium beliefs. Game theorists have developed
a literature on refinements of the sequential equilibrium itself, which attempts to provide ways of put'6To see this, let a; (D) represent the strategy of the manager of the
type-i firm, with i E {gb,u}, so that the probability with which this
manager chooses debt D is ai (13). Define the sequence of strategies,

faND), n = 1,2,..., c I, as follows:

1 ifD=0
cr; (D) = 10 if D > 0 for every n
E (D)

and

5n>0 for every n

(D) = , -n1 ifD=0

f(D)/n ifD> 0

where f(D) is some (arbitrary) probability density function over

(0, co) that is finite and nonnegative over (0, cc). Since f(D) is a
co
probability density, f f(D)dD = 1. Note that this sequence of

Strictly speaking, 1 have defined a Perfect Bayesian Nash equilibrium. Sequential equilibrium adds a third condition that the equilibrium strategies and beliefs should be the limit of some sequence of
(Bayesian) rational beliefs and completely mixed strategies (see
Kreps and Wilson 1361 and Rasmusen [50]). However, the equilibria
I study henceforth are sequential Nash equilibria. Note that a
"strategy" in this context is defined as a probability distribution over
all possible moves that the informed player can make, i.e., a strategy
ies the probability with which each move will be made.
is the subgame perfection requirement.
On the other hand, if K < 1, then an equilibrium with D > 0 may
be sequential but will not survive many of the refinements of sequential equilibrium.

strategies is such that lim ,r(D)= (D) Vie {gb,u}, where

n-.00

a#(D) is the equilibrium strategy of the type-i firm's manager, and it

satisfies
1 D
a(D)
01 if D= > 0
for every i E {gb,u }. With this sequence of strategies, we can also
specify beliefs that satisfy Bayes' rule for out-of-equilibrium moves
as well as equilibrium moves (recall that Bayes' rule cannot be applied
to out-of-equilibrium moves, given just the equilibrium strategies).
To see this, let P;i(i ID) be the market's posterior belief that the firm

THAKOR/GAME THEORY IN FINANCE

81

ting further sensible restrictions on beliefs off the equi

librium path. One such refinement is known as the
intuitive criterion of Cho and Kreps [16]. An equilibrium passes the intuitive criterion if: (1) it is sequential and (ii) it satisfies the following additional restriction on beliefs and strategies off the equilibrium
path.
Suppose the uninformed observe an out-of-equilibnum move, say in, in a game with N possible
types of the informed agent. If there does not exist
any belief on the part of the uninformed that
would support a best response on their part such
that a subset of player types, say S C N, would
defect with move m, then the uninformed should
assign zero probability (in their posteriors) to the
event that the defector's type belongs to S. Then,
for a sequential equilibrium to survive the ChoKreps intuitive criterion, for every type remaining in MS, there exists at least one probability
distribution of beliefs of the uninformed, concentrated on the remaining types MS, such that
the accompanying best response of the uninformed does not induce that informed agent type
to strictly prefer to defect from the equilibrium.
This criterion is indeed very intuitive. It says that we
should eliminate sequential equilibria that can only be
supported by out-of-equilibrium beliefs that put positive probability weight on types that would never wish
is of type i, having observed move D as part of the nth element of the
sequence of strategies {o'(D)}. Then, using Bayes' rule
PAfirm is ugly' some specific D>0) {fiDyn} 11-0E-00

ff(D)In} 11-0g-00 + 0

PAfirm is ugly I D=0) = 1 Vn

[1 -411[1figBbl

PP(firm is bad I D=0)

([1Ti] If Og 00+ Og+ 01/1

El

P;(firm is good I D=0)

08-0b1

[1 {1 Og 0b}]
B
[1

It is apparent that

_____O
__________
1

[1T1 {1-0gOb}l
lim

1D>0)=

0 if

= g orb b

{1-0g eb}l

l ifi
Um P' (i ID=0) = ',boo

an d

n-0)

with the assumed sequence of strategies converges to the equilibrium

beliefs.

to defect from the equilibrium, when reactions by

the uninformed to defections by the informed are
governed by the sequential equilibrium best response
(rationality) requirement.17
We can now apply the Cho-Kreps intuitive criterion
to eliminate the pooling sequential equilibrium.
Define D as a debt level such that
= a/4g arFb(D).

(10)

We continue to assume that a player will not defect

from the equilibrium unless he/she is strictly better off
for doing so. Thus, D is a debt level such that the bad
firm will not wish to defect from the equilibrium even
if the uninformed believe with probability one that the
defector is good. Thus, there does not exist any belief
on the part of the uninformed that would induce defection by the manager of the bad firm, since he'll do
strictly worse with any belief that puts less than one
probability weight on the defector being good. Because
of the FOSD relationship among the types, it is easy to
see that
> aug 7iFi,(D)

(11)

so that the manager of the ugly firm strictly prefers to

stay with his pooling equilibrium strategy, and
< aug ;WO')

(12)

so that the manager of the good firm strictly prefers to

defect from the pooling equilibrium.
We have thus established that there does not exist
any belief such that either the ugly or the bad firm's
manager will wish to defect with the debt level D.
According to the Cho-Kreps intuitive criterion, we
should set at zero the probability that the defector is
either the ugly or the bad firm. Thus, the only admissible posterior belief is that the defector is the good
firm with probability one. Given this belief, we now
require that the (competitive) market's price response
be a best response. That is, the defector should be
priced as if it is a good firm. And if this is the case, we
have already established that the manager of the good
firm will indeed strictly prefer to defect by issuing debt

=u

for every L Thus the sequence of Bayesian rational beliefs associated

"That is, the uninformed must react to a defection in a manner

rationalized by at least some belief on their part.

82

FINANCIAL MANAGEMENT/SPRING 1991

of D . Hence, the pooling sequential equilibrium does

not satisfy the Cho-Kreps intuitive criterion.
There is, however, another sequential equilibrium
that does survive the Cho-Kreps intuitive criterion,
given appropriate assumptions regarding the market's
prior beliefs about types. This is a partly pooling equilibrium in which the good firm's manager signals with
a debt level Dg > 0 and the managers of the bad and
ugly firms pool at a debt level of zero.
For equilibrium moves, the a priori uninformed
investors should revise their beliefs in accordance with
Bayes' rule. Since they know that only the bad and the
ugly firms are pooling at a zero debt level, when they
see a firm that issues zero debt, they should believe that
there is a probability Ob/[1 Og I that this firm is bad
and a probability [1 Os Bb]/[1 Os] that this firm is
ugly. Thus, in equilibrium the common pooling price
of firms that issue zero debt should be
Vp

1 e gaub + [{1 9 g Ob} fl e gau u ( 13 )

The equilibrium expected utility of the manager of

either a bad or an ugly firm is
(14)

A firm that issues debt of Dg will be priced at,us. So the

equilibrium expected utility of the manager of the good
firm is

apg 7rFg (DO .

(15)

For incentive compatibility we need

a Vp Aug 7rFb (13g) .

(16)

so that the manager of the bad firm does not mimic that
of the good firm. As usual, (16) holds tightly in equilibrium, so that we obtain the equilibrium debt level
D; = F171 ({apg aVp } )

(17)

Given (16) as an equality, it is straightforward to verify

that

3r.F.(D; )

( 18)

aVp< apg aFg (D; ) .

(19)

aVp >
and

Thus, all the incentive compatibility conditions are

satisfied.
I'll leave it up to the reader to verify that this is a
Nash equilibrium. To check that this is a sequential
equilibrium, consider a defection D E (0, Di*). Suppose investors assign the belief Pr(defector is ugly f D
E (0, Dg) observed) = 1. Then the defector is priced at
Pu. The expected utility of the defecting manager is
aP u 7r Fi ( D) Vi E { g b , u } . N o t e t h a t s i n c e
V p >,u u , aVp > a,u u aFi (D) Vi E{b,u} so that the
managers of the bad and ugly firms don't defect. And
since
aVg .7r.Fg(4) > aVp > Au, aFg (D),

the manager of the good firm doesn't defect. Thus, this

is a sequential equilibrium. Clearly, no manager will
defect with D > Dg even if the firm is priced atps.
Let us now verify that this equilibrium satisfies the
Cho-Kreps intuitive criterion. Consider some defection D E (0, Dg*). Is there any type that can be ruled
out as a potential defector, regardless of the market's
beliefs in response to the defection? Consider the
manager of the ugly firm. Although (18) is satisfied,
there may exist a debt level, say D: < 14, such that
aVp = Aug grF, (14)

(20)

So if D < Di, the manager of the ugly firm will defect,

conditional on the market believing that the defector
is the good firm with probability one. Thus, if the
manager of the bad firm wishes to defect with a debt
level such that he can convince the market that his firm
is not ugly, he must set this debt level D AI; with this
debt level, the manager of the ugly firm will never
defect.
Suppose now that a defection D E
DO is observed. By the above logic, we can rule out the ugly firm
as the defector. For the equilibrium to survive the
intuitive criterion, for each type there must exist at
least one belief concentrated over the bad and good
firms such that the associated market best response
does not cause that type of firm to strictly prefer to
defect. Assume now that the prior beliefs of the market
are such that