The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Working Paper Series
Optimal Banking Contracts and Financial
Fragility
WP 15-06
Huberto M. Ennis
Federal Reserve Bank of Richmond
Todd Keister
Rutgers University
This paper can be downloaded without charge from:
http://www.richmondfed.org/publications/
Optimal Banking Contracts and Financial Fragility
Huberto M. Ennis
Research Department
Federal Reserve Bank of Richmond
Huberto.Ennis@rich.frb.org
Todd Keister
Department of Economics
Rutgers University
todd.keister@rutgers.edu
Federal Reserve Bank of Richmond Working Paper 15-06
June 5, 2015
Abstract
We study a finite-depositor version of the Diamond-Dybvig model of
financial intermediation in which the bank and all depositors observe
withdrawals as they occur. We derive the constrained efficient allocation
of resources in closed form and show that this allocation provides liquidity
insurance to depositors. The contractual arrangement that decentralizes this
allocation resembles a standard bank deposit in that it has a demandable
debt-like structure. When withdrawals are unusually high, however,
depositors who withdraw relatively late experience significant losses. This
contractual arrangement can be fragile, admitting another equilibrium in
which depositors run on the bank by withdrawing funds regardless of their
liquidity needs.
JEL Classification Numbers: G21,G01, D82
Keywords: Bank runs; demand deposits; sequential service; liquidity
insurance
We thank seminar participants at the European Central Bank, the IESE Business School, the University of
Iowa, the Federal Reserve Banks of Philadelphia and Richmond, the 2010 Winter Meetings of the Econometric
Society, and the 2012 Meetings of the Society for Economic Dynamics for useful comments. We are especially
grateful to Ed Green and Krishna B. Athreya for helpful discussions about some of the ideas in this paper. The
views expressed in this paper are those of the authors and do not necessarily reflect the position of the Federal
Reserve Bank of Richmond or the Federal Reserve System.
1 Introduction
Banks and other financial intermediaries engage in maturity transformation by issuing shortterm liabilities while investing in assets whose full return is realized over a longer time horizon. In
fact, much of an intermediary’s liabilities are usually payable on demand, meaning that a holder
is entitled to request immediate repayment. These liabilities are also debt-like in the sense that
the value of a claim is typically not contingent on the precise timing of withdrawal; during normal times, all holders receive the “face value” of their claim. During episodes of high withdrawal
demand, however, claims that are repaid late in the process can be subjected to considerable discounts. Such intermediation arrangements often appear to be fragile in the sense of being susceptible to runs – events in which liability holders rush en masse to redeem their claims and, as a result,
resources are allocated inefficiently. In this paper, we provide a model of financial intermediation
that accounts for all of these features simultaneously. We follow Green and Lin (2003), Peck and
Shell (2003), and others in that we place no restrictions on financial arrangements or the allocation
of resources other than those explicitly specified as part of the environment.
Arrangements based on maturity transformation and debt-like liabilities are not limited to retail banking; they are widespread and integral parts of the modern financial system. A number of
such arrangements experienced events resembling a run during the recent financial crisis, including
wholesale deposits, brokerage accounts, repurchase agreements, and money market mutual funds,
to name a few.1 Regulatory reform efforts in the wake of the crisis have focused on limiting maturity transformation by requiring financial institutions to more closely align the maturity structure of
their assets with that of their liabilities and to limit the debt-like features of some contracts. Evaluating the desirability of such reforms requires having a theory of financial arrangements based on
maturity transformation that explains why these arrangements arise and what social benefits they
offer.
Diamond and Dybvig (1983) laid the foundations for such a theory by identifying important
elements of an economic environment in which financial intermediaries play a socially valuable
role while at the same time being susceptible to runs. In their model, high-return investment takes
time to mature and agents have random needs for immediate liquidity. A risk-sharing arrangement resembling a bank can improve the allocation of resources in this setting. The fact that a
depositor’s liquidity needs are private information requires the bank to allow depositors to choose
For descriptions of some of these events, see Baba et al. (2009), Duffie (2010), McCabe (2010), Gorton and
Metrick (2012), and Yorulmazer (2014).
1
1
when to withdraw their funds. This ability to choose, in turn, creates the potential for inefficient
withdrawals and fragility. However, private information alone is not enough to explain the fragility
of financial arrangements as an equilibrium outcome. Diamond and Dybvig hint at two other elements that would be needed for this theory to successfully explain fragility: (i) some form of
first-come, first-served (or sequential service) constraint and (ii) a degree of aggregate uncertainty
about the total demand for liquidity. While the role of these additional elements was discussed
informally in Diamond and Dybvig’s seminal work, the precise details of the arguments were left
mostly unexplored.
Starting with Wallace (1988), a literature has developed that attempts to explicitly model these
elements of the theory and investigate the extent to which fragility is an inherent feature of banking
and other financial arrangements.2 The approach in this literature is to fully specify the physical
environment and to derive the predicted outcomes by solving an allocation problem with no further
restrictions on what can be attained. Two important lessons have emerged. First, fragility obtains
only if the payouts associated with some of an institution’s short-term liabilities are relatively
insensitive to the state of aggregate liquidity demand. In addition, the information available to
depositors when making withdrawal decisions interacts with the payout structure in subtle ways to
determine the fragility of an arrangement.
The relative insensitivity of payouts in these models is a direct consequence of sequential service. A bank learns about the level of aggregate liquidity demand by observing individual depositors’ actions. In the absence of sequential service, the bank would fully observe this demand
before redeeming any liabilities and would set payouts accordingly. Under standard assumptions
on preferences, the resulting allocation would give depositors no incentive to run. However, when
withdrawals happen sequentially and depositors must be served as they arrive, the bank has less
information when payouts are made and, hence, these payouts are necessarily less responsive to
underlying conditions.
As more withdrawals occur, the bank will gradually learn the level of aggregate liquidity demand and adjust payouts to withdrawing depositors accordingly. In such a setting, information
about a depositor’s position in the withdrawal order becomes critical to her decision. With no information, each depositor will consider it equally likely that she occupies each possible position
in this order. All depositors will then make the same comparison between the expected utilities of
2
See, for example, the survey in Ennis and Keister (2010b) and the papers cited therein.
2
withdrawing early and of withdrawing late. If each depositor has some information about her own
position, on the other hand, each will face a different decision problem. The information structure
thus determines the nature of the strategic interaction between depositors and thereby influences
the scope for fragility.
Previous work has derived different results from different combinations of assumptions about
sequential service and the information structure. Green and Lin (2003) present one reasonable formulation and show that, under certain conditions, the optimal banking arrangement is not fragile.
In their model, all depositors report to the bank sequentially and either withdraw their deposit or
communicate that they do not currently wish to withdraw. Depositors have some information about
their likely position in the withdrawal order when they make this decision. The bank conditions the
payout to a withdrawing depositor on all previous reports: Subsequent withdrawals are adjusted
downward if a depositor makes an early withdrawal and upward if she instead reports that she will
not withdraw early. In this way, the payouts to depositors may be highly sensitive to the pattern of
withdrawal decisions. Green and Lin (2003) show that this sensitivity can be sufficient to rule out
bank runs as an equilibrium outcome. Peck and Shell (2003) consider the case where depositors
have no information about their place in the withdrawal order and assume that only those depositors intending to withdraw report to the bank. Under this alternative specification, they show that
bank runs can arise in equilibrium.3
In this paper, we propose an alternative environment that shares many features of these earlier
specifications but modifies some that we regard as extreme and perhaps unrealistic. In contrast
to Green and Lin (2003), we assume that depositors only report to the bank when they wish to
withdraw. In this respect, we follow Peck and Shell (2003), who state that “[i]t is hard to imagine
people visiting their bank for the purpose of telling them that they are not interested in making any
transactions at the present time.” Additionally, we depart from the previous literature by assuming that each depositor can observe how many withdrawals have already occurred when deciding
whether to withdraw. As a result, they can calculate exactly how much they will receive if they
choose to withdraw, instead of facing uncertainty that is only resolved after the withdrawal decision
has been made, as in both Green and Lin (2003) and Peck and Shell (2003).
Peck and Shell (2003) show that their result does not depend critically on whether depositors report to the bank
when they do not wish to withdraw; bank runs are possible in equilibrium under either specification. The key assumption is that depositors have no information about their position in the withdrawal order. Peck and Shell (2003) also
use a different specification of preferences from Green and Lin (2003), but Ennis and Keister (2009) show that this difference is not essential for their results; see also Sultanum (2014) and Bertolai et al. (2014).
3
3
We show that this new specification of the environment induces some properties in the banking
arrangement that resemble those observed in reality. In particular, the bank’s payout schedule is
initially almost completely insensitive to the pattern of withdrawal decisions and, in this sense,
displays the debt-like property that is common in demand deposit contracts. Furthermore, in situations where withdrawal demand becomes unusually high, depositors who are late to withdraw
suffer significant discounts relative to the face value of their deposits, a type of “partial suspension
of convertibility” that Wallace (1990) argues was a common feature of historical banking panics.4
We also show that the frictions in our environment restrict the flow of information sufficiently
to make a run on the bank consistent with equilibrium. The key ingredient for this result is that
when the bank only observes the actions of depositors who withdraw, it is relatively slow to react
to a situation in which aggregate withdrawal demand is high, as occurs when depositors run on
the bank. This slow reaction is anticipated by depositors and creates an incentive for those who
have the opportunity to be relatively early in the withdrawal order to withdraw regardless of their
liquidity needs. Based on these findings, we conclude that several common features of banking
– the face-value property of demand deposits, sharp discounts during crises, and fragility – may
have the same fundamental source: the gradual revelation of information inherent in a withdrawal
process that takes place sequentially.
The paper is organized as follows. Section 2 describes the environment, including our specification of sequential service and the information structure. Section 3 derives the efficient allocation of
resources in the absence of private information, which serves as a useful benchmark in the rest of
the analysis. Section 4 returns to the case of private information and describes a banking arrangement that aims to implement this allocation. Section 5 studies financial fragility and demonstrates
that a bank run can arise in our environment. Section 6 offers some concluding remarks.
2 The environment
There are two time periods, indexed by ∈ {0 1}, and a finite number of depositors. There is
a single good that can be consumed in each period. Let = (0 1 ) ∈ R2+ denote the consumption
of depositor in each period. A depositor’s preferences depend on her type ∈ {0 1} If = 0
the depositor is impatient and only cares about consumption in period 0. If = 1 the depositor is
For discussions of the specific features of historical suspension schemes, see Friedman and Schwartz (1963, pp.
160-8, 328-30), Selgin (1993), and Dwyer and Hasan (2003).
4
4
patient and cares about the sum of her consumption in the two periods. Depositor ’s utility level
is given by
¡
¢
0 + 1 =
¢1−
1 ¡ 0
+ 1
1−
(1)
As in Diamond and Dybvig (1983), we assume the coefficient of relative risk aversion is greater
than unity. Each depositor’s type is an independent draw from a Bernoulli distribution, where
is the probability of being impatient, and is private information. Let = (1 ) denote
the vector of types for all depositors, and let Ω denote the set {0 1}, so that we have ∈ Ω and
∈ Ω
We use () to denote the total number of patient depositors in the profile that is
() =
X
=1
Let () denote the probability of the set of profiles in which there are exactly patient depositors. Since types are independent across depositors, is a binomial random variable and we
have
() = ( ) (1 − ) −
(2)
where is the standard combinatorial function
( ) =
!
! ( − )!
(3)
There is also a bank that has an endowment of units of the good at the beginning of date 0
and aims to distribute these resources to maximize depositors’ expected utility.5 A depositor can
withdraw from the bank either in period 0 or in period 1 but not both. Depositors are isolated
from each other and goods must be consumed immediately after withdrawal. These assumptions
imply that no markets exist in which depositors could trade after withdrawing; a depositor simply
consumes what she receives from the bank (see Wallace, 1988, on this point). Each unit of the
good that is not consumed in period 0 is transformed into 1 units of the good in period 1
Depositors’ opportunities to withdraw in period 0 arrive sequentially in a randomly determined
fashion, with each depositor equally likely to occupy each position in the sequence. When her
opportunity arrives, a depositor is able to observe how many withdrawals have already been made.
It would be straightforward to add an earlier time period to the model in which individuals have private endowments
and choose whether to deposit in the bank, as in Peck and Shell (2003) and others. We bypass this step for simplicity.
5
5
She does not, however, observe the decisions of any depositors who have already chosen not to
withdraw in period 0. Likewise, the bank only observes depositors’ actions when they withdraw.
In other words, the bank always knows how many withdrawals it has already processed, but it has
no information about depositors who may have already chosen not to withdraw in period 0, nor
about depositors who have not yet had an opportunity to withdraw.
When a depositor withdraws, the amount she receives can depend only on the information
currently available to the bank. This sequential service constraint limits the ability of the bank
to make payouts contingent on the demand for early withdrawals. Consider, for example, the
first depositor to withdraw in period 0 When she arrives, the bank only knows that at least one
depositor is withdrawing; it has no other information that can be used to make inferences about the
total number of early withdrawals that will take place. The consumption this depositor receives
must, therefore, be the same for all outcomes in which at least one withdrawal occurs; let 1 denote
this amount. Similarly, let denote the amount of consumption received by the depositor to
withdraw in period 0, which must be the same for all outcomes in which at least early withdrawals
occur. The restrictions imposed by the sequential service constraint thus imply that the period-0
actions of the bank can be fully described using a payout schedule = { }=1 . This schedule
must satisfy the feasibility constraints
X
≤
and
≥ 0 for all
=1
(4)
Notice that the sequence is a contingent plan; it specifies the period-0 payouts the bank will
make in all possible scenarios, including the one where all depositors withdraw early. If fewer
than depositors withdraw in period 0 some of these payouts will not be made. After all depositors
have had an opportunity to withdraw in period 0 the bank observes that the period has ended and
the economy moves to period 1 At this point, the bank knows how many depositors have not yet
withdrawn. Since depositors are risk averse, efficiency requires that the bank divide the matured
assets in period 1 evenly among these depositors. As a result, the operation of the bank in our
environment is completely summarized by the period-0 payout schedule
Our formulation of the sequential service constraint differs from that in Wallace (1988, 1990)
and Green and Lin (2003). In those papers, each depositor contacts the bank in period 0 and announces whether or not she wishes to withdraw. In such a setting, the period-0 payout received
by a depositor can depend on the entire sequence of withdrawal decisions made up to that point.
6
We instead follow Peck and Shell (2003) in assuming that the bank only observes the decisions
of those depositors who have chosen to withdraw. This formulation slows down the flow of information about the level of total withdrawal demand to the bank and, in so doing, places stronger
restrictions on the payout schedule as described above. In the sections that follow, we show how
this slower flow of information to the bank has important implications for the form of optimal
banking arrangements and for financial fragility.
We also depart from most of the existing literature by assuming that each depositor is able to
observe the number of withdrawals that have already taken place when her opportunity to withdraw
arises.6 In Green and Lin (2003), a depositor observes a signal correlated with her position in the
sequence of withdrawal opportunities before making her decision. However, because she does not
observe actions taken by the depositors before her in this sequence, she will typically be uncertain
about the number of withdrawals that have already been made and, hence, about the amount she
would receive from the bank if she chose to withdraw. In Peck and Shell (2003), a depositor receives no information about her position in the sequence or the actions of other depositors. Since
the bank necessarily observes the actions of those depositors with whom it interacts, both of these
approaches imply that a depositor has less information than the bank about the withdrawal history.
This asymmetry raises the question of whether the bank might choose to communicate its information to a depositor before she makes her choice or whether a depositor could change her decision
after seeing the amount offered by the bank. (See Nosal and Wallace, 2009, for an analysis of
the former issue.) Under our approach, in contrast, depositors and the bank observe exactly the
same information about the withdrawal history; there is no scope for the bank to either provide or
withhold information about this history to depositors.
3 The efficient allocation
We begin our analysis by deriving the efficient allocation of resources in an environment that
is identical to the one just described, but where the preference type can be observed by the
bank when depositor has her opportunity to withdraw. This allocation will be a useful benchmark
in subsequent sections when we study the equilibria of a game played by depositors with private
information.
Andolfatto et. al (2007) extend the main result of Green and Lin (2003) in a setting where depositors observe
withdrawal decisions as they are made. Gu (2011a) studies informational cascades and herding in a banking model
in which depositors can observe the withdrawals of others as they occur, but does not study self-fulfilling runs as
we do here.
6
7
Given the form of preferences (1) and the return 1 on investment, efficiency clearly requires
that a depositor consume in period 0 if and only if she is impatient. When the bank can observe
types therefore, it will only permit impatient depositors to withdraw in the early period. Let
denote the amount of resources that the bank has remaining after serving depositors in period 0
under the payout schedule so that 0 = and
= −
X
for = 1 − 1
=1
The bank’s objective function can then be written as
!
à −1
à −
!
³ ´
X
X
X
−
+ (0)
()
( ) +
( ) + (−1 )
=1
=1
=1
(5)
(6)
The efficient payout schedule is the sequence {∗ } that maximizes (6) subject to the feasibility
constraints in (4) and the definition in (5). To solve for the efficient schedule, we reformulate (6)
as a dynamic programming problem. Define the following conditional probabilities:
= Prob [ − () ≥ | − () ≥ − 1]
After the bank has encountered − 1 impatient depositors in period 0 is the probability that it
will meet at least one more. Using the binomial density in (2), this conditional probability can be
written as
=
−
P
()
=0
−+1
P
()
=0
We can then derive the efficient payout schedule as a function of these probabilities
Proposition 1 The efficient payout schedule sets
∗
=
∗
−1
1
for = 1
( ) + 1
where the sequence {∗ } is defined by using {∗ } in (5) and { } is defined recursively by = 0
and
´
³
1
= +1 +1 + 1 + (1 − +1 ) ( − ) 1−
for = 1 − 1.
8
The proposition shows how the fraction of the remaining resources −1 that the impatient
depositor will receive depends on the remaining conditional probabilities +1 +2 etc., as well
as on the parameters and A proof of the proposition is given in the appendix .
Example. Figure 1 plots the efficient payout schedule for the parameter values = 11 = 6
and = 05 when there are 20 depositors (in panel a) and 200 depositors (in panel b). The lower
curve in each panel represents, for each value of , the consumption ∗ that the impatient
depositor will receive in period 0 The upper curve represents the level of consumption that all
patient depositors will receive in period 1 if there is a total of − 1 impatient depositors. The fact
that this latter curve lies everywhere above the former has the following interpretation. Consider
the last depositor in the sequence of period-0 withdrawal opportunities, and let the number of
impatient depositors before her be given by − 1 If she is impatient, she will receive ∗ , from
the lower curve in the figure; if she is patient, she will receive the consumption allocated to patient
depositors when there are a total of −1 impatient depositors, which is the corresponding point on
the upper curve. The figure shows that the last depositor in the sequence always consumes more
when she is patient than when she is impatient, regardless of the types of the other depositors.
Notice that this feature does not necessarily hold for other depositors. The first depositor in the
sequence, for example, consumes 1 if she is impatient and can end up receiving any point on
the upper curve if she is patient, depending on the total number of impatient depositors. In panel
(a), for example, she will consume more than 1 when she is patient if the number of impatient
depositors turns out to be less than 12 but otherwise she will consume less.
Using the solution given in Proposition 1, we can derive some properties of the efficient payout
schedule. First, we establish that this schedule offers depositors liquidity insurance in the sense
that the benefit of the return on investment is shared by all depositors, even those who consume
before this return is realized. In a model with no aggregate uncertainty, Diamond and Dybvig
(1983) showed how 1 implies that the efficient level of consumption for all impatient depositors is greater than the per-capita value of the bank’s assets in period 0 When there is aggregate
uncertainty, as in our model here, the definition of liquidity insurance must be adjusted because
different impatient depositors consume under different circumstances. In this case, we say that a
payout schedule offers liquidity insurance to the impatient depositor if
−1
−+1
9
(7)
Figure 1: The efficient payout schedule
that is, if the amount given to this depositor is larger than the value of the bank’s remaining resources divided by the number of depositors who have not yet withdrawn. The following result
shows that ∗ satisfies this expression for all values of up to − 1. A proof is given in the
appendix.7
Proposition 2 The efficient payout schedule ∗ offers liquidity insurance in the sense of (7) for
= 1 − 1
This result has important implications for the shape of the efficient payout schedule as depicted
in the lower lines of Figure 1. When more depositors consume in the early period, before the
return is realized, providing liquidity insurance for this group becomes more costly in terms of
lowering the consumption of patient depositors as a group. The bank’s efficient response to an
increase in the likely number of impatient depositors is, therefore, to decrease the consumption of
all remaining depositors. As additional withdrawals take place in period 0 the bank becomes more
pessimistic about the total number of impatient depositors. As shown in the figure, the bank adjusts
to this new information by setting the payout ∗+1 lower than ∗ The following proposition shows
that this monotonicity is a general feature of the efficient payout schedule.
Note that Proposition 2 only applies for ≤ −1 because the feasbility constraint (4) implies that the payout made
to the last depositor if everyone is impatient, cannot be greater than the bank’s remaining resources. In other words,
it is infeasible for condition (7) to hold for =
7
10
Proposition 3 The efficient payout schedule is strictly decreasing, with ∗+1 ∗ for =
1 − 1
Proof: From Proposition 1, we know that the efficient payout schedule satisfies
∗ =
∗
−1
and
1
∗
∗+1 =
+ 1
1
+1 + 1
where
∗
∗ = −1
− ∗
Combining these expressions yields
1
∗+1
=
∗
1
+1 + 1
Using the second inequality in Lemma 1 (in the appendix), it follows immediately that ∗+1 ∗
¥
for = 1 − 1
Figure 1 also indicates that the schedule ∗ is initially quite flat, with ∗1 very close in value to
∗2 ∗3 and several more payouts. Only as the number of early withdrawals becomes larger do the
downward adjustments in ∗ become visible. In other words, depositors withdrawing in period 0
are initially treated similarly, each receiving close to what might be considered the “face value”
of their deposit. This property emerges from the fact that, initially, withdrawals provide relatively
little information to the bank about the total number of impatient depositors and, hence, have little
impact on the efficient payout.
To understand this property better, consider the example in panel (a) of the figure. The total
number of impatient depositors in this example is likely to be close to 10 since each of the 20
depositors has a 12 probability of being impatient. When the first early withdrawal takes place,
the bank learns that at least one depositor is impatient, which rules out the (extremely unlikely)
event that all 20 depositors are patient. The bank’s belief about the distribution of is only slightly
changed by this information, which leads it to set the payout for the next impatient depositor, ∗2
only slightly different from ∗1 More generally, the continuation probability +1 is very close to
1 for small values of which implies that the bank is almost certain that additional withdrawals
will be made. Since depositors are risk averse, the efficient plan will approximately equalize the
values of all payments that are nearly certain to be made. As the number of depositors increases,
11
the information content of an additional withdrawal when is small decreases because the fraction
of depositors who are patient is less likely to deviate significantly from its mean. As a result, the
payout schedule is initially flatter when is larger, as illustrated in panel (b) of the figure.8
As the proportion of the population that has withdrawn in period 0 increases, additional withdrawals become more informative about the value of After 10 withdrawals have taken place in
panel (a), for example, the continuation probability 11 has fallen to 07. At this point, there is significant uncertainty about whether or not an additional early withdrawal will be made. If another
depositor arrives to withdraw, the bank’s belief about changes significantly and, as a result, the
payout ∗11 is noticeably lower than ∗10 . As illustrated in the figure, these changes become larger
and larger as increases toward the total number of depositors This part of the curves resembles what Wallace (1990) calls a “partial suspension of convertibility,” in which some depositors
receive significantly less than the face value of their deposits when the realized demand for early
withdrawal is high.
4 Banking
We now return to the environment in which preference types are private information. In this
setting, the only way for the bank to make a depositor’s consumption contingent on is to allow
her to choose the period in which she withdraws. We consider the following banking arrangement:
Each depositor chooses whether to withdraw early or late based on her own preference type and on
the number of withdrawals that have already been made when her opportunity arrives, and the bank
makes payouts according to the schedule ∗ derived in Proposition 1. This arrangement creates a
withdrawal game for depositors; the remainder of the paper is devoted to studying the equilibria
of this game.9
4.1
The withdrawal game
A depositor chooses a withdrawal strategy, which assigns a withdrawal period (either 0 or 1) to
each combination of her preference type and the number representing the opportunity to
Sultanum (2014) derives the efficient payout schedule in an environment with aggregate uncertainty and a continuum of depositors. Under some specifications, this schedule is initially very flat (see his Figure 2 on p. 99)
for the same reason ours is here.
9
In general, a withdrawal game can be defined based on any payout schedule Our focus here, however, is exclusively on the properties of the withdrawal game generated by the efficient payment schedule ∗
8
12
make the withdrawal in period 0,
= Ω × {1 } → {0 1}
We use to denote a profile of strategies, one for each depositor, and − to denote the strategies
of all depositors except
Our interest is in the Bayesian Nash equilibria of the withdrawal game, where each depositor
chooses the strategy that maximizes her expected utility while correctly anticipating the strategies − of other depositors.10 We study symmetric equilibria, in which all depositors follow the
same strategy. Different depositors may still take different actions, of course, as their preference
types and withdrawal opportunities will differ.
As described above, impatient depositors only care about consumption in period 0 and, hence,
the individual best response to any strategy profile − will satisfy (0 ) = 0 for all In other
words, we only need to check what a depositor will choose to do when she is patient.
Consider the decision faced by a patient depositor who has an opportunity to make the withdrawal in period 0. If she chooses to withdraw, she will receive ∗ . If she waits, the amount she
receives at = 1 will depend on the total number of withdrawals that take place in period 0 which
is not yet known. Let b
denote the number of depositors who wait until = 1 to withdraw in this
case, including herself. Then − b
will be the total number of early withdrawals. Note that b
is a random variable that depends on both the number of patient depositors and the profile of
³
´
withdrawal strategies followed by the other depositors, − Let b
; − denote the posterior
probability this depositor, using Bayes’ rule, assigns to the event that exactly b
depositors (including herself) wait until = 1 to withdraw. Given this belief, she computes the expected utility of
waiting to withdraw, denoted (; − ) as
³
´ µ ¶
; − −
b
(; − ) ≡
b
=1
X
(8)
An equilibrium of the withdrawal game is a strategy profile ∗ satisfying
¡
¢ ¾
½
∗
0
if = 0 or (∗ ) ;
∗
−
¢
¡
for all for all
( ) =
∗
1
if = 1 and (∗ ) ; −
While the structure of the game implies that there are sequential moves by the players, it is not hard to see that
the extensive form representation of the game has no proper subgames.
10
13
In other words, each depositor is optimally choosing when to withdraw given her beliefs about the
total number of early withdrawals, and those beliefs are consistent with the actions prescribed in the
equilibrium strategy profile ∗ Impatient depositors always withdraw early, and patient depositors
withdraw early when the value of the available payout, (∗ ) exceeds the expected value of
¡
¡
¢
¢
∗
∗
waiting, ; −
Note that when a patient agent is indifferent, because (∗ ) = ; −
either action ( = 0 or = 1) is consistent with equilibrium.
4.2
Incentive compatibility
One strategy of particular interest is the no-run strategy, in which a depositor withdraws early if
and only if she is impatient
0 ( ) = for all
(9)
Let 0 denote the no-run strategy profile. If 0 is an equilibrium of the withdrawal game, this
equilibrium achieves the efficient allocation of resources derived in Section 3.
In this case, the number of depositors who wait to withdraw, b
is the same as the number
of patient depositors A depositor’s initial belief about is given by (2). To illustrate how the
¡ 0 ¢
posterior beliefs ; −
are formed, suppose that = 1 meaning that no withdrawals have
occurred yet. This situation could arise because the depositor is first in the sequence of withdrawal
opportunities, in which case it would convey no information about . However, it also could arise
because the depositor is later in the sequence but all of the depositors before her were patient, in
which case is likely to be high. The depositor weighs the relative likelihood of these different
situations in updating her belief about according to Bayes’ rule.
More generally, suppose a patient depositor observes that − 1 withdrawals have already been
made when her opportunity to withdraw arrives. Define −1 to be the set of all type profiles in
which there is a patient depositor in the position with exactly − 1 impatient depositors in the
first − 1 positions, that is,
−1 =
(
: = 1 ∧
−1
P
=1
)
(1 − ) = − 1
The depositor knows that the realized profile must lie in this set for some value of . The
following proposition derives her posterior belief about the number of patient depositors ; a proof
of the result is given in the appendix.
14
Proposition 4 A patient depositor who anticipates that all other depositors are following (9) and
who has the opportunity to make the withdrawal in period 0 will assign probability
¶
µ
P
P
()
I{∈ −1 }
¡ 0 ¢
=
∈{:()=}
¶
µ
; − = (; ) ≡
P
P
()
I{∈ −1 }
=
∈Ω
to the event that exactly of the depositors are patient, where I{} is the indicator function for
the set
¡ 0 ¢
¡
¢
0
we can obtain the value of ; −
from equation (8). The
Using the distribution ; −
no-run strategy profile is an equilibrium of the withdrawal game if and only if:
¡
¢
0
(∗ ) ≤ ; −
for = 1
(10)
The left-hand side of these inequalities represents the value of withdrawing early and receiving
the payout ∗ for sure. The right-hand side is the expected utility of waiting to withdraw when all
other depositors are following (9). When these inequalities are satisfied for all values of a patient
depositor will find it optimal to withdraw in period 1 regardless of the number of withdrawals that
have taken place when her opportunity arrives in period 0. As a result, there is an equilibrium of
the withdrawal game in which all depositors follow the strategy (9) and the efficient allocation of
resources obtains. In this case, we say that the efficient allocation is incentive compatible.
Our focus in the rest of the paper is on situations in which the efficient allocation is incentive
compatible. We follow an approach that is common in the literature by first solving for the efficient
payout schedule ∗ without imposing the incentive compatibility conditions, using Proposition 1,
and then verifying that the solution satisfies (10). Note that the constraints imposed by private information are not binding on the efficient allocation of resources in these cases. This fact illustrates
the benefits of banking arrangements that resemble demand deposit contracts, in which depositors
are allowed to choose when to withdraw their funds from the bank. As in Diamond and Dybvig
(1983) and others, such an arrangement allows the bank to potentially achieve the same allocation
of resources it would choose if it could directly observe each individual depositor’s consumption
preferences when she arrives. However, such arrangements may also open the door to financial
fragility in the sense of admitting other equilibria in which some patient depositors “run” on the
bank and withdraw in period 0 leading to an inferior allocation. In the next section, we investigate
whether such fragility can arise in our model and what forms it may take.
15
5 Fragility
In this section, we investigate whether the withdrawal game defined above can have other
Bayesian Nash equilibria, in which some depositors choose to withdraw in period 0 even when
they are patient. If such equilibria exist, the bank is fragile in the sense that attempting to implement the efficient allocation of resources using the arrangement described above could lead to
an inefficient outcome that resembles a run on the bank. We first show that there cannot be an
equilibrium in which all depositors attempt to withdraw early. We then construct an example of a
partial run equilibrium, in which some patient depositors withdraw early but others do not.
5.1
No full-run equilibrium
The type of bank run studied in most of the literature has all depositors attempting to withdraw
their funds in the early period.11 We refer to this strategy profile,
( ) = 0 for all for all
(11)
as a full run. It is fairly easy to see that a full run cannot be part of an equilibrium in our model.
Consider a depositor with = meaning that all of the other − 1 depositors have already
withdrawn when her opportunity to withdraw in period 0 arises. If she chooses to withdraw in
∗
period 0, she will receive all of the bank’s remaining resources, −1
If she waits until period 1 to
∗
withdraw, however, she will receive the matured value of these resources −1
Because 1 a
patient depositor in this situation would always strictly prefer to wait to withdraw. In other words,
the full-run strategy (11) is strictly dominated by a strategy with (1 ) = 1 and, hence, cannot
be part of an equilibrium. The following proposition records this result.
Proposition 5 There is no equilibrium in which depositors play the strategy profile (11).
This aspect of our model is similar in spirit to one of the key features of Green and Lin (2003).
In their setting, depositors receive a signal about their position in the sequence of withdrawal
opportunities. If a depositor’s signal indicates that she is very likely to be last in this sequence, she
faces a decision that is similar to the one described above and will strictly prefer to wait until period
See, for example, Diamond and Dybvig (1983), Cooper and Ross (1998), and Peck and Shell (2003). Some
recent exceptions are Ennis and Keister (2009, 2010a), Gu (2011b), Azrieli and Peck (2011), Keister (2014), and
Keister and Narasiman (2015).
11
16
1 if she is patient. Depositors in our model do not observe their place in the sequence; however,
a depositor who observes that everyone else has already withdrawn can readily infer she is in the
last position. For this reason, the Green-Lin logic for this last depositor applies in our setting as
well and a full-run equilibrium cannot exist. Green and Lin (2003) then use a backward-induction
argument to show that the no-run strategy profile is the unique equilibrium in their setting; no type
of run equilibrium can exist. In the next subsection, we show this stronger result does not hold in
our environment.
5.2
Partial-run equilibria
Proposition 5 shows that if a run equilibrium exists in our environment, it must be partial, with
only some depositors participating. One possibility is for depositors to run until the total number
of withdrawals reaches some critical level and then for the run to stop. Consider the following
class of strategy profiles
̄
( ) =
½
0
≤ ̄
for
̄
for some 1 ≤ ̄ ≤ − 1
for all
(12)
Note that setting ̄ = 0 would correspond to a no-run strategy profile in (9), while setting ̄ =
would correspond to a full-run strategy profile in (11). For values of ̄ between 1 and − 1 this
profile represents a partial run.
Recall that b
denotes the total number of depositors who wait until period 1 to withdraw. Under
a partial-run strategy profile, b
is bounded above by min[ − ̄] The realized value of b
will be
less than the number of patient depositors if one or more of the first ̄ depositors in the decision
order is patient and, following (12), withdraws in period 0
To determine if (12) is consistent with equilibrium for some positive value of ̄, we need to
compare the expected utility a depositor receives from following this strategy to the expected utility
associated with deviating. In particular, for ≤ ̄ we need to consider whether a patient depositor
would be better off waiting until period 1 to withdraw, while for ̄ we need to consider whether
a patient depositor would be better off withdrawing in period 0 We address these two cases in turn.
(i) ≤ ̄ : If the depositor follows the strategy in (12), she will withdraw early and receive ∗ . If
she deviates, she will receive the period-2 payout associated with b
depositors waiting to withdraw.
Comparing the expected utility of these two outcomes requires deriving the depositor’s belief about
b
under the assumption that () she is patient, () all other depositors follow the strategy profile
17
in (12), and () she deviates from (12) and waits until period 1 In this case, the value of fully
reveals the depositor’s position in the sequence of withdrawal opportunities: She must be the
depositor to decide. It does not, however, give her any information about the profile of types .
The depositor knows that at least ̄ depositors will withdraw early, and that the remaining − ̄−1
depositors will each withdraw early if and only if they are impatient.12 Her posterior belief about
b
is then given by
and
³
´
̄
b
; − = 0
for b
= 0 or b
− ̄ and ≤ ̄
³
´
³
´
̄
; −
= − ̄ − 1 b
− 1 −̄− (1 − )−1
b
(13)
(14)
for b
= {1 − ̄} and ≤ ̄
where is the combinatorial function (3). In other words, the depositor knows that b
will be at
least 1 (herself) and can be up to − ̄ if all of the other depositors who do not participate in the run
³
´
̄
b
are patient. The types of these other depositors are i.i.d. Bernoulli trials. Notice that ;
−
is independent of for all ̄; the depositor’s precise position in the sequence of withdrawals
does not affect her posterior belief about b
in this case.
¡
¢
̄
The expected utility of deviating from (12) and withdrawing in period 1 is given by ; −
³
´
̄
b
as defined in (8), using the distribution ; − presented in (13) – (14). A necessary condition
for (12) to be consistent with equilibrium is that
¢
¡
̄
(∗ ) ≥ ; −
for ≤ ̄
¢
¡
̄
is independent of for ≤ ̄ while Proposition 3 establishes that (∗ ) is
Notice that ; −
strictly decreasing in . Therefore, the condition above will be satisfied for all ≤ ̄ if and only
if it is satisfied for ̄, that is,
¡
¢
̄
(∗̄ ) ≥ ̄; −
(15)
When this condition holds, a depositor who expects all other depositors to follow the strategy in
(12) and observes a value of less than or equal to ̄ will also be willing to follow (12). In other
Notice that the ̄ withdrawals that constitute the partial run will take place regardless of whether this individual
depositor chooses to participate in the run. If she deviates from (12), other depositors will continue the run until
̄ is reached.
12
18
words, (15) is a necessary condition for the strategy profile ̄ to be an equilibrium. We record this
result, which is useful for constructing the examples below, as a proposition.
¡
¢
̄
Proposition 6 If the strategy profile ̄ is an equilibrium, then (∗̄ ) ≥ ̄; −
(ii) ̄ : When the depositor’s opportunity to withdraw arrives later, after the run has ended, she
can no longer be certain about her position in the sequence of withdrawal opportunities. Instead,
she must form beliefs about how many of the − ̄ no-run depositors are impatient and what her
position within the sequence of these depositors might be. This inference problem is similar in
structure to the one discussed in the context of incentive compatibility in Section 3. In this case,
the depositor’s posterior belief about b
is given by
³
´
³
´
̄
b
; −
= −̄ b
; − ̄
for b
= {1 − ̄} and ̄
(16)
³
´
The expression −̄ b
; − ̄ is the probability distribution presented in Proposition 4 applied
to the subset of − ̄ depositors who do not participate in the run. Note that the index on this
distribution is adjusted from to − ̄ , since the depositor in question observes − 1 − ̄
withdrawals from the − ̄ no-run depositors.
A patient depositor following the strategy in (12) would wait until period 1 to withdraw and
³
´
¡
¢
̄
̄
would have expected utility given by ; −
; −
now
from (8), with the distribution b
given by (16). If she instead deviates and withdraws early, she will receive ∗ for sure. The
strategy profile ̄ is consistent with a best response for these depositors if
¡
¢
̄
(∗ ) ≤ ; −
for = ̄ + 1
(17)
If conditions (15) and (17) both hold for some ̄ with 1 ≤ ̄ ≤ − 1, the partial-run strategy
profile (12) based on that value of ̄ is a Bayesian Nash equilibrium of the withdrawal game. Our
main result in this section is that, for some parameter values, such a partial run equilibrium exists.
Proposition 7 There exist parameter values such that the efficient payout schedule ∗ both ()
generates an incentive compatible allocation and () admits a partial run equilibrium with 1 ≤
̄ ≤ − 1
The proof is by example. We present the example in the next subsection, followed by a detailed
discussion of the intuition behind it.
19
5.3
An Example
Our example is based on the same parameter values that were used to illustrate the efficient allocation in panel (a) of Figure 1: = 20, = 11 = 6 and = 05.13 Figure 2 plots three
curves, two of which are simply the utility associated with the consumption levels plotted in the
¡
¢
earlier figure. The third curve is an auxiliary function ; −
This curve gives the expected
utility from waiting to withdraw for a patient depositor who has the opportunity to make the
withdrawal in period 0 under the assumption that all of the depositors before her are running but
all of the depositors after her are not. In other words, this curve measures the value of deviating
from the strategy profile for the last depositor who participates in this partial run.
Figure 2: A partial run equilibrium
¡
¢
̄
Note that the necessary condition identified in Proposition 6 – that (∗̄ ) ≥ ̄; −
hold
– is satisfied for any potential cutoff value ̄ between 7 and 16. Let us concentrate attention on
the partial run strategy with the largest possible threshold, in this case ̄ = 16. The fact that the
necessary condition holds implies that depositors with ≤ 16 will all chose to play according
to the partial run strategy profile 16 Hence, to show that 16 is an equilibrium strategy profile,
we only need to show that depositors with 16 will also prefer to follow this strategy, which
specifies that they wait until period 1 if patient. This is done in Figure 3, where the upper curve
plots the incentive to run
¡
¢
16
(∗ ) − ; −
We focus on panel (a) rather than panel (b) because the computational burden of calculating the probabilities
( − ) defined in Proposition 4 increases rapidly with the number of depositors.
13
20
for all between 1 and 20.14 The discussion above demonstrated that this value must be positive
for between 1 and 16; this fact is also reflected in Figure 3. The new information in the figure is
that the expression is negative for = 17 through 20 which verifies that a patient depositor who
has any one of these opportunities to withdraw in period 0 would choose to wait, in accordance
with the candidate equilibrium strategy profile. Hence, the figure establishes that strategy profile
(12) with ̄ = 16 comprises an equilibrium of the withdrawal game.
Figure 3: Individual incentive to run
Figure 3 also verifies that the efficient allocation is incentive compatible in this example. The
lower curve in the figure plots
¡
¢
0
(∗ ) − ; −
;
that is, the gain in expected utility from running for each value of under the assumption that all
other depositors follow the no-run strategy. The fact that this line is negative everywhere demonstrates that the no-run strategy profile is also an equilibrium of the withdrawal game. Together, the
two curves in Figure 3 thus establish the result in Proposition 7. It is worth emphasizing that there
is nothing special about the parameter values used in this example; it is easy to construct similar
examples using a wide range of parameter values.
Notice that the term in this curve is different from that plotted in Figure 2 Here, the strategies of other depositors
16
are being held fixed at −
as varies from 1 to 20
14
21
5.4
Intuition
To gain intuition for why the partial-run strategy profile in (12) is an equilibrium, it is useful to
examine the behavior of the “critical” depositor whose opportunity to withdraw in period 0 is ̄.
This depositor follows the run strategy even though she believes that no one after her will run.
This type of behavior is inconsistent with equilibrium under Green and Lin’s (2003) formulation
of the sequential service constraint, a fact that is crucial for their backward-induction argument.
Understanding why it can arise here thus illustrates a key difference between our environment and
theirs.
Suppose we compare this depositor’s equilibrium belief about the number of early withdrawals
with the beliefs used by the bank in designing the payout schedule ∗ . The efficient payout ∗16
depends on the probability that there will be a 17 early withdrawal (and an 18 , and so on).
Given that the bank expects only impatient agents to withdraw early, and each depositor has an
independent, one-half probability of being impatient, the bank considers a 17 early withdrawal
fairly unlikely. When faced with a 16 early withdrawal, the bank believes that the four depositors
it has not yet seen are likely to have already had their opportunities to withdraw and decided to
wait because they are patient. In this sense, the bank is “optimistic” that the 16 early withdrawal
will be the last one and the payout ∗16 is chosen based on this optimism.
In the partial run equilibrium, however, the depositor making the 16 withdrawal recognizes
that a run is underway and the number of early withdrawals is thus likely to be much larger than the
number of impatient depositors. Importantly, she knows that four depositors have not yet had the
opportunity to contact the bank in period 0. She expects that, on average, two of these depositors
will be impatient and, hence, she realizes that the early withdrawals are unlikely to end with her.
Any additional withdrawals will further deplete the bank’s resources, lowering the consumption
she would receive if she were to wait and withdraw in period 1. Her more pessimistic belief about
the number of early withdrawals thus makes running – and accepting the payout based on the
bank’s optimistic belief – an attractive strategy.15
Notice how the limited flow of information implied by our formulation of the sequential service
It is easy to see why the incentive to run for depositors making withdrawals before the 16 are even stronger,
as indicated in Figure 3. The payment that the depositor making the first withdrawal receives, for example, is based on
the expectation that the number of early withdrawals will be, on average, around 10. In equilibrium, however, this
depositor anticipates that there will be at least 16 and on average 18 early withdrawals. This large gap in beliefs
makes running attractive.
15
22
constraint makes this divergence in beliefs possible. When a depositor makes the 16 withdrawal,
the bank does not know that she is the 16 depositor to have an opportunity to withdraw. In fact,
the payout schedule ∗ is based on the belief that if a 16 impatient depositor withdraws, she is
likely to be the last depositor to decide. In Green and Lin (2003), in contrast, patient depositors are
expected to announce to the bank that they will not withdraw when their turn in the order comes.
Under this specification, both the bank and the depositor will know that there are four depositors
who still have an opportunity to contact the bank in period 0. Since types are independent, if these
depositors will report truthfully, then the bank and the depositor must have the same belief about
the number of additional early withdrawals regardless of what strategies the earlier depositors
have followed.16 Because of this agreement in beliefs, the payout offered to each depositor in the
Green-Lin setup appears “appropriate” given her beliefs and, as a result, she will choose to report
truthfully. This reasoning is central to the unique implementation result in Green and Lin (2003).
In contrast, the divergence in beliefs in our model arises naturally whenever one or more depositors
follow a non-truthful strategy. The example presented here shows how this divergence in beliefs
can be strong enough to generate a run equilibrium in the withdrawal game.
5.5
Discussion
In earlier work (Ennis and Keister, 2009), we showed that partial run equilibria can arise in the
model with Green and Lin’s (2003) formulation of the sequential service constraint when depositors’ preference types are correlated. The intuition behind this earlier result is similar to that
described for our example above. In particular, in both cases there is a critical depositor who
chooses to run even though she expects everyone after her to report truthfully. The payout to this
depositor is designed based on the expectation that her withdrawal will likely be the last one in
period 0, while the critical depositor believes that additional early withdrawals are very likely.
As described above, this divergence in beliefs makes withdrawing early, and receiving the payout
based on the more optimistic belief, an attractive choice. The difference in beliefs was generated
by the correlation structure of types in our earlier work, while in the present paper it is generated
by the fact that the bank does not observe the actions of depositors who choose not to withdraw.
This insight about the importance of divergent beliefs is likely to prove useful for studying
fragility in other environments. Suppose for example, that both the bank and depositors observe
16
On this point, see also Andolfatto et al. (2007) and Nosal and Wallace (2009).
23
some measure of the “time” that elapses between two withdrawals and use this information to make
inferences about the number of depositors who have already chosen not to withdraw. Depending
on the distribution of depositors’ arrival times at the bank, agents may or may not have significantly
more information on which to base their decisions in such a setting. The efficient allocation would
be much more complex: A depositor’s consumption would in general depend not only on her
position in the withdrawal order, but also on the precise time she arrives at the bank and the times
at which all previous withdrawals were made. A banking arrangement that implements this type of
allocation might bear less resemblance to a standard demand-deposit contract than the arrangement
we study here. Nevertheless, it seems likely that the degree to which depositors’ beliefs about
the number of additional early withdrawals during a run diverge from those used by the bank in
designing payouts will be critical for determining whether a run equilibrium exists. Extending our
analysis to such alternative environments is an interesting avenue for future research.
One feature our model shares with much of the existing literature is the fact that a run is observationally equivalent to an event that occurs with positive probability under the no-run strategy
profile (that is, an event in which many depositors are impatient). In other words, the bank can
never actually be sure that a run is taking place, even ex post.17 There is, however, a difference in
how quickly the bank is able to infer that something unusual is happening. In our setting, a run
is initially equivalent to almost all events under the no-run strategy profile. Suppose depositors
follow the partial-run strategy profile in (12). As the first few depositors arrive to withdraw, the
bank is only able to infer that at least a few depositors are withdrawing, an event that is very likely
to happen under the no-run strategy profile as well. In the Green and Lin (2003) specification of
the sequential service constraint, in contrast, the bank will observe the fact that no depositors are
choosing to wait to withdraw, which is unlikely to occur if depositors are following the no-run
strategy profile. Relative to the Green-Lin approach, our specification slows down the flow of information about depositors’ strategies to the bank. This slower flow of information makes the bank
slower to react to a surge in early withdrawals, which in turn tends to increase the incentive for
individual depositors to participate in a run.
Our approach of allowing a depositor to observe the number of withdrawals that have already
been made when her opportunity to withdraw arrives gives the withdrawal game a more dynamic
Ennis and Keister (2010a) present a model based on limited commitment in which a partial run may occur and
the bank is eventually able to infer that a run is underway and react to this information.
17
24
flavor than much of the existing literature. It also potentially opens the door to issues of signaling,
in which a depositor takes into account the effect her action will have on the actions of subsequent
depositors (see Andolfatto et al., 2007). Such effects play a minimal role in our setting, however,
because depositors only observe withdrawals and receive no information when another depositor
chooses not to withdraw. Under the strategy profile ̄ , for example, there is no way for the
depositor with the first opportunity to withdraw to “signal” that she is deviating from this profile. If
she chooses not to participate in the run, this action is unobserved and the next depositor, observing
that no withdrawals have taken place yet, will incorrectly infer that she has the first opportunity to
withdraw. Similarly, if a patient depositor deviates from the no-run strategy profile by withdrawing
early, other depositors will infer that she was impatient rather than suspecting a deviation. For these
reasons, many of the complications typically associated with dynamic games do not arise in our
setting.
We have followed a common approach in the literature by assuming the bank makes payouts
according to the efficient schedule ∗ and asking whether the resulting withdrawal game admits
a run equilibrium. If the bank placed positive probability on the event of a run, it might follow a
different plan. For example, it might choose a payout schedule that is less efficient than ∗ when
depositors follow the no-run strategy profile 0 but mitigates the effects of a run if one occurs. If
the probability of a run is large enough, the bank may further alter the payout schedule to make
it “run proof” in the sense that the only equilibrium strategy profile is 0 (see Cooper and Ross,
1998, and Ennis and Keister, 2006). Alternatively, the bank might consider other, more complex
contractual arrangements that attempt to solicit additional information from depositors about play
in the withdrawal game (see, for example, Andolfatto et al., 2014, and Cavalcanti and Monteiro,
2011). In this sense, our result that the withdrawal game based on the efficient payout schedule
∗ has a (partial) run equilibrium can serve as a starting point for further investigation of financial
fragility and its implications in the environment we have introduced here.
6 Concluding Remarks
We study a model of financial intermediation in the tradition of Diamond and Dybvig (1983)
with some novel features. In our environment, a depositor’s actions are observed by both the bank
and other depositors, but only when she withdraws. We show this specification of sequential ser-
25
vice is tractable and generates results that are intuitive and perhaps more realistic than those in the
existing literature. Depositors, for example, are able to learn before making a decision what payout
they would receive by withdrawing. Moreover, even though our approach allows for very complex
patterns of payouts, the optimal arrangement resembles in many ways a traditional banking contract. Under this arrangement, depositors who withdraw early in the course of events all receive
approximately the face value of their deposits. If the number of withdrawals becomes unexpectedly large, however, depositors begin experiencing significant discounts in what they receive from
the bank.
In addition to highlighting the debt-like features of the optimal contract, we show that this
arrangement is fragile in the sense of being susceptible to a self-fulfilling run by depositors. A run
in our setting is necessarily partial, with only some depositors participating. The previous literature
has shown that in environments similar to ours, fragility may not arise if contractual arrangements
are sufficiently flexible. Our specification of the environment suggests that the debt-like features
of banking contracts and financial fragility may have a common origin in the gradual revelation of
information that is inherent in banking arrangements.
Appendix A. Proofs
Proposition 1: The efficient payout schedule sets
∗
=
∗
−1
1
for = 1
( ) + 1
where the sequence {∗ } is defined by using {∗ } in (5) and { } is defined recursively by = 0
and
for = 1 − 1.
³
´
1
= +1 +1 + 1 + (1 − +1 ) ( − ) 1−
Proof: Let denote the sum of the expected utilities of all depositors who have not yet consumed
when the bank encounters the impatient depositor, conditional on the bank dividing the available resources −1 efficiently among these depositors. These values must satisfy the following
26
recursive equation:
(−1 ) = max
{ }
for = 1
⎧
⎪
⎪
⎨
( )1−
1−
+ +1 +1 (−1 − ) +
⎫
⎪
⎪
⎬
³
´1− ⎪
⎪
⎪
⎪
⎩ (1 − +1 ) ( − ) 1 (−1 − )
⎭
1−
−
(18)
If all depositors are impatient, the bank will give all of the remaining resources to the last
depositor when she reports. We therefore have the following terminal condition
(−1 ) =
1
(−1 )1−
1−
The combination of this equation, the initial condition 0 = and equation (18) constitutes the
dynamic programming problem whose solution gives the efficient payout schedule.
Consider the decision problem faced by the bank if it faces an ( − 1) impatient depositor.
Given −2 the maximization problem in (18) reduces to
max
{−1 }
( (−2 − −1 )) 1−
(−1 )1−
(−2 − −1 )1−
+
+ (1 − )
1−
1−
1−
The solution to this problem sets
where
−2
−1 = ¡
¢ 1
−1 + 1
−1 ≡ + (1 − ) 1−
(19)
Substituting the solution back into the objective function and doing some straightforward algebra
yields the value function
−1 (−2 ) =
´
¢1
(−2 )1− ³¡
−1 + 1
1−
The function −1 captures the utility of the last two depositors to report to the bank in the event that
at least − 1 depositors are impatient. In this case, the ( − 1) depositor to report is necessarily
impatient. The may also be impatient, reporting in period 0 or patient, in which case she will
report in period 1 The probabilities of these events (given by ) are contained in the constant
−1
27
It is straightforward to use this same procedure to show that, for any the solution to the
maximization problem in (18) sets
−1
=
where
1
+ 1
³
´
1
= +1 +1 + 1 + (1 − +1 ) ( − ) 1−
Note that condition (19) emerges naturally from (20) using the “terminal” value = 0.
(20)
¥
Proposition 2: The efficient payout schedule ∗ offers liquidity insurance in the sense of (7) for
= 1 − 1
Proving this proposition requires showing that
∗ =
∗
−1
1
( ) + 1
∗
−1
−+1
or, equivalently,
1
−
holds for = 1 − 1 These inequalities are established as part of the following lemma:
Lemma 1: The inequalities
( − )
1−
1
1
+1 + 1 ≤ −
(21)
hold for = 1 − 1.
Proof of the lemma: The proof is by backward induction. First, consider the case of = − 1
Proposition 1 defines = 0 and
−1 = + (1 − ) 1−
Note that 1 and 1 implies 1− 1 Together with the fact that 0 1 these
definitions thus imply
1−
1
1
−1 + 1 = 1
which shows that (21) holds for = − 1
28
Next, suppose (21) holds for some ≤ − 1; we need to show that the set of inequalities then
also holds for − 1 Together with
1−
1 the first inequality in (21) implies
( − + 1)
1−
1
+ 1
From the definition of in Proposition 1, we have
³ 1
´
−1 = + 1 + (1 − ) ( − + 1) 1−
Since 0 1 these last two expressions imply
( − + 1)
1−
−1
or
( − + 1)
1−
³ 1
´
+ 1
1
1
−1 + 1
which establishes that the first two inequalities in (21) hold for − 1 as well. Turning to the third
inequality in (21), the assumption that this inequality holds for
1
+1 + 1 ≤ −
combined with the definition of in Proposition 1 implies
≤ +1 ( − ) + (1 − ) ( − ) 1−
+1 ( − ) + (1 − ) ( − )
= ( − )
The inequality ( − ) can be rewritten as
1
+ 1 − ( − 1)
which establishes that the third inequality in (21) also holds for − 1 as desired.18
18
Notice that this argument also establishes that the third inequality in (21) is strict for any ≤ − 2
29
¥
Proposition 4: A patient depositor who anticipates that all other depositors are following (9) and
who has the opportunity to make the withdrawal in period 0 will assign probability
¶
µ
P
P
()
I{∈ −1 }
¡ 0¢
=
∈{:()=}
µ
¶
; = (; ) ≡
P
P
()
I{∈ −1 }
=
∈Ω
to the event that exactly of the depositors are patient, where I{} is the indicator function for
the set
Proof: Let denote the depositor’s position in the sequence of withdrawal opportunities in period
0; recall that this position is not observed by the depositor. The depositor is concerned with two
independent random events: the ordered sequence of preference types and her position in the
order, which is drawn from the uniform distribution on {1 } The probability of a pair ( )
is given by
( ) =
()
where
() = (1 − )() −()
Recall that we have defined −1 as the set of all type profiles in which there is a patient
depositor in the position with exactly − 1 impatient depositors ahead of her in the order. Now
define the events
−1 =
©
ª
( ) : ∈ −1
= {( ) : = }
and
= −1
−1
T
is the set of pairs ( ) in which the depositor in question is in the position
The event −1
with exactly − 1 impatient depositors ahead of her in the order. Note that since the events −1
30
and are independent and ( ) = 1 for all values of we have
The union
¢
¡
¡ −1 ¢ −1
P () I{∈−1 }
=
=
∈Ω
−1 =
S
=
−1
contains all of the profiles that the depositor in question considers to be possible together with the
positions that she could conceivably occupy in each profile. Because the sets −1
are disjoint
for different values of we have
¡
¡
¢
¢
P
−1 =
−1
=
P
P () I{∈−1 }
=
= ∈Ω
I{∈ −1 }
P
P
()
=
=
∈Ω
(22)
In other words, the probability of the set −1 can be obtained by summing the probabilities of all
profiles each weighted by the fraction of the positions in the order that match the observed
criteria (that is, have a patient depositor with exactly − 1 impatient depositors ahead of her in
the order). Notice that any profile that is not in the set −1 for some value of receives zero
weight in this sum.
Next, define
= {(b
) :
b = }
The set contains all of the pairs ( ) that correspond to a particular profile of preference types
It is straightforward to see that the prior probability of this set is given by () The posterior
probability we want to calculate is
¡
¢
| −1
that is, the probability of a particular profile conditional on the depositor observing that there are
− 1 impatient depositors ahead of her in the order. Applying Bayes’ rule yields
¢ (−1 | ) ( )
¡
| −1 =
(−1 )
31
(23)
By definition
¡
¢
¡
T ¢
−1 | ( ) = −1 =
µ
For a given profile we have
−1 T
⎧
⎨ ∅
∅
=
⎩
if
¶
¡¡
¢T ¢
S
−1 T
(24)
=
⎫
−1 = ∅ ⎬
∈
−1
⎭
∈ −1
Again using the fact that is independent of and that ( ) = 1 for all , this implies
P
I{∈ −1 }
¡ −1 T ¢
=
= ( )
(25)
In other words, the probability of the intersection of the sets −1 and is equal to the probability
of multiplied by the fraction of the positions in the order that the depositor could conceivably
occupy when the realized profile is
Substituting (22), (24), and (25) into expression (23) yields the posterior belief over profiles
P
I{∈ −1 }
()
¡
¢
=
−1
¶
µ
|
=
P
P
()
I{∈ −1 }
∈Ω
=
This expression shows that the posterior probability of a profile is proportional to its prior probability multiplied by the fraction of positions in the order the depositor could conceivably occupy
if were the true profile.
Finally, the posterior probability distribution over the number of patient depositors, is given
by
(; ) =
P
{:()=}
32
¡
¢
| −1
¥
References
Andolfatto, David and Ed Nosal (2008) “Bank incentives, contract design, and bank runs,” Journal
of Economic Theory 142, 28-47.
Andolfatto, David, Ed Nosal, and Bruno Sultanum (2014) “Preventing bank runs,” Working Paper
2014-21, Federal Reserve Bank of St. Louis.
Andolfatto, David, Ed Nosal, and Neil Wallace (2007) “The role of independence in the Green-Lin
Diamond-Dybvig Model,” Journal of Economic Theory 137, 709-715.
Azrieli, Yaron and James Peck (2012) “A bank runs model with a continuum of types,” Journal of
Economic Theory 147, 2040-2055.
Baba, Naohiko, Robert N, McCauley, and Srichander Ramaswamy (2009) “US dollar money market funds and non-US banks,” BIS Quarterly Review, March, 64-81.
Bertolai, Jefferson D.P, Ricardo de O. Cavalcanti, and Paulo K. Monteiro (2014) “Run theorems
for low returns and large banks,” Economic Theory 57, 223-252.
Cavalcanti, Ricardo de O. and Paulo K. Monteiro (2011) “Enriching information to prevent bank
runs,” Economics Working Papers No. 721, Getulio Vargas Foundation.
Cooper, Russell and Thomas W. Ross (1998) “Bank runs: liquidity costs and investment distortions,” Journal of Monetary Economics 41, 27-38.
Diamond, Douglas W. and Phillip H. Dybvig (1983) “Bank runs, deposit insurance, and liquidity,”
Journal of Political Economy 91, 401-419.
Duffie, Darrell (2010) “The failure mechanics of dealer banks,” Journal of Economic Perspectives
24, 51-72.
Dwyer, Gerald P., Jr., and Iftekhar Hasan (2007) “Suspension of payments, bank failures, and the
nonbank public’s losses,” Journal of Monetary Economics 54, 565–580
Ennis, Huberto M. and Todd Keister (2010a) “Banking panics and policy responses,” Journal of
Monetary Economics 57, 404-419.
Ennis, Huberto M. and Todd Keister (2010b) “On the fundamental reasons for bank fragility,”
Federal Reserve Bank of Richmond Economic Quarterly 96, 33-58.
Ennis, Huberto M. and Todd Keister (2009) “Run equilibria in the Green-Lin model of financial
intermediation,” Journal of Economic Theory 144, 1996-2020.
Ennis, Huberto M. and Todd Keister (2006) “Bank runs and investment decisions revisited,” Journal of Monetary Economics 53, 217-232.
Friedman, Milton and Anna J. Schwartz (1963) A Monetary History of the United States, Princeton
University Press, Princeton, NJ.
Gorton, Gary and Andrew Metrick (2012) “ Securitized banking and the run on repo,” Journal of
Financial Economics 104, 425-451.
33
Green, Edward J. and Ping Lin (2003) “Implementing efficient allocations in a model of financial
intermediation,” Journal of Economic Theory 109, 1-23.
Green, Edward J. and Ping Lin (2000) “Diamond and Dybvig’s classic theory of financial intermediation: What’s missing?” Federal Reserve Bank of Minneapolis Quarterly Review 24, 3-13.
Gu, Chao (2011a) “Herding and bank runs,” Journal of Economic Theory 146, 163-188.
Gu, Chao (2011b) “Noisy sunspots and bank runs,” Macroeconomic Dynamics 15, 398-418.
Jacklin, Charles (1987) “Demand deposits, trading restrictions, and risk sharing,” in: E. Prescott
and N. Wallace, eds., Contractual arrangements for intertemporal trade, University of Minnesota Press, Minneapolis, MN, 26-47.
Keister, Todd (2014) “Bailouts and financial fragility,” Department of Economics Working Paper
2014-01, Rutgers University.
Keister, Todd and Vijay Narasiman (2015) “Expectations vs. fundamentals-based bank runs: When
should bailouts be permitted?” Review of Economic Dynamics, forthcoming.
McCabe, Patrick E. (2010) “The cross section of money market fund risks and financial crises,”
FEDS Working Paper No. 2010-51.
Nosal, Ed and Neil Wallace (2009) “Information revelation in the Diamond-Dybvig banking model,”
Federal Reserve Bank of Chicago Policy Discussion Paper 2009-7, December.
Peck, James and Karl Shell (2003) “Equilibrium bank runs,” Journal of Political Economy 111,
103-123.
Selgin, George (1993) “In defense of bank suspension,” Journal of Financial Services Research 7,
347-364.
Sultanum, Bruno (2014) “Optimal Diamond-Dybvig mechanism in large economies with aggregate uncertainty,” Journal of Economic Dynamics and Control 40, 95-102.
Wallace, Neil (1988) “Another attempt to explain an illiquid banking system: the Diamond and
Dybvig model with sequential service taken seriously,” Federal Reserve Bank of Minneapolis
Quarterly Review 12 (Fall), 3-16.
Wallace, Neil (1990) “A banking model in which partial suspension is best,” Federal Reserve Bank
of Minneapolis Quarterly Review 14 (Fall), 11-23.
Yorulmazer, Tanju (2014) “Case studies on disruptions during the crisis,” Federal Reserve Bank of
New York Economic Policy Review 20 (1), 17-28.
34
@FedFRASER