Dynamic Global Games of Regime Change

8/13/2019 Dynamic Global Games of Regime Change

1/46

Econometrica, Vol. 75, No. 3 (May, 2007), 711756

DYNAMIC GLOBAL GAMES OF REGIME CHANGE: LEARNING,MULTIPLICITY, AND THE TIMING OF ATTACKS

BYGEORGE-MARIOSANGELETOS, CHRISTIANHELLWIG,ANDALESSANDROPAVAN1

Global games of regime changecoordination games of incomplete information inwhich a status quo is abandoned once a sufficiently large fraction of agents attack ithave been used to study crises phenomena such as currency attacks, bank runs, debtcrises, and political change. We extend the static benchmark examined in the literatureby allowing agents to take actions in many periods and to learn about the underlyingfundamentals over time. We first provide a simple recursive algorithm for the charac-terization of monotone equilibria. We then show how the interaction of the knowledgethat the regime survived past attacks with the arrival of information over time, or withchanges in fundamentals, leads to interesting equilibrium properties. First, multiplic-ity may obtain under the same conditions on exogenous information that guaranteeuniqueness in the static benchmark. Second, fundamentals may predict the eventualfate of the regime but not the timing or the number of attacks. Finally, equilibrium dy-namics can alternate between phases of tranquilitywhere no attack is possibleandphases of distresswhere a large attack can occureven without changes in funda-mentals.

KEYWORDS: Global games, coordination, multiple equilibria, information dynamics,crises.

1. INTRODUCTION

GAMES OF REGIME CHANGEare coordination games in which a status quo is

abandoned, causing a discrete change in payoffs, once a sufficiently large num-ber of agents take an action against it. These games have been used to model avariety of crises phenomena: an attack against the status quo is interpreted asspeculation against a currency peg, as a run against a bank, or as a revolutionagainst a dictator.

Most applications of these games to crises have been confined to staticframeworks: they abstract from the possibility that agents take multiple shotsagainst the status quo and that their beliefs about their ability to induce regimechange vary over time.2 Yet, these two possibilities are important from both

1Earlier versions of this paper were entitled Information Dynamics and Equilibrium Multi-plicity in Global Games of Regime Change. We are grateful to a co-editor and three anonymousreferees for suggestions that helped us improve the paper. For useful comments, we also thank

Andy Atkeson, Daron Acemoglu, Pierpaolo Battigalli, Alberto Bisin, V. V. Chari, Lars Hansen,Patrick Kehoe, Alessandro Lizzeri, Kiminori Matsuyama, Stephen Morris, Hyun Song Shin, IvnWerning, and seminar participants at Berkeley, Bocconi, Bologna, British Columbia, Chicago,MIT, Northwestern, NYU, Pompeu Fabra, Princeton, UCL, UCLA, UPenn, Yale, the Minneapo-lis FRB, the 2003 and 2004 SED meetings, the 2005 CEPR-ESSET, the 2005 IDEI conference intribute to J. J. Laffont, the 2005 NBER Summer Institute, the 2005 Cowles workshop on coordi-nation games, and the 2005 World Congress of the Econometric Society.

2This is particularly true for recent applications that introduce incomplete information. SeeMorris and Shin (1998)for currency crises; Goldstein and Pauzner (2005)and Rochet and Vives

711
http://www.econometricsociety.org/http://www.econometricsociety.org/


2/46

712 G.-M. ANGELETOS, C. HELLWIG, AND A. PAVAN

an applied and a theoretical perspective. First, crises are intrinsically dynamicphenomena. In the context of currency crises, for example, speculators canattack a currency again and again until they induce devaluation, and their ex-

pectations about the ability of the central bank to defend the currency in thepresent may naturally depend on whether the bank has successfully defendedit in the past. Second, learning in a dynamic setting may critically affect thelevel of strategic uncertainty (i.e., uncertainty about one anothers actions) andthereby the dynamics of coordination and the determinacy of equilibria.

In this paper, we consider a dynamic global game that extends the staticbenchmark used in the literature so as to incorporate precisely the two pos-sibilities highlighted in the foregoing discussion.3 There is a large number ofagents and two possible regimes, the status quo and an alternative. The gamecontinues as long as the status quo is in place. In each period, each agent caneitherattackthe status quo (i.e., take an action that favors regime change) or

not attack. The net payoff from attacking is positive if the status quo is aban-doned in that period and negative otherwise. Regime change, in turn, occursif and only if the fraction of agents who attack exceeds a critical value Rthat parameterizes the strength of the status quo. The variable captures thecomponent of the payoff structure (the fundamentals) that is never commonknowledge; as time passes, agents receive more and more private informationabout.

We first provide an algorithm for the characterization of monotone equilib-ria, based on a simple recursive structure. A difficulty with extending globalgames to dynamic settings is the need to keep track of the evolution of the

cross-sectional distribution of beliefs. Our framework overcomes this difficultyby summarizing the private information of an agent about at any given pe-riod in a one-dimensional sufficient statistic and by capturing the dynamics ofthe cross-sectional distribution of this statistic in a parsimonious way. We thenapply this algorithm to examine the effects of learning on the determinacy ofequilibria and the dynamics of coordination.

Multiplicity

We find that multiple equilibria can exist in our dynamic game under the

same conditions on the precision of exogenous private and public informa-tion that would guarantee uniqueness in the static benchmark that is the back-bone of most recent applications of global games (Morris and Shin (1998,2001, 2003)). Multiplicity originates in the interaction between the endoge-nous learning induced by the knowledge that the regime survived past attacks

(2004) for bank runs; Morris and Shin(2004) and Corsetti, Guimaraes, and Roubini(2005) fordebt crises; Atkeson (2001)and Edmond(2005) for riots and political change.

3Global games are incomplete-information games that often admit a unique, iteratively dom-inant equilibrium; see Carlsson and Van Damme (1993)and Morris and Shin(2003). The appli-cations cited in footnote2are all based on static global games.


3/46

DYNAMIC GLOBAL GAMES 713

and the exogenous learning induced by the arrival of new private informationover time.

Iterated deletion of dominated strategies ensures that equilibrium play is

uniquely determined in the first period: an attack necessarily takes place forevery, but succeeds in triggering regime change if and only if is sufficientlylow. In any subsequent period, the knowledge that the status quo is still inplace makes it common certainty that it is not too weak (i.e., that is highenough) and ensures that no agent ever again finds it dominant to attack. Asa result, there always exists an equilibrium in which no attack occurs after thefirst period. This would actually be the unique equilibrium of the game if agentsdid not receive any information after the first period.

When instead new private information about arrives over time, this hastwo effects on posterior beliefs about the strength of the status quo and hence

on the agents incentives to attack. On the one hand, it dilutes the upwardshift in posterior beliefs induced by the knowledge that the regime survivedthe first-period attack, which contributes to making further attacks possible.On the other hand, it reduces the dependence of posterior beliefs on the com-mon prior, which in general has an ambiguous effect. When the prior mean ishigh (i.e., favorable to the status quo), discounting the prior also contributesto making a new attack possible; the opposite is true when the prior mean islow. A high prior mean thus ensures the existence of an equilibrium where asecond attack occurs once private information becomes sufficiently precise.

More generally, we show that when the prior mean is sufficiently high, thearrival of private information over time suffices for the existence of arbitrarilymany equilibria, which differ in both the number and the timing of attacks.

Dynamics of Coordination

The aforementioned multiplicity result does not mean that anything goes:equilibrium outcomes in any given period depend critically on available infor-mation and the history of past play.

The learning induced by the knowledge that the status quo survived past at-tacks introduces a form of strategic substitutability across periods: the more ag-gressive the agents strategy in one period, the higher the threshold in below

which regime change occurs in that period, but then the larger is the upwardshift in posterior beliefs induced by the knowledge that the regime survived thisattack and, hence, the lower is the incentive to attack in subsequent periods.When an aggressive attack takes place in one period but fails to trigger regimechange, then a significant increase in the precision of private information isnecessary to offset the endogenous upward shift in posterior beliefs and makea new attack possible in equilibrium. As a result, dynamics take the form of se-quences of periods in which attacks cannot occur and agents only accumulateinformation, followed by periods in which an attack is possible but does notmaterialize, eventually resulting in a new attack.


4/46


Moreover, although it is possible that attacks continue indefinitely as longas new information arrives over time, strategic uncertainty significantly limitsthe size of attacks. There exists a region of in which the status quo survives

in all equilibria of the incomplete-information game, but in which it wouldhave collapsed in certain equilibria of the complete-information version of thegame.

Implications for Crises

These results translate to novel predictions for the dynamics of crises.First, fundamentals may determine eventual outcomese.g., whether a cur-

rency is devaluedbut not the timing and number of attacks.Second, an economy can transit from phases of tranquility, where the

unique possible outcome is no attack, to phases of distress, where a signifi-cant change in outcomes can be triggered by a shift in market sentiments.Finally, the transition from one phase to another can be caused by a small

change in information or, in a later extension, by a small change in fundamen-tals.

These predictions strike a delicate balance between two alternative viewsof crises. The first associates crises with multiple self-fulfilling equilibria: largeand abrupt changes in outcomes are attributed to shifts in market sentimentsor animal spirits (Obstfeld (1996)). The second associates crises with a dis-continuity, or strong nonlinearity, in the dependence of the unique equilibriumto exogenous variables: large and abrupt changes in outcomes are attributedto small changes in fundamentals or in agents information (Morris and Shin(1998,2001)). Our results combine a refined role for multiple self-fulfilling ex-pectations with a certain discontinuity in equilibrium outcomes with respect toinformation and fundamentals.

Extensions

The benchmark model focuses on the arrival of private information as theonly exogenous source of change in beliefs. In an extension we show how theanalysis can accommodate public news about the underlying fundamentals.

This only reinforces the multiplicity result. Moreover, equilibrium dynamicscontinue to be characterized by phases of tranquility and phases of distress,but now the transition from one phase to another can be triggered by publicnews.

The benchmark model also deliberately assumes away the possibility thatthe critical size of attack that triggers regime change may vary over time. Thispermits us to isolate the effect of changes in information (beliefs), as opposedto changes in fundamentals (payoffs), on the dynamics of coordination. Nev-ertheless, introducing shocks to fundamentals is important for applications, aswell as for understanding the robustness of our results.


5/46


We first examine the case in which the shocks are perfectly observable.Shocks then provide an additional driving force for dynamics: a transition fromtranquility to distress may now be triggered by a deterioration in fundamentals.

Moreover, a sufficiently bad shock can push the economy into a phase wherean attack becomes inevitablea possibility absent in the benchmark model.We next consider the case in which the shocks are unobservable (or observed

with private noise). The novel effect is that the uncertainty about the shocksnoises up the learning induced by the knowledge that the regime survivedpast attacks: whereas in the benchmark model this knowledge leads to a trun-cation in the support of posterior beliefs about the strength of the status quo,here posterior beliefs retain full support over the entire real line. Thus, in con-trast to the benchmark model, agents with very low signals may now find itdominant to attack in every period and, other things equal, a unique equilib-rium outcome may obtain in any given period when private information in thatperiod is sufficiently precise. Nevertheless, our results are robust as long asthe noise in learning is small: any equilibrium of the benchmark model is ap-proximated arbitrarily well by an equilibrium of the game with shocks as thevolatility of the shocks vanishes.

Thus, what sustains the multiplicity of equilibria and the structure of dynam-ics identified in this paper is the combination of exogenous changes in informa-tion or fundamentals with the endogenous learning induced by the knowledgethat the regime survived past attacks. The fact that, in the benchmark model,this learning takes the sharp form of a truncation in the support of beliefs sim-plifies the analysis but is not essential for the results. What is essential is that

this learning implies a significant change in common beliefs about the strengthof the status quo.

Related Literature

This paper contributes to the literature on global games by highlighting theimportance of learning from past outcomes for equilibrium determinacy. Inthis respect, it shares with Angeletos, Hellwig, and Pavan (2006)who con-sider the signaling effects of policy interventions in a static environmenttheidea that natural sources ofendogenousinformation may qualify the applica-

bility of global-game uniqueness results, while at the same time reinforcingthe more general point that information is important for coordination. In ourframework this leads to novel predictions that would not have been possiblewith either common knowledge or a unique equilibrium.4

The paper also contributes to a small but growing literature on dynamicglobal games. Morris and Shin (1999) consider a dynamic model whose stage

4Information is endogenized also in Angeletos and Werning (2006), Hellwig, Mukherji, andTsyvinski (2006), and Tarashev (2005), where financial prices aggregate and publicize disperseprivate information, and in Edmond (2005), where a dictator manipulates the distribution ofprivate signals.


6/46


game is similar to ours, but where the strength of the status quo follows a ran-dom walk and is commonly observed at the end of each period. This reducesthe analysis to a sequence of essentially unrelated static games, each with a

unique equilibrium. Chamley (1999) also considers a global game in whichfundamentals change over time and where aggregate activity is perfectly ob-servable but where the latter is informative in equilibrium only when the fun-damentals enter the dominance regions. Heidhues and Melissas (2006) andGiannitsarou and Toxvaerd (2003) establish uniqueness results for dynamicglobal games on the basis of dynamic strategic complementarities. Dasgupta(2006) examins the role of noisy social learning in a two-period investmentmodel with irreversible actions. Levin (2001) considers a global game withoverlapping generations of players. Goldstein and Pauzner (2004) and Gold-stein (2005) consider models of contagion. Frankel and Pauzner (2000) ex-amin a dynamic coordination game where uniqueness is obtained by combining

aggregate shocks with idiosyncratic inertia. Abreu and Brunnermeier (2003)consider a setting in which speculators become gradually and asymmetricallyaware of the mispricing of a financial asset. All these papers feature multi-period coordination problems, but none of them features the form of learningthat is the center of our analysis.5 Our methodological approach is also quitedifferent: instead of forcing uniqueness, we wish to understand how a naturalform of learning sustained by repeated play may affect both the determinacyof equilibria and the structure of dynamics.

Finally, this paper shares with Chari and Kehoe (2003) the motivation thatinformation is important for understanding crises: our benchmark model of-

fers a theory where changes in information are the sole source for the dy-namics of crises. However, there are two important differences. First, Chariand Kehoe focus on the effect of herding in an environment without strate-gic complementarities. In contrast, we focus on the effect of learning on thedynamics of coordination. The coordination element is crucial for the predic-tion that there is a phase of distress during which an attack is possible butdoes not necessarily take place, as well as for the prediction that attacks occuras sudden and synchronized events. Second, the main learning effect in Chariand Kehoe is the negative information about the fundamentals revealed by thechoice by some agents to attacka form of learning that generates build-upor snowballing effects. In contrast, the main learning effect in our bench-mark model is the positive information revealed by the failure of an attackto trigger regime changea form of learning that is crucial for our predic-tion that phases of distress are eventually followed by phases of tranquility. InSection5.2we discuss an extension of our benchmark model in which agentsobserve noisy signals about the size of past attacks. This extension combinesour cycles between phases of distress and tranquility with snowballing effectssimilar to those stressed in the herding literature.

5Chamley(2003) also considers learning in a dynamic coordination game. However, his modelis not a global game; all information is public and so is learning.


7/46


Layout

The rest of the paper is organized as follows. Section 2reviews the staticbenchmark and introduces the dynamic model. Section3characterizes the set

of monotone equilibria. Section4establishes the multiplicity result and exam-ines the properties of equilibrium dynamics. Section5considers a few exten-sions of the benchmark model and examines robustness. Section6 concludes.Proofs omitted in the main text are provided in theAppendix.

2. A SIMPLE GAME OF REGIME CHANGE

2.1. Static Benchmark

Model setup

There is a continuum of agents of measure 1, indexed by i and uniformlydistributed over[0 1]. Agents move simultaneously, choosing between twoactions: they can either attack the status quo (i.e., take an action that favorsregime change) or refrain from attacking.

The payoff structure is illustrated in TableI. The payoff from not attack-ing (ai=0) is zero, whereas the payoff from attacking (ai=1) is 1c >0if the status quo is abandoned (R = 1) andc


8/46


iN(0 1/)is noise, independent and identically distributed across agentsand independent of. The Normality assumptions allow us to parameterizethe information structure parsimoniously with(z), that is, the precision

of private information, the precision of the common prior, and the mean of thecommon prior.

Interpretation

Although the model is highly stylized, it admits a variety of interpretationsand possible applications. The most celebrated examples are self-fulfilling bankruns, currency attacks, and debt crises. In these contexts, regime change oc-curs, respectively, when a large run forces the banking system to suspend itspayments, when a large speculative attack forces the central bank to abandonthe peg, and when a country/company fails to coordinate its creditors to roll

over its debt and is hence forced into bankruptcy. The model can also be in-terpreted as one of political change, in which a large number of citizens decidewhether or not to take actions to subvert a repressive dictator or some otherpolitical establishment. (For references, see footnote2.)

Equilibrium analysis

Note that the cumulative distribution function of an agents posteriorabout is decreasing in his private signal x. Moreover, it is strictly domi-nant to attack for sufficiently low signals (namely for x < x, where x solvesPr( 0|x) = c) and not to attack for sufficient high signals (namely forx >x,wherex solves Pr( 1|x) = c). It is thus natural to look atmonotoneBayesianNash equilibria in which the agents strategy is nonincreasing in x.

Indeed, suppose there is a thresholdx R such that each agent attacks ifand only ifxx. The measure of agents who attack is then decreasing in and is given by

A() = Pr(x x|) = (

(x ))where is the cumulative distribution function of the standard Normal. Itfollows that the status quo is abandoned if and only if, where solves

=A(

), or equivalently

= ((x ))(1)By standard Gaussian updating, the posterior about conditional on the pri-vate signal xis Normal with mean

+ x + + zand precision + . It followsthat the posterior probability of regime change is simply

Pr(R = 1|x) = Pr( |x)

= 1

+

+

x +

+

z


9/46


Because the latter is decreasing inx, an agent finds it optimal to attack if and

only ifx x, wherexsolves Pr( |x) = c, or equivalently

1 + + x + + z = c(2)

A monotone equilibrium is thus identified by a joint solution (x) to(1) and(2). Such a solution always exists and is unique for all z if and onlyif 2/(2). Moreover, iterated deletion of strictly dominated strategiesimplies that when the monotone equilibrium is unique, there is no other equi-librium.

PROPOSITION1: In the static game, the equilibrium is unique if and only if

2/(2) and is in monotone strategies.

In the limit as for given, the thresholdconverges to 1 c,and the size of attack A() converges to 1 for all < and to 0 for all > .Hence, when the noise in private information is small and is in the neighbor-hood of, a small variation in can trigger a large variation in the size ofattack and in the regime outcome. This kind of discontinuity, or strong nonlin-earity, in the response of equilibrium outcomes to exogenous variables under-lies the view of crises advocated by most global-game applications.6

2.2. Dynamic Game

We modify the foregoing static game in two ways: first, we allow agents toattack the status quo repeatedly; second, we let agents accumulate informationover time.

Time is discrete and is indexed by t {1 2 }. The game continues as longas the status quo is in place and is over once the status quo is abandoned. Wedenote byRt= 0 the event that the status quo is in place at the beginning ofperiod t, by Rt= 1 the alternative event, by ait {0 1} the action of agent i,and by At [0 1] the measure of agents attacking at date t. Conditional onthe regime being in place at the beginning of period t (Rt=0), the regimeis abandoned in that period (Rt+1=1) if and only ifAt, where againrepresents the strength of the status quo. Agent is flow payoff for period t(conditional on Rt=0) is thus it=ait(Rt+1 c), while his payoff from theentire game is i=

t=1

t1(1Rt)it, where (0 1) is the discount factor.

6A related strong nonlinearity emerges in the response of equilibrium outcomes to noise inpublic information; see the discussion of the publicity multiplier in Morris and Shin ( 2003) andthat of nonfundamental volatility in Angeletos and Werning (2006).


10/46


As in the static model, is drawn at the beginning of the game fromN(z 1/), which defines the initial common prior, and never becomes com-mon knowledge. Private information, however, evolves over time. In each pe-

riod t 1, every agent i receives a private signalxit= + it about , whereitN(0 1/t)is independent and identically distributed across i, indepen-dent of, and serially uncorrelated. Letxti= {xi}t=1 denote agent is historyof private signals up to period t. Individual actions and the size of past attacksarenot observable; hence the public history in period tsimply consists of theinformation that the regime is still in place, whereas the private history of anagent is the sequence of own private signals and own past actions. Finally, welett

t=1 and assume that

> t 2/(2) t and limt

t=

As shown in the next section,tparameterizes the precision of private infor-mation accumulated up to period t. The assumptions we make here for tensure (i) that the static game defined by the restriction that agents can moveonly in periodthas a unique equilibrium for everytand (ii) that private infor-mation becomes infinitely precise only in the limit.

REMARK: Although this dynamic game is highly stylized, it captures two im-portant dimensions that are absent in the static benchmark: first, the possibilityof multiple attacks against the status quo; second, the evolution of beliefs about

the strength of the status quo. By assuming that per-period payoffs do not de-pend on past or future actions and by ignoring specific institutional details,the model may of course fail to capture other interesting effects introduced bydynamics, such as, for example, the role of wealth accumulation or liquidity incurrency crises. However, abstracting from these other dimensions allows us toisolate information as the driving force for the dynamics of coordination andcrises.

Equilibrium

In what follows, we limit attention tomonotone equilibria, that is, symmetricperfect Bayesian equilibria in which the probability an agent attacks in period t,which we denote byat(x

t), is nonincreasing in his private signalsxt and inde-pendent of his own past actions. Restricting attention to this class of equilibriasuffices to establish our results.

3. EQUILIBRIUM CHARACTERIZATION

Let at: Rt [0 1]denote the strategy for period tand let at = {a}t=1 de-

note the strategy up to period t, with a

= {a

}

=1 denoting the complete


11/46


strategy for the dynamic game. Becausext is independent and identically dis-tributed across agents conditional on , for any given strategy a the size ofattack and the regime outcome in period tdepend only on . Thus let pt(; at)denote the probability that the status quo is abandoned in period twhen allagents follow the strategy at, conditional on the status quo being in place atthe beginning of period tand the fundamentals being . Finally, let1(|x1)denote the cumulative distribution function of the posterior beliefs in period 1,while for anyt 2, lett(|xt; at1)denote the cumulative distribution func-tion of the posterior beliefs in period tconditional on the knowledge that thestatus quo is still in place (i.e., Rt=0) and that agents have played in pastperiods according toat1.

Because neither individual nor aggregate actions are observable and becauseRt= 0 is always compatible with any strategy profile at any t, no agent can de-tect out-of-equilibrium play as long as the status quo is in place.

7

It followsthat beliefs are pinned down by Bayes rule in any relevant history of the game.Furthermore, as long as the status quo is in place, payoffs in one period do notdepend on own or other players actions in any other period and, hence, strate-gies are sequentially rational if and only if the action prescribed for any givenperiod maximizes the payoff for that period. We conclude that the strategya

is part of an equilibrium if and only if the following hold: at t= 1, for allx1,

a1(x1) arg maxa[01]

p1(; a1) d1(|x1) c

a

;(3)

at anyt 2, for allxt,

at(xt) arg max

a[01]

pt(; at) dt(|xt; at1) c

a

(4)

Next, definextandtrecursively by

xt=t1t

xt1 +t

txt and t= t1 + t

withx1=x1and 1= 1. By standard Gaussian updating, the distribution ofconditional onxt = {x}t=1 is Normal with mean t+t xt+ +t z and precisiont+ . It follows that xt is a sufficient statistic for xt with respect to and,hence, with respect to the event of regime change as well. As we show sub-sequently (and further discuss in Section 5.5), this ability to summarize pri-vate information into a one-dimensional sufficient statistic greatly simplifiesthe analysis.

7Indeed, the regime always survives any attack for > 1 and no realization of the private signalrules out > 1.


12/46


Clearly, condition (3) implies that in any equilibrium of the dynamic game,agents play in the first period exactly as in the static game in which they canattack only att= 1. Hence, by Proposition1,equilibrium play is uniquely de-termined in the first period and is characterized in terms of thresholds forx1 and . The following lemma shows that a similar property holds for sub-sequent periods.8

LEMMA 1: Any monotone equilibrium is characterized by a sequence{xt t}t=1, with xt R {}, t(0 1), and tt1 for all t2, suchthat:

(i) at anyt 1,an agent attacks ifxt< xt and does not attacks ifxt> xt;(ii) the status quo is in place in periodt 2if and only if > t1.

PROOF: We prove the claim by induction. Fort= 1, the result follows fromProposition1. Consider next any t2 and suppose that the result holds forany t1. Because at is nonincreasing inxt, the size of attack At() isnonincreasing in , implying that either At() < (and therefore Rt+1=0)for all > t1, in which case

t=t1, or there exists t > t1 such that

At( ) < if and only if > t . In the former case, the posterior probability of

regime change is 0 for all xtand hence xt= . In the latter, the posterior

probability of regime change is given by

pt(; at) dt(| xt

;at1)

=Pr(

t

|xt >

t

1)(5)

= 1

t+ [txt+zt+ t]

t+ [ txt+zt+ t1]

Because this is continuous and strictly decreasing in xt, and converges to 1as xt and to 0 as xt +, there exists xt R such that Pr(t|xt >

t1) = cfor xt= xt, Pr( t|xt > t1) > cfor xt< xt, and Pr(

t|xt > t1) < c for xt> xt. In either case, At() 0 implies thatt (0 1)for all t, which

completes the proof. Q.E.D.

Clearly, given that the status quo cannot be in place in one period withoutalso being in place in the previous, the sequence{t}is nondecreasing. On theother hand, the sequence {xt} is nonmonotonic in general: periods where someagents attack (xt >) may alternate with periods where nobody attacks(xt= ).

8To simplify the notation, we allow forxt= and xt= +, with which we denote the casewhere an agent attacks for, respectively, none and every realization of his private information.


13/46


As mentioned previously, the first period in our dynamic game is similar tothe static game, but any subsequent period is very different. In any t 2, thefact that the status quo is still in place makes it common certainty that > t1.

9

Becauset1 1> 0, there always exist equilibria in which nobody attacks inperiodt 2 (in which casext= and t= t1). In particular, there existsan equilibrium in which an attack takes place in period one and never there-after. When this is the unique equilibrium, the possibility to take repeated ac-tions against the regime adds nothing to the static analysis and the equilibriumoutcome in the dynamic game coincides with that in the static benchmark. Inwhat follows we thus examine under what conditions there also exist equilibriawith further attacks.

Lemma1rules outxt= +(situations where everybody attacks). This fol-lows directly from the fact that the status quo always survives for >1 and

hence it is dominant not to attack for xt sufficiently high. We thus look forequilibria in whichxt R.

The size of attack is then given by At() = Pr(xt xt|) = (

t(xtt)),

which is continuous and strictly decreasing in , while the probability of regimechange for an agent with statistic xt is given by (5), which is continuous andstrictly decreasing inxtif

t >

t1. It follows that, in any equilibrium in which

an attack occurs in period t,t andxt solve

t= At(t) and Pr( t|xt > t1) = c

or equivalently

t= (

t(xt t))(6)

1

t+ ( tt+ xt+ t+ z t)

t+ ( tt+ xt+ t+ z t1)= c(7)

Conditions (6) and(7) are the analogues in the dynamic game of conditions(1) and(2) in the static game: (6) states that the equilibrium size of an attackis equal to the critical size that triggers regime change if and only if the funda-

mentals are t, while(7) states that an agent is indifferent between attackingand not attacking if and only if his private information is xt.

An alternative representation of the equilibrium conditions is also useful.

Define the functions u :R[0 1]RR2+ R [c 1c], X: [0 1]R+

9Clearly, the knowledge that the regime is in place in period t is a form of public informa-tion. However, this is very different from the type of information conveyed by additive publicsignals of (Morris and Shin(2001,2003), Hellwig(2002)). First, the information here is endoge-nous, for it depends on the particular equilibrium being played; second, it leads to a first-orderstochastic-dominance shift in beliefs. We introduce additive public signals about in Section5.1.


14/46


R, andU: [0 1] RR2+ R [c 1 c]as follows10:u(x 1z)

1

+ ( + x + + z )

+ ( + x + + z 1)

c if> 1,c if 1,

X ( ) + 1

1()

U( 1z)

limx

u(x 1z) if= 0,

u(X() 1z) if (0 1),

limx+

u(x 1

z) if

=1.

These functions have a simple interpretation: u(xt 1 tz)is the net

payoff from attacking in periodtfor an agent with statisticxtwhen it is knownthat > 1 and that regime change will occur if and only if ;X ( t)is the thresholdx such that, if agents attack in periodtif and only ifxt x,thenAt() if and only if ;U ( 1 tz)is the net payoff fromattacking for the marginal agent with signal x= X ( t)when it is knownthat > 1.

Solving (6) forxt givesxt= X (t t). Substituting the latter into (7) gives

U(t t1 tz) = 0(8)Condition (8) is central to the characterization results hereafter; it repre-sents the indifference condition for the marginal agent in period t for t2.As for t= 1, because the regime has never been challenged in the past,the corresponding indifference condition is U(1 1z) = 0. Clearly,U( z)coincides with the payoff of the marginal agent in the staticbenchmark.

We can thus characterize the set of monotone equilibria as follows.

PROPOSITION2: The strategy{at()}t=1 is a monotone equilibrium if and onlyif there exists a sequence of thresholds{xt t}t=1 such that:

(i) for allt,at(xt) = 1ifxt< xt andat(xt) = 0ifxt> xt;

(ii) fort= 1,1 solvesU (1 1z) = 0and x1= X(1 1);(iii) for anyt 2,eithert= t1> 0and xt= ,ort > t1is a solution

toU (t t1 tz) = 0and xt= X (t t).

A monotone equilibrium always exists.

10With a slight abuse of notation, we let (+)= 1, ()=0, 1(1)= +, and1(0)

= .


15/46


Proposition 2 provides an algorithm for constructing the entire set ofmonotone equilibria: Start with t= 1 and let 1 be the unique solution toU(1 1z) = 0. Next, proceed to t= 2. If the equation U (2 1 2z)=0 admits no solution set 2=1; if it admits a solution, either let 2be such a solution or simply set2= 1 . Repeat for all t 3 the same step asfort= 2. The set of sequences{t}t=1 constructed this way, together with theassociated sequences{xt}t=1, gives the set of monotone equilibria.

This recursive algorithm is based on the property that equilibrium learningtakes the simple form of a truncation in the support of beliefs about : Theknowledge that the regime has survived past attacks translates into the knowl-edge that is above a threshold t1. In Section 5we examine how this propertymay or may not extend to richer environments. Note also that the foregoingcharacterization is independent of whether the horizon is finite or infinite: it isclearly valid even if the game ends exogenously at an arbitrary periodT 1, lim U(

1 z)

=

,where

1

c.

(ii) Lettbe the unique solution to U(t tz) = 0.A solution to(8)exists only ift1 ,a solution to(8)does not exist fortsufficiently high.(iv) Ift1< ,a solution to(8)necessarily exists fortsufficiently high.(v) Ift1 is the highest solution to (8)for period t 1, there exists > t1

such that (8)admits no solution for any period tsuch that< .

To understand why the function U( 1 )is nonmonotonic in when-ever 1 (0 1), recall that the higher is , the higher is the thresholdx

=X ( t) such that if agents attack in period t if and only if xt

x,

then At() if and only if. When < 1, the threshold x is solow that the size of attack is smaller than for all > 1. The marginalagent then attaches zero probability to regime change, which explains whyU( 1 )= c for < 1. When instead > 1, the threshold x ishigh enough that regime change occurs for a positive measure of > 1. Themarginal agent then attaches positive probability to regime change, which ex-plains whyU ( 1 ) > cfor> 1. Finally, when 1,x andhence the probability that the marginal agent attaches to the event that > 1converges to 1. The probability he attaches to regime change then convergesto zero, which explains whyU (

1

)

cas

1.


16/46


FIGURE1.The payoff of the marginal agent.

We thus have that Uis flat atcfor < 1, then increases with , andeventually decreases with , and converges again toc as 1. This isillustrated by the solid curve in Figure 1. Any intersection of this curve withthe horizontal axis corresponds to a solution to (8).11 The dashed line insteadrepresents the payoff of the marginal agent in the static game in which agentscan attack only in periodt; whentis sufficiently high, this is monotonic in

.

While the monotonicity of the payoff of the marginal agent in the static gameensures uniqueness, the nonmonotonicity in the dynamic game leaves open thepossibility for multiple equilibria.

Next, to understand why Udecreases with z, note that an increase in theprior mean implies a first-order stochastic-dominance change in posterior be-liefs about : the higher is z, the lower is the probability of regime changefor any given monotone strategy and, hence, the lower is the net payoff fromattacking for the marginal agent.

Similarly, because an increase in 1 also corresponds to an upwardshift in posterior beliefs about , U also decreases with 1. This implies

that, at any t 2, the payoff of the marginal agent is always lower thanU( tz), that is, lower than the payoff in the static game wherethe precision of private information is t. This in turn explains why the static-

11It can be shown that U( 1 ) is single-peaked in when 1 1/2. Numerical simu-lations suggest that this is true even when 1 < 1/2, although we have not been able to proveit. Single-peakedness of U implies that (8) admits at most two solutions (generically none ortwo). When there are two solutions, as in the case of the solid line in Figure 1,the lowest onecorresponds to an unstable equilibrium, while the highest one corresponds to a stable equilib-rium. None of these properties, however, is needed for our results. All that matters is that U isnonmonotonic inwhen1

(0 1), with a finite number of stationary points.


17/46


game thresholdt(which corresponds to the intersection of the dashed linewith the horizontal axis in Figure1) is an upper bound for any solution to(8).

To understand (iii) and (iv), note that as t , the impact on pos-terior beliefs of the knowledge that > t1 vanishes for any xt > t1.By implication, as t , the difference between U( t1 tz) andU( tz) also vanishes for any > t1. Combined with the factthatU ( tz) ast , this implies that, for t suffi-ciently high, (8) necessarily admits at least one solution ift1< , and no so-

lution ift1> , where= limttis the limit of the equilibrium thresh-old in the static game for .

Finally, to understand (v), suppose that the largest possible attack (that is,the one that corresponds to the highest solution to (8)) is played in one pe-riod and is unsuccessful. Then the upward shift in posterior beliefs induced by

the observation that the status quo survived the attack is such that if no newinformation arrives, no further attack is possible in any subsequent period. Bycontinuity then, further attacks remain impossible as long as the change in theprecision of private information is not large enough.

4. MULTIPLICITY AND DYNAMICS

Part (v) of Lemma2highlights that the arrival of new private information isnecessary for further attacks to become possible after period 1. Whether thisis also sufficient depends on the prior mean.

When z is sufficiently low (aggressive prior), discounting the prior con-tributes to less aggressive behavior in the sense thattdecreases with tand,hence,t1) is not willing to attack inany periodt 2 if he expects all other agents to play as if no attack occurredprior to period t(i.e., as in the equilibrium of the static game where attacking isallowed only in periodt). The anticipation that other agents will also take intoaccount the fact that the regime survived past attacks then makes that agenteven less willing to attack. Therefore, when z is low, the game has a uniqueequilibrium, with no attack occurring after the first period.12

When, instead, z is sufficiently high (lenient prior), discounting the priorcontributes to more aggressive behavior in the sense thattincreases witht.This effect can offset the incentive not to attack induced by the knowledgethat the regime survived past attacks, making new attacks eventually possible.Indeed, Lemma 2 implies that when 1 < (which is the case for z highenough), a second attack necessarily becomes possible once tis large enough.

Such an example is illustrated in Figure2. The dashed line represents thepayoff of the marginal agent in period 1. Its intersection with the horizontal

12In this case, the unique monotone equilibrium is also the unique equilibrium of the game.


18/46


FIGURE2.Equilibria with multiple attacks.

axis defines 1 < . The payoff of the marginal agent in period 2 is repre-sented by the dotted line. That in period 3 is represented by the solid line.Clearly, 2 is low enough that no attack is possible in period 2. In contrast,3 is high enough that a new attack is possible. Thus, there exist at least threeequilibria in this example: one in which t=1 for all t, another in which2= 1 and t= 3 for all t 3, and a third in which 2= 1 and t= 3 forallt 3, where3 and3 correspond to the two intersections of the solid linewith the horizontal axis.

In the example of Figure2, both3 and3 are lower than. By Lemma2,

then, a third attack also becomes possible at some future date. More gener-ally, if z is sufficiently high, any solution to (8) is strictly less than in allperiods, which ensures that a new attack eventually becomes possible after anyunsuccessful one. Hence, for z sufficiently high, not only are there multipleequilibria, but any arbitrary number of attacks can be sustained in equilibrium.

THEOREM1: There exist thresholds z z zsuch that:(i) ifzz , there is a unique monotone equilibrium and it is such that anattack occurs only in period one;

(ii) ifz (zz),there are at mostfinitely manymonotone equilibria and thereexistst < such that,in any of these equilibria,no attack occurs after periodt;

(iii) ifz > z,there areinfinitely manyequilibria;if in additionz > z,for anytandNthere is an equilibrium in which Nattacks occur after period t.

Finally,z = z = zwhenc 1/2,whereasz z < zwhenc > 1/2.

PROOF: Recall that1=1and, for allt 2,t


19/46


there exist thresholdsz z z (possibly functions of1 and ) with the fol-lowing properties:t1 for all t ifz z;1 () if and onlyz () z;andt< for alltif and only ifz > z.

(i) Consider firstz z. Thent1= 1 for alltand, hence, by part (ii)of Lemma2,(8) admits no solution at anyt 2. The unique monotone equi-librium is thust= 1 for allt.

(ii) Next, consider z(z z), in which case 1=1 > , but we cannotrule out the possibility that there exists a period t2 such that t>1 andU( 1 tz) = 0 admits a solution. Nevertheless, because t1 1> for all t, by part (iii) of Lemma2and the fact that t as t , thereexistst < such that (8) admits no solution for tt. Moreover, because (8)admits at most finitely many solutions for any t z, in which case 1 < . Then, by part (iv) of

Lemma2,there exists a t< such that U( 1 tz) = 0 admits a so-lution for all t t. Hence, for any t t, there is a monotone equilibrium inwhich =1 for < t, t solves U(t 1 tz)=0 and =t for all > t. That is, there exist (countably) infinitely many equilibria, indexed by thetime at which the second attack occurs.

When z (zz), the second attack may lead to a threshold t > , in whichcase a third attack might be impossible. If, however, z > z , thent< foralltand, hence, by part (ii) of Lemma2,t < , for allt. By part (iv), a new

attack then eventually becomes possible after any unsuccessful one. It followsthat, for anyt 1 and anyN 1, there exist increasing sequences{t2 t N}and{2 N}, with t2t, such that U(2 1 t2 z)=0, U(3 2 t3 z) = 0, and so on. The following is then an equilibrium: = 1 for < t2,= jfor {tj t j+1 1}and j {2 N 1}, and= N for tN.That is, for any t 1 and any N 1, there exists an equilibrium in which Nattacks occur after periodt. Q.E.D.

The existence of infinitely many equilibria in the case z > zrelies on the as-sumption that the game continues forever as long as the status quo is in place:if the game ended for exogenous reasons at a finite date, there would exist onlyfinitely many equilibria. Nevertheless, as long as z > zand tas t ,then, for anyM, there exists a finite T such that the game has at leastMequi-libria if it ends at date T. Moreover, even when T= 2, the game has multipleequilibria if2is sufficiently high andz > z.

In the remainder of this section, we identify equilibrium properties that seemuseful in understanding the dynamics of crises.

COROLLARY 1: Suppose > and z > z. The status quo survives in anymonotone equilibrium.Nevertheless,there existst andz > z, a new attack eventuallybecomes possible after any unsuccessful one, the preceding result implies thatdynamics may take the form of cycles in which the economy alternates fromphases of tranquility to phases of distress, eventually resulting in a new attack,without any change in the underlying fundamentals. Once again, this would not

have been possible in our framework if were common knowledge or if therewere a unique equilibrium.13

Also note that the set of equilibrium outcomes for any given period exhibits adiscontinuity with respect to the precision of private information in that period:a transition from a phase of tranquility, where nobody attacking is the uniqueequilibrium outcome, to a phase of distress, where the size of attack associatedwith any solution of (8) is bounded away from zero, can be triggered by a smallchange in t. As we will see in Section5.3,a similar discontinuity emerges withrespect to shocks that affect the strength of the status quo: a transition fromone phase to another can then be triggered by an arbitrarily small change in

fundamentals (that is, in payoffs).Finally, note that these results raise some interesting possibilities for policy

in the context of currency crises. On the one hand, because an increase in c

13Broner (2005) considered a model that combines a common-knowledge coordination prob-lem la Obstfeld(1996) with a negative trend in fundamentals la Krugman (1979). The firstfeature delivers multiplicity, while the second ensures that devaluation is eventually inevitable forexogenous reasons. His analysis thus shares with ours the property that it may be easier to predictthe eventual outcome than the timing of attacks, but it does not share our predictions about therepeated succession of phases of tranquility and phases of distress nor our focus on changes ininformation, rather than changes in fundamentals, as the source of dynamics.


21/46


shiftsUdownward, a central bank might be able to prevent a transition froma phase of tranquility to a phase of distressand thus eliminate the risk ofa speculative attackby raising interest rates or otherwise increasing the op-

portunity cost of attacking. On the other hand, the level of policy interventionrequired to achieve this may increase over time as speculators become moreinformed about the underlying fundamentals and may eventually become pro-hibitively expensive. Thus an interesting possibility is that certain defense poli-cies succeed in postponing but not in escaping a crisis.14

5. EXTENSIONS

In this section, we consider a few extensions of the benchmark model. Thepurpose of these extensions is to show how the analysis can accommodate ad-

ditional elements that the benchmark model has deliberately abstracted from,but which can be relevant for applications. At the same time, these extensionsshow robustness to alternative information assumptions and further clarify thedriving forces behind the results.

5.1. Public News

To capture the effect of public news, we modify the game as follows. In addi-tion to their private signals, agents observe in each period t 1 a public signalzt= +t, where tis common noise, normally distributed with zero mean andprecision

zt >

0, serially uncorrelated, and independent of

and the noise inthe agents private information. These signals may represent, for example, theinformation generated by news in the media, publication of government sta-tistics, or announcements by policy makers. We also allow for the possibilitythat the game ends for exogenous reasons at a finite date and we denote thehorizon of the game with T, where eitherT {2 3 }or T= .

The common posterior about conditional onzt {z}t=1 is Normal withmeanztand precisiont, where

zt=t1t

zt1 +ztt

zt t= t1 + zt

with(z0 0) = (z). However, because equilibrium play in past periods nowdepends on the realizations of past public signals, zt is not a sufficient sta-tisticconditional on the event that the regime is still in place. We thus allowagents to condition their actions on the entire sequencezt or, equivalently, onzt {z}t=1. Apart from this modification, the set of monotone equilibria canbe constructed following the same algorithm as in the benchmark model.

14Another possibility is that such defense measures themselves convey valuable information;Angeletos, Hellwig, and Pavan (2006) examined such signaling effects in a static global game.


22/46


PROPOSITION 3: In the game with public signals, the strategy{at()}Tt=0 isa monotone equilibrium if and only if there exists a sequence of functions{xt() t()}Tt=1,withxt:Rt Randt: Rt (0 1),such that:

(i) for allt,at(xt

zt

) = 1ifxt< xt(zt

)and at(xt

zt

) = 0ifxt> xt(zt

);(ii) at t= 1, 1(z1) solves U(1 1 1 z1)= 0 and x1(z1)=X (1(z1) 1);

(iii) at anyt 2,eithert(zt)solves

U(t t1(z

t1) t t zt) = 0(9)

andxt(zt) = X (t(zt) t)ort(zt) = t1(zt1)and xt(zt) = .

As in the benchmark model without public signals, there always exist equi-

libria in which attacks cease after any arbitrary period. However, becausefor any (t1(zt1) t t), (9) admits a solution if and only ifztzt, where

zt=z(t1 t t) is always finite, there also exist equilibria in which an at-tack occurs in periodtfor sufficiently low realizations ofzt, which proves thefollowing.

THEOREM2: In the game with public signals,there always exist multiple equi-libria.

This result extends and reinforces Theorem1: multiplicity now emerges re-

gardless of the mean z of the prior, the precisions{t t}Tt=1 of private andpublic information, and the horizon Tof the game. This stronger version ofmultiplicity relies on the combination of two properties: that sufficiently lowrealizations ofztmake an attack possible in every period and that the lowerdominance region is eliminated in all periods t 2 so that no attack also re-mains possible in every period after the first.

Consider now how the introduction of public news affects the ability of aneconometrician to predict the regime outcome and/or the occurrence of anattack in any given period. For any t, any (0 1), and any t1(zt1) < , con-dition (9) admits a solution higher thanif and only ifztis low enough, imply-

ing that, conditional on, the probability that the status quo is abandoned inany given period is strictly between 0 and 1. It follows that an econometricianwho can observe but cannot observe zt necessarily faces uncertainty aboutthe event of regime change. On the other hand, if he also knows z t, he may beable to predict the regime outcome in a given period for some combinations ofand z t, without, however, being able to predict whether an attack will occuror not. For example, take any t 2, let1(z1)andt(zt)be the lowest and thehighest solutions to U( 1 1 z1) = 0 and U( 1(z1) t t zt) = 0,respectively, and assume that >t(zt) > 1(z1). There is no equilibrium inwhich the status is abandoned in period t, but there exist both an equilibrium in


23/46


which an attack occurs and one in which no attack takes place in that period.15

Therefore, the combination of fundamentalsandpublic information may helppredict regime outcomes but not the occurrence of attacks, as in the bench-

mark model.Also note that the thresholdzt, below which(9) admits a solution, decreaseswith t1. Hence, an unsuccessful attack, other things being equal, causes a dis-crete increase in the probability that the game enters a phase during which noattack is possible. In this sense, the prediction of the benchmark model thatequilibrium dynamics are characterized by the alternation of phases of tran-quility and phases of distress survives the introduction of public news; the nov-elty is that the transition from one phase to another is now stochastic, becauseit depends on the realization ofz t.

5.2. Signals about Past AttacksIn the analysis so far, agents learn from the outcomeof past attacks but re-

ceive no information about the size of these attacks. For many applications,however, it seems natural to allow agents to observe noisy private and/or pub-lic signals about the size of past attacks.

In the online supplement (Angeletos, Hellwig, and Pavan (2007)) we showhow this can be done without any sacrifice in tractability. The key is to main-tain the Normality of the information structure. The algorithm for monotoneequilibria then remains the same as in Proposition3, except for the fact thatthe sequence

{t t

}t

=1 is now part of the equilibrium: the precisions of pri-

vate and public information in any periodt 2 depend on whether an attackoccurred in the previous period.

As for the structure of equilibrium dynamics, the novelty is that the upwardshift in posterior beliefs caused by an unsuccessful attack may now be dilutedby the information about conveyed by the size of the attack. This in turnmay lead to situations where new attacks become possible immediately afterunsuccessful ones, even without any exogenous arrival of information. As aresult, equilibrium dynamics may now feature snowballing effects reminiscentof the ones highlighted in herding models of crises (e.g., Chari and Kehoe(2003)).

5.3. Observable Shocks: Changes in Fundamentals as a Source of Dynamics

In this section we introduce shocks to the sustainability of the regime. In par-ticular, we modify the benchmark model as follows. The regime is abandoned

15Because in any equilibrium necessarily t1(zt1) 1(z1), from part (i) in Lemma1,t(zt)

is an upper bound for t (zt). This implies that if >t(zt) > 1(z1), there is no equilibrium

in which the regime is abandoned in period t. On the other hand, because U ( 1(z1) t tzt)=0 admits a solution =t(zt) > 1(z1), there exists an equilibrium in which the secondattack occurs exactly in period t.


24/46


in period t if and only ifAth( t). The variable continues to repre-sent the strength of the status quo, while t is an exogenous disturbance,independent of and independent and identically distributed over time, with

absolutely continuous cumulative distribution functionFand supportR

. Thescalar > 0 parameterizes the volatility of these disturbances. Finally, for sim-plicity, the function h is assumed to be linear, with h(t) = + t. Wedenote this game with (), nesting the baseline model as = 0.

We assume here that t is publicly observable, which may be relevant forcertain applications of interest. In the case of currency attacks, for example, may represent the type of the central banker, whereas tmay capture therole of interest rates, financial prices, and other macroeconomic variables thatare readily observable by economic agents and that may affect the willingnessor ability of the central banker to defend the peg.16

As we show next, observable shocks are easy to incorporate in the analysis,because they affect equilibrium dynamics without introducing any noise in thelearning about. The exercise here is thus useful not only for applications, butalso for separating the role of shocks as drivers of equilibrium dynamics fromtheir role as additional sources of noise in learning.

Equilibrium characterization

Given that the shocks are observable, the strategy of the agents in period tiscontingent on bothxt and t {1 t}. Accordingly, the regime outcomein periodtis contingent on both andt. A monotone equilibrium can thus

be characterized by a sequence of function{xt() t()}t=1 such that an agentattacks in period tif and only ifxt< x

(t) and the status quo is in place inperiodt+ 1 if and only if > t(t).

Note that sufficiently negative shocks reintroduce the lower dominance re-gion: whenevert1(

t1) + t< 0, it is dominant for agents with sufficientlylow xtto attack. Otherwise, the structure of equilibria and the algorithm forconstructing them are similar to those in the benchmark model.

PROPOSITION 4: In the game with observable shocks, the strategy{at()}t=1is a monotone equilibrium if and only if there exists a sequence of functions

{xt() t()}t=1 withxt:Rt R andt:Rt (0 1)such that:(i) for allt,at(x

t t) = 1ifxt< xt(t)and at(xt t) = 0ifxt> xt(t);(ii) fort= 1, 1(1) solves U(1(1) + 1 1z+ 1) = 0 and

x1(1) = X (t(1) + 1 1) 1;(iii) for anyt 2,eithert(t)solves

U

t(t) + t t1(t1) + t tz + t

= 0(10)16An alternative would have been to allow for shocks in the opportunity cost of attacking by

lettingcdepend ont.


25/46


and xt(t) = X (t(t) + t t) t,ort(t) = t1(t1) t and

xt(t) = .

This result can be understood as follows. Although the critical size of attackthat is necessary for regime change is constant in the benchmark model, hereit varies over time as a consequence of shocks. However, because shocks areobservable, the structure of beliefs remains the same apart from a change ofvariables in the following sense. Letht + tbe the critical size of attackfor periodt, letxt xt+ t, and letz t z + t. The distribution ofhtcon-ditional onxt is Normal with mean t/(t+ )xt+ /(t+ )ztand precisiont + , while the knowledge that > t1(t1)is equivalent to the knowledgethatht>

t1(

t1) + t. It follows that the net payoff from attacking for themarginal agent in periodt isU (t(

t) + t t1(t) + t; tz + t).The result then follows from the same arguments as in the proof of Proposi-tion2.

Multiplicity and dynamics

Interesting new effects can emerge because of the interaction of informa-tion and shocks. The equilibrium dynamics again feature phases of tranquility,where an attack is impossible, and phases of distress, where an attack is pos-sible but does not necessarily take place. However, shocks provide a secondchannel through which a transition from one phase to another can occur. Inparticular, a transition from distress to tranquility may now be triggered ei-

ther by an unsuccessful attack or by an improvement in fundamentals (a pos-itivet), and a transition from tranquility to distress can be caused either bythe arrival of new private information or by a deterioration in fundamentals(a negative t). What is more, the economy can now enter a phase where anattack is inevitablea scenario that was impossible in the benchmark model,but becomes possible here because sufficiently bad shocks reintroduce a lowerdominance region.17

If the benchmark game (0)admits multiple equilibria, then the game withshocks ()also admits multiple equilibria, regardless of; to see this, it suf-fices to consider realizations oftclose enough to zero. Moreover, becausethe impact of shocks on the conditions that characterize the equilibrium dy-namics clearly vanishes as 0, the following equilibrium convergence resultholds: for anyT > 0, any >0, and any equilibrium{xt t}t=1 of the bench-mark game (0), there exists a >0 such that for all


26/46


None of these results, however, should be surprising given that shocks do notinterfere with the learning process. Indeed, what is important for the results ofthis section is not the absence of uncertainty about the shocks, but the fact that

the shocks do not introduce noise in learning.To see this, consider the case that t is unobservable in period t but be-comes commonly known at the beginning of period t+1. Then agents faceadditional uncertainty about the regime outcome but the knowledge that theregime has survived past attacks still translates into common certainty thatisabove a certain threshold. That is, the form of learning remains as sharp as inthe benchmark model. Not surprisingly then, the aforementioned equilibriumconvergence result extends to this case as well.19

We conclude that with respect to robustness the question of interest iswhether equilibrium convergence obtains in situations where shocks also in-troduce noise in learning. To examine this question, we next turn to the casethattremains unobservable in all periods.

5.4. Unobservable Shocks: Noisy Learning

We now modify the game with shocks examined in the previous sectionby letting tbe unobservable. The unobservability of shocks noises up thelearning process and ensures that the updating of beliefs caused by the knowl-edge that the regime is still in place never takes the form of a truncationagents posteriors have full support in Rin all periods. The case of unobserv-able shocks that we examine in this section is therefore most significant from a

theoretical perspective.In what follows, we first explain how unobservable shocks affect the algo-

rithm for the construction of equilibria. We then show how the equilibria inthe benchmark model can be approximated arbitrarily well by equilibria of theperturbed game as forsmall enough. It follows that the key qualitative prop-erties of the equilibrium dynamics identified in the benchmark modelthemultiplicity and the succession of phases of tranquility and distressextend tothe case with unobservable shocks provided that the volatility of these shocksis small enough and that we reinterpret a phase of tranquility as one where atmost an (arbitrarily) small attack is possible.

Equilibrium characterization

Becausetaffects the regime outcome and is unobserved, the possibility tocharacterize the set of monotone equilibria in terms of a sequence of trunca-tion points for is lost. Nevertheless, as long as private information can besummarized by a sufficient statisticxtR, we can still characterize monotoneequilibria as sequences of thresholds{xt}t=1 such that an agent attacks in pe-riodtif and only ifxt xt, wherext R.

19This case is also examined in the online supplement (Angeletos, Hellwig, and Pavan ( 2007)).


27/46


To see this, consider an arbitrary monotone strategy, indexed by the se-quence of thresholds{xt}t=1, such that an agent attacks in periodtif and onlyif xt 0, learning is smoother in the sense that t(;xt1) is strictly positiveand continuous over the entire real line.

Finally, consider payoffs. For anyt 1, x R, andxt Rt, letvt (x;xt)de-note the net expected payoff from attacking in periodtfor an agent with suffi-

cient statisticsx

when all other agents attack in period

tif and only if theirsufficient statistic in is less than or equal tox. This is given by

vt (x;xt) = +

pt(;xt)t(|x;xt1) d c

where t(|x;xt1) denotes the density of the private posterior in period t.(The latter is computed applying Bayes rule to the common posteriorst(;xt1).) Note that vt (x;xt) depends on both the contemporaneous thresh-old

xt and the sequence of past thresholds

xt1; the former determines the


28/46


probability of regime change conditional on , whereas the latter determines

the posterior beliefs about. Next, for anyt 1 andxt Rt, let

Vt (xt)

limx+ v

t (x;xt

) ifxt=+,vt (xt;xt) ifxtR,lim

xvt (x;xt) ifxt=.

(11)

The functionVtis the analogue of the function Uin the benchmark model: itrepresents the net payoff from attacking in period tfor the marginal agent withsufficient statisticxt=xt.

In LemmaA2in theAppendix,we prove that, for any > 0,Vt (xt)is con-

tinuous in

xt for any

xt

R

t1

R, which we use to establish the existence, and

complete the characterization, of monotone equilibria.

PROPOSITION5: For any > 0, the strategy{at()}t=1 is a monotone equilib-rium for () if and only if there exists a sequence of thresholds {xt}t=1such that:

(i) for allt,at(xt) = 1ifxt< xt andat(xt) = 0ifxt> xt;

(ii) fort= 1,x1 R andV1 (x1) = 0;(iii) for any t 2,eitherxt= and Vt (xt) 0orxt Rand Vt (xt) = 0.

A monotone equilibrium exists for any > 0.

The equilibrium algorithm provided by Proposition5clearly applies also to

= 0 and is similar to the one in Proposition2:start with t= 1 and let x1 bethe unique solution to V1 (x

1)=0; proceed to t=2 and either let x2=

ifV2 (x1) 0 or let x2 be the solution to V2 (x1 x2) = 0; repeat for any

t 3. The difference is that here at each step twe need to keep track of theentire sequence of past thresholds xt1, while in the algorithm in Proposition2the impact ofxt1 on period-tbeliefs was summarized byt1.

Multiplicity and dynamics

As 0, the dependence of the regime outcome on the shock t van-ishes. By implication, the posteriors in any period t

2 converge pointwise

to truncated Normals as in the benchmark model. The pointwise convergenceofpt and

tin turn implies pointwise convergence of the payoff of the mar-

ginal agent: fort= 1 and anyx1,V1 (x1) V01(x1) U(1(x1) 1z);similarly, for anyt 2,xt1 andxt> ,

Vt (xt) V0t (xt) U

t(xt)t1(xt1) tz

Pointwise convergence of payoffs, however, can fail for t 2 atxt= . Tosee why, note that in the presence of shocks, an agent with sufficiently low xtmay attach probability higher than c to regime change in period t

2 even


29/46


if he expects no other agent to attack in that period. When this is the case,a positive measure of agents may attack in every period in the perturbed game,unlike the benchmark model.

Nevertheless, the pointwise convergence of V

t (xt

) for any xt

such thatxt> ensures that this dominance region vanishes as 0. It also ensuresthat givenxt1, wheneverV0t (xt)has an intersection with the horizontal axis atxt=xt,Vt also has a nearby intersection for > 0 small enough. These prop-erties imply that any equilibrium in the benchmark game can be approximatedarbitrarily well by an equilibrium in the perturbed game, except for knife-edgecases whereV0t (or equivalentlyU) is tangent to the horizontal axis instead ofintersecting it.

THEOREM3: For any > 0and anyT < ,there exists( T) > 0such thatthe following is true for all < (T ):For any equilibrium{xt}t=1of (0)such thatxt / argmaxx V0t (xt1 x)for allt {2 T },there exists an equilibrium{xt}t=1 of ()such that,for allt T,|xt xt| < ifxt Randxt < 1/ifxt= .

The result is illustrated in Figures3and4for an example where T= 2 andwhere (0) admits multiple equilibria. The solid line in Figure3representsthe probability density function of the common posterior in period 2 gener-ated by equilibrium play in period 1 in the game without shocks ( = 0). Thisis simply the initial prior truncated at 1=1(x1), where x1 is the uniquesolution to V

0

1(x1)=0 (or equivalently where 1 is the unique solution toU(1 1z) = 0). The other two lines represent the equilibrium com-mon posteriors 2 (; x1 ) for the game with shocks ( >0), where x1 is theunique solution to V1 (x

1 ) = 0; the dotted line corresponds to a relatively high

and the dashed one to a low . Because the support oftis the entire real line,the probability of regime change is less than 1 for any and therefore 2 isstrictly positive for all . However, as becomes smaller, x1 converges to x

1

and the probability of regime change at t=1 converges to 1 for < 1 andto 0 for > 1 . By implication, the smooth posterior of the perturbed game inperiod 2 converges to the truncated posterior of the benchmark model.

In Figure4, the solid line represents the payoffV0

2(x

1 x

2)

of the marginalagent in period 2 for = 0, whereas the other two lines represent the payoffV2 (x

1 x2)for > 0.

20 Note that forx2 small enough, V0

2 is negative butV

2 ispositive, which implies that nobody attacking in period 2 is part of an equilib-rium in the benchmark model but not in the game with shocks.21 Moreover,

20To illustrate V2 over its entire domain, the figure depicts V

2 (x1 x2) against f (x2) rather

than x2, where f is a strictly increasing function that maps R onto a bounded interval (e.g.,f= ).

21In this example an agent finds it dominant to attack in period 2 for sufficiently lowx2. How-ever, this need not be the case ifthas a bounded support and is small enough.


30/46


FIGURE3.Common posteriors with and without shocks.

when is high (dotted line), V2 is monotonic in x2 and therefore has a singleintersection with the horizontal line, in which case the equilibrium would beunique if the game ended in period 2. When, instead, is sufficiently small(dashed line), V2 is nonmonotonic and has three intersections, which corre-spond to three different equilibria for the two-period game with shocks. Themiddle and the highest intersections approximate the two intersections of thesolid line, while the lowest intersection is arbitrarily small, thus approximatingx2= . Along with the fact that x1 converges to x1, this implies that anyequilibrium of the two-period game without shocks can be approximated by anequilibrium of the perturbed game.

At the same time, it is important to recognize that uniqueness is ensured inthe alternative limit as2 for given > 0. To see this, note that, for anygiven >0, the common posterior in period 2 has a strictly positive density

FIGURE4.Payoff of marginal agent with and without shocks.


31/46


over the entire real line and hence over a connected set ofthat includes bothdominance regions ( < 0 and > 1). This in turn ensures that standard global-game uniqueness results apply: uniqueness necessarily obtains in the limit as

the noise in private information vanishes (see Proposition 2.2 in Morris andShin (2003)).However, away from this limit, the impact of private information here is

quite different from that in the static Gaussian benchmark examined in theliterature (Section2.1). There, private information always contributes towarduniqueness.22 Here, instead, it can have a nonmonotonic effect on the deter-minacy of equilibria: forsmall enough, in period 2 uniqueness obtains for2either close to1or close to , while multiplicity obtains for intermediate2.

This nonmonotonic effect of private information in our dynamic game high-lights the interaction of private information with equilibrium learning. On the

one hand, more precise private information increases strategic uncertainty asin the static game; on the other hand, it dilutes the upward shift in posteriorbeliefs caused by the knowledge that the regime survived past attacks. Whereasthe first effect contributes to uniqueness, the second can contribute to multi-plicity in a similar fashion as in the benchmark game without shocks.

In conclusion, what sustains multiplicity in the dynamic game is the propertythat the knowledge that the regime survived attacks in the past provides rele-vant common information about the strength of the status quo in the present.That this knowledge resulted in posterior beliefs that assign zero measure tosufficiently low in the benchmark model is not essential. What is important is

that the effect of this information on posterior beliefs is not diluted too mucheither by a significant change in fundamentals (sufficiently high ) or by a sig-nificant increase in strategic uncertainty (sufficiently high).

5.5. Private Information about Shocks: Long- versus Short-Lived Agents

In the environments examined in the preceding subsections, agents had noprivate information about the shocks. Relaxing this assumption may compro-mise tractability by removing the ability to summarize the history of privateinformation with a one-dimensional sufficient statistic.

To see this, consider the following variation of the game with shocks. Letht denote again the critical size of attack that triggers regime change inperiod t and assume that{ht}t=1 are jointly Normal with nonzero correla-tion across time. To simplify, think ofhtfollowing a Gaussian random walk:h1=N(z 1/) and ht= ht1+t for t2, with tN (0 1) being

22This is true in two senses. First, a higher makes it more likely that the game satisfies 2/(2), in which case the equilibrium is unique. Second, whenever < 2/(2), the range ofz for which (1) and(2)admit multiple solutions shrinks with , and the distance between thelargest and the smallest solutions for any givenz also diminishes with.


32/46


independent and identically distributed across time and independent of.23

Next, let the private signals agents receive in period tbext= ht+ t, wheretis independent and identically distributed across agents and time, and is inde-

pendent ofhsfor anys.Consider firstt= 1. Equilibrium play is the same as in the static benchmark:there exist thresholds x1and h

1such that an agent attacks if and only ifx1 x1

and the status quo is abandoned if and only ifh1 h1. Consider next t= 2. Theposterior beliefs about h2 given the private signalsx1 andx2 alone are Nor-mal with mean x20+ 1x1+ 2x2 and variance 22 , for some coefficients(0 1 2 2). This may suggest that x2 can be used as a sufficient statisticfor(x1x2)with respect to h2. However, the posterior beliefs abouth2 condi-tional also on the event thath1> h

1 are not invariant in(x1x2)for givenx2;

the reason is that x2 isnot a sufficient statistic for (x1x2)with respect to h1.Thus, private information cannot be summarized inx2and equilibrium play inperiod 2 is characterized by a function f2 :R

2 Rsuch that an agent attacksif and only if f2(x1x2)0 (and a corresponding function g2 :R2 R suchthat regime change occurs if and only ifg2(h1 h2) 0). Similarly, equilibriumplay in any period t2 is characterized by a function ft:Rt R such thatat(x

t) = 1 if and only ifft(xt) 0.Contrast this with the formalization in the previous section. There, in each

period, we had to solve an equation where the unknown was a real numberxt.Here, instead, we need to solve each period a functional equation where theunknown is a function ft with domain R

ta function whose dimensionality

explodes witht. Clearly, this is far less tractable, if at all feasible.Moreover, it is not clear if this alternative formalization brings any substan-tial gain from a theoretical perspective. Both formalizations ensure that thecritical size of attack ht(and hence the payoff structure) may change over time,that agents have asymmetric information about htin each period, and that thecommon posterior about ht is continuous over a connected set that includesboth ht 1 (and hence that dominance regions are possible forboth actions). In these respects, they both seem to be appropriate extensionsof global games to a dynamic setting.24

Nevertheless, this second formalization may be more appropriate for certainapplications. One way then to restore tractability is to assume that agents areshort-lived. In particular, consider the foregoing game in which ht follows aGaussian random walk, with the following modification. As long as the statusquo is in place, a new cohort of agents replaces the old one in each period.Each cohort is of measure 1 and lives exactly one period. Agents who are born

23Note that this is the same as ht= + t, where t 1+ + t; that is, the same asin the game with shocks in Sections5.3and5.4,but with the shocks correlated across time. Forsimplicity, such correlation was not allowed in Sections5.3and5.4.

24Note that global-game results do not require that agents have private information about allpayoff-relevant variables or that uncertainty vanishes in the limit forallpayoff-relevant variables.


33/46


in period treceive private signals xt=ht+ t about ht, where t is Normalnoise with precision t, independent and identically distributed across agentsand independent ofhs for anys t.

Given thathtis correlated across time, the knowledge that the regime sur-vived past attacks is informative about the strength of the regime in the present,as in the case with long-lived agents. However, unlike that case, agents whoplay in period thave no private information other than xt. Together with thefact that htalone pins down the cross-sectional distribution ofxt, this prop-erty ensures that monotone equilibria can again be characterized by sequences{xt ht}t=1 such that in period tan agent attacks if and only ifxt< xtand thestatus quo is abandoned if and only ifht< h

t.

The characterization of equilibria then parallels that in the previous sec-tion. To see this, let t(ht;xt1)denote the cumulative distribution function

of the common posterior in periodtabout

ht, when agents in earlier cohorts at-tacked in periods t1 if and only ifx(x),where(x) is the solution to (

(x h)) = h. Therefore, for any t 2,

t(ht;xt1)is recursively defined by

t(ht;xt1) =+

t1(xt1)((ht ht1)/)dt1(ht1;xt2)1 t1(t1(xt1);xt2)

(12)

with1 (h1) = (

(h1 z)). Next, lett(ht|x;xt1)denote the cumulativedistribution function of private posteriors about ht; this is obtained by applyingBayes rule to (12). Then the expected net payoff from attacking in period tforan agent with signalx is v1 (x;x1) = 1 (1(x1)|x) cfort= 1 and

vt (x;xt) = t(t(xt)|x;xt1) c(13)fort 2. Finally, letVt denote the payoff of the marginal agent, as defined incondition (11), but using the functionvtas in (13). With V

t defined this way,

the equilibrium algorithm of Proposition5applies to the environment exam-ined here as well. What is more, because beliefsand hence payoffsagainconverge to their counterparts in the benchmark game as

0, Theorem3

also applies. (See the online supplement Angeletos, Hellwig, and Pavan (2007)for details.)

The game with short-lived agents thus permits one to examine environmentswhere agents have private information about the innovations in fundamentalswhile maintaining tractability.

6. CONCLUSION

This paper examined how learning influences coordination in dynamic globalgames of regime change. Our results struck a delicate balance between the ear-


34/46


lier common-knowledge and the more recent global-games literature: the dy-namics featured both a refined role for multiplicity and a certain discontinuityof outcomes with respect to changes in information or payoffs. They also led

to novel predictions, such as the possibility that fundamentals predict eventualoutcomes but not the timing and number of attacks, or that dynamics alter-nate between phases of tranquility, during which agents accumulate informa-tion and no attacks are possible, and phases of distress, during which attacksmay occur but do no necessarily take place.

From a methodological perspective, our results offer two lessons with regardto the recent debate about uniqueness versus multiplicity in coordination envi-ronments. First, that equilibrium learning can be a natural source of multiplic-ity in a dynamic setting, despite the heterogeneity of beliefs. Second, and mostimportantly, that this debate may dilute what, at least in our view, is the cen-

tral contribution of the global-games approach: the understanding of how thestructure of beliefs can lead to interesting and novel predictions about equilib-rium behavior well beyond equilibrium determinacy.

From an applied perspective, on the other hand, the predictions we derivedmay help one to understand the dynamics of currency attacks, financial crashes,political change, and other crises phenomena. With this in mind, in Section5we sought to give some guidance on how the analysis can be extended to ac-commodate certain features that were absent in the benchmark model but maybe important for applications. The scope of these extensions, however, was lim-ited to changes in information or in fundamentalswe remained silent aboutother dynamic effects (such as those introduced by irreversible actions or liq-uidity constraints), as well as about the role of large players (such as that of aSoros or a policy maker). Extending the analysis in these directions presentsa promising line for future research.

Dept. of Economics, MIT, 50 Memorial Drive, Cambridge, MA 02142, U.S.A.;and National Bureau of Economic Research;[email protected],

Dept. of Economics, UCLA, Box 951477, CA 90095-1477, U.S.A.;[email protected],

andDept. of Economics, Northwestern University, 2001 Sheridan Road, Evanston,

IL 60208-2600, U.S.A.;[email protected] received December, 2004; final revision received July, 2006.

APPENDIX: PROOFSOMITTED IN THEMAINTEXT

PROOF OFPROPOSITION1: Solving(1) forx givesx =+1/21(). Sub-stituting this into (2) gives a single equation in,

Ust(

;z)

=0(14)
mailto:[email protected]:[email protected]:[email protected]:[email protected]:[email protected]:[email protected]:[email protected]:[email protected]


35/46


where

Ust(

;z)

1

+ 1()

+

(z

)

c(15)

Note that Ust(; ) is continuous and differentiable in (0 1), withlim0 Ust() = 1 c > 0 and lim1 Ust() = c < 0. A solution to (14) there-fore always exists. Next, note that

Ust(; )

=

+

+

1() +

(z )

1

(1())

Given that min(01)[1/(1())] =

2, the condition 2/(2) is bothnecessary and sufficient for Ust to be monotonic in , in which case themonotone equilibrium is unique. Finally, for the proof that only this equilib-rium survives iterated deletion of strictly dominated strategies, see Morris andShin (2001,2003). Q.E.D.

PROOF OF PROPOSITION 2: Necessity follows from the arguments in themain text. For sufficiency, take any sequence{xt t}t=1 that satisfies condi-tions (ii) and (iii), let 0

= , suppose all other agents follow strategies as

in (i), in which case Rt=0 if and only if > t1 for all t1, and considerthe best response for an agent. Ift= t1, in which case t 2, t1> 0, andxt= , then Pr(Rt+1= 1|xt Rt= 0) = Pr( t|xt > t1) = 0 for all xtand therefore not attacking is indeed optimal. If instead t >

t1, in which

caseU (t t1 tz) = 0 and xt= X (t t), then by the monotonicity of

the private posterior in

Dynamic Global Games of Regime Change

Documents

Transcript of Dynamic Global Games of Regime Change