World Bank Document

WPS 2-2 7

POLICY RESEARCH WORKING PAPER 2267 Public Disclosure Authorized

Do High Interest Rates No - there is no systematic associationbetween interest Defend Currencies during ratesand the outcome of SpeculativeAttacks? speculativeattacks. Public Disclosure Authorized

Aart Kraay Public Disclosure Authorized Public Disclosure Authorized The World Bank Development Research Group Macroeconomicsand Growth U January 2000 a tPQmI.CY RESEARCH WORKING PAPER 2267

Summary findings

Drawing on evidence from a large sample of speculative The lack of clear empirical evidence on the effects of attacks in industrial and developing countries, Kraay high interest rates during speculative attacks mirrors the argues that high interest rates do not defend currencies theoretical anmbiguitieson this issue. against speculative attacks. In fact, there is a striking lack of any systematic association between interest rates and the outcome of speculative attacks.

This paper - a product of Macroeconomics and Growth, Develcpment Research Group - is part of a larger effort in the group to study the causes and consequences of financial crises. Copies of the paper are available free from the World Bank, 1818 H Street,N'W, Washington, DC 20433. Please contact Rina Bonfield, room MC3-354, telephone 202-473-1248, fax 202-522-3518, email address [email protected]. Policy Research Working Papers are also posted on the Web a!t www.worldbank.org/research/workingpapers. The author may be contacted at akraay@

The Policy Research Working Paper Series dissemninatesthe findings of work in progress to encourage the exchange of ideas about | development issues. An objective of the series is to get the findings out quickiy, even if the presentations ar-eless than f7lly polished. Th2e papers carry the names of the authors and should be cited accordingly. The findings, interpretations, aisd conclusions expressed in this paper are entirely those of the authors. They do uot necessa-ily represent the view of the World Bank, its Executive Directors, or the countries they represent.

Produced by the Policy Research Dissemination Center Do High Interest Rates Defend Currencies During Speculative Attacks?

Aart Kraay The World Bank

1818H Street, N.W. Washington,DC 20433, (202)473-5756, [email protected]. The opinionsexpressed in this paperare the author's,and do not reflectthose of the World Bank, its executivedirectors, or the countriesthey represent. I would liketo thank Alan Drazen,Ilan Goldfajn,Patrick Honohan, Vickie Kraay, MariaSoledad MartinezPeria, Sergio Schmukler, Jakob Svensson, Jaume Ventura and seminar participantsat MIT, the TinbergenInstitute, and the World Bankfor helpfulcomments.

1. Introduction

Accordingto conventionalwisdom, currencies that comeunder speculative attack can be defendedwith high interestrates. By raising interestrates high enough, the conventionalwisdom argues that the monetaryauthority can make it prohibitively costlyfor speculatorsto take short positionsin the currencyunder attack. High interest rates are often alsosaid to conveya positivesignal regardingthe commitmentof the monetaryauthority to maintaininga fixed exchangerate. To the extentthat this signal alters the expectationsof foreign exchangemarket participants,high interestrates can serve to strengthenthe domesticcurrency. A classicexample in supportof the conventionalwisdom is the responseto the attackon the Swedishkrona in the summer of 1992,shown in the top panelof Figure 1. BetweenJuly and August,speculative pressuresagainst the krona resultedin a loss of nearlyone-quarter of the reservesof the Swedishcentral bank. To stemthe outflow,the centralbank's marginal lending rate was raisedto an incredible500 percenton September17 and 18, and hoveredin the vicinity of 50 percentfor the nextweek. Reservelosses were promptlyhalted, and the krona'speg was maintained.

Recently,a contrarianview of the effects of high interestrates duringspeculative attackshas emerged,which calls into questionboth tenets of the conventionalwisdom. First, it notesthat interestrates haveto be increasedto very high annualizedrates in order to enticeeven risk-neutralinvestors to hold local currency-denominatedassets in the face of a small expecteddevaluation over a short horizon,and suchextremely high interestrates are rarely observedin practice.' Second,this view notesthat the signaling value of high interestrates is unclear. Althoughsignals mustbe costly in orderto be credible,often they imposecosts that are too high for the monetaryauthority to take in stride. If marketparticipants know that the monetaryauthority is concernedabout the contractionaryeffects of high interestrates on domesticeconomic activity, they are unlikelyto believethat rates will be kept high enough,and for long enough,to deter speculation.Worse, as the costsof high interestrates mount,the monetaryauthority's signalcan becomeless credibleover time, raisingdevaluation expectations. A vicious

1For example,a risk-neutralinvestor expecting only a 0.5 percentovernight depreciation would requiie an overnightannualized rate of returnof 500 percenton domesticcurrency to compensatefor the expected devaluation.

1 spiral can result,as expectationsof a devaluationforce higher interestrates, which in turn imposegreater costs on the economy.2An exampleconsistent with this contrarian view of the effects of high interestrates is Koreain the secondhalf of 1997,shown in the lower panelof Figure 1. As the EastAsian financial crisis spread from Thailandand Malaysia,speculative pressures against the Koreanwon intensifiedand the reservesof the Koreancentral bank fell from 35 billionto 25 billion US dollars betweenJune and November.Although the overnightcall rate was raisedfrom around 12 percentin early Novemberto over 30 percentby the end of December,the won fell by over 50 percent duringthis period.

In light of thesetheoretical ambiguities and conflictinganecdotes, this paper asks whetherthere is any systematicempirical evidence in supportof the conventional wisdomregarding the effects of high interestrates during speculativeattacks. To answerthis question,I study the behaviourof interestrates around a large numberof successfulspeculative attacks (i.e. attacksthat end in a sharp nominaldevaluations) andfailed speculativeattacks (i.e.attacks that did not end in a devaluation)in a sample of 75 developedand developingcountries over the period 1960-1997. I examine whetherinterest rates rise duringfailed speculaltiveattacks (i.e. whetherraising interest rates is necessaryto preventa speculativeattack from endingin a devaluation),and whetherraising interest rates increasesthe probabilitythat an attack fails (i.e.whether raisinginterest rates is sufficientto preventa speculativeattack from endingin a devaluation).

This empiricalexercise faces threedifficulties: measuring the policy responseto a speculativeattack, accountingfor possiblenon-linearities in the effectsof the policy response,and controllingfor the endogeneityof the policy response. First, it is difficult to disentanglethe monetarypolicy response to a given speculativeattack from other sourcesof variationin observedmarket interest rates duringthe attack. For example, increasesin marketinterest rates duringa speculativeattack might reflectboth a tighteningof domesticcredit by the monetaryaLuthority, and also an increasein the

2Drazen and Masson(1994) developa modelin whichsignals become less credibleover time. Bensaid and Jeanne(1997) formalize devaluation spirals. Radeletand Sachs(1998) and Furmanand Stiglitz (1998)discuss other reasonswhy tighter monetarypolicy can weaken,rather than strengthen,the currency underattack.

2 devaluationexpected by marketparticipants. In order to obtain a direct measureof the monetarypolicy responseto speculativepressures, I rely primarilyon changesin interestrates underthe controlof the monetaryauthority (i.e. centralbank discount rates) as a measureof policy. A drawbackof this measureis that discountrates are only one of many instrumentsthat the monetaryauthorities have at their disposalto resist speculativepressures. I thereforealso check the robustnessof the resultsusing a variety of other noisierindicators of the stance of monetarypolicy.

Second,there may be importantnon-linearities in the effects of interestrates on speculativepressures, and ultimatelyon the outcomeof the attack. For example,the credibilityof the monetaryauthority's signal of its intentto defend the currencymay dependon the economy'sability to withstandthe contractionaryeffects of tight monetary policy,or on the quantityof reservesheld by the monetaryauthority. In this case, simplecorrelations between measures of monetarypolicy and the outcomeof speculativeattacks may obscureany effectsof policy presentonly in certain subsamplesof speculativeattacks. I take into accountthe possibilityof episode-specific variationin the effects of monetarypolicy by splittingthe samplealong various dimensions,and by interactingmeasures of monetarypolicy with episode-specific characteristics. 3

Third and perhapsmost important, the policy decisionsof the monetaryauthority are themselvesendogenous, and are likely to dependon both episode-specific characteristicsthat determinespeculative pressures, and on speculativepressures themselves.Consider an economythat is vulnerableto a speculativeattack, perhaps becauseits real exchangerate is overvaluedor its reservesare low relativeto its short- term obligations. If attackson vulnerablecurrencies are both more likelyto succeed, and also are more likely to provokea stronginterest rate defense on the part of a "tough"monetary authority committed to maintainingthe fixed exchangerate, onemight expectto find large increasesin interestrates during successfulattacks, and conversely,small increasesin interestrates duringfailed attacks. This endogeneity

3A secondpossible source of non-linearitiesis in the time dimension,if, for example,the signalingvalue of tight monetarypolicy becomes less credibleover time. Since I will be relyingon the relativelylow-frequency monthlydata availablefor this largesample of speculativeattacks, there is unlikelyto be enoughtime seriesvariation in each episodeto identifynon-linearities over time in the effectsof monetarypolicy.

3 problemmay obscurethe positiveeffects of high interestrates on investorconfidence and the probabilitythat an attackfails. It is also possiblethat the endogeneitybias exaggerates,rather than obscures,the conventionalwisdom regarding effects of high interestrates. For example,if the monetaryauthority is "realistic"and determinesthat it is futile to try to defenda highlyovervalued currency, but is willing to vigourouslydefend the currencywhen it believesfundamentals are sound,there may be a positive associationbetween high interestrates and failed attacksdriven by common fundamentals.In this paper,I presenta simplemodel which formalizesthis endogeneity problem,and motivatespossible instruments for the monetarypolicy response. I then use theseto controlfor the endogeneityof policy in a probitspecification which expressesthe probabilitythat a speculativeattack fails as a non-linearfunction of policy, episode-specificcharacteristics, and interactionsbetween the two.

The empiricalresults are not very supportiveof the conventionalwisdom that high interestrates defendcurrencies during speculative attacks. I find no evidencethat interestrates systematicallyincrease during failed speculativeattacks, nor that raising interestrates increasesthe probabilitythat a speculativeattack fails. I obtain the same resultsif I consideralternative measures of monetarypolicy, as well as possiblenon- linearitiesin the effectsof monetarypolicy due to differencesin a varietyof episode- specificcharacteristics. The lack of evidenceorn the efficacyof monetarypolicy during speculativeattacks persists even after I controlfor possiblebiases induced by the endogeneityof policy. Althoughthere appearsto be little evidencein supportof the conventionalwisdom, there is also little evidencein supportof the contrarianview that raising interestrates weakens currencies under speculativeattack. In fact, the main findingof this paper is the strikinglack of any associationwhatsoever between changes in variousmeasures of monetarypolicy and the outcomeof speculativeattacks.

This evidencecontributes to a small but growingempirical literature on the role of monetarypolicy during speculativeattacks. 4 Goldfajnand Gupta (1999) focus on the

Thereis of coursea largeliterature on the effectivenessof interventionsin foreignexchange markets (see Edison(1993) for a survey).Various authors have also appliedVAR methodologiesto estimatethe effects of monetarypolicy shocks on exchangerates. Thesepapers, which focus on normaltimes as opposedto the periodsof speculativepressures considered in this paper,find mixedresults. Eichenbaumand Evans (1995)and Cushmanand Zha (1997)find that positiveinnovations to monetarypolicy lead to depreciations of the domesticcurrency for the US and for Canada,respectively. In contrast,Sims (1992) and Grilli and Roubini(1995) find mixedevidence in the G5 and G7 economies,respectively, with positivemonetary

4 role of interestrates in the aftermathof largedevaluations that result in an undershootingof the real exchangerate. They ask whetherhigh interestrates following a devaluationincrease the likelihoodthat realexchange rate equilibriumis restored througha nominalappreciation rather than throughhigher inflation.They find that high interestrates are effectivein this senseonly in countrieswith strong bankingsectors. Furmanand Stiglitz(1998) examinedaily data on interestrates and exchangerates in a sampleof nine developingcountries during the 1990s to identifyepisodes of sustained high interestrates, and then ask whetherthese werefollowed by an appreciationof the domesticcurrency. They find little evidencethat this is the case. In contrast Dekle, Hsiao andWang (1999a,b) study the relationshipbetween interest rates and exchange rates usingweekly data for Korea,Malaysia and Thailand during 1997and 1998,and argue in favourof the conventionalview. The maindifficulty with all of these papersis that they simply documentreduced-form (partial) correlations between interest rates and exchangerates. Without controllingfor the endogeneityof the monetarypolicy response,it is difficultto infer anythingregarding the effects of high interestrates from these papers. This paper makesa first attemptto take seriouslythe identification problem,drawing on a much largersample of successfuland failed speculativeattacks. 5

The remainderof this paperproceeds as follows. In Section2, I describethe data andthe methodologyused to identifysuccessful and failed speculativeattacks. In Section3, I presentsome descriptiveresults, which providescant evidenceof any associationbetween changes in interestrates andthe outcomeof speculativeattacks. In Section4, I developa simplemodel to illustratethe endogeneityproblem, and I use this to motivatea set of probit regressionsexpressing the probabilitythat speculative attacksfail as a non-linearfunction of policiesand fundamentals. After instrumenting for the endogeneityof policy,I againfind no evidenceof a significantimpact of high interestrates on the outcomeof speculativeattacks. Section5 offers concluding remarks. shocksleading to appreciationsin somecountries and depreciationsin others. Finallythere is a large empiricalliterature documenting the propertiesof macroeconomicvariables around speculative attacks (e.g.example Eichengreen, Rose and Wyplosz(1994,1995,1996)), which to date has not focusedon the policyand non-policydeterminants of successfuland failed attacks. This concernwith the endogeneityof monetarypolicy is of coursenot new,and is a recurringtheme in the literatureon the effectsof monetarypolicy during normal times (as opposedto periodsof speculative pressures).See for examplethe discussionin Bemankeand Mihov(1998) and Christiano,Eichenbaum and Evans(1998).

5 2. IdentifyingSpeculative Attacks

I identifysuccessful speculative attacks as large nominaldepreciations preceded by relativelyfixed nominalexchange rates. I beginwith an unbalancedpanel of monthly observationson nominalexchange rates (expressedin local currencyunits per US dollar) and non-goldreserves. The sampleconsists of 75 middle-and high-income countrieswith populationsgreater than I million,over the period January1960 to April 1999. Detailsof the data can be found in the Appendix. I first ideritifyall episodesin whichthe one-monthdepreciation rate (i.e. the increasein the nominalexchange rate) exceeds10%, which is roughlytwo standarddeviations above the meanmonthly depreciationrate for the entirepooled sample of monthlyobservations. In orderfor these large depreciationsto be meaningfullyconsidered successful speculative attacks, it is necessarythat the exchangerate be relativelyfixed prior to the depreciationitself.6 Accordingly,for eachobservation I constructan averageover the previoustwelve monthsof the absolutevalue of percentagechainges in the nominalexchange rate. I then eliminateall large depreciationepisodes fcr which this averageexceeded 2.5%, or aboutone half of one standarddeviation from the mean for the entiresample. In order to avoid double-countingprolonged crises in which the nominalexchange rate depreciatessharply for severalmonths, I further eliminatesuccessful attacks that were precededby successfulattacks in any of the prior twelve months. Finally,I discardall speculativeattack episodes for whichthere is no data availableon any of the variablesI will useto measurethe monetarypolicy responseto the attack. This resultsin 105 usablesuccessful speculative attack episodes.

I identifyfailed speculativeattacks using two indicatorsof speculativepressures: sharpreserve losses, and sharp increases in nominalmarket interest rates. Specifically, I considerall episodesin which the monthlydecline in non-goldreserves measured in US dollars(the increasein the nonminalmoney niarket rate spread overthe US Federal Fundsrate) exceeds20% (exceeds5%), which is abouttwo standarddeviations above the meanchange for the entiresample. Enorder to restrict attentionto speculative

6I only requirethe exchangerate to be 'relatively"fixed prior to the devaluationfor two reasons. First,this enablesme to identifythe abandonmentof narrowtarget zone exchangerates regimes as well as of fully fixedexchange rate regimes. Second,this allowsme to identifycurrencies that are peggedagainst currenciesother than the US dollarwhose value relativeto the US dollardoes not fluctuatemuch (e.g.the Germanmark).

6 pressuresagainst relatively fixed exchangerates, I eliminateall those episodesfor whichthe same movingaverage of absolutevalues of changesin the nominalexchange rate as beforewas greaterthan 2.5%. Next,to avoid double-countingsuccessful attacks,I excludeall episodesin whichthe changein the nominalexchange rate in the same monthor any of the threefollowing months was greaterthan 10%. I definethese episodesas failed speculativeattacks and, as before, I eliminateall failed attacksthat are precededby a failed attackin any of the twelveprevious months, and those episodesfor which indicatorsof the monetarypolicy responseare not available.This results in 203 instancesof failed speculativeattacks.

Relyingon reservelosses and increasesin market interestrate spreadsto identifyfailed speculativeattacks is problematic,because these indicatorspotentially confoundspeculative pressures and the policy responseto these pressures.For examplepublished data on reservelosses does not permitme to distinguishbetween transactionsof the monetaryauthorities to accommodatethe increasedspeculative demandfor their reserves,and directsales of reservesby the monetaryauthority in orderto supportthe currency. Similarly,increases in observednominal interest rate differentialsmay reflectboth increasesin market participants'devaluation expectations as well as policy interventionsin the moneymarket, as notedin the introduction. However,to the extentthat these considerationsare important,the resultswill be biased towardsfinding that tighteningmonetary policy makes speculativeattacks more likely to fail, simply becausethe definitionof failed attacksin part reflectsthe presenceof tight monetarypolicy. It is interestingto note that despite this obvioussource of bias in favourof the conventionalwisdom, I find little evidenceof this view.

Table 1 lists the full sampleof 308 episodes,sorted by successand failure,and by year. The sampleincludes a numberof familiar episodes,as documentedin Table2. The recentspate of currencycrises in EastAsia in 1997 are all representedas successfulspeculative attacks, with the exceptionof Malaysiawhere the largestmonthly depreciationof the ringgit in August1997 (6.6 percent)was not large enoughto qualify as a successfulattack according to my definition. Table2 also lists severalspeculative attacksassociated with the turmoilin the ERMin 1992, and comparesthe datingof these attackswith that of Eichengreen,Rose and Wyplosz(1994). My criterion identifiesseveral of well-knownfailed attacksduring this period, includingthose on the

7 Danishkroner, French franc, Irish punt andthe Spanishpeseta, as well as the successfulattacks on the Britishpound, the Swedishkrona and the Finnishmarkkaa in the fall of 1992.7 In most cases,the datingof eventscorresponds fairly closelyto thaltof Eichengreen,Rose and Wyplosz(1994). The only large discrepancyis in the caseof France,where they identifya speculativeattack in September1992 which I do not. T his is becauseFrance's reserve losses of 8 percentin that month are not large enough accordingto my definition,while its muchlarger reserve losses in the fall of 1993are.

Table 1 neverthelessalso includesa numberof more questionableepisodes. Upon closerinspection, many of the episodesoccur in country-periodobservations characterizedby underdevelopedfinancial markets and/or restrictions on capital movementsof varioussorts. It is unlikelythat the dynamicsof speculativeattacks in such distortedenvironments will be comparablewith those occurringin countrieswith relativelydeveloped financial markets and free capital mobility. In orderto ensurethat the resultsare not tainted by these questionableepisodes, I also define a subsampleof eventswhere domestic credit to the privateseclor as a share of GDP (a summary indicatorof financialdevelopment) averages more than 20% in the five years prior to the attack, andthe blackmarket premiumon foreignexchange (a summaryindicator of de facto currencyconvertibility) averages less than 10% in the five years prior to the attack. I refer to this subsampleas the financially-developedsubsample, and the countries includedin it are indicatedwith asterisksin Table 1.

In Figure2, I providea graphicaloverview of successfuland failed speculative attacks,plotting the evolutionof the nominalexchange rate and reservesduring "typical" attacks. To constructthis figure, I computethe mediangrowth rate in the exchangerate and reservesover all successfuland failed attacksfor every month in a two-yearwindow centeredon the date of the crisis,and then cumulatethese mediangrowth rates on a baseof 100 oneyear prior to the crisis. By construction,successful attacks are marked by sharpnominal depreciations preceded by 12 monthsof very stableexchange rates. In fact, the medianchange in the nominalexchange rate prior to these episodesis zero. Reservesdecline steadily over the entireperiod leading up to the collapseof the

Theinitial attack on the Swedish krona in thesurnmer of 1992mentioned in theintroduction does not qualifyas an unsuccessfulattack accordingto my definitionsince it was followedby a depreciationwithin three months.

8 exchangerate, indicatingthat speculativepressures emerge in advanceof the collapse in the exchangerate itself, and reservesrecover fairly quickly afterwards. Failedattacks are also by constructionpreceded by very stablenominal exchange rates, and feature sharpreserve losses in the monthof the attack. As with successfulattacks, reserves recoverquickly following failed attacks.

The mainquestion of interestis whetherraising interestrates -- or more generally,tightening monetary policy -- preventsspeculative attacks from endingin a devaluationof the currency. To addressthis question,I requiremeasures of the stance of monetarypolicy around the speculativeattack episodes identified above. I primarily rely on the real centralbank discountrate (the nominaldiscount rate deflatedby contemporaneousannualized monthly inflation) as a measureof the policy instrument most directlyunder the controlof the monetaryauthority. 8 To the extentthat the monetaryauthority uses this instrumentduring a given episode,this variableprovides a good measureof the policy responseto the speculativeattack. However,as notedin the introduction,the monetaryauthorities in these manyspeculative attack episodes have a wide varietyof instrumentsat their disposal. Attemptingto identifythe mix of instrumentsactually employed during each of the 308 episodesin the sample,and hencethe appropriateepisode-specific measure of the stanceof monetarypolicy, would be ambitiousto say the least.9 1instead use two other measuresas crude "outcome" indicatorsof the stanceof monetarypolicy to check the robustnessof the results: real domesticcredit growth,and the reservesof deposit moneybanks held in the central bank. To the extent that the monetaryauthorities tighten monetarypolicy using other measures(e.g. open market operations,raising reserve requirements, etc.), this will be reflectedin a reductionin realdomestic credit and/or increases in bankreserves.'°

8 An unfortunatedrawback of this measureis that centralbank discountrates are reportedby the IMFon an end-of-periodbasis only, so that intra-monthlyfluctuations in this variableare ignored.Also, there is of courseconsiderable debate over howto proxyfor expectedinflation when constructingreal interestrates. The resultspresented here do not changesubstantially if I deflateusing eitherpast or futureinflation rates, or if I simplyconsider changes in nominaldiscount rates. Evenfor the UnitedStates which has beenthe subjectof decadesof intensiveresearch, there is noclear consensuson howprecisely to measurethe stanceof monetarypolicy. See for examplethe discussionin Bemankeand Mihov(1998). See also Borio(1997) for a descriptionof the bewilderingarray of instruments availablein a set of developedcountries. 10An obviousobjection to the domesticcredit growth measure is that it does not distinguishbetween shifts in the supplyand shifts in demandfor domesticcredit. To alleviatethis concernI havealso definedtight

9 For each speculativeattack episode, it is necessaryto determinewhether these measuresof monetarypolicy tightened or not, relativeto a suitablebenchmark. For failed attacks,I considerthe increasein real discountrates, the decreasein real domesticcredit growth, and the increasein bank reservesin the monthof the attack relativeto the monthprior to the attack. That is, I ask whethertightening monetary policy in responseto a suddenreserve outflow serves to arrestfurther reservelosses and maintainthe value of the currency. Giventhat speculativepressures appear in advanceof the actualdevaluation during successful attacks, I considerthe changein each measureof monetarypolicy in the month prior to the aftackrelative to the previous month. That is, I ask whethertightening monetary policy in responseto mounting speculativepressures serves to preventattacks from succeeding.I do not includethe month of the attackitself, so as not to captureany post-devaluationpolicy responses which may be quitedifferent from those undertakenin defenseof the currencyprior to the devaluation.11

(loose)monetary policy as periodswhere both domesticcredit growthfell (increased)and the discountrate increased(fell), with substantiallysimilar results. A similarobjection holds for the bank reservesmeasure. The effectivenessof monetarypolicy in the aftermathof devaluationsis studiedby Goldfajnand Gupta (1999).

10 3. Descriptive Results

In this sectionI presentsome simpledescriptive statistics on the incidenceand meanvalue of changesin the stanceof monetarypolicy around successful and failed speculativeattacks. The siimplestpossible graphical overview of the evidenceis in Figure3, which reportsthe frequencydistribution of changesin realdiscount rates during successfuland failed speculativeattacks. The strikingfeature of this graph is that there is no apparentdifference in the directionor magnitudeof changesin discount rates duringsuccessful and failedspeculative attacks. Moreformally, Table 3 uses contingencytables to summarizethe changesin monetarypolicy during successful and failed speculativeattabks. The three panelsof Table 3 correspondto the three different measuresof tightermonetary policy: increasesin real discountrates, decreases in real domesticcredit growth, and increasesin bank reservesas a shareof domesticcredit. Each panel reportsa contingencytable, with the columnscorresponding to successful and failed attacks,and the rows correspondingto whethermonetary policy tightened or eased. Basedon thesetables, I report severalstatistics of interest. I first reportthe conditionalprobability that monetarypolicy tightensgiven that a speculativeattack fails. If the conventionalwisdom is correct andtightening monetary policy is a necessary conditionto preventspeculative attacks from endingin a devaluation,one wouldexpect this probabilityto be near one. In fact, it rangesfrom 0.33 to 0.57, dependingon the measureof policy. In each case, the upper boundof a 95% confidenceinterval extends to no morethan 0.65. In fact, for the first two measuresof policy,the 95% confidence intervalincludes 0.5, so that it is not even possibleto rejectthe null hypothesisthat tighter and looser monetarypolicy during failed attacksare equallylikely. This casts doubt on the notionthat tighteningmonetary policy is necessaryto ensurethat speculativeattacks fail.

I also reportthe conditionalprobability that a speculativeattack fails giventhat monetarypolicy tightens. If the conventionalwisdom is correctand tighteningmonetary policy is a sufficientcondition for speculativeattacks to fail, one wouldexpect this probabilityalso to approachone. In this sample,the estimatedprobability ranges from 0.60 to 0.72,with a 95% confidenceinterval extending to at most 0.80. This calls into

1 1 questionthe idea that raisingdiscount rates is sufficientto ensure that speculative attacksfail.

More formally,I also report the p.-valuelor a chi-squaredtest of independence betweenchanges in monetarypolicy and the successor failure of speculativeattacks. For the first two measures,it is not possibleto rejectthe null hypothesisof independenceat conventionalsignificance levels, suggesting that there is no relationshipwhatsoever between changes in the stance of monetarypolicy and the successor failure of speculativeattacks. For the third measure,the null is (barely) rejectedat the 95%significance level, but this does not constituteevidence in favourof the conventionalview. To see this, note that the probabilitythat an attackfails conditionalon tightenedmonetary policy is 0.60, while the unconditionalprobability that an attackfails is 177/263=0.67- in otherwords, speculative attacks are significantly less likelyto fail when monetarypolicy is tighterthan in the sample as a whole. The rejectionof the null of independencetells us that this differencein probabilitiesis (barely)statistically significant.

In Table4, I repeatthe analysis,but restrictingthe sample to the financially- developedsubsample described in the previoussection. One might expectthat any evidenceon the efficacyof a high interestrate defensewould be more apparentin this smallerset of observations. However,the results in Table4 showthat this is notthe case. Althoughthe conditionalprobabilities (of tighter monetarypolicy conditional on failure,and of failure conditionalon tighter monetarypolicy) are generallya little higher in this sample,they are still far from one, and in no case can I rejectthe null hypothesis that changesin the stance of monetarypolicy and the outcomeof speculativeattacks are independent.

These resultshave beensubjected to a wide varietyof robustnesschecks that are not reportedfor brevity. Theseinclude: (1) restrictingthe sampleof eventsto those for which all measuresof monetarypolicy are available,(2) using a three-monthinstead of a one-monthwindow over whichto measurechanges in the various indicatorsof monetarypolicy aroundthe speculativeattack, (3) varying the timing of changesin monetarypolicy relative to the date of the attack, and (4) definingtightened monetary policy as episodeswhere both realdiscount rates increasedand real domesticcredit

12 growthfell. The main conclusionthat the stanceof monetarypolicy and the outcomeof the speculativeattack are independentis robustto all of these variants.

While Tables3 and 4 providea concisesummary of the availableevidence, they discardpotentially useful information by treatingchanges in policy as binaryevents, i.e. interestrates, domestic credit growth or bankreserves increase or decreaseonly. I relaxthis restrictionin Tables5 and 6, which presenttwo sets of statisticsfor the full sampleof eventsand the financially-developedsubsample, respectively. In the first three columnsof bothtables, I computethe meanchange in each measureof monetary policyduring successfuland failedspeculative attacks and test the null hypothesisthat they are equal. This may be thoughtof as a weaker versionof the tests of necessityin Tables3 and 4, in the sensethat a rejectionof this null hypothesisconstitutes evidence that tighteningmonetary policy is necessaryto preventspeculative attacks from succeeding.'2 In the next two columns,I estimatethe marginalimpact of the changein eachmeasure of monetarypolicy on the probabilitythat a speculativeattack fails, estimatedfrom a probit regressionincluding a constantterm. I report the estimated marginaleffect, andthe t-statisticcorresponding to the null hypothesisthat the underlyingcoefficient on the policymeasure is zero. This may be thoughtof as a weaker versionof the earliertests of sufficiency,in the sensethat a positiveimpact suggeststhat tighteningmonetary policy raises the probabilitythat an attackfails. In the final columnI reportthe numberof observationsincluded in eachtest. The three panelsof eachtable again correspondto the three measuresof changesin the stance of monetarypolicy. 13

A further drawbackof the previousresults is that they do not allowfor the possibilitythat the effectsof monetarypolicy may dependon fundamentalswhich vary acrossspeculative attack episodes. In orderto take these possiblenon-linearities into account,the rows of Table 5 reportresults for various subsamplescorresponding to "good"values of suchfundamentals. I first distinguishfurther betweenfinancially

12 A strongertest wouldalso requirethe meantightening in monetarypolicy to be positiveduring failed attacksand negativeduring successful attacks. 13Unlike the data descriptionin the previoustables, these summarystatistics are not robustto extreme outliersin the measuresof monetarypolicy. I thereforedrop a smallnumber of episodesoccurring during periodsof very high inflationwhere measured changes in real discountrates and realdomestic credit growthare greaterthan 100 percentin absolutevalue.

13 developedand less-developedepisodes by restrictingthe sampleto the OECD,and to the 1980sand 1990s. Followingthe suggestiornof Goldfajnand Gupta (1999)that interestrate defensesare only successfulwhen the bankingsystem is strong,I restrict the sampleto those episodesthat were not precededby a bankingcrisis in any of the previousfive years in the rows labelled"No BankingCrisis". Since one might expect that tighteningmonetary policy will only be effectiveif the exchangerate is not too overvalued,I constructa crudeindicator of real exchangerate overvaluationas the trend growthrate of the real CPI-weightedexchange rate versusthe US in the previous twelve months. In the rows labelled"No Real Cvervaluation",I restrict the sampleto those episodeswhere this growthrate is belowthe medianfor the entiresample. To capturethe notionthat a given defensemay be more credibleif the monetaryauthority can backup its commitmentto a fixedexchange rate with a large stock of foreign currencyreserves, I also dividethe samplein half accordingto non-goldreserves relativeto imports,and consideronly the high-reservessubsample in the rows labelled "High Reserves". I also proxyfor the overallweakness of the country'sexternal paymentsposition using the averageover the previoustwelve monthsof that country's borrowingfrom the InternationalMonetary Fund, expressed as a share of its quotain the organization.In the rows labelled"Low Quota Drawings",I consideronly those episodeswhere the country has no obligationsto the IMF accordingto this measure. Finally,I considerthe argumentthat it is easier to defend againsta speculativeattack duringa boomingeconomy than duringa recession,presumably because the domestic economyis better able to withstandany of the adverseeffects of high interestrates duringthe high point in the businesscycle. I measurethis as the deviationof real per capitaGDP growth in a countryfrom its average,in the five precedingyears, and then I dividecountries in two at the medianvalue of this deviationand consideronly the boomingeconomies in the rows labelled'High FPointin Cycle".

The resultsin Tables5 and 6 are not very supportiveof the conventionalview that tighteningmonetary policy lowers the probabilitythat a speculativeattack endsin a devaluationof the currency. In the vast majorityof cases,the meanchange in monetary policy is not significantlydifferent during failed and successfulattacks, and changesin the stanceof monetarypolicy are not statisticallysignificant predictors of the outcomeof the speculativeattack. Only in five cases are the estimatedeffects statistically significantat the 95% level, notablyin the OECDsubsample. However,when one

14 considersthat there are 96 separatehypothesis tests in Tables5 and 6, I shouldexpect aroundfive rejectionsat the 95%significance level even if the stance of monetarypolicy and the outcomeof speculativeattacks were independent. Moreover, even the few significantresults in the OECD subsampleare to a large extent driven by a handfulof successfulattacks in Greece,Turkey and Portugal.

At first glance,the descriptiveevidence presented in this sectionis hardly consistentwith the view that tighteningmonetary policy is effective duringspeculative attack episodes. At the same time, it is also hardlyconsistent with the alternativeview that tighter monetarypolicy has the perverseeffect of weakeningthe currencyunder attack. Rather,this descriptiveevidence suggests a strikingabsence of any systematic relationshipbetween the stanceof monetarypolicy andthe outcomeof speculative attacks.

15 4. The Endogeneity of Policy

Althoughuseful as data description,the evidencein the previoussection can provideonly limited informationabout the effects of policyduring speculativeattacks. Since policyis itself likelyto respondendogenously to the samefundamentals that drive speculation,and alsoto the strengthof speculativepressures themselves, it is difficultto infer any structuralrelationship frorn the correlatiornsof the previoussection. In this section,I presenta simple modelwhich formalizes this issueand illustratesits ambiguousimplications for the evidenceof the previcussection. 14 I then empirically addressthe endogeneityproblem by estimatingan instrumentalvariables probit model that expressesthe probabilitythat a given speculativeattack ends in a devaluationas a non-linearfunction of fundamentalsand measuresof monetarypolicy, treating monetary policyas endogenous.After controlling for endogeneityof policyin this way, I still find no evidencethat raisinginterest rates either lowersor raisesthe probabilitythat a speculativeattack ends in a devaluationof the currency.

A Simple Model

I considera one-periodmoclel of a srnall open economythat fixes its exchange rate and comesunder speculative attack. The economyis populatedby a continuumof identicalatomistic speculators of mass one, and a rnonetaryauthority. The monetary authoritysets the domesticinterest rate, i, at the beginningof the period,and at the end of the perioddecides whether or not to devaluethe currencyby an exogenously-given and knownamount, s. Speculatorsattack the currencyby shortingit, i.e. by takingout loans in local currencyat the interestrate set.by the monetaryauthority at the beginning of the period,selling the proceedsto the monetaryauthority in exchangefor US dollars at the beginning-of-periodexchange rate, and then unwindingtheir positionsat the end- of-periodexchange rate. Speculatorsdetermine their demand,S, for the reservesof

14 See Drazen(1999), Lahiri and Vegh (1997,1999) and Lall (1997)for other modelswhich focus specificallyon the role of interestrates as a defenseduring speculativeattacks. 15In practice,shorting the domesticcurrency during speculativeattacks is generallydone usingforward contracts,rather than domesticcurrency loans. However,the substanceof the analysisis not changedby this complication.See Goldsteinet. al. (1993),Garber and Svensson(1995), and Lall (1997)for details.

16 the monetaryauthority, R, by maximizingtheir profits net of borrowingcosts, which I assumefor convenienceto be quadraticin the volumeof speculation:

(1) max~~7* S - 2~~

where xt denotesthe representativespeculator's perception of the probabilitythat the currencywill be devalued.16Solving this optimizationproblem and aggregatingover all speculatorsresults in a speculativedemand for local currencyS(it,i)=

The monetaryauthority decides whether or not to devaluethe currencyby weighingthe costsand benefitsof maintaininga fixed exchangerate. There are two costs to fixing: the monetaryauthority must spend a fraction (R' of its reserves to R defendthe exchangerate, and in orderto maintaina desiredlevel of reserves,it may needto set domesticinterest rates higherthan it would otherwisedo in the absenceof speculativepressures.'7 These costs are summarizedin the followingloss functionof the monetaryauthority:

(2) L(7ri, *) = 57 i 0 * g R where for simplicityI haveassumed that the monetaryauthority's disutility of raising interestrates is linear in the interestrate, with 0* measuringthe strengthof its aversion to high domesticinterest rates. The parameter0* is not knownto speculators,who

16 This convenientformulation of speculativebehaviour is used by Drazen(1999). In the absenceof such adjustmentcosts, risk-neutral speculators will take infinite short (long)positions in the currencyunder attack if the expectedreturn to shortingis positive(negative). At the costof complicatingthe algebra,one can also motivatea continuousspeculative demand for loans by assumingthat speculatorsare risk averse.

17 I followthe conventional(implicit) assumption that the monetaryauthority dislikes reserve losses and devalueswhen these losses are excessive.However, it is naturalto ask why this shouldbe the case. One mightalso imaginethat the monetaryauthority does not value reservesper se, but ratherdislikes the capital lossesit suffersfollowing a devaluationwhen it restoresits targetlevel of reservesby purchasing themat the depreciatedexchange rate. In this case larger reservelosses make devaluations more costly. Moreover,raising interest rates may have the perverseeffect of raisingthe rationally-expectedprobability of a devaluationby makingdevaluations less costlyto the monetaryauthority.

17 sharea commonbelief that it is equalto i. Let ,Bdienote the benefitsof maintainingthe fixed exchangerate regime. Thesebenefits are also not knownto speculators,who correctlyperceive D to be uniformlydistributed on the unit interval. Speculatorsdo know that if the costs of maintaininga fixed exchangerate exceedthe benefits,the monetary authoritywill devaluethe currencyto 1+E.

Speculatorsrationally form their beliefsregarding the probabilitythat the monetaryauthority will devalue,given their perceptionsof the "type"of the monetary authority,0, and given the interestrate set by the mronetaryauthority. In particular, speculatorsunderstand that 7t= Prob[L(ir,i,9)> f3],so that the rationally-perceived devaluationprobability is: 1 8

~~(3) 7 9 R . 2 R l-£i-F~~

I plotthis probabilityas a functionof the interestrates as a bold line in the top panelof Figure4. At low levels of the interestrate, the perceiveddevaluation probability is decreasingin i. Overthis range,speculation against the currencyis intense,and the marginalbenefit of raisinginterest rates (in terms oi reducingreserve losses S) outweighsthe perceivedmarginal cost to the domesticeconomy (as measuredby the parameteri). As a result, raisinginterest rates lowersthe monetaryauthority's disutility of maintainingthe fixed exchangerate, making a devaluationis less likely. In contrast, when interestrates are high,the marginalbenefit of further increasesin interestrates is smallerthan the marginalcost to the domesticeconomy. Over this range,increases in the interestrate raisethe disutilityof the fixed exchangerate regime,and so raisethe probabilitythat the currencywill be devalued.

18To simplifythis calculation,I assumethat L(7r,i,O) < 1, so that Prob[L(j,i,0) > P] = L(r,i,i) . It is straightforwardto verify that this holds in equilibriumprovided that the followingparameter restriction is satisfied:R -. + < 1. This restrictionwill hcld providedthat the devaluationrate e is smallenough and/or the amountof reservesR is largeenough, which together ensure that the speculative demandfor reservesis nevertoo large.

18 The questionof interestin this paper is the slope of 7c(i),i.e. whetherraising interestrates raisesor lowersthe probabilitythat a speculativeattack ends in a devaluationof the currency. However,estimating Tl(i) using the data on speculative attack episodesdescribed in the previoussections is complicatedby two factors. First, for a given interestrate, the slopeof 7r(i)will dependon episode-specificcharacteristics. This nonlinearityis illustratedin the lower panelof Figure4, which considerstwo speculativeattack episodes that are alike in every respect,except that in the secondthe level of reservesis higherthan in the first. Not surprisingly,the probabilityof a devaluationis everywherelower in the secondepisode than in the first, sincethe monetaryauthority has more reservesat its disposalto defend the exchangerate. More important,at the same level of the interestrate (indicatedby the verticalline), a small increasein interestrates in the first episodewill lower the probabilityof a devaluation, while in the secondepisode it raisesthe probabilityof a devaluation.

The seconddifficulty is thatthe monetaryauthority's choice of interestrates is endogenous,and dependson the strengthof speculativepressures against the currency. In orderto illustratethis endogeneitywithin the confinesof a very simple model,I assumethat the monetaryauthority sets interestrates to minimizethe costsof maintaining a fixedexchange rate. In particular,I assumethat the monetaryauthority choosesi to minimizeEquation (2), taking into accountthe dependenceof 7E(i)as given by Equation(3). The optimalinterest rate chosen by the monetaryauthority is:

( ) R ( ;* and has a very naturalinterpretation. Other things equal,the higher is the devaluation rate e or the lowerare reservesR, the greateris the volumeof speculationand the higheris the interestrate set by the monetaryauthority to deter this speculation.The greateris the monetaryauthority's aversion to high interestrates (the higheris G*),the lower is the optimalinterest rate. Finally,the more speculatorsthink the monetary

~~19 authoritydislikes high interestrates (the higher is 0), the higherthe monetaryauthority needsto raiseinterest rates to reducespeculation.' 9~~

The importantpoint is of coursethat the inteirestrate chosenby the monetary authorityin Equation(4) dependson the samefundamentals as speculators'perceived probabilityof devaluationin Equation(3). In Figure5, I illustratehow this endogeneity problemcan eitherobscure or accentuatethe effectsof tighter monetarypolicy during speculativeaffacks. In the top panel, I again considertwo episodesthat are alike in everyrespect, except that in the latterthe reservesof the monetaryauthority are higher than in the former. At the equilibriumin the first episodeat A, 7T(i)is decreasingin i, so that a small increasein interestrates has the conventionaleffect of loweringthe perceivedprobability of a devaluation.In the high reservescase, the speculators' rationally-perceiveddevaluation probabilities are lowerthan before(shown as a downwardsshift in n(i)), whilethe monetaryauthority reacts to these devaluation perceptionswith a lower interestrate since it has a larger "cushion"of reserves. In this episode,the equilibriumis at B with a lower interestrate and a lowerdevaluation probability.Simply comparing these two episodes,one mighteasily be led to the mistakenconclusion that raisinginterest rates raisesthe probabilityof a devaluation, while preciselythe converseis true (sinceboth A and B fall on the downward-sloping portionof 7rt(i)).

Similarly,the endogeneityproblem may also lead to the conclusionthat raising interestrates hasthe conventionaleffect of loweringthe probabilityof a devaluation when in fact the oppositeis true. 11illustrate this possibilityin the bottompanel of Figure 5. 1again consider two identicalepisodes, which now differ only in the monetary authority'sdistaste for interestrates (0*) and speculators'beliefs regarding this parameter(0). The dashedlines correspondto an episodewhere both 9* and 0 are lowerthan in the episodeshown irn solid lines. Not surprisingly,the monetaryauthority sets a higherinterest rate, and sincespeculators believe that the monetaryauthority is "tough",the devaluationprobability is lower for every interestrate i (shownas a

19 assumethat the monetaryauthority knows speculators' perceptions regarding its type,i.e. the monetary authorityknows 0. The main pointof the modelregarding the endogeneityof policyis unaffectedif I insteadassume that the monetaryauthority does not know0 Ibutinstead takes speculators'perceived devaluationprobabilities as given when minimizingEquation (2).

I0 downwardsshift in 7c(i)).Comparing the equilibriaA (witha high devaluationprobability and a low interestrate) and B (witha low devaluationprobability and a high interest rate),one might easilyconclude that raisinginterest rates lowersthe probabilityof a devaluationwhen the converseis true (sinceboth A and B fall on the upward-sloping portion of 7c(i)).

This discussionillustrates how the endogeneityof policycan bias the estimated effectsof policyin unknowndirections. To the extentthat the fundamentalsthat drive both speculativepressures and the policy responseare not fully observable,partial correlationsbetween policy andthe outcomeof speculativeattacks will not correctly identifythe effects of policy. To achieveidentification, I requirean exogenoussource of variationin the interestrate set by the monetaryauthority that can be used as an instrumentfor policy. In this stylizedmodel, the monetaryauthority's private information aboutits "type"(0*) playsthis role, since changesin 0* shift the monetaryauthority's reactionfunction without shifting speculators' rationally-perceived devaluation probabilities.More generally, any privateinformation of the monetaryauthority which influencesits choiceof interestrates can in principleserve to identifythe effectsof interestrates on speculators'beliefs that an attackwill end in the devaluationof the currency.

~~EmpiricalSpecification~~

I nowturn to the empiricalspecification motivated by this simple model. The objectiveis to estimatethe impactof monetarypolicy on probabilitythat a speculative attackfails. Althoughthis probabilityis not observable,I do observea binaryindicator of whethera speculativeattack fails or not. I can thereforeestimate the marginaleffects of policyon probabilitythat an attackfails usinga probitmodel, with this indicatoras the dependentvariable. The first implicationof the theoryis that this probabilitywill be a non-linearfunction of fundamentalsand the monetarypolicy response. Althoughthe simplemodel discussed above is too stylizedto take the exact functionalform impliedby Equation(3) literally,it does suggestthat the explanatoryvariables in the probit equation shouldinclude not only measuresof policyand fundamentals,but also interactions betweenthe two. Accordingly,I considerthe followingnon-linear probit specification:

~~21 yi =PO+ l ii + P2'fj + 3' fj *ij +u~~

(~ ~ ~ ~~~~~) ,if Yj* > yi O, ifyj*

where yj* is an unobservedlatent variable; yj is an indicatorvariable taking the value 1 if speculativeattack j ends in a devaluation;ij is a measureof the stanceof monetary policy;fj is a vector of episode-specificfundamentals; and u;is a normally-distributed disturbanceterm. I considerthe samethree measuresof the stanceof monetarypolicy

~~(i1) as in the previoussection: increasesin real discountrates, decreases in domestic creditgrowth, and increasesin the reservesof the bankingsystem, and five of the~~

measuresof fundamentals(f 1) discussedin the previoussection: the presenceof bankingcrises, the extent of realovervaluation, the adequacyof reserves,indebtedness to the IMF, and the point in the businesscycle prior to the speculativeattack.

The secondimplication of the theory is that ij is endogenousand reactsto the samefundamentals that drivespeculative pressures. To the extent that the observed fundamentalsincluded in fj do not captureall of these factors,the error term in Equation (5) will be correlatedwith policy. It is thereforenecessary to instrumentfor both ij and f, ij in the above regression.The theory indicatesthat variablesthat are the private informationof the monetaryauthority and influenceits choiceof monetarypolicy are candidateinstruments. I rely on two instruments,both of which exploit informational asymmetriesthat are likely to exist betweenspeculators and the monetaryauthorities. The first is the changein reserves(expressed in monthsof imports)in the monthof the attack. Sincein mostcountries the monetaryauthority publishes data on its reserves only with a lag, it will have informationon the extent of aggregatespeculative pressures reflectedin reservelosses in advanceof market participantswho do not havetimely accessto this data. To the extentthat the monetaryauthority bases its policyresponse on its observedreserve losses, this instrumentwill be correlatedwith policy. To the extentthat these reservelosses are unknownto speculators,they will affect speculative pressuresonly throughtheir effectson policy,and hence this measureis a valid instrument.

22 The secondinstrument is the ex post availableinformation on the country's borrowingfrom the InternationalMonetary Fund. If a countrycomes under speculative attack,it may seek resourcesfrom the IMF for temporarybalance of paymentssupport. To the extent that the IMF placesconditions on the stanceof monetarypolicy prior to agreeingto such support,changes in observedIMF borrowingwill be correlatedwith the indicatorsof monetarypolicy aroundthe speculativeattack episodein question. To the extentthat speculatorshave imperfectinformation as to the substanceof the country's negotiationswith the IMF, the IMF's influenceover monetarypolicy will be knownto the monetaryauthority, but not to speculators. I thereforeproxy for the presenceof IMF involvementby the changein a country'sborrowing from the IMF in the three months followingthe speculativeattack relative to the three monthsprior to the attack,and use this variableas an instrumentfor policy.

Obviouslythese instrumentsare imperfect. First, they may not be truly exogenous.There may be unobservedepisode-specific characteristics which both raise the probabilitythat an attackfails and also acceleratereserve losses or trigger IMF involvement. Second,despite caricatures to the contrary,the involvementof the IMF may not be significantlycorrelated with the subsequentstance of monetarypolicy. The first objectionis easilyaddressed if the instrumentspass tests of overidentifying restrictions.The secondobjection concerns the strengthof the instruments.As is well known(Nelson and Startz (1990),Staiger and Stock (1997)),if the instrumentsare only weakly correlatedwith the endogenousvariables, two-stage least squarescoefficient estimateswill be biasedtowards the probabilitylimits of their uninstrumented counterpartsin finite samples. That is, instrumentingwith weak instrumentswill not correctthe problemof endogeneity.This caveat shouldbe kept in mind,since in many casesthe explanatorypower of the instrumentsis not as large as I wouldlike.

I estimateEquation (5) usingAmemiya's (1978) generalized least squares estimatorfor probitmodels with endogenousregressors. This is a two-stageprocedure, in which the observeddependent variable yj and the endogenousvariables are first all regressedon the exogenousvariables and the instruments.Amemiya's insight is that, providedthat the modelis over-identified, the structuralparameters in Equation(5) can be retrievedfrom a GLS regressionof the reducedform parametersof the first-stage regressioninvolving the dependentvariable on the reduced-formparameters from the

23 remainingfirst-stage regressions. As shownby Newey(1987), this methodis asymptoticallyequivalent to a minimumchi-squared estimator and is the mostefficient methodto extractstructural from reduced-formparaimeter estimates. Finally,Lee (1991)provides a test of overidentifyingrestrictions for this model.

~~Results~~

The resultsof this instrumentalvariables probit specification are presentedin Tables7 and 8 for the full sampleand the financially-developedsubsample, respectively.Each table reportsthe resultsof 15 probit regressions(three measures of policytimes five measuresof fundamentals),with a dummyvariable taking on the value one if the attackfails as the dependentvariable. For each regression,I report the estimatedcoefficients, their standarderrors, and the correspondingmarginal effect of an increasein the right-handside variableon the probabilitythat a speculativeattack fails. In orderto assessthe validityof the instruments,I reportthe p-valueassociated with the test of overidentifyingrestrictions (which tests the null hypothesisthat the instrumentsaffect the outcomeof the attackonly throughtheir effectson policy)and the p-valueassociated with a test of the null hypothesisthat the instrumentsare jointly significantin the first-stageregression of the policyvariable on the instruments.

The resultsin Tables7 and 8 providevery little evidenceof the efficacyof tight monetarypolicy as a defenseagainst speculative attacks. In all but two regressions,the coefficienton the policyvariable is not statisticallysig nificantly different from zero, and the estimatedmarginal effects are generallytiny. The onlyexception is when policyis measuredas the changein bankreserves, and the fundamentalsare proxiedby the presenceof bankingcrises or indebtednessto the IMF. For these two cases,the estimatedeffect of policy is negative-- a tighteningof monetarypolicy lowersthe probabilitythat a speculativeattack fails. However,this evidencein favourof "perverse" effectsof tighter monetarypolicy shouldnot be taken too seriously,given the large numberof otherspecifications which do not corroboratethis finding.

An importantcaveat regarding Tables 7 and 8 concernsthe validityof the instruments.Although in all casesthe instrumentscomfortably pass tests of overidentifyingrestrictions, in manycases they haverather weak explanatorypower for

24 policy. This impliesthat the estimatesare likely to be biasedtowards the probability limits of their uninstrumentedcounterparts, and hence may still be taintedby endogeneitybias. While this is unfortunate,it is not clear a priori whetherthis will result in a systematicoverstatement or understatementof the effectsof monetarypolicy, given the previousdiscussion that the directionof the endogeneitybias is theoretically ambiguous.

~~25 5. Conclusions~~

Do high interestrates help to defend exchangerates that comeunder speculativeattack? The evidenceconsidered in this paper suggeststhat the answeris no. Althoughproponents of the view that high interestrates can supporta currency underattack can pointto episodessuch as Swedenin the summerof 1992,while proponentsof the contrarianview that high interest ratesweaken currencies can pointto Koreain the fall of 1997 as supportiveanecdotes, a systematicexamination of interest rates arounda large numberof historicalspeculative attack episodesindicates a striking lack of evidencethat the stance of monetarypolicy is correlatedwith the outcomeof speculativeattacks. In particular,I find no evidencethat interestrates systematically increaseor decreaseduring failedspeculative attacks, nor that raisinginterest rates lowersor raisesthe probabilitythat a speculativeattack fails. This basicfinding is robustto alternativemeasures of the stance of monetarypolicy, to interactionswhich controlfor differencesin fundamentalsacross speculativeattack episodes, and to controllingfor the endogeneityof the policy responseto a speculativeattack.

Nevertheless,several shortcomings of this papersuggest that it may be prematureto concludethat monetarypolicy is entirely ineffectivein duringspeculative attacks. In the interestsof coveringa sampleof spieculativeattacks large enough to includeinteresting variation in the outcomeof speculativeattacks, the policyresponse to the speculativeattack, and the fundamentalsthat are likelyto determineboth the outcomeof the attackand the efficacy of the policy response,I havemade several compromiseswith regardsto data and methodology. Three suchcompromises, and possiblestrategies to avoidthem in future research,deserve mention.

First, I haverelied on readily-availablebut relativelylow-frequency monthly data to identifyspeculative attacks and the responseof policy. This is unfortunategiven that much of the economicallyinteresting variation during speculativeattack episodes is likely to occur at muchhigher daily,or even hourly,frequencies. The use of monthly data also precludesmodeling the likely path-dependencein the effectsof interestrates on speculativepressures, a point emphasizedby Drazen(1999). Movingto high- frequencydata for the more limitedsample of speculativeattacks for which suchdata is

~~26 availablemay uncoverevidence of the effectsof monetarypolicy that are obscuredby the low frequencyand absenceof dynamicsin the presentpaper.~~

Second,in this paper I haverelied on the very crudeindicators of monetary policy that can readilybe constructedfrom availablemonthly data. However,as noted earlier, monetaryauthorities have a wide varietyof instrumentsat their disposal, includingopen marketoperations, direct interventionsin foreign exchangemarkets, impositionof creditceilings, etc. Disentanglingthese interventionsfrom the observed fluctuationsin observablehigh-frequency data, and modelingthe choice between instrumentsover time and acrossepisodes, is essentialto obtaininga better understandingof the role of monetarypolicy during speculative attacks.

Third, in this paper I haverelied on what turn out to be rather weak instruments to extractthe exogenouscomponent monetary policy. While it is theoreticallyunclear howthis will systematicallybias the results-- giventhat the directionof the endogeneity bias is ambiguous-- it is neverthelessunsatisfying if one is interestedin understanding the effectsof monetarypolicy during speculative attacks. One possibilityfor progress on this front is that by switchingto higherfrequency data, the more pronounced informationalasymmetries between speculators and the monetaryauthority will resultin more robustinstruments. Another is to carefullyinvestigate the institutionalpeculiarities of the monetaryauthority during individual speculative attack episodes in the hopesof identifyingchanges in these institutionswhich might serveas valid instruments.

Implementingthese improvementsfor a sufficientlylarge set of speculative attackepisodes that spanthe relevantrange of countryexperiences will take time. Until then, however,it seemsthat the burdenof proof for both the conventionalwisdom that raisinginterest rates strengthenscurrencies under speculativeattack, and also the contrarianview that it weakensthem, lies with the proponentsof these views.

~~27 References~~

~~Amemiya,Takeshi (1978). "The Estimationof A SimultaneousEquation Generalized ProbitModel". Econometrica.46(5):1193-11205.~~

~~Bensaid,B. and 0. Jeanne(1997). "The Instabilityof FixedExchange Rate Systems When Raisingthe NominalInterest Rate is Costly". EuropeanEconomic Review. 41:1461-1478.~~

~~Bernanke,B. and 1.Mihov (1998). "MeasuringMonetary Policy". QuarterlyJournal of Economics. 113(3):869-902.~~

~~Borio,C. (1997). "MonetaryPolicy Operating Procedures in IndustrialCountries", in Bank for InternationalSettlements. Implementationand Tacticsof Monetary Policy. Basle.~~

~~Christiano,L, M. Eichenbaumand C. Evans(1998). 'Monetary PolicyShocks: What HaveWe Learnedand to What End?". NationalBureau of EconomicResearch Working PaperNo. 6400.~~

~~Cushman,D. and T. Zha (1997). "IdentifyingMonetary Policy in a SmallOpen EconomyUnder Flexible Exchange Rates". Journalof MonetaryEconomics. 39:433-448.~~

~~Caprio,Gerard and DanielaKlingebiel (1997). "Bank Insolvency:Bad Luck, Bad Policy or Bad Banking?". Paperpresented at the Annual Bank Conferenceon DevelopmentEconomics 1996, Washington, DC.~~

~~Demirguc-Kunt,Asli and EnricaDetragiache (1997). "The Determinantsof Banking Crises: Evidencefrom Industrialand DevelopingCountries". World Bank Policy ResearchWorking Paper No. 1828.~~

~~Drazen,A. and P. Masson(1994). "Credibilityof Policiesversus Credibilityof Policymakers".Quarterly Journal of Econon,ics. 109(3):735-754.~~

~~Drazen,A. (1999). "InterestRate DefenseAgainst SpeculativeAttack Under AsymmetricInformation". Manuscript. Univiersity of Marylandand NBER.~~

Edison,H. (1993). "The Effectivenessof Central Bank Intervention:A Surveyof the LiteratureAfter 1982". SpecialPapers in InternationalEconomics No. 18, InternationalFinance Section, Department of Economics,Princeton University.

~~Eichenbaum,M. and CharlesL. Evans(1995). "Some EmpiricalEvidence on the Effectsof Shocksto MonetaryPolicy on ExchangeRates". QuarterlyJournal of Economics. 110(4):975-1010.~~

~~Eichengreen,B., A. Roseand C. Wyplosz(1994). "SpeculativeAttacks on Pegged ExchangeRates: An EmpiricalExploration with SpecialReference to the EuropeanMonetary System". NBERWorking Paper No. 4898.~~

~~28 Eichengreen,B., A. Roseand C. Wyplosz(1995). "ExchangeMarket Mayhem: The Antecedentsand Aftermathof SpeculativeAttacks". EconomicPolicy. 21:249- 312.~~

~~Eichengreen,B., A. Roseand C. Wyplosz(1996). "ContagiousCurrency Crises". CEPRDiscussion paper No. 1453.~~

~~Furman,Jason and JosephStiglitz (1998). "EconomicCrises: Evidenceand Insights from EastAsia". Paperpresented at the BrookingsPanel on EconomicActivity, September.~~

~~Garber, P. and L. Svensson(1995). "The Operationand Collapseof Fixed Exchange Rate Regimes",in Grossman,G. and K. Rogoff, eds., Handbookof International Economics.Amsterdam: NorthHolland.~~

~~Grilli,V. and N. Roubini(1995). "Liquidityand ExchangeRates: Puzzling Evidence From the G-7 Countries". New York University,Salomon Brothers Working Paper:S/95131~~

~~Goldfajn,Ilan and PoonamGupta (1999). "Overshootingand Reversals: The Roleof MonetaryPolicy". InternationalMonetary Fund WorkingPaper No. 99/42.~~

Goldstein,M., D. Folkerts-landau,P. Garber, L. Rojas-Suarezand M. Spencer(1993). InternationalCapital Markets: ExchangeRate Managementand International CapitalFlows. Washington: InternationalMonetary Fund.

~~Lahiri,A. and C. Vegh(1997). "KrugmanBalance of PaymentsCrises: Are They For Real?". Manuscript,UCLA.~~

~~Lahiri,A and C. Vegh (1999). "OutputCosts, BOPCrises, and OptimalInterest Rate Policy". Manuscript,UCLA.~~

~~Lail, S. (1997). "SpeculativeAttacks, Forward Market Interventionand the ClassicBear Squeeze". IMF WorkingPaper No. 97/164.~~

Lee, Lung-Fei(1991). "Amemiya'sGeneralized Least Squaresand Test of Overidentificationin SimultaneousEquations Models with Qualitativeor Limited DependentVariables." Universityof MinnesotaCenter for EconomicResearch DiscussionPaper No. 262.

~~Nelson,Charles R. and RichardStartz (1990). "Some FurtherResults on the Exact Small SampleProperties of the InstrumentalVariable Estimator". Econometrica. 58:4:967-976.~~

~~Newey,Whitney K. (1987). "EfficientEstimation of LimitedDependent Variable Models with EndogenousExplanatory Variables". Journalof Econometrics.36:231-250.~~

~~Radelet,S. andJ. Sachs(1998). "The EastAsian FinancialCrisis: Diagnosis, Remedies,Prospects". BrookingsPapers on EconomicActivity. (1):1-90.~~

~~29 Sims, C. (1992)."Interpreting the macroeconomicTime SeriesFacts: The Effects of MonetaryPolicy". EuropeanEconomic Review. 36:975-1000.~~

~~Staiger,Douglas and James H. Stock(1997). "InstrumentalVariables Regression with Weak Instruments". Econometrica.65(3):557-586.~~

~~30 Appendix: Data Sources~~

~~The monthlydata employedin this paperare drawn from the International FinancialStatistics of the InternationalMonetary Fund, as follows:~~

* Nominalexchange, local currencyunits per US dollar, period average(IFS LinerF). * Non-goldreserves, US dollars(IFS Line 1I.d) * Moneymarket rate, percent(IFS Line 60b) * Discountrate, percentper year, end of period (IFS Line 60). For France,Singapore Swedenand the Netherlandsafter December1993, I use repurchaserates (IFS Line 60a) * Domesticcredit, local currencyunits (IFS Line 32) * Reservesof deposit moneybanks, local currencyunits (IFS Line 20) * Consumerprice index (IFS Line 64). For Australiaand Ireland,I use the wholesale priceindex (IFS Line 63) * Imports,c.i.f., US dollars (IFS Line71) * Total IMF credits and loans outstanding(IFS Line 2tl) * IMF quota (IFS Line 2t)

~~Lower-frequencydata correspondingto the speculativeattack episodes are drawnfrom varioussources.~~

* Domesticcredit to the privatesector as a shareof GDPis drawn from the World BankWorld Tables (FS.AST.PRVT.GD.ZS) * The blackmarket premium is drawnfrom Easterlyand Levine(1997) * Annualreal GDPgrowth is constantprice local currencyGDP figures drawn from the World BankWorld Tables (NY.GDP.MKTP.KN) * Dataon bankingcrises are basedon the IMF's May 1998 edition of the World EconomicOutlook, and augmentedusing the originalsources in Caprioand Klingebiel(1997) and Demirguc-Kuntand Detragiache(1997).

~~The sampleof countriesconsists of all countrieswith per capita GNPgreater than $800 in 1995 at Atlas exchangerates and with populationsgreater than one million,i.e. the~~

31 World Bank'sdefinition of middle-and upper-incomecountries in that year. From this sample,I dropChile, Algeria, Panama,Romania, Saudi Arabia, El Salvadorand Turkmenistanbecause requisite data on discountrates was not available.

The daily datain Figure1 are drawn from Bloomberg:SEK -- Swedishkrona/US dollar, spot,mid-rate, SWBRMRGN -- marginallending rate, mid-rate,KRW, won/US dollar,spot, mid-rate,and KWCR1T,overnight call rate, mid-rate.

~~32 Table 1: Successful and Failed Speculative Attacks~~

~~SuccessfulAttacks FailedAttacks~~

~~Large ReserveLosses LargeSpreads~~

ISR 62:2 BWA 84:7 BOL 62:1' PER 76:5 PHL 87:9' OEU 73:3- COL 65:9- PRY 84:3 DEU 62:1- PRT 75:10 TTO 87:1 ESP 74:5' ARG 66:11' THA 84:11' GTM 62:6 DNK 76:7' ZAF 87:11 MYS 74:6' ESP 67:11' VEN 84:2 PER 62:7' DOM 76:1 GAS 88:5' DNK 76:4' FIN 67:10' ZAF 84:7' COL 64:7- FRA 76:4- GRC 88:3 ESP 76:7- GBR 67:12' AUS 85:2' MAR 64:6 GBR 76:4' GTM 88:7 ITA 76:3' IRL 67:12' DOM 85:1 SYR 64-8 JAM 76:5 KWr 88:7 NLD 76:8' ISR 67:11' ECU 85:12 CHE 65:1' MAR 76:4' BOL 89:12' GBR 781' JAM 67:12' TTO 85:12 GBR 65:1' SWE 76:10' DOM 89:7 NLD 78:10' MUS 67:12 GTM 8686 GTM 85:5* SYR 78:3' MAR 89:3' CAN 79:1' PER 67:9' IDN 86:9 DOM 68:6 TUR 76:10' GRC 90:3' DNK 79:9' TTO 67:12 PRY 86:12 CHE 67:1' ZAW 76:2' MEX 90:3 ARG 80:10' FRA 69:8' VEN 86:12 COL 67:1 CHE 77:1- FIN 91:5' CAN 80:4' ARG 70:6' DOM 87:6' ESP 67:3' DNK 77:12' JOR 918' BEL 81:4' ECU 70:8 PER 87:11 SYR 67:12 GAB 771' TUN 91:4' MYS 81:10' PHL 70:3 SYR 88:1 CHE 68:7' MUS 77:8' ZAP 91:12' NOR 81:12 TUR 70:8' GTM 89:11' CRI 68:11 NOR 77:11' CAN 92:11' ESP 82:5' ISR 71:8 ISR 89:1 DNK 68:10' CAN 782' DEU 92:10' ESP 842' KOR 71:7 PRY 89:3 ECU 68:3 CRI 78:10' ESP 92:9' MYS 84:10 URY 71:12 VEN 89:3 FRA 68:6' MUS 78:11 GAB 929' MEX 85:3 BOL 72:11 DOM 90:4 GAB 68:2 ZAF 78:12' IRL 92:9' TUR 86:7' JAM 731' HUN 91:1 PHL 68:11 CAN 79:5' NAM 92:4 TUR 89:11' AUS 74:10' POL 91:6 TTo 68:11 DOM 79:7 TTO 92:1 IDN 90:12 ISR 74:11 TUR 91:3' DEU 69:1' GAB 79:3 BOL 93:1' PRT 90:10' KOR 74:12' BWA 92:7 FIN 69:5' JAM 79:8 ONK 931' POL 91:2' ARG 75:1 ECU 92:9' URY 69:10' SYR 79:8' FRA 93:11' ARG 92:7 PER 75:10 FIN 92:9' CRI 70:7' URY 79:3' BOL 94:2' KWT 92:11 ZAF 75:10' GBR 92:10' DOM 70:5 OL 80:4' DOM 948 NOR 92:9' AUS 76:12' LBN 92:3 ECU 70:1 ONK 80:2' MEX 94.4' PRT 92:9' MEX 76:9' SWE 92:11' ITA 70:7' PER 80:1' SVK 94:7 POL 93:6 ESP 77:7' TTO 93:4 SYR 70:12 PRT 80:2' TTO 94:5' URY 93:6 ISR 77:11' GAB 94:1' TWN 70:7 AUS 81:9' ZAP 94:3' ARG 94:12 PER 77:11 MEX 94:12' ARG 71:10 CAN 81:7' ARG 95:3 LVA 94:5 PRT 77:3' VEN 95:12' DNK 71:4' GTM 81:6 GAB 95:r MAR 94:1' IDN 78:11' BGR 96:5 ECU 71:9 MEX 81:6' ZAF 95:4' POL 94:9 JAM 78:5 GTM 97:1' IDN 71:12' MUS 81:3' BGR 96:1 COL 95:4' TUR 78:3 IDN 97:8' KOR 71:12 MAR 82:6' GRC 96:5' THA 951' BOL 79:12' KOR 97:11' LBN 71:2' MUS 82:6 NAM 96:9 DOM 96:6 MUS 79:11' MKO 97:7 MAR 71:10' PHL 82:1' SWE 96:11' LVA 96:2 TUR 79:6 PHL 97:9' CRI 72:2 URY 82:1' FIN 97:11' RUS 98:9 KOR 80:1 THA 97:7' GBR 72:7' GAB 83:11' IRL 97:4' URY 96:9 ARG 81:2' BWA 98:7 GAB 73:6 CAN 84:6' NOR 97:12' BRA 97:11 CRI 81:1' MOA 98:10 SYR 73:2 DNK 84:12' RUS 97:11 EST 97:11 MUS 81:10' MEX 98:9' ZAF 73:11' COL 65:1 BRA 98:9 POL 979 BOL 82:2 NAM 98:7 DOM 74:1 JOR 85:3' CAN 98:8' UKR 97:12 ECU 82:5 RUS 98:9 ITA 74:2' PHL 85:10' HKG 98:8' FIN 82:10' UKR 98:9 JAM 74:5' CAN 86:3' URY 98:9 MEX 82:2' ZAP 98:7' KOR 74:7' DOM 88:5' PRT 82:6 BRA 99:1 MUS 74:5 FIN 86:8' SWE 82:10' KAZ 99:4 URY 74:4' JAM 86:10 URY 82:12' AUS 75:12' KOR 86:1' GRC 83:1 8OL 75:9' BOL 87:3 ION 83:4' CRI 75:10' CAN 87:4' JAM 83:11' GAB 75:1' GAB 87:1' PHL 83:10 IDN 75:3' KOR 87:12 ITA 75:7'

~~Note: * indicatesattacks in the financially-developedsubsample.~~

~~33 Table 2: Selected Speculative Attack Episodes~~

~~Country Date Classification % Growthin: of Attack ExchangeRate Reserves Asia1997~~

~~Indonesia 97:8 Succeed 11.2% -4.7% Korea 97:11 Succeed 11.3% -19.6% Malaysia1/ 97:8 n/a 6.6% 1.4% Philippines 97:7 Succeed 10.4% 6.1% Thailand 97:7 Succeed 17.6% -6.1%~~

~~Europe1992-93~~

Belgium2/ 92:11 n/a 7.2% -14.3% Denmark3/ 93:1 Fail 1.9% -28.6% France41 93:11 Fail 2.8% -22.7% Ireland5/ 92:09 Fail 0.3% -3.8% Italy6/ 92:10 n/a 11.6% 7.7% UnitedKingdom 7/ 92:10 Succeed 11.6% -4.4% Spain8/ 92:09 Fail 4.5% -20.0% Sweden 92:11 Succeed 11.5% -24.6% Finland 92:11 Succeed 11.6% 1.9%

11Nominal devaluation too smallto qualifyas successfulattack. 2/ Nominaldevaluation too smailto qualifyas successfulattack. ERWdate is 92:9. 3/ ERWdate is 92:9. 41ERW date is 92:9. 5/ ERWdate is 92:11. 61ERW dateis 92:9. Exchangerate toovolatile prior to attackto qualifyas successfulattack. 7/ ERW dateis 92:8. 8/ Subsequentdevaluation of pesetain 92:10too smailto preventclassification as failedattack.

~~ERW: Eichengreen,Roseand Wyplosz (1994).~~

~~34 Table 3: Changes In Monetary Policy During Speculative Attacks (Full Sample)~~

~~Discount Rates SpeculativeAttack: Succeeds Fails Total Monetary Policy Tightens 37 95 132 Eases 40 72 112 Total 77 167 244~~

~~Estimate 95% ConfidenceInterval~~

~~P[Tightensi Fails] 0.57 0.49 0.65 P[FailsI Tightens] 0.72 0.64 0.80~~

~~P-Valuefor Independence 0.20 Domestic Credit Growth SpeculativeAttack: Succeeds Fails Total MonetaryPolicy Tightens 41 83 124 Eases 59 101 160 Total 100 184 284~~

~~Estimate 95% ConfidenceInterval~~

~~P[TightensI Fails] 0.45 0.38 0.52 P[FailsI Tightens] 0.67 0.58 0.75~~

~~P-Valuefor Independence 0.50 Bank Reserves SpeculativeAttack: Succeeds Fails Total MonetaryPolicy Tightens 39 58 97 Eases 47 119 166 Total 86 177 263~~

~~Estimate 95%Confidence Interval~~

~~P[TightensI Fails] 0.33 0.26 0.40 P[FailsI Tightens] 0.60 0.50 0.70~~

P-Valuefor Independence 0.05 Notes:The contingency tables report the distribution of speculativeattacks according to a two-way classificationof whethermonetary policy tightened or notusing the indicated measure of monetarypolicy, and whetherthe attack succeeded or failed.P(Tightens I Fails] reports the conditional probability that monetary policytightens during failed attacks, and P[Fails I Tightens]reports the conditional probability that the attack failsduring episodes where monetary policy tightens. P-Value for Independencereports the p-value associatedwith a chi-squaredtest of independenceof the rows and columns of thecontingency table.

~~35 Table 4: Changes In Monetary Polilcy During Speculative Attacks (Financially-DevelopedSample)~~

~~Discount Rates SpeculativeAttack: Succeeds Fails Total MonetaryPolicy Tightens 13 47 60 Eases 15 40 55 Total 28 87 115~~

~~Estimate 95%Confidence Interval~~

~~P[Tightensj Fails] 0.54 0.43 0.65 P[FailsI Tightens] 0.78 0.68 0.89~~

~~P-Valuefor Independence 0.48 Domestic CrediltGrowth SpeculativeAttack: Succeeds Fails Total MonetaryPolicy Tightens 14 35 49 Eases 22 49 71 Total 36 84 120~~

~~Estimate 95%Confidence Interval~~

~~P[TightensI Fails] 0.42 0.31 0.52 P[FailsI Tightens] 0.71 0.59 0.84~~

~~P-Valuefor Independence 0.78 Bank Reserves SpeculativeAttack: Succeeds Fails Total MonetaryPolicy Tightens 17 29 46 Eases 18 53 71 Total 35 82 117~~

~~Estimate 95%Confidence Interval~~

~~P[TightensI Fails) 0.35 0.25 0.46 P[FailsI Tightens] 0.63 0.49 0.77~~

~~P-Valuefor Independence 0.18~~

Notes: The contingencytables reportthe distributionof speculativeattacks according to a two-way classificationof whethermonetary policy tightened or not using the indicatedmeasure of monetarypolicy, and whetherthe attacksucceeded or failed. P[TightensI Failsj reportsthe conditionalprobability that monetary policytightens during failed attacks,and P[Fails I Tightens]reports the conditionalprobability that the attack fails duringepisodes where monetary policy tightens. P-Valuefor Independencereports the p-value associatedwith a chi-squaredtest of independenceof the rowsand columns of the contingencytable.

~~36 Table 5: WeakerTests of Necessity and Sufficiency (Full Sample)~~

~~Increasein RealDiscount Rate~~

~~Failed Successful Marginal Numberof Attacks Attacks P-Value Effect t-Statistic Observations~~

FullSample 1.342 -3.111 0.085 0.003 1.752 239 OECD 1.173 -8.492 0.040 0.006 2.163 88 1980sand 1990s -1.378 -2.847 0.644 0.001 0.471 136 No BankingCrises 1.655 -3.127 0.123 0.003 1.601 188 No RealOvervaluation 1.652 -1.279 0.437 0.002 0.861 113 HighReserves 1.067 -4.784 0.123 0.003 1.509 124 LowQuota Drawings 1.411 1.021 0.892 0.000 0.119 112 HighPoint in Cycle 0.391 -2.443 0.526 0.001 0.677 116

~~Decreasein RealDomestic Credit Growth~~

FullSample -3.657 0.404 0.482 -0.001 -0.717 230 OECD -1.779 13.801 0.056 -0.002 -1.623 75 1980sand 1990s -2.043 -4.808 0.734 0.000 0.348 131 No BankingCrises -4.032 4.096 0.204 -0.001 -1.298 184 No RealOvervaluation 0.499 7.143 0.452 -0.001 -0.790 109 HighReserves -1.515 1.907 0.677 0.000 -0.437 123 LowQuota Drawings 1.628 -3.868 0.501 0.001 0.674 100 HighPoint in Cycle -6.527 1.848 0.316 -0.001 -1.001 125

~~Increasein BankReserves~~

FullSample -0.522 -0.140 0.073 -0.033 -1.777 263 OECD -0.295 0.038 0.131 -0.071 -1.147 77 1980sand 1990s -0.399 -0.200 0.450 -0.017 -0.726 154 No BankingCrises -0.652 -0.328 0.128 -0.033 -1.445 204 No RealOvervaluation -0.536 0.031 0.017 -0.052 -1.841 126 HighReserves -0.544 -0.099 0.172 -0.033 -1.403 142 LowQuota Drawings -0.483 -0.380 0.741 -0.009 -0.321 114 HighPoint in Cycle -0.519 -0.208 0.279 -0.021 -0.950 143

Notes: The firsttwo columnsreport the meanvalue of the indicatedpolicy variable during failed and successfulattacks, in the indicatedsubsample of events. The thirdcolumn reports the p-valueassociated with a test of the null hypothesisthat the meansare equalin the two samples. Thefourth columnreports the estimatedmarginal effect of policy in a probitregression expressing the probabilitythat a speculativeattack fails as a functionof a constantand the indicatedpolicy variable, in the indicatedsubsample of events. The fifthcolumn reports the t-statisticassociated with the estimateof the underlyingslope coefricient.The final columnindicates the number of observationsfor whichthe policyvariables are availablein the indicated subsampleof events.

~~37 Table 6: Weaker Tests of Necessity and Sufficiency (Financially-Developed Sample)~~

~~Increasein RealDiscount Rate~~

~~Failed Successful Marginal Numberof Attacks Attacks PI-Value Effect t-Statistic Observations~~

Full Sample -0.155 -4.464 0.177 0.004 1.409 115 OECD 0.725 -11.236 0.016 0.008 2.540 71 1980sand 1990s -1.671 -0.972 0.836 -0.001 -0.214 65 No BankingCrises 1.181 -4.713 0.127 0.005 1.535 84 No RealOvervaluation 1.728 -1.163 0.555 0.003 0.640 56 HighReserves -0.267 -4.331 0.260 0.004 1.130 71 LowQuota Drawings 0.785 -1.590 0.556 0.002 0.515 66 HighPoint in Cycle -0.664 -6.405 0.242 0.005 1.258 57

~~Decreasein Real DomesticCredit Growl.h~~

FullSample -7.339 -3.195 0.561 -0.001 -0.542 108 OECD -5.725 6.347 0.103 -0.002 -1.191 63 1980sand 1990s -8.812 -4.203 0.635 -0.001 -0.438 61 No BankingCrises -6.166 3.530 0.233 -0.001 -1.083 83 No RealOvervaluation 1.780 -2.841 0.635 0.001 0.458 54 HighReserves -7.870 -0.305 0.457 -0.001 -0.751 67 LowQuota Drawings -3.078 -1.327 0.806 0.000 -0.170 62 HighPoint in Cycle -11.119 0.831 0.280 -0.001 -0.986 55

~~Increasein Bank Reserves~~

Full Sample -0.247 0.280 0.117 -0.050 -1.654 117 OECD -0.340 -0.081 0.246 -0.046 .-0.757 64 1980sand 1990s -0.099 0.220 0.339 -0.033 -0.837 68 No BankingCrises -0.380 -0.066 0.149 -0.065 -1.161 87 No RealOvervaluation -0.246 0.017 0.255 -0.064 -t).808 57 HighReserves -0.454 0.304 0.141 -0.064 -1.512 65 LowQuota Drawings -0.391 0.056 0.273 -0.052 -1.107 65 HighPoint in Cycle -0.071 0.350 0 224 -0.040 C.0.994 61

Notes:The first two columns report the mean value of the indicatedpolicy variable during failed and successfulattacks, in the indicatedsubsample of events.The third column reports the p-value associated witha testof thenull hypothesis that the means are equal in thetwo samples. The fourth column reports the estimatedmarginal effect of policyin a probitregression expressing the probability that a speculativeattack failsas a functionof a constantand the indicated policy variable, in theindicated subsample of events.The fifthcolumn reports the t-statistic associated with the estimateof theunderlying slope coefficient. The final columnindicates the number of observations for whichthe policyvariables are available in theindicated subsampleof events.

~~38 Table7: Controllingfor the Endogeneityof Policy (Full Sample)~~

~~Fundamental(f): Banking Real Reserves/ IMF Growth Crisis Overvaluation Imports Borrowing Deviation~~

~~Increase in Real Discount Rate~~

i 01 0.015 0.009 -0.142 0.037 -0.052 se(13) 0.152 0.462 0.188 0.129 0.220 dP[Fail]/d13 0.005 0.003 -0.017 0.012 -0.012 ixf 13 -0.070 0.000 0.090 0.000 0.013 se(p) 0.244 0.040 0.095 0.001 0.031 dP[Faildtdp -0.024 0.000 0.011 0.000 0.003 f 1 -0.221 0.011 -0.192 -0.002 -0.061 se(1) 1.955 0.265 0.295 0.004 0.308 dP[Failydo -0.074 0.004 -0.023 -0.001 -0.014 p-OlD 0.830 0.985 0.628 0.702 0.918 p-FSR 0.031 0.134 0.102 0.015 0.145 Numberof Observations 202 199 191 202 180

~~Decreasein RealDomestic Credit Growth~~

i13 -0.105 -0.093 0.118 -0.065 0.001 se(p) 0.134 0.110 0.226 0.084 0.081 dP[Fail]/d13 -0.011 -0.008 0.011 -0.011 0.000 ixf , 0.137 -0.007 -0.038 0.000 -0.002 se(f3) 0.162 0.009 0.052 0.000 0.017 dP[FaillJd 0.014 -0.001 -0.004 0.000 -0.001 f 1 0.652 0.019 0.215 -0.002 0.051 se(13) 1.538 0.055 0.432 0.003 0.187 dPVFaill/d13 0.066 0.002 0.021 0.000 0.017 p-OlD 0.716 0.848 0.759 0.559 0.915 p-FSR 0.329 0.762 0.477 0.682 0.075 Numberof Observations 200 197 191 199 182

~~Increasein BankReserves~~

i 13 -1.339- -2.266 -4.834 -2.122* -1.298 se(1) 0.574 5.489 3.419 1.191 1.725 dP[Failydo -0.254 -0.229 -0.506 -0.263 -0.213 ixf p 2.465 -0.065 0.850 0.004 -0.083 se(D) 1.919 0.245 0.764 0.006 0.129 dP[FailIl/dp 0.467 -0.007 0.089 0.001 -0.014 f .1 1.345 -0.029 0.451 0.004 -0.022 se(13) 0.841 0.168 0.496 0.003 0.160 dP[Faill/d13 0.255 -0.003 0.047 0.000 -0.004 p-OlD 0.957 0.952 0.833 0.736 0.730 p-FSR 0.051 0.155 0.633 0.357 0.087 Numberof Observations 229 226 218 228 215

Notes: This table reports the results of a probit regression which expresses the probabiiity that a speculative attack fails as a function of the monetary policy response (i), a fundamental (f), and the interaction between the two (ixf), for the indicated measures of policies and fundamentals. All regressionsalso include a constant. The rows labelled ,B,se(,B), and dP[Fail]/d,Breport the estimated coefficient, its standarderror, and the estimated marginal effect. The policy variable and its interaction with the fundamental are treated as endogenous,and are instrumentedusing the contemporaneouschange in reserves and the change in IMF lending, and their interactionswith fundamentals. The model is estimated using Amemiya's two-stage GLS procedure for instrumentalvariables probit models. p-OlD reports the p-value for the test of overidentifying restrictions,and p-FSR reports the p-value for the null hypothesisthat the instruments are jointly insignificant in the first-stage regressionof the policy variable on the instruments. * (*) denotes significanceat the 10% (5%) level.

~~39 Table 8: Controlling for the Endogeneityof Policy (Financially-DevelopedSample)~~

~~Fundamental(f): Banking Real Reserves/ IMF Growth Crisis Overvaluation Imports Borrowing Deviation~~

~~Increasein Real Discount Rate~~

i 3 -0-039 0.105 -0.004 0.333 -0.005 se(,) 6.700 0.167 0.381 0.395 0.796 dP[Failyd13 -0.010 0.020 -0.001 0.032 -0.001 ixf 0 -0.053 -0.009 0.022 -0.001 0.001 se(p) 9.661 0.010 0.081 0.002 0.097 dP[Failydp -0.013 -C.002 0.006 0.000 0.000 f p -0.907 -0.029 0.024 0.000 0.034 se(p) 28.351 0.069 0.378 0.008 0.757 dP[Failyd3 -0.222 -0.006 0.007 0.000 0.010 p-OlD 0.993 0.779 0.915 0.886 0.961 p-FSR 0.366 0.091 0.386 0.479 0.228 Numberof Observations 105 104 104 105 104

~~Decreasein Real Domestic Credit Growth~~

i 13 0.007 -0.044 -0.010 -0.979 0.055 se(p) 0.628 0.062 0.114 12.762 3.053 dPEFail]/d13 0.002 -0.008 -0.003 -0.013 0.008 ixf 1 0.010 0.004 0.004 0.004 -0.011 se(p) 1.114 0.006 0.025 0.046 0.605 dPtFailyd1 0.003 0.001 0.001 0.000 -0.002 f p -0.217 0.018 0.060 -0.006 -0.106 se(p) 8.346 0.060 0:379 0.078 7.385 dP[Fail]1d13 -0.067 0.003 0.019 0.000 -0.016 p-OlD 0.957 0.705 0.716 1.000 0.999 p-FSR 0.306 0.591 0.466 0.259 0.228 Numberof Observations 98 97 98 98 96

~~Increase in Bank Reserves~~

i 1 -2.946 -2.590 -2.811 -0.864 -6.951 se(p) 1.824 9.620 2.661 12.202 162.283 dP[Failydp -0.320 -0.261 -0.648 -0.193 -0.251 ixf ,B 2.080 0.116 0.674 -0.001 -0.681 se(13) 5.179 0.921 0.861 0.055 15.232 dP[Failyd13 0.226 0.012 0.155 0.000 -0.025 f 1 1.834 0.0118 0.163 0.002 -0.042 se(p) 1.847 0.420 0.286 0.062 2.365 dP[Faillyd 0.199 0.002 0.038 0.000 -0.002 p-OlD 0.712 0.903 0.649 0.956 0.996 p-FSR 0.620 0.965 0.612 0.347 0.897 Numberof Observations 105 104 104 105 105

Notes: This table reports the results of a probit regressionwhich expresses the probability that a speculative attack fails as a function of the monetary policy response (i), a fundamental (f), and the interaction between the two (ixt), for the indicated measures of policies and fundamentals. All regressions also include a constant. The rows labelled 1, se(3), and dP[Faillid13report the estimated coefficient, its standard error, and the estimated marginal effect. The policy variable and its interactionwith the fundamental are treatedas endogenous, and are instrumented using the contemporaneouschange in reserves and the change in IMF lending, and their interactionswith fundamentals. The model is estimated using Amemiya's two-stage GLS procedure for instrumentalvariables probit models. p-OlD reports the p-value for the test of overidentifying restrictions, and p-FSR reports the p-value for the null hypothesisthat the instruments are jointly insignificant in the first-stage regression of the policy variable on the instnrments. . * () denotes significanceat the 10% (5%) level.

40 log scale - o Co cn o cn Co cA o cn Co 0 CO C) C0 C) 02-Jun-97 -,J 0 F 1-Jun-92 . 12-Jun-97 m 11-Jun-92 22-Jun-97 O1- 23-Jun-92 02-Jul-97 og 121-Ju-97 3-Jul-92 22-Jul-97 15-Ju-92 201-Au-97 ~ ^ 27-Jul-92 - 01-Aug-97 6-Aug-92

~~210-Au-97 .- 18-Aug-92 31-Aug-97 6 28-Aug-92-. c~~

~~21 -Aec-97 s 8-Sep-929 : 0 20-Sep-97 21~~~~~~~~~~-Sep-92 0. 20-Sep-97 1-Oc-92 -0 20-Oct-97 23-Ot-9 10-oct-97 13-b)OcM(F2aocn >to -92~~

~~319-Aug-97 2 8-Aug-92z 30-Oct-97 4-Nov-92 09-Nov-97 ' 19-Nov-97 ~~~~~~~~~~~~~16-Nov-92 o ~~~~~~~ ~ oOctn92~ ~~~~ co 29-Nov-97 26Nv9 0.~~

~~09-Dec-97 8~~~~~~~~~~2-Dec-92 c 2 19-Dec-97 18-Dec-92 : C: (D~~~~~C~~

~~29-Dec-97 183-Dec-92 - I~~

~~0 -~~ CJ~~~~ -~~~ f~~3 ~~~) 0 CWJ rth. ui 0) -4j 0) 0 0 01 ~~ ~~~0CO 0 0 0 ~~~00 0) 0 Krona/$ Won/$ 0 Figure 2: Exchange Rates and Reserves During Successful and Failed Speculative Attacks~~

~~Successful Attacks~~

~~120 --~~

~~115- /~~

~~110 I~~

~~0, 105-~~

~~II~ ~~~~~~~9~~

~~85 -~~

~~-12 -10 -8 -6 -4 -2 C 2 4 6 8 10 12 Months Since Attack~~

~~Failed Attacks~~

~~120 -~~

~~110 -~~

~~80 -~~

~~70 -~~

~~,,I , , ,I , -60 , r , -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12 Months Since Attack~~

Notes: This figure shows the evolution of the median nominal exchange rate and reserves during successfuland failed speculativeattacks. The figuresare constructedby cumulatingthe median(across all episodes)growth rate of the indicatedvariables to a base of 100twelve months prior to the attack.

~~42 Figure 3: Changes in Real Discount Rates During Successful and Failed Speculative Attacks~~

~~0.25~~

~~Mean Change During Successful Attacks = 0.45 0.2 - eDSuccessful Attacks MeanChange During l Failed Attacks Failed Attacks = 0.30 as01~~

~~0 0.1 l~~

~~-25 -20 -15 -10 -5 0 5 10 15 20 25 >25 Percent Changes in Real Discount Rates Less Than:~~

Notes: This figure shows the frequency distribution of percentage changes in real discount rates during successful and failed speculative attacks. The mean changes during successful and failed attacks are based on changes in real discount rates less than 25% in absolute value.

~~43 Figure 4: Devaluation Probabilities as a Function of Interest Rates~~

~~DevaluationProbability~~

~~0.9~~

~~0.8~~

~~°0.7 - 2(i)~~

~~> 0.6 -~~

~~o 0.5~~

~~3. 0.4 2 0.3 0.2~~

~~0.1~~

~~0-I 0 0.05 0.1 0.15 0.2 0.25 0.3 InterestRate Non-LinearEffects of Policy 1~~

~~0.9~~

~~0.8 - 0 70(i),Low Reserves o 0.6 -~~

~~0 0.2 -~~

~~0.4 - ~~~~0.4 ~~~~~~~~~~ir(i), High Reserves 2 0.3- 0.2O-~~

~~0.1 0 0 0.05 0.1 0.15 0.2 0.25 0.3 InterestRate~~

~~44 Figure 5: The Endogeneity of Policy~~

~~Case 1: Endogeneity Bias Obscures Conventional View~~

~~0.9~~

~~0.8 . 0 ~0.7-~~

~~> 0.6~~

~~0.5~~

~~.0Q~~

~~0.4~~

~~0.2-~~

~~0.1 I~~

~~0 i 0 0.05 0.1 0.15 0.2 0.25 0.3 InterestRate Case 2: Endogeneity Bias Exaggerates Conventional View 1~~

~~0.9 0.8~~

~~~0. 0.2~~

~~0.4 - l l l l l l~~

~~00.5~0.2 - 0 0 0 0~~

~~20.3~ ~ ~ ~ ~~4~~

~~Interest Rate~~

~~Policy Research Working Paper BSries~~

~~Contact Titie Author Date for paper~~

~~WPS2252 Productivity Growth, Capital Ejaz Ghani December 1999 N. Mensah Accumulation, and the Banking Vivek Suri 80546 Sector: Some Lessons from Malaysia~~

~~WPS2253 Revenue Recycling and the Welfare Ian W. H. Parry December 1999 R. Yazigi Effects of Road Pricing Antonio Miguel R. Bento 37176~~

~~WPS2254 Does "Grease Money" Speed Up Daniel Kaufmann December 1999 H. Sladovich the Wheels of Commerce? Shang-Jin Wei 37698~~

~~WPS2255 Risk and Efficiency in East Asian Luc Laeven December 1999 R. Vo Banks 33722~~

~~WPS2256 Geographical Disadvantage: Anthony J. Venables December 1999 L. Tabada A Heckscher-Ohlin-von Thunen Nuno Limao 36896 Model of International Specialization~~

~~WPS2257 Infrastructure, Geographical Nuno Limao December 1999 L. Tabada Disadvantage, and Transport Costs Anthony J. Venables 36896~~

~~WPS2258 Market Access Bargaining in the J. Michael Finger December 1999 L. Tabada Uruguay Round: Rigid or Relaxed Ulrich Reincke 36896 Reciprocity? Adriana Castro~~

~~WPS2259 Predicting Currency Fluctuations Daniel Kaufmann December 1999 E. Khine and Crises: Do Resident Firms Gil Mehrez 37471 Have an Informational Advantage? Sergio Schmukler~~

~~WPS2260 Regional Integration Agreements: Anthony J. Venables December 1999 L. Tabada A Force for Convergence or 36896 Divergence?~~

~~WPS2261 is Knowledge Shared within Kaushik Basu December 1999 M. Mason Households? Ambar Narayan 30809 Martin Ravallion~~

~~WPS2262 How Inadequate Provision of Public Ritva Reinikka December 1999 H. Sladovich infrastructure and Services Affects Jakob Svensson 37698 Private Investment~~

~~WPS2263 When Is Growth Pro-Poor? Evidence Martin Ravallion December 1999 J. Israel from the Diverse Experiences of Gaurav Datt 85117 India's States~~

~~WPS2264 Do More Unequal Countries Branko Milanovic December 1999 P. Sader Redistribute More? Does the Median 33902 Voter Hypothesis Hold? Policy Research Working Paper Series~~

~~Contact Title Author Date for paper~~

~~WPS2265The PoliticalEconomy of Distress PaolaBongini January2000 R. Vo In East Asian FinancialInstitutions Stijn Claessens 33722 GiovanniFerri~~

~~WPS2266The Impactof Adult Deathson MarthaAinsworth January2000 S. Fallon Children'sHealth in Northwestern InnocentSemali 38009 Tanzania~~