Currency Crashes in Emerging Markets: An Empirical Treatment

Boardof Governorsof the FederalReserveSystem InternationalFinanceDiscussionPapers Number534 January 1996 CURRENCYCRASHESIN EMERGINGMARKETS: AN EMPIRICALTREATMENT JeffreyA. Frankeland AndrewK. Rose NOTE: InternationalFinanceDiscussionPapersare preliminarymaterialscirculatedto stimulate discussionand criticalcomment. Referencesin publicationsto InternationalFinance DiscussionPapers(otherthanan acknowledgmentthat the writerhashadaccessto unpublishedmaterial)shouldbe clearedwiththe authoror authors.

We usea panelof annualdata for overone hundreddevelopingcountriesfrom 1971through 1992to characterizecurrencycrashes. Wedefinea currencycrashas a largechangeof the nominalexchangeratethat is alSOa substantialincreasein the rate of changeofthe nominal depreciation. Weexaminethe compositionof the debtas wellas its Zeve/,anda varietyof othermacroeconomic,externaland foreignfactors. Ourfactorsare significantlyrelatedto crashincidence.especiallyoutputgro~h, the rateof changeof domesticcredit,and foreign interestrates. A lowratioof FDIto debt is consistentlyassociatedwith a highlikelihoodof a crash.

Currency Crashes in Emerging Markets: An Empirical Treatment JeffreyA. Fran.keland AndrewK. Rose* . l. Introductlou Are currencycrashesin developingcountriesall theresultof similarpolicymistakes? Or arethey insteadthe resultof a myriadunfortunateshocks? Cantheybe predictedexanre with standardeconomicindicators? Do differentcountriesexpost reactto crashesin similar fashion,or dothe policyresponsesvary by countryandovertime? In short:Are currenc)’ crashesaIlalike? The objectiveof this studyis to lookat a largesampleof developingcountryexperiences,andto arriveat a broad-brushstatisticalcharacterizationof currencycrashes. It is not an attemptto formulateor test specifictheoriesof whatcausesthese crashes. Somemay havebeencausedby idiosyncraticshockswhichare betterviewedas bad luckthananything else. Othersmayhaveresultedfrompoor fundamentalsor poorpolicy. Weexaminea varietyof potentialcausesof crashes,especiallythose thataddto a country’svulnerabilityto a crash. Wealso lookat theeffectsof currencycrashes. * FrankelisProfessorofEconomicsintheEconomicsDepartmenatttheUniversityofCalifornia, Berkeley,andDirectorof theNBER’sInternationalFinanceandMacroeconomicpsrogram.Rose is AssociateProfessorandChairof EconomicAnalysisand Policyinthe Haas Schoolof Businessatthe Universityof California,Berkeley,a ResearchAssociateof theNBER, and a Research Fellowof the CEPR. For comments,we thank:AndrewAtkeson,Jon Faust MarvinGoodfriend,EnriqueMendoza, JohnRogers,RalphTryon,CarlosVegh,seminarparticipantsattheInternationalFinanceDivisionofthe Board of Governors of the Federal Reserve System and conferenceparticipants at the Center for InternationalEconomicsConferenceonSpeculativeAttacksintheGlobalEconomy. WealsothankBill Easterly,BarryEichengreen,andAlanStockmanfordiscussionandencouragement,andtheWorldBank forfinancialsupport. TheSTATA4.0datasetandprogramsareavailableuponreceiptoftwoformatted 3.5”diskettesanda self-addressed,stampedmailer. ThisworkwaspetiormedinpartwhileRosewasa Visiting Scholarin thelntemationalFinanceDivision.Thispaperrepresentstheviewsoftheauthors and shouldnotbe interpretedas reflectingthe viewsof the Boardof Governorsof the FederalReserve Systemor membersof its staff.

Weclassi~ thevariablesin whichwe are interestedintofourcategories:I)joreign variables1ikeNortherninterestratesand output;2) domes(icmacroeconomicindicators,such as output,monetaryandfiscalshocks;3) exfernalvariablessuchas the over-valuation,the currentaccountandthe 1evelof indebtedness;and4) the compositionof the debt. Wefocus on the lastset ofvariables,as theyhave attractedincreasedinterestin the afiermathof the 1994Mexicancrash. Ourworkis non-structural,andtakesthe formof univariategraphical analysisand multivariatestatisticalanalysis. In sectionII,we discussour definitionof a currencycrash;the variablesthat we examine are analyzedin the sectionwhichfollows. SectionIV is the heartof the paper. It analyzes the movementsof ourvariablesaroundcwency crashesusingbothunivariategraphical techniquesanda multivariatestatisticalapproach. Thepaperendswith a briefconclusion. . l he D~lonl of a Currencv Crash l - Wedefinea currencycrashas a depreciationof the nominalexchangerateof at least 25 per centthat is alsoat leasta 10per cent increasein the rateof nominaldepreciation. Fourconceptualissuesimmediatelyarise. First,shouldcurrencycrashesbe limitedto episodesthat endin a largefall in the valueof the currency? Second,how big a changein the exchangerate is neededto quali&? Third,how shouldtheexchangerate be measured? Fourth,how doesonedealwith high-inflationcountriesthatroutinelyundergolargechanges in the exchangerate? Eichengreenet.al. (1995)definea cwency crisisto includeboththe largedepreciations thatwe considerhere,andalso speculativeattacksthataresuccessfullywardedoffby

the authorities. Theymakethe ideaof an unsuccessfulspeculativeattackoperationalby searchingfor suddenfallsinreserves and/orincreasesin interestrates. Sincewe focuson developingcountriesin this paper,it is muchmoredifficultto identi&successfuldefenses againstspeculativeattacks. Reservemovementsare notoriouslynoisymeasuresof exchange marketinterventionforalmostall countries. In addition,fewof our countrieshavemarketdeterminedshort-terminterestratesfor longperiodsof time. Thestandarddefensesagainst speculativeattacksof interestratehikesand reserveexpendituresmayalso be lessrelevantin thesecountriesthan suddentighteningof reserverequirements.emergencyrescuepackages fromthe IMFor otherforeigninstitutions,and especiallythe impositionof formalor informal controlson capitaloutflows. It is extremelydifficultto measuresuchpolicyactions,andwe leavethis taskto futureresearchers. But whileextendingthe analysisto take accountof preemptivedevaluationsand successfuldefensesis important,it mayalso be of intrinsicinterest to 1ookat successfulattacks. Wedefinea cumencycrashas a decreasein the vaiueof the localcumencyof at least 25 per cent. Thiscut-offpointis clearl>’arbitrary;sensitivityanalysis has assuredus thatthe exactfigureis not important. Thethirdquestionis howthe exchangerate shouldbe measured. We use the percentage change in the natural logarithm of the nominal bilateral dollar exchange rate. Until the 1970s, devaluations werediscretechangesin the exchangerate,whichwereeasilyidentified expost. However,developingcountrieshave morerecentlytakenadvantageof moreflexible exchangerate arrangements,includingcrawlingpegs,targetzones,and evenglidingbands. This forcesus to use a techniquewhichcan accommodatea diverseset of underlyingex-

changerate regimes. Also.manyof the countrieswe consider(notody LatinAmerican.but EastAsianas well)usethe U.S.dollarto definetheirexchangerate;Frankeland Wei (1994,1995). Henceouruse of a simplestatisticalcriterionusingdollarbilateralrates. Thefourthquestionis howto deal withcountriesthatmeetour criterion-- changesin the exchangerateof 25 per centor more-- yearafieryear. Thesearecountrieswith high ifiation rates and correspondinglyhigh ratesof depreciation. To ensurethat we do not considereachof thesedepreciationsto registeras an independentcrash,we requirethat the changein the exchangerate,not onlyexceed25 per cent,but exceedthepreviousyear’s changein the exchangerateby a marginof at least 10per cent. Wealsodefinea three-year “window”aroundcrashes,as explainedin the empiricalsectionbelow. . . . he Var@les of Interest We groupthe domesticvariablesinto fourcategories:internaldomesticmacroeconomic variables,factorspertainingto the levelof internationalindebtednessandotherexternaZ variables,those pertainingto the compositionof the debtstock,and~oreignvariables. Macroeconomic Indicators The academicliteratureon “speculativeattacks”is relevantto ouranalysis,eventhough empiricaltests are as yet rathermeager,and largelylimitedto currencycrisesamongindustrializedcountries. Krugman(1979)is the classictheoreticalmodelof currencycrisesas speculativeattacks. The originalpaperassumedthat the pre-crisisregimewas literallya fixed exchange

rate, but the model has beenextendedto crawlingpegs (Connolly,1986)and cumencybands (hgman and Rotemberg,1991). The speculativeattackmodeldeliversseveralfactorsthat shouldbe importantin predictingcwency crashes:monetaryand fiscalexpansions,declining competitiveness,currentaccountdeficits,and losses in internationalreserves. Whilesomeof the predictionsof thesemodelshavebeenborneout empirically,some speculativeattackshavetakenpIacewithoutlargeapparentmonetaryandfiscalimbalances. Theresponsehas beena “secondgeneration”of multiple-equilibriummodelswhichgenerate self-fulfillingattacks:Eichengreen,Roseand Wyplosz(1995)providea review, Thesemodels tendto focuson politicalfactors,suchas the politicalcost of highunemploymentor foregoneoutputwhichresultfroma toughdefenseagainsta speculativeattack. Weexaminesix variablesrelevantto the speculativeattackliterature:the rate of growth of domesticcredit(a crudemeasureof monetarypolicy),the governmentbudgetas a fraction of GDP(a crudemeasureof fiscalpolicy),the ratio of reservesto imports,the currentaccountas a percentageof GDP,the growthrate of real output,and the degreeof over-valuation. External Variables Externalvariablesarecriticalto our analysis. Weusethe ratioof debt to GNPas our prim~ measureforthe levelof internationaldebt. Wealsousethe ratioof foreignexchangereservesto monthlyimports,the ratio of the currentaccountto GDP,and the real 5

exchangerate(whichmeasurescompetitiveness) as additional measures of vulnerability to extemaishocks.c All have been widely used in the literature. Composition Thecompositionof bothcapitalinflowsandthe stockof debt has receivedmuchattention recently. RelevantindicatorsincludeForeignDirectInvestment(FDI)vs.portfolio flows,long-termVS.short-tern portfoliocapital,fixed-ratevs. floating-ratebonowing,and domestic-currencyVS.foreign-currencydenomination. Thesevariablesare a centralfocusof this study. ThehypothesisregardingForeignDirectInvestmentis that FDI is a saferwayto financeinvestmentthan is potifolioinvestment. Oneargumentis that FDI is directlytiedto real investmentin plant.equipmentandinfrastructure;whereasbomowingcango to finance consumption. Borrowingto financeconsumptiondoes not help add to the productivecapacity necessaryto generateexportearningsto servicethe debt in the fiture. ButFDItids may be fungible;an FDIsurplus in the capitalaccountis no guaranteeof high investment. Thestrongerargumentin favorof FDIisthat of stability. In the eventof a crash, investorscan suddenlydumpsecuritiesand bankscan refuseto roll overloans,but multinationalcorporationscannotquicklypackup theirfactoriesand go home. Yetthisargument too hasbeenquestioned. Dooleyez.al. (1995)havefoundthat a high levelof FDIseemsto ?. A variety ofotherratioshavealsobeenproposedintheliteraturaesproxiesforthelevelofthedebt burden.Theseincludet:he interestioutputratio;thedebtiexporrattio;theinterestiexporarttio;thedebtservice/exportratio;andthecumentaccountiexportratio. 6

be associatedwith higher variabilityin capitalflows,not lower. Thisprobablyreflectsmultinationalcorporationsmovingmoneyin and out of the country,throughtransfersbetween subsidiaryandparent,withgreatereasethan canbe doneoutsidethe corporatewalls. It makesthe FDIhypothesisworthtesting. Tworelevantaspectsof the compositionof capitalinflowsarethe fractionof debt whichis confessionaland the fractionthat comesfromrnulti!ateraldeve[opmen!banks. In both cases, the capital is both easier to service and fm less likely to depart quickly in times of trouble than is the case for private market-rate debt. Indeed, the inflows from these sources may even increase in a crash. Within portfolio capital, the maturi~~structureis perhaps the most important of the composition issues, followed closely by the question of variubZe-ra/earrangements. In the high-inflation 1970s, there wasa worldwide movement toward shorter maturities and nominal interest rates that were indexed to short-term interest rates such as LIBOR to protect creditor. The debt crisis that erupted in 1982 was clearly exacerbated by the fact that so much internationaldebt was tied to short-tern nominai interest rates. In the Mexican crash of 1994, the problem took the form of a heavy concentration of short-term debt, which describes the tesobonos as well as the CETESandajustobonos. This not onlyraisedthe costof borrowing in linewithU.S.interestrate increasesin 1994,but alsoresultedin difficultiesassociated withroilingoverthe debt lateron. In otherwords,shortmaturitiesapparentlyposeproblems of defaultrisk aboveand beyondthoseproblemsof interestrateriskthat they sharewith floating-ratedebt; Cole and Kehoe(1995)providemoreanalysis. Bothcompositionquestions,short-termvs. Iong-tem and floating-ratevs. fixed-rate,seemworth investigating.

Weare alsointerestedin the distinctionbetweensecuritiessalesandcommercialbank bomowing. Syndicatedcommercialbank loanswerethe preferredvehicleof international financein the 1970s,butthe 1982crisischangedthat. In the 1990s,theirplacehasbeen largeiybeentakenbyportfoliomanagersand institutionalinvestorsbuyingstocksandbonds (as wasthe normbeforeWwI). Somehavearguedthatcrashesin the 1990sare likelyto be f= lesscostlyto the bomowingcountriesthan wasthe crisisof the 1980s,becausecountries needno longerdealwithbanksto the samedegree. Also,equitiesare a moreefficientvehicle forrisk-sharingthaneitherloansor conventionalbonds. Withequities,unlikebondsor bankloans,the costof the obligationdoesnot stay fixedwhenthe abilityof the countryto earnexportrevenuefalls. Foreign Variables It is critical to look not only at individual country variables, but at the global financial environment as well. Globalvariablespotentiallyincludeworldeconomicactivity,commodity prices,real interestrates,andotherfinancialmarketshocks. The debtcrisisof 1982.and subsequentdebtordevduations.wereto a largeextenttriggeredby the tightNorthernmone- ~ policywhichresultsin highinterestratesand a global recession. It is quite striking that most of the econometric studies that were undertaken in 1993-94 on the causes of renewed large capitalinflowsto LatinAmericaand EastAsiain theearly 1990sconcludedthat externalfactorswerea majorcause,perhapsthe majorcause. Calve, Leidermanand Reinhart(1993,p. 136-137)foundthat “foreignfactorsaccountfora sizable fraction(about50 per cent)of the monthlyforecasterrorvariancein thereal exchange

rate...[and]...alsoaccountfora sizablefractionof the forecasterrorvariancein monthly reserves.” Theywarnedthat “Theimportanceof externalfactorssuggeststhat a reversalof thoseconditionsmayleadto a fiture capitaloutflow.” Chuhan,Claessensand Marningi (1994)estimatedthat U.S.factorsexplainedabouthalfof pofifolioflowsto LatinAmerica, thoughtheyexplainedlessthan countryfactorsin the caseof EastAsia. Femandez-Arias (1994)foundthatthe fall in U.S.returnswasthe key causeof the changein capitalflowsin the 1990s. Dooley,Femandez-Ariasand Klet.zer(1994),studiedthe determinantsof the increasein secondarydebt prices among 18 countries since 1986 and concluded that “Intemational interestratesare the key underlying factor.” The steep rise in American interest rates during 1994 constituted a test of the warning which most of these studieshad carried[explicitly or implicitly], that an adverse shifi in world financial conditionscould leadto an abrupt halt to the inflows and a new crisis on the order of 1982. In this paper, we focuson two important foreign variables: short-term Northern interest rates and real OECD output growth.3 The Data Set Mostof ourdata set was extractedfromthe 1994WorldBank”sWorZdDazacd-rem . It consistsof annual observationsfrom 1971through1992forone hundredand fivecountries.4 The samplewas selected(withrespectto choiceof bothcountryand time)to maxi- 3 Wehaveaddedboththe levelandthepercentagechangeintheIMF’sDevelopingCountryCommodity PriceIndex,butneitheriseversignificant. 4 The countrieswe includeare:Algeria;Argentina;Bangladesh;Barbados;Belize;Benin;Bhutan; Bolivia;Botswana;Brazil;BurkinaFaso;Burundi;Cameroon;CapeVerde;CentralAfricanRepublic;Chad: Chile;China;Colombia;Comoros;Congo;CostaRica;Coted’Ivoire;Djibouti;DominicanRepublic:Ecuador; ArabRepublicofEgypt;ElSalvador;EquatorialGuinea;Ethiopia;F~i;Gabon;TheGambia;Ghana;Grenada: 9

mizedataavailability. However,numerousobservationsare missingfor individualvariables. Wecheckedthe dataviaboth simpledescriptivestatisticsand graphicaltechniques. We have alsoused exchangeratesandinterestratesfromthe IMF’sInternationalFinancial Sfafisticscd-rem,andaggregaterealoutputfromthe OECD. Weexaminesevendifferentcharacteristicsof thecompositionof capitalinflowsor the debt. Each is expressedas a percentageof the total stockof externaldebt. Thevariablesare: 1)the amountof debtlentby commercialbanks;2) the amountwhichis confessional,3)the amountwhichis variable-rate;4) the amountwhichis publicsector,5) the amountwhichis short-term;6) the amountlentby multilateraldevelopmentbanks(this includesthe World Bankand regionalbanks,but not the InternationalMonetaryFund;and 7) the flowof Foreign DirectInvestment(FDI)expressedas a percentageof the debtstock As measuresof vulnerabilityto externalshocks,we examine:1)the ratioof totaldebt to GNP;2) the ratio ofresemesto monthlyimportvalues;3) the currentaccountsurplus(+) or“deficit(-) expressedas a percentageof domesticoutput;and 5) the degreeof overvaluation. We definethe lattersimplyas the deviationfromPurchasingPowerParity(measuringthe latteras thecountry-specificaveragebilateralreal exchangerate overthe periodin question.) Guatemala;Guinea;Guinea-BissauG;uyana;Haiti;HondurasH; ungary;India;1ndonesial;slarnicRepublicof 1ran;Jamaica;Jordan;Kenya;RepublicofKorea;LaoPeople’sDemocraticRepublic;Lebanon;Lesotho;Liberia; Madagascar;Malawi;MalaysiaM; aldives;Mali;Malta;MauritaniaM; auritiusM; exico;Morocco;Myanmar; Nepal;Nicaragua;Niger;Nigeria;Oman;Pakistan;Panama;PapuaNewGuinea;Paraguay;Peru:Philippines: Pomgal;Romania;Rwanda;St.VincentandtheGrenadines;SaoTomeandPrincipe;Senegal;Seychelles;Sierra Leone;SolomonIslands:Somalia:SriLanka;Sudan;Swaziland;SyrianArabRepublic;Tanzania;Thailand; Togo;TrinidadandTobago;Tunisia;Turkey;Uganda;Umguay;Vanuatu;Venezuela:WesternSamoa;Republic of Yemen;FederalRepublicof Yugoslavia;Zaire;Zambia;andZimbabwe. 10

Formacroeconomicpurposes,we examine: 1)the totalgovernmentbudgetsurplus(+) or deficit(-) (again.expressedas a percentageof GDP);2) the domesticcreditgrowthrate; and 3) the growthrateofreal GDPper capita. Finally,we usethe percentagegrowthrateof realOECDoutput(in Americandollars, at 1990exchangeratesandprices)as our measureof Northerndemand. Weconstructthe “foreigninterestrate”asthe weightedaverageof short-terminterestrates forthe United States,Germany,Japan,France,the UnitedKingdomand Switzerland;the weightsare proportionalto the fractionsof debtdenominatedin the relevantcurrencies.sThereis a gooddeal of heterogeneityby country(within-year)in foreigninterestrates. However,they generally movetogether,risinginthe mid-1970s,the early 1980sandthe early 1990s. IV: Red Event Study Methodology We begin our investigation by characterizing the behavior of countries suffering from a currencycrash.Ourmethodologyis that usedby Eichengreen,Rose,and Wyplosz(1995). As noted.wedefinea crashas an observationwherethe nominaldollarexchangerate increasesby at least25°/0in a yearand has increasedby at least 10°/0morethan it did inthe previous year. We excludecrasheswhichoccurredwithinthreeyearsof eachotherto avoid countingthe samecrashtwice. Our definitionof a currencycrashyields 117differentcrashes(74 crashesare deleted becauseof the three-year“windowing.) Theseare spreadovera largenumberof countries, 5 WeuseIFSline60b,moneymarketinterestrates. Usinglendingrates(IFSline601)doesnotchange anyresults. 11

but havea slighttendencyto be clusteredin the early-to-mid1980s.Thusthe observations probablyshouldnot be treatedas independentobservations.b Non-crashobservationswhicharenot withinthreeyearsof a crashconstitutea sample of “tranquil”observations(someof theseobservationsoccurin countriesthat neverhada crashthroughoutthe sampleunderstudy). Weuse theseas a controlsample,andcompare behavioraroundcrashepisodeswithbehaviorduringperiodsof tranquility. The figureis a setof sixteen“smallmultiple”graphics,eachan “eventstudy”ofthe sort used in finance.Eachof the graphicsportraysthe movementin a variableof interest beginningthreeyearsbeforethe crashandcontinuingthroughthe crash(markedby a vertical bar) untilthreeyearsafterwards.Thus,the “seeds”of crashescan be examined,alongwith their afiermath. The averagesforperiodsof tranquilityareexplicitlymarkedwitha horizontal line,makingit easyto comparebehavioraroundcrashesto that duringmore“typical” 6 Theactuallistis:Argentina1975;Argentina1981;Argentina1987;Burundi1984;Benin1981; BurkinaFaso 1981;Bangladesh197S;Bolivia1973;Bolivia1982;Brazil1979;Brazil 1983;Brazil1987;Brazil 1992;Bhutan1991;Botswana1985;Cen~l AfricanRepublic1981;Chile 1973;Chile 1982;Coted’lvoire1981; Cameroon1981;Congo 1981;Comoros1981:CostaRica1981;CostaRica1991;DominicanRepublic1985; DominicanRepublic1990;Algeria1991;Ecuacior1983;Egypt1979;Egypt1990;Ethiopia199Z;Gabon1981; Ghana1978;Ghana1983;Guinea1986;Gambia,The1984;Guinea-Bissau1984;Guinea-Bissau1991;EquatorialGuinea1981;Guatemala1986;Guatemala1990;Guyana1987;Guyana1991;Honduras1990;Indonesia1979; 1ndonesia1983;India1991;Jamaica1978;Jamaica1984;Jamaica1991;Jordan1989;Laos1976;Laos1980; Laos1985;Lebanon1984;Lebanon1990;SriLanka1978;Lesotho1984;Morocco1981;Madagascar1981; Madagascar1987;Maldives1975;Maldives1987;Mexico1977;Mexico1982;Mexico1986;Mali1981; Myanmar1975;Malawi1992;Niger1981;Nigeria1986;Nigeria1992;Nicaragua1979;Nicaragua1985;Peru 1976;Peru 1981;Peru 1985;Philippines1983;Paraguay1984;Romania1973;Romania1990;Rwanda1991; Sudan 1982;Sudan1988:Senegal1981;Sien Leone1983;Siem Leone1989;ElSalvador1986;El Salvador 1990;Somalia1982;.Somalia1988;SaoTomeandPrincipe1987;SaoTomeandPrincipe1991;Swaziland 1984;SyrianArabRepublic1988;Chad 1981;Togo 1981;Trinidad& Tobago1986;Turkey 1978;Turkey1984: Turkey 1988;Tanzania1984;Tan=nia 1992;Uganda1981;Uruguay1975;Uruguay1983;Uruguay1990; Venezuela1984;Vanuatu1981;Zaire 1976:Zaire 1983;Zaire 1987;Zaire 1991;Zambia 1983;Zambia1989; Zimbabwe1983;andZimbabwe1991. 12

periodsof tranquility.’ Thescalesof individualpanelsare not comparableacrossvariables. nor is the sampiesize(becauseof data availabilityproblems). Meanvaluesare provided, alongwitha banddelimitingplus and minustwo standarddeviations.8 A graphicalapproachlikethis has disadvantages. The graphsare informal. More importantly,they are intrinsicallyunivariafe. They encouragereadersto examineindividual variablesby themselves,whereasthe norm in econometricsis to lookat the marginalcontributionof eachvariableconditionalon the others. But graphicalmethodsalsohaveadvantages.Theyimposeno parametricstructureon the da~ and imposefewof the assumptionswhichare sometimesnecessaryfor statistical inferenceor estimationbutare fiequentl>’untenable.This is especiallyappropriatein a nonstructuralexplorationof the data. They are ofienmoreaccessibleand informativethantables of coefficientestimates. Forthesereasons,we use our graphscautiously. Wealso verifyour ocularanalysiswithmorerigorousstatisticaltechniques,usingprobitmodelsestimatedwith maximumlikelihoodto checkour results. Graphical Analysis The resultsin the figureare essentiallyas hypothesized. Countriesexperiencingcurrency crashestendto have:highproportionsof theirdebt lentby commercialbanks(compared, as always,to tranquilobservations),high proportionsof theirdebton variable-ratetermsand 7 A +/-2confidenceinternalforthetranquilmeanistickedontheordinate,centeredaroundthetranquil mean. 8 Thesemaynotrepresentwell-definedconfidenceinternals,giventheissueofpotentialnon-independence. 13

in shortmaturities;and relativelylow fractionsof debtthatareconfessional,lentby the multilateralorganizationsor lentto the publicsector.Crashcountriestendto experience disproportionatelysmallinflowsof FDI(i.e.,relativelyhigh“hotmoney”portfolio)flows. Foreigninterestratestend to be highin the periodprecedingcurrencycrashes,exceedingtranquilforeigninterestratesby over one percentage point. This corroborates tie COrnmonly-heldviewthat foreigninterestratesare an importantsourceof currencycrashes. Also, Northerngrowthis muchlowerin the periodsaroundcrashes. Countriesexperiencingcrashesalsotendto haveexchangeratesthat are over-valuedby overten per cent. Unsurprisingly,debt burdensforcrashingcountriesare highandrising. Internationalreservesare also lowand falling. Thus,externalconditionsforcrashingcountries aregenerallyweak. Thereis one impo~t exception. While the cu~ent account is in deficit,this deficitis small(comparedwithtranquilobservations)and shrinking. Curiously enough,the governmentbudgetsituationis very similarto thatof the cment account;small stinking deficitswhichdo not varysignificantlyfromtimesof tranquility. Ournegativeresultson the currentaccountand fiscalsideare in strikingcontrastwith the literature. Theyare especiallyinterestingin lightof the strongresultswe findelsewhere on the domesticmacroeconomicside. For instance,domesticcreditgrowthis noticeably high,consistentwith the classicspeculativeattackmodel. Mostvariables(except,trivially,the realexchangerate)tendto movevery sluggishlyin the yearssurroundingcurrencycrashes.9This leadsoneto expectthat it will be difficult to predictthe exacttimingof a currencycrashwithprecision. Thenotableexception is the 9 Anyrevaluationeffectsontradeflowsappearto besmail. 14

growthrateof realoutputper capita,whichdips significantly(in boththe economicand statisticalsenses)belowthe tranquilnom in the yearof the crash. Of course.the directionof causalityis unclear(especiallyat the annualfrequency)sincethe crashmaybe precipitatedin part by slowgrowth,butmayalsoitselfinducerecession. Theflowof FDI variesdramatically acrossepisodesimmediatelyafierthe crash. Regression Analysis The“eventstud}.”a’nalysisis bothnaiveand intrinsicallyunivariate. Moreconfirmation can be providedby simpleregressionwork. In particular,weestimateprobitmodels linkingourbinarycrashmeasureto ourvariables. Wehavesevendebt-compositionregressors,eachexpressedas percentagesof total debt: 1)commercialbankdebt;2) confessionaldebt; 3) variable-ratedebt;4) short-termdebt; 5) FDI;6) publicsectordebt;and 7) multilateraldebt.Ourlist of externalvariablesincludes: 1)the ratioof internationalreservesto monthlyimports;2) the currentaccountas a percentage of GDP;3) the externaldebtas a percentageof GNP;and 4) realexchangerate divergence(over-valuation).Asdomesticmacroeconomicvariables,we include:1)the governmentbudgetas a percentageof GDP;2) the percentagegrowthrateof domesticcredit;and 3) the percentagegrowthrateof realoutputper capita. Wealso includethe foreigninterestrate andthe Northerngrowthrate,both in percentagepoints. Weuse a multivariatemodelwhereall the variablesare employedsimultaneously. Throughout.we pool allthe availabledataacrossboth countriesandtimeperiods.and 15

estimateprobitmodelsusingmaximumlikelihood. Combiningthe effectsof the variablestogetherintoa singlemodelreducesthe samplesizedramatically. Ourbenchmarkresultsaretabulatedinthe middleof Table 1. Sinceprobitcoefficients are not easilyinterpretable,we reportthe effectsof one-unitchangesin regressorson the probabilityof crash(expressedin percentagepoints),evaluatedat the meanof the data. We also tabulatethe associatedz-statisticswhichtest the nullhypothesisof no effect. Diagnostic statisticsfollowat the bottomofthe table,includingactualandpredictedcrashcross-tabulations,andjoint hypothesistests forthe significanceof debtcomposition,external.macroeconomic,andall effects. Mostof the debt compositionvariablesdo not havestatisticallysignificantcoefficients, thoughsome(likethe confessionalvariable)are closeto significant. The somewhatweak resultsare probablythe resultof multicollinearitybetweenourdifferentdebtcharacteristics.10 Thecoefficientsfor commercialbankandpublicsectorproportionsof debtare inappropriately.thoughinsignificantlysigned. Wealso notethat theproportionof short-termdebt hasan insignificanteffecton crashincidence. Onthe otherhand,the proportionof externaldebt accountedforby FDI is consistentlystronglyand significantlyassociatedwith crashincidence;a fallin FDI inflowsby onepercentof the debtis associatedwithan increasein the probabilityof a crashby .3%. Thedebtcompositionvariableshavea weakbut non-negligible effecton crashincidenceoverall. Interestingly,neitherthe currentaccountnor the budgetdeficithasthe predictedsign, thoughneithereffectis statisticallysignificantat conventionallevels(consistentwith the 10 Wehaveexperimentedwithfactoranalysis,andfoundthatasinglefactoraccountsformuchvariation inthedebtcompositionvariables. 16

graphicalresults). Butthe externaleffectsexerta strongand sensibleinfluenceon the likelihood of crashincidence. Higherdebt,lowerresemes,anda moreover-valuedrealexchange rateall seemto raisethe oddsof crashincidence. Eachof theseeffectshavemarginally significantindividualeffectswhicharejointly significant. The domesticmacroeconomiceffectsare quitestrong. As noted,the fiscalstanceis incomectlysignedandof marginalsignificance. But high domesticcreditgrowthanda recessionbothcoincidewith an increasedprobabilityof a crash. Finally,increasesin Northerninterestratesincreasethe likelihoodof a crashby an amountwhichis bothstatisticallyand economicallysignificant. A onepercentagepoint increasein the foreigninterestrateraisestheprobabilityof a crashby overone percent, holdingall otherinfluencesconstant. ButNorthernreal outputgrowthhas littleeffecton crashlikelihood,onceothereffectshavebeentakeninto account On the right-handsideof TableI, we tabulateanalogousresultsin whichall the regressorsare lagged. Thisamountsto a cmdetest of the abilityof the regressorsto predictcrashes,preciselyone year in advance. Interestingly enough, the results are mostly stronger than those in the contemporaneous regression. Thejoint effects of debt composition, external, and internal effects are now all significant. Low fractions of debt which is either confessional or accounted for by FDI or a high &action which is public-sector, all raise the probability of a fiture crash. Low reserves and over-valuation are also crash predictors, as are high foreign interest rates or high domestic credit growth. 17

Sensitivity Anaiysis Table2 performsa varietyof robustnesschecks. The firstreportsthe resultsof weighted estimation,wheretheweightsare proportionalto realoutputper capita(usingweights proportionalto the actualexchangeratejump doesnot significantlychangeour benchmark results). The secondperfo~s the estimationonlyon Latincountries;thethirdonlyanalyzes post 1982data. Ourreportsare somewhatsensitiveto the exactway inwhichthe datais usedforestimation. Butour most importantresultscomethroughrelativelyclearly. Low FDI flows,highdomesticcreditgrowth,low outputgrowthandhighforeigninterestratesare all associatedwithcurrencycrashes. Currentaccountandbudgetdeficitsremaininsignificant determinantsof crashincidence. Table3 providesmore sensitivi~ analysis. Threeperturbationsof the modelareexamined. The firstreplacesthe foreigninterestwith interactiveeffectsbetweenthe levelof foreigninterestratesandsuchdomesticvariablesas the debtioutputratio,the variable-rate proportionof debt,andthe short-termproportionof debt. Onlythe productof the interest rate withthe debtiGDPratio is statisticallysignificant;itseffectseemssensible. A secondperturbationinvolvesaddinga variableto reflectthe “currencyexposure”of debtorsto fluctuationsin the exchangeratesamongthe dollar,yen,francandothermajor currencies. Wedefinedthe currencyexposurevariablefor a givendebtorto be a weighted averageof the changesin the dollarexchangeratesof the majorcurrencies,wherethe weights werethe sharesof that debtor’sliabilitiesdenominatedin the cwencies in question. Thusa countrywitha healy shareof yen-denominateddebtwouldshowa highvulnerabilityin a yearwhenthe yen appreciatedsharplyagainstthe dollar. The cumencyexposurevariable 18

enters the regression with high statistical significance, but the wrong sign Thisresultis dominated by the yetidollar exchange rate: countries with a lot of debt in the ever-appreciating yen did betterin the samplethan others. EastAsiancountrieshavethe heavyshareof yendebt,and haveprobablydone wellfor otherreasons,so our findingmaybe spurious. A thirdcheckaddscontinentdummyvariables;noneof the importantresultsareaffected. To sumup: our majorresultsappearnotto dependstronglyon the exacteconometric methodologywe employ, v: summarv and Con~ion Muchof the literatureon speculative attacks focuses on a few episodes. In this paper we searchforthe stylizedfactsassociated with currency crashes -- large currency depreciations -- in a broad group of emerging markets. We use annual data fromoveronehundred developing countries and over two decades. Our empirical results stem from a non-structural investigation of the data,andmostly come from a grossly over-parameterized statistical model. Thus we eschewstructuralinterpretations. Nevertheless, we find that currency crashes can be characterizedin whatappears to be a sensible way. Crashes tend to occur when FDI inflows dry up, when reserves are low,whendomesticcreditgrowthis high,whenNortherninterestratesrise,and whenthe real exchangeratehas beenover-valued. Theyalsotendto be associatedwith sharprecessions,thoughthe causallinkagesare veryunclear. Curiously,neithercurrentaccountnor 19

governmentbudgetdeficitsappearto play an importantrole in a typicalcrash. Crasheswere alsopredictableon the basisof thesecharacteristics. Wethinkof this as an encouragingstartingpoint for fiture research. 20

eference~ Calve,Guillermo,Leo Leidermanand CarmenReinhart(1993)“CapitalInflowsand Real ExchangeRateAppreciationin LatinAmerica:TheRoleof ExternalFactors,”IMFSlaff Papers40, no.1,March,pp.108-150. Chuhan,Punam,StijnClaessensandNlanduMamingi(1994)“Equity and BondFiowsto LatinAmericaand Asia:TheRoleof GlobalandCountryFactors,”WorldBankmimeo. Cole,HaroldL. and TimothyJ. Kehoe(1995)“Self-FulfillingDebtCrises”mimeo. Comolly, MichaelB. (1986)“TheSpeculativeAttackon the Peso and the RealExchange Rate,”Journalof InternationalMoneyand Finance5, supplement,I17-130. Dooley,Michael,EduardoFemandez-Arias,andKemeth Kletzer(1994) “RecentPrivate CapitalInflowsto DevelopingCountries:Is the DebtCrisisHistory?”NBER WPNo. 4792. Edwards,Sebastian(1988)ExchangeRate MisalignmentinDevelopingCountries,Johns HopkinsUniversityPress,Baltimore. Edwards,Sebastian(1994)“Realand MonetaryDeterminantsof RealExchangeRateBehavior: Theoryand EvidencefromDevelopingCountries,”Chapter3 in EstimatingEquilibrium ExchangeRa/es,J. Williamson,cd.,InstituteforInternationalEconomics,Washington,DC. Eichengreen,Barry,AndrewRoseand CharlesWypiosz(1995)“ExchangeMarketMayhem: TheAntecedentsand Aftermathof SpeculativeAttacks,”EconomicPoZic}~. 21

Femandez-Arias,Edumdo(1994)“TheNew Waveof PrivateCapitalInflows:Pushor PuI1?” PolicyResearchWorkingPaper 1312,Debtand InternationalFinanceDivision,International EconomicsDepartment,TheWorldBank,June. Frankel,Jefiey, and Shang-JinWei(1994)“YenBlocor DollarBloc?ExchangeRatePolicies of the EastAsianEconomies” in MacroeconomicLinkages:Savings,&change Rates, and CapitalFlows,NBER- EastAsiaSeminaron Economics,Volume3, TakatoshiItoand We Krueger,editors,Universityof ChicagoPress,Chicago. Frankel,Jeffrey,and Shang-JinWei(1995)“Isa YenB1OCEmerging?”Joinl U.S.-Korea AcademicSymposium,atthe Universityof Califomi%Berkeley;in Volume5, Economic Cooperationand Challenges in the PaciJc, editedby RobertRich,KoreaEconomicInstitute of tierica: Washington.D.C.. Krugman.Paul(1979)“AModelof Balanceof PaymentsCrises,”Journal of Money,Crediz and Banking11, 311-325. Krugman,Paul.andJulio Rotemberg(1991)“SpeculativeAttackson TargetZones,”in P. KrugmanandM. Miller.eds. TargefZonesand CurrencyBands,OxfordUniversityPress, Oxford. Obstfeld.Maurice(1994)“TheLogicof Cmency Crises,”CahiersEconomiqueset A40netaires,Banquede France,Paris, 189-213.

Table 1: Probit Estimates Default Predictive 8F(x)16x Iz/ 8F(.x.)/6x Izl Comm’1Bank/Debt -.07 0.57 .03 0.21 A Confessional -.10 1.74 -.14 2.10 r I VariableRate .03 0.21 -.03 0.22 1 1 ShortTerm .04 0.34 .23 1.97 i FDI/Debt -.33 2.88 -.31 2.47 I I PublicSector/Debt .11 1.32 .19 2.18 I Multilateral/Debt -.03 0.46 -.06 0.81 1 Debt/GNP .03 1.33 -.04 1.71 i * Reserves/Imports -.01 1.99 -.01 3.39 A CurrentAccount .10 1.03 ,02 0.22 1 Over-VaIuation .05 1.51 .08 2.53 Gov’tBudget .27 1.90 .16 1.06 DomesticCredit .13 4.78 .10 3.24 I Growth Rate -.38 3.13 -.16 1.29 I NorthernGrowth .55 0.98 -.85 1.50 1 ForeignInterest 1.27 4.50 .80 2.60 SampleSize 803 780 Pseudo-R2 .20 P-Val .17 P-Val , Ho:Slopes=O;x2(16) 93.6 .00 81.2 .00 Ho:DebtEffects=O;x2(7) 14.2 .05 25.5 .00 r Ho:ExternalEffects=O;x2(4) 8.8 .07 16.5 .00 r \ Ho:MacroEffects=O;x2(3) 32.9 .00 12.3 .01 r Ho:ForeignEffects~; x2(2) 21.5 .00 15.4 .00 Probitslopeder]vatlves(xl00 10convertIntopercentages)andassociatedZ-statlst]cs(forhypothesisofno effect). Slopessignificantlyd~fferentfromzeroatthe.05valueinbold.Modelestimatedwitha constant,by maximumlikelihood.PredictiveModellagsallregressorsoneyear. DefaultModel:Goodnessof Fit Tranquility Crash Total PredictedTranquility 727 65 792 PredictedCrash 6 5 11 Total 733 70 803 PredlctlveModel:Goodnessof Flt Tranquility Crash Total I PredictedTranquility 707 64 771 PredictedCrash 4 5 9 , Total 71I 69 780 23

Event: de>25%, de-deC-11>10%. Tranquil Averages Marked. —. — . . — Data from 105 LDCS, 1971-1992. Scales and Data Vary by Panel. aaJ I ~ So~ ao“* 40“ 199 90 I \ ls $ Commercial Bank/Oabt Public/Debt to 1 I 104 I A Short Term/Debt FDI/Debt Debt/GNP ao o -s0 Foreian Intereat Rata Real OECD Growth Reoervas/Monthly Importa Over-valuation Currant Account/GDP Bov’t BudgetOGNP Domsat3c Credit Growth OUtpUt ~rowth p/C Mean plus two standard deviation band; all fi9ures are Percentages- Movements 3 Years Before and After (117) Crashes 26

International Finance Discussion Papers IFDP U* - thor(s) 1996 534 CurrencyCrashesinEmergingMarkets:An Jeffrey A. Frankel EmpiricalTreatment Andrew K. Rose 533 RegionalPatternsintheLawof OnePrice: CharlesEngel TheRolesofGeographyvs.Cumencies JohnH. Rogers 532 AggregateProductivityandtheProductivity SusantoBasu ofAggregates JohnG. Femald 531 ACenturyofTradeElasticitiesforCanada,Japan, JaimeMarquez andtheUnitedStates 530 ModellingInflationinAustralia Gordonde Brouwer Neil R. Ericsson 529 Hyperinflation and Stabilisation: Cagan MarcusMiller Revisited LeiZhang 528 On the Inverseof theCovarianceMatrixin GuyV.G. Stevens PortfolioAnalysis 527 InternationalComparisonsof the Levels of Unit PeterHooper Labor Costs in Manufacturing ElizabethVrankovich 526 Uncertainty, Instrument Choice, and the Uniqueness DaleW. Henderson of Nash Equilibrium: Macroeconomic and Ning S.Zhu Macroeconomic Examples 525 Targeting Inflation in the 1990s: Recent Challenges RichardT. Freeman JonathanL. Willis 524 Economic Development and intergenerational MuratF. Iyigun Economic Mobility 523 Human Capital Accumulation, Fertility and MuratF. Iyigun Growth: A Re-Analysis 522 ExcessReturnsand Riskat the LongEndof the AlIanD. Brunner TreasuryMarket: An EGARCH-M-Approach DavidP. Simon Pleaseaddressrequestsforcopiesto InternationalFinanceDiscussionPapers,Divisionof InternationalFinance,Stop24,Boardof Governorsof the FederalReserveSystem, Washington,DC 2055!. 27

International Finance Discussion Papers IFDP - J995 521 The MoneyTransmissionMechanismin Mexico MartinaCopelman AlejandroM. Werner 520 When is MonetarypolicyEffective? John Ammer AlIanD. Brunner 519 CentralBankIndependence,Inflationand PrakashLoungani Growth in TransitionEconomies Nathan Sheets 518 AlternativeApproachesto RealExchangeRates Hali J. Edison and RealInterestRates: ThreeUp andThreeDown WilliamR. Melick 517 Productmarketcompetitionand the impactof VivekGhosal priceuncertaintyon investment: someevidence PrakashLoungani fromU.S. manufacturingindustries 516 BlockDistributedMethodsfor Solving Jon Faust Multi-countryEconometricModels RalphTryon 515 Supply-sidesourcesof inflation: evidence PrakashLoungani fromOECD countries PhillipSwagel 514 CapitalFlightfromthe CountriesinTransition: Nathan Sheets SomeTheoryand EmpiricalEvidence 513 Bank Lendingand EconomicActivityinJapan: AlIanD. Brunner Did “FinancialFactors”Contributeto theRecent StevenB. Kamin Downturn? 512 EvidenceonNominalWageRigidityFroma Panel VivekGhosal of U.S. ManufacturingIndustries PrakashLoungani 511 Do TaxesMatterfor Long-RunGrowth?: Harberger’s EnriqueG. Mendoza SupemeutralityConjecture Gian MariaMilesi-Ferretti PatrickAsea 510 Options,Sunspots,andtheCreationof Uncefiainty DavidBowman Jon Faust 509 Hysteresisina SimpleModelof CurrencySubstitution MartinUribe 508 ImportPricesandthe CompetingGoodsEffect PhillipSwagel 507 Supply-sideEconomicsin a GlobalEconomy EnriqueG. Mendoza LindaL.Tesar 28