Can Forecast Errors Predict Financial Crises? Exploring the Properties of a New Multivariate Credit Gap

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Can Forecast Errors Predict Financial Crises? Exploring the Properties of a New Multivariate Credit Gap Elena Afanasyeva 2020-045 Please cite this paper as: Afanasyeva, Elena (2020). “Can Forecast Errors Predict Financial Crises? Exploring the Properties of a New Multivariate Credit Gap,” Finance and Economics Discussion Series 2020-045. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2020.045. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Can Forecast Errors Predict Financial Crises? Exploring the Properties of a New (cid:73) Multivariate Credit Gap ThisVersion:May15,2020 ElenaAfanasyeva∗ 20thStreetandConstitutionAvenueNW,Washington,D.C.,20551,USA Emailaddress:elena.afanasyeva@frb.gov Abstract Yes,theycan.Iproposeanewmethodtodetectcreditboomsandbustsfrommultivariatesystems–monetaryBayesian vectorautoregressions. Whenobservedcreditissystematicallyhigherthancreditforecastsjustifiedbyrealeconomic activity variables, a positive credit gap emerges. The methodology is tested for 31 advanced and emerging market economies. Theresultingcreditgapsfithistoricalevidencewellanddetectturningpointsearlier, outperformingthe credit-to-GDPgapsinsignalingfinancialcrises,especiallyatlongerhorizons. Theresultssurviveinrealtimeandcan shedlightonthedriversofcreditbooms. Keywords: CreditBoom,CreditGap,BayesianVAR,ConditionalForecasts,EarlyWarning,FinancialCrises JELclassification: C11,C13,C53,E51,E58 1. Introduction Robust and adequate identification of credit cycles is a challenging and important task. Unsustainable credit expansionscanendupinadverseeconomicoutcomes. SchularickandTaylor(2012)amongothersdemonstratethat systemicfinancialcrisesareessentiallycreditboomsgonebust.Evenwhencreditboomsdonotendinfinancialcrises, theyarestillcostly. Jordaetal.(2013)showthatafteracreditboom,recessionsaremorepainful,evenifafinancialor bankingcrisisdoesnotoccur.Creditboomsleavelargesectorsoftheeconomyoverleveraged,whichimpairsfinancial intermediationduringthesubsequentrecovery. Thereforecreditboomscanbegoodpredictorsofcreditlessrecoveries (Dell’Aricciaetal.,2012). Giventhelinkbetweencreditboomsandfinancialcrises,measuresofexcessivecreditare (cid:73)IthankTobiasAdrian, PooyanAmirAhmadi, AlexBick, MatthiasBurgert, SteffenElstner(discussant), MatteoFalagiarda(discussant), DomenicoGiannone,LucaGuerrieri,LarsHansen,GaziKara,DimitrisKorobilis,MicheleLenza,OscarJorda,AkiraKohsaka,EstebanPrieto (discussant),SigridRoehrs,theparticipantsofnumerousconferencesandseminars.SamJerowprovidedexcellentresearchassistance.Allremainingerrorsaremine. ∗ElenaAfanasyevaisEconomistintheDivisionofFinancialStabilityoftheBoardofGovernorsoftheFederalReserveSystem. Disclaimer: ThematerialheredoesnotrepresenttheviewsoftheBoardofGovernorsoftheFederalReserveSystem.

usedasguidelinesfortighteningthemacroprudentialtools.1 Accordingly,themeasuresofexcessivecreditneedtobe adequateandrobust. IproposeanewmethodologytodetectcreditboomsandbustsfromBayesianvectorautoregression(VAR)forecast errors. Tothatend,Iemployadifferentoperationaldefinitionofcreditbooms,viewingthemasdeparturesfromthe fundamentally-justified levels rather than deviations from a univariate trend, such as Hodrick-Prescott (HP) trend. Whentheactuallevelofcreditissystematicallyhigher(lower)thanthelevelofcreditjustifiedbythefundamentals, this indicates a credit boom (bust) in the economy. Similarly to the intuition behind the credit-to-GDP ratios, I use realeconomicactivityvariablesasthefundamentalsforcreditandoperationalizethisideabyconstructing(pseudo) out-of-sampleforecastsforcredit, whichareconditionedonthepathofrealeconomicactivityfouryearsahead. To perform these forecasting exercises, I estimate medium-scale monetary Bayesian vector autoregressions (BVARs), therebyemployingafullymultivariateapproach. Thenewmethodisappliedto31advancedandemergingmarketeconomiesandtheresultsarecomparedwiththose ofcredit-to-GDPgaps–anestablishedandwell-performingbenchmarktestedonlargepanelsofcountries(Mendoza andTerrones(2008), Gourinchasetal.(2001), andDrehmannetal.(2010)). TheresultingBVAR-basedcreditgaps areconsistentwithhistoricalevidenceacrossthepanelofcountries. Elevatedcreditgapsprecedecrises,turnnegative atorshortlyafteracrisisonset, andstayinthenegativeterritoryforthepost-crisisbustperiod, inlinewithvarious crisis chronologies and economic history studies. The BVAR-based credit gaps robustly outperform credit-to-GDP gaps as crisis predictors, especially at longer horizons of 3- to 4 years ahead. The reason is that BVAR-based gaps detect turning points earlier. First, in credit-to-GDP ratios, the link with the real activity is by construction largely static, confinedtotheratiowithinoneperiod, whichaltersthedynamicsofthecredit-to-GDPtimeseriesandshifts theidentifiedpeakofthefinancialcycleayearortwoforward. Bycontrast,theVARmachineryandtheconditioning on the path of real activity when forming credit forecasts retain a dynamic link between these variables, taking the lead-lagrelationshipsintoaccount. Second, becauseoftimevariationinparameters, theimpliedtrendintheBVAR methodologyismuchmoreflexiblethananearlylinearHodrick-Prescotttrendofcredit-to-GDPratios.Hence,turning pointsleadingtobothboomsandbustsareoftenidentifiedfasterwithaBVAR,andthisfeaturesurvivesinrealtime. Thispaperprimarilyspeakstotheliteraturethatidentifiesandstudiescreditcycles. Theprevalentfilteringmethod –univariatedetrendingofcredit-to-GDPratioswiththeHPfilter–quitesuccessfullypointstosystemiccrisesinlarge panels butalso hasdrawbacks. Forinstance, Hamilton(2018) discusseshow thefilter itself caninduce inconsistent dynamics. Edge and Meisenzahl (2011) show how end-point characteristics of the HP filter impair the reliability ofgapestimatesinreal-time,whenitcomestothedetrendingofcredit-to-GDPratios. Inmyapproach,Iproposean 1Infact,theMacroVariablesTaskForceoftheBaselCommitteehasproposedmeasuresofexcessivecreditgrowth(inparticular,deviationsof thecredit-to-GDPratiofromitstrend)asareferencepointtodeterminetheneedtotightenbankcapitalrequirements. 2

alternativewayofmultivariatefiltering,departingfromtheHPfilter(similarlytoHamiltonsproposal)andyetretaining the useful linkage of credit with economic fundamentals, when identifying credit cycles (similar to the underlying idea of credit-to-GDP ratio detrending). The combination of these features improves early warning properties and eliminatesspuriouscreditcyclepeaksaftercrises,alsoinreal-timeconditions.2 Incontrasttounivariatedetrendingmethods,theBVAR-basedapproachinadditionallowstoshedsomelighton thenatureofcreditbooms. Forinstance,severalstudiesemphasizetheroleofshort-termratesinthebuild-upofcredit booms. Inparticular,“toolowfortoolong”rates(Taylor,2007)maystimulaterisk-takingbehavior(BorioandZhu, 2012;AdrianandShin,2010). IshowthatthecreditboomintheU.S.andtheeuroareacountriesintheyearsbefore totheGreatRecessioncanbe,toalargeextent(thoughnotcompletely),explainedbythepathofthemonetarypolicy rates. Finally,thispaperalsocontributestotheongoingdebateof,whichfinancialindicatorsaremoreinformativeabout thebuildupofsystemicfinancialvulnerabilitiesaheadofcrises. Thereisliteratureonthelinkofassetpricesswings tocrises(BorioandLowe(2002),Shiller(2008),andreferencestherein)aswellasontheroleofbroadmoneygrowth incomparisonwithcreditgrowth(SchularickandTaylor,2012). Icomputethegapsforassetpricesandbroadmoney usingtheBVAR-basedmethodology. Runninganearly-warninghorseraceexercisebetweenthesegapsandBVARbasedcreditgaps,Ifindthatcreditgapstendtodominateasearlywarningindicatorsatallhorizons. Thereasonisthat broadmoneygaps-whenelevated-oftentendtoreflectmonetarypolicyactions,suchasquantitativeeasingpolicies amongothers,whichleadstosizeableandprolongedpositivegapsafterthecrisisonset. Forassetpricegaps,therate offalsepositivesissubstantiallyhigherthanthatofBVAR-basedcreditgapsbecauseofmuchhighervolatilityofthe former. Therestofthepaperisorganizedasfollows. Section2presentstheapproachtocreditboomidentificationinmore detail, describesthedataandjustifiesthechoiceoftheeconometricmethodology. Section3discussesthehistorical plausibilityofresultsandrunsahorse-raceexerciseofBVAR-basedcreditgapsagainstcredit-to-GDPgapstocompare theircrisispredictionqualities. Section4illustratestheperformanceofthegapsinrealtime,whileSection5studies the role of short-term interest rates during credit booms. Section 6 discusses the properties of asset price gaps and moneygaps. Section7illustratestherobustnessofthemethodologytothebatteryofextensiverobustnesschecks,and Section8providesconcludingremarksandoutlinesdirectionsforfurtherwork. 2. Approach,Methodology,andData Creditboomsareusuallydefinedasepisodesofparticularlyrapidgrowthofcredittotheprivatesectorrelativeto thegrowthoftherealeconomy.Toidentifysuchepisodes,actualdevelopmentsinthecredit-to-GDPratioaretypically 2TherearemoredeparturesfromHPfilteringintheliterature,whenitcomestomeasuringexcesscredit. Jordaetal.(2013)resorttogrowth rates,forexample. 3

comparedwith thelong-runsmooth trendcomputed withtheHP filter, forinstance. In contrastwiththe detrending approach, Iapplyadifferentoperationaldefinitionofcreditbooms, viewingthemasdeparturesfromfundamentally justified levels. Similarly to the credit-to-GDP gap, I employ real economic activity variables as fundamentals for credit. When actual credit growth is persistently higher than the credit growth supported by the fundamentals, the economyisinacreditboom(andviceversaforacreditbust). Theintuitionisstraightforward: Whentheeconomyis notgeneratingenoughincomerelativetotherapidgrowthofdebt,thosedebtseventuallybecomehardertorepayor lesssustainable.3 The second feature of my approach is that it is multivariate and accounts for endogenous interactions between credit and the other variables in the economy that would be relevant for the credit cycle. Given that credit booms are general equilibrium phenomena, it appears desirable to detect them from multivariate systems rather than single timeseries. ThispointisalsostressedbyBorioandLowe(2002): “...itisthecombinationofeventsthatmattersfor detectingproblemsinfinancialstability: Itisnotjustcreditgrowth,oranassetpriceboom,theinteractionsbetween credit, asset prices, and real economy should not be ignored...” A further advantage of a multivariate approach is in overcoming the endogeneity and simultaneity biases, which might distort the results of univariate regressions. ThereforeIuseamonetaryVARasacreditboomdetectiontool. Thisapproachallowsforlinkingthevaluesofcredit to the fundamentals without resorting to direct normalization of credit by real activity measures, such as credit-to- GDPratios. Inparticular,basedontheVARestimation,Iconstructpseudoout-of-sampleforecastsforcredit,which are conditioned on real activity. These forecasts represent the fundamentally justified levels of credit and serve as a comparisonbenchmarkfortheactualcreditvalues. Inwhatfollows,Idiscusstheoperationalizationofthisapproach inmoredetail. On methodological grounds, the benchmark model is a medium-sized monetary VAR, which is estimated at a monthlyfrequencyforthesetofadvancesandemergingmarketeconomies. Thevariablesinthesystemaretypically usedinmonetaryVARs(Giannoneetal.(2019),andBanburaetal.(2010))andcontain: realeconomicactivity(representedbyindustrialproduction, unemploymentrate, orotherbusinesscycleindicatorsdependingonthecountry), broadconsumerpriceindexes,monetarypolicyrates(orothershort-termratesandspreads,dependingonthecountry), asset prices (broad stock market indexes), and monetary variables (credit and narrow and broad money aggregates). Forsmallopeneconomiesinthesample,thesetofvariablestypicallycontainseithernominalexchangeratesrelative totheUSDorthebroadrealexchangerate. TheprimarysourceforthecreditvariableistheInternationalFinancial Statistics(IFS)database,whichprovidesdataonbankclaimsondomesticprivatesector(firmsandhouseholds)ata 3ThefinancialinstabilityhypothesisofMinsky(1986)alsoprovidesmotivationforrealactivityvariablesasthechoiceoffundamentalsfor credit. AccordingtoMinsky,thereisaremarkableasymmetryintheloancreationprocess: Whileloansaregrantedbasedonexpectedprofits (orexpectedrealeconomicactivity), loansareeventuallypaidoutoftherealizedprofits(orrealizedrealeconomicactivity). Oncetheprofit expectationsareelevated,theeconomygoesintoafinanciallyunstable,fragilestatesincetherealizedprofitsarenotsufficienttopayoutthedebts. Asaresult,anunsustainablecreditboomoccurs. 4

monthly frequency (IFS line 32d). In many cases, however, these data are subject to large structural breaks due to definition changes following the transition of the country to new accounting standards, for example. In those cases, thebroadmeasuresofcredittothenonfinancialsectorprovidedbytheBankforInternationalSettlements(BIS)are usedinstead. Thefullsetoftimeseries,theirtransformations,anddatasourcesaregiveninAppendixA.Overall,all thedataseriesarechosentoreflectthesameeconomicconceptsacrosscountriesascloseaspossibleandtoensurea samplethatislongenoughtorenderVARestimationandforecastingexercisesfeasible. AseparateVARisestimated foreachcountry. SinceIbasethedetectionofcreditbooms(busts)onpseudoout-of-sampleconditionalforecasts,theforecasting performanceoftheVARisacrucialcomponentofthedetectiontechnology. Thetradeoffhereisasfollows. Ifthe forecastingperformanceisgenerallyverypoor,alotofdeviationsfromnormalitywouldbedetectedandonlysome portion of them would reflect credit booms or busts, the rest being just poor forecasting. At the same time, if the forecastingperformanceofthesystemisreasonablygood, theideaisthataVARwouldcapturetypicalinteractions betweenthevariablesonacertainsampleandthenprojectthemintotheconditionalforecast.Inthiscase,thedeviation of the forecast from the observed variable would also reflect a deviation from this typical variable co-movement capturedbytheVAR–acreditboomoracreditbust. Inpractice,thedeviationofconditionalforecastfromobserved levels could generally reflect both sides of this trade off: inability to forecast and deviation from a typical behavior. Inordertomitigatetheproblemsassociatedwithpoorforecastingperformance,theestimationapproachisBayesian rather than classical (“frequentist”). Bayesian methodology contains a useful tool – prior shrinkage – to deal with overfitting in estimation of densely parameterized systems, such as VARs. The use of priors makes the forecasting performance of Bayesian VARs (BVARs) more efficient. In particular, I estimate the VAR in log-levels rather than differencesinordertoavoidlosinginformationcontainedinlevelsofvariables. AsnotedbyGiannoneetal.(2019), theassessmentoflevel-relationshipsisparticularlyimportantinmonetaryanalysis.4 The benchmark model is a linear BVAR. The system is estimated with the prior distribution of Sims and Zha (1998),whichisimposedonastructuralVARoftheform: (cid:88)p y A =d+(cid:15),t=1,...,T, (1) t−l l t l=0 whereT isthesamplesize,y isthevectorofobservations,A isthecoefficientmatrixofthelthlag, pisthemaximum t l lag,disavectorofconstants,and(cid:15) isavectorofi.i.d. structuralGaussianshockswith: t (cid:16) (cid:17) E (cid:15)T(cid:15)|y ,s>0 = I t t t−s E((cid:15)|y ,s>0)=0,∀t. t t−s 4AnalternativeapproachwouldbetoestimateaBVARindifferencesasproposedbyVillani(2009). However,theforecastingpropertiesof suchmodelsoftendependonassumptionsaboutthesteadystateoftheVAR(JarocinskiandSmets,2008). 5

Although I estimate an identified VAR under the Cholesky identification, the ordering of variables does not affect the distribution of conditional forecasts and therefore is irrelevant for what follows. A formal proof of this result is presented in Waggoner and Zha (1999) (see Proposition 1). The maximum lag order is set to 13 months. This lag structureofapproximatelyoneyearisstandardintheliteratureformonthlydata(Giannoneetal.(2019),andBanbura etal.(2010)). Therateofdecayforthelagorderweightsisdeterminedbytheprior. Asalreadynotedabove,VARsaredenselyparameterizedsystemsthataregenerallypronetooverfitting–thatis, goodin-samplefitbutpoorout-of-sampleforecastingperformance. InBayesianestimation,thepriordeterminesthe degreeofshrinkage. Thereforethechoiceofpriorhyperparametersisquiteimportantfortheforecastingperformance ofthemodel. Thetradeoffhereisasfollows. Whentheprioristooloose(thatis,veryuninformative),themodelgeneratesdispersedforecastsbecauseofhighestimationuncertainty.Whentheprioristootight,theestimatedcoefficients willbeveryclosetothevaluesdeterminedbytheprior,whichislikelytoleadtopoorforecastsaswell. Thereforethe goalistochoosethe“right”amountofshrinkagewhenonesetsthehyperparametersoftheprior(Giannone,Lenza, andPrimiceri(2015)andCanova(2006)formorediscussionofthisargument). HereIfollowoneoftheapproaches intheBVARliteratureandobtainthevaluesofpriorhyperparametersbymaximizingthemarginallikelihoodoverthe trainingsample.5 Intheforecastingexercise,forecastdensitiesaresimulatedwiththeGibbssamplerofWaggonerandZha(1999) imposinghardconditions–thatis, theforecastsareconditionedontheexactpathoftherealactivityvariablerather thanitspossibleranges.6 TheVARisestimatedinarollingwindowapproach,wherethesizeoftherollingwindow isbetween9and15years.7 Therollingwindowallowsonetostudytheatypicaldeviationsthroughoutthesampleas wellastocapturetimevariationinparametersinanefficientway.8 Timevariationinparameters, inturn, hasdirect implicationsfortrendflexibilityandinteraliaallowsonetobettercapturephenomenalikefinancialdeepeningwhen comparedwithconstant-parametermodels. The forecasting exercise for each rolling window proceeds as follows. After estimating the BVAR up to t −4 years within the sample, I construct out-of-sample forecasts of credit that are conditional on a path of future values of the real economic activity variables. To be precise, these forecasts should be considered pseudo out-of-sample, sinceIconditionontheknownfuturevaluesofthebusinesscyclevariable. Ifurthercomparetheseforecastswiththe observedvaluesofcredit. Suppose,forinstance,theactual(observed)valueofcreditissubstantiallyhigherthanits conditional forecast. Conditional on the model specification, this deviation indicates that there is more credit in the 5ThemarginallikelihoodiscomputedasinChib(1995),andoptimizationisperformedviaagridsearch.Forcountries,wheredatasamplesare shorter,IadoptthevaluesfromtheoriginalpaperofSimsandZha(1998)andconductasensitivityanalysiswithrespecttothehyperparameters. 6Foreachrollingwindow60000drawsareproduced,thefirst15000ofthemarediscardedasburn-in. 7Thechoiceoftherollingwindowsizeeffectivelydeterminestheflexibilityorsmoothnessofthe‘trend’andisdeterminedbasedondata availabilityforaparticularcountryratherthanoptimized. Optimizingthelengthoftherollingwindowforeachcountry(similartochoosingan optimalsmoothingparameterfortheHPfilter)mayimproveearlywarningpropertiesofBVAR-basedgapsfurther. 8Bauwensetal.(2015)showthatrolling-windowmodelsoftenforecastbetterthanfull-fledgedtime-varyingparametermodels. 6

economythanjustifiedbythefundamentalvariable–apossibleindicationofacreditboom.Moreformally,Icompute thefollowingdeviation,oraforecasterror: dev =yact−ym, t,h t,h t,h where index t refers to a particular point in time and index h refers to the respective estimation window (h = 1 corresponds to the most recent forecast for this point in time t, h = H corresponds to the forecast from the earliest estimation window for this point in time), and ym is the point forecast.9 Note that the deviation defined above is t,h sign-preserving: itispositiveforpotentialboomepisodesandnegativeforpotentialbustepisodes.10 Thebaselineforecastinghorizonissettofouryears,andtheestimationproceedsinrollingwindows. Accordingly, ateachpointintimethereareseveralforecastsandthereforeseveraldeviationsavailablefromdifferentrollingwindows. The most recent rolling window would yield a short-term forecast and a respective deviation, while the most distant rolling window would yield a longer horizon forecast and a respective deviation for this particular point in time. Itwouldbehardtomotivatethechoiceofoneparticularforecasthorizonthatwouldberelevanttodetectcredit booms. Furthermore, choosing only one forecast horizon, for instance, the one with the smallest prediction error, could potentially bias the results. Therefore I use the forecasts and the deviations from all forecasting horizons and pool(average)themasfollows: (cid:80)H dev t,h gap = h=1 , (2) t H whereH-isthemaximumforecasthorizon,histheindexfortherespectiveestimationwindow. Inthenextsection,Iapplythemethodologytoasetof31advancedandemergingmarketeconomiesinorderto evaluatethequalityoftheresultingcreditgapmeasure. 3. HistoricalPlausibilityandCrisisPrediction Creditgapsarelatentvariables. Therearenoactualdatatocomparethemagainstinordertoassesstheirquality. However, a comparison against historical evidence is possible and useful. Similarly to the output gaps, one could expectacreditgaptobuildupandstaypositiveduringaboomphase,forexample,precedingacrisis,andthentoturn atcrisisonsetandstaynegativeduringacreditbustorcreditcrunch,forexample,followingacrisis. Figures1–through4showtheBVARcreditgapsfor31advancedandemergingmarketeconomiesplotted along with the vertical bars that signify the onset of systemic banking crises, financial crises or events relevant for macroprudential policy. The timing of these events is taken from the following chronologies: Laeven and Valencia 9Iusethemeanasthepointestimateinthebaselinespecifications.Medianpointestimatedeliverssimilarresults. 10Anearlierversionofthiscreditgapconsidersdeviationsbeingsignificantoncetheysurpassathreshold-theupward(downward)bandaround thepointforecastfortheboom(bust)(seeAfanasyeva(2013)formoredetails).Thisfeatureoftenhelpsreducethenoiseintheresultingcreditgap seriesbyremovingsmaller,lessprofoundepisodes. Atthesametime,however,imposingthresholdscanworsenearlywarningqualityforsome countries,assmallerupwarddeviationsattheonsetofthecreditboomarezeroedout. 7

(2013)(LVhenceforth),SchularickandTaylor(2012)(SThenceforth),Baronetal. (2018)(BVXhenceforth),andLo Ducaetal. (2017)(LDhenceforth).11 As figures 1 – through 4 illustrate, the resulting BVAR-based credit gaps are largely consistent with historical evidencefortherespectivecountries. Majorcrises,suchastheGlobalFinancialCrisisintheU.S.andEurope(Aliber and Kindelberger (2015)), the Heisei bubble burst of the late 1980s in Japan (Shiratsuka (2005)), the Savings and LoanCrisisintheU.S.(Elliott,Feldberg,andLehnert(2013)),theAsianfinancialcrisisof1998inSouthKoreaand Malaysia (Aliber and Kindelberger (2015)), were preceded by credit expansions and overheating, and it is indeed reflected in persistent and sizeable positive values of the corresponding gap measures. At the crisis onset date or shortlythereafter,thecreditgapstendtodeclineprecipitously,markingthepost-crisiscreditbustperiod. Thesebusts andcreditcrunchesarealsowelldocumentedintheliterature. Togiveoneillustrativeexample,intheU.S.,theBVAR credit gap adequately reflects credit crunches, such as the one in the aftermath of the 1974 banking crisis, the S&L crisis and the GLobal Financial Crisis (Bordo and Haubrich (2010)) as well as the prolonged credit crunch of the early1990s(BernankeandLown(1991)). Interestingly,theBVARcreditgapsalsoreflectsomewhatlessprominent episodes in the financial cycle – episodes not uniformly identified as crises preceded by major credit booms by the chronologies. Forinstance, inthecaseofJapan, aminorshort-livedboomepisodein2001emerges aperiodwhen Bank of Japans QE policies were announced and had a profound effect on bank lending (Bowman et al. (2015)). The reason is that, in contrast to filters with a built-in length of a cycle (such as HP filter or bandpass filter), this approachdoesnotaprioriexcludecyclesofacertainlength. Hence,theselessprominentepisodesarenotnecessarily interpretableasnoise,butreflecttheabilityoftheBVARgaptocapturebothlargerandsmallerwavesofthefinancial cycle. Inordertoquantifythesequalitativeobservationsandinordertoputtheresultsincontext,Iconductahorse-race exercisetocompareearlywarningpropertiesoftheBVARcreditgapwiththeestablished,well-performingbenchmark onlargepanelsofcountries–thecredit-to-GDPgap(seeDrehmannetal.(2010),forexample). Thelatterversionof thecreditgapisobtainedbydetrendingthelogarithmofthecredit-to-GDPratiowiththeHPfilterthatextractslowerfrequencymovements(byapplyingthesmoothingparameterof400.000). IcompareBVARgapswithbothone-and two-sidedversionsofthecredit-to-GDPgaps. Whiletheinformationsetoftheone-sidedversioniscomparablewith theoneoftheBVARgap(alsoaone-sidedmeasure),two-sidedcredit-to-GDPgapsareoftenusedandexamined,as theyencompassmoreinformation. To visualize the cyclical differences between the measures, I first construct the Burns-Mitchell diagrams for the BVAR gaps and credit-to-GDP gaps around systemic crises (figures 5 and 6).12 To plot a Burns-Mitchell diagram, 11Ifocusonfinancialcrisesandexcludesovereigndebtcrises,wheretheyfallintothespanofthesample,astheunderlyingcreditseriesdonot containgovernmentdebtandarehencenotcapabletoreflecttherelevantsovereigndebtdynamicsbyconstruction. 12Afullsetofcountry-specificgraphscomparingBVARcreditgapswithtwo-sidedcredit-to-GDPgapsispresentedinAppendixB. 8

muigleB)c( airtsuA)b( ailartsuA)a( elihC)f( adanaC)e( lizarB)d( ecnarF)i( dnalniF)h( cilbupeRhcezC)g( ecnarF-ailartsuA:seirtnuoCssorcaspaGPDG-ot-tiderCdediS-enOdnaRAVB:1erugiF :seigolonorhcgniwollofehtotgnidroccasesircfotesnoehtyfingissenillacitreV.stniopegatnecrepnidesserpxeeraspagtiderC.sretrauqnisisixalatnozirohehtnoemiT:setoN tnevelaudiserasetacidni∗,ygolonorhcDLfoesacnI.)7102(.lateacuDoL-DL,)3102(aicnelaVdnaneveaL-VL,)9102(.latenoraB-XVB,)2102(rolyaTdnakciraluhcS-TS .)6102(lebanhcSdnareiemrennurBrofsdnatsSBdna)5102(regrebeldniKdnarebilArofsdnatsZA,ailartsuAroF.)sisirccimetsysanahtrehtar( 9

dnalerI)c( gnoKgnoH)b( ynamreG)a( aeroKhtuoS)f( napaJ)e( ylatI)d( ocixeM)i( aisyalaM)h( gruobmexuL)g( ocixeM-ynamreG:seirtnuoCssorcaspaGPDG-ot-tiderCdediS-enOdnaRAVB:2erugiF :seigolonorhcgniwollofehtotgnidroccasesircfotesnoehtyfingissenillacitreV.stniopegatnecrepnidesserpxeeraspagtiderC.sretrauqnisisixalatnozirohehtnoemiT:setoN tnevelaudiserasetacidni∗,ygolonorhcDLfoesacnI.)7102(.lateacuDoL-DL,)3102(aicnelaVdnaneveaL-VL,)9102(.latenoraB-XVB,)2102(rolyaTdnakciraluhcS-TS .)sisirccimetsysanahtrehtar( 10

dnaloP)c( yawroN)b( dnalaeZweN)a( cilbupeRkavolS)f( eropagniS)e( lagutroP)d( nedewS)i( niapS)h( acirfAhtuoS)g( nedewS-dnalaeZweN:seirtnuoCssorcaspaGPDG-ot-tiderCdediS-enOdnaRAVB:3erugiF :seigolonorhcgniwollofehtotgnidroccasesircfotesnoehtyfingissenillacitreV.stniopegatnecrepnidesserpxeeraspagtiderC.sretrauqnisisixalatnozirohehtnoemiT:setoN tnevelaudiserasetacidni∗,ygolonorhcDLfoesacnI.)7102(.lateacuDoL-DL,)3102(aicnelaVdnaneveaL-VL,)9102(.latenoraB-XVB,)2102(rolyaTdnakciraluhcS-TS .)sisirccimetsysanahtrehtar( 11

yekruT)b( dnalreztiwS)a( aciremAfosetatSdetinU)d( modgniKdetinU)c( ASU-dnalreztiwS:seirtnuoCssorcaspaGPDG-ot-tiderCdediS-enOdnaRAVB:4erugiF :seigolonorhcgniwollofehtotgnidroccasesircfotesnoehtyfingissenillacitreV.stniopegatnecrepnidesserpxeeraspagtiderC.sretrauqnisisixalatnozirohehtnoemiT:setoN tnevelaudiserasetacidni∗,ygolonorhcDLfoesacnI.)7102(.lateacuDoL-DL,)3102(aicnelaVdnaneveaL-VL,)9102(.latenoraB-XVB,)2102(rolyaTdnakciraluhcS-TS .)sisirccimetsysanahtrehtar( 12

(a)LVChronology (b)BVXChronology (c)STChronology (d)LDChronology Figure5:AverageBurns-MitchellDiagramsAroundCrises:BVARCreditGapandOne-SidedCredit-to-GDPGap. Notes:Timeonthehorizontalaxesisinquarters.Creditgapsareexpressedinfractionsoftheirrespectivepeakvaluesduringthe creditcyclesurroundingthecrisis.Verticallinessignifytheonsetofcrisesaccordingtothefollowingchronologies:ST- SchularickandTaylor(2012),BVX-Baronetal.(2019),LV-LaevenandValencia(2013),LD-LoDucaetal.(2017). I normalize the time series by its peak value around the critical event as timed by the crisis chronology. I employ allfourchronologiesinordertousemaximalcross-countryinformation(alargernumberofcrisisepisodes)andfor robustness,ascrisistimingscanvaryconsiderably. Forinstance,theonsetofthesavingsandloancrisisisdatedwith adifferenceoffouryearsinvariouschronologies,andtherearemoreanalogousexamplesinthesampleofcountries. Furthermore,thereisconsiderabledisagreementamongchronologiesastowhetherornotacrisisevenoccurred(see AppendixD).Thereasonsaredifferentqualifyingcriteriaacrosschronologies.Typically,thechoiceisbasedonasetof qualitative(narrative,subjective)and/orquantitativeindicators.Baronetal.(2018)pointoutthatchronologiesplacing considerableweightonsubjectiveinformationmaysufferfromlook-backbias. IthereforealsoincludetheBVXcrisis chronology,whichprimarilyreliesonbankequityreturnsdatainidentifyingtheonsetofacrisis. Inaddition,theLD chronologydiffersfromtheotherthree(LV,STandBVX),asitprovidesdatesforsystemiccrisesaswellasresidual events – episodes that do not meet the criteria of systemic crises and yet are relevant for macroprudential or other regulatorypolicies. TheBurns-Mitchelldiagramsinfigures5and6displaytypical,averagedcreditcyclesacrossthesetofcountries thathavecrisisobservationsaccordingtothecorrespondingchronology.13 13Intheseexercises,Iexcludecrisisepisodes,whicharenotprecededbyatleastfouryearsofcreditgapestimatesinordertohaveanequal numberofcreditgapobservationsacrossepisodesattherelevanthorizons:zerotofouryearsahead.Thisselectiongivesafairchancetopredictive 13

(a)LVChronology (b)BVXChronology (c)STChronology (d)LDChronology Figure6:AverageBurns-MitchellDiagramsAroundCrises:BVARCreditGapandTwo-SidedCredit-to-GDPGap. Notes:Timeonthehorizontalaxesisinquarters.Creditgapsareexpressedinfractionsoftheirrespectivepeakvaluesduringthe creditcyclesurroundingthecrisis.Verticallinessignifytheonsetofcrisesaccordingtothefollowingchronologies:ST- SchularickandTaylor(2012),BVX-Baronetal.(2019),LV-LaevenandValencia(2013),LD-LoDucaetal.(2017). Thediagramsillustratesubstantialdifferencesacrossthemethodologies. First,BVARcreditgapsarepersistently positive four years prior to the crisis event for most chronologies. One-sided credit-to-GDP gaps become positive abouttwoyearsbeforethecrisis,whereastwo-sidedcredit-to-GDPgapsmostlystayinthenegativeterritoryandturn positive a year before the crisis. These findings suggest that BVAR credit gaps detect booms earlier and can have goodearlywarningqualities. Second,therearesubstantialdifferencesbetweenBVARgapsandcredit-to-GDPgaps inthepost-crisisquartersaswell. WhileBVARgapsturnnegativeanddeclineprecipitouslyattheonsetofthecrisis orshortlythereafter, credit-to-GDPgapshaveadelayedpeakingfeatureandtypicallyturnnegativeabouttwoyears aftertheonsetofthecrisisforone-sidedcredit-to-GDPgapsandsubstantiallylaterfortwo-sidedcredit-to-GDPgaps. ThisfeatureagainstressesabetterabilityoftheBVARgaptodetectaturningpoint,nowfrompositiveintonegative territory. The policy implications of these differences can be quite profound. Should macroprudential policy react tosuchafalsepositivesignalbyeitherraisingthecapitalrequirementsornotreleasingthemintime, theactioncan haveprofoundnegativeeffectsonbothlendingandtherealeconomy(seeEdgeandMeisenzahl(2011)fornumerical examples). Finally, the BVAR gaps rebound back into positive territory faster than credit-to-GDP gaps that tend to staynegativethereformanyyearsafteracrisis(seeAustraliainfigure1a,asanexample). abilitiesofcreditgapsacrossthesehorizonsforapanelofcountries.ThecorrespondinglistofcountriesandepisodesispresentedinAppendixD. 14

ThereareprimarilytworeasonswhyBVARgapsdetectturningpointsearlierandaresubstantiallylessproneto delayed peaking. The first reason is the link between credit and the real economic activity fundamental: it is rather staticforcredit-to-GDPgapsandmuchmoredynamicforBVARgaps. Credit-to-GDPgapsarebasedonaratio–a timeseriesthatgivesrisetoitsowndynamicswithpeaksinthepost-crisisquarters. ThereasonisthatGDPtendsto fallfasterthancreditduringcrises(KrishnamurthyandLi(2020)illustratethisstylizedfactacross41financialcrises), whichraisestheratiointhoseyearssubstantially,therebyshiftingtheimpliedpeakofthefinancialcycleseveralyears intothefuture. Bycontrast, theBVARcreditgaplinkscredittorealactivityvariablesviaconditionalforecastsina multivariate system, accounting for lead-lag relationships between credit and real activity. The second reason is the degree of trend flexibility. Due to time variation in parameters from rolling window estimation, the trend implied bytheBVARgapsisquiteflexibleandresponsivetonewobservations. Bycontrast, forcredit-to-GDPgaps, ahigh smoothingparameterimpliesanearlylinear,inflexibleHPtrend. Whenfacedwiththeartificialpeakgeneratedbythe normalization to GDP, this trend does not adjust upwards, thus generating a persistent series of positive credit gaps after a crisis. The rigidity of the trend also explains why credit-to-GDP gaps stay in the negative after the crisis for manyyears.Allthesefeaturesbecomeevenmorepronouncedfortwo-sidedcredit-to-GDPgaps,asthismethodyields anadditionalreasonfortrendrigidity. Duringaprolongedcreditexpansion,thetwo-sidedtrendislikelytoabsorba lotoftheacceleratingdynamics,sothattheresultingcyclicalcomponentiseithermildlypositiveorevennegative. The previous qualitative discussion suggests that BVAR-based gaps can have good early warning properties and perhapsevenoutperformtheircredit-to-GDPcounterparts. Notethatbothmethodsaresuppliedwiththesamecredit measuresonthesametimespantoensurecomparabilityofresults. Inordertoformallytestthesignificanceofthese observeddifferences,IrunpooledlogitregressionsforBVARcreditgapsandbothversionsofcredit-to-GDPgaps,for eachofthecrisischronologiesseparately. Thegoalistotestwhethercreditgapscanpredicttheonsetofthecrisisup tofouryearsinadvanceandhowwelltheycanpredictit. Toevaluatethelatterpartofthequestion,Icomputethearea underthereceiveroperatingcharacteristic(ROC)curve. Thiscurveplotsatruepositiverateagainstafalsepositive rate. Avalueof0.5wouldindicatediagnosticpropertiesofafaircoin: itwouldcorrectlypredictacrisisin50percent ofcases,andthecorrespondingROCcurvewouldbeadiagonalline. Inotherwords,thefurtherawayfrom0.50the ROCvalue,thebetterthediagnosticqualities. Table1illustratestheresultsfortherelevantcountriesandafullsample. ThepooledlogitcoefficientsforBVAR credit gaps are highly statistically significant at all horizons, regardless of chronology, while ROC values fluctuate around0.70. One-sidedcredit-to-GDPgapsareequallygoodascontemporaneouspredictorsofcrises(althoughtheir ROCvaluesareabithigheratthishorizon),signalcrisescomparablywell(albeitwithasomewhatlowerROCvalue) onetotwoyearsahead,andaresubstantiallyworsethanBVARgapsatlongerhorizonsofthreetofouryears.Forthese longerhorizons, theROCvaluesarelargerbyevenmorethanfor1-2yearsahead, sothatwecanrejecttheNullof equaldiagnosticperformancebetweenBVARgapsandcredit-to-GDPgapsathighstatisticallevels. Two-sidedcredit- 15

to-GDPgapsaremarginallycompetitivewithBVARgapsonlycontemporaneously. Atallotherhorizons, statistical significanceofthecorrespondingcoefficientsfadesaway,ROCvaluesdecrease,andthep-valuesrejectingtheNullof equaldiagnosticperformancebetweenBVARgapsandcredit-to-GDPgapsaremostlylowerthan0.01. Table1:CrisisPredictabilityResultsacrossVariousChronologies,FullSample LVChronology(981observation,12crisisepisodes) Horizon BVARcreditgap Credit-to-GDPgap(one-sided) Credit-to-GDPgap(two-sided) Coeff. ROC Coeff. ROC ROCp-val. Coeff. ROC ROCp-val. contemporaneously 8.72*** 0.63 8.29*** 0.70 0.43 6.18*** 0.67 0.75 1yearahead 12.73*** 0.69 7.42** 0.68 0.84 4.25 0.62 0.18 2yearsahead 12.17*** 0.70 6.55** 0.68 0.16 2.23 0.57 0.00 3yearsahead 11.78** 0.69 5.85* 0.65 0.02 0.70 0.53 0.00 4yearsahead 11.78*** 0.70 6.07** 0.65 0.02 0.04 0.51 0.00 BVXChronology(1037observations,12crisisepisodes) Horizon BVARcreditgap Credit-to-GDPgap(one-sided) Credit-to-GDPgap(two-sided) Coeff. ROC Coeff. ROC ROCp-val. Coeff. ROC ROCp-val. contemporaneously 12.14*** 0.71 9.78*** 0.73 0.83 6.88*** 0.68 0.00 1yearahead 13.03*** 0.74 7.64*** 0.71 0.12 3.27 0.58 0.00 2yearsahead 12.11*** 0.73 7.03*** 0.69 0.03 1.15 0.52 0.00 3yearsahead 11.52*** 0.72 6.22*** 0.67 0.00 -0.24 0.51 0.00 4yearsahead 11.37*** 0.72 6.41*** 0.67 0.00 -0.82 0.52 0.00 STChronology(866observations,10crisisepisodes) Horizon BVARcreditgap Credit-to-GDPgap(one-sided) Credit-to-GDPgap(two-sided) Coeff. ROC Coeff. ROC ROCp-val. Coeff. ROC ROCp-val. contemporaneously 12.01*** 0.69 9.57*** 0.70 0.90 7.72*** 0.70 0.96 1yearahead 12.15*** 0.73 7.91*** 0.70 0.36 4.59** 0.59 0.01 2yearsahead 11.40*** 0.72 7.53*** 0.69 0.22 2.75 0.54 0.00 3yearsahead 11.12** 0.71 6.79*** 0.67 0.07 1.78 0.53 0.00 4yearsahead 11.42** 0.71 7.18*** 0.68 0.10 1.36 0.52 0.00 LDChronology(860observations,14crisesandresidualevents) Horizon BVARcreditgap Credit-to-GDPgap(one-sided) Credit-to-GDPgap(two-sided) Coeff. ROC Coeff. ROC ROCp-val. Coeff. ROC ROCp-val. contemporaneously 5.64** 0.63 6.71*** 0.71 0.34 4.61*** 0.60 0.83 1yearahead 7.15*** 0.67 5.55*** 0.68 0.79 1.98 0.54 0.02 2yearsahead 7.93*** 0.69 4.52** 0.65 0.20 -0.16 0.52 0.00 3yearsahead 7.92*** 0.69 4.07 0.63 0.01 -1.51 0.56 0.00 4yearsahead 7.64*** 0.68 3.61 0.62 0.00 -2.31 0.59 0.00 Notes: Standarderrorsofpooledlogitregressioncoefficientsareclusteredatcrisisepisodes. ∗∗∗,∗∗,∗ standfor 1%,5%,and10%statisticalsignificancelevels,respectively. ROCpvaluesreflect,atwhichlevelofstatistical significancetheNullofequalROCbetweenBVARcreditgapandthecorrespondingcredit-to-GDPgapcanbe rejected. ThepreviousresultsmaydeliverconservativeestimatesonthedifferencesbetweenBVARgapsandcredit-to-GDP gaps. First, there are episodes within those full samples that count as a false positive for the BVAR gaps and boost thediagnosticqualitiesofcredit-to-GDPgaps. ConsidertheexampleofAustralia(figure1a). Allfourchronologies 16

account for only one crisis for this country – the one in the late 1980s. In particular, there is no crisis in 2008. At thesametime,AliberandKindelberger(2015)aswellasBrunnermeierandSchnabel(2011)describetheearly2000s in Australia as an example of a credit boom that was eventually landed successfully due to timely interventions of monetary and macroprudential policies. The BVAR-based credit gap detects this credit boom in a timely manner, whereasone-sidedcredit-to-GDPgapstaysinthenegativeterritory. Thepreviouspooledlogitregressionassignsthis episodeasafalsepositivefortheBVARgap,loweringitsROC,andasasuccessforthecredit-to-GDPgap. Second, manycountrysamplesendin2018and2019,whereasnotallcriseschronologiesareupdatedthatfarintothefuture. Hence, BVAR credit gaps, which tend to rebound faster after the Global Financial Crisis, might receive even more potentiallyinadequatefalsepositivesfortheseyearsaswell. In order to clean the exercise from these potential biases, I next concentrate on systemic crises only and run the pooled logit regressions on 10-year windows around the systemic crises rather than the full samples as before. The question now is, conditional on the presence of a systemic crisis, how do early warning qualities of credit gaps compare. The results are presented in table 2. The differences between BVAR-based gaps and credit-to-GDP gaps are even starker, as expected. One-sided credit-to-GDP gaps are still significant crisis predictors at horizons zero to two years, yet the corresponding ROC numbers are substantially lower relative to BVAR gaps. We can largely reject the Null of equal predictive qualities of BVAR gaps and one-sided credit-to-GDP gaps for horizons one to four years ahead. As for two-sided credit-to-GDP gaps, they remain good contemporaneous predictors and emerge as (marginally) significant predictors 4 years ahead, albeit with a negative sign. In other words, negative two-sided credit-to-GDPgapscanbestatisticallysignificantpredictorsofavulnerabilitybuild-upinthemediumterm. Afewcountrieswerenotprominentlyfeaturedinthepreviousregressions.Thefirstgrouparethecountries,where thecrisisonsetfallswithinthefirstfewquartersofavailablegapestimates,suchasItaly,Luxembourg,orIreland,for example(figures2d,2g,and2c). Thesecountriesdonothaveasufficientnumberofpre-crisiscreditgapobservations toparticipateintheearlywarningcomparisonexercise.Thepost-crisisbehaviorofBVARgapsinthesecasesdoesnot differfromtheoneoftheothercountrieswherethebuildupphaseisalsoavailable.BVARcreditgapsgointonegative territoryattheonsetdateandreboundfasterthantheircredit-to-GDPgapcounterparts. Second, thereisasubsetof countrieswherethechronologiesdonotassignanycrisesepisodes,suchasSouthAfrica(figure3g,Chile(figure1f), Singapore(figure3e), orCanada(figure1e). Inthesecases, wheneverbothmethodsidentifyacreditexpansion, for example,creditexpansionsintheearly2000s,BVAR-basedcreditgapstendtoidentifyitearlier. Theseobservations furthersupporttheconclusionthatBVAR-basedgapscandetectturningpointsfaster. Overall,theproposedBVARcreditgapmethodologydeliversreasonableresultsforbothadvancedandemerging market economies, even though in the latter case, data constraints are often binding.14 That said, there are several 14Forinstance,industrialproductiontimeseriesforSouthAfricahadtobesubstitutedbyasurvey-basedmeasureonbusinesscycle. 17

Table2:CrisisPredictabilityResultsacrossVariousChronologies,10-YearWindowsaroundtheCrises LVChronology(658observation,12crisisepisodes) Horizon BVARcreditgap Credit-to-GDPgap(one-sided) Credit-to-GDPgap(two-sided) Coeff. ROC Coeff. ROC ROCp-val. Coeff. ROC ROCp-val. contemporaneously 8.75*** 0.63 6.33** 0.64 0.93 3.97* 0.62 0.90 1yearahead 13.61*** 0.69 5.21 0.63 0.06 1.74 0.56 0.02 2yearsahead 13.55*** 0.71 4.03 0.60 0.00 -0.71 0.50 0.00 3yearsahead 13.65*** 0.71 2.98 0.57 0.00 -2.85 0.55 0.00 4yearsahead 14.24*** 0.72 3.09 0.57 0.00 -4.09 0.58 0.00 BVXChronology(658observations,12crisisepisodes) Horizon BVARcreditgap Credit-to-GDPgap(one-sided) Credit-to-GDPgap(two-sided) Coeff. ROC Coeff. ROC ROCp-val. Coeff. ROC ROCp-val. contemporaneously 10.94*** 0.70 7.62*** 0.66 0.43 4.38*** 0.63 0.49 1yearahead 12.57*** 0.74 4.96** 0.63 0.00 0.48 0.52 0.00 2yearsahead 12.26*** 0.74 4.02 0.61 0.00 -2.10 0.55 0.00 3yearsahead 12.20*** 0.73 2.79 0.57 0.00 -4.17 0.59 0.00 4yearsahead 12.63*** 0.74 2.79 0.57 0.00 -5.45* 0.62 0.00 STChronology(589observations,10crisisepisodes) Horizon BVARcreditgap Credit-to-GDPgap(one-sided) Credit-to-GDPgap(two-sided) Coeff. ROC Coeff. ROC ROCp-val. Coeff. ROC ROCp-val. contemporaneously 10.43*** 0.67 7.17*** 0.62 0.55 5.17*** 0.63 0.79 1yearahead 10.57** 0.71 4.82** 0.62 0.00 1.56 0.53 0.00 2yearsahead 9.95** 0.70 4.03 0.60 0.00 -0.79 0.52 0.00 3yearsahead 9.84** 0.69 2.88 0.57 0.00 -2.29 0.55 0.00 4yearsahead 10.38** 0.70 3.02 0.67 0.00 -3.21 0.57 0.00 LDChronology(688observations,14crisesandresidualevents) Horizon BVARcreditgap Credit-to-GDPgap(one-sided) Credit-to-GDPgap(two-sided) Coeff. ROC Coeff. ROC ROCp-val. Coeff. ROC ROCp-val. contemporaneously 4.33* 0.60 5.82*** 0.66 0.44 3.21* 0.56 0.79 1yearahead 5.80** 0.64 4.33** 0.63 0.83 0.36 0.50 0.02 2yearsahead 6.50*** 0.66 2.98 0.60 0.05 -2.14 0.58 0.01 3yearsahead 6.45*** 0.66 2.29 0.57 0.00 -4.03 0.62 0.22 4yearsahead 6.08*** 0.65 1.57 0.55 0.00 -5.48* 0.66 0.78 Notes: Standarderrorsofpooledlogitregressioncoefficientsareclusteredatcrisisepisodes. ∗∗∗,∗∗,∗ standfor 1%,5%,and10%statisticalsignificancelevels,respectively. ROCpvaluesreflect,atwhichlevelofstatistical significancetheNullofequalROCbetweenBVARcreditgapandthecorrespondingcredit-to-GDPgapcanbe rejected. 18

unsatisfactorycountry-specificresults,suchasthoseforAustria(figure1b),Portugal(figure3d)orTurkey(figure4b), whereBVAR-basedgapsdonotadequatelyreflectthebuild-upoffinancialvulnerabilitiesintherun-uptothecrises. Interestingly,inthesecases,BVARcreditgapsandcredit-to-GDPgapsarequitecorrelated–thatis,bothapproaches donotworkwell. Onereasonofthesefailurescouldbethatthechosendataseries(thecreditseries,mostimportantly) donotfullyreflectthedynamicsofthecreditcycleinthisparticulareconomy.15 Perhaps,amoredetailedlookatthe decompositionofcreditaggregatesbylenderandborrowercouldyieldbetteranswers. 4. CreditGapPerformanceinRealTime Theprevioussectioncomputedthegapsunderthe‘idealconditions’ofex-postreviseddata. Anothertestofthe creditgapmethodologyisitsreliabilityinrealtime–whenactualdecisionmakingoccurs. Twoconcernsariseinreal time: end-pointfilteringproblemsandtheuseofthereal-timedatavintagesperse. Iaddressboththesedimensionsin turnandillustratethemfortheU.S.economyasarepresentativeexample.16 Figure7:QuasiReal-TimeExercisefortheU.S.:CreditGaps(inPercentagePoints)fromVariousEstimationWindows(left)andtheCorresponding AverageRootMeanSquaredErrorsofConditionalCreditForecastbyHorizon(right). First,considerareviseddatavintageasofMarch2020,butinsteadofaggregatingalltheinformationinthecredit gap as of March 2020, let the computation pause each year in March. The resulting quasi real-time credit gaps are showninfigure7. Inthisexercise,asthedatastemfromthesamevintage(hencethequasireference),theend-point 15Forsomeoftheemergingmarkets,thesovereigndebtoftentriggerscrises. Governmentdebtisexcludedfromthecreditaggregatesstudied inthispaper. 16U.S.exampleischosenbecauseofitslongersampleandalargernumberofreal-timevintagesavailableforthetimeseriesoftheBVAR. 19

qualities of the filter come to light. On the one hand, the real-time version of the credit gap still correctly reflects the signs and overall dynamics of the full-sample time series. The only change in sign occurs around 2015, when thecreditgapisnearlyzeroandthereal-timeversionsareeitherslightlyaboveorbelowzero. Bycomparison,Edge andMeisenzahl(2011)findthatreal-timecredit-to-GDPgapschangesigns27percentofthetime. Also, theability to swiftly capture turning points is preserved for real-time BVAR gaps. For instance, while the estimate from the window ending in December 2007 shows a positive gap, the next computed estimate already displays a deep credit bustfollowingtheonsetoftheGreatRecession. Figure8:QuasiReal-TimeExercise:U.S.CreditGapsfromVariousEstimationWindowswithBaseline(left)andAdjustedWeights(right). Notes:Timeonthehorizontalaxesisinmonths.Creditgapsareexpressedinpercentagepoints. Ontheotherhand,theend-pointproblemofthefiltermanifestsitselfintheovershootingfeature: Thelast12-15 monthsoftherespectiveestimationwindow,althoughretainingthecorrectsign,tendtooverestimatethemagnitudeof thegapinabsoluteterms.Thisfeatureisespeciallyprominentatthepointsintimewherethegapincreasesordecreases atahighrate,suchas,forinstance,attheheightofthecreditboomprecedingtheGreatRecessionoratthetimeofa rapidreversalfollowingtheonsetoftheGlobalFinancialCrisis(figure7). Thereasonforsuchovershootingisthat,at theendoftherespectivedatawindow,forecastsatlongerhorizonsreceivedisproportionatelyhigherweights,asonly thoseforecastsareavailableandtendtobetheleastaccurate.Figure7illustratestheaverageroot-mean-squarederrors ofconditionalcreditforecastsbyhorizon: from1to48months. Theseerrorsdisplayaclearupward-slopingprofile. Within a full sample (as of March 2020, for instance), these forecasts would be pooled with shorter-term forecasts fromthenextestimationwindowsandsmoothedoutduetoaveragingateachpointintimewithequalweights(asin 20

Figure9:Real-TimeExercise:U.S.CreditGapsfromVariousEstimationWindowswithBaseline(left)andAdjustedWeights(right). Notes:Timeonthehorizontalaxesisinmonths.Creditgapsareexpressedinpercentagepoints. equation(2)). Inrealtime,theverylastforecast48monthsahead,wouldreceiveaweightofone(asopposedto1/n, withnbeingthenumberofforecastsavailableinfullsample),sinceitistheonlyforecastavailable. Theseobservationssuggestawaytomitigatetheovershootingproperty–byreducingthedisproportionatelyhigh weights of longer-term forecasts at the end of the sample. One simple weight adjustment rule illustrated in figure 8 reducestheweightsoflonger-horizonforecasterrorsattheendofthesample,sothattheyareequaltotheirwithinsample arithmetic average counterparts – (1/n). For instance, if the actual weight at the end of the sample is x, it is multipliedbythefactor1/(nx).17 Thisrulereducestheovershootingconsiderably,bringingthereal-timevintagesof creditgapveryclosetothefull-samplerevisedversion(rightpanelofFigure8). Finally, I add the second real-time data aspect to the exercise. Each estimation window is now based on the respectivereal-timedatavintage–thetimeseriesastheyareavailableatthatparticularpointintime(seeAppendix C for details on the data sources, the corresponding time series revisions, and timing of data vintages). The results withbaseline,orfull-sample,andadjustedweightingareinfigure9. Baselineweightingresultscloselyresembletheir quasireal-timecounterpartsfromthepreviousexercise(figure8,leftpanel). Hence,theprimaryquantitativesource ofreal-timediscrepanciesofthegapestimatestemsfromtheend-pointpropertiesofthefilterratherthantheuseof theactualreal-timedatavintages. Alsointhiscase,adjustedweightingschememitigatestheovershootingfeaturetoa similardegreeasinthequasireal-timeexercise(figure10). Averagedeviationsofreal-timegapsfromthefull-sample 17Alternatively,arulethatmakestheweightsaninversefunctionofRMSEsatthecorrespondinghorizoncouldbeconstructed. 21

Figure10:AverageAbsoluteDeviationsofQuasiReal-TimeCreditGaps(left)andReal-TimeCreditGaps(right)fromtheirRespectiveFull-Sample Counterparts,asaFractionoftheFull-SampleCreditGapStandardDeviation.Timeonthehorizontalaxesisinmonths. counterpartsareaboutthreetofourtimessmallerwithweightadjustmentandmeasureatmost31percentofthecredit gapstandarddeviationforthequasireal-timeversionand43percentforthereal-timeversion, respectively. Putting these numbers in perspective, Edge and Meisenzahl (2011) show that credit-to-GDP gap revisions in real time have a typical size of about one standard deviation of the gap and have a maximum value of slightly above two standard deviations. 5. CreditBoomsandShort-TermInterestRates Another important question besides timely identification of credit booms is the question about their nature. The advantageoftheVARcoreofmymethodologyisthatsuchquestionscanbestudied.HereIillustrateitfortheexample oftheroleofshort-terminterestratesduringthebuild-upofacreditboom. Oneofthehypothesesreferstotheorigin ofcreditboomsinalowinterestrateenvironment.Thereareseveraltheoreticalexplanationsforthiseffect:the“search for yield” theory by (Borio and Zhu, 2012) among others, and “income and valuation effects” by (Adrian and Shin, 2010;Adrianetal.,2010). Thereisalsoquiteabitofempiricalevidenceexaminingthesehypothesesinapplications tovariouscountriesintheyearsbeforetheGlobalFinancialCrisis.18 To study the role of short-term interest rates for credit booms, I perform a forecast exercise for credit where I condition on the path of the monetary policy rate (the federal funds rate in the U.S.) instead of the real economic 18Forinstance,Jimenezetal.(2014)andIoannidouetal.(2015)studythechannelusingloan-leveldata.Atthemacrolevel,Buchetal.(2014) andAfanasyevaandGuentner(2020)studythequestionusingtheFAVARapproach. 22

activity. Now conditional forecasts of credit can be seen as amounts of credit consistent with the current stance of monetarypolicyratherthanwiththerealeconomicactivity,aswasdonebefore.Theresultsshowthat,indeed,thesize ofthecreditgapsbeforethecrisisissubstantiallyreduced.Figure11illustratesthisfindingfortheU.S.creditvariable. In the left panel of figure 11, where the forecasts are conditional on industrial production, the model systematically under-predicts the actual growth rates of credit at all forecast horizons. Accordingly, credit gaps are positive, and a creditboomisdetectedintheseyears. Theonlychangeintherightpanelisthattheforecastsareconditionalonthe federalfundsratepath,whilethesameVARmodelisestimatedoverthesamerollingwindowasintheleftpanel.Now thegapbetweentheconditionalforecastsandtheobservedcreditvaluesdecreasessubstantially. Themeanforecasts evenresembletheshapeoftheactualcreditgrowthmuchcloser, andthebandsreflectingtheestimationuncertainty nowincludetheactualcreditgrowthstartingfrom2005. Resultsformanyoftheeuroareaeconomiesarequalitativelysimilarforthisepisode thecreditboompreceding the Global Financial Crisis. Conditioning on the short-term interest rates reduces the values of the credit gap or, in otherwords,itexplainsawaysomeportionsofthisparticularcreditboom. Figure11:CreditForecastsfor2004-2007intheU.S.:conditionedonindustrialproduction(leftpanel)andontheFederalFundsRate(rightpanel). Notes:Forecastbandscorrespondtothe16-thand84-thpercentilesandpointwisecontain68%oftheprobabilitymass.Forecasts oflog-levelsareconvertedtoyear-on-yeargrowthrates. Thesefindingssuggestthatmonetarypolicyratescouldhaveplayedanontrivialroleinthebuild-upofthecredit boombeforetheGreatRecession,feedingrisk-takingmotivesoffinancialintermediaries. Oneshouldnot, however, overestimatetheseresults. Asfigure11illustrates, monetarypolicyratescannotexplainthecreditboomof2003to 2007 completely as the gap between observed credit and its conditional forecast still remains positive, especially in the early phase of this credit boom. There are additional factors driving excessive lending behavior and excessive 23

Figure12:GapsforM2MonetaryAggregate(left)andAssetPrices(right)intheU.S. Notes:LVreferstothechronologyofLaevenandValencia(2013);BVXreferstothechronologyofBaronetal.(2018);STrefers tothechronologybySchularickandTaylor(2012).Gapsareexpressedinpercentagepoints.Timeonthehorizontalaxesisin quarters. risktakingoffinancialintermediariesduringcreditbooms. Forinstance,thestanceandimplementationofregulatory policies,marketsentiment,andoverheatedmacroeconomicconditionsduringcreditboomscouldbeadditionaldriving forces. Evaluatingthecontributionandrelativeimportanceofthesefactorsisanimportantquestionforfuturework.19 6. MoneyandAssetPriceGaps Broad monetary aggregates and asset prices have been linked to the emergence of financial imbalances in the academicliteratureandpolicydiscussions. IapplytheBVARmethodologytoconstructthecorrespondinggapsand comparetheirpropertieswiththoseofcreditgaps. 20 The role of money as an early warning indicator for financial imbalances has become a subject of academic discussionintheaftermathoftheGreatRecession. Oneofthereasonsisempirical. Beforethecrisisof2007-2008, large swings in broad money growth were observed in many advanced economies, including the U.S. and the euro area. Inparticular, aperiodofacceleratinggrowthratesbeforethecrisiswasfollowedbyanabruptandremarkable 19Infact,someoftheongoingrecentworkonthisquestionincludesKrishnamurthyandLi(2020),GortonandOrdonez(2020)andAfanasyeva etal.(2020). 20RecallthatrealeconomicactivityistheonlyvariableintheBVARthatisbeingconditionedonintheforecastingexercises. Hence,money andassetpricegapsarethebyproductsofthecreditgapcomputationandcanbecalculatedthesamewayascreditgaps. 24

fall in growth rates, as the crisis unfolded.21 Several studies investigate the link of money and credit dynamics to crisesandassetpricebooms.Forinstance,SchularickandTaylor(2012)contrastthecreditviewwiththemoneyview, studyingtrendsinalonghistoricaldataset. Thecreditboomdetectionliteraturehasalsostressedthenexusbetween creditboomsandassetpricesmovement. BorioandLowe(2002)findthatswingsinassetpricestendtogohandin handwiththecreditcycle. Icomputethegapsforbroadmoneyandassetpricesforthesamesetofcountriesasbeforeinordertotestwhether these gaps have better early warning qualities than credit gaps and what kind of episodes these gaps reflect. While detailed graphical results for each country are presented in Appendix E, consider an illustrative and representative exampleoftheU.S.infigure12. AlthoughaprolongedperiodofpositivemoneygapsprecedestheGlobalFinancialCrisis(figure12,leftpanel), moneygapsaremildlynegativebeforethesavingsandloancrisisregardlessofthechronologyconsidered. Furthermore, the largest spike in broad money gap dynamics occurs shortly after the Global Financial crisis between 2010 through the beginning of 2014 – the years, when QE 1, 2, and 3 were launched. More generally, I find similar tendencies for other countries. First, there are often substantial, persistent positive money gaps preceding the Global FinancialCrisis,butnotnecessarilybeforetheothercrisesinthesample(see,forexample,Sweden(figureE.3i)and Norway(figureE.3b)inAppendixE).Second, incontrastwithcreditgapsthatplummetintonegativeterritorywith thecrisisonset,moneygapstypicallystaypositiveforseveralquartersafterthecrisisonset,reflectinginmanycases consequencesofpost-crisismonetaryeasingpolicies. Thesefeaturesmaydiminishearlywarningqualitiesofbroad moneyinahorseraceagainstcreditgaps. Asset price gaps (figure 12, right panel) are larger in size than money gaps and are more volatile. At the onset date of both systemic crises in the U.S., these gaps tend to decline precipitously. The largest boom-bust dynamics is, however, observed around the dot-com bubble episode. Although the gap swings up slightly before the Global Financial Crisis, it rebounds quite quickly shortly thereafter. Again, these tendencies are representative for many countriesinthesample. Periodsofpositiveassetpricegapsprecedemanymajorcrisesepisodes(suchastheGlobal FinancialCrisisortheHeiseibubbleepisodeinJapaninthelate1980s,tonameafewprominentexamples),theyare generallyvolatile,spikingshortlythereafter. Thisfeaturecouldincreasethefalsepositiverate,whenitcomestothe earlywarningqualitiesofassetpricegapsrelativetocreditgaps. ToformalizetheearlywarninghorseracebetweenBVARcreditgapsandtheirrespectivemoneyandassetcounterparts,Icomparetheresultsfromthecorrespondingpooledlogitregressions(table3). Intheseregressions,Iconsider 21ThislatterfallinmoneygrowthwassosubstantialthatitrevolvedanalogiestothedownwarddynamicsobservedduringtheGreatDepression. Giannone,Lenza,Pill,andReichlin(2011)conductadetailedcomparisonofthesetwoepisodes. 25

eachgapasanearlywarningindicatorinisolation.22 Table3:CrisisPredictabilityResultsacrossVariousChronologiesandBVARGaps,FullSample LVChronology(981observation,12crises) BVXChronology(1037observations,12crises) Horizon Creditgap Moneygap Assetpr. gap Creditgap Moneygap Assetpr. gap Coeff. ROC Coeff. ROC Coeff. ROC Coeff. ROC Coeff. ROC Coeff. ROC contemp. 8.72∗∗∗ 0.63 3.60 0.60 -0.78 0.60 12.14∗∗∗ 0.71 3.70 0.63 −0.91∗ 0.61 1Yahead 12.73∗∗∗ 0.69 4.54 0.60 0.79 0.55 13.03∗∗∗ 0.74 4.29 0.63 0.54 0.54 2Yahead 12.17∗∗∗ 0.70 3.83 0.61 1.15 0.58 12.11∗∗∗ 0.73 3.62 0.61 0.90 0.58 3Yahead 11.78∗∗ 0.69 2.88 0.59 1.13 0.58 11.52∗∗∗ 0.72 2.76 0.60 1.00 0.58 4Yahead 11.78∗∗∗ 0.70 2.05 0.57 0.62 0.53 11.37∗∗∗ 0.72 1.86 0.57 0.62 0.55 STChronology(866observations,10crises) LDChronology(860observations,14crises) Horizon Creditgap Moneygap Assetpr. gap Creditgap Moneygap Assetpr. gap Coeff. ROC Coeff. ROC Coeff. ROC Coeff. ROC Coeff. ROC Coeff. ROC contemp. 12.01∗∗∗ 0.69 4.28 0.64 -0.35 0.55 5.64∗∗ 0.63 6.44∗∗ 0.73 −0.65∗∗ 0.60 1Yahead 12.15∗∗∗ 0.73 4.12 0.61 1.26 0.60 7.15∗∗∗ 0.67 9.13∗∗ 0.78 0.68 0.55 2Yahead 11.40∗∗∗ 0.72 3.26 0.59 1.69∗ 0.65 7.93∗∗∗ 0.69 11.36∗∗ 0.80 1.16 0.59 3Yahead 11.12∗∗ 0.71 2.02 0.57 1.59∗ 0.64 7.92∗∗∗ 0.69 12.11∗∗ 0.80 1.64∗∗ 0.63 4Yahead 11.42∗∗ 0.71 0.97 0.55 0.92 0.59 7.64∗∗∗ 0.68 11.27∗∗ 0.78 1.75∗∗ 0.64 Notes: Standarderrorsofpooledlogitregressioncoefficientsareclusteredatcrisisepisodes. ∗∗∗,∗∗,∗standfor1%, 5%,and10%statisticalsignificancelevels,respectively. Acrossthemajorityofchronologies(LV,ST,andBVX),thecoefficientsformoneygapsareinsignificant,andthe ROCnumbersformoneygapsaresmallerthanthoseofcreditgapsatallhorizons. TheLDchronologyisanotable exception,wherebroadmoneygapsarenotonlysignificantalongwithcreditgaps,butareactuallysomewhatbetterin theirearlywarningqualities.Thereasonforthisresultstemsfromthecountrysamplecompositionforthischronology. ThesampleprimarilyconsistsofEuropeancountrieswithobservationssurroundingtheGlobalFinancialCrisis. As discussedbefore,increasesinbroadmoneygapshavebeenquitesubstantialandsynchronousacrossthecountriesfor thisparticularcreditboom. Assetpricegapssignaltheonsetofthecrisescontemporaneously,astheyrapidlydecreaseatthisdate:Thepooled logit coefficients are marginally significant but negative. At longer horizons, due to a higher rate of false positives, the significance goes away and the ROC numbers are always smaller when compared with credit gaps. For the LD chronologysample,whichislargelycenteredaroundtheGlobalFinancialCrisis,thereissome(albeitnotverystrong) significanceofassetpricegapsaspredictorsathorizonsofthreeandfouryearsbeforetothecrisis. Alsointhiscase, 22Theresultsconveythesamemessage,whenallthreegapsareincludedintoonepooledregressionsimultaneously–thatis,thegapwiththe bestearlywarningqualitiesstillprevailsinthissetting,leavingtheothercompetitorsinsignificantandtheircontributiontotheincreaseinROC mostlymarginal. 26

however,creditgapsdominateintermsofstatisticalsignificanceandtheROCvalues. Theseresultsindicatethat,onaverage,BVARcreditgapsprovideamiddlegroundbetweenassetpricegapsthat aremorevolatileandhenceoftenpronetofalsepositivesandthecredit-to-GDPgapsthattargetcreditcyclesatalower frequency. Broadmoneygapsareusefulpredictorsforasubsetofepisodes. Thatsaid,therankingofearlywarning indicatorsmaydifferforaparticularcountrytakenindividually(see,forexampletheperformanceofassetpricegaps forSweden(figureE.3i)inAppendixE)andmaydependonatypeofassetpriceconsidered. 7. RobustnessExercises Allfindingspresentedpreviouslyareconditionalonthebaselinemodelspecificationandtheassumptionsofthe forecasting exercise. A first robustness check regards the credit measure in the system. I replace the broad total nonfinancialcreditmeasurefromtheflowoffundsbythetotalloansandleasesmeasureinthebaselineVARsystem and rerun the whole exercise. The latter measure is available at a monthly frequency and is narrower, as it includes credit to the nonfinancial sector on bank balance sheets only. The results presented in figure F.1 are quite intuitive. Majorboom-bustepisodesaroundthesystemicbankingcrisesarestillreflectedinforthealternativemeasure,although the bank-based credit gap predicts the savings and loan crisis, which primarily affected bank balance sheets, even earlier. Also, as to be exepcted, bank credit measure dips substantially more in 1975 to1979 – the years of a credit crunchfollowingthebankingcrisisof1974. Abroadermeasurefromtheflowoffundsreflectsthisbusttoo,butthe downturnissmaller(inabsoluteterms)quantitatively. Afinalnoteworthydifferenceisthedot-comepisode,wherethe bank-basedmeasurefluctuatesaroundzero,reflectingthewell-knownfactthatthisparticularboomepisodewasnot fundedbybankcredit(AliberandKindelberger(2015)amongothers). Several studies use real credit measures in order to detect credit booms. For instance, Mendoza and Terrones (2008) use real credit per capita as a benchmark measure of credit. Overall the results for the real per capita credit measure are quite similar to the nominal credit measure (figure F.2), although the credit boom preceding the Great Recessionissomewhatmoreprotractedfortherealpercapitacreditversion. Inextinspecttherobustnesswithrespecttothevariablesincludedinthebaselinemodelspecification.Therecould be potential considerations of omitted variable bias or, on the contrary, the presence of some variable, such as asset prices,drivingthemajorresults.Totestthesehypothesesmorethoroughly,Iconsiderbothasmallersystem(consisting ofindustrialproduction,short-terminterestrate,broadmeasureofmoney(M2)andcredit)andalargersystem(table F.2 for description of 16 variables included in it). Table F.3 sheds light on the forecasting qualities of the systems across rolling windows. We know from the VAR forecasting literature (see, for instance, Banbura, Giannone, and Reichlin(2010))thattheforecastingperformanceofBVARscanbeimprovedonceadditionalvariablesareincluded– thatis,inmediumandlargesystems(onceproperpriorshrinkageisapplied). Indeed,inmanyrollingwindowsandon averageacrossallwindows(althoughbyanarrowmargin),thelargerBVARsystemdeliverssmallerforecasterrors. 27

Theinspectionoftheresultingcreditgaps(figureF.3inAppendixF)reveals,however,thatthetimingoftheidentified eventsacrossthreesystemsisnearlyidentical.Theadditionofnewvariablesdoesimprovetheforecastingaccuracyof thesystemintheearly1990s(thecreditcrunchperiod)andtheearly2000s(thecreditboomperiod)consistentlywith theforecasterrorsintableF.3. Quantitativedifferencesinthegapestimatesare,however,small. Theonlyexception istheepisodeatthestartofthesampleforthelargestVARsystemin1989. Insmallersystems,boththereducedand thebaselineVAR,thegapismildlynegative,whereasforthelargersystemitismildlypositive. In a similar vein, replacing the monetary policy rate by the shadow rate estimate of Wu and Xia (2016) during the zero lower bound period as well as including commodity prices do not affect the results substantially. I also included the national Freddie Mac housing price index for the U.S. BVAR and performed the forecasting exercises conditioningonit(ratherthanonrealactivityvariables). Inthiscase,thecreditboombeforetheGreatRecessionis firstdetectedaboutayearbeforethecrisisonset,whenthetrendinhousingpricesstartstoreverseitself. Thisexercise againunderscorestheimportanceofthevariablethatisconditionedon. Ashousingpricesthemselveswerethemajor driving force of the credit boom in the early 2000s, conditioning credit forecasts on them helps explain this boom ratherthanidentifyitintime.23 Continuingonthechoiceofthefundamentalvariable,Iconsideralternativesforrealactivityvariables. Forsome oftheadvancedeconomies,includingtheU.S.,baselinemonthlyVARsemployindustrialproductionasrealactivity variable. Arguably,broadermeasures,suchasGDP,couldbeamoreappropriatechoice. ThereforeIcomputecredit gapsusingmonthlymeasuresofrealGDPconstructedbyMarkWatsoninsteadofindustrialproduction. Theresults are quite similar to the baseline. Major episodes are identified at the same time; minor episodes, such as a small upward deviation in 2016-2017 in the U.S., become less pronounced. The results do not change substantially also whencreditforecastsareconditionalonboththeunemploymentrateandtheindustrialproductionindexinthelarger system. Finally,therobustnesswithrespecttothesizeoftherollingwindow,theforecastinghorizon,andthepriortightness isexamined.VaryingthesizeoftherollingwindowfortheU.S.between10and20yearsdoesnotaffecttheresultssignificantly. Applyingaforecasthorizonofoneyearproducescreditgapsthatidentifyonlyasubsetofepisodes. Large credit booms are still detected but the signal becomes noisier and less persistent. For instance, there are interrupted upwarddeviationsinthe2003to2007episode. Thereisabenefitinaccountingfortheforecastsatlongerhorizons, given that large credit booms build up gradually and are quite persistent phenomena. In the baseline specification, priorhyperparameters,includingtheoverallpriortightness,aresetbasedonmarginallikelihoodmaximization. Inthe robustnessexercise,Irelaxthetightnessofthepriortotesttowhatdegreethepriorisdrivingtheresults(figureF.4). Although the relaxation of prior tightness somewhat increases the errors in several episodes somewhat (as is to be 23Toconservespace,theseandsomeotherresultsinthissectionareomittedandavailableuponrequest. 28

expected,sincepriorshrinkageisnolonger‘optimal’andoverfittingbecomesmoreprominent),theoverallcontours of the credit gap dynamics still remain similar (the correlation between the two credit gaps is 0.89). The episodes identified as busts typically run even deeper with a looser prior, the timing and size of the booms are quite close, although a looser prior magnifies the credit expansion of the late 1990s somewhat. These observations suggest that thebaseline‘optimizedpriorhelpsidentifyingtheepisodesmorepreciselybutisnotsolelydrivingthemajorresults. 8. ConcludingRemarks ThispaperdevelopsanewmethodologytoidentifycreditboomsandbustsfromBVARforecasterrorsandtests theapproachfor31advancedandemergingmarketeconomies. Qualitatively,theresultsareintuitiveandfithistorical evidencewell. Quantitatively,thenewapproachcandetectturningpointsearlier,whichimpliesbetterearlywarning properties of these gaps ahead of crises or policy-relevant events. The VAR core of the methodology also allows to testhypothesesaboutthenatureoftheparticularcreditboomepisode, suchastheroleofmonetarypolicyrates, for example. Thereareseveralopenquestionsforfuturework. First,morehastobelearnedaboutthenatureofcreditbooms, thereasonsforcreditdeviationsfromvariousfundamentals. Thesemi-structuralexerciseinthispapersuggeststhat a monetary policy stance accounts for part of the explanation only, when it comes to the credit boom preceding the GlobalFinancialCrisis. Thecreditgapapproachinthispaperdoesnotmakefullystructuralidentifyingassumptions, by letting reduced-form evidence speak. A formal identification and joint structural modelling of the productive capacityoftheeconomyalongwiththefundamentallysustainablelevelofcreditcouldbeanextfruitfulstep. Second, while this paper focuses on the totality of nonfinancial credit to the private sector, a more detailed look into the decompositionofcreditcanbeausefulextension.Inthecontextofsomeemergingmarketsandperhapsalsoadvanced economies, accounting for the dynamics of government debt in the computation of credit gaps may lead to a fuller representation of financial vulnerabilities and credit cycle. Finally, the credit aggregates considered in this paper largelyleaveoutthefinancialintermediationthroughtheshadowbankingsystem–yetanotherimportantdimension ofcreditcyclethatneedstobeproperlyaccountedforgoingforward. 29

9. References Adrian, Tobias and Hyun Song Shin (2010). “Financial Intermediaries and Monetary Economics,” in: Benjamin M. Friedman&MichaelWoodford(eds.),HandbookofMonetaryEconomics,edition1,volume3,chapter12: 601– 650Elsevier. Adrian,Tobias,EmanuelMoenchandHyunSongShin(2010).“MacroRiskPremiumandIntermediaryBalanceSheet Quantities,”IMFEconomicReview58(1),179–207. Afanasyeva,Elena(2013).“AtypicalBehaviorofCredit: EvidencefromaMonetaryVAR,”IMFSWorkingpaperNo. 70,GoetheUniversityFrankfurt. Afanasyeva, Elena, Sam Jerow, Seung Jung Lee, and Michele Modugno (2020). “Sowing the Seeds of Financial Imbalances: TheRoleofMacroeconomicPerformance,”FEDSWorkingpaper2020-028. Afanasyeva, Elena, and Jochen Guentner (2020). “Bank Market Power and the Risk Channel of Monetary Policy,” JournalofMonetaryEconomics,forthcoming. Aliber,RobertZ.andKindelbergerCharlesP.(2015).“Manias,Panics,andCrashes: AHistoryofFinancialCrises,” SeventhEdition,PalgraveMacmillan,NewYork. Banbura,Marta,DomenicoGiannoneandLucreziaReichlin(2010).“LargeBayesianVectorAutoRegressions,”JournalofAppliedEconometrics25: 71–92. Baron, Matthew, Emil Werner, and Wei Xiong (2018). “Banking Crises Without Panics.” Working paper, Princeton mimeo. Bauwens, Luc, Gary Koop, Dimitris Korobilis, and Jeroen V.K. Rombouts (2015). “The Contribution of Structural BreakModelstoForecastingMacroeconomicSeries,”JournalofAppliedEconometrics30(4): 596–620. Bernanke, BenS.andCaraS.Lown(1991).“TheCreditCrunch,”BrookingsPapersonEconomicActivity2: 205– 239. Bordo,MichaelD.andJosephG.Haubrich(2010).“CreditCrises,MoneyContractions:AnHistoricalView,”Journal ofMonetaryEconomics,57(1): 1–18. Borio,ClaudioandPhilipLowe(2002).“AssetPrices,FinancialandMonetaryStability: ExploringtheNexus,”BIS WorkingPaperNo.114,July2002. Borio,ClaudioandHaibinZhu(2012).“CapitalRegulation,RiskTakingandMonetaryPolicy,”JournalofFinancial Stability8: 236–251. 30

Brunnermeier,MarkusK.andIsabelSchnabel(2016).“BubblesandCentralBanks:HistoricalPerspectives,”In”Central Banks at a Crossroads: What Can We Learn from History?”. Cambridge, UK: Cambridge University Press, 2016.Web. Buch,Claudia,SandraEickmeier,andEstebanPrieto(2014).“InSearchforYield? Survey-BasedEvidenceonBank RiskTaking,”JournalofEconomicDynamicsandControl43: 12–30. Bowman,David,FangCai,SallyDavies,andStevenKamin(2015).“QuantitativeEasingandBankLending:Evidence fromJapan,”JournalofInternationalMoneyandFinance,57: 15–30. Canova,Fabio(2006).“MethodsforAppliedMacroeconomicResearch.”PrincetonandOxford: PrincetonUniversity Press. Chib,Siddhartha(1995).“MarginalLikelihoodfromtheGibbsOutput,”JournaloftheAmericanStatisticalAssociation90(432): 1313–1321. Dell’Ariccia,Giovanni,DenizIgan,LucLaeven,andHuiTongwithBasBakkerandJeromeVandenbussche(2012). “PoliciesforMacrofinancialStability: HowtoDealwithCreditBooms,”IMFDiscussionNote,June2012. Drehmann,Mathias,ClaudioBorio,LeonardoGambacorta,GabrielJimenezandCarlosTrucharte(2010).“CountercyclicalCapitalBuffers: ExploringOptions,”BISWorkingPaperNo.317,July2010. Elliott, DouglasJ., GregFeldbergandAndreasLehnert(2013).“TheHistoryofCyclicalMacroprudentialPolicyin theUnitedStates,”FEDSWorkingPaperNo.2013-29. Edge, Rochelle M., and Ralf R. Meisenzahl (2011). “The Unreliability of Credit-to-GDP Ratio Gaps in Real Time: ImplicationsforCountercyclicalCapitalBuffers,”InternationalJournalofCentralBanking,Vol7(4): 261–299. Giannone,Domenico,MicheleLenzaandGiorgioPrimiceri(2015).“PriorSelectionforVectorAutoregressions,”The ReviewofEconomicsandStatistics97(2): 436–451. Giannone,Domenico,MicheleLenzaandLucreziaReichlin(2019).“Money,Credit,MonetaryPolicyandtheBusinessCycleintheEuroArea:WhatHasChangessincetheCrisis?”InternationalJournalofCentralBanking,15(5): 137–173. Giannone, Domenico, Michele Lenza, Huw Pill and Lucrezia Reichlin (2011). “Non-Standard Monetary Measures and Monetary Developments”, in: ”Lessons for Monetary Policy from the Financial Crisis,” eds. J. Chadha and S.Holly,CambridgeUniversityPress. Gorton,GaryandGuillermoOrdonez(2020).“GoodBooms,BadBooms,”JournaloftheEuropeanEconomicAssociation,forthcoming. 31

Gourinchas,Pierre-Olivier,RodrigoValdes,andOscarLanderretche(2001).”LendingBooms: LatinAmericaandthe World,”Economia,Spring: 47–99. Hamilton,JamesD.(2018).“WhyYouShouldNeverUsetheHodrick-PrescottFilter,”TheReviewofEconomicsand Statistics,100(5): 831–843. Ioannidou,Vasso,StevenOngena,andJose-LuisPeydro(2015).“MonetaryPolicy,Risk-TakingandPricing:Evidence fromaQuasi-NaturalExperiment,”ReviewofFinance,19(1): 95–144. Jarocinski, Marek and Bartosz Mackowiak (2017). “Granger Causal Priority and Choice of Variables in Vector Autoregressions,”TheReviewofEconomicsandStatistics99(2): 319–329. Jarocinski,MarekandFrankSmets(2008).“HousePricesandtheStanceofMonetaryPolicy,”ECBNo.891(April 2008). Jimenez,Gabriel,StevenOngena,Jose-LuisPeydro,andJesusSaurina(2014).“HazardousTimesforMonetaryPolicy:WhatDoTwenty-ThreeMillionBankLoansSayabouttheEffectsofMonetaryPolicyonCreditRisk-Taking?,” Econometrica82(2): 463–505. Jorda,Oscar,MoritzSchularickandAlanM.Taylor(2013).“WhenCreditBitesBack”JournalofMoney,Creditand Banking45(2): 3–28. Krishnamurthy, Arvind and Wenhao Li (2020). “Dissecting Mechanisms of Financial Crises: Intermediation and Sentiment.”WorkingPaper,StanfordGraduateSchoolofBusiness,mimeo. Laeven,LucandFabianValencia(2013).“SystemicBankingCrisesDatabase.”IMFEconomicReview61(2): 225– 270. Lo Duca, Marco, Anne Koban, Marisa Basten, Elias Bengtsson, Benjamin Klaus, Piotr Kusmierczyk, Jan Hannes Lang,CarstenDetken(ed.),andTuomasPeltonen(ed.)(2017).“ANewDatabaseforFinancialCrisesinEuropean Countries,”ESRBOccasionalPaperSeriesNo.13,July2017. Mendoza,EnriqueG.andMarcoE.Terrones(2008).“AnAnatomyofCreditBooms: EvidenceFromMacroAggregatesandMicroData,”NBERworkingpaper14049. Minsky,HymanP.(1986).“StabilizingtheUnstableEconomy,”YaleUniversityPress,NewHavenandLondon. Schularick, Moritz and Alan M. Taylor (2012). “Credit Booms Gone Bust: Monetary Policy, Leverage Cycles, and FinancialCrises,1870-2008,”AmericanEconomicReview102(2): 1029–1061. 32

Shiller,RobertJ.(2008).“UnderstandingRecentTrendsinHousePricesandHomeOwnership,”inHousing,Housing FinanceandMonetaryPolicy,JacksonHoleConferenceSeries,FederalReserveBankofKansasCity: 85–123. Shiratsuka, Shigenori (2005). “The Asset Price Bubble in Japan in the 1980s: Lessons for Financial and Macroeconomic Stability,” BIS Papers chapters, in: Bank for International Settlements (ed.), Real estate indicators and financialstability,volume21,pages42-62BankforInternationalSettlements. Sims,ChristopherA.andTaoZha(1998).“BayesianMethodsforDynamicMultivariateModels,”InternationalEconomicReview39(4): 949–968. Taylor, John B. (2007). “Housing and Monetary Policy,” panel intervention in the Annual Economic Symposium organizedbytheKansasCityFedinJacksonHole. Villani, Mattias (2009). “Steady-State Priors for Vector Autoregressions,” Journal of Applied Econometrics 24(4): 630–650. Waggoner,DanielF.andTaoZha(1999).“ConditionalForecastsinDynamicMultivariateModels,” ReviewofEconomicsandStatistics81(4): 639–651. Wu, Jing C. and Fan D. Xia (2016). “Measuring the Macroeconomic Impact of Monetary Policy at the Zero Lower Bound,”JournalofMoney,CreditandBanking48(2-3): 253–291. 33

AppendixA. Data TableA.1:DataandTransformationsUsedintheBaselineMonetaryVARModelsandCredit-to-GDPGapsacrossCountries. No. SeriesName Retrievedfrom Transformationa Seasonal Adjustmentb Australia MonetaryVAR:1978M2-2012M12 1 UnemploymentRate FRED Level yes 2 ConsumerPriceIndex IFS Log-Levela yes 3 CentralBankPolicyRate BIS Level no 4 MSCIAustralia,NationalCurrency MSCI Log-Level no 5 M1MonetaryAggregate,NationalCurrency FRED Log-Level yes 6 M3MonetaryAggregate,NationalCurrency FRED Log-Level yes 7 Deposit Money Banks, Claims on Private Sector, IFS Log-Level yes NationalCurrency Credit-to-GDPGap:1978Q2-2012Q4 8 GDP,CurrentPrices,NationalCurrency FRED Levelc yes Austria MonetaryVAR:1997M10-2017M12 1 IndustrialProductionIndex FRED Log-Level yes 2 ConsumerPriceIndex FRED Log-Level yes 3 3-MonthInterbankRate FRED Level no 4 MSCIAustria,NationalCurrencyd MSCI Log-Level no 5 M1MonetaryAggregate,NationalCurrencyd Oesterreichische Log-Level yes Nationalbank 6 M2MonetaryAggregate,NationalCurrencyd Oesterreichische Log-Level yes Nationalbank 7 TotalCredittoPrivateNon-FinancialSectorbyDo- Oesterreichische Log-Level yes mesticBanks,NationalCurrencyd Nationalbank Credit-to-GDPGap:1997Q4-2017Q4 8 GDP,CurrentPrices,NationalCurrencyd IFS Levelc yes Belgium MonetaryVAR:1996M12-2017M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 3-MonthInterbankRate FRED Level no 4 MSCIBelgium,NationalCurrencyd MSCI Log-Level no 5 M1MonetaryAggregate,NationalCurrencyd NationalBankof Log-Level yes Belgium 6 M2MonetaryAggregate,NationalCurrencyd NationalBankof Log-Level yes Belgium Continuedonnextpage 34

TableA.1–Continuedfrompreviouspage No. SeriesName Retrievedfrom Transformationa Seasonal Adjustmentb 7 Total Credit to Private Non-Financial Sector by FRED Log-Levelaa yes Domestic Banks Adjusted for Breaks, National Currencyd Credit-to-GDPGap:1996Q4-2017Q4 8 GDP,CurrentPrices,NationalCurrencyd FRED Levelc yes Brazil MonetaryVAR:1995M1-2017M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 CentralBankPolicyRate BIS Level no 4 YahooIBOVESPA YahooFinance Log-Level no 5 ExchangeRaterelativetotheUSD FRED Log-Level no 6 M3MonetaryAggregate,NationalCurrency FRED Log-Level yes 7 TotalCredittoPrivateNon-FinancialSectorbyDo- FRED Log-Levela yes mestic Banks Adjusted for Breaks, National Currency Credit-to-GDPGap:1996Q1-2017Q4 8 GDP,CurrentPrices,NationalCurrency FRED Levelc yes Canada MonetaryVAR:1970M1-2017M12 1 IndustrialProductionIndex FRED Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 CentralBankPolicyRate BIS Level no 4 MSCICanada,NationalCurrency MSCI Log-Level no 5 M1MonetaryAggregate,NationalCurrency FRED Log-Level yes 6 M3MonetaryAggregate,NationalCurrency FRED Log-Level yes 7 TotalCredittoPrivateNon-FinancialSectorbyDo- FRED Log-Levela yes mestic Banks Adjusted for Breaks, National Currency Credit-to-GDPGap:1970Q1-2017Q4 8 GDP,CurrentPrices,NationalCurrency IFS Levelc yes Chile MonetaryVAR:1987M12-2008M12 1 UnemploymentRate(populationaged15yearsand IFS Level yes older) 2 ConsumerPriceIndex IFS Log-Level yes 3 ExchangeRaterelativetotheUSD FRED Log-Level no 4 MSCIChile,NationalCurrency MSCI Log-Level no Continuedonnextpage 35

TableA.1–Continuedfrompreviouspage No. SeriesName Retrievedfrom Transformationa Seasonal Adjustmentb 5 M1MonetaryAggregate,NationalCurrency FRED Log-Level yes 6 M3MonetaryAggregate,NationalCurrency FRED Log-Level yes 7 DepositMoneyBanks: ClaimsonthePrivateSec- IFS Log-Level yes tor,NationalCurrency Credit-to-GDPGap:1996Q1-2008Q4 8 GDP,CurrentPrices,NationalCurrency FRED Levelc yes CzechRepublic MonetaryVAR:1994M6-2017M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 CommercialBankLendingRateoverDepositRate IFS Level no Spread 4 SharePriceIndex Bloomberg Log-Level no 5 M1MonetaryAggregate,NationalCurrency FRED Log-Level yes 6 M3MonetaryAggregate,NationalCurrency FRED Log-Level yes 7 TotalCredittoPrivateNon-FinancialSectorbyDo- FRED Log-Levela yes mestic Banks Adjusted for Breaks, National Currency Credit-to-GDPGap:1995Q1-2017Q4 8 GDP,CurrentPrices,NationalCurrency FRED Levelc yes Finland MonetaryVAR:1996M1-2016M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 3-MonthMoneyMarketRate IFS Level no 4 IndustrialSharePriceIndexd IFS Log-Level no 5 Finnish Contribution to the M1 Monetary Aggre- BankofFinland Log-Level yes gate,NationalCurrencyd 6 Finnish Contribution to the M2 Monetary Aggre- BankofFinland Log-Level yes gate,NationalCurrencyd 7 CredittoPrivateNon-FinancialSectorbyDomestic FRED Log-Levela yes BanksAdjustedforBreaks,NationalCurrencyd Credit-to-GDPGap:1996Q1-2016Q4 8 GDP,CurrentPrices,NationalCurrencyd IFS Levelc yes France MonetaryVAR:1993M4-2017M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes Continuedonnextpage 36

TableA.1–Continuedfrompreviouspage No. SeriesName Retrievedfrom Transformationa Seasonal Adjustmentb 3 CentralBankDiscountRate FRBOG/IF Level no 4 MSCIFrance,NationalCurrencyd MSCI Log-Level no 5 French Contribution to the M1 Monetary Aggre- BanquedeFrance Log-Level yes gate,NationalCurrencyd 6 French Contribution to the M3 Monetary Aggre- BanquedeFrance Log-Level yes gate,NationalCurrencyd 7 CredittoPrivateNon-FinancialSectorbyDomestic FRED Log-Levela yes BanksAdjustedforBreaks,NationalCurrencyd Credit-to-GDPGap:1993Q2-2017Q4 8 GDP,CurrentPrices,NationalCurrency(converted IFS Levelc yes fromUSDtoEUR) Germany MonetaryVAR:1970M1-1998M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 DiscountRate FRBOG/IF Level no 4 MSCIGermany MSCI Log-Level no 5 M1MonetaryAggregate,NationalCurrency IFS Log-Level yes 6 M2MonetaryAggregate,NationalCurrency IFS Log-Level yes 7 BankingInstitutionsClaimsonPrivateSector,Na- IFS Log-Level yes tionalCurrency Credit-to-GDPGap:1979Q1-1998Q4 8 GDP,CurrentPrices,NationalCurrency(converted FRED Levelc yes toDMfromEUR) HongKong MonetaryVAR:1997M1-2017M12 1 UnemploymentRate IFS Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 DiscountRate IFS Level no 4 HangSengIndex YahooFinance Log-Level no 5 ExchangeRaterelativetotheUSD FRED Log-Level no 6 M2MonetaryAggregate,NationalCurrency HongKongCensusand Log-Level yes StatisticsDepartment 7 Total Credit to Private Nonfinancial Sector, Ad- FRED Log-Levela yes justedforBreaks,NationalCurrency Credit-to-GDPGap:1997Q2-2017Q4 8 GDP,CurrentPrices,NationalCurrency IFS Levelc yes Continuedonnextpage 37

TableA.1–Continuedfrompreviouspage No. SeriesName Retrievedfrom Transformationa Seasonal Adjustmentb Ireland MonetaryVAR:1991M1-2017M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 3-MonthInterbankRate FRED Level no 4 YahooISEQd YahooFinance Log-Level no 5 ContributiontoM1MonetaryAggrgeate,National CentralStatistics Log-Level yes Currencyd Office:Ireland 6 ContributiontoM3MonetaryAggregate,National CentralStatistics Log-Level yes Currencyd Office:Ireland 7 Total Credit to Private Nonfinancial Sector, Ad- FRED Log-Levela yes justedforBreaks,NationalCurrencyd Credit-to-GDPGap:1991Q1-2017Q4 8 GDP,CurrentPrices,NationalCurrencyd IFS Levelc yes Italy MonetaryVAR:1999M1-2017M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 3-MonthInterbankRate FRED Level no 4 FTSEMIBIndexd Bloomberg Log-Level no 5 Italian Contribution to M1 Monetary Aggrgeate, BankofItaly Log-Level yes NationalCurrencyd 6 Italian Contribution to M3 Monetary Aggregate, BankofItaly Log-Level yes NationalCurrencyd 7 Total Credit to Private Nonfinancial Sector, Ad- FRED Log-Levela yes justedforBreaks,NationalCurrencyd Credit-to-GDPGap:1999Q1-2017Q4 8 GDP,CurrentPrices,NationalCurrencyd IFS Levelc yes Japan MonetaryVAR:1970M1-2016M12 1 IndustrialProductionIndex FRED Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 DiscountRate FRED Level no 4 MSCIJapan,NationalCurrency MSCI Log-Level no 5 M1MonetaryAggrgeate,NationalCurrency IFS Log-Level yes 6 M2MonetaryAggregate,NationalCurrency IFS Log-Level yes 7 Total Credit to Private Nonfinancial Sector, Ad- FRED Log-Levela yes justedforBreaks,NationalCurrency Credit-to-GDPGap:1970Q1-2016Q4 Continuedonnextpage 38

TableA.1–Continuedfrompreviouspage No. SeriesName Retrievedfrom Transformationa Seasonal Adjustmentb 8 GDP,CurrentPrices,NationalCurrency IFS Levelc yes Korea MonetaryVAR:1978M1-2012M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 DiscountRate IFS Level no 4 MSCIKorea,NationalCurrency MSCI Log-Level no 5 ExchangeRaterelativetotheUSD FRED Log-Level no 6 M2MonetaryAggregate,NationalCurrency FRED Log-Level yes 7 Deposit Money Banks, Claims on Private Sector, IFS Log-Level yes NationalCurrency Credit-to-GDPGap:1978Q1-2012Q3 8 GDP,CurrentPrices,NationalCurrency FRED Levelc yes Luxembourg MonetaryVAR:1999M1-2017M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 3-MonthInterbankRate FRED Level no 4 LuxembourgStockExchangeIndexd Bloomberg Log-Level no 5 M1Contribution,NationalCurrencyd BanqueCentraledu Log-Level yes Luxembourg 6 M2Contribution,NationalCurrencyd BanqueCentraledu Log-Level yes Luxembourg 7 Total Credit to Private Nonfinancial Sector, Ad- FRED Log-Levela yes justedforBreaks,NationalCurrencyd Credit-to-GDPGap:1999Q1-2017Q4 8 GDP,CurrentPrices,NationalCurrencyd IFS Levelc yes Malaysia MonetaryVAR:1981M1-2016M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 Base Lending Rate over 3-Month Time Deposit IFS Level no Rate 4 MSCIMalaysia,NationalCurrency MSCI Log-Level no 5 M1MonetaryAggregate,NationalCurrency IFS Log-Level yes 6 M2MonetaryAggregate,NationalCurrency IFS Log-Level yes 7 CredittoPrivateNonfinancialSectorbyDomestic FRED Log-Levela yes Banks,AdjustedforBreaks,NationalCurrency Continuedonnextpage 39

TableA.1–Continuedfrompreviouspage No. SeriesName Retrievedfrom Transformationa Seasonal Adjustmentb Credit-to-GDPGap:1991Q1-2016Q4 8 GDP,CurrentPrices,NationalCurrency IFS Levelc yes Mexico MonetaryVAR:1978M1-2008M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 CentralBankPolicyRate BIS Level no 4 TotalSharePriceIndexforMexico FRED Log-Level no 5 ExchangeRaterelativetotheUSD FRED Log-Level no 6 M3MonetaryAggregate,NationalCurrency FRED Log-Level yes 7 Deposit Money Banks, Claims on Private Sector, IFS Log-Level yes NationalCurrency Credit-to-GDPGap:1993Q1-2008Q4 8 GDP,CurrentPrices,NationalCurrency FRED Levelc yes NewZealand MonetaryVAR:1988M9-2010M12 1 Unemployment(inthsd.persons) FRED Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 CentralBankPolicyRate BIS Level no 4 MSCINewZealand,NationalCurrency MSCI Log-Level no 5 M1MonetaryAggregate,NationalCurrency FRED Log-Level yes 6 M3MonetaryAggregate,NationalCurrency FRED Log-Level yes 7 Deposit Money Banks, Claims on Private Sector, IFS Log-Level yes NationalCurrency Credit-to-GDPGap:1989Q1-2010Q4 8 GDP,CurrentPrices,NationalCurrency IFS Levelc yes Norway MonetaryVAR:1979M1-2014M12 1 IndustrialProductionIndex FRED Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 3-MonthInterbankRate FRED Level no 4 MSCINorway,NationalCurrency MSCI Log-Level no 5 ExchangeRaterelativetotheUSD FRED Log-Level no 6 M2MonetaryAggregate,NationalCurrency IFS Log-Level yes 7 Total Credit to Private Nonfinancial Corporations, BIS Log-Levela yes AdjustedforBreaks,NationalCurrency Credit-to-GDPGap:1979Q1-2014Q4 8 GDP,CurrentPrices,NationalCurrency FRED Levelc yes Continuedonnextpage 40

TableA.1–Continuedfrompreviouspage No. SeriesName Retrievedfrom Transformationa Seasonal Adjustmentb Poland MonetaryVAR:1997M1-2016M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 3-MonthInterbankRate FRED Level no 4 WIGIndex Bloomberg Log-Level no 5 ExchangeRaterelativetotheUSD FRED Log-Level no 6 M2MonetaryAggregate,NationalCurrency IFS Log-Level yes 7 Total Credit to Private Nonfinancial Corporations, FRED Log-Levela yes AdjustedforBreaks,NationalCurrency Credit-to-GDPGap:1997Q1-2016Q4 8 GDP,CurrentPrices,NationalCurrency IFS Levelc yes Portugal MonetaryVAR:1995M1-2017M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 3-MonthInterbankRate FRED Level no 4 PSI20Indexd Bloomberg Log-Level no 5 M1PortugueseContribution,NationalCurrencyd BankofPortugal Log-Level yes 6 M2PortugueseContribution,NationalCurrencyd BankofPortugal Log-Level yes 7 Total Credit to Private Nonfinancial Corporations, FRED Log-Levela yes AdjustedforBreaks,NationalCurrencyd Credit-to-GDPGap:1995Q1-2017Q4 8 GDP,CurrentPrices,NationalCurrencyd IFS Levelc yes Singapore MonetaryVAR:1991M1-2017M12 1 ManufacturingProductionIndex IFS Log-Level* yes 2 ConsumerPriceIndex IFS Log-Level yes 3 3-MonthDepositRate IFS Level no 4 MSCISingapore,NationalCurrency MSCI Log-Level no 5 ExchangeRaterelativetotheUSD FRED Log-Level no 6 M2MonetaryAggregate,NationalCurrency IFS Log-Level yes 7 Deposit Money Banks, Claims on Private Sector, IFS Log-Level yes NationalCurrency Credit-to-GDPGap:1991Q1-2017Q2 8 GDP,CurrentPrices,NationalCurrency IFS Levelc yes Continuedonnextpage 41

TableA.1–Continuedfrompreviouspage No. SeriesName Retrievedfrom Transformationa Seasonal Adjustmentb SlovakRepublic MonetaryVAR:1997M11-2017M12 1 RegisteredUnemploymentRate FRED Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 3-MonthInterbankRate FRED Level no 4 SKSMIndex Bloomberg Log-Level no 5 M1MonetaryAggregate,NationalCurrency SlovakNationalBank Log-Level yes 6 M3MonetaryAggregate,NationalCurrency SlovakNationalBank Log-Level yes 7 LoanstoHouseholdsandNon-ProfitOrganizations, SlovakNationalBank Log-Level yes NationalCurrency Credit-to-GDPGap:1997Q4-2017Q4 8 GDP,CurrentPrices,NationalCurrency FRED Levelc yes SouthAfrica MonetaryVAR:1974M6-2012M12 1 BusinessConfidenceIndex FRED Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 DiscountRate FRED Level no 4 SharePriceIndex:IndustryandCommerce IFS Log-Level no 5 ExchangeRaterelativetotheUSD FRED Log-Level no 6 M2MonetaryAggregate,NationalCurrency IFS Log-Level yes 7 Total Credit to Private Non-Financial Sector Ad- BIS Log-Levela yes justedforBreaks,NationalCurrency Credit-to-GDPGap:1974Q3-2012Q4 8 GDP,CurrentPrices,NationalCurrency FRED Levelc yes Spain MonetaryVAR:1993M1-2017M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 Short-TermInterbankRate FRED Level no 4 MSCISpain,NationalCurrencyd MSCI Log-Level no 5 Long-TermGovernmentBondYield FRED Level no 6 M2 Monetary Aggregate, Spain Contribution Na- CentralBankofSpain Log-Level yes tionalCurrencyd 7 Total Credit to Private Non-Financial Sector Ad- BIS Log-Levela yes justedforBreaks,NationalCurrencyd Credit-to-GDPGap:1995Q1-2017Q4 8 GDP,CurrentPrices,NationalCurrencyd FRED Levelc yes Continuedonnextpage 42

TableA.1–Continuedfrompreviouspage No. SeriesName Retrievedfrom Transformationa Seasonal Adjustmentb Sweden MonetaryVAR:1971M1-2016M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 CentralBankPolicyRate BIS Level no 4 MSCISweden,NationalCurrency MSCI Log-Level no 5 ExchangeRaterelativetotheUSD IFS Log-Level no 6 M3Index FRED Log-Level yes 7 Total Credit to Private Non-Financial Sector Ad- BIS Log-Levela yes justedforBreaks,NationalCurrency Credit-to-GDPGap:1980Q1-2016Q4 8 GDP,CurrentPrices,NationalCurrency IFS Levelc yes Switzerland MonetaryVAR:1985M1-2016M12 1 RegisteredUnemploymentRate FRED Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 CentralBankPolicyRate BIS Level no 4 MSCISwitzerland,NationalCurrency MSCI Log-Level no 5 M1MonetaryAggregate,NationalCurrency IFS Log-Level yes 6 M2MonetaryAggregate,NationalCurrency IFS Log-Level yes 7 Deposit Money Banks: Claims on Private Sector, IFS Log-Levela yes NationalCurrency Credit-to-GDPGap:1985Q1-2016Q4 8 GDP,CurrentPrices,NationalCurrency IFS Levelc yes Turkey MonetaryVAR:1988M1-2008M12 1 IndustrialProductionIndex IFS Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 DiscountRate FRED Level no 4 MSCITurke,NationalCurrency MSCI Log-Level no 5 ExchangeRateagainsttheUSD FRED Log-Level no 6 M2MonetaryAggregate,NationalCurrency IFS Log-Level yes 7 Deposit Money Banks: Claims on Private Sector, IFS Log-Levela yes NationalCurrency Credit-to-GDPGap:1998Q1-2008Q4 8 GDP,CurrentPrices,NationalCurrency FRED Levelc yes Continuedonnextpage 43

TableA.1–Continuedfrompreviouspage No. SeriesName Retrievedfrom Transformationa Seasonal Adjustmentb UnitedKingdom MonetaryVAR:1987M1-2016M12 1 IndustrialProductionIndex FRED Log-Level yes 2 ConsumerPriceIndex IFS Log-Level yes 3 CentralBankPolicyRate BIS Level no 4 MSCIUK,NationalCurrency MSCI Log-Level no 5 M1MonetaryAggregate,NationalCurrency FRED Log-Level yes 6 M2MonetaryAggregat,NationalCurrencye FRED Log-Level yes 7 Deposit Money Banks: Claims on Private Sector, IFS Log-Levela yes NationalCurrency Credit-to-GDPGap:1987Q1-2016Q4 8 GDP,CurrentPrices,NationalCurrency IFS Levelc yes UnitedStatesofAmerica MonetaryVAR:1959M1-2018M12 1 IndustrialProductionIndex FRED Log-Level yes 2 Consumer Price Index for All Urban Consumers: FRED Log-Level yes AllItems 3 EffectiveFederalFundsRate FRED Level no 4 S&P500 R.ShillerDataBase Log-Level no 5 M1MonetaryAggregate,NationalCurrency FRED Log-Level yes 6 M2MonetaryAggregate,NationalCurrency FRED Log-Level yes 7 Total Credit to the Nonfinancial Sector, National FlowofFunds Log-Levela yes Currency Credit-to-GDPGap:1959Q1-2018Q4 8 GDP,CurrentPrices,NationalCurrency FRED Levelc yes aIf marked, the time series is originally available at quarterly frequency and is interpolated to monthly frequency with cubic splines. bIfthecorrespondingmacroeconomictimeseriesisnotseasonallyadjustedintheoriginalsource,Iremoveseasonalitywiththe X12procedureoftheCensusBureau.Financialtimeseriesarenotseasonallyadjusted. cToconstructthecredit-to-GDPgaps,Iformthecredit-to-GDPratiousingthesamecreditmeasureasinmonetaryBVAR.This ratioisthenloggedandHP-filteredwithasmoothingparameterλ=400000. dNationalcurrencyreferstotheeuroforthesecountries. DataSources Banca d’Italia Monetary Aggregates statistics, https://infostat.bancaditalia.it/inquiry/home?spyglass/taxo:CUBESET =/PUBBL 00/PUBBL 00 02 01 04/PUBBL 00 02 01 04 03&ITEMSELEZ=AGGM0300—true&OPEN=false /&ep:LC=EN&COMM=BANKITALIA&ENV=LIVE&CTX=DIFF&IDX=2&/view:CUBEIDS=/&timestamp= 1582823484790. 44

BancodeEspan˜a: MonetaryAggregatesstatistics,https://www.bde.es/webbde/en/estadis/infoest/bolest1.html BancodePortugal: MonetaryAggregatesstatistics,https://www.bportugal.pt/en/page/estatisticas. BankofFinlandStatistics,https://www.suomenpankki.fi/en/Statistics/mfi-balance-sheet/. BanqueCentraleduLuxembourg,Statistics,http://www.bcl.lu/en/statistics/series statistiques luxembourg/index.html. BanquedeFranceMonetaryStatistics,http://webstat.banque-france.fr/en/browse.do?node=5384975. Bank of International Settlements (BIS): credit to the non-financial sector database, https://www.bis.org/statistics/totcredit.htm BankofInternationalSettlements(BIS):centralbankpolicyrates,https://www.bis.org/statistics/cbpol.htm. BloombergDatabase,retrievedfromtheterminalattheBoardofGovernorsoftheFederalReserveBoard,Washington, D.C. CentralStatisticsOffice: Ireland,https://www.cso.ie/en/statistics/keyeconomicindicators/. FederalReserveBankofSt.Louis,FRED,https://fred.stlouisfed.org. FinancialAccountsoftheUnitedStates,FlowofFunds,https://www.federalreserve.gov/releases/z1/. HongKongCensusandStatisticsDepartment,https://www.censtatd.gov.hk/hkstat/. International Financial Statistics (IFS), IMF Data Set, https://data.imf.org/?sk=4C514D48-B6BA-49ED-8AB9- 52B0C1A0179B.AccessedfromtheBoardofGovernorsoftheFederalReservedSystem. MorganStanleyCapitalInternational(MSCI)SharePriceIndexes,https://finance.yahoo.com/quote/MSCI/history/. NationalBankofBelgium,statisticsonmonetaryaggregates: https://stat.nbb.be/index.aspx?queryid=90. Oesterreichische Nationalbank, statistics on monetary aggregates: https://www.oenb.at/en/Statistics/Standardized- Tables/OeNB–Eurosystem-and-Monetary-Indicators-/Monetary-Aggregates–Consolidated-MFI-Balance-Sheet- .html. OnlinedataarchiveupdatedandmaintainedbyRobertShiller,www.econ.yale.edu/shiller/data.htm. Slovak National Bank, monetary aggregates statistics, https://www.nbs.sk/en/monetary-policy/macroeconomicdatabase/macroeconomic-database-chart. YahooFinanceStockPriceIndexData,https://finance.yahoo.com/quote/YHOO/history?p=YHOO. 45

AppendixB. BVARCreditGapsandTwo-SidedCredit-to-GDPGaps 46

muigleB)c( airtsuA)b( ailartsuA)a( elihC)f( adanaC)e( lizarB)d( ecnarF)i( dnalniF)h( cilbupeRhcezC)g( ecnarF-ailartsuA:seirtnuoCssorcaspaGPDG-ot-tiderCdediS-owTdnaRAVB:1.BerugiF :seigolonorhcgniwollofehtotgnidroccasesircfotesnoehtyfingissenillacitreV.stniopegatnecrepnidesserpxeeraspagtiderC.sretrauqnisisixalatnozirohehtnoemiT:setoN tnevelaudiserasetacidni∗,ygolonorhcDLfoesacnI.)7102(.lateacuDoL-DL,)3102(aicnelaVdnaneveaL-VL,)9102(.latenoraB-XVB,)2102(rolyaTdnakciraluhcS-TS .)6102(lebanhcSdnareiemrennurBrofsdnatsSBdna)5102(regrebeldniKdnarebilArofsdnatsZA,ailartsuAroF.)sisirccimetsysanahtrehtar( 47

dnalerI)c( gnoKgnoH)b( ynamreG)a( aeroKhtuoS)f( napaJ)e( ylatI)d( ocixeM)i( aisyalaM)h( gruobmexuL)g( ocixeM-ynamreG:seirtnuoCssorcaspaGPDG-ot-tiderCdediS-owTdnaRAVB:2.BerugiF :seigolonorhcgniwollofehtotgnidroccasesircfotesnoehtyfingissenillacitreV.stniopegatnecrepnidesserpxeeraspagtiderC.sretrauqnisisixalatnozirohehtnoemiT:setoN tnevelaudiserasetacidni∗,ygolonorhcDLfoesacnI.)7102(.lateacuDoL-DL,)3102(aicnelaVdnaneveaL-VL,)9102(.latenoraB-XVB,)2102(rolyaTdnakciraluhcS-TS .)sisirccimetsysanahtrehtar( 48

dnaloP)c( yawroN)b( dnalaeZweN)a( cilbupeRkavolS)f( eropagniS)e( lagutroP)d( nedewS)i( niapS)h( acirfAhtuoS)g( nedewS-dnalaeZweN:seirtnuoCssorcaspaGPDG-ot-tiderCdediS-owTdnaRAVB:3.BerugiF :seigolonorhcgniwollofehtotgnidroccasesircfotesnoehtyfingissenillacitreV.stniopegatnecrepnidesserpxeeraspagtiderC.sretrauqnisisixalatnozirohehtnoemiT:setoN tnevelaudiserasetacidni∗,ygolonorhcDLfoesacnI.)7102(.lateacuDoL-DL,)3102(aicnelaVdnaneveaL-VL,)9102(.latenoraB-XVB,)2102(rolyaTdnakciraluhcS-TS .)sisirccimetsysanahtrehtar( 49

yekruT)b( dnalreztiwS)a( aciremAfosetatSdetinU)d( modgniKdetinU)c( ASU-dnalreztiwS:seirtnuoCssorcaspaGPDG-ot-tiderCdediS-enOdnaRAVB:4.BerugiF :seigolonorhcgniwollofehtotgnidroccasesircfotesnoehtyfingissenillacitreV.stniopegatnecrepnidesserpxeeraspagtiderC.sretrauqnisisixalatnozirohehtnoemiT:setoN tnevelaudiserasetacidni∗,ygolonorhcDLfoesacnI.)7102(.lateacuDoL-DL,)3102(aicnelaVdnaneveaL-VL,)9102(.latenoraB-XVB,)2102(rolyaTdnakciraluhcS-TS .)sisirccimetsysanahtrehtar( 50

AppendixC. Real-TimeDataSetandTiming: U.S.Example InthebaselineVAR,industrialproduction(IP),consumerpriceindex(CPI),bothmonetaryaggregatesaswellas thetotalnonfinancialcreditaggregatearesubjecttorevisionsinrealtime. TherevisionsaredisplayedinFiguresC.1 - C.5. I use the Archival Federal Reserve Economic Data (ALFRED) database to retrieve the real-time vintages for IP,CPI,M1andM2. ALFREDalsoprovidesamorerecentportionofvintagesofthetotalnonfinancialcreditseries produced by the Flow of Funds. These vintages start in 2014. In addition, I use internal databases of the Federal Reserve Board to retrieve vintages of the credit variable starting from 1996. The Federal Funds Rate and the stock marketpricesarenotrevised. FigureC.1:Year-on-YearGrowth(Percent)oftheIndustrialProductionIndexacrossVariousDataVintages. Quantitatively,therevisionsaremostsizeableforrealactivityvariablesandfortheFlowofFundscreditaggregate. Therevisionsstemfromupdatesinaggregatecalculationsaswellasdefinitionalchanges.FortheFlowofFundscredit aggregates,definitionalchangescanleadtoquitesizeablerevisionsinsomequarters,forinstance,whenthecoverage of financial institutions expands. Flow of Funds publishes many of the changes in definitions and computational methodologyoftheiraggregatevariableshere: https://www.federalreserve.gov/apps/fof/FOFHighlight. aspx. Asforthetiming, FlowofFundscreditaggregatesarereleasedatquarterlyfrequencyandwithadelayofabout one quarter. For instance, the vintage covering the data through the fourth quarter of a respective year, is typically releasedinthesecondweekofMarchofthenextyear. AllothervariablesintheVARareupdatedsoonerandoften athigherfrequencythanthecreditvariable. Inchoosingthecorrespondingreal-timevintageofthosevariables,Ipick thevintagereleasedclosesttothecorrespondingdateoftheFlowofFundsrelease. Thatway,allvariablesatacertain pointintimewouldadequatelyreflectthe(policymakers)informationsetatthattime. 51

FigureC.2:Year-on-YearGrowth(Percent)oftheConsumerPriceIndexacrossVariousDataVintages. FigureC.3:Year-on-YearGrowth(Percent)oftheM1MonetaryAggregateacrossVariousDataVintages. 52

FigureC.4:Year-on-YearGrowthoftheM2MonetaryAggregateacrossVariousDataVintages. Notes:Timeisinmonths.Growthratesareinpercent. FigureC.5:LevelsandYear-on-YearGrowthRatesoftheTotalNonfinancialCreditacrossVariousDataVintages. Notes:Timeisinmonths.Creditvolumeisexpressedinlevels(BillionsofUSD)intheleftpanelandinyear-on-yeargrowthrates inpercentintherightpanel. 53

AppendixD. CountriesinBurns-MitchellDiagramsandPooledLogitRegressionsbyChronology Thecountriesbelowhavecriseswithintheirsamplestimedaccordingtothecorrespondingchronologyandhave atleastfouryearsofcomputedcreditgapsprecedingthecrisisonset. Theresultingselectionisillustartedinthetable below. Country LVChronology BVXChronology STChronology LDChronology Austria x x Australia x x Belgium x x x x CzechRepublic x Finland x France x x x x Germany x Japan x xx x Malaysia x x Norway xx Poland x Portugal x x x x SlovakRepublic x SouthKorea x x Spain x x x x Sweden xx xx xx xx Switzerland x x x UK x x x x USA xx xx xx Notes: LVreferstoLaevenandValencia(2013),BVXreferstoBaronetal. (2018),STreferstoSchularickandTaylor(2012), andLDreferstoLoDucaetal. (2017). ’x’indicatesacrisisepisode,’xx’indicatestwoseparatecrisisepisodesthatsatisfythe selectioncriteria. 54

AppendixE. BVAR-BasedGapsforBroadMoneyandAssetPrices 55

muigleB)c( airtsuA)b( ailartsuA)a( elihC)f( adanaC)e( lizarB)d( ecnarF)i( dnalniF)h( cilbupeRhcezC)g( ecnarF-ailartsuA:seirtnuoCssorcasecirPtessAdnayenoMdaorBrofspaGRAVB:1.EerugiF :seigolonorhcgniwollofehtotgnidroccasesircfotesnoehtyfingissenillacitreV.stniopegatnecrepnidesserpxeeraspagtiderC.sretrauqnisisixalatnozirohehtnoemiT:setoN tnevelaudiserasetacidni∗,ygolonorhcDLfoesacnI.)7102(.lateacuDoL-DL,)3102(aicnelaVdnaneveaL-VL,)9102(.latenoraB-XVB,)2102(rolyaTdnakciraluhcS-TS .)6102(lebanhcSdnareiemrennurBrofsdnatsSBdna)5102(regrebeldniKdnarebilArofsdnatsZA,ailartsuAroF.)sisirccimetsysanahtrehtar( 56

dnalerI)c( gnoKgnoH)b( ynamreG)a( aeroKhtuoS)f( napaJ)e( ylatI)d( ocixeM)i( aisyalaM)h( gruobmexuL)g( ocixeM-ynamreG:seirtnuoCssorcasecirPtessAdnayenoMdaorBrofspaGRAVB:2.EerugiF :seigolonorhcgniwollofehtotgnidroccasesircfotesnoehtyfingissenillacitreV.stniopegatnecrepnidesserpxeeraspagtiderC.sretrauqnisisixalatnozirohehtnoemiT:setoN tnevelaudiserasetacidni∗,ygolonorhcDLfoesacnI.)7102(.lateacuDoL-DL,)3102(aicnelaVdnaneveaL-VL,)9102(.latenoraB-XVB,)2102(rolyaTdnakciraluhcS-TS .)sisirccimetsysanahtrehtar( 57

dnaloP)c( yawroN)b( dnalaeZweN)a( cilbupeRkavolS)f( eropagniS)e( lagutroP)d( nedewS)i( niapS)h( acirfAhtuoS)g( nedewS-dnalaeZweN:seirtnuoCssorcasecirPtessAdnayenoMdaorBrofspaGRAVB:3.EerugiF :seigolonorhcgniwollofehtotgnidroccasesircfotesnoehtyfingissenillacitreV.stniopegatnecrepnidesserpxeeraspagtiderC.sretrauqnisisixalatnozirohehtnoemiT:setoN tnevelaudiserasetacidni∗,ygolonorhcDLfoesacnI.)7102(.lateacuDoL-DL,)3102(aicnelaVdnaneveaL-VL,)9102(.latenoraB-XVB,)2102(rolyaTdnakciraluhcS-TS .)sisirccimetsysanahtrehtar( 58

yekruT)b( dnalreztiwS)a( aciremAfosetatSdetinU)d( modgniKdetinU)c( ASU-dnalreztiwS:seirtnuoCssorcasecirPtessAdnayenoMdaorBrofspaGRAVB:4.EerugiF :seigolonorhcgniwollofehtotgnidroccasesircfotesnoehtyfingissenillacitreV.stniopegatnecrepnidesserpxeeraspagtiderC.sretrauqnisisixalatnozirohehtnoemiT:setoN tnevelaudiserasetacidni∗,ygolonorhcDLfoesacnI.)7102(.lateacuDoL-DL,)3102(aicnelaVdnaneveaL-VL,)9102(.latenoraB-XVB,)2102(rolyaTdnakciraluhcS-TS .)sisirccimetsysanahtrehtar( 59

AppendixF. RobustnessResults TableF.2:DataandDataTransformationsintheExtendedMonetaryVAR(U.S.) Variable Transformation IndustrialProduction Log-Level ConsumerPriceIndex(CPI) Log-Level UnemploymentRate Level ProducerPriceIndex(PPI) Log-Level FederalFundsRate(FFR) Level OilPrice Log-Level StockPrices(S&P500composite) Log-Level PrimeLoanRate Level 1YearBondRate Level 3YearsBondRate Level 5YearsBondRate Level 10YearsBondRate Level M1 Log-Level MZM Log-Level M2 Log-Level TotalLoansandLeases Log-Level Notes:AlldataseriesareretrievedfromtheFRED. 60

TableF.3:RootMeanSquaredErrors(RMSE)aofForecastsinTotalLoansandLeases(U.S.) 1988b 1989 1990 1991 1992 (1989-1992)c (1990-1993) (1991-1994) (1993-1995) (1993-1996) 4VAR 3.93 11.13 8.96 2.39 8.38 7VAR 3.13 10.69 11.00 3.60 10.91 16VAR 3.78 11.26 7.37 2.50 9.64 1993 1994 1995 1996 1997 (1994-1997) (1995-1998) (1996-1999) (1997-2000) (1998-2001) 4VAR 6.18 3.13 2.94 1.35 2.13 7VAR 4.81 1.58 4.86 1.52 2.25 16VAR 4.80 1.44 1.97 1.37 4.34 1998 1999 2000 2001 2002 (1999-2002) (2000-2003) (2001-2004) (2002-2005) (2003-2006) 4VAR 3.01 2.01 4.18 11.64 5.55 7VAR 3.11 2.73 5.41 9.66 5.74 16VAR 2.34 3.81 3.25 11.35 3.49 2003 2004 2005 2006 allwindows (2004-2007) (2005-2008) (2006-2009) (2007-2010) (average) 4VAR 12.50 6.34 3.16 1.87 5.30 7VAR 12.93 4.22 2.83 4.17 5.53 16VAR 9.79 2.64 1.26 2.58 4.68 (cid:115) aReportedRMSEsaresomputedwithrespecttotheforecastmeanacrossallhorizons,i.e.RMSE = t (cid:80) = H 1 (yt −yˆt)2 , H whereHistheforecasthorizonandyˆ istheforecastmean. Replacingthemeanbythemediandoesnotchange t theresultssubstantially. ThesmallestRMSEsareinbold. bTheyeardenotesthelastyearinthe15yearestimationrollingwindow. cTheyearspandenotestheforecastingperiodoftheparticularrollingwindow. 61

FigureF.1:CreditGapValuesforTotalLoansandLeases(U.S.)andtheBaselineFlowofFundsTotalNonfinancialCreditMeassure. Notes:Positivevaluesindicateatypicalcreditexpansions.Negativevaluesindicateatypicalcreditcontractions.Timeisinmonths, creditgapisexpressedinpercentagepoints. 62

FigureF.2: CreditGapValuesforTotalLoansandLeases(U.S.)forBaselineSpecificationandforanAlternativeSpecificationwithRealTotal LoansandLeasesPerCapitaasCreditMeasure. Notes:Positivevaluesindicateatypicalcreditexpansions.Negativevaluesindicateatypicalcreditcontractions.Timeisinmonths, creditgapisexpressedinpercentagepoints. FigureF.3:CreditGapValuesforTotalLoansandLeases(U.S.)withfortheBaseline,SmallerandLargerVARSystems. Notes:Positivevaluesindicateatypicalcreditexpansions.Negativevaluesindicateatypicalcreditcontractions.Timeisinmonths, creditgapisexpressedinpercentagepoints. 63

FigureF.4:CreditGapValuesforTotalLoansandLeases(U.S.)forBaseline(Optimized)PriorHyperparametersandforaLooserPrior. Notes:Positivevaluesindicateatypicalcreditexpansions.Negativevaluesindicateatypicalcreditcontractions.Timeisinmonths, creditgapisexpressedinpercentagepoints. 64