AIandExchangeRatePredictability

AminIzadyar

ImperialCollegeBusinessSchool,ImperialCollegeLondon

Email:a.izadyar23@imperial.ac.uk

July3,2025

Abstract

Irevisittheexchangeratedisconnectpuzzle,firstdocumentedby

MeeseandRogoff

(

1983

),usinggenerativeartificialintelligence(AI)toforecastcurrencyreturnsbasedoneconomicfundamentals.UsingChatGPTandDeepSeek,Ianalyzeacomprehensivedatasetofeconomicdatareleasesformajorcurrencypairsandmeasurethefunda-mentalstrengthofeachcurrency.TheseAI-poweredfundamentalsexhibitsignificantcross-sectionalpredictivepower.AsimpletradingstrategythatgoeslongcurrencieswithstrongfundamentalsandshortcurrencieswithweakfundamentalsgeneratesaSharperatioexceeding0.7perannum.Theexcessreturnsofthisstrategyremainsignificantaftercontrollingfortraditionalcurrencyfactors.Tomitigateconcernsoflook-aheadbias,IrunmultipleexercisestoensurethatpredictabilitystemsfromAIreasoningratherthanmemorization.Finally,Iexplorethepotentialsourcesofpre-dictabilityandfindevidencethattheTaylorruleframework,generallyusedbycentralbankstosetinterestrates,isakeymechanismconnectingexchangeratestoeconomicfundamentals.

Keywords:ForeignExchange,ReturnPredictability,LargeLanguageModels,ChatGPT,ArtificialIntelligence

JELClassification:C53,F31,F37,G12,G15.

1

1Introduction

Theabilityofeconomicfundamentalstoforecastexchangeratesremainselusive,sincemod-elsbasedonfundamentalsareoftenoutperformedbyasimplerandomwalk,aphenomenonknownasthe“exchangeratedisconnect”puzzle(e.g.,

MeeseandRogoff

,

1983

).Althoughtherecentliteraturehasidentifiedafeweconomicvariablesthatappeartohavepredictivepower,theanswertothisempiricalpuzzleremainsunresolved(e.g.,

Mark

,

1995

;

Engeland

West

,

2005

;

Rossi

,

2013

).

Againstthisbackdrop,theemergenceofartificialintelligence(AI)offersnewopportunitiestore-examinethispuzzle.AI’srecentadvancementshaveenabledittosolveproblemsonceconsideredtoocomplexordata-intensivefortraditionalmethods.Motivatedbythesedevel-opments,IleverageAI’sreasoningpowerandproficiencyinhandlinglargedatasetstostudyacomprehensivedatasetofeconomicdatareleases,coveringover500indicatorsfrom1996to2024formajoreconomies.Thisstudyaimstofindalinkbetweenexchangeratesandeco-nomicfundamentals,thusenhancingourunderstandingofpricediscoveryinthelargestanddeepestfinancialmarketintheworld.Topreviewmyresults,IfindevidencethatAI-derivedfundamentalscanpredictfutureexchangeratereturnsandthemostimportantpredictorsareInflationdata,Employmentdata,andBroadeconomicactivityindicators.

Usinglargelanguagemodels(LLMs)likeGPT-4oandDeepSeek-V3,Ianalyzealargedatasetcomprisingrealizedvalues,previousfigures,andconsensusforecastsofkeyeconomicdatareleases,suchasGDPreports,employmentstatistics,inflationindices,andcentralbankdecisions.IinteractwiththeAImodelusingastructuredpromptandaskittogenerateaconciseanalysisandadirectionalsignalindicatingwhetherthedatareleaseimpliesthecurrencywouldSTRENGTHEN,WEAKEN,orhasanINSIGNIFICANTORUNCERTAINimpact.Notably,theinputprovidedtothemodelincludesonlytherealized,previous,andforecastvalues,alongwiththenameofthecurrencyassociatedwiththedatarelease,exclud-inganyinformationaboutthetimeordateoftherelease.UsingAI’soutput,Ithenconstructasimplemeasure,calledtheAIFXindex,thatcapturesthenetfundamentalstrengthofeachcurrency.Specifically,foreachcurrency,theAIFXindexisdefinedasthedifferencebetweenthenumberofpositive(directionalsignalas“STRENGTHEN”)andnegative(“WEAKEN”)

2

signals,dividedbythetotalnumberofsignals,overagivenlookbackwindow.

TheAIFXindexexhibitssignificantcross-sectionalpredictivepower.Asimpletradingstrat-egy,calledtheAIFXstrategy,thatgoeslongcurrencieswithastrongAIFXindexandshortcurrencieswithaweakAIFXindexproducesanannualizedSharperatiolargerthan0.7.Moreover,aftercontrollingfortraditionalcurrencyfactorslikedollar,dollarcarry,carry,momentum,andvalue,Iuncoverastatisticallysignificantalphathataccountsfor74%oftheAIFXstrategy’saveragereturn.Ifurthervalidatetheresultthroughapanelregressionexercise,showingthattheAIFXindexeffectivelypredictsnextmonthexchangeratereturns.Takentogether,thesefindingssuggestthatAIcanhelpuncoverpreviouslyunderexploredsourcesofreturnpredictabilityintheFXmarket,thussheddinglightontheroleofeconomicfundamentals,asadvocatedbythetheoreticalliterature.

Iconductmultiplerobustnesscheckstoensurethereliabilityofthebaselineresults.First,whilethemainanalysisusesGPT-4otointerpretdatareleases,IreplicatetheentireexerciseusingDeepSeekV3totestthesensitivityofthecorefindingstothechoiceofAImodel.Theresultsremainconsistent,withbothmodelsexhibitingverysimilarperformance.Second,Iconstructanalternativemeasure,theWeightedAIFXindex,whichassignsaweighttoeachdirectionalsignalbasedonitsestimatedlevelofimportance.Iusethisweightedindexasthesignalinacross-sectionaltradingstrategyandfindthatthecoreresultsremainrobust.Third,Ialsoimplementatime-seriestradingstrategyasanalternativetothecross-sectionalstrategyusedinthebaselinespecification.Thetime-seriesstrategyalsogenerateseconomi-callymeaningfulSharperatios,anditsperformanceremainssignificantaftercontrollingforbenchmarkcurrencyfactors.

AmajorconcernwhenusingLLMsforpredictiontasksislook-aheadbias,whichoccurswhenamodelistrainedusinginformationnotavailableatthetimeoftheprediction.Asaresult,themodel’sperformancemaylookbetterthanitwouldbeinreal-time.Tomitigatethisconcern,Iimplementfourdifferentexercises.Inthefirstone,IinvestigatewhethertheAImodelmayimplicitly“remember”thetimingofadatarelease.Specifically,IusethesamedatathatwasfedtotheAImodeltogeneratethedirectionalsignals,butthistimeIaskittoindicatetheyear(nottheexactdate)whenthedatawasreleased.IftheAImodelcanrecallthetimingoftherelease,weshouldexpectittocorrectlyidentifythereleaseyear

3

inahighproportionofcases.However,thedistributionofyearsguessedbytheAImodeldiffersmarkedlyfromthetruedistributionofdatareleasesinthedataset.Ishowthatonly5.6%ofthemodel’sguessesarecorrectonaveragewithineachyear.Inthesecondexercise,Iexploitthefactthattheknowledgecut-offdateforGPT-4oisOctober2023,whileforGPT-3.5itisSeptember2021.Thistwo-yeargapprovidesanopportunitytoexaminewhethertherelativeperformanceofthetwomodelsdifferssignificantly.Specifically,Icomparetheirperformanceduringtheperiodfrom1996to2021,coveredbybothmodels’trainingsets,totheperiodfrom2021to2023,whichonlyGPT-4owastrainedon.Iflook-aheadbiaswerepresent,wewouldexpectasharpdeclineintherelativeperformanceofGPT-3.5afteritstrainingperiodends,comparedtoGPT-4o.Totestthis,Iconductadifference-in-differencesanalysisandfindthattherelativeperformanceofthetwomodelsdoesnotdiffersignificantlyacrossthetwoperiods.Inthethirdexercise,ItestforthepossibilitythattheAImodelmayhaveamemoryoftheoverallrelationshipbetweenexchangeratereturnsandcertainmacro-variables.Forexample,ifthemodelwastrainedduringaperiodwheninflationandcurrencyreturnswerepositivelycorrelated,itmightpredicthigherexchangeratesinresponsetoris-inginflation,basedonmemoryandnotreasoning.Therefore,IinvestigatewhethertheAImodelhasamemoryoftherealizedcorrelationofmacrovariableswithnextmonthcurrencyreturnsoverthesampleperiod,butfindnoevidenceindicatingso.Inthefourthexercise,IconstructaportfoliobasedonwhattheAImodelcanrememberaboutmonthlycurrencyreturnsduringthesampleperiod,referredtoasthepurehindsightportfolio,anduseitasacontrolfactor.IfindthatthereturnoftheAIFXstrategyisorthogonaltothereturnofthepurehindsightportfolio.Overall,thesefindingscollectivelysuggestthattheAImodel’sperformanceisunlikelytobedrivenbylook-aheadbias,andshouldreflectgenuinereasoningbasedontheinformationavailableatthetimeofprediction.

AfterestablishingthepredictivepoweroftheAI-derivedvariables,Iinvestigatetheunder-lyingsourcesofthispredictability.Thisstepiscrucialfromaneconomicstandpoint,asitshedslightonthepossiblemechanismsthroughwhichfundamentalsinfluenceexchangeratemovements.First,IfindthatInflationdata,Employmentdata,andBroadeconomicactivityindicatorsarethemostimportantcategoriesforforecastingexchangerates.Thesevariablesarecloselylinkedtothemonetarypolicyframeworkproposedby

Taylor

(

1993

),

4

suggestingthatcentralbanks,policyresponsestoeconomicconditionsplayapivotalroleinexchangeratedetermination.Thisinterpretationissupportedbypriorempiricalstud-iessuchas

ClaridaandWaldman

(

2007

),

MolodtsovaandPapell

(

2009

),and

Engeland

Wu

(

2024

),whichdocumentthepredictivepowerofTaylor-rulefundamentalsforexchangerates.Second,Ifindthatthepredictivesignalislargelydrivenbypositivenews(newsim-plyingcurrencyappreciation)ratherthannegativenews(implyingdepreciation).Furtheranalysisrevealsthatnegativenewstendstotriggerastrongerimmediatemarketreactionthanpositivenews,consistentwiththefindingsof

Andersenetal.

(

2003

).Asaresult,neg-ativenewsmayleavelessroomfordelayedexchangerateadjustments,therebyreducingitspredictivecontentatlongerhorizons.Aplausibleinterpretationisthatthisasymmetryinmarketreactionispartlydrivenbythewaycentralbanksimplementmonetarypolicy.Inparticular,severalstudieshavedocumentedthatmonetaryauthoritiestendtorespondmoreaggressivelytonegativeoutputgapsthantopositiveones,leadingtoageneralbiastowardlowerinterestrates(

Juanetal.

,

2004

;

BrüggemannandRiedel

,

2011

;

Hofmannand

Bogdanova

,

2012

;

Komlan

,

2013

).Thepoliticaleconomyofmonetarypolicyalsoreinforcesthisasymmetrictendency.Ratehikescanbepoliticallyunpopularastheymayslowtheeconomyorincreaseborrowingcosts.Thisasymmetricpolicystancecaninfluenceinvestors,expectations,promptingstrongerimmediatereactiontonegativenewsandcontributingtotheasymmetricpredictivepowerdocumentedinthisstudy.Takentogether,theempiricalfindingspointtotheimportanceoftheTaylorruleandmonetarypolicyinexplainingex-changeratemovements.Nonetheless,alternativeexplanationscannotbedefinitivelyruledout.

Thisresearchcontributestotwostrandsofliterature.Thefirstinvolvesthewell-known“ex-changeratedisconnect”puzzle,firstobservedby

MeeseandRogoff

(

1983

).Theirfindings,seenasshockingatthetime,promptedalargeliteraturethatre-examinedtherobustnessoftheresults(

Mark

,

1995

;

Kilian

,

1999

;

Cheungetal.

,

2005

;

MolodtsovaandPapell

,

2009

).However,theearlyempiricalstudieswereinconclusiveinaddressingthepuzzle.Anotablecontributioninthiscontextis

EngelandWest

(

2005

),whoofferapotentialresolution.Theydemonstrateanalyticallythatexchangeratescanbeconsistentwithpresentvalueassetpricingmodelsandfollowaprocessarbitrarilyclosetoarandomwalkifcertainconditions

5

aremet.Followingthis,

Engeletal.

(

2007

)presentadefenseofexchangeratemodelsbyarguingthatarandomwalkmodelisatoughbenchmarktobeatandproposealternativemethodsforevaluatingtheperformanceofexchangeratemodels.Inaddition,recentempiri-calstudiessuggestaconnectionbetweencurrencyreturnsandcountries’externalimbalances(

GourinchasandRey

,

2007

;

DellaCorteetal.

,

2012

,

2016

),sovereignrisk(

Augustinetal.

,

2020

;

DellaCorteetal.

,

2022

,

2023

),theoutputgap(

Colacitoetal.

,

2020

),macroeconomicuncertainty(

BergandMark

,

2018

;

DellaCorteandKrecetovs

,

2024

),andunemployment(

Nucera

,

2017

).Notably,thepatternofpredictabilitydocumentedinthispaperisconsistentwiththefindingsof

DahlquistandHasseltoft

(

2020

),whoexaminehowpasttrendsinkeymacroeconomicindicators,referredtoaseconomicmomentum,canpredictcurrencyreturns.ThispaperadvancestheexistingliteraturebyusinganovelAI-poweredmethodologytoan-alyzeanexpandedsetofeconomicindicators.Thisinnovativeapproachhelpsuncovernewpredictabilitypatternsandprovidesnewinsightsintoexchangeratemovements.

Second,thisprojectaddstothebodyofliteratureonnovelresearchmethodsinfinancialeconomicsthatleveragegenerativeAI.Forexample,

EisfeldtandSchubert

(

2024

)conductacomprehensivesurveyofhowthisemergingtechnologycandecreasethetimeandcostsasso-ciatedwithtraditionalresearchdesignsinfinancewhileenablingnovelanalyticalapproaches.Emergingapplicationsincludegeneratingdataembeddings(

Gabaixetal.

,

2024

;

Kimetal.

,

2024

),textclassification(

Changetal.

,

2024

;

Krockenbergeretal.

,

2024

),retrieval-augmentedgeneration(

Bartiketal.

,

2024

;

ChenandWang

,

2024

),simulatingagentbehavior(

Horton

,

2023

;

Fedyketal.

,

2024

;

Hewittetal.

,

2024

),andhypothesisgeneration(

Sietal.

,

2024

;

LudwigandMullainathan

,

2024

).Specifically,thepromptingtechniqueemployedinthisstudyismostsimilartotheapproachesusedinthefollowingpapers.

Bybee

(

2023

)usesAItogenerateeconomicexpectationsfromhistoricalnewsdataspanning120years.Inaddi-tion,

Lopez-LiraandTang

(

2024

)and

Chenetal.

(

2024

)exploretheabilityofgenerativeAI,specificallyChatGPT,topredictstockpricemovementsbasedonsentimentsextractedfrombusinessnewsheadlines.Incontrasttothesepapers,whichfocusonsentimentextractionfromtextualdata,thisstudydoesnotaimtoextractsignalsfromtext.Instead,itreliesonstructured,numericaldata,andtheAImodelispromptedtogenerateanalysisbasedonthenumericalvaluesofeconomicdatareleases.Overall,thispapercontributestothe

6

existingliteraturebydemonstratingAI’scapabilitytoanalyzelargevolumesofstructureddatainthecontextofcurrencymarkets.Inaddition,itintroducesnewtechniquestoaddresslook-aheadbias.

Theremainderofthepaperisorganizedasfollows;

Section2

presentsthedata;

Section3

outlinestheconstructionoftheAI-poweredvariables;

Section4

evaluatesthepredictivepowerofthesevariables;

Section5

investigatestheissueoflook-aheadbias;

Section6

ex-plorestheunderlyingmechanismsdrivingthepredictability;and

Section7

concludes.Aseparate

InternetAppendix

providesadditionalresultsnotincludedinthemainbodyofthispaper.

2Data

IfocusonG-10currenciesthatincludetheUnitedStatesdollar(USD),Euro(EUR),Japaneseyen(JPY),Britishpoundsterling(GBP),Swissfranc(CHF),Canadiandollar(CAD),Australiandollar(AUD),NewZealanddollar(NZD),Swedishkrona(SEK),andNorwegiankrone(NOK).Ilimitmyfocustothesecurrenciesbecauseofthelonghistoryofeconomicdatareleasesavailable.Tomeasuretheeconomicfundamentalsofeachcurrency,IhavecollectedtheeconomiccalendardatafromI.Theeconomiccalendaraggregateskeyeconomicdatareleases,suchasGDPreports,employmentstatistics,infla-tionreadings,centralbankdecisions,andothereconomicindicators(544uniqueindicators),acrossmultiplecountries.Itprovides,foreachdatarelease,therealizedvalue,thepreviousfigure,andtheconsensusforecast.ThedatasetspansfromJanuary1996toOctober2024andcomprisesatotalof174,820datapoints.

Table1

displaysthenumberofobservationscollectedforeachcurrency.Inaddition,Ihavecollectedend-of-day(Londontime)exchangeratesandone-monthforwardratesfromBloomberg.Notably,therearenineexchangeratesinthecross-section,andallratesaredefinedastheamountofU.S.dollars(USD)requiredtopurchaseoneunitofforeigncurrency(FCU).

Table2

reportssummarystatisticsforexchangeratereturns.

Table1

and

Table2

abouthere

7

3AnalysingDataReleases

3.1AIasafinancialAnalyst

Foreachdatareleaseintheeconomiccalendardata,Ifeedstructuredprompts,asin

Prompt1

,toGPT-4ousingAPIs.Thepromptcontainsthetitle,realizedvalue,previousfigure,andtheconsensusforecastofthedatareleaseandinstructstheAImodeltogenerateaconciseanalysisandadirectionalsignalindicatingwhetherthedatareleaseimpliesthecurrencywouldSTRENGTHEN,WEAKEN,orhasanINSIGNIFICANTORUNCERTAINimpact.In

Prompt1

,{currency}willcorrespondtothecurrencyassociatedwiththedatarelease.Notably,Iexcludeanyinformationaboutthetimeordateoftherelease.

Prompt:Forgetallpreviousinstructions.Youareafinancialanalyst.Youwillbegivenanewsheadline.Pleaseanalyzethenews’seffecton{currency}andprovideabriefresponse.Thenindicatewhetherthisheadlineimplies{currency}willSTRENGTHEN,WEAKEN,orhaveanINSIGNIFICANTORUNCERTAINeffect.Generatetheoutput

inthisformat:{(ANALYSIS:shortanalysisdiscussingthechannel),(DIRECTION:oneofSTRENGTHEN,WEAKEN,INSIGNIFICANTORUNCERTAIN)}

Prompt1

Asillustratedin

Example1

,

Example2

,and

Example3

,theoutputexplainsthemechanismofhowthereleasecouldaffecttheassociatedcurrencyandindicatesthepotentialdirectionofchange.

?Currency:AUD

?Input:Newsheadline:TradeBalance(monthly),Actual:4.502B,Forecast:6.100B,Previous:7.180B

?Output:Analysis:Thetradebalancefigurecameinlowerthanboththeforecastandthepreviousmonth,indicatingasmallerthanexpectedsurplus.Thiscouldsuggestthatexportshavedecreasedorimportshaveincreasedmorethananticipated,potentiallysignalingweakereconomicactivity.AlowertradebalancecannegativelyimpactthecurrencyasitimpliesthatfewerAustraliandollarsaredemandedininternationalmarketstopurchaseAustraliangoodsandservices.,Direction:WEAKEN.

Example1

8

?Currency:CAD

?Input:Newsheadline:GDP(MoM),Actual:0,Forecast:0.001,Previous:0.006

?Output:Analysis:TheCanadianGDPmonth-on-monthgrowthrateisrecordedat0%,whichisbelowtheforecastof0.1%andsignificantlylowerthantheprevi-ousmonth’s0.6%.ThisstagnationsignalsaslowdownineconomicgrowthandmaydampeninvestorsentimenttowardstheCanadiandollar.AsGDPisakeyindicatorofeconomichealth,alowerthanexpectedperformancecanleadtoexpectationsofamoreaccommodativemonetarypolicyfromtheBankofCanada,potentiallyputtingdownwardpressureontheCAD.,Direction:WEAKEN.

Example2

?Currency:GBP

?Input:Newsheadline:CPI(MoM),Actual:0.004,Forecast:0.003,Previous:0.002

?Output:Analysis:TheConsumerPriceIndex(CPI)measuresthechangeinthepricelevelofabasketofgoodsandservices.Ahigher-than-expectedCPIindicatesincreasedinflationarypressures.TheactualCPIof0.004surpassesboththeforecastof0.003andthepreviousvalueof0.002.Thissuggestsstrongerinflationarytrends,potentiallyleadingtoexpectationsoftightermonetarypolicybytheBankofEngland.Higherinterestratesgenerallyleadtocurrencyappreciationastheyattractforeigninvestmentseekinghigherreturns.,Direction:STRENGTHEN.

Example3

Toclarifytheterminology,throughoutthispaper,IrefertodatareleaseswiththedirectionlabeledasSTRENGTHENintheoutputaspositivenews,thoselabeledasWEAKENasnegativenews,andthoselabeledasINSIGNIFICANTORUNCERTAINasneutralnews.

Table1

presentsadetailedbreakdownofthecountandpercentageofpositive,negative,andneutralnewsforeachcurrencyduringthesampleperiod,providinginsightsintothedistributionofoutputsfromtheAImodel.

3.2VariableConstruction

IfollowasimpleandintuitiveapproachtoconstructthreeAI-poweredvariables.Supposeweareattimetandletτdenotethelookbackperiod.Therefore,thetimeintervalL=(t-τ,t]wouldbethelookbackwindowattimet.Basedonthis,Iconstructthefollowingvariables

9

forcurrencyc:

AIFXc;t;τ=Strengthc;t;τ-Weaknessc;t;τ(3)

TheStrengthratiocapturestheproportionofpositivenews,whiletheWeaknessratiomea-surestheproportionofnegativenews.TheAIFXindexisdefinedasthenetbalancebetweenpositiveandnegativenews,providingasinglecompositemetricofimpliedcurrencystrengthderivedfromAI-classifieddata.

4PerformanceEvaluation

Inthissection,IanalyzethepredictivepoweroftheAI-derivedvariables.Ibeginbycon-structingcross-sectionaltradingstrategiesthatusetheAIFXindexasthesignal,evaluatedacrossarangeoflookbackperiods.Theperformanceofthesestrategiesisthenassessedrelativetocommoncurrencyfactors.ToformallytestthestatisticalsignificanceoftheAIFXindexinpredictingfuturereturns,Iestimatepanelregressions.Toensurerobustness,theentireanalysisisreplicatedusingDeepSeek-V3,aleadingalternativetothebaselinemodelGPT-4o.Next,IintroduceanalternativespecificationoftheAI-poweredvariablesbyweightingeachdatareleaseaccordingtoitsestimatedeconomicimportance.Finally,Iim-plementatime-seriesstrategybasedonthesamesignalandfurtherdecomposethepredictivecomponenttoisolatetheroleofU.S.dollarfundamentals.

4.1Cross-sectionalTradingStrategy

IassessthepredictivepoweroftheAIFXindex(asdefinedin

Equation(3)

),usingstylizedcross-sectionaltradingstrategies.Notably,toevaluatethesensitivityofperformancetothelengthoflookbackwindow,Iconsiderlookbackperiodsof1to60months.Specifically,foreachchoiceoflookbackperiod,currenciesaresortedbytheirAIFXindexattheendofeachmonthandItakelongpositionsinthetoptwocurrencieswiththehighestAIFXindexandshortpositionsinthebottomtwowiththelowest.Irefertothisstrategyasthe

10

AIFXstrategy.

Figure1

presentstheannualizedSharperatiosofthestrategyacrossdifferentlookbackperiods.

Figure1

abouthere

TheAIFXstrategyconsistentlyyieldspositiveSharperatiosacrossalllookbackperiods,witheconomicallysignificantperformanceinmostcases.Predictivepowerappearsparticularlystrongforlookbackperiodsof36to60months.Additionally,

Table3

reportstheperformancestatisticsoftheAIFXstrategy,includingthemean,standarddeviation,skewness,excesskurtosisandfirst-orderautocorrelationofreturns.

Figure2

illustratesthedollarvalueofaninitial$1investmentintheAIFXstrategy

1

.Avisualinspectionofthefigureindicatesageneralupwardtrendinperformance,withgainsdistributedrelativelyevenlythroughoutthesampleperiod.

Figure3

showstheportfoliocompositionovertime.Thestrategyexhibitsmoderateturnover,implyingthattransactioncostsareunlikelytosignificantlyerodereturns.

Table3

,

Figure2

and

Figure3

abouthere

Tofurtherexaminetheperformance,Iruncontemporaneousregressionsbasedon:

RXt;τ=ατ+β1;τDollart+β2;τDollarCarryt+β3;τCarryt+

β4;τMomentumt+β5;τValuet+∈t;τ(4)

whereRXt;τdenotesthemonthlyexcessreturnoftheAIFXstrategywithlookbackperiodτ;Dollarisalong-onlyportfoliothattakesequal-weightedlongpositionsinallcurrencies;DollarCarryisadirectionalstrategythatgoeslong(short)allcurrencieswhentheaverageforwarddiscountispositive(negative);Carry,Momentum,andValuearecross-sectionalstrategiesthatrankcurrenciesbytheirforwarddiscount,previousmonth’sreturn,andfive-yearreturn,respectively.Theconstructionofcurrencyfactorsisfurtherdetailedin

Internet

AppendixA

andtheirperformancestatisticsarereportedin

InternetAppendixTableA.1

.

Ireporttheregressionresultsof

Equation(4)

in

Table4

.

1Tosavespace,

Figure2

alsodisplaysthecumulativereturnofastrategythatusesStrengthratio(definedin

Equation(1)

)asthesignal.Thisstrategywillbediscussedin

Section6.2.1

.

11

Table4

abouthere

ThefindingsindicatethattheAIFXstrategy’sreturnsarenotfullyexplainedbythecommoncurrencyfactors,andthisconclusionholdsacrossdifferentlookbackperiods.Forexample,witha48-monthlookbackperiod,thestrategyyieldsastatisticallysignificantalphathataccountsfor74%oftheaveragereturnofthestrategy,suggestingthatonlyaboutone-quarterofthereturnissubsumedbytraditionalcurrencyfactors.Notably,forlookbackperiodsof54and60months,thealphabecomesonlymarginallysignificant,astheValuefactorgainsmoreexplanatorypower.Overall,theanalysispresentedinthissectionprovidesstrongevidencethattheAIFXindexpossessessignificantpredictivepower.ThesefindingssuggestthatAIcanhelpuncoverpreviouslyunderexploredsourcesofreturnpredictabilityintheFXmarket,contributingtoarenewedconnectionbetweenexchangeratesandunderlyingeconomicfundamentals.

4.2PanelRegression

ToassessthestatisticalsignificanceofthepredictivepoweroftheAIFXindex,Iestimatethefollowingpanelregressionmodelforeachchoiceoflookbackperiod:

Rc;t+1=αt;τ+βτAIFXc;t;τ+∈c;t;τ(5)

In

Equation(5)

,Rc;t+1representsth