央行的人工智能-國際清算銀行（英）

BISBulletinNo84ArtificialintelligenceincentralbankingDouglasAraujo,SebastianDoerr,LeonardoGambacortaandBrunoTissot23January2024BISBulletinsarewrittenbystaffmembersoftheBankforInternationalSettlements,andfromtimetotimebyothereconomists,andarepublishedbytheBank.Thepapersareonsubjectsoftopicalinterestandaretechnicalincharacter.TheviewsexpressedinthemarethoseoftheirauthorsandnotnecessarilytheviewsoftheBIS.TheauthorsaregratefultoBryanHardyandGaloNu?oforcomments,IlariaMatteiandKrzysztofZdanowiczforexcellentresearchassistance,andtoLouisaWagnerforadministrativesupport.TheeditoroftheBISBulletinseriesisHyunSongShin.ThispublicationisavailableontheBISwebsite().?BankforInternationalSettlements2024.Allrightsreserved.Briefexcerptsmaybereproducedortranslatedprovidedthesourceisstated.ISSN:2708-0420(online)ISBN:978-92-9259-738-2(online)DouglasAraujoSebastianDoerrLeonardoGambacortaBrunoTissotDouglas.Araujo@Sebastian.Doerr@Leonardo.Gambacorta@Bruno.Tissot@ArtificialintelligenceincentralbankingKeytakeaways

Centralbankshavebeenearlyadoptersofmachinelearningtechniquesforstatistics,macroanalysis,paymentsystemsoversightandsupervision,withconsiderablesuccess.Artificialintelligencebringsmanyopportunitiesinsupportofcentralbankmandates,butalsochallenges–somegeneralandothersspecifictocentralbanks.Centralbankcollaboration,forinstancethroughknowledge-sharingandpoolingofexpertise,holdsgreatpromiseinkeepingcentralbanksatthevanguardofdevelopmentsinartificialintelligence.Longbeforeartificialintelligence(AI)becameafocalpointofpopularcommentaryandwidespreadfascination,centralbankswereearlyadoptersofmachinelearningmethodstoobtainvaluableinsightsforstatistics,researchandpolicy(Doerretal(2021),Araujoetal(2022,2023)).Thegreatercapabilitiesandperformanceofthenewgenerationofmachinelearningtechniquesopenupfurtheropportunities.Yetharnessingtheserequirescentralbankstobuildupthenecessaryinfrastructureandexpertise.Centralbanksalsoneedtoaddressconcernsaboutdataqualityandprivacyaswellasrisksemanatingfromdependenceonafewproviders.ThisBulletinfirstprovidesabriefsummaryofconceptsinthemachinelearningandAIspace.Itthendiscussescentralbankusecasesinfourareas:(i)informationcollectionandthecompilationofofficialstatistics;(ii)macroeconomicandfinancialanalysistosupportmonetarypolicy;(iii)oversightofpaymentsystems;and(iv)supervisionandfinancialstability.TheBulletinalsosummarisesthelessonslearnedandtheopportunitiesandchallengesarisingfromtheuseofmachinelearningandAI.Itconcludesbydiscussinghowcentralbankcooperationcanplayakeyrolegoingforward.OverviewofmachinelearningmethodsandAIBroadlyspeaking,machinelearningcomprisesthesetoftechniquesdesignedtoextractinformationfromdata,especiallywithaviewtomakingpredictions.Machinelearningcanbeseenasanoutgrowthoftraditionalstatisticalandeconometrictechniques,althoughitdoesnotrelyonapre-specifiedmodeloronstatisticalassumptionssuchaslinearityornormality.Theprocessoffittingamachinelearningmodeltodataiscalledtraining.Thecriterionforsuccessfultrainingistheabilitytopredictoutcomesonpreviouslyunseen(“out-of-sample”)data,irrespectiveofhowthemodelspredictthem.Thissectiondescribessomeofthemostcommontechniquesusedincentralbanks,basedontheregularstocktakingexercisesorganisedinthecentralbankingcommunityundertheumbrellaoftheBISIrvingFisherCommitteeonCentralBankStatistics(IFC).Tree-basedmethodsareflexiblemachinelearningalgorithmsthatcantackleawiderangeoftasks.Decisiontreesgroupindividualdatapointsbysequentiallypartitioningdataintofinercategoriesaccordingtospecificcharacteristicsofinterest.Forexample,atreemayfirstsorthouses(theinputdata)intothosewithmorethanthreeroomsandthosewithatmostthree,andthenpartitionhousesineachoftheseBISBulletin1subgroupsintothosebuiltbefore1990andthosebuiltafter,andsoon.Theresultingfinerpartitioningofhousescanthenbecomparedwithaparticulardimensionofinterest(theoutput)toseehowwellthepartitioningmatchesanattributeofinterest.Forinstance,capturinghowhousepricesvaryacrossthefinerpartitioningwouldbeawaytogroupsimilarhousesintermsoftheirprice.Randomforestscombineseveraltreestrainedondifferentslicesofthesamedatatoimprovepredictionoutofsamplewhileguardingagainsttheriskofoverfittingthetrainingdatasample.Randomforestsandrelatedmodelscanbeseenasamoreflexibleformofregressionanalysis,astheypredictoutputfromtheexplanatoryvariablesofinterest(AtheyandImbens(2021)).Inaddition,tree-basedmethodscanserveasanexploratorytooltogleanpatternsinthedatawithoutimposingamodelstructure.Forinstance,theycanclassifydatapointsintosimilarcategories.Inthesamespirit,forestscanbedeployedinidentifyingoutliersbymeansofisolationforests,amethodthatsinglesoutthedatapointsthatcanbeisolatedfromothers.Neuralnetworksareperhapsthemostimportanttechniqueinmachinelearning,withwidespreadusesevenforthelatestgenerationofmodels.Theirmainbuildingblocksareartificialneurons,whichtakemultipleinputvaluesandtransformtheminanon-linearwaytooutputasinglenumber–likelogisticregressions.Theartificialneuronsareorganisedtoformasequenceoflayersthatcanbestacked:theneuronsofthefirstlayertaketheinputdataandoutputanactivationvalue.Subsequentlayersthentaketheoutputofthepreviouslayerasinput,transformitandoutputanothervalue,andsoforth.Thisway,similartoneuronsinthehumanbrain,anartificialneuron’soutputvalueisakintoanelectricalimpulsetransmittedtootherneurons.Anetwork’sdepthreferstothenumberoflayers.Eachneuron’sconstantandweightsattachedtotheoutputofpreviouslayers’neuronsarecollectivelycalledparameters;theydeterminethestrengthofconnectionsacrossneuronsandlayers.Theseparametersareimprovediterativelyduringtraining.Deepernetworkswithmoreparametersrequiremoretrainingdatabutpredictmoreaccurately.NeuralnetworksarebehindfacerecognitionorvoiceassistantsinmobilephonesandunderliethemostsignificantrecentinnovationsinAI.Transformers,unveiledin2017,drasticallyimprovedtheperformanceofneuralnetworksinnaturallanguageprocessing(NLP)andenabledtheriseoflargelanguagemodels(LLMs).Ratherthanjustrelatingawordtothosenearit,transformersattempttocapturetherelationshipbetweenthedifferentcomponentsofatextsequence,eveniftheyarefarapartinthesentence.Thisallowsthemodeltobetterunderstandthecontextandhencedifferentmeaningsawordcanhave.Forexample,themeaningoftheword“bank”differswhenitappearsinthesentence“I’llswimacrosstherivertogettotheotherbank”versus“Icrossedthestreettogotothebank”.TransformersunlockedusecasesofNLPthatrequiredealingwithlongstreamsoftextandgaverisetothemostrecentadvancesinLLMs,suchasChatGPT.LLMsunderlietherapidriseofgenerativeAI(“genAI”),whichgeneratescontentbasedonsuitableprompts,andcanperformtasksbeyondlanguagerecognition.LLMsareneuralnetworksthataretrainedtopredictthenextwordinagivensequenceoftext.Toperformthistask,LLMslearntoabsorballthewrittenknowledgeonwhichtheyweretrained.Asaresult,theirpredictionisusuallyaccurateevenfortextsthatrequirenuanceorfieldknowledge.LLMscanbefine-tunedforspecifictaskswithspecialiseddata.Forexample,ChatGPTisbasedonanLLMrefinedwithhumanfeedbacktogeneratemoreusefulresponses.KeycharacteristicsofgenAIarethatitcanbeusednotjustbyasmallsetofspecialistsbutbyvirtuallyeverybodyandthatitcaneasilyextractinsightsfromunstructureddata.MachinelearningandAIincentralbanks:usecasesWhatarethecurrentusecasesofmachinelearningandAIincentralbanks?Theycanbestbeorganisedbyscope:(i)informationcollectionandstatisticalcompilation;(ii)macroeconomicandfinancialanalysistosupportmonetarypolicy;(iii)oversightofpaymentsystems;and(iv)supervisionandfinancialstability.Thissectionprovidesrelevantexamplesineacharea.Moreinformationontheselectedexamples,aswellasabroaderlistofusecases,canbefoundintheannex.2BISBulletinInformationcollectionEnsuringtheavailabilityofhigh-qualitydataasinputsforeconomicanalysisandforstatisticscompilationandproductionisamajorchallengeforcentralbanks.Issuesincludedatacleaning,sampling,representativenessandmatchingnewdatatoexistingsources.Thesteadilyincreasingvolumeandcomplexityofdatanecessitateefficientandflexibledataqualitytools.Toprovidehigh-qualitymicrodata,centralbanksareprogressivelyusingmachinelearningtechniques.Isolationforestsareparticularlysuitableforthelargeandgranulardatasetstypicalofcentralbanks,owingtotheirscalabilityandabilitytoidentifyoutliersregardlessoftheshapeofthedata’sdistribution.Therearealsobenefitstoatwo-stepapproach:initially,amodelautonomouslyidentifiespotentialoutliers,whicharethenreviewedbyexpertswhoprovidefeedbacktorefinethealgorithm.Thisapproachbalancesthevalueofdomainexpertisewiththecostsofhumaninputs.Byanalysingdifferentmethodstoexplaintheoutlierclassification,thisapproachcanovercometheissueof“blackbox”machinelearningmodelslacking“explainability”,whichisdiscussedbelow.Moreover,explainablemachinelearningmethodsprovideexpertswithguidanceonwhichdatapointswarrantmanualverification.MacroeconomicandfinancialanalysistosupportmonetarypolicyCentralbanksrelyextensivelyonmacroeconomicandfinancialanalysistosupportmonetarypolicy.Inacomplexenvironment,asignificantchallengeistoefficientlyextractinformationfromawidearrayoftraditionalandnon-traditionaldatasources.Machinelearningoffersvaluabletoolsinthisarea.Neuralnetworkscan,forexample,breakdownservicesinflationintodifferentcomponents,revealinghowmuchinflationisduetopastpriceincreases,inflationexpectations,theoutputgaporinternationalprices.Suchmodelscanprocessmoreinputvariablesthantraditionaleconometricones,allowingcentralbankstousegranulardatasetsinsteadofmoreaggregatedata.Anotheradvantageisneuralnetworks’abilitytoreflectcomplexnon-linearitiesinthedata,whichcanhelpmodellerstobettercapturenon-linearities,fromthezerolowerboundtounequalassetholdingsandshiftsininflationdynamics.Otherusecasesareobtainingreal-timeestimates(nowcasts)ofinflationexpectationsorsummarisingeconomicconditionsovertime.Forexample,randomforestmodelscanidentifysocialmediapoststhatarerelatedtopricesandthenfeedthemintoanotherrandomforestmodelthatclassifieseachpostasreflectinginflation,deflationorotherexpectations.Thedifferenceinthedailycountsofsocialmediapostsforhigherversuslowerinflationgaugesinflationexpectations.Similarly,socialmediapostscanbeusedtotrackthecredibilityofcentralbankmonetarypolicywiththewiderpublic.AnotherexampleistheuseofopensourceLLMsfine-tunedwithfinancialnewstosummariseeconomicconditionnarrativesoveralongtimespan.Modelscanprocesseganecdotaltextsfrominterviewswithentrepreneurs,economistsandmarketexpertstoproduceatimeseriesoftheir(positiveornegative)sentimentvalue.ThesentimentindexcanthenbeusedtonowcastGDPorpredictrecessions.AdaptingLLMstocentralbankingterminologycanbringfurthergains,asshownbythecentralbanklanguagemodels(CB-LM)projectdevelopedattheBIS(Gambacortaetal(2024)).ThisapproachusesthousandsofcentralbankspeechesandresearchpaperscompiledbytheBISCentralBankHubtoadaptwidelyusedopensourcefoundationLLMsissuedbyGoogleandMeta.Thisadditionaltrainingfocusedoncentralbankingtextsincreasedaccuracyfrom50–60%to90%ininterpretingcentralbankterminologyandidioms.IthasalsoimprovedperformanceintaskssuchasclassifyingFederalOpenMarketCommitteepolicystancesandpredictingmarketreactionstomonetarypolicyannouncements.OversightofpaymentsystemsWellfunctioningpaymentsystemsarefundamentaltothestabilityofthefinancialsystem,yetthevastamountoftransactiondata,oftenwithahighlyskeweddistribution,poseschallengesindistinguishinganomaloustransactionsfromregularones.CorrectlyidentifyinganomalouspaymentsiscrucialtoBISBulletin3addressingissuessuchaspotentialbankfailures,cyberattacksorfinancialcrimesinatimelymanner.Moneylaundering,inparticular,underminestheintegrityandsafetyoftheglobalfinancialsystem.TheBISInnovationHub’sProjectAurorausessyntheticmoneylaunderingdatatocomparefraudulentpaymentidentificationbyvarioustraditionalandmachinelearningmodels(BISIH(2023)).Themodels,whichincludeisolationforestsandneuralnetworks,undergotrainingwithknown(synthetic)moneylaunderingtransactionsandthenpredictthelikelihoodofmoneylaunderinginunseendata.Machinelearningmodelsoutperformtherule-basedmethodsprevalentinmostjurisdictionsortraditionallogisticregressions.Graphneuralnetworks,whichtakepaymentrelationshipsasinput,identifysuspecttransactionnetworksparticularlywell.Thesemodelscanfunctioneffectivelyevenwithdatapoolingthatsafeguardsconfidentiality,suggestingthatcooperationtojointlyanalysemultipledatabasescanbesecureandbeneficial.Thisillustratesthepotentialformorecooperationbetweenauthorities.Anotherapproachforoverseeingpaymenttransactionsinvolvestheuseofunsupervisedlearningmethodstoautomaticallysingleouttransactionsthatareworthcloserinspection.Forexample,auto-encodermodels,neuralnetworkswhereboththeinputandoutputlayerslookatthesamedata,distinguishtypicalfromanomalouspaymentsandcandetectnon-lineardynamicssuchasbankruns.Insimulations,thesemodelseffectivelyidentifiedpatternsofsignificantbankdepositwithdrawalsoverseveraldays.Auto-encodersalsoidentifiedarangeofreal-lifeanomaliesinpaymentsystems,includingoperationaldisruptionsamongimportantdomesticbanks.SupervisionandfinancialstabilitySupervisorsanalyseabroadrangeofdatasourcestoefficientlyoverseefinancialinstitutions.Thesesourcesincludetextdocumentssuchasnewsarticles,internalbankdocumentsorsupervisoryassessments.Siftingthroughthiswealthofinformationtoextractrelevantinsightscanbetime-consuming,andwiththeeverincreasingvolumeofdataitbecomesnearlyinsurmountable.Moreover,analysesrelatedtoclimateandcyberriskshaveemergedassupervisorypriorities,buttheylackthecomprehensivedatainfrastructurealreadyinplaceformore“traditional”risks.Oneavenuepursuedbymanycentralbanksistoconsolidatethewealthofinformationinoneplaceandhelpsupervisoryanalysisofunstructureddata.Forexample,modelsfine-tunedonsupervisorycontenttogetherwithNLPtechniquescanclassifypublicandsupervisorydocuments,undertakesentimentanalysesandidentifytrendingtopics,asdoneintheECB’splatformAthena.Trainingmodelsonalargebodyoftextcombinedwithanexpert-definedlexiconofrelevantwordsandclausescanalsohelpautomatethediscoveryofexcerptscontaininginformationondifferentrisks.Suchmodels,forexampletheFederalReserve’sLEX,facilitatesupervisors’accesstorelevantinformationscatteredacrossmillionsofdocumentsandreducethetimespentreviewingdocumentsubmissions.Classificationmodels,leveragingtree-basedtechniquesorneuralnetworks,canalsohelpidentifyindividualborrowersforwhichlendersunderestimatepotentialcreditlosses,ataskforwhichtheCentralBankofBrazilcreatedADAM.Neuralnetworksthatincludethefirstlayersofatrainednetworkcanimproveidentificationofborrowerswithhighexpectedlosses.Supervisorscanthenrequirefinancialinstitutionstoprovisionexposuresthatarenotsufficientlycovered.BalancingopportunitiesandchallengesTheaboveexamplesillustratetheopportunitiesformachinelearningandAItotackleproblemsattheheartofcentralbankmandates.Yettherearealsonewchallenges,somemoregeneralandothersmorespecifictocentralbanks.Ageneralchallengeistheconflictbetweenaccuracyand“interpretability/explainability”.Sophisticatedmachinelearningmodelscanbecomenearperfectatprediction.Butsincemanyvariablesinteractincomplexandnon-linearways,itcanbedifficulttointerprethowimportantdifferentinputvariablesarefortheresult.Goodpredictioncanhencecomeatthecostofacceptingthattheunderlying4BISBulletinmodelisa“blackbox”.Thiscan,forexample,makeitchallengingtoassessdiscriminatorybiasesinalgorithms,especiallywhenthesehavebeentrainedonbiaseddatasets.Limitedexplainabilityfurthermeansthatitisdifficulttoexplainmodelbehaviourinhumanterms;forexample,whyinflationispredictedtogouporwhyamortgageapplicationwasrejected.ForgenAImodels,theissuegoesevenfurther,astheysufferfromthe“hallucinationproblem”.Thesemodelsmightpresentafactuallyincorrectanswerasifitwerecorrect.ThehallucinationproblemimpliesthatLLMsneedhumansupervision,especiallyintasksrequiringlogicalreasoning(Perez-CruzandShin(2024)).Forcentralbanks,theuseofunstructureddatacanoffervaluableinformationthatcanhelpsolvepreviouslyintractableproblems.Manuallyconvertingunstructureddata,inparticulartext,intostructuredformistime-consuming,pronetohumanerrorandinfeasibleatalargerscale.Astheaboveexamplesmakeclear,LLMscanhelpcentralbanksanalyseawiderangeoftextualdata,suchassocialmediaactivity,financialnewsandcentralbanks’ownreports(confidentialorpublic).Theuseofunstructuredandoftenpersonaldata,however,posesnewchallengesintermsoflegalframeworksanddataprivacy.Traditionally,mostdatawerecollectedandhostedwithinpublicinstitutionswithclearlydefinedaccessrightsandsounddataqualityassuranceprocesses.Butnow,largeswathesofdataarecreatedbyindividualsandfirmsandresidewiththeprivatesector,sometimeswithlittledocumentationpubliclyavailable.Trainingorfine-tuningLLMsmayrequiresignificantamountsofdata,whichcanbeobtained,forexamplebywebscrapinginformationfrommarketplatformsorsocialmedia,butforwhichlegalframeworksoftenremainunclearabouthowandforwhatpurposestheycanbeused.Theavailabilityofunstructuredpersonaldataalsoraisesconcernsaboutethicsandprivacy.Citizenshavearighttoprivacyandmightfeeluncomfortablewithcentralbanksscrutinisingtheirdata.Whileprivacy-enhancingtechnologiesaresteadilyimproving,theyarenotyetadefaultinAImodels.GreateruseofAIcouldalsohaveprofoundimplicationsforcentralbanks’investmentsininformationtechnology(IT)andhumancapital.Providingadequatecomputingpowerandsoftware,aswellastrainingexistingstaff,involveshighupfrontcosts.Meanwhile,hiringnewstafforretainingexistingstaffwiththerightmixofeconomicunderstandingandprogrammingskillscanbechallenging:thereishighdemandforthisresource,andpublicinstitutionsoftencannotmatchprivatesectorsalariesfortopdatascientists.However,theseinvestmentscould,overtime,yieldincreasedproductivity.TheaboveexamplessuggestthattheuseofmachinelearningandAIcanmarkedlyraisestaffproductivity–inparticularinsometime-intensivetasksthatrequirecognitivework,suchassummarisingandextractinginformationfromtext(Brynjolfssonetal(2023),NoyandZhang(2023)).Forexample,AIsystemscouldactas“co-pilots”tohumansupervisoryteamsbylearningfromacombinationofregulatorydata,priorsupervisoryactionsandbroadermarketdevelopments.AIcouldalsoimproveanalysisbyfreeingupeconomists’timeforinterpretingdataratherthancollectingandcleaningit.YetAIwillnotmakehumansobsolete.Incorporatingexpertfeedbackcanimprovemodelsandmitigatethehallucinationproblem.Thebusinessexpertiseofstaffhelpstoidentifywheremodelsaddthemostvalueaswellashowtoadaptthemtocentralbank-specifictasks.Finally,theriseofLLMsandgenerativeAIhasrenewedconcernsaboutdependenceonafewexternalproviders.Largeeconomiesofscalemeanthatthemostpowerfulfoundationmodelsareprovidedbyjustafewlargetechnologyfirms.Beyondthegeneralrisksthatmarketconcentrationposestoinnovationandeconomicdynamism,thishighconcentrationofresourcescouldcreatesignificantfinancialstability,operationalandreputationalrisks.Forexample,greaterrelianceonLLMsandgenAIbyjustafewcompaniesmakesthefinancialsystemsusceptibletospilloversfromITfailuresorcyberattacksontheseproviders.Outagesamongproviderscouldalsoleadtooperationalrisksforcentralbanksandhaverepercussionsfortheirabilitytofulfiltheirmandates.Theriskofoperationalproblemsleadingtoreputationalcostsloomslargeascentralbanks’greatestassetisthepublic’strust(Doerretal(2022)).Atthesametime,ifmanyinstitutionsadoptthesamefewbestinclassalgorithms,theirbehaviourduringstressepisodesmightlookincreasinglyalikeandleadtoundesirablephenomenasuchasliquidityhoarding,interbankrunsandfiresales(DanielsonandUthemann(2023)).BISBulletin5Theselessonsunderscorethebenefitsofcooperationamongcentralbanksandotherpublicauthorities.Knowledge-sharingandthepoolingofexpertisearewellestablishedinthecentralbankingcommunity,andcentralbanks’publicpolicymandategivesconsiderablescopeforcooperation,aswellastoestablishacommunityofpracticeformachinelearningandAI.CentralbankcollaborationandthesharingofexperiencescouldalsohelpidentifyareasinwhichAIaddsthemostvalueandhowtoleveragesynergies.Datastandardscouldfacilitatetheautomatedcollectionofrelevantdatafromvariousofficialsources,therebyenhancingthetrainingandperformanceofmachinelearningmodelsthatusemacroeconomicdata(Araujo(2023)).Additionally,thesharingofcodeorpre-trainedmodelsholdmuchpromise.Centralbankingisparticularlywellsuitedfortheapplicationofmachinelearningtechniquesgiventheavailabilityofstructuredandunstructureddataaswellastheneedforrigorousanalysisinsupportofpolicy.Thesynergiesbetweenmachinelearningandcorecentralbankingdisciplinessuchaseconomics,statisticsandeconometricsarelikelytoplacecentralbanksatthevanguardofadvancesinAI.ReferencesAraujo,DKG(2023):“gingado:amachinelearninglibraryfocusedoneconomicsandfinance”,BISWorkingPapers,no1122.Araujo,DKG,GBruno,JMarcucci,RSchmidtandBTissot(2022):“Machinelearningapplicationsincentralbanking:anoverview”,IFCBulletin,no57.———(2023):“Datascienceincentralbanking:applicationsandtools”,IFCBulletin,no59.Athey,SandGImbens(2021):“Machinelearningmethodsthateconomistsshouldknowabout”,AnnualReviewofEconomics,no11,pp685–725.BISInnovationHub(BISIH)(2023):ProjectAurora:thepowerofdata,technologyandcollaborationtocombatmoneylaunderingacrossinstitutionsandborders,May.Brynjolfsson,E,DLiandLRaymond(2023):“GenerativeAIatwork”,NBERWorkingPapers,no31161.Danielson,JandAUthemann(2023):“Ontheuse