版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
Theuseofartificialintelligenceandmachinelearningbymarketintermediaries
andassetmanagers
FinalReport
TheBoard
OFTHE
InternationalOrganizationofSecuritiesCommissions
FR06/2021
September2021
PAGE\*roman
iii
Copiesofpublicationsareavailablefrom:
TheInternationalOrganizationofSecuritiesCommissionswebsite
©InternationalOrganizationofSecuritiesCommissions2021.Allrightsreserved.Briefexcerptsmaybereproducedortranslatedprovidedthesourceisstated.
Contents
Chapter
Page
1
Executivesummary
1
2
Backgroundandscope
4
3
HowfirmsareusingAIandMLtechniques
6
4
IdentifiedrisksandharmsposedbytheuseofAIandML
9
5
Firms’responsetothepotentialrisksarisingfromtheuseofAIandML
14
6
Guidance
17
A1
HowregulatorsareaddressingthechallengescreatedbyAIandML
22
A2
Guidancepublishedbysupranationalbodies
34
A3
FeedbackStatement
38
PAGE
27
Chapter1-ExecutiveSummary
Background
ArtificialIntelligence(AI)andMachineLearning(ML)areincreasinglyusedinfinancialservices,duetoacombinationofincreaseddataavailabilityandcomputingpower.TheuseofAIandMLbymarketintermediariesandassetmanagersmaybealteringfirms’businessmodels.Forexample,firmsmayuseAIandMLtosupporttheiradvisoryandsupportservices,riskmanagement,clientidentificationandmonitoring,selectionoftradingalgorithmsandportfoliomanagement,whichmayalsoaltertheirriskprofiles.
Theuseofthistechnologybymarketintermediariesandassetmanagersmaycreatesignificantefficienciesandbenefitsforfirmsandinvestors,includingincreasingexecutionspeedandreducingthecostofinvestmentservices.However,thisusemayalsocreateoramplifycertainrisks,whichcouldpotentiallyhaveanimpactontheefficiencyoffinancialmarketsandcouldresultinconsumerharm.Theuseof,andthecontrolssurrounding,AIandMLwithinfinancialmarketsis,therefore,acurrentfocusforregulatorsacrosstheglobe.
IOSCOidentifieditsworkontheuseofAIandMLbymarketintermediariesandassetmanagersasakeypriority.TheIOSCOBoardapprovedamandateinApril2019forCommittee3onRegulationofMarketIntermediaries(C3)andCommittee5onInvestmentManagement(C5)toexaminebestpracticesarisingfromthesupervisionofAIandML.
1
Thecommitteeswereaskedtoproposeguidancethatmemberjurisdictionsmayconsideradoptingtoaddresstheconductrisksassociatedwiththedevelopment,testinganddeploymentofAIandML.
PotentialrisksidentifiedintheConsultationReport
IOSCOsurveyedandheldroundtablediscussionswithmarketintermediariesandconductedoutreachtoassetmanagerstoidentifyhowAIandMLarebeingusedandtheassociatedrisks.ThefollowingareaswerehighlightedintheConsultationReportreleasedinJune2020
2
wherepotentialrisksandharmsmayariseinrelationtothedevelopment,testinganddeploymentofAIandML:
Governanceandoversight;
Algorithmdevelopment,testingandongoingmonitoring;
Dataqualityandbias;
Transparencyandexplainability;
Outsourcing;and
Ethicalconcerns.
1 BoardPriorities-IOSCOworkprogramfor2019,March25,2019,availableat:
/library/pubdocs/pdf/IOSCOPD625.pdf
2 Theuseofartificialintelligenceandmachinelearningbymarketintermediariesandassetmanagers,IOSCOBoardConsultationReport,June2020,availableat:
/library/pubdocs/pdf/IOSCOPD658.pdf
IOSCOGuidance
BasedontheresponsesreceivedtotheConsultationReport,thisfinalreportprovidesguidancetoassistIOSCOmembersinsupervisingmarketintermediariesandassetmanagersthatutiliseAIandML.
TheguidanceconsistsofsixmeasuresthatreflectexpectedstandardsofconductbymarketintermediariesandassetmanagersusingAIandML.Althoughtheguidanceisnotbinding,IOSCOmembersareencouragedtoconsiderthesemeasurescarefullyinthecontextoftheirlegalandregulatoryframeworks.IOSCOmembersandfirmsshouldalsoconsidertheproportionalityofanyresponsewhenimplementingthesemeasures.
TheuseofAIandMLwilllikelyincreaseasthetechnologyadvances,anditisplausiblethattheregulatoryframeworkwillneedtoevolveintandemtoaddresstheassociatedemergingrisks.Therefore,thisreport,includingthedefinitionsandguidance,maybereviewedinthefuturetoremainuptodate.
Measure1:Regulatorsshouldconsiderrequiringfirmstohavedesignatedseniormanagementresponsiblefortheoversightofthedevelopment,testing,deployment,monitoringandcontrolsofAIandML.Thisincludesadocumentedinternalgovernanceframework,withclearlinesofaccountability.SeniorManagementshoulddesignateanappropriatelyseniorindividual(orgroupsofindividuals),withtherelevantskillsetandknowledgetosignoffoninitialdeploymentandsubstantialupdatesofthetechnology.
Measure2:RegulatorsshouldrequirefirmstoadequatelytestandmonitorthealgorithmstovalidatetheresultsofanAIandMLtechniqueonacontinuousbasis.ThetestingshouldbeconductedinanenvironmentthatissegregatedfromtheliveenvironmentpriortodeploymenttoensurethatAIandML:
behaveasexpectedinstressedandunstressedmarketconditions;and
operateinawaythatcomplieswithregulatoryobligations.
Measure3:Regulatorsshouldrequirefirmstohavetheadequateskills,expertiseandexperiencetodevelop,test,deploy,monitorandoverseethecontrolsovertheAIandMLthatthefirmutilises.Complianceandriskmanagementfunctionsshouldbeabletounderstandandchallengethealgorithmsthatareproducedandconductduediligenceonanythird-partyprovider,includingonthelevelofknowledge,expertiseandexperiencepresent.
Measure4:Regulatorsshouldrequirefirmstounderstandtheirrelianceandmanagetheirrelationshipwiththird-partyproviders,includingmonitoringtheirperformanceandconductingoversight.Toensureadequateaccountability,firmsshouldhaveaclearservicelevelagreementandcontractinplaceclarifyingthescopeoftheoutsourcedfunctionsandtheresponsibilityoftheserviceprovider.Thisagreementshouldcontainclearperformanceindicatorsandshouldalsoclearlydeterminerightsandremediesforpoorperformance.
Measure5:RegulatorsshouldconsiderwhatlevelofdisclosureoftheuseofAIandMLisrequiredbyfirms,including:
RegulatorsshouldconsiderrequiringfirmstodisclosemeaningfulinformationtocustomersandclientsaroundtheiruseofAIandMLthatimpactclientoutcomes.
RegulatorsshouldconsiderwhattypeofinformationtheymayrequirefromfirmsusingAIandMLtoensuretheycanhaveappropriateoversightofthosefirms.
Measure6:RegulatorsshouldconsiderrequiringfirmstohaveappropriatecontrolsinplacetoensurethatthedatathattheperformanceoftheAIandMLisdependentonisofsufficientqualitytopreventbiasesandsufficientlybroadforawell-foundedapplicationofAIandML.
Chapter2-BackgroundandScope
PreviousIOSCOworkinthisarea
IOSCOhasundertakenseveralworkstreamsontheuseofAIandMLinfinancialmarkets,including:
CommitteeonEmergingRisks(CER):TheCERundertookamandateontheuseofnoveltechnologiesdeployedbyregulatorstoincreasetheefficiencyandeffectivenessofsupervisoryandoversightprogramsandpublishedareportinFebruary2017.
3
CERexaminedtheregulatoryuseoftoolssuchasbigdataanalyticsanddatavisualisationtechnologies;AIandML,anddeeplearningtechnologies;anddistributedledgertechnologies.
CommitteeonRegulationofSecondaryMarkets(C2):C2publishedareportinApril2013onTechnologicalChallengestoEffectiveMarketSurveillanceIssuesandRegulatoryTools.
4
Thereportmaderecommendationstohelpmarketauthoritiesaddressthetechnologicaldifficultiesfacingeffectivemarketsurveillance.
IOSCOFintechNetwork:TheIOSCOFintechNetworkwasestablishedinMay2018tofacilitatethesharingofknowledgeandexperiencesamongIOSCOmembers.TheIOSCOFintechNetworkconsideredtheethicalimplicationsoftheuseofAIandMLtechnologies.
IOSCOMandate
Buildingonitspreviouswork,IOSCOreleasedaConsultationReportontheuseofAIandMLbymarketintermediariesandassetmanagersinJune2020,proposingguidancetoaddressthepotentialrisksandharmsthatmaybecausedbytheuseofAIandMLbythesemarketintermediariesandassetmanagers.Theproposedguidancelookedtohelpensurethatmarketintermediariesandassetmanagershave:
appropriategovernance,controlsandoversightframeworksoverthedevelopment,testing,useandperformancemonitoringofAIandML;
staffwithadequateknowledge,skillsandexperiencetoimplement,oversee,andchallengetheoutcomesoftheAIandML;
robust,consistentandclearlydefineddevelopmentandtestingprocessestoenablefirmstoidentifypotentialissuespriortofulldeploymentofAIandML;and
appropriatetransparencyanddisclosurestotheirinvestors,regulatorsandotherrelevantstakeholders.
3 IOSCOResearchReportonFinancialTechnologies(Fintech),February2017,availableat:
/library/pubdocs/pdf/IOSCOPD554.pdf
4 TechnologicalChallengestoEffectiveMarketSurveillanceIssuesandRegulatoryTools,August2012,availableat:
/library/pubdocs/pdf/IOSCOPD389.pdf
ThisFinalReportconfirmstheguidanceproposedintheConsultationReport,amendingittotakeaccountofresponsesreceivedasappropriate.
DefiningthetermsAlandMLforthisreport
ArtificialIntelligence
ThetermArtificialIntelligence,firstcoinedbydatascientistJohnMcCarthy
5
in1956,isdefinedas“thescienceandengineeringofmakingintelligentmachines”,orsimply,thestudyofmethodsformakingcomputersmimichumandecisionstosolveproblems.AIincludestaskssuchaslearning,reasoning,planning,perception,languageunderstandingandrobotics.AIinthefinancialservicesindustryisstillinitsrelativeinfancyandispoisedtobecomemorecommon,andwiththatwillcomelegal,ethical,economicandregulatorychallenges.
MachineLearning
ThetermMachineLearningisasubsetandapplicationofAI,whichfocusesonthedevelopmentofcomputerprograms-designedtolearnfromexperiencewithoutbeingexplicitlyprogrammedtodoso.
TherearethreecategoriesofMLalgorithms–supervisedlearning,unsupervisedlearningandreinforcementlearning.Thesecategoriesareusedbasedonthetypeofdataavailableandthelevelofhumaninterventionrequiredinprovidingfeedback.DeepLearninginvolvestrainingneuralnetworks(computingsystems)withmanylayersofunits,inspiredbythestructureofthehumanbrainandcanalsoincludeanyofthesecategories:
6
Supervisedlearning:thealgorithmisfedaninitialsetofdatathathasbeenlabelled.Basedonthistrainingset,thealgorithmwilllearnclassificationrulesandpredictthelabelsfortheremainingobservationsinthedataset.
Reinforcementlearning:thealgorithmisfedaninitialsetofdatathathasnotbeenlabelledandisaskedtoidentifyclustersofobservationsunderpinnedbysimilarcharacteristics.Asitchoosesanactionforthedatapoints,itreceivesfeedbackthathelpsitlearn.
7
Unsupervisedlearning:thealgorithmdetectspatternsinthedatabyidentifyingclustersofobservationsunderpinnedbysimilarcharacteristics–ituncoversthestructureofthedataonitsown.
5 WhatisAI?availableat:
/artificial-intelligence/index.html
6Deeplearningisamethodthatanalysesdatainmultiplelayers,startingwithlearningaboutsimpleconceptsthenlearningmorecomplexconcepts.Deeplearningcanbeusedforallthreecategoriesofmachinelearningalgorithms.
Adaptedfrom‘WhatisMachineLearning?’at
/cloud/learn/machine-
learning.
Afamousexampleofthisisthe“move37”inthegameofGo:whenGoogle’sAlphaGo5algorithmwaspittedagainstprofessionalGoplayerLeeSedolinMarch2016,makingamoveonthe37thturnthatwaspreviouslyunimaginable.Thisalgorithmuseddeeplearning,aformofMLtechniquethatefficientlylearnsassociationsandstatisticalpatternsfromaverylargedataset.
7 RSutton,ABarto,ReinforcementLearning:anintroduction,MITPress,1998.
Chapter3–HowfirmsareusingAIandMLtechniques
AIandMLusebymarketintermediariesandassetmanagers
Marketintermediariesandassetmanagers’useofAIandMLisgrowing,astheirunderstandingofthetechnologyanditsutilityevolves.TheriseintheuseofelectronictradingplatformsandtheincreaseofavailabledatahaveledfirmstoconsidertheuseofAIandMLinseveralareas,includingfortheirtradingandadvisoryactivities,aswellasforriskmanagementandcompliance.
TheIOSCOfirmengagementrevealedthatwithinfinancialmarkets,AIandMLarebeingadoptedtoaugmentexistingprocessesandactivities,withaviewtoreducingcostandincreaseefficiency.AIandMLarefreeingupresourcestofocusonmorecognitiveaspects,suchasstrategy,portfolioselectionandgeneratinginvestmentideas.Marketintermediariesaredeployingthistechnologyin:
Advisoryandsupportservices;
Riskmanagement;
Clientidentificationandmonitoring;
Selectionoftradingalgorithm;and
Assetmanagement/Portfoliomanagement.
TheuseofAIandMLbyassetmanagersappearstobeinitsnascentstagesandismainlyusedtosupporthumandecision-making.AIandMLtechniquesarebeingusedto:
Optimiseportfoliomanagement;
Complementhumaninvestmentdecision-makingprocessesbysuggestinginvestmentrecommendations;and
Improveinternalresearchcapabilities,aswellasbackofficefunctions.
SomeassetmanagersarealsobeginningtouseAIandMLfororderexecution,brokerselectionandorderroutingpurposes(includingthroughmethodssuchasalgo-wheels).
8
Advisoryandsupportservices
AccordingtoIOSCO’sindustryengagement,mostrobo-advisorsorautomatedinvestmentadvisorsusesimple,rule-based(i.e.,deductive)algorithms,althoughsomearebeginningtoutilisepredictiveMLalgorithms.WhereMLisusedtoprovideadvisoryservices,mostfirmshavemanualinterventionprocesses.Theautomatedadvicesystemis,therefore,usuallylimitedtogeneratingpotentialadviceorassetallocationfortheinvestmentadvisertoreview.TheinvestmentadvisercanthenusethisAI-generatedadviceasappropriateand,wheresuitable,tomakearecommendationtotheclient.
8Algo-wheelsmayperformdifferentfunctionsindifferentpartsoftheworld.Inthiscontext,wedefinealgo-wheelstomeanasoftware/modelthataggregatesdatatoselectthestrategyandbrokerthroughwhichtorouteordersbeforegeneratingareportthatsetsoutthereasonbehindhowandwhereaparticulartradewasmade.
Riskmanagement
Riskmanagementinvolvesusingdatatopriceandmanageexposure,includingcredit,market,operationalandliquidityrisk.MarketintermediariesareharnessingML-basedriskmanagementsystemsforcreditriskmonitoring,whichcouldhelpprovideanearly-warningindicatorofpotentialcustomerdefaultsandcanhelpcreateadynamicmeasurementofacustomer’sriskprofiletobetterunderstand,forexample,whentowriteoffadebt.
MLisimprovingtheefficiencyofbackofficeprocessingandreportingfunctionswithinmarketintermediaries.Itisalsoincreasinglyusedtovisualisemarketriskbyanalysingvolatilitytrends,andtogaugeliquidityriskbyanalysingmulti-dimensionalriskandexposuredata.MLalgorithmsareincreasinglyusedtomonitorstaffe-mailsbyleveragingadvancedpatternrecognition.
Somemarketmakers,whoprovideliquiditytomarketparticipantsbyfacilitatingtransactionsareadoptingMLmodelsandreinforcementlearningtominimisetheirinventoryriskandmaximisetheutilityoftheirbalancesheet.
Similarly,someassetmanagersareseekingtoharnesstheadvantagesofthesetechniquesinriskmanagement.Somehedgefundsandassetmanagersareautomatingriskmanagementandcomplianceprocessesbytrackingthebehaviourofindividualportfoliomanagers,automatingexecutionqualityreportsandassessingmarketliquidityrisk.
Clientidentificationandmonitoring
MLhasallowedmarketintermediariestoautomatetheirclientonboarding,frauddetection,moneylaunderingandcyber-attackmonitoring.MarketintermediariesshouldgenerallyundertakeKnowYourCustomer(KYC)checksbeforeonboardingclientsandsellingthemproductsandservices.KYCentailscollectionandverificationofcomprehensivepersonalinformationfrompotentialclients,involvingprocessingunstructuredmetadata.
InductivereasoningalgorithmshelpaccuratelyidentifyfakephotoIDs,whilerecognisingdifferentphotosofthesameperson.MLcanalsobeusedforscreeningandmonitoringclientsandtransactionsagainstsanctionsorotherlists,todetectevidenceofpossiblemoneylaundering,terroristfinancingandotherfinancialcrimes.
Selectionoftradingalgorithms
Manymarketintermediariescurrentlyofferasoftwaresolutiontotheirclientsthatselectsanappropriatetradingstrategyand/orabrokerdependingonthemarketsituationandtradingobjectivesforbestexecutionpurposes,oftennamedanalgowheel.Algo-wheelsseektoclassifyhistoricaltradingandperformance,predicttheperformanceofstrategiesandbrokeralgorithms,andrecommendwhentouseaparticularalgorithm.Usingalgo-wheelstooptimallyexecutesimplerordersallowsthetradertofocusonmorecomplextradeflows.
Predictivedataanalyticsareenablingtheidentificationofpotentialmarketconditionsconducivetoaflashcrashtypeevent.
Assetmanagement/Portfoliomanagement
Supervisedlearning,whereafunctionisinferredfromlabelledtrainingdata,hasbeenusedforsmall-scalepatternrecognitionandsimplepredictionmodelstoaidtradingdecisionswithinassetmanagersandmarketintermediariesforseveralyears.
Marginpressureandcompetitionisdrivinginnovationamongstassetmanagers.Tocompete,someactivemanagersthattraditionallyemphasisedtheirfundamentalresearchcapabilitiesarebeginningtoexpandalreadyexistingquantitativeapproachesbyleveragingdiversifieddatasources–suchassocialmedia,geospatialdata,andothermetadatatoenhanceinternalresearch.
Incertaincases,thesemethodsarebeingusedforasset-allocationandpricingsuchasidentifyingrelationshipsinmetadatawhichcouldbeusedtogeneratetradeideasoralphasignalsandforecastassetpricesbasedonhistoricalpricesaswellascurrenttrends.Moreover,assetmanagersmayapplythesetechniquestopricenewinvestmentproductscompetitively.Theydosobyusingdataonexistinginvestmentproductswithsimilarstructuresand/orconstituentassets.Otherapplicationsincludeinvestmentcompliancechecks,transferagencyactivitiesandclientservicing.
Chapter4–IdentifiedPotentialRisksandHarmsPosedbytheUseofAIandML
IOSCO’sindustryengagementrevealedthattheevolutionandincreasingadoptionofAIandMLmayraise(eitherintentionallyorunintentionally)anumberofconductconcernsformarketintermediariesandassetmanagers,regarding:
Governanceandoversight;
Algorithmdevelopment,testingandongoingmonitoring;
Dataqualityandbias;
Transparencyandexplainability;
Outsourcing;and
Ethicalconcerns.
Governanceandoversight
FirmsimplementingAIandMLmostlyrelyonexistinggovernanceandoversightarrangementstosignoffandoverseethedevelopmentanduseofthetechnology.Inmostinstances,theexistingreviewandseniorleadership-levelapprovalprocesseswerefollowedtodeterminehowrisksweremanaged,andhowcompliancewithexistingregulatoryrequirementswasmet.AIandMLalgorithmsweregenerallynotregardedasfundamentallydifferentfrommoretraditionalalgorithmsandfewfirmsidentifiedaneedtointroducenewormodifyexistingproceduralcontrolstomanagespecificAIandMLrisks.
Somefirmsindicatedthatthedecisiontoinvolveseniorleadershipingovernanceandoversightremainsadepartmentalorbusinesslineconsideration,ofteninassociationwiththeriskandITordatasciencegroups.Therewerealsovaryingviewsonwhethertechnicalexpertiseisnecessaryfromseniormanagementparticipatinginandoverseeingcontrolfunctionssuchasriskmanagement.Despitethis,mostfirmsexpressedtheviewthattheultimateresponsibilityandaccountabilityfortheuseofAIandMLwouldliewiththeseniorleadershipofthefirm.
SomefirmsnotedthatthelevelofinvolvementofriskandcompliancetendstofocusprimarilyondevelopmentandtestingofAIandMLratherthanthroughthelifecycleofthemodel(i.e.,implementationandongoingmonitoring).Generally,onceimplemented,somefirmsrelyonthebusinesslinetoeffectivelyoverseeandmonitortheuseoftheAIandML.Respondentsalsonotedthatrisk,complianceandauditfunctionsshouldbeinvolvedthroughoutallstagesofthedevelopmentofAIandML.
ManyfirmsdidnotemployspecificcompliancepersonnelwiththeappropriateprogrammingbackgroundtoappropriatelychallengeandoverseethedevelopmentofMLalgorithms.Withmuchofthetechnologystillatanexperimentalstage,thetechniquesandtoolkitsatthedisposalofcomplianceandoversight(riskandinternalaudit)functionscurrentlyseemlimited.Insomecases,thisiscompoundedbypoorrecordkeeping,resultinginlimitedcompliancevisibilityastowhichspecificbusinessfunctionsarereliantonAIandMLatanygivenpointintime.
Algorithmdevelopment,testingandongoingmonitoring
Itisimportantthatfirmshaverobustandwellunderstooddevelopmentandtestingframeworksinplace,regardlessofwhethertheyareusingAIandMLortraditionalalgorithms.
Overall,IOSCO’sengagementshowedthatinmostcasesthereisnotanestablishedframeworkforspecificallydevelopingAIandML.Instead,manyfirmsusethesamedevelopmentandtestingframeworksthattheyusefortraditionalalgorithmsandstandardsystemdevelopmentmanagementprocesses.
Firmsthatusealgorithmsshouldconsidermaintaininganappropriatedevelopmentandtestingframework,whichisconsistentlyappliedacrossallrelevantaspectsofthebusiness.ThisisparticularlyimportantwherefirmsareusingAIandMLwithintheiralgorithmictradingstrategies.
Algorithmsrelyonqualitydataandmostfirmsrecognisetheneedforqualitydatainputs.Excessiveimmaterial,or“noisy”dataisunwanteddatathatdoesnotcontributetoarelationshipandmaycauseMLalgorithmstomissthesignalinthedataandbehaveunexpectedly.
Robustdevelopmentandtestingcontrolsarenecessarytodistinguishsignalsandstatisticallysignificantdatafromthenoise.Unliketraditionalalgorithms,asmoredataisprocessedbyMLalgorithms,theymaybehaveunpredictablyastheyareexposedtonewdatapatterns.MLalgorithmsshouldthereforealsobecontinuouslymonitoredthroughouttheirdeploymenttohelpensuretheydonotbehaveinexplicablyowingtoasubtleshiftintheoperatingconditionsorexcessivenoise.Whilesomefirmsnotedtheyreviewandreviseexistingmodelsasnecessary,somefirmsfocuslessonmanagingmodelsinthepost-productionphasesothattheyperformastheyshouldovertime.Unliketraditionalalgorithms,MLalgorithmscontinuallylearnanddevelopovertime.Itisimportantthattheyaremonitoredtoensurethattheycontinuetoperformasoriginallyintended.
Dataqualityandbias
TheperformanceofAIandMLisinherentlydependentonthequalityofthedatasetparticularlywhenbuildingthemodel.AccordingtoafewoftherespondentstotheConsultationReport,thequalityofthedatasetsusedinthelearningphasecanhaveamaterialimpactonthepotentialoutcomesandperformanceofAIandMLapplications.Intheviewoftherespondents,assuchitisakeyrisk.Accordingtosomerespondents,anotherriskrelatedtodatasetsisthesourcingofsufficientlylargedatasetstotraintheapplications,particularlywhenrelatedtoinvestmentdecisions.
Learnedbiasinthedatasetcanimpactthedecisionsmadebysuchalgorithmsandmayresultindiscriminativedecisionsandprovideundesirableoutcomestomarketparticipants.Forexample,askingquestionsphrasedinacertainwayorinacertainsequencemayleadtoaresponsethatintroducesimplicitorexplicitbiasfromtherespondents.Suchadataset,whereabiasmayhavebeenintroducedbyeitherthequestionerorbytherespondents,willinfluencetheconclusionsreachedbythealgorithm.Anyoutputbasedonsuchabiaswilllikelydegrade
theperformanceofthealgorithmmorequicklyovertimeandcouldresultinconsumerdetriment.
9
Biasmayalsobeinadvertentlyintroducedduringdatacleansing–apre-processingstepoftennecessarytoimprovedataquality.CleaningdatabeforeapplyingMLcanincreasethesignal-to-noiseratioallowingformoremeaningfulinterpretationstobederived.However,cleansingdatainvolvessubjectivedecisions,whichmayinadvertentlyintroduceotherbiases.
Transparencyandexplainability
TheeffectiveuseandadoptionofAIandMLrequirealgorithmsthatarenotonlyaccuratebutarealsounderstandablebyfirms(includingfrontline,complianceandriskpersonnel),marketcounterparties,clientsandregulators.Itispossiblethatrisksareintroducedifoutcomescannotbefullyexplainable.Whileincreasedtransparencyinfirms’useofAIandMLcouldimprovepublicunderstandingandconfidenceintheuseofthetechnology,excessivetransparencycouldcreateconfusionoropportunitiesforindividualstoexploitormanipulatethemodels.Theleveloftransparencywillalsodifferdependingontheaudience;forexample,aregulatormayrequiremoredetailedinformationthanaclient.TheseconsiderationsneedtobebalancedindeterminingtheappropriateleveloftransparencyintheuseofAIandML.
Itisimportantthatfirmsappropriatelydiscloseinformationabouttheirserviceofferings,tohelpclientsunderstandthenatureandrisksofproductsandserviceofferings,sothattheycanmakeinformeddecisions.ApplyingunexplainableMLalgorithmstorefineatradingstrategycouldexposethefirmtounacceptablelevelsoflegalandregulatoryrisk.Firmsofferingautomatedinvestmentservices,includingroboadviceandautomatedportfolioconstructionshouldappropriatelydisclosethenatureoftheuseofMLandtheautomationintheirserviceoffering.
SomeMLmodelsoperateasa“blackbox”withlimitedclarityonthereasoningbehindt
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 第19课 科学技术的重大成果课件
- 2024年专业电工施工协议典范
- 中国特色社会主义基本原理(上)
- 2024年度层电梯厅装潢协议模板
- 2024年无薪实习劳动协议
- 2024年仓库租赁法律协议细则
- 2024年型车辆采购协议
- 2024届安徽省合肥高升学校高三八校第一次适应性考试数学试题试卷
- 2024建筑业劳务施工协议文本
- 2024年全职劳务雇佣协议范本
- 2022高中学业水平考试信息技术会考知识点归纳总结(复习必背)
- 2022秋国开公共关系学形考任务3试题及答案
- 对外汉语教学趋向补语练习题
- 油茶栽培(普通油茶)课件
- 高锰酸钾安全使用说明书
- 建筑废弃材料回收利用公司创业项目计划书
- 2021年12月医院临床药师培训理论考核试题(呼吸专业)
- 圆管钢立柱柱吊装施工方案
- 高等天气学讲座--高空急流的次级环流及其与锋面系统的耦合课件
- 登高车检查表
- 5陶行知的学前教育思想课件(42页PPT)
评论
0/150
提交评论