agent-based models: understanding the economy from the ... · topical articles agent-based models:...
TRANSCRIPT
Topical articles Agent-based models: understanding the economy 173
Agent-based models: understanding the economy from the bottom upBy Arthur Turrell of the Bank’s Advanced Analytics Division.(1)
• Agent-basedmodellinghaslongenjoyedsuccessinthenaturalsciences,providinginsightsintoeverythingfromcancertotheeventualfateoftheUniverse.
• Itissuitedtomodellingcomplexsystemssuchastheeconomy,particularlythoseinwhichdifferentagents’interactionscombinetoproduceunexpectedoutcomes.
• Ineconomics,agent-basedmodelshaveshownhowbusinesscyclesoccur,howthestatisticsobservedinfinancialmarkets(suchas‘fattails’)arise,andhowtheycanbeausefultoolinformulatingpolicy.
Overview
Agent-basedmodelsexplainthebehaviourofasystembysimulatingthebehaviourofeachindividual‘agent’withinit.Theseagentsandthesystemstheyinhabitcouldbetheconsumersinaneconomy,fishwithinashoal,particlesinagas,orevengalaxiesintheUniverse.
Thestrengthofthesemodelsisthattheyshowhowevenverysimplebehaviourscancombinefromthe‘bottomup’torecreatethemorecomplexbehavioursobservedintherealworld.Anexamplewouldbehowthedecisionsofeachindividualfishcreatetheseeminglyorganisedandunpredictablemovementsoftheshoal.
This‘bottom-up’approachisincontrasttomodelswhichare‘topdown’,andwhichpresumehowagents’behaviourswillcombinetogether,sometimesbyassumingthatallagentsareidentical.Thedifferentapproacheshavedifferentstrengths.
Theagent-basedapproachtoproblem-solvingbeganinthephysicalsciencesbuthasnowspreadtomanyotherdisciplinesincludingbiology,ecology,computerscienceandepidemiology.Inrecentyears,agent-basedmodelshavebecomemorecommonineconomics,includingattheBankofEngland.
Therearechallengestotheiruse,includingtheneedforadvancedprogrammingskills,theneedtocarefullyinterprettheirresults,andhowtobestselecttheappropriatebehavioursfortheagents.Inparticular,therearenotalwaysobviouscriteriaforchoosingwhichbehavioursarethemost
realistic.Theseissueshavebeenabarriertotheirmorewidespreadadoptionineconomics.
Despitebeinglesswidelyused,agent-basedmodelshaveproducedmanyimportantinsightsineconomics,includinghowthestatisticsobservedinfinancialmarketsarise,andhowbusinesscyclesoccur.
Recently,theBankofEnglandhasdevelopedagent-basedmodelsoftwomarkets:corporatebondsandhousing.Theincreasedavailabilityofdataandcomputationalpowermeanthatagent-basedmodellinglookssettogainimportanceasatoolforbothunderstandingtheeconomy,andforexploringtheconsequencesofpolicyactions.
(1) TheauthorwouldliketothankDavidBholatandChrisCaifortheirhelpinproducingthisarticle.
ENVIRONMENT
Agents Agent-agent interactions Agent-environment interactions
Summary figure Schematic of the typical elements of an agent-based model
174 Quarterly Bulletin 2016 Q4
Introduction
Thisarticleconsidersthestrengthsofagent-basedmodellingandthewaysthatitcanbeusedtohelpcentralbanksunderstandtheeconomy.ThesemodelsprovideacomplementtomoretraditionaleconomicmodellingwhichwascriticisedfollowingtheGreatRecession.(1)
Agent-basedmodelshavedifferentstrengthsandweaknessestootherapproachesineconomics.Theyhaveadvantagesindescribinghowthedifferentactionsandpropertiesofindividualagentscombinetodrivetheoverallbehaviourofsystems.
Thisarticleexplainsthemotivationbehinddevelopingmodelsandhowtheycanbeusedtobetterunderstandtheworld,beforediscussingmoredetailsofwhatmakesagent-basedmodelsdifferent.Itthendescribeshowthesemodelswerefirstdevelopedandusedindisciplinesoutsideofeconomics.
Thegeneralstrengthsandweaknessesofagent-basedmodelsarediscussed,withexamplesofhowtheiradvantageshavebeenusedtoimprovetheunderstandingofcertainmarkets.
Thepenultimatesectionisanoverviewofhowagent-basedmodelsarebeingappliedineconomicsingeneralandattheBankofEnglandspecifically.Finally,somethoughtisgiventohowthislineofmodellingmightbeusedinthefuture.
Agent-basedmodelsarecalledbydifferentnamesindifferentdisciplines,includingMonteCarlosimulations(inthephysicalsciences),individual-basedmodels(inbiologyandecology),agent-basedcomputationaleconomicsmodels(ineconomics)andmulti-agentsystems(incomputerscienceandlogistics).Thisarticleuses‘agent-basedmodel’torefertoanymodelinwhichtheinteractionsandbehavioursofalargenumberofheterogeneousagentsaresimulated,manuallyorbyacomputer.
ModellingModellingisubiquitousacrossacademicdisciplines,governmentsandtheprivatesector.Mostmodelsattempttoisolatetheunderlyingcausesofthebehaviourofsystems,removingextraneousdetailandfocusingonwhatmatterstothehypothesis,questionorpolicyunderconsideration.Modelsmaybeassimpleasthoughtexperimentsbut,inquantitativesubjects,ofteninvolvemathematicsandsimulation.
Usinganartificialandsimplifiedversionoftheworldallowsresearchersandpolicymakerstoexplorewhatmighthappenincertainscenarios.Inmacroeconomics,datamaybescarceandexperimentscanrarely,ifever,beperformedintherealworld,andthismakesmodelsespeciallyuseful.Differentmodelsare
goodforansweringdifferentquestionsandsoawiderangeofthemarerequired.
Allmodelsshouldbeabletoreproduce,asmuchaspossible,therealworldobservablestheyseektoexplain.Italsohelpsiftheyareeasytouseandinterpret,andiftheycanexplainthephenomenaassimplyaspossible.Overtime,modelswhichexplainrealitymoreadequatelyandmoreconciselyarefavoured,replacingthosewhichexplainitlesswell.
Agent-basedmodelsaresuitedtostudyingproblemsinwhichthecombinationoftheinteractionsofmanyagentsdrivestheoverallbehaviourofthesystem.Theysolveproblemsfromthe‘bottomup’ratherthanthroughrulesimposedfromthe‘topdown’.Typically,creatinganagent-basedmodelrequiresknowledgeofmathematics,statistics,andcomputerscience,aswellasthedisciplineinwhichitisbeingapplied.
Thesemodelsgettheirnamebecausetheyinvolvesimulatingalargenumberof‘agents’.Eachagentisaself-containedunitwhichfollowsitsownbehaviouralrules.Mostoften,thisisachievedwithinacomputersimulationbutitneednotbe.
Agentscouldrepresenttheconsumersinaneconomy,fishwithinashoal,particlesinagas,orevengalaxiesintheUniverse.Thebehavioursorrulesthatagentsfollowdependonthequestionofinterest.Somemodelshavemanydifferenttypesofagent,forinstancefirms,workersandgovernments.Thesemaythemselvesdiffer;forinstance,eachworkermighthaveadifferentproductivity,eachfirmadifferentsize.Agent-basedmodelshaveilluminatedasurprisinglywiderangeofsubjectsincludingmilitaryplanningandbattlefieldanalysis,operationalresearch,computerscience,biology,ecology,epidemiology,economics,socialsciencesandthephysicalsciences.
Withthem,suchvariedproblemsastheseparationofBrazilnutsfromothermixednuts,thebehaviouroftradersinthestockmarket,theflockingofbirds,thefallofancientcivilisations,thespreadofdisease,andtheeventualfateoftheUniversehavebeenexamined.
Inafewcases,resultswhichcouldonlyhavebeenobtainedthroughagent-basedmodellinghavesignificantlychangedthecourseofhistory:calculationsofthewaythatparticlesaretransportedthroughapileoffissilematerialinanuclearreactororweaponwouldbeprohibitivelydifficulttodoinanyotherway.
Theiruseineconomicscomeswithsomeparticularnuanceswhichareexploredinthisarticle.Inthenaturalsciences,agentbehavioursaretypicallymuchmoreconstrainedthanin
(1) SeeHaldane(2016).
Topical articles Agent-based models: understanding the economy 175
economics.Becauseofthis,agent-basedmodelsineconomicstypicallyproduceinsightsratherthanquantitativeforecasts.Theyaretypicallyqualitativeratherthanquantitative,andtheyaregoodfordeterminingwhatscenariosmightoccurratherthanexactlywhatwilloccur.
Anagent-basedmodelineconomicswouldnotnormallybeappropriateforforecastingthepriceofaparticularasset,forexample.Butitcangiveanideaofwhatactionsbytradersmightmovethepricesofassets,orwhythesupplyofanassetismuchmorevolatilethanthedemandforit.Otherphenomenawhichtheyhaveexplainedindifferentcontextsincludecycles,bubbles,clusteredvolatility,firesalesofassets,andtheonsetof‘bear’and‘bull’markets.(1)Despitebeingmoresuitedtoinsightsthanquantitativeprediction,therehavebeensomesuccessfulexamplesofforecastingwithagent-basedmodelsineconomics;forexample,forthedemandforelectricityortherepaymentrateofmortgages.(2)
TheGreatRecessionprofoundlychallengedtheeconomicsprofession,particularlyeconomicmodelling.Itdemonstratedthattheeconomyiscomplex,andnotalwaysatastableequilibrium.Theafter-effectsofthecrisisarestillbeingfelt.Establishedtools,suchas‘dynamicstochasticgeneralequilibrium’models(3)havebeencriticised(4)fornothavingenoughtosayaboutthedynamicsofcrises.Agent-basedmodelsareoneresponsetothechallengeandthisarticleexplorestheirpotentialinaidingourunderstandingoftheeconomy.
Theremainderofthissectionillustrateshowthesemodelshavedevelopedandwhatusestheyhavepreviouslybeenputto.Then,thearticleturnstoaspecificexampleofagent-basedmodellingineconomicswhichdemonstratessomeoftheirgeneralfeatures.
Theoriginsofagent-basedmodellingTheinitialspurfordevelopingagent-basedmodelscameinthe1930swhenphysicistEnricoFermiwastryingtosolveproblemsinvolvingthetransportofneutrons,asub-atomicparticle,throughmatter.Neutrontransportwasdifficulttomodelaseachstepinaneutron’sjourneyisprobabilistic:thereisachancetheparticlewilldirectlyinteractwith,scatteroff,orpass-byotherparticles.Previousmethodshadtriedtocapturetheaggregatebehaviourofalltheneutronsatonce,buttheimmensenumberofdifferentpossibilitiesforneutronpathsthroughmattermadetheproblemverychallenging.
Fermidevelopedanewmethodtosolvetheseproblemsinwhichhetreatedtheneutronsindividually,usingamechanicaladdingmachinetoperformthecomputationsforeachindividualneutroninturn.Thetechniqueinvolvedgeneratingrandomnumbersandcomparingthemtotheprobabilitiesderivedfromtheory.Iftheprobabilityofaneutroncollidingwere0.8,andhegeneratedarandomnumbersmallerthan0.8,
healloweda‘simulated’neutrontocollide.Similartechniqueswereusedtofindtheoutgoingdirectionoftheneutronafterthecollision.Bydoingthisrepeatedly,andforalargenumberofsimulatedneutrons,Fermicouldbuildupapictureoftherealwaythatneutronswouldpassthroughmatter.Fermitookgreatdelightinastonishinghiscolleagueswiththeaccuracyofhispredictionswithout,initially,revealinghistrickoftreatingtheneutronslikeagents.(5)
Theagent-basedtechniquesFermiandcolleaguesdevelopedwentontoplayanimportantroleinthedevelopmentofnuclearweaponsandnuclearpower.AtaroundthesametimethatFermiwasdevelopinghistechnique,thefirstelectroniccomputerswerebecomingavailableattheworld’sleadingscientificinstitutions.Computingpowerremainskeytoagent-basedmodellingtoday,withsomeoftheworld’ssupercomputersbeingharnessedforevermoredetailedsimulations.(6)
By1947scientistshaddevelopedanameforthistechniquewhichreflecteditsprobabilisticnature:theMonteCarlomethod.ThestorygoesthatthenamewasinspiredbyStanislawUlam’suncle,whowouldoftenasktoborrowmoneybysayinghe‘justhadtogotoMonteCarlo’.In1949,MetropolisandUlampublishedapapertogetherentitledThe Monte Carlo Method(7)whichexplainedthemanyusesofthenewtechniqueofusingrandomnumberstotackleproblems.Notallofthesewereagent-basedmodelsbutallreliedonusingartificiallygeneratedrandomnumberstosolveproblems.ThismoregeneralMonteCarlotechniquehasbeenappliedverywidely,forinstancetocalculatingsolutionstoequationswithmanyparameters,tothemanagementofriskandcatastrophes,andtoinvestmentsinfinance.
ThemoregeneralMonteCarlomethodhasthestrengththatitcanveryefficientlyexplorealargenumberofpossibilities.Forinstance,theusualwayforFermi’sneutronproblemtohavebeentreatedwouldhavebeentocreateagridofeverysinglepossibilityandthenfillinwhathappensforeachofthem.Thismeansthatevenimplausiblyunlikelyscenariosarecomputed.MonteCarloinsteadfocusesonthemostlikelyoutcomes.Thispropertycanmakethedifferencetowhetheraparticularproblemissolvableornot.TheMonteCarlomethodcanalsodealwithdistributions,forinstanceacrossincome,whicharenotdescribedbyanormaldistribution.(8)
(1) SeeZeeman(1974).(2) SeeGeanakoploset al(2012).(3) Aclassofmodelsinwhichmarketsareassumedtosimultaneouslyclear;see
Burgesset al(2013).(4) SeeMankiw(2006);Haldane(2016);Ascari,FagioloandRoventini(2015).(5) SeeMetropolis(1987).(6) LawrenceLivermoreNationalLaboratory(2013),‘Recordsimulationsconductedon
LawrenceLivermoresupercomputer’,availableathttps://www.llnl.gov/news/record-simulations-conducted-lawrence-livermore-supercomputer.
(7) SeeMetropolisandUlam(1949).(8)Normaldistributionsarewhatphysicalproperties,suchasheight,tendtofollowand
theyhaveawell-knownmathematicaldescription.TheyarealsoknownasGaussiandistributionsor‘bell-curves’.Manypropertiessuchaswealth,income,orfirmsizedonotfollowanormaldistribution.
176 Quarterly Bulletin 2016 Q4
Thesearestrengthswhichagent-basedmodelsinheritfromthemoregeneraltechnique.However,thisarticlefocusessolelyonMonteCarlosimulation,alsoknownasagent-basedmodelling.Thisseekstodescribeasystemofinteractingagentsandtheevolutionofthatsystem,ratherthancalculatethesolutiontoasingleequation.Inthisarticle,MonteCarlosimulationandagent-basedmodellingareusedsynonymously.
Agent-basedmodellinghasbeenusedacrossawiderangeofdisciplines,asdiscussedintheboxonpage177.Beforeembarkingonthestrengthsandweaknessesofagent-basedmodels,asimpleexampleispresentedwhichillustratessomeoftheirgeneralfeatures.ThisexampleisanearlyandinfluentialmodelbyeconomicsNobelMemorialprizelaureateThomasSchelling.
An early agent-based model in economicsInthelate1960sandearly70s,ThomasSchelling(1)developedamodelthatseekstounderstandtheeffectsofagents’preferencesaboutwheretheylive.ThedescriptionbelowdoesnotexactlyfollowSchelling’soriginalexample,butretainsitsmostsalientfeatures.
ImaginetherearetwospeciesnamedEconsandHumanswhoco-exist.Econsarealwaysrational.Humansareemotionalandsometimesmakemistakes.AlthoughEconsandHumanspeacefullyco-existandliveinthesamecity,theyeachhaveaslightpreferenceforlivingclosertothesamespecies.
Thispropensitytowanttobenearothersoftheirtypecanbecharacterisedbyanumberf,whichcanbethoughtofastheirstrengthofpreference.Itrepresentsthefractionofneighboursthattheyideallywishtobeofthesamespecies,withanfof1meaningthattheywillonlybehappyifalloftheirneighboursareofthesamespecies.
Ifagentsofeithertypeareunhappy,theycanchoosetomovehouseand,atrandom,aregivenanewproperty.Overtime,moreandmoreagentswillbehappywithwheretheyliveandstopmoving.Usingsimulations,Figure 1showsaninitialneighbourhoodofHumansandEconswhichismixed.
WhathappensiftheHumansandEconsarenowallowedtomovearounduntilalmostallofthemare‘happy’accordingtothevalueoff?Figure 2showsoneexamplesimulationwithf=25%;theagentsremaingenerallyinter-mixed.WhatissurprisingishowquicklythemixofHumansandEconsbecomessegregatedasfincreases.Figure 3showsanexampleofafinaldistributionwithf=26%.
Theagentsarenowclearlysegregated,eventhoughthechangeintheirpreferenceswasverysmallfromFigure 2.Thisisanexampleofa‘tippingpoint’,alsoknowninthephysicalsciencesasa‘phasetransition’.Itisasudden,emergentchangeintheoverallsystem.Tippingpointslikethiscanoccur
insystemswhicharecoupledtogetherbytheiragents;here,asmallchangeinfcanmeanthatalmosteveryonehastomoveinordertosatisfytheirnewpreferences.Asaresult,theneighbourhoodcanlookverydifferentafterthechangeinf.
Humans Econs
Figure 1 The initial distribution of Humans and Econs
Humans Econs
Figure 2 The final distribution of Humans and Econs with f=25%
Humans Econs
Figure 3 The final distribution of Humans and Econs with f=26%
(1) SeeSchelling(1969,1971);foranexampleseewww.jeromecukier.net/projects/models/segregate.html.
Topical articles Agent-based models: understanding the economy 177
Agent-based modelling across disciplines
Agent-basedmodellingsoonbecameaverypopulartechniqueinthephysicalsciences;(1)anearlypaperhasreceivedover30,000citationsbyotherresearchers.(2)Deepinsightshaveemerged,forinstancethatthestructureoftheUniversemaybeflatratherthancurved,(3)meaningthattheUniverseisunlikelytoendina‘BigCrunch’withallmatterconcentratedatasinglepointinspace.
Therangeofapplicationsinthephysicalsciencestodayisbroadindeedandincludesplasmaphysics,(4)particle-basedcancertherapies,materialsscience,crystallisation,magnetismandnuclearfusion.(5)Oneunexpectedapplication,publishedwiththetitleWhy the Brazil Nuts Are on Top,(6)simulatedhowmixednutssegregateovertimeinabag,researchwhichhashadimportantimplicationsforindustriesdealingwithparticulatemattersuchaspharmaceuticalsandmanufacturing.
Computersciencehasalsoseenheavyuseofthetechnique.ThepolymathJohnvonNeumannisbestknownineconomicsforhisworkongametheorybuthewasalsoaformidablecomputerscientist.Ina1948lecturehedescribeshisideafor‘cellularautomata’.Theseareartificialmodelsof‘cells’,whichlooklikefilledsquaresonaboard(similartotheSchellingmodeldiscussedinthisarticle).(7)VonNeumannfoundthat,contrarytointuition,verysimplerulesforthecellagentsgaverisetopuzzlinglycomplexbehaviourwhichlookslife-like.Thisisaso-calledemergentbehaviourbecauseitisalmostimpossibletopredictbasedontheruleswhichtheindividualagentsfollow.
Themilitarywereearlyadoptersofagent-basedmodelling.Wargameshadbeenplayedwithboardanddiceformanyyears,buttherewasaswitchtocomputationalagentsinthe1960s.Aswellasbeingusedtounderstandthedynamicsofpastbattles,suchashowoptimalthesearchingoperationsoftheBritishArmywereindetectingGermanU-boatsintheBayofBiscay,(8)theyaretodayusedbytheworld’slargestarmiestoformulatemilitarystrategy.Inrecenttimes,theautomatedbehaviourofagentsinwarmodelshasspilledoverintorealitywiththedevelopmentofautonomousweaponsandvehicles.
Thistrendisalsoreflectedincivilapplicationsofautonomousrobots.Mostrecently,therehasbeenmuchworkonmultipleautonomousroboticssystemsinwhichtheagentsarerobotsthatneedtodecidehowtomoveaboutintherealworld,forinstanceindriverlesscars.Itisextremelyusefultobeablemodelthebehaviourofautonomousrobotsbeforetryingthemoutinthefield.Forsometasks,centralisedcontrolofallrobotsisnotfeasiblesodesigningindividualbehaviourswhichproducetheoveralldesiredoutcomeisimportant.
Biologistsandecologistbegansimulatingthebehaviouroforganismswithinanenvironmentusingagent-basedmodelsinthe1980s.Thesemodelswereextendedtoincludemorecomplexphenomena,suchasagent-environmentinteractions.Anexampleistheresearchonmarineorganismswhichincludesbehaviourssuchasswimming,feeding,beingpreyedupon,andorganisms’interactionswithoceancurrents.Oneofthemanyapplicationsofthissortofmodelcouldbeexamininghoworganismsfairwithhigheroceantemperatures,orhigherconcentrationsofCO2.
(9)
Successesofthesemodelsincludepredictingthespawninglocationsoffish,explaininghowthetroutinLakeMichigan,USA,becamecontaminated,anddescribinghowsociallylearnedbehavioursleadtodistinctculturalgroupsamongmammals.Inthesecasestheadvantageofanagent-basedmodelwastobeabletoincludemanydifferentagentsandagentbehavioursinasinglemodelwhichcouldberunmillionsoftimestodeterminethemostprobableoutcomes.
Recentapplicationshavefocusedonconservationandmigration,includingthemanagementofforests,thetimingofanimalmigrations,andthepopulationpressuresonendangeredspecies.Forinstance,onemodelisoftheendangeredtigerpopulationinNepal;itusestheobservedbehaviourofindividualtigerstoexplorefutureconservationscenarios.(10)
Animportanthealth-relatedapplicationofagent-basedmodelsisinepidemiology—thestudyofthespreadofdiseases.Agoodunderstandingofthecomplexdynamicsofepidemicscouldsavemillionsoflives.Agent-basedmodelshaveallowedcountry-specificinformationsuchasgeographicaldata,commutingpatterns,agedistributionsandothercensusinformationtobetakenintoaccount.(11)
Inbiology,agent-basedmodelsarenowbeingusedtosimulateentireanimalcells,cancers,bacteriaandeventheeffectsofnewdrugsonpatients.Otherapplicationsincludeairtrafficcontrol(inwhichagentsrepresentaircraftandco-ordinatetominimisefueluse),transportationsystems(matchingagentstodestinations),crowdcontrolinemergencysituations,shoppingpatterns,predictinglandusepatterns(12)andnon-playeragentsincomputergames.
(1) SeeHaldane(2016)formore.(2) SeeMetropoliset al(1953).(3) SeeDaviset al (1985).(4) SeeTurrell,SherlockandRose(2015).(5) SeeTurrell(2013).(6) SeeRosatoet al(1987).(7) SeevonNeumann(1951);anexamplemaybefoundat
www.jeromecukier.net/projects/models/ca.html.(8) SeeCarl(2015);ChampagneandHill(2009).(9) SeeWerneret al(2001).(10)SeeCarteret al(2015).(11) SeeDegliAttiet al (2008).(12)SeeHeppenstallet al(2011).
178 Quarterly Bulletin 2016 Q4
Figure 4showswhatcanhappenwhenallagentshaveverystrongpreferences.
Schelling’smodelisverysimplebutitshowshowevenmildchangesinpreferencescanleadtosignificantchangesatthemacro-level.
Figure 5 showsthatstrikinglysimilarpatternsmaybeseenintheUSCensusdataforChicago.However,therearemanydifferencesbetweenthesimplemodelpresentedandtherealworld;preferencesmaynotbesymmetricacrossgroupsasisassumedinthebasicSchellingmodel,therearemanypracticalbarrierstomoving(suchasbudgetconstraints)andwhatwaslabelledas‘preference’islikelytobemuchricherinreality,reflectingthecomplicatedsocio-economichistoricalrelationshipsbetweengroups.
What is agent-based modelling good (and bad) for?
Theprevioussectiongaveaspecificexampleofthekindsofinsightswhichcomeoutofagent-basedmodels.Acrossmanyoftheirapplications,thereareasetofrecurringstrengthsandweaknesseswhichareexploredinthissection.Particularreferenceismadetotheapplicationsofthesemodelstoproblemsineconomicsandthesocialsciences.
StrengthsEmergentbehaviourThesinglemostpowerfulfeatureofagent-basedmodellingisthattheindividualactionsoftheagentscombinetoproducemacroscopic(1)behaviour.
Agoodexampleofthisistheherdingofsheeportheflockingofbirds.(2)Individualbehaviourscombinetoproduceaneffectwhichlooksorganisedevenwhentherulesforeachagentareincrediblysimple.(3)Trafficjamsareanotherfamiliarandunwelcomeexample;modelsandexperimentshaveshownthatjamscanresultevenwhenthereisnoimpedimenttotraffic.(4)Peoplecanherdintheireconomicactionsandexpectationstoo,forinstanceintheirexpectationsaboutinflation.Thishasdirectconsequencesfortheeconomy.
ThemostimportantexampleofemergentbehaviourineconomicsisAdamSmith’smetaphoroftheinvisiblehand:howtheself-interestedactionsofrealagentsintheeconomycombinetoproducesociallyoptimaloutcomes.Oneofthestrengthsofagent-basedmodellingisthatthisinvisiblehandismadevisibleanditsworkingsmaybeexamined.Thisisincontrasttosomeothermodelapproachesinwhichtheactionsofmanyindividualsareassumedtoleadtoaparticularoutcome,oftenusingasinglerepresentativeagent.Thissimplificationisvalidinsomecasesbutnotallcombinationsofbehaviourscanberepresentedbytheactionsofasingleagent.(5)
HeterogeneityAsindividualagentsaremodelled,itispossibletoexploretheconsequencesofagentsbeingheterogeneous;thatisagentsbeingdifferentinsomeway,perhapsbyincome,preferences,educationorproductivity.
Incorporatingheterogeneityallowsforthemodellingofmuchricherbehaviour.Inequalityisagoodexample—aggregatewealthcanincreasebutifitisonlyasmallfractionofthepopulationdrivingthisphenomenonitwouldsuggestvery
Humans Econs
Figure 4 The final distribution of Humans and Econs with f=70%
Figure 5 US Census data for Chicago showing segregation(a)
Source:http://demographics.coopercenter.org/DotMap/.
(a) NateSilver,538.com;seehttp://fivethirtyeight.com/features/the-most-diverse-cities-are-often-the-most-segregated/.
(1) Atthelargescale,inthiscasethesizeofthesystemwhichtheagentsinhabit.(2) SeeMacyandWiller(2002).(3) Anexamplemaybefoundatwww.tjansson.dk/2012/11/yabi-yet-another-boids-
implementation-simulation-of-flocking-animals/.(4)Anexampleofanagent-basedmodeloftrafficjamscanberuninyourinternet
browseratwww.traffic-simulation.de/.(5) Thisisanexampleofthe‘fallacyofcomposition’.
Topical articles Agent-based models: understanding the economy 179
differentunderlyingeconomicreasonsthaniftheentirepopulationwerebecomingwealthier.
StylisedfactsPerhapsthegreatestsuccessofagent-basedmodelsineconomicshasbeenexplainingthestylisedfactsobservedinassetmarkets.(1)(2)Thereareanumberofphenomenaobservedempiricallyinthemarketsforassetssuchasbondsorequitieswhicharenotexplainedbytraditionaleconomictheory.Someofthetwomostwidelyseenacrossmarketsare:
• Clusteredvolatility,inwhichthestandarddeviationofreturnsonanassetexhibitstrendswhichare‘clustered’intime.(3)
• ‘Fattails’,inwhichextremeeventssuchaslargepricechangesoccurmorefrequentlythanwouldbeexpectedifanormal(orGaussian)distributionwasassumedandfarbeyondwhatwouldbeexpectediftraderswerebehavingrationally.AnexampleisshowninChart 1.
TheSanta Fe Artificial Stock Market(4)isagoodexampleofhowtradingactivitiescanaffectaggregatemarketstatistics.Thismodelissimilarinmanyrespectstoatraditionaleconomicmodelbutwiththeimportantdistinctionthatagentshaveheterogeneousexpectationsofreturns.Thismakesitimpossibleforoneagenttoknowwhatotheragents’expectationsare,andthusimpossibletoformanunambiguousandrationalexpectationofprice.Instead,agentsadaptivelylearnusingarangeofexpectationmodels.Anevolutionaryprocessoccursonagents’strategies;theyareconstantlyupdatedinlightoftheactualpathofthemarket.Interestingly,themodelproducestworegimes—onewhichlookslikestherationalexpectationsworldwithamarketpriceatthefundamentalpriceofthefinancialproduct,andoneinwhichclusteredvolatilityand‘fattails’occur.Rapidevolution
ofstrategiescausesthenon-rationalmarketswhicharemorelikethoseobservedinreality.
Manyagent-basedmodelshaveshownthatchartistortrend-basedtrading(inwhichtradersfollowthetrenddirectionofprices),andalsoleverage,(5)cancontributetopriceovershoots,andthatthiscandriveclusteredvolatility,hightradingvolumesandfattails.(6)
RealisticbehavioursThegenerationofrealisticbehaviour,basedonobservedbehaviour,canbeastrengthofagent-basedmodels.Researchinbehaviouraleconomicshasshownthatpeopleoftenuseheuristics(7)whenmakingdecisionsandthattheyarenotfullyrational.Thesebehaviourscanbecollectivelydescribedas‘boundedrationality’.Anexampleisthatpeoplereactmorenegativelytolossthantheydopositivelytogain,aphenomenonknownaslossaversion.Thereareseveralmodelswhichexplorewhathappenswhenpurelyrationaloptionsarenotavailableoraretoocostly,orwhenagents’environmentschangeovertime.(8)
ExploringthepossibilitiesOneoftheadvantagesofagent-basedmodellingisthatitcanveryefficientlyexplorealargenumberofpossibilities.Probabilisticrulesappliedtoeachindividualagentinturncanbeasimplerwayofexploringscenariosthanworkingouthowtheentirepopulationofagentsshouldbehavetogether.Anexampleisinthetransportofparticles,andwasthepromptforFermitooriginallydeveloptheMonteCarlomethod.Anotherisinepidemiology,whichcansimilarlybemodelledeitherwithsetsofequationswhichtrytosummariseallbehaviouratonce,oranagent-basedmodel.Theequationapproachquicklybecomescomplicatedasmoreandmoreagentpropertiesthatarerelevanttodisease(suchashealth,age,andevencommutingpattern)areintroduced.Oragent-basedmodelscanbeappliedtoconflictinwhicharelativelyinflexiblesetofequationsmodellingtherateofchangeofsizeofarmiescanbereplacedwithagent-basedmodelswhichcancapturethefullheterogeneityofcombatantsinmodernwarfare.
(1) SeeHongandStein(1999).(2) SeeCutler,PoterbaandSummers(1989);LuxandMarchesi(1999,2000).(3) SeeCont(2007).(4) SeeLeBaron,ArthurandPalmer(1999).(5) SeeThurner,FarmerandGeanakoplos(2012).(6) SeeTesfatsion(2002).(7) Heuristicsare‘rulesofthumb’formakingdecisionswhichsimplifytheprocessbut
maynotalwaysgivetheoptimaldecisioninallcircumstances.(8) SeeGodeandSunder(1993);Farmer,PatelliandZovko(2005);Challetand
Zhang(1997);Hommes(2006);Axelrod(2006).
Equity prices
Best Gaussian fit
40 20 0 20 40– +
10-0
10-1
10-2
10-3
10-4
Probability density
Month-on-month change, per cent
Chart 1 Not normal: changes in the price of equities have a fat-tailed distribution (1709–2016)
Sources:Hills,ThomasandDimsdale(2016)andBankcalculations.
180 Quarterly Bulletin 2016 Q4
Complexity,non-linearityandmultipleequilibriaAnotherstrengthofagent-basedmodelsisthattheycandescribecomplexsystems.Complexsystemsarecharacterisedbyhavingmanyinterconnectedparts,havingvariableswhichcanchangedramaticallyandwhichcandemonstrateself-organisation.Additionally,complexsystemsundergosudden,dramatictransitions,sometimescalledphasetransitions.Recentworkonagent-basedmodelsofthemacroeconomyhasdescribedphasetransitionsbetweenlowandhighunemployment.(1)Theeconomydisplaysmanyofthecharacteristicsassociatedwithcomplexsystems.
WeaknessesToomuchfreedom?Whileitistruethatagent-basedmodellinghasmanystrengths,italsopresentschallenges.
Inmanyways,thegreateststrength—theflexibilitytomodelsuchavastrangeofscenarios—isalsothegreatestweakness.Thesheerextentofchoiceinconstructingagent-basedmodelsascomparedtomoretraditionaleconomicmodelsmeansthatmodellersfacetheproblemofselectingtherightcomponentsfortheproblemathand.Simulationresultscanvarydramaticallydependingonwhichassumptionsareused,somodellersmusttakegreatcareinchoosingthem.Furtherworkisneededtodevelopobjectivemeansforchoosingthemostappropriateassumptions.
TheLucascritiqueThehugerangeofbehavioursavailableforagentsmeansthatagent-basedmodelscanbevulnerabletotheLucascritique.Thiscritiquecentresaroundthefactthatagents’choicesmaynotfollowhistoricallyobservedrelationshipswhenpolicyinterventionsaremadewhicharepremisedonthoserelationships.Inprinciple,agentbehaviourscanbedesignedtorespondtochangingcircumstancesbutthereisatrade-offbetweencreatingagentsthatwillalwaysfollowtheiroptimalcourseofactionandbuildingsimple,understandablemodels.Thisiswhyfullyrationalbehaviourisusefulasanapproximation:itgivesanunambiguousandrelativelysimplesetofrulesforhowagentscanalwaysactintheirownbestinterests.
Eachagent-basedmodelshouldbeasLucas-critiqueproofaspossiblebutoftenthemostinterestingbehaviours—suchasboundedrationality—aretheoneswhicharethemostdifficulttomakerobusttothecritique.
DifficulttogeneraliseTheproliferationofchoicesinconstructingagent-basedmodelsleadstoanotherweakness,whichisthattheytendtobebespoke.Forinstance,theagent-basedmodelwhichtellsushowbondsaretradedisunlikelytobeveryhelpfulforansweringquestionsaboutthehousingmarket.
CalibrationandinterpretationCalibrationinvolvesadjustingthemodeltofitwiththeknownfacts,forinstanceinitialisingitwithempiricaldata.Validationischeckingthattheoutputofthemodelisreasonablegivenwhatisknown,andperhapscross-checkingitwithothermodelsorvariationsintheassumptions.Ifamodelcontainsmanydifferentoptionsinhowitisconstructed,itcanspuriouslyreproducedatathatlooksimilartoempiricaldata.Thisisknownasoverfitting,anditisaproblemforallmodels.Calibrationisevenmoredifficultinagent-basedmodels,however,becausetheytypicallyproducestylisedfactsratherthanquantitativeforecastsandtherearevariouswaystheagreementwithempiricallyobservedstylisedfactscouldbeassessed.
Theresultsofagent-basedmodelscanalsobedifficulttocommunicatebecausetheymustbepresentedalongsidetheassumptionsusedtocreatethem.Althoughtrueofallmodelstosomeextent,thisproblemislessacutewithmodelsbasedonhistoricaldataaloneastheyusecommonstatisticaltechniques.Norisitthecasewithmodelsbasedonrationalexpectations;ifagentscanonlyactperfectlyrationallythenthereisastrongconstraintwhichiseasilyunderstoodacrossallsimilarmodels.Whenusingagent-basedmodellingtoinformthechoicesofpolicymakers,thisweaknesscanbeabarrier.
Finally,itcanbedifficulttounderstandhowchangingmodelinputsaffectthemodeloutput.Thisisanunavoidablefeatureofcomplexsystems,andoftherealworlditself.Itisincontrasttomoreanalyticallytractablemodelsinwhichtheeffectsofchangingaparametermayhaveamuchclearereconomicinterpretation.Althougheveryagent-basedmodeldoeshaveauniquemathematicalrepresentation,theequationswouldbedifficulttotranscribefromacomputerprogramme.
Despitethesechallenges,agent-basedmodelsprovideanimportanttoolforunderstandingtheworld,andonewhichhasdeliveredinanastonishingrangeofscenarios.
Wenowturntohowthesemodelscanimproveourunderstandingoftheeconomyasawhole,atopicwhichisespeciallyrelevanttocentralbanks.
Macroeconomic agent-based modelling
Themacroeconomyischaracterisedbyhavingbusinesscycles:fluctuationsinthegrowthofGDParounditslong-termtrend.Instandardmacroeconomicmodelling,thesefluctuationsaretheresultofunexplained(‘exogenous’)shocks.(2)Forinstance,inflationwouldbeforcedtosuddenlychangebutnotasa
(1) SeeGualdiet al(2015).(2) Ashockcouldbedescribedasanunexpectedanddiscontinuouschangeinavariable.
Topical articles Agent-based models: understanding the economy 181
consequenceofanyemergentphenomenawithinthemodel.Thesearecalledexogenousshocks.
Intherealworld,thefluctuationsinGDParelikelytobeendogenous—thatis,generatedbytheeconomyitself.Agent-basedmodelsprovideawayofmakingthesefluctuationsendogenous,andthereforealsoprovidearoutetounderstandingtheircausesandwhatpoliciesmightaffectthem.TheGreatRecessionisacompellingexampleofareasonwhypolicymakersneedtounderstandwhatdrivesthesefluctuations.
Severalagent-basedmodelshavebeenputtingtogetherthedifferentelementsandagentswhicharerequiredtorealisticallyreproducethestylisedfactsoftheeconomyofanentirecountry,orevenseveralcountriesinterlinkedbytrade.Theseelementsincludetheentryandexitoffirms,(1)endogenousinnovation,monetarypolicyandfiscalpolicy.
Oneofthesemacroeconomicmodelsespeciallyaddressesmonetarypolicy.(2)Init,therearefirms,consumers,andpriceswhicharechangedaccordingtosimpleexpectations.Businesscyclessimilartothoseobservedinrealityaregenerated.Asimulatedexperimentshowsthatanactivemonetarypolicy,formulatedaccordingtoasimplerule,(3)reducesthesizeofthefluctuationsinGDPrelativetohavingastaticpolicyrule.Thisoccursbecausefirms’demandforcreditfallswhenthecentralbankraisesinterestratesaccordingtotherule.
Anothermodel(4)featuresaneconomycomposedofheterogeneouscapitalandconsumption-goodfirms,alabourforce,banks,agovernmentandacentralbank.Capital-goodfirmsperformresearchandproduceheterogeneousmachinetools.Consumption-goodfirmsinvestinnewmachinesandproduceahomogeneousconsumptiongood.Consumption-goodfirmsfinancetheirproductionandinvestmentsprimarilywiththeirliquidassetsand,ifrequired,bankcredit.Capital-goodfirmsproduceusingcashadvancedbytheirconsumers,ratherthanusingbanks.
Byincorporatingthefinancialsector,thismodelisabletoreproducemanyfeaturesseeninempiricalmacroeconomics,includingthecyclesinGDP,investmentandconsumption,aswellasthevolatilityofthesethreevariablesrelativetoeachother.Bankingcrisesarealsoanemergentphenomenaofthemodel;ashighproductionandinvestmentlevelsraisefirms’debt,thefirms’networthdecreases,increasingtheircreditrisk.Banksthenrationcreditandforcefirmstocurbproductionandinvestment,withthepotentialtotriggerarecession.Bankfailuresemergefromtheaccumulationofloanlossesonbanks’balancesheets.Themodelallowsforabetterunderstandingofthechainofeventswhichleadtobankingcrises.Confidencethattheseeventsdoreflecttherealworldsituationcanbegainedfromthereproductionof
stylisedfactssuchasthedistributionofbankingcrisisdurationsbeingveryclosetotheempiricalone.
ExperimentsonmonetarypolicywiththismodelsuggestthatadualmandatetotargetbothinflationandunemploymentresultinahigheraveragegrowthinGDPwithlowervolatilitythantargetinginflationaloneachieves.
Othermacroeconomicagent-basedmodels(5)havelookedathowinterestratesareset,(6)(7)athowliquiditytrapscanbeendogenously(8)generated(9)andattheeffectsofunconventionalmonetarypolicy.(10)
Thecostincomplexityofthesemacromodelsisbalancedbytheinsightswhicharegenerated.Theycanreproduceanimpressivelistofmacroeconomicstylisedfacts:businesscycles;theprocyclicalityofproductivity,nominalwages,firms’debt,bankprofitsandinflation;thecountercyclicalityofunemployment,(11)prices,mark-ups,andloanlosses;andtheappearanceoffattailsinthedistributionofoutputgrowth.Alternativeandcomplementarymodels,suchasdynamicstochasticgeneralequilibriummodels,havenotbeenabletogenerateallofthesephenomenaendogenously.(12)Thesemodelstherebyaidtheunderstandingofhowcomplexmacro-levelphenomenaemergefromunderlyingmicro-levelphenomena.
Thesestrongempiricalcredentialslendconfidencetotheconclusionsofpolicyexperimentsundertakenwithagent-basedmodels.Theflexibilityofthemodelsmeansthatextremelyfine-tunedregulationcanbe‘tested’usingthem.Examplesmightincludetheeffectofregulationonliquidityandprofitability,theinteractionofmicroandmacroprudentialpolicies,orhowcreditnetworkscangiverisetobusinesscyclesandfinancialcrises.(13)Asanexample,theNASDAQstockexchangehasusedanagent-basedmodeltodesignregulationwhicheliminatesloopholesthatcouldbeabusedbyitsusers.(14)
Thenextsectiondiscussestwoagent-basedmodelswhichhavebeendevelopedintheBankandwhichattempttoreplicatethestylisedfactsobservedintwodifferentmarkets:corporatebondsandhousing.
(1) SeeEliasson(1991).(2) SeeGattiandDesiderio(2015).(3) TheTaylorrule;seeTaylor(1993).(4) SeeDosiet al (2015);thismodelincorporatesaspectsofKeynesian,Schumpeterian
andMinskianeconomics.(5) SeeDilaver,JumpandLevine(2016).(6) SeeDelliGattiet al(2005).(7) SeeRaberto,TeglioandCincotti(2008).(8) Havinganinternalcauseratherthanbeingtheresultofanexternalshock.(9) SeeOeffner(2008).(10)SeeCincotti,RabertoandTeglio(2010).(11) SeeGualdiet al(2015).(12)SeeAscari,FagioloandRoventini(2015).(13)SeeGatti,GaffeoandGallegati(2010).(14)SeeBonabeau(2002).
182 Quarterly Bulletin 2016 Q4
Agent-based modelling in the Bank of England
Trading in corporate bonds by open-ended mutual funds(1)
Anagent-basedmodeldesignedtocapturesomeofthedynamicsoftradingincorporatebondsbyopen-endedmutualfundswasdevelopedwithintheBank.Themodelaimedtobeasparsimoniousaspossiblewhilereproducingrealisticbehaviourforthemarket.
Theassetsunderthemanagementofcorporatebondfundshavemorethandoubledsincethe2008financialcrisis.Atthesametime,concernsaboutthefragilityoffixed-incomemarketshavegrown.Despitethemarketbeinglarger,thereareworriesthattherehasbeenareductioninmarketliquidity,sothatlargeordershavemoreofaneffectonprices.
Thedynamicsofthismarkethaveimportantimplicationsforfinancialstability.Overshootingduringadjustmentsinthepriceofcorporatebondsmayunnecessarilyreducetheabilityofsomecompaniestoservicerefinanceddebt,threateningtheirsolvency.Somefirmsmayalsobedeterredfromraisingnewfinancing.Inextremis,thiscouldcauseanimpairmentofthesupplyofcredittotherealeconomy.
AstylisedpictureofthemodelcanbeseeninFigure 6.Therearearepresentativepoolofinvestors,asinglemarketmakerthroughwhichalltradesaremade,andthreedistincttypesoffund.Themodelendogenouslyreproducesoneoftheimportantstylisedfactsobservedinthecorporatebondmarket:thedistributionofdailylog-pricereturns.ThisisshowninChart 2,wheretheempiricalobservationsarefromaUSinvestment-gradecorporatebondindex.Theverytailendsofthedistributiondonotmatchtheempiricaldata;asituationwhichcouldpotentiallybeimprovedbysacrificingsomeoftheparsimonyofthemodel.However,theoverallfitofthedistributionisagoodmatchtodata.
Themodellooksathowinvestorsredeemingthecorporatebondsheldforthembyopen-endedmutualfundscancausefeedbackloopsinwhichbondpricesfallfurther.
Forexample,ifinterestratesweretorise,existingcorporatebondsmightbecomelessattractiveandredemptionscouldtakeplaceassomeinvestorspulltheirwealthout.Investorswhoredeemtheirbondsfirstgetagoodpricewhenfundssellthem,butpoorliquiditymaycausebondpricestofall.Thesepricefallscouldpromptremaininginvestorstoredeemtheirbondstoo,sothatfundshavetoselloffmorebondsandpricesfallevenfurther.Afeedbackloopofredemptionstakesplaceinwhichwealthisdestroyedandthosewhomaketheinitialredemptionsenjoyafirst-moveradvantage.Thisissimilartothechainofeventswhichcausedbankrunsbeforetheadventofdepositinsurance.
Simulatedexperimentswereundertakenwiththismodel.Theaimwastounderstandhowchangesinthebehavioursoftradersinbondsmightaffecttheextentofdislocationinpriceandyieldfollowingashock.
Theshockusedwastofunds’expectationsaboutthefractionofcompaniesthatwillfailinagivenyear,alsoknownastheexpectedlossrate.Thiscreatesanegativefeedbackloopinthepriceforbonds.Iffundsexpectmorecompaniestofail,theyarelikelytodemandhigheryieldsfrombondstocompensatethemforthis—soasuddenchangeintheexpectedlossratepushesdownonprices.Theaftermathofthisisseeninovershootsinbothpriceandyieldbeforetheysettledowntheirvalues.ThestepsinthefeedbackloopareshowninFigure 7.
(1) SeeBraun-Munzinger,LiuandTurrell(2016).
Corporate bond market
$
Value traders
Momentumtraders Passive funds
Investor
Market makerNoise
Figure 6 Schematic of an agent-based model of the corporate bond market
Daily log-price return
8 6 4 2 0 2 4 6 8
10-0Probability densitySimulation
Pre-crisis
– +
10-1
10-2
10-3
10-4
x10-3
CrisisPost-crisis
Chart 2 Reproducing stylised facts: the distribution of daily log-price returns
Topical articles Agent-based models: understanding the economy 183
Giventheshock,anditsknowneffectsin‘normal’times,scenariosinwhichthetradingbehaviourwasdifferentcouldbeexplored.
Inoneofthesescenarios,theeffectofanincreaseinthefractionoffundswithpassivetradingstrategieswasexplored.Thisledtoasurprisingresult;thereisatailriskofsevereovershootswhenthepresenceofpassiveinvestmentfundsincreases.Anotherfindingisthatifallfundsweretoincreasethetimewindowoverwhichredemptionsweremadebyinvestors,theextentofpricedislocationwouldbesignificantlyreducedintimesofcrisis.ThisisbecausethefeedbackloopshowninFigure 7isstrongerwheninvestorswithdrawanamountofwealthonasingledayasopposedtowithdrawingitoveralongertimeperiod.
An agent-based model of the UK housing market(1)
Thehousingmarkethasoftenbeenasourceoffinancialstressandcrisis,lookingacrossawiderangeofcountriesandawidespanoftime.Themarketexhibitsclearandsignificantcyclicality.Capturingthesecyclicaldynamicsisnotstraightforwardandonepotentialreasonisthatthehousingmarketcomprisesmanytypesofparticipant—forinstancerenters,first-timebuyers,andbuy-to-letlandlords.Theseagentsareheterogeneousbyincome,gearingandlocation,sotheyhavedifferentincentives.Thecombinationoftheiractionsgivesrisetothecyclicaldynamics.
Inadditiontoabankingsector(mortgagelender)andacentralbank,themodelcompriseshouseholdsofthreetypes:
• renterswhodecidewhethertoattempttobuyahousewhentheirrentalcontractendsand,ifso,howmuchtobid;
• owner-occupierswhodecidewhethertoselltheirhouseandbuyanewoneand,ifso,howmuchtobid/askfortheproperty;and
• buy-to-letinvestorswhodecidewhethertoselltheirrentalpropertyand/orbuyanewoneand,ifso,howmuchtobid/askfortheproperty.Theyalsodecidewhethertorentoutapropertyand,ifso,howmuchrenttocharge.
Thebehaviouralrulesofthumbthathouseholdsfollowwhenmakingthesedecisionsarebasedonfactorssuchastheirexpectedrentalpayments,housepriceappreciationandmortgagecost.Thesehouseholdsdiffernotonlybytype,butalsobyage,income,bankbalances,rentalpaymentandotherproperties.
Animportantfeatureofthemodelisthatitincludesanexplicitbankingsector,whichprovidesmortgagecredittohouseholdsandwhichsetsthetermsandconditionsavailabletoborrowersinthemortgagemarket.Thebankingsector’slendingdecisionsarethemselvessubjecttoregulationbyacentralbank,whichsetsloantoincome(LTI),loantovalue(LTV),andinterestcoverratiopolicies(ICR).Thevariousdifferentagents,andtheirinterlinkages,areshowninFigure 8.
ThismodelwascalibratedusinghousingmarketdatasourcesandhouseholdsurveyssothattheagentsinthemodelhavecharacteristicswhichmatchthoseintheUKpopulationoveraparticularperiodoftime.Itincludesbehaviouralcharacteristicssuchashowoftenandbyhowmuchthepriceofahouseisreducedifitremainsunsold.
Funds’ demand forbonds is reduced
Price falls as market makersees reduced demand
Returns to investors fall as aconsequence of price drop
Investors reduce allocationof cash to funds
Shock toexpectedloss rate
Funds’ wealth is reduced
Figure 7 Capturing non-linear relationships: the feedback loop following a shock to funds’ expected loss rate
(1) SeeBaptistaet al(2016).
Central bank
sets caps on LTV, LTIand ICR ratios, andaffordability tests
Ownershipmarket
Rental market
Buy-to-letinvestors
Renters
Households
SocialhousingOwner-occupiers
Bank
givesmortgages
Figure 8 Agents and interactions in the housing market model
184 Quarterly Bulletin 2016 Q4
Oneofthekeyfeaturesofthemodelisthatitisabletoendogenouslygeneratereal-worldhousepricecyclesgeneratedbythemodelinChart 3.ItalsoreproduceskeyaspectsoftheUKhousingmarket,suchastheempiricaldistributionoftheshareofloansbyloantoincomebandbasedontheProductSalesDatabase(PSD)ofUKmortgages.ThisisshowninChart 4.
Themodelwasusedtolookatseveralscenariosforthehousingmarket.Inone,alargerbuy-to-letmarketwasfoundtocausemuchlargerswingsinhousepricesduringcycles.Inanother,theeffectofamacroprudentialpolicythatlimitslenderstomakingamaximumof15%ofitsmortgagesatloantoincomeratiosgreaterthan3.5wasexplored(thepre-existingBankpolicyisaratioof4.5).Withthislimitinplace,theendogenouscyclesinhousepricegeneratedbythemodelaredampened.
The future for agent-based modelling in economics
Agent-basedmodelsineconomicsthriveondata.Moredatameansfewerassumptionsneedtobemade,andmoreofthestructureofthemodelcanbebasedonwhatisobserved.Fortunatelyformodellers,theamountofmicroeconomicdatawhichisavailableisincreasingconsiderablyandopportunitiestorefinemodelsarelikelytobeinreadysupply.
Therelentlessimprovementsincomputerprocessingpoweralsoprovidenewopportunitiestomodeltheeconomyonevermoregranularscales,andwithmorecomplexbehaviours.
Inthefuture,theabilitytoincorporaterealisticbehaviourscouldbeexploitedfurtherasartificialintelligencematures.Thetechniqueknownasmachinelearningiswidelyusedbytechcompaniesto,forinstance,suggestthefilmsthataconsumermightliketowatch.(1)Inthefuture,thesetechniquescouldbeusedto‘train’anartificialagenttobehavelikearealconsumerorfirm.Advancedartificialintelligencecouldmakeagent-basedmodelsmoreLucas-critiqueproofbyhavingagentsrespondrealisticallytonewcircumstances.Aglimpseintowhatispossibleinthisrespectwasgivenrecentlywhenanartificialintelligencerepeatedlybeatatrainedfighterpilotinanair-to-aircombatsimulation(2)—anextremelydemandingscenarioforacomputer.Recentprogressbycomputersatvariousgames,includingGoandJeopardy!addweighttothisargument.
Therearelessexoticwaysinwhichprogressmightbeachieved,forinstancewiththedevelopmentofastandardsetoftoolsforcalibratingandvalidatingagent-basedmodels.Anotherwaytomeetsomeofthecriticismlevelledatthesemodelsisforresearcherstoinvestigatethemostparsimoniousmodelspossiblewhichstillreproduceobservedstylisedfacts.
Withrespecttocentralbanks,therearethreeparticularlypromisingareasofdevelopmentforagent-basedmodelling.Thefirstistheongoingapplicationofmacroeconomicagent-basedmodelstomonetarypolicy.Severalmodelswhichexplicitlyincludecentralbankshavenowbeenestablishedandareonhandtoexaminepolicyquestions.Thesecondisinmodellingthebankingandfinancialsector,todeterminehowfinancialstressistransmittedthroughthesystemasawhole.Third,researchingthepotentialimpactoftheintroductionofacentralbankdigitalcurrencycouldbeexploredusinganagent-basedmodel.Asrecentlydescribed,(3)somespecificationsforacentralbankissueddigitalcurrencycouldhaveconsequencesforfinancialstability.Anagent-basedmodelmaybesuitedforanalysingtheeffectsof
Simulation time (years)0 5 10 15 20 25 30 35 40
House price index
0.90
0.95
1.00
1.05
1.10
1.15
1.20
1.25
1.30
Chart 3 A benchmark run of the model showing boom and bust cycles in the house price index
0
5
10
15
20
25
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 More
Share of total loans, per cent
Loan to income band
Model simulation
PSD data
Chart 4 The model can reproduce the loan to income distribution for the United Kingdom
(1) SeeFriedman,HastieandTibshirani(2001).(2) SeeErnestet al(2016).(3) SeetheBankofEngland’sresearchquestionsoncentralbankdigitalcurrencies,
availableatwww.bankofengland.co.uk/research/Documents/onebank/cbdc.pdf.
Topical articles Agent-based models: understanding the economy 185
thedifferentpossiblespecificationsofacentralbankdigitalcurrency.
Conclusion
Sincetheirinception,agent-basedmodelshavebroughtawealthofinsightstothoseproblemstowhichtheyaresuited.Theyhavefounduseinanastonishinglydiverserangeofsubjects;fromtigerstotraders,fromparticlestopeople.
Thisincludescomplexsystemsgenerally,butparticularlythoseinwhichoutcomesemergefromindividuals’choicesoragentsareheterogeneous.Thishasobvious,usefulapplicationstomodellingtheeconomyandmanyexampleshavedemonstratedhowagent-basedmodelsineconomicscanaidtheunderstandingofempiricallyobservedphenomena.
A‘bottom-up’perspectiveontheeconomyisacompellingcomplement(1)tothe‘top-down’approacheswhicharealreadywidelyused.
Theagent-basedapproachcomeswithdifficultiesofitsown,especiallyaroundtheassumptionsandvalidity,butthefutureholdspromiseintheserespects,partlyduetoeverimprovingtechniquesbutalsobecauseoftheavailabilityofrichdatasetstocalibratethemodelsagainst.
Agent-basedmodelshaveanimportantplaceinunderstanding,andevendesigning,markets,andprovideauniqueplatformforaugmentingpolicymakers’judgementsabouttheeconomy.
(1) SeeFagioloandRoventini(2012).
186 Quarterly Bulletin 2016 Q4
References
Ascari, G, Fagiolo, G and Roventini, A (2015),‘Fat-taildistributionsandbusiness-cyclemodels’,Macroeconomic Dynamics, Vol.19(2),page465.
Axelrod, R M (2006),The evolution of co-operation,BasicBooks.
Baptista, R, Farmer, J D, Hinterschweiger, M, Low, K, Tang, D and Uluc, A (2016),‘Macroprudentialpolicyinanagent-basedmodeloftheUKhousingmarket’,Bank of England Staff Working Paper No. 619,availableatwww.bankofengland.co.uk/research/Documents/workingpapers/2016/swp619.pdf.
Braun-Munzinger, K, Liu, Z and Turrell, A (2016),‘Anagent-basedmodelofdynamicsincorporatebondtrading’,Bank of England Staff Working Paper No. 592,availableatwww.bankofengland.co.uk/research/Documents/workingpapers/2016/swp592.pdf.
Bonabeau, E (2002),‘Agent-basedmodeling:methodsandtechniquesforsimulatinghumansystems’,Proceedings of the National Academy of Sciences, Vol.99(3),pages7,280–287.
Burgess, S, Fernandez-Corugedo, E, Groth, C, Harrison, R, Monti, F, Theodoridis, K and Waldron, M (2013),‘TheBankofEngland’sforecastingplatform:COMPASS,MAPS,EASEandthesuiteofmodels’,Bank of England Working Paper No. 471,availableatwww.bankofengland.co.uk/research/Documents/workingpapers/2013/wp471.pdf.
Carl, C R G (2015),Search theory and U-boats in the Bay of Biscay,PicklePartnersPublishing.
Carter, N, Levin, S, Barlow, A and Grimm, V (2015),‘Modelingtigerpopulationandterritorydynamicsusinganagent-basedapproach’,Ecological Modelling,Vol.312,pages347–62.
Challet, D and Zhang, Y C (1997),‘Emergenceofcooperationandorganizationinanevolutionarygame’,Physica A: Statistical Mechanics and its Applications, Vol. 246(3),pages407–18.
Champagne, L E and Hill, R R (2009),‘Asimulationvalidationmethodbasedonbootstrappingappliedtoanagent-basedsimulationoftheBayofBiscayhistoricalscenario’,The Journal of Defense Modeling and Simulation: applications, methodology, technology, Vol.6(4),pages201–12.
Cincotti, S, Raberto, M and Teglio, A (2010),‘Creditmoneyandmacroeconomicinstabilityintheagent-basedmodelandsimulatorEurace’,Economics: the Open-Access, Open-Assessment E-Journal, Vol.4.
Cont, R (2007),‘Volatilityclusteringinfinancialmarkets:empiricalfactsandagent-basedmodels’,Long memory in economics,SpringerBerlinHeidelberg,pages289–309.
Cutler, D M, Poterba, J M and Summers, L H (1989),‘Whatmovesstockprices?’,Journal of Portfolio Management,Vol.15(3),pages4–12.
Davis, M, Efstathiou, G, Frenk, C S and White, S D (1985),‘Theevolutionoflarge-scalestructureinauniversedominatedbycolddarkmatter’,The Astrophysical Journal,Vol.292,pages371–94.
Degli Atti, M L C, Merler, S, Rizzo, C, Ajelli, M, Massari, M, Manfredi, P, Furlanello, C, Tomba, G S and Iannelli, M (2008),‘MitigationmeasuresforpandemicinfluenzainItaly:anindividualbasedmodelconsideringdifferentscenarios’,PLoSONE,3(3),e1790.
Delli Gatti, D, Di Guilmi, C, Gaffeo, E, Giulioni, G, Gallegati, M and Palestrini, A (2005),‘Anewapproachtobusinessfluctuations:heterogeneousinteractingagents,scalinglawsandfinancialfragility’,Journal of Economic Behavior and Organization,Vol.56(4),pages489–512.
Dilaver, O, Jump, R and Levine, P (2016),‘Agent-basedmacroeconomicsanddynamicstochasticgeneralequilibriummodels:wheredowegofromhere?’,Discussion Papers in Economics,DP01/16.
Dosi, G, Fagiolo, G, Napoletano, M, Roventini, A and Treibich, T (2015),‘Fiscalandmonetarypoliciesincomplexevolvingeconomies’,Journal of Economic Dynamics and Control,Vol.52,pages166–89.
Eliasson, G (1991),‘Modelingtheexperimentallyorganizedeconomy:complexdynamicsinanempiricalmicro-macromodelofendogenouseconomicgrowth’,Journal of Economic Behavior and Organization,Vol.16(1–2),pages153–82.
Epstein, J M and Axtell, R (1996),Growing artificial societies: social science from the bottom up,BrookingsInstitutionPress.
Topical articles Agent-based models: understanding the economy 187
Ernest, N, Carroll, D, Schumacher, C, Clark, M and Cohen, K (2016),‘Geneticfuzzybasedartificialintelligenceforunmannedcombataerialvehiclecontrolinsimulatedaircombatmissions,JDefManag,6(144).
Fagiolo, G and Roventini, A (2012),‘MacroeconomicpolicyinDSGEandagent-basedmodels’,Revue de l’OFCE,Vol.5,pages67–116.
Farmer, J D, Patelli, P and Zovko, I I (2005),‘Thepredictivepowerofzerointelligenceinfinancialmarkets’,Proceedings of the National Academy of Sciences of the United States of America,Vol.102(6),pages2,254–259.
Friedman, J, Hastie, T and Tibshirani, R (2001),The elements of statistical learning,Springer,Berlin:SpringerSeriesinStatistics,Vol.1.
Gatti, D D and Desiderio, S (2015),‘Monetarypolicyexperimentsinanagent-basedmodelwithfinancialfrictions’,Journal of Economic Interaction and Coordination,Vol.10(2),pages265–86.
Gatti, D D, Gaffeo, E and Gallegati, M (2010),‘Complexagent-basedmacroeconomics:amanifestoforanewparadigm’,Journal of Economic Interaction and Coordination,Vol.5(2),pages111–35.
Geanakoplos, J, Axtell, R, Farmer, D J, Howitt, P, Conlee, B, Goldstein, J, Hendrey, M, Palmer, N M and Yang, C Y (2012),‘Gettingatsystemicriskviaanagent-basedmodelofthehousingmarket’,The American Economic Review,Vol.102(3),pages53–58.
Gode, D K and Sunder, S (1993),‘Allocativeefficiencyofmarketswithzero-intelligencetraders:marketasapartialsubstituteforindividualrationality’,Journal of Political Economy,pages119–37.
Gualdi, S, Tarzia, M, Zamponi, F and Bouchaud, J P (2015),‘Tippingpointsinmacroeconomicagent-basedmodels’,Journal of Economic Dynamics and Control,Vol.50,pages29–61.
Haldane, A G (2016),‘Thedappledworld’,availableatwww.bankofengland.co.uk/publications/Documents/speeches/2016/speech937.pdf.
Heppenstall, A J, Crooks, A T, See, L M and Batty, M (eds)(2011),Agent-based models of geographical systems,SpringerScienceandBusinessMedia.
Hills, S, Thomas, R and Dimsdale, N (2016),‘Threecenturiesofdata—version23’,availableatwww.bankofengland.co.uk/research/Pages/onebank/threecenturies.aspx.
Hommes, C H (2006),‘Heterogeneousagentmodelsineconomicsandfinance’,Handbook of Computational Economics,Vol.2,pages1,109–86.
Hong, H and Stein, J C (1999),‘Aunifiedtheoryofunderreaction,momentumtrading,andoverreactioninassetmarkets’,The Journal of Finance,Vol.54(6),pages2,143–84.
LeBaron, B, Arthur, W B and Palmer, R (1999),‘Timeseriespropertiesofanartificialstockmarket’,Journal of Economic Dynamics and Control,Vol.23(9),pages1,487–516.
Lux, T and Marchesi, M (1999),‘Scalingandcriticalityinastochasticmulti-agentmodelofafinancialmarket’,Nature,Vol.397(6719),pages498–500.
Lux, T and Marchesi, M (2000),‘Volatilityclusteringinfinancialmarkets:amicrosimulationofinteractingagents’,International Journal of Theoretical and Applied Finance,Vol.3(4),pages675–702.
Macy, M W and Willer, R (2002),‘Fromfactorstoactors:computationalsociologyandagent-basedmodeling’,Annual Review of Sociology,pages143–66.
Mankiw, N G (2006),‘Themacroeconomistasscientistandengineer’,The Journal of Economic Perspectives,Vol.20(4),pages29–46.
Metropolis, N (1987),‘ThebeginningoftheMonteCarlomethod’,Los Alamos Science,Vol.15(584),pages125–30.
Metropolis, N, Rosenbluth, A W, Rosenbluth, M N, Teller, A H and Teller, E (1953),‘Equationofstatecalculationsbyfastcomputingmachines’,The Journal of Chemical Physics,Vol.21(6),pages1,087–92.
Metropolis, N and Ulam, S (1949),‘TheMonteCarlomethod’,Journal of the American Statistical Association,Vol.44(247),pages335–41.
188 Quarterly Bulletin 2016 Q4
Oeffner, M (2008),‘Agent–basedKeynesianmacroeconomics—anevolutionarymodelembeddedinanagent–basedcomputersimulation’,MPRA Paper 18199.
Raberto, M, Teglio, A and Cincotti, S (2008),‘Integratingrealandfinancialmarketsinanagent-basedeconomicmodel:anapplicationtomonetarypolicydesign’,Computational Economics,Vol.32(1-2),pages147–62.
Rosato, A, Strandburg, K J, Prinz, F and Swendsen, R H (1987),‘WhytheBrazilnutsareontop:sizesegregationofparticulatematterbyshaking’,Physical Review Letters,Vol.58(10),page1,038.
Schelling, T C (1969),‘Modelsofsegregation’,The American Economic Review,Vol.59(2),pages488–93.
Schelling, T C (1971),‘Dynamicmodelsofsegregation’,Journal of Mathematical Sociology,Vol.1(2),pages143–86.
Taylor, J B (1993),‘Discretionversuspolicyrulesinpractice’,Carnegie-Rochester Conference Series on Public Policy,North-Holland,Vol.39,pages195–214.
Tesfatsion, L (2002),‘Agent-basedcomputationaleconomics:growingeconomiesfromthebottomup,Artificial Life,Vol.8(1),pages55–82.
Thurner, S, Farmer, J D and Geanakoplos, J (2012),‘Leveragecausesfattailsandclusteredvolatility’,Quantitative Finance,Vol.12(5),pages695–707.
Turrell, A E (2013),‘Processesdrivingnon-Maxwelliandistributionsinhighenergydensityplasmas’,Doctoralthesis,ImperialCollegeLondon.
Turrell, A E, Sherlock, M and Rose, S J (2015),‘Ultrafastcollisionalionheatingbyelectrostaticshocks’,Nature Communications,Vol.6.
von Neumann, J (1951),‘Thegeneralandlogicaltheoryofautomata’,Cerebral Mechanisms in Behavior,Vol.1(41),pages1–2.
Werner, F E, Quinlan, J A, Lough, R G and Lynch, D R (2001),‘Spatially-explicitindividualbasedmodelingofmarinepopulations:areviewoftheadvancesinthe1990s’,Sarsia,Vol.86(6),pages411–21.
Zeeman, E C (1974),‘Ontheunstablebehaviourofstockexchanges’,Journal of Mathematical Economics,Vol.1(1),pages39–49.