.
StochasticCellularAutomataModelforStockMarketDynamics
arXiv:cond-mat/0311372v2 [cond-mat.dis-nn] 24 Nov 2005M.Bartolozzi1andA.W.Thomas1
1
SpecialResearchCentrefortheSubatomicStructureofMatter(CSSM),
UniversityofAdelaide,Adelaide,SA5005,Australia
(Dated:February2,2008)
Abstract
Inthepresentworkweintroduceastochasticcellularautomatamodelinordertosimulatethedynamicsofthestockmarket.Adirectpercolationmethodisusedtocreateahierarchyofclustersofactivetradersonatwodimensionalgrid.Activetradersarecharacterisedbythedecisiontobuy,σi(t)=+1,orsell,σi(t)=−1,astockatacertaindiscretetimestep.Theremainingcellsareinactive,σi(t)=0.Thetradingdynamicsisthendeterminedbythestochasticinteractionbetweentradersbelongingtothesamecluster.Extreme,intermittentevents,likecrashesorbubbles,aretriggeredbyaphasetransitioninthestateofthebiggerclusterspresentonthegrid,wherealmostalltheactivetraderscometosharethesamespinorientation.Mostofthestylizedaspectsofthefinancialmarkettimeseries,includingmultifractalproprieties,arereproducedbythemodel.AdirectcomparisonismadewiththedailyclosuresoftheS&P500index.
PACSnumbers:05.45.Pq,52.35.Mw,47.20.Ky
Keywords:ComplexSystems,Percolation,StochasticProcesses,Multifractality,Econophysics
1
I.INTRODUCTION
SincethesuccessfulapplicationoftheBlack-Scholestheoryforoptionpricing[1]in1973moreandmorephysicistshavebeenattractedbytheideaofunderstandingthebehaviourofthemarketdynamicsintermsofcomplexsystemtheory,whereself-organizedcriticality[2,3]andstochasticprocesses[4,5]playimportantroles.Theaimofthemicroscopicmodelsproposedsofar(forgeneralreviews[6,7])istoreproducesomestylizedfacts[8]concerningthetemporalfluctuationsofthepriceindices,P(t).Inparticular,thelogarithmicpricereturns
R(t)=lnP(t+1)−lnP(t),
andthevolatility,definedinthepresentworkas
v(t)=|R(t)|,
(2)(1)
havebeenstudiedextensively[4]fromanempiricalpointofview.Theresultshaveshownthatwhilelongtimecorrelationsarepresentinthevolatility,aphenomenonknownasvolatilityclustering,theycannotbefoundinthetimeseriesofreturns.Moreoverthelattershowanintermittentbehaviourthatrecallsinsomeaspectshydrodynamicturbulence[9,10,11],characterizedbypowerlawtailsintheprobabilitydistributionfunction(pdf).Microsimulationshavedemonstratedthatthiskindofbehaviourcanoriginatebothasastochasticprocesswithmultiplicativenoise[12,13,14]andasapercolationphenomenon[15,16,17].
Inordertoreproducethesefeaturesofrealmarketsweintroduceastochasticcellularautomatamodel,representinganopenmarket.Thatis,amarketwherethenumberofactivetraders,definedascellswithspinstatedifferentfrom0,namelyσi(t)=±1,evolveintimeaccordingtoapercolationdynamics.Thepercolationdynamicsischoseninordertosimulatetheherdingbehaviourtypicalofinvestors[28].Accordingtothis,activetradersgatherinclustersornetworkswhere,followingastochasticexchangeofinformation,theyformulatethetradingstrategyforthenexttimestep.TheresultsobtainedbythesimulationsarethencomparedwiththetimeseriesofdailyclosuresoftheS&P500index[4,28,29,30,31,34,40,41,45]overaperiodofabout50years.
Moreover,recently,thefractalproperties[18]ofthepricefluctuationshavealsobeen
2
investigatedfordifferentmarkets[19,20,21,22].Acommonfeaturefoundinthesestudiesistheexistenceofanonlinear,multifractalspectrumthatexcludesthepossibilityofefficientmarketbehaviour[23].Theoriginofthemultifractalityofinthefinancialtimeserieshasalsobeenatthecenterofdiscussions[24,25,26,27].Inthispaperweconsiderthemulti-fractalspectrumofthepricefluctuationsasastylizedfactofthemarkettimeserieswithoutaddressinganyquestionabouttheunderlyingprocessabletogenerateit.Themultifractalspectrumisusedasafurthertestforourmodel.
Aparallelbetweenmultifractalandthermodynamicalformalismhasalsobeeninvesti-gated.Wefound,inagreementwiththepreviousworkofCanessa[45],thattheanaloguespecificheatcanprovideagoodtooltocharacterizeintermittency,thatisfinancialcrashesorbubbles,fromathermodynamics-equivalentpointofview.
II.THEMODEL
Inthepresentworkwesimulatethefinancialmarketdynamicsviaastochasticcellularautomatamodel.Theagentsofthemarketarerepresentedbycellsonatwodimensionalgrid,512x128.Theithagentatthediscretetimesteptischaracterizedbythreepossiblestatesorspinorientations,σi(t)=0,±1.Thevalueσi(t)=+1isassociatedwiththepurchaseofastockwhileσi(t)=−1withselling.Theformerstatesarecalledactive.Thecellswithspinvalueσi(t)=0areinactivetraders.Theactivetradersherdinnetworksorclustersviaadirectpercolationmethodrelatedtoaforestfiremodel[15].Theinformationcarriedbytheactivetraders,thatistheirspinstate,issharedwiththeothermembersofthecluster.Thepercolationdynamicsallowatimedependentherdingbehaviourandthemarketcanbeinterpretedasanopensystemnotboundedbyconservationlaws.Theclusteringprocesswillbediscussedindetailinthenextsubsection.
Thetradingdynamicsisinsteadrelatedtothesynchronousupdateofthespinsoftheactivetraders,ruledbyastochasticexchangeofinformationbetweenthem,similartoarandomIsingmodel[12,13].Aparticularfeatureofthepresentsimulationisthattheinformationisnotspreadalloverthegrid,asinothermulti-agentscenarios[12,13],butitislimitedbytheclustersofinteractionpreviouslydefined.Themechanismforthespindynamicsisexplainedin“stochastictradingdynamics”subsection.
3
A.PercolationClustering
Oneoftheaimsofourcellularautomatamodelistoreproducetheherdingbehaviourofactivetraders[28].Werefertoherdingbehaviourasthetendencyofpeopleinvolvedinthemarkettoaggregateinnetworksorclustersofinfluence.Thetradersthenusetheinformationobtainedbytheirnetworkinordertoformulateamarketstrategy.Evenifthetopologicalstructureofthesenetworksofinformationisnotimportant,sinceseveralkindsoflongrangeinteractionareavailablenowadays[13],thenumberofconnectionsforeachtradermustbe,inanycase,finiteandnotextendedoverthewholemarket.Inthisframeworkadirectpercolationmethodisusedtosimulateherdingdynamicsbetweenactivetraders.IfweassumethattheneighboursofinfluencearethoseofvonNeumann(up,down,left,right),thepercolationisfixedbythefollowingparameters:
ph:theprobabilitythatanactivetradercanturnoneofhisinactiveneighboursintoanactiveoneatthenexttimestep,σi(t)=0→σi(t+1)=±1.Thissimulatesthefactthatcertaininformationpossessedbyatradermayinduceapotentialtradertojointhemarketdynamics.
pd:theprobabilitythatanactivetraderdiffusesandsobecomesinactive,σi(t)=±1→σi(t+1)=0,duetoeachofhisinactiveneighbour.Thismimicsthefactthatonlytradersatthebordersofanetwork,thatistheweakerlinks,canquitthemarket.
pe:theprobabilitythatanontradingcellspontaneouslydecidestoenterthemarketdynamics,σi(t)=0→σi(t+1)=±1.
Thevaluesoftheadimensionalparameters,ph,pdandpe,influencethestabilityofthesystemandthepercentageofactivetradersonthegrid.Inordertotestdifferentmarketactivitieswefixthevaluespd=0.05andpe=0.0001whilewetunetheparameterph.Atthebeginningofthesimulationthegridisloadedrandomlywithasmallpercentageofactivetradersandthenthesystemispermittedtoevolveaccordingtothepreviousrules.Ifweareinastablerangeoftheparameterph,afteratransientperiodthatdependsbothontheparametervaluesandtheinitialnumberofactivecells,thenumberofactivetradersonthegridbeginstofluctuatearoundacertainaverage,asshowninFig.1(Top).Themarketcanbeconsideredopensincethenumberofagentschangesdynamicallyintime.Inthisregime,thecompetitionbetweenherdinganddiffusionproducesapowerlawdistributionoftheclustersize,asshowninFig.1(Bottom),ρ(S)≈S−λ,whereSistheclusterdimension,
4
20000Number of Active Traders15000ph=0.0493ph=0.0490ph=0.0488ph=0.0485ph=0.047510000500000200040006000Time Steps8000100001000λ∼1.1100ρ101110S1001000FIG.1:Top:DifferentvaluesoftheparameterHproducedifferentactivitiesoftradersonthegrid.Bottom:Clustersizedistributionforph=0.0485att=9000.
definedasthenumberofactivecellsbelongingtothesamecluster,andλ>0,creatingahierarchyofnetworks.Thishierarchyisnecessaryifwerearetotakeintoaccountarealaspectofthemarket,namelythatdifferenttradersalsohavedifferenttradingpowers.Areasonableassumptionisthatpeoplehavingalargernumberofsourcesofinformation,sobelongingtogreaterclusters,canbeassociatedwithprofessionalinvestorsthat,mostlikely,areabletomoveagreateramountofstockscomparedtotheoccasionalinvestor.Usingthisassumptionweareabletodefineaproperweightforthetradingpowerofdifferentcells,aswewilldiscussinthenextsubsection.
Asimilarpercolationmodelhasalsobeenusedtoreproducesomestatisticalandgeo-metricalfeaturesofsolaractivity[32,33].
B.StochasticTradingDynamics
k
Thedynamicsofthespinsoftheactivetraders,σi(t)=±1fori=1,...,Nk(t)(wherethe
superscriptk,fromnowon,referstothekthclusterofthegridconfigurationattimestept)followsastochasticprocessthatmimicsthehumanuncertaintyindecisionmaking[13].
kTheirvaluesareupdatedsynchronouslyaccordingtoalocalprobabilisticrule:σi(t+1)=+1kkkwithprobabilitypkiandσi(t+1)=−1withprobability1−pi.Theprobabilitypiis
5
determined,byanalogytoheatbathdynamicswithformaltemperaturekbT=1,by
pki(t)=
1
Nk(t)
Nk(t)
kk
Akijσj(t)+hi.
(4)
j=1
k
TheAkij(t)aretimedependentinteractionstrengthsbetweenagentsandhi(t)isanexternal
fieldreflectingtheeffectoftheenvironment[13].Theinteractionstrengthsandtheexternal
kkkfieldchangerandomlyintimeaccordingtoAkij(t)=Aξ(t)+aηij(t)andhi(t)=hζi(t).
Thevariablesξk(t),ηij(t),ζik(t)arerandomvariablesuniformlydistributedintheinterval(-1,1)withnocorrelationintimeorspace.Themeasureofthestrengthsofthepreviousterms,A,aandh,areconstantandcommonforallthegrid.
Inthiscontestthedynamicsofthepriceindex,P(t),canbeeasilyderivedifweassumethattheindexvariationisproportionaltothedifferencebetweendemandandsupply,
dP
1050-5-1019501050-5-10200040006000Time Steps800010000196019701980Decimal Years19902000FIG.2:Top:NormalizedlogarithmicreturnsfortheS&P500.Bottom:Timeseriesofreturnsreproducedwiththesimulationwithph=0.0493,A=1.8andh=0.
analyzediscomposedbyofthedailyindicesfrom3/1/1950to18/7/2003foratotalof13468data.Thetimeseriesoftheindexprices,P(t),isconvertedintothelogarithmicreturns(1)andthennormalizedoverthetimeinterval,T,
r(t)=
R(t)−R(t)T
rrThestrengthAalsoplaysanimportantroleinthetradingdynamics.Withphfixed,thisparameterisrelatedtotheintermittencyofthesystem.Bothforlargevaluesoftheactivity(A>10)andforA→0weobserveanapproachofthepdftowardaGaussian-likeshape.Thatisvery,largefluctuationsbecomemoreandmorerare,andAcanberegardedasatemporalscaleforthesystem,similartotheactivityparameterintheCont-Bouchaudmodel[16].Inspiteofthissomelargefluctuationscanbestillidentified.ThisisprobablyoneofthemaindifferencesbetweentheCont-Bouchaudmodelandthepresent.InfactMonteCarlosimulationsoftheformer[17]showthatanincreaseoftheactivitybringsarapidconvergencetowardaGaussiandistributionbecausealargenumberofclustersaretradingatthesametimefollowingarandomprocedureofdecisionmaking[16,17]:therearenoclustersthatcaninfluencethemarketmorethanothersandsotheresultingglobalinteractionisnoise-like.Inourmodelfluctuationsarealwaysallowedbecauseoftheheatbathdynamics.Clustersofactivetraderscanalwaysbesubjectedtophasetransitions,independentlyofthestateofotherclusters,creatingadisplacementbetweendemandandsupply.
Copyright © 2019- yrrf.cn 版权所有 赣ICP备2024042794号-2
违法及侵权请联系:TEL:199 1889 7713 E-MAIL:2724546146@qq.com
本站由北京市万商天勤律师事务所王兴未律师提供法律服务