Exploring the Nature of Trading Intuition

Transcription

1 ExploringtheNatureof TradingIntuition AntoineJBruguier 1,StevenRQuartz 1 andpeterlbossaerts 1,2 1CaliforniaInstituteofTechnology,Pasadena,CA91125,USA 2EPFL,CH 1015Lausanne,Switzerland Thisversion:28July2008 Abstract TheEfficientMarketsHypothesis(EMH)andtheRationalExpectations Equilibrium(REE)bothassumethatuninformedtraderscancorrectlyinfer informationfromthetradingprocess.experimentalevidencestartinginthemid 80shasconsistentlyconfirmedtheabilityofuninformedtraders,evennovices, todoso.here,wehypothesizethatthisabilityrelatestoaninnatehumanskill, namely,theoryofmind(tom).withtom,humansarecapableofdiscerning maliciousorbenevolentintentintheirenvironment.weconfirmourhypothesis inevidencethatparticipationinmarketswithinsidersengagesthe(evolutionary relativenew)brainregionsthatspecializefunctionallyintom,andinevidence thatskillinpredictingpricechangeswhenthereareinsidersiscorrelatedwith scoresontwotraditionaltomtests.sincetomisgenerallyunderstoodtorelyon patternrecognition(whichrecentlyhasbeenconfirmedformallyinthecontext ofstrategicgames),wesearchedforfeaturesinthedatawithwhichonecould identifythepresenceofinsiders.wediscoveredthatgarch likepersistencein absolutepricechangesincalendartimecharacterizesmarketswithinsiders. 1

2 I.Introduction Thispaperreportsresultsfromexperimentsmeanttoexplorehowuninformed tradersmanagetoreadinformationfromtransactionpricesandorderflowin financialmarketswithinsiders.sincetheseminalexperimentsofcharlesplott andshyamsunderintheearly1980s[plottandsunder(1988)],ithasbeen repeatedlyconfirmed(andwewilldosoheretoo)thatuniformedtradersare quitecapableofquicklyinferringthesignalsthatinformedtraders(insiders) receiveaboutfuturedividends,despitetheanonymityofthetradingprocess, despitelackofstructuralknowledgeofthesituation,anddespiteoftheabsence oflonghistoriesofpastoccurrencesofthesamesituationfromwhichtheycould havelearnedthestatisticalregularities. Itisstrikingthatsolittleisunderstoodabouttheabilityoftheuninformedto inferthesignalsofothers.thisabilityconstitutesthebasisoftheefficient marketshypothesisemh[fama(1991)],whichstatesthatpricesfullyreflectall availableinformation.underlyingemhistheideathattheuninformedwilltrade onthesignalstheymanagetoinfer,andthat,throughtheordersofthe uninformed,thesesignalsareeffectivelyamplifiedinthepriceformation.inthe extreme,priceswillfullyreflectallavailableinformation.withoutabetter understandingofhowuninformedpracticallymanagetoreadinformationin prices,emhremainsahypothesiswithoutawell understoodfoundation. Thefeedbackfromtradingbasedoninferredinformationtopriceformationhas beenformalizedintheconceptoftherationalexpectationsequilibrium(ree) [Green(1973),Radner(1979)].Foreconomists,REEformsthetheoretical justificationofemh.butagain,reetakestheabilityoftheuninformedto correctlyreadinformationfrompricesasgiven,ratherthanexplainingit.as such,likeemh,reelacksawell articulatedfoundation. Thegoaloftheexperimentsthatwereportonherecanbeexpressedinamore mundaneway,asanattempttobetterdefinewhatismeantwith trading intuition, andtounderstandwhysometradersarebetterthanothers.books havebeenwrittentoelucidatetradingintuition[fenton O'Creevy,Nicholson, 2

3 SoaneandWillman(2005)]butattemptsatformalizingthephenomenonhaveso farfailed. Whatdistinguishesmarketswithinsidersisthepresenceofawinner scurse: salesareoftensuccessfulonlybecausepriceshappentobetoolow(relativeto theinformationoftheinsiders),whilepurchasesmayoccuronlyatinflated prices.eitherway,theuninformedtraderishurt.whilethewinner scurseis usuallyassociatedwithstrategic,single sidedauctions,italsoappliesto competitive,double sidedmarkets,andindeed,thewinners curseisnotonly implicitinthetheoryofreebutalsoverymuchofconcerninreal worldstock markets[biais,bossaertsandspatt(2008)] Fromthepointofviewoftheuninformedtrader,thewinner scurseconjuresup animageofpotentialmalevolenceinthetradingprocess.humansareactually uniquelyendowedwiththecapacitytorecognizemalevolence(aswellas benevolence)intheirenvironment,acapacitythatpsychologistsrefertoas TheoryofMind[GallagherandFrith(2003)].TheoryofMindisthecapacityto readintentionorgoal directnessinpatterns,throughmereobservationofeye expression[baron Cohen,Jolliffe,MortimoreandRobertson(1997)],or movementofgeometricobjects[heiderandsimmel(1944)],inthemovesofan opponentinstrategicplay[mccabe,houser,ryan,smithandtrouard(2001), Gallagher,Jack,RoepstorffandFrith(2002),Hampton,BossaertsandO'Doherty (2008)],orinactionsthatembarrassothers( faux pas )[Stone,Baron Cohen andknight(1998)].theoryofmindisthustheabilitytoreadbenevolenceor malevolenceinpatternsinone ssurroundings.thiscontrastswithpredictionof outcomesgeneratedpurelyby,say,physicallaws,whichmayharmorbe beneficialbutwithoutintention. 1 PerhapsthemosttellingexampleofhowTheoryofMindconcernsdetectionof intentionalityinpatternsconcernsacasewheresimplephysicallawsofmotion areviolated.subjectsarefirstshownanobject(say,ablock)attractedtoatarget (ball)butencountersaphysicalbarrier(awall)thatitovercomesbymoving aroundit(seefigure1,top).subsequently,thebarrierisre movedandtwo situationsareshown.inthefirstone,theblock,freedoftheobstacle,moves 3

4 alongastraightpathtowardstheoriginaltargetlocation(figure1,middle);in thesecondone,theblockcontinuesalongtheoriginaltrajectory(figure1, bottom).thefirstsituationaccordswithstandardphysics;thesecondoneis unusual(fromaphysicspointofview)andmayalertthesubjectthattheoriginal patharoundthebarrierwasintentional,i.e.,thatitreflectedgoal directed behavior.ithasbeenrepeatedlyobservedthathumans(includingone yearold infants[gergely,nadasdy,csibraandbiro(1995)]andsomenon human primates[ullerandnichols(2000)]spendmoretimegazingatthesecond situationthanthefirstone,despiteitsnovelty(theballtakesadifferentpath), suggestingthatthesuspectedintentionalitydrawstheirattention. Thefirstgoalofourstudywas,throughexperiments,todeterminetowhat extenttradingintuitionandtomarerelated.tom,asmentionedbefore,isthe abilitytoreadbenevolenceormalevolenceinpatternsinone senvironment.so, itwasreasonabletoconjecturethatitappliedtoorderflowsinfinancialmarkets withinsidersaswell.specifically,weaskedwhethertheuninformedengagetom brainregionswhenshowntheorderflowofamarketwithinsiders,andwhether ToMskillscorrelatewithsuccessinforecastingmarketpriceswhenthereare insiders. Theanswerstobothquestionsareaffirmative.Wefoundthatinsideinformation engaged(only)functionallyspecifictomregionsinthebrainoftheuninformed, andwefoundthatabilitytocorrectlyforecastpricechangesinmarketswith insiderscorrelatedsignificantlywithperformanceinstandardtomtasks. ThesecondgoalofourstudywastostartformalizingtheconceptofToMtothe extentthatitappliedtofinancialmarketswithinsiders.tomisaboutpattern recognition whichisintuitivelyclearfromtheexampleinfigure1,andwas recentlyconfirmedformallyforplayinstrategicgames[hampton,bossaertsand O'Doherty(2008)].Formarkets,however,itisnotevenknownwhetherthere areanyfeaturesatallinthedatathatwouldallowonetomerelyidentifythe presenceofinsiders(inanalogywiththeviolationofphysicallawsinthebottom paneloffigure1thatpromptsmanytoconcludethattheremaybe 4

5 intentionality).existenceofsuchfeaturesisanimportantcomponentwhen arguingthattommayapply. Here,wereportresultsfromastatisticalinvestigationoftradeflowsinour experimentalfinancialmarketsthroughwhichthepresenceofinsiderscouldbe discriminated.wefindthatgarch likefeaturesemergewhenthereareinsiders. Specifically,autocorrelationcoefficientsofabsolutepricechangesincalendar timearesignificantlymoresizeable. ItseemstousthatthelinkbetweenGARCH likefeaturesandinsidertradingisa majorfindingthatdeservesfurtherinvestigationirrespectiveofitsimportfor theplausibilityoftominthecontextoffinancialmarkets. Remainderisorganizedasfollows.SectionIIdescribestheexperimentsand discussestheresults.sectioniiipresentsthefindingsfromstatisticalcontrasts oftradeflowsinsessionsinourexperimentswhenthereareinsidersandwhen therearenone.furtherdiscussionisprovidedinsectioniv. II.DescriptionOfTheExperimentsAndTheResults Here,weprovidebriefexplanationsoftheexperimentalset up(afulldiscussion canbefoundintheappendix)andwepresentthemainempiricalresults. Wefirstranamarketsexperimentforthesolepurposeofgeneratingorderand tradeflowinacontrolledsetting,tobeusedinthemainexperiments.thelatter consistedof:(i)abrainimagingexperiment,meanttodiscernhowsubjects attemptedtointerpretthedatabylocalizingareasofthebrainthatareactive duringre playofthemarkets;(ii)abehavioralexperiment,wherewetestedfor subjects abilitytopredicttransactionprices,andtoascertaintheirgenerictom skills(wealsoranatestoflogicalandmathematicalthinking,forareasontobe discussedlater). II.A.Step1:TradingDataCollectionExperiment Twenty(20)subjectsparticipatedinaparameter controlledmarketexperiment that used an anonymous, electronic exchange platform (jmarkets; see 5

6 times; each replication will be referred to as a session. Subjects were endowed with notes, cash, and two risky securities, all of which expired at the end of a session. The two risky securities ( stocks ) paid complementary dividends between 0 and 50 : if the first security, called stock X, paid x cents, then the secondsecurity,calledstockz,wouldpay50 xcents.thenotesalwayspaid50. Allocationofthesecuritiesandcashvariedacrosssubjects,butthetotalsupplies of the risky securities were equal; hence, there was no aggregate risk, and, theoretically as well as based on observations in prior experiments[bossaerts, Plott and Zame (2007)], prices should converge to levels that equal expected payoffs; that is, risk neutral pricing should arise. Subjects could trade their holdings for cash in an anonymous, continuous open book exchange system. Subjects were not allowed to trade security Z, however. Consequently, risk aversesubjectsthatheldmoreofxthanofzwouldwanttosellatrisk neutral prices; the presence of an equal number of subjects with more of Z than of X allowedmarketstoclear,inprinciple.inmostofthesessions,tobereferredtoas testsessions,anumberofsubjects(the insiders )weregivenanestimateofthe dividendintheformofa(common)signalwithin$0.10oftheactualdividend.all subjects were informed when there were insiders; only the insiders knew how many insiders there were. Subjects were paid in cash according to their performance and made $55 on average. More details can be found in the appendix. Figure2displaystheevolutionoftransactionpricesthroughouttheexperiment. Trading was brisk, independent of the type of session; on average, traders entered or cancelled an offer every 0.7s and one transaction took place every 3.2s. During test sessions, the uneven distribution of information evidently skewedthetransactionprices.overall,theevolutionofpricesisconsistentwith priorexperiments[whichdidnotcontrolforaggregaterisk,however;plottand Sunder(1988)]. As mentioned before, the sole purpose of the markets experiment was to generate order flow and transaction data from markets with tight control over endowments,informationandexchangeplatform.theresultingdataformedthe 6

7 inputforthebrainimagingandbehavioralexperiments.forthisreason,wewill notelaborateonthemarketsexperiment.wewillcomebacktorelevantfeatures ofthetradingflowinthenextsection,however. II.B.Step2:fRMIExperiment Eighteen(18)newsubjectswereshownareplayofthe13previouslyrecorded marketsessionsinrandomorderwhilebeingscannedwithfmri.thesesubjects played the role of outsiders who did not trade. Subjects were given the instructions of the markets experiment, so that they were equally informed as theoutsidersinthatexperiment.subjectsfirstchosewhethertheywouldbeton stock X or Z, after being told whether there were insiders in the upcoming session or not. Subsequently, the order flow and transaction history of stock X was replayed in a visually intuitive way (see Figure 3 and Video 1). During replay, subjects were asked to push a button each time they saw a trade, indicatedbya500ms(millisecond)changeincolorofthecirclecorrespondingto thebestbidorask. Our design had three main advantages. First, the subjects did not trade during thereplay(theydidhavetotakepositionsbeforethereplay).whilethequestion ofhowthehumanbrainmakesfinancialdecisionsisinteresting,weneededfirst to understand how humans perceive a stock market. By not introducing decision making,weavoidedaconfoundingfactor.second,theperiodswithout insiders were perfect controls. Since the trading data acquisition method, the display screens, and the number of traders were the same, the two types of sessions were identical in every respect except for the presence or absence of insiders. Third, by adding a blind bet, we elicited a feeling of randomness. Indeed,ifwehadforcedsubjectstochoosestockXforeverysession,thepayoff wouldhavebeenthesamefixednumberforeverysubject.moreover,wecould nothaveseparatedanincreaseinstockpricefromahigherexpectedreward,as thesetwosignalswouldhavebeenperfectlycorrelated.instead,byintroducinga blindbet,weorthogonalizedexpectedrewardandstockprice. 7

8 Toidentifyareasofsignificantbrainactivationintheinsidersessionsrelativeto the control sessions, we used a standard approach as implemented in the packagebrainvoyager(braininnovation,maastricht,thenetherlands).wefita General Linear Model (GLM) to the (filtered, motion corrected) activation data for each voxel (a cubic volume element of 27mm 3 ) which included several auxiliary predictors (to capture motion and visual effects) as well as the expectedrewardforthesubjectandfourtask relevantpredictors(figure4).the task relevant predictors were as follows. First, we constructed a parametric regressor that would quantify the effect of the insiders (if present) on stock prices.weusedtheabsolutevalueofthedifferencebetweenthestock strading price and 25. Separate parametric predictors were thus computed for the sessions with insiders and without insiders. This way, the sessions without insiderscouldbeusedascontrols,againstwhichthesessionswithinsiderscould becompared.second,weaddedblockregressors(dummyvariables),toidentify sessionswithandwithoutinsiders. All predictors were convolved with the standard hemodynamic response function, inordertoaccountforthelagbetweenactivationattheneuronallevel and the fmri signal. Areas of significant brain activation contrast were then recovered as those areas where the difference (across insider and no insider sessions)intheslopecoefficientsofthepredictorsweredeemedtobesignificant usingstandardstatisticaltests. Our way of modeling brain activation had several advantages. First, as mentionedabove,westrippedouttheeffectofdecision makinginthebrainand focused on the perception of the market. Second, the predictors we built were orthogonaltoexpectedrewardasweusedtheabsolutevalueofthedistanceto 25 ratherthanthe(signed)differencewith25.third,bycontrastingsessions with and without insiders and making sure that these two types of sessions displayednoobviousdifferencesintradingactivity,wecontrolledfordifferences invisualactivationandisolatedtheeffectofinsidersonsubjects perceptionof themarket. 8

9 Wefoundasignificantcontrastfortheparametricregressorsinonelargeregion ofparacingulatecortex(pcc;figure5 aandtablei).wealsofoundsignificant activationsinasmallerregioninthefrontalpartoftheanteriorcingulatecortex ( 14; 23; 39; 5 voxels; Figure 5 a and Table I). Finally, we found strong differentialactivationsinrightamygdala(21; 10; 12;5voxels;Figure5 band TableI)andleftinsula( 30; 7;11;5voxels;Figure5 candtablei). Significantcontrastsfortheblockpredictorsshowedupinalargeareaoflingual gyrus ( 9; 65; 6; 25 voxels; Figure 6 and Table II), as well as a small area of cerebellum( 13; 58; 30;9voxels;Figure6andTableII).Nootherareaswith fiveormorevoxelsexhibitedsignificantcontrasts. OurresultsprovidesupportforthehypothesisthatTheoryofMindisinvolved when subjects are facing markets with insiders. We found that the same brain areas are activated as during traditional ToM tasks. PCC(paracingulate cortex) activationisstandardintasksinvolvingtom[gallagherandfrith(2003)],andin strategic games in particular [Gallagher, Jack, Roepstorff and Frith (2002)], [McCabe,Houser,Ryan,SmithandTrouard(2001)].Similaractivationwasalso reported in choice vs. belief game theory experiments [Bhatt and Camerer (2005)]. The PCC has also been observed in tasks that involve attribution of mental states to dynamic visual images, such as intentionally moving shapes [Castelli,Happe,FrithandFrith(2000)]. Wealsofoundthatactivationintherightamygdalaandtheleftanteriorinsula increased as transaction prices deviated from the uninformed payoff. While some studies have reported the involvement of these structures in ToM tasks [Baron Cohen, Ring, Wheelwright and Bullmore (1999), Critchley, Mathias and Dolan (2001)], they are more typically regarded as involved in processing affective features of social interaction. Specifically, the amygdala is a critical structure in the recognition of facial emotional expressions of others [Phillips, Young, Scott, Calder, Andrew, Giampietro, Williams, Bullmore, Brammer and Gray (1999), Phan, Wager, Taylor and Liberzon (2002), Morris, Ohman and Dolan(1998)]whiletheanteriorinsulaisthoughttoplayacriticalrolebothin subjective emotional experience [Bechara and Damasio (2005)] and in the 9

10 perception/empathetic response to the emotional state of others [Singer, Seymour, O'Doherty, Kaube, Dolan and Frith (2004)]. A complementary interpretation of these activations is that they may also reflect subjects own emotionalresponsestomarketactivity,whichisconsistentwithrecentpsychophysiological evidence that financial market participation engages somatic marker(emotional)circuitryduringheightenedmarketvolatility[loandrepin (2002)].Futureresearchshouldshedmorelightonthispotentiallinkbetween marketparticipationandemotions. We also found activations in the frontal part of the ACC (anterior cingulate cortex). Our activation is in a slightly more posterior and dorsal location than when ToM is used in strategic and non anonymous, simple two person games [Gallagher,Jack,RoepstorffandFrith(2002),McCabe,Houser,Ryan,Smithand Trouard(2001)]. The increased activation of lingual gyrus in the presence of insiders provides further support for the role of ToM in market perception. For example, the lingual gyrus is involved in the perception of biological motion, a key cue for mentalizing [Servos, Osu, Santi and Kawato (2002)]. However, increased activation of lingual gyrus may also be related to recent accounts that this structureisinvolvedincomplexvisualtaskswheresubjectsareaskedtoextract global meaning despite local distractors [Fink, Halligan, Marshall, Frith, Frackowiak and Dolan (1996), Fink, Halligan, Marshall, Frith, Frackowiak and Dolan(1997)].Whentherearenoinsiders,subjectscanconcentrateonthetask we imposed, namely, to track trades. In our display, transactions were a local feature, indicated by changes in color of circles in the middle of the screen. In contrast, when there are insiders, the entire list of orders may reflect information with which to re evaluate the likely payoff on the securities but at the same time, subjects are still asked to report all transactions, which now amounts to a local distraction. Our interpretation of lingual gyrus activation is inconsistent with theories [Easley and O'Hara (1992), Glosten and Milgrom (1985)] that predict that only the trade prices are relevant to update beliefs. If thesetheorieswereright,otherfeaturesofthedisplayneednotbewatched,and simultaneously keeping track of transactions is not a distraction. Rather, our 10

11 finding is consistent with experimental evidence that insiders do not exploit theirsuperiorinformationbyexclusivelysubmittingmarketorders[barner,feri andplott(2005)].futureresearchshoulddeterminetowhatextentlingualgyrus activation reflects ToM (through motion of objects) or the proverbial conflict betweenthe forest (insiderinformation)andthe trees (trades). We did not observe any significant activation in brain areas related to formal mathematical reasoning. In particular, there was no evidence of estimation of probabilities [Parsons and Osherson (2001)] or arithmetic computation [Dehaene,Spelke,Pinel,StanescuandTsivkin(1999)].Wealsodidnotobserve any significant activation in brain areas related more generally to problemsolving or analytical thought [Newman, Carpenter, Varma and Just (2003)] or reasoning[acuna, Eliassen, Donoghue and Sanes(2002)]. This finding has also beenhighlightedinarecentstudyoftominstrategicgames[coricelliandnagel (2008)].WeshallelaborateinthenextSection. ItisimportanttonotethatengagementofToMbrainstructureswasnotsimply inresponsetoacomplexgraphicalrepresentationofthemarket,asbothcontrol andtestsessionsgeneratedapproximatelyequivalentmarketactivity.similarly, in both conditions the market activity was the result of human interactions. Hence,itwasnotmerelythepresenceofhumanactivitythatresultedinToM,as this was equivalent in both conditions. Rather, the salient difference between session type was the fact that, during sessions with insiders, trading activity providedinformationthatcouldbeextractedwiththeappropriateinferences. II.C.Step3:BehavioralExperiment The behavioral experiment was meant to correlate subjects ability to predict tradepricesinmarketswithinsidersandtheirgeneraltomskillasreflectedin traditional ToM tests. We added a test for mathematical and logical thinking, promptedbytheabsenceofactivationinregionsofthebrainusuallyassociated withformalmathematicalandprobabilisticreasoninginourtask. 11

12 Forty three (43) new subjects were given a series of four tasks that were administeredinrandomorder.thefirsttaskwasafinancialmarketprediction (FMP)taskinwhichthetradingactivitywasreplayedtothesubjectsatoriginal speedandpausedevery5seconds.duringhalfofthepauses,weaskedsubjects to predict whether the next trade was going to occur at a higher, lower, or identicalpriceastheprevioustradethathadhappened.fortheotherhalfofthe pauses, we reminded subjects of their predictions and informed them of their success. The second task was a ToM task based on eye gaze [Baron Cohen, Jolliffe,MortimoreandRobertson(1997)].ThethirdtaskwasaToMtaskbased on displays of geometric shapes whose movement imitated social interaction [HeiderandSimmel(1944)].AswiththeFMPtask,wepausedthemovieevery fivesecondsandaskedsubjectstopredictwhethertwooftheshapeswouldget closer,farther,orstayatthesamedistance.fortheotherhalfofthepauses,we reported the outcomes and the subjects predictions. The fourth task (the M task) involved a number of standard mathematics and probability theory questions[andfrequentlyusedinwallstreetjobinterviews seecrack(2004)]. ThebehavioralexperimentconfirmedtheuseofToM.Weobservedasignificant correlation(p=0.048andp=0.023)betweenthefmptaskandbothtomtasks (Figure7,toppanels)butnosignificantcorrelation(p=0.22)betweentheFMP andthemtasks(figure7,bottomleft).surprisingly,wedidnotobserveany correlationbetweenthetwotomtests(figure7,bottomright).self reports duringthebehavioralexperimentdidnotshowanyevidenceofpersonalization. Wedidnotobserveanysignificantgenderdifferences. Our behavioral experiments corroborate the relationship between trading successandtomthatourfmriresultssuggest.thecorrelationbetweenthetwo ToMtestsandtheabilitytopredictpricechangesinafinancialmarketconfirms thatsuccessfultradersusetom.wealsofoundthatthesubjectscoresonthetwo ToMtestswerenotcorrelated(Figure7 D).ThisfindingsuggeststhatToMisa multi dimensionalsetofabilities,apossibilitythathasnotbeenraisedyetinthe ToMliterature:eachtestrepresentsonedimensionofToMabilities,butfinancial abilitiescorrelatewithbothdimensions. 12

13 Lackof(significant)correlationbetweenthescoresonthetwoToMtests providesindirectconfirmationthatgeneralintelligenceorstateofattentiveness cannotexplainthesignificantcorrelationsbetweenscoresonthefmpandtom tests. IV.TheoryofMindinMarketswithInsiders:WhatPatternsToAttendto? Formally,ToMremainsaratherelusiveconcept.Itisoftendefinedonlyvaguely, andmostlyintermsofspecifictasks[gallagherandfrith(2003)],orintermsof activationofparticularbrainregions[mccabe,houser,ryan,smithandtrouard (2001),Gallagher,Jack,RoepstorffandFrith(2002)].Itisgenerallyaccepted, however,thattomconcernspatternrecognition,butonlyinsimpleexamplesis itimmediatelyobviouswhatpatternsareinvolved.figure1(bottompanel) providesanexample:themotionofanobjectviolatesphysicallawsandhenceis readilyrecognizedasrevealingintention. Inthecontextofstrategicgames,however,onerecentstudyhassuccessfully identifiedthepatternsinthemovesofone sopponentonwhichtombuilds. Thus,incollaborationwithAlanHamptonandJohnO Doherty[Hampton, BossaertsandO'Doherty(2008)],oneofusestablishedthat,whenasubjectis engagedinplayingtheinspectiongame,brainactivationintomregionsofthe humanbrainreflecttheencodingofanerrorinpredictinghowone sopponent changesmovesasaresultofone sownactions. 2 TheimportofthisfindingisthatToMingameplaycanbeconcretizedintermsof precisemathematicalquantitiesthatcharacterizeanopponent sactualplay.as such,tomconcernsconcrete, online learningofgameplay.thisisconsistent withthepropositionthattominvolvespatternrecognition detectionofthe natureofintentionalityinone senvironment. ItdeservesemphasisthatToMthereforecontrastswithNashreasoning,where playerswouldsimplyhypothesizethatopponentschoosenashequilibrium strategiesandthattheywouldsticktothem.unliketom,nashreasoningis offline: itworksevenifoneneverseesanymoveofone sopponent;itisalso 13

14 abstract:itconcernswhattheopponentcouldrationallydoandhowtooptimally respond.consistentwiththeideathattomandnashthinkingarenotrelated, brainregionsthatareknowntobeengagedinabstractlogicandformal mathematics[dehaene,spelke,pinel,stanescuandtsivkin(1999),newman, Carpenter,VarmaandJust(2003),Acuna,Eliassen,DonoghueandSanes(2002)], donotdisplaysignificantactivationduringgameplay[coricelliandnagel (2008)]. Inthecontextofourmarketswithinsiders,itisbyfarclearonwhichpatterns ToMcouldbuild.Thisisbecauseithasnotevenbeenestablishedwhetherthere areanypatternsthatdistinguishmarketswithandwithoutinsiders.butour findingthatsubjectsengageintomtocomprehendinsidertradingandthe generallyacceptedviewthattomconcernspatternrecognitionsuggeststhat theyexist. Consequently,wesetouttodeterminewhethertherearefeaturesofthedata thatdifferentiatedmarketswithinsiders.inthephysicalmovementofobjects, violationofphysicallawssignalsthepresenceofintentionality(seefigure1).in markets,wehopedtoidentifystatisticalpropertiesofthedatathatarenot presentwhentherearenoinsiders.wefocusedonanexaminationofthetrade flow. Welookedatahostoftimeseriespropertiesofthetradeflowsinthemarkets experimentthatformedthebasisofourstudy,suchasdurationbetweentrades orskewnessintransactionpricechanges.intheend,itwaspersistenceinthe sizeoftransactionpricechangesovercalendartimeintervalsthatprovidedthe onlystatisticallysignificantdiscrimination.assuch,garch likefeaturesappear todistinguishmarketswithandwithoutinsiders. Specifically,wecomputedtransactionpricechangesoverintervalsof2s(as mentionedbefore,thereisatradeevery3.7sonaverage,socalendartimetick sizewaschosentobeslightlyshorterthanaveragedurationbetweentrades). Wefollowedstandardpracticeandtookthelasttradedpriceineach2sinterval asthenewprice,andiftherewasnotradeduringaninterval,weusedthe transactionpricefromthepreviousinterval. 14

15 Wethencomputedthefirstfiveautocorrelationsinthesize(absolutevalue) 3 of transactionpricechanges.figure8 aplotstheresultsfortwoadjacentperiods, Periods7and8.AscanbeinferredfromFigure2,therewere14insiders(outof 20subjects)inPeriod7,whiletherewerenoneinPeriod8.Thepatternsinthe autocorrelationsoftheabsolutepricechangesinthetwoperiodsarevery different.thereissubstantialautocorrelationatalllagsforperiod7(whenthere areinsiders)whilethereisnoneforperiod8(whentherearenoinsiders). Figure8 bshowsthatthisisageneralphenomenon.plottedisthesumofthe absolutevaluesofthefirstfiveautocorrelationcoefficientsagainstthenumberof insiders.werefertotheformeras GARCHintensity becauseitmeasuresthe extenttowhichthereispersistenceinthesizeofpricechanges.garchintensity increaseswiththenumberofinsiders;theslopeissignificantatthe5%leveland ther squared,at0.32,isreasonablyhighgiventhenoiseinthedata. Consequently,itappearsthatGARCH likefeaturesintransactionpricechanges provideonewaytorecognizethepresenceofinsiders,andhence,afoundation onwhichtominmarketswithinsiderscouldbuild.assuch,wehaveshownthat thereindeedexistpatternsinthedatathatreflectintentionality,agenerally acceptedconditionfortomtoapply.fromthisperspective,ourfindingsthat participationinmarketswithinsidersengagestombrainregionsandthattom skillsandpriceforecastingperformancearecorrelatedmakesense. WehavenotidentifiedyetwhataspectoftheseGARCH likefeaturessubjectsare exploitingtoinferinsideinformation,butourresultsopenupmanypromising avenuesforfutureresearch.ourapproachofcombiningmarketsexperiments withbrainimagingandindividualbehavioraltestingmaythuseventuallyleadto aformalunderstandingofhowhumansmanagetocomprehendasocial institutionascomplexasafinancialmarket. V.FurtherDiscussion Ourimagingresultsformarketswithinsidershaveatleastonemoreaspectin commonwithrecentneurobiologicalstudiesoftominstrategicgames,namely, theabsenceofactivationinregionsofthebraininvolvedinformalmathematical 15

16 reasoning.theresultssuggestthatabstractreasoningskillsareunrelatedto performanceineitherdomain.withrespecttostrategicgames,itisnotknown whetherperformanceiscorrelatedwithmathematicalskill.withrespectto marketswithinsiders,wesetouttoinvestigatetheissue.specifically,aspartof thethird,behavioralexperiment,wetestedoursubjectsontheirabilitytosolve mathematicsandlogicproblemsandcorrelatedtheirscoreswithperformance onthemarketpredictiontest.wefocusedonasetofformalproblemsthathave becomepopularinrecruitmentinthefinancialindustry[crack(2004)].figure6 (BottomLeftPanel)showsthat,whilethecorrelationwaspositive,itwas insignificant.thebehavioralresultsarethereforeinlinewiththeimaging results. Our study is not only relevant to finance. It contains at least two major contributions for psychology and neuroscience. We are the first to report activationintombrainregionsduringtheperceptionofanonymousmulti agent systems, significantly extending the scope of ToM. Moreover, along with Hampton, Bossaerts and O'Doherty (2008) and Coricelli and Nagel (2008), we arethefirsttoquantifytom.inparticular,whentherewereinsiders,themore transaction prices deviated from the uninformed payoff of 25, the more evidence outsiders had that there was important information to be inferred, makingmentalizingincreasinglyimportant. Personalization(oranthropomorphization)isoftensuggestedasanexplanation whytomsometimesapplieseveninsituationsthatdonotdirectlyinvolve personalcontactwithanotherhumanbeing,suchasintheheidermovieusedin oneofourtomtests[heiderandsimmel(1944)].inthecontextoffinancial markets,personalizationemergesattimesinspeech,whenmarketsare describedasbeing exuberant, jittery, or anxious [Oberlechner(2004)]. Answersontheexitsurveyfollowingourbehavioralexperiment,however,do notrevealanyevidenceofpersonalization. Hence,itappearsthatthepersonalcomponentisnotimportant.Intentionality thatcanbeinferredfrompattersinone senvironmentis,however.thisshould 16

17 explainwhytomalsoappliestolarge scaleanonymoussocialinteraction(such asourfinancialmarkets). Theimportanceofintentionalityalsoclarifiesrecentevidence[McCabe,Houser, Ryan,SmithandTrouard(2001),Gallagher,Jack,RoepstorffandFrith(2002)] thattombrainregionsareactivatedwhenplayingstrategicgameswithhumans andnotwhenplayingagainstapre programmedcomputer.theevidenceshould beinterpretednotasindicatingthatthehumancomponentpersewasimportant, butassuggestingthatpotentialmalevolencewascrucial.indeed,thecomputers werepre programmedinawaythatwascompletelytransparenttosubjects,and didnotinvolveanyintentiontoharmorexploit;theyfollowedspecific laws justliketheobjectinfigure1,middlepanel,obeyedphysicallaws. Ourfindingshaveimmediateimplicationsforhiringpracticeinfinance.They suggestthatthefinancialindustryshouldtestcandidatetradersonsocialskills, especiallythoseinvolvingtom,suchastheabilitytorecognizeintentinother people seyes,orrecognizemalevolenceinthemovementofgeometricobjects, and,becauseithasrecentlybeenshowntoalsoinvolvetom,skillinplaying strategicgames.ouradviceisfarmorespecificthantheonetocomeoutofa recentanalysisofoveronehundredprofessionaltraders,whichrevealedthat tradercompensation(aratherindirectmeasureoftradingsuccess)correlated positivelywithexperienceandrank(perhapsnotsurprisingly)andnegatively withpsychologicalmeasuresofemotionality,mostofwhomarerelevantfor generaleverydayproblemsolvingratherthanbeingspecifictofinancialmarket trading[fenton O'Creevy,Nicholson,SoaneandWillman(2005)]. Ourfindingsshouldalsohelpimprovevisualrepresentationoforderandtrade flow.sincehumansoftenarebestinrecognizingthenatureofintentionin moving(animateorinanimate)objects[castellicastelli,happe,frithandfrith (2000),HeiderandSimmel(1944)],wesuggestthattradersmaybemorelikely tosuccessfullydetectinsidertradingwhenorderandtradeflowsarepresented inamovingdisplay,asopposedtothepurelynumericallistingscommonlyfound intheindustry.infact,onemaywonderwhetherthesuccessofour(untrained) subjectsinpredictingpricechangesinmarketswithinsidersmaybeattributable 17

18 toourusingagraphicalinterfacewhereorderandtradeflowaretranslatedinto movementofcirclesofvarioussizesandcolors. Inthesamevain,onemayconjecturethatToMisbehindthesuccessof continuousdoubleauctionsoverone shotcallclearingsystemsinfacilitating equilibration[plottandvernon(1978)].intheformermechanism,the continuousflowofordersmayinducetom,thusenhancingtrustthatthereisno insideinformation.inacallmarket,absentorderflowinformation,assessment ofthesituationcannotrelyontomandhencemustinvoketraders abilitiesto maketherightconjecturesaboutendowments,preferencesandinformationof others.thepracticalsuperiorityofthecontinuousauctionsystemoverthecall marketsextendstosituationswherethereareinsiders[plottandsunder (1988)],evenifthelattermayintheorybebetter[Madhavan(1992)]. 18

19 Appendix In this appendix, we describe in detail the three experiments of this paper. We used college students or college educated subjects for every experiment. Both CaltechandUCLAethicsboardsapprovedtheexperiments. A.1.PriorMarketsExperiment We collected trading data with a markets experiment. We asked 20 subjects to tradewiththehelpofananonymouscomputerizedtradingsystem.theytraded for13independentsessions.attheendoftheexperiment,wepaidthesubjects withcashaccordingtotheirperformance. Subjectsownedtwotypesofstocks.Theycouldbuyandsellthefirsttype( stock X )forapricebetween0 and50.attheendofeachsession,thisstockpaidan amount of money that we drew randomly between 0 and 50 (uniformly distributed). We called this amount of money dividend. The subjects learned thevalueofthedividendonlyattheendofeachsession.subjectscouldnottrade thesecondtypeofstock( stockz ).Itpaidadividendthatwascomplementary tothedividendofstockx(thetwodividendsaddedupto50 ).Forexample,if werevealedattheendofasessionthatthedividendofstockxwas38,thenthe dividendofstockzwas12. For each session, the payment was determined as followed. Before trading started, we endowed each subject with a different mix of stock X, stock Z, and cash.duringtrading,subjectscouldexchangestockxforcashastheysawfit.as aresult,attheendoftrading,asubjectgenerallyownedadifferentcombination ofstockx,stockzandcash.aftermarketsclosed,werevealedthevalueofthe dividend, and subjects learned the amount of their payoff for the session. For example,ifatraderowned10unitsofstockx,5unitsofstockz,andhad87 in cash,whenwerevealedthatthedividendforstockxwas38,thetraderlearned thatherpayoffwas (50 38 )+87 =$5.27.Weadded$5.27tothe trader searnings,andthenwerestartedanewtradingsessionindependently. 19

20 Sincethesubjectsdidnotknowthedividenduntilthetradingsessionwasover, the expected payoff of one unit of a stock was 25. In addition, risk averse subjects could neutralize risk by balancing their holdings of the (complementary) stocks X and Z. Since the total endowment of stock X and Z acrossallsubjectswasthesame,everyonecouldintheorybalanceholdings.asa result of this absence of aggregate risk, risk neutral pricing should obtain. Consequently, without insider information, the predicted equilibrium price is 25. Weexplainedindetailthepayoffschedulebutdidnottellthesubjectsthatthe equilibrium price of the stock was 25. During the trading, the price moved as predicted by the theory, confirming earlier experiments [Bossaerts, Plott and Zame (2007)]. Essentially, when the price was below 25, subjects tended to purchasethestock.indeed,ifastockthatpaysoffonaverage25 weretrading at, say, 17, a subject that would purchase it would earn on average 8. The purchase would push the price of the stock higher, until it reached the equilibrium,at25.similarly,traderswouldsellastocktradingat,say,33 until thepricewentdownto25. The setup described above constituted the two control sessions of our experiments. For the 11 other sessions (test sessions), we introduced an additionalfactor. For these test sessions, we split the traders into two randomly chosen groups: the insiders and the outsiders. At the beginning of each session, we gave additionalinformationtotheinsidersintheformofa signal. Thesignalwasa number chosen at random within 10 of the actual dividend of X (uniformly distributed).forexample,ifwetoldtheinsidersthatthesignalwasat14,they knewforsurethatattheendofthetradingsessionwewouldrevealadividend between4 and24.wedidnotgiveanyinformationtotheoutsidersexceptthe factthattherewereinsidersinthemarket. Theunevendistributionofinformationskewedthetrading.Atthebeginningof eachsession,thetwogroupsattacheddifferentvaluestothestocks.forexample, ifwetoldtheinsidersthatthesignalwas14,thentheywouldvaluestockxat 20

21 14. The outsiders did not have this additional information and could only estimate the value of stock X to be 25. Insiders and outsiders had to act carefully. Insiders could use their information to make additional profits. For example,ifthesignalwasat14,theycouldsellastocktoanoutsiderwhowas willing to pay 25. The insider would have received 25 for a stock that would havepaidatmost24,andonaverage14 (i.e.hemadeanaverageprofitof11 ). However, by selling, insiders would lower the price of the stock, and thus they wouldrevealtotheoutsiderstheirknowledgeofthesignal.theoutsiders,who knew that there was a signal but did not know its value, had to observe the marketcarefullyandattempttoinferthesignalfromthetradingactivity.ifthe outsiders were successful at estimating the signal from the trade, they would makeextraprofits.vice versa,theinsidershadtotradeasdiscreetlyaspossible toavoidrevealingtheirknowledgeofthesignaltotheoutsiders.theanonymity of the trading interface helped conceal their information. At the same time, however, insiders had to trade before the other insiders in order to make additionalprofits. The dynamics of such markets are extremely complicated. While researchers have attempted to model them[admati(1985), Grossman and Stiglitz(1980)], there is no exact description of how the traders (insiders or outsiders) do behave.thus,weusedfmritoimproveourunderstandingofsuchmarkets.we used the periods without insiders as control and the periods with insiders as test. A.2.StimulusSetforfMRIExperiment We used 18 new subjects and replayed the order and trade flow while we recordedbrainactivity.first,weexplainedtothesubjectshowwehaveacquired thedataandmadesurethattheyunderstoodtheexperimentbyadministeringa quiz. We also reminded the subjects that they were not going to trade in the market but that they would only observe the replay of a previously recorded market.still,theyhadaninterestinpayingattentionastheirpayoffsdepended 21

22 on what happened in the market (in a way that we describe below). We also instructedthemthattheterm insider didnotrefertoillegal insidertrading. Wedisplayedeachofthe13sessionsinadifferentrandomorderforeachsubject (Figure3andVideo1).Eachsessionbeganwitha blindbet: weaskedsubjects to choose between the stock X and the (complementary) stock Z. After the subjects had made a choice, we replayed the market activity for stock X, regardlessoftheirchoice,atdoublethespeed(2minutesand30secondsinstead of5minutes).finally,weinformedthesubjectsofhowmuchtheyearnedduring thatsession. Foreachsession,thepaymentwasdeterminedasfollows.Whensubjectsplaced ablindbet,weendowedthemwith10unitsofthestocktheychose.aftertheend ofthetradingreplay,wepaidthemaccordingtothedividendofthestockthey had chosen. For example, if a subject had placed a bet on stock X before the tradingreplayandifwerevealedafterthetradingreplaythatthedividendwas 45,weadded10 45 =$4.50tothesubject searnings.ifthesubjecthadchosen to bet on stock Z, he would have received 10 (50 45 )=$0.50. Thus, the subjectsforthefmriexperimentsplayedtheroleofoutsiderswhodidnottrade. Theywereoutsidersbecausetheydidnotknowthesignalandtheydidnottrade becausetheyonlysawareplayoftradesthathadoccurredinthepast. Thestockpricewasanindicatoroftheexpectedrewardforthesubjectsonlyin the case of a session with insiders. For example, an increasing price indicated thattheinsidersmayhaveskewedthemarketbecausethesignalwashigh.thus, if a subject had placed a blind bet on stock X, she could have expected her earningsfortheperiodtobehigh.ifshehadplacedablindbetonstockz,she could have expected her earnings for the period to be low. If there were no insiders,thestockpricedidnotcontainanyinformationontheexpectedearning for the period. We illustrated this computation of the expected reward on the thirdrowoffigure4. This design had three main advantages. First, the subjects did not trade. While thequestionofhowthehumanbrainmakesfinancialdecisionsisinteresting,we needed first to understand how humans perceive a stock market. By not 22

23 introducing decision making, we avoided a confounding factor. Second, the periodswithoutinsiderswereperfectcontrols.sincethetradingdataacquisition method,thedisplayscreens,andthenumberoftraderswerethesame,thetwo types of sessions were identical in every respect except for the presence or absence of insiders. Third, by adding a blind bet, we elicited a feeling of randomness. Indeed, if we had forced subjects to choose stock X for every session, the payoff would have been the same fixed number for every subject. Moreover, we could not have teased apart an increase in stock price with a higher expected reward, as these two signals would have been perfectly correlated. Instead, by introducing a blind bet, we orthogonalized expected rewardandstockprice. We replayed the market activity (section (iii) of Figure 3) with a simplified interface(video1).werepresentedthepricelevelsfortheofferstobuyandsell ( bids and asks )withacircle.anumberinsideeachcircleindicatedtheprice incentsandthediameterofthecirclerepresentedthenumberofunitsoffered. Blue circles represented the bids and red circles represented the asks. When a trade occurred at a certain price, we turned the corresponding circle green for 500ms. We rearranged the circles dynamically to reflect the changes in prices andtrades.thelocomotionbetweensessionswithandwithoutinsidersdidnot displayanyobviousdifferences.finally,inordertomonitorattention,weasked subjectstopressakeyeverytimeatradeoccurred. Despitethehighlysalientandvividnatureofourgraphicaldisplayoftheorder andtradeflow,itshouldbeunderscoredthatitisnotwhatcausedthetombrain activation that we report. This is because we contrast brain activation in a treatment with and without insiders using the same interface. Likewise, the human factor behind the market in itself cannot be the cause of the ToM brain activations,becausethereisanequalamountofhumaninteractioninthecontrol treatment. A.3.Constructionofpredictors 23

24 ToanalyzethefMRIdata,weconstructedfourpredictors(Figure4).Wewanted tousethesessionswithoutinsiders(secondandfourthcolumnsoffigure4)as controlsandcomparethemwiththesessionswithinsiders(othercolumns).we constructedtwoblockpredictors(lasttworows)andtwoparametricpredictors (fourthandfifthrows). As described above, we could compute the expected reward of one session by takingintoaccountthestockprice,thepresenceorabsenceofinsiders,andthe blind bet (second row of Figure 4). Indeed, during sessions without insiders, neithertheblindbetnorstockpricescarriedinformation;theexpectedreward wasthemeanvalueofthedividend,namely25.ifinsiderswerepresent,thena higherstockpriceforstockxindicatedthatthedividendwaslikelytobehigher. Thus,ifasubjecthadplacedablindbetonstockX,shewouldhaveexpecteda higher return. Conversely, if she had placed a bet on stock Z, she would have expectedalowerreturn. We quantified the insiders activity using the absolute value of the difference between the trading price and 25 (Figure 4, third row). Indeed the more insiders skewed the market, the further the price would go from 25. We computed this value and built two parametric predictors. One parametric predictor modeled this effect during sessions with insiders (fourth row). To control for other effects, we built a similar predictor for the sessions without insiders(fifthrow).bycontrastingthesetwoparametricpredictors,weisolated theeffectofinsidersonsubjects perceptionofthemarket.wealsocreatedtwo block predictors. They captured the mean brain activation during the sessions withorwithoutinsiders. Thismodelinghadthreeadvantages.First,asmentionedabove,westrippedout the effect of decision making in the brain and focused on the perception of the market.second,thepredictorswebuiltwereorthogonaltoexpectedrewardas weusedtheabsolutevalueofthedistanceto25 andratherthanthedistanceto 25.Similarly,ourpredictorswereorthogonaltorisk,measuredbythebid ask spread [Glosten and Milgrom (1985)]. Third, by contrasting sessions with and withoutinsidersandmakingsurethatthesetwotypesofsessionsdisplayedno 24

25 obvious differences in trading activity, we controlled for differences in visual activationaswell. A.4.BehavioralExperiment This experiment is necessary to confirm the fmri results because ToM related brainareasmayactivateduringnon ToMsituations. Theexperimentconsistedoffourtasksthatweadministeredinarandomorder to43newsubjects.duringthefirsttask,wereplayedthetradingactivityforfour sessions with insiders. We used the same interface as for the fmri experiment except that we played the market at the original speed (5 minutes) and we paused the replay of the market every 5s. During these pauses, we alternated predictionsandoutcomes.specifically,forhalfofthepauses,weaskedsubjects to predict whether the next trade was going to occur at a higher, lower, or identicalpriceastheprevioustradethathadhappened.fortheotherhalfofthe pauses, we reminded subjects of their prediction and informed them of the outcome.werewardedsubjectsforcorrectpredictionsbutdidnotpunishthem for incorrect ones. However, we only gave subjects 5 seconds to make a predictionandpenalizedthemforindecision.thus,guessingwasalwaysbetter thannotanswering. ThesecondtaskwasaToMtest.Usingacomputerizedinterface,werecordedthe scores on an eye test[baron Cohen, Jolliffe, Mortimore and Robertson(1997)]. While several advanced tests were available, such as the faux pas test [Stone, Baron CohenandKnight(1998)],wedecidedtousethisparticularonebecause itprovidedacontinuousscaleoftomabilitiesandwasdifficultenoughtoreveal differencesinabilitiesbetweenhealthyadults.inaddition,thefaux pastestmay not be appropriate as social codes are culturally dependent. We also timeconstrainedtheeyetest,andguessingwasbetterthannotanswering. ThethirdtaskwasalsoaToMtestbasedondisplaysofmovinggeometricshapes [Heider and Simmel (1944)]. We played two Heider movies and paused the replayevery5s.forhalfofthepauses,weaskedsubjectstopredictwhethertwo 25

26 oftheshapeswouldgetcloser,farther,orstayatthesamedistance.fortheother halfofthepauses,wereportedtheoutcomesandthesubjects predictions.the better the subjects understood the social interactions of the shapes, the better theywereatpredictingmovements. Weusedthefourthtasktomeasuremathematicalabilities.Withacomputerized interface, we asked subjects to solve seven mathematical puzzles (Table III) under time constraints and without the use of paper. Subjects typed their answers on the keyboard, and guessing was better than not answering. 26

27 Figure1.Top:Object(magneticblockatright)attractedtotarget(ballatleft),flowingoverwall in the middle; Middle: wall is removed and block moves straight to target, as expected under physicallaws;bottom:samesituation,butnowblockcontinuestomakeacurveevenifattracted totarget,suggestingintention(tocurve).(adaptedfrom:ullerandnichols(2000).) 27

28 Figure2.EvolutionoftransactionpricesinthemarketofstockX.Sessionsaredelineatedwith solid vertical separators; dotted vertical separators indicated end of trading. Sessions with insiders are marked ins. Signal levels are indicated with green horizontal line segments; red line segments denote final dividend (of X) for the session. i denotes number of insiders (all insiders receive the same signal); K# indicates whether all subjects ( All ) know how many insiderstherewere,ornone( 0 ),oronlytheinsiders( Ins ).Allsubjectsalwaysknewwhether therewereinsiders. 28

29 (i) choice (iii)replay (v) outcome There are insider s Choose X or Z? The stockyou picked paid $0.45 blank (ii) blank (iv) blank blank (ITI) 10s 10s 2 min 30s 10s 3s 13s repeated 13 times Figure3.fMRIexperiment.(i)Atthestartofeachsession,subjectswereinformedofwhetheror not the session contained insiders and were instructed to make a choice(a blind bet) between stockxandstockz.theywerethenendowedwith10unitsofthestocktheychose.(ii)ablank screen was then presented for 10 seconds, (iii) a market session was then replayed at double speed,(iv)followedbya10secondblankscreenafterwhich(v)subjectsareinformedoftheir paymentforthatsession(10xtheirchosenstock sdividend). 29

30 Figure 4. Construction of predictors in the GLM used in the analysis of brain activation data. Columns 1 5 represent five fictive periods of different combinations of presence/absence of insidersandwhetherthesubjectchosestockxorz.a)theevolutionofthechosenstockpriceis displayed.b)thefirstpredictorwastheexpectedreward(er),computedfromthestockprice, thepresence/absenceofinsiders,andtheblindbet.duringsessionswithoutinsiders(2 nd and4 th sessions), neither the blind bet nor the stock prices carried any information and the expected reward was the mean value of the dividend, namely 25. During sessions with insiders (remainingsessions)ahigherpriceforstockxwastakentoindicatethatthedividendwaslikely tobehigher,resultinginahigherexpectedrewardinthecaseofablindbetonstockxanda lower expected reward on stock Z. C) Insider activity was proxied by the absolute value of the differencebetweenthetradingpriceand25.fromthisproxy,twoparametricpredictorswere constructed.d)first,aparametricpredictormodeledthiseffectduringsessionwithinsiders.e) Second,asimilarpredictorwasconstructedforsessionswithoutinsiders.Inaddition,weadded thefollowingpredictors:f)blockpredictorcapturingmeanbrainactivationduringsessionswith insiders;g)blockpredictorcapturingmeanbrainactivationduringsessionswithoutinsiders. 30

31 Figure5.Locationofsignificantcontrastbetweenslopecoefficientstotheparametricregressor betweeninsiderandno insidersessions(p<0.001,uncorrected,randomeffects).(a)sagittalview oftheactivationinparacingulatecortex(talairachcoordinates 9;41;36;Brodmannareas9/32; extends for 22 voxels). (b) Amygdala activation in coronal view ( 14; 23; 39; extends for 5 voxels).(c)activationoftheleftinsulainaxialview( 30; 7;11;extendsfor5voxels). 31

32 Figure 6. Location of significant contrast between slope coefficients to the block regressor between insider and no insider sessions. We use the threshold p<0.001 (uncorrected, random effect).weobservetwoclustersofvoxelsactivated:thelingualgyrusandcerebellum. 32