EUROPEANCENTRALBANK

EUROSYSTEM

WorkingPaperSeries

StefanoBorgioli,GiampieroM.Gallo,ChiaraOngari

Financialreturns,sentimentandmarketvolatility.Adynamic

assessment.

No2999

Disclaimer:ThispapershouldnotbereportedasrepresentingtheviewsoftheEuropeanCentralBank(ECB).TheviewsexpressedarethoseoftheauthorsanddonotnecessarilyreflectthoseoftheECB.

ECBWorkingPaperSeriesNo29991

Abstract

In1936,JohnMaynardKeynesproposedthatemotionsandinstinctsarepivotalin

decision-making,particularlyforinvestors.Bothpositiveandnegativemoodscaninfluencejudgmentsanddecisions,extendingtoeconomicandfinancialchoices.Intuitions,emotionalstates,andbiasessignificantlyshapehowpeoplethinkandact.Measuringmoodorsen-timentischallenging,butsurveysanddatacollectionmethods,suchasconfidenceindicesandconsensusforecasts,offersomesolutions.Recently,theavailabilityofwebdata,includ-ingsearchenginequeriesandsocialmediaactivity,hasprovidedhigh-frequencysentimentmeasures.Forexample,theItalianNationalStatisticalInstitute’sSocialMoodonEconomyIndex(SMEI)usesTwitterdatatoassesseconomicsentimentinItaly.Therelationshipbe-tweenSMEIandfinancialmarketactivity,specificallytheFTSEMIBindexanditsvolatility,isexaminedusingatrivariateVectorAutoregressivemodel,takingintoaccounttheimpactoftheCOVID-19pandemic.

Keywords:VAR,GrangerCausality,sentimentanalysis,financialmarket,forecastingJELCodes:C1,C32,C53,G4

ECBWorkingPaperSeriesNo29992

1Non-technicalsummary

Bothpositiveandnegativeemotionscanimpactjudgmentsandeconomicdecisions.Althoughmeasuringmoodorsentimentischallenging,toolslikesurveysandconfidenceindiceshelp.Infactonlinedata,suchassearchenginequeriesandsocialmediaactivity,providemorefrequentsentimentmeasurements.Forinstance,theItalianNationalStatisticalInstitutecreatedtheSocialMoodonEconomyIndex(SMEI)usingTwitterdatatogaugeeconomicsentimentinItaly.ThisstudyexaminestherelationshipbetweentheSMEIandfinancialmarketactivity,particularlytheFTSEMIBindex,usingamodelthatconsidersalsopossibleeffectsoftheCOVID-19pandemic.

Twomaindatasourcesareused:theSMEI,whichmeasuresdailyItalianeconomicsentimentthroughtweets,andtheFTSEMIB,whichrepresentstheperformanceof40majorItalianstocksandincludesvolatilitydata.TheanalysisusesVectorAuto-regressiveModels(VAR)tostudytherelationshipsbetweentheSMEI,FTSEMIBreturns,andvolatilityfromFebruary10,2016,toMarch8,2020.Grangercausalitytestsarethenconductedtodetermineifpastvaluesofonevariablecanpredictcurrentvaluesofanother,revealingpotentialbidirectionalinfluences.

“Evenapartfromtheinstabilityduetospeculation,thereistheinstabilityduetothecharac-teristicofhumannaturethatalargeproportionofourpositiveactivitiesdependonspontaneousoptimismratherthanonamathematicalexpectation,whethermoralorhedonisticoreconomic.Most,probably,ofourdecisionstodosomethingpositive,thefullconsequencesofwhichwillbedrawnoutovermanydaystocome,canonlybetakenasaresultofanimalspirits–ofaspontaneousurgetoactionratherthaninaction,andnotastheoutcomeofaweightedaverageofquantitativebenefitsmultipliedbyquantitativeprobabilities.”

JohnMaynardKeynes,GeneralTheoryOfEmployment,InterestAndMoney.

2Introduction

JohnMaynardKeynesintroducedasearlyas1936theideathatemotionsandinstincts(the“animalspirits”)ratherthanmererationalanalysis,playacrucialroleindecisionmaking,particularlyamonginvestors.

ECBWorkingPaperSeriesNo29993

Moods,betheynegativeorpositive,affectjudgmentanddecision-making,evenwhenpromptedbyunrelatedevents.Thisappliestoeconomicandfinancialdecisions,aswell:

Kahneman

(2011)

hasbecomethereferenceofchoiceonhowintuitions,emotionalstatus,andbiasesshapejudg-mentsandaffectthinking,behavior,anddecisions.Therefore,informationisnottheonlyfactoratplay,andrationalityisnotalwaysthemainenginebehindactions.

Ifthetheoreticalreasoningisclear,thewaymoodorsentimentaremeasuredissomewhatofachallenge;intheattempttoisolateanaggregateresult,severalsuggestionsarepresentedintheliterature.Oneoptionistoconductsurveysandcollectdataabouthoweconomicagentsjudgetheevolutionofthegeneraleconomicconditionsortheirown.Thisisthecase,forexample,ofthevariousconfidenceindices(bothbusinessandconsumer)releasedmonthlybymostcentralstatisticaloffices;oralsoofthevariousconsensusforecastexercisesconductedpollingseveralresearchinstitutesdeliveringmedianforecasts,thespreadofopinions,andthechangesrelativetothepreviousreleaseoftheexercise(cf.

Galloetal.,

2002,onthewayparticipantsdynamically

influence,andpossiblybias,eachother).

Morerecently,thewidespreadavailabilityofinformationonthewebhasspurredahostofindicatorsderivedfromsearchenginesearches:

D’AmuriandMarcucci

(2017)isanexample

usingGooglesearchestobuildaleadingindicatoronunemployment(awaytoprojectsentimentaboutjobsecurity).Moreover,thediffusionofsocialforumshasfosteredathrivinglineofresearchbasedontextualanalysisofthecontentofopinionsshared,andreactionstoeconomicormarketnews.Thisisconducivetoexaminingboththelevel(akintoconfidence)andthechange(moodswings),andithastheoverwhelmingadvantageofbeingavailableatthedailylevel,whichismostrelevantwhenanalyzingfinancialdata.

Twitter(nowX)seemstobethenaturaloutletforthisexpressionofsentiments,helpedbyalargenumberofsinglemessages,thepossibilityofreplyingtooneanother,andofclassifyingthecontentbytheuseof“tags”.

Angelicoetal.

(2022)documentthetimelinessandaccuracy

ofderivingameasureofinflationexpectationsfromamassiveamountof“tweets”releasedinItaly(initially,millionsthatboildowntoseveralhundredsofthousandsafterprocessingandcleaning).Ingeneral,theultra-highfrequencynatureofthistextualdataoffersaveryrichpoolofinformationtobeextracted(withtheawarenessthatallsortsofmanipulationarepossibleonthatforumindirectingopinions).

StartingfromOctober2018,theItalianNationalStatisticalInstitute(ISTAT)ispublishingahigh-frequencyindexcomputedinreal-timefromItalianTwitter’spublicstreamdata,the

ECBWorkingPaperSeriesNo29994

“SocialMoodonEconomyIndex”(SMEI),providingadailymeasureofthesentimentabouttheItalianeconomy.Theindexiscalculatedonanaverageof26,000tweetsperday.ThisexperimentalstatisticisupdatedquarterlywiththetimeseriesstartinginFebruary2016.

Financialmarketactivityisinterpretedasanexpressionofbeliefsandsentimentsinpro-ducingequilibriumpricesandreturnsfromtrading;bythesametoken,marketvolatility(i.e.thevariabilityofreturns)canbeseenasinverselyrelatedtotheconsensusonhowinformationreachingthemarketpointstotheevolutionofthemarketitself.Typically,adownturninthemarketisareactiontobadnewsandischaracterizedbyhighvolatility.Sincetheearly1990s,amarket-basedmeasureofvolatilityextractedfromtheimpliedvolatilitiesofputandcallop-tions(atthemoney-30daystoexpiration)onamarketindexcametobeknownasthe“fearandgreedindex”(theVIXisderivedfromtheS&P500,

Whaley,

1993,butotheroption-based

volatilityindicesareavailable).

Financialinvestment,beingdrivenbyprofitincentives,isaninterestingfieldinwhichitispossibletoanalyzethepropertiesoftheSMEI,thatis,itscapabilitytorepresentarelevantfactorinteractingwithfinancialvariablesaboutthestockmarketactivityinItaly.Thelatterisrepresentedbytwovariables:thereturnsonanaggregateindex,theMilanoIndicediBorsa(FTSEMIB–thebenchmarkstockmarketindexfortheBorsaItaliana,madeupofthe40most-tradedstocks),anditsvolatility(representedbyarange-basedvolatilitymeasure,

Garmanand

Klass,

1980)

.

Afteradiscussionoftheexistingliterature(Section

3),wediscussthefeaturesofthevari

-ablesused(Section

4),documenting,inparticular,thecontentofavailablevolatilitymeasures

andtheSMEI(arelativelynovelindex).TherelationshipbetweentheSMEIandthemarketbehaviorrepresentedbytheFTSE-MIBisdiscussedinSection

5.

Wesuggest(Section

6)asim

-pletrivariateVectorAutoregressivemodelbetweenthreevariables(marketreturnsontheindex,thevolatilityandtheSMEI)toinvestigatewhichvariablesarein-samplerelevanttoincreasetheforecastingcapability(asimpleGranger-causalitytestonaugmentingthebenchmarkunivariateARmodel).Wesplittheanalysisbetweenin-sample(Subsection

6.1),andout-of-sample,where

weperformaDiebold-Marianotest(DieboldandMariano,

1995)toassesswhentheVARhas

asuperiorperformancethantheARandforwhichvariables.SomeconsiderationoftheimpactoftheCOVID-19pandemicontheserelationshipsisinordersinceourin-sampleperiodendswiththewideoutbreakofthevirusinMarch2020.

Themainfindingscanbesummarizedasfollows.TheresultsindicatethatbeforeCOVID-19,

ECBWorkingPaperSeriesNo29995

marketvolatilitywastheonlyvariablesignificantlyinfluencedbypastvaluesofothervariables.However,duringthepandemic,therelationshipsshifted:pastvolatilityinfluencedbothSMEIandreturns,pastreturnsimpactedvolatility,andSMEIalsobegantoaffectvolatility,thoughlesssignificantly.Thissuggeststhatthepandemicsignificantlyalteredthedynamicsamongthesevariables.ThestudyusesthentheDiebold-Marianotesttocomparethepredictiveabil-itiesofaunivariateautoregression(AR)modelandaVARmodelinanout-of-samplecontext.Byconductingrollingregressionsandgeneratingforecasts,theresultsshowthattheonlyvari-ableforwhichtheVARispredictivelysuperiortotheARmodelistherange-basedvolatility,indicatingthatbothlaggedSMEIandreturnsarevaluableinformationforforecastingmarketactivityturbulence.Movingforward,theISTATSMEIindex,trackingsocialmoodfromshortmessagesonsocialplatformslikeX,isavaluabletoolforunderstandingmarketdynamics,es-peciallyduringunexpectedevents.Thepapersuggestsalsofurtherresearchintohowsentimentindicesrelatetomarketreturnsandvolatilityandhighlightsthepotentialuseofothermarketactivitymeasures,likeaVIX-typevolatilityindex,toenhanceanalysis.Additionally,theuniqueavailabilityofSMEIdataonweekendsandholidaysraisesquestionsaboutitsimpactonmarketactivityatthestartofthetradingweek,whichmaybeworthbeingaddressed.

3Earliercontributionsandissues

JohnMaynardKeynes,asearlyas1936,introducedtheideathatemotionsandinstincts,thefamed“animalspirits”,mayplayacrucialroleindecision-making,especiallyamonginvestors.Morerecently,studieshaveexploredthepotentialconnectionsbetweenpublicsentimentindicesandeconomicandfinancialvariables.

Astreamofpapersindeedfoundthatsentimentandstockmarketdynamicscanbehighlycausalrelated.Forinstance,

BrownandCliff

(2004)examinetherelationshipbetween“in

-vestorsentiment”andstockmarketreturns.Theyfirstbuildasentimentmeasurestartingfromsurveydataoninvestorsentiment(likebullishinvestorexpectationsofabove-averagereturns)andusingKalmanfilterandprincipalcomponentanalysisasmeansofextractingcompositeunobservedsentimentmeasures.Theythenexplorethebi-directionalrelationbetweentheseinvestorsentimentmeasuresandthenear-termstockreturnsinavectorautoregression(VAR)framework.Theyfindthatchangesinthecompositemeasuresofinvestorsentimentarehighlycorrelatedwithcontemporaneousmarketreturns,butthiscorrelationdoesnotdirectlyreveal

ECBWorkingPaperSeriesNo29996

thecausalrelationbetweensentimentandthemarket.ThentheVARanalysisrevealsthatmarketreturnsclearlycausefuturechangesinsentiment.However,verylittleevidencesuggestssentimentcausessubsequentmarketreturns.

Similarresultsaredisplayedin

Wangetal.

(2006),whotestwhethersentimentisusefulfor

volatilityforecastingpurposes.Infact,theyfindthatmostofthesentimentmeasurestheyusearecausedbyreturnsandvolatilityratherthanviceversa.Inaddition,theyfindthatlaggedreturnscausevolatility.Finally,allsentimentvariableshaveextremelylimitedforecastingpoweroncereturnsareincludedasaforecastingvariable.

Tetlock

(2007)exploresinsteadhowmedia

contentinfluencesinvestorsentimentand,consequently,stockmarketmovementsbyusingdailycontentfromapopularWallStreetJournalcolumn.Hefoundthathighmediapessimismpredictsdownwardpressureonmarketpricesfollowedbyareversiontofundamentals,andunusuallyhighorlowpessimismpredictshighmarkettradingvolume.

Inthesamevein,

GilbertandKarahalios

(2010)showhowestimatingemotionsfromweblogs

providesnovelinformationaboutfuturestockmarketprices.Fromadatasetofover20millionLiveJournalposts,theyconstructametricofanxiety,worryandfearcalledtheAnxietyIndex.UsingthenaGrangercausalityframework,theyfindthatincreasesinexpressionsofanxietypredictdownwardpressureontheS&P500index.ThesefindingsarethenconfirmedviaMonteCarlosimulationsandshowhowthemoodofmillionsinalargeonlinecommunity,evenonethatprimarilydiscussesdailylife,cananticipatechangesinaseeminglyunrelatedsystem.

Zhang

etal.

(2011)explorehowsentimentandactivityonTwittercanbeleveragedtounderstand

investorbehaviorandpredictstockmarketindicatorsfromDowJones,NASDAQ,andS&P500.Startingfromarandomizedsampleoftweetstheymeasureddaily“collectivehopeandfear”andanalyzedthenthecorrelationbetweentheseindicesandthestockmarketindicators.TheanalysisshowsthatTwittersentimentissignificantlycorrelatedwithstockmarketmove-ments.PositivesentimentonTwitterisoftenassociatedwithrisingstockprices,whilenegativesentimentcorrelateswithdecliningprices.ThevolumeofTwitteractivityisalsofoundtobeausefulpredictor,withhighertweetvolumesindicatingincreasedmarketattentionandpotentialvolatility.Also

Bollenetal.

(2011)startfromTwittertoinvestigatewhetherpublicmoodstates

derivedfromfeedsarecorrelatedtothevalueoftheDowJonesIndustrialAverage(DJIA)overtime.Theyanalyzedover9.8milliontweetsfrom2.7millionusersoversixmonthstoassessthesentimentofeachtweetaseitherpositiveornegative.Then,aGrangercausalityanalysisandaSelf-OrganizingFuzzyNeuralNetworkareusedtoinvestigatethehypothesisthatpublicmood

ECBWorkingPaperSeriesNo29997

statesarepredictiveofchangesintheDJIAclosingvalues.Theauthorsfindthattheinclu-sionofcertainpublicmooddimensionsindeedimprovestheaccuracyofstandardstockmarketpredictionmodels.Twitterdataareusedalsoby

RaoandSrivastava

(2012)toinvestigatehow

sentimentanalysiscanbeemployedtopredictstockmarketmovements.Thestudy,basedonmorethan4milliontweetsbetweenJune2010toJuly2011,findsthatthereisasignificantcorrelationbetweenTwittersentimentanddiscussionsandstockmarketmovements.Positive

sentimentisoftenassociatedwithrisingstockprices,whilenegativesentimentcorrelateswithfallingprices.Thevolumeoftweetsisalsofoundtobeausefulpredictor,withhighervolumesindicatinggreatermarketattentionandpotentialvolatility.

Daetal.

(2015)insteadavailof

dailyInternetsearchvolumefrommillionsofhouseholdstoinvestigatetherelationshipbetweeninvestormarket-levelandassetprices,particularlyhowfear-basedsentimentimpactsmarketdynamics.ByaggregatingthevolumeofqueriesrelatedtohouseholdconcernstheyconstructaFinancialandEconomicAttitudesRevealedbySearch(FEARS)indexasanewmeasureofinvestorsentiment.TheFEARSindexwasthenfoundtohavesignificantpredictivepowerregardingfuturemarketreturns.Infact,between2004and2011,theyfoundthatFEARS(i)predictshort-termreturnreversals,(ii)predicttemporaryincreasesinvolatility,and(iii)predictmutualfundflowsoutofequityfundsandintobondfunds.Adifferentapproachisappliedby

AggarwalandMohanty

(2018)whomakeuseofprincipalcomponentanalysis(PCA)tobuild

sentimentindexasaproxyforIndianstockmarketsentimentsoveratimeframefromApril1996toJanuary2017.Threetypesofvariablesenterthecalculationoftheindex:indirectmarketmeasures(mostlyindicatorslikeforinstancepricetoearningratios,dividendyieldsorpricetobookratios)andIndianandUSmacrovariables.TheindexisthenusedtoestimateviaOLSregressionstheimpactofIndianinvestorsentimentsoncontemporaneousstockreturnsofBombayStockExchange,NationalStockExchangeandvarioussectoralindices.Thestudyfindsthatthereisasignificantpositivecorrelationbetweenthesentimentindexandstockindexreturns.

Chenetal.

(2019)investigatewhethersentimentanalysisofsocialmediapostscould

beusedtopredictthedirectionofstockpricemovements.TheauthorsapplysevendifferenttechniquesofdataminingtopredictstockpricemovementofShanghaiCompositeIndexfortheperiodApril2016toMay2018.Thefindingssuggestthatsentimentanalysisofsocialmediapostscouldprovidevaluableinsightsintothepotentialdirectionofstockpricemovements;forinstancesentimentderivedfromEastmoney,asocialmediaplatformfortheChinesefinancialcommunity,furtherenhancesmodelperformances.

ECBWorkingPaperSeriesNo29998

Nymanetal.

(2021)investigatestheinfluenceofnewsandnarrativesinfinancialsystems,

especiallyinutilizingbigdataforevaluatingsystemicrisk.Thepaperexaminestheapplicationoftextualanalysistechniquesandbigdataanalyticstoextractvaluableinsightsfromnewsandnarrativesources,assistinginidentifyingandassessingsystemicriskswithinfinancialsystems.Theirresultshighlighthowourmeasuresofsentimentandnarrativeconsensuscorrelatewellwith,andinsomecasesevenappearto‘cause’,certaineconomicandfinancialvariables.Also

Huangetal.

(2019),usecomputationaltextanalysistechniquetoconstructsentimentindicesfor

20countriesfrom1980to2019.Theauthorsthenassesswhetherthesesentimentindicestriggerearlywarningindicators(EWIs)aheadoffinancialcrises.Foreachsentimentindex,anEWSistriggeredeachtimethereisaspike,i.e.whentheindexvalueisabove2standarddeviationsfromabackward-lookingaverageof24months.Theyfindthat,foreachcountryinoursample,atleastoneoftheindicatorswouldhavesuccessfullyanticipatedmostcrisesinawindowof24months.Asregardstechniquestoanalysetextdata,

LoughranandMcDonald

(2011)have

investigatedhowtextualanalysistechniquesareutilizedtointerpretfinancialreports.Theauthorsparticularlyfocusonalargesampleof10-KfilingssubmittedtotheSecuritiesandExchangeCommission(SEC)from1994to2008.Relevanttoourstudy,theyalsofindsignificantrelationsbetweenthesentimentmeasurestheybuildandeconomicandfinancialvariables;forinstance,theyfoundthatsomemeasuresaresignificantinregressionsestimatingabnormalsharestradingvolumes.

TurningtothestatisticalpropertiesoftheSMEIindex,

Righietal.

(2020)analyzethe

relationshipsofthismetricwithsomedailyandmonthlymacroeconomicindicatorscomingfromtraditionalandnon-traditionalsources.Theyuseseveralnon-traditionalsourcestoproducetimeseriestorelatetotheSME,suchasthedailynumberofCOVID-19deathsandnewpositivecasesreportedbytheCivilProtectionDepartmentormacroeconomicindicatorscomingfromTarget2andBI-COMPseries(onPOSandATMtransactions),butalsotheBankofItalyelectroniccardtransactionande-commercetransactionmonthlyseriesandtheconsumerconfidenceindicators.TheyfoundthatthemonthlyaverageofthedailyseriesoftheleveloftheSMEindexshowsalowcontemporaneouscorrelationandaweakpredictivepoweroftheSMEindexforthetraditionalmonthlyindicators.Ontheotherhand,theyobservedapositivecorrelationbetweentheSMEIandtheBI-COMPPOSdailytransactionseries.

ECBWorkingPaperSeriesNo29999

AuthorsPeriodMarkets-includedSentimentIndexMethodResults

GilbertandKara-

halios

(2010)

2008S&P500”AnxietyIndex”(anaggregate

measureofanxietyandworry)es-timatedstartingfromadatasetof

GrangercausalityandMonteCarlosimulations

Thepaperstatisticallyshowsthatabroadindexofmoodfromanonlinecommunityhasnovelpredictiveinformationaboutthestockmarket.

RaoandSrivastava

(2012)

Bollenetal.

(2011)

Zhangetal.

(2011)

LoughranandMc-

Donald

(2011)

Nymanetal.

(2021)

Tetlock

(2007)

AggarwalandMo-

hanty

(2018)

Daetal.

(2015)

Huangetal.

(2019)

Chenetal.

(2019)

BrownandCliff

(2004)

Wangetal.

(2006)

June2010-

July2011

February

2008-Decem-ber2008

March2009-September2009

1994–2008

January1984-September1999

April1996-January2017

2004-2011

1980to2019

April2017-

May2018

March1965-December1998

February1990-December2001

DJIAandNASDAQ-100

DJIA

DJIA,NASDAQ-100andS&P500

NASDAQ

DIJA

Bombaychange,

StockEx-National

StockExchange,var-iousIndiansectoralindices

S&P500,NASDAQ-100,Russell2000

ShanghaiCompositeIndex.

S&P500,Russell2000

S&P100

over20millionpostsmadeonthesocialnetworkingserviceLiveJour-nal.(dailyfrequency)

Twitterfeeds(dailyfrequency)

Twitterfeeds(dailyfrequency)

Twitterfeeds(dailyfrequency)

Sentimentmeasuresbuiltviatex-tualanalysisstartingfrom10-Kre-portforms(monthlyfrequency)

BoEdailymarketsreports;brokerresearchreports;Thomson-ReutersNewsarchive(dailyfrequency).

WallStreetJournal’s(WSJ’s)“AbreastoftheMarket”columnonU.S.stockmarketreturns(dailyfrequency).

PCAtobuildsentimentindex,thenOLS(monthlyfrequency).

Ameasureofinvestorsentiment(AttitudesRevealedbySearch(FEARS))builtbyaggregat-ingdailyInternetsearchesfrommillionsofhouseholds(dailyfre-quency)

NewsarticlesfromtheFinancialTimes(dailyfrequency).

SentimentderivedfromEastmoney,asocialmediaplatformforthefi-nancialcommunityinChina(dailyfrequency).

Twosurvey-basedinvestorsenti-mentmeasures.Indirectsentimentmeasuressuchasadvance-declineratio,shortinterest,andclosed-endfunddiscounts(weeklyfrequency).

S&P100(OEX)put–calltrad-ingvolumeratio(PCV);OEXput–callopeninterestratio(PCO);NYSEARMSindex.Twosurvey-basedsentimentratiosprovidedbyinvestmentinformationproviders.(dailyfrequency)

causalityMining

ExpertSystem

GrangerModel

(EMMS)

GrangercausalityandSelf-OrganizingFuzzyNeuralNetwork

Correlationanalysis

Logitregression

Acombinationofnaturallanguageprocessing(NLP)andmachinelearningalgo-rithms.

VAREstimates

Acompositesentimentin-dexforIndianstockmar-ket,builtwithprincipalcomponentanalysis,start-ingfromindirectmarketmeasures(suchaspricetoearningsratio,pricetobookratio,dividendyield...)andmacrovariablesforIndianandUSmarkets.

RegressionAnalysis

variousmachinelearningmodels,suchaslogisticregression,supportvectormachines(SVM),andneuralnetworks

PrincipalComponent,Kalmanfilter,VAR

Grangercausality/regres-sions

Highcorrelation(upto0.88forreturns)betweenstockpricesandTwit-tersentiments

TheaccuracyofDJIApredictionscanbesignificantlyimprovedbytheinclusionofspecificpublicmooddimensions

Emotionaltweetpercentagesignificantlynegativelycorrelated