ContentslistsavailableatScienceDirect
Journal
of
International
Financial
Markets,
Institutions
&
Money
j o u r n a l h o m e p a g e :w w w . e l s e v i e r . c o m / l o c a t e / i n t f i n
Language,
news
and
volatility
夽
Hans
Byström
∗LundUniversity,DepartmentofEconomics,Box7082,22007Lund,Sweden
a
r
t
i
c
l
e
i
n
f
o
Articlehistory:Received16June2015
Accepted5March2016
Availableonline14March2016
Keywords: Newsaggregator Language Volatility Stockmarket Chinese
a
b
s
t
r
a
c
t
IuseGoogleNewstostudytherelationbetweennewsvolumesandstockmarketvolatilities. Morethanninemillionstockmarket-relatednewsstoriesinEnglishandChineseare col-lectedandthedynamicsofthenewsvolumeandthestockmarketvolatilityiscompared.I findthatthestockmarketvolatilityandthenumberofpubliclyavailableglobalnewsstories arestronglylinkedinbothlanguages.Furthermore,thedirectionallinkbetweennewsand volatilityratherisfromnewstovolatilitythanviceversa.Inout-of-sampleevaluationsof volatilityforecastsIfindnewsvolumestoimproveforecasts,regardlessoflanguage.
©2016TheAuthor.PublishedbyElsevierB.V.ThisisanopenaccessarticleundertheCC BY-NC-NDlicense(http://creativecommons.org/licenses/by-nc-nd/4.0/).
1. Introduction
InthispaperItrytoshedsomenewlightontheoldquestionsofwhethertheamountofpubliclyreportedstock market-relatednewsislinkedtothevolatilityinthestockmarketand,ifso,ifitisnewsthatcausesvolatilityorviceversa.Iuse GoogleNewsTM,thenewsaggregator,tocapturetheactualmonth-to-monthdynamicsoftheglobalnewsvolume.Tomake
thestudymoreinclusive,Ilookatnewswritteninthetwomostimportantgloballanguages,i.e.EnglishandChinese. Largemovementsinthestockmarketareoften(expost)explainedasthemarket’sreactiontosuddenimportantnews arrivals.Atothertimes,however,marketsmoveseeminglywithoutanyevidenceofimportantnewsarriving.Acomparison oftworecentso-called“flashcrashes”canbeusedtoexemplifythis.Whilethe“TwitterCrash”ofApril2013,whenS&P500 lost$120billioninmarketvalueinseconds,wascausedbyfaketweets(i.e.news)aboutexplosionsattheWhiteHouse,the “FlashCrash”ofMay2010,when$1trillioninmarketvaluetemporarilywaslost,isnormallynotconsideredtohavebeen causedbythearrivalofnews.Inotherwords,itisnotobviousthatpricemovementsalwaysarereactionstonewinformation (news)arrivinginthemarket.
Themaincontributionofthispaper,comparedtothetypicalstudylinkingnewsandvolatility,isitsuniqueproxyforthe totalamountof(global)stockmarketrelatednewsincirculation.Byusinganautomatedweb-basednewsaggregator,inmy caseGoogleNews,Iamablenotonlytocollectasignificantshareofallgloballyavailablemarket-drivingpublicinformation but,throughthecontinuousdatacollectionprocess,Iamalsoabletocapturetheactualmonth-to-monthdynamicsofthis newsdissemination.Thatis,insteadofmerelylookingatspecificnewsevents,Ilookatthedynamicsoftheoverallflowof publicinformation.Moreover,byfocusingonthebulkoftherelevantinformationflow(eachmonth,Icollectallavailable
夽 ThisworkwassupportedbyTheMarianneandMarcusWallenbergFoundation[MMW2012.0019];andHandelsbankensForskningsstiftelser
[P2013-0209:1].
∗ Tel.:+46462229478;fax:+46462224118.
E-mailaddress:[email protected] http://dx.doi.org/10.1016/j.intfin.2016.03.002
1042-4431/© 2016 The Author. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
newsstorieswhereagenericphrasesuchas“stockmarket”ismentioned)Iefficientlyavoidanyundueemphasizeonnews storiesthat,expost,turnouttohavehadasignificanteffectonthevolatilityinourparticularmarkets.
Intotal,Icollectmorethanninemillionstockmarket-relatednewsstoriespublishedbymajornewspapersandother newssourcesworldwideoveraneight-yearlongperiod.Toputthisamountofnewsintoperspectiveitcanbecompared tothe120,000Reuter’sNewsServicenewsreleasescollectedbyBerryandHowe(1994),the752,647WallStreetJournal andBroadtapestoryheadlinescollectedbyMitchellandMulherin(1994)andthe129,737DowJonesandReutersnews announcementscollectedbyJohnsonandMarietta-Westberg(2004).Furthermore,tohighlightthetrulyglobalnatureof bothnewsandstockmarketsIhavecollectednewsinbothofthetwomajorgloballanguages,i.e.EnglishandChinese;eight millionofthenewsstoriesareinEnglishandonemillionareinChinese(Mandarin).Ibelievethistobethefirsttimeanyone looksatnewswritteninChinese,i.e.,arguably,thesecond-mostimportantlanguageintheworld,inconnectiontomarket volatility.TheEnglish-Chineselanguage-pairisalsoparticularlyinterestinginthelightofthetwolanguages’significant semanticandlinguisticdifferencesandduetothefactthatfewstockmarketparticipantsactuallyreadnewsbothinEnglish andinChinese.
Stockreturnvolatilityvarieswidelyacrosstime.Italsotendstobepersistentandtoexhibitso-calledvolatilityclustering, whereperiodsofhighvolatilityarefollowedbyhighvolatilityandviceversa.Althoughitissomethingofastylizedfact thatnewinformationreachingthemarket,i.e.news,isthemaincontributorofthisvolatility,andalthoughseveralstudies havelookedintotherelationshipbetweenmarketvolatilityandnewsdissemination,theempiricalevidenceisnotasstrong asonewouldexpect.Infact,evenwhenalinkbetweennewsandmarketmovementsisfound,thestrengthofthelinkis oftenquestioned.OneofthefirststudiesonnewsandvolatilitywasFrenchandRoll(1986)whocomparesthevolatility intheUSstockmarketduringexchangetradinghoursandnon-tradinghoursandconcludesthatthedifferenceintheflow ofinformation,particularlyprivateinformation,explainsthedifferenceinvolatility.Inotherwords,theyconcludethatit isthevariabilityintheflowofprivateinformationthatexplainsmostofthevariabilityinvolatility.MitchellandMulherin (1994),inturn,looksatpublicnewsandfindsapositiveandstatisticallysignificant,albeitweak,relationshipbetweenthe variability(absolutereturn)intheUSstockmarketandthenumberofpublicnewsannouncements,measuredasthedaily numberofstoryheadlinesreportedbyDowJones&Company(WallStreetJournalandBroadtape).BerryandHowe(1994), ontheotherhand,doesnotfindthatpublicinformationisstatisticallyrelatedtostockvolatilityintheUSintradaystock market.BerryandHowe(1994)measurespublicinformationflowasthenumberofnewsreleasesbyReuter’sNewsServices. Theinterestrateandforeignexchangemarketsalsoexhibittime-varyingvolatilityandEderingtonandLee(1993)shows thattheimpactofscheduledmacroeconomicnewsannouncementshasanimmediateeffectonpricesandalonglasting effectonpricevolatility.Inthestockmarket,anevenlongerlastingperiodofelevatedvolatilityafterannouncementsis foundbyPatellandWolfson(1984).Inamorerecentstudy,JohnsonandMarietta-Westberg(2004)findsthatincreasesin idiosyncraticstockreturnvolatilityarepositivelyrelatedtoincreasesintheamountoffirm-specificpublicnews.Andin Byström(2009,2011)IusethesameGoogleNewsmethodologyofcollectingnewsasinthisstudybutacrossamuchshorter sampleandlimitingtheanalysistosimplecross-correlations.1TheresultsinByström(2009,2011),evenifmerelytentative, indicateapositivelinkbetweenstockmarketvolatilityandnewsvolumes.
Arelatedstrandofliteraturefocusesoninvestorattention,ratherthanonnewsdigestion.Here,anotherGoogleproduct, GoogleTrendsTM,hasbeenemployedrecently.Inthisliterature,Googlesearchfrequencies(SearchVolumeIndex(SVI))
areusedasaproxyforinvestorattention.DimpflandJank(2015)proxyinvestorattentionwithGooglesearchfrequencies andfindsastrongcorrelation,usingdailydata,betweenthesearchqueryvolumeandUSstockmarketvolatility.Theyalso showthatsearchqueriesimprovevolatilityforecasts.Similarly,Vozlyublennaia(2014)collectsGooglesearchfrequencies onaweeklybasisusingGoogleTrendsforarangeofdifferentfinancialmarkets,includingstockmarkets,butfindsarather weakrelationshipbetweeninvestorattention(Googlesearchfrequency)andvolatility.Finally,Daetal.(2011),although notexplicitlyfocusingonvolatility,findsapositiverelationshipbetweenabnormalabsolutereturnsintheUSstockmarket andtheGoogleSearchVolumeIndex.Likethecurrentpaper,allthesestudiesemploymodernweb-basedtools,butrather thancollectingnewsvolumestheycollectsearchfrequencies.
Moststudiesoninformationflowsandstockmarketreactionshavewrestledwithvariousdata-relatedissues.Some studieshavenotbeenabletodifferentiatebetweengood,badandneutralnewsandmanyhavenotbeenabletomeasure theimportanceofaparticularpieceofnews.Otherstudiesisolatespecificeventsandthereforelackacontinuousmeasure ofthenumberofavailablenewsstories.Thisistypicallyanissuewhenmacroeconomicnewsisstudied,andwithouta continuoussamplingoftheamountofnewsincirculation,thedynamicsofnewsvolumescannotbestudied.Ibelievethat severaloftheseissuesareavoidedinourresearchsetup.First,sinceGoogleNewsallowstheusertospecifyexactlywhich wordstringstocrawl,Iam,byconstruction,abletogroupnewsaccordingtosentiment(neutralnewsorbadnews).2The
1 Byström(2009,2011)looksatthelinkbetweennewsandvolatilityusingGoogleNewsbutotherwisethosestudiesdiffersignificantlyfromthisstudy. InthispaperIincludenewsinChinese(Mandarin)aswellasinEnglish,Ilooknotonlyatcontemporaneouscorrelationsbutfocusinsteadprimarily onregressionsbetweencurrentvolatilityandlaggednewsvolumes,Irunlead-lagregressionstotellinwhichdirectioninformationflows,Ilookatthe volatilityforecastingperformanceofnews,Ilookatchangesinadditiontolevels,Ilookatatime-periodthatisalmostthreetimesaslong,Iusemonthly non-overlappingnewsvolumeobservationsratherthandailyoverlappingones,IlookattwelvemajorstockindexesinboththeEnglish-languagedominated worldandintheChinese-languagedominatedpartoftheworldandinarobustnesssectionIinvestigatewhetherextremenewsobservations,thecrisis period,missingobservationsortheexactwordingofthenewssearchstringisdrivingtheresults.
wordstringsinthisparticularpaperarestockmarketandstockmarketcrashintheUSeditionand (stockmarket)and (approximatelytranslatedintoEnglishasstockmarketcollapseandstockmarketcrash)intheChina edition.3Second,byfocusingontheamountofnewsratherthanthemereexistenceornotofnewsIamautomaticallyable totellwhethertheunderlyingactionsoreventsthatshapethenewsareimportantornot.Third,throughthecontinuous datacollectionprocess,Iamabletocapturethetime-seriesdynamicsofthenewsincirculation.Finally,thecomprehensive natureoftheGoogleNewsgeneratednewsdatabaseandthesheernumberofnewsstoriesstrengthenstheresults.
SincethenewsthatIamcollectingusingGoogleNewsisglobalinnature,inthesensethatitiswrittenintwolanguages thattogetherarereadbyamajorityoftheworld’smarketparticipants,Ihavechosentolookatgloballyimportantstock marketsandthevolatilityinthemajorstockindexesinthesemarkets.Duetothedual-languagefocusoftheresearchIhave chosenhalfofthestockindexesfromtheEnglish-speakingworldandhalffromtheChinese-speakingworld.Theformer areMSCIWorld,S&P500,DJIA,Nasdaq,Russell2000andFTSE100andthelatterareShanghaiA,ShanghaiB,ShenzhenA,
ShenzhenB,HangSengChinaEnterprisesandHangSengChina-AffiliatedCorporations.Intotal,IlookattheimpactofGoogle News-generatedglobalnewsvolumesontwelvemajorstockindexes.
Tomyknowledge,thisisthefirsttimeanewsaggregatorisemployedintheforecastingoffinancialmarketvolatility and,furthermore,thefirsttimerelativelystrongevidenceisfoundthatnewsvolumescanactuallypredictstockmarket volatility.Inadditiontorobustandsignificantpositive contemporaneouscorrelationsbetweentheamountofnewsin circulationandthevolatilityinvariousmajorstockmarkets,lead-lagregressionstellusthatthedirectionallinkbetween newsandvolatilityratherisfromnewstovolatilitythanviceversa.Ialsofindevidenceofnewsvolumespredicting (one-monthahead)volatility.Thelatterfindingissupportedbysignificantandeconomicallyrelevantnewsvolumeregression parametersandeconomicallymeaningfulout-of-sampleforecastingerrorreductions.Theaverageimpactofaone-standard deviationchangeinnewsvolumeonnextmonth’sstockindexvolatilityis11basispoints(0.11%).Albeitnotlarge,the impactiseconomicallymeaningfulwhencomparedtothemeanofthetwelvestockindexes’unconditionalvolatilityacross thesampleperiod(coveringtheveryvolatilecredit-andeuro-areasovereigncrises)whichis123basispoints.Moreover, whenIlookatchanges,ratherthanlevels,theeconomicsignificanceiseven moresignificant(9basispointsversusa meanof44basis points).Theout-of-samplemeanabsoluteforecastingerror(MAE),inturn,isonaveragereduced by 15%whenlaggednewsvolumesareaddedtopastvolatilitieswhenpredicting(one-monthahead)volatility.Itshouldbe stressedthatnofine-tuningoftheforecastsaremade,theestimationwindowisforexamplesetratherarbitrarily,andfor somemarket/language/searchstringcombinationstheeconomicsignificanceofincludingnewsvolumesinthevolatility predictionismuchlarger.Inotherwords,thereisscopeforamoresubstantialforecastingimprovementwhenallowingfor systematicdatamining.IconcludetheempiricalstudywitharobustnesssectionwhereIfindtheresultstoberobustto theremovalofextremenewsvolumeobservations,creditcrisisobservationsandmissingobservationsaswellastoslight changestothenewsvolumecollectionprocess.
2. Newsvolumecollectionanddatadescription
Newsiswritteninhundredsofdifferentlanguages,somereadandunderstoodbyglobalaudiencesofmillionsofpeople, othersreadsolelybylocalsorcognoscenti.Atanypointintimetherearethousandsofnewspiecesavailabletomarket participantsandanyofthisnewsmayaffectthemarketinonewayoranother.Inthispaper,Ifocusonthevolatilityinthe majorstockmarketsandonhowthisvolatilityisrelatedtotheamountofnewsavailable.Ithereforefocusprimarilyonnews thatislikelytohaveanimpactonstockmarketvolatility.Ihavealsochosentofocusonthetwo,arguably,mostimportant globallanguages,EnglishandChinese.Englishisthelinguafrancaoftodaywithuptoabillionnativeand non-native speakers.Chinese,ontheotherhand,isthemostcommonlyspokennativelanguageintheworldwitharoundonebillion nativespeakers.Comparedtomanyotherfinanciallyimportantlanguage-pairs,suchasFrenchandEnglishorSpanishand Portuguese,thereisalsoverylittleoverlapinthereadershipofEnglishandChinesenews.Infact,speakersofonelanguage oftendonotunderstandasinglewordintheotherlanguage.Forus,thisisofimportancesincethismakestheEnglishand ChineseGoogleNews-generatednewsvolumesmoredistinctiveandmorelikelytohavetheirownuniquerelationtomarket volatility.
IcollectwhatIdeemtobestockmarket-relatednewsvolumesusinganEnglishlanguageedition(theUSedition)as wellasaChineselanguageedition(theChinaedition)ofthenewsaggregatorGoogleNews.Thisnewsaggregatormakes itpossibletocollectasignificantamountofthemanythousandsofavailablenewspiecesaroundtheworldselectedand groupedbytopic.Bycollecting,onamonthlybasis,thenumberofnewsstoriespresentedbyGoogleNewsIgetanestimate ofthedynamicsofthenewsvolume,i.e.thedynamicsoftheoverallflowofpublicinformation,ratherthanjustsnapshots ofthevolumearoundcertainchosenevents.ThenewsvolumeiscollectedusingtwodifferenteditionsofGoogleNewsas wellastwodifferentlanguages,EnglishandChinese(Mandarin).4Thetotalnumberofseparatenewsstoriescollectedin
3Ofcourse,anysearchforthephrasestockmarketalsoincludesbadnewslikestockmarketcrash.Onaverage,however,Iassumethatgoodandbadnews averageoutwhensearchingforthe“neutral”termstockmarket.
4ApreliminaryanalysisindicatesthatthesearchresultsareverysimilarwhenthenewsdataiscollectedfromtheHongKongeditionofGoogleNews insteadoftheChinaedition.
Table1
NewsVolume–DescriptiveStatistics.InthisTableIpresentsomedescriptivestatisticsforthe105monthlyEnglishnewsvolumeobservations(numberof
monthlynewsstories)fromSeptember11,2006toSeptember1,2014andforthe51monthlyChinesenewsvolumeobservationsfromNovember1,2010
toSeptember1,2014.
Stockmarket Stockmarket
crash (stockmarket) (stockmarket “collapse”) (stockmarket “crash”) Mean 88,600 1900 19,100 630 890 Standarddeviation 59,100 2500 15,400 390 720 Max 307,000 23,400 54,700 1900 4600 Min 4100 400 1500 190 200
thiswayoverthetime-period2006–2014ismorethanninemillion,ofwhicheightmillionarewritteninEnglishandone millionarewritteninChinese.5
GoogleNewsisanautomatednewsaggregator(computer-generatednewsservice)thatusescomputeralgorithmsto collect,presentandsortwebnewsintocategories.GoogleNewsaggregatesnewsfrommorethan50,000newssources worldwide.Itthengroupssimilarstoriestogether,anddisplaysthemaccordingtoeachreader’sinterests(Google,2014). Newsstoriesarecollectedfromnewspagesonthewebwiththegeographicallocationofthenewssourcesdependenton theedition.GoogleNewsincludesnewsthatappearedonanyoftheselectedwebpagesduringthepast30daysandsince nohumaneditorsareusedthepolitical/ideologicalbiasisminimized(Google,2014).Theactualnewssourcesareprobably notknownoutsidetheGoogleplex(theGoogleHQ)butunverifiedrumoursonthewebclaimthatthelargestcontributorsto theGoogleNewsflowareNewYorkPost,WashingtonPost,HoustonChronicleandBloombergintheUS,andTheGuardian, BBCNewsandTheTimesintheUK.
InordertomaketheexercisefeasibleIfocussolelyonnewspiecesdeemedrelevanttostockmarketparticipants.I thereforelimitmyGoogleNewssearchtothesearchstringsstockmarketandstockmarketcrashintheUSeditionand
(stockmarket)and (approximatelytranslatedintoEnglishasstockmarketcollapseandstockmarket
crash)intheChineseedition.6DespitethislimitationinnewscoverageIstillmanagetocollectmorethanninemillion separatenewsstoriescontainingthewordstockmarketovertheeight-yearperiod2006–2014.7
Thedata,i.e.theamountofnewspubliclyavailableworldwideoverthelast30days,iscollectedonamonthlybasisfrom September11,2006toSeptember1,2014fortheEnglishnewsandfromNovember1,2010toSeptember1,2014forthe Chinesenews.Moreexactly,thedataiscollectedmanuallybytheauthoreveryfourthMondayatapproximately9:00a.m. CentralEuropeanTimeatthesamelocation(theofficeoftheauthor)andwithoutbeingsignedintoanyGoogleAccount. OnahandfulofoccasionswhentheauthorwastravellingonthescheduledMonday,thedatacollectionwasdoneonthe followingTuesdayorWednesdayoratanotherlocation.8,9Onanyparticularday,theGoogleNewsaggregatorcollectsdata fromthelast30-dayperiod.Thismeansthatmydefinitionofamonth(asa4-weekor28-dayperiod)differsslightlyfrom Google’sdefinitionofamonth(a30-dayperiod).Thedifferenceissmall,2days,andsincetheadditional2daysofnews crawlingalwaysconstitutesaweekend(whenmarketsareclosedandthereislessmarket-relatednewsavailable)four weeksagoIbelievethatitbiasestheresultsminimally.Inanycase,thisdiscrepancyshouldprobablybiastheresultsagainst
usfindingalinkbetweennewsvolumeandvolatility.10
InTable1,Ipresentsomedescriptivestatisticsonthemonthlysamplednewsvolume.Themonthlynumberofnews storiesvariesnotonlyacrosstimebutamongthesearchstringsusedintheGoogleNewssearch.Whilethemoregeneral searchstringstockmarket(inEnglish)returnsonaverage89,000separatenewsstoriespermonth,themorenarrowsearch stringstockmarketcrash(again,inEnglish)returnsonaverage1900newsstoriespermonth.Thepatternissimilarforthe Chinese-languagenewsvolumeswhere (stockmarket)onaveragereturns19,000newsstoriespermonthand
5 Inmid-May2012,thenumberofnewsstoriesreportedbyGoogleNewsincreasesdramatically(mostlikelybytheinclusionofadditionalnewssources)
andIhavethereforechosentoadjustthenumbersfromJune2012onwards.Thenumbersarenormalizedsothatthefirstobservationafterthechange
isidenticaltothelastobservationbeforethechange.Reportednumbersarealwaysnormalizedonesandtheyarethereforenotdirectlycomparableto currentGoogleNewsvolumes.Thetotalnumberofnewsstoriesreportedinthispaperisthereforealsounder-reported.Someofthecorrelationanalysis inthispaperhasbeenredonewithdataupuntilthechangewithroughlyunchangedresults.
6 Thereasonforchoosingtwodifferent“pessimistic”searchstringsinChineseistocontrolforanypotentiallanguage-differenceintheinterpretationof theword“crash”.
7 Here,Iassumethatnonewsstoryisreprintedagainatalaterstageandthatnotwonewsstoriesareexactlyidentical.
8 The“December”searchresultissometimesmissingduetoChristmasandNewYear’sEve.Therearealsosomemissingobservationsonotherdays
randomlyscatteredthroughouttheyearsandtheseaswellasthemissingDecemberobservationsareallreplacedbythelastavailabledatapoint.Inorder toseewhetherthesemissingobservationsaffecttheresults,adummyisaddedtotheregressions,andcorrelationsarerecalculatedwiththedatesofthe
missingobservationsremoved,withalmostunchangedresults.
9 Inordertomakesurethatthenewsvolumeisnotplatform-dependentthenumberofnewsstorieswasoccasionallycollectedatseverallocationsthe
sameday(atrandomlychosendaysthroughoutthesampleperiod)withveryminordifferences.
10 AccordingtothehomepageofGoogleNewsthenewsaggregatorincludesnewsarticlesthathavebeencrawledwithinthelast30days.However,a
carefulstudyofthesearchresultssometimesrevealsafewnewsstoriesthatare(afewdaysorweeks)olderthan30days.However,ontheseoccasions thenumberofnewsstoriesthatareolderthan30dayshavebeenfoundtobefewcomparedtothetotalnumberofstoriesandthereforelesslikelyto significantlybiastheresults.
Fig.1. English-languagenewsvolumeandMSCIWorldstockmarketvolatility.ThisgraphshowstheEnglish-language(USedition)GoogleNewsvolumes forthesearchstring“stockmarketcrash”togetherwiththeMSCIWorldstockreturnvolatility.Boththenewsvolumeandthestockvolatilityarenormalized
tostartatoneandaresampledonamonthlybasisbutsmoothedusingathree-month(quarterly)smoothingwindow.
Fig.2.Chinese-languagenewsvolumeandMSCIWorldstockmarketvolatility.ThisgraphshowstheChinese-language(Chinaedition)GoogleNews
volumesforthesearchstring“stockmarketcrash”togetherwiththeMSCIWorldstockreturnvolatility.Boththenewsvolumeandthestockvolatilityare
normalizedtostartatoneandaresampledonamonthlybasisbutsmoothedusingathree-month(quarterly)smoothingwindow.
(stockmarketcrash)onaveragereturns890newsstoriespermonth(acrossashortertime-period).Thetime-seriesvariation ofthestockmarketcrashnewsvolumes,inEnglishandinChinese,respectively,isgraphicallypresentedinFigs.1and2.Not surprisingly,theamountofstockmarket-relatednewsincirculationwasatitshighestaroundtheperiodoftheLehmann Brotherscollapse.Thenumberofnewsstoriescontainingthewordsstockmarketwashigherthan200,000(atroughly2½ timesthemonthlyaverage)bothinSeptemberandOctoberin2008.Twoyearsearlier,inSeptemberandOctober2006, thenumberofnewsstorieswasaround70,000permonth.TheChinesenewswasnotcollectedatthetimeoftheLehman Brotherscollapseandthepeakinthenumberofnewsstoriescontainingtheword (stockmarket)wasreachedinthe springof2012whenmorethan30,000newsstorieswerereleasedeachmonthfromMarchtoJune.
Thestockmarketdata,inturn,iscollectedforthesametime-periodasthenewsvolumes,i.e.September11,2006to September1,2014.ThedataisdownloadedfromDatastreamandallthestockindexesaredenominatedintheir home-currency.Iincludetwelvedifferentstockindexesinmyanalysis.Sincemyaimistostudytheeffectoflanguageonthe news-volatilitylink,IhavechosenhalfoftheindexesfromthemainlyEnglish-speakingsphere,i.e.theUS,theUKandthe globalcommunity,andhalffromthemainlyChinese-speakingsphere,whichIdefineasChinaincludingHongKong.As thestockindexrepresentingtheglobalAnglophonecommunityIhavechosentheMSCIWorldstockindexwhichincludes securitiesfrom23developedstockmarketsaroundtheworld.FromtheUSIhaveincludedtheS&P500-,DJIA-, Nasdaq-andRussell2000indexes.Themainmotivationbehindincludingthelasttwoindexesistheirfocusonsmall-capstocks.It ispossiblethatthenews-volatilitylinkisdifferentforsmallstockswherethebalancebetweensmallretail-investorsand largeinstitutionalinvestorsisdifferent.TheUK,finally,isrepresentedbytheFTSE100indexwhichcoversthe100largest companiesontheLondonStockExchange.
WhilemostreadersarefamiliarwiththeAnglophoneworldanditsstockmarkets,theChinese-speakingsphereandthe stockmarketsdominatedbyChinese-speakersprobablyneedssomeintroduction.SincemyfocusisMandarin,themain languagespokeninChina,IhaveonlychosenstockindexesthatcontainChinesestocks.TheChinesestockmarketishighly segmentedwithdifferentmarketsaimedatdifferentinvestors:
•A-shares:A-sharesareRMB-denominatedsharesissuedbydomesticcompaniesregisteredinmainlandChinaandlisted ontheShanghaiStockExchangeortheShenzhenStockExchange.A-sharescanonlybepurchasedbydomesticChinese investorsorholdersofQualifiedForeignInstitutionalInvestor(QFII)licenses.
•B-shares:B-sharesaredollar-denominatedsharesissuedbydomesticcompaniesregisteredinmainlandChinaandlisted ontheShanghaiStockExchange(US$)ortheShenzhenStockExchange(HK$).B-sharescanonlybepurchasedbyforeign investorsorbydomesticinvestorswithforeigncurrencyholdings,andcapitalcontrolsrestrictChineseresidents’ability topurchaseBshares.
•H-shares:H-sharesareHK$-denominatedsharesissuedbycompaniesincorporatedinmainlandChinabutlistedonthe HongKongStockExchange.H-sharescannotbepurchasedbydomesticChineseinvestors.
•RedChip-shares:RedChip-sharesareHK$-denominatedsharesissuedbyChinesecompaniesincorporatedinHongKong andlistedontheHongKongStockExchange.RedChip-sharescannotbepurchasedbydomesticChineseinvestors.
WhiletheA-share marketisaimedmainlyatdomesticinvestors,theB-sharemarketis aimedpredominantly at for-eigninvestorsandtheH-shareandRedChip-sharemarketsareaimedsolelyatforeigninvestors.Thisgradualincreasein segmentationfacilitatesthestudyoftheeffectofmarketparticipantandlanguageonthenewsvolume—volatilitylink.
IlookatChinesesharestradedonthreedifferentexchanges;A-andB-sharestradedontheShanghaiStockExchange, A-andB-sharestradedontheShenzhenStockExchange,andH-andRedChip-sharestradedontheHongKongStockExchange. TheactualstockindexesaretheShanghaiSEAIndex,theShanghaiSEBIndex,theShenzhenSEAIndex,theShenzhenSEB
Index,theHangSengChinaEnterprisesIndexandtheHangSengChina-AffiliatedCorporationsIndex.
Eachmonth,i.e.everyfourthMonday,thepastmonth’sstockmarketvolatilityiscalculatedasthestandarddeviation ofthedailystockindexreturnsoverthelastfourweekssothatthetime-periodforthevolatilityestimatematchesthe time-periodforthenewsvolumecollection(exceptforthe2daysdiscussedabove).Thepossibilityofmatchingthenews collectionperiod(onemonth)withthevolatilitycomputationperiod(onemonth)isonehugeadvantageofusingamonthly frequencyintheanalysis.Anotheradvantageisthattheexacttimestampofthenewsreleaseisnotrequired(aproblem facedbyseveralpreviousstudies,forinstanceDimpflandJank(2015))andthatthetimezonedifferencesaroundtheworld, mostnotablybetweenChinaandtheUS,haveaminimaleffectontheresults.
3. Therelationbetweennewsvolumesandstockmarketvolatility
Inefficientfinancialmarkets,pricemovementsaretheresultsofmarketparticipantsreactingonmarket-relatednews. Asaresult,themorenewsthatreachesthemarketoveracertaintime-periodthehigherthepricevolatilityinthemarket islikelytobe.InthisstudyofstockmarketsworldwideIthereforeexpectthestockreturnvolatilitytobepositivelylinked totheamountofstockmarket-relatednewsworldwide.Indeed,aninitialvisualinspectionofFigs.1and2,wheremonthly stockmarketvolatilitiesandnewsvolumesarepresentedonamonthlybasisforEnglishandChinesenewsstories(smoothed usingaquarter-of-a-yearlongwindowandnormalizedtostartatone),motivatesustoinvestigatethislinkfurther.
3.1. Correlationanalysisofnewsvolumesandstockmarketvolatility
Tostartwith,Ipresentsimplecontemporaneouscorrelationsbetweennewsaggregatorgeneratednewsvolumesand stockmarketvolatilities.IstudytwelvedifferentstockmarketindexesandcollectnewsinEnglishaswellasinChinese. Allvariablesaresampledonamonthlybasis.InadditiontolevelsIalsolookatchangesinnewsvolumeandvolatility. Forthechanges,IfollowMitchellandMulherin(1994),whotakedifferencesfromamulti-daymovingaverage,bytaking differencesfroma12-monthmovingaverageofthepastnewsvolumeandvolatility,respectively.JustlikeMitchelland Mulherin(1994),Itakemulti-perioddifferencestoavoidthelossofinformationaroundtheoccasionalclusteringofhigh newsvolumesandhighvolatilitylevelsandtoreducetheinfluenceofpossiblemonth-of-the-yeareffects.
ThecorrelationresultsarepresentedinTable2,whichisdividedintotwoparts,oneforlevelsandoneforchanges.With veryfewexceptions,thecorrelationcoefficients(basedontheentiresample)amongnewsvolume-andvolatilitylevelsare large,positiveandstatisticallysignificant.Mostcorrelationslieinthe0.3–0.8rangeandtheonlynon-significantcorrelations arethoseinvolvingtheChinese-language (stockmarketcrash)newsstories.Thesecorrelationsaretypicallythe lowest,regardlessofstockmarket,andinthemainlandChinastockmarketsinShanghaiandShenzhenthecorrelationsare occasionallyevenslightlynegative.Evenifthenegativecorrelationsaresmallandnotstatisticallysignificant,itisinteresting thatthelinktoChinesenewsisfoundtobeweakestintheChinesestockmarketregardlessoftheactualGoogleNewssearch string.
WhenIturntochangesratherthanlevels,thecorrelationcoefficientsarestilllargelypositiveandstatisticallysignificant. Thestatisticalsignificance,i.e.thesize,ofthecorrelationsisgenerallysomewhatweakerforchanges,however,andthe correlationsareagainlowestwhenChinesenewsormainlandChinastockmarketsareinvolved.Forchanges,though,the linkbetweenmainlandChinastockmarketsandnewsisweakacrosstheboard,i.e.regardlessofthelanguageofthenews. OnepossibleexplanationfortheratherweakrelationshipbetweennewsandstockvolatilityinmainlandChina(i.e.excluding stockstradedinHongKong)isthatthesemarketsaredominatedbyinvestorsthatdonotread,oratleastdonottradeon, (traditional)news.ThishypothesisisstrengthenedbythefactthatthecorrelationisparticularlyweakintheA-sharemarket
Table2
CorrelationsbetweenNewsVolumeandVolatilityInthisTableIpresentNewsVolume–Volatilitycorrelationsforthetwelvestockindexeswhennews
volumesarecollectedbyGoogleNewsisinEnglishandChinese,respectively.Resultsarepresentedbothforlevelsandchanges.
NewsinEnglish(Levels)
Stockmarket Stockmarketcrash
MSCI 0.50*** 0.77*** S&P500 0.57*** 0.75*** DJIA 0.57*** 0.77*** NASDAQ 0.55*** 0.74*** Russell2000 0.48*** 0.66*** FTSE100 0.60*** 0.79*** ShanghaiA 0.59*** 0.34*** ShanghaiB 0.47*** 0.33*** ShenzhenA 0.49*** 0.26*** ShenzhenB 0.50*** 0.33*** HongKongH 0.66*** 0.73***
HongKongRedChip 0.66*** 0.69***
NewsinChinese(Levels)
(stockmarket) (stockmarket“collapse”) (stockmarket“crash”)
MSCI 0.47*** 0.75*** 0.22* S&P500 0.34*** 0.71*** 0.26** DJIA 0.32** 0.71*** 0.27*** NASDAQ 0.33*** 0.67*** 0.25** Russell2000 0.35*** 0.67*** 0.19* FTSE100 0.45*** 0.70*** 0.18 ShanghaiA 0.23** 0.40*** 0.00 ShanghaiB 0.33*** 0.42*** −0.02 ShenzhenA 0.23** 0.37*** −0.06 ShenzhenB 0.37*** 0.54*** −0.03 HongKongH 0.29** 0.69*** 0.14
HongKongRedChip 0.33*** 0.67*** 0.17
NewsinEnglish(Changes)
Stockmarket Stockmarketcrash
MSCI 0.19** 0.71*** S&P500 0.18** 0.62*** DJIA 0.16* 0.62*** NASDAQ 0.15* 0.61*** Russell2000 0.17** 0.58*** FTSE100 0.23*** 0.62*** ShanghaiA 0.09 0.17** ShanghaiB 0.03 0.18** ShenzhenA 0.02 0.07 ShenzhenB 0.04 0.19** HongKongH 0.21** 0.53***
HongKongRedChip 0.27*** 0.54***
NewsinChinese(Changes)
(stockmarket) (stockmarket“collapse”) (stockmarket“crash”)
MSCI 0.09 0.63*** 0.37*** S&P500 0.06 0.61*** 0.35*** DJIA 0.06 0.60*** 0.35*** NASDAQ 0.10 0.59*** 0.35*** Russell2000 0.05 0.58*** 0.27** FTSE100 0.02 0.54*** 0.28** ShanghaiA −0.24** 0.26** 0.04 ShanghaiB 0.01 0.27** 0.10 ShenzhenA −0.08 0.21* −0.01 ShenzhenB −0.05 0.39*** 0.11 HongKongH −0.07 0.54*** 0.23**
whichisaimedmainlyatdomestic(retail)investors.Infact,thelinkbetweenthevolatilityintheChinesestockmarketand theamountofstockmarket-relatednewsreachingthemarketparticipantsisstrongerinmarketswheretherearefewer mainlandChineseretailinvestors.ThelinkisstrongestintheH-shareandRedChip-sharemarketsinHongKongwhere Chineseretailinvestorsarebannedfromtrading.ThelinkissomewhatweakerinthetwoB-sharemarketswherefew Chineseretailinvestorsinvestanditisweakestintheretail-dominatedA-sharemarkets.
Tosomeextent,thesefindingsonthesegmentationandinformationprocessingintheChinesestockmarketareinline withthelimitedamountofresearchthatexistsonthetopic.InapreviousstudyusingGoogleNews,Byström(2011)finds astronglinkbetweenEnglish-languagenewsvolumesworldwideandtheglobalstockmarketvolatility,andaweakerlink betweentheChinesestockmarketandthesamesetofworldwide(English)news.PoonandFung(2000)looksathow informationflowsamongthevariousChina-andHongKong-basedmarketsandfindsreturnandvolatilityspillovereffects amongsecuritieslistedonthedifferentmarkets.TheRedChipmarketisfoundtoprocessinformationfasterthantheother markets.ThesegmentationandinformationflowsamongChinesestockmarketshasalsobeenstudiedbyYang(2003)who findstheforeigninvestorswhodominatetheB-sharemarkettobebetterinformedthanthedomesticinvestorsinthe A-sharemarket.AndinastudyrelatingnewsandstockmarketvolatilityinthesegmentedChinesestockmarket,Suand Fleisher(1999)findsthevolatilityintheA-sharemarketinthelate90stobesignificantlyhigherthanthatintheB-share marketandtheytrytoexplainthisfactusingargumentsrelatedtonewsflowsanddifferentinvestorbases.
Lookingattheoverallpicture,however,thelinkbetweennewsandvolatilityisstrong.Thosemonthswhenalotofnews isreleasedaretypicallyalsothosemonthswhenthestockmarketvolatilityishigh,andmonthswithsignificantrelative increases/decreasesinnewsvolumes(relativetothelasttwelvemonths’volume)arealsomonthswithasignificantrelative increase/decreaseinmarketactivity(i.e.volatility).Additionalevidenceofthestrongassociationbetweennewsandvolatility isgiveninFigs.1and2wheretherelativesizeofthemonthlymovementsinnewsvolumeisquitesimilartotherelative sizeofthemovementsintheMSCIWorldvolatility.Thatis,newsvolumesandstockvolatilitiesarenotonlymovinginthe samedirectionatthesametimebuttheactualsizeofthechangesissimilaraswell.
3.2. Regressionanalysisofnewsvolumesandstockmarketvolatility
Thenextstepistorununivariatelead-lagOLSregressionsbetweennewsvolumeandstockmarketvolatility.The regres-sionsallowustoassessthepredictiveabilityofnewsvolumesandtoevaluatetheeconomicsignificanceoftheinter-temporal news-volatilitylinkinthestockmarket.Iruntwosetsofregressions.Inthefirstset,thedependentvariableisthe(monthly) volatilityandtheexplanatoryvariablesare(one-month)laggednewsvolumeand(one-month)laggedvolatility.Inthe sec-ondset,Ireversetheregressionandthedependentvariableisthe(monthly)newsvolumeandtheexplanatoryvariables are(one-month)laggedvolatilityand(one-month)laggednewsvolume.Toaccountforthepossibleinfluenceofmissing newsvolumevaluesIincludeadummyforthemissingvaluesinallreportedregressions.Theregressionsincludeonlyone lagofvolatilityandnewsvolumeandthereasonisthatIlookatmonthlydata;anyempiricalrelationshipfoundbetween stockmarketvolatilitythismonthandnewsreleasedseveralmonthsagoislikelytobespurious.Indeed,Tetlock(2007), wholooksatdailynewsreleasesandincludeslagsuptofivedays,findsareversalintheinitialreactiontonewsalready fourorfivedaysafterthenewsrelease.AndVozlyublennaia(2014),whocollectsGooglesearchqueriesonaweeklybasis asameasureofinvestorattention,arguesthatherweaklinkbetweenattentionandvolatilitycouldbecausedbyherusing weekly,ratherthandaily,datasincetheeffectofattentiononvolatilitycoulddisappearalreadyindays,ratherthaninweeks. Inotherwords,ifthesameappliestonewsandtoitseffectonvolatility,thereisnoreasontoexpectalinkbetweenstock marketvolatilitythismonthandnewsreleasedseveralmonthsago.
Tables3and4summarizetheregressionresults.Sincemyfocusisprimarilyonthenewsvolume/volatilitycoefficients, theinterceptintheregressionequationandthemissingvaluedummycoefficient(whichisrarelystatisticallysignificant) areleftout.Table3presentstheresultsfortheEnglish-languagenewsandTable4presentstheresultsfornewsinChinese. IntheupperregressionslabelledNews,thenewsvolumeisthedependentvariableandinthelowerregressionslabelled
Volatility,thestockmarketvolatilityisthedependentvariable.Inbothregressions,thefirstexplanatoryvariable,ˇ1,is
alwaystheone-monthlaggedvalueoftheothervariable(i.e.volatilityintheupperregressionandnewsvolumeinthe lowerregression)andthesecondexplanatoryvariable,ˇ2,isalwaystheone-monthlaggedvalueofthedependentvariable.
IntheanalysisthatfollowsIthereforefocusonˇ1.Resultsforchangesfollowafterthoseforlevelsand*,**and***represent
significanceatthe10%,5%and1%level,respectively.
Overall,ˇ1isoften,butnotalways,positive.ItismuchmoreoftenpositiveintheVolatilityregression(inallbutfive
cases),i.e.laggednewsvolumeismorelikelyto(Granger)causevolatilitythanviceversa.FortheVolatilityregression,ˇ1
isstatisticallysignificantforroughlyhalfthemarket/language/searchstringcombinationsinTables3and4,andthereare 42caseswhennewsvolume(Granger)causesvolatilityandvolatilitydoesnotcausenewsvolume,butonly7caseswhen volatilitycausesnewsvolumeandnewsvolumedoesnotcausevolatility.Interestingly,theselatterseveninstancesareall foundinmainlandChinawhenthenewsisinEnglish.ThisisanindicationofEnglish-languagenewsnotreachingmainland Chinesestockmarketparticipants,oratleastthisnewsdoesnothelpinpredictingvolatility.Meanwhile,Chinesenews (Granger)causesvolatilityinmainlandChinajustasitdoesinmostotherstockmarkets.Asforˇ2,finally,itiswell-known
thatvolatilityispersistentandthatlaggedvolatilitypredictsfuturevolatilityandthisisevidentinmyregressionsinthe (almost)unanimouslysignificantˇ2coefficient.Interestingly,newsvolumeseemstobeaspersistentasvolatility.
Table3
ResultsofEnglish-LanguageNewsVolume–VolatilityRegressionsInthisTableIpresentresultsfromtheEnglish-languageregressions.Intheupper
regressions(News)thenewsvolumeisthedependentvariableandinthelowerregressions(Volatility)thestockmarketvolatilityisthedependent
variable.Inbothregressions,thefirstexplanatoryvariable,ˇ1,isalwaystheone-monthlaggedvalueoftheothervariable(i.e.volatilityintheupper
regressionandnewsvolumeinthelowerregression)andthesecondexplanatoryvariable,ˇ2,isalwaystheone-monthlaggedvalueofthedependent
variable.Eachregressionhasanunreportedinterceptandanunreporteddummyformissingvalues(whichisrarelysignificant).*,**and***represent
significanceatthe10%,5%and1%level,respectively.Theregressionsarebasedon105monthlyobservationsfromSeptember11,2006toSeptember1,
2014andtheresultsarepresentedbothforlevelsandforchanges.
NewsinEnglish(Levels)
Stockmarket Stockmarketcrash
Dependentvariable ˇ1 ˇ2 ˇR2 ˇ1 ˇ2 ˇR2 MSCI News −0.11* 0.88*** 0.70 0.12 0.42*** 0.25 Volatility 0.058 0.49*** 0.62 0.11* 0.42*** 0.62 S&P500 News −0.080 0.88*** 0.70 0.22 0.35*** 0.26 Volatility 0.083 0.61*** 0.65 0.14* 0.54*** 0.65 DJIA News −0.090 0.88*** 0.70 0.17 0.38*** 0.25 Volatility 0.086* 0.54*** 0.64 0.14** 0.47*** 0.64 NASDAQ News −0.066 0.87*** 0.70 0.20 0.36*** 0.26 Volatility 0.083 0.57*** 0.61 0.14** 0.49*** 0.62 Russell2000 News −0.080 0.87*** 0.70 0.092 0.45*** 0.25 Volatility 0.12** 0.69*** 0.63 0.22*** 0.58*** 0.64 FTSE100 News −0.080 0.88*** 0.70 0.186 0.37*** 0.26 Volatility 0.057 0.49*** 0.55 0.080 0.45*** 0.55 ShanghaiA News 0.10* 0.80*** 0.70 0.22** 0.44*** 0.29 Volatility 0.064 0.50*** 0.56 0.061 0.50*** 0.56 ShanghaiB News 0.079 0.82*** 0.70 0.24*** 0.43*** 0.29 Volatility 0.10 0.56*** 0.39 0.090 0.56*** 0.39 ShenzhenA News 0.083 0.82*** 0.70 0.18** 0.47*** 0.27 Volatility 0.10* 0.40*** 0.41 0.087 0.41*** 0.41 ShenzhenB News 0.066 0.83*** 0.70 0.21** 0.44*** 0.28 Volatility 0.055 0.43*** 0.38 0.055 0.42*** 0.38
HongKongH News −0.049 0.87*** 0.70 0.12 0.43*** 0.25
Volatility 0.13 0.71*** 0.51 0.023 0.76*** 0.50
HongKongRedChip News −0.035 0.86*** 0.70 0.15 0.41*** 0.25
Volatility 0.097 0.64*** 0.52 0.010 0.67*** 0.51 NewsinEnglish(Changes)
Stockmarket Stockmarketcrash
DependentVariable ˇ1 ˇ2 ˇR2 ˇ1 ˇ2 ˇR2 MSCI News −0.038 0.52*** 0.26 0.049 0.39*** 0.17 Volatility 0.058 0.34*** 0.39 0.11* 0.27*** 0.40 S&P500 News −0.014 0.52*** 0.26 0.071 0.38*** 0.17 Volatility 0.063 0.33*** 0.35 0.13** 0.26*** 0.37 DJIA News −0.028 0.52*** 0.26 0.032 0.40*** 0.17 Volatility 0.060 0.33*** 0.36 0.13** 0.26*** 0.38 NASDAQ News −0.019 0.52*** 0.26 0.056 0.39*** 0.17 Volatility 0.056 0.27*** 0.31 0.13** 0.21*** 0.34 Russell2000 News −0.007 0.52*** 0.26 0.007 0.42*** 0.16 Volatility 0.08* 0.30*** 0.37 0.14*** 0.22*** 0.40 FTSE100 News −0.011 0.52*** 0.26 0.056 0.39*** 0.17 Volatility 0.044 0.28*** 0.33 0.079 0.25*** 0.34 ShanghaiA News 0.13* 0.50*** 0.28 0.079 0.41*** 0.17 Volatility 0.026 0.11*** 0.12 0.029 0.10*** 0.12 ShanghaiB News 0.11 0.51*** 0.27 0.14 0.40*** 0.18 Volatility 0.060 0.12*** 0.09 0.035 0.12*** 0.08 ShenzhenA News 0.048 0.52*** 0.26 0.048 0.42*** 0.17 Volatility 0.043 0.11*** 0.12 0.017 0.10*** 0.11 ShenzhenB News 0.12 0.51*** 0.27 0.10 0.41*** 0.18 Volatility 0.025 0.12*** 0.11 0.022 0.11*** 0.11
HongKongH News 0.022 0.51*** 0.26 −0.035 0.44*** 0.17
Volatility 0.043 0.24*** 0.31 0.056 0.22*** 0.31
HongKongRedChip News 0.031 0.51*** 0.26 0.006 0.42*** 0.16
Table4
ResultsofChinese-LanguageNewsVolume-VolatilityRegressionsInthisTableIpresentresultsfromtheChinese-languageregressions.Intheupper regressions(News)thenewsvolumeisthedependentvariableandinthelowerregressions(Volatility)thestockmarketvolatilityisthedependent variable.Inbothregressions,thefirstexplanatoryvariable,ˇ1,isalwaystheone-monthlaggedvalueoftheothervariable(i.e.volatilityintheupper
regressionandnewsvolumeinthelowerregression)andthesecondexplanatoryvariable,ˇ2,isalwaystheone-monthlaggedvalueofthedependent
variable.Eachregressionhasanunreportedinterceptandanunreporteddummyformissingvalues(whichisrarelysignificant).*,**and***represent significanceatthe10%,5%and1%level,respectively.Theregressionsarebasedon51monthlyobservationsfromNovember1,2010toSeptember1,2014 andtheresultsarepresentedbothforlevelsandforchanges.
NewsinChinese(Levels)
(stockmarket) (stockmarket “collapse”) (stockmarket “crash”) Dependent Variable ˇ1 ˇ2 ˇR2 ˇ1 ˇ2 ˇR2 ˇ1 ˇ2 ˇR2 MSCI News 0.055 0.81*** 0.69 0.59*** 0.21 0.56 0.002 0.30** 0.09 Volatility 0.092** 0.26*** 0.60 0.11** 0.22*** 0.60 0.11** 0.29*** 0.61 S&P500 News 0.005 0.84*** 0.69 0.52*** 0.29** 0.54 0.012 0.30** 0.09 Volatility 0.098* 0.26*** 0.44 0.22*** 0.14** 0.51 0.17*** 0.24*** 0.54 DJIA News −0.0002 0.84*** 0.69 0.51*** 0.30** 0.53 0.013 0.30** 0.09 Volatility 0.084* 0.23*** 0.45 0.17*** 0.14** 0.51 0.15*** 0.22*** 0.54 NASDAQ News −0.002 0.84*** 0.69 0.49*** 0.33** 0.54 0.008 0.30** 0.09 Volatility 0.12** 0.23*** 0.36 0.24*** 0.11 0.45 0.20** 0.22*** 0.47 Russell2000 News 0.023 0.83*** 0.69 0.46*** 0.35** 0.52 −0.054 0.31** 0.09 Volatility 0.14* 0.33*** 0.39 0.37*** 0.13 0.55 0.25*** 0.34*** 0.51 FTSE100 News 0.063 0.81*** 0.69 0.43*** 0.36** 0.50 −0.001 0.30** 0.09 Volatility 0.12*** 0.21*** 0.55 0.14*** 0.16*** 0.53 0.076* 0.25*** 0.50 ShanghaiA News 0.10 0.82*** 0.70 −0.051 0.69*** 0.40 −0.12 0.30** 0.10 Volatility 0.035 0.050 0.02 0.074* 0.030 0.07 0.029 0.060 0.01 ShanghaiB News 0.10 0.81*** 0.70 0.14 0.60*** 0.42 −0.033 0.30** 0.09 Volatility 0.13* 0.072 0.09 0.067 0.081 0.04 0.006 0.11* 0.02 ShenzhenA News 0.066 0.85*** 0.70 −0.13 0.71*** 0.42 −0.13 0.29** 0.11 Volatility 0.052 0.029 0.09 0.067 0.029 0.05 −0.016 0.039 0.00 ShenzhenB News 0.092 0.81*** 0.70 0.19 0.56*** 0.43 −0.065 0.30** 0.09 Volatility 0.11* 0.17*** 0.26 0.11* 0.14** 0.26 0.009 0.21*** 0.20
HongKongH News 0.074 0.82*** 0.70 0.26* 0.48*** 0.44 −0.075 0.31** 0.10
Volatility 0.082 0.33*** 0.43 0.27*** 0.16** 0.57 0.084 0.34*** 0.43
HongKongRedChip News 0.071 0.82*** 0.70 0.24* 0.50*** 0.44 −0.073 0.31** 0.10
Volatility 0.082 0.30** 0.43 0.27*** 0.15** 0.57 0.084 0.32*** 0.43
NewsinChinese(Changes)
(stockmarket) (stockmarket “collapse”) (stockmarket “crash”) Dependent Variable ˇ1 ˇ2 ˇR2 ˇ1 ˇ2 ˇR2 ˇ1 ˇ2 ˇR2 MSCI News 0.025 0.61*** 0.39 0.45*** 0.23 0.36 0.18 0.009 0.00 Volatility 0.039 0.25*** 0.39 0.16*** 0.16*** 0.48 0.19*** 0.19*** 0.57 S&P500 News −0.020 0.61*** 0.39 0.43*** 0.25* 0.35 0.14 0.025 0.00 Volatility 0.051 0.26*** 0.27 0.29*** 0.09 0.46 0.33*** 0.15*** 0.64 DJIA News −0.020 0.61*** 0.39 0.41*** 0.27* 0.34 0.14 0.027 0.00 Volatility 0.049 0.26*** 0.28 0.25*** 0.11** 0.44 0.31*** 0.16*** 0.61 NASDAQ News −0.011 0.61*** 0.39 0.44*** 0.25* 0.36 0.13 0.028 0.00 Volatility 0.059 0.20*** 0.20 0.27*** 0.06 0.40 0.31*** 0.11** 0.57 Russell2000 News −0.005 0.61*** 0.39 0.35** 0.31** 0.31 0.030 0.067 0.00 Volatility 0.051 0.24*** 0.22 0.33*** 0.05 0.52 0.31*** 0.15*** 0.61 FTSE100 News 0.026 0.61*** 0.39 0.34** 0.34** 0.31 0.12 0.042 0.00 Volatility 0.05 0.22*** 0.35 0.14*** 0.14*** 0.43 0.17*** 0.16*** 0.52 ShanghaiA News 0.034 0.62*** 0.39 −0.13 0.55*** 0.24 −0.041 0.077 0.00 Volatility −0.074** 0.025 0.08 0.049 0.030 0.02 0.049 0.04 0.03 ShanghaiB News 0.076 0.61*** 0.39 0.097 0.49*** 0.24 0.10 0.065 −0.02 Volatility −0.039 0.090* 0.04 0.019 0.083 0.03 0.048 0.083* 0.04 ShenzhenA News 0.038 0.62*** 0.39 −0.19 0.56*** 0.26 −0.081 0.075 −0.03 Volatility −0.058* 0.03 0.07 0.037 0.026 0.03 0.016 0.04 0.01 ShenzhenB News 0.046 0.62*** 0.39 0.14 0.46*** 0.24 0.065 0.068 −0.03 Volatility −0.008 0.14*** 0.14 0.069 0.11** 0.18 0.061 0.13*** 0.17
HongKongH News 0.037 0.62*** 0.39 0.19 0.41*** 0.25 0.019 0.071 −0.03
Volatility 0.011 0.26*** 0.44 0.18*** 0.16** 0.59 0.14*** 0.23*** 0.57
HongKongRedChip News 0.048 0.61*** 0.39 0.16 0.43*** 0.24 0.021 0.069 −0.03
Asfortheeconomicsignificanceoftheresults,andhereIfocussolelyontheimpactofnewsonvolatility,thesizeofthe regressionparameterstellsusthat,forlevels,theaverageimpactofaone-standarddeviationchangeinnewsvolumeon nextmonth’sstockindexvolatilityis11basispoints(0.11%).Albeitnotlarge,theimpactiseconomicallymeaningfulwhen comparedtothemeanofthetwelvestockindexes’unconditionalvolatilityacrossthesampleperiod(coveringthevery volatilecredit-andtheeuro-areasovereigncrises)whichis123basispoints.Ofcourse,forsomemarket/language/search stringcombinationstheeconomicsignificanceoftheregressionparametersismuchlarger.FortheRussell2000indexofsmall USfirms,forinstance,theaverageimpactofaone-standarddeviationchangeinthevolumeofnewscontainingtheChinese word“ i.e.stockmarketcollapse”,onnextmonth’sstockindexvolatilityis37basispoints.Forchanges,theaverage impactofaone-standarddeviationchangeinnewsvolume,comparedtothe12-monthaverage,onnextmonth’schangein stockindexvolatilityis9basispoints.11Althoughsmaller,theimpactforchangesisactuallymoreeconomicallymeaningful thanthatforlevelswhencomparedtothesmallermeanofthemonthlychangeinthestockindexes’unconditionalvolatility whichis44basispoints.
Theregressionresultsinthissectionrevealthattheamountofnewsthismonthpredictsthevolatilitynextmonth.Thisis possiblyanindicationofnewsdissipatingfromnewssourcetomarketparticipantratherslowlyanditmirrorsresultsfound byEderingtonandLee(1993)intheinterestrateandforeignexchangemarkets.Likethestockmarket,thesetwomarketsalso exhibittime-varyingvolatilityandEderingtonandLee(1993)showsthatscheduledmacroeconomicnewsannouncements haveanimmediateeffectonpricesbutalonglastingeffectonpricevolatility.Inthestockmarket,anevenlonger-lasting periodofelevatedvolatilityafterannouncementsisfoundbyPatellandWolfson(1984).Thisgradualdigestionofnewsby themarketcouldliebehindthepredictabilityofvolatilityusingnews.
Arecurringfindinginthispaperisthatpessimistic(negative)newshasasomewhatstrongerconnectiontostockreturn volatilitythanneutralnews.ThecorrelationspresentedinTable2aregenerallylargerfornegativenews,regardlessof language,andthesameholdsfortheregressioncoefficient,ˇ1,inTables3and4.Ihavenoexplanationforthisotherthan
thepossibilitythatrisk-aversemarketparticipantsaremorepronetoreacttonegativethantoneutralnews.
Toconclude,themainfindinginthissectionisthattheinter-temporallinkbetweennewsandvolatilityisstatisticallyand economicallysignificantandthatitseemstobestrongerinthedirectionfromnewstovolatilitythanviceversa.Asecondary findingisthatthepatternisdifferentinmainlandChinawherethereislessofapositiveinter-temporallinkbetweennews andvolatilityandwhere,particularlyEnglish-language,newsdoesnothelpinpredictingvolatility.
3.3. Out-of-sampleforecastingofstockmarketvolatilityusingnewsvolumes
Theregressionsintheprevioussectionshowthatnewscausesvolatilityandthatnewsvolumespossiblycouldbeusedto predictfuturestockmarketvolatility.Whileanyforecasting-assessmentbasedontheregressionsabovearemereindications basedonin-sampleevidence,inthissectionItrytoassesswhethernewshastrueout-of-samplepredictiveabilities.Since Ididnotfindvolatilityto(Granger)causenewsvolumeIfocussolelyonvolatilityprediction.Idothisthrougharolling windowestimationoftheOLSregressionparametersintheprevioussectionwhereonlypastinformationisusedtopredict future(one-monthahead)volatility.Thesampleisdividedintotwo(essentially)equallylongperiods,oneestimationperiod andoneevaluationperiod.12Thevolatilityisthenforecastedintwoways;(i)basedsolelyonpastvolatility(Withoutnews) or(ii)basedonpastvolatilitytogetherwithpastnewsvolume(Withnews).
Theforecastingperformance isassessedusingtwodifferentlossfunctions;themeanabsoluteerror(MAE)andthe quasi-likelihood(QL)lossfunction.Thetwolossfunctionsaredefinedas
MAE=
realized,t+1−forecast,t (1)and QL= realized,t+1 forecast,t −log
realized,t+1 forecast,t −1 (2)whereforecast,tistheforecastatmonthtofthevolatilityinmontht+1usinginformationavailableuptoandincluding montht.
Forecastingresults,forlevelsaswellasforchanges,arepresentedinTables5and6.13WhileTable5presentresultsfor English-languagenews,Table6presentsthesameresultsfornewsinChinese.Thesmallest(best)forecastingerror/lossis typedinboldandmoreoftenthannot,theinclusionofnews(Withnews)inthepredictionimprovesthevolatilityforecast. ForEnglishnewstheforecastisimprovedforeachandeverystockindexwhentheamountofnewsisacknowledged.This strongresultholdsbothforlevelsandforchanges.Meanwhile,theinclusionofChinese-languagenewsislessusefulfor
11Whennolaggedvolatilityisincludedintheregression,unreportedregressionsshowthattheaverageimpactofaone-standarddeviationchangein newsvolumeonnextmonth’sstockindexvolatilityis28basispointsandtheaverageimpactforchangesis16basispoints.
12Somewhatarbitrarily,IhavechosentolettheevaluationperiodfortheEnglishnews-basedforecastingstartatthesamedateastheChinesenews collectionstarts,i.e.November1,2010.Thisleaves51monthsintheevaluationperiodand54monthsintheestimationperiod.TheChinesenews-based forecastingstartsonOctober29,2012whichleaves25monthsintheevaluationperiodand26monthsintheestimationperiod.
Table5
ResultsofVolatilityForecastsusingEnglish-LanguageNewsVolumesInthisTableIpresentstockreturnvolatilityforecastingresultsforEnglish-language newsvolumes.Intotal,51monthlyforecastsaremadeineachcategoryandallforecastsarebasedonregressionparametersestimatedusingarolling
windowof54monthlyobservationsstartingfromNovember1,2010toSeptember1,2014.Thenumbersintherowslabelled(Withoutnews)arethe
oneswhentheone-monthlaggednewsvolumeisnotincludedinthepredictionandthenumbersintherowslabelled(Withnews)aretheoneswhenthe
one-monthlaggednewsvolumeisincludedinadditiontoone-monthlaggedstockreturnvolatility.MAEisthemeanabsoluteerrorandQListheQLloss function.Thesmallest(best)forecastingerror/lossistypedinboldandtheresultsarepresentedbothforlevelsandforchanges.
NewsinEnglish(Levels)
Stockmarket Stockmarketcrash
MAE(×10−5) QL MAE(×10−5) QL
MSCI Withoutnews 2.13 0.0244 2.13 0.0244
Withnews 1.80 0.0183 1.56 0.0142
S&P500 Withoutnews 2.91 0.0322 2.91 0.0322
Withnews 2.42 0.0235 1.77 0.0135
DJIA Withoutnews 2.45 0.0249 2.45 0.0249
Withnews 2.13 0.0195 1.40 0.0093
NASDAQ Withoutnews 4.37 0.0646 4.37 0.0646
Withnews 3.74 0.0505 3.51 0.0456
Russell2000 Withoutnews 4.69 0.0470 4.69 0.0470
Withnews 4.00 0.0363 2.88 0.0207
FTSE100 Withoutnews 2.46 0.0247 2.46 0.0247
Withnews 1.88 0.0154 1.66 0.0123
ShanghaiA Withoutnews 3.78 0.0225 3.78 0.0225
Withnews 2.82 0.0134 2.95 0.0146
ShanghaiB Withoutnews 6.37 0.0650 6.37 0.0650
Withnews 4.83 0.0418 5.59 0.0529
ShenzhenA Withoutnews 6.90 0.0483 6.90 0.0483
Withnews 6.09 0.0393 6.39 0.0426
ShenzhenB Withoutnews 6.46 0.1678 6.46 0.1678
Withnews 5.96 0.1505 5.91 0.1487
HongKongH Withoutnews 2.81 0.0111 2.81 0.0111
Withnews 2.39 0.0082 1.38 0.0029
HongKongRedChip Withoutnews 2.96 0.0166 2.96 0.0166
Withnews 2.38 0.0112 2.02 0.0084
NewsinEnglish(changes)
Stockmarket Stockmarketcrash
MAE(×10−3) QL MAE(×10−3) QL
MSCI Withoutnews 1.24 0.1056 1.24 0.1056
Withnews 1.15 0.0863 0.77 0.0320
S&P500 Withoutnews 1.91 0.2377 1.91 0.2377
Withnews 1.81 0.2019 1.08 0.0507
DJIA Withoutnews 1.95 0.2602 1.95 0.2602
Withnews 1.85 0.2202 1.21 0.0679
NASDAQ Withoutnews 3.92 0.3364 3.92 0.3364
Withnews 3.83 0.3126 3.66 0.2711
Russell2000 Withoutnews 3.25 0.3327 3.25 0.3327
Withnews 3.16 0.3038 2.68 0.1841
FTSE100 Withoutnews 1.08 0.0901 1.08 0.0901
Withnews 1.04 0.0814 0.39 0.0083
ShanghaiA Withoutnews 0.69 0.0865 0.69 0.0865
Withnews 0.66 0.0749 0.35 0.0164
ShanghaiB Withoutnews 1.14 0.1074 1.14 0.1074
Withnews 1.01 0.0774 0.96 0.0676
ShenzhenA Withoutnews 2.78 0.7350 2.78 0.7350
Withnews 2.66 0.6132 2.62 0.5785
ShenzhenB Withoutnews 4.83 0.2569 4.83 0.2569
Withnews 4.78 0.2481 4.70 0.2362
HongKongH Withoutnews 0.96 0.0767 0.96 0.0767
Withnews 0.91 0.0665 0.37 0.0083
HongKongRedChip Withoutnews 1.14 0.1092 1.14 0.1092
Table6
ResultsofVolatilityForecastsusingChinese-LanguageNewsVolumesInthisTableIpresentstockreturnvolatilityforecastingresultsforChinese-language newsvolumes.Intotal,25monthlyforecastsaremadeineachcategoryandallforecastsarebasedonregressionparametersestimatedusingarolling windowof26monthlyobservationsstartingfromOctober29,2012toSeptember1,2014.Thenumbersintherowslabelled(Withoutnews)aretheones whentheone-monthlaggednewsvolumeisnotincludedinthepredictionandthenumbersintherowslabelled(Withnews)aretheoneswhenthe one-monthlaggednewsvolumeisincludedinadditiontoone-monthlaggedstockreturnvolatility.MAEisthemeanabsoluteerrorandQListheQLloss function.Thesmallest(best)forecastingerror/lossistypedinboldandtheresultsarepresentedbothforlevelsandforchanges.
NewsinChinese(Levels)
(stockmarket) (stockmarket “collapse”)
(stockmarket “crash”)
MAE(×10−5) QL MAE(×10−5) QL MAE(×10−5) QL
MSCI Withoutnews 5.82 0.0400 5.82 0.0400 5.82 0.0400
Withnews 4.91 0.0301 5.47 0.0361 4.71 0.0280
S&P500 Withoutnews 7.49 0.0473 7.49 0.0473 7.49 0.0473
Withnews 6.82 0.0405 6.56 0.0380 7.73 0.0497
DJIA Withoutnews 5.90 0.0329 5.90 0.0329 5.90 0.0329
Withnews 5.31 0.0276 5.30 0.0275 6.35 0.0372
NASDAQ Withoutnews 13.6 0.1210 13.6 0.1210 13.6 0.1210
Withnews 12.7 0.1104 12.7 0.110 14.9 0.1375
Russell2000 Withoutnews 13.7 0.0823 13.7 0.0823 13.7 0.0823
Withnews 14.6 0.0905 13.3 0.0787 13.1 0.0769
FTSE100 Withoutnews 5.90 0.0325 5.90 0.0325 5.90 0.0325
Withnews 4.59 0.0212 6.53 0.0386 3.06 0.0103
ShanghaiA Withoutnews 7.61 0.0221 7.61 0.0221 7.61 0.0221
Withnews 7.70 0.0225 7.04 0.0192 7.87 0.0234
ShanghaiB Withoutnews 7.55 0.0267 7.55 0.0267 7.55 0.0267
Withnews 7.72 0.0277 7.98 0.0294 6.44 0.0203
ShenzhenA Withoutnews 10.9 0.0318 10.9 0.0318 10.9 0.0318
Withnews 11.2 0.0330 11.0 0.0322 8.70 0.0215
ShenzhenB Withoutnews 9.10 0.0998 9.10 0.0998 9.10 0.0998
Withnews 8.42 0.0889 9.17 0.1009 4.79 0.0898
HongKongH Withoutnews 10.6 0.0326 10.6 0.0326 10.6 0.0326
Withnews 9.64 0.0276 10.1 0.0297 10.6 0.0322
HongKongRedChip Withoutnews 7.65 0.0252 7.65 0.0252 7.65 0.0252
Withnews 8.08 0.0276 6.69 0.0199 4.94 0.0116
NewsinChinese(Changes)
(stockmarket) (stockmarket “collapse”)
(stockmarket “crash”)
MAE(×10−3) QL MAE(×10−3) QL MAE(×10−3) QL
MSCI Withoutnews 1.99 0.0571 1.99 0.0571 1.99 0.0571
Withnews 1.86 0.0480 2.52 0.1043 1.77 0.0425
S&P500 Withoutnews 5.19 0.6378 5.19 0.6378 5.19 0.6378
Withnews 5.17 0.6245 5.23 0.6542 6.10 1.3033
DJIA Withoutnews 5.21 0.6697 5.21 0.6697 5.21 0.6697
Withnews 5.15 0.6381 5.35 0.7447 6.02 1.2770
NASDAQ Withoutnews 13.0 3.2917 13.0 3.2917 13.0 3.2917
Withnews 13.0 3.2835 12.9 3.1819 14.0 7.0598
Russell2000 Withoutnews 9.88 1.8001 9.88 1.8001 9.88 1.8001
Withnews 9.93 1.8603 9.50 1.4474 9.21 1.2339
FTSE100 Withoutnews 2.28 0.0983 2.28 0.0983 2.28 0.0983
Withnews 2.22 0.0919 3.06 0.2260 0.50 0.0031
ShanghaiA Withoutnews 3.17 1.2796 3.17 1.2796 3.17 1.2796
Withnews 3.13 1.2003 3.65 2.9913 4.23 19.926
ShanghaiB Withoutnews 0.91 0.0114 0.91 0.0114 0.91 0.0114
Withnews 0.56 0.0040 1.60 0.0416 0.62 0.0050
ShenzhenA Withoutnews 6.06 0.9781 6.06 0.9781 6.06 0.9781
Withnews 6.74 1.6757 6.57 1.4629 5.36 0.5843
ShenzhenB Withoutnews 8.53 0.1660 8.53 0.1660 8.53 0.1660
Withnews 9.25 0.2118 9.01 0.1955 6.46 0.0783
HongKongH Withoutnews 3.06 0.2712 3.06 0.2712 3.06 0.2712
Withnews 3.35 0.3636 3.54 0.4415 3.47 0.4124
HongKongRedChip Withoutnews 2.52 0.1358 2.52 0.1358 2.52 0.1358
Table7
CorrelationsbetweenNewsVolumeandVolatility:RobustnessAsarobustnesstest,inthisTableIpresentNewsVolume–Volatilitycorrelationsforthe MSCIWorldstockindexwhennewsvolumesarecollectedbyGoogleNewsinEnglishandChinese,respectively.InthefirstpartoftheTableItestthe robustnessoftheresultsinthepapertotheremovalofextremeobservations,crisisobservationsandmissingobservationsandinthesecondpartItest therobustnessoftheresultstoslightchangestothenewsvolumecollectionprocess.Resultsarepresentedbothforlevelsandforchanges.
NewsinEnglish(Levels)
Stockmarket Stockmarketcrash
MSCI
ExcludingExtremes 0.18** 0.35***
ExcludingCrisis −0.04 0.53***
ExcludingMissing 0.47*** 0.79***
NewsinEnglish(Changes)
Stockmarket Stockmarketcrash
MSCI
ExcludingExtremes 0.20** 0.50***
ExcludingCrisis 0.11 0.50***
ExcludingMissing 0.18** 0.72***
NewsinChinese(Levels)
(stockmarket) (stockmarket“collapse”) (stockmarket“crash”)
MSCI ExcludingExtremes 0.44*** 0.57*** 0.07
MSCI ExcludingMissing 0.46*** 0.77*** 0.26**
NewsinChinese(Changes)
(stockmarket) (stockmarket“collapse”) (stockmarket“crash”)
MSCI ExcludingExtremes 0.05 0.27** 0.24**
MSCI ExcludingMissing 0.06 0.65** 0.40***
Levels
“stockmarketcrash” Globalstockmarketcrash
(stockmarket“slump”)
MSCI 0.67*** 0.71*** 0.54*** S&P500 0.66*** 0.69*** 0.49*** DJIA 0.68*** 0.71*** 0.51*** NASDAQ 0.64*** 0.66*** 0.46*** Russell2000 0.56*** 0.60*** 0.41*** FTSE100 0.71*** 0.75*** 0.55*** ShanghaiA 0.29*** 0.30*** 0.30** ShanghaiB 0.31*** 0.32*** 0.33*** ShenzhenA 0.21** 0.23*** 0.27** ShenzhenB 0.31*** 0.33*** 0.40*** HongKongH 0.69*** 0.72*** 0.38***
HongKongRedChip 0.65*** 0.69*** 0.47***
Changes
“stockmarketcrash” Globalstockmarketcrash
(stockmarket“slump”)
MSCI 0.60*** 0.73*** 0.48*** S&P500 0.51*** 0.65*** 0.41*** DJIA 0.50*** 0.65*** 0.42*** NASDAQ 0.49*** 0.61*** 0.37*** Russell2000 0.48*** 0.59*** 0.34*** FTSE100 0.50*** 0.68*** 0.43*** ShanghaiA 0.11 0.16* 0.26** ShanghaiB 0.18** 0.20** 0.24*** ShenzhenA 0.03 0.07 0.25** ShenzhenB 0.18** 0.25*** 0.33*** HongKongH 0.43*** 0.61*** 0.35***