BIS-中央銀行家的演講-加布里埃爾·馬克盧夫介紹性聲明-財(cái)政、公共支出和改革聯(lián)合委員會(huì)和Taoiseach

AlexanderAl-Haschimi,ApostolosApostolou,AndresAzqueta-Gavaldon,MartinoRiccingPaperSeriesUsingmachinelearningtomeasurefinancialriskinChinauaryDisclaimer:ThispapershouldnotbereportedasrepresentingtheviewsoftheEuropeanCentralBank(ECB).TheviewsexpressedarethoseoftheauthorsanddonotnecessarilyreflectthoseoftheECB.ECBWorkingPaperSeriesNoJanuary1AbstractWedevelopameasureofoverallfinancialriskinChinabyapplyingmachinelearningtechniquestotextualdata.Apre-definedsetofrelevantnewspaperarticlesisfirstselectedusingaspecificconstellationofrisk-relatedkeywords.Then,weemploytopicalmodellingbasedonanunsupervisedmachinelearningalgorithmtodecomposefinancialriskintoitsthematicdrivers.TheresultingaggregatedindicatorcanidentifymajorepisodesofoverallheightenedfinancialrisksinChina,whichcannotbeconsistentlycapturedusingfinancialdata.Finally,astructuralVARframeworkisemployedtoshowthatshockstothevariablesinChinaandabroad.KeywordsChinafinancialrisk;textualanalysis;machinelearning;topicmodelling;LDA.JEL-Classi?cation:C32,C65,E32,F44,G15.ECBWorkingPaperSeriesNo2767/January20232Non-technicalsummarySincetheGlobalFinancialCrisis(GFC),financialrisksinChinahavebeenaccumulating.Credittothenonfinancialsector,asapercentageofGDP,hasincreasedatamorerapidpacethaninanyothermajoremergingmarket,anditsoaredabovethatofmajoradvancedeconomies.Risingdebtlevelshavebeenobservedbothinthehouseholdandinthecorporatesectorwithleveragebeingparticularlyhighintherealestatesector.Risksalsoarisefromover-relianceonshort-termfundinginsomesectors,theopacityoftheshadowbankingsector,andasinternationaltradedisputes,theCOVID-19pandemic,geopoliticaltensions,policyuncertaintyregardingthefuturegrowthmodelforChina,andregulatorychangesinkeysectorsfurthercomplicatethelandscapeforfinancialmarketparticipants.Inthisenvironment,materialisationofrisksinChina’sfinancialsystemrecursperiodically.Recentexamplesaretheequitypricecollapseexperiencedin2015,theinsolvencyproblemsandsubsequentdefaultofBaoshangBankin2019-2020,thedefaultsofseveralmajorrealestatedevelopersin2021,andastockmarketaloutflowsfromChinainearlyThislastepisodefollowedanincreaseinregulationofthefast-growingtechnologysectorandrisinggeopoliticalrisksassociatedwiththewarinUkraine.Theseolatilitywhilethepotentialtriggersofriskarediverse.onomyremainsachallengingtask.Theriseoftheshadowbankingsectorandtheentryoftechnologypliesthatthesourceofriskisrapidlychanging,asarefinancialregulationsandfinancialreportingrequirements.Theavailabilityofconsistentfinancialdatathathassufficienthistorytoallowforstandardtimeseriesanalysisisoftenlacking,thusposingchallengestoquantifyingthevariousaspectsoffinancialrisksinthesystemusingtraditionalmeasures.opingameasureoffinancialriskbyapplyingmachinelearningtechniquestoalargewspaperarticlesSpecificallywerelyontextbasedanalysistoidentifymajorepisodesoffinancialrisksinChina,quantifythemanddisentanglethedifferentsourcesoffinancialrisks.WedothisusingtheLatentDirichletAllocation(LDA)algorithm,amachinelearningtechniquewhichallowsfortopicmodelling.Subsequently,weuseastructuralvectorautoregressive(SVAR)model,ascommoncialriskroECBWorkingPaperSeriesNoJanuary20233andfinancialvariables.1IntroductionFinancialrisksinChinaareincreasing.Therealestatesectorisunderseverestressfollowingastringofdefaultsbymajordevelopers,includingEvergrande.Thetensionsintherealestatesectorhadbeenrisingovertimeasleveragereachedlevelswhichpressuredregulatorstosignificantlytightenaccesstocredit,therebypromptingaliquiditysqueezeamongdevelopers.Morebroadly,financialrisksinChinahavebeenaccumulatingsincetheGlobalFinancialCrisis(GFC).CredittothenonfinancialsectorasapercentageofGDP,hasincreasedatamorerapidpacethaninanyothermajoremergingmarket,anditsoaredabovethatofmajoradvancedeconomies(Figure1a).Risingdebtlevelshavebeenobservedbothinthehouseholdandinthecorporatesectorwithleveragebeingparticularlyhighintherealestatesalsoarisefromoverrelianceonshorttermfundinginsomesectors,theopacityoftheshadowbankingsector,andhiddenrisksintraditionalbankingoperations.1Moreover,additionalheadwindssuchasinternationaltradethefuturegrowthmodelforChina,andregulatorychangesinkeysectorsfurthercomplicatethelandscapeforfinancialmarketparticipants.Inthisenvironment,materialisationofrisksinChina’sfinancialsystemrecursperiodically.In2015,equitypricescollapsedanderasedUSD3trillioninshareholdervalue(SongandXiong,2018).In2019,theinsolvencyproblemsandmarket.In2021,excessiveleverageintherealestatesector,incombinationwitheralmajorrealestatedevelopers,causingsignificantvolatilityanduncertaintyinChinesefinancialmarkets.In2022,aChinesestockmarketcorrectioncoincidedwithunusuallyhighcapitaloutflowsfromChina,followingincreasedregulationofthefast-growingtechnologysectorandrisinggeopoliticalrisksassociatedwiththewarlmarketsinChinaremainsubjecttovolatility,whilethepotentialtriggersofriskarediverse.Atthesametime,monitoringfinancialriskintheworld’ssecond-largesteconomyremainsachallengingtask.China’sfinancialsectorischangingrapidly:Whileitislargelyunderthepurviewofthepublicsectorandrelativelyclosedtointernationalprivatecapital,itisundergoingmanystructuralchanges,fromthepartialliberalisationofthecapitalaccounttoevolvingmonetaryandregulatoryregimes,1Shadowbankingreferstonon-bankfinancialinstitutionsprovidingbank-likefinancialservicesandinparticularcreditinstruments.ECBWorkingPaperSeriesNo2767/January20234andtheriseofnewfinancialmarketparticipants(SongandXiong,2018).Technologycompanies,inparticular,haveassumedasignificantlygreaterroleincatchuptomitigatenewriskscreatedbynewmarketparticipants.Figure1SourceBISandChinaStatisticsYearbook,NationalBureauofStatistics(NBS)ofChinaviaHaverAnalytics.Notes:Credittononfinancialcorporationisfromallsectorsandatmarketvalue,namelythepriceatwhichtheassetwouldchangehandsifsoldontheopenmarket.Itincludesbothcredittohouseholdsandtononfinancialcorporations.Inthispaper,weaugmenttheoftensparseavailableconventionaldataonfinancialngECBWorkingPaperSeriesNoJanuary20235techniquestoalargenumberofnewspaperarticles.Specifically,werelyontext-basedanalysistoidentifymajorepisodesoffinancialrisksinChina,quantifythemanddisentanglethedifferentsourcesoffinancialrisks.WedothisusingtheLatentwsfortopicmodelling.Subsequently,weuseastructuralvectorautoregressive(SVAR)model,ascommonintheliteratureonriskanduncertainty,toquantifytheimpactofrisingfinancialriskontheChineseandtheglobaleconomy.WefindthatanincreaseinthefinancialriskindexhasastatisticallysignificantnegativeimpactonbothChineseandglobalmacroandfinancialvariables.Therestofthepaperisstructuredasfollows.Section2providesaliteraturereviewanddiscussestheidentificationof?nancialrisksusingtext-basedanalysis;Sectionandvalidationofthefinancialriskindices;Section5discussestheSVARanalysesofthelinkbetweenfinancialriskandeconomicactivity;andSection6concludes.China’sgrowthslowdownaftertheGFCandthesubsequentsurgeincreditlevelshaveinducedconditionsunderwhichthefinancialsystemisshowingincreasingstrains.Torevitalisetheeconomy,thegovernmentimplementedamajorstimulusplanfocusingonlargeinvestmentprojectsfinancedinsignificantpartbycredit.Toprovisionofloans.By2010,acapontheceilingoftheloan-to-depositratioforcedthesebankstoincreasedeposits.Facingincreasedcompetitionfordeposits,smallerandmedium-sizebanks,startedtoissueso-calledWealthManagementProducts(WMP),whichareoff-balancesheetitemswithshortmaturitiesthataresubstitutesfordeposits(Acharyaetal.,2020).Acharyaetal.(2020)findthatWMPsincreasedriskintheChinesebankingsystem,andpresentempiricalevidencethatstockmarketinvestorspriceinrolloverrisksfromWMPsduringtimesofstressedfundingconditions.Sun(2019)findsthatshadowbankingactivitiesimpairtheeffectivenessofbankingregulationandcontributetotheaccumulationofsystemicriskinChina’sfinancialsystem.Non-bankfinancialactorsincreasedinprominenceovertime,inpartbymanagingtocircumventregulationsandspecificallylimitsonleverage.Forinstance,theFintechsectorcreatedshadow-financedaccountsenablingretailinvestorstotradeintheChinesestockmarketwithhigherleveragethanallowedusingregulatedbrokerageaccounts.In2015,regulationsbegantotightenonshadow-financedequitytrading,andexcessleverageinducedfire-salesthatcontributedtothemajorstockmarketcorrectioninmid-2015(Bianetal.,2018).ECBWorkingPaperSeriesNo2767/January20236dastringofbankfailureswiththePeoplesBankofChina(PBC)andthebankingregulatortakingcontrolofBaoshangBankinMay2019,citingseverecreditrisks.BaoshangBankrepresentedthemosthigh-profilebankfailurein20years.InJulyandAugustofthatyear,twomorebanksfailed,promptingthePBCtoconductliquidityinjectionstoavoidanescalationofsystemicrisk(Lo,2019).Theincidenthighlightedasourceofriskintheformofimplicitguarantees.Manyinvestorsdonotadequatelyaccountforrisksintheirinvestmentdecisions,believingthatthestatewillbailoutbanks,evenintheabsenceofformalguaranteeagreements.AsimilarconcernregardingimplicitguaranteesaffectsWMPsusedtoinvestintherealestatesector.Over40%ofoutstandingWMPshavematuritiesofthreemonthsorless,whilebeingusedasasourceforlonger-termlending.Thistypeofshort-termfundingforlonger-terminvestmentsexposestheseproductstoliquidityandrolloverrisks(Apostolouetal.,2021).However,investorsarenotadequatelyinternalisingtheserisksasinvestorsperceiveWMPsasbeingimplicitlycoveredbyguaranteesfromtheissuingbankorthegovernment,evenetal.,2019).In2021,excessiveleverageledtostressintherealestatesector.Precedingtherealestateturmoil,governmentregulatoryeffortshadimposedlimitsonleverageforationsintroducedasetofthresholdsforfinancialratios,theso-calledthreeredlines,whichwhenexceeded,wouldlimittheabilityofdeveloperstoraisedebt.Subsequently,liquidityinthesectordriedup,andseveraldevelopers,andmostnotablyEvergrande,defaultedontheirdebtin21-22.Again,thePBCintervenedwithliquidityinjectionstopreventthespilloverofriskstoothersectors.Nevertheless,thevolatilityintherealestatesector,andresultingslowdowninthehousingsector,weighedoninvestorsentimentandofsourcesofunderlyingfinancialrisksinChina’seconomy,relatingtoexcessiveleverage,thecircumventionofleveragelimitsviaoff-balancesheetfundingsources,andthemispricingofriskduetotheassumptionofimplicitguarantees,amongothers.Overall,thenatureofarapidlydevelopingfinancialsectorinChinapresentschallengesformacro-andmicro-prudentialpoliciestokeeppacewiththeongoingfinancialmarketsmostrecentlybylargeemergingfintechcompaniesAssuch,developingbroadmeasuresthataimtoquantifyawidevarietyofrisksinthefinancialsystemcanbeausefuladditionalmonitoringtool.Oneapproachthatcancaptureabroadmeasureofriskisbasedontextualanalysis.usedineconomicsinanumberofcontextsandGentzkowetal.(2019)providearichreviewoftheliterature.TextualanalysiscanbeconductedECBWorkingPaperSeriesNoJanuary20237arneredwideattentionwiththeEconomicPolicyUncertaintyindexbyBakeretal.(2016),entailscountingthenumberofnewspaperarticlesthatcontainasetofwordslinkedtoaspecifictopic.Resultshaveshownthatthecounting-methodproducesafairlyaccurateindexthatcanidentifyepisodesofincreaseduncertaintyabouteconomicUS.2Asimilarapproachhasrecentlybeenemployedtoconstructariskindicatorrelatingtogeopoliticaldevelopments.ThegeopoliticalriskindexbyCaldaraandIacoviello(2022)countsnewspaperarticlescitingwordsindifferentcategorieswarandterroristorganisations,amongothers.Theauthorsshowthattheirindexappearstocapturewellglobalepisodesofelevatedgeopoliticalrisks.bigdataliteratureAmongthemmachinelearningtoolsareusedtoanalysetextwithsupervisedandunsupervisedmachinelearningalgorithmsandneuralnetworks.Insupervisedmachinelearning,aso-calledtrainingsampleoftextdataisfirstclassi?edbyresearcherswhowould,forexample,notewhetherthesentimentofagivenarticleispositiveornegative.Subsequently,thealgorithmlearnstodistinguishbetweenarticleswithpositiveandnegativesentimentusingthistrainingsampleofclassi?edarticlesasaguide.trainingsamples.Onewell-knownapplicationofunsupervisedlearningistopicmodelling,tanceontheLatentDirichletAllocationLDAalgorithmasinBleietal(2003).Themethodallowsresearcherstoidentifythemaintopicsdiscussedinabodyoftextandtrackhowthefrequencyofthesetopicschangesovertime.Forexample,astudyofNorwegiannewsarticlesfoundthatthefrequencyofcertaintopicsdiscussingfinancialmarketdevelopments,andinparticularcreditandborrowing,canhavepredictivepowerfortheevolutionofkeyquarterlyeconomicvariablesandassetprices(LarsenandThorsrud,2019).Whileapplicationsoftextualanalysistomeasure?nancialriskareratherlimited,thereareafewpapersthatspecificallylinktext-basedindicatorstofinancialvariablesPikorecetaldevelopedanewscohesivenessindex,whichreflectsthesimilarityofwordfrequenciesacrossdifferentarticlesandfindthattheirvolatilityindicessuchastheVIX.Ormerodetal.(2015)developsatext-basedindexicatinganxietyrelatedsentimentexpressedintheThompson-Reutersnewsfeed.Theresultingindicatorisshowntohavepredictivepowerforone-quarteraheadUSGDPgrowth.TheauthorsalsofindthatthissentimentindexGrangercausesthefinancialstressindicesofthe2ThemeasurementandimpactofeconomicpolicyuncertaintyinChinahasbeenstudiedinsomedetail;see,e.g.,Davisetal.(2019),Heetal.(2020),HuangandLuk(2020),LiandWu(2020),LiuandZhang(2020),andShaetal.(2020),amongothers.ECBWorkingPaperSeriesNoJanuary20238ClevelandandStLouisFederalReserveBanks.Thefollowingsectionsdescribethedataandsourcesusedintheanalysis,aswellasriskmeasureinourcollectionofnewspaperarticles.Inordertoobtainadatasourcethatisconsistentoveranextendedtimeperiod,thetheSouthChinaMorningPost(SCMP),aHongKong-basednewspaperthatcoversmainlandChinaextensively,viatheDowJonesFactivadatabase.ThetwonewspapersprovideauthoritativecoverageofdevelopmentsinChinaandhavesufficienthistoryinthedatabaseusedtoaccessthearticles.Therearetwomainconcernsaroundtheconstructionofatext-basedindicatorontheChineseeconomyusingnewspaperarticles.Thefirstisrelatedtothepotentialmanipulationofinformationbyauthoritiesatthesource.Thisproblem,whichabstractsfromthedegreeofindependenceofagivenmediaoutlet,limitsthescopeofouranalysisasimportantnewscouldbecensoredordistortedbeforereachingthepublic.ThesecondrelatestothedirectinfluenceChineseauthoritiesexertonsomemedia.theimpactofcensorshipofmainlandChinesenewsmedia.Atthesametime,inHongKongbasedSMCPmayincreasinglyfaceeditorialpressuresToedriskindicator.Uponcloserinspection,webelievethatourindicatoris,byconstruction,relativelyrobusttoofficialorself-censorship.First,therearestudieswhichconstructedriskoruncertaintymeasuresforChinabasedontextualanalysis(e.gDavisetal.,2019;HuangandLuk,2020)usingmainlandnewspapers.This,presumably,exposesthemtoissuesconcerningeditorialindependence,reliabilityandcensorshipgiventhetightgripoftheChineseCommunistPartyonmedia.3However,thoseauthorsarguethatcensorshiphasnoqualitativeimpactontheirindex(HuangandLuk,2020).Second,weverifiedthatcomparingthefrequencyoftheriskincidenceinthenewsreportingisqualitativelythesamewhethertheSMCPisincludedinthesampleornot.Onereasonforthisisthatouranalyticalapproachdoesnotrelyonsentimentanalysis,wherebytheindexreactstotheintensityofpositiveornegativereporting,whichmaybeaffectedbyoverlyoptimisticcoverage3AccordingtoReportersWithoutBorders,aninternationalnon-profitorganizationwhichseekstodefendandpromotefreedomofinformation,Chinaisranked175thinover180countriesintheworldin2022forthefreedomofitspress.ECBWorkingPaperSeriesNoJanuary20239resultingfromcensorship.Instead,theriskindexreflectsthefrequencywithwhichcertainsubjectsarereported,whichisametricthatislesssensitivetothewayspecificnewsarecovered.Ourapproachthereforefollowsthatofrelatedstudies,includingChenandTillmann(2021),whoconducttextualanalysisonChinausingforeignmediacoverage.Thechoiceofprintoveronlinenewspapereditionsisdesignedinsteadtoguardagainstupwardtrendsintheamountofnewscoverageavailablethroughonlinemedia,whiletheeditorialprocessofprintnewspapersalsoreducesduplicationorupdatesofexistingarticles.Theuseofprintmediahasthecostofpotentiallyreducingthesamplesizeofavailablenewsarticlesgiventheproliferationofonlinemedia.Fortheregionalrestrictions,weemployedFactiva’sgeographicalfilters,selectingmainlandChina(i.e.,excludingHongKong).ThearticlesfilteredgeographicallyareeranaverageofaroundarticlespublishedinbothfortheSCMPthanfortheWSJ:ofthe10,000articlesonChina,around80%arepublishedintheSCMP.Toidentifyarticlesrelatingtofinancialrisks,wefilteredforasetofwordscontainedinthearticle.Toobtainthislistofwords,westartedwiththewords‘risk’and‘financial’aswellaspermutationsofthesetermssuchasrisks,riskiness,etc.Wethengeneratedalistofwordsthataresemanticallysimilarusingtheword2vecalgorithmcreatedbyMikolovetal.(2013).4Theword2vecalgorithmwasappliedtoatrainingsampleof1,000articlesfor2015,whichwasayearinwhichChinaexperiencedelevatedvolatilityinitsexchangerate,capitalflowsaswellasineconomicgrowthrates.Thealgorithmgenerated100wordsthatweremostsimilarialandjudgmentwasusedtoreducethislisttothosewordsthatwererelevantbothinmeaningandsentiment.Forthetermfinance,therelatedwordsidentifiedasrelevantcomprisebanking,lendingandcommerce.Wealsoincluded“shadowbanking”givenitsuniqueimportanceinChina’sfinancialsystem.termsislistedinTable1.5Thisrefinedsearchproducesaround2,000Chinarisk-relatedarticlesperyear.Interestingly,inthisrefinedsample,theshareofarticlesidentifiedintheWSJincreasestoaround35%(from20%ifweconsiderthegeographicalfilteringonly).ThisisexplainedbythefactthattheWSJfocusesmoreonfinancialmattersthanthemoregeneraltopicscoveredbytheSCMP.4WeusethePythonimplementationoftheword2vecalgorithm.5TableA.1.intheAppendixshowsthewordsobtainedwiththeworld2vecalgorithm.ECBWorkingPaperSeriesNoJanuary202310Table1:Finalsearchquery(chinaorchinese)and(?nanc$orbank$orlend$orcommerc$orshadowbank$)and(risk$orlever$orpanygeproblem$oropaqu$orpain$orunstableordanger$ordefault$orunwind$ordeterioat$orvulnerab$orvolatilornegativeorcontagioorimbalanceorhazardortriggerordeleverorbadoruncertain$orjeopard$orsystem$orserious$orconsequ$orshock$orbubbl$orstress$orthreat$orburden$)AllAuthorsAllSubjectsSource:DowJonesFactivadatabase.ectednumberofnewsarticles.Specifically,weusetheLatentDirichletAllocation(LDA),asdevelopedbyBleietal.(2003).Intuitively,thealgorithmstudiestheco-occurrencesofwordsacrossarticlestoframeeachtopicasadistributionofwordswithaspecificprobabilityofbelongingtoagiventopic.Eacharticle,inturn,isanunsupervisedlearningapproachinthatthealgorithmcomputesthetwolatentdistributionswithoutanypriorlabellingofthearticles.Theonlyparameterinputtothealgorithmisthegivennumberoftopics,K,thatexistwithinadocument.Thedatageneratingprocessisthereforedescribedbytwoprobabilitydistributions:topicsasadistributionofwords,andarticlesasadistributionoftopics.Themodelrecoversthesetwodistributionsbyobtainingtheparametersthatmaximisetheprobabilityofeachwordappearingineacharticle,giventhetotalnumberoftopicsKInthisrespecttheprobabilityofawordioccurringinanarticleis:P(i)=∑=1P(i|zi=j)P(zi=j)(1)whereziisalatentvariableindicatingthetopicfromwhichtheithwordwasdrawnandP(wi|zi=j)istheprobabilityofwordwibeingdrawnfromtopicj.P(zi=j)isECBWorkingPaperSeriesNoJanuary202311theprobabilityofdrawingawordfromtopicjinthecurrentarticle,whichwillvaryacrossdi?erentarticles.Intuitively,P(w|z)indicateswhichwordsareimportanttoevalenceofthosetopicswithinanarticleInitscompleteformat,theposteriordistributioncanberepresentedbyallitscomponents:p(z,K,θ|w,α,η),wherezisthetopicassignment(theprobabilityofchoosingagiventopicacrossthesetofarticles),Kisthenumberoftopics,andθisthearticle-topicdistribution.Moreover,thelistofwordsisgivenbyw,andthedistributionofwordstotopicsbeingmoreeven,whilealowlevelofηrepresentsfewerwordshavingamuchhigherprobabilityofdefiningthattopicthantherest.indicatearticlescontainingasimilartopicdistributionperarticlewhilelowlevelsofαindicateamoredispersedistribution.hodThismethodinvolvesestimatingempiricallythelikelihoodP(w|K)whichisthelikelihoodofthedata(words)foranygivenmodelspecifiedbydifferentvaluesofK.Wethenselectthemodel,andthusaitrequiressummingoverallpossibleassignmentsofwordstotopicsbutcanbeapproximatedusingtheharmonicmeanofasetofvaluesofP(w|z,K),whenzissampledfromtheposteriordistribution(Gri?thsandSteyvers,2004).WesampledomputedtheloglikelihoodscoregroupsoftopicsrangingfromN…,80.Figure2showstheresultingscoresandthataccordingtothisprocedure30istheoptimalnumberoftopicsforbothsamplesubsets.Specifically,increasingKthelikelihoodthatagivendocumentcontainsmorethan30distincttopics.-2510000-2515000-2520000-2525000-2530000-25350001020304050607080-5330000-5340000-5350000-5360000-5370000-5380000-5390000-5400000-5410000LogLikelihood1,000articlesSource:Authors’calculations.LogLikelihood2,000articles(rhs)ECBWorkingPaperSeriesNoJanuary202312Notes:Y-axesrepresentthelog-likelihoodscorelogP(w|K)for10,20,30,40,50,60,70and80topicsusingthepackagetopicmodelsinRandtheGri?thsandSteyvers,2004approach.TheLDAalgorithmdoesnotdefineatopicbeyondthewordsincludedinthedistribution.Thus,theclassificationoftopics,requiresjudgment.Asanexample,fiedtopicsFinallyaggregatingmonthlythefrequencyofeachtopicineacharticleoverallarticlesprovides30timeFigure3:RepresentativeTopicsSource:Authors’calculations.NotesThesizeofthewordsinthewordcloudsreflectthenumberofoccurrencesofthatwordinthetopicconsidered.ThefiguresalsopresentthetimeseriesrescaledbythetotalnumberofarticlespublishedonChinaeachmonthtoaccount–ascommonintextualanalysis-forresarellowthreestepsFirstweselect