X.Lietal./Knowledge-BasedSystems69(2014)14–2315
(1)Reductivedimension.Comparingwithtensofthousandsofwordsthatarecommonlyobservedwhileusingwordsasthefea-tures,thesentimentrepresentationlargelyreducesthedimensionstotheorderofhundreds,e.g.,HarvardIV-4psychologicaldictio-naryhas182dimensionsandLoughran–McDonald nancialsenti-mentdictionaryhas6dimensions;(2)Affectiveinterpretation.Itisusuallyhardtointerpretthemappinggeneratedinbag-of-wordsspace.Incontrast,sentimentscoresindifferentdimensionscansometimesgivepeoplemorestraightforwardfeelingsaboutthedocument.
Althoughdocumentsentimentanalysishasbeenproposedandemployedinmanyapplications,e.g.,recommendersystem[9],fewworkhasbeenreportedinthecrossdomainofcomputerscienceandalgorithmictrading.Inthispaper,we rstsetupagenericframeworkthatcantakemarketinformationsetsandplugindifferentcustomizedpredictionmodels.Wethenimplementthepredictionmodelsusingeithersentimentanalysisapproach,bag-of-wordsapproachorsentimentpolarityapproach,andcom-paretheirdailystockpricereturnpredictionaccuracyon veyearsofHongKongStockExchangemarketdata.Inordertoavoidanybiasesintroducedbytheframework,suchassentimentdictionary,thelabelingmethodandthecomparisonatdifferentmarketclassi- cationlevels,etc.,we(1)employtwosentimentdictionariesforcomparison;(2)discussandrigorouslytestdifferentinstancelabelingmethodsand(3)comparemodels’performanceatindivid-ualstock,sectorandindexlevelsintheexperiment.Theempiricalresultsshowthat(1)atindividualstock,sectorandindexlevels,themodelswithsentimentanalysisoutperformthebag-of-wordsmodelinbothvalidationandindependenttestingsets;(2)themodelswhichusesentimentpolaritycannotprovideusefulpredic-tions;(3)thereisaminordifferencebetweenthemodelsusingtwodifferentsentimentdictionaries.
Therestofthepaperisorganizedasfollows.InSection2,wereviewtherelatedworkinbothstockpricepredictionandsenti-mentanalysis.InSection3,weillustratethework owoftheframeworkanddoexploratoryinvestigationonthedatasetandsentimentdictionaryweuse.InSection4,weexplainthesetupoftheexperimentandshowtheparametertuningrelatedtotheframework.Andalso,wegivetheexperimentalresultsanddiscus-sions.InSection5,weprovideourconclusionandfutureworkdirections.
2.Relatedwork
Theconceptofsentimenthasbeenknownbypeopleforalongtime.Itreferstoaspeci cviewornotion[12].Sentimentanalysisreferstotheuseofnaturallanguageprocessing,textanalysisandcomputationallinguisticstoidentifyandextractsubjectiveinforma-tioninsourcematerials[16].Inthissection,wereviewtherelated
workaboutnewsimpactandsentimentanalysisinboth nanceandcomputersciencedomain.
2.1.Newssentimentanalysisin nancedomain
Thesentimentofnewsarticlesandtheirimpactsonstockpricereturnshavebeenstudiedin nancedomain.Niederhoffer[27]analyzesNewYorkTimesandclassi es20yearsofheadlinesinto19prede nedsemanticcategoriesfromextreme-badtoextreme-good.Healsoanalyzeshowthemarketsreacttothenewsofdiffer-entcategoriesand ndsthatmarketshaveatendencytooverreacttobadnews.Davisetal.[10]analyzestheeffectsofoptimisticorpessimisticlanguageusedinnewson rms’futureperformance.Theirconclusionhastwofolds:(1)thereisabiasbetweenthereaders’expectationandthewriters’intensionand(2)readersreactstronglytoboththecontentandtheaffectivesideofthereportswhichviolatetheirexpectations.Tetlock[36]extractsandquanti estheoptimismandpessimismofWallStreetJournalreports,andobservesthattradingvolumetendstoincreaseafterpessimismreportsandhighpessimismscoredreportstendtobefollowedbyadowntrendandareversionofmarketprices.Tetlocketal.alsouseHarvardIV-4psychologicaldictionaryintheirwork[37],whereonlypositiveandnegativedimensionsareexploited.TheyanalyzethefractionofnegativewordsinDowJonesNewsServiceandWallStreetJournalstoriesaboutS&P500 rmsfrom1980through2004.
2.2.Bag-of-wordsapproachincomputersciencedomain
Thebag-of-wordsapproachhasbeenappliedtonewsimpactanalysisin nancialmarketformanyyears.Seoetal.[35]buildaTextMinersystem(amulti-agentsystemforintelligentportfoliomanagement),whichcouldassesstheriskassociatedwithcompa-niesbyanalyzingnewsarticles.SchumakerandChen[34]buildAZFinTextsystemwhichisabletogivedirectionalforecastofpricesbasedon nancialnews.Bag-of-wordsapproachrepresentsthetextualnewsarticlesbytermvectorsandevaluatesthe‘‘importance’’ofeachtermastheirweights.Afterlearningthemappingfromthewordstatisticalpatternstotheoutcomelabels,theapproachmakespredictionsforfutureunseendata.2.3.Sentimentanalysisincomputersciencedomain
2.3.1.Sentimentdictionaryconstruction
Applicationsevaluateword’ssentimentmainlybyconstructingasentimentdictionary.Theconstructionapproachcouldbebrie ycategorizedassemi-automaticandmanualwhicharedescribedbelow:
Semi-automatic.Thedictionaryis rstconstructedbysomeseedwordsthataremanuallyselected.Thedictionaryisthenexpandedfromtheseedsbytherulesde nedbyapplication.Manual.Thedictionaryispurelyconstructedbylinguisticexperts.Thiskindofdictionaryisusuallysmallerinsizethantheoneconstructedsemi-automatically,butmoreaccurate.HatzivassiloglouandMcKeown[17]maketwohypotheses:(1)adjectivesthatareseparatedby‘‘and’’havethesamepolarityand(2)adjectivesthatareseparatedby‘‘but’’haveoppositepolar-ity.Theyuseseedwordsandclassifyadjectivesintopositiveandnegativegroups.Wiebe[40]alsoevaluatesadjectivesforpolarityclassi cation.Hegroupsadjectivesbyword’stoneandorientationclusters.KimandHovy[21]generatethepolaritydictionarybyusingtheWordNettoexpandtheselectedseedwords.Theymaketwoassumptions:(1)synonymshavethesamepolarityand(2)antonymshavetheoppositepolarity.Thestrengthofa

word
newspaper,.fundsandinvestmentinstitutionofthecountry...ChinabyStockDynamicAnalysis.Thesurveycan Fridayandthe...sentimentandthepriceofstockhavedynamicrelationship....
Good News or Bad News Let the Market Decide_...sentiment analysis has been the procurement of ...as having positive/negative impact on stock price...
Investor Sentiment and t... 36页 1财富值 Economic News and Stock ......Chapter 5 Analysis and Results: Empirical results 5.1 Price ?uctuations . ...
A Bayesian analysis of... 暂无评价 17页 免费 Price return auto-corr.....Baker and Stein (2002) show that trading volume is also a sentiment ...
sentiment and confidence in the market – Pre ...Stock return (% change in price) for firm i ...(minimum 3), news and analysis announcements in ...
The Bubbles of China Stock Market Based on Return Decomposition and Cumulative Return_专业资料。维普资讯 http://www.cqvip.com JunlfnhaUiest(n.E.V/2 ?...
Return threshold model analysis of two stock markets: Evidence study of Italy and Germany's st_专业资料。Jn21,ou ,.Srl.9 a.00Vlme9No1(ei 7)aNo...
Return threshold model analysis of two stock markets: Evidence study of Italy and Germany's st_专业资料。Jn21,ou ,.Srl.9 a.00Vlme9No1(ei 7)aNo...
热门文档