The Fed's Fine-Tune: Coarse Statements and Predictive Pressers

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) The Fed’s Fine-Tune: Coarse Statements and Predictive Pressers Ryan Byun, Bennett Fees, Margaret M. Jacobson, Todd B. Walker 2026-029 Please cite this paper as: Byun, Ryan, Bennett Fees, Margaret M. Jacobson, and Todd B. Walker (2026). “The Fed’s Fine-Tune: Coarse Statements and Predictive Pressers,” Finance and Economics Discussion Series 2026-029. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2026.029. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

THE FED’S FINE-TUNE: COARSE STATEMENTS AND PREDICTIVE PRESSERS * RyanByun† BennettFees‡ MargaretM.Jacobson§ ToddB.Walker¶ April2026 Abstract Centralbankcommunications,particularlyFOMCstatementsandpressconferences,playacrucialroleinshapingfinancialmarketexpectations. Usinglargelanguagemodelstoquantifycentral bankcontent,thispaperdemonstrateshowsentimentalignswithtraditionalmarket-basedmonetarypolicymeasures.Weshowthatpressconferencescorrelatewithfuturepolicytoagreaterextent thanothercommunications. WhileFOMCstatementscoarselysignalthecurrentstanceofpolicy, pressconferencesfine-tunethemessage,whichhelpsmarketparticipantsrevisetheirexpectations aboutfuturepolicy. Keywords:centralbankcommunication;largelanguagemodels;monetarypolicytransmission;empiricalmonetaryeconomics. JELCodes:E52,E58,E31,E32. *ThismaterialreflectstheviewsoftheauthorsandnotthoseoftheFederalReserveBoardofGovernors.Theauthorsthank KurtLewis,JoshGallin,DanLi,CynthiaDoniger,ConnorBrennan,BrentBundick,BenJohannsen,DonKim,KathrynHolston andstaffattendingtalksattheFederalReserveBoard,theReserveBankofAustralia,theFederalReserveBankofKansasCity, TheBureauofEconomicAnalysisforhelpfulcomments. †PukyongNationalUniversity;rwbyun@pknu.ac.kr ‡FederalReserveBoard;bennett.d.fees@frb.gov §FederalReserveBoard;Margaret.M.Jacobson@frb.gov,correspondingauthor. ¶IndianaUniversity,walkertb@iu.edu

1 INTRODUCTION Communication from central banks to the public presents a wealth of potential identified monetary policyeventstoestimatethetransmissionofmonetarypolicy. Becauseassetpricesobservedminutes beforetheseeventscontainallavailableinformation, assetpricesobservedimmediatelyafterarethus themarketreactiontomonetarynews,andtheirdifferencecanidentifyexogenousvariationinmonetary policy.WhilepreviousworkfocusedexclusivelyonthemarketreactiontoFOMCstatements,subsequent work—likeSwansonandJayawickrema(2024),NarainandSangani(2026),GordonandLunsford(2024), andJarocinskiandKaradi(2025)—hasstudiedspeeches,pressconferences,andofficialtestimoniesas typesofmonetaryevents. Unsurprisingly, market reactions can differ by type of communications, especially the press conference.1 Forexample, duringthe2022tighteningcyclethemarketreactiontothepressconferenceis sometimesoppositeoftheprecedingFOMCstatement. ThesereversalsinthemarketreactionarepuzzlingbecauseGómez-CramandGrotteria(2022)findthatthecontentofthepressconferenceissimilar tothatofthestatement,onaverage. Weexplainreversalsinthemarketreactiontothepressconferenceviacontentthatisabsentfrom thestatementandcorrelatesmorestronglywiththefuturepolicyrate—afindingwearethefirsttodocument,toourknowledge. WhileNarainandSangani(2026)documentreversalsinthemarketreaction andattributethemtovariationinthecontentofthepressconference,wetaketheiranalysisastepfurtherandshowthatthisvariationcorrelateswiththestanceoffuturepolicy.Ourfindingonthepredictive correlationsofthepressconference’scontentcomplementsthepredictivecontentofFOMCtranscripts andspeechesfoundbyCieslaketal.(2025b,2024b)andCieslaketal.(2025a,2024a). TounderstandwhythepressconferencehaspredictivecorrelationsaboveandbeyondthatofFOMC statementsandspeechesbytheChair,wecomparesentimentscoresfromaBERTlargelanguagemodel fine-tunedtoFedcommunication. WhiletheoverallsignalsofpressconferencesmatchthoseofFOMC statements,signalsfromstatementstendtoberelativelycoarser.Overthree-fourthsofsentimentscores from FOMC statements are either entirely hawkish, dovish, or neutral while only one-quarter of press conferencescoresaresostark. Thevariationincontentofthepressconference—somedovishcontent eventhoughitishawkishonnet,andviceversa—isthesourceofpredictivecorrelationswiththefuture federalfundsrateandhelpsaccountformarketreversals. Forexample,inthesummerof2023,record increasesinthefederalfundsratewereaccompaniedbyentirelyhawkishFOMCstatementsandpositive marketreactions.However,themarketreactiontothepressconferencewasoftennegative—oppositeof theFOMCstatements—withcontentthatwashawkishonnet,butcontainedsomedovishcontent. Becausethecurrentstanceofpolicyisnowlargelyanticipatedbymarketsandtherelationshipbetween the federal funds rate and sentiment is highly endogenous, we also study how high-frequency market reactions about expected future policy relate to the sentiment of federal reserve communications. Among all types ofcommunication studied, we findthat press conferences stand out ashaving thestrongestrelationshipbetweencontentandmarketreactions. Whileoverallhawkishcontentisof- 1Furthermore,manystudieshavedocumentedtheimportanceofspeechesandpressconferencefortheEuropeanUnion (Altavillaetal.,2025)andtheU.K.(Mumtazetal.,2024).

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION tenassociatedwithdovishsurprises,wecanexplaintheseseeminglyopposite-signedresponsesviathe pressconference’spredictivecorrelationswiththefuturefederalfundsrate. Essentiallymarketsrevise theirexpectationsaboutfuturepolicyinresponsetodovishcontentinpressconferencesthatisabsent fromFOMCstatements. Press conferences do not announce changes in fundamentals (like the federal funds rate) and insteadcommunicatenarrativeinformation,whereasFOMCstatementsannouncefundamentalsandnarratives. Consequently,wealsoassessmarketreactionstospeechesbytheFederalReserveChairtounderstandifothereventsthatcommunicatenarrativesbutnotchangesinfundamentalshavesimilarrelationshipsbetweencontentandmarketreactions.Asshownbyregressionanalysis,speechesdonot,on average,haveastrongrelationshipwithmarketreactions.Weinterpretthisasspeecheshavinganoisier signalrelativetowhatweobservefrompressconferences. We formalize the distinct role of press conferences in refining the statement’s message via an announcement variance ratio of trading volume and find that markets indeed perceive additional signal from press conferences after they were introduced in 2011. The perception of additional signal from pressconferenceswasnotimmediateandoccurredin2015,afewyearsaftertheirintroduction,asthe Fed prepared to raise interest rates from the effective lower bound (ELB). However, by the time Chair Powell implemented press conferences after every meeting in 2019, we detect that press conferences containmoresignalthanFOMCstatements.Takentogether,thisevidencesuggeststhatnewcommunicationtoolsevolveanditmaytaketimeforthemtobeunderstood. Giventhecentralroleofpressconferences, weconcludebyassessinghowtobestincludethemin theconstructionofhigh-frequencymonetarypolicyshocks. BecauseBrennanetal.(2025)andBundick andSmith(2020)notethatdifferenthigh-frequencyshockscanleadtodifferentestimatesofmonetary policy transmission, it is important to understand how differences arising from the inclusion of press conferencesaffectestimatesofmonetarypolicytransmission. WhileAcostaetal.(2025)andGrotteria andvanBinsbergen(2025)expandthetimewindowaroundFOMCstatementstoincludepressconferences,wefindthatthisconstructioncansometimesbenefitfromincludinglonger-termratestocapture thelonger-termforwardguidancesignaledviathepressconference. Weestimatemonetarypolicytransmissionviathevectorautoregression(VAR)ofGertlerandKaradi (2015)from1973topresent(insteadofendingin2019likeinotherwork)andfindthatpointestimates are largely aligned with existing empirical and theoretical evidence. In fact, the point estimates from theNakamuraandSteinsson(2018)monetarypolicyshockaremoststablewhenthepressconference isincludedinshockconstruction,otherwiseresponsestotheexcessbondpremiumareopposite-signed orambiguous. However, inthecaseoftheGürkaynaketal.(2005)monetarypolicyshockswherecurrentpolicyandforwardguidancearedistinctdimensions,wefindthataugmentingtheinstrumentset withlonger-termrateshelpsassurethecorrectly-signedresponseinforwardguidancewhenthepress conferenceisincludedinshockconstruction. 2

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION 2 LITERATURE REVIEW FederalReservepressconferencesarenowperceivedtobethesinglemostusefulchannelofFedcommunication,asnotedinWesselandBoocker’s(2024)survey.2 LamlaandVinogradov(2019)findthatmore consumersreceivenewsabouttheFedviathepressconference,whichcouldhelpaccountforitsgrowing importance. Marketreactionstocentralbankpressconferenceshavebeenstudiedsinceinceptionby Rosa(2016), EhrmannandFratzscher(2009a), Gómez-CramandGrotteria(2022), NarainandSangani (2026),Gorodnichenkoetal.(2023),CurtiandKazinnik(2023),Parle(2022),Altavillaetal.(2025). TounderstandwhymarketreactionstoFedpressconferencesaresometimesoppositethoseofthe FOMCstatements,weturntoaBERTlargelanguagemodeltounderstandifcontentexplainsthesedifferences.Advancesinlargelanguagemodelinghaveensuredthatnaturallanguageprocessingcancapture the nuance and soft information of central bank policy communication. Hansen and Kazinnik (2024) andDunnetal.(2024),showheightenedefficiencyofmoresophisticatedmethodslikeBERTcompared to dictionary-based predecessors such as Gati and Handlan (2023), Kypraios et al. (2024), Czudaj and Nguyen(2025),Calomirisetal.(2022),LuccaandTrebbi(2011),Apeletal.(2022),Gardneretal.(2022), andPicaultandRenault(2017),MundayandBrookes(2021),Parle(2022),CherryandTong(2023),Tobbacketal.(2017),Banerjeeetal.(2025).Becausedictionary-basedmethodsoftenrelyonkeywordmatchingorbag-of-wordsmodels,theirabilitytocapturethenuancedmeaningandcontextwithincomplex textcanbelimited. Morespecifically, thesemethodstreatwordsasisolatedunits, ignoringcrucialorderings,whichcanfundamentallyaltermeaning. Bycontrast,ourBERTmodel(DeBERTaV3;Heetal.(2020);Devlinetal.(2019);Heetal.(2023)can considerallwordsinasentencesimultaneouslywhenencodingthemeaningofwordandthusaccommodatelong-rangedependencies. Similartotransformer-basedAIchatbotslikeGPT,BERTmodelsare pretrainedonamassivecorpusoftextdataandthenfine-tunedtoaspecifictasklikesentimentanalysis ofFederalReservecommunications. WithanenhancedabilitytoanalyzeFederalReservecommunications at the paragraph or sentence level, we can quantify the stance on monetary policy conveyed in thetext. BERTmodelsareusedtostudyFederalReserveCommunicationsbyAlexopoulosetal.(2024), OsowskaandWojcik(2024),Shahetal.(2023),PfeiferandMarohl(2023),Gorodnichenkoetal.(2023), Gambacortaetal.(2024).3 AssessinghowquantifiedcontentofcentralbankcommunicationcorrelateswithcurrentandorfuturepolicyisstudiedbyDohetal.(2021,2025),AruobaandDrechsel(2026),Fischeretal.(2023),Cieslak etal.(2025b,2024b), Cieslaketal.(2025a,2024a), andothers. Ourstudyofthepredictivecorrelations ofthepressconferencecomplementsthesefindingsthatspanFedcommunicationssuchastranscripts, speeches,theBeigeBook,Tealbooks,andtheminutes.Becausepressconferencesareavailableimmedi- 2Seehttps://www.brookings.edu/articles/grading-fed-communications-a-2024-survey-of-fed-watchers/ 3OtherapplicationsoflargelanguagemodelsincludethosebyAruobaandDrechsel(2026),Dengetal.(2024),Gnanetal. (2025),Silvaetal.(2025),Schmanskietal.(2023),Fischeretal.(2023),Ahrensetal.(2025),Granzieraetal.(2025).Morespecific approachesincludeLatentDirichletAllocationtopicmodelingbyDePooter(2021)andHansenandMcMahon(2016);Hansen etal.(2017);semanticsimilaritybyEhrmannandTalmi(2020),Handlan(2022a),andAcostaandMeade(2015));topicmodeling byFraccarolietal.(2020);universalsentenceencodingbyDohetal.(2021,2025);ChatGPTbyHansenandKazinnik(2024);and neuralnetworksbyHandlan(2022b)andCurtiandKazinnik(2023);andmulti-agentLLMsimulationbyKazinnikandSinclair (2026). 3

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION atelywhiletranscriptsarereleasedwithafive-yearlag,weviewourcontributionasprovidingareal-time windowintothedeliberativecontentthattranscriptseventuallydisclose. Given the improved ability of large language models to accurately capture the content of central bankcommunication,wenextaskifourmodelcanaccountformarketreactions,specificallyreversals observedinthe2020stighteningcycle.Marketreactionsandnaturallanguageprocessingarefeaturedin Handlan(2022b),Fischeretal.(2023),Parle(2022),Gnanetal.(2025),Kypraiosetal.(2024),Shahetal. (2023),OsowskaandWojcik(2024).Morespecifically,Fischeretal.(2023)andGnanetal.(2025)demonstrate that text-based sentiment indicators can help explain errors in market policy expectations that traditionalTaylor-rulevariablesmiss. Wecomparethemarketreactionacrosstypesofcommunication (statements,pressconferences,andspeeches)andfindthatthepressconferencestandsoutforhaving thestrongestsentimentscoreandmarketreactions,particularlyinthe2022-2024tighteningepisode. Finally, we assess how including the press conference in the construction of high-frequency monetary policy shocks affects estimates of monetary policy transmission. Augmenting monetary policy shocks with additional forms of communication to estimate monetary policy transmission has been studied by Swanson and Jayawickrema (2024), Altavilla et al. (2025), Mumtaz et al. (2024), and others. Like An et al. (2025) we find that estimates of monetary policy transmission are largely stable as time windowsexpand,butshowthatincludinglong-termratesinshockconstructioncanaddadditionalinformationwhenthepressconferenceisincluded. 3 MARKET REACTIONS WedefinemarketreactionsaschangesinassetpricesaroundFederalReservecommunications. Specifically, we construct market reactions using high-frequency intraday changes in futures prices immediatelysurroundingcommunicationeventstocapturemarketparticipantsimmediatelyupdatingtheir beliefsaboutmonetarypolicy.Futurespricescorrespondtobasispointchangesinimpliedinterestrates suchthatpositivemarketreactionsindicateupwardsurprisesintheexpectedfuturepolicyrateandnegativereactionsindicatedownwardsurprises. We collect intraday futures data used in the construction of monetary policy shocks for scheduled andunscheduledfedcommunications: FOMCstatements, post-FOMCpressconferences, FederalReserveChairspeeches,andCongressionaltestimonies,from1995throughApril2025. Timingdetailsare confirmedusingSwansonandJayawickrema(2024),Gürkaynaketal.(2005),andGordonandLunsford (2024),supplementedbyFactivasearches. Mathematically,themarketreactionforanasset Aateventtimet isdefinedas: Marketreaction A,t =P A,t+∆t −P A,t−∆t , (1) whereP isthepriceofassetAattimet,andMarketreaction capturesthemarketreactionwithin A,t A,t anarrowwindowaroundthereleaseoftheFOMCstatement{+∆t,−∆t},whichistypically10minutes beforeand20minutesaftertheevent.Themarketreactionforeachtypeofcommunicationiscomputed separately,enablingustoassesstheirdistinctimpacts. 4

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION 3.1 INTRADAYFUTURES WeusethedatasetofBrennanetal.(2025)whichreliesonintradaydatafrom the CME Group Inc. DataMine (https://datamine.cmegroup.com/) at the Federal Reserve Board. AlthoughtheintradaytickdataisavailablestartinginJanuary1995,wesometimesrestrictoursampleto onethatstartsinJune1999whenFederalReservecommunicationsareavailableinformatsmorereadily usablefornaturallanguageprocessing. OurfullsampleendsinApril2025andcontains259statements afterFOMCmeetings(7ofwhichareunscheduled)85post-FOMCmeetingpressconferences,121Congressionaltestimonies,and707speechesbytheFederalReserveChair.4 SeeAppendixBformoredetails, especiallythoseondates,times,andwindowsizes. IntradayfuturesusedtoassessmarketreactionstoFOMCcommunicationsarefederalfundsfutures andEurodollar/SecuredOvernightFinancingRate(SOFR)futures. Thesefuturesrepresentthefirstyear of the term structure of interest rates and are commonly used to construct high-frequency monetary policy shocks as detailed in Brennan et al. (2025).5 Federal fund futures are monthly contracts such thatthefirstexpiringcontract,FF1,expiresattheendofthemonthofagivenFOMCcommunication. For example, for the March 19, 2014 FOMC statement FF1 expires at the end of March. The second, third, andfourthexpiringfederalfundsfuturesexpireone, two, andthreemonthsafteragivenFOMC statement,respectively. Forexample,FF4expiresattheendofJune2014fortheMarch19,2014FOMC statement. Eurodollar/SOFRfuturesarequarterlycontractswherethefirstexpiringcontractexpiresat theendofthequarterforaparticularFOMCcommunicationinthecaseofthelatter,butinthequarter followinginthecaseoftheformer. Figure1showsthedistributionofintradaychanges10minutesbeforeand20minutesafteranFOMC communicationforFOMCstatements,pressconferences,chairspeeches,andCongressionaltestimonies. Thetightnessofthedistributionsdecreasesasthematurityincreases,i.e.theshortestmaturityfutures— FF1—are the least volatile and the longest maturity futures shown—EDSR4—are the most volatile. Changesintheformerrepresentsurprisestoactualchangesinthefederalfundsratewhilechangesin thelatterrepresentsurprisestoforwardguidanceorcommunicationaboutthefuturepathofpolicy. Therearetwomainreasonsforthepositivecorrelationbetweenmaturityandvolatility.First,FOMC statementsaremuchlesslikelytobesurprisingtomarketsthaninpreviouseras. Asaresult, changes inFF1,theimmediatelyexpiringfutureandhencetheclosestmarketexpectationtothefederalfunds rate,aresmallinmagnitude.Second,communicationsotherthanFOMCstatementsareunlikelytohave surprisesinFF1orshortermaturitiesbecausethefederalfundsratecannotpossiblychangeonthese otherevents.Therefore,theseeventscanonlybeforwardguidance. Wealsonotethatmarketreactionstospeeches, asshowninpanel1caremuchmoremuted—and 4Althoughthereare707speechesbyaFederalReserveChairavailableinoursample,werestrictoursampletothe226Chair speechessince2008thatareobtainableinaformataccessibletoourmodel.SwansonandJayawickrema(2024)showthatthe chairspeechesaremorelikelytomoveintradayfuturesthanthoseofotherofficials.IncontrasttoSwansonandJayawickrema (2024)andSekkelandZhan(2024)whofiltertheirsampleofchairspeechestothosethatrelatetomonetarypolicy,wekeep allchairspeechesinoursamplefortworeasons. First,thetextofChairSpeechesistypicallyunreleasedbeforehandleaving fewwaysbesidesthetopicandlocationformarketstoknowthecontentsofaspeech. Second,textswithnomonetarypolicy contentshouldsimplyhaveasentimentscoreofzeroandnomarketreaction. 5BecauseEurodollarswerediscontinuedin2023andSOFRfuturesintroducedin2018,weswitchtoSOFRfuturesinJanuary 2022,asadvocatedbyAcostaetal.(2024).Wenotethattheswitchdatehaslittlematerialeffectontheseriesofintradaychanges inassetprices. 5

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION '' '' '' '' &%43 &%43 &%43 &%43     (a)FOMCstatements,basispoints '' '' '' '' &%43 &%43 &%43 &%43     (b)Pressconferences,basispoints '' '' '' '' &%43 &%43 &%43 &%43     (c)ChairSpeeches,basispoints '' '' '' '' &%43 &%43 &%43 &%43     (d)CongressionalTestimonies,basispoints Figure1: DistributionsofintradaychangesinfuturespricesaroundFederalReservecommunications, January1995toApril2025. Intradaychangesarecalculated10minutesbeforeand20minutesafteranFOMCcommunication. SeeAppendixBformore detailsontimewindowsizes. FFi fori =1,...,4istheithfederalfundsratefuturewherei correspondstothemonthahead, i =0isthecurrentmonth. TheithEurodollar/(i+1)thSOFRfuture(ESRi),whichcorrespondstotheexpectedpolicyratei quartersaheadfori=1,...,4. clustered closer to zero—than those for FOMC statements or press conferences. This arises because 76 percent of speeches have a market reaction that is less than 3 basis points, or quite close to zero. WhilesomespeechescanmovemarketsjustasmuchaspressconferencesorFOMCstatements,market movementsareonaveragequitesmall. 6

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION Figure2:Marketreactionstothepressconferencevs.FOMCstatements,basispoints. The figure shows the slope coefficient βˆ from the regression MarketReactiont,Press =α+βMarketReactiont,FOMC +ϵ t by Fedchairsforthe4thEurodollar/5thSOFRfuture(EDSR4),whichrepresentstheexpectedpolicyratefour-quartersahead. BernankeorYellenisApril2011toJanuary2018andPowellisfromFebruary2018toApril2025. Shadedbandsare90%error bands. Thesmallmarketreactiontospeecheshighlightshowspeechesmaybedistinctfromotherevents. WhileFOMCstatementsandpressconferencesoccuralongsidepolicyactions,likechangesinthefederal fundsrate,speechesdonot. Eventhoughallthesecommunicationscancontaininformationaboutthe futurestanceofpolicy,speechesmaynotundergothesamedegreeofmarketre-pricinggiventhatthey donotoccuralongsidechangesinfundamentals. Relatedly, themarketmayreactmoretostatements andpressconferencesbecausetheyarethefirstcommunicationtofollowa10-dayblackoutpreceding theFOMCmeetingasstudiedbyEhrmannandFratzscher(2009b). Given that market reactions to press conferences and the FOMC statements are roughly the same size,aretheyofteninthesamedirection?Figure2showsthemarketreactionofthepressconferenceregressedonthatoftheprecedingFOMCstatementforthefourthEurodollar/fifthSOFRfuture,whichrepresentstheexpectedpolicyratefourquartersahead.6 WhiletheslopeispositiveunderChairsBernanke andYellen(2011-2018),theslopeisinsignificantfromzerounderChairPowell(2018-2025).Infact,there aremanyobservationswherethepressconferencehastheoppositesigntothatoftheFOMCstatements underChairPowell. Message1. ThemarketreactiontothepressconferencesometimesreversesFOMCstatements,especially duringthe2022-2024tighteningcycle(underChairPowell). Figure3showsanexampleofamarketreversalonJune15, 2022wherethemarketreactiontothe FOMCstatementispositive,butisnegativetothepressconference. ThissuggeststhatmarketparticipantsreviseduptheirexpectationsforthefederalfundsrateayearaheadinreactiontotheFOMCstate- 6AppendixfigureB.1showsasimilarrelationshipbetweenthemarketreactiontothepressconferenceandtheFOMCstatementsbyFedchairforotherinterestratefuturesusedinconstructionofhigh-frequencymonetarypolicyshocks. 7

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION ment,butrevisedthemdowninreactiontothesubsequentpressconference. AppendixfigureB.2confirmsthatthesereversalsarenotone-offevents. Relatedly,AppendixfigureB.3showsthatthespeeches intheweekafteranFOMCstatementdidnothavethesamewide-spreadopposite-signedmarketreactionspriortotheintroductionofthepressconferencein2011. GiventhatpressconferenceshavereversedmarketreactionstoFOMCstatementsunderChairPowell,butnothispredecessors,wenextexplorewhy. WhileNarainandSangani(2026)attributethesereversalsinmarketreactionstodifferentcontentconveyedbytheQ&Aofthepressconferencerelativeto thatoftheFOMCstatement,Gómez-CramandGrotteria(2022)detectasimilaritybetweenFOMCand pressconferencepairs, albeitonasamplethatendsin2019. Weusealargelanguagemodeltoclarify thesesomewhatcontrastingfindingsbyshowingthatthecontentofpressconferencesrefinesthecoarse signalofFOMCstatements. Thisrefinedsignal, inturn, containspredictivecorrelationsforthefuture stanceofpolicytoagreaterextentthanFOMCstatements. Understandingthesourceofthesereversals isimportantforestimatingmonetarytransmissionasreversalscouldalterthesignofmonetarypolicy shocksiftheFOMCstatementandpressconferencearetreatedasasingleevent. Positive Negative surprise surprise (cid:122) (cid:125)(cid:124) (cid:123)(cid:122) (cid:125)(cid:124) (cid:123) ← PressConf. → FOMCStatement Figure3:Tradingpricesofthe5thSOFRfutureontheafternoonofJune15,2022,percent Thefigureshowsintra-daytradepricesforthe5thSOFRfuture(4-quarter-aheadinterestrateexposure)whentheFederal Reserveraisedinterestrates75basispointsonJune15,2022. ThesolidredlinerepresentstheFOMCstatementreleasewith thedashedredlinesindicatingthetradepricepulled10minutesbeforeand20minutesaftertheFOMCstatementrelease. Thesolidgreenlinerepresentsthestartofthepressconferenceandthedashedgreenlinesindicatethetradepricepulled10 minutesbeforeand20minutesaftertheconferenceends. 4 SENTIMENT AND MONETARY POLICY 4.1 LARGE LANGUAGE MODELING To quantify the content of Federal Reserve communications, we employ DeBERTa V3 (He et al., 2021, 2023), an iteration of the Bidirectional Encoder Representations fromTransformers(BERT)architectureDevlinetal.(2019)optimizedfornaturallanguageunderstand- 8

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION ing. Unlikegenerativemodels(e.g.,GPT)designedfortextsynthesis,BERT-basedarchitecturesexcelat producingcontextualizedwordembeddings,dynamicnumericalvectorsthatcapturethemeaningofa wordbasedonitsspecificusagewithinasentence. Whilestandardself-attention(Vaswanietal.,2017) collapsescontentandpositionintoasinglequery-keyproduct,DeBERTaseparatescontentandrelative position.7 Byisolatingrelativepositioningfromsemanticcontent,themodelachievesfinerdistinctions inmeaning. Usingthisadvancednaturallanguageprocessingtechnique,weanalyzethetextofFederalReserve communications at the paragraph or sentence level. By training text classification models based on BERTtorecognizepre-definedcategories,wequantifythemonetarypolicystanceandcontentrelatedto macroeconomicconditionsviasentimentscores.Weusetwodistinctclassificationapproachesbasedon BERT,whicharesubsequentlypassedthroughaclassificationheader—afullyconnectedneuralnetwork layer—tocomputetheprobabilityoftextbelongingtoapredefinedcategory. P(s∈C)=σ(BERT(s)), (2) wheresisatextsequence(sentenceorparagraph),Crepresentstheclasscategory,andσ(·)istheclassificationheader. Wefine-tunethesemodelsusingtrainingdataspecifictotwoclassificationtasks: • Sentimentclassification(BERT-Sentiment): Paragraph-levelclassificationintomonetarypolicy stances:Dovish,Hawkish,orNeutral.Theselabelspertaintothestanceofpolicy:dovishindicates loosening, hawkish indicates tightening, and neutral indicates no change. Each paragraph s is i assignedtothecategorywiththehighestpredictedprobability: s ∈CL, where L=arg max P(s ∈Cℓ ). (3) i i ℓ∈{Dove,Hawk,Neut} Thesentimentscoresareaggregatedas FedStanceDove/Hawk= 1 N (cid:88)c,t 1(s ∈CDove/Hawk), (4) c,t N i c,t i=1 whereN isthenumberofparagraphsincommunicationcattimet.Paragraphswithoutexplicit c,t policystanceorwithambiguousstatementsareclassifiedasNeutral. • Topic-basedclassification(BERT-Topic):Sentence-levelbinaryclassificationidentifyingmacroeconomic aspects influencing policy decisions: employment, inflation, financial conditions, economicgrowth,globaleconomicconditions,and“others"(includingfiscalpolicy,oilprices,etc.). 7Mathematically,theattentionweightAi,j betweenwordiandwordjisdecomposedintoasumofcomponentsrepresentingcontent-to-content,content-to-position,andposition-to-contentinteractions Ai,j =QcK c T+QcK r T+PrK c T+PrK r T whereQcandKcarecontent-basedqueryandkeymatrices,andQr andKr aretheirrelativepositioncounterparts. 9

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION Eachsentencecanbeclassifiedintomultipleoverlappingcategories,subjecttoa70%probability threshold TopicWeight T = 1 N (cid:88)c,t 1(s ∈T ), (5) c,t N i c,t i=1 whereT isthetopiccategory.Sentencesdiscussingmonetarypolicywithoutexplicitreferenceto economiccontextsareclassifiedseparately. Topic-specificsentimentscoresarecomputedas TopicStance T,Dove/Hawk= 1 N (cid:88)c,t(cid:163) 1(s ∈T )×1(s ∈CDove/Hawk) (cid:164) (6) c,t N i i c,t i=1 To determine which communication channel has the most significant impact, we analyze texts of 223 FOMC statements, 85 post-FOMC press conferences, 226 Federal Reserve Chair speeches, and 34 congressionaltestimoniesbytheFedChairfrom1999to2025. OursampleofChairspeechesandCongressional testimonies begins in 2008 due to availability and our sample of press conferences in April 2011 which is when they were first held. Footnotes, appendices, supplementary information, and administrativeaddendawereexcludedfromtheanalysis. AppendixCshowsthatrawsentimentscoresacrosstopicsandtypesofcommunications.Thevalues ofoursentimentscoresareinlinewiththoseofDunnetal.(2024)andFischeretal.(2023). Following Silvaetal.(2025),wecomputenetpolicysentimentscoresthatquantifiesthedirectionalstanceofmonetarypolicy—essentiallywhereitisandwhereitgoing.Foreachcommunicationtype j ∈{Pressconference,Speech,FOMCStatement},weaggregateparagraph-levelclassificationsfromequation(4)toobtainHawkish =FedStanceHawkandDovish =FedStanceDove. j,t j,t j,t j,t Hawkish −Dovish NetSentiment = j,t j,t (7) j,t Hawkish +Dovish j,t j,t WhenHawkish +Dovish =0(allneutral),wesetNetSentiment =0.NetSentiment ∈[−1,1]such j,t j,t j,t j,t thatavalueof-1indicatesentirelydovishsentiment,0isneutral,and1entirelyhawkish.Avalueof0.75, forexample,indicatespredominantlyhawkishsentiment,butwithsomedovishsentiment. As shown in figure 4, net sentiment at the paragraph-level is generally consistent with the stance of monetary policy: Hawkish (dovish) net sentiment, is consistent with tightening (loosening) cycles ofmonetary policy. The correlationbetweenthe netsentimentofthe press conference and thetarget federal funds rate is the most striking.8 The contemporaneous correlation of the net sentiment of the pressconferenceandthefederalfundsrateisonly0.37—muchlowerthanthatofthenetsentimentof theFOMCstatementandthefederalfundsrateof0.53. However,thecorrelationbetweencontemporaneousnetsentimentandthefuturefederalfundsratepeaksatabout0.87ninemeetingsaheadforthe pressconference.Bycontrast,thepeakcorrelationbetweennetsentimentofFOMCstatementsandthe federalfundsratecomesinlowerat0.79sevenmeetingsahead.Thehighercorrelationbetweennetsen- 8SeeAppendixGfortheconstructionofthetargetfederalfundsrate. Weusethemid-pointofthetargetrangeaftereach FOMCmeeting. 10

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION timentofthepressconferenceandthefuturefederalfundsratelikelyarisesbecausenetsentimentatthe paragraphlevelisfinerandmorevariedforpressconferencesthanFOMCstatements.WhileoverthreefourthsofallnetsentimentscoresfortheFOMCstatementare1,0,or-1—entirelyhawkish,neutral,or dovish—onlyone-fourthofnetsentimentscoresforthepressconferencearethatcoarse. AppendixfigureC.3isthesameasfigure4,butshowssentencelevelinsteadofparagraphlevelnetsentimentforthe FOMCstatementinthetoppanel. Atthisfinerfrequency,thereismorevariationthanthecoarseparagraphlevelanalysisofthestatement,butpeakcorrelationisstillonly0.8fivemeetingsahead,whichis stillbelowthatofthepressconference. fi '0.$
4UBUFNFOUT   )BXLJTI  fl
%PWJTI         fi $IBJS
4QFFDIFT   )BXLJTI  fl
%PWJTI         fi 1SFTT
$POGFSFODFT   )BXLJTI  fl
%PWJTI        



 'FEFSBM
GVOET
SBUF
QFSDFOU















       Figure4:Netsentimentscoresattheparagraphlevelandthetargetfederalfundsrate. Hawkish sentiment pertains policy tightening and dovish to loosening. For communication type j ∈ {Pressconference,Chairspeech,FOMCStatement}, NetSentimentj,t = (Hawkishj,t −Dovishj,t)/(Hawkishj,t +Dovishj,t). ThefederalfundsrateisthetargetaftereachFOMCmeeting,seeAppendixGfordetailsonitsconstruction.AppendixAdetails theavailabilityandsourcesofeachtypeofcommunication. Message2. Thepressconferencecorrelateswiththefuturepolicyratetoagreaterextentthanothercommunications. This is because the FOMC statement coarsely signals the current stance of policy and the pressconferencefine-tunesthemessagetosignalfuturepolicy. 4.2 LOCAL PROJECTIONS Toformalizethepredictivecorrelationsofsentimentandthefederalfunds rate,weestimatelocalprojections.Wechoosetostudytheactualfederalfundsrateintheseestimations, inlightofitsendogeneitytosentiment,becausethecommunicationsunderlyingoursentimentscores 11

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION areoftenaimedatcommunicatingthecurrentandfuturepathofthefederalfundsrate.9 Tothatend, ifthefederalfundsrateisconstrainedatitseffectivelowerbound, thenthecommunicationswillalso necessarilyreflectthisconstraint. HypotheticalratessuchastheshadowfederalfundsrateofWuand Xia(2016)thatproxyforwhatthefederalfundsratewouldbeifitcouldgonegative,necessarilyreflect policychoicesthatarenotreflectedincommunications. Wefirstestimateimpulseresponsesseparatelyforthestatementandthepressconference: FederalFundsRate t+h =βh 0 +βh 1 NetSentiment j,t +ϵ j,t (8) Wheret istheFOMCmeetingfrequency,h=0,..,24,and j ∈{FOMCstatement,Press}.10 Standarderrors arethoseofNeweyandWest(1987)withtwolags(findingsarerobustfortwotoeightlags)androbust toheteroskedasticity. ThesamplestartsinApril2011andonlyincludesmeetingswherethereisapress conference(everyothermeetinguntil2019andeverymeetingthereafter). AppendixfiguresC.4andC.5 mostimportantlyshowthatthepointestimatesforthenetsentimentofFOMCstatementsarealmost identicalwhenthesampleiseitherallFOMCmeetingsfromJanuary2011toApril2025orrestrictedto meetingsfollowedbypressconferences,asshowninthissection. Thet+h federalfundsrateincludes allmeetingssuchthatthe8-meeting-aheadpointestimatescorrespondstoaone-year-aheadhorizon. AppendixfigureC.6controlsforthelaggedvalueofthefederalfundsrate. Figure5showsthatanincreaseinnetsentiment(hawkishcontent)iscorrelatedwithahigherfederal fundsrateforboththenetsentimentofFOMCstatementsandpressconferences.Theresponsetopress conferences(dashedredline)hasamoredistincthump-shapeandpeaksabout8meetings(oneyear) ahead. While this suggests that an increase in net sentiment in press conferences is associated with a relativelyhigherfederalfundsratethanthesameincreaseinFOMCstatements,thesesentimentscores arehighlycorrelatedat0.8andtheirimpulseresponsecoefficientnotstatisticallysignificantfromeach other. Consequently,weexpandthespecificationinequation(8)toincludebothnetsentimentscores. We alsoincludeaspecificationwiththelaggedfederalfundsratetocontrolfor(1)observedpersistencein the federal funds rate and (2) the fact that current net sentiment likely depends on the recent policy stance.11 FederalFundsRate t+h =βh 0 +βh 1 NetSentiment FOMC,t +βh 2 NetSentiment Press,t +ϵ t (9) FederalFundsRate t+h =βh 0 +βh 1 NetSentiment FOMC,t +βh 2 NetSentiment Press,t +... αhFederalFundsRate t−1 +ϵ t (10) Figure6showstheimpulseresponsetospecifications(9)and(10)inpanels6aand6b,respectively.12 9The“FOMCPolicyonExternalCommunicationsofCommitteeParticipants"states,“TheFederalOpenMarketCommittee iscommittedtoprovideclearandtimelyinformationtothepublicabouttheCommittee’smonetarypolicyactionsandrationale forthosedecisions."https://www.federalreserve.gov/monetarypolicy/files/FOMC_ExtCommunicationParticipants.pdf. 10AppendixfigureC.8showsthattheimpulseresponsecoefficientsofthenetsentimentofspeechesislowerthanthatof eitherpressconferencesorFOMCstatements.Moreover,speechesarenecessarilyadifferentsample. 11SeeCrawleyetal.(2026)foradiscussionofgradualismandinertiainmonetarypolicyrules. 12Appendix figure C.7 shows point estimates of equations (9) and (10) that include an interaction term 12

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION            .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU /FU
4FOUJNFOU
1SFTT
$POGFSFODF Figure5: Impulseresponseofthefederalfundsratetonetsentiment,90%errorbandswithtwoNewey- Westlags. The figure plots βˆh 1 from the local projection, FederalFundsRate t+h = βh 0 +βh 1 NetSentimentj,t +ϵ j,t in equation (8). Where t istheFOMCmeetingfrequency, h =0,...,24, and j ∈{FOMCstatement,Press}. NetSentimentj,t =(Hawkishj,t − Dovishj,t)/(Hawkishj,t +Dovishj,t). ThesampleisfromApril2011toApril2025andisforFOMCmeetingsfollowedbypress conferences(everyotherFOMCmeetingfrom2011to2019andeverymeetingthereafter).AppendixfigureC.4comparespoint estimatesofnetsentimentforFOMCmeetingsonthesampleofallmeetings.Thet+hfederalfundsrateincludesallmeetings suchthatthe8-meeting-aheadpointestimatescorrespondstoaone-year-aheadhorizon. Panel6ashowsthatnetsentimentofFOMCstatementsandpressconferenceshavedistinctandcomplementarycorrelationswiththefederalfundsratewhenestimatedjointly. Thepointestimatesofnet sentimentofFOMCstatements(solidblueline)aremostlypositivesuchthatanincreaseinnetsentimentrepresentsmorehawkishcontentinFOMCstatementsandcorrelateswithahigherfederalfunds rateuptoroughlyeightmeetingsahead. Thereafter,thepointestimatesareinsignificantfromzero. By contrast,thepointestimatesofnetsentimentofpressconferences(dashedredline)havetheopposite pattern. Thatis,pointestimatesareinitiallyinsignificantfromzeroforaboutfourmeetingsaheadand thenturnpositiveandsignificantwithapeakeffectabout12meetingsahead. Becausenetsentiment of press conferences leads the federal funds rate, these point estimates can be interpreted as hawkish sentimenttodaycoincidingwithahigherfederalfundsrateatleast6monthsahead. Panel 6b shows that controlling for the pre-meeting value of the federal funds diminishes the predictive correlations of net sentiment of FOMC statements with the federal funds rate such that point estimatesareinsignificantfromzeromorethanfourmeetingsahead. Thissuggeststhatthatmuchof thevariationpreviouslyattributedtonetsentimentofFOMCstatementsmayinsteadreflecttherecent policystance.However,thepredictivecorrelationsofnetsentimentofpressconferenceswiththefederal fundsratearesimilartotheircounterpartsshowninpanel6a.Takentogether,thisevidencesuggestthat netsentimentofpressconferenceshasadistinctrelationshipwiththefuturefederalfundsratebeyond thatofFOMCstatementsandtherecentpolicystance. Thepositiveandsignificantpredictivecorrela- (NetSentimentFOMC,t ×NetSentimentPress,t). 13

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION tionsofthepressconferenceabout22meetingsahead(justunderthreeyears)suggeststhatpressconferencessignalthefuturestanceofpolicywellinadvanceofactualchangesinthefederalfundsrate.13 13AppendixfigureC.9followsCrawleyetal.(2026),CarlstromandFuerst(2014),andCoibionandGorodnichenko(2012)by addinginasecondlagofthefederalfundsrate,whichshortensthepositiveandsignificanthorizonofthepressconferenceto about19meetingsahead.AppendixfigureC.10showsthatlocalprojectionswiththesentence-levelFOMCstatementsinstead oftheparagraph-levelFOMCstatementsaresimilartothosedescribedinthissection.FigureC.11showsthatresultsarerobust tostandardmacroeconomiccontrols. 14

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION           .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU /FU
4FOUJNFOU
1SFTT
$POGFSFODF (a)Jointestimation           .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU /FU
4FOUJNFOU
1SFTT
$POGFSFODF (b)Jointestimationwithlaggedfederalfundsrate Figure6: Impulseresponseofthefederalfundsratetonetsentiment,90%errorbandswithtwoNewey- Westlags. Panel 6a plots βˆh 1 and βˆh 2 from the local projection, FederalFundsRate t+h = βh 0 + βh 1 NetSentimentFOMC,t + βh 2 NetSentimentPress,t +ϵ t inequation(9). Panel6bplotsβˆh 1 andβˆh 2 fromthelocalprojection, FederalFundsRate t+h = βh 0 +βh 1 NetSentimentFOMC,t +βh 2 NetSentimentPress,t +αhFederalFundsRatet−1 +ϵ t inequation(10). NetSentimentj,t = (Hawkishj,t −Dovishj,t)/(Hawkishj,t +Dovishj,t). ThesampleisfromApril2011toApril2025andisforFOMCmeetings followedbypressconferences(everyotherFOMCmeetingfrom2011to2019andeverymeetingthereafter). Thet+hfederal fundsrateincludesallmeetingssuchthatthe8-meeting-aheadpointestimatescorrespondstoaone-year-aheadhorizon. 15

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION 4.3 OUTOFSAMPLE Thepredictivecorrelationsofnetsentimentwiththefederalfundsrateshownin thelocalprojectionsinfigures5and6arenecessarilyaninsampleexercise.Wenextaskifnetsentiment cancorrelatewiththefederalfundsrateoutsampleandhowtheseforecastscomparewithprescriptions from the well-known Taylor (1999) rule and other simple benchmarks for predicting the federal funds rate.14 Thepredictionsshowninfigure7arethefollowing: Federa(cid:225)lFundsRate t+h =βˆ 0 t+h+βˆ 1 t+hNetSentiment FOMC,t +αˆt+hFederalFundsRate t−1 (11) Federa(cid:225)lFundsRate t+h =αˆ 0 t+h+αˆ 1 t+hFederalFundsRate t−1 (12) Federa(cid:225)lFundsRate t+h =2.5+1.5(π t+h −2)+2(u t ∗ +h −u t+h ) (13) (cid:179) (cid:180) Federa(cid:225)lFundsRate t+h =0.85FederalFundsRate t+h−1 +0.15 2.5+1.5(π t+h −2)+2(u t ∗ +h −u t+h ) (14) Eachhorizonish=1,..,8meetingsahead,whichcorrespondstooneyearahead. WeuseallFOMC meetingsfromJanuary1999toApril2025.15 Thenetsentimentandlaggedfederalfundsrateforecasts showninequations(11)and(12),respectively,areinitiallyestimatedfort ={1999,...,2001}andthenreestimatedforeachmeetingthroughApril2025. Thecoefficientsforthenon-inertialandinertialTaylor (1999)rulesshowninequations(13)and(14), respectively, arethewell-knownvaluesofTaylor(1999). Theinflationrateπ t+h istherealizedt+hmonthlypersonalconsumptionexpendituresindexlessfood andenergydetailedinAppendixG.Theunemploymentrateu t+h istherealizedt+hmonthlyunemploymentrateandthenaturalrateu ∗ istherealizedt+hnon-cyclicalrateofunemployment,alsodetailed t+h in Appendix G. The 0.85 weight on the lagged federal funds rate is from the Federal Reserve Bank of Atlanta’sTaylorRuleUtilityTool(https://www.atlantafed.org/research-and-data/data/taylor-rule.) Panel 7a shows that the t+1 forecast of the federal funds rate using just the net sentiment FOMC statements,asdetailedinequation(11),performsremarkablywellinforecastingthefederalfundsrate outsample. Infact,itperformsbetterthanthepredictionoftheTaylor(1999)ruledetailedinequation (13).However,theforecastwithnetsentimentisalsoquiteclosetotheprescriptionsoftheinertialTaylor (1999)ruleandaforecastbasedonthelaggedfederalfundsrateas,detailedinequations(14)and(12), respectively.16 Forthisreason,wecalculaterootmeansquarederrors8FOMCmeetingsaheadtobetter understand forecast performance. In equation (15), the federal funds rate is the target as detailed in AppendixGandthepredictedFederal(cid:225)FundsRate t+h isshownforequations(11),(12),and(14). (cid:118) (cid:117) (cid:117)1 (cid:88) h RMSE h =(cid:116) (FederalFundsRate t+h −Federal(cid:225)FundsRate t+h )2 (15) h t=1 14Seepages43-45oftheFederalReserve’sJune2025MonetaryPolicyReporttoCongressforaversionoftheTaylor(1999)rule weuse,https://www.federalreserve.gov/monetarypolicy/files/20250620_mprfullreport.pdf. 15Wedonotincludenetsentimentofpressconferencesinequation(11)toavoidestimatingadifferentspecificationforevery othermeetingfrom2011to2019. 16Crawleyetal.(2026)showthatdoubleinertialrulesaresometimesnecessarytoimproveuponthepredictivefitofTaylortyperules. 16

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION (a)Federalfundsratewithselectone-periodaheadpredictions,percent (b)Rootmeansquarederrorsforthepredictedfederalfundsrate Figure7:Outofsamplepredictionsofthefederalfundsrate. All data series are detailed in G, including the target federal funds rate. Panel 7a shows t+1 predictions. The net sentimentforecastisthatofequation(11),Federa(cid:225)lFundsRate t+h =βˆ 0 t+h+βˆ 1 t+hNetSentimentFOMC,t +αˆt+hFederalFundsRatet−1 andthelaggedfederalfundsrateforecastisthatofequation(12)Federa(cid:225)lFundsRate t+h =αˆ 0 t+hαˆ 1 t+hFederalFundsRatet−1. Theseforecastsareinitiallyestimatedfort ={1994,...,1997}andthenre-estimatedforeachmeetingthroughApril2025for h=1,...,8.TheTaylor(1999)ruleisequation(13),Federa(cid:225)lFundsRate t+h =2.5+1.5(π t+h −2)+2(u t ∗ +h −u t+h )whereinflation isthet+h realizedcorePCEindexandunemploymentrategapseriesistherealizedt+h unemploymentratelessitsnaturalrate,andtheinertialTaylor(1999)ruleisthatofequation(14),Federa(cid:225)lFundsRate t+h =0.85FederalFundsRate t−1+h + 0.15 (cid:179) 2.5+1.5(π t+h −2)+2(u t ∗ +h −u t+h ) (cid:180) whichusesthesameinflationandunemploymentrateseriesasitsnon-inertialcounterpart. The0.85weightonthelaggedfederalfundsrateisfromtheFederalReserveBankofAtlanta’sTaylorRuleUtilityTool, (https://www.atlantafed.org/research-and-data/data/taylor-rule).Therootmeansquarederrorsshowninpanel (cid:113) 7bisfromequation(15)RMSE h = h 1(cid:80)h t=1 (FederalFundsRate t+h −Federal(cid:225)FundsRate t+h )2forh=1,...,8. 17

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION Panel 7b shows that the root mean squared error of the net sentiment forecast is the lowest for all horizonsshown. WhiletheRMSEofthelaggedfederalfundsratespecificationislowerthanthatofthe inertialTaylor(1999)ruleatshorthorizons,asshownbyCrawleyetal.(2026),thenetsentimentforecast is the lowest shown at all horizons and significantly so as confirmed by Diebold and Mariano (1995) testsofforecastaccuracyshowninappendixtable1. Thesefindingsarerobusttoexcludingtheperiods whenthefederalfundsrateisatitseffectivelowerbound(December16,2008toDecember16,2015as well as March 15, 2020 to March 16, 2022). Taken together, this evidence suggests that net sentiment communicatedbytheFederalReserveisatleastasgoodofapredictoroffuturepolicyaswell-studied reactionfunctionsbasedonmacroeconomicaggregates. WhilethissectionshowsthatthenetsentimentoftheFOMCstatementsandpressconferencescontainpredictivecorrelationswiththefuturefederalfundsrate,thesespecificationsarehighlyendogenous. Tobetterunderstandtheeffectofsentimentonfuturepolicy,wenextturntohigh-frequencymarketreactionsthatallowfortheconstructionofmonetarypolicyshocks. 5 SENTIMENT AND MARKET REACTIONS Weregressmarketreactionsdiscussedinsection3onsentimentasconstructedinsection4foreachtype ofcommunicationtopartiallycontrolfortheendogeneityofthefederalfundsrate. Wefocusonthe4th Eurodollar/5thSOFRfuture,EDSR4,whichcorrespondstotheexpectedpolicyratefourquartersahead, foreaseofdiscussion. AppendixD.1confirmsthatallresultsforEDSR4holdwithallotherinterestrate futuresoflowermaturities. MarketReaction =α+βSentiment +ϵ (16) t,i,j t,i,j t,j,i Forcommunicationtype j ∈{Pressconference,Chairspeech,FOMCStatement}andrawsentiment i ∈{Hawkish,Dovish,Neutral}foreachmonetaryeventt. Message3. WhenregressingmarketreactionsonLLMsentiment,thepressconferencestandsoutforthe strongestrelationshipbetweensentimentandmarketreaction. Figure 8 shows the estimates of the slope coefficient βˆ from equation (16) estimated via OLS with Huber-Whiterobuststandarderrorsand95percenterrorbands.Thex-axisshownishawkishsentiment andthey-axisisthe4thEurodollar/5thSOFRfuturerepresentingtheexpectedpolicyratefourquarters ahead.AppendixD.1showstheresultsfordovishandneutralsentimentalongwithothermarketindicators.Themarketreactioncanvary,withsomecommunicationeventshavingalmostnomarketreaction andothershavingonethatisalmost20basispoints.Asinsection4,communicationeventscancontain dovish,hawkish,andneutralsentiment. Pressconferencesstandsoutashavingthestrongestrelationshipbetweenhawkishsentimentand marketreaction,consistentwithitspredictivecorrelationswiththefederalfundsrateshownintheprevious section. The slope coefficient βˆ is statistically significant from zero meaning that the stronger the hawkish content, the stronger the market reaction. By contrast, speeches by the chair and FOMC 18

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION statementshaveaslopecoefficientthatisstatisticallyinsignificantfromzerosuggestingnodiscernible relationshipbetweencontentandthesizeofthemarketreaction.Ifspeechesalsohadastatisticallysignificantslopecoefficientlikepressconferences,onecouldconcludethatmarketsreactmoretoFOMC communicationsthanactions. However,theresultsinfigure8suggestthatcommunicationsfollowing actionselicitthelargestmarketreactions,whichsuggeststheimportanceofcommunicationsrelatedto changesinfundamentalsandspecialroleforthepressconferencethatisdistinctfromspeeches. Figure8presentsapuzzle: themostnegativemarketreactionsofthepressconferencearethosein the 2022-2024 tightening cycle. Why was the market surprised to the downside during a rapid tightening cycle characterized by hawkish sentiment? We first note that Appendix figure D.4 confirms that thestrongestmarketreactionsareindeedthosethatreversethereactionstotheprecedingFOMCstatement. Wearguethatthesereversalsaretheresultofdovishsentimentinpressconferencesrefiningthe coarsehawkishsignalofFOMCstatementsaboutthefuturestanceofpolicy. Thatis,becausethepress conferencehasahighercorrelationwiththefuturefederalfundsrate,themarketisreactingtorefined signals. Figure8:Regressionofmarketsurprisesonhawkishsentiment. Thefiguresshowsequation(16)∆MarketReactiont,j =α+βSentimentj +ϵ t,j forhawkishsentimentforeachtypeofcommunication j∈{Pressconference,Speech,FOMCstatement}. Shadedbandsare95percentconfidenceintervals. Thefigureplots allmarketreactionswithpairedsentimentscores. Marketreactionsareintradaychangesoftradeprices20minutesaftera communicationeventrelativetotradeprices10minutesbeforefortheithEurodollar/(i+1)thSOFRfuture(ESRi),whichcorrespondstotheexpectedpolicyrateiquartersaheadfori=1,...,4.Theestimationsamplestartsin1999forFOMCstatements, in2008forChairspeeches,andin2011forpressconferences,whichisbasedontheavailabilityoftextsofthesecommunications. 19

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION While FOMC statements used to be the primary form of communication, our findings highlight a keyroleforthesentimentofthepressconferenceinexplainingmarketreactionsthatsometimesreverse thoseofthestatement,especiallyininterestratefuturesfourquartersahead. Wenexttestifthemarket doesindeedperceivethepressconferenceasclarifyinginformationthatshapesexpectationsaboutthe futurepathofpolicy. Afterall,NarainandSangani(2026)documentheightenedmarketvolatilityunder ChairPowell’spressconferences,particularlywhenaddressingforwardguidance,whichalignswithour findingthatpressconferencesrefinethestatement’smessage. 6 ANNOUNCEMENT VARIANCE RATIO Although elevated trading volume around FOMC statements is well known, less is understood about how long the announcement-related impulse remains informative. To measure the incremental signal associated with the press conference, we estimate an Announcement Variance Ratio (AVR) using anapproachsimilarinspirittothelocal-projectionsdifference-in-differencesframeworkinDubeetal. (2025).17 TheAVRcaptures,intheOLSsense,theshareofvariationinminute-leveltradingactivityattributabletotheFOMC-announcementtreatmentinteractionrelativetothevariationleftunexplained by the fitted model. It is therefore a partial/incremental measure of explained variance and inherits a standardvariance-decompositioninterpretation.ThisisbecauseOLSpartitionsthetotalsumofsquares intoacomponentexplainedbytheregressorsandaresidualcomponent,andtheAVRreportshowmuch theannouncementinteractioncontributestotheformerperunitofthelatter. Specifically,wemeasure themarginalvariationintradingactivityaroundanFOMCstatementrelativetothesamedayandtime one week before and one week after the release.We focus on scheduled meetings only in this section, butresultsaresimilarifunscheduledmeetingsareincluded. Pre-trendassumptionsnecessaryforthe differences-in-differencesanalysisarelesslikelytoholdforunscheduledmeetingsastheycouldbeproceededbyatypicaltradingactivity.18 Foreachfuturescontracti andeachhorizonh,weestimate (cid:104) (cid:105) Y τ i ,d =βi 0 ,h+βi 1 ,h1(d=FOMC)+βi 2 ,h1(−10≤τ≤h)+βi 3 ,h 1(d=FOMC)×1(−10≤τ≤h) +λ m×y +ε τ i, , h d , (17) whereYi isminute-τtradingvolumeforcontracti,scaledbythatday’s(d)totaltradingvolume,andτ= τd 0denotesthescheduledFOMCstatementreleasetime.Thefuturesstudiedi aretheithEurodollar/(i+ 1)thSOFRfuturesfori =1,..,4correspondingtothefirstyearofthetermstructureofinterestrates. The indicator 1(d =FOMC) equals one on FOMC statement days and zero on control days, defined as the sameweekdayoneweekbeforeandoneweekaftertheFOMCmeeting. Scalingbydailytradingvolume accountsfordifferencesinoveralltradingintensityacrossdays,whilemonth×yearfixedeffectsabsorb additionalvariation. 17AnnetteVissing-Jorgensen’sdiscussionofSwansonandJayawickrema(2024)conductsarelatedvariancedecomposition exerciseofstockmarketreturnsaroundFOMCevents. Shefindsthatspeechesarenoisier(lessattributablevariation)than statementsandpressconferenceshttps://drive.google.com/file/d/1GD8FK7QwA1VqSu1g1LdCtLeOqYQ1cujT/view. 18AppendixfigureE.1showstradingvolumesaroundFOMCannouncementsasashareofdailyvolumeforFOMCstatement releasesstudied. 20

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION Theestimationsampleisfixedoverτ∈[−70,170]. Foragivenhorizonh,thetreatedwindowisdefinedastheintervalfromtenminutesbeforethestatementreleasethroughminuteh,thatis,τ∈[−10,h]. Webeginwithh=−9,sotheinitialtreatedwindowcontainsonlythetwominutesimmediatelyprecedingthestatementrelease,andthenexpandthewindowoneminuteatatimeuptoh=170. Thisdesign preservescomparabilitywithstandardidentificationofhigh-frequencymonetarypolicyevents, which measurespricechangesrelativetothetradepricetenminutesbeforethestatementrelease. We focus on the incremental explanatory power of the interaction term as the treated window expands.Ateachhorizonh,wedefinetheAVRas SS(1(d=FOMC)×1(−10≤τ≤h)) AVR = , (18) h RSS h whereSS(·)denotestheANOVAsumofsquaresattributabletotheinteractiontermandRSS istheresidh ualsumofsquaresfromthefullmodelestimatedathorizonh. Ahighervalueofthestatisticindicates thatincludingminutehintheeventwindowaddsmoreFOMC-relatedsignalthanresidualnoise. Thisminute-by-minuteanalysisprovidesadirectwaytoassesshowquicklystatement-relatedtradingactivitybecomesinformativeandhowlongthatinformativenesspersists.Inparticular,avalueofzero whenh<0indicateslittledifferentialtradingactivitybeforethescheduledrelease,asshowninfigure9. Anincreaseinthestatisticatτ=0andcontinuingafterwardsuggeststhatstatement-relatedinformation arrivesatthescheduledreleaseandcontinuestounfoldduringthesubsequentpressconferenceperiod whichbeginsatτ=30. Message4. Thestatementaloneislessinfluentialcomparedtoafewyearsago. Marketparticipantsincreasingly wait for the press conference to fine-tune their expectations, as indicated by relatively higher announcementvariancedetectedfromthepressconferencerelativetothestatement. Figure9comparestheAVRaroundreleasesofFOMCstatementsintheperiodpriortothepressconferenceinpanel9a(1995toMarch2011)tothataftertheintroductionofthepressconferenceinpanel9b (April2011-2025)andshowsthatthemarketperceivesadditionalsignalfrompressconferences. While panel9ashowsthatsignaldetectedfromtradingvolumespeakedabout40minutesafterFOMCstatementsinallfourinterestratefuturesshown, panel9bshowsadifferentpatternaftertheintroduction ofthepressconference. Inpanel9b,signalpeaksabout20minutesaftertheFOMCstatementrelease, flattensandthenbeginstoclimbagainafterthestartofthepressconference,whichis30minutesafter thestatementrelease(denotedbytheseconddashedverticalline).Thesignalcontinuestoclimbduring pressconferencesandpeaksabout30minutesafterthestart,whichisaboutanhourafterthereleaseof theFOMCstatement. Theannouncementvarianceratiosarerelativelylowerinthepost-pressconferencesamplein9b(2011-2025)becauseof,inpart,relativelylowvaluesduringtheELBin2020and2021, asshowninfigure10.Byaveragingovereachyearinsteadofpre-andpost-pressconferenceperiods,figure10showsthatannouncementvarianceratioslendtorelativelymoresignalfromthepressconference thanthestatementby2019whenChairPowellinstitutesapressconferenceaftereverymeeting. Giventhatthepressconferencecontainsadditionalinformationaboveandbeyondthestatement, howshouldresearchersincorporatethepressconferenceintotheconstructionofhigh-frequencymon- 21

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION (a)Pre-pressconference,1995toMarch2011 (b)Post-pressconference,April2011to2025 Figure9:Announcementvarianceratiobypre-andpost-pressconferencesamples The figure plots the Announcement Variance Ratio, whichis the sumofsquares of the coefficientYi =βi,h+βi,h1(d = τ,d 0 1 (cid:104) (cid:105) FOMC)+β 2 i,h1(−10≤τ≤h)+βi 3 ,h 1(d =FOMC)×1(−10≤τ≤h) +λ m×y +εi τ , , h d . ZeroindicatestheminuteoftheFOMC statement(typically2:00PMEST),andthex-axisisminutessincethestatementrelease.Thepressconferencetypicallystarts 30minutesafterthestatementrelease,asindicatedbythesecondverticaldashedlineinpanel9b. Thesampleincludesall scheduledFOMCStatementsfromJanuary1995toApril2025.TheithEurodollar/(i+1)thSOFRfuture(ESRi)correspondsto theexpectedpolicyrateiquartersaheadfori=1,...,4. etarypolicyshocks? Afterall,focusingsolelyonstatementsmaymissthefullscopeofmonetarynews containedinthepressconference. 7 MONETARY TRANSMISSION Weconstructcommonlyusedmonetarypolicyshocksfromthepreviouslydiscussedintradaychanges infederalfundsfuturesandEurodollar/SOFRfutures. TheNakamuraandSteinsson(2018)shockseries isthescaledfirstprincipalcomponentoftheinstrumentset {MP1,MP2,EDSR2,EDSR3,EDSR4} using 30-minute windows around FOMC statements.19 We construct a second version of this series using the same instrument set, but 90-minute windows around FOMC statements when there is a post-meeting press conference as in Acosta et al. (2025) and Grotteria and van Binsbergen (2025). Using this same instrument set, Gürkaynak et al. (2005) extract and rotatetwofactors—thetargetandpath—whichcorrespondtothelevelandslopeoftheyieldcurvefor one-year-aheadinterestrates.Giventhatthenetsentimentofpressconferencescontainspredictivecorrelationswiththefuturefederalfundsratebeyondoneyear,weaugmenttheinstrumentsettoinclude 19MP1andMP2arecalculationsperformedonFF1throughFF4tocorrectfortimeaveraginginsettlementprices. See Brennanetal.(2025)andAppendixFfordetaileddescriptions. 22

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION Figure10:Announcementvarianceratiobyyear,2011-2025. The figure plots the Announcement Variance Ratio, whichis the sumofsquares of the coefficientYi =βi,h+βi,h1(d = τ,d 0 1 (cid:104) (cid:105) FOMC)+β 2 i,h1(−10≤τ≤h)+βi 3 ,h 1(d =FOMC)×1(−10≤τ≤h) +λ m×y +εi τ , , h d . ZeroindicatestheminuteoftheFOMC statement(typically2:00PMEST),andthex-axisisminutessincethestatementrelease.Thepressconferencetypicallystarts 30minutesafterthestatementrelease.ThesampleincludesallscheduledFOMCStatementsfromJanuary1995toApril2025. ThevolumeshownisforEDSR4whichisthe4thEurodollar/5thSOFRfuturecorrespondingtotheexpectedpolicyratefour quartersahead. high-frequencychangesin2-,5-,10-,and30-yearTreasuryyieldsasinBrennanetal.(2025). Appendix Fcontainsdetailsonshockconstruction. Theeffectofmonetarypolicyonthemacroeconomyisfrequentlyestimatedviaastructuralvector autoregressionframeworkatamonthlyfrequencybyusinghigh-frequencymonetarypolicyshockseries asexternalinstruments.20 WeusetheVARspecificationofBauerandSwanson(2022)whichisavariant ofGertlerandKaradi(2015)andextendittoApril2025insteadofstoppingin2019asinpreviousstudies. The external instrument imposes a second moment restriction to identify shocks, more specifically it replaces one column of the rotation matrix with predicted values from a regression of a reduced form VARinnovationontheexternalinstrument. Identificationviaanexternalinstrumenthingesonahigh-frequencymonetarypolicyshockseries εi satisfying relevance and exogeneity conditions to be an adequate external instrument for εmp the t t unobservabletruemonetaryshocks. E[εmpεi]̸=0 and E[εiε(cid:8)m (cid:8) p ]=0 t t t t Whereε isanyseriallyuncorrelatedstructuralshocksdrivingtheeconomyandε(cid:8)m (cid:8) p isasubsetofthese t 20SeeStockandWatson(2018)foradditionaldocumentationofVARswithexternalinstrumentsandMiranda-Agrippinoand Ricco(2023)foradiscussionofwhyitisthedominantspecificationinempiricalmacroeconomics. 23

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION shocksunrelatedtomonetarypolicy. Sincethetruevalueofεmp isunobserved,instrumentvaliditymustultimatelybejustifiedlogically. t Anyhigh-frequencymonetaryshockseriesshouldsatisfytherelevanceconditionastheycapturemonetarynewsconveyedbyFedcommunicationbyconstruction.TheexclusionconditionshouldalsobesatisfiedbecausethetightwindowsaroundFOMCstatementreleasesshouldpreventnon-monetarynews from moving markets. While Bauer and Swanson (2023) show that the exclusion condition is violated fortheNakamuraandSteinsson(2018)monetarypolicyshockseries,Brennanetal.(2025)findthatthat othershockserieslikethoseofBuetal.(2021)andKuttner(2001)areunpredictable,whichsuggeststhat predictabilityisafeatureofcertainshocks,nothigh-frequencymonetaryshocksperse.Relatedly,these shockseriesexhibitconventionally-signedresponseswhichcanalleviateconcernsaboutcontamination fromtheso-calledFedinformationeffect. ThespecificationforaVARwithexternalinstrumentsisgivenas: Y T =α+B(L)Y T−1 +s 1 Y T 2Y +u˜ T (19) WhereY isavectorcontainingfourmonthlyeconomicvariablesfromJanuary1973toApril2025: the T logoftheconsumerpriceindex(CPI),thelogofindustrialproduction(IP),theGilchristandZakrajšek (2012)excessbondpremium(asadvocatedbyCaldaraandHerbst(2019),andthetwo-yearzero-coupon Treasuryyieldatamonthlyfrequency. AppendixGdetailsthesourcesoftheseseries. Wealsonotethat thetwo-yearTreasuryyieldisthedailychangeobservedattheendofthemonthasintheexcelspreadsheetusedbyBauerandSwanson(2022).21 Theexcessbondpremiumcontrolsforfinancialfactorsand thetwo-yearTreasuryisameasureofthestanceofmonetarypolicy. AlthoughtheoriginalGertlerand Karadi(2015)specificationusestheone-yearTreasury,thetwo-yearhastheadvantageofbeinglessconstrainedattheELBandisusedbyBauerandSwanson(2022). Next,B(L)isthematrixpolynomialinthelagoperator. AlthoughBauerandSwanson(2022)follow GertlerandKaradi(2015)andRamey(2016)inusing12lags,wefollowBrennanetal.(2025)anduse8 lags.22 WhileBauerandSwanson(2022)constructaversionoftheNSshockseriesfrom1988to2019,our monetarypolicyshockseriesbeginsin1995whichisasclosetothe1994introductionofexplicitpolicy statementsasourintradaydataallowswithoutresortingtosourceswecannotreplicatefromintraday dataonactualtrades. WeextendtheseshockseriestopresentandestimatetheVARtopresentaswell withAppendixfiguresF.4-F.6showingthatresultsaresimilarifweinsteadstopin2019. Finally, ϵi istheinstrumentfor s Y 2y estimatedviatwo-stageleastsquaresandu˜ istheresidual. t 1 T T The sample of the external instrument εi does not have to be the same as that of the economic data. t Infact, thesampleusedforourshockseriesisfromJanuary1995toApril2025andthesampleforthe economicdataisfromJanuary1973toApril2025.Furthermore,followingtheliterature,ϵi isaggregated t to a monthly frequency by summation, ϵi = (cid:80) ϵi. We do not make any further adjustments for the T t t timingofshockswithinthemonthasRamey(2016)andMiranda-AgrippinoandRicco(2021)findthat theadjustmentsproposedbyGertlerandKaradi(2015)caninduceserialcorrelation. 21Seehttps://www.frbsf.org/wp-content/uploads/monetary-policy-surprises-data.xlsx?2026-03-24. 22AppendixfiguresF.1-F.3showthattheresultsarequitesimilarwith12lags,albeitwithwidererrorbandsinsomeinstances. 24

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION Figure 11 plots impulse responses from equation (19) obtained via the Canova and Ferroni (2022) toolboxwith68and90percenterrorbandsand20,000draws.Theimpulseresponsestoa25basispoint monetaryshockarequalitativelyinlinewithmacroeconomictheoryandsimilarinsignandshape,even thoughtheshockseriesusedinpanel11aisconstructedwithoutthepressconference,itscounterpartin panel11bincludesthepressconference,andpanel11ccontainsthepressconferenceandlonger-term Treasuriesintheinstrumentset.Theresponseofthetwo-yearTreasury,showninthebottomrowofeach panel,risesonimpacttoanormalizedinitialresponseof25basispointsandthenreturnstozeroabout 10monthsaftertheinitialimpulse. TheimpulseresponsesofbothindustrialproductionandCPItoacontractionarymonetarypolicy showninthefirstandsecondrowsoffigure11,respectively,aresignificantlynegative—asstandardNew Keynesiantheorypredicts.Differencesinpointestimatesamongtheshockseriesareonlyinmagnitude ratherthaninsign.A25basispointcontractionarymonetarypolicyshocksresultsinabouta1and0.25 p.p.reductioninindustrialproduction(IP)andtheconsumerpriceindex(CPI),respectively. Theexcessbondpremiumshowninrowthreeistheonlyvariablethatdisplaysqualitativedifferences inpointestimates. Whilethecoefficientsareambiguouslyzerofortheoriginal30-minutetimewindow showninpanel11a,theyarepositiveandstatisticallysignificantfromzeroonimpact,asshowninpanel 11b, and statistically negative in panel 11c. In Gertler and Karadi (2015), the excess bond premium is positiveandsignificant,whichisonlythecaseinpanel11bwheretheNakamuraandSteinsson(2018) monetarypolicyshockisconstructedtoincludethepressconference,butnotlong-termTreasuries. Figure12plotstheimpulseresponsesfromequation(19)usingthetargetshockofGürkaynaketal. (2005)constructedfromtherotatedfirstprincipalcomponentoftheinstrumentset {MP1,MP2,EDSR2,EDSR3,EDSR4} and the version augmented with intraday changes in 2-, 5-, 10-, and30-yearTreasuries. Thetargetshockrepresentunexpectedchangesinactualinterestratedecisions by the FOMC. Figure 12 shows point estimates that are aligned with theoretical predictions and have littledifferenceacrossmonetarypolicyshockconstructions.Infact,unliketheimpulseresponsestothe NakamuraandSteinsson(2018)shockshowninfigure11,thoseforthetargetshockofGürkaynaketal. (2005)infigure12areallpositiveandsignificantfortheexcessbondpremiumshowninrowthree.Taken together,theseresultssuggestthatexpandingtimewindowstoincludetheforwardguidancecontained inthepressconferenceanditseffectsonlong-termrateshaslittleeffectonunexpectedchangesinthe federalfundsrate,asonewouldexpect. Figure13plotstheimpulseresponsesfromequation(19)usingthepathshockofGürkaynaketal. (2005)constructedfromtherotatedsecondprincipalcomponentoftheinstrumentset {MP1,MP2,EDSR2,EDSR3,EDSR4} and the version augmented with intraday change in 2-, 5-, 10-, and30-yearTreasuries.Thefigureshowspointestimatesthatareeitheropposite-signedorinsignificant fromzerofortheoriginalshockconstructionandtheversionthatincludesinformationfromthepress conference(panels13aand13b,respectively). WhileauthorslikeBundickandSmith(2020)accountfor thesepointestimatesasinformationshocksaboutthestateoftheeconomy,SwansonandJayawickrema (2024)showthatincludingmarketreactionstospeechesinmonetarypolicyshockconstructioncanlead tosignsofpointestimatesthatareconventionally-signedandlooklikethoseinpanel13c. Additionally, 25

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION Adams and Barrett (2026) find point estimates similar to ours with long-term rates in panel 13c from syntheticmonetarypolicyshocksconstructedfromlong-horizonforwardguidance. dnoBssecxE raeY-2 PI IPC muimerP yrusaerT F-stat=4.08 F-stat=4.16 F-stat=1.77 OrignalNS(2018) NS(2018)w/Press NS(2018)w/Press TimeWindow Conference Conference+longrates 1 1 1 0 0 0 -1 -1 -1 -2 -2 -2 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.5 0.5 0.5 0 0 0 -0.5 -0.5 -0.5 -1 -1 -1 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.1 0.1 0.1 0 0 0 -0.1 -0.1 -0.1 -0.2 -0.2 -0.2 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.5 0.5 0.5 0 0 0 -0.5 -0.5 -0.5 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 (a)Originaltimewindow (b)Includespressconference (c) Includes press conference and long-termrates Figure11: Impulseresponsetoa25basispointNakamuraandSteinsson(2018)monetarypolicyshock, x-axisismonthsandy-axisispercentagepoints. Impulseresponsesareestimatesfromequation(19)YT =α+B(L)YT−1 +s1Y T 2Y +u˜T obtainedviatheCanovaandFerroni (2022)BayesianVARtoolboxwith68and90percenterrorbands(darkerandlighterbands,respectively),20,000draws,and 8lags. ThesampleofmonetaryshockseriesisfromJanuary1995toApril2025whilethesampleofeconomicdatastartsin January1973. Theshockseriesarethefirstprincipalcomponentoftheinstrumentset{MP1,MP2,EDSR2,EDSR3,EDSR4}. Theshockseriesinpanelaisconstructedfromthe30-minutechangeinthesefuturesaroundFOMCstatements. Theshock seriesinpanelbisconstructedfromthe90-minutechangeinthesefuturesaroundFOMCstatementswhenthereisapostmeetingpressconferencesothatthetimewindowincludesthepressconference.Panelcaugmentstheinstrumentsetinpanel bwithintradaychangesin2-,5-,10-,and30-yearTreasuryyields. IPistheindustrialproductionindex,CPIistheconsumer priceindex,excessbondpremiumisfromGilchristandZakrajšek(2012),andthetwo-yearTreasuryistheendofthemonth dailychangeinthezero-couponyield.AllsourcesofseriesaredetailedinAppendixG.AppendixTable2displaysthefirst-stage F-statistics. 26

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION dnoBssecxE raeY-2 PI IPC muimerP yrusaerT F-stat=3.71 F-stat=3.96 F-stat=3.42 Target,OriginalTimeWindow Targetw/Press Target+longrates Conference 1 1 1 0 0 0 -1 -1 -1 -2 -2 -2 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.5 0.5 0.5 0 0 0 -0.5 -0.5 -0.5 -1 -1 -1 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.1 0.1 0.1 0 0 0 -0.1 -0.1 -0.1 -0.2 -0.2 -0.2 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.5 0.5 0.5 0 0 0 -0.5 -0.5 -0.5 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 (a)Originaltimewindow (b)Includespressconference (c) Includes press conference and long-termrates Figure 12: Impulse response to a 25 basis point Gürkaynak et al. (2005) target monetary policy shock, x-axisismonthsandy-axisispercentagepoints. Impulseresponsesareestimatesfromequation(19)YT =α+B(L)YT−1 +s1Y T 2Y +u˜T obtainedviatheCanovaandFerroni (2022)BayesianVARtoolboxwith68and90percenterrorbands(darkerandlighterbands,respectively),20,000draws,and8 lags.ThesampleofmonetaryshockseriesisfromJanuary1995toApril2025whilethesampleofeconomicdatastartsinJanuary 1973.Theshockseriesistherotatedfirstprincipalcomponentoftheinstrumentset{MP1,MP2,EDSR2,EDSR3,EDSR4}.The shockseriesinpanelaisconstructedfromthe30-minutechangeinthesefuturesaroundFOMCstatements.Theshockseries inpanelbisconstructedfromthe90-minutechangeinthesefuturesaroundFOMCstatementswhenthereisapost-meeting pressconferencesothatthetimewindowincludesthepressconference.Panelcaugmentstheinstrumentsetinpanelbwith intradaychangesin2-,5-,10-,and30-yearTreasuryyields. IPistheindustrialproductionindex,CPIistheconsumerprice index,excessbondpremiumisfromGilchristandZakrajšek(2012),andthetwo-yearTreasuryistheendofthemonthdaily changeinthezero-couponyield. AllsourcesofseriesaredetailedinAppendixG.AppendixTable2displaysthefirst-stage F-statistics. 27

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION dnoBssecxE raeY-2 PI IPC muimerP yrusaerT F-stat=0.34 F-stat=0.21 F-stat=0.69 Path,OriginalTimeWindow Pathw/Press Path+longrates Conference 5 5 5 0 0 0 -5 -5 -5 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0 0 0 -2 -2 -2 -4 -4 -4 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 1 1 1 0 0 0 -1 -1 -1 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 1 1 1 0 0 0 -1 -1 -1 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 (a)Originaltimewindow (b)Includespressconference (c) Includes press conference and long-termrates Figure13: Impulseresponsetoa25basispointGürkaynaketal.(2005)pathmonetarypolicyshock,xaxisismonthsandy-axisispercentagepoints. Impulseresponsesareestimatesfromequation(19)YT =α+B(L)YT−1 +s1Y T 2Y+u˜T obtainedviatheCanovaandFerroni(2022) BayesianVARtoolboxwith68and90percenterrorbands(darkerandlighterbands,respectively),20,000draws,and8lags.The sampleofmonetaryshockseriesisfromJanuary1995toApril2025whilethesampleofeconomicdatastartsinJanuary1973. Theshockseriesaretherotatedsecondprincipalcomponentoftheinstrumentset{MP1,MP2,EDSR2,EDSR3,EDSR4}.The shockseriesinpanelaisconstructedfromthe30-minutechangeinthesefuturesaroundFOMCstatements.Theshockseries inpanelbisconstructedfromthe90-minutechangeinthesefuturesaroundFOMCstatementswhenthereisapost-meeting pressconferencesothatthetimewindowincludesthepressconference.Panelcaugmentstheinstrumentsetinpanelbwith intradaychangesin2-,5-,10-,and30-yearTreasuryyields. IPistheindustrialproductionindex,CPIistheconsumerprice index,excessbondpremiumisfromGilchristandZakrajšek(2012),andthetwo-yearTreasuryistheendofthemonthdaily changeinthezero-couponyield. AllsourcesofseriesaredetailedinAppendixG.AppendixTable2displaysthefirst-stage F-statistics. 28

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION 8 CONCLUSION ThisstudyexplainspatternsinmarketreactionstoFederalReservecommunicationswithinsightsfrom largelanguagemodeling.WedocumentempiricalrelationshipsshowingthatFederalReservepressconferencesrefinecoarsesignalsofFOMCstatementsandcontaincontentthataccountsforthefuturefederal funds rate and intraday changes in asset prices that are opposite of those in response to FOMC statements. By integrating large language models with high-frequency asset price data, we find that pressconferencesstandoutfortheirintensityofsentimentandmarketreactions. Weconfirmthedistinctroleofpressconferencesinrefiningthestatement’smessagebyshowingthatmarketparticipants indeedperceiveadditionalsignal. Finally,weshowthatincludingpressconferencesintheconstruction ofhigh-frequencymonetarypolicyshockscanaffectestimatesofmonetarypolicytransmission. 29

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION REFERENCES Acosta,MiguelandEllenMeade,“Hangingoneveryword: SemanticanalysisoftheFOMC’spostmeeting statement,” 2015. FEDS Notes. Washington: Board of Governors of the Federal Reserve System, September30,2015.https://doi.org/10.17016/2380-7172.1580. ,AndreaAjello,FrancescaLoria,andSilviaMiranda-Agrippino,“FinancialMarketEffectsofFOMC Communication:EvidencefromaNewEvent-StudyDatabase,”2025. WorkingPaper. ,ConnorM.Brennan,andMargaretM.Jacobson,“Constructinghigh-frequencymonetarypolicysurprisesfromSOFRfutures,”EconomicsLetters,2024,242,111873. Adams, JonathanJ.andPhilipBarrett, “What are Empirical Monetary Policy Shocks? Estimating the TermStructureofPolicyNews,”2026. Workingpaper. Ahrens, Maximilian, Deniz Erdemlioglu, Michael McMahon, Christopher J. Neely, and Xiye Yang, “Mindyourlanguage:Marketresponsestocentralbankspeeches,”JournalofEconometrics,2025,249, 105921. Alexopoulos, Michelle, Xinfen Han, Oleksiy Kryvtsov, and Xu Zhang, “More than words: Fed Chairs’ communicationduringcongressionaltestimonies,”JournalofMonetaryEconomics,2024,142,103515. Altavilla,Carlo,RefetGurkaynak,ThiloKind,andLucLaeven,“MonetaryTransmissionwithFrequent PolicyEvents,”2025. WorkingPaper. An,Phillip,KarlyeDiltsStedman,andAmazeLusompa,“HowHighDoesHighFrequencyNeedtoBe? AComparisonofDailyandIntradayMonetaryPolicySurprises,”2025. WorkingPaper. Apel, Mikael, MariannaBlixGrimaldi, andIsaiahHull, “How Much Information Do Monetary Policy CommitteesDisclose? EvidencefromtheFOMC’sMinutesandTranscripts,”JournalofMoney,Credit andBanking,2022,54(5),1459–1490. Aruoba, S.BoraganandThomasDrechsel, “Identifying Monetary Policy Shocks: A Natural Language Approach,”2026. Forthcoming,AmericanEconomicJournal:Macroeconomics. Banerjee,Shantanu,PaulCordova,MichielDePooter,andOlesyaV.Grishchenko,“GaugingtheSentimentofFederalOpenMarketCommitteeCommunicationsthroughtheEyesoftheFinancialPress,” 2025. WorkingPaper. Bauer, Michael D. and Eric T. Swanson, “A Reassessment of Monetary Policy Surprises and High- FrequencyIdentification,”in“NBERMacroeconomicsAnnual2022,volume37,”UniversityofChicago Press,2022. and , “An Alternative Explanation for the "Fed Information Effect",” American Economic Review, March2023,113(3),664–700. 30

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION Brennan,ConnorM.,MargaretM.Jacobson,ChristianMatthes,andToddB.Walker,“MonetaryPolicy Shocks:DataorMethods?,”TheB.E.JournalofMacroeconomics,2025,25(2),595–659. Bu,Chunya,JohnRogers,andWenbinWu,“AunifiedmeasureofFedmonetarypolicyshocks,”Journal ofMonetaryEconomics,2021,118,331–349. Bundick, BrentandA.LeeSmith, “The Dynamic Effects of Forward Guidance Shocks,” The Review of EconomicsandStatistics,122020,102(5),946–965. Caldara,DarioandEdwardHerbst,“MonetaryPolicy,RealActivity,andCreditSpreads: Evidencefrom BayesianProxySVARs,”AmericanEconomicJournal:Macroeconomics,2019,11(1),157–192. Calomiris, Charles W., Joanna Harris, Harry Mamaysky, and Cristina Tessari, “Fed Implied Market PricesandRiskPremia,”2022. WorkingPaper. Canova, FabioandFilippoFerroni, “A hitchhiker’s guide to empirical macro models,” 2022. Working Paper. Carlstrom,CharlesT.andTimothyS.Fuerst,“AddingDoubleInertiatoTaylorRulestoImproveAccuracy,”2014. FederalReserveBankofClevelandEconomicCommentaryEC2014-08. Cherry,RennaeandEricTong, “WordsoftheRBNZ:TextualAnalysisofMonetaryPolicyStatements,” 2023. WorkingPaper. Cieslak,Anna,MichaelMcMahon,andHaoPang, “Did Imake myself clear? The Fedandthe market underthe2020monetarypolicyframework,”2025. WorkingPaper. ,StephenHansen,andMichaelMcMahon,“ToughTalk:TheFedandtheRiskPremium,”2024. WorkingPaper. , , ,andHaoPang,“RiskManagementinMonetaryPolicy:ReviewwithAssetPricingImplications,” 2025. WorkingPaper. , , ,andSongXiao,“Policymakers’Uncertainty,”2024. WorkingPaper. Coibion, Olivier and Yuriy Gorodnichenko, “Why Are Target Interest Rate Changes So Persistent?,” AmericanEconomicJournal:Macroeconomics,May2012,4(4),126–62. Crawley,Edmund,LiamGoodwin,MargaretM.Jacobson,andFabianWinkler,“DoubleInertia,Taylor Rules,andMonetaryPolicyGradualism,”2026. WorkingPaper. Curti,FilippoandSophiaKazinnik,“Let’sfaceit:Quantifyingtheimpactofnonverbalcommunication inFOMCpressconferences,”JournalofMonetaryEconomics,2023,139,110–126. Czudaj, Robert L. and Bich Ngoc Nguyen, “ECB’s central bank communication and monetary policy transmission: predictabilityfromtext-basedsentimentindicators?,”MacroeconomicDynamics,2025, 29,e102. 31

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION DePooter,Michiel,“QuestionsandAnswers:TheInformationContentofthePost-FOMCMeetingPress Conference,”2021. FEDSNotes.Washington: BoardofGovernorsoftheFederalReserveSystem,October12,2021,https://doi.org/10.17016/2380-7172.2997. Deng, Yayue, Mohan Xu, and Yao Tang, “FMPAF: How Do Fed Chairs Affect the Financial Market? A Fine-grainedMonetaryPolicyAnalysisFrameworkonTheirLanguage,”2024. Devlin, Jacob, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova, “BERT: Pre-training of Deep BidirectionalTransformersforLanguageUnderstanding,”ProceedingsofNAACL-HLT,2019,pp.4171– 4186. Diebold, Francis and Roberto Mariano, “Comparing Predictive Accuracy,” Journal of Business & EconomicStatistics,1995,13(3),253–63. Doh,Taeyoung,SonghoSong,andShu-KueiYang, “DecipheringFederalReserveCommunicationvia TextAnalysisofAlternativeFOMCStatements,”2025. WorkingPaper. ,SungilKim,andShu-KueiYang,“HowYouSayItMatters: TextAnalysisofFOMCStatementsUsing NaturalLanguageProcessing,”2021. WorkingPaper. Dube,Arindrajit,DanieleGirardi,ÒscarJordà,andAlanM.Taylor, “ALocalProjectionsApproachto Difference-in-Differences,”JournalofAppliedEconometrics,2025,40(7),741–758. Dunn,Wendy,EllenE.Meade,NitishRanjanSinha,andRaakinKabir,“UsingGenerativeAIModelsto UnderstandFOMCMonetaryPolicyDiscussions,”2024.FEDSNotes.Washington:BoardofGovernors oftheFederalReserveSystem,December06,2024,https://doi.org/10.17016/2380-7172.3678. Ehrmann, Michael and Jonathan Talmi, “Starting from a blank page? Semantic similarity in central bankcommunicationandmarketvolatility,”JournalofMonetaryEconomics,2020,111,48–62. andMarcelFratzscher,“ExplainingMonetaryPolicyinPressConferences,”JournalofInternational CentralBanking,2009,5. and , “Purdah—On the Rationale for Central Bank Silence around Policy Meetings,” Journal of Money,CreditandBanking,2009,41(2-3),517–528. Fischer,Eric,RebeccaMcCaughrin,SakethPrazad,andMarkVandergon,“FedTransparencyandPolicyExpectationErrors:ATextAnalysisApproach,”2023. WorkingPaper. Fraccaroli,Nicolo,AlessandroGiovannini,andJean-FrancoisJamet,“CentralBanksinParliaments: A TextAnalysisoftheParliamentaryHearingsoftheBankofEngland,TheEuropeanCentralBank,and theFederalReserve,”2020. WorkingPaper. Gambacorta,Leonardo,ByeungchunKwon,TaejinPark,PietroPatelli,andSonyaZhu,“CB-LMs:languagemodelsforcentralbanking,”2024. 32

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION Gardner,Ben,ChiaraScotti,andClaraVega,“Wordsspeakasloudlyasactions:Centralbankcommunicationandtheresponseofequitypricestomacroeconomicannouncements,”JournalofEconometrics, 2022,231(2),387–409. SpecialIssue:TheEconometricsofMacroeconomicandFinancialData. Gati,LauraandAmyHandlan,“MonetaryCommunicationRules,”2023. WorkingPaper. Gertler,MarkandPeterKaradi,“MonetaryPolicySurprises,CreditCosts,andEconomicActivity,”AEJ: Macroeconomics,2015,7,44–76. Gilchrist,SimonandEgonZakrajšek,“CreditSpreadsandBusinessCycleFluctuations,”AmericanEconomicReview,June2012,102(4),1692–1720. Gnan,Phillipp,MaximilianSchleritzko,MaikSchmeling,andChristianWagner,“DecipheringMonetaryPolicyShocks,”2025. WorkingPaper. Gordon, Matthew V. and Kurt G. Lunsford, “The effects of the Federal Reserve Chair’s testimony on interestratesandstockprices,”EconomicsLetters,2024,235,111537. Gorodnichenko,Yuriy,ThoPham,andOleksandrTalavera,“TheVoiceofMonetaryPolicy,”American EconomicReview,February2023,113(2),548–84. Granziera,Eleanora,VegardH.Larsen,GretaMeggiorini,andLeonardoMelosi,“SpeakingofInflation: TheInfluenceofFedSpeechesonExpectations,”2025. WorkingPaper. Grotteria,MarcoandJulesvanBinsbergen,“MonetaryPolicyWedgesandtheLong-termLiabilitiesof HouseholdsandFirms,”2025. WorkingPaper. Gürkaynak,RefetS.,BrianP.Sack,andEricT.Swanson, “DoActionsSpeakLouderthanWords? The ResponseofAssetPricestoMonetaryPolicyActionsandStatements,”InternationalJournalofCentral Banking,2005,1,55–93. Gómez-Cram, Roberto and Marco Grotteria, “Real-time price discovery via verbal communication: MethodandapplicationtoFedspeak,”JournalofFinancialEconomics,2022,143(3),993–1025. Handlan, Amy, “FedSpeak Matters: Statement Similarity and Monetary Policy Expectations,” 2022. WorkingPaper. , “Text Shocks and Monetary Surprises: Text Analysis of FOMC Statements with Machine Learning,” 2022. WorkingPaper. Hansen, AnneLundgaardandSophiaKazinnik, “Can ChatGPT Decipher Fedspeak?,” 2024. Working Paper. Hansen,StephenandMichaelMcMahon,“Shockinglanguage: Understandingthemacroeconomiceffectsofcentralbankcommunication,”JournalofInternationalEconomics,2016,99,S114–S133. 38th AnnualNBERInternationalSeminaronMacroeconomics. 33

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION , ,andAndreaPrat,“TransparencyandDeliberationWithintheFOMC:AComputationalLinguistics Approach,”TheQuarterlyJournalofEconomics,102017,133(2),801–870. He, Pengcheng, JianfengGao, andWeizhuChen, “DeBERTaV3: Improving DeBERTa using ELECTRA- StylePre-TrainingwithGradient-DisentangledEmbeddingSharing,”CoRR,2023,abs/2111.09543. ,XiaodongLiu,JianfengGao,andWeizhuChen, “DeBERTa: Decoding-enhancedBERTwithDisentangledAttention,”CoRR,2020,abs/2006.03654. , , , and , “DeBERTa: Decoding-enhanced BERT with Disentangled Attention,” International ConferenceonLearningRepresentations(ICLR),2021. Jarocinski,MarekandPeterKaradi,“DisentanglingMonetaryPolicy,CentralBankInformation,andFed ResponsetoNews,”2025. WorkingPaper. Kazinnik,SophiaandTaraSinclair,“FOMCinSiico:AMulti-AgentSystemforMonetaryPolicyDecision Modeling,”2026. Workingpaper. Kuttner,KennethN.,“MonetaryPolicySurprisesandInterestRates: EvidencefromtheFedFundsFuturesMarket,”JournalofMonetaryEconomics,2001,47,523–544. Kypraios, Emmanuel, Juan Arismendi-Zambrano, and Alessia Paccagnini, “Moderating Uncertainty with Sentiment: The Case of the Federal Reserve Chair Communication,” Academy of Management Proceedings,2024,2024(1),14009. Lamla,MichaelJ.andDmitriV.Vinogradov,“Centralbankannouncements:Bignewsforlittlepeople?,” JournalofMonetaryEconomics,2019,108,21–38. Lee,Eunkyung,“TheTransmissionofMonetaryPolicytoCorporateInvestment:theRoleofLoanRenegotiation,”2026. Forthcoming,AEJ:Macroeconomics. Lucca,DavidO.andFrancescoTrebbi, “MeasuringCentralBankCommunication: AnAutomatedApproachwithApplicationtoFOMCStatements,”2011. WorkingPaper. Miranda-Agrippino,SilviaandGiovanniRicco,“TheTransmissionofMonetaryPolicyShocks,”AmericanEconomicJournal:Macroeconomics,July2021,13(3),74–107. and ,“IdentificationwithExternalInstrumentsinStructuralVARs,”JournalofMonetaryEconomics, 2023,135(C),1–19. Mumtaz,Haroon,JumanaSaleheen,andRoxaneSpitznagel,“KeepItSimple:CentralBankCommunicationandAssetPrices,”2024. WorkingPaper. Munday,TimandJamesBrookes,“MarkmyWords: theTransmissionofCentralBankCommunication totheGeneralPublicviathePrintMedia,”2021. WorkingPaper. 34

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION Nakamura, EmiandJónSteinsson, “High-Frequency Identification of Monetary Non-Neutrality: The InformationEffect,”QuarterlyJournalofEconomics,2018,133,1283–1330. Narain, Namrata and Kunal Sangani, “The Market Impact of Fed Communications: The Role of the PressConference,”2026. Forthcoming,InternationalJournalofCentralBanking. Newey,WhitneyK.andKennethD.West,“ASimple,PositiveSemi-Definite,HeteroskedasticityandAutocorrelationConsistentCovarianceMatrix,”Econometrica,1987,55(3),703–708. Osowska,EwelinaandPiotrWojcik,“PredictingthereactionoffinancialmarketstoFederalOpenMarketCommitteepost-meetingstatements,”DigitalFinance,2024,6,145–175. Parle,Conor, “ThefinancialmarketimpactofECBmonetarypolicypressconferences—Atextbased approach,”EuropeanJournalofPoliticalEconomy,2022,74,102230. Pfeifer, Moritz and Vincent P. Marohl, “CentralBankRoBERTa: A fine-tuned large language model for centralbankcommunications,”TheJournalofFinanceandDataScience,2023,9,100114. Picault,MatthieuandThomasRenault,“Wordsarenotallcreatedequal: AnewmeasureofECBcommunication,”JournalofInternationalMoneyandFinance,2017,79,136–156. Ramey,V.A.,“Chapter2-MacroeconomicShocksandTheirPropagation,”inJohnB.TaylorandHarald Uhlig,eds.,HandbookofMacroeconomics,Vol.2,Elsevier,2016,pp.71–162. Rosa,Carlo,“Fedspeak:WhoMovesU.S.AssetPrices,”2016. Schmanski,Bennett,ChiaraScotti,ClaraVega,andHediBenamar,“FedCommunication,News,Twitter,andEchoChambers,”2023. WorkingPaper. Sekkel,RodrigoandXuZhan,“Speeches,PressConferencesandMinutes: TheInternationalTransmissionofFederalReserveCommunication,”2024. WorkingPaper. Shah,Agam,SuvanPaturi,andSudheerChava,“TrillionDollarWords:ANewFinancialDataset,Task& MarketAnalysis,”2023. Silva, Thiago Christiano, Kei Moriya, and Romain M Veyrune, “From Text to Quantified Insights - A Large-ScaleLLMAnalysisofCentralBankCommunication,”2025. WorkingPaper. Stock,JamesandMarkWatson,“IdentificationandEstimationofDynamicCausalEffectsinMacroeconomicsUsingExternalInstruments,”EconomicJournal,2018,128(610),917–948. Swanson, Eric T. and Vishuddi Jayawickrema, “Speeches by the Fed Chair are more Important than FOMCAnnouncements:anImprovedHigh-FrequencyMeasureofU.S.MonetaryPolicyShocks,”2024. WorkingPaper. Taylor,JohnB., “Therobustnessandefficiencyofmonetarypolicyrulesasguidelinesforinterestrate settingbytheEuropeancentralbank,”JournalofMonetaryEconomics,1999,43(3),655–679. 35

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION Tobback,Ellen,StefanoNardelli,andDavidMartens,“BetweenHawksandDoves: MeasuringCentral BankCommunication,”2017. WorkingPaper. Vaswani, Ashish, NoamShazeer, NikiParmar, JakobUszkoreit, LlionJones, AidanNGomez, Łukasz Kaiser,andIlliaPolosukhin, “AttentionisAllYouNeed,” AdvancesinNeuralInformationProcessing Systems,2017,30,5998–6008. Wessel,DavidandSamBoocker,“GradingFedcommunications: A2024surveyofFedwatchers,”2024. BrookingsEconomicResearch,https://www.brookings.edu/articles/grading-fed-communications-a- 2024-survey-of-fed-watchers/. Wu,JingCynthiaandFanDoraXia,“MeasuringtheMacroeconomicImpactofMonetaryPolicyatthe ZeroLowerBound,”JournalofMoney,CreditandBanking,2016,48(2-3),253–291. 36

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION A APPENDIX: DATES AND TIMES WedeterminethedatesandtimesofthemonetarypolicyannouncementsfollowingSwansonandJayawickrema (2024). In particular, we collected the dates, times, and high-frequency, intraday asset price changesfromJune1995toApril2025for:i)unscheduledandscheduledFOMCstatements,ii)speeches by the Federal Reserve chair, iii) Congressional testimony by the Federal Reserve Chair, and iv) post- FOMC-meetingpressconferencesfromJanuary2011toApril2025. ThestartandendtimesofCongressionaltestimonystarttimeswereconfirmedandendtimesobtainedfromGordonandLunsford(2024). Transcribed text is readily available starting in 1999 for FOMC statements and 2008 for Congressional testimoniesandspeechesbytheFedChair. FOMCStatementsandMinutesReleases WecollectedtheschedulesofFOMCannouncementsandmeetingminutesreleasesfromJanuary1995to April2025throughtheFederalReserveBoard’spublicwebsite.TheFederalReserveannouncesdecisions onmonetarypolicyorregulationstothepress,andallsuchstatementscanbefoundinthepressrelease sectionoftheofficialwebsite.ThedatesandtimeswhenFOMCstatementswerereleasedandwhenthe minutesofeachmeetingareavailablefrom2006onwardontheFederalReserve’spublicwebsite. For events from 1995 to 2005, we obtained the dates through the website and supplemented the timinginformationfromGürkaynaketal.(2005)andSwansonandJayawickrema(2024).Anyremaining missinginformationwasverifiedthroughFactivasearchesofnewsarchives. FOMCstatementsareissuedonthelastdayofthemeeting.From1999to2010,withafewexceptions, these statements were announced at 2:15 PM Eastern Time. In 2011, the introduction of post-FOMC pressconferencesledtochangesinthistiming. Forthetwoyearsfrom2011to2012,ifapost-meeting press conference occurred, the FOMC issued a press release at 12:30 PM Eastern Time; for meetings withoutpressconferences,theFedmaintainedthe2:15PMreleasetime.From2013onwards,theschedulewasadjusted,andtheFOMCgenerallyannounceditsmonetarypolicydecisionsat2:00PMEastern Time. WedropallnotationalmeetingsincludingAugust27,2000,October4,2019,March11,2008,and August10,2007. Followingmuchoftheliterature,wedropthemeetingsafter9/11. WedroptheMarch 15,2020unscheduledmeetingasitoccurredonaSundayanditisdifficulttosourcetrades. Post-FOMCpressconferences All materials related to post-FOMC-meeting press conferences are available on the Federal Reserve’s publicw website. These press conferences began in April 2011, so unlike other events, the data exists only from 2011 onwards. Currently, press conferences are held after almost every FOMC meeting, resultingineightscheduledpressconferencesperyear. However,intheearlyyears(2011to2012),press conferenceswereheldeveryothermeeting,occurringfourtimesannually. From2011to2012,thepressconferencesstartedat2:15PM,andfrom2013,theybeganat2:30PM. TheFederalReserve’swebsitepostsvideosofpressconferencesalongwithtranscripts. Thedurationof eachconferencecanbeverifiedthroughthesevideos.Thepressconferenceslastbetween14minutesat theshortestand76minutesatthelongest,withanaveragedurationof53minutes. 37

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION SpeechesbytheFederalReserveChair TheschedulesforspeechesandCongressionaltestimoniesoftheFedChairareavailableontheFederal ReserveBoard’spublicwebsitefrom2006onward. Thewebsitepoststranscriptsofthespeeches,which include the time each document was released to the public. We extract the timing information from thesetranscripts. However,transcriptswerenotpostedforspeechesandCongressionaltestimonyfrom 2006to2009,makingitimpossibletoconfirmprecisetiming. Therefore,forspeechesandtestimonies prior to 2009, we verified the timing data using the FRASER library of the Federal Reserve Bank of St. Louis. Additionally, speech materials from before 2006 that are not available on the Federal Reserve’s websitecanalsobefoundinFRASER. FRASER contains transcripts of all speeches and testimonies by the Federal Reserve Chair and includethestarttimeofthespeech,fromwhichwewereabletoextractthetiminginformation.Occasionally,forCongressionaltestimoniesratherthanregularspeeches,thetimemaynotbeincluded. Inthese cases,weverifiedthetestimonytimethroughnewssourcesusingFactivasearches. 38

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION B APPENDIX: HIGH-FREQUENCY MARKET REACTIONS B.1 MARKETREACTIONS: PRESSCONFERENCEVS. FOMCSTATEMENT FigureB.1plotsthemarketreactionsforthefirstyearofinterestratefutures,EDSR1toEDSR4,whichcorrespondtotheithEurodollar/(i+ 1)thSOFRfuturesfori =1,...,4quartersahead. FigureB.1:Marketreactionstothepressconferencevs.FOMCstatement,basispoints. ThefigureshowstheregressionMarketReactiont,Press =α+MarketReactiont,FOMC +ϵ tbyFedchairsforthe1stEurodollar/2nd SOFRfuturethroughthe4thEurodollar/5thSOFRfuture.Marketreactionsareintradaychangesoftradeprices20minutesafter acommunicationeventrelativetotradeprices10minutesbeforefortheithEurodollar/(i+1)thSOFRfuture(ESRi),which correspondstotheexpectedpolicyratei quartersaheadfori =1,...,4. BernankeorYellenisApril2011toJanuary2018and PowellisfromFebruary2018toApril2025.Shadedbandsare90%errorbands. 39

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION B.2 MARKET DISAGREEMENTS Figure B.2 shows that the rate of disagreement is increasing in the maturity of interest rate futures. The first expiring federal funds future FF1 has no disagreements, as changesinthisfuturearedirectlylinkedtothefederalfundsrateanditisthereforenotaffectedbythe press conference. By contrast, EDSR4 has about one-fifth to one-third of all FOMC statement/press conferencepairsindisagreeinginthesignofthemarketsurprise. (a)Threebasispointthresholdfordisagreements (b)Fivebasispointthresholdfordisagreements FigureB.2:DisagreementinmarketreactionstoFOMCstatementsandpressconferences AdisagreementisanintradaychangethatismorethanthreeorfivebasispointswithdifferentsignsforanFOMCstatement andapressconference. Anagreementisallotherobservations. Marketreactionsarethetradeprices20minutesafterthe FOMCstatementreleaseorstartofthepressconferencelessthetradeprices10minutesprior,respectively.Thesampleisfrom April2011toApril2025.FFi fori=1,...,4istheithfederalfundsratefuturewherei correspondstothemonthahead,i=0is thecurrentmonth. TheithEurodollar/(i+1)thSOFRfuture(ESRi),whichcorrespondstotheexpectedpolicyratei quarters aheadfori=1,...,4. Marketreactionsareintradaychangesoftradeprices20minutesafteracommunicationeventrela- 40

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION tivetotradeprices10minutesbeforefortheithEurodollar/(i+1)thSOFRfuture(ESRi),whichcorrespondstotheexpectedpolicyratei quartersaheadfori =1,...,4. FigureB.3:Comparingdisagreementsinmarketreactionsforspeechesandthepressconferences Pre-2011istheperiodbeforetheintroductionofthepressconferenceinApril2011. Apre-2011disagreementisamorethan threebasispointintradaychangesurroundingaspeechbytheChairintheweekfollowinganFOMCstatementthatisdifferent insignfromtheintradaychangearoundtheprecedingFOMCstatement.Apost-2011disagreementisanintradaychangethat ismorethanthreeorfivebasispointswithdifferentsignsforanFOMCstatementandapressconference.Anagreementisall otherobservations. Thepre-2011samplehas32observationsandthepost-2011samplehas76observations. Thesampleis fromJanuary1995toApril2025.FFifori=1,...,4istheithfederalfundsratefuturewhereicorrespondstothemonthahead, i =0isthecurrentmonth. TheithEurodollar/(i+1)thSOFRfuture(ESRi),whichcorrespondstotheexpectedpolicyratei quartersaheadfori=1,...,4. 41

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION C APPENDIX: SENTIMENT SCORES C.1 RAW SENTIMENT SCORES ThisappendixshowsthesentimentscoresofFOMCstatements,press conferences, Congressional testimonies, and Fed Chair speeches. Figures C.1 and C.2 show the Fed’s statutory goals—inflation and unemployment—along with its main instrument in the form of monetary policy. The figure shows that the share of text devoted to these topics and the hawkish or dovish sentiment—that is, sentiment pertaining to tightening or loosening, respectively—varies across time. Theprevalenceofbothhawkishanddovishsentimentvariesacrosstypeofcommunicationandisdiscussedinthemaintext. ToassurethattheLLMcapturestheFed’smonetarypolicystancewenoteafewepisodes. First,the Fed’s2022-2024tighteningcycleaimedtobringinflationbacktoits2percentpolicyobjectiveandfigures C.1 and C.2 show that hawkish sentiment about inflation consisted of about 40 percent of the FOMC statement communication, the highest hawkish inflation sentiment recorded in our sample. Second, from 2008 to about 2015 about 30 percent of the FOMC statement focused on dovish unemployment sentiment, which is consistent with the Fed’s objective of returning the economy to full employment. Finally,monetarypolicyisdetectedashawkishduringthe2015-2018and2022-2024tighteningcycles. 42

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION Sentiment of FOlVIC Statements by Topic, Percent of Text 40 0 Inflation -40 40 -40 40 0 l\lionetary Policy -40 2008 2014 2019 2025 - Hawkish - Dovish FigureC.1:Sentimentbytopic,percentoftext. Hawkishsentimentpertainspolicytighteninganddovishtoloosening. They-axisisthepercentageofthetextthatisdevoted tothattopic. 43

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION FigureC.2:Sentimentbytopic,percentoftext. Hawkishsentimentpertainspolicytighteninganddovishtoloosening. They-axisisthepercentageofthetextthatisdevoted tothattopic. 44

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION C.2 NET SENTIMENT AND MONETARY POLICY, SENTENCE-LEVEL Thefigurebelowisthesameasfigure 4 in the main text except for the top panel that plots sentence-level net sentiment for the FOMC statementinsteadofparagraph-levelnetsentiment. fi '0.$
4UBUFNFOUT   )BXLJTI  fl
%PWJTI         fi $IBJS
4QFFDIFT   )BXLJTI  fl
%PWJTI         fi 1SFTT
$POGFSFODFT   )BXLJTI  fl
%PWJTI        



 'FEFSBM
GVOET
SBUF
QFSDFOU















       FigureC.3: NetsentimentscoresatthesentencelevelfortheFOMCstatementandtheparagraphlevel forthepressconferenceandspeechesalongwiththetargetfederalfundsrate. Hawkishsentimentpertainspolicytighteninganddovishtoloosening.Forcommunicationtypej∈{Pressconference,Speech, FOMCstatements},NetSentimentj,t =(Hawkishj,t −Dovishj,t)/(Hawkishj,t +Dovishj,t).Thefederalfundsrateisthetarget aftereachFOMCmeeting,seeAppendixGfordetailsonitsconstruction. AppendixAdetailstheavailabilityandsourcesof eachtypeofcommunication. 45

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION C.3 ROBUSTNESS OF LOCAL PROJECTION SAMPLES While press conferences now occur after every meeting,theyonlyoccurredaftereveryothermeetingfrom2011to2019. PanelC.4acomparesimpulse responsesfromnetsentimentofFOMCstatementstothefederalfundsrateforthesampleofallmeetingsfromJanuary2011toApril2025tothatshowninthemaintextinfigure5basedonthesampleof meetingsfollowedbypressconferences.ThefullsampleinpanelC.4ahas121meetings,whilethesample in the main text in figures 5-6 has 83. The impulse responses are nearly identical, which suggests that joint estimation conditioned on meetings with only press conference are robust to estimation on thefullsampleofmeetings.PanelC.4bextendsthestartofthesamplefrom2011to1999(whenwehave sentimentfortheFOMCstatements)andlikewiseshowspointestimatesthatarequitesimilartothose showninthemaintext.           .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU
0OMZ
.FFUJOHT
XJUI
1SFTT
$POGFSFODFT /FU
4FOUJNFOU
'0.$
4UBUFNFOU
"MM
.FFUJOHT (a) Comparingsamples           .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU

UP

0OMZ
.FFUJOHT
X
1SFTT
$POG /FU
4FOUJNFOU
'0.$
4UBUFNFOU

UP

"MM
.FFUJOHT (b) ComparingTimePeriodSamples FigureC.4:Impulseresponseofthefederalfundsratetonetsentiment,90%errorbandswithtwoNewey- Westlags. Thefigureplotsβˆh 1 fromthelocalprojection,FederalFundsRate t+h =βh 0 +βh 1 NetSentimentFOMC,t +ϵ FOMC,tinequation(8). WheretistheFOMCmeetingfrequency,h=0,...,24.NetSentimentj,t =(Hawkishj,t −Dovishj,t)/(Hawkishj,t +Dovishj,t).The sampleisfromApril2011toApril2025inpanelC.4a.Thet+hfederalfundsrateincludesallmeetingssuchthatthe8-meetingaheadpointestimatescorrespondstoaone-year-aheadhorizon. 46

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION Tocheckwhetherornotpredictivecorrelationsofnetsentimentwiththefederalfundsratedepend onoccurrenceofpressconferences,weextendthestartdateofthesamplefrom2011backto1999(when wehavesentimentfortheFOMCstatements). Theindicator1Press isequaltoonewhentherewasa t postmeetingpressconference. FederalFundsRate t+h =βh 0 +βh 1 NetSentiment FOMC,t +βh 2 1Press t +... (20) βh 3 (cid:161) NetSentiment FOMC,t ×1Press t (cid:162)+ε t+h Figure C.5 shows that the coefficient on net sentiment of the FOMC statement βˆh is positive and 1 significant and slightly higher in magnitude then its estimates shown in figures 5 and C.4. The coefficients on the press conference and its interaction with net sentiment—βˆh and βˆh, respectively—are 2 3 insignificantoronlyonthemarginofsignificancefrom0. Takentogether,theseestimatessuggeststhat theavailabilityofpressconferencesasanadditionalcommunicationtooldoesnotsignificantlyalterthe predictivecorrelationsoftheFOMCstatementwiththefederalfundsrate.           .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU *OUFSBDUJPO 1SFTT
$POG
*OEJDBUPS FigureC.5:Impulseresponseofthefederalfundsratetonetsentiment,90%errorbandswithtwoNewey- Westlags. Thefigureplotsβˆh j forj=1,...,3fromthelocalprojection,FederalFundsRate t+h =βh 0 +βh 1 NetSentimentFOMC,t +βh 2 1Presst + βh 3 (cid:161) NetSentimentFOMC,t ×1Presst (cid:162)+ε t+h in equation (8). Where t is the FOMC meeting frequency, h = 0,...,24 and NetSentimentj,t =(Hawkishj,t −Dovishj,t)/(Hawkishj,t +Dovishj,t). The sample is from January 1999 to April 2025 and 1Presst indicatesiftherewasapost-meetingpressconferenceornot. Thet+hfederalfundsrateincludesallmeetingssuch thatthe8-meeting-aheadpointestimatescorrespondstoaone-year-aheadhorizon. 47

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION C.4 ROBUSTNESSOFLOCALPROJECTIONSWITHTHELAGGEDFEDERALFUNDSRATE FigureC.6shows thatthepointestimatesinfigures5arerobusttocontrollingforthelaggedfederalfundsrate.           .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU /FU
4FOUJNFOU
1SFTT
$POGFSFODF Figure C.6: Impulse response of the federal funds rate to net sentiment with inertia, 90% error bands withtwoNewey-Westlags. Thefigureplotsβˆh 1 fromthelocalprojection,FederalFundsRate t+h =βh 0 +αhFederalFundsRatet−1 +βh 1 NetSentimentj,t + ϵ j,tforj∈{FOMCstatement,Press}.WheretistheFOMCmeetingfrequencyandh=0,...,24.NetSentimentj,t =(Hawkishj,t − Dovishj,t)/(Hawkishj,t +Dovishj,t). ThesampleisfromApril2011toApril2025andisforFOMCmeetingsfollowedbypress conferences(everyotherFOMCmeetingfrom2011to2019andeverymeetingthereafter).Thet+hfederalfundsrateincludes allmeetingssuchthatthe8-meeting-aheadpointestimatescorrespondstoaone-year-aheadhorizon. 48

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION C.5 ROBUSTNESSOFLOCALPROJECTIONSWITHINTERACTIONTERMS FigureC.7showsthepointestimatesofequations9and10withaninteractionterm(NetSentiment ×NetSentiment )in- FOMC,t Press,t cludedineachspecification.            .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU *OUFSBDUJPO /FU
4FOUJNFOU
1SFTT
$POGFSFODF (a)Jointestimationwithinteraction           .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU *OUFSBDUJPO /FU
4FOUJNFOU
1SFTT
$POGFSFODF (b)Jointestimationwithinteractionandinertia FigureC.7:Impulseresponseofthefederalfundsratetonetsentimentwithinteractionterms,90%error bandswithtwoNewey-Westlags. Panel C.7a plots {βˆh 1 ,βˆh 2 ,βˆh 3 } from FederalFundsRate t+h = βh 0 +βh 1 NetSentimentFOMC,t +βh 2 NetSentimentPress,t + βh 3 (NetSentimentFOMC,t ×NetSentimentPress,t)+ϵ t. Panel C.7b plots {βˆh 1 ,βˆh 2 ,βˆh 3 } from FederalFundsRate t+h = βh 0 + αhFederalFundsRatet−1 + βh 1 NetSentimentFOMC,t + βh 2 NetSentimentPress,t + βh 3 (NetSentimentFOMC,t × NetSentimentPress,t)+ϵ t. Where t is the FOMC meeting frequency and h = 0,...,24. NetSentimentj,t = (Hawkishj,t − Dovishj,t)/(Hawkishj,t +Dovishj,t). ThesampleisfromApril2011toApril2025andisforFOMCmeetingsfollowedbypress conferences(everyotherFOMCmeetingfrom2011to2019andeverymeetingthereafter).Thet+hfederalfundsrateincludes allmeetingssuchthatthe8-meeting-aheadpointestimatescorrespondstoaone-year-aheadhorizon. 49

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION C.6 ROBUSTNESS OF LOCAL PROJECTIONS WITH NET SENTIMENT OF SPEECHES FigureC.8showsthe pointestimatesinfigure5alongwithpointestimatesforspeechesinthelocalprojectioninequation(8): FederalFundsRate t+h =βh 0 +βh 1 NetSentiment j,t +ϵ j,t (8) Wheret istheFOMCmeetingfrequency,h=0,..,20,and j ∈{FOMCstatement,Press,Chairspeech}.           .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU /FU
4FOUJNFOU
1SFTT
$POGFSFODF /FU
4FOUJNFOU
4QFFDIFT FigureC.8: Impulseresponseofthefederalfundsratetonetsentimentwithspeeches,90%errorbands withtwoNewey-Westlags. Thefigureplotsβˆh 1 fromthelocalprojection,FederalFundsRate t+h =βh 0 +βh 1 NetSentimentj,t +ϵ j,t inequation(8).Wheret istheFOMCmeetingfrequency,h=0,...,24,andj∈{FOMCstatement,Press,Chairspeech}.NetSentimentj,t =(Hawkishj,t − Dovishj,t)/(Hawkishj,t +Dovishj,t). ThesampleisfromApril2011toApril2025andisforFOMCmeetingsfollowedbypress conferences(everyotherFOMCmeetingfrom2011to2019andeverymeetingthereafter).Thet+hfederalfundsrateincludes allmeetingssuchthatthe8-meeting-aheadpointestimatescorrespondstoaone-year-aheadhorizon. 50

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION C.7 DOUBLE INERTIA LOCAL PROJECTIONS FigureC.9includestwolagsofthefederalfundsrateinto equation, asadvocatedbyCrawleyetal.(2026), CarlstromandFuerst(2014), andCoibionandGorodnichenko(2012).Theresultarelittlechangedfromthoseshowninpanel6b,albeitwiththepositiveand significantpointestimatesofnetsentimentofthepressconferenceshorteningfromabout22meetings aheadto19meetingsahead. TheresultsaresimilarifFederalFundsRate t−1 −FederalFundsRate t−2 is includedinsteadofFederalFundsRate t−2 . FederalFundsRate t+h =βh 0 +βh 1 NetSentiment FOMC,t +βh 2 NetSentiment Press,t +... ...αh 1 FederalFundsRate t−1 +αh 2 FederalFundsRate t−2 +ϵ t (21)           .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU /FU
4FOUJNFOU
1SFTT
$POGFSFODF FigureC.9: Impulseresponseofthefederalfundsratetonetsentimentwithdoubleinertia, 90%error bandswithtwoNewey-Westlags. The figure plots βˆh 1 and βˆh 2 from the local projection, FederalFundsRate t+h = βh 0 +βh 1 NetSentimentFOMC,t + βh 2 NetSentimentPress,t +αh 1 FederalFundsRatet−1 +αh 2 FederalFundsRatet−2 +ϵ tinequation(10).WheretistheFOMCmeetingfrequencyandh=0,...,24.NetSentimentj,t =(Hawkishj,t −Dovishj,t)/(Hawkishj,t +Dovishj,t).ThesampleisfromApril 2011toApril2025andisforFOMCmeetingsfollowedbypressconferences(everyotherFOMCmeetingfrom2011to2019and everymeetingthereafter). Thet+hfederalfundsrateincludesallmeetingssuchthatthe8-meeting-aheadpointestimates correspondstoaone-year-aheadhorizon. 51

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION C.8 ROBUSTNESSOFLOCALPROJECTIONSWITHSENTENCE-LEVELFOMCSTATEMENTSENTIMENT FigureC.10showsestimatesinsection4.2andappendixCwithsentence-levelFOMCstatementsentiment.            .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU /FU
4FOUJNFOU
1SFTT
$POGFSFODF (a)Separateestimation                       .FFUJOHT
"IFBE .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU /FU
4FOUJNFOU
'0.$
4UBUFNFOU *OUFSBDUJPO /FU
4FOUJNFOU
1SFTT
$POGFSFODF /FU
4FOUJNFOU
1SFTT
$POGFSFODF (b)Jointestimation (c)Jointestimationwithinteraction                       .FFUJOHT
"IFBE .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU /FU
4FOUJNFOU
'0.$
4UBUFNFOU *OUFSBDUJPO /FU
4FOUJNFOU
1SFTT
$POGFSFODF /FU
4FOUJNFOU
1SFTT
$POGFSFODF (d)Jointestimationwithinertia (e)Jointestimationwithinteractionandinertia FigureC.10: Impulseresponseofthefederalfundsratetonetsentimentwithsentence-levelsentiment forthestatement,90%errorbandswithtwoNewey-Westlags. Panel C.10a plots βˆh 1 from the local projection, FederalFundsRate t+h = βh 0 +βh 1 NetSentimentj,t +ϵ j,t for j ∈ {FOMCstatementsentence-level,Press}.WheretistheFOMCmeetingfrequencyandh=0,...,24.PanelC.10bplotsβˆhandβˆh 1 2 fromthelocalprojection,FederalFundsRate t+h =βh 0 +βh 1 NetSentimentSentence-levelFOMC,t +βh 2 NetSentimentPress,t +ϵ t. Panel C.10c plots {βˆh 1 ,βˆh 2 ,βˆh 3 } from FederalFundsRate t+h = βh 0 + βh 1 NetSentimentSentence-levelFOMC,t + βh 2 NetSentimentPress,t +(NetSentimentFOMC,t ×NetSentimentPress,t)+ϵ t. Panel C.10d plots βˆh 1 and βˆh 2 from the local projection, FederalFundsRate t+h = βh 0 + αhFederalFundsRatet−1 + βh 1 NetSentimentSentence-levelFOMC,t + βh 2 NetSentimentPress,t +ϵ t. Panel C.10e plots {βˆh 1 ,βˆh 2 ,βˆh 3 } from FederalFundsRate t+h =βh 0 +αhFederalFundsRatet−1 + βh 1 NetSentimentSentence-levelFOMC,t +βh 2 NetSentimentPress,t +(NetSentimentFOMC,t ×NetSentimentPress,t)+ϵ t.Where t is the FOMC meeting frequency and h =0,...,20. NetSentimentj,t =(Hawkishj,t −Dovishj,t)/(Hawkishj,t +Dovishj,t). The sample is from April 2011 to April 2025 and is for FOMC meetings followed by press conferences (every other FOMC meetingfrom2011to2019andeverymeetingthereafter). Thet+h federalfundsrateincludesallmeetingssuchthatthe 8-meeting-aheadpointestimatescorrespondstoaone-year-aheadhorizon. 52

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION C.9 ROBUSTNESS OF LOCAL PROJECTIONS WITH LAGGED VIX AS A CONTROL While Lee (2026) uses laggedrealGDPgrowthandtheVIXvolatilityindexascontrolsformacroeconomicconditions, weare unabletousetheformerbecausethetimingofquarterlyGDPisdifficulttolineupattheFOMCmeeting frequency. Insteadweuselaggedmonthlygrowthinindustrialproductiontocontrolforoutput. Even with these macroeconomic controls, figure C.11 shows impulse responses that are similar to those in section 4.2 and appendix C. Results in figure C.11 are robust to using the average of the VIX five days beforeanFOMCannouncementinsteadofthevaluethedaybeforeanFOMCannouncement. 53

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION                .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU /FU
4FOUJNFOU
1SFTT
$POGFSFODF (a)Separateestimation                               .FFUJOHT
"IFBE .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU /FU
4FOUJNFOU
'0.$
4UBUFNFOU *OUFSBDUJPO /FU
4FOUJNFOU
1SFTT
$POGFSFODF /FU
4FOUJNFOU
1SFTT
$POGFSFODF (b)Jointestimation (c)Jointestimationwithinteraction                       .FFUJOHT
"IFBE .FFUJOHT
"IFBE /FU
4FOUJNFOU
'0.$
4UBUFNFOU /FU
4FOUJNFOU
'0.$
4UBUFNFOU *OUFSBDUJPO /FU
4FOUJNFOU
1SFTT
$POGFSFODF /FU
4FOUJNFOU
1SFTT
$POGFSFODF (d)Jointestimationwithinertia (e)Jointestimationwithinteractionandinertia FigureC.11: Impulseresponseofthefederalfundsratetonetsentimentwithmacrocontrols,90%error bandswithtwoNewey-Westlags. Panel C.11a plots βˆh 1 from the local projection, FederalFundsRate t+h = βh 0 +βh 1 NetSentimentj,t +γh 1 VIXτ−1 + γh 2 ∆logIPT−1 +ϵ j,t for j ∈ {FOMC,Press}. Where t is the FOMC meeting frequency, τ−1 is the day before FOMC meeting t, T −1 is the month prior to FOMC meeting t, and h = 0,...,24. Panel C.11b plots βˆh 1 and βˆh 2 from the local projection, FederalFundsRate t+h = βh 0 +βh 1 NetSentimentFOMC,t +βh 2 NetSentimentPress,t + γh 1 VIXτ−1 +γh 2 ∆logIPT−1 +ϵ t. PanelC.11cplots{βˆh 1 ,βˆh 2 ,βˆh 3 }fromFederalFundsRate t+h =βh 0 +βh 1 NetSentimentFOMC,t + βh 2 NetSentimentPress,t +(NetSentimentFOMC,t ×NetSentimentPress,t)+γh 1 VIXτ−1 +γh 2 ∆logIPT−1 +ϵ t. PanelC.11dplots βˆh 1 and βˆh 2 from the local projection, FederalFundsRate t+h =βh 0 +αhFederalFundsRatet−1 +βh 1 NetSentimentFOMC,t + βh 2 NetSentimentPress,t +γh 1 VIXτ−1 +γh 2 ∆logIPT−1 +ϵ t. Panel C.11e plots {βˆh 1 ,βˆh 2 ,βˆh 3 } from FederalFundsRate t+h = βh 0 +αhFederalFundsRatet−1 +βh 1 NetSentimentSentence-levelFOMC,t +βh 2 NetSentimentPress,t +(NetSentimentFOMC,t × NetSentimentPress,t)+γh 1 VIXτ−1 +γh 2 ∆logIPT−1 +ϵ t.NetSentimentj,t =(Hawkishj,t −Dovishj,t)/(Hawkishj,t +Dovishj,t). 54

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION C.10 DIEBOLD AND MARIANO (1995) TESTS OF FORECAST ACCURACY Table 1 shows that the out of samplenetsentimentfederalfundsrateforecastshowninequation(11)isstatisticallysignificantfrom predictionsgivenbyequations(12)-(14). Horizon,h Taylor(1999) InertialTaylor(1999) Laggedfederal fundsrate 1 −5.90 ∗∗∗ −5.54 ∗∗∗ −3.06 ∗∗∗ 2 −4.09 ∗∗∗ −3.54 ∗∗∗ −2.20 ∗∗ 3 −3.45 ∗∗∗∗ −2.52 ∗∗∗ −1.74 ∗∗ 4 −3.2 ∗∗∗ −2.29 ∗∗ −1.65 ∗∗ 5 −3.01 ∗∗∗ −2.03 ∗∗ −1.62 ∗ 6 −2.86 ∗∗∗ −1.87 ∗∗ −1.65 ∗ 7 −2.74 ∗∗∗ −1.81 ∗∗ −1.75 ∗∗ 8 −2.65 ∗∗∗ −1.77 ∗∗ −1.77 ∗∗ Table1:DieboldandMariano(1995)Statisticscomparingtheforecastofthefederalfundsrateusingnet sentimenttootherspecifications. *** p<0.01, ** p<0.05, * p<0.1. Horizon h = 1,...,8 is for FOMC meetings ahead. The net sentiment forecast is that of equation(11),Federa(cid:225)lFundsRate t+h =βˆ 0 t+h+βˆ 1 t+hNetSentimentFOMC,t +αˆt+hFederalFundsRatet−1 andthelaggedfederal funds rate forecast is that of equation (12) Federa(cid:225)lFundsRate t+h =αˆ 0 t+hαˆ 1 t+hFederalFundsRatet−1. These forecasts areinitiallyestimatedfort ={1994,...,1997}andthenre-estimatedforeachmeetingthroughApril2025forh=1,...,8. The Taylor(1999)ruleisequation(13),Federa(cid:225)lFundsRate t+h =2.5+1.5(π t+h −2)+2(u t ∗ +h −u t+h )whereinflationisthet+h realized core PCE index and unemployment rate gap series is the realized t+h unemployment rate less its natural rate, (cid:179) andtheinertialTaylor(1999)ruleisthatofequation(14),Federa(cid:225)lFundsRate t+h =0.85FederalFundsRate t−1+h +0.15 2.5+ 1.5(π t+h −2)+2(u t ∗ +h −u t+h ) (cid:180) which uses the same inflation and unemployment rate series as its non-inertial counterpart. The0.85weightonthelaggedfederalfundsrateisfromtheFederalReserveBankofAtlanta’sTaylorRuleUtilityTool, (https://www.atlantafed.org/research-and-data/data/taylor-rule).DataseriesusedaredetailedinappendixG. 55

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION D APPENDIX: MARKET REACTIONS AND SENTIMENT SCORES D.1 MARKET REACTION VS. SENTIMENT SCORES FigureD.1showsthatpressconferencesstandsout fortheirrelationshipbetweenhawkishsentimentandmarketreactionsforestimatesofequation(16). FigureD.1:Marketreactionsvs.Sentiment Thefiguresshowsequation(16)∆MarketReactiont,j =α+βSentimentj +ϵ t,j forhawkishsentimentforeachtypeofcommunication j∈{Pressconference,Chairspeech,FOMCstatement}. Shadedbandsare95percentconfidenceintervals. Thefigure plotsallmarketreactionswithpairedsentimentscores.Marketreactionsareintradaychangesoftradeprices20minutesaftera communicationeventrelativetotradeprices10minutesbeforefortheithEurodollar/(i+1)thSOFRfuture(ESRi),whichcorrespondstotheexpectedpolicyrateiquartersaheadfori=1,...,4.Theestimationsamplestartsin1999forFOMCstatements,in 2008forChairspeeches,andin2011forpressconferences,whichisbasedontheavailabilityoftextsofthesecommunications. 56

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION FigureD.2:MarketReactionvs.Sentiment,filledobservationscorrespondtomarketreactionsthatdiffer byatleast3basispointsforeachFOMCstatment/pressconferencepair. Thefiguresshowsequation(16)∆MarketReactiont,j =α+βSentimentj +ϵ t,j forhawkishanddovishsentimentforeachtype ofcommunication j ∈{Pressconference,FOMCstatement}. Shadedbandsare95percentconfidenceintervals. Thefigure plotsallmarketreactionswithpairedsentimentscores. Marketreactionsareintradaychangesoftradeprices20minutes afteracommunicationeventrelativetotradeprices10minutesbeforefortheithEurodollar/(i+1)thSOFRfuture(ESRi), whichcorrespondstotheexpectedpolicyratei quartersaheadfori=1,...,4. Theestimationsamplestartsin1999forFOMC statementsandin2011forpressconferences,whichisbasedontheavailabilityoftextsofthesecommunications. 57

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION FigureD.3:MarketReactionvs.NeutralSentiment Thefiguresshowsequation(16)∆MarketReactiont,j =α+βSentimentj +ϵ t,j forneutralsentimentforeachtypeofcommunication j∈{Pressconference,Chairspeech,FOMCstatement}. Shadedbandsare95percentconfidenceintervals. Thefigure plotsallmarketreactionswithpairedsentimentscores.Marketreactionsareintradaychangesoftradeprices20minutesaftera communicationeventrelativetotradeprices10minutesbeforefortheithEurodollar/(i+1)thSOFRfuture(ESRi),whichcorrespondstotheexpectedpolicyrateiquartersaheadfori=1,...,4.Theestimationsamplestartsin1999forFOMCstatements,in 2008forChairspeeches,andin2011forpressconferences,whichisbasedontheavailabilityoftextsofthesecommunications. ThefigureshowsthesameregressionasfigureD.1butfortheneutralsentimentscoreonly. 58

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION FigureD.4:MarketReactionvs.HawkishSentiment Thefiguresshowsequation(16)∆MarketReactiont,j =α+βSentimentj +ϵ t,j forhawkishsentimentforeachtypeofcommunicationj∈{PressconferenceFOMCstatement}.Shadedbandsare95percentconfidenceintervals.Thefigureplotsallmarket reactionswithpairedsentimentscores. Marketreactionsareintradaychangesoftradeprices20minutesafteracommunicationeventrelativetotradeprices10minutesbeforeforthe4thEurodollar/5thSOFRfuture(ESR4),whichcorrespondstothe expectedpolicyratefourquartersahead.Theestimationsamplestartsin1999forFOMCstatementsandin2011forpressconferences,whichisbasedontheavailabilityoftextsofthesecommunications.Filledobservationindicatearespectiveopposite signshockofatleast3basispoints. 59

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION E APPENDIX: ANNOUNCEMENT VARIANCE RATIO This appendix presents the average minute-level share of daily trades around FOMC statements with andwithoutapressconference. FigureE.1:IntradayTradeSharesAroundFOMCStatementsforthe4thEurdollar/5thSOFRFuture Thefigureplots,foreachminuterelativetotheFOMCstatement,theshareofthatday’stotaltradesthatoccurinthatminute. GraylinesshowindividualFOMCmeetingdays;coloredlinesshowtheaverageacrossmeeting-daygroups. 60

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION F APPENDIX: VAR F.1 MONETARYPOLICYSHOCKCONSTRUCTION ThedetailsaredrawnfromBrennanetal.(2025). NAKAMURAANDSTEINSSON(2018)MONETARYPOLICYSHOCK usesfuturestickdataaccessedviaCME GroupInc. DataMine(https://datamine.cmegroup.com/)attheFederalReserveBoard. Theconstruction of the shock series follows that of Gürkaynak et al. (2005) as described in Nakamura and Steinsson(2018)andBrennanetal.(2025). TheshocksarethefirstprincipalcomponentofchangesinhighfrequencyfederalfundsratefuturesandEurodollar/SOFRfutures:  Ds   (ff1 −ff1 ) if Ds−ds>7 MP1 = Ds−ds s,t+∆t s,t−∆t (22) s  ff2 −ff2 otherwise s,t+∆t s,t−∆t MP2 =     Ds D ′− s′ ds′ (cid:34) (ff s j ′,t+∆t −ff s j ′,t−∆t )− D ds s ′ ′ MP1 s (cid:35) if Ds′−ds′>7 (23) s   ff j+1 −ff j+1 otherwise s′,t+∆t s′,t−∆t  ED2 −ED2 if q<2022:Q1 ∆EDSR2 = q,t+∆t q,t−∆t (24) q SF3 −SF3 otherwise q,t+∆t q,t−∆t  ED3 −ED3 if q<2022:Q1 ∆EDSR3 = q,t+∆t q,t−∆t (25) q SF4 −SF4 otherwise q,t+∆t q,t−∆t  ED4 −ED4 if q<2022:Q1 ∆EDSR4 = q,t+∆t q,t−∆t (26) q SF5 −SF5 otherwise q,t+∆t q,t−∆t ′ LetsindexthemonthofthecurrentFOMCmeetingands indexthemonthofthenextmeeting.Forexample, s =March2014 and s ′ =April2014fortheMarch19, 2014FOMCmeetingwhere s and s ′ need not be consecutive months. We define t as the time of the FOMC statement release, typically 14:00 EST and t+∆t more precisely as 20 or 80 minutes after the FOMC statement release while t−∆t is defined as 10 minutes before the FOMC statement release. For the March 19, 2014 FOMC statement, which occurred at t = March19,14:00, t+∆ = March19,201414:20 or t = March19,201415:20 and t−∆t=March19,201413:50. Let ff j denote the duration j of the federal funds futures contract ff. For example, j =1 denotes thecontractexpiringinthecurrentmonth, j =2thecontractexpiringnextmonth,etc.Formonths,Ds andds arethenumberoftotaldaysinthemonthandthedayoftheFOMCmeeting, respectively. Ifa monetarypolicyannouncementoccursinthefirst23daysofthemonth,thenthatmonth’sfederalfunds future j =1 is used to calculate MP1 . Because the settlement prices are based on the average of the s effectiveovernightfederalfundsrateinmonthsratherthanthefederalfundsrateonaspecificday,one must correct for time averaging and scale by the inverse of the share of days remaining in the month, Ds . Otherwise,iftheFOMCannouncementoccursinthelastsevendaysofthemonth,nextmonth’s Ds−ds 61

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION future j =2isusedtocalculateMP1 . s MP2 capturestheunexpectedchangeinthefederalfundsfuturescontractsthatexpireattheend s ′ ofmonths whichisthemonthofthenextscheduledFOMCmeeting.Brennanetal.(2025)showthatin practicethenextorfollowingmonth’sfederalfundsfuture j =2,3isusedtocalculateMP2 . s Becausefederalfundsfuturesarehighlyliquidforcontractsexpiringinthenextthreemonthsbutless liquidforcontractsthereafter,researchersuseEurodollarandSOFRfuturestocovertheremainingfirst yearofthetermstructure. EurodollarfutureswerelistedquarterlyandmatureinMarch,June,September, andDecember. Theywereanagreementtoexchange, onthesecondLondonbusinessdaybefore thethirdWednesdayofthelastmonthofthequarter,thepriceofthecontractminusthethree-monthUS dollarBBALIBORinterestrate. BecausetheBBALIBORinterestrateisdiscontinued,Eurodollarfutures ceasedtradinginApril2023.SOFRfutureshavesuccessfullyreplacedEurodollarfuturesontheChicago MercantileExchange. TheSOFRrateisbasedonthecosttoborrowUSDovernightusingTreasurysecuritiesascollateral. BecauseSOFRfuturesaredesignedtoreplaceEurodollarfutures,theycanbespliced intoshockconstructionwhentheyareavailable.WefollowAcostaetal.(2024)anduseJanuary2022asa startdateforSOFRfutures. Letq indexthequarterofthecurrentFOMCmeetingandq+1indextheofthenextFOMCmeeting. Forexample,q=2014:Q1,q+1=2014:Q2,andq+2=2014:Q3fortheMarch19,2014FOMCmeeting. Themonetarypolicyshockisthenthefirstprincipalcomponentofexpressions(22)-(26)scaledso thatitseffectonone-yearnominalTreasuryyieldsisequaltoone. GÜRKAYNAKETAL.(2005)MONETARYPOLICYSHOCKS ThetargetandpathshocksofGürkaynaketal. (2005)areconstructedusingprincipalcomponentanalysisoverthesameinstrumentsetofNakamura andSteinsson(2018)—expressions(22)-(26). Ratherthanextractingjustthefirstprincipalcomponent, Gürkaynaketal.(2005)extractthefirsttwoprincipalcomponentsandthenrotatetheseprincipalcomponentssothatthesecondhasnoeffectonthefederalfundsrate.Thefirstrotatedprincipalcomponent iscalledthetargetshockandisnormalizedsothatitisone-for-onewiththefederalfundsrate.Thesecondisthecalledpathshockwhichisnormalizedtobeone-for-onewithEDSR4 inexpression(26)and q capturesallforwardguidancesurprises. 62

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION F.2 VAR ESTIMATES Thefigurebelowshowalternativelagsandsamplestothoseshowninthemain text. dnoBssecxE raeY-2 PI IPC muimerP yrusaerT F-stat=2.94 F-stat=2.90 F-stat=1.53 OrignalNS(2018) NS(2018)w/Press NS(2018)w/Press TimeWindow Conference Conference+longrates 1 1 1 0 0 0 -1 -1 -1 -2 -2 -2 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.5 0.5 0.5 0 0 0 -0.5 -0.5 -0.5 -1 -1 -1 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.1 0.1 0.1 0 0 0 -0.1 -0.1 -0.1 -0.2 -0.2 -0.2 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.5 0.5 0.5 0 0 0 -0.5 -0.5 -0.5 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 (a)Originaltimewindow (b)Includespressconference (c) Includes press conference and long-termrates FigureF.1:Impulseresponsetoa25basispointNakamuraandSteinsson(2018)monetarypolicyshock, x-axisismonthsandy-axisispercentagepoints.12lagsinsteadofeight. Impulseresponsesareestimatesfromequation(19)YT =α+B(L)YT−1 +s1Y T 2Y +u˜T obtainedviatheCanovaandFerroni (2022)BayesianVARtoolboxwith68and90percenterrorbands(darkerandlighterbands,respectively),20,000draws,and 12lags. ThesampleofmonetaryshockseriesisfromJanuary1995toApril2025whilethesampleofeconomicdatastartsin January1973. Theshockseriesarethefirstprincipalcomponentoftheinstrumentset{MP1,MP2,EDSR2,EDSR3,EDSR4}. Theshockseriesinpanelaisconstructedfromthe30-minutechangeinthesefuturesaroundFOMCstatements. Theshock seriesinpanelbisconstructedfromthe90-minutechangeinthesefuturesaroundFOMCstatementswhenthereisapostmeetingpressconferencesothatthetimewindowincludesthepressconference.Panelcaugmentstheinstrumentsetinpanel bwith2-,5-,10-,and30-yearTreasuryyields. IPistheindustrialproductionindex,CPIistheconsumerpriceindex,excess bondpremiumisfromGilchristandZakrajšek(2012),andthetwo-yearTreasuryistheendofthemonthdailychangeinthe zero-couponyield.AllsourcesofseriesaredetailedinAppendixG.AppendixTable2displaysthefirst-stageF-statistics. 63

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION dnoBssecxE raeY-2 PI IPC muimerP yrusaerT F-stat=2.51 F-stat=2.60 F-stat=2.33 Target,OriginalTimeWindow Targetw/Press Target+longrates Conference 1 1 1 0 0 0 -1 -1 -1 -2 -2 -2 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.5 0.5 0.5 0 0 0 -0.5 -0.5 -0.5 -1 -1 -1 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.1 0.1 0.1 0 0 0 -0.1 -0.1 -0.1 -0.2 -0.2 -0.2 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.5 0.5 0.5 0 0 0 -0.5 -0.5 -0.5 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 (a)Originaltimewindow (b)Includespressconference (c) Includes press conference and long-termrates Figure F.2: Impulse response to a 25 basis point Gürkaynak et al. (2005) targetmonetary policy shock, x-axisismonthsandy-axisispercentagepoints.12lagsinsteadofeight. Impulseresponsesareestimatesfromequation(19)YT =α+B(L)YT−1 +s1Y T 2Y+u˜T obtainedviatheCanovaandFerroni(2022) BayesianVARtoolboxwith68and90percenterrorbands(darkerandlighterbands,respectively),20,000draws,and12lags. ThesampleofmonetaryshockseriesisfromJanuary1995toApril2025whilethesampleofeconomicdatastartsinJanuary 1973. Theshockseriesaretherotatedfirstprincipalcomponentoftheinstrumentset{MP1,MP2,EDSR2,EDSR3,EDSR4}. Theshockseriesinpanelaisconstructedfromthe30-minutechangeinthesefuturesaroundFOMCstatements. Theshock seriesinpanelbisconstructedfromthe90-minutechangeinthesefuturesaroundFOMCstatementswhenthereisapostmeetingpressconferencesothatthetimewindowincludesthepressconference.Panelcaugmentstheinstrumentsetinpanel bwith2-,5-,10-,and30-yearTreasuryyields. IPistheindustrialproductionindex,CPIistheconsumerpriceindex,excess bondpremiumisfromGilchristandZakrajšek(2012),andthetwo-yearTreasuryistheendofthemonthdailychangeinthe zero-couponyield.AllsourcesofseriesaredetailedinAppendixG.AppendixTable2displaysthefirst-stageF-statistics. 64

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION dnoBssecxE raeY-2 PI IPC muimerP yrusaerT F-stat=0.41 F-stat=0.27 F-stat=0.15 Path,OriginalTimeWindow Pathw/Press Path+longrates Conference 5 5 5 0 0 0 -5 -5 -5 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0 0 0 -2 -2 -2 -4 -4 -4 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 1 1 1 0 0 0 -1 -1 -1 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 1 1 1 0 0 0 -1 -1 -1 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 (a)Originaltimewindow (b)Includespressconference (c) Includes press conference and long-termrates FigureF.3: Impulseresponsetoa25basispointGürkaynaketal.(2005)pathmonetarypolicyshock,xaxisismonthsandy-axisispercentagepoints.12lagsinsteadofeight. Impulseresponsesareestimatesfromequation(19)YT =α+B(L)YT−1 +s1Y T 2Y+u˜T obtainedviatheCanovaandFerroni(2022) BayesianVARtoolboxwith68and90percenterrorbands(darkerandlighterbands,respectively),20,000draws,and12lags. ThesampleofmonetaryshockseriesisfromJanuary1995toApril2025whilethesampleofeconomicdatastartsinJanuary 1973.Theshockseriesaretherotatedsecondprincipalcomponentoftheinstrumentset{MP1,MP2,EDSR2,EDSR3,EDSR4}. Theshockseriesinpanelaisconstructedfromthe30-minutechangeinthesefuturesaroundFOMCstatements. Theshock seriesinpanelbisconstructedfromthe90-minutechangeinthesefuturesaroundFOMCstatementswhenthereisapostmeetingpressconferencesothatthetimewindowincludesthepressconference.Panelcaugmentstheinstrumentsetinpanel bwith2-,5-,10-,and30-yearTreasuryyields. IPistheindustrialproductionindex,CPIistheconsumerpriceindex,excess bondpremiumisfromGilchristandZakrajšek(2012),andthetwo-yearTreasuryistheendofthemonthdailychangeinthe zero-couponyield.AllsourcesofseriesaredetailedinAppendixG.AppendixTable2displaysthefirst-stageF-statistics. 65

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION dnoBssecxE raeY-2 PI IPC muimerP yrusaerT OrignalNS(2018) NS(2018)w/Press NS(2018)w/Press TimeWindow Conference Conference+longrates 1 1 1 0 0 0 -1 -1 -1 -2 -2 -2 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.5 0.5 0.5 0 0 0 -0.5 -0.5 -0.5 -1 -1 -1 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.4 0.4 0.4 0.2 0.2 0.2 0 0 0 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.5 0.5 0.5 0 0 0 -0.5 -0.5 -0.5 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 (a)Originaltimewindow (b)Includespressconference (c) Includes press conference and long-termrates FigureF.4:Impulseresponsetoa25basispointNakamuraandSteinsson(2018)monetarypolicyshock, x-axisismonthsandy-axisispercentagepointsstoppingin2019. Impulseresponsesareestimatesfromequation(19)YT =α+B(L)YT−1 +s1Y T 2Y +u˜T obtainedviatheCanovaandFerroni (2022)BayesianVARtoolboxwith68and90percenterrorbands(darkerandlighterbands,respectively),20,000draws,and 8lags. ThesampleofmonetaryshockseriesisfromJanuary1995toApril2025whilethesampleofeconomicdatastartsin January1973. Theshockseriesarethefirstprincipalcomponentoftheinstrumentset{MP1,MP2,EDSR2,EDSR3,EDSR4}. Theshockseriesinpanelaisconstructedfromthe30-minutechangeinthesefuturesaroundFOMCstatements. Theshock seriesinpanelbisconstructedfromthe90-minutechangeinthesefuturesaroundFOMCstatementswhenthereisapostmeetingpressconferencesothatthetimewindowincludesthepressconference.Panelcaugmentstheinstrumentsetinpanel bwith2-,5-,10-,and30-yearTreasuryyields. IPistheindustrialproductionindex,CPIistheconsumerpriceindex,excess bondpremiumisfromGilchristandZakrajšek(2012),andthetwo-yearTreasuryistheendofthemonthdailychangeinthe zero-couponyield.AllsourcesofseriesaredetailedinAppendixG.AppendixTable2displaysthefirst-stageF-statistics. 66

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION dnoBssecxE raeY-2 PI IPC muimerP yrusaerT Target,OriginalTimeWindow Targetw/PressConference Target+longrates 1 1 1 0 0 0 -1 -1 -1 -2 -2 -2 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.5 0.5 0.5 0 0 0 -0.5 -0.5 -0.5 -1 -1 -1 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.4 0.4 0.4 0.2 0.2 0.2 0 0 0 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.5 0.5 0.5 0 0 0 -0.5 -0.5 -0.5 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 (a)Originaltimewindow (b)Includespressconference (c) Includes press conference and long-termrates Figure F.5: Impulse response to a 25 basis point Gürkaynak et al. (2005) targetmonetary policy shock, x-axisismonthsandy-axisispercentagepointsstoppingin2019. Impulseresponsesareestimatesfromequation(19)YT =α+B(L)YT−1 +s1Y T 2Y+u˜T obtainedviatheCanovaandFerroni(2022) BayesianVARtoolboxwith68and90percenterrorbands(darkerandlighterbands,respectively),20,000draws,and8lags.The sampleofmonetaryshockseriesisfromJanuary1995toDecember2019whilethesampleofeconomicdatastartsinJanuary 1973. Theshockseriesaretherotatedfirstprincipalcomponentoftheinstrumentset{MP1,MP2,EDSR2,EDSR3,EDSR4}. Theshockseriesinpanelaisconstructedfromthe30-minutechangeinthesefuturesaroundFOMCstatements. Theshock seriesinpanelbisconstructedfromthe90-minutechangeinthesefuturesaroundFOMCstatementswhenthereisapostmeetingpressconferencesothatthetimewindowincludesthepressconference.Panelcaugmentstheinstrumentsetinpanel bwith2-,5-,10-,and30-yearTreasuryyields. IPistheindustrialproductionindex,CPIistheconsumerpriceindex,excess bondpremiumisfromGilchristandZakrajšek(2012),andthetwo-yearTreasuryistheendofthemonthdailychangeinthe zero-couponyield.AllsourcesofseriesaredetailedinAppendixG.AppendixTable2displaysthefirst-stageF-statistics. 67

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION dnoBssecxE raeY-2 PI IPC muimerP yrusaerT Path,OriginalTimeWindow Pathw/PressConference Path+longrates 0 0 0 -5 -5 -5 -10 -10 -10 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 1 1 1 0 0 0 -1 -1 -1 -2 -2 -2 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.5 0.5 0.5 0 0 0 -0.5 -0.5 -0.5 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 0.5 0.5 0.5 0 0 0 -0.5 -0.5 -0.5 5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30 (a)Originaltimewindow (b)Includespressconference (c) Includes press conference and long-termrates FigureF.6: Impulseresponsetoa25basispointGürkaynaketal.(2005)pathmonetarypolicyshock,xaxisismonthsandy-axisispercentagepointsstoppingin2019. Impulseresponsesareestimatesfromequation(19)YT =α+B(L)YT−1 +s1Y T 2Y+u˜T obtainedviatheCanovaandFerroni(2022) BayesianVARtoolboxwith68and90percenterrorbands(darkerandlighterbands,respectively),20,000draws,and8lags.The sampleofmonetaryshockseriesisfromJanuary1995toDecember2019whilethesampleofeconomicdatastartsinJanuary 1973.Theshockseriesaretherotatedsecondprincipalcomponentoftheinstrumentset{MP1,MP2,EDSR2,EDSR3,EDSR4}. Theshockseriesinpanelaisconstructedfromthe30-minutechangeinthesefuturesaroundFOMCstatements. Theshock seriesinpanelbisconstructedfromthe90-minutechangeinthesefuturesaroundFOMCstatementswhenthereisapostmeetingpressconferencesothatthetimewindowincludesthepressconference.Panelcaugmentstheinstrumentsetinpanel bwith2-,5-,10-,and30-yearTreasuryyields. IPistheindustrialproductionindex,CPIistheconsumerpriceindex,excess bondpremiumisfromGilchristandZakrajšek(2012),andthetwo-yearTreasuryistheendofthemonthdailychangeinthe zero-couponyield.AllsourcesofseriesaredetailedinAppendixG.AppendixTable2displaysthefirst-stageF-statistics. 68

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION Seriesname F-stat RobustF-stat 8lags,April2025enddate Originaltimewindow,NS 4.08 1.73 Pressconferenceincluded,NS 4.16 1.83 Pressconference+long-rates,NS 1.77 1.49 Originaltimewindow,target 3.71 1.30 Pressconferenceincluded,target 3.96 1.41 Pressconference+long-rates,target 3.42 1.50 Originaltimewindow,path 0.34 0.14 Pressconferenceincluded,path 0.21 0.09 Pressconference+long-rates,path 0.69 0.24 12lags,April2025enddate Originaltimewindow,NS 2.94 1.35 Pressconferenceincluded,NS 2.90 1.41 Pressconference+long-rates,NS 1.53 1.29 Originaltimewindow,target 2.51 0.97 Pressconferenceincluded,target 2.60 1.02 Pressconference+long-rates,target 2.33 1.11 Originaltimewindow,path 0.41 0.17 Pressconferenceincluded,path 0.27 0.13 Pressconference+long-rates,path 0.15 0.05 8lags,2019enddate Originaltimewindow,NS 4.14 1.60 Pressconferenceincluded,NS 4.22 1.64 Pressconference+long-rates,NS 1.80 1.63 Originaltimewindow,target 4.33 1.32 Pressconferenceincluded,target 4.42 1.35 Pressconference+long-rates,target 3.58 1.38 Originaltimewindow,path 0.00 0.00 Pressconferenceincluded,path 0.02 0.01 Pressconference+long-rates,path 0.75 0.23 Table2:First-stageF-statistics. Estimatesfromequation(19)YT =α+B(L)YT−1 +s1Y T 2Y +u˜T obtainedviatheCanovaandFerroni(2022)BayesianVARtoolboxwith68percenterrorbands,20,000draws.ThesampleofmonetaryshockseriesisfromJanuary1995toApril2025,unless otherwisenoted,whilethesampleofeconomicdatastartsinJanuary1973.Theshockseriesofeachpanelareprincipalcomponentsoftheinstrumentset{MP1,MP2,EDSR2,EDSR3,EDSR4}withtheseriesconstructedfromlongratesaugmented withintradaychangesin2-,5-,10-,and30-yearTreasuries. The“NS"shockisthefirstprincipalcomponentandthe“target" and“path"shocksaretherotatedfirstandsecondprincipalcomponent,respectively. The“originaltimewindow"shockseriesisconstructedfromthe30-minutechangeinthesefuturesaroundFOMCstatements. The“pressconferenceincluded" shockseriesisconstructedfromthe90-minutechangeinthesefuturesaroundFOMCstatementswhenthereisapost-meeting pressconferencesothatthetimewindowincludesthepressconference.The“Pressconference+longrates"includesintraday changesin2-,5-,10-,and30-yearTreasuryyieldsintheinstrumentset.AllsourcesofseriesaredetailedinAppendixG. 69

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION G APPENDIX: DATA CONSTANTMATURITYTREASURYYIELDS aredailymarketyieldsonU.S.TreasuriesobtainviatheH.15 SelectedInterestRateReleasefromtheFederalReserveBoard. CONSUMERPRICEINDEX istheseasonallyadjustedmonthlyConsumerPriceIndexfromtheBureauof LaborStatistics(FRED: CPIAUCSL). EURODOLLAR FUTURES are available at an intraday tick frequency from 1995 to March 2023 via the CMEGroupInc.DataMine(https://datamine.cmegroup.com/)attheFederalReserveBoard. EXCESS BOND PREMIUM isamonthlycreditspreadindexfromGilchristandZakrajšek(2012)andis availablefromtheFederalReserveBoard (https://www.federalreserve.gov/econres/notes/feds-notes/ebp_csv.csv.) FEDERALFUNDSFUTURES areavailableatanintradaytickfrequencyfrom1995topresentviatheCME GroupInc.DataMineattheFederalReserveBoard(https://datamine.cmegroup.com/). FEDERAL FUNDS RATE isthetargetfederalfundsrateaftereachFOMCmeeting. Becauseachangein thefederalfundsrateisexecutedthedayafteranFOMCmeeting,thisseriesisconstructedbytakingthe valueforthedayafteranFOMCmeeting. TheFREDseriesDFEDTARisusedfrom1994toDecember15, 2008andthentheaverageofDFEDTARUandDFEDTARLisusedthereafter. INDUSTRIAL PRODUCTION is the seasonally adjusted monthly Industrial Production Index from the FederalReserveBoard(FRED: INDPRO). CORE PERSONAL CONSUMPTION EXPENDITURES INDEX is the seasonally adjusted monthly personal consumption expenditures index excluding food and energy, seasonally adjusted from the Bureau of EconomicAnalysis(FRED: PCEPILFE). SOFR FUTURES areusedatanintradaytickfrequencyfromJanuary2022toSeptember2024viathe CMEGroupInc.DataMine(https://datamine.cmegroup.com/)attheFederalReserveBoard. UNEMPLOYMENTRATE istheseasonallyadjustedmonthlyunemploymentratefromtheBureauofLaborStatistics(FRED: UNRATE). NON-CYCLICAL RATE OF UNEMPLOYMENT isthemonthlynotseasonallyadjustedseriesfromtheU.S. CongressionalBudgetOffice(FRED: NROU). VIX isthedailynotseasonallyCBOEVolatilityIndexofadjustedmarketexpectationofvolatilityconveyedbystockindexoptionsfromtheChicagoBoardofOptionsExchange(FRED: VIXCLS). 70

BYUN,FEES,JACOBSON,&WALKER: CENTRALBANKCOMMUNICATION ZERO-COUPONTREASURYYIELDS arecontinuouslycompoundedzero-coupondailyyields(mnemonic: SVENYXX)obtainedfromtheFederalReserveBoard (https://www.federalreserve.gov/data/yield-curve-tables/feds200628_1.htmlorasacsvfile). 71