What Determines Household Expectations?

Board of Governors of the Federal Reserve System International Finance Discussion Papers ISSN 1073-2500 (Print) ISSN 2767-4509 (Online) Number 1412 July 2025 What Determines Household Expectations? Anushka Mitra and Aditi Singh Please cite this paper as: Mitra, Anushka and Aditi Singh (2025). “What Determines Household Expectations?,” International Finance Discussion Papers 1412. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/IFDP.2025.1412. NOTE: International Finance Discussion Papers (IFDPs) 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 International Finance Discussion Papers Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers. Recent IFDPs are available on the Web at www.federalreserve.gov/pubs/ifdp/. This paper can be downloaded without charge from the Social Science Research Network electronic library at www.ssrn.com.

What Determines Household Expectations? Anushka Mitra and Aditi Singh* Abstract Thispaperexamineswhichmacroeconomicsignalsshapehouseholdexpectations andfindsthatunemploymentshocksplayamoreinfluentialrolethaninflationshocks. Usingdailydata,weidentifywhichannouncementsprompthouseholdstorevisetheir expectations. We construct two shock series—assuming households are either sophisticatedornaive—basedonthesurprisecomponentsofannouncements. Labor marketnewsstronglyinfluencesbothgeneraleconomicsentimentandinflationexpectations.Evenwheninflationrisesandunemploymentfalls,householdsrespondmore tounemploymentshocks.Mostchangesininflationexpectationsaredrivenbylabor marketshocks.Duringnegativesupplyanddemandshocks,unemploymentremains thedominantdriver. JELCodes:E70,D84,E30 Keywords: Householdexpectations,macroeconomicdatareleases,high-frequency identification,surveydata 1Mitra:FederalReserveBoardofGovernors 2Singh:ShivNadarInstitutionofEminence *WesincerelythankAnmolBhandari,SarojBhattarai,ChristophBoehm,JesseBruhn,OlivierCoibion, ArpitaChatterjee,FernandoDuarte,GautiEggertson,StefanoEusepi,AmyHandlan,YannKoby,Pascal Michaillat,AndreasMueller,JulioOrtiz,MarcelPeruffo,AysegulSahin,NeilThakral,andDavidWeilfor theirguidanceandsupport.WealsothankseminarparticipantsatBrownUniversity,UTAustin,andIIM, Bangalore,andconferenceparticipantsatDSEWinterSchool,IAES,ISI,IITDelhi,IEG,MidwestMacro, NASMES,PAAandWEAI.Theviewsexpressedherearesolelytheresponsibilityoftheauthorandshould notbeinterpretedasreflectingtheviewoftheBoardofGovernorsoftheFederalReserveSystemorofany otherpersonassociatedwiththeFederalReserveSystem. 1

1 Introduction "Whenmyinformationchanges,Ialtermyconclusions". -JohnMaynardKeynes Householdexpectationsarecentraltothetransmissionofmacroeconomicpolicy. They influenceconsumption,saving,wagebargaining,andresponsestomonetaryandfiscal interventions (D’Acunto & Weber 2024, Coibion et al. 2023, Mueller et al. 2021). In his famous1976paper(LucasJr1976),RobertLucaswroteabouttheimportanceofexpectations ininfluencingpolicy. Hewarnedthatifexpectationaleffectswerenottakenintoaccount properly,theeffectsofpolicywouldbedistorted. Despitetheimportanceofexpectations, weknowsurprisinglylittleaboutwhatinformationhouseholdsactuallyusewhenforming theseexpectations. Householdsareexposedtoavarietyofeconomicsignals,wheresome aremoresalientthanothers.1 Thispaperinvestigatesafundamentalquestion:information fromwhichmacroeconomicvariablescauseshouseholdstoupdatetheirexpectations? Wefindthathouseholdsprimarilyusetheunemploymentrateasasufficientstatisticfor expectationformation. Toarriveatthisresult,weproposeamodelofbeliefformation thatisolatestheunanticipatedcomponentor‘shock’inmacroeconomicannouncements. Weconstructtheseshocksbycomparingrealizedannouncementstoexpectationsagainst twobenchmarks: a)sophisticatedhouseholds,whoincorporateallavailableinformation andformforecastscomparabletoprofessionalforecasters,withforecasterrorscapturing thenewinformation;andb)naivehouseholds,whorelysolelyonpastdatawithoutincorporatingnewinformation,measuringtheshockasthedifferencebetweenrealizationsof consecutiveannouncements. Thesebenchmarksprovideupperandlowerboundsforthe truehouseholdresponse. Wethenusehigh-frequencylocalprojectionsusingdatafrom twolarge-scalesurveys: theGallupDailyTrackingPoll(2008–2017)(GallupInc.2017)and theMichiganSurveyofConsumerAttitudesandBehavior(1980–2019),toassesstheimpact oftheseshocksonhouseholdexpectations. Akeyidentificationchallengearisesfromdiscrepanciesbetweenthefrequencyofeconomicannouncementsandhouseholdsurveys. Whilehouseholdsreceivemultiplesignals each month, household expectations are typically measured only once a month.2 To 1Moreover,householdexpectationsareformedjointlyoverthesesignals(Andreetal.2022,Kamdar &Ray2024),whichmakesitimperativetounderstandwhichsignalsaremostinformativetohouseholds. Thisisbecausehouseholdexpectationsoftenreflectbroadereconomichealth,particularlyinassessing householdbehaviorduringrecessions.Understandingthedriversofhouseholdsentimentisvaluablefor policymakerstogaugegeneraleconomicsentimentandrespondeffectivelytodownturns. 2The two most popular sources of expectations in the US are the University of Michigan’s Survey of 2

addressthis,weutilizeGallup’sdailyexpectationdatatoobservechangesinexpectations aroundannouncementdatesforkeyvariableswithinanarrowwindow,allowingprecise identificationofresponsesaroundannouncementdates. Ouranalysisshowsthathouseholdexpectationsrespondprimarilytolabormarketshocks ratherthanshockstoothermacroeconomicvariablessuchasinflation,GDPgrowth,and housing starts. This pattern is consistent across different sample periods and surveys. To verify this is not merely a function of our sample period, we extend our analysis usingmicrodatafromtheMichiganSurveyofConsumers(MSC)from1980to2019. This longer-termanalysisconfirmsthatwhileexpectationsabouttheeconomydorespondto movementsininflation,theyreactmorestronglytounemploymentshocks,evenduring episodesofrisinginflation. Toformexpectations,householdscanprioritizedifferentvariablesatdifferentpointsin time. Forexample,duringperiodsofrapidlyrisingunemployment,suchastheCOVID-19 pandemic,unemploymentlikelybecomesthedominantdriverofexpectations. Conversely, inperiodsofhighinflation,suchastheearly1980s,inflationlikelytakesprecedence. We validatethisdynamicbydividingtheMSCsampleintofourscenariosbasedonchangesin unemploymentandinflation: (1)bothunemploymentandinflationareincreasing,(2)both aredecreasing,(3)onlyunemploymentisincreasing,and(4)onlyinflationisincreasing. Inallscenarios,unemploymentremainsastatisticallysignificantdriverofexpectations.3 Itisinterestingtonotethateveninthecaseswhereinflationisincreasing,itisshocksto unemploymentthathavealargereffectonhouseholdexpectationsthanshockstoCPI. Sofar,wehavetestedtheresponseofhouseholdexpectationsabouttheeconomy. However, MSC allows us to further test whether the same patterns holds true for inflation expectations. Inflationexpectationsareanimportantpolicytool,anditispossiblethat whilesentimentrespondstounemployment,inflationexpectationsdonot. Wefindthat shockstounemploymentaffectinflationexpectationsatleastasmuchas,andoftenmore than,shockstotheConsumerPriceIndex. Underthesophisticatedexpectationsmodel, households respond exclusively to unemployment shocks in scenarios conventionally associatedwithnegativesupplyanddemandshocks. Underthenaiveexpectationsmodel, householdsrespondmorebroadlybutstillprioritizeunemploymentinformation. Wethusmakeseveralcontributions. First,wedevelopamodeltoisolatetheunanticipated ConsumersandtheNYFed’sSurveyofConsumerExpectations,bothofwhicharemonthly. 3This is true regardless of whether households are assumed to be sophisticated or naive. The only scenariowheretheresponsetounemploymentisinsignificantisinthecaseofnaivehouseholdswhenboth unemploymentandinflationaredecreasing. 3

componentofmacroeconomicannouncements,constructingtwodistinctshockseries thatprovideboundsonthetrueprocessofexpectationadjustments. Second,wedemonstratethatlabormarketinformationsignificantlyinfluencesbothhouseholds’subjective economicexpectationsandinflationexpectations. Thisissurprisinggiventhefocuson inflationexpectationsintheliterature. Eveninperiodsofdecliningunemploymentand risinginflation,wefinditisshockstounemploymentthatsignificantlyaffecthouseholds’ sentiment. Analyzing inflation expectations directly, we find that even these are predominantlydrivenbyunemploymentshocksratherthanCPIshocks. WhileCPIshocks occasionallyaffectinflationexpectations,unemploymentshocksexertamoreconsistent influence. Lastly,CPIshocksaremoreinfluentialthanunemploymentshocksonlyduring positivesupplyordemandshocks,whereasduringnegativeshockstypicallyassociated withrecessions,unemploymentconsistentlydominateshouseholdexpectations. Weinterpretthesefindingsthroughthelensofrationalinattentionandbehavioralsalience. Labormarketindicatorsmaybemorecognitivelyaccessible,personallyrelevant,oremotionallysalientthanaggregateinflationdata,especiallyintimesofeconomicuncertainty. Finally,byextendingtheanalysisacrossmultipleeconomicepisodessuchashighandlow inflation,andsupplyanddemandshocks,weshowthatlabormarketnewsremainsakey anchorofexpectations. Ourresultssuggestthatunemploymentmaybemorethanjustan economicindicator: itisthedominantsignalthroughwhichhouseholdsmakesenseof theeconomy. OurpapercloselyrelatestoBinderetal.(2024)andMertensetal.(2020). Binderetal.(2024) useaneventstudywiththeSurveyofConsumerExpectations,examiningdailyhousehold expectationresponses. Althoughcomplementary,ourapproachdiffersnotablybyemployinglocalprojectionsandshockmeasures,capturingasymmetriesandnonlinearitiesfrom varyingdegreesofinformationalsurprise. Mertensetal.(2020)usedailyGallupdataand localprojectionsspecificallyformonetarypolicyannouncements. Webroadenthisscope byanalyzingmultiplemacroeconomicannouncements,developingshockseriestoisolate unanticipatedinformation,andextendingtheanalysisoveralongerperiodwithMichigan Surveymicrodata,capturingbotheconomicandinflationexpectationsacrossdifferent episodes. Ourstudycontributestoliteratureexaminingsurvey-basedexpectationsandhousehold behavior(Malmendier&Nagel2015,Kuchler&Zafar2015,Mianetal.2021). Mostprevious workfocusesonpointestimatesofexpectationsregardinginflation(Armantieretal.2015, Bachmannetal.2015,Coibionetal.2020),houseprices(Armonaetal.2018),orthelabor 4

market(Potter2020,Muelleretal.2021). Recentpapers(Kamdar&Ray2024,Ehrmannetal. 2017,Andreetal.2022,Roth&Wohlfart2019)suggesthouseholdsformjointexpectations about the aggregate economy. We present novel evidence that household expectations respondpredominantlytolabormarketnews,validatingthisfindingfurtherwithinflation expectations. Lastly,wecontributetotheliteratureonmacroeconomicannouncementpremiumsusing high-frequency events. Prior research demonstrates announcement impacts on spot exchangerates(Andersenetal.2003,Evans&Lyons2008),commodityreturns(Caporale etal.2016),futurescontracts(Balduzzietal.2001,Andersenetal.2007),globalassetprices (Boehm&Kroner2023),andmarketvolatility(Jiangetal.2014). Ourpaperalsorelatesto studiesonmonetarypolicyannouncementsaffectinglong-terminterestrates(Gürkaynak etal.2005)andhouseholdexpectations(Coibionetal.2022). Therestofthepaperisorganizedasfollows: Section2describesallthedatasetsthatwe use. Section3introducesthemodelofexpectationformationthatweproposeandderives an empirically testable result. Section 4 describes our empirical strategy to isolate the responseofhouseholdexpectationstovariousmacroeconomicannouncements. Section 5discussestheseresultsandSection6concludes. 2 Data Weusethreemaindatasourcesforourstudy. OurprimarydatasourceistheGallupDaily Tracking Poll, which provides us with daily data on household expectations. The high frequencynatureoftheGallupsurveyallowsforacleaneridentification. OurseconddatasourceistheMichiganSurveyofConsumers(MSC),amonthlysurveyof householdexpectationsintheUnitedStates. MSCreportsbothafuturesentimentindex aswellaspointestimatesforinflationexpectations. Themicrodatacontainsinterview dates,whichallowsforanalysisatadailyandweeklylevel. OurthirddatasourceisBloomberg’sUnitedStatesEconomicCalendar,whichreportsthe median expectations of professional forecasters prior to each macroeconomic release. Theseforecastshelpustocaptureameasureoftheunanticipatedcomponentofreleases. 5

2.1 Gallup’s US Daily Tracking Poll The US Daily Tracking Poll (Gallup Inc. (2017)), henceforth GDTP, is a repeated crosssectionalsurveyconductedbyGallup,apremierpollingandanalyticsfirm. Itwasfielded toabout1000individualseverydayfrom2008to2013,and500individualseverydayfrom 2013to2017. WeshowtheaveragenumberofrespondentseachdayinamonthinFigure 2banditisconsistentlyintherangeof450-500eachday.4 Thedataisrepresentativeatthe dailylevelanditmatchestargetsfromtheUSCensusBureaubyage,sex,region,gender, education, ethnicity, race, and population density of self-reported location. Appendix Table1displayssummarystatisticsofGallup’spoll.5 Themainvariableweareinterestedinisameasureofhouseholds’expectationsaboutthe futureoftheeconomy. Specifically,participantsareaskedthefollowingquestion: “Rightnow,doyouthinkthateconomicconditionsinthecountry,asawhole,aregettingbetter orgettingworse?" Participantscanchoosebetweenthreeoptions: gettingbetter,stayingthesame,orgetting worse. WedenotethisvariableasourExpectationsIndex. Theproportionofpeoplewho respondbysayingthattheeconomyisgoingtostaythesameislessthan5%fortheentire sample,thuswedropthem. Wearethereforeleftwithabinaryindexthattakesvalue1 whenpeopleareoptimistic(i.e. whentheyreportthattheeconomyisgoingtogetbetter) and 0 when people are pessimistic (i.e when they report that the economy is going to get worse). Higher values of the index indicate more optimism about the future of the economy,whilelowervaluesindicatemorepessimism. Weusethisquestionasameasure ofhouseholdexpectationsabouttheperformanceoftheaggregateeconomyinthefuture. Figure 1 represents the evolution of the Expectations Index over time. Since our index is binary, it can also be interpreted as the share of optimists.6 Figure 1 shows that the ExpectationsIndex,ortheproportionofpeoplewhoareoptimisticaboutthefutureof 4The survey is conducted for 350 days every year. The respondents are evenly divided between the Well-beingtrackandthePoliticsandEconomytrack.Certainvariables,suchasemploymentindicatorsand keydemographics,areaskedonbothtracks. 5Werestrictoursampletoindividualsbetweentheagesof18and90 6Here,wedefineoptimistsasthoseparticipantswhoreportthattheyexpecttheeconomytobegetting better,whilepessimistsasthoseparticipantswhoexpecttheeconomytobegettingworse. 6

Figure1: HouseholdExpectationsIndex Notes:TheGDTPExpectationsIndexisbasedonthefractionofrespondentsratingfutureeconomic conditions(‘Gettingbetter’or‘Gettingworse’).TheMSCshareofoptimistsiscalculatedfromthe fractionofrespondentsratingbusinessconditionsinthecountryasawholeduringthenexttwelve monthsasgoodtimesfinancially(relativetobadtimes).Thecorrelationcoefficientbetweenthese twoseriesis0.86.ThedataisfromGallupInc.andSurveyofConsumers,UniversityofMichigan. SurveyResearchCenter. theeconomy,hasrisenovertime. WealsoplottheshareofoptimistsfromtheMichigan SurveyofConsumers(MSC).7 TheexpectationsindicesfrombothGallupandMichigan displayhighco-movementwithacorrelationcoefficientof0.86,andhaveasimilartrend overtime. Thisisreassuring,sinceitindicatesthatbothindicescapturesimilareconomic expectations. AppendixTable8showsthechangeinourExpectationsIndexaroundmajoreventsthat 7ThiscorrespondstotheQuestionBUS12intheMSCQuestionnaire. 7

Table1: SummaryStatistics:Expectations Variable TotalObs Mean Std. Dev. Frequency (1) (2) (3) (4) MichiganSurveyofConsumers IndexofConsumerExpectations(ICE) 277,160 79.8 45.8 Daily 12-monthaheadInflationExpectations 209,744 5.4 5.5 Daily FractionofOptimists 231,304 51.9 50.0 Daily ChangeinFractionofOptimists 12,227 -0.03 22 Daily GallupDailyTrackingPoll FractionofOptimists 1,705,161 0.4 0.5 Daily ChangeinFractionofOptimists 3387 -0.009 3.8 Daily BloombergEconomicNews(1996-2019) Surprise(Unemployment) 273 -0.03 0.14 Monthly Surprise(CPI) 276 -0.01 0.12 Monthly Surprise(Housing) 257 1.67 78.63 Monthly Surprise(GDP) 84 0.01 0.71 Quarterly ActualEconomicVariables(1980-2019) Change(Unemployment) 480 -0.005 0.17 Monthly Change(CPI) 480 0.38 0.47 Monthly Change(Housing) 257 0.16 183 Monthly Change(GDP) 84 -0.04 2.1 Quarterly Notes:ThistablerecordssummarystatisticsforkeyhouseholdexpectationsforbothGDTPandMSC. WealsoreportSurprisesandActualVariablesfromtheEconomicNewsConsensusForecast,Bloomberg FinanceLP.ActualnewsaboutHousingandGDPisreportedfrom1997-2019.Surveyweightsused.Data accessedfromGallupInc.andSurveyofConsumers,UniversityofMichigan.SurveyResearchCenter. occurredinoursampleperiod. Column4reportsthedifferenceinexpectationsoneday aftertheeventtoonedaybeforetheevent. Thefirstrowreportsthechangeinhousehold expectationsthatoccurredwhentheLehmanBrothersfiledforbankruptcy,triggering the2008recession. Weobservethathouseholdexpectationsdecreasedby–0.22pointson average. Finally,weobservesubstantialheterogeneityinhouseholdexpectationsacross 8

demographicgroups. WediscusstheseinAppendixSectionA.1.1. 2.2 University of Michigan’s Survey of Consumers TheMichiganSurveyofConsumers(MSC)beganinJanuary1978andisthelongestrunning surveyofhouseholdexpectationsintheUnitedStates. Itcollectsbothqualitativeandquantitativeexpectationsbyinterviewingapproximately500individualseachmonth,selected toberepresentativeoftheUSpopulation. Participantsanswerseveralquestionscovering topicsrangingfromqualitativeassessmentsofchangesintheircurrenteconomicsituation andfutureexpectationstoquantitativeestimates,suchaspointforecastsofinflation. We usetheMSCmicrodatanotonlybecauseitisthelongest-runningsurveyonhousehold expectationsbutalsobecause,sinceJune1979,theinterviewdatesofrespondentshave been recorded and made publicly available, as noted by York (2023). This allows for a high-frequencyidentificationatthedailyaswellasweeklylevel. Wereporttheaverage numberofrespondentseachdayofamonthfortheMSCinFigure2a. Figure2: NumberofRespondentsEachDayoftheMonth (a)MichiganSurveyofConsumers (b)GallupDailyTrackingPoll Notes:ThisfigureshowstheaveragenumberofrespondentseverydayofthemonthintheSurveyof Consumers,UniversityofMichigan.SurveyResearchCenter,aswellastheGallupDailyTracking Poll. WeusetwomeasuresofexpectationsfromMSC-aqualitativemeasuredescribinghouseholdsexpectationsaboutfuturebusinesscondition,andaquantitativemeasureofinflation expectations. Letusstartbyexaminingthequalitativemeasurefirst. WhileMSCoffers several qualitative measures of different kinds of expectations, as well as a composite measure,wepicktheonethatasksaboutbusinessconditionsinordertogetascloseto 9

ourGallup’sExpectationIndexaspossible. 8 Specifically,thesurveyasks: Nowturningtobusinessconditionsinthecountryasawhole–doyouthinkthatduringthenext twelvemonthswe’llhavegoodtimesfinancially,orbadtimes,orwhat? Similar to GDTP, participants can pick one of the following three choices - good times, uncertain,orbadtimes. Herealso,theproportionofpeopleanswering’uncertain’isvery low, so we drop them from our sample, and we are once again left with a binary index thatisqualitativelysimilartoGDTP’sExpectationIndex. Changesinthisindexcanalsobe interpretedinasimilarway-anincreasedenotesariseinoptimism,whereasadecrease denotesafallinoptimismorariseinpessimism. Oursampleperiodcoverstheperiod fromJanuary1980toNovember2021. 9 Thesecondmeasureofexpectationsweuseisaquantitativeone: MSC’sinflationexpectations. Specifically,weusethefollowingquestionfromthesurvey: “Byaboutwhatpercentdoyouexpectpricestogoup/downontheaverage,duringthenext12 months?" Thisquestiongivesuspointestimatesofhouseholds’inflationexpectations. Respondents are asked to report a number from 0 to 100. We test our results using this quantitative measure,inadditiontothequalitativeexpectationsdescribedabove. Figure3ashowsthe evolutionofinflationexpectations. Inflationexpectationswereveryhighintheearly1980s, whichisalsothetimewhenactualinflationwasveryhighintheUS.Expectationsdeclined rapidlyasrealizedinflationcamedown,andtheyhavebeenmoreorlessstablesincethen. However,duringcertainepisodes,particularlyaroundthetimeofrecessions,inflation expectationshavegoneup. Figure3breportsthedistributionofinflationexpectations. Whiletherearesomeoutliers,mostofthedistributionliesinthe0to10%range. Appendix Table1providessummarystatisticsforbothourexpectationmeasures. 2.3 Bloomberg’s US Economic Calendar Bloomberg’sUSEconomicCalendarreportsdataforallmajormacroeconomicannouncements,10 aswellastheaverageexantemedianforecastofprofessionalforecasters,called 8Forrobustness,wealsocheckourresultsusingtheIndexofConsumerExpectations,whichisacomposite indexofferedbyMSC.Wegetsimilarresults,whicharereportedintheappendix. 9Althoughthesurveystartedin1978,forthefirsttwoyears,weonlyhavedataontheweekinwhich participantswereinterviewed,notthedate. Moreover,whilethereismonthlydataavailableafter2021, interviewdatesareonlyavailabletillNovember2021,whichiswhyoursampleendsthere. 10Bloombergalsoreportsanyrevisionstotheactualreleases.However,weonlylookattheinitialreported datapoint,sincethatcapturesnewinformationreleasedduringtheannouncement. 10

Figure3: 12-MonthAheadInflationExpectations (a)Time-Series (b)Distribution Notes:Thisfigureshows(a)theevolutionofthe12MonthAheadInflationExpectationsovertime and(b)thedistributionofthe12MonthAheadInflationExpectationsintheSurveyofConsumers, UniversityofMichigan.SurveyResearchCenter. theConsensusForecast. Beforeeverymacroeconomicannouncement,Bloombergsurveys economistsandasksthemwhattheyexpecttoseeintheupcomingannouncement. For this paper, we focus on four variables of policy relevance: unemployment rate, output as measured byGDP growth(Advance), inflation asmeasuredby the month-on-month consumerpriceindex(CPI)andhousingstarts. Newdataforallvariablesisreleasedevery month,exceptforGDP,whichisreleasedquarterly. Table1reportsthebasicsummary statisticsrelatedtothesevariables. 3 A Model of Expectation Formation To determine which macro variables are important to households in their expectation formationprocess,welookatwhennewinformationonthesevariablesisreleasedand investigatewhetherexpectationsadjustinresponsetothatinformation.Todothis,weneed toknowhowmuchoftheinformationreleasedduringmacroeconomicannouncementsis newtohouseholds,i.e.,weneedameasureofshock tohouseholds’informationset. Since thetrueexpectationformationprocessisunknown,findingashockbecomeschallenging. Toaddressthisproblem,wedevelopamodelofexpectationformationandexaminehow macroeconomicannouncementsfeatureinit. 11

3.1 Households’ Expectations Formation Process ConsideramacroeconomicannouncementX thatoccursondatet. LetZ denotesome t fundamental of the economy based on which households form expectations about the future. Oncetheannouncementismade,householdexpectationscanbewrittenas: E [Z] = p·g(X )+(1–p)·h(ψ ) t t t whereψ containsallinformationasidefromtheannouncementthatisavailabletoagents t forformingexpectations. Theparameterpindicatestheweightthathouseholdsgiveto theannouncementintheirbeliefformationprocess. Householdexpectationsbeforethe announcementismadecansimilarlybewrittenas: E t–1 [Z] = E t–1 [E t [Z]|X t–1 ,ψ t–1 ] = E t–1 [p·g(X t )+(1–p)·h(ψ t )|X t–1 ,ψ t–1 ] = p·E t–1 [g(X t )]+(1–p)·E t–1 [h(ψ t )] wherethefirstequalityfollowsfromtheLawofIteratedExpectations.11 Sincewehave dailydataonexpectations,wewillrestrictourattentiontocomparingexpectationsright aftertheannouncementwithexpectationsjustbefore. E t [Z]–E t–1 [Z] = p·[g(X t )–E t–1 [g(X t )]]+(1–p)·[h(ψ t )–E t–1 [h(ψ t )]] (1) A strong identifying assumption would be that the only new information households havebetweendayt andt–1iswhatisprovidedintheannouncement. Thisimpliesthat ψ t = ψ t–1 = ψ. Thus, E t–1 [h(ψ t )] = E t–1 [h(ψ t–1 )] = h(ψ t–1 ) = h(ψ). So the change in expectationssimplifiesto: E t [Z]–E t–1 [Z] = p·[g(X t )–E t–1 [g(X t )]]+(1–p)·[h(ψ)–h(ψ)] = p·[g(X t )–E t–1 [g(X t )]] However,thisisamuchstrongerassumptionthanwhatweneed. Weonlyneedthatno otherinformationisreleasedsystematicallywithanannouncementi.e. cov[g(X ),h(ψ )] = t t 0. Thisimpliesthatcov[g(X t )–E t–1 [g(X t )],h(ψ t )–E t–1 [h(ψ t )]] = 0,whichisthestandard OLSassumption. Weassumethatgisalinearfunctionoftheformg(X ) = aX +b. Equation t t 11Thisistrueaslongastheprobabilitydensityfunctioniswelldefinedandexpectationsareintegrable. 12

1thusbecomes: E t [Z]–E t–1 [Z] = p·a·[X t –E t–1 [X t ]]+(1–p)·[h(ψ t )–E t–1 [h(ψ t )]] a allows us to build in under-reaction or over-reaction into our expectation formation process. Forourpurposes,wewillassumea = 1. Hence,wegetthefollowingregression equation: E t [Z]–E t–1 [Z] = α+β·ShockX t +ϵ t (2) whereShockX t = X t –E t–1 [X t ]andϵ t = (1–p)·[h(ψ t )–E t–1 [h(ψ t )]]. ShockX t isjustthedifferencebetweentheactualvalueofthevariablereleasedintheannouncementandhouseholds’forecastofit. Ouridentificationassumptioncanbewrittenascov(ShockX ,ϵ ) = 0, t t which is the standard OLS assumption. The change in households’ expectations due to themacroeconomicannouncementdependsonthedifferencebetweentheinformation releasedintheannouncementandhouseholds’forecastofit,denotedbyShockX . t 3.2 Sophisticated versus Naive Models of Expectation Formation Wepositthathouseholds’forecastsareacombinationofbackwardlookingvariablesand forwardlookingexpectations: E t–1 [X t ] = (1–ω)X t–1 +ωE t–1 [X t ] (3) whereX t–1 isthevalueofthemacroeconomicvariableinthepreviousperiod,andE t–1 [X t ] isthefullinformationforecastofX attimet–1. Sincewecannotobservehouseholds’ t forecasts directly, we analyze two extreme cases - when the forecast is fully backward looking,andwhenitisfullyforwardlooking. Lookingatthesetwoextremeswillhelpus getarangeofhouseholds’forecasts. First,weconsiderthecaseofhouseholdsfollowinganaiveexpectationformationprocess. SincehouseholdsknowX t–1 whenmakingtheirforecastaboutX t ,weassumethatthey cannotdoanyworseintheirprediction. ThusthenaiveforecastofX issimplythevalue t ofthemacroeconomicvariablefromthepreviousannouncement,X t–1 . Underthenaive modelhouseholdsdonotmakeuseofanynewinformationbetweentwoannouncements ofthesamevariabletoupdatetheirbeliefs. Theirforecastisentirelybackwardlooking, 13

i.e. ω = 0. Theunanticipatedpartofcurrentannouncementinthengivenby: ShockX t = X t –E t–1 [X t ] = X t –X t–1 Thusinthiscase,theshockfromtheannouncementismeasuredsimplybythedifference thevalueofthemacroeconomicvariableinthecurrentannouncementtotheprevious. Ontheotherextreme,considerthecaseofhouseholdsfollowingasophisticatedexpectationformationprocess. Thesehouseholdsusealltheinformationavailabletothemto maketheirforecast. Thusinthiscase,weconsiderthathouseholdforecastsarethesame asthatofprofessionalforecasters,whichwetaketobeourbenchmarkfullinformation forecast. Forthesehouseholds,ω = 1. Theunanticipatedpartofcurrentannouncement inthiscaseisthengivenby: ShockX t = X t –E t–1 [X t ] = X t –E t–1 [X t ] = X t –E t P – F 1 [X t ] = SurpriseX t whereEPF[X ]isprofessionalforecasters’forecastofX beforetheannouncement. This t–1 t t iscommonlyknownintheliteratureasaSurprise(Gürkaynaketal.(2005)). Households’ expectations in this case can only be affected by the information they were not able to predict. Thisisthestandardrationalexpectationsformulation. In reality, households’ forecasts probably lie somewhere between these two extremes. Analyzingthesetwocaseshelpsusprovideaboundonthetrueresponseofhousehold expectationstonewmacroeconomicinformation. 4 Empirical Strategy With our model of expectations formation in hand, analyzing which macroeconomic variablesaffecthouseholdexpectationsisnowstraightforward. FollowingGürkaynaketal. (2005)andMertensetal.(2020),weproposethatifweestimatethechangeinexpectations withinanarrowwindowaroundthereleasedateofamacroeconomicannouncement,then wecanassignacausalclaimtoit. Inotherwords,bychoosingatightwindow,weassume thattheonlyeventoccurringinthattimeframeisthemacroeconomicannouncement,and 14

thereforeanychangeinexpectationsinthiswindowmustbeduetotheannouncement.12 To be precise, let the announcement occur on day t. We then consider the change in household expectations in the window [t –1,t +h], where h denotes days from t. Since peoplemaytakesometimetoupdatetheirexpectations,wevarythehorizonhfromone tofivedays. LetEi[Z]denoteexpectationsofindividualiondayt andShockX denotethe t t shockcomingfromnewinformationintheannouncement. Then,followingJordà(2005), theeffectoftheannouncementonexpectationscanbeestimatedusingthefollowinglocal projection: E t i +h [Z]–E¯ t–1 [Z] = α h +β h ·ShockX t +D i t+h +ϵi th (4) ThisfollowsfromEquation2. Di denotesdemographiccharacteristicsofpersoniattimet t andincludeage,education,income,gender,occupation,jobstatusand,stateofresidence. NotethatsincetheGalluppollisnotapanelsurvey,wecannottrackexpectationsofthe same person over time. Thus, we average expectations for day t –1 and subtract them. Since Gallup is representative at the daily level, E¯ t–1 [Z] denotes the expectations for a representativeagent. Wealsomatchonobservablesandsubtractthoseexpectations,in order to get close to the true expectations of person i at time t –1. Our results remain unchanged,sointherestofthepaper,wesubtracttherepresentativeagent’sexpectations. Althoughwedonotobservethetimeatwhichapersonissurveyed,Galluponlysurveys peopleafter5pmonweekdays. Sincemostannouncementscomeoutearlyinthemorning, we feel that it is safe to include responses obtained on day t as coming after the announcement.13 Ourresults,however,remainrobusttotheexclusionofdayt. Itisalsoimportanttotalkaboutthetimelineofmacroeconomicreleases. TheBLSjobs report is the first major macroeconomic release of every month, and it is released on thefirstFridayofeverymonth. ItisfollowedbyCPI,whichcomesoutinthemiddleof the month. Next is the housing report, which is released between the 15th and 20th of everymonth. Finally,theGDPreportisreleasedbetweenthe27thandthelastdayofevery month. Sinceweusethetimingofannouncementsforidentification,itiscrucialthatourrelease datesnotclashwithotherannouncements. Forthisreason,wedonotlookattheIndexof 12We check for overlaps of major events with macroeconomic releases and omit the days where any overlapoccurs. 13Thesurveyoccursfrom11amonweekends,butnoannouncementsaremadeonweekends. 15

IndustrialProduction(IIP)becauseitisoftenreleasedveryclosetothehousingreport. AsimilarissueispresentwiththeBLSjobsreport, whichcomesoutonthefirstFriday of every month. It is preceded by the jobless claims numbers that are released every Thursday. Furthermore, ADP Research Institute also usually releases its employment report on the first Wednesday of every month. It could thus be argued that the correct priortolookatfortheunemploymentratewouldbeTuesday,sinceWednesdaytoFriday arefilledwithnewinformationregardingthelabormarket. Appendixtable#looksatthis caseandfindstheresultstoberobust. We are using unemployment to proxy for BLS’s jobs report. However, several data is releasedinthejobsreport,suchaslaborforceparticipation,non-farmpayrolletc. While laborforceparticipationtendstobeacyclical,non-farmpayrollisveryprocyclicaland couldbeanothercandidatewithwhichtoproxythejobsreport. However,ADPResearch Institutealsoreleasesnumbersonnon-farmpayrollinitsreport,whichishighlycorrelated withthenon-farmnumbersintheBLS’sreport. SinceADP’sreportcomesoutbeforeBLS’s jobs report, we consider that non-farm payroll numbers are not actually new data and wouldalreadybeincorporatedinhousehold’sexpectationsatdayt–1priortothejobs report. Therefore,weuseunemploymentratetoproxytheBLS’sjobsreport,notnon-farm payroll. Theshocktoinformationcomingfrommacroeconomicannouncementswillvarydependingonwhichcaseweconsider. Insection3.2,weshowedthatinthecaseofsophisticated households,onlyunanticipatedchangesintheannouncementcaninfluenceexpectations. Sincehouseholdsareutilizingallavailableinformationtomaketheirforecast,weassume thattheirforecastisthesameasthatofprofessionalforecasters,whichwetaketobethe benchmark. ThisisalsoconsistentwithCarroll(2003),whoshowsthathouseholdexpectationsderivefromnewsreportsabouttheviewsofprofessionalforecasters. Weutilize datafromBloomberg’sConsensusForecasttogetinformationonprofessionalforecasters’ expectations. Beforeeveryannouncement,Bloombergasksexpertswhattheythinkwill occurintheupcomingannouncement. FollowingGürkaynaketal.(2005),wedefine: SurpriseX t = X t –E t P – F 1 (X t ) where X is the announcement that comes out on date t, and EPF(X ) is the forecast t t–1 t of announcement X made using information at time t–1 by professional forecasters. t Surprise thusgivesameasureoftheunanticipatedcomponentofeveryannouncement. t Sincehouseholdshavethesameforecastasexperts,Surprise servesasourmeasureof t 16

shocktohouseholds’informationset. Inthecasewhenhouseholdsarenaive,householdspredictthatthemacroeconomicvariablewilltakethesamevaluethatithadinthepreviousannouncement. Theshockinthis casewillbegivenbythedifferenceinthevalueofthevariablefromthisannouncementto theprevious. Therefore,theunanticipatedcomponentofeachmacroeconomicannouncementcanbesummarizedasfollows:  ∆X ifnaive t ShockX = t SurpriseX t ifsophisticated 5 Results WenowmoveontoestimateEquation4forbothsophisticatedandnaivehouseholdsfor Gallupdata. Wewillthenexplorewhetherpositiveandnegativeshocksaffecthousehold expectationsdifferently. Next,wewillmoveontotheMichigansurvey. Wewillperform thesameexercisethatwedidforGallupbyestimatingthebaselineresponseandresponse toasymmetricshocks. WewillthenmakeuseofthelongertimeperiodofMSCtostudy howtheresponseofexpectationsdependsontheco-movementbetweenunemployment andinflation. 5.1 Gallup Daily Tracking Poll InthissectionwediscussseveralkeyresultsfromtheGallupDailyTrackingPoll(GDTP). Onceweestablishourbaselineresults,wegodeeperintoanalyzingwhethershockshave asymmetriceffectsornot. Wethendiscussourmotivationtoexploitthemicrodatafrom theMichiganSurveyofConsumers. 5.1.1 BaselineResults WeestimatealocalprojectionasgiveninEquation4tostudywhetherhouseholdexpectationsrespondtomacroeconomicvariablesbyusingtheirannouncementstoidentify shockstothevariables. Weidentifytheeffectofanannouncementthatcomesoutondayt bycomparingthechangeinexpectationsthedaybeforetheannouncementtothedayafter. Theuseofnarrowwindowsallowsustogiveacausalinterpretationtothecoefficient. We varythehorizontolookatuptofivedaysaftertheannouncementintheregressiontables, anduptotwentyfivedaysintheimpulseresponses(seeappendix). Thisallowstimefor 17

householdstoupdatetheirexpectationsinresponsetonewinformation. However,the broader the window becomes, the less precise the estimate will be, because there is a greaterchancethathouseholdscouldbeexposedtoothernewinformationduringalonger timeperiod. Table2: ResponseofHouseholdExpectationstoMacroeconomicAnnouncements PanelA:SophisticatedHouseholdsModel y : ∆(ExpectationsIndex) Day0 Day1 Day2 Day3 Day4 Day5 t t Surprise(Unemp) -0.564∗ -0.562∗ -0.640∗ -0.374 -0.173 -0.568∗ (0.298) (0.316) (0.342) (0.299) (0.300) (0.313) Surprise(CPI) 0.433 -0.173 0.396 -0.129 -0.419 0.108 (0.290) (0.306) (0.308) (0.282) (0.288) (0.295) Surprise(GDP) 0.172 0.430 0.252 0.685∗∗ -0.0762 0.189 (0.324) (0.325) (0.324) (0.323) (0.341) (0.327) Surprise(Housing) 0.0103 -0.108 0.110 0.648∗∗ -0.214 0.745∗∗ (0.321) (0.331) (0.331) (0.328) (0.343) (0.337) PanelB:NaiveHouseholdsModel Day0 Day1 Day2 Day3 Day4 Day5 ∆(Unemp) -0.785∗∗∗ -0.633∗∗ -0.950∗∗∗ -0.387 -0.185 -0.582∗ (0.300) (0.316) (0.329) (0.305) (0.303) (0.307) ∆(CPI) 0.0741 -0.0804 0.672∗∗ 0.306 0.667∗∗ 0.156 (0.288) (0.292) (0.294) (0.290) (0.294) (0.296) ∆(GDP) 0.791∗∗ 0.818∗∗ 0.625∗ 0.960∗∗∗ 0.395 0.203 (0.343) (0.345) (0.351) (0.344) (0.352) (0.356) ∆(Housing) -0.485 -0.341 -0.171 0.569∗ -0.128 0.376 (0.337) (0.352) (0.350) (0.343) (0.344) (0.349) Notes:Thistablereportsestimatesofβ fromEquation4.Hereeachcellisthecoefficientfromaseparate h regressionequation.Wecontrolfordemographicssuchasage,income,education,race,gender,political affiliation,andstateofresidenceoftherespondent.Surveyweightsareused.Standarderrorsarepresented inparentheses.*p<0.10,**p<0.05,***p<0.01.DatausedfromGallup.IncandBloombergFinanceLP. Sophisticated Expectations Model. Panel A in Table 2 presents our primary results from Equation 4. The columns denote the number of days from t (i.e., from when the 18

announcement came out). Hence the dependent variable in column 1 is the change in expectationsbetweendaytandt–1,incolumn2isthechangeinexpectationsbetweenday t+1andt–1,andsoon. Eachcellisaseparateregression.14 Inthesophisticatedmodel,we findthatexpectationsrespondnegativelytounanticipatedchangesintheunemployment rate,butnottoothervariables. Aonestandarddeviationsurpriseinunemploymentcauses ourExpectationsIndextodeclineby0.5%. Inotherwords,aonestandarddeviationsurprise intheunemploymentratecausestheproportionofpeoplewhowereoptimisticaboutthe futuretodeclineby0.5%. BytheLawofLargeNumbers,thisimpliesthattheprobability thatoneagentwasoptimisticabouttheeconomydeclinedby0.5%. In Figure 4, we plot the dynamic response of household expectations over time. We report 95% confidence intervals. In the sophisticated model, households respond the most to unanticipated changes in unemployment rate. In contrast, for other variables, expectationsdonotrespondonimpactnoristhereanydynamicimpact. Thereisamild responsetohousingondays3and5,butitisnotpersistent. Unanticipatedshocksfrom outputgrowth,inflation,andhousingstartsdonotchangehouseholdexpectationsover timeinasignificantway. Ourresultsindicatethathouseholdsconsidertheunemploymentrateanimportantindicatorintheirexpectationformationprocess. Thefactthathouseholdsgiveimportancetothe labormarketwhileformingexpectationsisnotsurprising,sincelaborincomeisthelargest componentoftotalincomeformosthouseholds. Inaddition,beingunemployedhashuge negativeeffectsonthehealthandwell-beingofhouseholds(Sullivan&VonWachter2009, Blanchflower&Oswald2004,Lucasetal.2004,Michaillat&Saez2021). Carbone&Hey (2004) and Saporta-Eksten (2014) show that changes in labor markets influence households’decisions. Itisthereforenotsurprisingthatthelabormarketinfluenceshousehold expectationsabouttheeconomyaswell. Interestingly,householdsdonotadjusttheirexpectationsinresponsetounanticipated changesinanyoftheothervariables,includingoutput,inflation,andhousing. Thefact thathouseholdsdonotrespondtounanticipatedmovementsininflationisconsistentwith Andradeetal.(2023). Furthermore,itmustbenotedthatoursampleperiodfrom2008to 2017wasmostlyaperiodoflowinflation,whichmightalsocontributetothenon-response ofhouseholdexpectations. 14Wehavecombinedallregressionsintoonetableforeaseofviewing. Individualregressionscanbe foundintheAppendixTables9aand9b. 19

NaiveExpectationsModel. Wenowconsiderthenaivemodel. Inthisscenario,oncean announcementismade,householdsupdatetheirbeliefstothatvalue,whichanchorstheir beliefsuntilanewannouncementismade. Thus,ourmeasureofshocktohouseholds’ informationisthedifferencebetweenthevalueofthemacroeconomicvariableinthis announcementandthevalueinthepreviousone;thatis,ShockX t = ∆X t = X t –X t–1 . We followthesamehighfrequency–localprojectionapproachoutlinedintheprevioussection andestimateEquation4withtheshockbeing∆X . Onceagain,wevarythehorizonhto t lookatuptofivedaysaftertheannouncementintheregressiontable,anduptotwenty fivedaysintheimpulseresponses. PanelBinTable2reportstheresponseofhouseholdexpectationstonewinformationon macroeconomicvariablesforthenaivemodel. Wefindthatnaivehouseholdsupdatetheir expectationsinastatisticallysignificantwaytochangesintheunemploymentrateand outputgrowth,butnottochangesinothervariables. Aonestandarddeviationincreasein theunemploymentratecausestheshareofoptimistsintheeconomytodeclineby0.79%, whereasaonestandarddeviationincreaseinGDPgrowthrate(Advance)causestheshare ofoptimiststoincreaseby0.787%. InFigure4, weplotthedynamicresponseofhouseholdexpectationsovertimeforthe naivemodel. Analogoustothecaseofsophisticatedhouseholds,naivehouseholdsalso respondmosttoshocksintheunemploymentrate. Onceagain,outputgrowthandinflation donotrespondonimpactnordotheyhaveanydynamiceffectsonhouseholdexpectations. Therearesmalleffectsonhousingbuttheyarenotconsistent. Forunemployment,thecoefficient–0.79inthenaivecaseislargerinmagnitudethanthe coefficientinthesophisticatedcase(–0.568). Sincethesetwocasesprovidebounds,we canconcludethatthetruedecreaseinoptimismaboutthefuture(takingintoaccountthe confidenceintervals)aftertheunemploymentannouncementisbetween[–0.37%,–1.1%].15 Similarly, after the GDP announcement, expectations will change between [–0.16,1.1]. Sincethisintervalcontainszero,wecannotconcludethattheGDPannouncementhasa significanteffectonhouseholdexpectations. Thus,itisonlyshockstounemploymentthat affecthouseholdexpectations. 15Theconfidenceintervalinthesophisticatedcaseis[–0.37,–0.87],andtheconfidenceintervalinthenaive caseis[–1.1,–0.5]. 20

5.1.2 Asymmetry So far, we have looked at the response of household expectations to aggregate shocks. However, shocks in opposing directions could have different effects since they signify different information. Thus, it is possible that by reporting the net effect, we are missing out on differential movements caused by these shocks. To rectify this, we separate announcementsintotwocategories: thosewithpositiveshocksandthosewithnegative shocks. Weestimatetwoseparateregressions: Figure4: DynamicResponseofHouseholdExpectationstoMacroeconomicAnnouncements (a)Unemployment (b)CPI (c)GDPAdvance (d)HousingStarts Notes:Thisfigureshowsthecoefficientplotfor(a)Unemployment,(b)CPI,(c)GDPAdvanceand (d)HousingStartsannouncementsforNaive(Blacksolidcircle)andSophisticated(Bluesolid Diamond)expectationsmodelsfromtheGallupDailyTrackingPoll. Theshadedbaringreen betweentheNaiveandSophisticatedmodelcoefficientsistheregionwherethetruecoefficient exists.The95%confidenceintervalforNaive(gray)andSophisticated(lightblue)modelsarealso plotted.DatausedfromGallup.IncandBloombergFinanceLP. 21

E t i +h [Z]–E¯ t–1 [Z] = α 1h +β 1h ×(ShockX t |ShockX t > 0)+D i 1t+h +ϵ 1th (5) E t i +h [Z]–E¯ t–1 [Z] = α 2h ++β 2h ×(ShockX t |ShockX t < 0)+D i 2t+h +ϵ 2th (6) wheret isthedayoftheannouncement,hindicatesdaysfromt,Eτ indicatesexpectations formed by person i on day τ, Di denotes demographic information for person i, and t+h ShockX denotestheshocktohouseholds’informationset. ThevalueofShockX will,of t t course,bedifferentinourtwocases. Inthesophisticatedcase,ShockX = SurpriseX . A t t positive(negative)surprisemeansthattheactualvalueofthemacroeconomicvariable turnedouttobehigher(lower)thanforecasted.16 Inthenaivecase,ShockX = ∆X . Soa t t positive(negative)shocknowmeansthatthevariableincreased(decreased)invaluefrom thisannouncementtothelast. Given that our baseline results suggest a large role of unemployment in affecting expectations,wefocusourattentionontheunemploymentannouncementinthissection. However,giventhepervasivenessofinflationandit’saffectonhouseholdexpectations intheliterature,wewillcontinuestudyingtheinflationannouncementaswell. Table3 reportstheresultsforasymmetricresponseofhouseholdexpectations. PanelAreports resultsforsophisticatedhouseholds,whilepanelBreportsresultsfornaivehouseholds. Wefindthatpositiveshockstounemploymentaffecthouseholdexpectationsmuchmore thannegativeshocks. Notethatapositiveshocktounemployment(i.e.,whenunemploymentishigherthananticipated)indicatesaworseningofeconomicconditions,whereas a negative shock to unemployment indicates an improvement of economic conditions. Wefindthathouseholdsrespondasymmetrically: theybecomepessimisticuponreceivinginformationindicatingworseningeconomicoutcomes,butdonotnecessarilyturn optimistic upon receiving information indicating improving economic outcomes. The coefficientforapositiveshockisalsomuchlargerthanthecoefficientforanegativeshock in both the sophisticated and naive cases. These results are not surprising, especially whenweconsiderthatunemploymenthasbeendecliningsincetheGreatRecessionanda higherthanexpectedunemploymentratehasbeenassociatedwitharecession. Turning now to CPI, we observe that in the case of sophisticated households, only positivesurpriseshaveaneffectonhouseholdexpectations. However,inthecaseofnaive households,bothpositiveandnegativeshocksinfluenceexpectations,andinfact,negative 16Notethatapositivesurprisemightimplydifferentthingsfordifferentvariables.Forexample,apositive surpriseinunemploymentindicatesaworseningoftheeconomy,whereasapositivesurpriseinGDPgrowth impliesanimprovementintheeconomy. 22

shocks continue affecting expectations even five days after the announcement. Again, giventhatGDTPspansthedecadeafterthe2008recession,whichisatimewheninflation islowandmostlyfalling,itmakessensetoseeagreatereffectonnegativeshocks. 5.1.3 ExpandingtotheMichiganSurveyofConsumers Estimates from the Gallup survey reveal that households respond primarily to unemploymentannouncements,regardlessofwhethertheyaresophisticatedornaiveintheir forecasts,butnotnecessarilytoCPIannouncements. Thisfindingisparticularlystriking, given that most of the literature as well as conventional macroeconomic models focus predominantlyoninflationexpectations. Weidentifytwothreatstoexternalvalidity-i) thereissomethingspecificaboutthetimeperiodoftheGallupsurveythatisgivingthese results(forexample,inflationwasnotaconcerninthisdecade),andii)ourmeasureof expectationsisaqualitativesentimentbasedindex,andisnotnuancedenoughtocapture changesduetoothermacroeconomicvariables. Toaddressbothofthese,wemovetothe Michigansurvey. WhileGDTPprovidesacleanidentificationofhowhouseholdsrespondtoeconomicnews, the sample covers only the decade after the Great Recession, limiting the scope of our analysis. Tovalidateourfindingsoveralongerperiodandfurtherexplorethenuancesof householdexpectations,weturntomicrodatafromtheMichiganSurveyofConsumers (MSC), which spans from 1980 to 2019. Our use of the MSC microdata serves two key purposes: (1) to ensure that the household response estimates from the Gallup survey are not merely a function of the sample period, and (2) to extend the analysis beyond the traditional focus on inflation expectations and CPI announcements by examining howhouseholdexpectationsrespondtonewsaboutunemployment. Givenourbaseline findingsthathouseholdexpectationsreactmorestronglytolabormarketconditions,itis essentialtoassesstheirresponsesunderdifferentscenarios,includingtheco-movement ofinflationandunemployment,periodsofhighandlowinflationandunemployment,and positiveversusnegativeeconomicshocks. Thefollowingsectionsaddresstheseobjectives indetail,sheddinglightonthefactorsdrivinghouseholdexpectations. Estimates from the Gallup survey reveal that households respond primarily to unemploymentannouncements,regardlessofwhethertheyaresophisticatedornaiveintheir forecasts,butnotnecessarilytoCPIannouncements. Thisfindingisparticularlystriking, given that most of the literature as well as conventional macroeconomic models focus predominantlyoninflationexpectations. Weidentifytwothreatstoexternalvalidityofour results. Thefirstthreatcomesfromthepossibilitythatissomethingspecificaboutthetime 23

periodoftheGallupsurvey(2008-2017)thatisgivingtheseresults. Forexample,itcouldbe thatbecauseinflationisprettylowinthisperiod,householdsarenotpayingattentionto it,andthatiswhyshockstoinflationfailtoinfluencehouseholdexpectations. Toensure ourresultsarenotsensitivetotime,weturntotheMichiganSurveyofConsumers(MSC), whichisthelongestrunningsurveyofhouseholdexpectationsintheUS. Table3: AsymmetricResponseofHouseholdExpectationstoMacroeconomicAnnouncements PanelA:SophisticatedHouseholdsModel Day0 Day1 Day2 Day3 Day4 Day5 Surp(Unemp) > 0 -0.457 -1.373∗∗ -0.791 -0.0412 0.255 -0.147 t (0.588) (0.571) (0.579) (0.562) (0.568) (0.583) Surp(Unemp) < 0 -0.0543 0.262 -0.634 -1.489∗∗ -0.518 -1.334∗ t (0.728) (0.795) (0.938) (0.739) (0.716) (0.747) Surp(CPI) > 0 0.953∗∗ 0.325 1.028∗∗ 0.400 -0.0809 0.770∗ t (0.458) (0.510) (0.523) (0.446) (0.449) (0.451) Surp(CPI) < 0 0.101 -0.703 -0.0379 -0.675∗ -1.081∗∗∗ -0.234 t (0.430) (0.436) (0.434) (0.410) (0.419) (0.428) PanelB:NaiveHouseholdsModel Day0 Day1 Day2 Day3 Day4 Day5 ∆(Unemp) > 0 -1.209∗∗ -1.804∗∗∗ -2.275∗∗∗ -0.710 -0.426 -0.641 t (0.524) (0.526) (0.501) (0.516) (0.503) (0.508) ∆(Unemp) < 0 0.630 2.546∗∗ 1.876 -1.038 0.720 -1.465 t (1.037) (1.096) (1.223) (1.154) (1.065) (1.104) ∆(CPI) > 0 -1.386∗∗ -1.885∗∗∗ 0.0382 -0.928 -0.0708 -0.0560 t (0.628) (0.662) (0.677) (0.637) (0.639) (0.636) ∆(CPI) < 0 1.680∗∗∗ 0.859 3.078∗∗∗ 1.842∗∗∗ 1.173∗ 1.590∗∗∗ t (0.609) (0.620) (0.606) (0.586) (0.602) (0.616) Notes:Thistablereportstheestimatesofβ fromEquations5and6.Here,β ischangeintheexpectations h h duetopositiveornegativeshockintheBLSjobsreportandCPIannouncement,inthewindow[t–1,t+h] wheretisthedayoftheannouncementandh=0,1,2,3,4.Hereeachcellisthecoefficientfromaseparate regressionequation.Wecontrolfordemographicssuchasage,income,education,race,gender,political affiliation,andstateofresidenceoftherespondent.Surveyweightsareused.Standarderrorsarepresented inparentheses.*p<0.10,**p<0.05,***p<0.01.DatausedfromGallup.IncandBloombergFinanceLP. 24

Thesecondthreattoexternalvaliditycomesfromthefactthatourmeasureofexpectations in GDTP is a qualitative sentiment based index, and is not nuanced enough to capture changesduetoothermacroeconomicvariables. Weovercomethisbyusingpointestimates ofhouseholds’inflationexpectationsfromtheMSC.Sinceinflationexpectationsisalsoa keyvariableinseveralmacroeconomicmodels,itisimportanttostudyifitshowssimilar patterns as our qualitative indices/if it also reacts to the labor market/if our results go throughwithit. Figure5: AsymmetricResponseofHouseholdExpectationstoMacroeconomicAnnouncements (a)Unemployment:PositiveShock (b)Unemployment:NegativeShock (c)CPI:PositiveShock (d)CPI:NegativeShock Notes:Thisfigureshowsthecoefficientplotfor(a)positiveunemploymentshock,(b)negative unemploymentshock,(c)positiveCPIshocksand(d)NegativeCPIshockforNaive(Blacksolid circle)andSophisticated(BluesolidDiamond)expectationsmodelsfromtheGallupDailyTracking Poll.TheshadedbaringreenbetweentheNaiveandSophisticatedmodelcoefficientsistheregion wherethetruecoefficientexists.The95%confidenceintervalforNaive(gray)andSophisticated (lightblue)modelsarealsoplotted.DatausedfromGallup.IncandBloombergFinanceLP. 25

UsingMSCalsohasseveralotheradvantages. Sincethedataspanalmostfortyyearsfrom 1980to2019(westopbeforetheCovid-19pandemic),itallowsustostudythebehaviorof expectationsinawidearrayofeconomicconditions. ThelongertimeframeofMSCalso helps us in getting more power for our sub-sample regressions. GDTP contains twelve monthlyannouncementsovertenyearsforatotalof120announcements. Tocheckfor asymmetry,weneededtofurtherdividethese120announcementsintotwocategories, which reduces the power of our regressions. MSC alleviates this issue to a large extent becauseofthelongersampleperiod. 5.2 Michigan Survey of Consumers WefollowYork(2023)inusingmicrodatafromtheMichiganSurveyofConsumers(MSC) tocreateadailycrosssectionandre-estimateEquation4. However,thenumberofsurvey respondentseachdayinMSCisusuallylow,sincethesurveyhas500householdsspread overonemonth. Therefore,toensurethatourstatisticalanalysishasenoughpower,we createasevendaylongwindowaroundeachannouncementdateandexaminewhether expectations changed the week after the announcement compared to the week before. We report results for the first week in tables 4 and 6. We estimate the following local projection: E w i [Z]–E¯ w–1 [Z] = α w +β w ·ShockX t +D i w +ϵi tw (7) wheret isthedayoftheannouncement,w = [t,t+6]i.e. w indicatestheweek(sevendays) aftertheannouncement,w–1indicatestheweek(sevendays)beforetheannouncement,Ei τ indicatesexpectationsformedbypersonionweekτ,Di denotesdemographicinformation τ forpersonisurveyedinweekτ,andShockX denotestheshocktohouseholds’information t set. Similar to Gallup, since Michigan is also not a panel at the daily level, we take the averageexpectationsintheweekprecedingtheannouncementandsubtractthosefrom personi’sexpectation. Weusethesametwomeasuresofshockstohouseholdsthatwe havebeenusingsofar. Inthesophisticatedcase,ShockX = SurpriseX .17 Inthenaivecase, t t ShockX = ∆X .18 t t WeconsidertwomeasuresofexpectationsfromMSC.Thefirstisthefractionofoptimists, 17BloombergConsensusForecast,whichweusetocalculateourshockinthecaseofsophisticatedhouseholds,isonlyavailablepost1997. 18Tocalculatetheshockinthecaseofnaivehouseholds,wehaveuseddataontheinitialreleaseofeach announcement.However,wedon’thaveinformationoninitialreleasesfortheoldertimeseries(pre-1996). Therefore,weusethefinalreviseddatatocalculateournaiveshocksothatwecanhavetheseriesstarting allthewayfrom1980. Usuallytherevisionsareminor,andwedon’texpectthistoaffectourresultsina significantway.Wedorobustnessbycheckingtheeffectofinitialreleasespost1997,andfindsimilarresults. 26

calculatedfromthequestiononexpectationsabut12-monthaheadgeneralbusinessconditionsintheeconomy. ThisiscomparabletotheGDTPshareofoptimists. Foroursecond measureofexpectations,weusepointestimatesof12-monthaheadinflationexpectations. Sinceinflationexpectationsarebyfarthemostpopularmeasureofexpectationsinthe literature,wefocusonthemtotryandstudywhetherourresultscarryovertothemornot. 5.2.1 BusinessConditionsintheCountryasaWhole TobeabletocompareourresultswithGDTP,westartbystudyingMichigan’sbusinessconditionindex. SimilartoGallup,thisindexdenotestheshareofpeoplewhoareoptimistic about future business conditions. An increase in the value of the index indicates that morepeoplearebecomingoptimistic,whereasadecreaseindicatesthatfewerpeopleare optimisticormorepeoplearepessimistic. PanelAoftable4reportstheresultsforsophisticatedhouseholds.19 Column1reportsresultsaroundtheunemploymentannouncement andcolumn2reportsresultsaroundtheCPIannouncement. PanelAreportsresultsfor sophisticatedhouseholds. Ourbaselineresultisthatshockstobothunemploymentand CPI affect household expectations, but shocks to CPI have a larger effect, in line with conventionalwisdom. Thisrelationship,however,breaksdownonceweconsiderpositive andnegativeshocksseparately. Wefindthatpositivesurprisesaffectexpectationsmore thannegativesurprises. Apositivesurpriseoccurswhenthevalueofthemacroeconomic variableasreleasedintheannouncementturnsouttobelargerthanexpected. Households interpretthisasbadnewsinthecaseofunemploymentandinflation,anddecreasetheir expectations. Wealsoobservethatshockstounemploymentcauselargermovementsin householdexpectationsthanshockstoCPI,whichisconsistentwiththeresultsinGDTP. PanelBreportsresultsassumingthathouseholdsarenaive. Inthebaseline,wefindthat changesinunemploymenthavealargereffectonhouseholdexpectationthanchangesin CPIannouncement. Lookingatasymmetry,weonceagainfindthatpositivechanges(i.e. an increase in the variable from last period) have bigger effects than negative changes. Focusingonpositivechanges,wefindanincreaseinunemploymentaffectshousehold expectationsmorethananincreaseinCPI.However,thereverseistruefornegativeshocks. Thismakessensebecauseinmostofoursample,inflationwasverylow,andinfact,there wereevenfearsofdeflationduringsomeyears. 19Note that although we have data until 2021, we are currently reporting results for the pre-Covid-19 period.ThatisbecausetheCovid-19pandemicwasahugeshock(withstandarddeviation13timeshigher thanaverage),andincludingitbiasestheresults.WepresentresultsincludingtheCovid-19periodinthe appendix. 27

5.2.2 InflationExpectations Next,wemoveontostudyingtheeffectsofaquantitativemeasureofexpectations-the twelvemonthaheadinflationexpectations. Inflationexpectationsfeedintoalotofmacroeconomic models, so it is important to study them and see if they behave similarly to qualitativemeasuresofexpectations. InTable6,wepresenttheresponseofhouseholds’12-month-aheadinflationexpectations tomacroeconomicannouncements. PanelApresentsresultsforsophisticatedhouseholds whilepanelBpresentsresultsfornaivehouseholds. Wefindsimilarresultstothequalitative measures in MSC and GDTP. In the baseline, in the sophisticated case, inflation expectationsareonlysignificantlyaffectedbytheunemploymentannouncement,notby theCPIannouncement. Inthenaivecase,expectationsrespondtobothannouncements. Lookingatasymmetry,weonceagainfindthatpositiveshockshaveamuchlargereffect thannegativeshocks. Forsophisticatedhouseholds,onlythepositiveshockssignificantly affectinflationexpectations,andevenhere,unemploymentaffectsinflationexpectations more than CPI. In the naive case, inflation expectations respond to both positive and negative shocks to unemployment, but only to positive shocks to CPI. The response to unemploymentisalsoslightlylargerthantheresponsetoCPI.Theseresultsshowthat unemploymentisanimportantdeterminantofinflation. 5.2.3 SupplyversusDemandShocks GiventhatwehavedataforfortyyearsfromMSC,itispossibletodividethetimeperiod intosub-samplesbasedonvariousepisodesofco-movementofinflationandunemploymentandstudythoseseparately. Welookatfourscenarios-(1)bothunemploymentand inflation20 areincreasing,(2)botharedecreasing,(3)unemploymentisincreasingwhile inflationisdecreasing,and(4)inflationisincreasingwhileunemploymentisdecreasing. Table5reportstheproportionofoccurrencesofeachscenario. We can also interpret these scenarios as supply and demand shocks. Demand shocks typicallymoveoutputandinflationinthesamedirection,whichmeanstheymoveunemploymentandinflationinoppositedirections. Supplyshocksmoveoutputandinflationin oppositedirections,implyingtheymoveunemploymentandinflationinthesamedirection. Thusourfirstscenarioofbothunemploymentandinflationincreasingwouldcorrespond toanegativesupplyshock,andsecondscenarioofbothdecreasingwouldcorrespondtoa positivesupplyshock. Similarly,thethirdscenarioofincreasingunemploymentwith 20Weconsidermonthlychangeinunemploymentandmonthlychangeininflationtocreatethesescenarios. 28

Table4: ResponseofBusinessConditionsIndextoMacroeconomicAnnouncements y =BusinessOutlook X =U X =CPI t t t (1) (2) PanelA:SophisticatedHouseholdsModel1997-2019 Baseline SurpX -1.2*** -1.6*** t (0.32) (0.41) Asymmetry SurpX > 0 -13*** -6.5*** t (1.27) (1.22) SurpX < 0 -0.8 0.6 t (0.84) (1.34) Scenarios ∆U > 0,∆π > 0 -3.8*** -1* (1.07) (1.05) ∆U < 0,∆π < 0 2.7*** -3.5*** (0.96) (1.11) ∆U > 0,∆π < 0 -4.4** -0.4 (1.41) (0.91) ∆U < 0,∆π > 0 3.9*** -2.8*** (1.14) (0.94) PanelB:NaiveHouseholdsModel1980-2019 Baseline ChangeX -5*** -1.7*** t (0.24) (0.32) Asymmetry ChangeX > 0 -7*** -8*** t (0.45) (0.62) ChangeX < 0 -2.3*** 4*** t (0.64) (0.47) Scenarios ∆U > 0,∆π > 0 -17*** 0.5 (0.94) (0.56) ∆U < 0,∆π < 0 -0.2 -3.5*** (0.77) (0.93) ∆U > 0,∆π < 0 -5.7*** -3.2*** (0.93) (0.63) ∆U < 0,∆π > 0 -6.1*** -2.8*** (1.14) (0.90) Notes:Thistablereportsestimatesofβ fromEquation7.Wecontrolfor h demographicssuchasage,income,education,maritalstatus,gender, and region of residence of the respondent. Survey weights are used. Standarderrorsarepresentedinparentheses. *p < 0.10, **p < 0.05, ***p < 0.01. ThedataisfromtheSurveyofConsumers,Universityof Michigan.SurveyResearchCenter.SurprisesfromBloombergFinance LP. 29

decreasinginflationwouldcorrespondtoanegativedemandshock,whereasthefourth wouldcorrespondtoapositivedemandshock. ThesectionnamedScenariosinbothpanels oftable4andtable6presentstheresultsofthisexercise. Wefirstexaminetheresponseofourbusinessconditionsindextothesescenarios. When both unemployment and inflation increase, we find that both sophisticated and naive householdsreactstronglytoshockstounemployment,withshockstoCPIhavinglittleor nosignificanteffect. Whenbothunemploymentandinflationaredecreasing,shocksto unemploymentandtheCPIhavenearlythesameeffectforsophisticatedhouseholds,but onlyshockstotheCPIhaveasignificanteffectfornaivehouseholds. Whenunemployment andinflationmoveintheoppositedirection,householdsplacesignificantlymoreweight onshockstounemploymentinformingtheirexpectations. Thinkingintermsofsupply anddemandshocks,wefindthattheonlycaseinwhichshockstoCPImattermorethan shockstounemploymentisthecaseofpositivesupplyshocks. Figure6: ResponseofBusinessConditionsIndextoMacroeconomicAnnouncements (a)UnemploymentShocks (b)CPIshocks Notes: This figure shows the coefficient plot for response of business conditions index to (a) unemployment shocks and (b) CPI shocks Naive (Black solid circle) and Sophisticated (Blue solid Diamond) expectations models from the Survey of Consumers, University of Michigan. SurveyResearchCenter.TheshadedbarinlightbluebetweentheNaiveandSophisticatedmodel coefficientsistheregionwherethetruecoefficientexists.The95%confidenceintervalforNaive (solidlines)andSophisticated(dashedlines)modelsarealsoplotted.SurprisesfromBloomberg FinanceLP. Turningtoinflationexpectations,weobservesimilarpatterns. Whenbothunemployment andinflationincrease,wefindthatbothsophisticatedandnaivehouseholdsreactstrongly toshockstounemployment,withshockstoCPIhavinglittleornosignificanteffect. Inthe secondscenario,whenbothunemploymentandinflationdecrease,onlyunemployment 30

hasasignificanteffect,andthattooonlyinthecaseofnaivehouseholds. Whenunemployment is increasing while inflation is decreasing, we find that shocks to unemployment arethelargerdriverofinflationexpectations. Inthefinalcasewhenunemploymentis decreasing while inflation is increasing, both naive and sophisticated households only respondtoshockstoCPI.Thinkingintermsofsupplyanddemandshocks,wefindthat heretheonlycaseinwhichshockstoCPImattermorethanshockstounemploymentis thecaseofpositivedemandshocks. Table5: DistributionofScenarios Scenario NumberofDays %ofSample (1) (2) ∆U > 0,∆π > 0 68 20% ∆U < 0,∆π < 0 105 30% ∆U > 0,∆π < 0 89 25% ∆U < 0,∆π > 0 86 25% Total 348 100% Notes:Thistableshowstheproportionofoccurrenceofeach scenariobetween1980-2019. SurprisesfromBloombergFinanceLP. Overall, these findings highlight the important role of news about unemployment in shaping household expectations, not only about the general economy but also about inflation. This suggests that labor market conditions are a crucial driver of household expectationformation,eveninoutcomestraditionallyassociatedwithinflationdynamics. OurresultsareconsistentwithMasolo(2022),whofindthatnewsaboutbusinesscycle andlabormarketfluctuationsarerelated,implyingthatpeoplelookatthelabormarketto infermovementsinbusinessconditions. 31

5.3 InterpretingtheSalienceofLaborMarketInformationinHousehold Expectations Oneofthemoststrikingandconsistentfindingsofouranalysisisthathouseholdsrespond morestronglytolabormarketnews,particularlyunemploymentshocks,thantoanyother macroeconomic indicator, including inflation. This holds true not only for subjective economicsentimentbutalsoforquantitativeinflationexpectations,andisrobustacross differentidentificationstrategies,surveydatasets,andtimeperiods. Inthissection,we discusspotentialexplanationsforthisempiricalregularityandsituateourfindingswithin recenttheoreticalframeworks. Figure7:Responseof12MonthAheadInflationExpectationstoMacroeconomicAnnouncements (a)UnemploymentShocks (b)CPIshocks Notes:Thisfigureshowsthecoefficientplotforresponseof12MonthAheadInflationExpectations to(a)unemploymentshocksand(b)CPIshocksNaive(Blacksolidcircle)andSophisticated(Blue solid Diamond) expectations models from the Survey of Consumers, University of Michigan. SurveyResearchCenter.TheshadedbarinlightbluebetweentheNaiveandSophisticatedmodel coefficientsistheregionwherethetruecoefficientexists.The95%confidenceintervalforNaive (solidlines)andSophisticated(dashedlines)modelsarealsoplotted.SurprisesfromBloomberg FinanceLP. SalienceandPersonalRelevance. First,labormarketnewsmaybemoresalientthan inflation statistics for many households. As noted by Bordalo et al. (2020), individuals oftenover-weightinformationthatiscompellingorpersonallyrelevant. Unemployment announcements are closely linked to personal job security and income prospects, and arefrequentlyreportedusingemotionallyresonantlanguage(suchas“joblosses,”“mass layoffs,”“recordunemployment”). Incontrast,householdsoftenfindinflationdifficult tounderstandandfrequentlyconflateitwiththeoverallpricelevel,complicatingtheir interpretationofinflation-relatednews(Stantcheva2024,Weberetal.2022). 32

Table6: Responseof12MonthAheadInflationExpectationstoMacroeconomicAnnouncements y t = E t π t+12 X t =U X t =CPI (1) (2) PanelA:SophisticatedHouseholdsModel1997-2019 Baseline SurpX 0.1*** 0.05 t (0.02) (0.03) Asymmetry SurpX > 0 0.7*** 0.3*** t (0.11) (0.11) SurpX < 0 0.05 -0.07 t (0.06) (0.11) Scenarios ∆U > 0,∆π > 0 0.3*** -0.02 (0.09) (0.09) ∆U < 0,∆π < 0 -0.03 0.1 (0.07) (0.08) ∆U > 0,∆π < 0 0.3** 0.02 (0.11) (0.08) ∆U < 0,∆π > 0 -0.1 0.2** (0.08) (0.07) PanelB:NaiveHouseholdsModel1980-2022 Baseline ChangeX 0.2*** 0.3*** t (0.03) (0.04) Asymmetry ChangeX > 0 0.5*** 0.6*** t (0.06) (0.08) ChangeX < 0 -0.3*** 0.04 t (0.07) (0.05) Scenarios ∆U > 0,∆π > 0 0.6*** 0.2*** (0.13) (0.07) ∆U < 0,∆π < 0 -0.5*** 0.05 (0.09) (0.08) ∆U > 0,∆π < 0 0.7*** 0.2*** (0.12) (0.07) ∆U < 0,∆π > 0 0.1 0.5*** (0.11) (0.09) Notes:Thistablereportsestimatesofβ fromEquation7.Weconh trolfordemographicssuchasage,income,education,maritalstatus, gender, and region of residence of the respondent. Survey weightsareused.Standarderrorsarepresentedinparentheses.* p < 0.10,**p < 0.05,***p < 0.01. ThedataisfromSurveyofConsumers,UniversityofMichigan.SurveyResearchCenter.Surprises 33 fromBloombergFinanceLP.

TheGDTPallowsustoobservewhetherrespondentsresideincountieswithhighorlow localunemploymentrates. Wefindthatindividualslivingincountieswithhigherlocal unemployment are more sensitive to changes in the national unemployment rate. We furtherfindthatrespondentsresidinginhighlocalunemploymentregionsdonotrespond toshockstoCPI.TheseresultsarereportedinAppendixTable13. Theseresultssuggestthat householdexpectationsmaybeshapedbyimmediateeconomicenvironment,indicating anelementofstate-dependenceatplaywhichwedonotdiscussfurtherinthispaperand leaveforfutureresearch. RationalInattentionandInformationProcessingCosts. Fromatheoreticalstandpoint, ourfindingsareconsistentwithrationalinattentionmodels(Sims2003,Gabaix2014),in whichagentsoptimallyallocatelimitedattentionacrosscompetinginformationsources. Undertheseframeworks,householdsdevoteattentiontovariablesthatprovidethehighest informationalbenefitatthelowestcognitivecost. Iflabormarketinformationissimpler, moretimely,andmoredirectlyrelevanttopersonalwelfarethaninflationdata,households willoptimallyfocuslimitedcognitiveresourcesonunemploymentratherthanoncomplex priceindices. Householdsmightalsousetheunemploymentrateasasummarystatisticfor broadereconomichealth,encompassingdemandconditions,incomestability,andeven futureinflation. Thisperspectiveexplainswhyinflationexpectationsrespondsignificantly tounemploymentshocks,evenwhenCPIchangesthemselvesarelesspronounced. AccessibilityandCommunication. Thestructureandcommunicationoflabormarket statisticsmayfurtherenhancetheiraccessibilityandtrustworthiness. Theunemployment rateisastraightforward,easilyinterpretedmetricwithcleardirectionalmeaning—rising unemploymentindicatesdeterioratingconditions,whilefallingunemploymentsuggests improvement. Inflationreports,bycontrast,aremorecomplex,subjecttomethodological debates,andharderfornon-expertstointerpretintuitively. Moreover,householdsoften misunderstandinflation,confusingchangesinthepricelevelwithinflationitself,andgenerallyperceiveinflationnegatively,indicatingwidespreadmisunderstanding(Stantcheva 2024). Collectively,thesefactorssuggestthatthelabormarketservesastheprimarylensthrough whichhouseholdsinterpreteconomicdevelopments. Thisbehavioralanchoringonlabor conditions provides a robust explanation for our empirical findings and highlights the importanceofdevelopingpsychologicallyinformedtheoriesofexpectationformation. It also indicates that strategies aiming to manage household expectations, especially regardinginflation,shouldexplicitlyconsiderlabormarketnewsasakeyintermediary 34

channel. 5.3.1 UnderstandingtheAsymmetricResponseofHouseholdExpectations Ourresultsrevealanasymmetryinhowhouseholdsrespondtomacroeconomicshocks: negativeshockstounemploymentorinflationleadtolargershiftsinexpectationsthan similarlysizedpositiveshocks. Thisholdsacrossbothsubjectivesentimentandquantitativeinflationexpectations,andisconsistentacrossnaiveandsophisticatedexpectation models. Wediscusssomeconceptualframeworkstointerpretthisasymmetry,drawingon behavioraleconomicsandmacroeconomicexpectationtheory. Wefindthathouseholdstendtoreactmorestronglytonegativeeconomicnewsthanto positivedevelopments,apatternconsistentwithnegativitybiasandlossaversion. When unemployment rises or inflation exceeds expectations, households experience heightened concern about job security and eroding purchasing power. In contrast, improvements—suchasfallingunemploymentorlowerinflation—oftenelicitmutedresponses. Thisasymmetricsensitivityalignswithprospecttheory,wherelossescarrygreaterpsychologicalweightthanequivalentgains,andhelpsexplainwhyexpectationsshiftmore sharplyinresponsetobadnewsevenwhenunderlyingshocksaresymmetric. A related mechanism is reference dependence, where households interpret economic developmentsrelativetoamentalbenchmarkofwhatis“normal.”Deviationsbelowthis reference point, like rising unemployment, trigger stronger reactions than equivalent deviationsaboveit. Thisbehavioriscompoundedbylimitedattentionandrelianceon heuristics,leadinghouseholdstoanchorexpectationsonsalientindicatorssuchaspast unemploymentrates. 5.4 Robustness and Additional Exercises Weconductseveralrobustnesscheckstoensurethevalidityofourresults. Thissection provides a summary of these exercises, with further details available in the appendix. Onepotentialconcernisthatthetimingofannouncementsmaybedrivingourfindings, particularlybecausetheBLSemploymentsituationreportistypicallyreleasedonthefirst Friday of each month. If households update their expectations primarily based on the first major announcement of the month, our observed effects could simply reflect this sequencingratherthanthespecificimpactoftheemploymentreportitself. Toaddressthis concern,weadjustouranalysisforsubsequentannouncements,suchastheCPIrelease, bycomputingthechangeinexpectationsasEi [Z]–E¯ first[Z],whereE¯ first[Z]istheaverage t+h t–1 t–1 35

expectationsbeforetheemploymentsituationreleaseeachmonth. Thisexerciseallows us to calculate changes in expectation around each release relative to the expectations setbeforethefirstannouncementofthemonth. Ourresultsarerobusttothisexercise whichindicatesthatourestimatesarenotdrivenbythesequencingofmacroeconomic news. AcloselyrelatedexerciseanalysisexaminestheBLSannouncementwindow. While theBLStypicallyreleasesitsemploymentreportonthefirstFridayofeachmonth,other relevantlabormarketindicators,suchastheADPemploymentreportandjoblessclaims data,arereleasedearlierinthesameweek—onWednesdayandThursday,respectively. To accountforpotentialinformationspillovers,weredefinethechangeinexpectationsfor unemploymentannouncementsasEi [Z]–E¯ Tuesday[Z]whereE¯ Tuesday[Z]representsthe t+h t–1 t–1 expectationspriortotheseearlylabormarketreleases. Ourresultsremainrobustunder thisspecification;however,naivehouseholdsexhibitaweakerresponse,suggestingthat earlyannouncementsonnon-farmpayrollandjoblessclaimsprovidehouseholdswith additionallabormarketinformation,therebymoderatingtheirreactiontotheBLSreport. Next, we perform several additional exercises and are highlighting some of these for brevity. Weconsiderseveralscenariosbasedontheabsolutelevelsofunemploymentand inflation. Wecontrolforrecessionssincehouseholdsentimentsmightbemoreresponsive inrecessionsandfindthatsophisticatedhouseholdsaremoreresponsiveduringrecessions whilenaivehouseholdsarenot. Wealsoestimateseverallocalprojectionswithcontrolsforthestockmarkets(S&P500, Nasdaq). Although,giventhefrequencyofsurveyexpectations,itisnotentirelypossible todisentanglewhetherthehouseholdresponseisdrivenbythemacroannouncements directlyortheresponseofthefinancialmarketstothemacroannouncements,sincethese are all simultaneous events. However, our results hold when we control for financial marketsaswellasnewsmedia. We find that households respond more in times of high unemployment and inflation relative to when both of these are low. Additionally, we construct a synthetic panel by matchingrespondentsbasedondemographicobservablesandanalyzehowexpectations adjust within this panel. Our results remain robust across all these exercises, further reinforcingthevalidityofourfindings. 36

6 Conclusion Inthispaper,weanalyzewhatinformationhouseholdsusetoadjusttheirexpectations abouttheeconomyaswellasinflation. Usinghigh-frequencydatafromtheGallupDaily Tracking Poll and the Michigan Survey of Consumers, we identify systematic patterns inexpectationformation,revealingthatlabormarketconditions,particularlyshocksto unemployment, play a central role in influencing household expectations, often more sothanshockstoinflationorothervariablessuchasoutputgrowthandhousingstarts. Wedocumentthathouseholdsrespondmorestronglytonegativeeconomicshocksthan to positive ones. Moreover, we find that even when inflation is increasing, unemploymentremainsthedominantfactorinhouseholdadjustmentstobothgeneraleconomic expectationsandinflationexpectations. Ourstudymakesseveralcontributions. First,wedevelopaframeworktoisolatetheunanticipatedcomponentofmacroeconomicannouncements,allowingustoprovidebounds totheexpectationadjustmentprocess. Second,weshowthatlabormarketinformationis crucialnotonlyforsubjectiveeconomicexpectationsbutalsoforinflationexpectations. Third,eveninperiodsofrisinginflationanddecliningunemployment,shockstounemploymentsignificantlyinfluencehouseholdexpectations,highlightingtheimportanceof labormarketinformationinthisprocess. Finally,weshowthatinflationexpectationsare primarilyshapedbyunemploymentshocks,exceptincasesofpositivesupplyordemand shocks,whereshockstothepricelevelplayamoredominantrole. References Andersen, T. G., Bollerslev, T., Diebold, F. X. & Vega, C. (2003), ‘Micro effects of macro announcements: Real-time price discovery in foreign exchange’, American economic review93(1),38–62. Andersen,T.G.,Bollerslev,T.,Diebold,F.X.&Vega,C.(2007),‘Real-timepricediscovery inglobalstock,bondandforeignexchangemarkets’,JournalofinternationalEconomics 73(2),251–277. Andrade,P.,Gautier,E.&Mengus,E.(2023),‘Whatmattersinhouseholds’inflationexpectations?’,JournalofMonetaryEconomics138,50–68. Andre,P.,Pizzinelli,C.,Roth,C.&Wohlfart,J.(2022),‘Subjectivemodelsofthemacroe- 37

conomy: Evidencefromexpertsandarepresentativesample’,TheReviewofEconomic Studies. Armantier,O.,BruinedeBruin,W.,Topa,G.,VanDerKlaauw,W.&Zafar,B.(2015),‘Inflationexpectationsandbehavior:Dosurveyrespondentsactontheirbeliefs?’,International EconomicReview56(2),505–536. Armona,L.,Fuster,A.&Zafar,B.(2018),‘Homepriceexpectationsandbehaviour:Evidence fromarandomizedinformationexperiment’,TheReviewofEconomicStudies86(4),1371– 1410. Bachmann, R., Berg, T.O.&Sims, E.R.(2015), ‘Inflationexpectationsandreadinessto spend: Cross-sectionalevidence’,AmericanEconomicJournal: EconomicPolicy7(1),1–35. Balduzzi,P.,Elton,E.J.&Green,T.C.(2001),‘Economicnewsandbondprices: Evidence fromtheustreasurymarket’,JournaloffinancialandQuantitativeanalysis36(4),523–543. Binder,C.C.,Campbell,J.R.&Ryngaert,J.M.(2024),‘Consumerinflationexpectations: Dailydynamics’,JournalofMonetaryEconomicsp.103613. Blanchflower,D.G.&Oswald,A.J.(2004),‘Well-beingovertimeinbritainandtheusa’, Journalofpubliceconomics88(7-8),1359–1386. BloombergFinanceLP(1996-),‘Bloombergterminals(open,anywhere,anddisasterrecoverylicenses)’. Boehm,C.E.&Kroner,T.N.(2023),Theus,economicnews,andtheglobalfinancialcycle, Technicalreport,NationalBureauofEconomicResearch. Bordalo,P.,Gennaioli,N.&Shleifer,A.(2020),‘Memory,attention,andchoice’,TheQuarterlyjournalofeconomics135(3),1399–1442. Borgschulte, M. & Martorell, P. (2018), ‘Paying to avoid recession: Using reenlistment toestimatethecostofunemployment’,AmericanEconomicJournal: AppliedEconomics 10(3),101–27. Caporale, G. M., Spagnolo, F. & Spagnolo, N. (2016), ‘Macro news and stock returns in theeuroarea: avar-garch-in-meananalysis’,InternationalReviewofFinancialAnalysis 45,180–188. Carbone,E.&Hey,J.D.(2004),‘Theeffectofunemploymentonconsumption: Anexperimentalanalysis’,TheEconomicJournal114(497),660–683. 38

Carroll,C.D.(2003),‘Macroeconomicexpectationsofhouseholdsandprofessionalforecasters’,theQuarterlyJournalofeconomics118(1),269–298. Coibion,O.,Georgarakos,D.,Gorodnichenko,Y.&vanRooij,M.(2023),‘Howdoesconsumptionrespondtonewsaboutinflation? fieldevidencefromarandomizedcontrol trial’,AmericanEconomicJournal: Macroeconomics15(3),109–52. URL:https://www.aeaweb.org/articles?id=10.1257/mac.20200445 Coibion,O.,Gorodnichenko,Y.,Kumar,S.&Pedemonte,M.(2020),‘Inflationexpectations asapolicytool?’,JournalofInternationalEconomics124,103297. Coibion,O.,Gorodnichenko,Y.&Weber,M.(2022),‘Monetarypolicycommunicationsand theireffectsonhouseholdinflationexpectations’,JournalofPoliticalEconomy130(6),1537– 1584. D’Acunto,F.&Weber,M.(2024),‘Whysurvey-basedsubjectiveexpectationsaremeaningful andimportant’,AnnualReviewofEconomics. Ehrmann, M., Pfajfar, D. & Santoro, E. (2017), ‘Consumers’ attitudes and their inflation expectations’,InternationalJournalofCentralBanking. Evans,M.D.&Lyons,R.K.(2008),‘Howismacronewstransmittedtoexchangerates?’, JournalofFinancialEconomics88(1),26–50. Gabaix,X.(2014),‘Asparsity-basedmodelofboundedrationality’,TheQuarterlyJournalof Economics129(4),1661–1710. GallupInc.(2017),‘Usdailytrackingpoll’. RetrievedfromtheBrownUniversityLibrary. Gürkaynak,R.S.,Sack,B.&Swanson,E.(2005),‘Thesensitivityoflong-terminterestrates toeconomicnews: Evidenceandimplicationsformacroeconomicmodels’,American economicreview95(1),425–436. Jiang,G.J.,Lo,I.&Valente,G.(2014),High-frequencytradingaroundmacroeconomic newsannouncements: Evidencefromtheustreasurymarket,Technicalreport,Bankof CanadaWorkingPaper. Jordà, Ò. (2005), ‘Estimation and inference of impulse responses by local projections’, Americaneconomicreview95(1),161–182. Kamdar, R. & Ray, W. (2024), Attention-Driven Sentiment and the Business Cycle, CAEPR, CenterforAppliedEconomicsandPolicyResearch. 39

Kuchler, T. & Zafar, B. (2015), ‘Personal experiences and expectations about aggregate outcomes’,TheJournalofFinance. LucasJr,R.E.(1976),Econometricpolicyevaluation: Acritique,in‘Carnegie-Rochester conferenceseriesonpublicpolicy’,Vol.1,North-Holland,pp.19–46. Lucas,R.E.,Clark,A.E.,Georgellis,Y.&Diener,E.(2004),‘Unemploymentalterstheset pointforlifesatisfaction’,Psychologicalscience15(1),8–13. Malmendier, U. & Nagel, S. (2015), ‘Learning from inflation experiences’, The Quarterly JournalofEconomics131(1),53–87. Masolo,R.(2022),‘Mainlyemployment: survey-basednewsandthebusinesscycle’. Mertens,K.,Lewis,D.J.&Makridis,C.(2020),‘Domonetarypolicyannouncementsshift householdexpectations?’. Mian, A., Sufi, A. & Khoshkhou, N. (2021), ‘Partisan bias, economic expectations, and householdspending’,TheReviewofEconomicsandStatisticspp.1–46. Michaillat, P.&Saez, E.(2021), ‘Resolvingnewkeynesiananomalieswithwealthinthe utilityfunction’,ReviewofEconomicsandStatistics103(2),197–215. Mueller,A.I.,Spinnewijn,J.&Topa,G.(2021),‘Jobseekers’perceptionsandemployment prospects: Heterogeneity,durationdependence,andbias’,AmericanEconomicReview 111(1),324–63. Potter,T.(2020),‘Learningandjobsearchdynamicsduringthegreatrecession’,Journalof MonetaryEconomics. Roth, C. & Wohlfart, J. (2019), ‘How do expectations about the macroeconomy affect personalexpectationsandbehavior?’,ReviewofEconomicsandStatisticspp.1–45. Saporta-Eksten,I.(2014),‘Jobloss,consumptionandunemploymentinsurance’. Sims, C. A. (2003), ‘Implications of rational inattention’, Journal of monetary Economics 50(3),665–690. Stantcheva,S.(2024),‘Whydowedislikeinflation?’,BrookingsPapersonEconomicActivity 2004(1),1–46. Sullivan,D.&VonWachter,T.(2009),‘Jobdisplacementandmortality: Ananalysisusing administrativedata’,TheQuarterlyJournalofEconomics124(3),1265–1306. 40

UniversityofMichigan.SurveyResearchCenter(n.d.),‘SurveyofConsumers’. URL:http://new.sca.isr.umich.edu/ Weber,M.,d’Acunto,F.,Gorodnichenko,Y.&Coibion,O.(2022),‘Thesubjectiveinflation expectationsofhouseholdsandfirms: Measurement,determinants,andimplications’, JournalofEconomicPerspectives36(3),157–184. York, J. (2023), ‘Do household inflation expectations respond to macroeconomic data releases?’,AvailableatSSRN4648451. 41

A Appendix Inthissection,wereportseveralresultsfromrobustnesschecksaswellassomestatistics tobetterunderstandthedatainourstudy. A.0.1 SummaryStatistics Table7a: GeneralSummaryStatistics Variable GDTP MSC (1) (2) Age 47years 49years Female 51% 54% LowIncome 38% 43% MiddleIncome 47% 33% HighIncome 15% 25% White 73% NA Black 12% NA <HighSchool 11% 4% HighSchool 35% 6% SomeCollege 31% 28% N 1,705,158 277,160 This table records summary statistics for demographicvariablesforbothGDTPand MSC.Surveyweightsused. 42

A.1 Gallup A.1.1 HeterogeneityinHouseholdExpectations We observe substantial heterogeneity in household expectations across demographic groups. InFigure8a,wefindthatcollegegraduatesweresystematicallythemostoptimistic overtime. Thiscanbelinkedtojobstatus,sincecollegegraduatestendtohavethehighest employmentratesandthustendtobeconsistentlymoreoptimisticthantheunemployed (Figure 8b). Looking across age groups in Figure 8c, we find that younger respondents are consistently more optimistic than middle-aged and older respondents. While little differenceinoptimismexistsacrossgendersinmostyears(Figure8d),thereseemstobe asharpincreaseamongmenpost2016. Interestingly,wefindareversalwhenlookingatheterogeneityacrosspoliticalaffiliation andrace. AsFigure8fdemonstrates,households’optimismisproportionaltotheirparty affiliation,andchangesdependingontherulingparty(Mianetal.2021).21 Thisreversalis alsopresentwhenlookingatheterogeneitybyrace(Figure8e).22 21Atthestartof2008,whentheRepublicanpartyisinpower,weobservethathouseholdsaffiliatedwiththe RepublicanpartyaremoreoptimisticthanthoseaffiliatedwiththeDemocraticparty.Inthe2008elections, whentheDemocratswin,weseethatexpectationsofhouseholdsaffiliatedwiththemincrease,whilethose ofhouseholdsaffiliatedwiththeRepublicansdecline.Democratsstayconsistentlymoreoptimisticthan Republicansafterwinningthe2012election,butbecomepessimisticafterlosingin2016. 22Afterthe2008PresidentialelectionwhenBarackObamaiselectedasthefirstBlackpresidentofthe UnitedStates,Blackhouseholdsbecomesignificantlymoreoptimistic,evenexceedingtheproportionof whitehouseholdswhoareoptimistic.Incontrast,afterthe2016electionwhichbroughtDonaldTrumpto power,thereverseoccursandBlackhouseholdsbecomemorepessimisticthanWhitehouseholds. 43

(a)Education (b)JobStatus (c)Age (d)Gender (e)Race (f)PartyAffiliation Figure8: HeterogeneityinHouseholdExpectations 44

A.1.2 MajorEvents2008-2017 Table8: ChangeinExpectationsIndexaroundMajorEvents Date Event E_{t+1}-E_{t-1} 15Sep2008 LehmanBankruptcy -0.22 4Nov2008 USElection2008 0.27 25Nov2008 QuantitativeEasing -0.03 23Mar2010 AffordableCareAct -0.06 9Aug2011 ForwardGuidance 0.04 6Nov2012 USElection2012 0.11 1-17Oct2013 CongressShutdown -0.13 Nov2016 USElection2016 0.05 Notes:Thistablesummarizesthechangesinhouseholdexpectationsaroundsomemajorevents duringthesampleperiod2008-17. 45

A.1.3 DetailedRegressionTablesforGDTP Table9a: Baseline:SophisticatedHouseholds PanelA:SophisticatedHouseholds Day0 Day1 Day2 Day3 Day4 Day5 Surprise(Unemp) -0.568∗ -0.565∗ -0.642∗ -0.377 -0.170 -0.578∗ (0.298) (0.316) (0.341) (0.299) (0.300) (0.313) Observations 33363 32657 30264 32826 33950 31544 R 2 0.051 0.049 0.045 0.052 0.049 0.048 Surprise(CPI) 0.434 -0.168 0.391 -0.123 -0.419 0.111 (0.290) (0.306) (0.308) (0.282) (0.288) (0.295) Observations 32697 31409 31836 33342 33528 32386 R 2 0.049 0.047 0.051 0.051 0.055 0.055 Surprise(GDP) 0.168 0.427 0.252 0.690∗∗ -0.0763 0.189 (0.324) (0.325) (0.324) (0.323) (0.341) (0.327) Observations 32112 31029 29208 30264 29008 30338 R 2 0.056 0.051 0.049 0.048 0.045 0.057 Surprise(Housing) 0.0147 -0.107 0.115 0.650∗∗ -0.216 0.749∗∗ (0.321) (0.331) (0.331) (0.328) (0.343) (0.337) Observations 34746 32693 33347 33745 32740 31851 R 2 0.053 0.044 0.049 0.045 0.054 0.051 Controls Yes Yes Yes Yes Yes Yes Notes:Thistablereportsestimatesofβ fromEquation4forsophisticatedexpectationsmodel. h Hereeachcellisthecoefficientfromaseparateregressionequation.Wecontrolfordemographics suchasage,income,education,race,gender,politicalaffiliation,andstateofresidenceofthe respondent.Surveyweightsareused.Standarderrorsarepresentedinparentheses.*p<0.10,** p<0.05,***p<0.01.SurprisesfromBloombergFinanceLP. 46

Table9b: Baseline:NaiveHouseholds PanelB:NaiveHouseholds Day0 Day1 Day2 Day3 Day4 Day5 ∆(Unemp) -0.790∗∗∗ -0.640∗∗ -0.951∗∗∗ -0.392 -0.184 -0.591∗ (0.300) (0.315) (0.329) (0.305) (0.303) (0.307) Observations 32657 31955 29556 32156 33276 30858 R 2 0.053 0.051 0.047 0.053 0.050 0.050 ∆(CPI) 0.0657 -0.0737 0.663∗∗ 0.308 0.666∗∗ 0.157 (0.288) (0.293) (0.294) (0.289) (0.294) (0.296) Observations 32003 30699 31113 32589 32834 32386 R 2 0.049 0.048 0.053 0.052 0.056 0.055 ∆(GDP) 0.787∗∗ 0.817∗∗ 0.625∗ 0.963∗∗∗ 0.397 0.204 (0.343) (0.345) (0.351) (0.344) (0.352) (0.356) Observations 31690 30593 28761 29814 28569 29890 R 2 0.058 0.052 0.050 0.049 0.046 0.058 ∆(Housing) -0.484 -0.339 -0.167 0.571∗ -0.131 0.378 (0.337) (0.352) (0.350) (0.343) (0.344) (0.349) Observations 33937 31857 32493 32937 32617 31379 R 2 0.054 0.045 0.050 0.046 0.053 0.051 Controls Yes Yes Yes Yes Yes Yes Notes:Thistablereportsestimatesofβ fromEquation4forthenaiveexpectationsmodel.Here h eachcellisthecoefficientfromaseparateregressionequation. Wecontrolfordemographics suchasage,income,education,race,gender,politicalaffiliation,andstateofresidenceofthe respondent.Surveyweightsareused.Standarderrorsarepresentedinparentheses.*p<0.10,** p<0.05,***p<0.01.SurprisesfromBloombergFinanceLP. 47

A.1.4 Gallupweeklyresults Table10: GDTPWeeklyEstimates X = X =CPI X =GDP X =Housing t t t t 1 2 3 4 PanelA:SophisticatedHouseholds SurpX -3*** -2*** 0.07 -0.03 t (0.12) (0.12) (0.13) (0.13) N 222644 227598 215142 229071 R 2 0.047 0.046 0.043 0.045 Controls Yes Yes Yes Yes PanelB:NaiveHouseholds ChangeX -4.5*** -0.3** 1.8 2*** t (0.12) (0.12) (0.14) (0.14) N 217823 223661 212041 224846 R 2 0.053 0.045 0.045 0.047 Controls Yes Yes Yes Yes Thistableshowstheestimatesfromaweeklywindowaroundeach announcementintheGDTP,equivalenttotheweeklywindowsin theMSC.Standarderrorsarepresentedinparentheses.*p<0.10,** p<0.05,***p<0.01.SurprisesfromBloombergFinanceLP. 48

A.1.5 Takingt–1expectationsforallannouncementsastheexpectationsthedaybeforethefirstannouncementinthemonth,i.e.theunemploymentannouncement Table11: GDTP:Ei [Z]–E¯first[Z]=α +β first ·ShockX +Di +ϵi t+h t–1 h h t t+h th PanelA:SophisticatedHouseholds Day0 Day1 Day2 Day3 Day4 Day5 Surprise(CPI) 0.501∗ -0.000281 0.426 -0.0285 -0.224 0.384 (0.292) (0.307) (0.309) (0.272) (0.276) (0.283) Surprise(GDP) 0.0357 -0.0171 -0.183 0.350 -0.390 0.148 (0.412) (0.417) (0.363) (0.364) (0.387) (0.603) Surprise(Housing) 0.593∗ 0.341 0.818∗∗ 1.169∗∗∗ 0.209 1.496∗∗∗ (0.321) (0.332) (0.331) (0.328) (0.343) (0.337) PanelB:NaiveHouseholds Day0 Day1 Day2 Day3 Day4 Day5 ∆(CPI) -0.163 -0.121 0.507∗ 0.150 0.482∗ 0.0631 (0.289) (0.293) (0.294) (0.288) (0.291) (0.293) ∆(GDP) 1.030∗ 0.868 0.785∗ 0.961∗∗ 0.737 0.132 (0.566) (0.563) (0.456) (0.453) (0.491) (0.710) ∆(Housing) 0.345 0.337 0.685∗ 1.357∗∗∗ 0.637∗ 1.338∗∗∗ (0.337) (0.353) (0.351) (0.342) (0.344) (0.349) ThistableshowstheestimatesfromcomputingthechangeinexpectationsasasEi [Z]– t+h E¯first[Z],whereE¯first[Z]istheaverageexpectationsbeforetheemploymentsituationrelease t–1 t–1 eachmonth.Thisexerciseallowsustocalculatechangesinexpectationaroundeachrelease relativetotheexpectationssetbeforethefirstannouncementofthemonth.Standarderrors arepresentedinparentheses.*p<0.10,**p<0.05,***p<0.01.SurprisesfromBloomberg FinanceLP. 49

A.1.6 Takingt–3expectationsfortheunemploymentannouncementinsteadoft–1 Table12 PanelA:SophisticatedHouseholds Day0 Day1 Day2 Day3 Day4 Day5 Surprise(Unemp) -0.00519∗ -0.00600∗ -0.00683∗ -0.00588∗ -0.000968 -0.00425 (0.003) (0.003) (0.004) (0.003) (0.003) (0.003) Observations 30407 29659 27281 29719 30921 29203 R 2 0.052 0.049 0.044 0.054 0.048 0.047 PanelB:NaiveHouseholds Day0 Day1 Day2 Day3 Day4 Day5 ∆(Unemp) -0.00480 -0.00381 -0.00701∗∗ -0.00296 0.000810 -0.00153 t (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) Observations 30407 29659 27281 29719 30921 29203 R 2 0.052 0.049 0.044 0.054 0.048 0.047 ThistablereportsestimatesforEquation4takingthepriorexpectationtobe3daysbeforetheannouncement,insteadofonedaybefore.Standarderrorsarepresentedinparentheses.*p<0.10,** p<0.05,***p<0.01.SurprisesfromBloombergFinanceLP. 50

A.1.7 Localunemployment We now test whether people change their expectations differently depending on their local economic conditions. Both personal as well as local conditions can influence an individual’sexpectations.23 Peoplelivinginareaswithtraditionallyhigherunemployment couldbemoresensitivetomovementsintheunemploymentrate. Itcouldalsobethat whenunemploymentincreases,theshockisgreatesttopeopleinareaswithtraditionally lowerunemployment,sotheyrespondmore. Weexaminethesehypothesesempirically byestimating: E t i +h [Z]–E¯ t–1 [Z] = α 1h +β 1h ×(ShockX t |LocalU t > median(LocalU t ))+D i 1t+h +ϵ 1th (8) E t i +h [Z]–E¯ t–1 [Z] = α 2h ++β 2h ×(ShockX t |LocalU t > p75(LocalU t ))+D i 2t+h +ϵ 2th (9) wheret isthedayoftheannouncement,hindicatesdaysfromt,Eτ indicatesexpectations formedbyagentiondayτ,Di denotesdemographicinformationforpersoni,ShockX t+h t denotes the shock in information due to the announcement, LocalU denotes the local t unemploymentrateofthefipscodethatagentilivesin. Weclusterstandarderrorsbystate. Wefindthemedianandthe75thpercentilelocalunemploymentrateforallfipscodeevery month,andsplitareasaccordingtothatvalue. Wefindthatpeoplelivinginareaswith highlocalunemploymentpaymoreattentiontoshockstothenationalunemployment rate. Thisresult,however,doesnotholdforshockstoCPI. 23Borgschulte&Martorell(2018)usedataonmilitarypersonnelrecordsandtheyfindthatservicemembers wouldforgo1.5%inreenlistmentearningstoavoida1percentagepointincreaseinlocalunemployment rate. 51

Table13: ResponseofHouseholdExpectationstoUnemploymentShocksDependingonLocal AreaUnemployment PanelA:SophisticatedHouseholds Day0 Day1 Day2 Day3 Day4 Day5 Surp(U) ,High50 -0.678∗∗ -0.630∗∗ -0.649 -0.357 -0.201 -0.691∗∗ t (0.259) (0.260) (0.421) (0.329) (0.323) (0.287) Surp(U) ,Low50 0.0875 -0.637 -0.565 -0.00299 -0.0558 0.668 t (0.909) (0.583) (0.899) (1.002) (0.851) (0.740) Surp(U) ,High75 -0.650∗∗ -0.633∗∗ -0.516 -0.313 -0.327 -0.675∗∗ t (0.308) (0.271) (0.458) (0.385) (0.329) (0.268) Surp(U) ,Low75 -0.104 -0.610 -1.237 -0.690 0.577 -0.0739 t (0.912) (0.455) (0.708) (0.785) (0.808) (0.650) PanelB:NaiveHouseholds Day0 Day1 Day2 Day3 Day4 Day5 ∆(Unemp) ,High50 -0.753∗∗∗ -0.682∗∗ -0.893∗∗ -0.364 -0.135 -0.602∗∗ t (0.278) (0.284) (0.388) (0.349) (0.399) (0.297) ∆(Unemp) ,Low50 -0.721 -0.708 -1.420 -0.363 -0.669 -0.168 t (0.676) (0.732) (0.970) (0.856) (0.810) (0.797) ∆(Unemp) ,High75 -0.715∗∗ -0.742∗∗ -0.759∗ -0.255 -0.255 -0.623∗∗ t (0.316) (0.320) (0.411) (0.384) (0.428) (0.306) ∆(Unemp) ,Low75 -0.805 -0.562 -1.894∗∗ -0.953 0.107 -0.489 t (0.678) (0.459) (0.704) (0.678) (0.738) (0.638) Thistablereportstheestimatesofβ fromEquation8forshockstounemployment.Here,β h h ischangeintheexpectationsduetoashockintheunemploymentrateintheBLSjobsreport interactedwiththestate’sunemploymentrate,inthewindow[t–1,t+h]wheretisthedayof theannouncementandh=0,1,2,3,4.Inbothpanels,rows1and2indicateareaswithhighand lowlocalunemploymentdependingonmediancountylevelunemployment,androws2and4 indicateareaswithhighandlowlocalunemploymentdependingon75%percentilecountylevel unemployment.Standarderrorsarepresentedinparentheses.*p<0.10,**p<0.05,***p<0.01. SurprisesfromBloombergFinanceLP. 52

Table14: ResponseofHouseholdExpectationstoCPIShocksDependingonLocalAreaUnemployment PanelA:SophisticatedHouseholds Day0 Day1 Day2 Day3 Day4 Day5 Surp(CPI) ,High50 0.578∗ -0.0716 0.424 0.00657 -0.264 0.213 t (0.316) (0.333) (0.335) (0.307) (0.313) (0.323) Surp(CPI) ,Low50 -0.395 -0.587 0.451 -0.958 -1.169 -0.375 t (0.750) (0.803) (0.829) (0.748) (0.773) (0.756) Surp(CPI) ,High75 0.441 0.0685 0.369 0.110 -0.0578 0.267 t (0.329) (0.345) (0.348) (0.318) (0.327) (0.337) Surp(CPI) ,Low75 0.242 -0.937 0.224 -0.808 -1.505∗∗ -0.282 t (0.615) (0.662) (0.667) (0.615) (0.608) (0.620) PanelB:NaiveHouseholds Day0 Day1 Day2 Day3 Day4 Day5 ∆(CPI) ,High50 0.120 -0.0258 0.535∗∗ 0.245 0.535 0.432 t (0.326) (0.327) (0.261) (0.222) (0.348) (0.349) ∆(CPI) ,Low50 -0.0491 0.116 1.720 0.595 1.264∗ -1.238 t (0.678) (0.886) (1.031) (0.596) (0.701) (0.898) ∆(CPI) ,High75 -0.0687 0.190 0.534∗ 0.331 0.707∗ 0.406 t (0.353) (0.266) (0.285) (0.244) (0.362) (0.338) ∆(CPI) ,Low75 0.690 -0.938 1.220∗∗ 0.405 0.638 -0.465 t (0.489) (0.780) (0.564) (0.529) (0.707) (0.903) This table reports the estimates of β from Equation 8 for shocks to CPI. Here, β is h h changeintheexpectationsduetoashockintheunemploymentrateintheBLSjobsreport interactedwiththestate’sunemploymentrate,inthewindow[t–1,t+h]wheretistheday oftheannouncementandh=0,1,2,3,4.Inbothpanels,rows1and2indicateareaswith highandlowlocalunemploymentdependingonmediancountylevelunemployment, androws2and4indicateareaswithhighandlowlocalunemploymentdependingon75% percentilecountylevelunemployment.Standarderrorsarepresentedinparentheses.* p<0.10,**p<0.05,***p<0.01.SurprisesfromBloombergFinanceLP. 53

A.1.8 Prevspost2012 Table15: ResponseofExpectationsinRecessionversusNon-recessionYears PanelA:SophisticatedHouseholds Day0 Day1 Day2 Day3 Day4 Day5 Surp(U) ,Pre2012 -0.00269 -0.00893∗∗ -0.0104∗∗∗ -0.00570 -0.00324 -0.00762∗∗ t (0.003) (0.004) (0.004) (0.004) (0.004) (0.004) Surp(U) ,Post2012 -0.0111∗ 0.000518 -0.000430 -0.00192 0.00201 0.00209 t (0.006) (0.007) (0.007) (0.006) (0.006) (0.006) Surp(CPI) ,Pre2012 0.00280 -0.00192 0.00126 -0.00285 -0.00943∗∗∗ -0.00230 t (0.003) (0.004) (0.003) (0.003) (0.003) (0.003) Surp(CPI) ,Post2012 0.00961∗ 0.000798 0.00988 0.00115 0.00491 0.00930∗ t (0.006) (0.006) (0.006) (0.006) (0.005) (0.006) PanelB:NaiveHouseholds Day0 Day1 Day2 Day3 Day4 Day5 ∆(Unemp) ,Pre2012 -0.00697∗∗ -0.0109∗∗∗ -0.0153∗∗∗ -0.00776∗∗ -0.00526 -0.00587 t (0.003) (0.004) (0.004) (0.004) (0.004) (0.004) ∆(Unemp) ,Post2012 -0.0125∗ 0.00137 -0.000297 -0.00339 0.00801 -0.00122 t (0.007) (0.008) (0.008) (0.007) (0.007) (0.008) ∆(CPI) ,Pre2012 -0.00191 -0.00259 0.00521 0.00470 0.00536 0.00394 t (0.003) (0.003) (0.003) (0.003) (0.004) (0.004) ∆(CPI) ,Post2012 0.00445 0.00180 0.00858 -0.000261 0.00922∗ -0.00203 t (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) ThistablereportsestimatesfortwosubsamplesforGDTP-theperiodoftheGreatRecession(2008-2011),andthe non-recessionperiod(2012-2017).Standarderrorsarepresentedinparentheses.*p<0.10,**p<0.05,***p<0.01. SurprisesfromBloombergFinanceLP. 54

A.2 Michigan A.2.1 IndexofConsumerExpectations(ICE) ICEisacompositeindexofthreeforwardlookingsurveyquestions: 1. Nowlookingahead–doyouthinkthatayearfromnowyou(andyourfamilylivingthere) willbebetterofffinancially,orworseoff,orjustaboutthesameasnow? 2. Nowturningtobusinessconditionsinthecountryasawhole–doyouthinkthatduringthe nexttwelvemonthswe’llhavegoodtimesfinancially,orbadtimes,orwhat? 3. Lookingahead,whichwouldyousayismorelikely–thatinthecountryasawholewe’ll havecontinuousgoodtimesduringthenextfiveyearsorso,orthatwewillhaveperiodsof widespreadunemploymentordepression,orwhat? MSCcalculatesICEinthefollowingmanner:firstcomputestherelativescores(thepercent givingfavorablerepliesminusthepercentgivingunfavorablereplies,plus100)foreach of the three index questions. Each relative score is then rounded to the nearest whole X1+X2+X3 number. Then, {ICE = +2.0} where, the relative scores are divided by the 4.1134 1966baseperiodtotaltheaddedconstantistocorrectforsampledesignchangesfromthe 1950s. Thesethreequestionstakentogetherprovideameasureofhousehold’sexpectationsabout thefutureoftheeconomy,makingitqualitativelysimilartoGallup’sExpectationIndex. ChangesinICEcanalsobeinterpretedinasimilarway-anincreaseinICEdenotesarise inoptimism,whereasadecreasedenotesafallinoptimismorariseinpessimism. 55

Table16: MSCScenariosDependentonLevelofU andπ t t y =ICE X =U X =CPI t t t (1) (2) PanelA:SophisticatedHouseholds1997-2019 Baseline SurpX -0.6*** -1.2*** t (0.13) (0.2) Asymmetry SurpX > 0 -10.4*** -4.7*** t (0.47) (0.59) SurpX < 0 -0.5* 0.1 t (0.30) (0.62) Scenarios ∆U > 0,∆π > 0 -3.8*** -0.6 (0.45) (0.44) ∆U < 0,∆π < 0 2.6*** -2.6*** (0.28) (0.40) ∆U > 0,∆π < 0 -4.6*** -0.5 (0.48) (0.47) ∆U < 0,∆π > 0 3*** -2.4*** (0.46) (0.4) PanelB:NaiveHouseholds1980-2019 Baseline ChangeX -3.3*** -1.2*** t (0.09) (0.17) Asymmetry ChangeX > 0 -5.8*** -4.4*** t (0.17) (0.29) ChangeX < 0 -1*** 2.8*** t (0.23) (0.20) Scenarios ∆U > 0,∆π > 0 -13*** 0.02 (0.30) (0.26) ∆U < 0,∆π < 0 0.9*** -2.1*** (0.25) (0.40) ∆U > 0,∆π < 0 -7.1** -2.2*** (0.38) (0.33) ∆U < 0,∆π > 0 -3.6*** -3.2*** (0.45) (0.49) Thistableshowstheestimatesscenariosdependenton levelsofinflationandunemployment. Wedefinehigh unemploymenttobegreaterthan5%andhighinflation tobegreaterthan3%.*p<0.10,**p<0.05,***p<0.01. DatafromSurveyofConsumers,UniversityofMichigan. SurveyResearchCenter. SurprisesfromBloombergFinanceLP. 56

A.2.2 Levelcategorytables Table17: MSCScenariosDependentonLevelofU andπ t t y =BusinessOutlook X =U X =CPI t t t 1 2 PanelA:SophisticatedHouseholds HighU ,Highπ -2*** -2.8* t t (0.80) (0.98) LowU ,Lowπ -2.7*** -0.9 t t (0.70) (0.96) HighU ,Lowπ -1.6*** 0.9 t t (0.50) (0.74) LowU ,Highπ 0.08 -3.6*** t t (1.10) (1.08) PanelB:NaiveHouseholds HighU ,Highπ -5.7*** -7*** t t (0.43) (0.65) LowU ,Lowπ -1.2 -2.5*** t t (0.78) (0.84) HighU ,Lowπ -5*** 1.2*** t t (0.38) (0.45) LowU ,Highπ -5.4*** -2.7* t t (1.51) (1.55) Thistableshowstheestimatesforscenariosdependent on levels of inflation and unemployment. We define high unemployment to be greaterthan5%andhighinflationtobegreater than 3%. * p < 0.10, ** p < 0.05, *** p < 0.01. DatafromSurveyofConsumers,Universityof Michigan.SurveyResearchCenter. 57

A.2.3 MSC:ResponseofExpectationsduringUSRecessions Inthissectionweestimate E t i +h [Z]–E¯ t–1 [Z] = α h +β 1h ×ShockX t +β 2h ×1(Recession) +β 3h ×(ShockX t ×1(Recession))+D i t+h +ϵ th Table18: ShareofOptimistsinRecessions y =ShareofOptimists X =U X =CPI t t t 1 2 PanelA:SophisticatedHouseholds Surprise(X ) 1*** -1.6* t (0.36) (0.46) RecessionYear -23*** -24*** (1.13) (1.33) Surprise(X )×RecessionYear -3.5*** 1.2 t (0.86) (0.92) PanelB:NaiveHouseholds ∆X -1*** -0.8* t (0.41) (0.48) RecessionYear -21*** -24*** (1.40) (1.32) ∆X ×RecessionYear -1 0.3 t (0.83) (0.87) This table shows the estimates for change in share of optimistsduringUSRecessions. *p < 0.10,**p < 0.05, ***p<0.01.DatafromSurveyofConsumers,University of Michigan. Survey Research Center. Surprises from BloombergFinanceLP. 58

Table19: ResponseofHouseholdInflationExpectationsinRecessions y t = E t π t+12 X t =U X t =CPI 1 2 PanelA:SophisticatedHouseholds Surprise(X ) 0.003 0.05 t (0.025) (0.04) RecessionYear 1.2*** -1.6*** (0.10) (0.14) Surprise(X )×RecessionYear 0.18** -0.05 t (0.08) (0.09) PanelB:NaiveHouseholds ∆X 0.06** 0.03 t (0.03) (0.04) RecessionYear 1.1*** 1.6*** (0.13) (0.14) ∆X ×RecessionYear 0.03 0.1 t (0.08) (0.09) Thistableshowstheestimatesforchangeinhousehold inflationexpectationsduringUSRecessions.*p<0.10,** p<0.05,***p<0.01.DatafromSurveyofConsumers,UniversityofMichigan.SurveyResearchCenter.Surprises fromBloombergFinanceLP. 59

A.2.4 Michigandailyresults Table20a: MSCDailyEstimates PanelA:SophisticatedHouseholds1997-2019 Day0 Day1 Day2 Day3 Day4 Day5 Surprise(Unemp) 0.438 0.189 0.273 -0.257 -0.459 0.700 (0.860) (0.858) (0.862) (0.766) (0.806) (0.822) N 3208 3711 3302 4470 4295 3918 R 2 0.027 0.022 0.033 0.019 0.023 0.024 Surprise(CPI) -0.637 2.361∗∗ -2.049∗ 4.313∗∗∗ -2.255∗∗ 1.651 (1.048) (1.076) (1.184) (1.093) (0.999) (1.037) N 2284 2233 1941 2146 2280 2352 R 2 0.028 0.019 0.023 0.030 0.031 0.020 PanelB:NaiveHouseholds1980-2019 Day0 Day1 Day2 Day3 Day4 Day5 ∆(Unemp) -0.0389 0.373 0.711 -0.166 -1.288∗∗ 1.145∗ (0.635) (0.654) (0.647) (0.563) (0.604) (0.613) N 6357 6845 6315 8557 8468 7936 R 2 0.024 0.022 0.026 0.020 0.023 0.018 ∆(CPI) -1.380∗ 1.925∗∗ -0.663 1.571∗∗ -1.248∗ 2.589∗∗∗ (0.746) (0.773) (0.806) (0.707) (0.715) (0.843) N 4561 4228 3715 4026 4238 4408 R 2 0.026 0.016 0.023 0.019 0.025 0.024 ThistableshowstheestimatesfromadailywindowaroundeachannouncementintheMSC forthefullsample: 1980-2019. Standarderrorsarepresentedinparentheses. *p < 0.10, **p < 0.05,***p < 0.01. DatafromSurveyofConsumers,UniversityofMichigan. Survey ResearchCenter.SurprisesfromBloombergFinanceLP. 60

A.2.5 Michigandailyresults-gallupsubsample Table20b: MSCDailyEstimates:2008to2017 PanelA:SophisticatedHouseholds Day0 Day1 Day2 Day3 Day4 Day5 Surprise(Unemp) -1.476 1.204 -0.756 0.997 -2.839∗∗∗ 3.499∗∗∗ (0.964) (0.981) (1.038) (0.868) (0.875) (1.116) N 1459 1479 1262 1861 1773 1603 R 2 0.026 0.040 0.058 0.016 0.026 0.045 Surprise(CPI) -3.625∗∗ 2.620 -0.370 3.196∗ -4.093∗∗∗ 4.153∗∗∗ (1.649) (1.802) (1.869) (1.811) (1.444) (1.579) N 986 925 891 973 1020 1035 R 2 0.042 0.034 0.050 0.030 0.042 0.049 PanelB:NaiveHouseholds Day0 Day1 Day2 Day3 Day4 Day5 ∆(Unemp) -0.162 3.109∗∗∗ 0.0219 -1.894∗ -0.437 0.162 (1.015) (1.099) (1.153) (0.975) (1.033) (1.081) N 1681 1698 1434 2103 2040 1832 R 2 0.027 0.036 0.045 0.016 0.019 0.037 ∆(CPI) -2.212∗∗ 1.535 -0.528 0.638 -2.691∗∗∗ 3.519∗∗∗ (1.008) (1.066) (1.094) (0.979) (0.924) (1.194) N 1132 1086 1034 1095 1111 1132 R 2 0.031 0.024 0.056 0.022 0.034 0.037 ThistableshowstheestimatesfromadailywindowaroundeachannouncementintheMSC, equivalenttothedailywindowintheGDTPfor2008-17. Standarderrorsarepresentedin parentheses.*p<0.10,**p<0.05,***p<0.01.DatafromSurveyofConsumers,Universityof Michigan.SurveyResearchCenter.SurprisesfromBloombergFinanceLP. 61