Availableonlineatwww.sciencedirect.com
ScienceDirect
HOSTED BY
EconomiA18(2017)40–59
Inattention
in
individual
expectations
夽
Yara
de
Almeida
Campos
Cordeiro
a,
Wagner
Piazza
Gaglianone
b,
João
Victor
Issler
a,∗aGraduateSchoolofEconomics–EPGE,GetulioVargasFoundation,Brazil bResearchDepartment,BancoCentraldoBrasil,Brazil
Received3March2016;accepted9December2016
Availableonline19January2017
Abstract
Thispaperinvestigatestheexpectationsformationprocessofeconomicagentsaboutinflationrate.UsingtheMarket Expec-tationsSystem of CentralBank of Brazil, weperceive that agents do not update theirforecasts every period and that even agentswhoupdatedisagreeintheirpredictions.Wethenfocuson thetwomostpopulartypesofinattentionmodelsthathave been discussedinthe recent literature: sticky-informationand noisy-informationmodels. Estimating ahybrid modelwefind that,although formallyfitting theBrazilian data,it happensatthe costofamuchhigher degreeofinformationrigiditythan observed.
© 2017 The Authors. Production and hosting by Elsevier B.V. on behalf of National Association of Postgraduate Cen-ters in Economics, ANPEC. This is an open access article under the CC BY-NC-ND license (http://creativecommons. org/licenses/by-nc-nd/4.0/).
JELclassification: C14;C52;D84;E37
Keywords:Expectations;Inflation;Imperfectinformation;Rationalinattention
Resumo
Este artigo investigaoprocesso deformac¸ãodeexpectativasde inflac¸ãodeagentes econômicos.UtilizandooSistemade ExpectativasdeMercadodo BancoCentraldo Brasil, percebemosque osagentes nãoatualizam suasprojec¸õesem todosos períodosemesmoaquelesagentesqueofazemdiscordamsobreosvaloresprevistos.Nestesentido,investigamososdoistiposde modelosmaispopularessobreinatenc¸ãodiscutidosnaliteraturarecente:informac¸ãocomrigidezeinformac¸ãocomruído.Combase
夽 WeareverythankfultoValdemarRodriguesdePinhoNeto,MarcoAntonioCésarBonomoandCecíliaMachadoBerrielforthehelpful
suggestionsandcomments,totheInvestorRelationsandSpecialStudiesDepartment(Gerin)oftheBancoCentraldoBrasilforkindlyproviding
thedatausedinthispaperandtoPhilippeAndrade,whoprovidedhisGAUSScodestous.WethankalsotheexcellentresearchassistanceofMarcia
WaleriaMachadoandAndreaV.Machado.Theopinionsexpressedinthisarticlearethoseoftheauthorsanddonotnecessarilyreflectthoseofthe
BancoCentraldoBrasil.Anyremainingerrorsareours.
∗Correspondingauthorat:GraduateSchoolofEconomics,GetulioVargasFoundation,PraiadeBotafogo190,s.1104,RiodeJaneiro,RJ
22.253-900,Brazil.
E-mailaddresses:cordeiro.ya@gmail.com(Y.d.A.C.Cordeiro),wagner.gaglianone@bcb.gov.br(W.P.Gaglianone),
joao.issler@fgv.br(J.V.Issler).
PeerreviewunderresponsibilityofNationalAssociationofPostgraduateCentersinEconomics,ANPEC.
http://dx.doi.org/10.1016/j.econ.2016.12.002
1517-7580©2017TheAuthors.ProductionandhostingbyElsevierB.V.onbehalfofNationalAssociationofPostgraduateCentersinEconomics,
Y.d.A.C.Cordeiroetal./EconomiA18(2017)40–59 41
naestimac¸ãodeummodelohíbrido,concluímosque,emboraformalmenteomodelosejacapazdeseajustaraosdadosbrasileiros, talresultadoocorreaocustodeumgraumuitomaiorderigidezinformacionaldoqueoobservado.
© 2017 The Authors. Production and hosting by Elsevier B.V. on behalf of National Association of Postgraduate Cen-ters in Economics, ANPEC. This is an open access article under the CC BY-NC-ND license (http://creativecommons. org/licenses/by-nc-nd/4.0/).
Palavras-chave: Expectativas;Inflac¸ão;Informac¸ãoimperfeita;Inatenc¸ãoracional
1. Introduction
Theexpectationsformationprocessofeconomicagentsaboutmacroeconomicvariableshaslongbeenoneofthe mostdebatedquestionsinmacroeconomics.Neverthelessremainsanopenquestionhowexpectationsareformed,and howbesttomodelit.Inmuchclassicaltheory,thereisnoroomfordisagreementinexpectations,sinceitisusually assumed thatallagentsformexpectations conditionalonacommoninformationset.Howeverifnoteveryone has thesameexpectationsandtheinformationfrictionsarelargeandeconomicallysignificant,thedegreeofinformation rigiditymayhavesignificantimplicationsformacroeconomicdynamicsandoptimalpolicy.
What we aim todo is relatedto the recentempirical work tryingto determinethe nature of the expectations formationprocess.Rationalexpectationsmodels withinformationfrictionssuchas MankiwandReis(2002),Reis (2006a,b),Sims(2003)andWoodford(2003)havebeenassociatedtoagents’inattentiontonewinformation,dueto costsofcollectingandprocessinginformation.Thesemodelshavethekeyadvantageofparsimoniouslyexplaining somepatternsofindividualexpectationsobservedinthedata–suchasdisagreementacrossforecastersandpredictable forecasterrors–thatareconflictingwiththestandardhypothesisofperfectinformation.
Thesticky-informationmodelsproposedbyMankiwandReis(2002)andReis(2006a,b)arebasedontheassumption thattheagentsdonothaveaccesstoinformationinstantly.InMankiwandReis(2002),forinstance,itisassumedthatthe acquisitionofinformationfollowsaPoissonprocessinwhich,ateachdate,agentsfaceagivenandconstantprobability
λof beingabletogetnewinformation. Nevertheless,onceagentsupdatetheirinformationset,theyobtainperfect informationandformexpectations rationally.Thuswe refertoλastheattentiondegreeforthe sticky-information modeland(1−λ)canbeseenasthedegreeofinformationrigidity.
Theinfrequentupdateimpliesthat,eachperiod,onlyafractionoftheagentshasaccesstothelatestmacroeconomic news andthe expectationsandactionsof thosewhodidnot updatetheir informationsetscontinue tobe basedon theiroldinformation.Asaresult,agentswhoupdatedtheirinformationsetsinthesameperiodmustmakethesame forecastsandagentswhodidnothaveaccesstonewinformationshouldnotrevisetheirlastprediction.
On theotherhand,inthenoisy-informationmodelsdevelopedbySims(2003)andWoodford (2003),although agentscontinuouslytrackvariablesandupdatetheirinformationset,theyonlyobservenoisysignalsaboutthetrue state.Asagentsknowtheyhaveanimperfectaccesstothenewstheygetateachperiod,theydonotcompletelypassit ontotheirforecast.Moreprecisely,forecastsareaweightedaverageofthenewandthepreviousinformationreceived, sotheweightonpreviousbeliefsistakenasthedegreeofinformationrigidity.
FollowingAndradeandLeBihan(2013),wewillfocusonthesetwomostpopulartypesofinattentionmodelsthat havebeendiscussedintherecentliterature:sticky-informationandnoisy-informationmodels.Themodel–developed bytheseauthors–isahybridone:itassumesthat,ateachdate,everyforecasterfacesagivenprobabilityofbeing abletoupdatehisinformationsetandthat,whenupdating,hegetsanoisyperceptionofthestateoftheeconomy.The modelisthenestimatedbyaMinimumDistanceEstimation(MDE)procedure,whichallowsustotestifthemodelis capableofquantitativelyfittingthedata,particularlytheforecasterrorsandthedisagreementamongforecasters.
Asfarasweknow,weareoneofthefirstauthorstouseexpectationsdataoninflationinBrazilinordertotryto modelitbasedoninattentionmodels.1Astheresponseofagentstomacroeconomicdynamicsisstronglyimpacted bythewayindividualexpectationsareformed,modelingtheexpectationsformationprocessisimportantforbetter conductof economicpolicyandbetterunderstandingitsimplications.AnadvantageoverAndrade andLeBihan’s
1 Guillén(2008)investigatesasetofexpectationstheoriesusingBraziliandataandconcludesthatthemedianinflationforecastismorelikelyto
(2013)studyisthatforecastsonourdatabaseareobservedinadailyfrequency–theirsareonaquarterlyfrequency. Thismeansweareabletofollowthesequenceofforecastsmadebyeachagentmoreclosely.
Theremainderofthispaperisstructuredasfollows.InSection2wemakeabriefliteraturereviewandinSection3we presentsomebasicfactsaboutforecasterrorsanddisagreementbetweenforecasters.Section4presentsthemethodology weapplyandinSection5weintroducethedatabasewewilluse,namelytheMarketExpectationsSystemofCentral BankofBrazil.Finally,Section6presentstheresultsofourestimationandSection7concludes.
2. Literaturereview
Thesticky-informationPhillipscurveproposedbyMankiwandReis(2002)wasdevelopedinordertoobtainan alternativetothe newKeynesian Phillipscurve– asticky-pricemodel – thatfailedinexplainingsomeaspects of macroeconomicfluctuations.2Inthismodel,ineachperiod,ashareλofthefirmsupdatetheirinformationset3and computeanewpathofoptimalprices.Theonesthatdidnotreceivenewinformationcontinuetosetpricesbasedon oldplans.Afirm’soptimalpriceisgivenby(allvariablesexpressedinlogs):
p∗t =pt+αyt
wherept isoverallpricelevelandytoutputgap.Afirmthatupdateditsinformationsetjperiodsagosetstheprice
Et−j[p∗t]andaggregatepricelevelistheaverageofthepricessetbyallfirms:
pt=λ
∞
j=0
(1−λ)jEt−j(pt+αyt)
Thusinflationratecanberepresentedby:
πt = αλ 1−λ yt+λ ∞ j=0 (1−λ)jEt−1−j(πt+αyt).
Definingmoneysupply(oraggregatedemand)bym=p+yandtakingittobeexogenous,theyexaminehowoutput andinflationrespondtovariationofmunderthesticky-informationmodel,comparingittothedynamicpropertiesof thesticky-pricemodelandabackward-lookingmodel.Theresponsesofthesticky-informationmodelseemconsistent withwhathappenswheneconomiesareexposedtotheseshocks.
Inthesameline,Reis(2006b)andReis(2006a)studyrespectivelytheproblemofaproducerandtheconsumption decisionsofagentswhofacecostsofacquiringand/orinterpretingnewinformationandtrytounderstandthedynamic responseofpricestoshocksinordertoexplaininflationdynamic.In thissetup,the agentsrationallydecidetobe inattentivetonewinformationandchoosetheoptimallengthofinattentiveness.Butoncetheypayattentiontonew information,theybecomeawareofeverythingthatisrelevant.
Sims(2003),inturn, assumesthat people havelimited information-processing capacity– instead of assuming thatagentsupdatetheir informationsetonlysporadically.Here thelimitedcapacityisrepresentedbythefact that informationthatagentshaveaccessmaybecontaminatedwitharandomnoise.Moreprecisely,informationisthought ofas movingthrough achannel inwhichoneentersinputdata,andoutputdataemerges,possiblywitherror.The authortakesthenatureofthenoiseagentsfaceasexogenousandassumethatnoisechangessystematicallyaccording tochangesinthedynamicpropertiesoftheeconomy.UsingtheideaofafiniteShannoncapacity,4heanalysesthe implicationsofincludingtheseinformation-processingconstraintstothedynamicprogrammingproblemusedtomodel behavior.Theimplications–intermsofalteringpartofitsoutcomes–seeminlinewithobservedmacroeconomic behavior.
2 Forexample:thesticky-pricemodelcannotexplainwhyinflationissopersistentorwhyshockstomonetarypolicyhavegradualanddelayed
effectoninflation,andityieldsthatannouncedandcredibledisinflationleadstobooms.
3 Eachfirmhasthesameprobabilityofbeingoneofthefirmsthatreceivenewinformation,regardlessofhowlongithasbeensinceitslastupdate.
4 Ininformationtheory,theShannoncapacityisthemaximumamountofinformationthatcanbetransmittedbyachannelwithouterror.Itdepends
Y.d.A.C.Cordeiroetal./EconomiA18(2017)40–59 43 Following the approach proposed by Sims (2003), Woodford (2003), Moscarini (2004) and Mackowiak and Wiederholt(2009)modelinattentivenessbyprice-settersassumingthatagentshavelimitedcapacitytoacquireand/or processalltheinformationintheirenvironment.InMoscarini(2004),agentsalsochoosenottoupdatetheirinformation systematically.Inthissense,themodelisalsorelatedtoReis’(2006a,b).Buthere,oncetheyupdate,theydonotobtain aperfectsignalonthestateoftheeconomy.
Woodford(2003)assumesthateachprice-setteractstakingintoaccounthisownsubjectiveperceptionofthestate ofaggregatedemand,whichismodeledasanobservationofthetruevaluewithanidiosyncraticerror,andthathe formsoptimalestimatesoftheaggregatestatevariablesgiventhisimperfectinformation.Heassumesthattheestimates areupdatedinrealtimeusingaKalmanfilter.Theprice-setter(correctly)believesthattheeconomy’saggregatestate evolvesaccordingto
¯
Xt =¯c+M ¯Xt−1+mut
where¯candmarevectors,Mamatrix,utisassumedtorepresentamonetarypolicyshockand
¯ Xt = ⎡ ⎢ ⎢ ⎣ Xt ∞ k=1 ξ(1−ξ)k−1X(k)t ⎤ ⎥ ⎥ ⎦
withX(k)t beinghigher-orderexpectations –others’ expectationsof others’expectations...–andξ∈ (0,1)being
ameasureof strategiccomplementarity.∞k=1ξ(1−ξ)k−1X(k)t indicatesthat weightedaverageof expectationsand higher-orderexpectationsalsomatterforthedeterminationofprices,outputandoftheirfutureevolution.
Theonlyinformationreceivedbysupplieriinperiodt(theobservationequation)takestheform
zt(i)=e1X¯t+vt(i)
wherevt(i)isamean-zeroGaussianwhitenoiseerrorterm,distributedindependentlybothofthehistoryoffundamental
disturbancesandoftheobservationerrorsofallothersuppliers.ejrefertojthunitvector.
Thus,i’soptimalestimateofthestatevectorevolvesaccordingto: ¯
Xt|t(i)= ¯Xt|t−1(i)+k[zt(i)−e1X¯t|t−1],
wherekisthevectorofKalmangains.Andthedatetstateperceivedatt−1isgivenby ¯
Xt|t−1(i)=¯c+M ¯Xt−1|t−1(i).
Woodfordanalysestheresponsesofinflationandoutputtoamonetaryshockimpliedbyhismodelandcomparesit tothoseimpliedbythesticky-pricemodel.WhentheeffectofchangesonnominalGDPgrowthpresentconsiderable persistence,thenoisy-informationmodelmatchestheempiricalevidencebetter.
Someauthorshavealsoexploitedsurveyofforecastsofmacroeconomicvariablestoproducemicro-factsthat char-acterizetheformationofexpectationsand,then,triedtoassesswhichmodelbestdescribethebehaviorofforecasters.
Mankiwetal.(2004)investigatewhethertheamplitudeofdisagreementobservedinthedataaboutinflation expecta-tions,aswellasitsevolutionovertime,canbepredictedbythesticky-informationmodel.Theyfindthatthisapproach seemscapableofaccountingformanyaspectsoftheobserveddispersionandcentraltendencyduringtheperiodunder study,butitdoesnotconsistentlygenerateenoughtimevariationindisagreement.
Branch(2007)extendedMankiwandReis’s(2002)sticky-informationmodelallowingthedistributionof informa-tionacrossagentstovaryovertimeandfoundthat thismodelprovidesabetterfittothedistributionofthesurvey datausedthanthebasic/staticversiondoes.5PattonandTimmermann(2010),ontheotherhand,focusontherelation ofthesourceofdisagreementanddifferencesinagents’beliefs.Theyexploresurveydatacross-sectionaldispersion inforecasters’predictionsatseveralforecasthorizonsandfindevidencethatdifferencesinopiniondonotstemfrom differencesininformationsets,butfromheterogeneityinpriors.
5 Inthissense,modelsthatconsidertheinteractionbetweendispersedandstickyinformationalsoprovideanalternativeframeworktothestandard
BothCoibion and Gorodnichenko(2012a) andCoibion andGorodnichenko (2012b) find evidence against the nulloffullinformationconsistentwiththepresenceofinformationrigidities.CoibionandGorodnichenko(2012b)
useaggregatedsurveydatafromUSprofessionalforecasters, consumers,firms, andcentralbankers toexplorethe conditionalresponse of the averageforecast errorandof the disagreementacrossforecasters tovariousstructural shocks.Theyfindevidenceinfavorofthetwomodels.CoibionandGorodnichenko(2012a)proposeanapproachthat doesnotrequiretheidentificationofshocks.Theydocumentevidencethattheeconomicconditionsaffecthowmuch resourcesaredevotedtothecollectionandprocessingofinformation.Thedegreeofinformationrigidityseemstovary acrossmacroeconomicvariablesandthisvariationgoesinthemannerpredictedbynoisy-informationmodels.
AndradeandLeBihan(2013)andAndradeetal.(2013)alsousesurveydatatoanalyzeinformationrigiditymodels fortheeurozoneandtheUnitedStates.Theyfindevidenceonthedatainfavorofimperfectinformationmodelsand focusonstickyandnoisy-informationmodels.Anadvantageofbothstudiesisthattheyrelyonmultivariatemodels, whichtakeintoaccountthedynamicinteractionsacrossvariableswhenagentsformtheirexpectations.Andradeetal. (2013)useaggregatedsurveydatatostudythetermstructureofdisagreementforUSrealoutputgrowth,consumer priceindexinflationandthefederalfundsrate,encounteringthatdisagreementistimevaryingatallhorizons,including theverylongrunandthatthetermstructureofdisagreementdifferssignificantlyacrossvariables.Theirresultsindicate thatbothsticky-andnoisy-informationmodelsareabletocharacterizethedata.
AndradeandLeBihan(2013)exploitthe paneldimensionof theECBSurveyofProfessionalForecasters.The authors developa hybrid sticky-noisy-information model and testthe empirical performanceof the model. They assumethatthe economycanbesummarized byastate vectorZt made ofp lagsof 3variablesXt –yearonyear
inflationrate,changeinunemploymentrateandrealGDPgrowth–withassociatedinnovationt.Itsdynamicsare
describedby:
Zt=FZt−1+ηt,
whereZt =(XtXt−1 ...Xt−p+1),andηt =(t0...0)hasacovariancematrix η.
Letidenoteanindividualforecaster.Ateachdate,everyimayreceivenewrelevantinformationwithprobability
λ,andwhenupdating,agentsobserveanoisysignal(Yit)ofthetruestate(Zt),thatfollows
Yit=HZt+vit, vit∼i.i.d.(0, v)
whereHisassumedtobeequaltotheidentityandwith vadiagonalmatrix.Besides,becauseeveryperiodonlya
shareofthepopulationupdates,wheneveraggregatingtheforecasts–oranyothermeasure–oneshouldweighby
λ∞j=0(1−λ)j,asinMankiwandReis(2002),wherejdenotetheforecastersthatlastupdatedint−j.Theauthors detectthatalthoughthefactsobservedfromthedataqualitativelysupportsbothexpectationmodels,whenestimating them,theycannotquantitativelyreplicatethehighpersistenceandvarianceofforecasterrortogetherwiththelowlevel andtime-varianceofdisagreementobservedinthedata.
3. Whyinattentionmodels?
Twoparticular patternsof individualexpectations observedinthe datathat areinconflictwiththe assumption ofperfectinformationandthatcanbereproducedbymodelsof rationalinattentionarerelatedtothedisagreement amongforecastersandthepredictabilityofforecasterrors.Inthissection,wepresentsomebasicfactsfocusingon bothforecasterrorsanddisagreement.
3.1. Forecasterrors
Thepredictabilityofforecasterrorsisapropertythatarisesfrombothsticky-informationandnoisy-information models.Theinfrequentupdateofthesticky-informationmodelimpliesthat,indatet−1,theexpectationsofthosewho didnotupdatetheirinformationsetscontinuetobebasedontheinformationavailableatt−2.Accordingly,theforecast erroratdatetofagentswhodidnotupdatetheirinformationsetatt−1willbepredictable,basedoninformation availableatt−1.Ontheotherhand,inthenoisy-informationmodel,althoughhavingaccesstonewinformationevery period,agentsknowtheinformationtheygeteachperiodiscontaminatedwithnoise.Thustheydonotcompletely passitontotheirforecastsandforecasterrorwillalsobepredictable.
Y.d.A.C.Cordeiroetal./EconomiA18(2017)40–59 45
Fig.1.Timeseriesofrealizationofmonthlyinflation(purpledotted),averageforecasts(purplesolid)anddifference(blue).
Individualforecasterrorisdefinedas:
eit,T =xT −fit,T
wherexTdenotestherealizedinflationatdateTeventandfit,Tindicatestheforecastmadebyindividualiatdatetfor
thedateTevent.
Andtheaverageforecasterrorisgivenby:
et,T =xT −ft,T
where ft,T =(1/nt)
nt
i=1fit,T denotetheaverageforecastacrossagents,andntisthenumberofinstitutionstaken
intoaccount.
Fig.1exhibitsthetimeseriesofaverageBrazilianmonthlyinflation(asmeasuredbytheIPCA–TheBroadNational ConsumerPriceIndex)forecastspoolingthecurrentandthenext11monthaheadhorizons,usingmonthlyinflation realizationsfromJanuary2002toDecember2014.Onecaneasilyseethat thedifferencebetweenthetwoseries– inflationrealizationsminusforecasts–presentconsiderablypersistence.Periodsofunder-oroverestimationusually lastforsomeconsecutiveperiods,whichleadstopredictableforecasterrors.Also,thefirst-orderautocorrelationof theaverageforecasterror(e(1))is0.6366.
Allthissuggestpastforecasterrorscontaininformationthathasnotbeenexploitedwhenproducingpresentforecasts. Therefore,the predictabilityof forecast errors derivedfrom bothsticky andnoisy-information models– and also obtainedfromthehybridmodelfeaturingbothtypesofinattention–areobservedinourdata.
3.2. Disagreement
Disagreementamongforecastersisdefinedasthecross-sectionstandarddeviationofforecastsateachdate:
σt,T = (1/nt) nt i=1 (fit,T −ft,T)2
Itemergesfrominformationfrictionsandthefactthattheagentsdonotallhaveacommoninformationset,andcan alsobeexplainedbybothsticky-andnoisy-informationmodels.Fig.2exhibitsthetimeseries(fromJanuary2002to
Fig.2.Averagedisagreement.
November2015)ofaveragedisagreementacrossforecasterspoolingthecurrentandthenext11monthaheadhorizons. Noticeably,disagreementisalwaysgreaterthanzero,whichisevidenceinfavorofinformationrigidity.
However,thesetwomodelsmakesomedifferentpredictionsaboutdisagreement.First,sticky-informationmodels, asopposedtothebasicnoisy-information–whichassumesconstantnoiseinthesignalandnoheterogeneitybetween forecasters–admittimevariationindisagreement.Sinceinsticky-informationapproachesonlyafractionoftheagents updatetheirinformationset,thesizeoftheshockshittingtheeconomywilldeterminethevariationofdisagreement overtime.Moreprecisely,thedifferencebetweentheforecastsproducedbytheagentswhoupdatedtheirinformation andtheoneswhodidnotwillbemuchlargerunderlargethanundersmallshocks.Ontheotherhand,inbasic noisy-informationmodels,disagreementisnotaffectedbythemagnitudeoftheshocks,becausethedifferentperceptions ofrealityareresponsibleforthedifferenceinforecasts.Itthusdependsonlyonthevarianceofthenoise,whichis assumedtobeconstantovertime.
AlsoinFig.2,wecanseethat,apartfromthedisagreementbeingalwaysnon-zero,itvariesovertime.Therefore, wecanmakeananalysisoftheresponseofdisagreementtothedynamicsoftheeconomytoseeiftheyarecorrelated. Apossiblemeasureforthedynamicoftheeconomy–ortheshockhittingit–isthesquaredvariationinlastperiod inflation,i.e.,(πt−1)2,orthesquaredchangeinaverageforecast,i.e.,(ft)2.So,weregressdisagreementofforecasts
onthesemeasures.TheresultisshowninTable1:botharestatisticallysignificantat1%.Thisindicatesdisagreement
Table1
Disagreementamongforecastersofmonthlyinflation.Mean(σ)standsforthemeanofthedisagreementacross
datesandStd-Dev(σ)foritsstandarderror.(πt−1)2denotesthesquaredvariationinpastinflationand(ft)2
denotesthesquaredchangeinaverageforecast.Standarderrorsareinparentheses.
Disagreement,σ MonthlyIPCA
Descriptivestatistics Mean(σ) 0.091 Std-Dev(σ) 0.039 Regressionsofσon (πt−1)2 0.055***(0.014) (ft)2 0.827***(0.078)
Y.d.A.C.Cordeiroetal./EconomiA18(2017)40–59 47
Fig.3.Averagedisagreementamongrevisers.
respondtoshocks,whichisevidenceinfavorofthesticky-informationmodel.Butthisalsogoesinthemannerpredicted bythehybridmodel.Althoughmaintainingtheconstantnoiseinthesignal,inthehybridapproachwehaveashareof thepopulationthatupdatesitsinformationsetandasharethatdoesnot.Thisisenoughtogeneratetimevariancein disagreement.
Asecondpointofdivergencyregardstheanalysisofthedisagreementamongforecasterswhorevisetheirforecasts. Non-zerodisagreementinthiscaseisevidenceinfavorofthenoisy-informationapproach.Whilethesticky-information modelassumesthatagentswhoreceivenewinformationinthesameperiodhavethesameinformationset–andthus makethesamepredictions,whichshouldindicatezerodisagreementamongrevisers–thenoisy-informationmodel considersthateachagentreceivesnewinformationcontaminatedwithnoise.Therefore,evenagentswhoupdatetheir forecastsconcomitantlymaymakedifferentpredictions.Fig.3plotsdisagreementamongrevisersformonthlyIPCA forecasts.Asonecansee,disagreementisalwaysnon-zerobetween2002and2015–evidenceinfavourofthenoisy. Stillthisisalsopredictedbythehybridmodel.Asagentswhoupdatereceivenewinformationcontaminatedwithnoise, itisunlikelythatallforecastersthatupdatetheirinformationsetinagivendatemakeexactlythesamepredictions. Thisleadsdisagreementamongreviserstobenon-zero.
4. Methodology
4.1. Attention/inattentiondegree
Workingwithahybridmodelimplieswe’llderivetwomeasuresofattention/inattentiondegree,whicharerelated tothedegreeofinformationrigidityfacedbytheeconomy.AsmentionedinSection1,inthesticky-informationmodel theprobabilityofupdatingaforecastbetweentwoconsecutivedatescanbeseenasameasureofagents’attention tonewinformation.So,inordertoobtainourfirstmeasureoftheattentiondegree6–representedbyλ–wewantto estimate
P(fit,T =/ fit−1,T)
Fig.4.Average(sticky-information)degreeofattention,λ.
Table2
Degreeofattention(λ)–probabilityofupdatingaforecastofmonthlyIPCAinflationforadateTeventbetween
twoconsecutivemonths.
Degreesofattention,λ MonthlyIPCA
Descriptivestatistics
Mean( ˆλ) 0.4999
Std-Dev( ˆλ) 0.0657
wherefit,Tdenotesagenti’sforecastfortheIPCAinflationatdatetforthedateTevent.Assumingthattheprobability
ishomogeneousacrossinstitutionsandacrosshorizons,wecanestablishsomeempiricalcounterpartstothisattention degreefromthesurveydata:
ˆ
λt,T =(1/nt) nt
i=1
I(fit,T =/ fit−1,T)
whereI(·)isaindicatorfunction,thatassumesvalue1whentheindividual’sforecastfordateeventTchangesbetween
t−1andt,and0otherwise.
Inotherwords, ˆλt,T tellsusthepercentageofourpopulation–heretheinstitutionsrespondingthesurvey–that
updatedtheirforecastsandsupposedlyhadaccesstorelevantnewinformationbetweenthedatest−1andt.Thusthe greaterλ,thelowerthedegreeofinformationrigidityfacedbytheeconomy.
Fig.4showstheaverage(acrossinstitutionsandforecasthorizons)attentiondegreeforpredictionsmadeforeach dateTevent,specifiedinthehorizontalaxis.7Ascanbeseen,theattentiondegreeneverreaches1.Itinfactremains alwaysbetween0.3and0.8duringtheperiodunderstudyanditsmean(overtime)–denotedby ˆλanddescribedin
Table2–is0.4999.
7 Giventhestructureofourdatabase,toconstructasequenceofforecastsmadebythesameinstitutionforthesamedateT,wepoolthemonthly
Y.d.A.C.Cordeiroetal./EconomiA18(2017)40–59 49 Oursecondmeasureofinattentiondegreeisobtainedfromthenoisy-informationmodelandisrelatedtothevariance ofthemeasurementerror,i.e.,theerrorassociatedtothesignalaboutthestateoftheeconomyreceivedbytheinstitutions eachperiod–denotedby v.Ittellsushownoisyistheperceptionofthestateoftheeconomy–or howdifferent
fromtheactualinflationrealizationtheforecastcanbe–andthusthegreater v,thegreateristheinformationrigidity
facedbyagents.Thisparameterwillbeexplainedinmoredetailsbelow.
4.2. Themodel
DisagreementamongforecastersandpredictableforecasterrorsobservedinsurveydataanddocumentedinSection3
indicatethatbothsticky-andnoisy-informationmodelsmaydescribeexpectationformationinasatisfactoryway.So, we will implementthe methodologyproposedbyAndrade andLeBihan (2013). Thehybrid modelfeatures both approachesandinordertoassesstheempiricalperformanceofthisexpectationsmodel,itiscomparedtosomedata characteristics throughaMinimumDistanceEstimation(MDE)procedure.Thisprocedureallowstoanalyzeifthe modelcanreplicatetheforecasterrorandthedisagreementobservedinthesurveydata.
Wesketchbelowthe structureof themodel. WeassumethatZt isthestate that summarizesthe economy.The
followingAR(p)describeitsdynamics:
A(L)Zt=wt,
withA(L)=pk=0ρkLk,andwherewt haszeromeanandvarianceequaltoQ.
AsinMankiwandReis(2002),ateachdate,everyforecasterimayupdatehisinformationset,withagivenand constantprobabilityλ–modeledasaPoisson.Theagentswholastupdatedtheirinformationsetint−jarenamedthe generationj.Howeverwealsoconsideranoisyperceptionofthenewinformationwhenupdating,whichiscaptured bythesignalYitthatfollows
Yit=Zt+vit, vit∼i.i.d.(0, v),
wherevitisaprivateshockand vistheinattentionparameterofthenoisyapproach.
Theestimationprocedureinvolvesthreesteps.First,anAR(p)modelforinflationrate,withthevariablepreviously centered, is estimated.Second, takingthe AR(p) parameteras given, andfor a givensetof structural parameters (λ, v),fourmomentsaresimulatedusingtheKalmanfilterandthehybridmodel.Theselectedmoments–thesame
usedinAndradeandLeBihan(2013)–are:themeansquareofforecasterrors,forecasterrors’firstautocorrelation, disagreements’averagelevelanddisagreements’variance,respectivelydenotedby:E[e2],e(1),Mean[σ]andV[σ].
These momentsare relatedeither toforecast errors or disagreement, two key features that canbe reproduced by imperfectinformationmodelsandthat,inthishybridmodel,arefunctionsofthetwomeasuresofattention/inattention degree–λand v.
Next,thepreviousmodelisestimatedbytheMDEprocedure.Takethestructuralparametersvectorθ0=(λ, v),
thereducedformparametervectorμ0=(E[e2],e(1),Mean[σ],V[σ]),andh:R2→R4acontinuouslydifferentiable
function.TheMDEprocedureconsistsoffirstestimatingμ0byμ,ˆ andthenchoosinganestimatorθofθ0byminimizing thedistancebetweenμˆ andh(θ).Fromnowon,wedenoteh(θ)=μ(λ, v).Wethuscomputethedistancebetween
thevectorofdata-momentsμˆ andthecorrespondingmomentsgeneratedbythemodel,whichareafunctionofthe attention/inattentionparametersμ(λ, v).Theobjectivefunctionisminimizedoverthespaceofparameters.I.e.,one
mustminimize:
[μˆ −μ(λ, v)]ˆ−1[μˆ −μ(λ, v)],
where ˆisaconsistentestimatoroftheasymptoticvarianceofμ,ˆ i.e.,√T(μˆ −μ)→N(0,a )andμ(λ, v)takesthe
form: μ(λ, v)= ⎡ ⎢ ⎢ ⎢ ⎣ E[e2] e(1) Mean[σt] V[σt] ⎤ ⎥ ⎥ ⎥ ⎦= ⎡ ⎢ ⎢ ⎢ ⎢ ⎣ Et{Ej[(ejt,T) 2 ]} e(1) Et[σt,T] Vt[σt,T] ⎤ ⎥ ⎥ ⎥ ⎥ ⎦
where8 Ej[(ejt,T) 2 ]=ρ2(T−t)(1−Gt)2λ ∞ j=0 (1−λ)jPt|t−j−1+ρ2(T−t)G2t v+ρ 2(T−t)−1 ρ2−1 Q , e(1)=
E{[λ∞j=0(1−λ)j−1]2Ei(Zit|t−j−1)Ei(Zit|t−j−2)}
E{[λ∞j=0(1−λ)j−1]2Ei(Zit|t−j−1)2} and σt,T = λ∞ j=0 (1−λ)j ρ2(T−t){V i(Zit|t−j−1|j)+G2t[Vi(Zit|t−j−1|j)+ v]} +ρT−tEi(Zit|t−j−1)+ρT−tGtZt+ρT−tGtEi(Zit|t−j−1) − ⎛ ⎝λ∞ j=0 (1−λ)j ⎞ ⎠ ρT−tE i(Zit|t−j−1)+ρT−tGtZt+ρT−tGtEi(Zit|t−j−1)]] 2 } 5. Data
Inordertoascertainthenatureoftheexpectationsformationprocessandtoperformanempiricalassessmentof inattentionmodelswewilltakeprofessionalforecastsontheheadlinerateofinflation(π),IPCA,surveyedbytheMarket ExpectationsSystemdevelopedbytheCentralBankofBrazil.9 Professionalforecastersareparticularlyinteresting becausetheseagentsaresomeofthemostinformedintheeconomy.Asaresult,anyevidenceofinformationrigidity observedintheirexpectationsshouldbeindicativeofexistenceofinformationrigidityfacedbythewholeeconomy.
Since1999,theCentralBankofBrazilconductsadailysurveyofforecastsofthemainBrazilianmacroeconomic variables.Theso-called“Focussurvey”currentlycoversmorethan100institutions,includingcommercialbanks,asset managementfirms,consultingfirms,non-financialinstitutionsandotherlegalentities.Nonetheless,thedatasetusedin thispaperincludesforecastsmadeby254institutions,takingintoaccountthewholesampleperiodandtheinstitutions thatarecurrentlyactiveinthesystemandtheonesthatnolongerare.Wethusdealwithanunbalancedpanel.10
TheCentralBankofBrazilalsomakesavailablemonthlyrankingsincludingthe5bestforecasterswithregardto short,mediumandlongrun.Thisstimulatestheinstitutionstoupdatetheirforecastsregularlyandtoestimatethem withaccuracy,whichmakesthissurveydatamorereliable.Moreover,thesystemonlytakesintoaccountforecasts imputedinthelast30days,whichpreventsthestatisticstobeinfluencedbyforecaststhathaven’tbeenupdated.See
Marques(2013)forfurtherdetails.
Anotherimportantadvantageisthat,whilemostexpectations surveysconductedalloverare aggregated,which impliesoneisnotabletodirectlyobserveupdatesofforecastsinthemicrolevel,intheMarketExpectationsSystem wecantrackthesequenceofresponsesofaparticularsurveyparticipantovertime.11Finally,apartfrombeingableto exploitindividualdata,wehaveaccesstodailyresponses,whatmeanswecancloselyfollowthesequenceofforecasts madebyinstitutions.ThisisanadvantageoverAndradeandLeBihan’s(2013)study,sincetheECBSPFprovides forecastsonlyonaquarterlyfrequency.FortheIPCAinflation,weobserve(inadailybasis)5annualIPCAinflation forecasts,i.e.,forthecurrentcalendaryearandthenext4calendaryears,besidestheforecastforthetwelve-month
8 SeeAppendixAfordetailedcalculation.
9 ThecollectionandmanipulationofdatafromtheFocussurveyisconductedexclusivelybythestaffoftheCentralBankofBrazil.
10 Thesurveyrespondentsarefollowedthroughouttimewithareasonableturnover.Asnewparticipantsareoftenaddedtothesurvey,andothers
dropout,thepanelofsurveyforecastsisunbalanced.Thus,fromasetof254registeredinstitutionsinthesystem(inoursample),thereisasmaller
activegroupofaround100institutionsthatfrequentlyupdatetheirforecasts.Onaverage,87institutionsaredailysurveyedinourdataset.
Y.d.A.C.Cordeiroetal./EconomiA18(2017)40–59 51
Table3
Informationcriterion–LikelihoodRatio(LR),Akaike(AIC),Hannan-Quinn(HQIC)andSchwartzBayesian(SBIC)informationcriteria.
Selection-ordercriteria
Sample:May2002toDecember2014 Numberofobs=152
Lag LR AIC HQIC SBIC
0 0.961446 0.969527 0.98134
1 109.97* 0.251084* 0.267247* 0.290872*
2 0.14116 0.263313 0.287558 0.322995
3 0.07645 0.275968 0.308295 0.355544
4 0.05561 0.28876 0.329168 0.38823
*denotestheoptimallagineachcolumn.
aheadcumulatedinflationrateandmonthlyIPCAinflationforecastsforthepresentmonthandupto17monthsahead thecurrentdate.
In Brazil,theinflationrate asmeasuredbythe IPCAiscalculatedandreleasedonamonthlyfrequencybythe BrazilianInstituteofGeographyandStatistics(IBGE).Thus,inthisstudy,weusemonthlyIPCAinflationforecasts. TherangeofdatagoesfromJanuary2002toNovember2014.Theaggregationofthedailyintomonthlyresponseis donebyusingthedateofreferenceforIPCA,alsoknownasthe“criticaldate”,whichisthelastbusinessdaybefore theIPCA-15releasedate.12Wesimplytakethelastforecastmadeuntilthedateofreferenceasthemonthlyresponse, becausethisistheforecastusedbytheCentralBankofBraziltoformulatetheTop5rankingsthataremadepublicly available.So,weunderstandtheinstitutionshaveincentivestoupdatetheirforecastsclosetothisdate.
Weassumethattheprobabilityofupdatingaforecastisconstantacrosshorizonsandovertime.Havingestablished that,inordertoconstructasequenceofforecastsmadeforthesamedateTeventbythesameinstitution–sothatwe areabletocomparerevisionsofconsecutiveforecastsforthesameTandsamei–wepoolthemonthlyforecastsmade byeachinstitutionforthecurrentandthenext11months.Wethususuallyhaveasequenceof12forecasts–therefore 11revisions–foreachdateT.13
6. Estimationandresults
Asmentioned,firstwemodeltheinflationprocess.Theautoregressivemodelthatseemstobestdescribetheinflation rateistheAR(1).Thechoiceismadebytheanalysisofautocorrelationandpartialautocorrelationfunctionsandthe Schwartz’sBayesianInformationCriterion(SBIC)–althoughtheresultremainsthesamewithalternativeinformation criteria,seeTable3–consideringupto4lags.
Finally,theBreusch–Godofreytestofautocorrelationoftheresidualsdoesnotrejectthenullofnoautocorrelation (p-value=0.6855)andtheresidualsanalysiscorroboratethischoice.Thegraphicofthetimeseriesoftheresiduals canbeseeninFig.5,aswellastheoriginalIPCAandfittedIPCA.
Havingestablishedtheprocessthatrulesinflation,weproceedtothesimulationofthemoments,namelythemean squareofforecasterrors(E[e2]),forecasterrors’firstautocorrelation(e(1)),averagelevelofdisagreement(Mean[σt])
andvarianceofdisagreement(V[σt]).TheresultsobtainedfromsimulationarepresentedinTable4,Column(1).The
respectivesampledmoments–thatweideallywouldliketomatchwiththesimulatedmoments–arealsospecifiedin
Table4,Column(2).
FortheMDEprocedureweuseadiagonalmatrixasestimatorof(asymptoticcovariancematrix).Thevariance of each momentiscomputed utilizingNewey-Westestimatortoovercome autocorrelation inthe errorterms.The attention/inattentiondegreessimulatedbythehybridmodelare: ˜ v=0.1158and ˜λ=0.2256,whichisverydistant
fromtheλvalueof0.4999observedfromthedata–Table4.Infact,somethingonecanimmediatedrawfromthese
12ThecollectingperiodofIPCAisthecalendarmonth,whereasfortheIPCA-15itrangesfromthe16thofthepreviousmonthtothe15thofthe
referencemonth.
13Exceptforthefirst11andthelast11datesTsampled–i.e.,fromJanuary2002toNovember2002andfromJanuary2015toNovember2015–
Fig.5.TimeseriesoftheresidualsafterfittinganAR(1)toinflationprocess.
Table4
Simulatedandsampledattention/inattentiondegrees.λdenotesmean(overtime)probabilityofupdatingaforecastfordateTbetweentwoconsecutive
monthsand vstandsforthevarianceofthenoiseinthesignal.Firstcolumnspecifiesthemeasuressimulatedbythehybridmodelandsecond
columnthemeasureofλobtainedfromthedata,usingupto12monthsaheadforecasts.
Attention/inattentiondegrees
Simulated Sampled
λ 0.2256 0.4999
v 0.1158
Testofoveridentifyingrestrictions χ2
2statistic 2.3343
***Significantat1%;**Significantat5%;*Significantat10%.
numbersisthatinordertominimizethedistancebetweentheselectedmoments,themodelgeneratesamuchhigher degreeofinformationrigiditythantheobservedinthedata.
Using the MDE procedure, we can also perform a test of over-identifying restrictions – which null is that the parameters to be estimated (λ and v) can accurately describe the 4 selected moments, or alternatively,
that the distance between the simulated moments and the observed ones is zero. The test statistic has asymp-toticChi-squaredistributionwith2 degrees of freedom,χ22.Our statisticequals2.3343 (p-value=0.3113), which leadsus to not reject the null evenunder the statistical significance level of 10%. Although the over-identifying restrictions are not rejected by the data, as mentioned, that happens at the cost of a much lower frequency of updating(λ).14
Nowweanalyzehowthemomentsrespondtosomealternativecombinationsofλand v.Thisexerciseispresented
inColumns3and4ofTable5.Oncewefixλatthevalueobservedfromdata( ˆλ =0.4999)and vatthevaluegenerated
bythehybridmodel( ˜ v=0.1158)–Column3–thetwomomentsrelatedtotheforecasterrorandtime-variance
ofdisagreementbecome lowerandholdofffurtherfromthedata,andthemeanofdisagreementgetshigher.Ifon theotherhandwefixλat0.4999andraise vto0.9for example–Column4–whencomparedtoColumn3,we
observearaiseinthe firsttwomoments,relatedtoforecasterror, whilebothmomentsassociated todisagreement fall.
14 Lookingforalocalminimum,wecouldrestricttheintervalforpossiblevaluesofλaroundthevalueencounteredinthedata.Nevertheless,
Y.d.A.C.Cordeiroetal./EconomiA18(2017)40–59 53
Table5
Moments.E[e2]denotesmeansquareofforecasterrors,
e(1)forecasterrors’firstautocorrelation,Mean[σt]theaveragelevelofdisagreementand
V[σt]itstimevariance.Firstcolumnspecifiesthemomentssimulatedbythehybridmodelandsecondcolumnthemomentsobtainedfromthedata,
usingupto12monthsaheadforecasts.Columns3to5specifytheresultsunderalternativeconfigurations.
Moments (1) (2) (3) (4) (5) Simulated λ= ˜λ Data λ=0.4999 λ=0.4999 λ=0.4999 v= ˜ v v= ˜ v v=0.9 vfree(=0.4596) E[e2] 0.0685 0.1267 0.0536 0.0757 0.0686 e(1) 0.6271 0.6366 0.5919 0.6449 0.6280 Mean[σt] 0.0912 0.0912 0.1150 0.0752 0.0908 V[σt] 0.000414 0.0015 0.000154 0.01549×10−3 0.04305×10−3
Thesepatternsarewhatwould beexpectedfromthemodel.Accordingtothemodelanincreasein v (i.e.,an
increaseinthevarianceofthenoiseassociatedtothesignalreceivedbytheagents)oradecreaseinλ(i.e.,adecreasein theprobabilityofupdatingaforecastbetweentwoconsecutivemonths)arebothindicativeofanincreaseininformation rigidity.Itshouldthusraisebothvarianceandpersistenceofforecasterrors.Theeffectonaveragedisagreement,onthe otherhand,isambiguous.Finally,anincreasein vshoulddecreasetimevarianceofdisagreement,whileadecrease
inλshouldraiseit–seeAppendixAformoredetails.Theoppositeshouldhappenwhenwefaceanincreaseinλor adecreasein v.
Finally, in Column 5 of Table 5 we fix λ at the value observed from the data ( ˆλ =0.4999) and let v
free to be simulated by the model. We obtain v=0.4596. When compared to the case where all
parame-ters are freely simulated (Column 1), the two moments related to forecast errors and the mean of disagreement do not change too much. Nevertheless the variance of disagreement gets even lower, holding off from the data value.
Theresultsindicatethat,despiteoffittingthepersistenceofforecasterrorsandthelevelofdisagreementrelatively well,themodelcannotaccountforbothhighvarianceofforecasterrorsandofdisagreement.Infact,wheneverwefit sampledvarianceofdisagreement,forexample,wemustset vatamuchlowerlevel.Butthisinturnleadstomuch
higheraveragedisagreementthanobservedandevenlowervarianceofforecasterrors.Ontheotherhand,togetcloser tothevarianceofforecasterrorsweneedalowerλandamuchhigher v.Neverthelessitleadstohigherpersistence
offorecasterrorsandmuchlowerlevelandvarianceofdisagreement.
7. Conclusion
In thisstudywe usedthe MarketExpectations Systemof CentralBankofBraziltoanalyzehowgoodcanthe expectationsformationprocessofinflationforBraziliandatabemodeledbythehybridmodelfeaturingboth sticky-informationandnoisy-informationmodels.Themodelisbasedontheassumptionthattheagentsdonothaveaccessto informationinstantly/everyperiod,andthatwhenagentsupdatetheirinformationset,theyonlyobservenoisysignals aboutthetruestateandformexpectationsrationally.
Utilizing a Minimum Distance Estimation (MDE) procedure, we were able to test if the model is capable of quantitatively fitting the Brazilian data. Estimating the model we encounter that formally the model fits the data. Nevertheless, it happens at the cost of a much higher degree of information rigidity than observed in the data.
Infutureresearch,onepossiblewayofimprovingthefitofthemodeltothedatamaybeallowingthedegreeof attentionofthesticky-informationapproach(λ)tovaryacrossinstitutions–somethingthatisobservedinthedata.Or onecouldtrytoincludethestickinessasinReis(2006a,b).Thatis,theforecasterexplicitlychoosestobeinattentive, solvingaoptimizationproblemtodecidewhentoupdatehisinformationset–insteadofbeingsubjecttoanexogenous constantprobabilityofbeingabletogetnewinformation.Otherpossibleroutesaretotestdifferentmodels,following theapproachofCoibionandGorodnichenko(2012b),ortoconsiderdispersedinformationwithstrategicinteractions, inlinewiththemethodologyofMorrisandShin(2002).
AppendixA. Themodel
Thestatespacerepresentationisgivenby:
Yit=Zt+vit
Zt =ρZt−1+wt
(A.1)
wherevit=it+ηt∼i.i.d.(0, v)andwt∼i.i.d.(0,Q).
Thefirst(A.1)equationiscalledtheobservationequationandthesecondthestateequation. Thedatetstateoftheeconomyperceivedbyagentiatdatet−1is
Zi,t|t−1=E[Zt|Ii,t−1] (A.2)
whereIi,t−1=(Yi,t−1 ,Yi,t−2 ,...,Yi,1 )and
Pt|t−1=E[(Zt−Zi,t|t−1)(Zt−Zi,t|t−1)] (A.3)
Besides,
Yi,t|t−1=E[Yi,t|Ii,t−1] (A.4)
E[Zt]=ρE[Zt−1] (A.5)
E[Yi,t|Zt]=Zt (A.6)
Thus,
Yi,t|t−1=E[Yi,t|Ii,t−1]=E[E(Yi,t|Zt)|Ii,t−1]=E[Zt|Ii,t−1]A.2=Zi,t|t−1 (A.7)
TheoptimalforecastfordateT15eventmadebygenerations0andj–i.e.,theagentsthatupdatedtheirinformation setinthecurrentperiodandtheoneswhoupdatedjperiodsago–isrespectivelygivenby:
fit,T =E[ZT|Zi,t|t]=ρT−tZi,t|t (A.8)
fit−j,T =E[ZT|Zi,t−j|t−j]=ρT−tZi,t|t−j (A.9)
From(A.1)and(A.7)
Yi,t−Yi,t|t−1=Zt+vit−Zi,t|t−1=(Zt−Zi,t|t−1)+vit (A.10)
thus
E[(Yi,t−Yi,t|t−1)(Yi,t−Yi,t|t−1)]
=E[[(Zt−Zi,t|t−1)+vit][(Zt−Zi,t|t−1)+vit]]
=E[(Zt−Zi,t|t−1)(Zt−Zi,t|t−1)]+E[vitvit]=Pt|t−1+ v (A.11)
wherethesecondequalitycomesfromvitandZtbeingnon-correlated.
Y.d.A.C.Cordeiroetal./EconomiA18(2017)40–59 55 TheKalmanFiltersolutiontothissignalextractionproblemisgivenby:
Zi,t|t =Zi,t|t−1+Git(Yit−Yi,t|t−1) (A.12)
whereGitistheKalmangain.Weassumeittobehomogeneousacrossagents:Git=Gt =Pt|t−1(Pt|t−1+ v).
Usingthissolution,wecanrewritetheoptimalforecastas:
fit,TA.8= ρT−tZi,t|t=ρT−tZi,t|t−1+ρT−tGt(Yit−Yi,t|t−1) (A.13)
Theaverageforecastwithinagenerationjis:
Ei[fit−j,T|j] =ρT−tEi[Zi,t|t−j−1]+ρT−tGtEi
×[Zt−Zi,t|t−j−1+vit]A.1= ρT−t+1Ei[Zi,t−1|t−j−1]+ρT−tGt[Zt−ρEi(Zi,t−1|t−j−1)] (A.14)
andtheforecasterrorwithinagenerationjis:
ejt,T =ZT −Ei[fit−j,T|j]=ρT−tZt+ T−t
i=1
ρT−t−iwt+i−{ρT−tEi[Zi,t|t−j−1]+ρT−tGt[Zt−Ei(Zi,t|t−j−1)]}
=ρT−t(1−Gt){Zt−Ei[Zi,t|t−j−1]}+ T−t
i=1
ρT−t−iwt+i (A.15)
Theaverageforecastacrossgenerationsfollows:
Ej{Ei[fit−j,T|j]}=λ
∞
j=0
(1−λ)jEi[fit−j,T] (A.16)
andtheaverageforecasterrorisgivenby:
et,T =Ej[ejt,T]=ZT −Ej{Ei[fit−j,T|j]}=ρT−tZt+ T−t i=1 ρT−t−iwt+i−λ ∞ j=0 (1−λ)jEi[fit−j,T] =ρT−tZt+ T−t i=1 ρT−t−iwt+i−λ ∞ j=0 (1−λ)j{ρT−tEi[Zi,t|t−j−1]+ρT−tGt[Zt−Ei[Zi,t|t−j−1]} =ρT−t(1−Gt)λ ∞ j=0 (1−λ)j[Zt−Ei(Zit|t−j−1)]+ T−t i=1
ρT−t−iwt+i (A.17)
Wenowturntothe disagreementgenerated bythishybrid-model.Our measureofdisagreement– presentedin Section3.2–isthecross-sectionstandarddeviationateachdate.So,wederivethecorrespondingformulaobtained bythetheoreticalmodel.Inordertomakecalculationssimpler,sowedonothavetodealwithsquaredroots,wewill herelookatthecross-sectionvariance.
Notethatthetotalcross-sectionvarianceofpointforecastsacrossindividualsiindifferentgenerationsjcanbe decomposedintwoelements:firstthedifferencesofperceptionsofthestateoftheeconomywithinagivengeneration, weightedbyitsrelativeshareinthetotalpopulationandsecondthedifferencesinaverageperceptionofthestateof
theeconomybetweendifferentgenerations:
Vij(fit−j,T) = Ej[Vi(fit−j,T|j)]+Vj[Ei(fit−j,T|j)]
= λ ∞ j=0 (1−λ)jVi(fit−j,T|j)+λ ∞ j=0 (1−λ)j{Ei(fit−j,T|j)−Ej[Ei(fit−j,T|j)]}2 A.13= λ∞ j=0
(1−λ)j Vi[(ρT−tZit|t−j−1+ρT−tGt(Zt−Zit|t−j−1+vit))|j]+ {Ei(fit−j,T|j)
−Ej[Ei(fit−j,T|j)]}2 = λ ∞ j=0 (1−λ)j ⎧ ⎨ ⎩ρ2(T−t){Vi(Zit|t−j−1|j)+G2t[Vi(Zit|t−j−1|j)+ v]} + ρT−tEi(Zit|t−j−1)+ρT−tGtZt+ρT−tGtEi(Zit|t−j−1) − ⎛ ⎝λ∞ j=0 (1−λ)j ⎞ ⎠ ρT−tE i(Zit|t−j−1)+ρT−tGtZt+ρT−tGtEi(Zit|t−j−1) ⎤⎦2 ⎫ ⎪ ⎬ ⎪ ⎭ (A.18) So,disagreementgeneratedbythemodelisgivenby:
σt,T = Vij(fit−j,T) (A.19)
A.1. Selectedmoments
Basedonthepreviousequation,letsderivethefourmoments-themeansquareofforecasterrors,forecasterrors’ firstautocorrelation,disagreements’averagelevelanddisagreements’variance–usedintheprocedure.
First,wederivethemeanofsquaredforecasterror. Ej[(ejt,T) 2 ]=Ej{Ei[(ZT −(fit−j,T|j))2]} =Ej{Ei{[(ρT−t(1−Gt)(Zt−Zit|t−j−1))−ρT−tGtvit] 2 }+Ei T−t i=1 (ρT−t−iwt+i) 2 ! } =Ej{Ei[ρ2(T−t)(1−Gt)2(Zt−Zit|t−j−1)2]+ρ2(T−t)G2tEi[v2it]+ T−t i=1 ρ2(T−t−i) ! Ei[w2t+i]} =ρ2(T−t)(1−Gt)2Ej{Ei[(Zt−Zit|t−j−1)2]}+ρ2(T−t)G2t v+ρ 2(T−t)−1 ρ2−1 Q =ρ2(T−t)(1−Gt)2Ej[Pt|t−j−1]+ρ2(T−t)G2t v+ ρ2(T−t)−1 ρ2−1 Q =ρ2(T−t)(1−Gt)2λ ∞ j=0 (1−λ)jPt|t−j−1+ρ2(T−t)G2t v+ρ 2(T−t)−1 ρ2−1 Q (A.20)
Ourfirstmomentisthemean(overtime)ofmeansquaredforecasterror,i.e.,
E[e2]=Et{Ej[(ejt,T)
2
Y.d.A.C.Cordeiroetal./EconomiA18(2017)40–59 57 Theforecasterrors’firstautocorrelationisdetailedbelow.
e(1)= γ j 1 V(ejt,T) = E[(ejt,T −et,T)(ejt−1,T−et−1,T)] V(ejt,T) (A.22) where γ1j=E{[ρT−t(1−Gt)(Zt−Ei(Zit|t−j−1))+ T−t i=1 ρT−t−iwt+i−ρT−t(1−Gt)λ ∞ j=0 (1−λ)j[Zt−Ei(Zit|t−j−1)] −T−t i=1
ρT−t−iwt+i]× [ρT−t(1−Gt)(Zt−Ei(Zit|t−j−2))
+T−t i=1 ρT−t−iwt+i−ρT−t(1−Gt)λ ∞ j=0 (1−λ)j[Zt−Ei(Zit|t−j−2)]− T−t i=1 ρT−t−iwt+i]} =ρ2(T−t)(1−Gt)2E{(λ ∞ j=0 (1−λ)j−1) 2
Ei(Zit|t−j−1)Ei(Zit|t−j−2)} (A.23)
and V(ejt,T)=E{[ρT−t(1−Gt)(λ ∞ j=0 (1−λ)j−1)Ei(Zit|t−j−1)] 2 } (A.24) Thus e(1)=
E{[λ∞j=0(1−λ)j−1]2Ei(Zit|t−j−1)Ei(Zit|t−j−2)}
E{[λ∞j=0(1−λ)j−1]2Ei(Zit|t−j−1)2}
(A.25) Theaveragedisagreementis:
Mean[σt]=Et[σt,T] (A.26)
anddisagreement’stimevariance:
V[σt]=Vt[σt,T] (A.27)
whereσt,Tisgivenby(A.19).
A.2. Selectedmomentsandthedegreeofinformationrigidity
Weshallnowtakealookathowthesemomentsrelatetochangesinthedegreesofattention/inattention.Recallλ
istheprobabilityofupdatingaforecastafteroneperiodofnewinformation.Thusλisinverselyproportionaltothe degreeofinformationrigidity. v,ontheotherhand,describeshowvolatileisthenoiseoftheperceptionofthestate
oftheeconomy.Thisimpliesthatthegreater v,thehigherthedegreeofinformationrigidity.
Supposeanincreaseinλoradecreasein v,whicharebothequivalenttoadecreaseininformationrigidity.This
meansagentshaveaccesstomoreaccurateinformationandshouldconsistentlymakebetterpredictions.Everything elseremainingconstant,wewillfaceadecreaseinthevarianceof theforecasterrorsandthesamehappenstoits persistence.Itmeanstheresponseofthefirsttwomomentsgoesinthesamedirectionasthechangeininformation rigidity.
Withregardtothetimevarianceofdisagreement,anincreaseinλdecreasesthedifferenceofinformationbetween generations.Thusthearrivalofanewinformationsetthatisobservedonlybygenerationj=0andthatcancontain alotoforjustafewrelevantinformationunnotedbythepreviousgenerations–whichcanbeseenasthesizeofa shockhittingtheeconomyonthatdate–willhavealowerimpactonthemagnitudeofdisagreementandthatleads
tolowertimevarianceofdisagreement.Ontheotherhand,adecreasein vdecreasestheamountofnoiseaboutthe
stateoftheeconomyfacedbyagents.Thisimpliesnewinformationismoreinformativeandthusmoreweightisgiven tonewinformationwhenmakingaforecast.Thusitcanalsorisetimevarianceofdisagreement,sinceitwouldleadto disagreementbeinghigherwhenmorerelevantinformation(oralargershock)arisesandtolessdisagreementwhen fewrelevantinformation(orasmallershock)arises.
Finally,withrespecttotheresponseoftheaveragedisagreement,wehavetwoconcurrenteffectsforbothdegreesof (in)attention.Anincreaseinλ,forexample,decreasesthelengthbetweeninformationupdating–becausetheprobability toreceivenewinformationfacedbyeachagenteachperiodisnowgreater-,whichdecreasestheheterogeneitybetween forecastsmadebydifferentgenerationsandshouldimplyadecreaseindisagreement.Butitalsodecreasestheshareof generationsbasingtheirpredictiononoldinformation.Withinthegenerationsthathaven’tupdatedtheirinformation setfortoolong,theoptimalforecastissimplytheunconditionalmeanoftheinflationprocessanddisagreementtends tozero.Thusadecreaseintheshareofthesesgenerationscontributestoincreasedisagreement.
Considernowadecreasein v.Ononehand,itdecreasesthedifferentperceptionswithineachgenerationof
fore-casters(viathefallintheamountofnoisefacedbyagentsaboutthestateoftheeconomy)anddecreasesdisagreement. Ontheotherhand,adecreasein vimpliesagentstendtoincorporatemoreofthenewinformationintotheirforecast,
becausethesignalismoreprecise.Thisleadstoanincreaseinaveragedisagreement.
Therefore,whiletheeffectofachangeinthedegreeofinformationrigidityonthemomentsrelatedtoforecasterrors canbeeasilyanticipated,thedirectionofthemomentsrelatedtodisagreementdependonwhichforceswillprevail.
TableA.1summarizesit:
TableA.1
Responseofthemomentstochangesintheattention/inattentiondegrees.λstandsfortheprobabilityofupdatingaforecastbetweentwoconsecutive
monthsand vforthevarianceofthenoiseinthesignal.E[e2]denotesmeansquareofforecasterrors,e(1)forecasterrors’firstautocorrelation,
Mean(σt)theaveragelevelofdisagreementandV(σt)itstimevariance.
Moments
DecreaseinInformationRigidity
Increaseinλ Decreasein v
E[e2] ↓ ↓
e(1) ↓ ↓
Mean[σt] ambiguous ambiguous
V[σt] ↓ ↑
References
Andrade,P.,Crump,R.K.,Eusepi,S.,Moench,E.,2013December.Noisyinformationandfundamentaldisagreement.StaffReports655.Federal
ReserveBankofNewYork.
Andrade,P.,LeBihan,H.,2013.Inattentiveprofessionalforecasters.J.Monet.Econ.60(8),967–982.
Areosa,M.,Areosa,W.,2012.AssetPricesandMonetaryPolicy–Asticky-dispersedinformationmodel.WorkingPapersSeries285.CentralBank
ofBrazil.
Areosa,M.,Areosa,W.,Carrasco,V.,2012.ASticky-DispersedInformationPhillipsCurve:amodelwithpartialanddelayedinformation.Working
PapersSeries276.CentralBankofBrazil.
Branch,W.A.,2007.Stickyinformationandmodeluncertaintyinsurveydataoninflationexpectations.J.Econ.Dyn.Control31(January(1)),
245–276.
Coibion,O.,Gorodnichenko,Y.,2012December.InformationRigidityandtheExpectationsFormationProcess.IMFWorkingPapers12/296.
InternationalMonetaryFund.
Coibion,O.,Gorodnichenko,Y.,2012b.Whatcansurveyforecaststellusaboutinformationrigidities?J.Polit.Econ.120(1),116–159.
Guillén,D.,2008.EnsaiossobreExpectativasdeInflac¸ãonoBrasil(Master’sthesis).Puc-Rio.
Mackowiak,B.,Wiederholt,M.,2009.Optimalstickypricesunderrationalinattention.Am.Econ.Rev.99(June(3)),769–803.
Mankiw,N.G.,Reis,R.,2002.Stickyinformationversusstickyprices:aproposaltoreplacethenewKeynesianPhillipscurve.Q.J.Econ.117
(November(4)),1295–1328.
Mankiw,N.G.,Reis,R.,Wolfers,J.,2004October.Disagreementaboutinflationexpectations.In:NBERMacroeconomicsAnnual2003,NBER
Y.d.A.C.Cordeiroetal./EconomiA18(2017)40–59 59
Marques,A.B.C.,2013June.CentralBankofBrazil’smarketexpectationssystem:atoolformonetarypolicy.In:BankforInternationalSettlements,
IFCBulletins,vol.36,pp.304–324,Availableat:http://www.bis.org/ifc/publ/ifcb36t.pdf.
Morris,S.,Shin,H.S.,2002.Socialvalueofpublicinformation.Am.Econ.Rev.92(December(5)),1521–1534.
Moscarini,G.,2004.Limitedinformationcapacityasasourceofinertia.J.Econ.Dyn.Control28(September(10)),2003–2035.
Patton,A.J.,Timmermann,A.,2010.Whydoforecastersdisagree?Lessonsfromthetermstructureofcross-sectionaldispersion.J.Monet.Econ.
57(October(7)),803–820.
Reis,R.,2006a.Inattentiveconsumers.J.Monet.Econ.53(November(8)),1761–1800.
Reis,R.,2006b.Inattentiveproducers.Rev.Econ.Stud.73(3),793–821.
Sims,C.A.,2003.Implicationsofrationalinattention.J.Monet.Econ.50(April(3)),665–690.
Woodford,M.,2003.Imperfectcommonknowledgeandtheeffectsofmonetarypolicy.In:Aghion,P.,Frydman,R.,Stiglitz,J.,Woodford,M.