Noisy Macroeconomic Announcements, Monetary Policy, and Asset Prices *

Transcription

1 Noisy Macroeconomic Announcemens, Moneary Policy, and Asse Prices Robero Rigobon Sloan School of Managemen, MIT and NBER Brian Sack Macroeconomic Advisers, LCC June 8, 006 Absrac The curren lieraure has provided a number of imporan insighs abou he effecs of macroeconomic daa releases on moneary policy expecaions and asse prices. However, one puling aspec of ha lieraure is ha he esimaed responses are quie small. Indeed, hese sudies ypically find ha he major economic releases, aken ogeher, accoun for only a small amoun of he variaion in asse prices even hose closely ied o near-erm policy expecaions. In his paper we argue ha his apparen deachmen arises in par from he difficulies associaed wih measuring macroeconomic news. We propose wo new economeric approaches ha allow us o accoun for he noise in measured daa surprises. Using hese esimaors, we find ha asse prices and moneary policy expecaions are much more responsive o incoming news han previously believed. Our resuls also clarify he se of facs ha should be capured by any model aemping o undersand he ineracions beween economic daa, moneary policy, and asse prices. We hank John Campbell and Larry Meyer for valuable commens and suggesions, and Mike McMorrow for excellen assisance wih he analysis. 1

2 1. Inroducion The relaionship beween economic daa, on he one hand, and asse prices and moneary policy, on he oher, has become a widely sudied opic in he academic lieraure and for good reason. Macroeconomic condiions are a key facor deermining near-erm policy expecaions, and hose expecaions reverberae hroughou he financial sysem by influencing he reurns expeced on all asse classes. Bu despie being widely sudied, our curren knowledge of he ineracions beween economic news and asse prices has many shorcomings, and he resuls are puling in some dimensions. Perhaps mos imporanly, he esimaed effecs of daa releases on moneary policy expecaions and asse prices are found o be relaively small. This is he case even for hose asses ha are known o be very sensiive o near-erm moneary policy expecaions, such as eurodollar fuures and shor-erm Treasury securiies. This finding is surprising. Afer all, he lieraure over he pas wo decades has argued ha moneary policy o a large exen responds sysemaically o economic condiions. Indeed, he lieraure has made remendous progress esimaing moneary policy rules ha accoun for hese sysemaic responses in erms of low-frequency daa (such as quarerly daa). If moneary policy is so sysemaic, one would expec o see evidence of i also in he higher-frequency movemens in ineres raes and asse prices around daa releases. Tha is, he major economic daa releases would be expeced o explain an exensive amoun of he variaion in asses sensiive o near-erm policy expecaions. In our view, he pule of he deachmen of moneary policy expecaions and asse prices from he incoming economic news is parly relaed o he difficulies associaed wih measuring he surprise componen of ha news. Mos sudies o dae compue a surprise measure for a given release based on expecaions aken from a survey conduced ahead of he release. They hen regress changes in an asse price on his surprise measure, which we refer o as he sandard evensudy approach. The aemp o isolae he unexpeced componen of he release was a vas improvemen over earlier effors ha could no make such a separaion, as only he unexpeced componen should promp a marke reacion. However, his approach likely falls shor of accuraely measuring he marke effecs of he incoming news perhaps considerably. A problem wih he sandard evensudy approach is ha he macroeconomic news is likely o be measured very poorly, for several reasons. Firs, i is hard o accuraely measure wha he markes are expecing for a given release a he ime i comes ou, including he full disribuion of risks seen for he release. Second, even if one accuraely measured expecaions, he acual release may be seen as a noisy indicaor of he underlying rue fundamenal facor ha drives marke responses, And hird, he variable measured is usually only one componen of a repor. Afer all, mos of hese repors are complicaed, providing los of informaion of varying relevance. Thus, i is quie likely ha he macroeconomic surprise included on he righ-hand-side of he evensudy regressions is only a very rough measure of he rue incoming news. This paper focuses on measuring he reacion of asse prices and moneary policy expecaions o he rue economic news embedded in he major U.S. daa releases. Raher han aemping o beer measure he daa or he expecaions, we focus on developing economeric echniques ha will adequaely deal wih he measuremen problems associaed wih he daa surprises used in he exising evensudy lieraure.

3 Our effors ake us in wo direcions. Firs, we modify he sandard evensudy regression framework o accoun for he possibiliy ha he measured surprises conain error. The measuremen issues considered here lead o a classical error-in-variables problem of a sandard regression, one ha biases downward he esimaed sensiiviy of asse prices o he incoming daa. We develop a new esimaor ha allows for measuremen error and hence eliminaes his downward bias. The procedure could be used in oher applicaions o correc for he error-invariables problem. Second, we employ a principal componens approach ha removes he need o even ry o measure he daa surprises. In effec, he approach uses he observed marke reacions o infer wha he rue daa surprises were. Such an approach may have appeal if one regards he incoming daa as being complex and having many dimensions ha could affec asse prices condiions ha make i difficul o measure he daa surprise in he manner of he sandard evensudy exercise. The resuls provide us wih unbiased esimaes of he response of moneary policy expecaions and asse prices o he rue surprise conained in all of he major daa releases. They also allow us o recover he imporance of hose rue surprises. An imporan finding from he paper is ha macroeconomic daa releases maer o a much greaer exen han found in previous sudies ha is, hey accoun for a greaer porion of he flucuaions in marke ineres raes. Moreover, using hese esimaors, we are able o refine a se of paerns in he responses ha should be explained by any model addressing he ineracions beween economic variables, moneary policy, and asse prices.. Esimaing he Effecs of Macroeconomic Announcemens: Curren Mehods Researchers in boh macroeconomics and financial economics are very ineresed in undersanding he linkage beween moneary policy and asse prices. To ha end, one srand of lieraure has aemped o measure he response of asse prices o moneary policy shocks, or he erraic and unpredicable componen of moneary policy decisions. Bu such shocks are limied in sie and accoun for only a very small porion of he variaion in asse prices. Insead, mos of he movemen in shor-erm ineres raes likely represens he sysemaic response of moneary policy o economic developmens. Thus, i may be more relevan o invesigae he responses of moneary policy expecaions and asse prices o incoming news abou he economy. A siable lieraure has aken up his opic and has provided us wih some valuable resuls. The sudies o dae almos uniformly ake an approach ha is commonly referred o as evensudy..1. The Evensudy Specificaion Papers in he evensudy lieraure ypically proceed in a simple regression framework in which he reacion of a given asse price (or marke yield) is regressed on he surprise componens of he daa release, as in he following specificaion: s = γ + ε (1a) τ [ M ] = M E, (1b) 3

4 M is he released value of he macroeconomic announcemen and [ ] where M is a measure of he marke s expecaion ahead of he release. The specificaion assumes ha he only marke-moving informaion is he surprise componen of he release, and he parameer γ is he marke sensiiviy o ha surprise which is he primary ineres of his paper. The basic approach implici in specificaion (1) has no varied much over ime, bu he empirical implemenaion of he equaion has changed in wo dimensions. Firs, he measure of expecaions has improved. Early papers in his area had o model he marke s expecaions eiher as pas realied values of he macroeconomic variables or as he oucome of forecasing models ha do no necessarily perform very well. More recenly, researchers have increasingly relied on surveys o measure expecaions and o beer isolae he surprise componen of daa releases. Hence, he measuremen of he variable has likely improved over ime. Second, sudies have increasingly used a narrower window o measure he marke response o he daa release. Whereas earlier papers may have used monhly or quarerly daa, he evensudy lieraure has moved o using changes a a daily frequency (see, for example, McQueen and Roley (1993) and Gürkaynak, Sack, and Swanson (005)) or even in some cases on an inraday bases (see, for example, Fleming and Remolona (1997) and Baldui, Elon, and Green (001)). 1 The idea of using a narrower window is o reduce he influence of oher evens ha migh be affecing he asse price in addiion o he daa surprise. In erms of he equaion (1), i reduces he variance of he error erm ε, which should improve he accuracy of he esimae of he parameer γ. The evensudy approach has imporanly conribued o our undersanding of he manner in which moneary policy expecaions and asse prices reac o incoming economic daa. Indeed, as we will show below, his approach finds ha he marke reacion o a number of releases is saisically significan. Neverheless, in our view he evensudy approach has some shorcomings ha preven i from recovering he marke response o a rue macroeconomic daa surprise... The Economeric Problem: Noisy Daa Surprises The poenial problem ha arises wih he evensudy approach is ha he resuls will only be as good as he measure of daa surprises included on he righ-hand side of he equaion. Indeed, he model (1) implicily assumes ha he measured daa surprise ruly capures he rue macroeconomic news arising from he releases. If ha is no he case, he esimaed parameer γ will be biased. E τ 1 Several oher papers use inraday daa bu focus primarily on foreign exchange raes, including Andersen, Bollerslev, Diebold, and Vega (003) and Faus, Rogers, Wang, and Wrigh (003). As an example of a paper using lowfrequency daa, Culer, Poerba, and Summers (1989) aemp o measure he influence of macroeconomic shocks on equiy prices using a monhly VAR. Sock and Wason (003) provide a review of oher papers ha examine he relaionship beween financial variables and macroeconomic condiions a a monhly or quarerly frequency. Noe, however, ha he primary ineres in ha paper is measuring he predicive power of financial variables for economic oucomes raher han he effecs of economic oucomes on financial variables. 4

5 Even wih he improvemens noed above, i is a somewha dubious assumpion ha he variable is perfecly measured or ha i is even well measured. Insead, i is more plausible ha he variable conains considerable measuremen error, from a variey of sources. Firs, i is unlikely ha he survey measures used accuraely capure he marke expecaions a he ime of he release. In he resuls presened below, we collec hose expecaions from wo surveys and splice hem ogeher o creae a full ime series. Before Sepember 004, we use he median response from he Money Marke Services survey, which is a survey of professional forecasers aken he Friday before each release. Since hen, we insead use he median response from he regular survey aken by Bloomberg. This figure is he mos commonly discussed measure of consensus expecaions in he financial markes. Bu here are a number of reasons o believe ha he expecaions measured from hese surveys are no necessarily appropriae for gauging he marke response. The survey respondens are no he relevan marke paricipans whose expecaions maer. Moreover, he survey covers a variey of respondens wih very differen backgrounds and skill ses, raising quesions abou wheher cerain individual responses could disor he measures. I is no even clear ha he respondens have he correc incenive scheme, as we suspec ha hey may assign greaer uiliy o having an ou-of-consensus call ha comes in correc han having a consensus call ha comes in correc. And lasly, we arbirarily use he median from he panel, hough he argumen for using his over he mean or some oher measure is no clear-cu. In addiion o concerns abou he cross-secion of paneliss, we also have some concerns abou he iming of he surveys. Ideally, we would like o know he marke expecaions he momen before he daa release. The MMS survey is insead aken he Friday before he release, making i somewha sale. For hose releases ha come ou on a Friday (e.g., he employmen repor), ha leaves an enire week (and all he daa released ha week) for expecaions o evolve and move away from he survey response. And he siuaion for he Bloomberg expecaions is even worse. Those responses are submied a irregular imes. Mos respondens ener heir esimaes abou a week before he releases, bu many insead do i wo weeks in advance while ohers wai unil he week of he release. Anoher source of mismeasuremen of he macroeconomic surprise is he daa release iself. The released daa can be hough of as a noisy version of he rue economic fundamenal o which he marke responds. Researchers usually focus on jus one aspec of he release, and ofen ha one aspec can appear anomalous. A recen example was he advance GDP repor for he fourh quarer of 005, which came in well below he marke s expecaions. Tha surprise owed in large par o a puling drop in defense spending ha quarer, and hence Wall Sree analyss generally dismissed he implicaions of he repor. 3 Overall, we believe ha he measured daa surprises could be quie noisy. Marke expecaions are probably no measured paricularly well, as he survey used is a random variable ha a bes To ake an example, consider he employmen repor ha was released on November 5, 004. Of he 78 responses o he survey, 13 were submied more han wo weeks in advance. Mos of he responses, 39, came in he abou one week in advance (wih ohers coming in earlier ha week). And 4 respondens waied unil he week of he release o submi heir views. 3 For example, David Greenlaw from Morgan Sanley summaried he repor as follows: Much weaker han expeced repor. Boh final sales and invenories came in well below expecaions in Q4. However, we believe ha a significan porion of he downside is likely o be recouped in Q1 Defense [spending] plunged 13% in Q4. We suspec ha a leas some of his drop reflecs a iming quirk ha will be unwound in Q1. 5

6 can be considered o be unbiased bu no measured wihou error. And he acual release is likely o conain some noise relaive o he rue macroeconomic news ha affecs markes..3. The Bias in Evensudy Esimaes We sar wih he assumpion ha he macroeconomic surprises used in he evensudy lieraure are measured wih error for he reasons discussed above. In his case, he esimaes obained in he sandard lieraure are plagued wih error-in-variables bias. To provide some srucure for discussing he problem, we assume he asse price change immediaely around he release a ime is denoed by s. This marke reacion is driven by he rue macroeconomic news conained in he announcemen, which we denoe, according o he following equaion: s = γ + ε. () We are ineresed in measuring he sensiiviy of financial markes o he rue economic news, capured by he parameer γ. The residual ε capures movemens in he asse price in ha window ha are no driven by he daa surprise (or a leas no under his linear srucure). To esimae equaion (), mos researchers aemp o measure he rue macroeconomic news as he difference beween he released daa and he expecaion of ha daa, where he expecaion is ypically deermined from a survey aken in advance of he release. Bu, as discussed above, here are wo poenial problems wih ha measure ha he release may be seen as a noisy version of he rue relevance of he news, and ha he expecaions may be measured poorly. Considering his, we should perhaps ake he measured daa surprises o be a noisy represenaion of he rue economic news, as follows: = + η, (3) where denoes he measured daa surprise. In his case, he mismeasuremen of he rue daa surprise is capured in he variable η. Using his proxy for he rue macroeconomic news, researchers ypically resor o esimaing he following equaion: s = γ + υ, (4) using an ordinary-leas-squares (OLS) regression. However, given he srucure above, he error erm from he esimaed equaion is υ = ε γ η, (5) which is negaively correlaed wih he righ-hand-side variable in he regression. This correlaion, of course, resuls in he bias in he regression esimae of γ. 6

7 To quanify he bias, we assume ha he rue macroeconomic news has a variance of σ and ha he measuremen error is mean ero condiional on he rue surprise ( E [ η ] ) and has a variance of σ η. We also assume ha he porion of he asse price movemen no explained by he macroeconomic surprise ( ε ) is mean ero condiional on he rue news and he measuremen error ( E [ ε η, ] ) and has a variance of σ ε. Under hese assumpions, he esimae obained by an OLS regression is: σ η ˆ γ OLS = γ γ. (6) σ + σ This esimae has he sandard downward bias (owards ero), which is he sandard resul in he presence of an error-in-variables problem. Based on his consideraion, we argue ha he ypical evensudy esimaion may undersae he influence of macroeconomic news on asse prices. A his poin, i is useful o noe ha we have considered wo forms of mismeasuremen of he macroeconomic news one based on noise in our reading of he marke s expecaions, and one based on noise in he release iself. Boh forms are capured by equaion (3), and hence he bias in he OLS esimaes applies o boh of hem. Neverheless, he inerpreaion of he resuls is differen depending on which of he wo sources predominanly accouns for he mismeasuremen. If he mismeasuremen is in erms of measuring he marke s expecaions, hen he OLS esimaes are acually missing par of he marke reacion. If insead he noise is conained in he acual daa, hen he marke is reacing by less, as i is doing he signal exracion problem and discouning he value of he released daa. In ha case, he OLS esimaes are an accurae measure of he rue (bu limied) marke reacion o he released daa. We are ineresed in discovering he marke reacion o he rue surprise, adjusing for he measuremen error from hese wo sources. There are several poenial soluions. One is o find an insrumen, somehing ha is correlaed wih he rue macroeconomic news bu uncorrelaed wih he measuremen error. Bu such insrumens do no exis, leaving he problem of esimaion unresolved. Anoher soluion is o improve he daa iself, for example by beer measuring marke expecaions. In ha regard, he emergence of economic derivaives may be useful, in ha hey may provide a more accurae and imely reading of marke expecaions. Sill, given all of he above consideraions, i is no clear ha we will ever have a fully accuraely measure he macroeconomic news. In his paper we ake an alernaive approach in which we aemp o address he issue hrough economeric echnique. We will ulimaely develop wo mehods ha help us resolve some of hese issues and allow us o beer undersand he linkages from economic news o asse prices and moneary policy expecaions. η 3. Idenificaion hrough Censoring The problem of error-in-variables ha we discuss above is, in fac, a problem of idenificaion. To see ha, consider he case of measuring he effec of a single daa release on a single asse price. In ha siuaion, we can compue only hree saisics: he variance of he asse price, he variance of he macroeconomic news, and he covariance beween hem. The problem is ha 7

8 hese momens are deermined by four underlying parameers: γ, σ, soluion is no idenified, or here is a coninuum of soluions. σ η, and σ ε. Thus, he Above we noed ha an insrumenal-variables approach is one way of solving he problem, if one were able o find an appropriae insrumen. Noe ha he availabiliy of such an insrumen basically solves he idenificaion problem. For a variable ω o be a valid insrumen, i mus be correlaed wih he rue news bu uncorrelaed wih he measuremen error, as follows: = β ω + κ, (7) The availabiliy of his insrumen adds hree pieces of informaion (he variance of ω, is covariance wih he measured news, and is covariance wih he asse price response) while only adding wo unknown variables ( β and he variance of κ ). As long as β is differen from ero, hese addiional condiions resolve he idenificaion problem. However, as noed above, we canno hink of an insrumen ha is valid in he circumsances sudied in his paper. In he absence of a valid insrumen, he quesion is wheher we can solve he idenificaion problem hrough some oher means. We will do so by developing a new echnique ha we label idenificaion hrough censoring. 3.1 The Case of One Macroeconomic Announcemen To demonsrae he mehodology, we firs assume ha here is only one macroeconomic announcemen a a given ime. One special feaure of macroeconomic announcemens is ha hey occur a pre-specified days. This is imporan, because i implies ha we can find a sample of oher days (or imes) a which he magniude of he surprise variable is exacly ero. When he variable is exacly equal o ero, i means ha is error-in-variables is ero as well. This censoring of he measuremen error will provide he idenificaion. 4 Formally, his siuaion can be described by he following equaion: s γ + ε = ε D + 1 D (8) where D is he se of days (or imes) on which he announcemens ake place. We are assuming ha no announcemens ake place he day before hose included in D. Under he assumpion ha he disurbance ε is homoskedasic, we can use he variance of he asse price observed a ime - 1 as addiional informaion in he idenificaion. In ha case, he following equaions hold: 4 This inuiion comes from Goldberger (1991) who argues ha he variance of he error-in-variables in survey daa depends on he sie of he announcemen. He used he following example: If you ask how many cigarees a person smokes in a day, a non-smoker will answer ero and ha reply has no error-in-variables whasoever. Bu someone who smokes a pack and a half a day will probably have a siable error. In oher words, he magniude of he error depends on he magniude of he reply, wih complee censoring of he error a ero. 8

9 var( cov( s s 1 var( s var( ) ), ) ) = = = = γ ε σ + + γσ σ σ σ σ ε η (9) This is a sysem of four equaions and four unknowns ha can be solved for he parameers. Mos imporanly, he sensiiviy of he asse price o he incoming news can be solved as follows: var( s ) var( s 1) γ =. (10) cov( s, ) This esimaor is in he spiri of Rigobon and Sack (004), in which he esimaor depended on he change in he variance relaive o he change in he covariance. Here, he change in he covariance is jus he covariance iself, since he macroeconomic surprise has no variance when i is censored. The above compuaions rely on he assumpion ha he srucural shocks in he asse price equaion ( ε ) are homoskedasic. This is a fairly srong assumpion, and one ha is no necessary. To derive an esimaor like (10), all we need is a predicion of wha he variance would have been like in he absence of he macroeconomic news. Thus, we can incorporae heeroskedasiciy o he degree ha i is predicable. In oher words, he idenifying assumpion is ha he variance of ε is predicable. For example, suppose we observe a release a 8:30, and as a conrol window we use a 30- minue inerval from he previous afernoon a :30. The above assumes ha he variance of ε around 8:30 is he same as ha around :30 on he previous day. Bu even on days of no announcemens, his does no seem o be he case. Insead, we require a much weaker assumpion ha he shif in he variance of ε on announcemen days is he same as he shif on non-announcemen days, or ε, 1,8:30 ε,,:30 ε,,8:30 ε, 1,:30 σ σ = σ σ. (11) This assumpion allows for he daa o have heeroskedasiciy over our sample, as long as ha heeroskedasiciy looks he same on announcemen and non-announcemen days. In his case, he esimaor (10) sill works if we replace he variances wih he shif in he variances. This is he assumpion ha we will employ in he empirical resuls below. This esimaor eliminaes he bias coming from error-in-variables ha affecs he ypical OLS esimaes. However, he esimaor is only as good as is idenifying assumpions. The wo main idenificaion assumpions needed are ha he errors-in-variable are classical and ha he variance of he asse prices is predicable (so ha we can make an accurae judgmen of wha he variance would have been in he absence of he macroeconomic surprise). Condiional on hose idenifying assumpions, he coefficiens from his procedure are accurae. However, if eiher of he wo main assumpions is violaed, he esimaes are biased. We will reurn o hese issues below. 9

10 3. The Case of Muliple Macroeconomic Announcemens The BEA, BLS, and oher governmen agencies would make our lives easier if hey released one saisic a a ime. Unforunaely, his is no he case. Because differen releases follow differen schedules, ofen muliple imporan releases will randomly coincide in boh he dae and ime. If his problem were jus limied o coincidence, we could deal wih i by simply eliminaing hose days wih muliple releases. Unforunaely, some of he daa releases always coincide wih one anoher. This is he case for hose repors ha include muliple saisics ha have marke influence. For example, he employmen repor involves he simulaneous release of nonfarm payrolls, he unemploymen rae, and average hourly earnings each of which are found o have an independen effec on markes. In he OLS framework, we can deal wih his simulaneiy by simply puing he muliple releases ino a single regression. We can also address his issue in he idenificaion-hrough-censoring approach. To achieve idenificaion in such circumsances, i urns ou ha we simply have o incorporae more han one asse price. For simpliciy, we will show his poin for he case of wo announcemens. Also, for simpliciy le us assume ha he srucural shock ε is homoskedasic. In his case, he model has he following srucure: s 1,, = = = γ 1 1, 1,, + γ + η + η 1,,, + ε (1) where he errors in measuring he rue surprises ( η 1, and η, ) are likely o be correlaed. Noe firs ha he idenificaion is los. The covariance marix of he asse price and he wo measures of macroeconomic surprises provides six equaions, and he variance of he asse price when here are no surprises provides a sevenh momen. Bu he model has nine unknown parameers: γ 1, γ, σ ε, σ, σ, σ η, σ, 1 1 η he covariance beween 1 and, and he covariance beween η 1 and η. The under-idenificaion is even more severe in he case of hree simulaneous announcemens. The soluion o he problem is o consider addiional asse prices. If we consider wo asse prices, we have he following sysem of equaions: s s 1,, 1,, = = = = γ γ 1,1 1,,11, + γ + γ, 1, +, + 1, η η,, 1,, + ε + ε 1,, (13) where he srucural shocks ε 1, and ε, are possibly correlaed and he errors in he macroeconomic surprises are, as before, also correlaed. We have now achieved idenificaion. The variance-covariance marix of he asse prices and he macroeconomic surprises on boh 10

11 announcemen and non-announcemen days provides 13 momen condiions. These are sufficien o solve for he 13 unknown parameers. 5 Wha delivers he idenificaion? I comes from he fac ha he noise conained in our measures of he macroeconomic announcemens has o be he same independen of he asse price we are considering. Tha resricion allows he incorporaion of an addiional asse o bring new informaion for he idenificaion. 3.3 Implemenaion of he Esimaor In he resuls below, we will include five differen asse prices and will allow for as many as hree simulaneous releases. (All deails are described in he nex secion.) This se-up implies ha our esimaor is always over-idenified. To esimae he parameer values, we use a GMM esimaor ha seeks o minimie he squared deviaions of he errors for each momen condiion. 6 I can be shown ha his esimaor is consisen, and ha he esimaes are asympoically normal. 4. The Esimaed Effecs of Macroeconomic Surprises This secion begins by describing he daa ha we use and some of he specific decisions made in implemening he various approaches. I hen provides some resuls from boh he sandard evensudy esimaor and he idenificaion-hrough-censoring approach. 4.1 Daa In he resuls ha follow, we measure he reacion of five financial variables o incoming macroeconomic news. The se of financial variables is inended o capure he behavior of nearerm policy expecaions as well as broader asse prices. Specifically, we include several near-erm ineres raes ha are very sensiive o moneary policy. Eurodollar fuures raes are probably he mos useful, liquid insrumen for ha purpose. We herefore include he raes on he second and fourh eurodollar conracs o expire which will reflec changes in moneary policy expecaions roughly a horions of 6 and 1 monhs ahead. 7 We also include he wo-year Treasury yield, which is very sensiive o he expeced pah of moneary policy beyond he horion covered by he eurodollar conracs, and he en-year Treasury yield. Lasly, we include he S&P 500 index. 8 5 Adding he second asse price brings six new momen condiions is variance and is covariance wih he oher asse price on boh announcemen days and non-announcemen days, and is covariances wih he wo measures of surprise on announcemen days and is variance on non-announcemen days) while adding only four new parameers ( γ,1, γ,, σ ε, and he covariance beween ε 1 and ε ). 6 So ha he relaive imporance of he momen condiions is no influenced by he uni of measure, we normalie he movemens in each asse price by heir sandard deviaion. The resuls, however, are expressed in erms of basis poins for yields and percenage poins for equiies. 7 The second conrac will have beween hree and six monhs o expiraion (wih an average of 4.5). I is ied o he hree-monh Libor rae, which will be sensiive o he expeced average funds rae over hose hree monhs (wih an average of 1.5). Adding ogeher hese averages yields 6 monhs. Similar calculaions yield 1 monhs for he fourh conrac. We exclude he firs and hird conracs because we fel ha much of heir informaion would be redundan. In addiion, we worried ha he variaion in he expiraion of he firs conac from 0 o 3 monhs migh be more problemaic (given insiuional deails such as he spacing of meeings). 8 We had hoped o include exchange raes as well, bu our inraday daa did no exend back far enough o make i a useful sample. 11

12 For all of hese asse prices, we use inraday daa. This feaure alone provides a siable improvemen over daily evensudy exercises. As noed above, wih inraday daa we can look a a narrow window around he ime of he release an inerval ha includes he influence of daa releases a a given ime bu excludes mos oher marke-moving evens. In effec, we are shrinking he sie of he error erm ε relaive o he influence of he daa. The inraday daa slices we consider are 30-minues long, beginning 5 minues before he ime of an announcemen o avoid any complicaions from variaion in he precise iming of he quoes or of he releases. The daa releases ha we consider all ake place a eiher 8:30 am, 9:15 am, or 10:00 am, giving us slices ha run from 8:5 8:55 am, 9:10 9:40 am, and 9:55 10:5 am. For equiies, unforunaely, we only have inraday quoes from when he sock marke is open, from 9:30 am 4:00 pm. Thus, we have o modify our slices accordingly. For he 8:30 am and 9:15 am releases, we use he change in he S&P index from he previous close o 9:55 am. For he 10:00 am release, we can use he same slice ha we use for he ineres raes. The conrol window ha we use in each case is a 30-minue window around :30 pm on he previous afernoon. We use he variance-covariance marix in ha window o predic wha he variance-covariance marix would have been in he even window in he absence of he daa release. The advanage of using he inraday quoes is shown in Figure 1, which focuses on he response of he wo-year Treasury yield o he nonfarm payrolls saisic from he monhly Employmen Siuaion repor from he Bureau of Labor Saisics. This is he daa release ha, in recen years a leas, has commanded he mos aenion in financial markes. As can be seen, here is a clear posiive relaionship beween surprises in he payroll release and he movemen in he wo-year yield. Moreover, his relaionship ighens if we use inraday daa insead of daily daa. We invesigae he marke reacions o 13 differen daa releases. Those releases are shown in Table 1, along wih some informaion abou he frequency of he release and he sample over which we have a measure of marke expecaions. We generally begin our sample in 1994, hough he sample for he Chicago Puchasing Manufacurers Index (PMI) has a shorer sample because we do no have a measure of marke expecaions unil December Our lis includes nearly all of he major macroeconomic indicaors ha are generally seen as significan marke movers Evensudy Esimaes Even hough i may have he shorcomings discussed above, we sill view he sandard evensudy regression as a very useful exercise, one ha can ell us a lo abou how asse prices and moneary policy expecaions are affeced by incoming daa. The above discussion simply cauions ha he resuling coefficiens may have some downward bias, hus undersaing he imporance of he daa. We implemen he evensudy regression per release, using he daa described above. The resuls are shown in Table. One of he primary findings from his exercise is ha moneary policy expecaions reac significanly o incoming daa. The expeced pah of he federal funds rae (as capured in 9 In all of he resuls ha follow, we discard hose days for which we have muliple releases. For he wo series from he employmen repor (nonfarm payrolls and hourly earnings), we always consider heir effecs ogeher, as discussed above. 1

13 eurodollar fuures) generally shifs up significanly in response o boh srong daa on growh (such as reail sales) and high daa on inflaion (such as core CPI). Overall, we find ha 1 of he 13 macroeconomic variables considered promp a significan reacion in he eurodollar fuures raes. 10 A second finding is ha he effec of he daa releases coninues o be siable even as he mauriy of he insrumen is exended. Indeed, he wo-year yield ofen moves by abou he same amoun as he eurodollar fuures raes, suggesing ha any influence on moneary policy is seen as being very persisen. 11 The sensiiviy of marke yields exends all he way ou o he en-year Treasury noe. The magniude of is reacion is large enough ha i suggess ha even disan forward raes are reacing o he news, as found by Gürkaynak, Sack, and Swanson (005). 1 A final observaion from he evensudy resuls has o do wih he reacion of equiy markes. The deachmen issue seems paricularly problemaic for equiies, as even he mos imporan daa releases (such as nonfarm payrolls) do no promp a significan marke reacion. 13 Bu looking a he response of equiies o all of he releases provides us wih an imporan clue abou why ha may be he case. The likely explanaion for his finding is ha a release such as non-farm payrolls conains offseing forces on equiy prices. On he one hand, a srong repor would sugges more srengh in he economy and hence beer earnings prospecs, which should boos equiy prices. On he oher hand, i also raises long-erm ineres raes, which should lower equiy prices. These wo forces offse one anoher, leaving he ne effec on equiy prices insignificanly differen from ero. A similar sory could be old for all of he demand-side indicaors, which all have no effec on equiies. If his were in fac he case, hen we should more clearly see a negaive response of equiy prices o daa ha is direcly abou inflaion. The reason is ha here is no offseing news in ha case higher inflaion implies ha raes will be higher bu no ha growh will be higher. Thus, equiy prices should fall. Indeed, his is precisely wha we find. Indeed, he S&P index reacs negaively and significanly o posiive surprises in core CPI, core PPI, and hourly earnings every single inflaion measure considered. 14 Overall, he evensudy regressions provide an ineresing paern of marke responses o differen ypes of incoming news. Neverheless, he R-squared saisics from he regressions are relaively low, generally ranging from 0.15 o In oher words, he evensudy regressions ypically accoun for only a small porion of he variance of he marke reacions, even if we focus on he 10 Oher sudies, including Fleming and Remolona (1997) and Baldui, Elon, and Green (001), and Gürkaynak, Sack, and Swanson (005), have found ha marke ineres raes respond significanly o a wide range of macroeconomic daa releases. 11 A similar resul was found by Kohn and Sack (005). They noed a similar persisence in he response o FOMC saemens and inferred ha hose saemens may be seen as conveying informaion abou he sae of he economy in addiion o informaion abou he near-erm direcion of policy. 1 Tha paper looked explicily a disan forward raes and found ha hey ofen responded o daa in he same direcion as near-erm forward raes. The auhors developed he case ha his response refleced he fac ha long-erm inflaion expecaions in he Unied Saes are variable, a case srenghened by he fac ha similar sensiiviy is no observed in he U.K. perhaps because of is explici inflaion arge. 13 By conras, equiies do appear o reac significanly o moneary policy shocks, as shown by Bernanke and Kuner (003). 14 To our knowledge, his is no an empirical fac ha has been emphasied in he lieraure o dae. Fair (??) finds a posiive reacion of equiies o inflaion news. McQueen and Roley (1993) find a reacion ha differs across differen saes of he business cycle, wih negaive responses for some variables in some saes. 13

14 movemens in he 30-minue window brackeing he announcemen. This las observaion is he area in which we will see some improvemen under he new esimaor. 4.3 Idenificaion-hough-Censoring Esimaes Table 3 shows he esimaed responses under he idenificaion-by-censoring (IC) approach. Broadly speaking, he paerns of he responses are he same as in he evensudy (ES) exercise: sronger-han-expeced readings on growh or higher-han expeced readings on inflaion end o boos marke ineres raes. The sock marke response o incoming daa on growh is mixed and ofen insignifican, while i reacs negaively o incoming daa on inflaion. The primary difference beween he ES and IC approaches is he magniude of he marke responses. The IC coefficiens are ofen wo or hree imes as large as he ES coefficiens. This finding suggess ha he problem of deachmen is, o a large exen a leas, associaed wih he mismeasuremen of macroeconomic news. For example, a one-sandard-deviaion upward surprise o core CPI (nearly 0.1 percenage poin) is esimaed o increase yields 6 o 9 basis poins, raher han he response of o.5 basis poins found under ES. I is worh considering again how o inerpre his difference. The IC measure is capuring he marke response o a rue core CPI surprise, one ha marke paricipans are convinced has no measuremen error in i and one for which he marke expecaions are measured perfecly. The rue CPI release may be discouned if i is seen as conaining measuremen error (e.g., a higher-han-expeced reading driven by a single componen, such as he price of lodging away from home), or is esimaed effec under he ES may be downward biased if he marke s expecaions were measured improperly. One implicaion of he resuls is ha moneary policy expecaions and asse prices may be more sysemaically relaed o incoming daa han found under he ES approach. This conclusion accords wih our undersanding of moneary policy from he (lower-frequency) macroeconomic lieraure, including he view ha one way policy has been effecive over he pas decade is by sysemaically responding o changes in economic condiions. Our resuls provide a highfrequency version of ha conclusion. One issue is ha he resuls appear oo good in some sense. The esimaed amoun of noise in he daa announcemens, a saisic ha is also idenified in he IC procedure, ends o be very high for many of he releases. (This paern, of course, is direcly relaed o he fac ha he IC coefficiens are several imes larger han he ES coefficiens.) For example, he resuls sugges ha 31% of he variaion in he non-farm payrolls surprise is due o noise, while 77% of he variaion in he core CPI release is due o noise. I is somewha hard o grasp jus how much noise one would expec relaive o some acual ruh ha we never observe. However, some of he readings from Table are clearly implausible. For example, we doub ha 94% of he measured surprise associaed wih he ISM index is acually noise. The exen of he esimaed noise may raise some quesions abou wheher he idenificaion assumpions hold. We migh be paricularly concerned abou our effors o predic wha he variance of he asse prices would have been in he absence of he macroeconomic surprise, as needed in he IC procedure. Noe ha he esimaes of boh he sensiiviy of he marke response (γ ) and he amoun of noise in he surprises ( σ ε ) end o increase in he shif in he variance of 14

15 he asse price beween non-announcemen days and announcemen days. Hence, if we are underesimaing he variance ha would be presen in he absence of a macroeconomic announcemen, we would be overesimaing boh of hese parameers. One reason o suspec his paern is ha he macroeconomic surprises measured on he righhand-side of our equaions ofen coincide wih he release of oher daa ha migh move markes. 15 For example, he employmen repor no only includes he curren monh surprise o nonfarm payrolls, bu also revisions o payrolls in he wo previous monhs. Thus, even in he absence of a surprise regarding he curren monh payroll, one migh expec more marke volailiy han on a non-announcemen day because of he possible marke reacion o his oher informaion. If his were he case, he IC esimaes presened in Table 3 may have some upward bias. Bu noe ha his upward bias exiss because he daa release is acually more meaningful han capured by he surprise measure on he righ-hand-side of he equaion. Thus, i sill likely reflecs ha he macroeconomic news is more imporan han accouned for by he evensudy approach. We migh herefore wan o hink of an esimaor ha can beer incorporae ha addiional informaion. 5. A Principal Componens Exercise This las consideraion leads us in he direcion of a compleely differen bu complemenary approach. The IC esimaor was developed ou of concern ha he macroeconomic surprise variable may be measured poorly, inroducing oo much variaion ino ha measure. Bu perhaps he bigger problem is he opposie one ha he righ-hand-side variable does no capure enough of he surprise in a given daa release. This would be he case if he daa release conained marke-moving informaion oher han ha represened in he surprise measures considered above. To be sure, mos daa releases are complicaed and convey many pieces of informaion. I may be difficul o deermine a macroeconomic surprise measure ha capures all of ha informaion. 16 An alernaive approach ha avoids his difficuly is o le he financial marke daa iself deermine he daa surprise. Specifically, we again consider he movemens of he four ineres raes and equiy prices in he 30-minue window around a given release. Our idenificaion assumpion is ha he primary even driving he markes in hose windows is he daa release an assumpion ha is cerainly plausible for he narrow window ha we consider around he release. We are no ruling ou ha oher evens ake place in ha window; bu if here does appear o be one common even, we will assign is effecs o he daa release. The approach ha we use o implemen his assumpion is principal componens. For a given release, we sack he marke reacions ino a marix wih one row per observaion and one column for each asse price (he second and fourh eurodollar conracs, he wo-year Treasury yield, he en-year Treasury yield, and he S&P index). The principal componens exercise deermines a se 15 In addiion, he announcemen iself (even if i is on expecaion) could resul in some variance of he asse price, because i would presumably reduce uncerainy and cause invesors who had differen expecaions o adjus heir posiions and views. 16 Above we have he example of he payrolls release and he relevance of concurren revisions o pas monhs payrolls releases. Oher examples are quie ha reail sales ex-auos coincides wih oal reail sales (including auos), capaciy uiliaion coincides wih indusrial producion, and so on. 15

16 of orhogonal facors F (same dimensions as X) ha are linear combinaions of he original daa series: F = X A, (14) where A A = I. As a resul, he variance-covariance marix of he responses of he financial variables is given by F F = A Σ A, where Σ is a diagonal marix conaining he variances of he facors. The facors are ordered by he magniude of heir variances (wih he facor wih he highes variance lised firs). In his sense, he firs facor explains as much of he variaion across he observable variables as possible, he second facor capures as much addiional variaion as possible, and so on. 17 The loadings of he financial variables on each of he facors is given by he inverse of he marix A, or A. This approach is more general han he IC esimaor. I does no require he wo idenifying assumpions needed in ha case, and i can capure a broader se of informaion han measured by he surprise variables included in he IC and ES approaches. The poenial cos, however, is ha i could accidenally include some variaion no ruly associaed wih he daa release. A finding ha here is a srong co-movemen in he asse prices over he 30-minue window around he daa release would boos our confidence ha he procedure is picking up he effecs of ha release. As repored in Table 4, i urns ou ha a single facor explains he vas majoriy of he marke reacion o each release. This facor ypically accouns for 90% o 95% of he variaion in he asse prices in he 30-minue window. 18 I is his movemen ha we associae wih he daa release, since he release is presumably he dominan marke even in he window. 19 In his case, he firs principal componen provides a measure of he rue daa surprise, one ha incorporaes all of he marke-sensiive news included in a given release. As we would expec, hese daa surprises are somewha correlaed wih he survey-based surprises used above. The able shows ha he survey-based surprises accoun for as much as 50% of he variaion in he firs principal componen. Thus, clearly he surprises used in he ES exercise are an imporan componen of he oal news around a daa release. However, hey are no a complee measure of he marke-sensiive news conained in he release, as suggesed by he addiional (unexplained) variaion in he firs principal componen. Figure presens an example, ha for (ex-auo reail sales). On he horional axis is he surveybased surprise used above, and on he verical axis is he firs principal componen (normalied in a way o make i mos comparable o he reail sales release). Again, he wo measures are clearly relaed, bu hey are far from idenical. The PC-based surprise measure has more variaion han can be explained by he survey-based surprise measure, presumably capuring he addiional informaion in he release. 17 When we apply his echnique o he above daase, we normalie each variable by is sandard deviaion. 18 The able shows he variance of he firs facor relaive o all of he oher facors. Bu ha saisic is nearly idenical o he fracion of he variance of he marke ineres raes explained by he firs facor. 19 For comparison, if we conduc he same exercise in he non-even window considered above (he 30-minue window brackeing :30), we find ha he firs facor explains only 80% of he variance of he asse prices. Thus, i does appear ha he daa release window conains an even ha causes a co-movemen in he asse prices ha is larger han ha observed a oher imes. 16

17 The able also repors he loadings of he various asse prices on he PC-based surprise measure. For ease of inerpreaion, we have normalied each PC measure o have a uniary sandard deviaion, jus as we did wih he survey-based surprises used in he ES exercise. The coefficiens reain many of he ineresing paerns observed in he earlier resuls. The marke ineres raes considered have a siable response o he macroeconomic news, suggesing ha he news is affecing he expeced pah of moneary policy. Those responses are ypically also observed a longer-erm mauriies. One puling aspec of he resuls is ha he equiy marke no longer appears o have as large of a negaive reacion o incoming daa on inflaion. I is rue ha he firs facor explains a larger fracion of equiy price movemens for he inflaion-relaed daa releases han for oher releases, bu he response relaive o he ineres-rae response is smaller han in he above resuls. Insead, he facor analysis essenially finds a separae facor ha drives much of he movemens in equiy prices. We wonder wheher his finding in par reflecs ha we are forced o use a wider window for he equiy price movemens (17.5 hours insead of 30 minues!), which considerably weakens he idenificaion assumpion used in he PC exercise. Perhaps he mos imporan aspec of he PC exercise is is usefulness for assessing he amoun of variaion in yields ha can be aribued o macroeconomic daa. The PC exercise indicaes ha markes are much more sensiive o macroeconomic daa releases han suggesed by he ES approach. This is a similar finding as he IC esimaor used above. However, in his case he reason is no only ha we are accouning for he measuremen error in he survey-based surprise measure, bu also because we are accouning for any oher relevan informaion in he release. One useful aspec of he PC approach is ha, unlike he case for he IC esimaor, we recover a ime series of he rue macroeconomic news, as discussed above. This allows us o cumulae he effecs of each release on a paricular asse price. Figure 3 shows he cumulaive effecs of each he daa releases on he wo-year Treasury yield, where each line represens an individual release. (For example, one line represens he effecs of all reail sales releases over our sample.) The poin of he figure is no o focus on any paricular line, bu o ge a general sense of he oal variaion explained under he wo approaches. As can be seen, he movemens explained by he releases under he ES exercise are much smaller han hose under he PC exercise. Table 5 conains some saisics ha furher quanify he variaion explained under he wo approaches. I compues he absolue value of he changes aribuable o each release, expressed as basis poins per year. By his measure, he mos influenial daa release, by far, has been he employmen repor. Oher influenial releases include reail sales, he ISM index, he CPI, and he PPI. More imporanly, he PC measure accouns for much more variaion han he sandard evensudy approach. (This, of course, is simply a differen way of expressing ha he R-squared saisic from he regression increases significanly.) Indeed, his is he case for every single daa release considered. We can sum hese saisics across all of he releases o obain a measure of he oal variaion explained by incoming macroeconomic daa (or a leas by our releases). By ha measure, he PC approach has accouned for nearly wice as much of he variaion in he wo-year yield han he ES approach. Thus, he new mehodology makes an imporan sep owards beer undersanding he oal influence of macroeconomic daa on asse prices and moneary policy expecaions. 17