information aggregation around macroeconomic announcements: revisions matter

Contents lists available at ScienceDirect

Journal of Financial Economics

Journal of Financial Economics 101 (2011) 114–131

0304-40

doi:10.1

$ Thi

comple

Berkele

their co

insights

Jonatha

Hender

particip

the Uni

Empres

Rice Un

Dame,

2008 m

support

All erro� Tel.

E-m

URL

journal homepage: www.elsevier.com/locate/jfec

Information aggregation around macroeconomic announcements:Revisions matter$

Thomas Gilbert �

Michael G. Foster School of Business, University of Washington, PACCAR Hall, Box 353226, Seattle, WA 98195-3226, USA

a r t i c l e i n f o

Article history:

Received 6 May 2010

Received in revised form

27 September 2010

Accepted 19 October 2010Available online 24 February 2011

JEL classification:

G14

E44

Keywords:

Macroeconomic announcements

Revisions

Information precision

Price discovery

5X/$ - see front matter & 2011 Elsevier B.V.

016/j.jfineco.2011.02.013

s paper is based on the third chapter o

ted at the Haas School of Business at the Univ

y. I am deeply indebted to Rich Lyons and C

ntinuous help and support. Miguel Palacios p

, and I also thank two anonymous refere

n Berk, Greg Duffee, Nicolae Garleanu, Bo

shott, Christophe Perignon, Ed Rice, Johan W

ants at the University of California at Berkeley

versity of Oxford, IESE (Instituto de Estudio

a) Business School, HEC (Hautes Etudes Co

iversity, the University of Washington, the U

the Federal Reserve Board, and the French F

eeting for helpful comments and suggest

of the Dean Witter Foundation is gratefu

rs are my own.

: +1 206 616 7184.

ail address: [email protected]

: http://faculty.washington.edu/gilbertt/

a b s t r a c t

I show that an empirical relation exists between stock returns on macroeconomic news

announcement days and the future revisions of the released data but that this link

differs across the business cycle. Using three major macroeconomic series that undergo

significant revisions (nonfarm payroll, gross domestic product, and industrial produc-

tion), I present evidence that daily returns on the Standard & Poor’s 500 index and

revisions are positively related in expansions and negatively related in recessions. The

results suggest that revisions do matter, i.e., that investors care about the final revised

value of a macroeconomic series, that they infer accurate information from the release

of the preliminary inaccurate report, and that the more precise information is

aggregated into prices on the day of the initial announcement. The results are

consistent with the predictions of rational expectations trading models around public

announcements combined with well-established empirical results on the asymmetric

interpretation of information across the business cycle.

& 2011 Elsevier B.V. All rights reserved.

1. Introduction

Macroeconomic announcements undergo significantrevisions in the months and sometimes years following

All rights reserved.

f my dissertation

ersity of California,

hristine Parlour for

rovided invaluable

es, Bob Anderson,

b Goldstein, Terry

alden, and seminar

, McGill University,

s Superiores de la

mmerciales) Paris,

niversity of Notre

inance Association

ions. The financial

lly acknowledged.

their initial release. The revisions represent the additionof new information in the statistical agencies’ reports thatwas not available to them at the time of the originalannouncement. Rational investors should take this impre-cision into account when analyzing the initial report andact accordingly, filtering out the noise and responding tothe information conveyed by the preliminary announce-ment about the variable’s correct (revised) value and notonly its preliminary (unrevised) value. However, thisinference process has been overlooked by the previousliterature analyzing the impact of macroeconomicreleases on asset returns.

McQueen and Roley (1993), Fleming and Remolona(1997, 1999), and Andersen, Bollerslev, Diebold, and Vega(2007), among many others, test whether the release ofpublic macroeconomic information moves asset prices.While these papers have provided strong evidence thatannouncement surprises do move asset returns, theexplanatory power of the event study regressions istypically very small. Based on the magnitude of therevisions that macroeconomic variables undergo, I posit

www.elsevier.com/locate/jfec

dx.doi.org/10.1016/j.jfineco.2011.02.013

mailto:[email protected]

mailto:http://faculty.washington.edu/gilbertt/a4.3d

dx.doi.org/10.1016/j.jfineco.2011.02.013

T. Gilbert / Journal of Financial Economics 101 (2011) 114–131 115

that markets not only react to the contemporaneoussurprises but also focus on the information conveyed bythe initial release about the true underlying value thatwill ultimately be released much later. Because revisionsare almost guaranteed to occur, the inference that inves-tors make from the newly released data about the revisedvalue must be taken into account if we are to fullyunderstand the information aggregation process thatoccurs around public announcements.

In this paper, I ask whether the inherent inaccuracy ofmacroeconomic announcements matters, i.e., do revisionsmatter? Rational expectations models of trade aroundpublic announcements such as Kim and Verrecchia(1991), among others, yield the following predictions.First, the release of more precise announcements shouldhave a higher price impact (Hautsch and Hess, 2007).Second, announcement-day returns should be negativelyrelated to the future revision of the released announce-ment. The latter hypothesis stems from the fact that, inequilibrium, an above-average signal moves prices upeven though the revision is expected to be negative.1 Bytaking into account this inference process, which happenswhen imprecise public announcements are made, I testwhether markets care about the true value of initiallyinaccurate public information releases.

Empirically, I show that a strong relation existsbetween announcement-day returns and the future revi-sions of the released macroeconomic data but that thislink differs across the business cycle. Returns and revi-sions are positively related in expansions and negativelyrelated in recessions. In terms of magnitude, an expectedupward revision of 100,000 employees to the month’snonfarm payroll announcement by the Bureau of LaborStatistics (BLS) leads to an 18 basis point increase in theStandard & Poor’s (S&P) 500 index on announcement dayduring expansions and a 77 basis point decrease in theindex during recessions.2 In dollar terms, this announce-ment error of 100,000 jobs is equivalent to an averagedaily change in the market value of the index of $10billion in expansions and $43 billion in recessions.Furthermore, the addition of the revisions as an indepen-dent variable significantly increases the R2 of the eventstudy regressions explaining announcement-day returns(almost double in recessions).

Federal statistical agencies such as the BLS signifi-cantly revise macroeconomic variables during the monthsand even years following the initial announcements. Earlyrevisions (one to three months) mainly stem from reportsor surveys that are submitted late by firms or individuals.Each month, the agencies extrapolate from the sample ofreports received in order to obtain an economy-wideestimate. Annual revisions (one to five years) arisebecause of updates of the economy-wide benchmarks,such as the total number of workers or firms. Both earlyand annual revisions reflect the addition of new

1 The presence of noise in the signal makes the upward change in

price smaller than it would have been otherwise.2 In my sample, this is roughly equal to a one standard deviation

shock to nonfarm payroll revisions, which happens to be roughly equal

to the average announced value of monthly nonfarm payroll changes.

information in the agencies’ reports that was not availableto them at the time of the initial announcement. Infre-quent methodological revisions are put in place by sta-tistical agencies so as to improve the measurement ofeconomic trends. Their impact on the choice of datavintage is discussed in Section 4.

I define the sample-based early revisions of a macro-economic announcement as the difference between thenumber available one to two months after the initialrelease and the original announced value. By extrapolat-ing this definition, I define the final revision as thedifference between the number available after as manyrounds of revisions as possible (in May 2010 when theanalysis was conducted) and the original announcedvalue. The initial announcement surprise is defined asthe difference between the realization of the announce-ment and the market median (consensus) expectationcollected shortly prior to the announcement. It is impor-tant to note that final revisions of nonfarm payroll, grossdomestic product (GDP), or industrial production are ofthe same order of magnitude as initial surprises, both interms of mean and standard deviation.

I follow the previous literature in labeling a macro-economic announcement that is above expectations asgood news if it indicates an increase in economic activity.For instance, a positive surprise in nonfarm payroll(announcement of 90 thousand versus an expectation of75 thousand) or a positive surprise in GDP is defined asgood news. Similarly, an unexpected increase in unem-ployment (announcement of 9.8% versus an expectationof 9.6%) is defined as bad news.

Standard rational expectations models of trade aroundpublic announcements yield unambiguous predictionsabout the relation between expected future revisions ofthe information releases and announcement-day returns.The relevant theoretical literature is discussed in Section3. In these, rational investors know that public signals arenoisy estimates of the true underlying state. They there-fore take into account the noise in the releases: A releasethat is above expectations (above average) is expected tobe revised downward in the future, and vice versa for arelease that is below expectations. However, in equili-brium, the signals enter prices with the ‘‘right’’ (positive)sign: Good signals lead to an increase in prices, and viceversa, even though the increase is not as large as theactual release would suggest because rational investorsexpect the downward revision. The natural conclusion ofthese rational models in which good news moves pricesup is that announcement-day returns and future datarevisions are negatively related.

An increase in nonfarm payroll can lead to an increaseor a decrease in stock prices depending on the interpreta-tion of the information. For example, a monetary responseframework implies that, following good news, the FederalReserve might increase interest rates to avoid overheatingand control inflation, which would be bad news for stocks.Furthermore, such interpretation could vary across thebusiness cycle: In the depth of a recession, an increase inpayroll is typically good news because it indicates thatfirms are hiring in expectation of increased consumerdemand. As a result, whether good news moves stock

(footnote continued)

behind the noise created by uninformed investors who trade solely

T. Gilbert / Journal of Financial Economics 101 (2011) 114–131116

prices up or down, i.e., whether good news is truly goodnews or actually bad news, is inherently an empiricalquestion. Boyd, Hu, and Jagannathan (2005), amongothers, show that the release of good unemploymentnews, i.e., a decrease of the unemployment rate, leads toa decrease in stock prices. The relevant empirical litera-ture is discussed in Section 2. This ‘‘good news is badnews’’ result holds on average because the economy is inan expansionary cycle most of the time even though therelation is reversed in recessions: Good news is goodnews for stocks in bad times.

To test whether revisions matter and whether marketscare about the final revised value of initially inaccuratedata, I combine the above predictions of rational expecta-tions trading models with the empirical results on theinterpretation of macroeconomic information. The formerliterature predicts that announcement-day returns andfuture revisions should be negatively related. However,this is true only if good news moves prices up. The latterliterature shows that good news does move stock pricesup in recessions and down in expansions. As a result, twotestable hypotheses emerge: In expansions, announce-ment-day returns and future revisions should be posi-tively related, and the relation should be negative inrecessions.

The theoretical and empirical evidence presented inthis paper highlights an overlooked information aggrega-tion mechanism around public announcements. It sug-gests that analyzing the relation between surprises, asdefined above, and announcement-day returns might notbe sufficient to claim that the release of a particularmacroeconomic series does or does not move stockreturns. If revisions are economically significant andinvestors care about the true underlying value of nonfarmpayroll or GDP, for instance, then we should observe anempirical link between returns and future revisions. Thepresence of this link is consistent with the joint hypoth-esis that investors do care about the final value of a seriesand that some series, such as industrial production, domove stock returns when released even though no directrelation between surprises and returns was shown by theprior literature (Andersen, Bollerslev, Diebold, and Vega,2007, among others).

The above information aggregation mechanism, whichleads to accurate information being impounded intoprices on days that inaccurate information is releasedgiven that investors know that initial releases are inaccu-rate, can stem from two empirically indistinguishablereasons: (1) investors learn about the revision processesthat are used by statistical agencies and build a forecast ofthe final value using the initial release as one of manyinputs in the forecasting model or (2) investors haveprivate information about the true (revised) value andtrade only on this information on announcement day.Because the information analyzed in this paper is macro-economic information and not firm-level information, thelatter story seems unlikely.3 However, it is rational for

3 In a model such as Kyle (1985), the insider who knows the true

value trades on announcement day since he can hide his information

investors to care about the true final value of a macro-economic signal when making asset allocation decisions.This therefore implies that they exert effort to learn howstatistical agencies compile their data, and hence uncoverwhy revisions occur. I posit that investors account forsuch potential errors when preliminary reports arereleased, which leads them to forecast the revised finalvalue by using the initial inaccurate announcement as oneof many potential inputs in the forecasting model. Basedon this intuition and the theoretical framework discussedabove, we should therefore observe an empirical relationbetween announcement-day returns and future revisions.

There are two important points to note about thepaper. Although the results presented point to the factthat announcement-day returns can be used to forecastthe final value of the released data, the paper is not aforecasting exercise.4 In addition, the paper is not abouttrading strategies that may or may not be able to exploitthese results. The paper is focused on testing rationalexpectations hypotheses about the effect of inaccuracy inpublicly announced macroeconomic signals by buildingan empirical relation between revisions and announce-ment-day returns.

The remainder of the paper is organized as follows.After reviewing the related literature in the next section, Idevelop testable hypotheses using existing rational mod-els of trade and well-established results on informationinterpretation in Section 3. In Section 4, I use nonfarmpayroll as an example to describe the process of revisionsto macroeconomic series and define all variables used inthe empirical analysis. In Section 5, I analyze the informa-tion aggregation process around macroeconomicannouncements by pinning down the link between revi-sions and stock returns for nonfarm payroll, GDP, andindustrial production. Finally, I summarize and concludein Section 6.

2. Related literature

To the best of my knowledge, this is the first paper thattheoretically and empirically links returns to revisionsand thereby sheds light on the information aggregationprocess surrounding macroeconomic releases. However, itis closely related to two strands of literature. The firstaddresses the question of information aggregation (pricediscovery) and more precisely the impact of macroeco-nomic announcements on asset returns, and the secondtackles the issue of the forecastability of data revisions.

2.1. Macroeconomic announcements and asset returns

Andersen, Bollerslev, Diebold, and Vega (2007) analyzethe impact of announcement surprises of 20 monthly

based on the inaccurate public announcement. While the Kyle-type

insider also breaks his trades ahead of the public signal, this effect is not

the focus of this paper.4 Stock and Watson (2002) build a dynamic factor model using 215

predictors to forecast eight monthly macroeconomic time series.


macroeconomic announcements on high-frequency S&P500 futures returns. They find that only nonfarm payroll,new home sales, net exports, and inflation systematicallymove prices. Similarly, Flannery and Protopapadakis(2002) find that only inflation and money supply movedaily aggregate stock returns.5 Adams, McQueen, andWood (2004) confirm that inflation moves intradailyreturns of size-based portfolios.6 However, none of theseevent studies take into account the subsequent revisionsto the macroeconomic announcements, which couldexplain the low R2 of their tests and the surprisinglylow number of macroeconomic series that appear to moveprices.

McQueen and Roley (1993) and Boyd, Hu, andJagannathan (2005) find stronger relations betweenmacroeconomic announcements and stock returns afterconditioning on the business cycle: In good times, goodnews is bad news for stocks. That is, a decrease inunemployment during expansions pushes equity pricesdown. Andersen, Bollerslev, Diebold, and Vega (2007)follow this strategy and find much stronger results inboth their expansion and recession samples, even thoughthe latter contains only 21 observations. While thesepapers do not account for revisions, the conditionalinterpretation of the macroeconomic information isimportant for the validity of my empirical tests.

Savor and Wilson (2010) show that stock marketexcess returns are significantly higher on inflation,employment, and interest rate announcement days com-pared with all other days. Using an endowment economywith Epstein-Zin preferences, they argue that investorsshould demand compensation for facing information risk,even public information risk.7 However, they do not takeinto account the size of the announcement surprises andtheir subsequent revisions, which clouds the interpreta-tion of their results.

2.2. Predictability of macroeconomic revisions

The macroeconomics literature has a significant bodyof literature analyzing revisions in what is known as the‘‘news versus noise’’ debate. Under the noise hypothesis,the initial announcement is an observation of the truefinal number but is polluted with error, or noise, that isuncorrelated with the true value. Investors therefore facea filtering problem when observing the release as they tryto build an optimal estimate of nonfarm payroll, forinstance. Under the news hypothesis, the initialannouncement is optimal given all the information avail-able at that time and subsequent revisions reflect thearrival of new information. In this second case, the error

5 Stronger results have been found between announcement sur-

prises and foreign exchange or fixed income: Fleming and Remolona

(1997, 1999), Balduzzi, Elton, and Green (2001), Andersen, Bollerslev,

Diebold, and Vega (2003), and Green (2004), to cite just a few.6 Cenesizoglu (forthcoming) analyzes the impact of macroeconomic

news on both size- and book-to-market-sorted portfolios and finds

similar results.7 This result on public information risk complements Easley,

Hvidkjaer, and O’Hara (2002), who show that private information risk

is priced.

in the signal is correlated with the true value butuncorrelated with the initial release.

Aruoba (2008), following Mankiw, Runkle, and Shapiro(1984) and Mork (1987), presents the most up-to-dateanalysis of macroeconomic variables’ vintages. He showsthat revisions are not well behaved: They do not have azero mean, they are as volatile as initial announcements,and, furthermore, they are predictable using public infor-mation, not including prices, available at the time of theannouncement. Krueger and Fortson (2003) analyze therelations between the various rounds of revisions toemployment data. They find that the initial releases canbe used to forecast the future benchmark releases. How-ever, subsequent revisions are informative in the sensethat they incorporate new information in the govern-ments’ reports that was not available at the time of theinitial announcement. Faust, Rogers, and Wright (2005)show that revisions to GDP announcements are largelyunpredictable, which is in support of the news hypothesis.

Even though none of the above papers explicitly linkthe revisions to asset returns, Mankiw, Runkle, andShapiro (1984) note that, by including the Treasury billinterest rate and the rate of return on the stock market,they can marginally improve on the forecasting of finalrevised values of the quarterly growth of the moneystock.8 Faust, Rogers, and Wright (2005) also include thegrowth rate of equity prices and the three-month interestrate to forecast revisions to GDP, neither of them beingsignificant.

To summarize, the somewhat contradictory evidenceon the forecastability of revisions presented in the papersabove suggests that data revisions are probably a mixtureof news and noise. Importantly, the results of my paperand their interpretation do not depend on the outcome ofthis debate because I am focusing on the informationaggregation process on announcement day instead of thetime-series behavior of data vintages. Nevertheless, I doindirectly show that one can use announcement-daystock returns to significantly improve the forecastabilityof revisions but that this exercise is business cycledependent.

3. Theory: linking revisions and stock returns

In this section, I derive the theoretical relationbetween the revision of public signals and announce-ment-day returns in the classical rational expectationsframework of Kim and Verrecchia (1991). I then combineit with well-established empirical results about informa-tion interpretation throughout the business cycle in orderto develop testable hypotheses linking revisions andreturns.

8 Orazem and Falk (1989) point out that traditional announcement-

effect event studies can lead to misleading results if the government’s

announcements do not reflect all available market information and if the

market does not respond to the announcement itself but rather to the

information conveyed by the announcement about the (true) revised

value.

Fig. 1. Timing of the standard information aggregation model around public announcements. Si is the private signal to investor i. qit is the demand of

investor i at time t. pt is the price in period t. SG is the public signal. n is the final payoff of the risky security. ei and e are noise terms. For ease of

presentation, I have split demands and prices even though they are set simultaneously in a general equilibrium setting.


3.1. Standard model of public announcements

Rational expectations models with asymmetric infor-mation and public signals take many different forms, butall share a common structure. I use as main reference themodel of Kim and Verrecchia (1991) and Fig. 1 shows asimplified timeline of its information structure.9 In themodel, private signals precede the public signal and allsignals S are modeled as truth plus noise:

S¼ nþe, ð1Þ

where n is the true end-of-period payoff of the riskysecurity that is traded and e is an independent randomvariable with mean zero.10 A macroeconomic announce-ment made by a government, or private, agency is labeledas SG and, in such a setting, the public signal’s revision Rev

is defined as the difference between the final (true) valueand the initial announcement:

Rev� n�SG ¼�e, ð2Þ

which I relate to the equilibrium price change, p2�p1, asshown in Fig. 1.

Kim and Verrecchia (1991) present a three-date com-petitive trading model with a continuum of risk-averseand differentially informed agents. In the first period,agents receive independent noisy private signalsSi ¼ nþei of the end-of-period payoff n. In the secondperiod, a noisy public signal SG ¼ nþe is released. Agentswith precise private information (low signal variance)update their beliefs less if the public signal has a lowerprecision, but poorly informed agents update more. Thisdifferential updating generates trading volume at the timeof the public announcement. In equilibrium, an investor’sdemand in the second round of trading decreases in theprecision of his private information. Because he updatesless, a well-informed agent trades less than a poorlyinformed one when a noisy signal is released.11

9 Other similar models of trade around public announcements are

Grundy and McNichols (1989) and Kandel and Pearson (1995), among

others.10 Empirically, such additive information structure implies that

announcements are unbiased forecasts of the true final payoff.11 A feature of the equilibrium demands is that a well-informed

investor does not exercise his information advantage when the public

announcement is released. He knows how inaccurate the public signal is

and he, in addition, knows that the poorly informed investor is updating

using inaccurate information. Nevertheless, he does not take advantage

of his informational advantage in the second period because he has

already done so in the first period.

In terms of equilibrium prices, the model has thefollowing implication: The price change p2�p1 is nega-tively related to the future revision in the public signaln�SG. In the first period, optimal demands are based onthe private signals, which are noisy signals of n. Onaverage, the price p1 is therefore positively correlatedwith n. In the second period, the public signal enters theprice with the ‘‘right’’ signal: A positive signal moves theprice up and a negative signal moves the price down. Theprice p2 is therefore positively correlated with SG. Conse-quently, the price change p2�p1 is positively related toSG�n¼�ðn�SGÞ ¼�Rev.12

The economic intuition for the negative relationbetween price change and subsequent revision is givenby the information structure. Investors know that if thepublic signal is, for example, above average (that is, aboveexpectations), then it is likely to be partially due to apositive noise shock and, hence, that the revision will bedownward.13 Nevertheless, in equilibrium, the positivepublic signal will move the price up. These price dynamicsare independent of the price formation mechanism aslong as the public signal is unknown to all participantsbefore its release. Kim and Verrecchia (1991) have acompetitive model, but Pasquariello and Vega (2007)have a strategic two-date trading model with publicinformation, which has the same equilibrium prediction:Returns and revisions are negatively related.14

3.2. Positive news can be good or bad for stocks

Analyzing the link between macroeconomic informa-tion announcements and aggregate equity returnsdepends on the interpretation of the information, whichcould itself be linked to the business cycle. Informationfor equity markets can be split into two parts: news aboutcash flows and news about discount rates. Campbell andVuolteenaho (2004) follow this intuition in order to splitbeta into a cash flow beta and a discount rate beta.Positive news about future cash flows leads to an increasein price (numerator of the present value formula), but thisincrease can be offset if the market raises the discount

12 The negative relation between revision and return is directly

obtainable by rearranging the price reaction equation of Proposition 1 in

Kim and Verrecchia (1991).13 Empirically, such rational expectations models imply that market

expectations of public announcement are unbiased.14 Related theoretical papers that link noisy public signals to asset

returns are Veronesi (2000), Aruoba (2008), Cenesizoglu (2010), and Ai

(2010), among others.

15 It is released on Thursday if the Fourth of July falls on a Friday.

However, it is still released on Friday if that Friday is Good Friday and

markets are closed. In November 1998, the Employment Situation

Report was released by mistake 24 hours before schedule.16 The CES program is also known as the Payroll or Establishment

Survey.


rate of these cash flows (denominator in the present valueformula). As such, the theoretical impact of macroeco-nomic news on equity prices is unclear and it is thereforean empirical question.

Boyd, Hu, and Jagannathan (2005) empirically showthat an increase in unemployment is associated with amarginally significant rise in the S&P 500 index inexpansions (on announcement days). In contrast, anincrease in unemployment in recessions leads to a muchmore significant decrease in the stock market. Because theeconomy is in an expansionary cycle most of the time, thestock market on average rises when negative unemploy-ment news is released. The economic mechanism used torationalize why the stock market falls upon the release ofpositive news about real economic activity is a monetaryresponse function: To avoid overheating of the economyand ramping inflation, the Federal Reserve is expected tomake borrowing more costly by raising interest rates,which is bad news for stocks.

McQueen and Roley (1993) also find stronger relationsbetween macroeconomic announcements and stockreturns after conditioning on the business cycle. Usingfive-minute S&P 500 futures, Andersen, Bollerslev,Diebold, and Vega (2007) show that positive news aboutreal activity variables tends to move equity futures pricesup in recessions and down in expansions.

3.3. Testable implications

Combining the standard rational expectations modelwith the asymmetric interpretation of information fromthe two previous subsections yields the following predic-tion: Announcement-day returns and future revisions arepositively related. Indeed, if positive news pushes pricesdown, p2 is then negatively related to SG and, therefore,the price change p2�p1 is positively related to�SG�ð�nÞ ¼ n�SG ¼ Rev. Two clear empirical predictionsemerge. First, if positive news is good news for stocks,then there should be a negative relation betweenannouncement-day returns and revisions to the publicsignals. Second, if positive news is bad news for stocks,then the relation between announcement-day returns andrevisions should be positive.

The interpretation of the news for stocks does dependon the business cycle, as in Boyd, Hu, and Jagannathan(2005). Positive news is bad news in expansions and goodnews in recessions. As a result, one can reformulate theabove predictions in terms of the state of the economy:(1) In expansions, announcement-day returns and revi-sions should be positively related; (2) in recessions, therelation should be negative.

These relations between returns and revisions signifi-cantly deepen our understanding of the informationaggregation process around public announcements. Iftrue, they would show that announcement-day pricesaggregate not only information about the initial releasebut also information about the final revised value of theannounced variable. In turn, this would highlight the factthat revisions do matter and that investors rationally takethem into account when they observe inaccurate publicsignals.

4. Announcement surprises and revisions

In this section, I use the release of nonfarm payroll andunemployment by the Bureau of Labor Statistics as anexample to explain the mechanics of data announcementand revisions. I describe the data that are used in theempirical analysis and define all the necessary variables,such as the preliminary and final revisions.

4.1. Announcement methodology and revision process

Nonfarm payroll is released on a monthly basis as partof the Current Employment Statistics (CES) report by theBLS at 8:30 a.m. Eastern Standard Time (EST) on the thirdFriday after the conclusion of the reference week, which isthe week including the 12th of the month.15 Eachmonthly report tabulates the previous month’s estimatedchange in payroll, the previous two months’ revisedchanges in payroll, and the average changes during theprevious two quarters. It also provides a complete break-down of these nationwide numbers (split by industries,race, age, etc.).

The CES program is an estimate of the nation’s employ-ment based on the responses (surveys) from about160,000 businesses and government agencies represent-ing approximately 400,000 worksites.16 This sampleaccounts for about one-third of total nonfarm payrollemployment and is constructed to be as representativeas possible. The nationwide change in nonfarm payroll isestimated by extrapolating this sample using the latestavailable Unemployment Insurance (UI) tax numbers. TheUI ‘‘universe count’’ tabulates the total number ofemployees covered by UI laws and is available on a laggedquarterly basis. It contains individual employer recordsfor more than eight million establishments coveringnearly 97% of total nonfarm employment, which thereforeprovides a national benchmark for the sample-basedestimates.

For a particular month, the BLS releases its initialnonfarm payroll estimate based on the received surveysand the lagged UI numbers. However, for the month inquestion, the BLS continues to receive payroll informationfrom firms after the initial release. The result is that, inthe following month’s report, the BLS also releases arevision to the previous release based on the late surveysreceived. This process, called sample revisions, continuesfor two months.

Each year, the BLS revises the payroll estimates to analmost-complete count of the payroll employment usingthe new UI reports. This is called the benchmark revision.It also includes updated data about the birth and death offirms during the past year. Each annual benchmark revi-sion can affect the prior five years of data because

Fig. 2. This figure presents the history of revisions to the August 2000 release of nonfarm payroll. All numbers refer to the change in nonfarm payroll in

thousands of employees during the month of July 2000. The Money Market Services (MMS) expectation (square) was 74, and the initial release (circle)

reported a decrease of 108. There were two sample-based revisions (diamonds) in the subsequent two months, at �51 and �41. This was then followed

by five benchmark-based revisions (triangles) in the subsequent five years, leading to a final revised value of 176.


seasonal adjustment models are rerun based on the newbenchmark numbers.

On an irregular basis, the BLS also makes improvementsto the methods it uses to calculate the national nonfarmpayroll. For instance, between 2000 and 2004, it put inplace a probability-based sample design to replace itsoutdated quota sample-based design.17 When such meth-odological changes are put in place, historical time series ofmacroeconomic data must be reconstructed ‘‘to avoid seriesbreaks and provide users with continuous, comparable, y,time series suitable for economic analysis.’’18

Fig. 2 shows the history of revisions for one particularnonfarm payroll announcement (August 2000). The initialsurprise was negative, as evidenced by the differencebetween the positive market expectation (square) and thenegative initial announcement (circle). The first two sam-ple-based revisions (diamond) indicate that the initialrelease was too low, a trend subsequently reversed by thefirst two benchmark-based revisions (triangle). However,the last three benchmark revisions (triangle) pushed theestimate of the change in nonfarm payroll in July 2000 to apositive level, even higher than the initial expectation. The2003 revision was partly based on the North AmericanIndustry Classification System (NAICS) industry reclassifi-cation that the BLS fully converted to in 2002. Thisparticular revision history was chosen for its extremenature to highlight the process of revisions that macro-economic announcements undergo. It is not representativeof the average revision, as is shown in the next subsection.

In summary, revisions occur for three main reasons:late reports, benchmark updates, and methodologicalchanges. The key point is that revisions due to late reportsand benchmark revisions are informative in the sense that

17 Another example of a methodological update is the change in

base year that the Federal Reserve put in place in 1998 for the industrial

production series.18 See http://www.bls.gov/ces/cesregrevtec.htm.

they incorporate new information in the federal agencies’reports that was not available to them at the time of theinitial announcement. Infrequent methodological changesare not necessarily informative, even though they pre-sumably lead to a more accurate measurement of eco-nomic trends.

4.2. Data sample

I obtain the historical releases of the CES report for allmonths since 1985 from the BLS. The reports contain theinitial announcements of nonfarm payroll, unemployment,as well as more detailed information on hourly earnings,payroll distribution across regions and industries, etc.I focus on nonfarm payroll and the unemployment ratebecause of the wide attention they receive in the financialpress and on Wall Street.19 The sample and benchmarkrevisions to all nonfarm payroll announcements come fromthe real-time data set for macroeconomists collected by theFederal Reserve Bank of Philadelphia. This data set containscomplete vintages of the payroll data starting in 1964.

As is standard in the literature, market expectationsare from the Money Market Services (MMS) database. TheMMS data collected by Informa Global Markets, now asubsidiary of Standard & Poor’s, contains the announce-ment date, time, and market expectation of all nonfarmpayroll and unemployment releases since February 1985.The market consensus expectation is the median fromabout 40 money market economists’ expectations sur-veyed each Friday regarding the coming week’sannouncements. MMS expectations do contain informa-tion about the upcoming announcements and have sig-nificantly lower mean squared errors than autoregressive

19 Andersen, Bollerslev, Diebold, and Vega (2007) show that non-

farm payroll is the monthly macroeconomic announcement that has the

highest price impact on five-minute S&P 500 futures return.

http://www.bls.gov/ces/cesregrevtec.htm

Table 1Summary statistics of daily Standard & Poor’s (S&P) 500 index return employment announcement days.

This table presents the mean daily return, the standard deviation of returns, and the minimum and maximum returns of the S&P 500 index on days of

employment announcements and all other days between February 1985 and May 2005. Expansions and recessions are tabulated according to the

National Bureau of Economic Research business cycle indicator. N is the number of observations. All figures are in percentages.

N Mean Standard deviation Minimum Maximum

All days 5,128 0.033 1.072 �20.467 9.099

All announcement days 239 0.033 1.181 �6.768 3.673

Announcement days in expansions 221 0.068 1.170 �6.768 3.673

Announcement days in recessions 18 �0.396 1.258 �2.476 1.573

Table 2Summary statistics of nonfarm payroll variables.

This table presents summary statistics of the nonfarm payroll-related

variables released in the monthly employment report of the Bureau of

Labor Statistics. The surprise (Surt) is defined as the difference between

the announced value and the survey expectation. The revisions are

defined as the difference between the revised values and the initial

announcements. There are two sample revisions (Revt +1,t and Revt+ 2,t)

and one final (benchmark) revision (RevT,t), where T is May 2010. The

total surprise (TSurT,t) is the difference between the final revised value

available at T=May 2010 and the initial expectation. The total sample

period is from February 1985 to May 2005. Payroll is measured as a

change in thousands of employees. Expansions and recessions are

tabulated according to the National Bureau of Economic Research

business cycle indicator. N is the number of observations. n, nn, and nnn

denote significance at the 10%, 5%, and 1% level, respectively, in tests in

which each series has a mean of zero.

N Mean Standard

deviation

Minimum Maximum

Panel A: Nonfarm payroll surprise Surt

Whole

sample

239 �8.772 116.736 �318 409

Expansions 221 �2.595 116.364 �318 409

Recessions 18 �84.611nnn 94.691 �238 75

Panel B: First sample revision of nonfarm payroll announcement Revt +1,t

Whole

sample

239 0.289 53.652 �145 192

Expansions 221 �0.348 52.783 �145 192

Recessions 18 8.111 64.616 �110 130

Panel C: Second sample revision of nonfarm payroll announcement

Revt+ 2,t

Whole

sample

239 15.314nnn 63.702 �158 222

Expansions 221 14.348nnn 62.083 �158 222

Recessions 18 27.167 82.286 �112 172

Panel D: Final revision of nonfarm payroll announcement RevT,t

Whole

sample

239 12.502n 107.050 �443 293

Expansions 221 14.086n 108.461 �443 293

Recessions 18 �6.944 88.126 �133 177

Panel E: Nonfarm payroll total surprise TSurT,t

Whole

sample

239 3.730 100.736 �330 265

Expansions 221 11.491n 98.061 �330 265

Recessions 18 �91.556nnn 85.029 �296 90


models. However, they are not completely unbiased, asshown by both Pearce and Roley (1985) and McQueen andRoley (1993).

Because data revisions occur at different frequenciesand have different effects on the time series of announce-ments (sample revisions affect only two months of datawhereas benchmark revisions affect five years of data), itis important to ensure that all revisions are comparableacross the sample. As a result, I drop the last five years ofannouncements (June 2005–May 2010), which ensuresthat all revised values are mature enough.20 As robustnesschecks, I also repeat the analysis without this filter as wellas with a different definition of revisions, with the resultsbeing qualitatively unchanged.

By combining the release and expectations data togetherwith the above maturity-of-revisions filter, my commonsample of payroll announcements starts in February 1985and ends in May 2005. As in Boyd, Hu, and Jagannathan(2005), I define stock returns as the daily percentagechange in the S&P 500 index, which ignores dividenddistributions. Over the 243-month announcement period,four CES releases occurred on Easter Friday, when the stockmarket is closed. I define each announcement month as arecession or expansion month according to the NationalBureau of Economic Research (NBER) business cycle datingmethodology. Of 239 months, 221 are expansion monthsand 18 are recession months spanning three differentrecessions. Table 1 reports the average daily return as wellas the standard deviation of returns on all announcementdays and split according to the business cycle. Stock returnsare on average lower in recessions even though thedifference is not statistically significant.

4.3. Definitions of surprises and revisions

I define the announcement surprise Surt as the differ-ence between the actual announcement At at time t andthe market’s median expectation E[At]:

Surt � At�E½At �, ð3Þ

where the expectation is taken sometime before time t.Panel A of Table 2 shows that the average nonfarm payroll

20 Dropping the last five years of data also alleviates concerns about

possible time-trends in revisions. According to the BLS, the size of

revisions has somewhat decreased over the past five years, mainly due

to the increase in electronic submissions of reports, which allows for

greater accuracy and completeness. By excluding all releases since May

2005, I control for this potential issue.

surprise is highly negative in recessions. The initialannouncement At is the preliminary estimate that subse-quently undergoes sample and benchmark revisions.

For two months after its initial release, each payrollannouncement undergoes two sample-based revisions.I define the first sample revision Revt+1,t of the time t releaseas the difference between its revised announcement RAt+1,t

Table 3Correlations between nonfarm payroll variables.

This table presents correlations between nonfarm payroll-related variables released in the monthly employment report of the BLS. The surprise (Surt) is

defined as the difference between the announced value (At) and the survey expectation (E[At]). The revisions are defined as the difference between the

revised values and the initial announcements: There are two sample revisions (Revt+ 1,t and Revt +2,t) and one final (benchmark) revision (RevT,t), where T is

May 2010. The total surprise (TSurT,t) is the difference between the final revised value available at T=May 2010 and the initial expectation. The sample

period is from February 1985 to May 2005 (239 observations). p-values are shown in bracket under each estimated correlation.

E[At] At FT,t Surt Revt+ 1,t Revt +2,t RevT,t TSurT,t

E[At] 1.000

At 0.749 1.000

(0.000)

FT,t 0.797 0.800 1.000

(0.000) (0.000)

Surt 0.167 0.778 0.435 1.000

(0.010) (0.000) (0.000)

Revt+ 1,t 0.027 �0.124 0.121 �0.211 1.000

(0.676) (0.055) (0.062) (0.001)

Revt+ 2,t 0.041 �0.122 0.160 �0.221 0.895 1.000

(0.528) (0.059) (0.013) (0.001) (0.000)

RevT,t 0.003 �0.400 0.230 �0.598 0.386 0.443 1.000

(0.964) (0.000) (0.000) (0.000) (0.000) (0.000)

TSurT,t 0.196 0.477 0.749 0.524 0.167 0.215 0.370 1.000

(0.002) (0.000) (0.000) (0.000) (0.010) (0.001) (0.000)


available at time t+1 and the initial announcement that wasmade at time t:

Revtþ1,t � RAtþ1,t�At : ð4Þ

Panel B of Table 2 shows the average and standard deviationof the first nonfarm payroll sample revision since 1985.Similarly, the second sample revision is defined asRevtþ2,t � RAtþ2,t�At and its average is shown in Panel Cof Table 2.

I define the final revision of an announcement, RevT,t,as the difference between the final value FT,t of themacrodata available from the federal agencies at time T

and the original announcement At:

RevT ,t � FT ,t�At , ð5Þ

where T is May 2010. The average final revision of nonfarmpayroll announcements is shown in Panel D of Table 2. Thechoice of T as May 2010 is based on the macroeconomicliterature that defines the final published revision as themost accurate measure of the corresponding macroeco-nomic variable [see, for instance, Croushore (2008) andreferences therein]: FT,t is the government’s best estimate ofthe true underlying state variable at time t.

By comparing Panels A and D of Table 2, it should beclear that revisions are statistically and economically sig-nificant: They are on average of the same order of magni-tude as the initial surprises. The signs of the revisionssuggest the BLS systematically under-reports the change innonfarm payroll. In expansions, the final number tends to behigher than the initial release, while in recessions, the finalnumber tends to be lower than the initial release. Looking atthe incremental change in revisions across Panels B–D, onecan observe that the largest fraction of the revisions occursvia the second sample revision (about 55%) and via the

benchmark revisions (about 35%), and the first samplerevision is insignificant in comparison (about 10%).

Finally, I define the total surprise of an announcementTSurT,t as the difference between the final available valuein May 2010 and the market’s median expectation of theinitial announcement:

TSurT,t � FT�E½At� ¼ ðFT�AtÞþðAt�E½At�Þ ¼ RevT,tþSurt :

ð6Þ

Comparing Panels A and E of Table 2, total surprises areon average positive in expansions and negative in reces-sions, as should be expected.

Table 3 shows the correlations between the abovenonfarm payroll variables. Initial announcement surprisesSurt and final revisions RevT,t are strongly negativelycorrelated (�0.598), which suggests that the processleading to expectations is economically very differentfrom the process leading to revisions. The former is drivenby professional investors whose compensation partlydepends on the accuracy of their forecasts of short-termannouncements, whereas the latter is rooted in thestatistical agencies’ main purpose of providing precisedata using verifiable and transparent methods to inves-tors, policy markers, and researchers. Note also that thecorrelation between initial surprises and total surprises isabout 0.5, again suggesting that revisions are large and domatter.

On each nonfarm payroll announcement day, the BLSreleases three other relevant pieces of information: theunemployment rate and the sample revisions to theprevious two announcements. I define the unemploymentsurprises as SurUN

t and the prior two sample revisions asRAt,t�1 and RAt,t�2, and controlling for these releases inthe empirical tests is important.

22 Instead of splitting the sample between expansions and reces-

sions, one can also run joint tests with a dummy variable indicating the

state of the business cycle that multiplies all independent variables

except the constant. While the results are qualitatively similar with both

methods, I follow the previous literature in separating the sample, which

allows a cleaner interpretation of the results.23 I am grateful to an anonymous referee for pointing this out.24 Information releases should drive prices. However, if federal

agencies use the market’s reaction to their announcements to revise

the data in the future, this would imply that causality could run from

prices to revisions. During telephone conversations with analysts at the

BLS, I was assured that they do not use current market returns to revise


5. Empirical analysis

Using the data described in the previous section, I testmy main empirical hypotheses concerning the linkbetween announcement-day returns and revisions acrossthe business cycle. I first focus on nonfarm payrollbecause the previous literature has shown that it is thekey monthly macroeconomic release. To show the robust-ness of my analysis, I then turn my attention to GDPbecause it is the most important quarterly release. Finally,I analyze the release of industrial production because ofits wide use in cross-sectional asset pricing tests [seeChen, Roll, and Ross (1986), among many others].

5.1. Setup

Units of measurement differ across macroeconomic vari-ables (for instance, nonfarm payroll is in thousands ofemployees and the unemployment rate is in percentages).To alleviate this problem and facilitate comparison acrossannouncements, I follow the previous literature in usingstandardized surprises and revisions. These are constructedby dividing all surprises and revisions by their respectivehistorical sample standard deviation. For instance

Surt ¼Surt

sSur¼

At�E½At�

sSurð7Þ

and

RevT ,t ¼RevT,t

sRev¼

FT ,t�At

sRev, ð8Þ

where sSur is the sample standard deviation of the surprisesand sRev is the sample standard deviation of the finalrevisions. The same normalization procedure is applied toall other variables defined above, such as the samplerevisions. The sample standard deviations are different foreach series but, because they are constant for each, thenormalizations affect neither the statistical significance ofthe coefficients nor the fit of the regressions.21

To understand what drives announcement-day returnson the S&P 500 index (rt

SP), I first run two specifications,

labeled (i) and (ii), where the independent variablesrepresent different information sets released at time t:

rSPt ¼ aiþbiSurtþei,t ð9Þ

and

rSPt ¼ aiiþbiiSurtþgiiRAt,t�1þdiiRAt,t�2þliiSur

UN

t þeii,t ,

ð10Þ

where the regression coefficients are labeled i and ii inEqs. (9) and (10) and the normalized information vari-ables are as defined in the previous section. The previousliterature on the impact of macroeconomic releases hasfocused on Specification (i). In Specification (ii), I includeall other major pieces of information released in the samereport and hence at the same time as nonfarm payroll,

21 To allow the comparison of the estimated regression coefficients,

the normalization is done separately in the expansion versus recession

samples.

namely, the two sample-based revisions of the previoustwo announcements and the unemployment surprise.

I run the above regressions over the entire sample(Table 4), for the announcements that occur in expansions(Table 5), and for the announcements that occur inrecessions (Table 6). Following the results of Boyd, Hu,and Jagannathan (2005) and Andersen, Bollerslev,Diebold, and Vega (2007), I expect the b coefficient onthe nonfarm payroll surprise to be negative in expansions,positive in recessions, and negative over the entire sam-ple. Lastly, I expect the l coefficient on the unemploymentsurprise to be positive in expansions, negative in reces-sions, and positive over the entire sample.22

In the second and key set of tests labeled (iii) to (v), Iinclude the future revisions to the data just released asexplanatory variables, Revtþ1,t , Revtþ1,t , or RevT,t:

rSPt ¼ aiiiþxiiiInfotþyiiiAtþZiiiRevtþ1,tþeiii,t , ð11Þ

rSPt ¼ aivþxivInfotþyivAtþZivRevtþ2,tþeiv,t , ð12Þ

and

rSPt ¼ avþxvInfotþyvAtþZvRevT ,tþev,t , ð13Þ

where, for ease of presentation in the above equations, Isummarize the set of contemporaneous information vari-ables Surt , RAt,t�1, RAt,t�2, and Sur

UN

t in the matrix Infot . Ialso include the normalized announcement of nonfarmpayroll At released at time t as an independent variable.The reason for doing so is a potential omitted variable bias.Because the initial announcements and the revisions arenegatively correlated (�0.1 to �0.4 in Table 3), omitting theformer could bias the estimated coefficient on the latter.23

By combining the standard rational expectations model ofSection 3 with the asymmetric interpretation of informationacross the business cycle, the expectation is for returns andrevisions to be positively related in expansions ðZ40Þ andnegatively related in recessions ðZo0Þ.24

Empirically, an alternative hypothesis is that the priceimpact at the announcement is in response to the totalsurprise, not the standard surprise defined above.25

I therefore run two additional specifications ((vi) and(vii)) with TSurT,t as the main independent variable:

rSPt ¼ aviþzviTSurT ,tþevi,t ð14Þ

past data. It is therefore reasonable to assume that the causality of my

regressions is from information to prices.25 In that spirit, Roll (1984) shows that orange juice futures returns

contain information about the future error in the National Weather

Service forecast that will be released later on that day.

Table 4Impact of nonfarm payroll information on daily Standard & Poor’s (S&P) 500 index returns: full sample.

This table reports ordinary least squares regressions examining the information aggregation process on nonfarm payroll announcement days. The daily

return of the S&P 500 index (in percentages) is the dependent variable for all specifications. Sur t is the normalized surprise of the initial nonfarm payroll

announcement at date t. RAt,t�1 and RAt,t�2 are the normalized revised announcements at day t of the nonfarm payroll releases of the previous month

(t�1) and two months prior (t�2), respectively. SurUN

t is the normalized surprise to the simultaneous announcement of unemployment at date t.

Revtþ1,t , Revtþ2,t , and RevT ,t are the normalized revisions available one month later (t+1, first sample revision), two months later (t+2, second sample

revision), and at T=May 2010 (benchmark revision(s)), respectively, of the initial nonfarm payroll announcement released at date t. At is the normalized

change in nonfarm payroll announced at date t. TSurT ,t is the normalized total surprise of the initial nonfarm payroll announcement at date t. The sample

period is from February 1985 to May 2005. Heteroskedasticity-robust standard errors are reported in parentheses below the coefficients. n, nn, and nnn

denote significance at the 10%, 5%, and 1% level, respectively.

Regressand Specification

rt

S&P500(i) (ii) (iii) (iv) (v) (vi) (vii)

Constant 0.020 �0.122 �0.280 �0.297 �0.280 0.032 �0.056

(0.077) (0.098) (0.219) (0.217) (0.215) (0.076) (0.097)

Sur t �0.168nnn�0.175nn

�0.501nnn�0.499nnn

�0.423nn

(0.082) (0.081) (0.192) (0.193) (0.203)

TSurT ,t 0.028 0.049

(0.080) (0.082)

RAt,t�10.034 �0.092 �0.094 �0.100 0.056

(0.077) (0.105) (0.105) (0.104) (0.093)

RAt,t�20.138 �0.012 �0.005 �0.004 0.153

(0.109) (0.148) (0.146) (0.144) (0.117)

SurUN

t0.020 0.047 0.045 0.045 0.045

(0.068) (0.070) (0.069) (0.068) (0.071)

At 0.472nn 0.466nn 0.457n�0.099

(0.237) (0.237) (0.236) (0.114)

Revtþ1,t 0.087

(0.075)

Revtþ2,t 0.070

(0.074)

RevT,t 0.141

(0.089)

N 239 239 239 239 239 239 239

R2 0.020 0.039 0.068 0.066 0.072 0.001 0.022

26 While I have to acknowledge that the recession sample is small, I

experimented with different methods to add recession months, as in

Andersen, Bollerslev, Diebold, and Vega (2007), with similar qualitative

results.


and

rSPt ¼ aviiþzviiTSurT ,tþgviiRAt,t�1þdviiRAt,t�2þlviiSur

UN

t

þyviiAtþevii,t : ð15Þ

Due to the simple relation between surprise, revision, andtotal surprise highlighted in Eq. (6), one would expect thetotal surprise to encompass the effect of both the surpriseand the revision. However, given the strong negative correla-tion between surprises and revisions, it is likely that usingthe total surprise as explanatory variable hides the separateimpact of surprises and revisions. Furthermore, if investorsuse the initial surprise when forecasting the final revisedvalue, then using the total surprise would also not allow theeconometrician to separate the effect of the initial announce-ment surprise from the information aggregation processstemming from the revisions. As a result, the empiricalprediction is for the total surprise to come up insignificantin the tests ðz¼ 0Þ while either the surprise or the revision issignificantly related to announcement-day returns.

5.2. Main result: nonfarm payroll

Looking across Tables 4 (full sample), 5 (expansionsample), and 6 (recession sample), the results of Specifica-tion (i) confirm what is now a well-established fact: Macro-economic information releases do move prices and goodpayroll news is bad news for stocks in general, particularlyso in expansions. The R2 of such regressions is also very lowð � 2:5%Þ, consistent with the existing literature.26

Specification (ii) includes all other informationreleases that occur simultaneously on nonfarm payrollannouncement days: the release of two sample-basedrevisions and the release of the unemployment rate. Tothe best of my knowledge, this is the first paper to showthat the release of revised macroeconomic data has no

Table 5Impact of nonfarm payroll information on daily Standard & Poor’s (S&P) 500 index returns: expansion sample.

This table reports ordinary least squares regressions examining the information aggregation process on nonfarm payroll announcement days. The daily

return of the S&P 500 index (in percentages) is the dependent variable for all specifications. Surt is the normalized surprise of the initial nonfarm payroll

announcement at date t. RAt,t�1 and RAt,t�2 are the normalized revised announcements at day t of the nonfarm payroll releases of the previous month

(t�1) and two months prior (t�2), respectively. SurUN

t is the normalized surprise to the simultaneous announcement of unemployment at date t.

Revtþ1,t , Revtþ2,t , and RevT,t are the normalized revisions available one month later (t+1, first sample revision), two months later (t+2, second sample

revision), and at T=May 2010 (benchmark revision(s)), respectively, of the initial nonfarm payroll announcement released at date t. At is the normalized

change in nonfarm payroll announced at date t. TSur T,t is the normalized total surprise of the initial nonfarm payroll announcement at date t. The sample

period covers the National Bureau of Economic Research-dated expansion months between February 1985 and May 2005. Heteroskedasticity-robust

standard errors are reported in parentheses below the coefficients. n, nn, and nnn denote significance at the 10%, 5%, and 1% level, respectively.


rt


Constant 0.063 �0.076 �0.256 �0.284 �0.276 0.065 0.013

(0.078) (0.110) (0.232) (0.233) (0.233) (0.078) (0.112)

Sur t �0.202nn�0.196nn

�0.477nn�0.472nn

�0.392

(0.084) (0.082) (0.209) (0.209) (0.240)

TSur T ,t 0.027 0.078

(0.082) (0.084)

RAt,t�10.044 �0.038 �0.039 �0.056 0.081

(0.078) (0.101) (0.101) (0.100) (0.091)

RAt,t�20.121 �0.030 �0.022 �0.012 0.152

(0.110) (0.154) (0.153) (0.150) (0.117)

SurUN

t0.055 0.077 0.075 0.071 0.077

(0.069) (0.070) (0.069) (0.069) (0.073)

At 0.405n 0.395n 0.399n�0.152

(0.238) (0.237) (0.240) (0.103)

Revtþ1,t 0.149nn

(0.076)

Revtþ2,t 0.135n

(0.074)

RevT ,t 0.180nn

(0.083)

N 221 221 221 221 221 221 221

R2 0.030 0.049 0.080 0.078 0.082 0.001 0.032

27 Adding the final revision to the unemployment rate does not

affect the results. This should be expected because the unemployment

rate undergoes only minor revisions: 30% of the announcements are

never revised and the average absolute revision is 0.12%.


impact on stock returns, i.e., the coefficients on RAt,t�1 andRAt,t�2 are insignificant. An increase in unemployment isbad news for stocks in recessions, which is consistentwith the results of Boyd, Hu, and Jagannathan (2005).

In Specifications (iii)–(v), I analyze the impact of thefuture revisions of the data just released. Focusing first onSpecifications (iii) and (iv), it looks as though sample-based revisions have some explanatory power forannouncement-day returns only in expansions. The factthat the coefficients on Revtþ1,t and Revtþ2,t are mainlyinsignificant suggests that either investors do not attemptto forecast intermediate sample-based revisions when theinitial release is made or these revisions are not usefulforecasts of the final revised value.

Specification (v) shows the key result of the paper:Final revisions are significantly related to announcement-day returns but the relation differs across the businesscycle. The coefficient on the final revisions RevT,t issignificant in Tables 5 and 6 and has the predicted sign:positive in expansions and negative in recessions.The theoretical interpretation of this result crucially

depends on how nonfarm payroll information is inter-preted across the business cycle. Good news is good newsfor stocks in recessions and bad news in expansions asevidenced by the sign of the coefficient on the announce-ment surprise Surt . As in Specifications (iii) and (iv), thecoefficient on the initial surprise is statistically differentfrom the coefficient on the revision, the former beinginsignificant in both the expansion and the recessionsample.27

The economic significance of the result is large. First, theimpact coefficients on the final revision are of the sameorder of magnitude as, and sometimes larger than, the oneson the initial surprises. Second, a one standard deviationexpected shock in the final revision (� 107,000 difference inpayroll according to Table 2) leads to an 18 basis pointincrease in the S&P 500 index in expansions and a 77 basis

Table 6Impact of nonfarm payroll information on daily Standard & Poor’s (S&P) 500 index returns: recession sample.

This table reports OLS regressions examining the information aggregation process on nonfarm payroll announcement days. The daily return of the S&P

500 index (in percentages) is the dependent variable for all specifications. Sur t is the normalized surprise of the initial nonfarm payroll announcement at

date t. RAt,t�1 and RAt,t�2 are the normalized revised announcements at day t of the nonfarm payroll releases of the previous month (t�1) and two

months prior (t�2), respectively. SurUN

t is the normalized surprise to the simultaneous announcement of unemployment at date t. Revtþ1,t , Revtþ2,t , and

RevT ,t are the normalized revisions available one month later (t+1, first sample revision), two months later (t+2, second sample revision), and at T=May

2010 (benchmark revision(s)), respectively, of the initial nonfarm payroll announcement released at date t. At is the normalized change in nonfarm

payroll announced at date t. TSur T,t is the normalized total surprise of the initial nonfarm payroll announcement at date t. The sample period covers the

National Bureau of Economic Research-dated recession months between February 1985 and May 2005. Heteroskedasticity-robust standard errors are

reported in parentheses below the coefficients. n, nn, and nnn denote significance at the 10%, 5%, and 1% level, respectively.


rt


Constant �0.474 �0.731 �0.721 �0.658 �0.347 �0.723 �0.231

(0.420) (0.484) (0.431) (0.446) (0.374) (0.445) (0.356)

Sur t 0.087 0.256 0.394 0.455 0.354

(0.316) (0.270) (0.525) (0.512) (0.361)

TSurT ,t �0.489* �0.588**

(0.255) (0.228)

RAt,t�1�0.548 �0.326 �0.271 �0.279 �0.498

(0.419) (0.652) (0.682) (0.604) (0.497)

RAt,t�2 0.069 0.208 0.174 0.091 0.062

(0.375) (0.472) (0.385) (0.324) (0.335)

SurUN

t�0.452* �0.551* �0.604* �0.593* �0.543*

(0.231) (0.286) (0.304) (0.326) (0.277)

At 0.009 �0.009 �0.546 �0.115

(0.599) (0.619) (0.599) (0.349)

Revtþ1,t �0.471(0.444)

Revtþ2,t �0.537

(0.398)

RevT,t �0.768***

(0.205)

N 18 18 18 18 18 18 18

R2 0.005 0.238 0.296 0.311 0.454 0.151 0.325


point decrease in recessions.28 Over my sample, the averagemarket capitalization of the S&P 500 index is $5.548 trillion.So on average, this error of 100,000 jobs in the BLS employ-ment report is equivalent to a $10 billion daily change inmarket value during expansions and $43 billion duringrecessions. And third, the R2 of the regressions increaseswhen the final revision is included as an explanatoryvariable. While the increase is small in expansions, it is largein recessions where Specification (v) explains more than 45%of the variation in daily stock returns. This can be contrastedto Andersen, Bollerslev, Diebold, and Vega (2007), who, usingfive-minute futures returns as their dependent variable, canexplain only 18% of the variability in returns around theannouncement time during recessions.

Specifications (vi) and (vii) test the alternative hypoth-esis that the market reacts to the total surprise TSurT,t

28 This is comparable to the average announced change in nonfarm

payroll (not the surprise), which is � 139,000 employees over the

sample.

instead of to the initial surprise Surt. If this is the trueempirical relation and because the total surprise is equalto the initial surprise plus the final revision, TSurT,t=Surt +RevT,t, then the announcement-day return should reactequally to all three components. However, the resultsshow otherwise. In the full sample, the total surprise hasno explanatory power while the initial surprise does. Inthe expansion sample, the total surprise also has noexplanatory power while the final revision does. Finally,in the recession sample, the total surprise has someexplanatory power and the final revision has a largeexplanatory power for announcement-day returns. Also,the R2 of the total surprise regressions are significantlylower than when the initial surprise and final revisionsare used as separate independent variables.

By comparing Specifications (v) and (vii), it becomes clearthat using the total surprise as explanatory variable is notsufficient and actually hides the separate information aggre-gation mechanisms that can occur via the initial surprise andthe final revision. It is likely that investors pay attentionto the initial surprise not so much because it conveys

Table 7Impact of nonfarm payroll information on daily Standard & Poor’s (S&P)

500 index returns: revisions as five-year moving windows.

This table reports ordinary least squares regressions examining the

information aggregation process on nonfarm payroll announcement

days. The daily return of the S&P 500 index (in percentages) is the

dependent variable for all specifications. Sur t is the normalized surprise

of the initial nonfarm payroll announcement at date t. RAt,t�1 and RAt,t�2

are the normalized revised announcements at day t of the nonfarm

payroll releases of the previous month (t�1) and two months prior

(t�2), respectively. SurUN

t is the normalized surprise to the simultaneous

announcement of unemployment at date t. At is the normalized change

in nonfarm payroll announced at date t. Revtþ5y,t is the normalized

revisions available five years after each initial nonfarm payroll

announcement released at date t. The sample period is from February

1985 to May 2005. Expansions and recessions are tabulated according to

the National Bureau of Economic Research business cycle indicator.

Heteroskedasticity-robust standard errors are reported in parentheses

below the coefficients. n, nn, and nnn denote significance at the 10%, 5%,

and 1% level, respectively.

Regressand Full sample Expansion Recession

rt

S&P500

Constant �0.283 �0.237 �0.890

(0.215) (0.239) (0.534)

Sur t �0.440nn�0.271 �0.234

(0.208) (0.167) (0.453)

RAt,t�1�0.100 �0.040 �0.342

(0.103) (0.115) (0.675)

RAt,t�2 �0.001 0.116 0.024

(0.142) (0.100) (0.416)

SurUN

t0.044 0.042 �0.549

(0.068) (0.069) (0.349)

At 0.455n 0.280 �0.267

(0.236) (0.210) (0.679)

Revtþ5y,t 0.123 0.155n�0.303nn

(0.085) (0.086) (0.115)

N 239 221 18

R2 0.070 0.075 0.373


information in and of itself but rather because it helps themforecast the final revised value. The strong negative correla-tion of �0.598 between initial surprises and final revisionsshown in Table 3 provides further evidence that the twoprocesses should be separated in empirical tests if theeconometrician is to observe whether investors truly careabout final revised values of macroeconomic data.

The main empirical takeaway from this paper is thatreturns are linked to future revisions and that this relationis business cycle-dependent. Another way to see theresult is that announcement-day returns do containinformation about the final revision of the initial release.Practically speaking, the results suggest that revisionsmatter and that investors do care about the final truevalue of the released data and rationally aggregate thecorrect information when the initial (inaccurate)announcements are released. It is highly unlikely thatthis effect is driven by insiders who come into the marketon announcement day and trade on private informationabout the upcoming BLS announcement of the country’snonfarm payroll change over the last month. It seemsmuch more likely that investors rationally account for thefact that announcements are potentially inaccurate andcould undergo significant revisions over time. They, there-fore, have an incentive to forecast the final value and theinitial release could be one input in their forecastingmodel. As a result, future revisions are priced in at thetime of the initial release.

5.3. Robustness checks and extensions

In this subsection, I check the robustness of the mainresults. I first present several different definitions of finalrevisions. Second, I repeat the empirical analysis with twoother macroeconomic series, namely GDP and industrialproduction.

5.3.1. Definition of revisions and choice of T=May 2010

In all previous tests, the final revision of an announce-ment made at time t is defined as the difference betweenthe final value available at T=May 2010 and the initialrelease. As is explained in Section 4.1, the BLS releasessample-based revisions of each announcement for threemonths and furthermore performs benchmark-based revi-sions every March. These benchmark revisions can impactdata up to five years prior. As a result, the choice of T as May2010 implies that all benchmark revisions are included forthe announcements that occurred before May 2005 and theannouncements that occurred after that could have onlyundergone some revisions (benchmark or sample or both).To ensure that all revisions in my sample are matureenough, I stop the sample in May 2005. However, theresults are qualitatively similar, albeit somewhat weaker,if I extend the sample up to December 2009. This should beexpected because, according to Table 2, a large fraction ofthe revisions are benchmark revisions. So the inclusion ofthe later announcements that have undergone fewer, andsmaller, revisions should bias the tests on the expandedsample against finding any significant results.

Another way to define the revisions is to use a movingwindow, say, five years, for every announcement. As a

result, the final revision for announcement t, Revt+ 5y,t, isdefined as the difference between the value available fiveyears later and the initial release. I re-run Specification (v)in the full, expansion, and recession samples and theresults are shown in Table 7. Overall, the results areweaker but qualitatively and quantitatively similar whenusing this new definition of revisions. The estimatedcoefficient on Revt +5y,t is insignificant over the full sample,significantly positive in expansions, and significantlynegative in recessions.

The moving-window definition of revisions, however,does introduce issues of data consistency across the timeseries due to the existence of methodological revisions.The BLS has made some infrequent updates its samplingmethodologies, which leads to a complete reconstructionof the time series of nonfarm payroll. The reconstructionis necessary to avoid artificial breaks in the data.By defining revisions using a moving window, theseartificial breaks reappear. For instance, in 1995–1996,the BLS introduced some improvements to its seasonal

Table 8Impact of gross domestic product (GDP) information on daily Standard &

Poor’s (S&P) 500 index returns: expansion sample.


information aggregation process on GDP announcement days in expan-

sionary months. The daily return of the S&P 500 index (in percentages) is

the dependent variable for all specifications. SurAdv

t , SurPrel

t0 , and SurFin

t00 are

the normalized surprise of the advance, preliminary, and final GDP

announcements at date t, t0 , or t00 . RevAdv

T,t , RevPrel

T ,t0 , and RevFin

T,t00 are the

normalized revisions available at T=May 2010 of the advance, prelimin-

ary, and final GDP announcements released at date t, t0 , or t00 . Advance

GDP is released three weeks after the end of the reference quarter;

preliminary GDP is released one month later; and final GDP is released

two months later. The sample period covers the National Bureau of

Economic Research-dated expansion months between January 1990 and

December 2008. Heteroskedasticity-robust standard errors are reported

in parentheses below the coefficients. n, nn, and nnn denote significance at

the 10%, 5%, and 1% level, respectively.


rS&P500 (i) (ii) (iii) (iv) (v) (vi)

Constant �0.054 �0.043 �0.089 �0.089 0.120 0.116

(0.136) (0.135) (0.118) (0.119) (0.117) (0.118)

SurAdv

t�0.034n

�0.037(0.020) (0.135)

RevAdv

T,t0.098

(0.116)

SurPrel

t0�0.071 �0.079(0.092) (0.099)

RevPrel

T,t00.050

(0.122)

SurFin

t00�0.090 �0.108(0.102) (0.103)

RevFin

T,t000.126nn

(0.068)

N 65 65 64 64 63 63

R2 0.020 0.036 0.006 0.015 0.011 0.048

Table 9Impact of gross domestic product (GDP) information on daily Standard &

Poor’s (S&P) 500 index returns: recession sample.


information aggregation process on GDP announcement days in reces-

sionary months. The daily return of the S&P 500 index (in percentages) is

the dependent variable for all specifications. SurAdv

t , SurPrel

t0 , and SurFin

t00 are

the normalized surprise of the advance, preliminary, and final GDP

announcements at date t, t0 , or t00 . RevAdv

T,t , RevPrel

T ,t0 , and RevFin

T,t00 are the

normalized revisions available at T=May 2010 of the advance, prelimin-

ary, and final GDP announcements released at date t, t0 , or t00 . Advance

GDP is released three weeks after the end of the reference quarter;

preliminary GDP is released one month later; and final GDP is released

two months later. The sample period covers the National Bureau of

Economic Research-dated recession months between January 1990 and

December 2008. Heteroskedasticity-robust standard errors are reported

in parentheses below the coefficients. n, nn, and nnn denote significance at

the 10%, 5%, and 1% level, respectively.


rS&P500 (i) (ii) (iii) (iv) (v) (vi)

Constant 0.135 0.126 �0.018 �0.117 �0.199 �0.126

(0.335) (0.353) (0.251) (0.306) (0.390) (0.448)

SurAdv

t0.567nn 0.513n

(0.222) (0.222)

RevAdv

T,t�0.166

(0.211)

SurPrel

t00.464 0.566

(0.330) (0.343)

RevPrel

T,t0�0.285n

(0.158)

SurFin

t000.025 0.047

(0.300) (0.308)

RevFin

T,t00�0.414nn

(0.139)

N 11 11 11 11 11 11

R2 0.248 0.262 0.016 0.295 0.001 0.524


adjustment methodology. By using T=May 2010 as thefinal value for all announcements, the effect of thismethodological update is basically absent because theentire May 2010 vintage has undergone this revision.However, by using a five-year moving-window definition,all announcements prior to 1990 (five years before therevision occurred) have a revised data under the oldseasonal adjustment methodology while all otherannouncements are revised under the new methodology.This effect could be particularly strong in my samplebecause another important methodological changehappened between 2000 and 2003, namely, theswitch from Standard Industry Classification (SIC) toNAICS industry classification. This data consistency issueis the main reason for my choice of T=May 2010 as thebenchmark case, while it is comforting to find similar,although weaker, results using other definitions ofrevisions.

5.3.2. GDP announcements

Nonfarm payroll is the key release within the monthlyannouncement cycle and has been called the king of

announcements in the financial press. However, GDPis arguably the most famous quarterly release, and italso undergoes significant revisions over time. Threeweeks after the end of a particular quarter, advanceGDP is released by the Bureau of Economic Analysis(BEA) at 8:30 a.m. EST. One month later, this releaseis revised in the form of preliminary GDP. One monthlater, the number is revised once more and is releasedas final GDP. Finally, the BEA also performs somesignificant benchmark revisions over time and a fullyrevised GDP number is revealed potentially yearsafter the reference quarter is over. Landefeld, Seskin,and Fraumeni (2008) provides details on the constructionof GDP and its ensuing revisions. Andersen, Bollerslev,Diebold, and Vega (2007), among many others, showthat the release of advance GDP does move prices butthat the announcements of preliminary and final GDPdo not.

In Tables 8 and 9, I present the results of similar testsas with nonfarm payroll:

rSPt ¼ aiþbiSur

GDP

t þei,t ð16Þ

Table 10Impact of industrial production information on daily Standard & Poor’s

(S&P) 500 index returns.


information aggregation process on industrial production announce-

ment days. The daily return of the S&P 500 index (in percentages) is the

dependent variable for all specifications. SurIP

t is the normalized surprise

of the initial industrial production announcement at date t. SurCA

t is the

normalized surprise to the simultaneous announcement of capacity

utilization at date t. RevIP

T ,t is the normalized revisions available at

T=May 2010 of the initial industrial production announcement released

at date t. The sample period is from January 1990 to May 2007, and the

expansion versus recession dating is done using the National Bureau of

Economic Research economic cycle indicator. Heteroskedasticity-robust

standard errors are reported in parentheses below the coefficients. n, nn,

and nnn denote significance at the 10%, 5%, and 1% level, respectively.

Regressand Expansion Expansion Recession Recession

rt

S&P500

Constant 0.079 0.068 �0.094 �0.250

(0.077) (0.078) (0.255) (0.230)

SurIP

t0.010 0.018 0.432n 0.504nn

(0.117) (0.119) (0.219) (0.230)

SurCA

t0.021 0.070 0.437 0.558

(0.115) (0.115) (0.351) (0.356)

RevIP

T ,t0.070nn

�0.658nnn

(0.032) (0.193)

N 189 189 17 17

R2 0.001 0.047 0.105 0.341


and

rSPt ¼ aiiþbiiSur

GDP

t þZiiRevGDP

T ,t þeii,t , ð17Þ

where GDP is either the advance (Adv), preliminary (Prel),or final (Fin) release. Table 8 reports the results in theexpansion sample, and Table 9 covers the recessionsample. Due to data availability of market expectations,the sample starts in 1990 and I drop the last two years ofdata to ensure that revisions are mature enough. BecauseGDP releases are quarterly, the number of observations,especially in the recession sample, is small.29

Specification (i) in both tables confirms the fact thatadvance GDP does move stock returns. It also confirms the‘‘good news is bad news in good times and good news inbad times’’ hypothesis. Specifications (iii) and (v) alsoconfirm the previous literature’s finding that the releaseof preliminary and final GDP does not move stock returns.In Specifications (ii), (iv), and (vi), I add the final revision ofthe data just released as an independent variable.Announcement-day returns for preliminary and final GDPare systematically linked to the future revisions of the datajust released. In expansions, the relation is positive while itis negative in recessions, as expected. The R2 of all regres-sions is dramatically increased by the inclusion of thefuture revisions. These results support my claim of a

29 As with nonfarm payroll, I experimented with different methods

to extend the recession sample, as in Andersen, Bollerslev, Diebold, and

Vega (2007), with similar qualitative results.

significant, albeit overlooked, information aggregationchannel on macroeconomic announcement days wherebyinvestors impound information about the true final value ofmacroeconomic series on announcement day.

5.3.3. Industrial production

A puzzle that emerged from the macroeconomicannouncement literature is the fact that the release ofindustrial production by the Federal Reserve does notappear to impact stock returns [see Flannery andProtopapadakis (2002) and Andersen, Bollerslev, Diebold,and Vega (2007), among others] even though there isstrong evidence of a long-term relation between returnsand industrial production.30 The theoretical and empiricalframework presented in this paper provide one solutionto this puzzle: If rational investors care more about thetrue value of industrial production than about the poten-tially inaccurate initial release, then the econometricianshould not necessarily observe a relation betweenannouncement-day returns and surprises but couldobserve a relation between returns and revisions.

Table 10 shows the results of similar regressions to thenonfarm payroll analysis:

rSPt ¼ aiþbiSur

IP

t þliSurCA

t þei,t ð18Þ

and

rSPt ¼ aiiþbiiSur

IP

t þliiSurCA

t þZiiRevIP

T,tþeii,t , ð19Þ

where IP stands for industrial production and CA standsfor capacity utilization, which are both released in thesame report at 8:30 a.m. EST. As before, I report theresults separately for expansionary and recessionarymonths, as defined by the NBER business cycle commit-tee. Due to data availability of market expectations, thesample starts in 1990 and I drop the last three years ofdata to ensure that revisions are mature enough.

In expansions, I do not observe a significant linkbetween announcement surprises and returns, which isin agreement with the previous literature. However, whenincluding the future revision to the data just released,RevIP, it does significantly improve the fit of the regres-sion. In recessions, the release of industrial productionand capacity utilization does move prices in a wayconsistent with the ‘‘good news is good news in badtimes’’ story. Nevertheless, the addition of the futurerevisions more than doubles the R2 of the regression andthe coefficient has the expected negative sign.

Overall, the evidence on industrial production suggeststhat the previous literature on macroeconomic announce-ments overlooked a completely separate channel for theinformation to impact prices: Investors expect revisionsto occur and announcement-day returns are correlatedwith the future revision of the data just released. Eventhough we do not necessarily observe a statistical relationbetween returns and surprises, the information released

30 Fama (1981, 1990) shows that real activity variables, such as

industrial production, are linked to stock returns, but with a lag.

Realizations of the macroeconomic variables one to two quarters ahead

can explain a significant portion of this month’s return variance.


does move prices by allowing investors to price in thefuture revisions.

6. Conclusion

If initial macroeconomic reports are inaccurate andinvestors do care about the true underlying value, whichis only revealed later, then rational expectations modelsof trade around public announcements predict a relationbetween announcement-day returns and the futurerevision of the released signal. By combining this predic-tion with well-established facts about the interpretationof information across the business cycle, I establishtwo testable hypotheses: In expansions, revisions andreturns should be positively related, and in recessions,the relation should be negative. Using data on nonfarmpayroll, GDP, and industrial production, I show that theselinks do exist and, therefore, provide evidence thatrevisions do matter and that investors care about thetrue final value of a macroeconomic release. They respondto the information conveyed by the initial release aboutthe correct value and not only its preliminary estimate.The economic significance of the results is highlighted bythe fact that the regression coefficients on the finalrevisions are of the same order of magnitude and oftenlarger than the coefficients on the initial surprises.Furthermore, I can explain a sizable portion of theannouncement-day variability in returns by includingthe future revisions.

These empirical findings yield fresh light on the exten-sive event study literature that attempts to explain pricemovements using, among others, public information. Thefindings in Roll (1988) are consistent with the low R2’sfound in studies such as Andersen, Bollerslev, Diebold, andVega (2007): an average R2 close to 1% using five-minutereturn windows. It could be that these event studies havelow power because they do not take into account thefuture revisions of the data just released. Rational investorsexpect revisions to occur and information about the truevalue is impounded into prices at the time of theannouncement, hence creating price volatility that ispotentially uncorrelated to the preliminary releases.

My empirical results suggest an alternate reason forthe Veronesi (2000) result that there is no premium fornoisy signals. Even if the public signal is imprecise,rational investors impound information about theexpected revisions on announcement day. In equilibrium,announcement-day market returns include the correctinformation with respect to the final value of the macro-economic variable and hence no premium is required.However, this contradicts the results of Savor and Wilson(2010), who show that market returns are significantlyhigher on days of employment, inflation, and interest rateinformation releases compared with all other days, sug-gesting a premium for facing macroeconomic announce-ment risk. More research is needed to uncover the type ofuncertainty faced by investors on days of public informa-tion releases and hence the type of compensationdemanded in equilibrium.

Finally, it would be valuable to analyze more industry-specific announcements and measure the information

aggregation process using sector portfolio returns (hous-ing, health care, etc.). One could further analyze how theinformation aggregates across sectors and into the overallindex, perhaps shedding further light on the findings ofHong, Torous, and Valkanov (2007), who show that thereturns on some industries lead the aggregate stockmarket. Using firm-specific information releases, Tetlock(2010) presents evidence that public news resolves asym-metric information by inducing the aggregation of pri-vately held information, which is consistent with myfindings at the macroeconomic level.

References

Adams, G., McQueen, G., Wood, R., 2004. The effects of inflation news onhigh frequency stock returns. Journal of Business 77, 547–574.

Ai, H., 2010. Information quality and long-run risk: asset pricingimplications. Journal of Finance 65, 1333–1367.

Andersen, T.G., Bollerslev, T., Diebold, F.X., Vega, C., 2003. Micro effectsof macro announcements: real-time price discovery in foreignexchange. American Economic Review 93, 38–62.

Andersen, T.G., Bollerslev, T., Diebold, F.X., Vega, C., 2007. Real-time pricediscovery in stock, bond and foreign exchange markets. Journal ofInternational Economics 73, 251–277.

Aruoba, S.B., 2008. Data uncertainty in general equilibrium. Unpublishedworking paper, University of Maryland, College Park, MD.

Balduzzi, P., Elton, E.J., Green, T.C., 2001. Economic news and bondprices: evidence from the U.S. treasury market. Journal of Financialand Quantitative Analysis 36, 523–543.

Boyd, J.H., Hu, J., Jagannathan, R., 2005. The stock market’s reaction tounemployment news: why bad news is usually good for stocks.Journal of Finance 60, 649–672.

Campbell, J.Y., Vuolteenaho, T., 2004. Bad beta, good beta. AmericanEconomic Review 94, 1249–1275.

Cenesizoglu, T., 2010. The reaction of stock returns to news aboutfundamentals. Unpublished working paper, HEC Montreal, Canada.

Cenesizoglu, T. Size, book-to-market ratio and macroeconomic news.Journal of Empirical Finance, forthcoming.

Chen, N.-F., Roll, R., Ross, S.A., 1986. Economic forces and the stockmarket. Journal of Business 59, 383–403.

Croushore, D., 2008. Revisions to PCE inflation measures: implicationsfor monetary policy. Unpublished working paper. University ofRichmond, Richmond, VA.

Easley, D., Hvidkjaer, S., O’Hara, M., 2002. Is information risk a determi-nant of asset returns? Journal of Finance 57, 2185–2221

Fama, E.F., 1981. Stock returns, real activity, inflation, and money.American Economic Review 71, 545–565.

Fama, E.F., 1990. Stock returns, expected returns, and real activity.Journal of Finance 45, 1089–1108.

Faust, J., Rogers, J.H., Wright, J.H., 2005. News and noise in G-7 GDPannouncements. Journal of Money, Credit, and Banking 37, 403–419.

Flannery, M.J., Protopapadakis, A.A., 2002. Macroeconomic factors doinfluence aggregate stock returns. Review of Financial Studies 15,751–782.

Fleming, M.J., Remolona, E.M., 1997. What moves the bond market?Federal Reserve Bank of New York Economic Policy Review Decem-ber, 31–50.

Fleming, M.J., Remolona, E.M., 1999. Price formation and liquidity in theU.S. treasury market: the response to public information. Journal ofFinance 54, 1901–1915.

Green, T.C., 2004. Economic news and the impact of trading on bondprices. Journal of Finance 59, 1201–1233.

Grundy, B.D., McNichols, M., 1989. Trade and the revelation of informa-tion through prices and direct disclosure. Review of Financial Studies2, 495–526.

Hautsch, N., Hess, D., 2007. Bayesian learning in financial markets:testing for the relevance of information precision in price discovery.Journal of Financial and Quantitative Analysis 42, 189–208.

Hong, H., Torous, W., Valkanov, R., 2007. Do industries lead stockmarkets? Journal of Financial Economics 83, 367–396

Kandel, E., Pearson, N.D., 1995. Differential interpretation of publicsignals and trade in speculative markets. Journal of Political Econ-omy 103, 831–872.

Kim, O., Verrecchia, R.E., 1991. Trading volume and price reactions topublic announcements. Journal of Accounting Research 29, 302–321.


Krueger, A.B., Fortson, K.N., 2003. Do markets respond more to morereliable labor market data? A test of market rationality. Journal ofthe European Economic Association 1, 931–957.

Kyle, A.S., 1985. Continuous auctions and insider trading. Econometrica53, 1315–1335.

Landefeld, J.S., Seskin, E.P., Fraumeni, B.M., 2008. Taking the pulse of theeconomy: measuring GDP. Journal of Economic Perspectives 22,193–216.

Mankiw, N.G., Runkle, D.E., Shapiro, M.D., 1984. Are preliminaryannouncements of the money stock rational forecasts? Journal ofMonetary Economics 14, 15–27

McQueen, G., Roley, V.V., 1993. Stock prices, news, and businessconditions. Review of Financial Studies 6, 683–707.

Mork, K.A., 1987. Ain’t behavin’: forecast errors and measurement errorsin early GNP estimates. Journal of Business and Economic Statistics5, 165–175.

Orazem, P., Falk, B., 1989. Measuring market responses to error-riddengovernment announcements. Quarterly Review of Economics andBusiness 29, 41–55.

Pasquariello, P., Vega, C., 2007. Informed and strategic order flow in thebond markets. Review of Financial Studies 20, 1975–2019.

Pearce, D.K., Roley, V.V., 1985. Stock prices and economic news. Journalof Business 58, 49–67.

Roll, R., 1984. Orange juice and weather. American Economic Review 74,861–880.

Roll, R., 1988. R2. Journal of Finance 43, 541–566.Savor, P., Wilson, M., 2010. How much do investors care about macro-

economic risk? Evidence from scheduled economic announcements.Unpublished working paper, University of Pennsylvania, Philadel-

phia, PA.Stock, J.H., Watson, M.W., 2002. Macroeconomic forecasting using

diffusion indexes. Journal of Business & Economic Statistics 20,

147–162.Tetlock, P.C., 2010. Does public financial news resolve asymmetric

information? Review of Financial Studies 23, 3520–3557.Veronesi, P., 2000. How does information quality affect stock returns?

Journal of Finance 55, 807–837.

information aggregation around macroeconomic announcements: revisions matter

Documents