journal of financial economics - fuqua school of … › ~mbrandt › papers › published › ...at...

Contents lists available at ScienceDirect

Journal of Financial Economics

Journal of Financial Economics 117 (2015) 489–507

http://d0304-40

☆ WeFabio Fopants atBusinesCollegethe Uniindebte

n CorrUnited K

E-m

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

Distilling the macroeconomic news flow$

Alessandro Beber a,b,n, Michael W. Brandt c,d, Maurizio Luisi e

a Cass Business School, City University London, United Kingdomb CEPR, United Kingdomc Fuqua School of Business, Duke University, United Statesd NBER, United Statese Bloomberg LP, United Kingdom

a r t i c l e i n f o

Article history:Received 19 November 2013Received in revised form20 March 2014Accepted 4 April 2014Available online 27 May 2015

JEL classification:G12

Keywords:Macroeconomic newsForecastingNowcastingDisagreement

x.doi.org/10.1016/j.jfineco.2015.05.0055X/& 2015 Elsevier B.V. All rights reserved.

thank an anonymous referee, Daryl Caldwrnari, Amit Goyal, Ana-Maria Tenekedjieva,BlackRock, City University, the 2012 Asset Ps School, the Fall 2012 Inquire UK ConferenceHedge Fund Conference, the Stockholm Schoversity of York, for their comments andd to Inquire UK for financial support.esponding author at: Cass Business School, Ciingdom. Tel.: þ44 20 7040 8737; fax: þ44ail address: [email protected] (A

a b s t r a c t

We propose a simple cross-sectional technique to extract daily factors from economicnews released at different times and frequencies. Our approach can effectively handle thelarge number of different announcements that are relevant for tracking current economicconditions. We apply the technique to extract real-time measures of inflation, output,employment, and macroeconomic sentiment, as well as corresponding measures ofdisagreement among economists about these indices. We find that our procedureprovides more timely and accurate forecasts of future changes in economic conditionsthan other real-time forecasting approaches.

& 2015 Elsevier B.V. All rights reserved.

1. Introduction

Timely measurement of the state of the economy reliestraditionally on low-frequency observations of a few eco-nomic aggregates referring to previous weeks, months, oreven quarters. A prominent example is the advance esti-mate of Gross Domestic Product (GDP) released quarterlyabout a month after the end of the quarter. The lowfrequency and delayed observation of any such economicaggregate considered in isolation stands in sharp contrast

ell, Robert Darwin,and seminar partici-ricing Retreat at Cassin Bath, the Imperialol of Economics, andsuggestions. We are

ty University London,20 7040 8881.. Beber).

with the rich macroeconomic news flow that marketparticipants observe almost daily. This news flow containsinformation that agents use to learn about the economy inthe absence of private information. In particular, the financeliterature has identified a large cross-section of dozens ofdifferent news releases that have significant and immediateeffects on financial markets (e.g., Andersen, Bollerslev,Diebold, and Vega, 2003).

We distill the economic news flow observed by marketparticipants into a small set of indicators describing fourdistinct aspects of the economy: inflation, output, employ-ment, and macroeconomic sentiment. Specifically, wepropose a simple cross-sectional technique to extract dailyprincipal components from economic news releases asso-ciated with a given information type and observed atdifferent times and frequencies. Our approach is simple,robust (no numerical optimization is required), and caneffectively handle the large number of announcementsthat are relevant for tracking the evolution of economicconditions in real-time. At the same time, our empirical

www.sciencedirect.com/science/journal/0304405X

www.elsevier.com/locate/jfec

http://dx.doi.org/10.1016/j.jfineco.2015.05.005



http://crossmark.crossref.org/dialog/?doi=10.1016/j.jfineco.2015.05.005&domain=pdf



mailto:[email protected]


A. Beber et al. / Journal of Financial Economics 117 (2015) 489–507490

analysis shows that the output of our approach is moretimely and informative than more sophisticated but alsomore difficult-to-implement statistical techniques. Intui-tively, the potential disadvantage of a simpler modelingapproach is more than compensated for by the sheerquantity of data our approach can effectively incorporate.

Our paper relates to the literature on measuring thestate of the economy in a time-series setting based only onfundamental economic data [see Banbura, Giannone,Modugno, and Reichlin, 2013 for a survey], commonlyreferred to as “nowcasting.” There are two generalapproaches to this problem. The first approach is to use abalanced panel regression, along the lines of the seminalpaper of Stock and Watson (1989). The purpose of this firstapproach is to construct a coincident index of economicactivity using factor models on a large set of macroeco-nomic releases, which basically amounts to constructing aweighted average of several monthly or quarterly indica-tors. The advantage of this technique is that the resultingindex is based on many macroeconomic variables. How-ever, this advantage also results in a relatively low mea-surement frequency, because the econometrician has towait for the panel to be complete before the index can beconstructed. A second general approach is to modelmacroeconomic data using a state-space model (e.g.,Evans, 2005). The advantage of this second approach isto produce an indicator at a higher frequency, since astate-space model can effectively handle the sparse anddelayed reporting of economic data and missing informa-tion on non-release days. However, this technique isimpractical for large cross-sections of macroeconomicreleases. For example, Evans (2005) only considers theset of different (preliminary, advance, and final) GDPreleases. Aruoba, Diebold, and Scotti (2009) propose abusiness condition index, called the Arouba Diebold Scotti(ADS) index, constructed using four indicators at differentfrequencies, including a continuously observable financialmarkets variable. Finally, Giannone, Reichlin, and Small(2008) combine the two approaches by modeling factorsextracted from a balanced panel of monthly releases in astate-space setting.

Our goal is to measure the state of the economy with amethodology that broadly retains the advantages of bothapproaches without their respective limitations. Specifi-cally, we consider a large universe of macroeconomicannouncements. This is a crucial aspect of our methodol-ogy, given the evidence of many influential releases fromthe macroeconomic announcement literature. At the sametime, our approach can handle data released at differentfrequencies and missing observations to produce a real-time high-frequency measurement of the state of theeconomy.

Our methodology has several other differentiatingfeatures relative to the literature. First, we do not aim toestimate a real-time series of GDP, for example, but werather leave the macroeconomic factor(s) truly latent andunspecified. In this sense, we do not impose any structureon the estimation and thus do not take a stand on what isthe appropriate metric of the state of the economy. Wesimply let the data speak for itself. Second, our focus on alarge cross-section of economic news releases allows us to

extract factors from four subsets of macroeconomic news(e.g., inflation, output, employment, and macroeconomicsentiment). We use these subset indicators to learn aboutthe relations between different driving forces of theeconomy. Third, we utilize news flow data that are trulyreal-time and unrestated, as opposed to approximatelydated historical data that are often restated [e.g., Koenig,Dolmas, and Piger, 2003; see also Ghysels et al., 2012, foran illustration of the issues arising from restated macro-economic data]. Fourth, we refrain from using any finan-cial market-based data, as our aim is to objectivelymeasure the macroeconomic news flow absent any ofthe market's interpretation of the same. Finally, we alsoapply our methodology to the dispersion of economicforecasts as a newway to obtain a high-frequency measureof macroeconomic uncertainty based on the disagreementof a cross-section of economic experts. In summary, ourfairly simple and data-driven method delivers a real-time,daily, unbiased, and objective reading of the state of themacroeconomy, which can be used for a number ofpurposes, most notably to study the relation betweenfinancial market and economic dynamics.

We find that an economic activity factor (which com-bines output and employment information, as they arehighly correlated) as well as a macroeconomic sentimentfactor, both extracted from the large cross-section ofmacroeconomic news, have sensible dynamics. The great-est dips in both series are well aligned with the ex postdefined National Bureau of Economic Research (NBER)recession periods. The macroeconomic sentiment factor,obtained from consumer and business confidence releases,is highly correlated with economic activity, but appears tolead fundamentals especially around important turningpoints. Finally, our inflation factor exhibits dynamics thatseem only weakly correlated with growth, with muchmore erratic variation, and has an unclear pattern inexpansions versus recessions.

Our empirical proxy of economic uncertainty based oneconomic expert disagreement is interesting for at leasttwo reasons. First, it shows little correlation with theestimates of the latent economic activity, macroeconomicsentiment, and inflation factors, suggesting that they arelikely to contain different information. Second, and moreimportantly, macroeconomic uncertainty exhibits intri-guing asymmetric dynamics. The peaks of disagreementcorrespond to the final stages of recession periods, whileuncertainty is relatively subdued at the end of economicexpansions. This evidence suggests that economists tendto disagree mostly on recoveries from prior contractions,whereas everyone seems to see the end of an economicexpansion coming.

We formally relate a real-time factor of economicgrowth (which further aggregates the information relativeto economic activity by combining information relating tooutput, employment, and macroeconomic sentiment) tovintages of the Chicago Fed National Activity Index(CFNAI), constructed by the Chicago Federal Reserve Boardbased on Stock and Watson (1989), on CFNAI release datesat the monthly frequency, and to the vintage version of theADS index of Aruoba, Diebold, and Scotti (2009) at theweekly frequency. We find that our latent growth factor is

2 We emphasize the fact that we work with distinct announcementsbecause there are a lot more than 43 statistics if we included multipleversions of essentially the same data released in the same economicreport. For example, the CFNAI uses 13 industrial production statistics,resulting in 20% of the index being determined by a single release. Incontrast, we include in our analysis only the headline month-over-monthfigure.

3 The importance of using real-time versus final data in macroeco-nomic forecasting has been discussed extensively in the literature (e.g.,Koenig, Dolmas, and Piger, 2003). In our real-time framework, revisions

A. Beber et al. / Journal of Financial Economics 117 (2015) 489–507 491

strongly correlated to both of these alternative approaches.However, since our factor is constructed using informationfrom either a larger cross-section of news or in a moretimely manner, it turns out to have significant forecastingpower for both CFNAI and the ADS index beyond their ownlags. We also find that our growth factor has predictivepower for future actual GDP releases and is highly corre-lated with the quarterly GDP expectations in the Survey ofProfessional Forecasters (SPF). This is a remarkable featuregiven that, unlike the ADS index, our growth factor is notoptimally weighted to forecast GDP. The large correlationwith the quarterly releases of the SPF offers an intuitiveinterpretation of our growth factor as the high-frequencydaily reading of economist expectations about macroeco-nomic fundamentals.

We extend this empirical analysis to the real-timeinflation factor extracted from inflation-related announce-ments. This is a novel aspect of our analysis, as the extantapproaches generally ignore these releases to optimallyforecast GDP and growth.1 Our real-time inflation factorgenerally seems to lead the pattern of Consumer PriceIndex (CPI) actual releases and the inflation forecastcontained in the SPF, albeit in a relatively noisy fashion.More specifically, we find that our inflation factor observedon quarterly SPF release dates has predictive power for theupcoming actual CPI announcement, beyond CPI's ownlags and the median SPF inflation forecast.

Another intriguing finding is that our latent factorsobtained exclusively from macroeconomic information arehighly correlated with financial indicators, such as thedefault spread and the implied stock return volatility indexVIX. More specifically, we find that the combination of ourlatent growth factor and its dispersion can explain almostone-third of VIX levels. This is an important finding in lightof the documented difficulties for macroeconomic quan-tities to explain financial market volatility [see, for exam-ple, the seminal paper of Schwert, 1989].

Finally, we combine the information of the growthindicator and its dispersion extracted from economistdisagreement, and document very strong predictabilityfor future growth, from five days and up to six monthsahead. Given the illustrated relation of our macroeconomicindicator with financial variables and its extremely timelynature, this result suggests that our quantitative measureof the news flow could have predictive power for futurefinancial market dynamics.

The remainder of the paper proceeds as follows. InSection 2, we describe the macroeconomic news and wecarry out some preliminary analysis on macroeconomicannouncements. Section 3 explains our methodology forestimating in real-time the state of the economy and itsuncertainty. We present our empirical results in Section 4.Section 5 concludes with a summary of our findings.

1 For example, the 85 macroeconomic indicators used to constructthe CFNAI are drawn from production, employment, consumption, andsales categories, but none of them is drawn from a nominal inflation-related category.

2. Data and preliminaries

2.1. Macroeconomic news and forecasts

We obtain data on the dates, release times, and actualreleased figures for 43 distinct U.S. macro-economicannouncements covering the period from January 1997through December 2011, for a total of more than 8,000announcements over about 3,800 business days. Thesedata are obtained from Bloomberg through the EconomicCalendar screen, which provides precisely time-stampedand unrestated announcement data.2,3 We also collect dataon economist forecasts for each announcement. Bloom-berg surveys economists during the weeks prior to therelease of each indicator to obtain a consensus estimate.We work with the individual economist-level forecasts,rather than the aggregated consensus forecasts, in order toconstruct cross-sectional measures of disagreement foreach news release.

Bloomberg contains data for many of our series prior to1997, but those data are stored in historical fields which (a)are not associated with clear announcement dates andtimes (rather they are dated according to the period theyreference) and (b) are restated over time.4 We collect thismore problematic data for January 1985 through 1996 fortwo reasons. First, we use these historical data to constructan initial correlation matrix estimate, which is required byour methodology (see Section 3). Second, we use thesedata for a robustness check with a longer sample period(see Section 4.5). In order to date the releases prior to1997, we compute for each news series the median timebetween the reference period and the announcement. Forexample, the employment report is traditionally releasedfour days after the end of the month to which the reportrefers. We then apply this median reporting lag to thereference period of the older data in order to obtain anapproximate announcement date.

Since economist-level forecasts are not available priorto 1997, we instead collect data from the Survey ofProfessional Forecasters (SPF). The SPF is the oldest quar-terly survey of macroeconomic forecasts in the UnitedStates. The survey began in 1968 and was conducted by

and restatements could be used as new information that becomesavailable on the date of the restatement release. However, since restate-ments are typically announced contemporaneously with new initialreleases, we focus exclusively on the latter.

4 For example, there are monthly releases of quarterly GDP labeled“advance,” “preliminary,” and “final” all referring to the same quarter.Bloomberg's historical field for GDP is dated according to the referencedquarter, so that the advance release gets overwritten by the preliminaryrelease, which in turn gets overwritten by the final release. Historically,only the final releases are stored.

Conference Board Consumer ConfidenceChicago Purchasing Managers Index Inflation

University Michigan Consumer Survey EmploymentADP National Employment Report Output

ISM Manufacturing PMI Macro sentimentNonfarm Payrolls Total,Manufacturing + Unemployment Rate + Average Weekly HoursISM Non-Manufacturing PMI

Retail Sales + Retail Sales Less AutoImport Price Index

PPI + PPI CoreIndustrial Production + Capacity UtilizationEmpire State Manufacturing Survey

Manufacturing Trade InventoriesCPI + CPI Core

Durable Goods OrdersConference Board Leading Index

GDP + GDP Price IndexPersonal Income + Pers. Consum. Exp. + PCE Price Index

Manufacturers New Orders

24 26 28 30 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 1 3 5 7 9 11 13 15 17 19 21 23Reference

+MhtnoMMhtnoM 1 Month M +2

Fig. 1. This figure shows the typical reporting structure for a large cross-section of U.S. macroeconomic announcements. On the horizontal axis, werepresent the days of the reference month M and the subsequent two months. On the vertical axis, we list the macroeconomic releases in order ofreporting, highlighting in bold the typical reporting period. The macroeconomic announcements are broken down in the four aggregates of inflation news(underlined font), employment news (normal font), output news (bold font), and macro-sentiment news (italic font).

6 The economy is often separated into nominal and real sidesbecause shocks to the two should be treated differently from a policyperspective. For example, many argue, from the perspective of monetarypolicy, that nominal shocks should be minimized, whereas real shocksshould not be intervened upon. Other studies also suggest that a nominaland a real factor can jointly account for much of the observed variation in


the American Statistical Association and the NationalBureau of Economic Research. The Federal Reserve Bankof Philadelphia took over the survey in 1990. The SPF'sWeb page offers the actual releases, documentation, meanand median forecasts of all the respondents, as well as theindividual responses from each economist. The individualresponses are kept confidential by using identificationnumbers.

Most macroeconomic indicators are released on differ-ent days and at different frequencies, making it difficult toprocess the flow of information in a systematic andconsistent way. Fig. 1 shows that actual news releasesoccur with a variety of different lags with respect to themonth they are referencing. Furthermore, news on differ-ent indicators are frequently released simultaneously.5 Forexample, the employment report traditionally announcedon the first Friday of the month contains four differentindicators: nonfarm payrolls, nonfarm payrolls in themanufacturing sector, the unemployment rate, and aver-age weekly hours. Finally, the release frequency variesacross different economic aggregates. Data releases ofdifferent economic indicators are usually observed atdifferent frequency; e.g., GDP data are sampled quarterly,the nonfarm payrolls are released monthly, initial joblessclaims are sampled weekly, etc. These features of our largecross-section of macroeconomic news releases generate asparse matrix of data that our methodology will have totake up. The Appendix A describes in detail the set ofmacroeconomic news in our sample, including their fre-quency, source, and units of measurement.

5 On approximately 80% of days, there was at least one data release.Multiple data releases occurred much less frequently, on approximately60% of the days in the sample.

2.2. Categorizing the macroeconomic news flow

Our aim is to extract a set of factors describing the stateof the economy. Rather than relying on a statisticalprocedure to obtain orthogonalized factors that areincreasingly difficult to interpret with the order of thefactor, we impose a specific economically motivated struc-ture on the macroeconomic news flow. Based on bothempirical evidence and economic rationale, we first sepa-rate the aggregate economy into two broad dimensions:the nominal and the real side.6 In practice, we split theset of announcements into nominal inflation-relatedannouncements and news that relates to real growth.Growth data, in turn, come in two flavors — objectiverealizations of past economic activity and subjective, oftenforward-looking, views derived from surveys which welabel “macro sentiment.”7 Finally, economic activity can besplit one last time into information relating to outputversus employment.

Through this structure, we obtain two (inflation andgrowth), three (inflation, economic activity, and macrosentiment), or four (inflation, output, employment, and

major economic aggregates.7 The behavioral finance and economics literature tends to associate

the term sentiment with emotions that in a rational framework shouldnot affect decisions. We take a broader perspective and use the termsentiment to encompass agents’ subjective forward-looking interpreta-tion of the data as revealed through surveys.

j =1 j =2 … j =5 j =6 … j =N

1 ... ... ... Missing ... ... ...

… ... ... ... Missing ... ... ...

… ... ... ... Missing ... ... ...t -22 At-22,1 Not released ... Missing Not released ... ...

t -21 Not released At-21,2 ... Missing At-21,6 ... ...

… Not released Not released ... Missing Not released ... ...

t At,1 Not released ... At,5 Not released ... ...

t +1 Not released At+1,2 ... Not released At+1,6 ... ...

… Not released Not released ... Not released Discontinued ... ...

… ... ... ... ... Discontinued ... ...

T ... ... ... ... Discontinued ... ...

j =1 j =2 … j =5 j =6 … j =N

1 ... ... ... Missing ... ... ...

… ... ... ... Missing ... ... ...

… ... ... ... Missing ... ... ...t -22 At-22,1 E[At-22,2]=At-43,2 ... Missing E[At-22,6]=At-43,6 ... ...

t -21 E[At-21,1]=At-22,1 At-21,2 ... Missing At-21,6 ... ...

… E[A ...,1]=At-22,1 E[A...,2]=At-21,2 ... Missing E[A...,2]=At-21,6 ... ...

t At,1 E[A t,2]=At-21,2 ... At,5 E[A t,2]=At-21,6 ... ...t +1 E[At+1,1]=At,1 At+1,2 ... E[At+1,5]=At,5 At+1,6 ... ...

… E[A ...,1]=At,1 E[A ...,2]=At+1,2 ... E[A ...,5]=At,5 Discontinued ... ...

… ... ... ... ... Discontinued ... ...

T ... ... ... ... Discontinued ... ...

Fig. 2. This figure shows a stylized example of the actual macroeconomic announcement data, for N announcement types over a daily sample periodbetween 1 and T. The releases j¼1 and j¼2 are monthly indicators released on two different days of the month. The macroeconomic indicator j¼5 is anews release that did not exist at the beginning of the sample, but was included in the sample from day t onwards. The macroeconomic indicator j¼6 didexist at the beginning of the sample, but was subsequently discontinued. The top panel represents the matrix of the actual macroeconomic releases in real-time as it is constructed from the data. The bottom panel shows how our simple forward-filling algorithm is used to fill in the expectation of the indicatorwhen it is not released.


macro sentiment) factors:

� Inflation

� GrowthEconomic activity

OutputEmployment

(

Macro sentiment

8><>:

where, for example, the economic activity factor isobtained from the combined information relating to out-put and employment. In that sense, the information isnested from right to left.

More specifically, the inflation factor is extracted fromthe news flow of nine inflation-related releases: consumerprice index, CPI ex food and energy, employment costindex, GDP price index, import price index, nonfarmproductivity, personal consumption expenditure core priceindex, producer price index (PPI), and PPI ex food andenergy. For the output factor, we utilize information fromboth the supply and demand side of the economy in theform of news about advance retail sales, business inven-tories, capacity utilization, consumer credit, domesticvehicle sales, durable goods orders, durables ex-transpor-tation, factory orders, GDP, industrial production, Instituteof Supply Management (ISM) manufacturing, ISM non-manufacturing composite, personal consumption, personal

income, personal spending, retail sales less autos, andwholesale inventories. Employment news is captured byreleases of ADP payrolls, manufacturing payrolls, nonfarmpayrolls, continuing claims, initial jobless claims, and theunemployment rate. Finally, we extract the macro senti-ment factor from the information in ten macroeconomicsurveys: ABC consumer confidence, Chicago purchasingmanager, consumer confidence, Dallas Fed manufacturingactivity, Empire manufacturing survey, leading indicatorsindex, ISM-Milwaukee, Philadelphia Fed business outlooksurvey, Richmond Fed manufacturing index, and the Uni-versity of Michigan confidence index. The Appendix Asummarizes the assignment of announcements to the fourcategories: inflation, output, employment, and macrosentiment.

It is worth reiterating at this point that we do notinclude any market-based data (such as stock prices,interest rates, credit spreads, or VIX) in our analysis,unlike, for example, Aruoba, Diebold, and Scotti (2009)and Giannone, Reichlin, and Small (2008). While such dataare very timely and undoubtedly informative about thestate of the economy, they already represent the market'sinterpretation of the macroeconomic news flow. Our aim isto objectively summarize and describe the macroeconomicnews flow itself.


2.3. Transformation and temporal alignment

We examine the stationarity of each data series in twoways. First, we conduct a Dickey-Fuller test on each series.Second, we read the definition and description of eachstatistic to determine from an economic perspectivewhether it is a non-stationary index or a stationaryquarterly growth rate, for example. In a few cases wherethe conclusions from the two approaches differ, usuallybecause the available data are too short to examinestatistically, we rely more on the description to determinewhether the series is stationary. All series that are deemednon-stationary are first-differenced in news release time.The Appendix A contains more details.

The final data management task is to align the datatemporally by moving from announcement time to calen-dar time. We do this by populating the news releases in aT�N matrix where T denotes the total number of weekdays in our sample and N refers to the 43 announcementtypes. The data at this stage look like the top panelof Fig. 2.

There are two important aspects of the data to discuss.First, there are a vast number of missing values, as we canthink of each news series as a continuously evolvingstatistic that is observed only once per month or quarter.Second, not all announcements have a complete history.Some announcements are initiated in the middle of thesample and/or are terminated before the end of thesample. To solve the missing data problem, we simplyforward-fill the last observed release until the nextannouncement. Forward-filling can be rationalized asreplacing missing values with expected values under asimple independent random walk assumption for eachnews series. Of course, both independence in the cross-section and random walk dynamics through time aresimplifying assumptions that are rejected by the data (infact, the motivation for our methodology described belowis the cross-sectional correlation structure within newscategory). A more sophisticated approach for filling inmissing data would be to compute the expectation of themissing values given the full cross-section of previousreleases as well as the cross-sectional and intertemporalcorrelation structure of the data. An optimal solutionwould also allow for sampling error, which is the case inKalman filter or Bayesian data augmentation algorithms.However, there is a clear trade-off between statisticalcomplexity and ability to process a large cross-section ofnews series. Since the goal of our approach is to utilize theentire cross-section of news, we choose a very simplestatistical model for filling in missing observations. Afterforward-filling, the data look like the bottom plot of Fig. 2.8

Note that the second data issue, the fact that someseries do not span the entire sample period, cannot besolved with missing values imputation. It is instead expli-citly addressed in our methodology below.

8 The forward-filling could potentially accommodate data revisionsor restatements on the day they occur, if the restatements are notcontemporaneous with subsequent initial releases.

3. Methodology

3.1. Subset principal component analysis

Our goal is to extract from the cross-section of macro-economic news releases a set of factors that capture inreal-time the state of inflation, output, employment, andmacro sentiment, as well as the two more overarchingfactors measuring economic activity and growth. The mostobvious ways of accomplishing this, full data principalcomponents analysis (PCA) and forecasting regressions, donot appeal to us. First, with full data PCA we obtain factorsthat are mechanically orthogonal, whereas the dimensionsof the economic news flow we want to capture are likelycorrelated (e.g., output and employment are both high atthe peak and low at the trough of an economic cycle). Thisorthogonalization makes it practically impossible to assignan economic meaning to higher order factors. Second,trying to identify the factors through predictive regres-sions on a candidate variable in each category, such as finalGDP for output, would require us being able to identify asingle series that represents each category. While this is acommon approach in the nowcasting literature, it relies onex ante knowledge of the key statistic to track and assumesthat there is only one such statistic that does not changeover time (see also Stock and Watson, 1989).

Instead, we rely on our ex ante categorization of thenews and, within each category subset, let the data speakfor itself by extracting the first principal component of thatsubset of data. Specifically, on each day of our sample t, weobtain for each news category i the first principal compo-nent from the correlation matrix Ωt;i of the stationarynews series in category i. We work with the correlationmatrix to abstract from arbitrary scaling of data. Moreover,in order to obtain a real-time measure, we use a telescop-ing (with a common historical start date and rolling enddates) correlation matrix starting in 1980.9 We denote theNi � 1 principal component weights by ct;i, where Ni is thenumber of news series in category i. Consistent withextracting principal components from a telescoping corre-lation matrix, we standardize the news series using tele-scoping estimates of their means and standard deviations.

3.2. Economic new series correlation matrix

The key inputs to our methodology are the within newscategory correlation matrices Ωt;i. Specifically, we need tocalculate from historical data up through date t thecorrelation of all news series of category i that are “active”on that date, where active means that the news series waspreviously initiated and has not yet been terminated.There are two issues that need to be addressed in comput-ing these correlation matrices. First, the data are in theform of an unbalanced panel due to some of the seriesbeing initiated after the start date of the estimationwindow (e.g., series j¼5 in Fig. 2). Second, the data are

9 We also experimented with fixed-window-size rolling correlationmatrices for five, ten, 15, and 20 years. The results are qualitativelysimilar, particularly for the longer data windows.

Table 1Summary statistics.

Panel A shows correlations between daily observations of six real-time macroeconomic indices and their respective economist forecast dispersions. Panel B reports additional summary statistics and correlationsbetween the growth index, growth dispersion, and a set of financial variables. Specifically, Rmt�Rft denotes the log return on the S&P 500 index in excess of the 3-month Treasury-bill rate, VIX is the Chicago Boardof Options Exchange (CBOE) option-implied volatility index, lnðP=EÞ and lnðD=PÞ are the log price–earnings ratio and log dividend yield, Def is the default spread (Moody's BAA minus AAA corporate bond yields),Term is the term spread (10-year minus 3-month Treasury yields). The sample period is January 1997 to December 2011.

Panel A:

Index Dispersion

Inflation Output Employment Sentiment Economic activity Growth Inflation Output Employment Sentiment Economic activity Growth

Index Inflation 1.00 0.36 0.13 0.14 0.25 0.22 �0.08 �0.18 �0.16 �0.18 �0.19 �0.20Output 1.00 0.84 0.82 0.96 0.95 �0.53 �0.22 �0.34 �0.19 �0.26 �0.26Employment 1.00 0.75 0.96 0.92 �0.53 �0.12 �0.30 �0.15 �0.16 �0.16Sentiment 1.00 0.82 0.93 �0.50 0.07 �0.05 �0.07 0.06 0.05Economic activity 1.00 0.97 �0.55 �0.18 �0.34 �0.18 �0.23 �0.23Growth 1.00 �0.56 �0.10 �0.25 �0.15 �0.13 �0.14

Dispersion Inflation 1.00 0.26 0.15 0.14 0.25 0.25Output 1.00 0.52 0.40 0.98 0.96Employment 1.00 0.35 0.68 0.68Sentiment 1.00 0.42 0.54Economic activity 1.00 0.99Growth 1.00

Panel B:Growth Growth

Rmt�Rft Index Dispersion VIXln

PE

� �ln

DP

� �Def Term

Summary statistics Mean 0.63 �0.04 �0.00 0.23 2.98 0.55 1.03 1.68Std deviation 21.42 1.15 0.99 0.09 0.23 0.25 0.48 1.30Skewness �0.20 �1.30 1.40 1.78 0.05 0.48 2.82 �0.06Kurtosis 9.77 4.97 4.19 8.85 2.19 3.56 12.25 1.66

Correlation matrix Rmt�Rft 1.00 0.01 0.01 �0.13 0.04 �0.03 �0.01 0.01Growth index 1.00 �0.14 �0.51 0.55 �0.71 �0.84 �0.51Growth dispersion 1.00 0.25 0.19 0.09 0.14 0.02VIX 1.00 �0.14 0.31 0.63 0.22lnðP=EÞ 1.00 �0.85 �0.52 �0.28lnðD=PÞ 1.00 0.69 0.42Def 1.00 0.40Term 1.00

A.Beber

etal./

Journalof

FinancialEconom

ics117

(2015)489

–507495


naturally persistent, partly due to autocorrelation of thedata in announcement time, partly due to the cross-sectional misalignment of the news in calendar time, andlargely due to the forward-filling of missing data.

We address the first unbalanced panel issue by using acorrelation matrix estimator along the lines of Stambaugh(1997), who shows how to adjust first and secondmoments estimates for unequal sample lengths. The intui-tion of his approach is to use the observed data on thelonger series, along with a projection of the shorter seriesonto the longer ones estimated when both are observed, toadjust the moments of the shorter time series.

To correct for the persistence, we could use the stan-dard approach of Newey-West (1987), where due to thenature of the data we would account for up to one quarterof autocorrelation and cross-autocorrelation. Unfortu-nately, the kind of persistence in our data is not ideallycaptured by the nonparametric Newey-West approach fortwo reasons. First, we have daily data, so adjusting for upto a quarter of autocorrelation would involve approxi-mately 60 cross-autocorrelation matrices. Second, the(cross-) autocorrelations are not exponentially decayingas a typical AutoRegressive Moving Average (ARMA) modelmight predict. Instead, the data are locally constant, due tothe forward-filling, and over longer intervals only moder-ately (cross-) autocorrelated due to the statistical nature ofthe news series.

This peculiar correlation structure of economic newsforward-filled onto a daily calendar is actually identical tothat found in high-frequency asset prices, where asyn-chronous and infrequent trading creates a misaligned andlocally constant panel of observations. In that literature,Ait-Sahalia, Mykland, and Zhang (2005) propose a “two-scales realized volatility” estimator to handle this specificstructure of short-term constancy versus long-horizonweak dependence. Specifically, their estimator subsamplesthe data at a sufficiently low frequency that overcomes thelocal constancy and then averages over the set of allpossible estimators that start the subsampling schemesat different times.

We adopt exactly the same approach, except of courseour application is very different. Specifically, at date t wesubsample the forward-filled news series backward at amonthly frequency and then compute a Newey-Westestimate of the correlation matrix using four lags. Werepeat the same for monthly sampling starting at datesft�1; t�2;…; t�dþ1g (assuming d days per month) andthen average the resulting d correlation matrix estimates.

3.3. Level versus disagreement factors

Given the vector of principal component weights ct;iobtained with our methodology, we then construct foreach news category two time series. First, we sum at eachdate the product of the weights multiplied by the mostrecent releases to obtain our real-time level factors. Sec-ond, we sum the product of the same weights multipliedthis time by the cross-sectional standard deviations of theeconomist forecasts for the most recent releases to obtainour real-time disagreement factors. Throughout our samplenot every news series has economist-level forecast data

available. We therefore construct the disagreement factorusing the available data, re-normalizing first the principalcomponent weights to account for the proportion ofmissing data.

4. Results

We first describe empirically the dynamics of the real-time macroeconomic factors. To get a sense for how ourmethodology compares to other approaches, we thenrelate our growth factor to the vintage releases of theCFNAI and ADS index. We analyze whether our real-timegrowth factor actually predicts subsequent GDP releases,comparing it to the predictability by the correspondingSPF forecasts. Along the same lines, we examine our real-time inflation factor and analyze whether it predictssubsequent CPI releases relative to the SPF forecasts. Wethen examine the relation between the growth factor andits dispersion with volatility in financial markets. Thislatter analysis is motivated by the apparent lack of a strongrelation between real activity and financial market volati-lity (e.g., Schwert, 1989). Finally, we examine the jointdynamics of our real-time growth index, growth disper-sion, inflation index, and inflation dispersion, and we alsoextend the sample backward using a pseudo real-timeapproach as a robustness check.

4.1. Preliminaries

In Panel A of Table 1, we present correlations betweenthe seven real-time macroeconomic indices and theirrespective economist forecast dispersions, which we inter-pret as proxies for macroeconomic uncertainty. There are anumber of interesting observations. First, inflation isrelatively uncorrelated with the other macroeconomicindices. Its highest correlation is 0.36 with output. Incontrast, output, employment, and sentiment are highlycorrelated with each other (correlations ranging from 0.75to 0.84) and are each even more highly correlated with thecomposite indices for economic activity and growth. Thecorrelations with the growth index, in particular, rangefrom 0.92 to 0.95. We conclude from these high correla-tions that the growth index contains most of the informa-tion revealed by output, employment, and sentiment, andwe therefore focus on examining the aggregated growthindex and its dispersion going forward.

Second, the correlations between macroeconomicuncertainty mimics the general patterns we observe inthe indices, but at somewhat lower levels, particularly forsentiment. For example, the correlations between outputdispersion, employment dispersion, and sentiment disper-sion with growth dispersion are 0.96, 0.68, and 0.54,respectively.

Finally, we observe an interesting negative correlationbetween the levels and dispersions of our real-timemacroeconomic factors. The correlations are generallysmall in magnitude, except for inflation uncertainty, whichis �0.5 and more highly correlated with the level ofgrowth and its components (output, employment, andmacro sentiment). This suggests that at times of strong(weak) growth, the uncertainty about inflation is low


(high). In contrast, the state of inflation seems irrelevantfor the uncertainty about the other real-time macroeco-nomic indices.

In Panel B, we compute contemporaneous correlationsbetween excess stock market returns, the growth factor,the dispersion of growth forecasts, and a number offinancial market variables associated in the literature withthe state of the economy or macroeconomic uncertainty.10

There is no meaningful contemporaneous correlationbetween our real-time macroeconomic factor and stockmarket returns. In contrast, there are significant contem-poraneous correlations between the growth factor and anumber of financial variables, most notably the correlationwith VIX (�0.51), the dividend yield (�0.71), and thedefault premium (�0.84). Growth dispersion has a weakerrelation with the financial variables, but it still retains asignificant correlation with VIX (0.25) and with the price–earnings ratio (0.19). These descriptive results foreshadowthe link between our real-time growth factor and financialmarket volatility, proxied here by VIX, that we investigatemore thoroughly in Section 4.3.11

Figs. 3–6 provide graphical descriptions of our real-time macroeconomic indices. The upper panel of Fig. 3starts by plotting the estimated real-time output andemployment factors. The gray areas in the plots representNBER recessions. Output seems to anticipate employmentsomewhat, especially around business cycle turns, but thetwo factors are very highly correlated. For this specificcomparison, we extend the sample to the end of 2013 tostudy more closely the recovery out of the most recentrecession. The lower panel of Fig. 3 shows that theemployment factor lagged behind output during therecovery until the end of 2011, suggesting that growthwas occurring without a comparable improvement in jobs.Since then, however, employment caught up with theoutput factor in 2013. While it might be worthwhile totease apart the marginal information contained in thesetwo series for these kinds of analyses, for the purposes ofthis paper we collapse them into a single factor, labeledeconomic activity.

In the upper plot of Fig. 4, we relate this aggregatedeconomic activity index to our macroeconomic sentimentfactor. As in the previous figure, we observe a largecorrelation between the two series, with macro sentimentclearly anticipating economic activity around turningpoints. Following the same reasoning as above, we there-fore further aggregate the information into a single growthfactor (comprised now of the information contained inoutput, employment, and macro sentiment). Finally, in thelower panel of Fig. 4, we compare this aggregate growthfactor with our real-time inflation index. While these two

10 We obtain daily data on S&P 500 returns, the VIX index, thedividend yield, the price–earnings ratio, the default premium (as thedifference between Moody's BAA and AAA rated bond yields), and theterm premium (as the difference between 10-year and 3-month Treasuryyields) from Bloomberg and Datastream.

11 Our inflation index is only significantly correlated with VIX(�0.30). Inflation dispersion is related to VIX (0.26), the dividend yield(0.61), and the price–earnings ratio (�0.33). These results, which wereport for completeness but do not return to in later analyses, are nottabulated to preserve space.

series are also somewhat positively correlated, thestrength of correlation is far weaker, with the inflationseries behaving much more erratically. For the remainderof the paper we therefore keep the real-time growth factorseparate from the inflation factor.

4.1.1. Economist disagreementWe conclude this preliminary analysis with two figures

showing growth and inflation factors and economist dis-agreement. Specifically, in Fig. 5 we plot the real-timegrowth factor in the top chart and the economist disagree-ment about growth in the bottom chart. Not surprisingly,the growth index dips through the recession periods of2001 and 2008–2009. More interestingly, though, theforecast dispersion appears relatively low at the beginningand extremely high toward the end of recessions, suggest-ing that economists tend to agree on downturns butcannot foresee recoveries as clearly.

We further investigate this intriguing pattern of econ-omist disagreement in several ways. First, we find that thegrowth index and its volatility tend to have a stable andsignificantly negative correlation of �0.35 over all oursample period, suggesting that growth tends to be morevolatile and difficult to predict during economic contrac-tions (when the growth index is negative). Second, weselect periods identified as NBER-dated recessions and findthat the volatility of growth is 63% higher in these periodsthan it is in expansions. Along similar lines, we try toidentify more precisely the last part of recession phases byselecting periods when the growth index level is negative,but the growth index first difference is positive (usingeither a daily or monthly first difference). In these laterecession periods, the volatility of the growth index tendsto be about 20% higher than in the other business cyclephases. In summary, this empirical evidence suggests thatthe cyclicality of macroeconomic disagreement seems tobe also a result of the growth index becoming morevolatile in recessions. In these periods, macroeconomicannouncements generate larger innovations in the growthindex and this contributes to larger macroeconomicdispersion.

Note that the dispersion of growth forecasts is alsorelatively large and noisy at the beginning of our sample.While there might have been indeed a higher degree ofmacroeconomic uncertainty at that time, it is more likelythat this pattern is due to the small number of macro-economic news releases for which economist forecastswere available in the first year of our sample. Out of the 34variables used to construct the growth index, only 11 hadforecasts reported on Bloomberg in 1997 and for thosereleases, only an average of four economists were provid-ing their forecasts.

In Fig. 6 we show the real-time inflation factor in thetop plot and the economist disagreement about inflationin the bottom plot. The inflation index is more erratic thanthe growth index, but it still dips through the recessionperiods, especially in 2008–2009. The inflation forecastdispersion is extremely volatile and the only pattern tostand out is the very large disagreement characterizing theend of the last recession.

1997 1999 2001 2003 2005 2007 2009 2011-5

-4

-3

-2

-1

0

1

2Output and employment

2008 2009 2010 2011 2012 2013-5

-4

-3

-2

-1

0

1

2Output and employment (recent extended sample)

Fig. 3. The upper panel shows the real-time output (thin blue line) andemployment factor (thick red line) from 1997 to 2011. Grey areas denoteNBER recessions. The lower panel plots again the real-time output (thinblue line) and employment factor (thick red line) for the recent extendedsubsample 2008 to 2013. (For interpretation of the references to color inthis figure caption, the reader is referred to the web version of this paper.)

1997 1999 2001 2003 2005 2007 2009 2011

-4

-3

-2

-1

0

1

2Economic activity and macro-sentiment

1997 1999 2001 2003 2005 2007 2009 2011

-5

-4

-3

-2

-1

0

1

2

3

Growth and inflation

Fig. 4. The upper panel shows the real-time economic activity (thick blueline) and macro–sentiment (thin red line) factors. The lower panel plotsthe real-time growth (thick blue line) and inflation (thin red line) factors.Grey areas denote NBER recessions. (For interpretation of the referencesto color in this figure caption, the reader is referred to the web version ofthis paper.)


4.1.2. Comparison with the CFNAIThe CFNAI published monthly by the Chicago Federal

Reserve Bank of Chicago is a commonly used real-timeindicator of economic conditions in the finance andeconomics literature (e.g., Beber, Brandt, and Kavajecz,2011). The index, which evolved from the Stock andWatson (1989) coincident indicator, is generally preferredto NBER expansion and recession dates because it is timely(though at a monthly frequency) and continuous, asopposed to the discrete peak and trough NBER dates.Given its popularity, as well as because it utilizes a broadcross-section of economic indicators like our approach, theCFNAI is an obvious first benchmark for evaluating theperformance of our approach. Before we dive into thequantitative comparison, though, it is worthwhile high-lighting the differences between the CFNAI and ourapproach. First, the CFNAI is a weighted average ofcurrently 85 monthly indicators that is formed monthlyonce about two-thirds of the indicators have been updated(the remaining one-third are projected). Second, theweights are determined by PCA using a simple unadjustedmonthly correlation matrix. In contrast, our index is

formed daily, based on the most recent observations ofonly a subset of growth-related data series, and theweights are determined by PCA using an autocorrelation-adjusted daily correlation matrix.

There are two important details in setting up a faircomparison between our approach and the CFNAI. First, atany release date, the CFNAI is constructed for the wholehistory given the most recent PCA weights, restatedfigures, and subsequently realized (for the one-third pro-jected series) economic data, as opposed to keeping trackof a sequence of point-in-time measures. We thereforeobtain a panel of CFNAI vintages from the Chicago Fed'sWeb site. This allows us to construct a point-in-timeversion of the CFNAI that reflects not only unrestated orunobservable data, but also the relative weighting basedon changing correlation structure. The second detail is thetiming of the monthly releases. The CFNAI is normallyreleased toward the end of each calendar month. Based onthe last available publication dates, the data are on averagereleased on the 23rd day of the month. We thus match

1997 1999 2001 2003 2005 2007 2009 2011

-4

-3

-2

-1

0

1

2Growth index

1997 1999 2001 2003 2005 2007 2009 2011-2

-1.5

-1

-0.5

0

0.5

1

1.5

2

2.5

3

3.5Dispersion

Fig. 5. The upper panel shows the real-time growth factor. The lowerpanel is the dispersion of economist forecasts about upcoming growthnews releases. Grey areas denote NBER recessions.

1997 1999 2001 2003 2005 2007 2009 2011

-4

-3

-2

-1

0

1

2Inflation index

1997 1999 2001 2003 2005 2007 2009 2011-2

-1.5

-1

-0.5

0

0.5

1

1.5

2

2.5

3

3.5Inflation dispersion

Fig. 6. The upper panel shows the real-time inflation factor. The lowerpanel is the dispersion of economist forecasts about upcoming inflationnews releases. Grey areas denote NBER recessions.

2001 2003 2005 2007 2009 2011-4

-3

-2

-1

0

1

2CFNAI and growth index

CFNAIGrowth

Fig. 7. This figure shows the real-time growth factor and the CFNAI, bothobserved at the monthly frequency on the same day during the sampleperiod Feb-2001 to Dec-2011.


each monthly CFNAI release with our real-time growthindex on either the actual release dates, when available, orestimated release dates based on this average timing.

Fig. 7 plots the monthly CFNAI with matching monthlyobservations of our real-time growth factor. To ease thecomparison over the subsample for which both series areavailable, we re-standardize them to have mean zero andstandard deviation one in-sample. As is immediatelyapparent, the two series are very similar with a correlationof 0.94. More importantly, though, notice that the real-time growth index seems to anticipate the turning pointsof the CFNAI.

The high correlation between the two indices is notsurprising given the similarities in methodology. The secondobservation, that our real-time growth index seems to leadthe CFNAI, however, deserves closer inspection. For this, weset up a vector autoregression (VAR) model for the CFNAI, thereal-time growth factor, and their respective previous month'slags. In Panel A of Table 2 the model is constrained to bediagonal, whereas in Panel B it is unconstrained. The real-timegrowth index has significant predictive power for the CFNAI,beyond the lagged CFNAI. However, the opposite is not true,as the CFNAI is not a significant predictor of the real-timegrowth index beyond its own lag. In other words, there is

fairly strong evidence of Granger causality from our Growthindex to the CFNAI (with a t-statistic of 4.25), but not in theopposite direction.

Table 2CFNAI versus growth index.

This table shows estimates of the following vector autoregression:

Yt

Xt

" #¼ AþB

Yt�1

Xt�1

" #þϵt ;

where Yt is the CFNAI and Xt is the real-time growth index. We use thevalue of the real-time growth index observed on the day of the monthlyrelease of the CFNAI. The sample is monthly observations from February2001 (first available vintage value of the CFNAI) to December 2011. InPanel A the model is constrained to be diagonal, whereas in Panel B it isunconstrained. The R2 statistic is adjusted and in parentheses are robustNewey-West t-statistics.

Independent

Dependent Constant CFNAIt�1 Growtht�1 R2 ð%Þ

Panel A: ConstrainedCFNAIt �0.03 0.91 83.56

(1.89) (17.30)Growtht �0.01 0.96 91.77

(�0.74) (24.79)Panel B: UnconstrainedCFNAIt �0.08 0.41 0.39 86.86

(�2.08) (3.68) (4.25)Growtht �0.01 0.05 0.93 91.72

(�0.48) (0.52) (13.31)

2008.12 2009.6 2009.12 2010.6 2010.12 2011.6 2011.12-3

-2.5

-2

-1.5

-1

-0.5

0

0.5

1

1.5

2ADS and growth index

ADSGrowth

Fig. 8. This figure shows the real-time growth factor and the correspond-ing vintages of the ADS index during the sample period Dec-2008 to Dec-2011.


The forecasting power of the real-time growth indexfor the CFNAI can potentially be explained by a number offeatures. First, as we noted in footnote 2, our growth indexis constructed from a cross-section of distinct statistics,whereas the CFNAI uses multiple variations of the sameinformation. Second, some of the releases used in theconstruction of CFNAI (e.g., the six statistics on housing)are not directly related to growth and could potentiallyintroduce a pattern with different cyclicality or persis-tence. Third, the CFNAI release includes projected monthlyvalues for one-third of the series. Clearly, this can be animportant source of index predictability. Finally, theweighting factor for each of the CFNAI underlying seriesis re-estimated monthly, whereas it is updated daily in ourreal-time growth index, even if we sample our indexmonthly for comparison purposes in the VAR estimation.For these reasons, even if the CFNAI and our growth indexboth rely on a large cross-section of news, the informationcontent of the resulting index can still differ substantiallywith the use of different methods. This is true even whenwe constrain our growth index to be observed with thesame monthly frequency of the CFNAI.

4.1.3. Comparison with the ADS indexThe CFNAI is an obvious benchmark because like our

approach, it utilizes a large cross-section of data series. TheADS index, developed by Aruoba, Diebold, and Scotti(2009) and now published by the Philadelphia Fed, is anequally worthy candidate for comparison because, beingbased on a state-space model, it can be updated daily likeour approach (though in practice the ADS index is updatedweekly). For the ADS index it is even more critical to usevintage data, as for a given release the index time-series isfull-sample smoothed, using the Kalman filter algorithm,and therefore contains forward-looking information (in

addition to using restated or subsequently released datalike the CFNAI does). Only the end-point of the index seriesis therefore a valid point-in-time measure. Weekly vintagereleases of the ADS business conditions index are availablestarting at the end of 2008, resulting in a relatively shortsample of 283 observations. We match each weeklyrelease with our daily real-time growth factor observedon the release date.

Fig. 8 plots the ADS index and our growth factor, whereagain we re-standardize both for this subsample. The twoseries are also very similar with a correlation of 0.91. Thisobservation is a little more surprising. On one hand, theADS index is based on only six indicators as opposed to our34, which likely explains why the ADS index is consider-ably more noisy. On the other hand, through the state-space model used to construct the ADS index, the weight-ing of data is optimized to forecast GDP. The weights of ourreal-time growth index are instead optimized to explainthe correlation structure of the cross-section of newsreleases. The figure suggests that the principal componentof growth-related news is highly correlated with the bestpredictor for future GDP formed from a subset of the dataseries. We will return to the question of how well our real-time growth factor forecasts future growth in the nextsection.

Table 3 repeats the Granger causality analysis for theADS index and our real-time growth factor. In Panel A theVAR model is constrained and in Panel B it is uncon-strained. Similar to our findings for the CFNAI, we find astatistically significant Granger causal relation from ourgrowth factor to the ADS index, meaning that the growthfactor predicts future realizations of the ADS index beyondthe lagged ADS index (with a t-statistic of 4.9). Theopposite causal relation, from the ADS index to our growthfactor is insignificant, and of the wrong sign.

To better understand the reasons for the forecastingpower of our real-time growth factor for the ADS index, weconstruct an ADS replica using our method on the samemacroeconomic announcement series used in the originalADS construction, namely, weekly initial jobless claims,monthly payroll employment, industrial production,

Table 3ADS index versus growth index.

This table shows estimates of the following vector autoregression:

Yt

Xt

" #¼ AþB

Yt�1

Xt�1

" #þϵt ;

where Yt is the ADS index and Xt is the real-time growth index. We usethe value of the real-time growth index observed on the day of theweekly release of the ADS index. The sample is weekly observations fromDecember 2008 (first available vintage value of the ADS index) throughDecember 2011. In Panel A the model is constrained to be diagonalwhereas in Panel B it is unconstrained. The R2 statistic is adjusted and inparentheses are robust Newey-West t-statistics.

Independent

Dependent Constant ADSt�1 Growtht�1 R2 ð%Þ

Panel A: ConstrainedADSt �0.02 0.95 91.43

(�1.78) (52.98)Growtht 0.01 0.99 99.61

(0.41) (165.45)Panel B: UnconstrainedADSt 0.04 0.76 0.12 92.18

(2.11) (21.98) (4.90)Growtht 0.01 �0.02 1.01 99.61

(0.78) (�1.18) (68.46)

Table 4Predicting quarterly GDP releases.

This table shows estimates of the following predictive regression:

GDPt ¼ αþβ1GDPt�1Q þβ2Xt�2Mþϵt ;

where GDPt is a quarterly GDP release, GDPt�1Q is the previous quarter'sGDP release, and Xt�2M is the average forecast of quarterly GDP by theSurvey of Professional Forecasters (SPF) and/or our real-time growthindex, both observed on the same day about two months before the GDPrelease (when the SPF is released). The sample is quarterly observationsfrom the beginning of 1997 to the end of 2011. The R2 statistic is adjustedand in parentheses are robust Newey-West t-statistics.

Model specification

1 2 3 4

Constant �0.61 �0.15 �0.11 �0.08(�2.75) (�0.99) (�0.60) (�0.53)

GDPt�1Q 0.22 0.06 0.04 0.03(4.14) (1.10) (0.68) (0.57)

SPFt�2M 0.57 0.44(3.25) (1.68)

Growtht�2M 0.58 0.20(2.73) (0.75)

R2 ð%Þ 30.28 44.60 41.18 44.25

Marginal R2 ð%Þ of Xt�2M 20.54 15.63 20.04


personal income less transfer payments, manufacturingand trade sales, and quarterly real GDP. The ADS index andits replica have a 0.95 correlation, suggesting that thedifferent methodologies alter the output only modestly.We then replace the original ADS index with our replicaand re-estimate the VAR model of Table 3. We find verysimilar results (not reported, for brevity), that is, a statis-tically significant Granger causal relation from our growthfactor to the ADS replica index, meaning that the growthfactor predicts future realizations of the ADS replica indexbeyond the lagged ADS replica index. These findingssuggest that the forecasting power of the real-time growthindex originates mainly from the large cross-section ofmacroeconomic releases, rather than from the differentmethodology. Obviously, our method can easily deal withan arbitrarily large cross-section, whereas the ADS meth-odology is constrained in this dimension.

4.2. Forecasting future GDP and CPI releases

The last two subsections showed that our methodologyof extracting daily factors from economic news released atdifferent times and frequencies delivers a real-timegrowth factor that is highly correlated with existing now-casting indices, but provides potentially more timely andcertainly more frequent information. We specifically founda high correlation with the ADS index for which the dataare weighted to best forecast future GDP growth. Thisfinding begs the question of how well our real-timefactors, which are not explicitly constructed to forecast,can nevertheless be used for forecasting future fundamen-tals (both growth and inflation).

Since the CFNAI and ADS index focus exclusively ongrowth and their vintage histories are limited, especiallyfor the more forecasting-oriented ADS index, we instead

use as forecasting benchmarks the much longer historiesof quarterly growth and inflation forecasts from the Surveyof Professional Forecasters (SPF) carried out by the Phila-delphia Fed. More specifically, we use the average forecastsof the annualized nominal GDP growth rate for the nextquarter as well as of the annualized percent change in theCPI over the next year. The survey results are releasedaround the end of the second month of the quarter, andwe match the timing of our real-time growth and inflationfactors to the survey release dates. Our focus on releasedates, as opposed to survey dates, allows us to preserve thereal-time nature of the comparison. Since our methodol-ogy does not involve a collection and dissemination lag, itcan potentially use a larger information set than the SPFsurvey. For the actual statistics to be forecasted, we useadvance GDP growth, which is announced about onemonth after the end of the quarter, and headline CPIchange, which is typically released two weeks after theend of the quarter. For example, we forecast the 1997 first-quarter GDP and CPI using the SPF mean forecasts of 4.90%and 3.01%, respectively, released on February 26, 1997 andthe real-time growth and inflation factors of 0.79 and�0.45, respectively, obtained on the same day. The actualrelease of CPI came out about three weeks later on March19, 1997 at 3.00% and GDP was announced two monthsafter the survey on April 30, 1997 at 5.60%.

Table 4 shows the results for growth forecasting. Wefind that the mean forecasts of the SPF and our real-timegrowth factor contain about equally useful information forpredicting subsequent GDP releases beyond lagged GDP.The model R2's are large at 45% and 41% and the moreinformative marginal R2's (measuring the incrementalforecasting ability of the additional regressor relative to asimpler autoregressive model) are 20% and 15%, respec-tively. Furthermore, the correlation between the real-timegrowth factor and quarterly SPF forecasts is very high at

1997 1999 2001 2003 2005 2007 2009 2011

-4

-3

-2

-1

0

1

2

3GDP, SPF, and real-time growth index

GDP actualSPF medianGrowth index

Fig. 9. This figure shows the real-time growth factor, the medianprojection of nominal GDP growth rate from the Survey of ProfessionalForecasters (SPF) on the same dates, and the actual GDP release for thesame quarter, at quarterly frequency during the sample Jan-1997 to Dec-2011.

Table 5Predicting monthly CPI releases.

This table shows estimates of the following predictive regression:

CPIt ¼ αþβ1CPIt�1Mþβ2Xt�1Mþϵt ;

where CPIt is the monthly CPI release, CPIt�1M is the previous month'sCPI release, and Xt�1M is the average forecast of year-on-year CPI by theSurvey of Professional Forecasters (SPF) and/or our real-time inflationindex, both observed quarterly on the same day about one month beforethe CPI release (when the SPF is released in the second month of eachquarter). The sample is quarterly observations from the beginning of 1997to the end of 2011. The R2 statistic is adjusted and in parentheses arerobust Newey-West t-statistics.

Model specification

1 2 3 4

Constant 0.57 �0.04 0.63 0.08(3.54) (�0.18) (3.90) (0.47)

CPIt�1M 0.77 0.48 0.75 0.50(11.11) (3.03) (10.45) (3.98)

SPFt�1M 0.54 0.47(2.55) (3.12)

Inflationt�1M 0.21 0.17(1.69) (1.96)

R2 ð%Þ 66.48 69.84 69.04 71.48

Marginal R2 ð%Þ of Xt�1M 10.02 7.64 14.92

1997 1999 2001 2003 2005 2007 2009 2011-6

-5

-4

-3

-2

-1

0

1

2

3

4CPI, SPF, and real-time inflation index

CPIInflation indexSPF median

Fig. 10. This figure shows the real-time inflation factor, the medianprojection of CPI from the Survey of Professional Forecasters (SPF) onthe same dates, and the actual CPI release of the following month, atquarterly frequency during the sample Jan-1997 to Dec-2011.


0.89. This suggests that the real-time growth index can beinterpreted as a higher frequency reading of economicgrowth expectations with the same properties as thelower frequency SPF forecasts. Alternatively, perhaps theprofessional forecasters deploy nowcasting models, eitherexplicitly or, more likely for the historical data, implicitly.This observation is consistent with Liebermann (2014),who finds that (a different approach to) nowcasting iscomparable to the SPF at the date of release but superiorprior (when no SPF is available) and shortly after, as itupdates.

Fig. 9 illustrates these points graphically. The highcorrelation between our growth factor, SPF consensus, and

subsequent GDP releases, is immediately apparent from theplot. This is particularly the case around the shaded NBERrecession periods.

Table 5 shows the results for inflation forecasting.Again we find that both the mean forecasts of the SPFand the real-time inflation factor contain about equallyuseful information to predict the subsequent CPI releases,beyond lagged CPI. The model R2's are even higher at 70%and 69%, respectively, but a larger fraction of this predict-ability comes simply from the higher persistence of infla-tion. The marginal R2 relative to the autoregressivebenchmark model is 10% for the SPF forecasts and 8% forthe real-time inflation factor. By that metric, the forecast-ing ability of both predictors is weaker compared to GDPforecasting. Moreover, the correlation between the pre-dictors is also significantly weaker at only 0.21. It appearsfrom these results that it is relatively more difficult topredict inflation from the intra-quarter news flow, whichmay partly be attributed to the fact that there is lessinflation-relevant news (only nine distinct releases onseven days).

Fig. 10 presents these results graphically. Although ourreal-time inflation factor and the SPF consensus forecastsare clearly correlated and seem to anticipate actual CPIreleases, particularly around the NBER recession periods,the real-time inflation factor exhibits a more erraticbehavior. This reflects again the relatively sparse inflationnews flow.

The results in Tables 4 and 5 demonstrate not only theability of our real-time factors to predict subsequentrealizations of economic fundamentals but, equally inter-estingly, how similar these factors are to the SPF consensusforecasts. This is consistent with the findings ofLiebermann (2014) and by no means diminishes therelevance of our real-time factors, since they have thedistinct advantage of being available daily for weeks beforeand incorporating new information daily for months afterthe quarterly SPF is released. To complete the comparisonof our real-time factors with the SPF, however, we can also

1997 1999 2001 2003 2005 2007 2009 2011

-4

-2

0

2

Gro

wth

Inde

x (in

verte

d sc

ale)

0.05

0.2

0.35

0.5

0.65

VIX

Fig. 11. This figure shows the real-time growth factor (inverted left-scale,thin line) and VIX (right-scale, bold line) during our sample period.

1997 1999 2001 2003 2005 2007 2009 2011-2

-1

0

1

2

Gro

wth

Dis

pers

ion

1997 1999 2001 2003 2005 2007 2009 20110.05

0.2

0.35

0.5

0.65

VIX

Fig. 12. This figure shows the dispersion of economist forecasts aboutgrowth news (left-scale, thin line) and VIX (right-scale, bold line) duringour sample period.

Table 6Explaining financial market volatility.

This table shows estimates of the following contemporaneous regression:

VIXt ¼ αþβXtþϵt ;

where Xt is our real–time growth index and/or dispersion of economistforecasts about growth news. The sample is daily observations fromJanuary 1997 through December 2011 in Panel A and January 2000through December 2011 in Panel B. Robust Newey-West t-statistics arereported in parentheses.

Model specification

1 2 3

Panel A: 1997–2011Constant 0.23 0.23 0.23

(34.46) (39.45) (40.79)Growth index �0.0387 �0.0368

(�5.79) (�5.42)Growth dispersion 0.0217 0.0159

(3.50) (3.40)

R2 ð%Þ 25.70 5.98 28.85

Panel B: 2000–2011Constant 0.23 0.21 0.20

(27.41) (34.64) (30.26)Growth index �0.0534 �0.0622

(�7.37) (�6.04)Growth dispersion 0.0402 �0.0212

(3.84) (�1.93)

R2 ð%Þ 40.83 9.68 42.41


relate their respective second moments. Specifically, wecompare our measures of uncertainty surrounding growthand inflation, which capture the disagreement of econo-mists about the various components that make up ourreal-time factors, with the dispersions of SPF forecasts,which capture the disagreement among economists aboutfuture growth and inflation directly. It is reasonable toexpect that when economists disagree on recent economicdata, the same or similar economists will also disagreeabout the future path of the economy. Consistent with thisintuition, we find that our measure of uncertainty aboutgrowth and the dispersion of SPF growth forecasts has acorrelation of 0.55. The corresponding correlation forinflation is 0.39. Although these correlations for thesecond moments are not as strong as for the first, we stillconclude that our proxies for macroeconomic uncertaintycapture, at a daily frequency, similar uncertainty as thatreflected in the dispersion of SPF forecasts.

4.3. Macroeconomic conditions and financial marketvolatility

One of the differentiating aspects of our methodology isthat it produces a daily reading of the state of the economythat does not rely on information from financial markets,unlike the approaches of Giannone, Reichlin, and Small(2008) and Aruoba, Diebold, and Scotti (2009), for exam-ple. We can therefore use our real-time factors to inves-tigate the link between macroeconomic conditions andfinancial market dynamics, particularly stock market vola-tility. We focus on stock market volatility for two reasons.First, volatility is easier to measure than expected returns.Second, but related, the apparent disconnect betweenstock market volatility, which is easily measured, andeconomic fundamentals, the improved measurement ofwhich is the purpose of our methodology, is one of thelongest standing puzzles in finance. In a seminal paper,Schwert (1989) finds that the standard deviations of a hostof macroeconomic variables and a recession dummyexplain only a small fraction of stock market volatility.More recently, Engle and Rangel (2008) refer to therelation between the macro economy and stock marketvolatility as the central unsolved problem of 25 years ofvolatility research.

We measure stock market volatility using the forward-looking option-implied volatility index VIX, rather than ameasure of backward-looking realized volatility. Realizedvolatility could be mechanically correlated with our real-time factors because large economic surprises invoke largestock market responses. The empirical question is notwhether the stock market responds contemporaneouslyto economic data, there is plenty of evidence it does (e.g.,Flannery and Protopapadakis, 2002), but rather whetherbusiness-cycle-related changes in economic conditionslead to persistent changes in future stock market volatility.

We first provide some graphical evidence of the relationbetween the VIX index, our real-time growth factor, andgrowth dispersion. Specifically, we plot the VIX index alongwith the growth factor in Fig. 11, where we invert the axis forthe growth factor to highlight the strong negative correlation(�0.51 from Table 1) between the two series. We plot the VIX


index along with growth dispersion in Fig. 12. The correlationbetween these two series is lower (0.25 from Table 1), butincreases somewhat to 0.31 when we start the sample in2000 when growth dispersion is less noisy (recall the discus-sion surrounding Fig. 5).

We extend this bivariate analysis in Table 6, where weregress the VIX index contemporaneously on our real-timegrowth factor and/or growth dispersion. Panel A shows theresults for the full sample, and Panel B is for the less noisy2000 onward subsample. We will focus the discussion onpanel B. In the first two model specifications both regres-sors are by themselves strongly statistically significant.Our real-time growth factor explains 41 percent of thevariation in the VIX index, and growth dispersion explainsabout ten percent stand-alone. Combined, in the thirdmodel specification, the adjusted R2 increases to 42percent with the growth factor being highly significantand growth dispersion being borderline significant.Beyond statistical significance, though, the economiceffects implied by the coefficient estimates are large. Aone standard deviation deterioration in growth results inmore than a five percentage point increase in the VIXindex, which is about a quarter increase relative to a baselevel of 23 percent. A one standard deviation increase in

Table 7Growth and inflation dynamics.

This table shows estimates of the following vector auto-regression:

Yt ¼ AþBYt�Lþϵt� L;

where Yt is a vector containing our real-time growth index, growth dispersion, insample is daily observations from January 2000 to December 2011. The R2 statistmarginal R2 represents the proportion of variance explained beyond the first la

Dependent Yt Independent Yt�L

Growth Growth Inflatindex dispersion inde

Growth factorL¼5 1.0033 0.0348 0.01

(169.46) (3.68) (2.3L¼20 1.0006 0.1269 0.05

(46.40) (3.95) (1.6L¼60 0.9198 0.2349 0.10

(14.12) (2.47) (0.9Growth dispersion

L¼5 �0.0475 0.8949 0.03(�5.10) (55.11) (1.9

L¼20 �0.2149 0.5270 0.12(�6.10) (9.43) (2.1

L¼60 �0.4366 �0.0005 0.12(�6.05) (�0.01) (1.3

Inflation factorL¼5 �0.0182 �0.0026 0.84

(�1.79) (�0.17) (38.4L¼20 �0.0677 0.0481 0.32

(�2.13) (1.07) (5.7L¼60 �0.0990 0.0696 0.09

(�2.24) (1.12) (1.4Inflation dispersion

L¼5 �0.0477 0.0559 0.01(�3.05) (2.00) (0.5

L¼20 �0.2501 0.0813 0.09(�4.73) (0.90) (0.9

L¼60 �0.6148 �0.3779 0.05(�4.74) (�2.32) (0.3

growth dispersion is associated with a four percent pointincrease in VIX.

In summary, contrary to Schwert (1989) and much ofthe subsequent literature, we present evidence of a stronglink between macroeconomic conditions and stock marketvolatility. We find that the level of growth, i.e., businesscycles, are more important than the uncertainty aboutgrowth, though the latter still plays a significant role, bothstatistically and economically in magnitude. This suggeststhat better real-time measurement of economic funda-mentals may help resolve the long-standing disconnectbetween the macro economy and financial stock marketvolatility.

4.4. Real-time growth and inflation dynamics

Table 7 describes the joint dynamics of the real-timegrowth index, growth dispersion, the real-time inflation index,and inflation dispersion. We estimate three first-order vectorautoregression (VAR) models with one-period lag lengths offive, 20, or 60 business days, respectively. The estimates arebased on the sample starting in 2000 (when dispersionmeasures are less noisy), using overlapping daily observations,and the standard errors used to compute the t-statistics are

flation index, and inflation dispersion. L represents the lag in the VAR. Theic is adjusted, and in parentheses are robust Newey-West t-statistics. Theg of the dependent variable.

R2 ð%Þ Marginal R2 ð%Þ

ion Inflationx dispersion

93 �0.0089 98.75 2.802) (�2.13)63 �0.0233 93.42 6.389) (�1.38)01 �0.0309 74.48 5.916) (�0.53)

15 0.0066 91.10 4.335) (0.77)68 0.0143 67.16 18.976) (0.50)08 0.0258 51.82 37.791) (0.61)

59 �0.0001 73.02 0.610) (�0.00)32 �0.0497 15.66 3.255) (�1.98)49 �0.0489 8.94 6.791) (�1.16)

67 0.8656 84.87 2.701) (37.98)40 0.4555 48.38 11.280) (9.29)07 0.0835 33.26 26.443) (0.86)

0 10 20 30 40 50 60 70 80 90-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

Horizon (days)

Cha

nge

in g

row

th in

dex

Top quartileAbove medianBelow medianBottom quartileUnconditional

Fig. 13. This figure shows the median first difference in the growth indexfor different horizons (in days), unconditionally and conditionally oncurrent dispersion about growth news being above (below) the samplemedian and in the top (bottom) quartile.


autocorrelation-adjusted. Since the results are fairly consistentacross specifications, we mainly focus our discussion on theintermediate 20-day horizon.12

All four series are persistent, especially at shorter horizons,as evidenced by the magnitude and statistical significance ofthe own lag terms, as well as by the differences between theR2 and the marginal R2 that exclude the impact of the own lagterms. Growth is highly persistent at all three horizons,whereas the autocorrelation of the other three series dropssharply as the lag length increases. This finding is visuallyconsistent with the behavior of the series in Figs. 5 and 6.

We also observe an interesting lead-lag interactionbetween the growth index and growth dispersion. Highergrowth dispersion is associated with higher future growthwhereas, in the opposite direction, higher growth isassociated with subsequently lower growth dispersion.Fig. 13 illustrates graphically the first cross-autocorrela-tion, from growth dispersion to the growth level. It showsthe median change in growth at different horizons uncon-ditionally and following realizations of growth dispersionabove or below median and in the top or bottom quartile.Periods of high dispersion, and especially those in the topquartile, are clearly followed by acceleration in growthover the subsequent weeks and months. The second cross-autocorrelation in Table 7, from the growth level to growthdispersion, is even stronger both in magnitude (recall thedata are standardized so coefficients can be directly inter-preted) and statistical significance (t-statistics around sixand marginal R2 of almost 20%). This result is consistentwith our prior observation that economists seem to agreeon the end of an economic expansion (following highgrowth, dispersion is low), but not on the end of aneconomic contraction (following low growth, dispersionis high).

The cross-autocorrelations between the inflation indexand inflation uncertainty are largely small and insignificant.However, higher growth seems to lead to lower uncertaintyabout inflation, with strongest results at longer horizons. Thisfinding is consistent with Fig. 6, where the uncertainty aboutinflation appears relatively larger at the end of the recessionsin our sample.

4.5. Extending the sample backwards

Our sample is limited by the availability of preciselydated and unrestated economic news releases. In thissection, we extend our sample backward to the beginningof 1985 using the median reporting lag for each releasetype and inferring the release date.13 While the use of

12 We also estimate the VAR model in subsamples corresponding tofour distinct business cycle phases: early expansion (positive growthindex and positive growth index monthly first difference), late expansion(positive growth index and negative growth index monthly first differ-ence), early recession (negative growth index and negative growth indexmonthly first difference), and late recession (negative growth index andpositive growth index monthly first difference). The results are largelysimilar in all four subsamples.

13 To get a sense for the accuracy of our procedure of dating theannouncements based on reporting lags, we partially cross-check ourinferred release dates with a database of Reuters news. More specifically,for a subsample of 15 announcements on the total of 43 considered news

potentially misdated and restated data weakens the real-time interpretation of our macroeconomic indices, thelonger sample period that spans one more business cycleserves as a useful robustness check.

In Fig. 14 we plot our “real-time” growth index togetherwith NBER ex-post determined recession dates and theexpectations for current quarter GDP growth in the SPF.The growth index behaves the same during the 1990–1991recession as it does for the other two recessions that arecovered by our original sample. We observe the suddendrop in the growth index and the subsequent gradualrecovery. Moreover, in the new 1985 to 1997 period, thegrowth index tracks the low-frequency growth expecta-tions of the SPF even more closely, suggesting again thatour approach captures the same information but at a dailyfrequency.

For our measure of macroeconomic uncertainty, thesample cannot be extended back because the panel ofeconomist forecasts we use to construct the disagreementabout growth measures are not available before 1997.Nevertheless, to see what a longer growth dispersionseries might look like, we apply our methodology to thefive disagreement measures about growth that can beobtained from the SPF (namely, disagreement about GDP,corporate profits, employment, unemployment, and pro-ductivity growth). Fig. 15 shows the results. We first noticea large correlation between our daily measure of macro-economic uncertainty and the quarterly measure of dis-agreement from the SPF over the original sample period.More interestingly, the backdated SPF-based measure ofuncertainty corroborates our earlier observation thatuncertainty peaks toward the end of recessions and ismore subdued at the end of expansions.

(footnote continued)items and for a shorter sample period going back to 1990, we find that91% of the estimated release days are less than two days off from theactual release days.


5. Conclusions

We proposed a simple cross-sectional technique toextract daily factors from economic news released atdifferent times and frequencies. Our approach can effec-tively handle the large number of different announce-ments that are relevant for tracking current economicconditions. We applied the technique to extract real-timemeasures of inflation, output, employment, and macro-economic sentiment, as well as corresponding measures ofdisagreement among economists about these indices. Ourprocedure provides more timely and accurate forecasts offuture changes in economic conditions than other real-time forecasting approaches. At the same time, both thelevel and dispersion measures are highly correlated with

1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011-5

-4

-3

-2

-1

0

1

2

Growth (1985-2011)

Fig. 14. This figure shows our real-time growth index constructed fromeconomic releases backfilled to January 1985. The bold line indicates thequarterly expectation of GDP growth for the current quarter contained inthe Survey of Professional Forecasters. Grey areas denote NBERrecessions.

1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011-2

-1

0

1

2

3

4

Dispersion and SPF disagreement (1985-2011)

Fig. 15. This figure shows our daily factor for the dispersion of economistforecasts about growth news (thin line) and the disagreement measure inthe quarterly Survey of Professional Forecasters (SPF) (bold line). Greyareas denote NBER recessions.

corresponding statistics from the SPF, suggesting theycapture the same information except our approach doesso at a daily instead of quarterly frequency. Finally, incontrast to much of the extant literature, our real-timegrowth factor and corresponding disagreement measure,both constructed entirely from macroeconomic data,explain a remarkable fraction of financial volatilitydynamics.

The purpose of our method is to obtain a real-time,daily, unbiased, and objective reading of the state of themacroeconomy, using an approach that lets the data speakas much as possible. Our forecasting results demonstratethat a fairly simple and unstructured method still deliversa very sensible and timely measurement of the state of theeconomy. A real-time daily reading of macroeconomicfundamentals that is reliable can be used for a number ofpurposes, most notably to study the relation betweenfinancial market and economic dynamics.

Appendix A. Macroeconomic news

The following table summarizes the main features ofthe macroeconomic news releases we work with. Thenews Category is either inflation (Inf), employment(Emp), output (Out), or sentiment (Sen). If the sampleseries is stationary in our sample, we make no adjustment(Adj¼0), otherwise we use first differences with respect tothe previous period (Adj¼1). We also indicate Units,Frequency (M for monthly, W for weekly, Q for quarterly),and the Source of the release.

Category
Release name Adj Units Freq Source
Inf
US Import Price Indexby End Use All MoM
0
Rate M Bureau LaborStatistics
Inf
US PPI FinishedGoods Total MoM
0
Inf
US PPI FinishedGoods Except FoodsEnergy
0
Inf
US CPI UrbanConsumers MoM
0
Inf
US CPI UrbanConsumers Less FoodEnergy
0
Inf
BLS Employment CostCivilian Workers QoQ
0
Rate Q Bureau LaborStatistics
Inf
US GDP Price IndexQoQ SAAR
0
Rate Q BureauEconomicAnalysis
Inf
US Personal Cons.Expenditure CorePrice Index MoM
0
Rate M BureauEconomicAnalysis
Inf
US Output Per HourNonfarm BusinessSector QoQ
0
Rate Q Bureau LaborStatistics
Emp
ADP NationalEmployment ReportPrivate NonfarmChange
0
Volume M AutomaticDataProcessing
Emp
US Initial JoblessClaims
1
Volume W Departmentof Labor
Emp
US Continuing JoblessClaims
1
Volume W Departmentof Labor
Emp
US Employees onNonfarm PayrollsTotal Net Change
0
Value M Bureau LaborStatistics


Emp
US Employees onNonfarm PayrollsManufact Net Change
0
Value M Bureau LaborStatistics
Emp
US UnemploymentRate Total in LaborForce
1
Emp
US Average WeeklyHours All TotalPrivate
1
Volume M Bureau LaborStatistics
Out
ISM ManufacturingPMI
0
Value M InstituteSupplyManagement
Out
US ManufacturersNew Orders TotalMoM
0
Rate M U.S. CensusBureau
Out
US Auto SalesDomestic Vehicles
1
Volume M Bloomberg
Out
ISM Non-Manufacturing NMINSA
0
Value M InstituteSupplyManagement
Out
Federal ReserveConsumer Credit NetChange
1
Value M FederalReserve
Out
MerchantWholesalersInventories Change
0
Out
Adjusted Retail FoodServices Sales Change
0
Out
Adjusted Retail SalesLess Autos Change
0
Out
US IndustrialProduction MoM2007¼100 SA
0
Rate M FederalReserve
Out
US CapacityUtilization of TotalCapacity
0
Rate M FederalReserve
Out
US ManufacturingTrade InventoriesTotal
0
Out
US Durable GoodsNew OrdersIndustries
0
Out
US Durable GoodsNew Orders ExTransp.
0
Out
GDP US Chained 2005Dollars QoQ SAAR
0
Out
GDP US PersonalConsumptionChained Change
0
Out
US Personal IncomeMoM
0
Out
US PersonalConsumptionExpend. NominalDollars
0
Sen
Bloomberg USWeekly ConsumerComfort Index
1
Price W Bloomberg
Sen
University MichiganSurvey ConsumerConfidence
1
Price M U. ofMichiganSurveyResearch
Sen
Empire StateManufact. SurveyBusiness Conditions
1
Value M FederalReserve
Sen
Conference Board USLeading Index MoM
0
Rate M ConferenceBoard
Sen
Philadelphia FedBusiness OutlookGeneral Conditions
1
Price M PhiladelphiaFed
Sen
Conference BoardConsumer ConfidenceSA 1985¼100
1
Rate M ConferenceBoard
Sen
Richmond FedReserveManufacturingSurvey
0
Rate M RichmondFed
Sen
US ChicagoPurchasing ManagersIndex SA
1
Price M KingsburyIntern.
Sen
ISM MilwaukeePurchasersManufacturing Index
1
Rate M NAPM –
Milwaukee

Sen
Dallas Fed Manufact.Outlook BusinessActivity
1
Rate M Dallas Fed
References

Ait-Sahalia, Y., Mykland, P.A., Zhang, L., 2005. How often to sample acontinuous-time process in the presence of market microstructurenoise. Review of Financial Studies 18, 351–416.

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

Aruoba, S.B., Diebold, F.X., Scotti, C., 2009. Real-time measurement ofbusiness conditions. Journal of Business and Economic Statistics 27,417–427.

Banbura, M., Giannone, D., Modugno, M., Reichlin, L., 2013. Now-castingand the real-time data flow. In: Elliott, G., Timmermann, A. (Eds.),Handbook of Economic Forecasting, vol. 2a. , North Holland, Amster-dam, pp. 195–237.

Beber, A., Brandt, M.W., Kavajecz, K.A., 2011. What does equity sectororderflow tell us about the economy? Review of Financial Studies 24,3688–3730.

Engle, R.F., Rangel, J.G., 2008. The spline-GARCH model for low-frequencyvolatility and its global macroeconomic causes. Review of FinancialStudies 21, 1187–1222.

Evans, M., 2005. Where are we now?: real-time estimates of the macroeconomy The International Journal of Central Banking 1, 127–175.

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

Ghysels, E., Horan, C., Moench E., 2012. Forecasting through the rear-viewmirror: data revisions and bond return predictability. Federal ReserveBank of New York Staff Reports No. 581.

Giannone, D., Reichlin, L., Small, D., 2008. Nowcasting: the real timeinformational content of macroeconomic data releases. Journal ofMonetary Economics 55, 665–676.

Koenig, E., Dolmas, S., Piger, J., 2003. The use and abuse of real-time datain economic forecasting. Review of Economics and Statistics 85,618–628.

Liebermann, J., 2014. Real-time nowcasting of GDP: factor model versusprofessional forecasters. Oxford Bulletin of Economics and Statistics76, 783–811.

Newey, W.K., West, K.D., 1987. A simple, positive semi-definite, hetero-skedasticity and autocorrelation consistent covariance matrix. Econ-ometrica 55, 703–708.

Schwert, G.W., 1989. Why does stock market volatility change over time?Journal of Finance 44, 1115–1153.

Stambaugh, R.F., 1997. Analyzing investments whose histories differ inlength. Journal of Financial Economics 45, 285–331.

Stock, J.H., Watson, M.W., 1989. New indexes of coincident and leadingeconomic indicators. In: Blanchard, O.J., Fischer, S. (Eds.), NBERMacroeconomics Annual, pp. 352–394.

http://refhub.elsevier.com/S0304-405X(15)00076-8/sbref1











































journal of financial economics - fuqua school of … › ~mbrandt › papers › published › ...at...

Documents