Is the Intrinsic Value of Macroeconomic News Announcements Related to their Asset Price Impact?

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Is the Intrinsic Value of Macroeconomic News Announcements Related to their Asset Price Impact? Thomas Gilbert, Chiara Scotti, Georg Strasser, and Clara Vega 2015-046 Please cite this paper as: Gilbert, Thomas, Chiara Scotti, Georg Strasser, and Clara Vega (2015). “Is the Intrinsic Value of Macroeconomic News Announcements Related to their Asset Price Impact?,” Finance and Economics Discussion Series 2015-046. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2015.046r1. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Is the Intrinsic Value of Macroeconomic News ∗ Announcements Related to their Asset Price Impact? Thomas Gilbert Chiara Scotti Foster School of Business Board of Governors University of Washington Federal Reserve System Georg Strasser Clara Vega DGR Monetary Policy Research Board of Governors European Central Bank Federal Reserve System December 8, 2016 Abstract The literature documents a heterogeneous asset price response to macroeconomic news announcements. We explain this variation with a novel measure of the intrinsic value of an announcement – the announcement’s ability to nowcast GDP growth, inflation, and the Federal Funds Target Rate – and decompose it into the announcement’s relation to fundamentals, a timeliness premium, and a revision premium. We find that differences in intrinsic value can explain a significant fraction of the variation in the announcements’priceimpactonTreasurybondyields. Theannouncements’timeliness and relation to fundamentals are the most important characteristics in explaining this variation. Keywords: Macroeconomicannouncements, pricediscovery, learning, forecasting, nowcasting JEL classification: G14, E37, E44, E47, C53, D83 ∗Forcommentsandsuggestionsonanearlierandpreliminarydraft,wethankananonymousreferee,Torben Andersen, Tim Bollerslev, Dean Croushore, Eric Ghysels, Refet Gu¨rkaynak (associate editor), Michael McCracken, Ricardo Reis (editor), Barbara Rossi, Jonathan Wright, the participants of the (EC)2 Conference, the ECB Workshop on Forecasting Techniques, the Applied Econometrics and Forecasting in MacroeconomicsandFinanceWorkshopattheSt. LouisFed,theConferenceonReal-TimeDataAnalysis,Methods and Applications at Federal Reserve Bank of Philadelphia, the Conference on Computing in Economics and Finance, the Humboldt-Copenhagen Conference on Financial Econometrics, the HITS workshop on Advances in Economic Forecasting, and the seminar participants at the University of Nu¨rnberg and the Ifo Institute. We thank Domenico Giannone, Lucrezia Reichlin, and David Small for sharing their computer codewithus,andMargaretWaltonforoutstandingresearchassistance. Anyviewsexpressedrepresentthose oftheauthorsandnotnecessarilythoseoftheEuropeanCentralBank,theEurosystem,theFederalReserve System, or the Board of Governors. Corresponding author: Clara Vega, Board of Governors of the Federal Reserve System, 1801 K Street N.W., Washington, D.C. 20006, USA, +1-202-452-2379, clara.vega@frb.gov.

1. Introduction An extensive literature has linked macroeconomic news announcements to movements in stock, government bond, and foreign exchange returns.1 Some of these papers have highlighted the heterogeneous response of asset prices to news: Some announcements have a strong impact on asset prices, but some do not. However, there are no papers that sys- 5 tematically investigate what causes this heterogeneous response. In this paper, we help fill in the void by (i) proposing, estimating and decomposing a novel empirical measure of announcements’ intrinsic value, and (ii) relating differences in the U.S. Treasury bond market’s responses to differences in our novel measures of announcement characteristics. First, motivated by economic theory, we define and estimate the intrinsic value of an 10 announcement as its importance in nowcasting the following primitives or fundamentals: the U.S. Gross Domestic Product (GDP), the GDP price deflator, and the Federal Funds Target Rate (FFTR). More precisely, the intrinsic value is the nowcasting weight placed on the macroeconomic announcement at the time of its release. Next, using the same nowcasting framework, we decompose this intrinsic value into 15 three components that capture the announcement’s relation to fundamentals, timing, and revisions. While the previous literature has discussed each of the last two characteristics in isolation, our contribution is to formally define all three announcement characteristics coherently within a single nowcasting framework. Our definition of the announcement’s relation to fundamentals is its importance in nowcasting our three primitives independent 20 of the announcement’s release time and revisions. We define the announcement’s timeliness premium as the change in its nowcasting weight due to its release time. Similarly, we define 1Papers that study the government bond market response to macroeconomic announcements include Fleming and Remolona (1997, 1999), Balduzzi et al. (2001), Goldberg and Leonard (2003), Gu¨rkaynak et al. (2005), Beechey and Wright (2009), and Swanson and Williams (2014). Papers that study the foreign exchange market response include Almeida et al. (1998), Andersen et al. (2003), and Ehrmann and Fratzscher (2005). See Neely and Dey (2010) for a review of the literature on foreign exchange response to macroeconomicannouncements. StudiesofthestockmarketresponseincludeFlanneryandProtopapadakis (2002), Ehrmann and Fratzscher (2004), Bernanke and Kuttner (2005), and Bekaert and Engstrom (2010). Boyd et al. (2005), Faust et al. (2007), Bartolini et al. (2008), among others, study multiple asset classes simultaneously. 1

the announcement’s revision premium as the change in its nowcasting weight due to its future revisions. Finally, we relate an announcement’s intrinsic value, timeliness, revision, and relation 25 to fundamentals to the announcement’s asset price impact. We find that using GDP as the nowcasting target is more useful in explaining the price impact of announcement surprises than using the GDP deflator or the FFTR. When using GDP as the nowcasting target, our intrinsic value measure explains between 12 and 19 percent of the variation in the heterogeneous response of asset prices to macroeconomic news announcements. When we estimate 30 the importance of the three individual announcement characteristics separately, we find that our novel measures of timeliness and relation to fundamentals are the most important characteristics in explaining the announcement’s price impact. Note that our novel measure of intrinsic value explains the heterogeneous response of asset prices to macroeconomic announcements better than other variables discussed in the previous literature, such as the 35 reporting lag of the announcement and the magnitude of its revisions. Since our focus is on understanding the U.S. Treasury bond market’s response to macroeconomic news announcements, we choose nowcasting primitives that are consistent with this literature. In particular, Beechey and Wright (2009) group macroeconomic announcements into three broad categories: news about real output, news about prices, and 40 news about monetary policy.2 The primitives we choose, namely GDP, GDP price deflator, and the FFTR, are representative of each of these categories. When studying the response of other asset classes to macroeconomic announcements, researchers should consider other primitives: For example, in the case of foreign exchange markets, the primitives should include both domestic and foreign monetary policy rates. 45 Our paper contributes to the literature by showing that the price response to a particular type of announcement cannot be analyzed in isolation.3 The effect that announcements 2Since nominal Treasury bond prices embody inflation expectations and expected future real interest rates, news about prices, real output, and monetary policy are natural choices of primitives. 3Recent studies by Ehrmann and Sondermann (2012) and Lapp and Pearce (2012) further support this view. 2

have on asset prices crucially depends on the information environment. When studying the link between asset prices and macroeconomic fundamentals, researchers need to account not only for the surprise component of an announcement but also for the announcement’s intrin- 50 sic value, its relation to fundamentals, its timeliness, and its revisions, all relative to other announcements. For example, researchers who analyze the effect that final GDP announcements have on asset prices are likely to find that they have no impact and may therefore wrongly conclude that there is a disconnect between asset prices and macroeconomic fundamentals. We show that asset prices do not react to final GDP announcements because, 55 even though its relation to fundamentals is high, the timeliness of the GDP final release is poor and, as a result, the price impact of GDP final announcements relative to other announcements is small. Importantly, our analysis shows that the relationship between the intrinsic value of an announcement and its asset price impact is not perfect. In particular, we find that nonfarm 60 payroll has the biggest impact on U.S. Treasury bond yields, yet it is not the announcement with the biggest intrinsic value. This raises the possibility that there may be an overreaction to certain announcements, such as nonfarm payroll, because of the coordination value of public information beyond its intrinsic value, as in the model of Morris and Shin (2002). Another possibility is that our definition of the intrinsic value of macroeconomic announce- 65 mentsneedstobefurtherrefined. Forexample, onecouldconsiderotherprimitives, liketerm premia. Furthermore, even though our method allows announcements to vary in their importance over time, one could impose more structure to better estimate this time-variation, as in Bacchetta and van Wincoop (2013) and Goldberg and Grisse (2013), for example. Another extension would be to control for regime switches driven by, for instance, Alan Greenspan’s 70 2004 statement that nonfarm payroll numbers are more informative than the unemployment numbers.4 We leave these extensions to future research. 4Gu¨rkaynak and Wright (2013) show that Greenspan’s statement shifted the market’s attention to nonfarm payroll and away from the unemployment rate. This may be because investors became convinced that nonfarm payroll is indeed more informative about the state of the economy. Or it may be because investors learned what the Federal Reserve pays attention to it, allowing them to predict future policy actions. 3

2. Macroeconomic and Financial Data We collect data on 36 U.S. macroeconomic series, listed in Table 1, covering a broad set of real activity, prices, consumption, and investment variables. For each of these, we have 75 announcement dates and times, (median) market expectations, initial (actual) released values, and final (revised) values. Each announcement an is uniquely identified by the index p,t number n of the announcement series in Table 1, by the date and time t of its release, and by its reference period p. Nonfarm payroll released in early February, for example, has January as its reference period. Table 1 also provides the announcement unit used in both the agency 80 reports and the market expectations, the time(s) of the announcements, and the number of observations for each quarterly, monthly or weekly variable. For a given reference month p, the release of macroeconomic information follows a relatively stable and predictable schedule. Figure 1 shows, for instance, that the University of Michigan (UM) consumer confidence index is almost always released first, and nonfarm 85 payroll is always released on the first Friday of month p + 1 at 8:30 am ET. Following Andersen et al. (2003), the variables in Table 1 are presented in the order of their release date within each group (real activity, forward looking, etc.). Most of our macroeconomic data is from Bloomberg: announcement dates, times, reference periods, market expectations, final revised values and actual released values. The 90 Bloomberg data covers the sample from January 1997 to the present. We augment this with historical data from Money Market Services (MMS). The variables in the MMS dataset, however, start at different times. Many variables go back to the 1980’s, but initial jobless claims, consumer confidence, and GDP price deflator start in 1991; core CPI and core PPI start in 1992; and the University of Michigan consumer confidence index, the Chicago PMI, 95 and the Philadelphia Fed manufacturing index are not part of the MMS data. The final (revised) numbers, covering the period from 1990 to 2015 for all variables, were collected in May 2016 from Bloomberg, the various statistical agencies (BLS, BEA, etc.) and the FRED database. 4

Because we have actual release dates, times, expectations, and values for all variables 100 starting only in January 1997, we begin nowcasting in that month and analogously use January 1997 through December 2015 as sample in the event study. This choice is made for consistencybetweentheconstructionoftheannouncementcharacteristicsandtheassetprice impacts we aim to explain. However, since we have actual announcements or final values (or both) for all macroeconomic variables starting in 1990, we utilize the 1990-1996 sample 105 to estimate the transition matrices required in the nowcasting exercise. We also collect data for the Federal Funds Target Rate (FFTR) and its release dates. Our financial data are from the Federal Reserve Board and consist of daily changes in yields for the constant maturities 6-month, 1-, 2-, and 5-year U.S. Treasury bonds.5 We focus on the bond market as opposed to the equity or foreign exchange markets because, 110 as shown by the previous literature, e.g., Andersen et al. (2007), the link between Treasury bond price movements and macroeconomic news announcements is theoretically simpler and empirically stronger. 3. Asset Price Response to Macroeconomic Announcements In this section, we discuss the relationship between an announcement’s price impact and 115 what we label as its intrinsic value, timeliness, revisions, and relation to fundamentals within the context of a noisy rational expectations model. We also document the heterogeneous response of Treasury yields to 36 major macroeconomic announcements over the period 1997 through 2015. 3.1. Theoretical Framework 120 To provide a framework for defining an announcement’s price impact, its intrinsic value, and the effect of its underlying characteristics, we briefly discuss a stylized noisy rational 5We use daily changes instead of changes from a shorter time window around the announcement time (e.g., 5minutes)toaccountforthepricedriftsaheadofseveralmacroeconomicannouncementsdocumented in Kurov et al. (2016). Nevertheless, our conclusions are similar if we relate announcements’ characteristics to 5-minute price impacts. Daily data are from the Federal Reserve H.15 Selected Interest Rates (Daily) release. 5

expectationsmodelofpricereactionstopublicsignals, similartoKimandVerrecchia(1991a) andKimandVerrecchia(1991b). ThedetailsofthemodelareintheOnlineAppendix. Every period, the equilibrium price of a traded asset is a function of the representative investor’s 125 expectation of the asset’s final payoff. When a noisy public signal about this final payoff is received, the investor updates her expectation in a Bayesian manner. As a result, the price change is equal to the surprise component of the signal times a constant. We can label this constant as the price impact of the announcement because it is the coefficient one obtains when regressing price changes on the surprise component of the announcement. We can also 130 label this constant as the intrinsic value of the announcement because, in the model, it is equal to the weight placed by the investor on the signal when she is updating her belief about the asset’s payoff. In the empirical analysis that follows, we allow the intrinsic value of an announcement tobedifferentfromtheitspriceimpact. Toestimatetheintrinsicvalueoftheannouncement, 135 we assume that the asset’s payoff is related to the state of the economy, as proxied by GDP, GDP price deflator, or the FFTR. We further assume that the investor uses a Kalman filter tonowcastthestateoftheeconomy, andwedefinetheintrinsicvalueoftheannouncementas theweighttheinvestorputsontheannouncementwhennowcastingthestateoftheeconomy. Following previous studies, in the next sub-section, we estimate the price impact of 140 the announcement by regressing daily U.S. Treasury bond yield changes on macroeconomic news surprises (e.g., Fleming and Remolona (1997, 1999), Balduzzi et al. (2001), Goldberg and Leonard (2003), Gu¨rkaynak et al. (2005), Beechey and Wright (2009), and Swanson and Williams (2014)). The first main objective of our paper is to relate the intrinsic value of the announcement, the weight the investor puts on the announcement when nowcasting the 145 state of the economy, to the price impact of the announcement.6 The model makes several clear and intuitive predictions about the effect of an an- 6We are implicitly assuming that the expectations hypothesis holds. For this reason, we focus on shortterm bonds (6-month, 1-, 2- and 5-year maturities). In fact, we observe that our measure of intrinsic value, which does not take into account the impact of macroeconomic news announcements on the term premia, explainsahigherfractionofthevariationinpriceimpactfortheseshortermaturitiesthanfor10-and30-year maturity bonds (not tabulated in the paper). 6

nouncement’s characteristic – either its relation to fundamentals, timeliness, or revisions – on its intrinsic value and thus on its price impact (see the Online Appendix for details). A more timely announcement, an announcement that is more highly correlated with the 150 payoff of the risky asset, and an announcement that undergoes smaller revisions, has a higher intrinsic value and therefore has a higher price impact. To ensure consistency with our novel measure of the intrinsic value of the announcement, we define and estimate these characteristics within the nowcasting framework as well. The second main objective of our paper is then to assess which characteristic is most highly related to the price impact of the 155 announcement. 3.2. Price Impact of Announcements Followingtheliterature,wedefinethesurprisecomponentofamacroeconomicannouncement as the difference between its actual realization an and its corresponding market expectation p,t µn based on the information available before its release. The realization an is the value 160 p,t p,t of the macroeconomic variable n referring to period p, which is released at time t. Market expectations are measured as the median expectation across the set of Bloomberg/MMS forecasts. Also following the literature, the surprises are standardized by dividing each of them by their sample standard deviation in order to make the units of measurement comparable across macroeconomic variables. The standardized news surprise associated 165 with the release of macroeconomic variable n with reference period p at time t is therefore an −µn sn = p,t p,t (1) p,t σn s where σn is the sample standard deviation of an −µn based on all (initial) release times of s p,t p,t the respective macroeconomic variable n. We estimate the impact of a given macroeconomic announcement n on asset prices by 7

estimating the following equation 170 ∆y = α +β sn +(cid:15)n (2) t n n p,t t where ∆y is the daily change in Treasury yields (in basis points) and the intercept α is a t n time-invariant, variable-specific announcement return.7 Since σn is constant for any variable s n, the standardization in equation (1) does not have an impact on the statistical significance of the response estimates nor on the fit of equation (2).8 Table 2 reports the results of equation (2) for each of the 36 macroeconomic variables 175 across the four different Treasury bond maturities for the 1997–2015 sample period. Our measures of each variable’s price impact are the slope coefficient β on the standardized surn prise, whichrepresentsbasispointsperstandarddeviationofsurprise, andthecorresponding R2 of the regression. Consistentwiththepriorliterature, wefindlargedifferencesinslopecoefficientsandR2 180 across announcements. For instance, while the releases of nonfarm payroll and the Institute for Supply Management (ISM) PMI have large and significant price impacts, the releases of housing starts, durable goods orders, and the PPI have insignificant price impacts. It is this wide heterogeneity in asset price impact that we aim to explain in this paper.9 Consistent with the above model and the findings in Fleming and Remolona (1997), 185 Andersen et al. (2003), and Hess (2004), among others, we find that, within a general category of macroeconomic indicators, announcements released earlier tend to have greater impact than those released later. An obvious example is that of GDP, where the advance (first) release has the highest price impact. Similarly, the preliminary announcement of the 7For a nice review of the literature on event studies, including its caveats and limitations, please refer to Gu¨rkaynak and Wright (2013). 8By using identification through censoring, Rigobon and Sack (2008) estimate the share of the surveybased surprise due to noise. We choose not to follow their procedure because we allow the impact of news to vary with its noise. If we purged the noise from the announcement, we would underestimate the effect of noise on the price impact. 9In the Online Appendix, we present results for the sample period excluding the Federal Reserve’s zero lower bound, starting in December 2008. Consistent with the findings of Swanson and Williams (2014), the assetpriceimpactsaresomewhatstrongerpriortothezerolowerboundperiod, inparticularfortheshorter maturity bonds. 8

University of Michigan’s (UM) consumer confidence index (released around the middle of the 190 reference month) has a bigger effect on asset prices than the final announcement (released just before the end of the reference month). Other studies highlight the importance of the timeliness of an announcement. Hess and Niessen (2010) show that the price impact of the German Ifo business indicator diminished substantially after the creation of the German ZEW business indicator, because the ZEW 195 index is released before the Ifo index. Andersson et al. (2008) show that the reason for the small reaction of German bond prices to the aggregate German Consumer Price Index (CPI) announcement lies in the earlier release of CPI data for the individual German states. In a similar spirit, Ehrmann et al. (2011) show that there is no significant market reaction to Euro area macroeconomic announcements because all individual country releases are already 200 known (money supply being the only counter-example since it is only measured at the Euro area level). However, the results in Table 2 make it clear that timeliness is not the only characteristic that is related to the price impact of an announcement. For instance, even though the unemployment rate and nonfarm payroll are released simultaneously and early, surprises in 205 nonfarm payroll have a much larger price impact than surprises in the unemployment rate (more than 20 percent R2 versus 2 percent R2). Similarly, core CPI has a higher price impact than headline CPI. In light of the model above, it may be that nonfarm payroll and core CPI have a bigger price impact because they either undergo smaller revisions after their initial release or because they are more “useful” to investors in forecasting a fundamental variable 210 of interest, such as GDP, GDP deflator or the FFTR. In the following, we define our novel measures of announcement characteristics and we investigate how these characteristics help explain the heterogeneity in price impact of macroeconomic announcements. 4. Measuring and Decomposing the Intrinsic Value of Announcements In this section we describe the methodology for consistently measuring an announcement’s 215 intrinsic value and its components: timeliness, revisions, and relation to fundamentals. We 9

start by setting up a nowcasting framework, which we subsequently use to define these four characteristics. 4.1. Nowcasting GDP Growth, Inflation, and the Federal Funds Target Rate We propose and estimate a novel empirical measure of an announcement’s intrinsic value and 220 its components. We define the intrinsic value of an announcement as its importance in nowcasting three primitives: GDP advance, GDP price deflator advance, and the FFTR.10 We generate nowcasts based on a dynamic factor model, because this class of models parsimoniously captures the evolution of the high-dimensional vector of macroeconomic announcements. Whenever new information arrives, the Kalman filter provides an estimate (nowcast) 225 of the current state vector, which we then use to forecast the current level of the primitive of interest. Repeating this procedure every time new information arrives, generates a timeseries of Kalman gains and regression coefficients, which forms the basis of our measures of intrinsic value, timeliness premium, revision premium, and relation to fundamentals.11 Our approach to nowcasting is similar to Evans (2005) and Giannone et al. (2008). We 230 assume that the state vector of the economy, Φ , follows a VAR(1) process, captured at p,t time t by the state equation Φ = B Φ +C ν , (3) p,t t p−1,t t p,t where ν ∼ WN(0,I ). Note that there are two time subscripts, p and t. The state of the p,t 2×2 economy evolves at a monthly frequency, indexed by the reference period p. The subscript t governs how much information is available about the current and the past state vectors, 235 and identifies specific times within the month. This setup naturally maps the ever-evolving information set – with its missing values, revisions, and irregular announcement dates – into our data structure. Because the dataset changes with each data release, the state space 10OurprimaryreasonforfollowingtheKalmanfilter-basednowcastingapproachisthatitsdatastructure lends itself to traceable counterfactual exercises. Macroeconomic forecasting with mixed-frequency data has received considerable attention in recent years, e.g., Andreou et al. (2010). Nevertheless, the Kalman filter remains the method of choice in terms of accuracy, at the cost of being computationally more demanding than, for instance, mixed data sampling (MIDAS) regressions (Bai et al., 2013). 11SectionO.2.intheOnlineAppendixprovidesextensivedetailsondatamanagement,timingconventions, and the nowcasting procedure. 10

model is re-estimated at each data release time t. The corresponding observation equation for a given information set t is 240 A = D Φ +ε , (4) p,t t p,t p,t (cid:2) (cid:3)(cid:48) where ε ∼ WN(0,V ), and A = a1 ,...,aN is the monthly vector of N macroecop,t p,t p,t p,t p,t nomic variables containing the values an available at time t. The variable an contains only p,t p,t values announced on or before time t for the macroeconomic announcement n. The 36 macroeconomic announcements listed in Table 1 and the FFTR series, which are assumed to jointly capture the state of the U.S. economy, are used in the nowcasting 245 exercise, either in their original reporting units or transformed in order to approximate a linear relationship with the forecasting object. For variables reported in percent or percent changes, the original reporting unit is used, while variables reported in levels are transformed into percent changes. For example, the retail sales series, reported as a percent change, is not transformed, while the new home sales series is transformed from levels to percent change. 250 For indexes, we use the original reporting unit.12 Weestimatethestatespacerepresentationgivenbyequations(3)and(4)withthetwostep procedure of Giannone et al. (2008).13 The estimation proceeds in four steps, which we repeat for each announcement release time t. We use an expanding window from January 1990 until time t, starting with the window ending on t = January 1st, 1997. 255 First, we consolidate variables that are released piece by piece, namely GDP (advance, preliminary, final), GDP price deflator (advance, preliminary, final), and the University of Michigan consumer confidence index (preliminary, final) into one series, respectively. Thus we have N = 32 consolidated macroeconomic time series. However, for determining the 12Moredetailsontheoriginalreportingunitsandpossibletransformationofeachmacroeconomicvariable are collected in Section O.3. of the Online Appendix. 13Such “partial” models, specifying the target variable separately from the model of the predictors, are widely used in policy institutions (Ban´bura et al., 2013). For our sample, this two-step procedure outperformed the one-step procedure in nowcasting GDP in terms of RMSFE. Further, the two-step approach allows us to tailor the second step to the forecasting target, which we exploit when replacing equation (4) for the FFTR by an ordered probit specification. 11

intrinsicvalue,wekeeptrackofeachobservation’soriginaldesignation(advance,preliminary, 260 or final). Given t, each time-series is standardized to zero mean and unit standard deviation. Second, we define a five-dimensional state vector based on five principal components Φ extracted from the balanced part of the sample. Two principal components are based p,t on all announcement series. Three further principal components are based on the subsets of real, nominal, and forward-looking announcement series, respectively. The matrix C collects 265 t the five eigenvectors, linking the factors Φ with the announcements A .14 p,t p,t Third, the Kalman filter is estimated given information available up until time t and the Kalman gains assigned to the announcements at the end of the sample are retrieved. Specifically, to construct the time-series of the intrinsic value of announcement n, only the gains kn at the time of a new release of macroeconomic variable n are used. 270 t Fourth, given the information at time t, we (Kalman-)smooth the latent factors. Then weusethesefactorstofitaforecastingmodelforthenowcastingtargetvariables, analogously to equation (4). For the nowcasting targets GDP and the GDP price deflator, a linear model at quarterly frequency is used, whereas for the FFTR an ordered probit specification at monthly frequency is employed. For each forecasting target, indexed by j, we estimate 275 coefficients (marginal effects for the FFTR) D ˜j on the latent factors at each point in time. t The absolute value of the product w(j)n = |D ˜jkn| of this coefficient (row) vector with the t t t respective column of the Kalman gain matrix is the weight on announcement n at time t for nowcasting the variable j.15 When these weights are derived from actual data released according to the actual 280 release schedule, we refer to them as w (j)n. In order to estimate the effect of timeliness A t and revisions, we create counterfactual datasets and apply the same nowcasting procedure 14We extract two factors from all announcements because for GDP such a model performs notably better at nowcasting and at forecasting 1-month-ahead GDP than one factor. For GDP deflator and FFTR, the performance is similar across different numbers of factors. 15Wetakeabsolutevaluestocapturethedirection-freeimpactofanannouncement. Becausewedetermine this weight by a two-step procedure, it differs from the weights implicitly assigned to observations within the Kalman filter as in, e.g., Koopman and Harvey (2003) and Ban´bura and Ru¨nstler (2011). In contrast, in our paper, the weight combines the gains determined by the Kalman filter with the coefficients from a separate forecasting regression, and captures the empirical relevance of only the most recent announcement release. 12

on these new datasets. These datasets differ from the original one in terms of release timing, revisionstatus, orboth. Wemodifytherespectivepropertyofonlyonemacroannouncement series n per nowcasting exercise. 285 Tocontrolforreleasetiming,wecounterfactuallyreorderthedata. Todoso,weidentify theearliestannouncementforeachreferenceperiodandsetthecounterfactualannouncement time of the variable of interest to one second before this previously earliest announcement. Applying the nowcasting procedure to these reordered actual datasets yields the weight series w (j)n. 290 RA t To control for revision status, we counterfactually replace all releases of the variable of interest by the final revised values. By subjecting the original data to both this counterfactual replacement with final values and the counterfactual time reordering, the nowcasting procedurewiththiscounterfactualdatasetofreordered final announcementsyieldstheweight series w (j)n. 295 RF t 4.2. Intrinsic Value and its Decomposition We define the intrinsic value I(j)n of macroeconomic variable n with respect to target varit able j (GDP, GDP deflator or FFTR) as the natural logarithm of the nowcasting weight put on macroeconomic variable n at the time t of its announcement, I(j)n ≡ log[w (j)n]. The t A t intrinsic value can therefore be thought of as the importance placed on the announcement 300 when nowcasting the state of the economy. Columns 1, 5, and 9 of Table 3 report the time-series average of our novel measure of intrinsic value of each macroeconomic variable for the three nowcasting targets. Note that because the weights, w (j)n, turn out to be between zero and one, the intrinsic value, the A t logarithmoftheweight, isnegative. Thismeansthatanannouncementwithasmallnegative 305 number has large intrinsic value, and an announcement with a large negative number has very little intrinsic value. Based on this metric, Table 3 indicates that forward-looking announcements such as the consumer confidence indices and the PMI indices have large intrinsic values (small negative numbers) when nowcasting GDP and the FFTR. Similarly, 13

price variables such as CPI and PPI appear to have large intrinsic value when nowcasting 310 the GDP price deflator. We decompose the intrinsic value I(j)n of each macroeconomic variable n for a given t target variable j into the announcement’s relation to fundamentals F(j)n, a timeliness pret mium T(j)n, and a revision premium R(j)n: t t I(j)n ≡ F(j)n +T(j)n +R(j)n, (5) t t t t where each component is defined using the nowcasting weights defined in the previous sub- 315 section:16 (cid:20) (cid:21) (cid:20) (cid:21) w (j)n w (j)n log[w (j)n] ≡ log[w (j)n]+log A t +log RA t . (6) A t RF t w (j)n w (j)n RA t RF t Each term in equation (6) reflects one of the announcement characteristics in equation (5): • The intrinsic value, I(j)n ≡ log[w (j)n], is the nowcasting weight placed on the actual t A t macroeconomic announcement at the time of its release. • The relation to fundamentals, F(j)n ≡ log[w (j)n], is the nowcasting weight placed 320 t RF t on the macroeconomic announcement independent of its timing and its revisions. • The timeliness premium, T(j)n = log[w (j)n]−log[w (j)n], is the difference between t A t RA t the nowcasting weight placed on the actual macroeconomic announcement at the time of its release and the nowcasting weight placed on the actual announcement when it is reordered to be the first release in each forecasting period. 325 • The revision premium, R(j)n ≡ log[w (j)n]−log[w (j)n], is the difference between t RA t RF t the nowcasting weight placed on the actual announcement when it is reordered to be the first release in each forecasting period and the nowcasting weight placed on the announcement when it is reordered and replaced by its final revised value. 16Starting with the factorization w A (j)n t ≡w RF (j)n t w w R A A ( ( j j ) ) n t n t w w R R F A( ( j j ) ) n t n t we obtain equation (6) by taking the natural logarithm of this identity. 14

Wenowdiscusseachcomponentoftheintrinsicvalueinturn, whicharepresentedinTable3, 330 and compare them with some alternative na¨ıve measures. 4.3. Relation to Fundamentals In the noisy rational expectations model, market participants put more weight on announcements that are more closely related to fundamentals. The above definition, F(j)n ≡ t log[w (j)n], captures this idea since it is the nowcasting weight placed on the announce- 335 RF t ment that has been replaced with its final revised value (to remove the impact of revisions) and reordered so that it is the first release in each reference cycle (to remove the impact of timing). The times-series average of this novel measure of relation to fundamentals is reported in columns 2, 6, and 10 of Table 3 for each macroeconomic variable. As for the intrinsic 340 value, an announcement with a small negative number has a large relation to fundamentals, and an announcement with a large negative number has a small relation to fundamentals. Intuitively, GDP announcements are closely related to fundamentals when nowcasting GDP, as well as nonfarm payroll and forward looking indicators. GDP deflator announcements, as well as CPI and PPI announcements, are most closely related to fundamentals when 345 nowcasting the GDP price deflator. A mix of real activity and inflation announcements have a high relation to fundamentals when nowcasting the FFTR. Alternatively, onecouldmeasuretherelationtofundamentalsbylookingatthecorrelationofeachannouncementwithGDP,theGDPpricedeflator, andFFTR.Thesecorrelations are reported in columns 13-15 of Table 3. Note that the correlations between our novel mea- 350 sures and these alternative measures are 0.7, 0.6 and 0.5, when nowcasting GDP, the GDP price deflator and FFTR, respectively. 4.4. Timeliness Premium In the noisy rational expectations model, market participants put more weight on announcements that are more timely. The definition of this premium, T(j)n = log[w (j)n]− 355 t A t 15

log[w (j)n], captures this idea because thereby it is the difference between the actual now- RA t casting weight and the reordered nowcasting weight. This difference should be negative and small for timely announcements, but large and negative for announcements that are released late and whose re-ordering improves their nowcasting ability. The time-series average of this novel measure of timeliness is reported in columns 3, 7, 360 and 11 of Table 3. Looking at GDP announcements, our timing premium is higher (smaller negative number) for the timelier variable, GDP advance, than for GDP final. Forward looking variables that are released early, such as the confidence indices, have very high timeliness premia. The previous literature (e.g., Fleming and Remolona (1997)) uses the reporting lag 365 as a measure of timing discount, which is the difference between the announcement date and the end of the reference period.17 The time-series average of each variable’s reporting lag (measured in days) is shown in column 16 of Table 3. We call reporting lag a timing discount because the larger the number the worse the timing of the announcement. Thus the correlation between our timing premium and reporting lag should be negative. Indeed, we 370 find the correlations to be -0.47, -0.52, and -0.37 when the target variables are GDP, GDP price deflator, and the FFTR, respectively. One drawback of the announcement’s reporting lag as a measure of timeliness is that it isalinearfunctionoftime, soanimprovementintimelinessof, say, sixdaysisthesameforan early and a late announcement. However, we expect a 7-day reporting lag announcement to 375 gainmorefrommovingupitsreleasedatesixdaysthana21-dayreportinglagannouncement moving up six days. This is because the 7-day reporting lag announcement will now be the first announcement while the 21-day reporting lag will be the 15th announcement, and it is likely that the earlier releases have already conveyed sufficient information. The novel measure we propose explicitly takes into account the position of the announcement when 380 17There is a difference between the end of the reference period and the end of the survey period. For instance, at the Bureau of Labor Statistics, “employment data refer to persons on establishment payrolls who received pay for any part of the pay period that includes the 12th of the month” (http://www.bls.gov/web/cestn1.htm). This means that taking the end of the month as the end of the reference period is not exact, because the surveying stopped much earlier in the month. 16

computing the nowcasting gain in timeliness. This is the reason why two announcements releasedatthesametime,liketheunemploymentrateandnonfarmpayroll,canhavedifferent timeliness premia. 4.5. Revision Premium In the noisy rational expectations model, market participants put more weight on announce- 385 ments that undergo smaller revisions. The above definition of this premium, R(j)n ≡ t log[w (j)n]−log[w (j)n], captures this idea since it is the difference between the now- RA t RF t casting weight of the actual announcement minus the weight of its final revised value, both independent of the timing of the announcement (reordered). This number should be negative and small for announcements that are not heavily revised, but large and negative for 390 announcements that are heavily revised and their revisions improve their nowcasting ability. The times-series average of this novel measure of revisions is reported in columns 4, 8, and 12 of Table 3. Overall, there is significantly less variation in revision premium across announcements compared to the other characteristics. Many numbers are even positive, which indicates that the final revised values do worse in nowcasting the given primitive than 395 the actual releases. This is consistent with the findings in Orphanides (2001) who shows that a Taylor rule with real-time macroeconomic announcements performs better than a Taylor rule with final revised numbers. The previous literature (e.g., Gilbert (2011)) uses an alternative measure of revision noise, namely the absolute value of the difference between the final revised value and the 400 initial release. This measure captures the magnitude of the revisions that an announcement undergoes.18 This definition includes both sample and benchmark revisions and assumes that the last available value reflects the “true” situation.19 In the last column of Table 3, 18Inordertonormalizetheunitofmeasurementacrossmacroeconomicseries,wenormalizethisalternative measure of revision magnitude (cid:12) (cid:12) (cid:12)an −an (cid:12) p,∞ p,t σ |an −an | p,∞ p,t where t is the time of the initial release of an and an is the final revised value. p,t p,∞ 19As a robustness check, we also use the first-available sample revisions for the variables available in the 17

we report the time-series average of this measure of revision magnitude. The correlation between our novel measure of revision premium and the alternative 405 revision magnitude (discount) measure is on average -0.10 for the three nowcasting targets (GDP, GDP deflator and FFTR). This occurs because the revision magnitude does not take into account the possibility that the revised (final) number is less useful in nowcasting target variables than the original (first-released) number. This measure only captures the magnitude of the revision but not the relevance of a revision, which our nowcasting measure 410 does capture. For example, the UM consumer confidence index is heavily revised, and hence its preliminary release has a big revision magnitude shown in the last column of Table 3. However,wefindthatthepreliminaryreleasehasarevisionpremiumofzerowhennowcasting the FFTR, which suggests that the final revised value does no better than the initial released value. 415 5. Relating the Price Impact to the Announcements’ Characteristics In this section, we relate our novel measure of the announcements’ intrinsic value, as well as its components (relation to fundamentals, timeliness premium, and revision premium) to their price impact. We first examine whether our measures affect the impact of announcement surprises on asset prices using the full sample. Then we investigate whether 420 our measures explain the cross-section of price impact. All results are presented for the full sample period, but qualitatively similar results using the period excluding the Federal Reserve’s zero lower bound period are presented in the Online Appendix. 5.1. Direct Impact on Asset Returns Toassesstheimportanceoftheannouncements’characteristics,theeventstudyexercisefrom 425 Section 3. is repeated with the intrinsic value, relation to fundamentals, timeliness premium, and revision premium added into the regressions. However, rather than estimating the price Federal Reserve Bank of Philadelphia’s Real-Time Data Set and Bloomberg. The results are qualitatively similar. 18

impactseparatelyforeachannouncement(aswedoinequation(2)andTable2), weestimate an average price impact β(j) across announcements, and only allow this price impact to vary across announcements according to the announcements’ characteristics X(j) . More 430 p,t precisely, we estimate the following equation separately for each target variable j: ∆y = β (j)+β(j)s +β (j)s X(j) +(cid:15)(j) , (7) t 0 p,t x p,t p,t t where ∆y are the daily changes in U.S. Treasury bond yields in basis points around the t macroeconomic releases and s are the surprise components of all the macroeconomic anp,t nouncements pooled together, defined as in equation (1).20 The interaction term, s X(j) , p,t p,t allows the price impact of the announcement to vary across the announcements’ charac- 435 teristics, which are either the intrinsic value (I), relation to fundamentals (F), timeliness premium (T), revision premium (R), or a vector with all three characteristics.21 Westandardizeandsmoothourmeasureofintrinsicvalueoftheannouncement. Specifically, we divide each characteristic by its standard deviation estimated across all announcementsandalltimes. Thiseasestheinterpretationofthecoefficientestimates. Inaddition, we 440 smooth the weights by taking a 12-month backward-looking moving average. The assumption is that, in calculating the importance of an announcement, investors take the average importance over the past year. There is one table of results per nowcasted primitive j: Table 4 for GDP, Table 5 for the GDP price deflator, and Table 6 for the FFTR. Columns 2 to 5 in all three tables 445 show the results with each different characteristic included in the regression in isolation, and column 6 shows all three characteristics competing against each other. Column2showsthat, forallnowcastingtargets, theintrinsicvalueofanannouncement 20We change the sign of the surprise of two announcements, the unemployment rate and initial jobless claims, so that positive surprises are associated with either higher economic activity or higher inflation than expected. 21Wedonotincludeamaineffectfortheannouncementcharacteristicbecausethenoisyrationalexpectations model predicts that the announcement characteristic only affects the price impact, and does not affect the yield change. Consistent with this view, when we include a main effect for the announcement, the main effect is not statistically significant and our results are qualitatively similar. 19

has an economically and statistically significant effect on the asset price impact of that announcement. The sign of the coefficient is consistent with theory: the bigger the intrinsic 450 value of the announcement is, the bigger is its price impact. For example, a one-standard deviation surprise in an announcement with an average intrinsic value of zero increases the 6-month Treasury yields by about 1 basis point when the nowcasting target is GDP (Table 4, column 2). If we increase the intrinsic value of this announcement by one standard deviation, a surprise on this announcement will increase 6-month bond yields by about 1.2 455 basis point (1.022+0.216), which is a 20 percent increase in the impact on yields. Repeating this calculation, we see that the increase in price impact due to intrinsic value is about 15 percentacrossmaturitieswhenthenowcastingtargetistheGDPpricedeflatorortheFFTR. Columns 3 through 6 suggest that, across forecasting targets, the relation to fundamentals and timeliness premium are the most relevant announcement characteristics; and 460 revision noise is, most of the time, statistically insignificant. Column 6 suggests that increasing the timing of an announcement by one standard deviation increases the impact of the surprise by about 10 to 20 percent, when the nowcasting variable is GDP, while increasing the relation to fundamentals by one standard deviation increases the impact of the surprise by about 20 to 30 percent. The sign of these effects is consistent with the theoretical model 465 summarized in Section 3.1. The importance of the timeliness premium suggests that financial markets indeed learn in a Bayesian manner. Imprecise, but early, information can be as useful from a nowcasting perspective as precise, but late news. 5.2. Determinants of Average Surprise Impact 470 In the previous sub-section, we examined whether our novel measures affect the impact of announcement surprises on asset prices using the full sample. We now investigate whether our measures explain the cross-section of price impact and how they compare with the alternative announcement characteristics previously used in the literature, such as reporting lag. In this cross-sectional analysis, we take our estimates of the asset price impact, namely 475 20

the R2 from equation (2) and Table 2, and estimate the following equation: R2(j) = α (j)+α (j)X (j)+(cid:15) (j), (8) n 0 x n n where X is the time-series average of our announcement characteristics. Table 7 shows the n results where X is the announcement’s intrinsic value for all three nowcasting targets j, n namely GDP, GDP price deflator and FFTR. Table 8 shows the results for GDP only, but where X is the announcement’s relation to fundamentals, timeliness premium, revision 480 n premium, as well as the alternative measures of these components used by the previous literature: correlation with GDP, reporting lag, and revision magnitude. We include each of these characteristics separately because our sample is small, with only 36 observations (one estimate of price impact per announcement). Lookingacrosscolumns1through3inTable7, wefindthatourintrinsicvaluemeasure, 485 when using GDP or FFTR as our nowcasting targets, explains a significant fraction (6 to 18 percent) of the variation in the price impact of announcement surprises, as measured by the R2.22 In contrast, using GDP deflator as the nowcasting target is not useful at all. This finding may be an artifact of the sample period we analyze, during which inflation was relatively low and inflation expectations may not have played a big role in nominal U.S. 490 Treasurybondprices.23 UsingGDPasthenowcastingtargetisalsomoreusefulinexplaining the variation than using the FFTR. This may not be surprising because the impact of news about the FFTR on nominal U.S. Treasury bonds includes offsetting effects on real and inflation components, as shown by Beechey and Wright (2009). Columns 2 through 4 of Table 8 further confirm that an announcement’s relation to 495 fundamentals and timeliness premium are more important in explaining the asset price impactofmacroeconomicnewsannouncementsthantherevisionpremium. Timelinessexplains from 6 to 14 percent of the variation in asset price impact coefficients, and relation to funda- 22We obtain qualitatively similar results if we use the slope coefficients β as measure of price impact. n 23Indeed we find that prior to the “Zero Lower Bound‘” period the GDP Deflator target is much more relevant – similar in magnitude to the FFTR. The Online Appendix reports these results. 21

mentals explains 8 to 12 percent of the variation in asset price impact coefficients. However, the revision characteristic explains only 0.8 to 4 percent of this variation. Overall, column 1 500 shows that our novel measure of intrinsic value explains the largest fraction of the variation in price impact when compared to its three components and their alternative measures. Amongst the alternative measures in columns 5 through 7, correlation with GDP is mostly insignificant but reporting lag is significant and explains a sizeable fraction of the variation in asset price impact. Interestingly, revision magnitude is statistically significant, 505 but the sign is the opposite of what our theoretical model would predict: announcements that undergo larger revisions have a higher price impact. The counter-intuitive sign suggests that one should not consider the magnitude of the revisions in isolation; instead, one should consider both the magnitude of the revision and the relevance of the revision, which our nowcasting framework does. 510 6. Conclusion In this paper, we propose and estimate a novel measure of the intrinsic value of macroeconomic announcements. Our definition is based on the announcement’s ability to nowcast GDP growth, the GDP price deflator, and the FFTR. We decompose this intrinsic value into three separate announcement characteristics: relation to fundamentals, timeliness, and 515 revisions. We find that timeliness and relation to fundamentals are the most significant characteristics in explaining the variation in the announcements’ asset price impact on U.S. Treasury bonds. Our study offers two additional takeaways for policy makers and future research. First, the price response to a particular type of announcements cannot be analyzed in isolation. 520 The effect that announcements have on asset prices crucially depends on the information environment. Second, our analysis shows that the relationship between the intrinsic value of anannouncementanditsassetpriceimpactisnotperfect. Inparticular,wefindthatnonfarm payroll has the biggest impact on U.S. Treasury bonds, yet it is not the announcement with the biggest intrinsic value. This raises the possibility that there may be an overreaction to 525 22

certain announcements. 23

References Almeida, A., Goodhart, C., Payne, R., 1998. The effects of macroeconomic news on high frequency exchange rate behavior. Journal of Financial and Quantitative Analysis 33 (3), 383–408. Andersen, T. G., Bollerslev, T., Diebold, F. X., Vega, C., 2003. Micro effects of macro announcements: Real-time price discovery in foreign exchange. American Economic Review 93 (1), 38–62. Andersen, T. G., Bollerslev, T., Diebold, F. X., Vega, C., 2007. Real-time price discovery in global stock, bond and foreign exchange markets. Journal of International Economics 73 (2), 251–277. Andersson, M., Ejsing, J. W., von Landesberger, J., 2008. How do Euro area inflation expectations evolve over time?, European Central Bank. Andreou, E., Ghysels, E., Kourtellos, A., 2010. Regression models with mixed sampling frequencies. Journal of Econometrics 158 (2), 246–261. Bacchetta, P., van Wincoop, E., 2013. On the unstable relationship between exchange rates and macroeconomic fundamentals. Journal of International Economics 91 (1), 18–26. Bai, J., Ghysels, E., Wright, J. H., 2013. State space models and MIDAS regressions. Econometric Reviews 32 (7), 779–813. Balduzzi, P., Elton, E. J., Green, T. C., 2001. Economic news and bond prices: Evidence from the U.S. treasury market. Journal of Financial and Quantitative Analysis 36 (4), 523–543. Ban´bura, M., Giannone, D., Modugno, M., Reichlin, L., July2013.Now-castingandtherealtime data flow. In: Elliott, G., Timmermann, A. (Eds.), Economic Forecasting. Vol. 2A of Handbooks in Economics. North Holland, Ch. 4, pp. 195–236. Ban´bura, M., Ru¨nstler, G., 2011. A look into the factor model black box: Publication lags and the role of hard and soft data in forecasting GDP. International Journal of Forecasting 27 (2), 333–346. Bartolini, L., Goldberg, L., Sacarny, A., 2008. How economic news moves markets. Current Issues in Economics and Finance 14 (6). Beechey, M. J., Wright, J. H., 2009. The high-frequency impact of news on long-term yields and forward rates: Is it real? Journal of Monetary Economics 56 (4), 535–544. Bekaert, G., Engstrom, E., 2010. Inflation and the stock market: Understanding the “Fed Model”. Journal of Monetary Economics 57 (3), 278–294. Bernanke, B. S., Kuttner, K. N., 2005. What explains the stock market’s reaction to Federal Reserve policy? Journal of Finance 60 (3), 1221–1257. 24

Boyd, J. H., Hu, J., Jagannathan, R., 2005. The stock market’s reaction to unemployment news: Why bad news is usually good for stocks. Journal of Finance 60 (2), 649–672. Ehrmann, M., Fratzscher, M., 2004. Taking stock: Monetary policy transmission to equity markets. Journal of Money, Credit, and Banking 36 (4), 719–737. Ehrmann, M., Fratzscher, M., 2005. Exchange rates and fundamentals: New evidence from real-time data. Journal of International Money and Finance 24 (2), 317–341. Ehrmann,M.,Fratzscher,M.,Gu¨rkaynak,R.,Swanson,E.,2011.Convergenceandanchoring of yield curves in the euro area. Review of Economics and Statistics 93 (1), 350–364. Ehrmann, M., Sondermann, D., 2012. The news content of macroeconomic announcements – what if central bank communication becomes stale? International Journal of Central Banking 8 (3), 1–53. Evans,M.,2005.Wherearewenow? Real-timeestimatesofthemacroeconomy.International Journal of Central Banking 1 (2), 127–175. Faust, J., Rogers, J. H., Wang, S.-Y. B., Wright, J. H., 2007. The high-frequency response of exchange rates and interest rates to macroeconomics announcements. Journal of Monetary Economics 54, 1051–1068. Flannery, M. J., Protopapadakis, A. A., 2002. Macroeconomic factors do influence aggregate stock returns. Review of Financial Studies 15 (3), 751–782. Fleming, M. J., Remolona, E. M., 1997. What moves the bond market? Economic Policy Review 3 (4), 31–50. Fleming, M. J., Remolona, E. M., 1999. Price formation and liquidity in the U.S. treasury market: The response to public information. Journal of Finance 54 (5), 1901–1915. Giannone, D., Reichlin, L., Small, D. H., 2008. Nowcasting: The real-time informational content of macroeconomic data. Journal of Monetary Economics 55 (4), 665–676. Gilbert, T., 2011. Information aggregation around macroeconomic announcements: Revisions matter. Journal of Financial Economics 101, 114–131. Goldberg, L., Leonard, D., 2003. What moves sovereign bond markets? The effects of economic news on U.S. and German yields. Current Issues in Economics and Finance 9 (9). Goldberg, L. S., Grisse, C., 2013. Time variation in asset price responses to macro announcements. Staff Reports 626, Federal Reserve Bank of New York. Gu¨rkaynak,R.,Wright,J.,2013.Identificationandinferenceusingeventstudies.TheManchester School 81 (S1), 48–65. Gu¨rkaynak, R. S., Sack, B. P., Swanson, E. T., 2005. The excess sensitivity of long-term interest rates to economic news: Evidence and implications for macroeconomic models. American Economic Review 95 (1), 425–436. 25

Hamilton, J.D., Jorda`, O., 2002.Amodelofthefederalfundsratetarget.JournalofPolitical Economy 110 (5), 1135–1167. Hess, D., 2004. Determinants of the relative price impact of unanticipated information in U.S. macroeconomic releases. Journal of Futures Markets 24 (7), 609–629. Hess, D., Niessen, A., 2010. The early news catches the attention: On the relative price impact of similar economic indicators. Journal of Futures Markets 30 (10), 909–937. Kim, O., Verrecchia, R. E., 1991a. Market reaction to anticipated announcements. Journal of Financial Economics 30, 273–309. Kim, O., Verrecchia, R. E., 1991b. Trading volume and price reactions to public announcements. Journal of Accounting Research 29, 302–321. Koopman, S. J., Harvey, A., 2003. Computing observation weights for signal extraction and filtering. Journal of Economic Dynamics and Control 27 (7), 1317–1333. Kurov, A., Sancetta, A., Strasser, G., Wolfe, M. H., 2016. Price drift before U.S. macroeconomic news: Private information about public announcements? Working Paper 1901, European Central Bank. Lapp, J. S., Pearce, D. K., 2012. The impact of economic news on expected changes in monetary policy. Journal of Macroeconomics 3 (2), 362–379. Morris, S., Shin, H. S., 2002. Social value of public information. American Economic Review 92, 1521–1534. Neely, C. J., Dey, S. R., 2010. A survey of announcement effects on foreign exchange returns. Federal Reserve Bank of St. Louis Review 92 (5), 417–463. Orphanides, A., 2001. Monetary policy rules based on real-time data. American Economic Review 91 (4), 964–985. Rigobon,R.,Sack,B.,January2008.Noisymacroeconomicannouncements,monetarypolicy, and asset prices. In: Campbell, J. Y. (Ed.), Asset Prices and Monetary Policy. NBER Chapters. NBER, Ch. 8, pp. 335–370. Swanson, E., Williams, J., 2014. Measuring the effect of the zero lower bound on mediumand longer-term interest rates. American Economic Review 104 (10), 3154–3185. Wallis, K. F., 1986. Forecasting with a econometric model: The “ragged edge” problem. Journal of Forecasting 5 (1), 1–13. 26

UM Consumer Confidence Index Philadelphia Fed Index Conference Board Consumer Confidence Index Chicago PMI and ISM/NAPM PMI Nonfarm Payroll + Unemployment Rate + Average Hourly Earnings Retail Sales + Retail Sales Less Auto PPI + PPI Core Industrial Production + Capacity Utilization CPI + CPI Core Housing Starts Government Budget Deficit Durable Goods Orders GDP + GDP Price Index (quarterly) New Home Sales Personal Income + Personal Consumption Expenditures Index of Leading Indicators Factory Orders Construction Spending Consumer Credit Business Inventories Trade Balance 22 25 28 31 3 6 9 12 15 18 21 24 27 30 2 5 8 11 14 17 20 23 Reference Month p Month p+1 Month p+2 Fig. 1. Macroeconomic announcement calendar. Note: This figure shows the usual calendar timing of U.S. macroeconomic announcements across the month. The reference month is labeled as p with most variables released in the subsequent month and some released up to six weeks later. Each GDP series (advance, preliminary, or final) is released on a quarterly basis. Not represented in the figure is the initial jobless claims announcement, which is released weekly on Thursday for the previous week. The University of Michigan releases a final version (not shown) of their consumer confidence index two weeks after their preliminary one.

Table 1 Characteristics of Macroeconomic Announcements. n Announcement Unit Release Time Obs. Quarterly Announcements Real Activity 1 GDP advance (first estimate) % change 8:30 76 2 GDP preliminary (second estimate) % change 8:30 76 3 GDP final (third estimate) % change 8:30 76 Prices 4 GDP price deflator advance % change 8:30 76 5 GDP price deflator preliminary % change 8:30 76 6 GDP price deflator final % change 8:30 76 Monthly Announcements Real Activity 7 Unemployment rate % 8:30 228 8 Nonfarm payroll employment change 8:30 228 9 Retail sales % change 8:30 228 10 Retail sales less automobiles % change 8:30 227 11 Industrial production % change 9:15 228 12 Capacity utilization % 9:15 228 13 Personal income % change 8:30/10:00 228 14 Consumer credit change 15:00 228 Consumption 15 Personal consumption expenditures % change 8:30 228 16 New home sales level 10:00 227 Investment 17 Durable goods orders % change 8:30/9:00/10:00 227 18 Construction spending % change 10:00 227 19 Factory orders % change 10:00 227 20 Business inventories % change 8:30/10:00 228 Government Purchases 21 Government budget deficit level 14:00 228 Net Exports 22 Trade balance level 8:30 228 Prices 23 Average hourly earnings % change 8:30 228 24 Producer price index (PPI) % change 8:30 228 25 Core PPI % change 8:30 228 26 Consumer price index (CPI) % change 8:30 228 27 Core CPI % change 8:30 228 Forward Looking 28 U. Michigan (UM) consumer confidence preliminary index 9:55/10:00 200 29 Philadelphia Fed manufacturing index index 10:00 227 30 UM consumer confidence final index 9:55/10:00 200 31 Conference Board (CB) consumer confidence index 10:00 228 32 (ISM-)Chicago Purchasing Managers Index (PMI) index 10:00 226 33 ISM Manufacturing PMI index 9:15/10:00 228 34 Housing starts level 8:30 226 35 CB leading economic index % change 8:30/10:00 228 Weekly Announcements 36 Initial jobless claims level 8:30 992 Note: The table displays the 36 U.S. macroeconomic variables analyzed in the paper, along with the announcementunitusedinboththeagencyreportsandthemarketexpectations,thetimeoftheannouncement release (Eastern Time), and the number of available data releases. The sample covers January 1997 to December2015. ISMstandsforInstituteforSupplyManagement,formerlyNationalAssociationofPurchasing Management (NAPM).

2 elbaT .sdleiY yrusaerT .S.U no sesirpruS tnemecnuonnA cimonoceorcaM fo tceffE yrusaerT raeY-5 yrusaerT raeY-2 yrusaerT raeY-1 yrusaerT htnoM-6 .sbO 2R .ffeoC 2R .ffeoC 2R .ffeoC 2R .ffeoC tnemecnuonnA n 67 810.0 820.1 860.0 **136.1 940.0 *149.0 450.0 **767.0 ecnavda PDG 1 67 000.0 820.0 000.0 000.0 000.0 930.0- 210.0 892.0 yranimilerp PDG 2 67 100.0 381.0- 000.0 790.0 300.0 722.0 200.0 131.0lanfi PDG 3 67 310.0 768.0 910.0 768.0 500.0 882.0 400.0 122.0 ecnavda rotafled ecirp PDG 4 67 160.0 **545.1 610.0 156.0 410.0 504.0 000.0 650.0yranimilerp rotafled ecirp PDG 5 67 400.0 243.0- 000.0 280.0 820.0 076.0 220.0 794.0 lanfi rotafled ecirp PDG 6 822 200.0 443.0- 710.0 **550.1- 020.0 **957.0- 230.0 ***577.0etar tnemyolpmenU 7 822 012.0 ***786.3 172.0 ***812.4 552.0 ***437.2 381.0 ***548.1 tnemyolpme lloryap mrafnoN 8 822 311.0 ***433.2 521.0 ***580.2 790.0 ***231.1 860.0 ***757.0 selas liateR 9 722 480.0 ***220.2 380.0 ***107.1 260.0 ***609.0 330.0 ***035.0 selibomotua ssel selas liateR 01 822 210.0 *256.0 140.0 ***132.1 600.0 004.0 000.0 510.0noitcudorp lairtsudnI 11 822 430.0 ***280.1 960.0 ***606.1 220.0 **757.0 200.0 991.0 noitazilitu yticapaC 21 822 000.0 011.0- 100.0 281.0- 100.0 901.0- 000.0 520.0 emocni lanosreP 31 822 300.0 883.0- 300.0 553.0- 000.0 280.0- 000.0 450.0 tiderc remusnoC 41 822 400.0 183.0 400.0 453.0 610.0 *135.0 500.0 782.0 serutidnepxe noitpmusnoc lanosreP 51 722 910.0 **787.0 600.0 414.0 500.0 782.0 900.0 353.0 selas emoh weN 61 722 110.0 336.0 600.0 064.0 800.0 483.0 800.0 533.0 sredro sdoog elbaruD 71 722 300.0 123.0 700.0 064.0 000.0 460.0 300.0 662.0 gnidneps noitcurtsnoC 81 722 300.0 313.0 100.0 071.0 100.0 721.0 200.0 361.0 sredro yrotcaF 91 822 100.0 891.0 000.0 610.0 000.0 840.0 000.0 110.0seirotnevni ssenisuB 02 822 800.0 684.0- 410.0 *185.0- 410.0 *334.0- 510.0 *454.0ticfied tegdub tnemnrevoG 12 822 610.0 *707.0 200.0 302.0 000.0 240.0- 100.0 490.0ecnalab edarT 22 822 250.0 ***048.1 420.0 **062.1 710.0 **417.0 600.0 533.0 sgninrae ylruoh egarevA 32 822 900.0 555.0 200.0 842.0 300.0 981.0 400.0 312.0 IPP 42 822 030.0 ***499.0 900.0 435.0 510.0 *164.0 600.0 842.0 IPP eroC 52 822 600.0 345.0 410.0 *796.0 210.0 825.0 400.0 613.0 IPC 62 822 810.0 **629.0 730.0 ***31.1 520.0 **377.0 010.0 084.0 IPC eroC 72 002 830.0 ***643.1 420.0 **510.1 810.0 *976.0 610.0 *775.0 yranimilerp ecnedfinoc remusnoc MU 82 722 080.0 ***487.1 370.0 ***575.1 230.0 ***967.0 300.0 552.0 xedni deF aihpledalihP 92 002 100.0 802.0 000.0 050.0 000.0 110.0- 100.0 080.0lanfi ecnedfinoc remusnoc MU 03 822 720.0 **300.1 420.0 **678.0 330.0 ***627.0 930.0 ***666.0 xedni ecnedfinoc remusnoc BC 13 622 290.0 ***228.1 280.0 ***546.1 260.0 ***330.1 530.0 ***407.0 IMP ogacihC 23 822 461.0 ***638.2 471.0 ***665.2 131.0 ***966.1 250.0 ***280.1 IMP MSI 33 622 100.0 102.0 100.0 251.0 100.0 211.0 000.0 680.0 strats gnisuoH 43 822 700.0 205.0 000.0 570.0 400.0 072.0 220.0 **757.0 xedni cimonoce gnidael BC 53 299 720.0 ***630.1- 930.0 ***251.1- 30.0 ***367. 0- 910.0 ***855.0smialc sselboj laitinI 63 tnemecnuonna cimonoceorcam dezidradnats no segnahc dleiy dnob yliad fo snoisserger yduts tneve laudividni fo stluser eht stroper elbat ehT :etoN dna ,5 ,1 a tneserper * dna ,** ,*** dna ,desu era srorre dradnats etihW .5102 rebmeceD ot 7991 yraunaJ morf doirep eht srevoc elpmas ehT .sesirprus .ylevitcepser ,ecnacfiingis fo level %01

3 elbaT .stnemecnuonnA cimonoceorcaM fo scitsiretcarahC gnitsacwoN noisiveR gnitropeR htiw noitalerroC RTFF )vda( rotafleD ecirP PDG )vda( PDG edutingaM gaL RTFF .feD PDG R T F I R T F I R T F I tnemecnuonnA n )71( )61( )51( )41( )31( )21( )11( )01( )9( )8( )7( )6( )5( )4( )3( )2( )1( ytivitcA laeR 62.1 92 22.0 00.0 00.1 00.0 24.1- 79.3- 93.5- 00.0 60.1- 92.5- 63.6- 82.0- 24.1- 65.1- 52.3ecnavdaPDG 1 13.1 85 12.0 20.0 69.0 00.0 35.1- 79.3- 05.5- 00.0 81.1- 92.5- 74.6- 00.0 35.1- 48.1- 73.3yranimilerpPDG 2 03.1 78 41.0 20.0 49.0 10.0 55.1- 89.3- 25.5- 12.0 81.1- 25.5- 94.6- 13.0 45.1- 51.2- 73.3lanfiPDG 3 secirP 12.1 92 20.0 00.1 00.0 20.0 58.0- 73.5- 91.6- 66.0 12.1- 49.3- 94.4- 00.0 75.1- 69.3- 25.5ecnavdarotafledecirpPDG 4 42.1 85 10.0 89.0 20.0 30.0 08.0- 83.5- 61.6- 00.0 12.1- 03.3- 05.4- 40.0 16.1- 00.4- 65.5yranimilerprotafledecirpPDG 5 11.1 78 31.0 79.0 10.0 20.0- 28.0- 33.5- 71.6- 81.0 22.1- 74.3- 15.4- 60.0- 26.1- 98.3- 85.5lanfirotafledecirpPDG 6 ytivitcA laeR 18.0 5 60.0 52.0 02.0 20.0- 91.1- 79.3- 81.5- 00.0 91.0- 95.4- 97.4- 50.0- 96.1- 30.2- 77.3etartnemyolpmenU 7 41.1 5 13.0 11.0 06.0 50.0 51.1- 20.4- 21.5- 71.0- 08.0- 32.5- 91.6- 50.0 41.1- 59.1- 40.3tnemyolpmelloryapmrafnoN 8 62.1 31 91.0 10.0 94.0 21.0 18.1- 64.4- 51.6- 23.0 50.1- 47.5- 74.6- 51.0 76.1- 92.2- 18.3selasliateR 9 22.1 31 82.0 72.0 25.0 42.0 96.1- 24.4- 78.5- 03.0 01.1- 67.5- 65.6- 43.0 25.1- 83.2- 65.3selibomotuasselselasliateR 01 82.1 61 72.0 70.0 66.0 41.0 28.1- 63.4- 50.6- 50.0- 55.1- 84.5- 70.7- 20.0 57.1- 41.2- 78.3noitcudorplairtsudnI 11 08.1 61 20.0 72.0 03.0 70.0- 16.1- 60.4- 47.5- 40.0- 09.0- 30.5- 69.5- 21.0- 19.1- 70.2- 01.4noitazilituyticapaC 21 18.0 92 10.0 90.0 42.0 20.0 81.2- 68.4- 20.7- 20.0 33.1- 93.6- 96.7- 10.0 12.2- 87.2- 89.4emocnilanosreP 31 32.1 73 00.0 30.0 33.0 10.0 20.2- 04.4- 14.6- 40.0 81.1- 79.5- 11.7- 10.0 71.2- 53.2- 05.4tidercremusnoC 41 noitpmusnoC 32.1 92 41.0 61.0 84.0 50.0 61.2- 75.4- 76.6- 52.0- 64.1- 74.5- 81.7- 00.0 20.2- 63.2- 93.4serutidnepxenoitpmusnoclanosreP 51 21.1 72 31.0 11.0 11.0 80.0- 34.2- 31.6- 36.8- 92.0- 97.1- 74.6- 45.8- 70.0- 52.2- 77.3- 90.6selasemohweN 61 tnemtsevnI 60.1 62 01.0 51.0 71.0 60.0 81.2- 47.4- 78.6- 11.0 36.1- 87.5- 92.7- 70.0 40.2- 95.2- 65.4sredrosdoogelbaruD 71 42.1 23 51.0 30.0 24.0 01.0- 62.2- 27.4- 80.7- 12.0- 47.1- 48.5- 97.7- 51.0- 02.2- 35.2- 78.4gnidnepsnoitcurtsnoC 81 90.1 43 41.0 30.0 53.0 91.0 31.2- 66.4- 06.6- 81.0 46.1- 86.5- 51.7- 22.0 60.2- 55.2- 83.4sredroyrotcaF 91 32.1 44 31.0 33.0 44.0 28.0- 59.1- 93.4- 71.7- 15.0- 23.1- 37.5- 65.7- 39.0- 21.2- 84.2- 45.5seirotnevnissenisuB 02 sesahcruP tnemnrevoG 71.0 51 10.0 71.0 03.0 51.0 88.1- 88.6- 16.8- 30.0- 99.0- 40.7- 50.8- 01.0 20.2- 98.4- 28.6ticfiedtegdubtnemnrevoG 12 stropxE teN 31.1 34 30.0 34.0 01.0 52.1- 33.2- 17.5- 92.9- 74.0- 27.1- 61.7- 63.9- 89.0- 34.2- 36.3- 30.7ecnalabedarT 22 secirP 98.0 5 21.0 12.0 40.0 22.0 76.1- 14.6- 78.7- 51.1- 53.0- 06.4- 01.6- 33.0 10.2- 31.4- 08.5sgninraeylruohegarevA 32 89.0 51 51.0 83.0 82.0 02.0 14.0- 66.4- 88.4- 10.0- 92.0- 62.3- 65.3- 20.0- 48.0- 34.3- 92.4- IPP 42 29.0 51 10.0 04.0 11.0 47.0 55.0- 00.6- 08.5- 01.0 57.0- 26.3- 62.4- 42.0 35.1- 43.4- 36.5- IPPeroC 52 83.1 71 91.0 55.0 43.0 51.0 55.0- 16.4- 10.5- 90.0 25.0- 42.3- 76.3- 00.0 91.1- 13.3- 15.4- IPC 62 72.1 71 90.0 33.0 30.0 14.0 48.0- 19.5- 33.6- 80.0 80.1- 24.3- 34.4- 02.0- 46.1- 58.3- 96.5- IPCeroC 72 gnikooL drawroF 02.1 71- 12.0 50.0 84.0 00.0 42.0- 36.3- 78.3- 00.0 72.0- 16.4- 78.4- 00.0 72.0- 86.1- 59.1yranimilerpecnedfinocremusnocMU 82 32.1 21- 55.0 00.0 16.0 20.0 35.0- 99.3- 05.4- 20.0 14.0- 07.4- 90.5- 20.0- 83.0- 18.1- 12.2xednideFaihpledalihP 92 11.0 3- 91.0 40.0 94.0 00.0 68.0- 36.3- 05.4- 00.0 38.0- 06.4- 34.5- 00.0 30.1- 86.1- 17.2lanfiecnedfinocremusnocMU 03 88.0 3- 11.0 31.0 34.0 10.0 08.0- 50.4- 48.4- 20.0- 12.0- 07.4- 39.4- 00.0 82.1- 31.2- 14.3xedniecnedfinocremusnocBC 13 92.1 1- 24.0 31.0 35.0 50.0 51.1- 01.4- 02.5- 40.0- 89.0- 97.4- 18.5- 10.0- 99.0- 39.1- 39.2- IMPogacihC 23 02.1 2 05.0 01.0 16.0 20.0 42.1- 29.3- 41.5- 80.0 89.0- 57.4- 56.5- 00.0 20.1- 97.1- 08.2- IMPMSI 33 30.1 81 11.0 50.0 62.0 71.0 92.2- 25.5- 36.7- 80.0 95.1- 31.6- 46.7- 90.0 11.2- 01.3- 21.5stratsgnisuoH 43 98.0 32 41.0 34.0 81.0 87.0- 02.2- 91.4- 71.7- 44.0- 97.1- 60.5- 92.7- 07.0- 21.2- 98.1- 17.4xednicimonocegnidaelBC 53 10.1 6 81.0 71.0 83.0 11.0- 60.0- 13.4- 74.4- 20.0 01.0 26.5- 05.5- 50.0- 50.0- 53.2- 54.2smialcsselbojlaitinI 63 eht ,)F( slatnemadnuf ot noitaler eht :stnenopmoc sti dna )I( eulav cisnirtni s’elbairav eht fo egareva seires-emit eht syalpsid elbat eht ,elbairav cimonoceorcam hcae roF :etoN debircsedsadetupmoceradna,RTFFehtro,rotafledecirpPDGeht,PDGgnitsacwonnodesaberascitsiretcarahcesehT.)R(muimerpnoisiverehtdna,)T(muimerpssenilemit ehtyleman,stnenopmoceerhtehtfoserusaemevitanretlatroper)71(-)31(snmuloC .sngisevitagenehtecneh,smhtiragollarutanera)21(-)1(snmulocnisrebmunehT.4noitceSni .5102rebmeceDot7991yraunaJmorfsielpmasatadehT.4noitceSnidenfiedsa,edutingamnoisiverehtdna,galgnitropereht,tegratgnitsacwonehthtiwnoitalerroc

4 elbaT .stluseR lennahC PDG yrusaerT raeY-1 yrusaerT htnoM-6 )6( )5( )4( )3( )2( )1( )6( )5( )4( )3( )2( )1( notneicffieoC ***425.1 ***845.0 ***158.0 ***082.1 ***874.1 ***935.0 ***711.1 ***973.0 ***575.0 ***399.0 ***220.1 ***973.0 S )602.0( )750.0( )501.0( )102.0( )981.0( )060.0( )012.0( )460.0( )690.0( )981.0( )391.0( )750.0( ***513.0 ***612.0 eulaVcisnirtnI×S )450.0( )950.0( ***542.0 ***552.0 ***002.0 ***112.0 slatnemadnuFotnoitaleR×S )450.0( )560.0( )150.0( )350.0( ***851.0 ***191.0 **8290.0 **021.0 muimerPssenilemiT×S )650.0( )550.0( )440.0( )050.0( *011.0 080.0 220.0 300.0muimerPnoisiveR×S )560.0( )940.0( )640.0( )550.0( ***013.0- ***613.0- ***313.0- ***613.0- ***013.0- ***813.0- ***763.0- ***173.0- ***863.0- ***963.0- ***663.0- ***173.0tnatsnoC )940.0( )050.0( )160.0( )950.0( )840.0( )050.0( )150.0( )040.0( )440.0( )830.0( )540.0( )050.0( 020.0 510.0 710.0 810.0 020.0 510.0 210.0 900.0 010.0 210.0 210.0 900.0 2R yrusaerT raeY-5 yrusaerT raey-2 )6( )5( )4( )3( )2( )1( )6( )5( )4( )3( )2( )1( notneicffieoC ***103.2 ***899.0 ***604.1 ***529.1 ***692.2 ***579.0 ***872.2 ***368.0 ***892.1 ***398.1 ***942.2 ***148.0 S )652.0( )890.0( )951.0( )572.0( )032.0( )101.0( )962.0( )480.0( )131.0( )332.0( )342.0( )560.0( ***344.0 ***274.0 eulaVcisnirtnI×S )470.0( )470.0( ***223.0 ***623.0 ***453.0 ***163.0 slatnemadnuFotnoitaleR×S )770.0( )680.0( )370.0( )370.0( ***022.0 ***462.0 ***132.0 ***972.0 muimerPssenilemiT×S )170.0( )970.0( )080.0( )460.0( ***632.0 **691.0 ***622.0 ***381.0 muimerPnoisiveR×S )480.0( )090.0( )960.0( )760.0( ***691.0- **302.0- ***102.0- ***502.0- **691.0- ***702.0- ***072.0- ***872.0- ***672.0- ***972.0- ***172.0- ***282.0tnatsnoC )760.0( )080.0( )660.0( )270.0( )870.0( )670.0( )570.0( )570.0( )870.0( )650.0( )270.0( )070.0( 520.0 120.0 220.0 320.0 520.0 020.0 620.0 120.0 220.0 320.0 620.0 020.0 2R tnemecnuonna ruo htiw detcaretni sesirprus dna sesirprus orcam no ,)1 nmuloc( sesirprus orcam no segnahc dleiy dnob yliad gnisserger fo stluser syalpsid elbat ehT :etoN era scitsiretcarahC .)6 nmuloc( ecno ta lla scitsiretcarahc tnemecnuonna htiw detcaretni sesirprus dna sesirprus orcam no dna ,)5 dna ,4 ,3 ,2 snmuloc( yletarapes scitsiretcarahc deppartstooB .snoitavresbo 595,7 no desab si noisserger hcae dna ,5102 rebmeceD ot 7991 yraunaJ morf snur elpmas atad ehT .PDG rof esicrexe gnitsacwon eht morf devired .ylevitcepser,ecnacfiingisfolevel%01dna,5,1atneserper*dna,**,***dna,desuerasrorredradnats

5 elbaT .stluseR lennahC rotafleD ecirP PDG yrusaerT raeY-1 yrusaerT htnoM-6 )6( )5( )4( )3( )2( )1( )6( )5( )4( )3( )2( )1( notneicffieoC ***406.1 ***735.0 ***416.0 ***764.1 ***934.1 ***935.0 ***801.1 ***573.0 ***693.0 ***270.1 ***709.0 ***973.0 S )122.0( )250.0( )980.0( )202.0( )062.0( )150.0( )981.0( )840.0( )850.0( )122.0( )132.0( )250.0( ***622.0 **331.0 eulaVcisnirtnI×S )060.0( )450.0( ***522.0 ***512.0 ***361.0 ***061.0 slatnemadnuFotnoitaleR×S )150.0( )240.0( )930.0( )150.0( *011.0 880.0 330.0 020.0 muimerPssenilemiT×S )950.0( )360.0( )050.0( )340.0( 4100.0 410.0- 920.0- 430.0muimerPnoisiveR×S )850.0( )060.0( )640.0( )240.0( ***513.0- ***813.0- ***513.0- ***813.0- ***413.0- ***813.0- ***073.0- ***173.0- ***073.0- ***173.0- ***963.0- ***173.0tnatsnoC )640.0( )550.0( )940.0( )850.0( )050.0( )450.0( )440.0( )640.0( )450.0( )540.0( )640.0( )740.0( 810.0 510.0 510.0 810.0 810.0 510.0 110.0 900.0 900.0 010.0 010.0 900.0 2R yrusaerT raeY-5 yrusaerT raey-2 )6( )5( )4( )3( )2( )1( )6( )5( )4( )3( )2( )1( notneicffieoC ***689.1 ***769.0 ***741.1 ***427.1 ***279.1 ***579.0 ***831.2 ***838.0 ***389.0 ***209.1 ***710.2 ***148.0 S )493.0( )960.0( )911.0( )714.0( )553.0( )890.0( )413.0( )180.0( )501.0( )803.0( )823.0( )180.0( ***052.0 ***592.0 eulaVcisnirtnI×S )180.0( )080.0( **391.0 *371.0 ***362.0 ***642.0 slatnemadnuFotnoitaleR×S )580.0( )490.0( )660.0( )960.0( ***512.0 **102.0 ***091.0 ***661.0 muimerPssenilemiT×S )770.0( )580.0( )460.0( )850.0( 930.0- 560.0- 900.0- 330.0muimerPnoisiveR×S )770.0( )680.0( )770.0( )280.0( ***102.0- ***702.0- **102.0- **702.0- **302.0- ***702.0- ***772.0- ***282.0- ***772.0- ***282.0- ***772.0- ***282.0tnatsnoC )770.0( )270.0( )680.0( )680.0( )180.0( )360.0( )080.0( )260.0( )860.0( )270.0( )360.0( )570.0( 220.0 120.0 120.0 120.0 220.0 020.0 220.0 020.0 020.0 120.0 220.0 020.0 2R tnemecnuonna ruo htiw detcaretni sesirprus dna sesirprus orcam no ,)1 nmuloc( sesirprus orcam no segnahc dleiy dnob yliad gnisserger fo stluser syalpsid elbat ehT :etoN era scitsiretcarahC .)6 nmuloc( ecno ta lla scitsiretcarahc tnemecnuonna htiw detcaretni sesirprus dna sesirprus orcam no dna ,)5 dna ,4 ,3 ,2 snmuloc( yletarapes scitsiretcarahc .snoitavresbo 595,7 no desab si noisserger hcae dna ,5102 rebmeceD ot 7991 yraunaJ morf snur elpmas atad ehT .rotafled ecirp PDG rof esicrexe gnitsacwon eht morf devired .ylevitcepser,ecnacfiingisfolevel%01dna,5,1atneserper*dna,**,***dna,desuerasrorredradnatsdeppartstooB

6 elbaT .stluseR lennahC RTFF yrusaerT raeY-1 yrusaerT htnoM-6 )6( )5( )4( )3( )2( )1( )6( )5( )4( )3( )2( )1( notneicffieoC ***074.1 ***045.0 ***408.0 ***433.1 ***494.1 ***935.0 ***390.1 ***873.0 ***055.0 ***930.1 ***620.1 ***973.0 S )433.0( )3550.0( )311.0( )512.0( )142.0( )660.0( )592.0( )360.0( )301.0( )962.0( )022.0( )160.0( ***232.0 ***751.0 eulaVcisnirtnI×S )350.0( )150.0( **441.0 ***361.0 **711.0 ***531.0 slatnemadnuFotnoitaleR×S )160.0( )640.0( )950.0( )250.0( ***361.0 ***981.0 *001.0 *321.0 muimerPssenilemiT×S )450.0( )060.0( )060.0( )560.0( 950.0 820.0 62700.0 810.0muimerPnoisiveR×S )350.0( )550.0( )050.0( )650.0( ***313.0- ***713.0- ***513.0- ***613.0- ***313.0- ***813.0- ***863.0- ***173.0- ***963.0- ***073.0- ***863.0- ***173.0tnatsnoC )750.0( )260.0( )050.0( )740.0( )350.0( )050.0( )840.0( )840.0( )440.0( )150.0( )150.0( )050.0( 810.0 510.0 710.0 610.0 810.0 510.0 010.0 900.0 010.0 010.0 010.0 900.0 2R yrusaerT raeY-5 yrusaerT raey-2 )6( )5( )4( )3( )2( )1( )6( )5( )4( )3( )2( )1( notneicffieoC ***602.2 ***089.0 ***222.1 ***879.1 ***721.2 ***579.0 ***891.2 ***548.0 ***851.1 ***679.1 ***151.2 ***148.0 S )073.0( )490.0( )801.0( )873.0( )123.0( )080.0( )583.0( )080.0( )821.0( )543.0( )113.0( )060.0( ***082.0 ***913.0 eulaVcisnirtnI×S )770.0( )170.0( ***112.0 ***602.0 ***422.0 ***332.0 slatnemadnuFotnoitaleR×S )370.0( )870.0( )570.0( )960.0( 931.0 **771.0 **581.0 ***622.0 muimerPssenilemiT×S )490.0( )380.0( )370.0( )370.0( *951.0 511.0 331.0 580.0 muimerPnoisiveR×S )380.0( )990.0( )280.0( )270.0( **102.0- ***502.0- **502.0- **502.0- **202.0- ***702.0- ***572.0- ***182.0- ***972.0- ***082.0- ***672.0- ***282.0tnatsnoC )480.0( )560.0( )380.0( )580.0( )780.0( )470.0( )660.0( )970.0( )260.0( )770.0( )560.0( )560.0( 220.0 120.0 120.0 120.0 220.0 020.0 320.0 020.0 120.0 120.0 320.0 020.0 2R tnemecnuonna ruo htiw detcaretni sesirprus dna sesirprus orcam no ,)1 nmuloc( sesirprus orcam no segnahc dleiy dnob yliad gnisserger fo stluser syalpsid elbat ehT :etoN scitsiretcarahC .)6 nmuloc( ecno ta lla scitsiretcarahc tnemecnuonna htiw detcaretni sesirprus dna sesirprus orcam no dna ,)5 dna ,4 ,3 ,2 snmuloc( yletarapes scitsiretcarahc 595,7 no desab si noisserger hcae dna ,5102 rebmeceD ot 7991 yraunaJ morf snur elpmas atad ehT .etaR tegraT sdnuF laredeF eht rof esicrexe gnitsacwon eht morf devired era .ylevitcepser,ecnacfiingisfolevel%01dna,5,1atneserper*dna,**,***dna,desuerasrorredradnatsdeppartstooB .snoitavresbo

Table 7 Price Impact and Intrinsic Value. 6-Month Treasury 1-Year Treasury Nowcasting Target Nowcasting Target GDP GDP GDP Deflator FFTR GDP Deflator FFTR Coefficient on (1) (2) (3) Coefficient on (1) (2) (3) Intrinsic Value 0.0123** 0.0024 0.0099* Intrinsic Value 0.0196** 0.0060 0.0163** (0.0060) (0.0027) (0.0051) (0.0088) (0.0040) (0.0075) Constant 0.0570** 0.0281** 0.0602** Constant 0.0884** 0.0516*** 0.0961** (0.0235) (0.0124) (0.0263) (0.0341) (0.0185) (0.0384) R2 0.117 0.004 0.067 R2 0.139 0.013 0.086 2-Year Treasury 5-Year Treasury Nowcasting Target Nowcasting Target GDP GDP GDP Deflator FFTR GDP Deflator FFTR Coefficient on (1) (2) (3) Coefficient on (1) (2) (3) Intrinsic Value 0.0266** 0.0064 0.0209** Intrinsic Value 0.0248** 0.0070 0.0191** (0.0099) (0.0055) (0.0086) (0.0094) (0.0052) (0.0083) Constant 0.1180*** 0.0605** 0.1230*** Constant 0.1120*** 0.0625** 0.1150** (0.0383) (0.0250) (0.0439) (0.0362) (0.0244) (0.0421) R2 0.187 0.011 0.103 R2 0.178 0.014 0.094 Note: The table displays results of regressing the estimated R2 coefficients in equation (2) on the announcement’sintrinsicvaluederivedfromnowcastingGDPadvance,theGDPpricedeflatoradvance,and the Federal Funds Target Rate. The sample covers the period from January 1997 to December 2015, and each regression is based on 36 observations. White standard errors are used, and ***, **, and * represent a 1, 5, and 10% level of significance, respectively.

Table 8 Price Impact and Macroeconomic Announcement Characteristics. 6-Month Treasury NowcastMeasuresof AlternativeMeasures Intrinsic Relationto Timeliness Revision Correlation Reporting Revision Value Fundamentals Premium Premium withGDP Lag Magnitude (1) (2) (3) (4) (5) (6) (7) Coefficient 0.0123** 0.0100** 0.0122 0.00453 0.0363* -0.00626* 0.00383 (0.0060) (0.0048) (0.0073) (0.0044) (0.0206) (0.0036) (0.0085) Constant 0.0570** 0.0483** 0.0408** 0.0191*** 0.00510 0.0265*** 0.0145 (0.0235) (0.0188) (0.0179) (0.0059) (0.0043) (0.0092) (0.0089) R2 0.117 0.089 0.060 0.008 0.083 0.059 0.001 1-Year Treasury NowcastMeasuresof AlternativeMeasures Intrinsic Relationto Timeliness Revision Correlation Reporting Revision Value Fundamentals Premium Premium withGDP Lag Magnitude (1) (2) (3) (4) (5) (6) (7) Coefficient 0.0196** 0.0135* 0.0227** 0.0120** 0.0500 -0.0111** 0.0206** (0.0088) (0.0070) (0.0111) (0.0055) (0.0312) (0.0054) (0.0098) Constant 0.0884** 0.0673** 0.0686** 0.0287*** 0.00879 0.0413*** 0.00484 (0.0341) (0.0276) (0.0268) (0.0085) (0.0065) (0.0134) (0.0088) R2 0.139 0.076 0.098 0.026 0.074 0.087 0.016 2-Year Treasury NowcastMeasuresof AlternativeMeasures Intrinsic Relationto Timeliness Revision Correlation Reporting Revision Value Fundamentals Premium Premium withGDP Lag Magnitude (1) (2) (3) (4) (5) (6) (7) Coefficient 0.0266** 0.0195** 0.0293** 0.0138* 0.0703* -0.0168*** 0.0427*** (0.0099) (0.0080) (0.0128) (0.0070) (0.0376) (0.0058) (0.0122) Constant 0.118*** 0.0922*** 0.0880*** 0.0363*** 0.00863 0.0558*** -0.0120 (0.0383) (0.0315) (0.0306) (0.0099) (0.0087) (0.0151) (0.0112) R2 0.187 0.115 0.120 0.025 0.107 0.146 0.052 5-Year Treasury NowcastMeasuresof AlternativeMeasures Intrinsic Relationto Timeliness Revision Correlation Reporting Revision Value Fundamentals Premium Premium withGDP Lag Magnitude (1) (2) (3) (4) (5) (6) (7) Coefficient 0.0248** 0.0162** 0.0302** 0.0159** 0.0547 -0.0168*** 0.0388*** (0.0094) (0.0077) (0.0126) (0.0063) (0.0361) (0.0058) (0.0115) Constant 0.1120*** 0.0824** 0.0894*** 0.0362*** 0.0142 0.0556*** -0.00797 (0.0362) (0.0305) (0.0297) (0.0095) (0.0084) (0.0144) (0.0097) R2 0.178 0.088 0.139 0.036 0.071 0.162 0.047 Note: ThetabledisplaysresultsofregressionsoftheR2 fromequation(2)inTable2onthemacroeconomicannouncement’s intrinsicvalueanditscomponents(relationtofundamentals,timelinesspremium,andrevisionpremium)derivedfromnowcastingGDPadvance. Thetablealsodisplaystheresultsofsimilarregressionsusingalternativemeasuresforthethreecomponents, namelycorrelationwithGDP,reportinglag,andrevisionmagnitude. ThedatasampleisfromJanuary1997toDecember2015, andeachregressionisbasedon36observations. Whitestandarderrorsareused,and***,**,and*representa1,5,and10% levelofsignificance,respectively.

Online Appendix to: Is the Intrinsic Value of Macroeconomic News Announcements Related to their Asset Price Impact? Thomas Gilbert, Chiara Scotti, Georg Strasser, Clara Vega July 2016 In this online appendix, we present supplementary material on our analysis of the link betweentheintrinsicvalueandthepriceimpactofmacroeconomicannouncements. First, we present the details of the noisy rational expectations model of the price response to expected public announcements. Second, we provide details on the nowcasting procedure and data management. Third, we present the exact transformations of the macroeconomic variables we use in our tests. Fourth, we present results of our analysis without the Federal Reserve’s zero lower bound period. 1

O.1. A Noisy Rational Expectations Model In this appendix, we provide the details of the noisy rational expectations model we use to motivate and frame the relationship between an announcement’s price impact and its intrinsic value, timeliness, revisions, and relation to fundamentals. For more details on this class of models, we refer the reader to, among many others, Grundy and McNichols (1989), Kim and Verrecchia (1991a,b), Kandel and Pearson (1995), Veronesi (2000), Hautsch and Hess (2007), and Hess and Niessen (2010). O.1.1. Model Setup We consider a discrete-time and finite-horizon model where a representative investor trades a claim on future consumption. The terminal payoff of this traded asset is a random variable, which depends on the underlying state of the economy. Every period, the investor updates her belief about the asset’s payoff as she receives public (macroeconomic) information and trades accordingly. This setup maps into our empirical analysis by thinking of the traded asset as U.S. Treasury bonds and by viewing the timeline as one specific reference period in actual data, i.e., for a reference period p, the investor receives a sequential set of macroeconomic signals, trades as the information is received, and the final payoff is realized at the end of the calendar of announcements referring to that period. Before observing any information at time t = 0, the investor assumes that the asset’s terminal payoff X is normally distributed with mean µ and precision (inverse of variance) 0 ρ . At each release time t, the investor observes a noisy signal an of X, where the subscript n 0 t indicates the announcement type (e.g., nonfarm payroll, industrial production).1 This signal is equal to the asset payoff plus noise, an = X +εn, where εn is normally distributed with t t t zero mean and precision ρ . an t The representative investor maximizes her expected final consumption (wealth W) based on negative exponential utility with constant absolute risk aversion γ: E [U(W)] = E [−e−γDX], (E-1) t t where D is the (optimal) amount of the traded asset held in that period. For simplicity, we assume that γ = 1 and abstract away from private information and heterogeneous prior beliefs. Thelatterisrequiredtogeneratetradingvolume(KimandVerrecchia, 1991b). Thus prices move without any trading in our model. 1In the empirical analysis, announcements have the reference period p as additional subscript. Because the model in this appendix studies the information updating for a specific reference period, we drop the p subscript here. 2

Because all payoffs and signals are normally distributed and i.i.d., the first-order condition consistent with the above negative exponential utility function is standard, and the investor’s demand D for the traded asset at time t is a linear function of the asset’s price t p at that time:2 t E [X]−p D = t t (E-2) t Var [X] t At each information release t, the rational investor estimates the conditional mean and variance of the asset’s payoff based on all available information (current and past signals). Since all signals are public, there is nothing additional to be learned from the price; hence the agent needs to condition only on the signals themselves. Using Bayes’ rule, the asset’s conditional expected payoff after information release t is (cid:32) (cid:33) t (cid:88) E [X] ≡ µ = ρ−1 ρ µ + ρ an , (E-3) t t t 0 0 an i i i=1 where ρ = ρ + (cid:80)t ρ is the conditional precision of the investor’s posterior at this time. t 0 i=1 an i When updating her belief about the state of the economy, the investor places a weight of ρ an t on signal an. ρt t The negative exponential utility function implies linear demand functions. Imposing the market-clearing condition that demand must equal an exogenous supply of the (normally distributed) traded asset, it is straightforward to show that at each time prices equal the conditional expected payoffs, i.e. that p = E[X] and p = E [X]. Thus the price changes 0 t t around macroeconomic announcements according to ρ p −p = an t (an −µ ). (E-4) t t−1 ρ t t−1 t The price change around the public release of information is therefore equal to a constant times the announcement’s surprise. O.1.2. Intrinsic Value and Price Impact The previous literature refers to the constant ρ an t in equation (E-4) as the price impact of ρt announcement an. It is the weight that the representative investor places on that announcet ment when updating her belief about the state of the economy, which we therefore refer to as the intrinsic value of the announcement. Because prices in this stylized model react only to information, the price impact and the intrinsic value are equal. Empirically, this is not 2The investor’s coefficient of absolute risk aversion enters the demand function in the denominator, i.e., ceteris paribus, the higher the risk aversion, the lower the demand for the risky asset. Higher risk aversion dampens the equilibrium asset price response. 3

the case. Whether an announcement matters for forecasting a variable of interest and how it affects asset prices depends on how the market reacts to this underlying information. O.1.3. Timeliness To capture the effect of the announcement’s timeliness on the intrinsic value, we consider announcements which are released at different times but are equally precise, i.e., they have the same precision ρ ≡ ρ . We can re-write equation (E-4) as an a t ρ p −p = a (an −µ ). (E-5) t t−1 ρ +tρ t t−1 0 a Clearly, ρa decreases in t. Therefore an early surprise has a bigger price impact than an ρ0+tρa equally large surprise later on. O.1.4. Revisions To capture the effect of the announcement’s revisions on intrinsic value and price impact, we now consider the case of multiple announcements being released at the same time. For simplicity, suppose that M announcements are released simultaneously at time t = 1 and that these announcements differ in their precision ρ , where n = 1,...,M. We therefore an 1 have (cid:80)M ρ (an −µ ) p −p = n=1 an 1 1 0 . (E-6) 1 0 ρ + (cid:80)M ρ 0 n=1 an 1 The weight on the ith announcement released at time t = 1 is therefore ρ ai 1 , which ρ0+ (cid:80)M n=1 ρ an 1 increases in the announcement’s precision ρ . Among announcements released at the same ai 1 time, the more precise announcement has a bigger price impact.3 Importantly, the precision of a noisy announcement combines two components, the announcement’s relation to fundamentals and its revisions. Indeed, macroeconomic announcements undergo revisions following their initial release (Croushore, 2011), but even the most carefully revised macroeconomic announcements are imperfect proxies for fundamentals because of measurement error. We can therefore decompose the precision of an announcement into these two components: (cid:16) (cid:17) ρ = ρ∞ − ρ∞ −ρ , (E-7) an t an t an t an t where ρ∞ is the announcement’s relation to fundamentals, i.e., the precision of the fully an t revised announcement, and ρ∞ − ρ is the revision noise. We assume that ρ increases an t an t an t monotonically in t with each revision, converging to ρ∞ < ∞ in the limit. By this definition, an t 3No additional intuition is gained from generalizing this to time t, but the equilibrium return is significantly more cumbersome. 4

the revision noise shrinks to zero over time, whereas the relation to the fundamental is a (finite) constant. For a set of announcements with the same relation to fundamentals, it follows from equations (E-6) and (E-7) that the announcement weight decreases with revision noise. Ceteris paribus, less revised announcements have a bigger price impact. O.1.5. Relation to Fundamentals Per equation (E-7), the relation to fundamentals captures the noise component that never goes away. It is the precision of the final revised value an . ∞ Forasetofannouncementswiththesamerevisionnoise, itfollowsfromequations(E-6) and (E-7) that the announcement weight increases in ρ∞. Ceteris paribus, announcements an t with a larger relation to fundamentals have a bigger price impact. 5

O.2. Methodology This appendix provides additional details on the definition of our three nowcasting targets, the nowcasting procedure, and the definition of nowcasting weights in actual and counterfactual settings. O.2.1. Nowcasting Target and Data Management Mirroring the monthly evolving state of the economy, the data matrix at time t captures the latest known value of each macroeconomic announcement in each reference month. Figure F-1 shows the data structure and its sequential filling. reference t=October 6th, 2015 t=October 20th, 2015 month p a1 a2 a3 a4 aN a1 a2 a3 a4 aN p,t p,t p,t p,t p,t p,t p,t p,t p,t p,t XXXX X XXXX X Jun 15 q q q q q q XX X X XX X X Jul 15 q q q q q q X X X XX X X Aug 15 > q q q q q q X X X Sep 15 q q q q q q X Oct 15 q q q q q q Nov 15 q q q q q q Fig. F-1. Data Structure We reestimate our model completely each time a new announcement is released. This iterative method requires rebuilding the dataset at every t, because past values might have been revised. Our raw dataset contains 36 macroeconomic announcement series and the Federal Funds Target Rate (FFTR). For each of these series we record the release times, the initially published values, the reference periods and the latest revised values. We transform these announcements as described in Appendix O.3. to ensure stationarity. We consolidate variables that are released by installments, namely GDP (advance, preliminary, final), GDP price deflator (advance, preliminary, final), and the University of Michigan (UM) consumer confidence index (preliminary, final) into one series, respectively. That is, we maintain only a single time series of GDP, GDP price deflator, and UM consumer confidence, and replace preliminary values in real time by revised ones as they become available. In terms of Figure F-1 this means that GDP appears as a single column, and 6

that earlier values (in boxes marked with “X”) are overwritten by later releases for the same reference period. This reduces the 37 raw announcement series to N = 32 consolidated series. Several of our macroeconomic series refer to periods different from a calendar month. These are variables that are released weekly, quarterly, or irregularly. We convert them to monthly frequency in the following way: The only weekly series in our dataset, initial jobless claims, is measured in headcounts, which we simply add up. If at time t claims are known for only a part of a given reference month, then we scale them up to the full month, assuming the unknown later part of the month will have the same headcounts as its known part, and revise these values as additional weeks become known. We fill quarterly values into all months of the respective quarter and apply mean-invariant smoothing for compounding growth rates to avoid jumps between quarters. The only irregular series is the FFTR. We specify the monthly FFTR vector to contain the FFTR on the 15th of each month at 23:59:59. We further assume that an FOMC announcement pins down the FFTR until the next scheduled FOMC meeting. We allow any FFTR entry to change again if there is another FOMC meeting before the next 15th of a month. If there are several meetings within a month, then only the FFTR of the last meeting before the 15th of each month at 23:59:59 will remain in the data matrix going forward. All other FFTR rates appear only temporarily, and are eventually overwritten by the value announced at that last meeting. The monthly FFTR change is accordingly the difference between its value on the 15th of the current month and its value on the 15th of the previous month. Our nowcasting target variables are GDP advance, the GDP price deflator advance, and the Federal Funds Target Rate (FFTR). In the case of quarterly nowcasting targets, i.e. GDP and GDP price deflator, we switch to the next forecasting quarter when their advance estimate is released. In the case of FFTR, we switch on the 15th of any given month to forecasting the next month, in line with our assignment of FFTR announcements to reference periods. This also implies that the change in the FFTR is the difference between its value on the 15th of the current month and its value on the 15th of the previous month. Our dataset covers the period from January 1990 until December 2015. We base our estimates of the intrinsic value on an expanding window beginning in January 1990. We start the nowcasting exercise with the window ending in January 1997. Our choice of the starting date has two reasons. First, we need initial observations to estimate the system matrices reliably. Second, for some series real-time data is not available for some or all of the years before 1997. When real-time data is not available, in particular for the Chicago Purchasing Manager Index and the Philadelphia Fed Index in the early years, we use instead final values during these years. 7

O.2.2. State Space Model As discussed in the main text, we work with the VAR(1) state equation Φ = B Φ +C ν , (E-8) p,t t p−1,t t p,t where ν ∼ WN(0,I ). The state of the economy evolves at a monthly frequency, indexed p,t 2×2 by reference period p. The subscript t governs how much information is available about the current and the past state vectors, and identifies specific times within the month. The announcement series end in different reference periods, which we denote by p¯n. At time t t the last reference period with the complete set of data available is p¯ = min (p¯n). t n t We use a 5-dimensional state vector, Φ , consisting of two common factors, one real p,t factor, onenominalfactor, andoneforward-lookingfactor. Thecommonfactorsarebasedon all N = 32 consolidated announcement series. The real factor is based on 19 announcement series: unemployment rate, durable goods orders, housing starts, trade balance, nonfarm payroll, advance retail sales, capacity utilization, industrial production, business inventories, construction spending, factory orders, new home sales, personal consumption, personal income, monthly budget statement, consumer credit, initial jobless claims, GDP, and retail saleslessautos. Thenominalfactorisbasedonsixannouncementseries: ConsumerPriceIndex, Producer Price Index, CPI ex food and energy, PPI ex food and energy, average hourly earnings, and GDP price deflator. The forward-looking factor is based on nine announcementseries: indexofleadingindicators, consumerconfidenceindex, ISMPMI,ChicagoPMI, Philadelphia Fed index, UM consumer confidence, durable goods orders, housing starts, and factory orders. The corresponding observation equation for a given information set t is A = D Φ +ε , (E-9) p,t t p,t p,t (cid:2) (cid:3)(cid:48) where ε ∼ WN(0,V ), and A = a1 ,...,aN is the monthly vector of N macroecop,t p,t p,t p,t p,t nomic variables containing only values announced on or before time t. O.2.3. Nowcasting Procedure Because past values are revised, the state space model (E-8) and (E-9) must be re-estimated at each data release. We use the two-step procedure of Giannone et al. (2008), because it permits forecasting the FFTR by an ordered probit specification.4 The estimation proceeds 4Such “partial” models, specifying the target variable separately from the model of the predictors, are widely used in policy institutions (Ban´bura et al., 2013). For our sample, this two-step procedure outperformed the one-step procedure in nowcasting GDP in terms of mean squared forecasting error. 8

in four steps, which we repeat for each announcement release time t. We use an expanding window from January 1990 until time t, starting with the window ending on t = January 1st, 1997. First, we consolidate variables that are released piece by piece (GDP, GDP price deflator, University of Michigan consumer confidence index) into one series, respectively, as described. For determining the intrinsic value later, we keep track of each observation’s original designation (advance, preliminary, or final). Given t, each time series is standardized to zero mean and unit standard deviation. Second, we define the five-dimensional state vector based on five principal components Φ extracted from the balanced part of the sample from January 1990 to p¯. Two principal p,t t components are based on all, three further principal components are based on only real, nominal, and forward-looking announcement series, respectively. The matrix C collects the t five eigenvectors, linking the factors Φ with the announcements A . The diagonal matrix p,t p,t V = diag(v1,v2,...,vN) contains the estimate of the idiosyncratic component, that is, the t t t t residual variance from projecting separately each an series on the factors Φ by ordinary ·,t ·,t least squares. We modify V to account for observations of A which are missing or which t p,t cover only a fraction of the month. Denoting the share of a given reference month covered by information about macroeconomic variable n in reference period p available at time t with χn we define p,t  vn/χn if an missing or incomplete, vn = t p,t p,t (E-10) p,t vn otherwise. t If, for example, the monthly observation an is missing, then vn = ∞, or, in the actual p,t p,t implementation, it is set to a very large number. For a weekly series χn is the share of days p,t of month p for which data has already been released by time t. These values are collected in the diagonal matrix V = diag(v1 ,v2 ,...,vN ). p,t p,t p,t p,t Third, the system matrices B and C of the VAR in equation (E-8) are estimated by t t ordinary least squares, the Kalman filter is initialized by the principal component estimates for the first period, and the initial variance is set equal to the unconditional variance of the common factors. For a given information set (indexed by t), the Kalman filter returns a sequence of Kalman gain matrices, K . Consider now a specific release time t. Because p,t the matrices B , C , D , and V are constant within the balanced part of the sample, t t t p,t K converges until the last reference period p¯ = min (p¯n) with the complete set of data p,t t n t available within that information set. For p > p¯ some announcements are missing, reflecting t the “ragged edge” problem (Wallis, 1986). In effect, V varies over time, and therefore K p,t p,t fluctuates for p > p¯. For each forecasting target and each information set t, the Kalman t 9

filter produces a Kalman gain matrix K for each reference month p. p,t   k11 ... kN1 p,t p,t  . .  K =  . . . . . (E-11) p,t   k15 ... kN5 p,t p,t In a balanced sample, the Kalman gain of interest would obviously be the gain in the very last period. Standard results show that the Kalman gain converges to a constant matrix as p becomes large. In our case, the most recent period with all announcements available is usually two months earlier, and more recent months contain only a subset of the announcements in varying composition. The composition does not follow a strict monthly or quarterly periodicity, because the sequence of announcements changes due to calendar effects specific to each month. It is further complicated by idiosyncratic events such as government shutdowns. The convergence result for Kalman gains does therefore not apply for this most recent period.5 To construct the time series of the intrinsic value of announcement n, only the gains at the time of a new release of macroeconomic variable n are used, i.e. in period p¯n.6 Therefore we can refer to the column n of K corresponding to this announcement t p¯n t ,t series (sampled in the periods with new releases of n) as kn. Here we keep the release times t of the advance, preliminary and final releases of GDP, GDP price deflator, and UM consumer confidence separate in order to assess their impact separately. Fourth, given the information at time t, we refine the in-sample estimates of the latent factors by Kalman smoothing, which improves the estimates of their past realizations by accounting for subsequently (but not after time t) revealed information. Then we use these ˜ smoothed factors Φ to fit a forecasting model for the nowcasting targets GDP and GDP p,t price deflator, Aj = D ˜jΦ ¯ +ε , (E-12) p,t t p,t p,t at quarterly frequency by ordinary least squares, where ε ∼ WN(0,v˜j). To account for the p,t t ¯ quarterly frequency, Φ contains the arithmetic average of the estimated monthly factors p,t 5Ban´bura and Ru¨nstler (2011) impose the ragged edge pattern at the end of the sample of the final data vintage (i.e. of their complete dataset) on the end of each subsample in the recursive estimation. This is justified if the rugged edge pattern does not vary with t. Unfortunately, this is not satisfied in U.S. macroeconomic announcement data, even if the day of the month was held fixed. In fact, several important macroeconomic announcements contribute to time variation in the ragged edge pattern. As a further complication, our approach requires us to reestimate the filter before every release, i.e. multiple times per month, and so the ragged edge varies by construction additionally within each month. 6In our setup with unbalanced data, the last converged Kalman gain (from the very last period before someannouncementsaremissing)isanex-postmeasureofgain. Instead,weusetheKalmangaininthemost recentmonthforwhichtherespectivevariablehasdata. ThereforeperiodforKalmangaincalculationdiffers between variables. Both Kalman gain vectors would be identical if a given variable was always announced last. 10

˜ Φ during the respective quarter. For the discrete nowcasting target FFTR we use the p,t ordered probit model AFFTR = A if α < A∗ ≤ α (E-13) p,t i i,t p,t i+1,t A∗ = D ˜ Φ ˜ +ε , (E-14) p,t t p,t p,t following Hamilton and Jorda` (2002) at monthly frequency. Here α are the cutoff points i,t which map the latent variable A∗ into FFTR steps and ε ∼ WN(0,v˜FFTR). To account p,t p,t t for the discreteness of the FFTR, we round FFTR changes to 0.25% and define as many ordered probit categories A as needed at any given time t. i For each forecasting target, indexed by j, we estimate coefficients D ˜j on the latent t factors at each point in time. In the discrete choice model for the FFTR we use the marginal effect instead. The absolute value of the product, w(j)n = |D ˜jkn|, of this coefficient (row) t t t vector with the respective column of the Kalman gain matrix is the weight on announcement n at time t for nowcasting the variable j. We take absolute values to capture the directionfree impact of an announcement. Repeating these four steps recursively at each announcement time t in our sample gives us a sequence of weights. Based on equations (E-12) to (E-14), one can forecast the factors (or states) out-ofsample for τ > t. The root mean squared forecasting error (RMSFE) of our nowcast of GDP is 1.4 during the period from 1997 to 2015, much lower than that of a random walk forecast of 2.1. The RMSFE for the GDP price deflator is 0.8, which is also lower than that of a random walk forecast with 1.3. For FFTR, the RMSFE is 0.18, which is also better than a random walk with 0.24. Nevertheless, obtaining an optimal nowcast is not a goal of this paper. It is for us just a means to evaluate the impact of announcement characteristics consistently. We assume that agents with rational expectations care about the best case scenario, i.e., the intrinsic value when the announcement is just released. These are (ex-post) weights on the standardized, transformed macroeconomic variables at announcement time.7 Our measure of intrinsic value is the (logarithm of) the weights. Because these weights are derived from actual data released according to the actual release schedule, we refer to them as w (j)n. A t 7TheseKalmangainvectorsare,ofcourse,columnsofKalmangainmatrices,butaretakenfrommatrices calculated for, in general, different reference periods p – the most recent period p for which the respective variable had data at time t. Note that we are interested in the most recent weight, not in the cumulative weight that the filter assigns to all past realizations of that announcement. 11

O.2.4. Counterfactual Announcement Times and Revision Status In order to measure the impact of an announcement while controlling for timing and noise, we create counterfactual datasets. These datasets differ from the original dataset in the release timing, the revision status, or both. We modify the respective property of only one macro announcement series n per nowcasting exercise. Tocontrolforreleasetiming,wecounterfactuallyreorderthedata. Todoso,weidentify theearliestannouncementforeachreferenceperiodandsetthecounterfactualannouncement time of the variable of interest to one second before this previously earliest announcement. The earliest announcements are typically initial jobless claims and UM consumer confidence preliminary. Applying the nowcasting procedure to the reordered dataset yields the weight series w (j)n. RA t To control for revision status, we counterfactually replace all releases of the variable of interest by final revision values. In cases where no final value is available, we keep the value of the initial release. Subjecting the original data to both this counterfactual replacement with final values and the counterfactual reordering, and feeding this into the nowcasting procedure yields the weight series w (j)n. RF t 12

O.3. Data Preparation The dataset covers the reference months from January 1990 until December 2015. The realtime series of the Chicago PMI begins with the release for reference period November 1996, and the real-time series of the Philadelphia Fed Index with the release for reference period January 1997. Furthermore, some real-time values up until January 1992 are missing for consumer confidence, initial jobless claims, CPI ex food and energy, PPI ex food and energy, and the GDP deflator (advance, preliminary, final). The 36 macroeconomic announcements listed in Table 1 in the main paper and the FFTR series, which are assumed here to jointly capture the state of the U.S. economy, are used in the nowcasting exercise, either in their original reporting units or transformed in order to approximate a linear relationship with the forecasting object. For indexes and variables reported in percent or percent changes, the original reporting unit is used, while variables reported in levels are transformed into percent changes. For example, the retail sales series, reported as a percent change, is not transformed, while the new home sales series is transformed from levels to percent change. We transform the raw data to ensure that all time-series available as of December 31, 2015 are stationary. More precisely, we transform the macroeconomic series, i.e., the dependent variable A in the observation equation (4), in order to approximate a linear p,t relationship with the forecasting object. Table T-1 summarizes the transformations. We do not modify published data by, for instance, removing or replacing outliers with fitted values. Instead, we treat them as features of the data that our estimates should capture. 13

Table T-1 Transformations of Macroeconomic Announcements. n Announcement Original Unit Transformation Real Activity 1 GDPadvance %change Original 2 GDPpreliminary %change Original 3 GDPfinal %change Original Prices 4 GDPpricedeflatoradvance %change Original 5 GDPpricedeflatorpreliminary %change Original 6 GDPpricedeflatorfinal %change Original Real Activity 7 Unemploymentreport % Original 8 Nonfarmpayrollemployment change Original/NFPPopulation 9 Retailsales %change Original 10 Retailsaleslessautomobiles %change Original 11 Industrialproduction %change Original 12 Capacityutilization % Original 13 Personalincome %change Original 14 Consumercredit change %change Consumption 15 Personalconsumptionexpenditures %change Original 16 Newhomesales level %change Investment 17 Durablegoodsorders %change Original 18 Constructionspending %change Original 19 Factoryorders %change Original 20 Businessinventories %change Original Government Purchases 21 Governmentbudgetdeficit level %change Net Exports 22 Tradebalance level %change Prices 23 Averagehourlyearnings %change Original 24 Producerpriceindex %change Original 25 Coreproducerpriceindex %change Original 26 Consumerpriceindex %change Original 27 Coreconsumerpriceindex %change Original Forward Looking 28 UMconsumerconfidencepreliminary index Original 29 PhiladelphiaFedindex index Original 30 UMconsumerconfidencefinal index Original 31 CBconsumerconfidenceindex index Original 32 ChicagoPMI index Original 33 ISMPMI index Original 34 Housingstarts level %change 35 CBleadingeconomicindex %change Original 36 Initialjoblessclaims level Original/NFPPopulation Note: This table reports, for each of the 36 announcements, the original unit used in both original agency reports and Bloomberg expectations, and the transformation used in this paper. 14

O.4. Results Without the Zero Lower Bound Period In this appendix, we present replications of the paper’s main results during the sub-sample prior to the Federal Reserve’s zero lower bound period that started in December 2008. Consistent with the findings of Swanson and Williams (2014), the zero lower bound does weaken the findings, especially for the shorter maturities bonds, but our overall conclusions are qualitatively unchanged.

2-T elbaT .dnuoB rewoL oreZ ot roirP – sdleiY yrusaerT .S.U no sesirpruS cimonoceorcaM fo tceffE yrusaerT raeY-5 yrusaerT raeY-2 yrusaerT raeY-1 yrusaerT htnom-6 .sbO 2R .ffeoC 2R .ffeoC 2R .ffeoC 2R .ffeoC tnemecnuonnA n 84 610.0 598.0 350.0 646.1 470.0 *870.2 290.0 **562.2 decnavda PDG 1 84 600.0 354.0- 300.0 863.0- 100.0 091.0- 000.0 520.0yranimilerp PDG 2 74 000.0 60.0- 110.0 976.0- 110.0 897.0- 000.0 120.0lanfi PDG 3 84 000.0 831.0- 100.0 072.0 200.0 863.0 600.0 526.0 ecnavda rotafled ecirp PDG 4 84 720.0 051.1 920.0 064.1 120.0 443.1 220.0 864.1 yranimilerp rotafled ecirp PDG 5 74 000.0 010.0 000.0 210.0 000.0 680.0 310.0 107.0 lanfi rotafled ecirp PDG 6 441 300.0 473.0- 210.0 069.0- 910.0 *193.1- 730.0 **908.1troper tnemyolpmenU 7 441 521.0 ***890.2 961.0 ***360.3 702.0 ***539.3 342.0 ***069.3 tnemyolpme lloryap mrafnoN 8 441 250.0 ***982.1 680.0 ***499.1 201.0 ***603.2 631.0 ***764.2 selas liateR 9 341 940.0 ***673.1 950.0 ***718.1 560.0 ***320.2 070.0 ***059.1 selibomotua ssel selas liateR 01 441 520.0 *588.0 040.0 **203.1 940.0 ***566.1 640.0 ***375.1 noitcudorp lairtsudnI 11 441 970.0 ***635.1 590.0 ***189.1 880.0 ***871.2 380.0 ***190.2 noitazilitu yticapaC 21 341 800.0 355.0- 400.0 834.0- 200.0 923.0- 200.0 983.0emocni lanosreP 31 441 020.0 *398.0- 900.0 717.0- 700.0 307.0- 200.0 103.0tiderc remusnoC 41 341 910.0 407.0 110.0 326.0 900.0 546.0 210.0 507.0 serutidnepxe noitpmusnoc lanosreP 51 341 150.0 ***500.1 240.0 **270.1 030.0 **500.1 720.0 **089.0 selas emoh weN 61 341 310.0 545.0 710.0 097.0 210.0 967.0 400.0 274.0 sredro sdoog elbaruD 71 441 200.0 112.0- 000.0 711.0- 000.0 590.0 200.0 062.0 gnidneps noitcurtsnoC 81 441 800.0 705.0 120.0 *319. 510.0 848.0 700.0 425.0 sredro yrotcaF 91 441 000.0 330.0 100.0 942.0 100.0 172.0 000.0 320.0 seirotnevni ssenisuB 02 441 810.0 776.0- 110.0 216.0- 600.0 474.0- 10.0 475.0ticfied tegdub tnemnrevoG 12 441 610.0 467.0 700.0 885.0 600.0 016.0 000.0 671.0 ecnalab edarT 22 441 330.0 **742.1 330.0 **345.1 030.0 **817.1 820.0 **445.1 sgninrae ylruoh egarevA 32 441 300.0 262.0 100.0 161.0 000.0 341.0- 000.0 410.0xedni ecirp recudorP 42 441 920.0 **487.0 010.0 685.0 400.0 224.0 100.0 091.0 xedni ecirp recudorp eroC 52 441 440.0 **371.1 840.0 ***314.1 50.0 ***684.1 30.0 **430.1 xedni ecirp remusnoC 62 441 890.0 ***448.1 301.0 ***851.2 011.0 ***603.2 470.0 ***107.1 xedni ecirp remusnoc eroC 72 611 110.0 895.0 520.0 *790.1 320.0 431.1 820.0 *013.1 yranimilerp ecnedfinoc remusnoc MU 82 241 830.0 **642.1 370.0 ***578.1 090.0 ***802.2 180.0 ***310.2 xedni deF aihpledalihP 92 511 210.0 345.0 010.0 495.0 100.0 422.0 300.0 723.0 lanfi ecnedfinoc remusnoc MU 03 341 600.0 744.0 110.0 327.0 510.0 059.0 040.0 **474.1 xedni ecnedfinoc remusnoc BC 13 141 580.0 ***438.1 641.0 ***465.2 251.0 ***748.2 261.0 ***769.2 IMP ogacihC 23 441 691.0 ***794.2 302.0 ***130.3 091.0 ***653.3 112.0 ***572.3 IMP MSI 33 441 000.0 530.0 100.0 381.0 200.0 682.0 200.0 762.0 strats gnisuoH 43 341 600.0 535.0 100.0 381.0 000.0 360.0 000.0 550.0 xedni cimonoce gnidael BC 53 426 310.0 ***126.- 220.0 ***669.- 220.0 ***240.1- 620.0 ***311.1smialc sselboj laitinI 63 ehT .sesirprus tnemecnuonna cimonoceorcam no segnahc dleiy dnob yliad fo snoisserger yduts tneve laudividni fo stluser eht stroper elbat ehT :etoN * dna ,** ,*** dna ,desu era srorre dradnats etihW .8002 rebmeceD-dim ot 7991 yraunaJ morf ,dnuob rewol orez eht ot roirp doirep eht srevoc elpmas .ylevitcepser ,ecnacfiingis fo level %01 dna ,5 ,1 a tneserper 16

3-T elbaT .dnuoB rewoL oreZ ot roirP – stluseR lennahC PDG yrusaerT raeY-1 yrusaerT htnoM-6 )6( )5( )4( )3( )2( )1( )6( )5( )4( )3( )2( )1( notneicffieoC ***382.2 ***118.0 ***161.1 ***431.2 ***250.2 ***897.0 ***017.1 ***575.0 ***308.0 ***046.1 ***134.1 ***575.0 S )333.0( )880.0( )781.0( )592.0( )392.0( )890.0( )123.0( )180.0( )061.0( )962.0( )013.0( )801.0( ***524.0 ***092.0 eulaVcisnirtnI×S )980.0( )010.0( ***624.0 ***654.0 ***643.0 ***463.0 slatnemadnuFotnoitaleR×S )390.0( )490.0( )590.0( )580.0( 341.0 **232.0 770.0 541.0 muimerPssenilemiT×S )010.0( )690.0( )960.0( )090.0( 901.0 011.0 000.0 100.0muimerPnoisiveR×S )280.0( )480.0( )470.0( )870.0( ***664.0- ***974.0- ***374.0- ***074.0- ***564.0- ***084.0- ***465.0- ***475.0- ***965.0- ***665.0- ***465.0- ***475.0tnatsnoC )660.0( )880.0( )970.0( )280.0( )270.0( )070.0( )380.0( )770.0( )570.0( )760.0( )270.0( )380.0( 030.0 220.0 420.0 820.0 920.0 220.0 810.0 310.0 410.0 810.0 710.0 310.0 2R yrusaerT raeY-5 yrusaerT raey-2 )6( )5( )4( )3( )2( )1( )6( )5( )4( )3( )2( )1( notneicffieoC ***508.2 ***881.1 ***965.1 ***136.2 ***206.2 ***561.1 ***061.3 ***381.1 ***856.1 ***149.2 ***029.2 ***551.1 S )543.0( )3790.0( )391.0( )263.0( )913.0( )811.0( )874.0( )421.0( )102.0( )023.0( )623.0( )311.0( ***784.0 ***895.0 eulaVcisnirtnI×S )790.0( )201.0( ***964.0 ***105.0 ***965.0 ***016.0 slatnemadnuFotnoitaleR×S )401.0( )901.0( )631.0( )501.0( 651.0 ***852.0 *891.0 ***123.0 muimerPssenilemiT×S )211.0( )890.0( )011.0( )001.0( *781.0 *781.0 **032.0 **132.0 muimerPnoisiveR×S )111.0( )001.0( )011.0( )190.0( **362.0- ***872.0- ***172.0- ***862.0- **262.0- ***972.0- ***233.0- ***153.0- ***143.0- ***833.0- ***033.0- ***153.0tnatsnoC )611.0( )401.0( )401.0( )290.0( )501.0( )201.0( )811.0( )190.0( )490.0( )680.0( )590.0( )101.0( 130.0 720.0 720.0 030.0 130.0 620.0 630.0 820.0 920.0 430.0 530.0 720.0 2R tnemecnuonna ruo htiw detcaretni sesirprus dna sesirprus orcam no ,)1 nmuloc( sesirprus orcam no segnahc dleiy dnob yliad gnisserger fo stluser syalpsid elbat ehT :etoN era scitsiretcarahC .)6 nmuloc( ecno ta lla scitsiretcarahc tnemecnuonna htiw detcaretni sesirprus dna sesirprus orcam no dna ,)5 dna ,4 ,3 ,2 snmuloc( yletarapes scitsiretcarahc deppartstooB .snoitavresbo 3264 no desab si noisserger hcae dna ,8002 rebmeceD-dim ot 7991 yraunaJ morf snur elpmas atad ehT .PDG rof esicrexe gnitsacwon a morf devired .ylevitcepser,ecnacfiingisfolevel%01dna,5,1atneserper*dna,**,***dna,desuerasrorredradnats 17

4-T elbaT .dnuoB rewoL oreZ ot roirP – stluseR lennahC rotafleD ecirP PDG yrusaerT raeY-1 yrusaerT htnoM-6 )6( )5( )4( )3( )2( )1( )6( )5( )4( )3( )2( )1( notneicffieoC ***558.1 ***397.0 ***780.1 ***707.1 ***868.1 ***897.0 ***634.1 ***075.0 ***017.0 ***483.1 ***702.1 ***575.0 S )425.0( )401.0( )231.0( )853.0( )504.0( )9390.0( )624.0( )4090.0( )451.0( )634.0( )462.0( )601.0( ***372.0 **161.0 eulaVcisnirtnI×S )001.0( )960.0( 491.0 **222.0 481.0 *891.0 slatnemadnuFotnoitaleR×S )721.0( )980.0( )211.0( )811.0( ***152.0 ***962.0 701.0 521.0 muimerPssenilemiT×S )380.0( )770.0( )660.0( )580.0( 590.0- 111.0- 311.0- *221.0muimerPnoisiveR×S )010.0( )690.0( )370.0( )760.0( ***174.0- ***974.0- ***964.0- ***184.0- ***574.0- ***084.0- ***075.0- ***375.0- ***965.0- ***575.0- ***175.0- ***475.0tnatsnoC )680.0( )670.0( )080.0( )290.0( )570.0( )860.0( )550.0( )470.0( )470.0( )370.0( )960.0( )270.0( 620.0 220.0 520.0 320.0 520.0 220.0 510.0 410.0 410.0 410.0 410.0 310.0 2R yrusaerT raeY-5 yrusaerT raey-2 )6( )5( )4( )3( )2( )1( )6( )5( )4( )3( )2( )1( notneicffieoC ***333.2 ***651.1 ***465.1 ***831.2 ***714.2 ***561.1 ***763.2 ***741.1 ***095.1 ***241.2 ***745.2 ***551.1 S )645.0( )721.0( )661.0( )806.0( )864.0( )811.0( )136.0( )141.0( )871.0( )975.0( )584.0( )911.0( **913.0 ***553.0 eulaVcisnirtnI×S )521.0( )611.0( 691.0 832.0 791.0 242.0 slatnemadnuFotnoitaleR×S )841.0( )851.0( )951.0( )841.0( ***843.0 ***173.0 ***483.0 ***504.0 muimerPssenilemiT×S )211.0( )890.0( )211.0( )301.0( 802.0- *922.0- 261.0- *481.0muimerPnoisiveR×S )351.0( )531.0( )221.0( )901.0( ***472.0- ***872.0- ***072.0- ***972.0- ***772.0- ***772.0- **962.0- ***772.0- ***562.0- **672.0- ***572.0- ***572.0tnatsnoC )080.0( )580.0( )001.0( )680.0( )201.0( )380.0( )701.0( )590.0( )290.0( )311.0( )690.0( )490.0( 030.0 720.0 920.0 720.0 820.0 620.0 130.0 820.0 030.0 820.0 030.0 720.0 2R tnemecnuonna ruo htiw detcaretni sesirprus dna sesirprus orcam no ,)1 nmuloc( sesirprus orcam no segnahc dleiy dnob yliad gnisserger fo stluser syalpsid elbat ehT :etoN era scitsiretcarahC .)6 nmuloc( ecno ta lla scitsiretcarahc tnemecnuonna htiw detcaretni sesirprus dna sesirprus orcam no dna ,)5 dna ,4 ,3 ,2 snmuloc( yletarapes scitsiretcarahc .snoitavresbo 3264 no desab si noisserger hcae dna ,8002 rebmeceD-dim ot 7991 yraunaJ morf snur elpmas atad ehT .rotafled ecirp PDG rof esicrexe gnitsacwon a morf devired .ylevitcepser,ecnacfiingisfolevel%01dna,5,1atneserper*dna,**,***dna,desuerasrorredradnatsdeppartstooB

5-T elbaT .dnuoB rewoL oreZ ot roirP – stluseR lennahC RTFF yrusaerT raeY-1 yrusaerT htnoM-6 )6( )5( )4( )3( )2( )1( )6( )5( )4( )3( )2( )1( notneicffieoC ***430.3 ***208.0 ***169.0 ***230.3 ***521.2 ***897.0 ***982.2 ***575.0 ***366.0 ***182.2 ***954.1 ***575.0 S )875.0( )301.0( )731.0( )425.0( )443.0( )111.0( )594.0( )7490.0( )841.0( )675.0( )283.0( )5290.0( ***423.0 **612.0 eulaVcisnirtnI×S )480.0( )780.0( ***834.0 ***344.0 ***543.0 ***833.0 slatnemadnuFotnoitaleR×S )111.0( )101.0( )401.0( )901.0( 5510.0 931.0 420.0- 470.0 muimerPssenilemiT×S )390.0( )580.0( )011.0( )011.0( 250.0 840.0 200.0- 400.0muimerPnoisiveR×S )370.0( )460.0( )670.0( )860.0( ***764.0- ***084.0- ***674.0- ***764.0- ***964.0- ***084.0- ***465.0- ***475.0- ***275.0- ***465.0- ***665.0- ***475.0tnatsnoC )270.0( )960.0( )870.0( )580.0( )470.0( )270.0( )180.0( )470.0( )380.0( )470.0( )470.0( )860.0( 820.0 220.0 320.0 820.0 620.0 220.0 710.0 310.0 310.0 710.0 510.0 310.0 2R yrusaerT raeY-5 yrusaerT raeY-2 )6( )5( )4( )3( )2( )1( )6( )5( )4( )3( )2( )1( notneicffieoC ***753.3 ***271.1 ***942.1 ***723.3 ***453.2 ***561.1 ***760.4 ***561.1 ***313.1 ***140.4 ***538.2 ***551.1 S )866.0( )821.0( )091.0( )576.0( )494.0( )231.0( )176.0( )501.0( )871.0( )337.0( )674.0( )990.0( **092.0 ***014.0 eulaVcisnirtnI×S )711.0( )411.0( ***544.0 ***824.0 ***185.0 ***275.0 slatnemadnuFotnoitaleR×S )621.0( )231.0( )521.0( )141.0( 350.0- 170.0 720.0- 431.0 muimerPssenilemiT×S )221.0( )731.0( )101.0( )101.0( 0080.0 870.0 211.0 901.0 muimerPnoisiveR×S )711.0( )211.0( )121.0( )211.0( **762.0- ***972.0- ***772.0- ***662.0- ***862.0- **972.0- ***633.0- ***253.0- ***843.0- ***433.0- ***733.0- ***153.0tnatsnoC )311.0( )690.0( )301.0( )290.0( )201.0( )411.0( )990.0( )311.0( )290.0( )901.0( )190.0( )090.0( 920.0 620.0 620.0 920.0 820.0 620.0 330.0 720.0 720.0 230.0 030.0 720.0 2R tnemecnuonna ruo htiw detcaretni sesirprus dna sesirprus orcam no ,)1 nmuloc( sesirprus orcam no segnahc dleiy dnob yliad gnisserger fo stluser syalpsid elbat ehT :etoN era scitsiretcarahC .)6 nmuloc( ecno ta lla scitsiretcarahc tnemecnuonna htiw detcaretni sesirprus dna sesirprus orcam no dna ,)5 dna ,4 ,3 ,2 snmuloc( yletarapes scitsiretcarahc 3264 no desab si noisserger hcae dna ,8002 rebmeceD-dim ot 7991 yraunaJ morf snur elpmas atad ehT .etaR tegraT sdnuF laredeF eht rof esicrexe gnitsacwon a morf devired .ylevitcepser,ecnacfiingisfolevel%01dna,5,1atneserper*dna,**,***dna,desuerasrorredradnatsdeppartstooB .snoitavresbo

Table T-6 Price Impact and Intrinsic Value – Prior to Zero Lower Bound. 6-Month Treasury 1-Year Treasury Nowcasting Target Nowcasting Target GDP GDP GDP Deflator FFTR GDP Deflator FFTR Coefficient on (1) (2) (3) Coefficient on (1) (2) (3) Intrinsic Value 0.016** 0.005 0.010** Intrinsic Value 0.024** 0.013** 0.017** (0.006) (0.004) (0.005) (0.009) (0.006) (0.006) Constant 0.075*** 0.048** 0.070*** Constant 0.109*** 0.090*** 0.106*** (0.025) (0.018) (0.025) (0.036) (0.029) (0.033) R2 0.142 0.018 0.056 R2 0.166 0.060 0.075 2-Year Treasury 5-Year Treasury Nowcasting Target Nowcasting Target GDP GDP GDP Deflator FFTR GDP Deflator FFTR Coefficient on (1) (2) (3) Coefficient on (1) (2) (3) Intrinsic Value 0.030*** 0.017** 0.021*** Intrinsic Value 0.026** 0.017** 0.018** (0.011) (0.008) (0.007) (0.011) (0.008) (0.007) Constant 0.140*** 0.114*** 0.134*** Constant 0.124*** 0.114*** 0.119*** (0.041) (0.038) (0.040) (0.041) (0.040) (0.039) R2 0.190 0.067 0.083 R2 0.151 0.077 0.066 Note: The table displays results of regressing the estimated R2 coefficients in equation (2) on the announcement’s intrinsic value derived from nowcasting GDP, the GDP price deflator, and the Federal Funds Target Rate. The sample covers the period from January 1997 to mid-December 2008, and each regression is based on 36 observations. White standard errors are used, and ***, **, and * represent a 1, 5, and 10% level of significance, respectively. 20

Table T-7 Price Impact and Macroeconomic Announcement Characteristics – Prior to Zero Lower Bound. 6-Month Treasury NowcastMeasuresof AlternativeMeasures Intrinsic Relationto Timeliness Revision Correlation Reporting Revision Value Fundamentals Premium Premium withGDP Lag Magnitude (1) (2) (3) (4) (5) (6) (7) Coefficient 0.016** 0.016** 0.011 0.006* 0.046* -0.008* 0.011 (0.006) (0.006) (0.007) (0.003) (0.023) (0.004) (0.015) Constant 0.075*** 0.072*** 0.046** 0.027*** 0.0093 0.037*** 0.014 (0.025) (0.023) (0.018) (0.007) (0.006) (0.011) (0.014) R2 0.142 0.132 0.044 0.018 0.095 0.079 0.008 1-Year Treasury NowcastMeasuresof AlternativeMeasures Intrinsic Relationto Timeliness Revision Correlation Reporting Revision Value Fundamentals Premium Premium withGDP Lag Magnitude (1) (2) (3) (4) (5) (6) (7) Coefficient 0.0237** 0.0209** 0.0193* 0.0109*** 0.0608* -0.0144** 0.0308* (0.00922) (0.00899) (0.0107) (0.00385) (0.0355) (0.00659) (0.0171) Constant 0.109*** 0.0974*** 0.0714** 0.0380*** 0.0141 0.0549*** 0.00238 (0.0356) (0.0334) (0.0262) (0.00929) (0.00833) (0.0149) (0.0132) R2 0.166 0.122 0.068 0.032 0.087 0.119 0.034 2-Year Treasury NowcastMeasuresof AlternativeMeasures Intrinsic Relationto Timeliness Revision Correlation Reporting Revision Value Fundamentals Premium Premium withGDP Lag Magnitude (1) (2) (3) (4) (5) (6) (7) Coefficient 0.0304*** 0.0290*** 0.0234 0.0110* 0.0802* -0.0212*** 0.0562** (0.0109) (0.0105) (0.0139) (0.00577) (0.0435) (0.00700) (0.0213) Constant 0.140*** 0.131*** 0.0890*** 0.0482*** 0.0171 0.0735*** -0.0156 (0.0412) (0.0388) (0.0320) (0.0112) (0.0120) (0.0171) (0.0182) R2 0.190 0.164 0.070 0.023 0.106 0.178 0.079 5-Year Treasury NowcastMeasuresof AlternativeMeasures Intrinsic Relationto Timeliness Revision Correlation Reporting Revision Value Fundamentals Premium Premium withGDP Lag Magnitude (1) (2) (3) (4) (5) (6) (7) Coefficient 0.0259** 0.0242** 0.0205 0.00959* 0.0606 -0.0215*** 0.0562*** (0.0110) (0.0103) (0.0146) (0.00511) (0.0420) (0.00697) (0.0195) Constant 0.124*** 0.115*** 0.0816** 0.0459*** 0.0222* 0.0718*** -0.0178 (0.0407) (0.0380) (0.0324) (0.0106) (0.0124) (0.0166) (0.0162) R2 0.151 0.125 0.059 0.019 0.066 0.202 0.087 Note: ThetabledisplaysresultsofregressionsoftheR2fromequation(2)inTableT-2onthemacroeconomicannouncement’s intrinsicvalueanditscomponents(relationtofundamentals,timelinesspremium,andrevisionpremium)derivedfromnowcastingGDPadvance. Thetablealsodisplaystheresultsofsimilarregressionsusingalternativemeasuresforthethreecomponents, namelycorrelationwithGDP,reportinglag,andrevisionmagnitude. ThedatasampleisfromJanuary1997tomid-December 2008,andeachregressionisbasedon36observations. Whitestandarderrorsareused,and***,**,and*representa1,5,and 10%levelofsignificance,respectively.