Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
Distilling the Macroeconomic News Flow∗
Alessandro Beber†, Michael W. Brandt‡, Maurizio Luisi§
February 2013
Abstract
We propose a simple cross-sectional technique to extract daily latent factors fromeconomic news releases available at different dates and frequencies. Our approachcan effectively handle the large number of heterogeneous announcements that arerelevant for tracking current economic conditions. We apply the technique to extractreal-time measures of inflation, output, employment, and macroeconomic sentiment,as well as corresponding measures of disagreement among economists about thesedimensions of the data. We find that our procedure provides more timely andaccurate forecasts of the future evolution of the economy than other real-timeforecasting approaches in the literature.
Keywords: macroeconomic news, nowcasting, disagreement.
JEL classification: G12
∗We thank Fabio Fornari, Amit Goyal, and seminar participants at BlackRock, City University, the 2012 AssetPricing Retreat at Cass Business School, the Fall 2012 Inquire UK Conference in Bath, the Imperial College HedgeFund Conference, and the University of York, for their comments and suggestions. We are indebted to InquireUK for financial support.†Cass Business School, City University London, and CEPR‡Fuqua School of Business, Duke University, and NBER§Quantitative Investment Solutions
1 Introduction
Timely measurement of the state of the macroeconomy relies traditionally on low-frequency
observations of a few economic aggregates that refer to previous weeks, months, or even quarters.
A prominent example is the advance estimate of GDP released quarterly about a month after
the end of the quarter. The low frequency and delayed observation of any such economic
aggregate considered in isolation stands in sharp contrast with the rich economic news flow that
market participants observe daily. This news flow contains information that agents use to learn
about the economy in the absence of private information. In particular, the macroeconomic
news literature has identified a large cross-section of dozens of different news releases that have
significant and immediate effects on financial markets (e.g., Andersen et al., 2003).
We propose to distill the economic news flow observed by market participants into a set
of indicators measuring four distinct aspects of the economy: inflation, output, employment,
and macroeconomic sentiment. Specifically, we describe a simple cross-sectional technique
to extract daily principal components from economic news releases associated with a given
information type and observed at different times and frequencies. Our approach is simple,
robust (no numerical optimization is required), and can effectively handle the large number
of announcements that are relevant for tracking the evolution of macroeconomic conditions in
real-time. At the same time, our empirical analysis shows that the output of our approach is as
if not more timely and informative than more sophisticated but also more difficult to implement
statistical techniques. Intuitively, the potential disadvantage of a simpler modeling framework
is more than compensated by the sheer quantity of data our approach can effectively process.
Our paper relates to the literature that attempts to measure the state of the economy in
a time-series setting based only on fundamental economic data (see Banbura et al., 2012, for
a survey), commonly referred to as “nowcasting.” There are basically two general approaches
to this problem. The first approach is to use a balanced panel regression, along the lines of
the seminal paper of Stock and Watson (1989). The goal is to construct a coincident index of
economic activity using factor models on a large set of macroeconomic releases, which basically
amounts to constructing a weighted average of several monthly indicators. The advantage of this
approach is that the indicator relies on many macroeconomic variables. However, this advantage
also results in a relatively low frequency of the measurement, because the econometrician has
to wait for the panel to be complete before the index can be evaluated. A second general
approach is to model macroeconomic data using a state-space framework (e.g., Evans, 2005).
The advantage of this second approach is to produce an indicator at a higher frequency, since
a state-space model can effectively handle the sparse and delayed reporting of macroeconomic
data and missing information on non-release days. However, this technique is impractical for
large cross-sections of macroeconomic releases. For example, Evans (2005) only considers the
set of different (preliminary, advance, final) GDP releases. In follow-up work, Arouba et al.
(2009) accommodate four indicators at different frequencies including a continuously observable
financial markets variable. Finally, Giannone et al. (2008) combine the two approaches by
1
modeling factors extracted from a balanced panel in a state-space setting.
Our goal is to measure the state of the economy with a methodology that retains the
advantages of both approaches without their respective limitations. Specifically, we propose a
simple and economically intuitive cross-sectional dimension reduction technique that can easily
accomodate a large set of macroeconomic releases. This is a crucial aspect of our approach,
given the evidence in the literature that the information relevant for describing the state of the
economy is high dimensional. At the same time, our methodology is designed to handle data
released at different frequencies and missing observations to produce a real-time (daily or even
intradaily) high-frequency measurement of the state of the economy.
Our methodology has several other differentiating features relative to the literature. First,
we do not aim to estimate a real-time series of GDP, or any other single observable variable,
but we rather leave the factor(s) truly latent and unspecified. In this sense, we do not impose
any structure on the estimation and thus do not take a stand on what is the appropriate metric
of the state of the economy. We simply let the data speak for itself. Second, our focus on a
large cross-section of economic news releases allows us to extract factors from four subsets of
macroeconomic news (e.g., inflation, output, employment, and macroeconomic sentiment). We
use these subset indicators to learn about the relations between different driving forces of the
economy. Third, we utilize news flow data that is truly real-time and unrestated, as opposed to
approximately dated historical data that is notoriously restated (e.g., Koening et al., 2003; see
also Ghysels et al., 2012, for an illustration of the issues arising from restated macroeconomic
data). Finally, we also apply our methodology to the dispersion of forecasts. This is a new way
to obtain a high-frequency measure of macroeconomic uncertainty based on the disagreement
of a cross-section of economic experts.
We find that an economic activity factor (which combines output and employment information,
as they are highly correlated) as well as a macroeconomic sentiment factor, both extracted
from the large cross-section of macroeconomic news, have sensible dynamics. The greatest
dips in both series are well aligned with the ex-post defined NBER recession phases. The
macroeconomic sentiment factor, obtained from consumer and business confidence releases, is
highly correlated with economic activity, but appears to lead fundamentals especially around the
most important turning points. Finally, our inflation factor exhibits dynamics that seem only
weakly correlated with growth, with much more erratic variation, and has an unclear pattern
in expansion versus recession phases.
Our empirical proxy of economic uncertainty based on economist disagreement is interesting
for at least two reasons. First, it shows little correlation with the estimates of the latent economic
activity, macroeconomic sentiment, and inflation factors, suggesting that they are likely to
contain different information. Second, and more importantly, macroeconomic uncertainty exhibits
intriguing asymmetric dynamics. The peaks of disagreement correspond to the final stages of
recession periods, while uncertainty is relatively subdued at the end of economic expansions.
This evidence suggests that economists tend to disagree mostly on recoveries from economic
contractions, whereas everyone seems to see the end of an expansion coming.
2
We formally relate a real-time factor of economic growth (which further aggregates the
information relative to economic activity by combining information relating to output, employment
and macroeconomic sentiment) to the CFNAI index, which is constructed by the Chicago
Federal Reserve Board based on Stock and Watson (1989), on CFNAI release dates at the
monthly frequency and to the vintage version of the ADS index of Arouba et al. (2009) at
the weekly frequency. We find that our latent growth factor is strongly correlated to both of
these alternative approaches. However, since our factor is constructed using information from
either a larger cross-section of news or in a more timely manner, it turns out to have significant
forecasting power for both CFNAI and ADS beyond their own lags. In a related empirical
exercise, we find that our growth factor has predictive power for future actual GDP releases
and is highly correlated with the quarterly GDP expectations in the Survey of Professional
Forecasters (SPF). This is a remarkable feature given that, unlike the CFNAI and ADS indices,
our growth factor is not optimally weighted to forecast GDP. The large correlation with the
quarterly releases of the SPF offers an intuitive interpretation of our growth factor as the high-
frequency daily reading of economist expectations about macroeconomic fundamentals.
A very intriguing finding is that the latent factor obtained exclusively from macroeconomic
information turns out to be highly correlated with financial indicators, such as the default spread
or the implied stock return volatility index VIX. More specifically, we find that the combination
of our latent growth factor and its dispersion can explain almost one-third of VIX levels. This
is an important finding in light of the documented difficulties for macroeconomic quantities to
explain financial volatility (see, for example, the seminal paper of Schwert, 1989).
Finally, we combine the information of the growth indicator and its dispersion and document
very strong predictability for future growth, from five days and up to six months ahead. Given
the illustrated relation of our macroeconomic indicator with financial variables and its extremely
timely nature, this result suggests that our quantitative measure of the news flow could have
predictive power for future financial market dynamics.
The remainder of the paper proceeds as follows. In Section 2, we describe the macroeconomic
news and we carry out some preliminary analysis on economic announcements. Section 3
explains our methodology for estimating in real-time the state of the economy, as captured
by four dimensions of the data (inflation, output, employment, and macroeconomic sentiment),
and ex-ante uncertainty revealed by the cross-section of economist forecasts. We present our
empirical results in Section 4. Section 5 concludes with a summary of our findings.
2 Data and Preliminaries
2.1 Macroeconomic news and forecasts
We obtain data on the dates, release times, and actual released figures for 43 U.S. macroeconomic
announcements covering the period from January 1997 through December 2011, for a total
of more than 8,000 announcements over about 3,800 working days. This data is obtained
3
from Bloomberg through the Economic Calendar screen, which provides precisely time-stamped
and unrestated announcement data.1 We also collect data on economist forecasts for each
announcement. Bloomberg surveys economists during the weeks prior to the release of each
indicator to obtain a consensus estimate. We work with the individual economist level forecasts,
rather than the aggregated consensus forecasts, in order to construct cross-sectional measures
of disagreement for each news release.
Bloomberg contains data for many of our series prior to 1997, but those data are stored in
historical fields which (a) are not associated with clear announcement dates and times (rather
they are dated according to the period they reference) and (b) are restated over time.2 We
collect this more problematic data for January 1985 through 1996 for two reasons. First, we
use this historical data to construct an initial correlation matrix estimate, which is required by
our methodology (see Section 3). Second, we use this data for a robustness check with a longer
sample period (see Section 4.5). In order to date the releases prior to 1997, we compute for each
news series the median time between the reference period and the announcement. For example,
the employment report is traditionally released four days after the end of the month to which
the report refers. We then apply this median reporting lag to the reference period of the older
data in order to obtain an approximate announcement date.
Since economist-level forecasts are not available prior to 1997, we instead collect data from
the Survey of Professional Forecasters. The Survey of Professional Forecasters is the oldest
quarterly survey of macroeconomic forecasts in the United States. The survey began in 1968 and
was conducted by the American Statistical Association and the National Bureau of Economic
Research. The Federal Reserve Bank of Philadelphia took over the survey in 1990. The Survey of
Professional Forecasters’ web page offers the actual releases, documentation, mean and median
forecasts of all the respondents as well as the individual responses from each economist. The
individual responses are kept confidential by using identification numbers.
Most macroeconomic indicators are released on different days and at different frequencies,
making it difficult to process the flow of information in a systematic and consistent way. Figure
1 shows that actual news releases occur with a variety of different lags with respect to the
month they are referring to. Furthermore, news on different indicators are frequently released
simultaneously.3 For example, the employment report traditionally announced on the first
Friday of the month contains four different indicators: non-farm payrolls, non-farm payrolls
in the manufacturing sector, unemployment rate, and average weekly hours. Finally, the
release frequency varies across different economic aggregates. Data releases of different economic
indicators are usually observed at different frequency; e.g., GDP data are sampled quarterly,
1The importance of using real-time versus final data in macroeconomic forecasting has been discussedextensively in the literature (e.g., Koenig et al., 2003).
2For example, there are monthly releases of quarterly GDP labeled “advance,” “preliminary” and “final” allreferring to the same quarter. Bloomberg’s historical field for GDP is dated according to the referenced quarter,so that the advance release gets overwritten by the preliminary release, which in turn gets overwritten by thefinal release. Historically only the final releases are stored.
3On approximately 80 percent of days, there was at least one data release. Multiple data releases occurredmuch less frequently, on approximately 60 percent of the days in the sample.
4
the non-farm payrolls are released monthly, initial jobless claims are sampled weekly, etc. These
features of our large cross-section of macroeconomic releases generate a sparse matrix of data
that our methodology will have to take up.
The Appendix describes in detail the set of macroeconomic news in our sample, including
their frequency, source, and units of measurement.
2.2 Categorizing the macroeconomic news flow
Our aim is to extract a set of factors describing the state of the economy. Rather than relying on
a statistical procedure to obtain a set of orthogonalized factors that are increasingly difficult to
interpret with the order of the factor, we impose a specific economically motivated structure on
the macroeconomic news flow. Based on both empirical evidence and economic rationale, we first
separate the aggregate economy into two broad dimensions: the nominal, and the real side.4
In practice, we split the set of announcements into nominal inflation-related announcements
and news that relates to real growth. Growth data, in turn, come in two flavors – objective
realizations of past economic activity and subjective often forward-looking views derived from
surveys which we label “macro sentiment.” Finally, economic activity can be split one last time
into information relating to output versus employment.
Through this structure, we obtain two (inflation and growth), three (inflation, economic
activity, and macro sentiment), or four (inflation, output, employment, and macro sentiment)
factors:• Inflation
• Growth
Economic Activity
Output
Employment
Macro Sentiment
where, for example, the Economic Activity factor is obtained from the combined information
relating to Output and Employment. In that sense, the information is nested from right to left.
More specifically, the inflation factor is extracted from the news flow of 10 inflation-related
releases: consumer price index, CPI ex food and energy, the employment cost index, GDP price
index, import price index, nonfarm productivity, PCE core, PPI ex food and energy, producer
price index, and unit labor costs. For the output factor, we utilize information from both the
supply and demand side of the economy in the form of news about advance retail sales, business
inventories, capacity utilization, consumer credit, domestic vehicle sales, durable goods orders,
durables ex-transportation, factory orders, GDP, industrial production, ISM manufacturing,
ISM non-manufacturing composite, personal consumption, personal income, personal spending,
retail sales less autos, and wholesale inventories. Employment news is captured by releases
of ADP payrolls, manufacturing payrolls, non-farm payrolls, continuing claims, initial jobless
4The economy is often separated into the nominal and real sides because shocks to the two should be separatedand treated differently. For example, many argue, from the perspective of monetary policy, nominal shocks shouldbe minimized, whereas real shocks should not be intervened upon. Other studies also suggest that a nominal anda real factor can account for much of the observed variation in major economic aggregates.
5
claims, and the unemployment rate. Finally, we extract the macro sentiment factor from
the news flow generated by 10 macroeconomic surveys: ABC consumer confidence, Chicago
purchasing manager, consumer confidence, Dallas Fed manufacturing activity, Empire manufacturing
survey, leading indicators index, NAPM-Milwaukee, Philadelphia Fed business outlook survey,
Richmond Fed manufacturing index, and the University of Michigan confidence index. The
Appendix includes a summary on how the 43 announcements are assigned to the four categories:
inflation, output, employment, and macro sentiment.
It is worth reiterating at this point that we do not include any market based data (such as
stock prices, interest rates, credit spreads, or VIX) in our analysis, unlike, for example, Arouba
et al. (2009) and Giannone et al. (2008). While such data are very timely and undoubtedly
informative about the state of the economy, they represent already the market’s interpretation
of the macroeconomic news flow. Our aim is to objectively summarize and describe the
macroeconomic news flow itself.
2.3 Transformation and temporal alignment
We examine the stationarity of each data series in two ways. First, we conduct a Dickey-Fuller
test on each series. Second, we read the definition and description of each statistic to determine
from an economic perspective whether it is a non-stationary index or a stationary quarterly
growth rate, for example. In a few cases where the conclusions from the two approaches differ,
usually because the available data is too short to examining statistically, we rely more on
the description to determine whether the series is stationary. All series that are deemed non-
stationary are then first-differenced in news release time. The Appendix provides more details.
The final data management task is to align the data temporally by moving from announcement
time to calendar time. We do this by populating the news releases in a T×N matrix where T
denotes the total number of days in our sample and N refers to the 43 announcement types.
The data at this stage looks like the top panel of Figure 2.
There are two important aspects of the data to discuss. First, there are a vast number
of missing values as we can think of each news series as a continuous evolving statistic that
is observed only once per month or quarter. Second, not all announcements have a complete
history. Some announcements are initiated in the middle of the sample and/or are terminated
before the end of the sample. To solve the missing data problem, we simply forward fill the last
observed release until the next announcement. Forward filling can be rationalized as replacing
missing values with expected values under a simple independent random walk assumption for
each news series. Of course, both independence in the cross-section and random walk dynamics
through time are simplifying assumptions that are rejected by the data (in fact, the motivation
for our methodology described below is the cross-sectional correlation structure within news
category). A more sophisticated approach for filling in missing data would be to compute the
expectation of the missing values given the full cross-section of previous releases as well as the
cross-sectional and intertemporal correlation structure of the data. An optimal solution would
6
also allow for sampling error, which is the case in Kalman filter or Bayesian data augmentation
algorithms. However, there is a clear trade-off between statistical complexity and ability to
process a large cross-section of news series. Since the goal of our approach is to utilize the
entire cross-section of news, we choose a very simple statistical model for filling in missing
observations. After forward filling, the data looks like the bottom plot of Figure 2.
Note that the second data issue, the fact that some series do not span the entire sample
period, cannot be solved with missing values imputation. It is instead explicitly addressed in
our methodology below.
3 Methodology
3.1 Subset principal component analysis
Our goal is to extract from the cross-section of macroeconomic news releases a set of factors
that capture in real time the state of inflation, output, employment, and macro sentiment, as
well as the two more overarching factors measuring economic activity and growth. The most
obvious ways of accomplishing this, full data principal components and forecasting regressions,
do not appeal to us. First, with full data principal components (or factor analysis) we would
obtain factors that are mechanically orthogonal, whereas the dimensions of the economic news
flow we want to capture are likely correlated (e.g., output and employment are both high at the
peak and low at the trough of an economic cycle). This orthogonalization makes is practically
impossible to assign an economic meaning to higher order factors. Second, trying to identify
the factors through predictive regressions on a candidate variables in each category, such as
final GDP for output, would require us being able to identify a single series that represents
each category. While this is a common approach in the nowcasting literature (e.g., Stock and
Watson, 1989), it relies on ex-ante knowledge of the key statistic to track and assumes that
there is only one such statistic that does not change over time.
Instead, we rely on our ex-ante categorization of the news and, within each category subset,
let the data speak for itself by extracting the first principal component of that subset of data.
Specifically, on each day of our sample t, we obtain for each news category i the first principal
component from the correlation matrix Ωt,i of the stationary news series in category i. We
work with the correlation matrix to abstract from arbitrary scaling of data. Moreover, in order
to obtain a real-time measure, we use a telescoping (with a common historical start date and
rolling end dates) correlation matrix starting in 1980.5 We denote the Ni×1 principal component
weights by ct,j , where Ni is the number of news series in category i.
5We also experimented with fixed window size rolling correlation matrices for 5, 10, 15, and 20 years. Theresults are qualitatively similar, particularly for the longer data windows.
7
3.2 Economic new series correlation matrix
The key inputs to our methodology are the within news category correlation matrices Ωt,i.
Specifically, we need to calculate from historical data up through date t the correlation of all
news series of category i that are “active” on that date, where active means that the news
series was previously initiated and has not yet been terminated. There are two issues that
need to be addressed in computing these correlation matrices. First, the data is in the form
of an unbalanced panel due to some of the series being initiated after the start date of the
estimation window (e.g., series j = 5 in Figure 2). Second, the data is naturally persistent,
partly due to autocorrelation of the the data in announcement time, partly due to the cross-
sectional misalignment of the news in calendar time, and partly due to the forward filling of
missing data.
We address the first unbalanced panel issue by using a correlation matrix estimator along
the lines of Stambaugh (1997), who shows how to adjust first and second moments estimates
for unequal sample lengths. The intuition of his approach is to use the observed data on the
longer series, along with a projection of the shorter series on the longer ones estimated when
both are observed, to adjust the moments of the shorter time series.
To correct for the persistence in the data, we could use in principle the standard approach
of Newey-West (1987), where due to the nature of the data we would like to adjust for up to
one quarter of autocorrelation and cross-autocorrelation. Unfortunately, the nature of the
persistence in our data is not ideally captured by the non-parametric Newey-West (1987)
approach for two reasons. First, we have daily data, so adjusting for up to a quarter of
autocorrelation would involve approximately 60 cross-autocorrelation matrices. Second, the
(cross-) autocorrelations are not exponentially decaying as a typical ARMA model might predict.
Instead, the data is locally constant, due to the forward filling, and over longer intervals
only moderately (cross-) autocorrelated due to the statistical nature of the news series. This
peculiar correlation structure is actually identical to that found in high-frequency asset prices,
where asynchronous and infrequent trading creates a misaligned and locally constant panel
of observations. In that literature, Ait-Sahalia, Mykland, and Zhang (2005) propose a “two-
scales realized volatility” estimator to handle this specific structure of short versus long-horizon
dependence. Specifically, their estimator subsamples the data at a sufficiently low frequency
that overcomes the local constancy and then averages over the set of all possible estimators that
start the subsampling schemes at different times.
We adopt exactly the same approach, except of course our application is very different.
Specifically, at date t we subsample the forward filled news series backward at a monthly
frequency and then compute a Newey-West estimate of the correlation matrix using four lags.
We repeat the same for monthly sampling starting at dates t− 1, t− 2, ..., t− d+ 1 (assuming
d days per month) and then average the resulting d correlation matrix estimates.
8
3.3 Level versus disagreement factors
Given the vector of principal component weight ct,j , we then construct for each news category
two times-series. First, we sum at each date the product of the weights multiplied by the
most recent releases to obtain our real-time level factors. Second, we sum the product of the
same weights multiplied this time by the cross-sectional standard deviations of the economist
forecasts for the most recent releases to obtain our real-time disagreement factors. Throughout
our sample not every news series has economist level forecasts data available. We therefore
construct the disagreement factor using the available data, renormalizing first the principal
component weights to account for the proportion of missing data.
4 Results
We first provide some descriptive evidence about the dynamics of our real-time macroeconomic
factors. Next, to have a sense for how our methodology compares empirically with other
approaches in the literature, we focus on the real-time growth factor and compare it with
the vintage versions of the CFNAI index and of the ADS index of Aruoba et al. (2009).
We then analyze whether our real-time factors are significant predictors of subsequent GDP
releases, in a horse race with the popular forecasts from the Survey of Professional Forecasters
(SPF) of the Federal Reserve Bank of Philadelphia. We then examine the relation between the
real-time growth factor, its dispersion, and volatility in financial markets to establish a tighter
relation between financial volatility and macroeconomic dynamics than previously observed
in the literature (e.g., Schwert, 1989). Finally, we analyze the predictability of our real-time
growth factor for different forecasting horizon and we extend the sample backward using a
pseudo real-time approach as a robustness check.
4.1 Preliminaries
In panel A of Table 1, we compute correlations between seven underlying factors describing
different dynamics of the U.S. economy and their respective proxies for macro uncertainty. We
observe a number of interesting facts. First, there is low correlation between latent factors
and their dispersion measures, suggesting that they are likely to contain different kinds of
information. For example, the growth factor has a -0.14 correlation with growth dispersion.
Second, the inflation and growth factors also exhibit low correlation, confirming our ex-ante
view that it makes sense to consider them separately. Third, the macroeconomic sentiment
factor happens to be strongly related to economic activity, suggesting that the various confidence
readings are largely based on the observed strength of the economy. Inflation information, in
contrast, seems unrelated to sentiment.
In panel B, we compute contemporaneous correlations between the growth and growth
dispersion factors and a number of financial market variables that in various cases have been
associated with the state of the economy or macroeconomic uncertainty. There is basically no
9
contemporaneous correlation between our latent factors and stock market returns. In contrast,
there are important contemporaneous correlation between the growth factor and a number of
financial variables, most notably the correlation with VIX (-0.51), the dividend yield (-0.71),
and the default premium (-0.84).6 The growth dispersion factor has a somewhat weaker relation
with financial variables, but it still retains a significant correlation with VIX (0.25) and with
the price/earning ratio (0.19). These descriptive results foreshadow the strong link between our
macroeconomic factors and financial volatility documented more thoroughly in Section 4.3.
Figure 3, 4, and 5 provide the graphical counterpart of the unconditional correlations in
Table 1. Figure 3 shows the estimated real-time output and employment factors. Output seems
to anticipate employment somewhat, especially around turning points, but the two factors are
very highly correlated. Given this high correlation, we proceed to work with the more aggregated
economic activity factor for the remainder our our empirical analysis.
We compare our real-time measurements of economic activity (which combines output and
employment releases) with macro sentiment in the upper panel of Figure 4. As in the previous
figure, we observe a large correlation between the two real-time factors, with macro sentiment
clearly anticipating economic activity around the most important turning points. While this
divergence could be useful in a stock market predictability exercise, in the remainder of this
paper we will focus on the growth factor, which is constructed using both economic activity
and macro sentiment releases.
In the lower panel of Figure 4, we compare the inflation factor with the growth factor that
represents well its subset factors. We notice that while the two series are somewhat positively
correlated, the inflation series is much more erratic and thus represents different dynamics, as
we expected from the unconditional correlation reported in Table 1. For this reason, we will
keep the growth factor separated from the inflation factor in the remainder of the paper.
We conclude this graphical presentation of results in Figure 5, where we illustrate the
estimate of the real-time growth factor in the upper panel and the growth disagreement factor
in the lower panel. As can be readily seen, the growth index dips in the recession periods of 2001
and 2008-2009. The dispersion index exhibits the taller peaks of disagreement corresponding to
the terminal stages of the recession periods, suggesting that economists tend to disagree mostly
on the way out of the economic contractions rather than at the end of economic expansions.
4.1.1 Comparison with the CFNAI index
Recent economics and finance literature have frequently utilized the CFNAI index as an indicator
of economic conditions (e.g., Beber et al., 2011). The CFNAI index, which is based on the work
6We obtain daily data on S&P500 returns, the VIX index, the dividend yield, the price earning ratio, thedefault premium (as the difference between Moody’s BAA - AAA corporate bond yield), and the term premium(as the difference between 10-year and 3-mo Treasury yields) from Bloomberg and Datastream. We obtain ameasure of expected realized volatility ERV using 5-minute S&P500 returns computed with data from TickDataand the econometric model outlined in Section 4.3. The volatility risk premium is the difference between theVIX index and the expected realized volatility ERV .
10
of Stock and Watson (1989), is preferred to the traditional NBER expansion and recession dates,
because it is generally available promptly. The CFNAI is a very interesting comparison for our
high-frequency indicator, because it is a weighted average of 85 monthly indicators and, as such,
utilizes information for a large cross-section, as we do. However, the CFNAI indicator remains
at the low monthly frequency level and does not deal with information released on different
days over the month and with missing observations.
An effective comparison needs to rely again on information that was available in real-time
and that was not subsequently revised. We thus obtain CFNAI vintages from the CFNAI
website maintained by the Chicago Fed. This is particularly important for the CFNAI indicator,
since at the time of the release, the actual value of a large part of the indicators is not available
and is projected from the last available value and subsequently restated when information
becomes available. For the same reasons, we also carefully consider the CFNAI release date.
The Chicago Fed states that CFNAI is normally released toward the end of each calendar month.
Based on the last available actual release dates, the release date is on average on the 23rd day
of the month. We thus match each monthly CFNAI release with our estimated growth factor
on the same day.
Figure 6 plots the CFNAI monthly indicator and our real-time growth factor that have been
both standardized for this subsample to facilitate the comparison. As can be readily seen, the
two series are again very similar. More specifically, the correlation between the two series is
0.94. We also notice that our growth index anticipates the turning points of the CFNAI index.
We prove this observation more formally by regressing the CFNAI index on its own lag
and the previous month’s real-time growth factor. Panel A of Table 2 shows that the growth
factor has significant predictive power for the CFNAI index, beyond lagged CFNAI. This both
surprising and reassuring, since our growth factor is constructed very differently. Rather than
specifically geared toward predicting GDP, as the CFNAI index is, our growth factor is simply
the first principle component of all growth related news. When we perform the opposite exercise
in Panel B, that is, we regress the growth factor on lagged growth and lagged CFNAI, the
lagged CFNAI is not a significant predictor. Finally, it is worthwhile reiterating that even if
our growth index did not lead the CFNAI index chronologically (which it does) but was only
contemporaneously highly correlated with it, the benefit remains that our index is observed
daily and is based on truly real time data, as opposed to being available at a monthly frequency
and based on questionably timed and potentially restated data.
4.1.2 Comparison with the ADS index
As explained in detail in the description of the methodology above, we depart from the existing
literature mainly because we are able to consider a large cross-section of the news flow, much in
the same way a professional investor would observe the daily releases of macroeconomic news.
It is then important to compare the features of our estimated factor to a more traditional real-
time factor obtained from a small set of releases but through a more complicated statistical
11
modeling framewprl.
We obtain vintages of the ADS business condition index from the website of the Federal
Reserve Bank of Philadelphia. It is crucial to use the vintages of the index as opposed to the
smooth full-sample time-series. In fact, our aim is to compare our real-time growth factor with
ADS on common grounds, that is, using calendar time for information that became actually
available to market participants. In other words, we do not want any information to matter
that was not available in real-time, especially revisions of previous announced figures.
Vintage data for the ADS index are only available starting from the end of 2008 and are
released weekly. The vintage sample (283 observations) is much shorter than the full-sample,
but it is still very informative for our comparative exercise. We thus match each weekly released
vintage of ADS with our rolling daily growth factor for the same date.
Figure 7 plots the ADS index and the growth factor that have both been standardized for
this subsample to facilitate the comparison. As can be readily seen, the two series are very
similar. More specifically, the correlation between the two series is 0.91. We also notice that
our growth index is much smoother and seems to anticipate somewhat the turning points of the
ADS index.
We prove this visual intuition more formally, when we regress the ADS index on its own lag
and the previous week’s growth index. Panel A of Table 3 shows that our real-time growth factor
has significant predictive power for the ADS index, beyond lagged ADS. When we perform the
opposite exercise in panel B, that is, we regress the growth factor on lagged growth and lagged
ADS, the lagged ADS is not a significant predictor.
In summary, in this last subsection, we have provided empirical evidence that our methodology
of extracting information from a large cross-section of news releases at daily frequency delivers
an estimate of a growth factor that is correlated with existing indices, but seems to provide
more timely and frequent information. Since our index is correlated with these other indices
that are explicitly designed to forecast GDP, checking how well our index forecasts GDP is a
logical progression of our analysis.
4.2 Forecasting actual GDP
We examine the relation between our real-time growth factor and the actual releases of GDP
figures. Unlike other approaches to obtain economic indices (e.g., the construction of the
CFNAI), our methodology does not rely on optimizing the weights on the releases to obtain
the best GDP predictions. On one hand, we would favor our unstructured approach given that
GDP is measured with error and there is some disagreement on what is the most appropriate
measure of the state of the economy at any point in time (e.g., Stiglitz et al., 2010). On the other
hand, we need to make sure that a real-time growth factor obtained with such an unstructured
approach resembles an actual economic aggregate.
We set this empirical exercise in a forecasting framework using as a benchmark for our
growth factor the nominal GDP growth quarterly forecasts from the Survey of Professional
12
Forecasters (SPF) carried out by the Federal Reserve Bank of Philadelphia. More specifically,
we use the average projection of nominal GDP annual growth rate for the current quarter that
is usually released around the end of the second month of the quarter and match it with our
growth factor for the same day. We use both the SPF projection and our growth index to
forecast the actual GDP growth for the same quarter, which will be released about one month
after the end of the quarter. For example, for the GDP in the first quarter of 1997, we use the
SPF median projection of 4.73 percent released on February 26, 1997 and our growth index of
0.7877 on the same day to forecast the actual release of GDP that occurred on April 30, 1997
at 5.60 percent.
Table 4 shows the results. We find that both the mean projection of the SPF and our real-
time growth factor contain useful information to predict the following actual GDP releases, with
large R2 of about 45 percent and 41 percent, respectively. Furthermore, the correlation between
the real-time growth factor and quarterly SPF projections is large and equal to 0.89, suggesting
that our growth index can essentially be read as an high-frequency (daily) reading of growth
expectations with very similar features of the ones obtained from the Survey of Professional
Forecasters. Figure 8 makes these points more clear in a graphical comparison, where the large
correlation between our growth factor, SPF expectations, and subsequent GDP actual releases,
can be easily appreciated. This observation is consistent with Liebermann (2001) who finds
that (a different approach to) nowcasting is comparable to the SPF at the date of release but
superior prior (when no SPF is available) and shortly after as it updates.
A related topic is the relation between our measure of real-time macroeconomic uncertainty
and the disagreement about GDP growth in the Survey of Professional Forecasters. We thus
match our indicator of dispersion of economic growth with the disagreement about current
quarter GDP growth in the SPF, both recorded on the day of the SPF release. The two series
have an unconditional correlation of 0.55 and peak at similar times on the way out of recession
phases. Although the correlation is not as strong as for the average expectation and the growth
factor, our proxy for macroeconomic uncertainty can still be interpreted as a high-frequency
reading of disagreement dynamics of professional forecasters.
4.3 Macroeconomic factors and financial volatility
One of the most intriguing aspects of our methodology is to produce a real-time daily reading
of the state of the economy that does not rely on information from financial markets, unlike
Aruoba et al. (2009) and Giannone et al. (2008). As a result, investigating the link between
the macroeconomic state and financial market dynamics is a natural next step. The existence
of this link is far from obvious. In the seminal paper of Schwert (1989), a number of volatilities
obtained from a host of macroeconomic variables and a recession dummy could explain only
a tiny part of stock market volatility for different sample periods. More recently, Engle and
Rangel (2008) define the relation between the state of the economy and aggregate financial
volatility the central unsolved problem of 25 years of research in volatility.
13
Figure 9 and 10 provide some graphical evidence of the strong link between our real-time
macro factors and financial volatility. In Figure 9, the real-time growth factor (on an inverted
scale) and VIX are almost indistinguishable. In Figure 10, our proxy for growth uncertainty
obtained from economist disagreement also correlates well with the VIX index.
In Table 5, we test this contemporaneous link more formally. More specifically, in panel
A, we regress the VIX index on the growth factor and its dispersion. Both variables are
strongly statistically significant explanatory variables for VIX, with dispersion alone explaining
six percent of VIX variation, the growth factor explaining about a quarter, and both explaining
29 percent of VIX variation.
At this point, given that the VIX index is a risk-neutral measure of expected volatility
embedding the effects of investor preferences, we decompose it into its primitive components,
namely an expected realized volatility measure and a volatility risk premium, to better understand
the origin of the macro variables explanatory power. Specifically, we obtain high-frequency
returns for the S&P500 stock index from Tickdata. We construct a measure for daily realized
variance RVt as the summation of the 78 five-minute squared returns covering the normal trading
hours, as in Bollerslev et al. (2009).7
We compute the daily variance risk premium V RPt as the difference between daily implied
variance IVt and the current expectation of RVt for the next 22 trading days. We forecast future
realized variance using information from both realized and implied volatility, as in Drechsler and
Yaron (2011). More specifically, we obtain the expectation of realized variance as the optimal
forecast of future variance based on current realized and implied volatility as observed both
at time t and on average during the last month. Using lagged volatility terms measured on
different horizons is consistent with some promising volatility forecasting Heterogeneous Auto-
Regressive models posited in Corsi (2009) and used in Corsi, Fusari, La Vecchia (2012) and
Mueller, Vedolin, and Yen (2012). Formally,
V RPt = IVt − Et [RVt,t+22] (1)
Et [RVt,t+22] =(β1IV
0.5t−22,t + β2RV
0.5t−22,t + β3IV
0.5t + β4RV
0.5t
)2, (2)
where the vector of β parameters are estimated with a rolling one-year window so as to avoid
any look-ahead bias in the following regression:
RV 0.5t,t+22 = α+ β1IV
0.5t−22,t + β2RV
0.5t−22,t + β3IV
0.5t + β4RV
0.5t + εt. (3)
This forecasting model is sufficiently parsimonious and delivers large explanatory power in
the forecast of future volatility, mainly thanks to the high-frequency measurement of realized
volatility and the degree of persistence imposed by the use of implied volatility. More involved
7As a robustness check, we also obtain high-frequency five-minute returns for the shortest maturity futurescontracts written on the same indices. The measures of realized variance obtained from futures returns arevirtually indistinguishable from the ones obtained from the underlying stock index returns.
14
specifications with further lags and volatility measured on different horizons improve the explanatory
power only marginally.
In Panel B of Table 5, we examine the volatility risk premium. The growth index and
its dispersion have explanatory power for the volatility risk premium much in the same way
they explain the VIX index. This evidence suggests that the real-time macro conditions affect
investor preferences and more specifically the compensation that investors want to receive to
take on volatility risk.
In Panel C of Table 5, we now examine the expected realized volatility. In this case, the
results are somewhat different, with growth dispersion still explaining a comparable fraction of
expected realized volatility, but with the state of the economy explaining almost half of a share
compared to the VIX index or the VRP. This analysis suggests that the state of the economy
affects preferences and, to a lesser extent, the volatility of stock returns, while macroeconomic
uncertainty has comparable influence on both preferences and realized volatility.
4.4 Forecasting the real-time growth factor
We now explore whether growth and macroeconomic uncertainty have predictive power for the
future levels of the growth factor. We plot in Figurefs 11 the median change in the growth index
at different horizons unconditionally and following states of growth dispersion above/below
median and in the top/bottom quartile. It is clear from the plot that periods of high dispersion,
and especially those in the top quartiles, are followed by large increases in the growth index in
the following days and months.
Table 6 confirms this graphical observation. Both current growth and dispersion have
statistical significant forecasting power for future growth at different horizons. While the role
of current growth in predictability is not surprising at short horizons given the persistence in
the growth factor, the significance of dispersion is intriguing. The dispersion index is already
significant at the five-day horizon and maintains its forecasting ability with increasing horizons
up to six months. The level of growth today is also strikingly important for growth up to six
months later.
4.5 Extending the sample backwards
Our sample period is constrained by the availability of real-time news releases, that is, macroeconomic
announcements for which the exact release day and time together with the released figure figure
was available. In this section, we extend our sample backward to the beginning of 1985 using
the median reporting lag for each news item and inferring the release date. Wherever possible,
we also try to avoid revisions/restatements of the original releases.
To have a better sense for the accuracy of our procedure of backward-filling based on
reporting lags, we partially cross-check our inferred release dates with a database of Reuters
news. More specifically, for a subsample of 15 announcements on the total of 43 considered
15
news items and for a shorter sample period going back to 1990, we find that 91 percent of the
estimated release days are less than two days off from the actual release days.
While in this empirical exercise, the real-time feature of our approach is weakened and thus
there is less value from an investment strategy point of view, a longer sample is desirable to
understand the dynamics of the macro factors in different economic environments. In Figure 12,
we plot our growth indicator together with NBER ex-post determined expansions and recessions
as gray-shaded areas and the expectations for current GDP growth in the quarterly Survey of
Professional Forecasters. During the recession period in 1990-1991, we observe the sudden drop
in the growth index and the subsequent gradual recovery. Furthermore, we still observe a very
large correlation with the low-frequency growth expectations in the SPF, suggesting that our
growth factor construction procedure can still be interpreted as a high-frequency reading of
professional forecaster macroeconomic views. In summary, we have suggestive indications here
that our estimate of the growth factor, even with a less reliable identification of the release day,
is effectively picking up well U.S. growth dynamics.
It would be nice to extend the sample series for our proxy for macroeconomic uncertainty
too. Unfortunately, the economist forecasts we use for the construction of our disagreement
measure are not available before 1997. For illustration purposes, we apply our methodology
to the five disagreement measures about growth in the economy (GDP, corporate profits,
employment, unemployment, and productivity growth) released quarterly in the Survey of
Professional Forecasters. After 1997, we complement these five low-frequency series with the
economist disagreement about upcoming scheduled growth releases.
Figure 13 shows the results. We first notice a large correlation between our measure of
macroeconomic uncertainty and the quarterly measure of disagreement from the SPF, meaning
that we can interpret our measure of dispersion as a high-frequency reading of professional
forecasters disagreement. Furthermore, we do notice again also in the earlier part of the sample
that uncertainty seems to peak on the way out of recessions and is subdued at the end of
expansion phases, confirming the pattern observed in the last 15 years.
5 Conclusions
We use a novel and extremely simple technique to extract daily latent factors from macroeconomic
releases available at different times and frequencies. This approach can effectively handle the
large cross-section of announcements that are relevant to track macroeconomic conditions. Our
methodology is implemented in real-time and is used to extract measures of inflation, output,
employment, and macroeconomic sentiment, as well as corresponding measures of disagreement
among economists about these indices.
We find that our procedure improves on existing approaches as it provides a more accurate
and timely measurement of the state of the economy. In contrast to the extant literature,
the real-time growth factor and its disagreement constructed using our approach using only
macroeconomic information explains a remarkable fraction of financial volatility dynamics.
16
6 Appendix: Macroeconomic News
The following table summarizes the main features of the macroeconomic releases considered
in our sample. Category is either inflation (INF), employment (EMP), output (OUT), or
sentiment (SEN). If the sample series is stationary in our sample, we make no adjustment
(Adj.=0), otherwise we use first differences with respect to previous period (Adj.=1). We also
indicate Units, Frequency (M for monthly, W for weekly, Q for quarterly), and the source of
the release.
Category Name Macro Release Adj. Units Freq Source
INF US Import Price Index by End Use All MoM 0 Rate M Bureau Labor Statistics
INF US PPI Finished Goods Total MoM 0 Rate M Bureau Labor Statistics
INF US PPI Finished Goods Except Foods Energy 0 Rate M Bureau Labor Statistics
INF US CPI Urban Consumers MoM 0 Rate M Bureau Labor Statistics
INF US CPI Urban Consumers Less Food Energy 0 Rate M Bureau Labor Statistics
INF BLS Employment Cost Civilian Workers QoQ 0 Rate Q Bureau Labor Statistics
INF US GDP Price Index QoQ SAAR 0 Rate Q Bureau Economic Analysis
INF US Personal Cons. Expenditure Core Price Index MoM 0 Rate M Bureau Economic Analysis
INF US Output Per Hour Nonfarm Business Sector QoQ 0 Rate Q Bureau Labor Statistics
EMP ADP National Employment Report Private Nonfarm Change 0 Volume M Automatic Data Processing
EMP US Initial Jobless Claims 1 Volume W Department of Labor
EMP US Continuing Jobless Claims 1 Volume W Department of Labor
EMP US Employees on Nonfarm Payrolls Total Net Change 0 Value M Bureau Labor Statistics
EMP US Employees on Nonfarm Payrolls Manufact Net Change 0 Value M Bureau Labor Statistics
EMP US Unemployment Rate Total in Labor Force 1 Rate M Bureau Labor Statistics
EMP US Average Weekly Hours All Total Private 1 Volume M Bureau Labor Statistics
OUT ISM Manufacturing PMI 0 Value M Institute Supply Management
OUT US Manufacturers New Orders Total MoM 0 Rate M U.S. Census Bureau
OUT US Auto Sales Domestic Vehicles 1 Volume M BLOOMBERG
OUT ISM Non-Manufacturing NMI NSA 0 Value M Institute Supply Management
OUT Federal Reserve Consumer Credit Net Change 1 Value M Federal Reserve
OUT Merchant Wholesalers Inventories Change 0 Rate M U.S. Census Bureau
OUT Adjusted Retail Food Services Sales Change 0 Rate M U.S. Census Bureau
OUT Adjusted Retail Sales Less Autos Change 0 Rate M U.S. Census Bureau
OUT US Industrial Production MoM 2007=100 SA 0 Rate M Federal Reserve
OUT US Capacity Utilization of Total Capacity 0 Rate M Federal Reserve
OUT US Manufacturing Trade Inventories Total 0 Rate M U.S. Census Bureau
OUT US Durable Goods New Orders Industries 0 Rate M U.S. Census Bureau
OUT US Durable Goods New Orders Ex Transp. 0 Rate M U.S. Census Bureau
OUT GDP US Chained 2005 Dollars QoQ SAAR 0 Rate Q Bureau Economic Analysis
OUT GDP US Personal Consumption Chained Change 0 Rate Q Bureau Economic Analysis
OUT US Personal Income MoM 0 Rate M Bureau Economic Analysis
OUT US Personal Consumption Expend. Nominal Dollars 0 Rate M Bureau Economic Analysis
SEN Bloomberg US Weekly Consumer Comfort Index 1 Price W BLOOMBERG
SEN University Michigan Survey Consumer Confidence 1 Price M U. of Michigan Survey Research
SEN Empire State Manufact. Survey Business Conditions 1 Value M Federal Reserve
SEN Conference Board US Leading Index MoM 0 Rate M Conference Board
SEN Philadelphia Fed Business Outlook General Conditions 1 Price M Philadelphia Fed
SEN Conference Board Consumer Confidence SA 1985=100 1 Rate M Conference Board
SEN Richmond Fed Reserve Manufacturing Survey 0 Rate M Richmond Fed
SEN US Chicago Purchasing Managers Index SA 1 Price M Kingsbury Intern.
SEN ISM Milwaukee Purchasers Manufacturing Index 1 Rate M NAPM - Milwaukee
SEN Dallas Fed Manufact. Outlook Business Activity 1 Rate M Dallas Fed
17
18
Tab
le1:
Su
mm
ary
Sta
tist
ics
Pan
elA
show
sco
rrel
atio
ns
bet
wee
nd
iffer
ent
aggre
gate
sof
macr
oec
onom
icn
ews,
as
econ
om
icin
dic
esor
fore
cast
dis
per
sion
.P
an
elB
rep
ort
s
add
itio
nal
sum
mar
yst
atis
tics
and
corr
elat
ion
sb
etw
een
the
gro
wth
ind
ex,
the
gro
wth
dis
per
sion
,an
da
set
of
fin
an
cial
vari
ab
les.
Sp
ecifi
call
y,
Rm
t−R
ft
den
otes
the
log
retu
rnon
the
S&
P500
inex
cess
of
the
3-m
oT
-bil
lra
te,IV
isth
eV
IXin
dex
,ERV
isth
eex
pec
tati
on
of
reali
zed
vola
tili
tyb
ased
ona
tim
e-se
ries
mod
elw
ith
lagged
IV
an
dre
ali
zed
vola
tili
tyco
nst
ruct
edfr
om
five
-min
ute
retu
rns,VRP
isth
eva
rian
ceri
sk
pre
miu
mdefi
ned
asVIX
2−ERV
2,
logP
t
Et
and
logD
t
Pt
are
the
log
of
the
pri
ce-e
arn
ing
rati
oth
ediv
iden
dyie
ld,DEF
isth
ed
efau
ltsp
read
(Mood
y’s
BA
A-
AA
Aco
rpor
ate
bon
dyie
lds)
,TERM
isth
ete
rmsp
read
(10-y
ear
-3-m
oT
reasu
ryyie
lds)
.T
he
sam
ple
per
iod
exte
nd
sfr
om
Jan
uar
y19
97to
Dec
emb
er20
11.
19
Pan
elA
:
Ind
exD
isp
ersi
on
Infl
Sent
EconAct
Growth
Infl
Sent
EconAct
Growth
Ind
ex:
Infl
1.00
0.1
40.2
50.2
2-0
.08
-0.1
8-0
.19
-0.2
0
Sent
1.0
00.8
20.9
3-0
.50
-0.0
70.0
60.0
5
EconAct
1.0
00.9
7-0
.55
-0.1
8-0
.23
-0.2
3
Growth
1.0
0-0
.56
-0.1
5-0
.13
-0.1
4
Dis
per
sion
:
Infl
1.0
00.1
40.2
50.2
5
Sent
1.0
00.4
20.5
4
EconAct
1.0
00.9
9
Growth
1.0
0
20
Pan
elB
:
Rm
t−R
ft
GRO
DISP
VRP
VIX
RV
log
(P E
)lo
g(D P
)DEF
TERM
Su
mm
ary
stati
stic
s
Mean
0.63
-0.0
4-0
.00
0.0
40.2
30.1
42.9
80.5
51.0
31.6
8
Std
ev.
21.4
21.
150.9
90.0
50.0
90.0
50.2
30.2
50.4
81.3
0
Skew
-0.2
0-1
.30
1.4
04.9
61.7
80.9
70.0
50.4
82.8
2-0
.06
Ku
rt9.
774.
974.1
936.5
78.8
54.3
82.1
93.5
612.2
51.6
6
Corr
elati
on
matr
ix
Rm
t−R
ft
1.00
0.01
0.0
1-0
.13
-0.1
3-0
.09
0.0
4-0
.03
-0.0
10.0
1
GRO
1.00
-0.1
4-0
.46
-0.5
1-0
.37
0.5
5-0
.71
-0.8
4-0
.51
DISP
1.0
00.1
50.2
50.2
00.1
90.0
90.1
40.0
2
VRP
1.0
00.8
80.3
4-0
.30
0.4
00.6
50.1
9
VIX
1.0
00.7
0-0
.14
0.3
10.6
30.2
2
RV
1.0
00.1
10.0
70.3
90.1
2
log
(P/E
)1.0
0-0
.85
-0.5
2-0
.28
log
(D/P
)1.0
00.6
90.4
2
DEF
1.0
00.4
0
TERM
1.0
0
21
Table 2: Comparing CFNAI with our Real-Time Growth Index
This table shows estimates of the following equation
Yt = α+ β1Yt−1 + β2Xt−1 + εt,
where Yt is the CFNAI index and Xt is our real-time growth index (Panel A) and Yt is our real-time growth
index and Xt is the CFNAI index (Panel B). Robust Newey-West t-statistics are reported in parentheses.
Panel A: CFNAI on Real-Time Growth Index
Dep. Var.: CFNAI 1 2
Constant -0.03 -0.08
(-1.89) (-2.08)
CFNAIt−1 0.91 0.41
(17.30) (3.68)
REALTIMEt−1 0.39
(4.25)
Adj.R2(%) 83.56 86.86
Panel B: Real-Time Growth Index on CFNAI
Dep. Var.: REALTIME 1 2
Constant -0.01 -0.01
(-0.74) (-0.48)
REALTIMEt−1 0.96 0.93
(24.79) (13.31)
CFNAIt−1 0.05
(0.52)
Adj.R2(%) 91.77 91.72
22
Table 3: Comparing ADS with our Real-Time Growth Index
This table shows estimates of the following equation
Yt = α+ β1Yt−1 + β2Xt−1 + εt,
where Yt is the ADS index and Xt is our real-time growth index (Panel A) and Yt is our real-time growth
index and Xt is the ADS index (Panel B). Robust Newey-West t-statistics are reported in parentheses.
Panel A: ADS on Real-Time Growth Index
Dep. Var.: ADS 1 2
Constant -0.02 0.04
(-1.78) (2.11)
ADSt−1 0.95 0.76
(52.98) (21.98)
REALTIMEt−1 0.12
(4.90)
Adj.R2(%) 91.43 92.18
Panel B: Real-Time Growth Index on ADS
Dep. Var.: REALTIME 1 2
Constant 0.01 0.01
(0.41) (0.78)
REALTIMEt−1 0.99 1.01
(165.45) (68.46)
ADSt−1 -0.02
(-1.18)
Adj.R2(%) 99.61 99.61
23
Table 4: Forecasting GDP
This table shows estimates of the following equation:
GDPt = α+ β1GDPt−1Q + β2Xt−2mo + εt,
where GDPt are actual GDP quarterly releases, GDPt−1Q is the previous GDP quarterly release, and Xt−2mo
is either the average GDP nominal growth quarterly forecast of the Survey of Professional Forecasters (SPF)
or our real-time growth factor, both observed on the same day about two months before the GDP release.
Robust Newey-West t-statistics are reported in parentheses.
Dep. Var.: GDPt 1 2 3 4
Constant -0.61 -0.15 -0.11 -0.08
(-2.75) (-0.99) (-0.60) (-0.53)
GDPt−1Q 0.22 0.06 0.04 0.03
(4.14) (1.10) (0.68) (0.57)
SPFt−2mo 0.57 0.44
(3.25) (1.68)
GROWTHt−2mo 0.58 0.20
(2.73) (0.75)
Adj.R2(%) 30.28 44.60 41.18 44.25
24
Table 5: The Growth Factor, its Dispersion, and VIX, VRP, E(RV)
This table shows estimates of the following equation
Yt = α+ β1DISPt + β2GROWTHt + εt,
where Yt is either V IXt (Panel A), V RPt (Panel B) and Et [RVt+1,t+22] (Panel C), respectively. In column
(3), each of the variable is orthogonalized with respect to the other. All of the regression are based on
daily observations. The sample period extends from January 1997 to December 2011. Robust Newey-West
t-statistics are reported in parentheses.
Panel A: VIX1 2 3
Constant 0.2269 0.2252 0.2253
(34.4636) (39.4552) (40.7793)
DISPt 0.0217 0.0221
(3.4952) (4.8659)
GROWTHt -0.0387 -0.0394
(-5.7853) (-5.8185)
Adj.R2(%) 5.9840 25.6984 28.8520
Panel B: VRP1 2 3
Constant 0.0582 0.0571 0.0571
(13.9758) (16.6792) (16.9750)
DISPt 0.0102 0.0104
(2.7046) (3.9473)
GROWTHt -0.0245 -0.0249
(-4.4586) (-4.4493)
Adj.R2(%) 3.2898 25.6129 26.9065
25
Panel C: ERV1 2 3
Constant 0.0304 0.0302 0.0303
(38.2076) (40.4112) (41.1817)
DISPt 0.0022 0.0023
(2.8041) (3.2592)
GROWTHt -0.0035 -0.0036
(-4.7526) (-4.9391)
Adj.R2(%) 4.0389 13.6034 15.9232
26
Table 6: Growth Predictability Regressions
This table shows estimates of the following equation:
GROWTHt+lead = α+ ρGROWTHt + βDISPt + εt,
where Growtht is the real-time growth factor and DISPt denotes the dispersion of analyst expectations about
macroeconomic growth releases. All of the regression are based on daily observations. We do not report
α to save space. The sample period extends from January 1997 to December 2011. Robust Newey-West
t-statistics are reported in parentheses.
lead (days) 5 20 40 60 80 100 120
Growtht 0.9954 0.9737 0.9383 0.8956 0.8456 0.7943 0.7422
(289.8041) (72.7546) (32.6357) (20.0954) (14.3388) (11.1513) (9.1673)
DISPt 0.0097 0.0417 0.0843 0.1179 0.1576 0.1902 0.2188
(3.1894) (3.6303) (3.8682) (3.8856) (3.9698) (4.0097) (3.9724)
Adj.R2(%) 98.9051 94.1737 87.0181 79.1716 70.7449 62.7566 55.4122
Adj.R2(%)AR1 98.8985 94.0485 86.5029 78.1608 68.9290 60.1036 51.8990
27
References
Ait-Sahalia, Yacine, Mykland, Per A., and Zhang, Lan, 2005, How often to sample a continuous-time process in the presence of market microstructure noise, Review of Financial Studies 18,351–416.
Andersen, Torben G., Tim Bollerslev, Francis X. Diebold, and Clara Vega, 2003, Micro effectsof macro announcements: Real-time price discovery in foreign exchange, American EconomicReview 93, 38–62.
Andersen, Torben G., Tim Bollerslev, Francis X. Diebold, and Clara Vega, 2005, Real-timeprice discovery in stock, bond and foreign exchange markets, Working Paper, NorthwesternUniversity, Duke University, University of Pennsylvania and University of Rochester.
Aruoba, S.B., Diebold, F.X. and Scotti, C., 2009, Real-Time Measurement of BusinessConditions, Journal of Business and Economic Statistics, 27, 417–427.
Banbura, Marta, Giannone, Domenico, Modugno, Michele and, Lucrezia Reichlin, 2012, Now-casting and the real-time data flow, CEPR Discussion Papers 9112.
Baker, Malcolm, Wurgler, Jeffrey and Yu Yuan, 2012, Global, local, and contagious investorsentiment, Journal of Financial Economics 104, 272–287.
Beber, Alessandro, Michael W. Brandt, and Kenneth A. Kavajecz, 2011, What Does EquitySector Orderflow Tell Us about the Economy?, Review of Financial Studies 24, 2011, 3688-3730.
Bollerslev, Tim, George Tauchen, and Hao Zhou, 2009, Expected Stock Returns and VarianceRisk Premia, Review of Financial Studies 22, 4463–4492.
Corsi, Fulvio, 2009, A Simple Approximate Long-Memory Model of Realized Volatility, Journalof Financial Econometrics 7, 174–196.
Corsi Fulvio, Nicola Fusari, and Davide La Vecchia, 2012, Realizing Smiles: Pricing Optionswith Realized Volatility, Journal of Financial Economics forthcoming.
Drechsler, Itamar, and Amir Yaron, 2011, What’s Vol Got To Do With It, Review of FinancialStudies 24, 1–45.
Engle, Robert F., and Jose Gonzalo Rangel, 2008, The Spline-GARCH Model for Low-FrequencyVolatility and Its Global Macroeconomic Causes, Review of Financial Studies 21, 1187–1222.
Evans, Martin, 2005, Where Are We Now?: Real-Time Estimates of the Macro Economy, TheInternational Journal of Central Banking.
Ghysels, Eric, Casidhe Horan, and Emanuel Moench, 2012, Forecasting through the Rear-ViewMirror: Data Revisions and Bond Return Predictability, Working Paper, University of North-Carolina.
Giannone, Domenico, Reichlin, Lucrezia, and David, Small, 2008, Nowcasting: the real timeinformational content of macroeconomic data releases, Journal of Monetary Economics 55,665–676.
Koenig, E., S. Dolmas, and J. Piger, 2003, The use and abuse of real-time data in economicforecasting, Review of Economics and Statistics 85, 618–628.
28
Mueller, Philippe, Andrea Vedolin, and Yu-Min Yen, 2012, Bond Variance Risk Premia, WorkingPaper, London School of Economics.
Newey, Whitney K.; West, Kenneth D., 1987, A simple, positive semi-definite, heteroskedasticityand autocorrelation consistent covariance matrix, Econometrica 55, 703–708.
Stock, J.H. and M.W. Watson, 1989, New Indexes of Coincident and Leading EconomicIndicators, in O.J. Blanchard and S. Fischer (eds.), NBER Macroeconomics Annual, 352–394.
Schwert, G William, 1989, Why Does Stock Market Volatility Change over Time?, Journal ofFinance 44, 1115–1153.
Stambaugh, Robert F., 1997, Analyzing investments whose histories differ in length, Journal ofFinancial Economics 45, 285–331.
Stiglitz, J., J. Fitoussi and A. Sen, 2010, Mismeasuring Our Lives: Why GDP Doesnt Add Up,New York, The New Press.
29
Conference Board Consumer ConfidenceChicago Purchasing Managers Index Inflation
University Michigan Consumer Survey EmploymentADP National Employment Report Output
ISM Manufacturing PMI Macro SentimentNonfarm Payrolls Total,Manufacturing + Unemployment Rate + Average Weekly HoursISM Non-Manufacturing PMI
Retail Sales + Retail Sales Less AutoImport Price Index
PPI + PPI CoreIndustrial Production + Capacity UtilizationEmpire State Manufacturing Survey
Manufacturing Trade InventoriesCPI + CPI Core
Durable Goods OrdersConference Board Leading Index
GDP + GDP Price IndexPersonal Income + Pers. Consum. Exp. + PCE Price Index
Manufacturers New Orders
24 26 28 30 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 1 3 5 7 9 11 13 15 17 19 21 23referencemonth M month M+1 month M+2
Figure 1: This figure shows the typical reporting structure for a large cross-section ofU.S. macroeconomic announcements. On the horizontal axis, we represent the days of thereference month M and the subsequent two months. On the vertical axis, we list themacroeconomic releases in order of reporting, highlighting in bold the typical reporting period.The macroeconomic announcements are color-coded in the four aggregates of inflation news,employment news, output news, and macro-sentiment news.
30
j=1 j=2 … j=5 j=6 … j=N1 ... ... ... missing ... ... ...
… ... ... ... missing ... ... ...
… ... ... ... missing ... ... ...
t-22 At-22,1 not released ... missing not released ... ...
t-21 not released At-21,2 ... missing At-21,6 ... ...
… not released not released ... missing not released ... ...
t At,1 not released ... At,5 not released ... ...
t+1 not released At+1,2 ... not released At+1,6 ... ...
… not released not released ... not released discontinued ... ...
… ... ... ... ... discontinued ... ...
T ... ... ... ... discontinued ... ...
j=1 j=2 … j=5 j=6 … j=N1 ... ... ... missing ... ... ...
… ... ... ... missing ... ... ...
… ... ... ... missing ... ... ...
t-22 At-22,1 E[At-22,2]=At-43,2 ... missing E[At-22,6]=At-43,6 ... ...
t-21 E[At-21,1]=At-22,1 At-21,2 ... missing At-21,6 ... ...
… E[A ...,1]=At-22,1 E[A...,2]=At-21,2 ... missing E[A...,2]=At-21,6 ... ...
t At,1 E[A t,2]=At-21,2 ... At,5 E[A t,2]=At-21,6 ... ...
t+1 E[At+1,1]=At,1 At+1,2 ... E[At+1,5]=At,5 At+1,6 ... ...
… E[A ...,1]=At,1 E[A ...,2]=At+1,2 ... E[A ...,5]=At,5 discontinued ... ...
… ... ... ... ... discontinued ... ...
T ... ... ... ... discontinued ... ...
Figure 2: This figure shows a stylized example of the actual macroeconomic announcement data,for N announcement types over a daily sample period between 1 and T . The releases j = 1 andj = 2 are monthly indicators released on two different days of the month. The macroeconomicindicator j = 5 is a news release that did not exist at the beginning of the sample, but wasincluded in the sample from day t onwards. The macroeconomic indicator j = 6 did exist atthe beginning of the sample, but was subsequently discontinued. The top panel represents thematrix of the actual macroeconomic releases in real-time as it is constructed from the data. Thebottom panel shows how our simple forward filling algorithm is used to fill in the expectationof the indicator when it is not released.
31
1997 1999 2001 2003 2005 2007 2009 2011-5
-4
-3
-2
-1
0
1
2Output and Employment
Figure 3: The graph shows the real-time output (blue thicker line) and employment factor (redline). Grey areas denote NBER recessions.
32
1997 1999 2001 2003 2005 2007 2009 2011
-4
-3
-2
-1
0
1
2Economic Activity and Macro Sentiment
1997 1999 2001 2003 2005 2007 2009 2011
-5
-4
-3
-2
-1
0
1
2
3
Growth and Inflation
Figure 4: The upper panel shows the real-time economic activity (blue thicker line) and macrosentiment (red line) factors. The lower panel plots the real-time growth (blue thicker line) andinflation (red line) factors. Grey areas denote NBER recessions.
33
1997 1999 2001 2003 2005 2007 2009 2011
-4
-3
-2
-1
0
1
2Growth Index
1997 1999 2001 2003 2005 2007 2009 2011-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
3
3.5Dispersion
Figure 5: The upper panel shows the real-time growth factor. The lower panel is the dispersion ofeconomist’ forecasts about upcoming growth news releases. Grey areas denote NBER recessions.
34
2001 2003 2005 2007 2009 2011-4
-3
-2
-1
0
1
2CFNAI and Growth Index
CFNAIGrowth
Figure 6: This figure shows the real-time growth factor GROWTH and the real-time CFNAIindex, both observed at the monthly frequency on the same day during the sample periodFeb-2001 to Dec-2011.
35
2008.12 2009.6 2009.12 2010.6 2010.12 2011.6 2011.12-3
-2.5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2ADS and Growth Index
ADSGrowth
Figure 7: This figure shows our real-time growth factor GROWTH and the correspondingvintages of the ADSindex during the sample period Dec-2008 to Dec-2011.
36
1997 1999 2001 2003 2005 2007 2009 2011
-4
-3
-2
-1
0
1
2
3GDP, SPF, and real-time Growth Index
GDPactSPFGROWTH
Figure 8: In this figure we plot the estimate of the latent growth factor GROWTH, the medianprojection of nominal GDP growth rate from the Survey of Professional Forecasters (SPF) onthe same dates, and the actual GDP release for the same quarter, at quarterly frequency duringthe sample Jan-1997 to Dec-2011.
37
1997 1999 2001 2003 2005 2007 2009 2011
-4
-2
0
2
GR
OW
TH (i
nver
ted
scal
e)
1997 1999 2001 2003 2005 2007 2009 20110.05
0.2
0.35
0.5
0.65
VIX
Figure 9: In this figure we plot the estimate of the latent growth factor GROWTH (invertedleft-scale) and VIX (right-scale) during our sample period.
38
1997 1999 2001 2003 2005 2007 2009 2011-2
-1
0
1
2
DIS
P
1997 1999 2001 2003 2005 2007 2009 20110.05
0.2
0.35
0.5
0.65
VIX
Figure 10: In this figure we plot the dispersion of economists’ forecast about growth newsreleases (left-scale) and VIX (right-scale) during our sample period.
39
0 10 20 30 40 50 60 70 80 90-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
horizon (days)
Forecasting Growth Index Returns
aboveMEDbelowMEDtopQbotQunconditional
Figure 11: This figure shows the median first difference in the growth index for different horizons(in days), unconditionally and conditionally on current dispersion about the growth rate beingabove (below) the sample median and in the top (bottom) quartile.
40
1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011-5
-4
-3
-2
-1
0
1
2
Growth (1985-2011)
Figure 12: This figure shows our estimated growth index for the U.S. economy constructed oneconomic releases backfilled to January 1985. The red line indicates the quarterly expectation ofGDP growth for the current quarter contained in the Survey of Professional Forecasters. Greyareas denote NBER recessions.
41
1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011-2
-1
0
1
2
3
4
Dispersion and SPF Disagreement (1985-2011)
Figure 13: This figure shows our daily factor for the dispersion of economist growth forecasts(in red) and the disagreement measure in the quarterly Survey of Professional Forecasters (inblue). Grey areas denote NBER recessions.
42