Realized and Anticipated Macroeconomic Conditions Forecast

53 Pages·2014·0.94 MB·English

Realized and Anticipated Macroeconomic
Conditions Forecast Stock Returns∗
Alessandro Beber†, Michael W. Brandt‡, Maurizio Luisi§
November 2014
Abstract
We construct daily real-time indices capturing the public information on realized and
anticipated economic activity. The one-month change in realized fundamentals predicts
U.S.stockreturnsacrosshorizonswithstrongestresultsbetweenamonthandaquarter.
The information in anticipated fundamentals that is orthogonal to the realized data
predictsreturnsevenmorestrongly,particularlyatlongerhorizonsofuptotwoquarters.
Splittingthesampleintotimesofhighversuslowuncertainty, asmeasuredbythecross-
sectional dispersion of economist forecasts, we show that the predictability is largely
concentrated in high-uncertainty times. Finally, extending the analysis internationally,
we find similar results that are curiously stronger when U.S. data are used as predictors
rather than global composites or local data.
Keywords: stock market predictability, state of the economy, macroeconomic uncertainty.
JEL classification: G12
∗Earlier versions of this paper were circulated under the title “Economic Cycles and Expected Stock Returns.”
We thank Inquire UK for financial support. We thank Daryl Caldwell, Robert Darwin, Fabio Fornari, Amit Goyal,
Ana-MariaTenekedjieva,andseminarparticipantsatBlackRock,CityUniversity,PanAgora,the2012AssetPricing
Retreat at Cass Business School, the Fall 2012 Inquire UK Conference in Bath, the Imperial College Hedge Fund
Conference, the London Quant Group Conference, the Stockholm School of Economics, and the University of York,
for their comments and suggestions.
†Cass Business School, City University London, and CEPR
‡Fuqua School of Business, Duke University, and NBER
§Bloomberg L.P.
1 Introduction
The risk premium on stocks varies, both through time and across countries. Characterizing this
variationempiricallyandmodelingtheunderlyingeconomicmechanismstheoreticallypreoccupiesa
substantialpartoftheacademicfinanceprofession. Despitesharingabroadobjective, however, the
empiricalandtheoreticalliteraturescanappeardisjointed. Empiricalpaperstendtouseforecasting
variables whose variation is primarily driven by financial markets (e.g., valuation ratios, interest
rates, option implied volatilities), whereas theoretical models are based on economic fundamentals
(e.g., GDP growth, consumption, inflation). This disconnect arises because the fundamental data
empiricists observe is infrequent, backward-looking, and often restated after the initial release,
whereasmarketbasedpredictorscanbemeasureddailyandreflectcurrentforward-lookinginformation.
Empiriciststhereforetendtofavormarketbasedpredictorsasimplicitproxiesforeconomicfundamentals,
butmarketpricesalsoreflectotherthings,suchasaggregatepreferencesandpotentialmisvaluation.
We construct daily measures of economic activity based on the almost continuous flow of
macroeconomic data releases. Our measures are meant to summarize in real-time the public
informationabouttheeconomyavailabletomarketparticipants. Weparticularlyfocusoninformation
about economic growth (abstracting from information about inflation, housing, trade, and the
public sector) and explicitly differentiate between (i) realized ex-post measures, such as quarterly
GDP releases, and (ii) anticipating ex-ante information, namely data from surveys of consumers
and firm managers. The resulting economic indices overcome some of the concerns about using
economic data for stock return predictability: the indices are measured daily and capture the
information of the entire news flow, not just a few select data series; they are both backward- and
forward-looking, so we can potentially capture leading information as market variables do; and
they are based on carefully dated and unrestated data to alleviate concerns about look-ahead bias
through restatements. We use these real-time economic indices to reexamine whether and, if so,
to what extent stock market returns are predictable by economic fundamentals through time and
across four major economies.
We find that the one-month change in realized growth, constructed from backward-looking
and delayed data, still predicts U.S. stock market returns one to four months into the future. In
addition, the part of anticipated growth, constructed from forward-looking data, that is orthogonal
to realized growth predicts market returns at horizons of two to six months. Due to the orthogonal
construction of the two predictors, the results are roughly additive in a multivariate regression,
leading to return predictability that far exceeds what is generated by the usual suspects: valuation
ratios (D/P and E/P), interest rate spreads (term and default), and the difference between option-
1
implied and realized stock market volatility.1 Moreover, since realized and orthogonal anticipated
growth predict returns at different horizons – the realized growth results are strongest at one and
two month return horizons whereas the orthogonal anticipated growth results are strongest at
one and two quarter horizons – the joint predictions are substantially improved across horizons
compared to univariate specifications.
Our second set of empirical results shows that the predictability by economic fundamentals,
both realized and orthogonal anticipated, is state dependent. The results are much stronger during
timesofgreaterdisagreementamongeconomistsabouteconomicgrowth. Wemeasuredisagreement
as the weighted average, across economic news series, of the cross-sectional, across economists,
standard deviation of forecasts, where the weights are the same as for our real-time growth index,
reflecting the relative importance and correlation structure of different economic series. We find
that this disagreement measure is highly counter-cyclical, so that return predictability is also much
stronger in economic recessions than in expansions, where we define recessions as below average
growth periods. When we combine greater disagreement and being in a recession, the conditional
predictability is further strengthened, suggesting that each conditioning contains at least some
separate information. For example, unconditionally the bivariate specification has an R2 of 2.6
percent at the monthly horizon. Conditioning on greater disagreement or on being in a recession
raises the R2 to 6.5 and 5.1 percent, respectively. Conditioning on both further raises it to 10.3
percent.
We extend the evidence to international data, specifically equity indices in Europe, Japan, and
the U.K.. We find that economic fundamentals predict equity returns as strongly if not more so
internationally. More interestingly, though, at least for Japan and the U.K., global aggregates and
even just the U.S. factors are more predictive than local versions. For Europe, both are relevant.
This finding is consistent with those of Ang and Bekaert (2007) and Bollerslev et al. (2012), who
also find global aggregates to have stronger forecasting power than local predictors. To quantify
the economic significance of these international results, we construct a global equity market timing
and country selection portfolio based on realized and anticipated global growth. Consistent with
the strong statistical evidence, the Sharpe ratio of this investment strategy is 0.7 annualized, as
compared to a Sharpe ratio close to zero for being long world equities over the same time period.
The value of forecasting with economic fundamentals is as strong during the first half of our sample
as during the second, suggesting that our results are not entirely driven by the 2008-2009 financial
crisis when global economic fundamentals fluctuated the most since the 1930s. Finally, we find that
1Thefollowingisapartiallistofacademicpapersthatdocumentvariousdegreesofreturnpredictabilityandthe
variables they use: Bollerslev et al. (2009, 2012), variance risk premium; Campbell (1987), term spread; Campbell
and Shiller (1988a, 1988b), dividend yield; Cochrane (2008), dividend yield; Fama and French (1988, 1989), default
spread,dividendyield,termspread;FamaandSchwert(1977),Treasurybillyield;FersonandHarvey(1991),default
spread,dividendyield,laggedreturns,termspread,Treasurybillyield;KeimandStambaugh(1986),defaultspread,
trend; Lamont (1998), dividend-to-earnings ratio.
2
the international data corroborates our results on conditioning. In all three regions, the forecasting
ability of realized and orthogonal anticipated growth is much stronger during times of greater
economist disagreement and recessions.
Our paper blends two literatures: that on stock return predictability surveyed partially in
footnote1andthatonmeasuringthestateoftheeconomybasedoneconomicnewsdata,commonly
referred to as “nowcasting” (see Banbura et al., 2012, for a survey). There are two general
approaches to nowcasting. The first approach is to use a balanced panel regression, along the
lines of the seminal paper of Stock and Watson (1989), now the Chicago Federal National Activity
Index (CFNAI). This first approach uses a large set of news releases but results in a relatively low
measurement frequency because the econometrician has to wait for the panel to be complete before
the index can be constructed. The second approach to nowcasting is to model macroeconomic data
using a latent state-space model (e.g., the ADS business conditions index of Arouba et al., 2009).
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 economic data, but
this technique is impractical for large cross-sections of news releases.2 Our approach to nowcasting
uses a large cross-section of news data but, by forward-filling missing data and making appropriate
correlation matrix adjustments, still produces high frequency indicators. We show in Beber et
al. (2014) that the resulting real-time indices are highly correlated with the CFNAI and ADS
business condition index, but that they appear to be more timely and informative about future
economic fundamentals.
The two closest papers to ours are Ludvigson and Ng (2009) and Bai (2010). Both papers use a
large panel of macroeconomic variables, among other things, to forecast bond returns (Ludvigson
and Ng, 2009) and stock returns (Bai, 2010). Their emphasis is on return predictor selection
and combination by searching over a very large set of not only macroeconomic but also financial
variables. Our focus, in contrast, is on documenting the link between economic fundamentals
and equity returns. We do so by taking a two-step approach. We first extract real-time indices
that best capture the dynamics of economic fundamentals and only then examine how these factors
forecastfuturestockreturns. Weexplicitlyexcludefinancialvariablesfromthefactors,althoughwe
show that the predictability we uncover is not subsumed by the usual suspect financial predictors.
Finally, unlike the vast majority of predictability and nowcasting papers, we exclusively work with
precisely date- and time-stamped initial data releases, as opposed to restated macroeconomic data.
Ghysels et al. (2012) demonstrate the importance of using unrestated data in the case of bond
return predictability.
Thepaperproceedsasfollows. InSection2,wedescribethedataandcarryoutsomepreliminary
2For example, Evans (2005) only considers the set of different (preliminary, advance, and final) GDP releases.
Aroubaetal. (2009)constructtheirbusinessconditionindexusingfourindicatorsatdifferentfrequencies,including
a continuously observable financial markets variable.
3
analyses. Section 3 explains our methodology for constructing our real-time economic factors as
well as an empirical proxy for uncertainty surrounding those factors. We present our empirical
results in Section 4, and Section 5 concludes.
2 Data
2.1 Macroeconomic news and forecasts
We obtain data on the dates, release times, and actual released figures for 43 distinct U.S. macro-
economicannouncementscoveringtheperiodfromJanuary1997throughDecember2011,foratotal
of more than 8,000 announcements over about 3,800 business days.3 This data is obtained from
Bloomberg through the economic calendar screen (i.e., “ECO <Go>”), which provides precisely
time-stamped and unrestated announcement data.4 We also collect data on economist forecasts
for each announcement. Bloomberg surveys economists during the weeks prior to the release of
each major 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.5 We collect this
more problematic data from January 1990 through 1996 simply to construct an initial correlation
matrixestimate, whichisrequiredbyourmethodology(seeSection3). Inordertodatethereleases
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.
We complement the U.S. data with equivalent information for the Eurozone, the U.K., and
Japan. Specifically, we obtain information for 183 European macro releases, for 43 U.K. releases,
and for 45 Japanese releases, over the same sample period. The surveyed economist forecasts are
only available for a slightly shorter sample period, starting in June 1997 for Europe and the U.K,
3Weemphasizethefactthatweworkwithdistinctannouncementsbecausetherearealotmorethan43statistics
if we included multiple versions of essentially the same data released in the same economic report. For example, the
CFNAI uses 13 industrial production statistics, resulting in 20 percent of the index being determined by a single
release. In contrast, we include in our analysis only the headline month-over-month figure.
4Theimportanceofusingreal-timeversusfinaldatainmacroeconomicforecastinghasbeendiscussedextensively
in the literature (e.g., Koenig et al., 2003, or Ghysels et al., 2012).
5For example, there are monthly releases of quarterly GDP labeled “advance,” “preliminary” and “final” all
referring to the same quarter. Bloomberg’s historical field for GDP is dated according to the referenced quarter,
so that the advance release gets overwritten by the preliminary release, which in turn gets overwritten by the final
release. Historically only the final releases are stored.
4
and starting in May 2000 for Japan.
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
showsthatactualnewsreleasesoccurwithavarietyofdifferentlagswithrespecttothemonththey
are referencing. Furthermore, news on different indicators are frequently released simultaneously.6
For example, the employment report traditionally announced on the first Friday of the month
contains four different indicators: nonfarm payrolls, nonfarm payrolls in the manufacturing sector,
the 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 frequencies; e.g., GDP data are sampled quarterly, the nonfarm payrolls are released
monthly, initial jobless claims are sampled weekly, etc. These features of our large cross-section of
macroeconomic news releases generate a sparse matrix of data that our methodology will have to
takeup. TheAppendixdescribesindetailthesetofU.S., Europe, U.K., andJapanmacroeconomic
news in our sample, including their frequency, source, and units of measurement.
2.2 Market returns and other predictors
We collect daily returns data on four major equity indices: the S&P 500 index for the U.S., the
EURO STOXX 50 index for the Eurozone, the FTSE 100 index for the U.K., and the Nikkei 225
index for Japan. The sample period is January 1997 through December 2011. We compute daily
excess returns using corresponding LIBOR rates from Datastream.
We also collect daily data on a number of return predictors used in the literature. Specifically,
we obtain price-to-earnings (P/E) and dividend-to-price (D/P) ratios for the four equity indices
from Datastream. We calculate a default spread variable as the difference between Moody’s BAA
and AAA corporate bond spreads for U.S. issuers and use this variable as a proxy for the default
spread of the other countries as well. We also compute a term spread variable as the difference
between the ten-year and three-month zero coupon Government bond yields, using U.S., German,
U.K., and Japanese data obtained from Datastream.
Finally, following Bollerslev et al. (2009), we construct for each market a daily variance risk
premium VRP as the difference between the one-month ahead option implied variance and the
t
expectation of realized variance for the next month:
VRP = IV −E [RV ], (1)
t t t t,t+d
where d denotes the number of days in a month, IV denotes the implied variance from date t to
t
6For example, for the U.S., there was at least one data release on approximately 80 percent of days. Multiple
datareleasesoccurredmuchlessfrequently,onapproximately60percentofthedaysinthesample. Obviouslyinthe
case of Europe, with over four times as many data series, coincident releases occur much more frequently.
5
t+d, and RV = (cid:80)d RV is the one-month ahead realized variance. The notation reflects
t,t+d s=1 t
the fact that the implied variance is known at date t whereas the corresponding realized variance
is partially unknown until date t+d. Both variance measures are assumed to be expressed in the
same units of time – daily, monthly, or annualized.
For the implied variance IV , we use the square of the VIX volatility index for the U.S., of the
t
VDAX volatility index for the Eurozone, of the VFTSE volatility index for the U.K., and of the
VXJ volatility index for Japan. All of these volatility indexes are based on highly liquid equity
index options and are constructed by the exchanges with the same model-free calculation approach
of Britten-Jones and Neuberger (2000).
We construct daily realized variances RV as the sum of the five-minute squared returns over
t
normal trading hours of each market, as in Bollerslev et al. (2009). The required high frequency
data is obtained from Tickdata. We then forecast one month ahead realized volatility using the
information contained in both realized and implied volatility, as in Drechsler and Yaron (2011).
Specifically, we regress the future realized volatility from t to t + d on the current (i.e., date t)
realized and implied volatilities, the realized volatility from t−d to t−1, and the average implied
volatility over that same period:7
RV 1/2 = α+β RV 1/2+β IV 1/2+β RV 1/2 +β IV 1/2 +(cid:15) . (2)
t,t+d 1 t 2 t 3 t−d,t−1 4 t−d,t−1 t
WeestimatethecoefficientsofthisregressionbyOLSusingatrailingone-yearwindowsoastoavoid
any look-ahead bias in the following regression. 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. We also explored more involved specifications with further lags and volatility
measured over different horizons, but those specifications improve the explanatory power only
marginally.8
2.3 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
statisticalproceduretoobtainorthogonalizedfactorsthatareincreasinglydifficulttointerpretwith
theorderofthefactor,weimposeaspecificeconomicallymotivatedstructureonthemacroeconomic
newsflow. Basedonbothempiricalevidenceandeconomicrationale,wefirstseparatetheaggregate
7Usinglaggedvolatilitytermsmeasuredondifferenthorizonsisconsistentwiththeheterogeneousauto-regressive
volatility forecasting models proposed by Corsi (2009) and used in Corsi, Fusari, La Vecchia (2013) and Mueller,
Vedolin, and Yen (2012).
8Our realized volatility forecasting model has an average (or median) R2 of 37 (33) percent. Adding one-week
realized and average implied volatilities as predictors increases the explanatory power by only one percent.
6
economy into two broad dimensions: the nominal and the real side.9 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 “anticipated growth.”
Finally, realized growth can be split one last time into information relating to output versus
employment.
Through this structure, we can potentially obtain two (inflation and growth), three (inflation,
realized growth, and anticipated growth), or four (inflation, output, employment, and anticipated
growth) factors:
• Inflation
 (cid:40)
Output

 Realized Growth
• Growth Employment

 Anticipated Growth
where,forexample,therealizedgrowthfactorcombinesinformationrelatingtooutputandemployment.
In that sense, the information is nested from right to left.
We use information from different subsets of news to construct the different macroeconomic
factors. For example, we extract the U.S. anticipated growth factor from the news flow generated
by 10 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. For completeness, the Appendix lists the assignments of
all macroeconomic announcements for the U.S., Europe, U.K., and Japan to the four categories:
inflation, output, employment, and anticipated growth.
Itis worthreiterating atthispoint that wedo notincludeanymarket-baseddata (suchas stock
prices, interest rates, credit spreads, or implied volatilities) in our analysis, unlike, for example,
Arouba et al. (2009) and Giannone et al. (2008). While such data are very timely and undoubtedly
informativeaboutthestateoftheeconomy,theyrepresentalreadythemarket’sinterpretationofthe
macroeconomic news flow. Our aim is to objectively summarize and describe the macroeconomic
news flow itself, so that we can relate the actual state of the economy to expected stock market
returns.
2.4 Transformation and temporal alignment
We examine the stationarity of each data series in two ways. First, we conduct a Dickey-Fuller test
oneachseries. Second, we readthe definition anddescriptionof eachstatistictodetermine froman
9Theeconomyisoftenseparatedintonominalandrealsidesbecauseshockstothetwoshouldbetreateddifferently
from a policy perspective. For example, many argue, from the perspective of monetary policy, that nominal shocks
should be minimized, whereas real shocks should not be intervened upon.
7
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 first-differenced in
news release time. The Appendix contains more details.
The final data 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 week days in our sample and N refers, for example, to the 43 announcement types
for the U.S.. 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 continuously 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 also allow for sampling
error, which is the case in Kalman filter or Bayesian data augmentation algorithms. However, there
isacleartrade-offbetweenstatisticalcomplexityandabilitytoprocessalargecross-sectionofnews
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 anticipated growth, as well as
8
the two more overarching factors measuring realized growth and growth. As we already discussed,
the most obvious ways of accomplishing this, full data principal components analysis (PCA) and
forecasting regressions, do not appeal to us. First, with full data PCA we 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
meaningtohigherorderfactors. Second,tryingtoidentifythefactorsthroughpredictiveregressions
on 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, 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 (see also Stock and Watson, 1989).
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 Ω of the stationary news series in category i. We work
t,i
with the correlation matrix to abstract from arbitrary scaling of data. Moreover, in order to obtain
areal-timemeasure, weuseatelescoping(meaning, withacommonhistoricalstartdateandrolling
end dates) correlation matrix starting in 1990.10 We denote the N ×1 principal component weights
i
by c , where N is the number of news series in category i. Consistent with extracting principal
t,i i
components from a telescoping correlation matrix, we standardize the news series using telescoping
estimates of their means and standard deviations.
3.2 Economic new series correlation matrix
ThekeyinputstoourmethodologyarethewithinnewscategorycorrelationmatricesΩ . Specifically,
t,i
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 data in
announcement time, partly due to the cross-sectional misalignment of the news in calendar time,
and largely due to the forward filling of missing data.
We address the first unbalanced panel issue by using a correlation matrix estimator along the
linesofStambaugh(1997),whoshowshowtoadjustfirstandsecondmomentsestimatesforunequal
sample lengths. The intuition of his approach is to use the observed data on the longer series, along
10Wealsoexperimentedwithfixedwindowsizerollingcorrelationmatricesfor5,10,15,and20years. Theresults
are qualitatively similar, particularly for the longer data windows.
9