Going granular: The importance of firm-level equity information in anticipating economic activity

41 Pages·2013·0.88 MB·English

Going granular: The importance of firm-level
equity information in anticipating economic
activity
F. di Mauro∗, F. Fornari†
December 9, 2013
Abstract
Theequityreturnsofindividualfirms, andtheir realizedvolatilities,
are shown to improve the in-sample and out-of-sample predictability of
the US business cycle, as measured by the IP index. In fact, significant
declines in the root mean squared errors (𝑅𝑀𝑆𝐸𝑠) are found when
these variables are added to aggregate financial variables and selected
macroeconomic indicators. Overall, to the aim of forecasting, there
is a noticeable swing in the relative importance of individual firms
across time, although firms that become key predictors of economic
activity in a given month continue to do so for around six months, on
average, bringing support to the idea that there is 𝑠𝑡𝑟𝑢𝑐𝑡𝑢𝑟𝑒 in the
information that they convey. Unconditionally, belonging to a given
sector does not boost the predictive power of firms, but we find that it
becomes important for example around periods of recessions. Balance
sheet data show that predictive ability of the firms is associated with
features as performance, liquidity, the size of the foreign activity. Firm
size also matters, as suggested by recent literature (Gabaix, 2011),
although it is not - as put forward there - the only indicator to prevail.
JEL Classification: C53; C58; F37; G15
∗European Central Bank. E-mail: filippo.di [e-mail]
†European Central Bank; e-mail: [e-mail]; corresponding author.
1
Keywords: Business cycle forecasting; granular shock; equity mar-
ket; realized volatility.
2
1 Introduction
The recent recession episode that started in the United States in December
2007 challenged our ability to anticipate the timing and the amplitude of busi-
ness cycle fluctuations. Throughout 2007, almost all the forecasts computed
by central banks, academics and market participants were not able to detect
the approaching sharp decline in real GDP, even when produced around
end-2008, right ahead of remarkably negative GDP growth figures. The highly
coincident and sharply negative GDP growth rates recorded almost worldwide
among industrial countries through the recession, and especially in 2008Q4
and 2009Q1, contribute to make the failure in forecasting even more serious
and call, at the very least, for a critical review of the mainstream forecasting
methodologies. This paper aims to make some steps in this direction.
So far, economic activity has been predicted almost exclusively via the
aggregate information conveyed either by i) macro variables (labor market
conditions, money, credit, lagged growth, to mention a few), ii) financial
indicators (aggregate stock market returns and variances, slope of the yield
curve, credit spreads) or iii) confidence (households or business) indicators
and/or outcomes of surveys. Focusing on models including aggregate financial
variables, which are also considered in the present paper, a broad conclusion
reached by analyses carried out so far is that their predictive power is rather
unstable over time and also that the set of indicators which are key to forecast
business cycle developments tends to change composition over time.
Fornari and Mele (2013) provide a detailed assessment of the out-of-sample
forecasting ability of univariate linear and non linear models which rely on
financial indicators. Overall, their conclusion is that the term spread, together
with a time-varying measure of stock market volatility, does a rather good job
in anticipating the rates of change in the US post-War industrial production
index. However, nearly all of the combinations of variables investigated
by the authors have their short-lived moment of popularity, i.e. what is
judged to be the best model for one business cycle episode is not necessarily
the best in another. This finding cannot but confirm that recessions are
intrinsically different, both as concerns their roots and the way in which
the originating shock propagates across the economy. But, if recessions are
different and shocks transmit both domestically and internationally in a
3
time varying fashion, should not we employ a broader set of regressors - and
potentially models - to better track this variability across time? For example,
recent approaches to forecasting consider pooling the individual forecasts
stemming from a large number of models, each differing from the other as
concerns for example the lag specification, the sample over which estimation
is carried out or the number of variables included. This has been the way in
which the so-called uncertain instabilities have been dealt with in weather
forecasting, an approach which has recently spilled over to macroeconomic
and financial forecasting (see, among others, Amisano and Geweke, 2009;
Clark and McCracken, 2006; Jore et al., 2008).
In this paper we come closer to this strategy as we take a granular, i.e.
a firm-level, perspective. In particular, we test the hypothesis that a linear
combination of past idiosyncratic shocks recorded by the equity price of
selected firms, and its volatility, helps improving the forecast of aggregate
business cycle fluctuations at selected horizons. To the best of our knowledge,
this is the first paper that examines empirically the validity of the granular
hypothesis put forward in Gabaix (2011; see below for more details) in the
context of forecasting and looking at detailed equity market information.
The way we present our results relies on the criterion we use to select, rank
and aggregate the firms we look at into a manageable collection of 𝑚𝑜𝑑𝑒𝑙𝑠,
although its influence on the ranking of the firms and their aggregation and
therefore on the forecasts that we take to be the most informative, is only
marginal.
Firm-level information has so far been scarcely used for macro forecasting
(see, however, Gilchrist et al., 2009, for an application in which firm-level
credit spreads are used for business cycle forecasting)1. The underlying reason
is that the idiosyncratic fluctuations of a given firm-level variable (e.g. the
equity price in our case) should in principle be irrelevant in an aggregate
economy characterized by a large number of firms. This assumption, however,
implies that firms’ returns or the logarithms of their sizes, are normally
distributed. However, a fat tailed distribution for sizes - and even more for
returns - may be a better proxy of reality, consistently also with the industrial
1This paper points out that not all corporate bond spreads help forecast business cycle
developments. Rather the forecasting power of corporate bonds with too high or too low
rating is poorer than for bonds with a ’average’ rating.
4
structure of modern economies, with large corporations and multinationals
significantly on the rise over at least the last two decades. It is exactly
under the latter conditions that Gabaix (2011) derived his so-called granular
explanation of aggregate fluctuations.2 Basically, his empirical evidence shows
that the aggregated shock to the rate of growth of the sales made by the
100 largest US firms anticipates the rate of growth of the US GDP over the
subsequent quarter and its power remains robust to the various controls that
he applies. We anticipate, however, that we do not find size to be the sole
reason behind the predictive power for aggregate fluctuations that we find to
be associated with the equity return of specific firms. We also show that this
predictive ability could not be found in random selections of the same firms.
On the contrary, superior predictive ability is a feature of a relatively small
number of firms at a given point in time.
What we do not deal with in this paper is why should firm-level infor-
mation help to predict business cycle changes. While we do not develop an
explicit model, we conjecture that the sequence of events behind firm-level
predictability could be as follows. Let us suppose that there is an unexpected
decline in the equity price of a firm. This could stem from the postponement
of some of the firms’ projects - due for example to lack of demand for its
products or tight credit availability - which some market analysts first - and
eventually the market as a whole - interpret as a bad signal about the future
profitability of the firm. Of course, being firm-specific, the shock to the
earnings expectations of the selected firm, as perceived by the analyst, and as
transmitted to the firm’s equity price as a consequence of such a revision to
the expected earnings, will be irrelevant for most of the remaining firms as well
as for the aggregate economy in the moment in which it is realized (see also
Barksy and Sims, 2012, for a reasoning along the same lines). Nonetheless, it
may be capturing the first signs of macroeconomic or financial shocks that
later on will eventually spread through the whole economy.
Thefactthatourregressionsevidencethatthepredictabilityofthechanges
in the industrial production indices peaks at long horizons, between 12 and
2SimilarlytowhatGabaixproposes, Carvalho(2009), showsthatnetworkeffectsamong
sectors generate significant propagation effects. There is also an established literature
exploring the impact of microeconomic shocks on aggregate fluctuations, as Jovanovic
(1987); Durlauf (1993); Horvath (1998, 2000); Conley and Dupor (2003).
5
18 months, rather than at short ones (we present results for the 12-month
horizon only and evidence on other horizon can be obtained upon request),
would suggest that also shocks to a given equity price are almost orthogonal
to current growth, while anticipating future developments in business cycle
conditions over more distant horizons.3
Thepaperisorganizedasfollows. InSection2weillustratetheeconometric
methodology that we adopt. In section 3 we describe the data we use in the
paper and we report some unconditional evidence of the relationships between
real activity, aggregate financial information and firm-level equity returns and
variances. This evidence is intended to give a preliminary flavor of the results
presentedintheremainderofthepaper. Section4investigatestherelevanceof
granular information through an out-of-sample econometric exercise. Section
5 looks at the sectoral composition of the predictive distribution of the firms
as well as it analyzes whether some characteristics of the firms, as captured
by key balance sheet items, are related to their predictive power for business
cycle developments. Section 6 looks at some robustness issues, while Section
7 concludes.
2 Methodology
The hypothesis that we want to test is whether real economic activity in the
United States - proxied by industrial production - can be better anticipated
when firm level information4 is added to aggregate information. Beyond
lagged industrial production, our aggregate variables include the term spread
(𝑇𝑒𝑟𝑚) and the return and the variance of the composite stock market index
(𝑀𝑘𝑡𝑅𝑒𝑡 and 𝑀𝑘𝑡𝑉𝑎𝑟). A larger set of aggregate variables is considered in
Section 6, although results remain unchanged, so confirming that firm-level
information is orthogonal to a wide set of aggregate variables. Although
3Always with reference to equity price shocks, Beaudry et al., (2011) analyze the
international spillover of news shocks and conclude that a news shock in a large country
can create national business cycles and international business cycles, thereby providing
motivation for our research, although in their analysis the spillover of the news shock is
related to a concept of geographical proximity.
4The firm level variables that we use are the return and the variance of selected equity
prices(𝑅𝑒𝑡𝑖,𝑉𝑎𝑟𝑖,forevery𝑖𝑡ℎ firminthesample),whichmatchtheaggregateinformation
we look at.
6
Stock and Watson (2003) are frequently reported as evidence against the
existence of predictive power in financial variables, we rely on them especially
as the results in Espinoza et al. (2011) point to financial information i) being
not useless when one takes an out of sample standpoint and ii) being more
important in improving the forecasts in periods characterized by financial
turbulence.
We forecast developments in the growth rate of the Industrial Production
index in the United States ℎ months ahead through the following simple
univariate regression:
𝑚 𝑚
∑ ∑
Δ ln(𝑖𝑝) = 𝛼+ 𝛽 Δ ln(𝑖𝑝) + 𝛽 𝑇𝑒𝑟𝑚
ℎ 𝑡 1,𝑗 ℎ 𝑡−𝑓(ℎ) 2,𝑗 𝑡−𝑓(ℎ)
𝑗=1 𝑗=1
𝑚 𝑚
∑ ∑
+ 𝛽 𝑀𝑘𝑡𝑅𝑒𝑡 + 𝛽 𝑀𝑘𝑡𝑉𝑎𝑟 (1)
3,𝑗 𝑡−𝑓(ℎ) 4,𝑗 𝑡−𝑓(ℎ)
𝑗=1 𝑗=1
𝑚 𝑚
∑ ∑
+ 𝛾𝑖 𝑅𝑒𝑡𝑖 + 𝛾𝑖 𝑉𝑎𝑟𝑖 +𝜀
1,𝑗 𝑡−𝑓(ℎ) 2,𝑗 𝑡−𝑓(ℎ) 𝑡
𝑗=1 𝑗=1
with 𝑠𝑖𝑧𝑒(𝑓(ℎ)) = 𝑚 and where ℎ, the forecast horizon, is equal, in turn,
to 6, 12, 18 or 24 months and 𝑓(ℎ) = (ℎ+6,ℎ+12,ℎ+18) represents the
lag structure chosen for the regressors; 𝑖𝑝 is the Industrial Production index,
𝑡
𝑇𝑒𝑟𝑚, 𝑀𝑘𝑡𝑅𝑒𝑡 and 𝑀𝑘𝑡𝑉𝑎𝑟 are - respectively - the term spread (the 10-year
T-Bond yield minus the 3-month T-bill rate) and the return and the realized
variance of the composite stock market index5. As said, 𝑅𝑒𝑡𝑖 and 𝑉𝑎𝑟𝑖 are
the return and the realized variance of the equity price of selected firms6.
Given the overlapping nature of the data induced by the lag structure of the
regressors, regressions are always corrected via a heteroskedasticity consistent
Newey and West estimator based on a window of data which is a function of
ℎ. The choice for 𝑓(ℎ) made above is of course arbitrary in our regressions.
We tried different combinations and always reached the conclusion that long
lags are needed to significantly improve forecasts (see also, concerning this
5In order to enhance our forecasting scheme, we have tested the inclusion of the return
and variance of the respective sectoral index return and variance, for each firm, either
individually or jointly with the stock market index. The impact of such alternative setting
is negligible and, to save space, we do not report the results.
6The next Section briefly discusses the computation of realized variances.
7
choice, the shape of the impulse responses of GDP to an uncertainty shock as
presented in Bloom, 2009).
As said, the inclusion of the term spread and the stock market return
and volatility among the regressors is motivated by the remarkable success
of these variables reported in the literature (see Estrella, 2005; Fornari
and Mele, 2013). Overall, they convey information about financial risk,
economic risk premiums and monetary policy. During expansions, market
participants exhibit increasing risk appetite and the risk premiums for long
term investments declines. For this reason, and also because monetary policy
is typically counter-cyclical, the term spread is expected to be negatively
correlated to the economic activity. Stock market volatility, on the other
hand, conveys information about the riskiness of financial markets and, more
generally, of the overall macroeconomic environment. A riskier environment
typically leads firms to under-invest and under-hire (Bloom, 2009), ultimately
leading to a deceleration of economic activity, so that higher stock market
volatility is expected to lead to lower economic growth. Households are
also typically found to postpone spending decisions at times of heightened
uncertainty, thereby increasing the strength of such a negative relationship.
The composite stock market return is also included in the regressions
to filter out the part of a firm equity return that stems from its systematic
co-movement with the market. In fact, what we look at are the idiosyncratic
movements of the firms’ returns and returns volatilities, relative to the market
index. Basically, rather than pre-filtering firms’ returns with the market
return and firms’ variances with the market variance, and having therefore to
deal with the problems induced by generated regressors, we directly insert
the aggregate market return and variance in the above equation (more on
this aspect is in the Robustness section at the end of the paper). It is key
to point out here that we insert information about one firm, 𝑖, at a time in
equation (1), in order to assess the significance and extent of every marginal
piece of information added by individual firms. In this way we ignore the
possibility that the return, say, of firm 𝑖, 𝑟𝑖, could be powerful to forecast the
𝑡
industrial production index only because the return of another firm, say 𝑗, 𝑟𝑗,
𝑡
anticipates it and therefore this latter firm is the key anticipator of business
cycle developments. Of course, it is impossible to control for this possibility,
8
given the large number of firms in the sample. However, it is typical that
𝑐𝑜𝑣(𝑟𝑖,𝑟𝑗 ) ≃ 0 for 𝑘 larger than a few days and for nearly all 𝑖 and 𝑗, so
𝑡 𝑡−𝑘
that as far as returns are concerned, we can rule this possibility out. By
contrast, contemporaneous correlations between two equity returns can be
efficiently handled by the presence in equation (1) of the composite market
return and variance.
3 Data and In Sample Evidence
3.1 Data
The firm-level information that we consider comes from the equity prices of a
large set of firms sharing the following characteristics: i) they are listed in US
stock exchanges; ii) they have been continuously listed in US stock exchanges
since 1973, the first year for which Thomson Reuters provides historical data.
This results in a set of 𝑛 = 280 firms. They are allowed to come from any
US industrial sector, so to maximize our chances of detecting firms with high
forecasting power. We consider the so-called level-3 Industry Classification
Benchmark - the standard company classification system developed by Dow
Jones and FTSE - i.e. a 10-sector classification. Table 1 provides a brief
description of the sectoral structure in the dataset.
For each firm we collect the daily equity price and build, at the end of
each month in the sample, i.e. between January 1973 and December 2012,
realized returns and realized volatilities over various horizons (6, 12, 18 and
24 months). The use of realized volatilities builds on the large literature
initiated by Andersen et al. (2003), and basically boils down to cumulating
daily absolute equity returns within each calendar month or longer backward-
looking horizons. It is important to highlight here that the set of firms we look
at certainly suffers from a survivorship bias. However, from the standpoint
of a forecasting exercise we could only improve upon the results we present
in this paper if we were to consider additional firms, for example those that
defaulted or started their business after January 1973. Yet, the survivorship
bias could lead us to miss some important factors behind the forecasting
power of firms when we try to figure out a structural explanation for it. For
example, if firms were to increase or decrease their forecasting power for
9
Sector US
1)Oil&Gas 22
2)BasicMaterials 20
3)Industrials 65
4)ConsumerGoods 41
5)HealthCare 16
6)ConsumerServices 27
7)Telecommunications 4
8)Utilities 35
9)Financials 37
10)Technology 13
Total 280
Table 1: Firms distribution across sectors.
subsequent IP developments when their default risk reached critical values,
we could have a hard time trying to nail down such a relationship, as these
firms would likely be about to leave, or would have already left, the composite
equity index we look at.
As said before, the industrial production index (IP) is our measure of
real activity. We also collect the daily composite stock market index7, from
which we compute end-month realized returns and realized variances (in the
same way as for individual firms), as well as the term spread (the difference
between the ten-year government bond yield and the three-month T-bills or
eurodeposit rate, when the former is not available).
3.2 In-sample evidence
To anticipate the kind of results we will get, we present here the unconditional
relationshipbetweenrealactivity, aggregateinformationandfirm-levelreturns
and realized variances. In practice, we estimate model (1) throughout the
whole sample and look at the firms’ performance as summarized by the
regressions’ adjusted 𝑅2, which is reported in Figure 1 for the 12-month
horizon only.8 The horizontal line represents the adjusted 𝑅2 from the
7We use the index provided by Thomson Reuters. It is constructed as the average of
the domestic equity returns, weighted by the firms capitalization. Such index turns out to
have a correlation near to one with the MSCI US index.
8We report adjusted 𝑅2 coefficients but the difference in regressors between the specifi-
cation with aggregate information only and with aggregate and firm-specific information is
not particularly large as only one firm at a time is considered and therefore the additional
variables are only two, each with three lags, which represents a negligible difference in
10