Documente Academic
Documente Profesional
Documente Cultură
stock returns
Ningzhong Li
University of Texas at Dallas
ningzhong.li@utdallas.edu
Scott Richardson
London Business School
srichardson@london.edu
İrem Tuna
London Business School
ituna@london.edu
Abstract
We are grateful to Eli Amir, Mary Barth, Victor DeMiguel, George Foster, Francisco Gomes, Jeff
Hales, Trevor Harris, Chris Higson, Andrew Jackson, Peter Joos, Rabih Moussawi, Paul Pacter,
Stephen Penman, Peter Pope, Tjomme Rusticus, Lakshmanan Shivakumar, Florin Vasvari, Franco
Wong, an anonymous referee and seminar participants at American Century Investments, University of
California, Berkeley, Columbia University, INSEAD Business School, London Business School,
London School of Economics, Manchester Business School, Rotterdam School of Management, and
Stanford University for helpful discussion and comments. Any errors are our own.
(macroeconomic) exposure is useful for forecasting firm fundamentals and stock returns. While
the link between firm operating and investing decisions and broader macroeconomic features
seems relevant for forecasting, surprisingly little archival, empirical research has examined these
landscape is important.
The rapid change in the relative economic importance of countries around the world
suggests that attention to a given company’s geographic exposure should be useful to an investor
seeking to forecast future cash flows and associated risks for the purpose of security valuation.
The potential usefulness of macroeconomic information from the perspective of security analysis
The set of potential macroeconomic variables to consider is large. We consider how each
company is exposed to its home country and other countries. This is a natural choice given that
operating and investing choices that span across countries is likely to be a primary mechanism by
which macroeconomic factors affect firm performance. If all firms operated in the same country,
then dispersion in macroeconomic factors across countries would not be relevant. We identify
country exposures via the geographic segment disclosures included in annual reports. Second,
we rely on information external to the firm via country level forecasts. We use (i) forecasts of
real GDP growth from Consensus Economics (CE), and (ii) country level stock returns from the
several reasons. Given our primary measure of country exposures is geographic segment data,
there is likely to be measurement error due to the subjective manner in which countries are
disaggregated across companies and also due to the country exposures being primarily driven by
sales data (a data limitation with geographic segment reporting). The cost exposures across
countries are missing from our measure, thereby limiting our ability to capture the full set of
exercise. We not only have to measure company to country exposures well, but we must also
have a meaningful forecast of relative performance across those countries. While we use
forecasts from CE and country level stock returns as our measures of expected country level
performance, we note that any errors in these forecasts will feed directly into our forecasts of
firm fundamentals.
Country exposures may not be useful to improve forecasts of sell-side analysts or stock
returns for reasons in addition to the measurement error and compounded forecasting challenge
described above. Specifically, analysts are likely to use macroeconomic information in their
earnings forecasts, target prices and stock recommendations. Likewise, stock prices are likely to
efficiently incorporate this information on a timely basis. However, the extent of geographic
exposures for large multi-national companies and the challenges in systematically incorporating
this information into firm specific forecasts, suggest it is an open empirical question as to
1
Collins (1976), Silhan (1983) and Roberts (1989) find that the incremental contribution of earnings relative to sales
data at the segment level was quite small in terms of improving earnings forecasts, suggesting that revenues is most
important for forecasting. In part, the similarity in predictive power from using either sales or earnings based
geographic exposures, may be attributable to the subjectivity in cost allocation and transfer pricing for large multi-
national companies.
For a sample of 381,030 firm-years for US and non-US firms over the 1998-2010 time
period, we find that combining country exposures with country level forecasts improves
forecasts of firm fundamentals. Given our exposures are sales based, we assess forecasts of
ROA (income to average assets) as well as SOA (sales to average assets). The predictive power
is evident in annual cross-sectional regressions which suggest that a one percentage point
increase in real GDP growth or recent country level returns translates to an additional 10 to 20
(10 to 30) basis points of ROA (SOA) over the next year. Further, we show that out-of-sample
find that sell-side analyst earnings and sales forecasts appear to be slow to incorporate this
information. Specifically, we find that analyst earnings and sales revisions are associated with
information contained in current country exposures and country level forecasts for the next 3
months.
We also show evidence that stock returns appear to incorporate the information in
country exposures with a lag. This is supported in cross-sectional regressions of equity returns
where the country exposures combined with country level forecasts are able to explain cross-
sectional variation in equity returns for the next 3 months, after controlling for known
determinants of equity returns (e.g., momentum, size, beta, accruals, earnings-to-price, and book-
to-price). Further, time series tests based on portfolios formed using country exposures and
country level forecasts achieve statistically significant Sharpe ratios that are not explained by
standard risk factors. We note that the economic significance of the stock return predictability
across size portfolios is quite strong: a long/short portfolio exploiting current country exposures
3
and country level forecasts generates 85 to 87 (136 to 172) basis points per month for the largest
(smallest) quintile of firms. This return relation is robust across various groups of firms based on
Our sample includes ‘domestic’ firms with exposure only to their home country, and
‘non-domestic’ firms that have exposures to multiple countries. For both groups of firms it is
relevant to jointly consider the usefulness of country exposures and country level forecasts to
improve forecasts of firm fundamentals. Including both ‘domestic’ and ‘non-domestic’ firms
together increases the power of our fundamental forecasts, as we do not expect the information
content of country exposures and country level forecasts to be different across ‘domestic’ and
‘non-domestic’ firms in terms of forecasting firm profitability. However, for our empirical
analysis exploring the speed with which information about country exposures and country level
forecasts is incorporated into analyst revisions and stock returns, we examine these groups
separately as there are valid reasons to expect differential speed of price discovery for ‘domestic’
We find that macroeconomic information is useful for both purely ‘domestic’ firms as
well as ‘non-domestic’ firms. Our purely ‘domestic’ firms are smaller and less liquid, so finding
stronger relations between macroeconomic information and future analyst revisions and future
stock returns may not be too surprising. However, for a reduced sample of ‘non-domestic’ firms
that have either (i) a meaningful fraction of foreign sales (i.e., greater than 20 percent), or (ii) a
greater number of geographic segments (i.e., greater than three), we find strong evidence of a
lagged relation between macroeconomic information and future analyst revisions and future
stock returns. While these sub-samples comprise only 23% and 11% respectively of the full
sample, they represent 43% and 34% respectively of aggregate market capitalization. These
4
reduced samples contain the largest and most economically important firms. For these sub-
samples, the information content of country level exposures comes from either the home country
exposure and/or the set of foreign country exposures. We find that home country and foreign
country exposures are useful in predicting future analyst revisions and future stock returns.
current country exposures and country level forecasts differs across multiple measures of
‘saliency’ (see e.g., Hong and Stein, 1999 and Hirshleifer and Teoh, 2003). For example, we
find that price discovery is faster for securities included in the MSCI country indices. This is
consistent with market segmentation where investors trading on the basis of macroeconomic
information are likely to trade baskets of securities (e.g., MSCI country indices) which have
greater sensitivity to that macroeconomic information. This result of faster price discovery for
MSCI constituents is found for both ‘domestic’ and ‘non-domestic’ firms. We also find that the
speed of price discovery for macroeconomic information is slower for more ‘complicated’ firms,
using the number of reported geographic segments as a proxy for geographic complexity.
Finally, we show that the return predictability of information contained in current country
exposures and country level forecasts is greater after periods of increased dispersion in real GDP
growth forecasts across countries. Collectively, these results are consistent with theories of
limited attention and limited arbitrage where investors are slow to incorporate new value relevant
The primary contribution of this paper is to introduce a simple framework to identify and
exploit linkages between firm performance and potential macroeconomic drivers of that
performance. Our approach is similar in spirit to recent research exploiting economic linkages
across firms. Examples include: (i) Cohen and Frazzini (2008) who show that incorporating
5
information about linkages between firms along the supply chain improves forecasts of firm
fundamentals and stock returns, (ii) Menzly and Ozbas (2010) who show that incorporating
information about linkages across industries, using data from the Bureau of Economic analysis,
improves forecasts of firm and industry level fundamentals and stock returns, and (iii) Cohen and
Lou (2012) who show that decomposing multi-segment firms into separate ‘pure plays’
The rest of the paper is structured as follows. Section 2 lays out a framework for linking
country exposures to forecasts of country performance and describes our economic hypotheses.
Section 3 describes our measures of country exposures and country forecasts that are used in our
empirical tests. Section 4 presents our empirical analysis and section 5 concludes.
A large literature in accounting and finance has explored the determinants of firm
profitability. Some classic papers include Penman (1991) and Fama and French (2000) where
the focus is on documenting a strong mean reversion in profitability. Such mean reversion is not
unexpected as competitive forces will erode firms with above ‘normal’ profitability and the
discipline of the market will remove firms with below ‘normal’ profitability. A vast literature
has expanded the set of determinants of firm profitability to exploit: (i) accruals vs. cash flows
(Sloan, 1996 and Xie, 2001), (ii) margins vs. turnover (Fairfield, Whisenant and Yohn, 2001 and
Soliman 2008), (iii) earnings volatility (e.g., Dichev and Tang, 2009), (iv) domestic vs. foreign
earnings (e.g., Thomas, 2000), and (v) the impact of accounting distortions attributable to
6
A common feature of the majority of past research is that it does not explicitly
incorporate information external to the firm itself. While it is possible that disaggregating
earnings into components will identify, in a reduced form, links to such external drivers of firm
profitability, they are not explicit with respect to these external drivers. Our focus is on first
principles to identify potential factors outside the firm’s direct control that will have an impact
on profitability. As noted in the introduction, this is potentially a very large set of variables.
Examples could include: (i) currency movements, (ii) commodity movements, and (iii) financial
market variables such as aggregate credit spreads and sovereign yield curves.
Rather than attempt to construct a general equilibrium model linking a set of primitive
macroeconomic variables to firm profitability, we have deliberately reduced the focus of our
empirical analysis to macroeconomic exposures that are both intuitive and measurable by the
researcher. We recognize that firms operating across countries are exposed to cross-country
differences in a variety of factors (including, but not limited to, those mentioned above) that will,
in part, determine their profitability. Not all firms share the same set of exposures across
countries at a point in time and not all firms keep their cross country exposures constant through
time. For example, Burberry Group PLC specializes in the design, manufacture and distribution
of apparel and accessories via retail and wholesale channels. As of December 31, 2011,
Burberry has market capitalization of 5.2 billion pounds and total revenues of 1.5 billion pounds.
Burberry’s revenue is sourced from around the world as follows: (i) Europe 33.8 percent, (ii)
Asia Pacific 30.4 percent, (iii) Americas 25.7 percent, and (iv) other 10.0 percent. In contrast,
Mulberry Group PLC designs, manufactures and retails fashion accessories and clothing. It
operates a retail and design division and as of December 31, 2011, Mulberry has market
capitalization of 0.9 billion pounds and total revenues of 121 million pounds. Mulberry’s
7
revenue is sourced from around the world as follows: (i) Europe 81.5 percent, (ii) Asia 12.7
percent, (iii) North America 4.3 percent, and (iv) other 1.5 percent. Clearly, the geographic
footprint of these two luxury good specialists is different and this difference in geographic
exposures is likely to be a key determinant of the difference in profitability into 2012 and beyond
Our empirical strategy is to identify for each firm the geographical source of its revenues.
In section 3.1, we describe in detail the source of the geographic segment data we use for this
purpose, along with the data choices necessary to make these disclosures cross-sectionally
comparable.
2.2 Prior research linking macroeconomic (country) exposures to firm level profitability
Prior accounting literature has explored the potential forecasting benefit of industry (line
of business) segment disclosure information. Pacter (1993) categorizes this literature into papers
that explore the effect of industry segment data on (i) investor assessments of expected returns,
and (ii) investor assessments of risk and cost of capital. There are several older papers
examining the usefulness of line-of-business segment data to improve forecasts of earnings at the
parent company level. These studies typically use very small samples from the early 1970s when
the SEC first introduced segment disclosure requirements (e.g., Collins, 1975; Collins, 1976;
Kinney, 1971; and Foster, 1975). In addition, there are more recent papers examining the
company level (e.g., Balakrishnan, Harris and Sen, 1990; and Roberts, 1989). Again, these
papers use very small samples (e.g., 89 firms for Balakrishnan, Harris and Sen, 1990; and 78
firms for Roberts, 1989) making generalizability difficult. Furthermore, prior research finds
mixed evidence that segment data at the line-of-business or geographic level improves out-of-
8
sample forecasts of firm profitability. With the exception of Collins (1976), none of these papers
examine the speed with which this information is incorporated into security prices or capital
In the more recent literature in financial economics, there are several streams of related
papers. However, none of them directly subsume our analysis. Thomas (2000) shows that
decomposing earnings changes into a domestic and foreign component generates superior
forecasts of future earnings growth and that the stock market fails to appreciate this in a timely
manner. However, this approach does not make any use of information external to the firm.
Thomas (2000) uses realizations of domestic and foreign earnings growth without any attempt to
forecast the domestic economy relative to foreign economies nor to decompose the foreign
exposure into its country level composition. In unreported analyses we have separated current
profitability into its domestic and foreign components and this separation takes nothing away
factors including (i) inflation expectations (e.g., Chordia and Shivakumar, 2005; Basu, Markov
and Shivakumar 2010; and Konchitchki, 2011), (ii) foreign currencies (e.g., Bartov and Bodnar,
1994; and Bartram and Bodnar, 2012), and (iii) general macroeconomic state variables (e.g.,
Cochrane 2000; Cochrane, 2010; Liew and Vassalou, 2000; Vassalou, 2003; and Li, Vassalou
and Xing, 2006). A common limitation to all of this prior research is that the sensitivities are
generally estimated statistically and they do not make any attempt to incorporate forecasts of the
respective macroeconomic variables. Our empirical analysis differs from these papers in several
key respects. First, we estimate our exposures using ‘priors’, as opposed to statistical estimation
which is known to be an imprecise estimate of latent sensitivities (see e.g., Scholes and Williams,
9
1977, and Dimson, 1979 for a discussion of estimation errors for ‘beta’). Second, we exploit the
full set of country exposures provided in the geographical segment disclosures and do not limit
our analysis to a comparison of ‘foreign’ to ‘domestic’ effects. Third, and perhaps most
importantly, we utilize forecasts of the expected performance of each country that a given
company is exposed to. We are thus able to answer the question whether knowledge of macro-
content in the combination of country exposures and forecasts of country level performance for
(i) forecasting firm profitability, and (ii) testing whether this information is reflected in analyst
We use forecasts of real GDP growth from Consensus Economics (CE) and country level
stock returns from MSCI as our measures of expected country level performance. Here we
describe how we combine the company level geographic segment data with these country level
forecasts.
For each firm-year observation we disaggregate total sales into country level sales based
on the geographic segment data in the most recent annual report. We retain companies with a
purely domestic footprint (i.e., those companies with zero foreign sales) because our empirical
analysis is based on a global set of firms and retaining domestic firms allows us to more cleanly
assess the importance of macro-economic information. For example, if firm A has 50 percent of
its sales in Germany and 50 percent of its sales in Greece and Firm B has 100 percent of its sales
in Greece, and Greece is expected to outperform Germany, then holding all else equal, the ‘best’
portfolio exposure to express that view would be via Firm B, the purely domestic firm. As
10
discussed in the introduction, grouping ‘domestic’ and non-domestic’ firms together will
increase the power of our tests forecasting future profitability. However, for our empirical
analysis using analyst forecast revisions and stock returns we expect differences across these two
groups of firms due to differences in firm size, liquidity and complexity. Hence we examine
After gathering the sales data for firm i, for each country c, at each point in time t,
,, , we standardize these sales measures so that they sum to one. We then use our
forecasts of expected country level performance for each county c at each point in time t,
we label , . This measure captures both cross sectional and time series variation in firm
We conduct three sets of empirical analyses. First, we assess the relative (out-of-sample)
performance of forecasts of firm fundamentals (income on average assets and sale on average
assets) that include , . Second, we assess the ability of , to forecast sell-side
analysts’ earnings and sale forecast revisions. To the extent that our measure is able to forecast
analyst revisions, it is consistent with analysts failing to incorporate this information in a timely
manner. A benefit of the analyst revision tests is that it can allow us to attribute any stock return
relation to mispricing rather than risk (e.g., Bradshaw, Richardson and Sloan, 2001). Third, we
11
2.4.1 Firm fundamentals
P1: Combining country level exposures with expectations of country level performance is useful
to forecast future firm profitability.
forecasts of firm profitability. The benchmark forecasting model for firm level profitability is a
modified random walk that acknowledges profitability is mean reverting, and also exploits
various firm characteristics that isolate differences in persistence of profitability (see e.g., Core,
Guay and Rusticus, 2006; Fama and French, 1995; So, 2012; and Hou, van Dijk and Zhang,
2012). Specifically, we run the following regression for each quarter (firm subscripts, i, dropped
Equation (1) is estimated using two measures of profitability: return on assets (ROA) and
sales on assets (SOA). Income is more important for security valuation; however our country
exposures are based on sales data so we also examine sales based measures of fundamentals.
( ) is return on assets (sales) computed as income before extraordinary items
(sales) divided by average total assets, is as defined previously, ( ) is return
(sales) on assets for the previous twelve months, )* is book-to-price measured as the book
value of common equity divided by market capitalization using data available at the start of the
period for which we examine future profitability, ,- is the log of market capitalization (in
USD to ensure cross-sectional comparability), /0 is the change in net operating assets as
measured in Richardson, Sloan, Soliman and Tuna (2005), /_3 is an indicator variable equal
12
to one for firms reporting a loss in year t, and zero otherwise, /_/,5 is an indicator variable
equal to one for firms paying a dividend in year t, and zero otherwise, and /,5_7,8 is the
dividend yield for year t. We estimate this regression each year for the 12 sector groupings
identified in Fama and French (1997). This ensures that we have sufficient sample size for each
sector-year group. Standard errors are computed using the time series of the sector-year
regression coefficients with a correction for serial correlation. Inferences are unchanged if we
instead run pooled regressions with standard errors clustered for industry (or firm) and year. We
expect profitability to be mean reverting so our priors are for %& to be less than one and greater
than zero. We expect firms with greater growth opportunities, as measured (inversely) by )* ,
to have high levels of profitability after controlling for current profitability, so we expect a
negative %( coefficient. As originally noted in Fama and French (1995), we expect smaller firms
to exhibit lower levels of future profitability controlling for current profitability, so we expect a
We expect loss making firms to have lower profitability (i.e., %1 < 0 ) and firms paying
dividends to have higher profitability (i.e., %4 > 0 and %6 > 0). Finally, we expect a positive
coefficient for our primary variable of interest, . The greater the exposure of a firm to
countries that are expected to do well, the greater we expect future profitability to be, controlling
for other known determinants of profitability. Figure 1 illustrates the timing of our variable
measurement for the estimation of equation (1). For simplicity we discuss a December year-end
firm, but the timing convention carries over to all fiscal year ends.
In addition to the descriptive analysis of firm profitability from estimating equation (1),
estimating equation (1) for each sector grouping every year using an expanding window. This
13
provides a set of sector-year coefficients which are then combined with the current realizations
estimate a second forecast of firm level profitability that excludes that variable,
<=AB >?@ [ ]. We then compare the relative accuracy of these two forecasts with the
actual future profitability, . The resulting errors are defined as follows:
<= >?@ = C
<= >?@ [ ] − C (2a)
<=AB >?@ = C
<=AB >?@ [ ] − C (2b)
We test the differences of these error distributions to assess the out-of-sample predictive
P2: Sell-side analysts do not efficiently incorporate information about country level exposures
and expectations of country level performance into their earnings (and sales) forecasts.
Prior literature has shown that analyst forecasts appear to be slow in incorporating a
variety of information (e.g., Hughes, Liu and Su, 2008 for past stock returns; Bradshaw,
Richardson and Sloan, 2001 and 2006 for measures of accruals and external financing; So, 2012
for a variety of other measures). In contrast, Kadan, Maduereira, Wang and Zach (2012) note
that sell-side analysts are respected for their industry, sector and general market knowledge.
Indeed, analysts are rated most highly on this industry and market knowledge in surveys. Thus,
14
country level exposures and expectations of country level performance into their earnings (and
sales) forecasts.
We test P2 directly by examining the speed with which analysts incorporate the
information contained in E into their firm level earnings and sales forecasts. Specifically,
we estimate the following regressions every month (again firm subscripts, i, dropped for the sake
of brevity):
Equation (3) is estimated for the next three months (i.e., k = 1 to 3). As with our
fundamental forecasts discussed in section 2.4.1, we track analysts’ revisions of earnings and
sales forecasts. 5,,F is the monthly revision in consensus sell-side analyst forecasts. To
ensure cross-sectional comparability of sell-side analyst earnings and sales forecasts across firms
with different fiscal year ends, we first take a calendar weighted average of one year ahead,
[
1, ] , and two-year ahead earnings and sales forecasts,
[
2, ], where the weight is a
linear function of the number of months to the end of the next fiscal year, M. We label the
December year end firm we place 9/12 weight on the forecast for the 2010 fiscal year and 3/12
weight on the forecast for the 2011 fiscal year. The consequence of this choice is that our
resulting earnings or sales forecast is twelve months ahead for all firms. Finally, we compute
5,,F as:
[O&P,QRS ]
5,,,F = ln [O& (4)
P,QRSTU ]
15
Given that we use the natural logarithm operator we restrict our firms to those where the
calendar weighted forecasts across both months are strictly positive.2 Prior literature has shown
that analyst forecast revisions are highly serially correlated (e.g., Hughes, Liu and Su, 2008).
We therefore expect %& to be positive. )* is as defined previously. 0G/ is the ratio of net
income before extraordinary items to market capitalization at the start of the month. We expect
both %( and %+ to be negative, as firms with high expectations of earnings growth should, on
average, deliver that earnings growth (and changing expectations of growth). IJ is
the recent six month stock return. We include this variable as prior research has shown that sell
side analyst forecasts reflect expectations embedded in stock price with a lag (e.g., Hughes, Liu
and Su, 2008). Consistent with prior research, we expect %. to be positive. We also expect %1 to
be negative as sell-side analysts fail to incorporate the lower persistence of accruals into their
forecasts (e.g., Bradshaw, Richardson and Sloan, 2001). Finally, we expect % to be positive if
analysts are slow to incorporate information about company level geographic exposures and
It is important to note the risk of ‘throwing the baby out with the bath water’ in equation
(3). We have included market price via three variables, )* , 0G/ , and IJ . To
the extent that stock prices have efficiently incorporated all information, then any predictive
content of other information will be reduced. Figure 2 illustrates the timing of our variable
We also make additional empirical predictions for our analyst revision tests as follows:
2
In unreported results we have measured 5,,F using firms with positive and negative twelve month ahead
VO&P,QRS WXVO&P,QRSTU W
earnings forecasts as follows: 5,,,F = . Our inferences are unchanged with
(CO&P,QRS CCO&P,QRSTU C)/&
this alternative measure.
16
P3a: P2 is expected to vary based on the natural variation in expectations of country level
performance.
P3b: P2 is expected to vary based on differences in saliency of macroeconomic information.
P3a is motivated on the basis that the information content of macroeconomic information
is naturally greater when then there is greater dispersion in expected performance across
countries. We test whether analysts appreciate this greater information content by examining
whether the lagged association between analyst revisions and MACRO is greater after periods of
increased dispersion in expected performance across countries. P3b is motivated on the basis
that limited attention can impede individuals ability to incorporate information for (i) smaller
firms, (ii) less followed firms, and (iii) more complicated firms.
P4: Stock prices do not efficiently incorporate information about country level exposures and
expectations of country level performance.
tests to assess the relation, if any, between future stock returns and the information contained in
For our cross sectional characteristic tests, we run the following regression every month
*F = # + % + %& )* + %( 0G/ + %+ )I + %. ,- + %1
* +
Equation (5) is estimated for the next three months (i.e., k = 1 to 3). To simplify the
interpretation of the results, we examine each month separately (i.e., the stock returns,
*F ,
are not cumulated across K months, but instead focus on the Kth month). The relevant test is
17
whether % = 0 , and finding % > 0 is consistent with stock returns failing to efficiently
incorporate company level geographic exposures and country level performance. Of course, this
inference is conditional on our ability to control for known risk attributes in the cross sectional
regression model. Building on Fama and French (1992 and 2008) we include firm characteristics
known to be associated with future returns: 0G/ and )* . We expect both to be positively
associated with future returns. )* is as defined previously. 0G/ is computed as net income
before extraordinary items divided by market capitalization as at the end of the most recent fiscal
period. We also include measures of firm size, ,- , as defined in Section 2.4.1, and )I ,
measured as the single factor CAPM beta, using monthly data from the last 60 months for each
two measures of recent stock returns. First is I , which is the return for the most recent month.
Given prior research has documented a short term reversal effect (e.g., Jegadeesh, 1990) we
has shown a continuation in stock returns over the medium term, we expect the coefficient on
previous research has shown to be strongly negatively associated with future returns. We also
include an indicator for loss making firms, /_3 , but note that prior research has found mixed
results with this variable (e.g., positive in Fama and French, 1992 and marginally negative in
We estimate equation (5) twice for each cross-section and then report test statistics using
the time series variation in the regression coefficients. First, we report value weighted cross
sectional regressions. This weighting approach allows us to assess the strength of any cross
sectional relation across firm size. If the return result is attributable to smaller (and potentially
18
less liquid and riskier) securities, value weighting will reflect this. Second, we report ‘risk’
optimally combining forecasts of expected return and expected risk (see e.g., French, Schwert
and Stambaugh, 1987; Lang, Stulz and Walkling, 1991; and Pontiff, 2006). Indeed, much
current research emphasizes the importance of constructing portfolios where the portfolio
weights are inversely related to measures of volatility (e.g., Maillard, Roncalli, and Teiletche,
2010; and Kirby and Ostdiek, 2012). Therefore, for our equity security portfolios, ignoring any
correlation structure across common factors, portfolio weights should be inversely proportional
residual monthly returns (using a single market factor model over the last 24 months). An added
benefit of this alternative weighting scheme is that it will (i) help correct for microstructure
issues in returns because volatility is positively associated with the likelihood of bid-ask bounce
(e.g., Blume and Stambaugh, 1983 and Asparouhova et al., 2010), and (ii) help correct for
expected transaction costs because volatility is positively associated with expected trading costs
(see e.g., Korajczyk and Sadka, 2004; and Richardson, Tuna and Wysocki, 2010, p 444, for a
related discussion).
For our portfolio level analyses we conduct two sets of tests. First, following the
suggestion of Fama and French (2008) and Lewellen (2010), we sort each cross-section into five
quintiles based on market capitalization, . We then sort firms within each based on
. This allows us to quantify the relation between and future returns holding
firm size constant. This analysis helps us make inferences about economic significance. If a
return result is only evident in the smallest securities, then the economic significance of the
relation is weak. Second, we perform time series regressions where the zero-cost hedge portfolio
19
return, Z
/[
, (a portfolio that is long (short) the securities in the top (bottom) quintile of
) is projected onto a set of changes in macro-economic state variables (e.g., Chen, Roll
and Ross, 1986) and standard factor-mimicking portfolio returns (e.g., Fama and French, 1992
following regression:
Z
/[
= # + % \* + %& ) +%( Z3 +%+ + %. 8 + %1 8* + %4 8G + (6)
), Z3, and are the factor-mimicking portfolio returns from Ken French’s
website. As our empirical analyses uses a global set of securities we use the global factor returns
based on stock level data from 23 developed markets that correspond closely to our sample
composition. \* is the excess return to the global market portfolio. 8 is the change in
corporate risk premium, measured as the change in the default spread (the difference between the
Moody’s Seasoned BAA Corporate Bond Yield and the 10 year US Treasury constant maturity
rate). 8* is the change in term structure, measured as the change in the difference between the
10 year US Treasury constant maturity rate and the 2 year US Treasury constant maturity rate.
8G is the percentage change in Industrial Production for the month. To the extent that factor-
mimicking portfolio returns and the changes in our selected macro-economic state variables
reflect compensation for changes in risk profile, we control for time series variation in risk in our
analysis by including these variables. The relevant test is then whether the intercept in this time
Similar to our analyst revision empirical predictions, we also make additional predictions
P5a: P4 is expected to vary based on the natural variation in expectations of country level
performance.
20
P5b: P4 is expected to vary based on differences in saliency of macroeconomic
information.
The motivation for these additional predictions is the same as that for P3a and P3b.
We extract geographic exposure data from the annual fundamental file created by
Compustat for US firms, and the annual fundamental file created by FactSet Fundamentals for
non-US firms. We capture the geographic exposure data from annual reports for firms with
positive sales. We use the geographic sales data because the coverage of geographic earnings
data is very limited. SFAS 131 ‘Disclosures about segments of an enterprise and related
information’ is the relevant standard in effect for US firms for our sample period. This standard
requires companies to disclose detailed segment data using segment definitions based on a
‘management approach’. This means that the identification of operating segments for the
purpose of external financial reports needs to be consistent with the segment basis used by the
firm’s key operating decision makers. While this creates considerable flexibility in the
identification of operating segments across firms (i.e., some firms may elect to identify operating
segments on a product basis or an industry basis, while others may adopt a geographic basis),
there is still a clear geographic disclosure requirement for US firms. Paragraph 38a of SFAS 131
customers, and assets located in the country of domicile and those in foreign countries as a part
assets located in a single foreign country are material, then they need to be reported separately.
21
For non-US firms that followed international accounting standards, the relevant
accounting standard for the period 1998 to 2008 was IAS14. IAS14 required firms to make
separate disclosures for geographic segments. A geographic segment is based on either where
the enterprise’s assets or customers are located (paragraph 13). Materiality thresholds determine
the identification of a unique segment (typically 10 percent of the enterprise value). Unique
geographies are identified until a 75 percent total threshold is met (paragraph 37), and the
smaller segments are typically aggregated together (paragraph 36). Paragraphs 51 to 67 outline
in considerable detail the required disclosures for each geographic segment. For fiscal years
ending after 2009, IFRS8 is in effect for firms following international standards (it replaced
IAS14 effective January 1, 2009). IFRS8 is virtually identical to FAS131 in its segment
disclosure requirements. Thus, for the majority of our sample period (1998 to 2009), it appears
that the requirement for geographic disclosures for US firms is less detailed than that for non-US
firms that were following the international accounting standards. This suggests that there will be
greater measurement error in the identification of geographic exposures for US firms relative to
this group of non-US firms. In later analysis we find that the predictive content of geographic
Our final sample covers 381,030 unique firm-years, spanning 56 countries over the 1998
to 2010 period. Panel A of Table 1 provides a breakdown of the country headquarters for the
firms included in our sample. US firms make up 21 percent of our sample. The next most
important countries are Japan (13.6 percent), UK (6.6 percent), China (6.3 percent), and India
(4.7 percent). The average firm in our sample reports $862 million in annual sales, $1.89 billion
in total assets and has a market capitalization of $0.9 billion. In contrast, the median firm in our
sample has $93 million in annual sales, $146 million in total assets and has a market
22
capitalization of $87 million. All of these amounts are expressed in USD. We have translated
balance sheet (income statement) amounts reported in local currency to USD using fiscal year
end (average) foreign exchange spot rates. Our sample contains some of the largest multi-
national companies in the world, but also contains a large number of the smaller firms. The
average (median) firm in our sample has a )* value of 1.08 (0.75) and reports virtually zero
profitability. 49.5 (28.4) percent of sample firms pay dividends (report losses). The sample
covers the main economic sectors with the greatest concentration in money and finance (19.0
We keep firm-years which do not have any foreign sales (i.e., ‘domestic’ firms). As
reported in panel B of Table 1, our primary sample contains about 75 percent purely domestic
firms. We retain these firm-year observations in our primary sample as our aim is to assess the
level exposures to matter for both ‘domestic’ and ‘non-domestic’ firms, we include both sets of
firms to increase the power of this fundamental forecast. However, for our empirical analysis
testing the relative efficiency of analyst forecasts and stock prices with respect to
expect those groups to differ in systematic ways (e.g., firm size, liquidity and complexity).
The disclosure practices of firms related to segment disclosures vary considerably, both
across time and across firms. There is little homogeneity in how firms choose to describe the
geographic regions in which they source their revenues. This creates a challenge for accurately
mapping geographic regions to countries. We use a standard tree structure that maps various
23
geographic regions to member countries. For companies that report sales at an aggregated
regional level, we allocate these sales across the member countries using a GDP weighting for
that respective year (consistent with Roberts, 1989). This approach exploits the relative
importance of economic activity across countries within that region by allocating more sales to
the more important member countries. Undoubtedly, this choice introduces measurement error
into our country level sales exposures for firms that have targeted certain countries within a
geographic region. However, absent reliable data we cannot do more than this. Balakrishnan,
Harris and Sen (1990) note that for their sample of 89 firms there is a close mapping between
actual country specific sales disclosures and implied country specific sales (using a GNP
weighting across countries within a particular region), suggesting that the measurement error
may not be that large for our sample. We then standardize the country level sales data such that
they sum to one for each firm year. A detailed example for Mulberry Group PLC is shown in
Appendix I.
We use two measures of expected country level performance: (i) country level real GDP
forecasts from Consensus Economics (CE), and (ii) recent country level stock returns from the
MSCI indices. CE was founded in 1989 and they have been collecting survey data from over
700 economists since that time. Each month, CE surveys the economists to collect views on
expected growth across a large set of countries. The surveyed economists typically provide a
forecast of real GDP growth (and components) for the next two calendar years. A key benefit of
this data source is that it is ‘point-in-time’: the forecasts of economists are included in the CE
datasets and they are never changed. In addition, prior research has shown that, with few
exceptions, the CE forecasts are less biased and more accurate in terms of mean absolute error
24
and root mean square error relative to forecasts from the OECD and IMF (Batchelor, 2001). We
use the average GDP forecast across the CE survey participants for each country. Similar to our
focus on 12-month ahead earnings and sales forecasts from sell side analysts, we combine the
one year ahead and two year ahead GDP growth forecasts by placing less (more) weight on the
one (two) year ahead GDP growth forecast as the forecasting month gets closer to the end of the
first year. This 12 month-ahead forecast of GDP growth has a natural economic interpretation as
it is measured in percentage points of expected growth. For our second measure of country level
performance we compute a rolling six month stock return momentum measure for each country.
Table 2 reports the distribution of both country level measures. We have real GDP
growth forecasts for all 56 countries but only have stock return data from MSCI for 34 countries.
The requirement of MSCI country level returns reduces our global sample by about 15 percent.
Not all countries are covered in all years for the CE measures as countries were added during the
1990s as CE was growing its coverage. Given that our time period spans the 1998 to 2010
period, it is not surprising to see that the countries with the highest average level are
concentrated in the developing markets (e.g., China and India). Across all countries, however,
As described in section 2.3, we combine the country level forecast with the firm level
geographic exposures (from the most recent fiscal year) to compute two measures labelled as
and respectively where the superscript reflects the information source
for the expected country level performance. A key difference between these two measures is that
reflects expectations only of real GDP growth, whereas reflects
aggregate expectations of changes in dividend growth (inclusive of real GDP growth and other
25
affects) and changes in discount rates. The latter measure is likely to be more responsive to
changes in the macroeconomic environment, but will be affected by both fundamental and non-
fundamental factors (see e.g., Cutler, Poterba and Summers, 1989) which may decrease its
Panel B of table 1 notes that the average value of ( ) is 3.23
(0.005) consistent with most countries experiencing GDP growth during this time period, but
muted stock returns over the same period. More important, however, is the large standard
deviation in these measures, 2.35 (0.036), and large inter-quartile range, 2.57 (0.028)
respectively. Thus, ex ante, there should be sufficient power to exploit both time series and cross
sectional variation in to help forecast firm fundamentals, analyst revisions and future
stock returns.
All of our fundamental data to compute the measures described in section 2.4 are derived
from annual (or interim) financial statements collected by Compustat for US firms and FactSet
Fundamentals for non-US firms. Analyst forecast data are sourced from I/B/E/S for both US and
non-US firms. Our market data are obtained from CRSP for US firms and Compustat Global for
non-US firms. We include all firms in our analysis with non-missing data to compute ,
and make no exclusions on the basis of industry membership. Our primary sample starts in 1998
due to our inability to obtain geographic segment data from FactSet Fundamentals prior to 1998.
4. Results
Table 3 reports the regression coefficient estimates of equation (1). We estimate this
regression separately each year for each of the twelve industry groups listed in table 1. The
26
reported coefficients are then averaged across years and industry groups. Standard errors are
based on the time series and cross sectional variation in industry-year estimates with corrections
for serial correlation. Inferences are similar if we instead estimate pooled regressions and
compute standard errors cluster by firm and year. We estimate equation (1) for both
and .
Consistent with prior research we see that profitability is mean reverting as evidenced by
the %& coefficient of about 0.582 (0.92) for the () forecasting equations. As expected,
we also see that the level of future profitability is decreasing (increasing) in )* (,- ). We
also document a strong negative relation between /0 and future profitability and a strong
positive relation between dividend payment status and dividend yield and future profitability.
All of these results are consistent with recent research (e.g., So, 2012 and Hou, van Dijk and
Zhang, 2012).
Consistent with P1 we find a positive and significance coefficient on both and
. The % coefficient of 0.002 (0.001) for one year ahead ROA (SOA) using
has a clear economic interpretation: a one percentage point change in real GDP
forecasts is associated with an additional 20 (10) basis points of () in the following
year controlling for other known determinants of profitability. Likewise, the % coefficient of
0.078 (0.321) for one year ahead ROA (SOA) using also has a clear economic
interpretation: a one percentage point increase in country level returns is associated with an
additional 8 (32) basis points of ROA (SOA) in the following year controlling for other known
determinants of profitability.
To make stronger inferences about the predictive content of , we compare the
absolute forecast errors described in equations (2a) and (2b). For each industry-year group we
27
estimate equation (1) with and without the respective variable. We re-estimate
equation (1) each year from 2006 to 2011 adding one additional year as we move forward in time.
We then compare differences in forecast errors on a pooled and industry grouping basis.
When we include into the forecast, we find that the average of the median absolute
forecast error of ROA (SOA) is significantly lower by 2 basis points (not statistically different).
When we include into the forecast, we find that the average of the median absolute
forecast error of ROA (SOA) is significantly higher by 2 basis points (significantly lower by 12
basis points). While the magnitude of the reduction in forecast error seems small in economic
terms, and is actually worse for the ROA forecasts when using , it is consistent with
previous research. For example, Fairfield and Yohn (2001) document that a forecasting model
for changes in return on net operating assets that includes profit margins and asset turnover
relative to a forecasting model that excludes this information, was more accurate by a magnitude
of 0.0003 (0.0002) for the average (median) paired difference. Further, Fairfield, Sweeney and
Yohn (1996) document that the median improvement in out-of-sample forecast accuracy by
between 5 and 10 basis points (relative to book equity) for a large sample of US firms over the
1981-1990 time period. Thus, the magnitude of forecast accuracy improvement we document is
For our tests using real GDP forecasts we also run additional analysis attempting to
decompose GDP into its components (untabulated for sake of brevity). Corporate profits are a
individual company level earnings forecasts to the country level we are able to form crude
28
aggregate all 12 month-ahead forecasts for each firm in a given country and then compute
monthly revisions in this aggregate earnings forecast. We use these country level corporate
A]^.O]A`
earnings growth forecasts to compute a new measure: . For our sample, the
average cross-sectional correlation in expected country performance across real GDP growth
forecasts and our imputed country level corporate earnings growth forecast is only 0.035.
Further, the average firm level cross-sectional correlation between and
A]^.O]A`
is 0.094. Thus, multicollinearity is not likely to be an issue when including
both measures. Including this new variable in our estimation of equation (1) does not affect the
magnitude, nor significance, of in the ROA specification. Assuming our aggregation
of firm level earnings forecasts to the country level is valid, this suggests that it is growth in
other components of GDP (e.g., employee wages and other aggregate sources of income) are
driving the improved forecasts of ROA. We note, however, that in the SOA specification
A]^.O]A`
is significant, but is not, suggesting that expectations of corporate
Table 4 panels A and B (C and D) reports the regression coefficient estimates of equation
(3) for earnings (sales) forecast revisions. We estimate these equations separately for each month
and reported coefficients are averaged across months. Standard errors are based on the time
Consistent with prior research we find that analyst revisions are strongly serially
correlated. The %& coefficient is about 0.1 for the next three months for both earnings and sales
forecast revisions. Likewise, analyst revisions are also strongly related to past returns (%. is
strongly significant for both earnings and sales revisions) and market expectations for growth (%(
29
and %+ are significant in most months). We also find a strong negative relation between /0
and future earnings revisions consistent with Bradshaw, Richardson and Sloan (2001).
Consistent with P2, we find a robust positive association between and future
analyst revisions for both earnings (panel A of Table 4) and sales (panel C of Table 4) for the
following three months. However, we only find a robust positive association between
and future analyst revisions of sales (panel D of Table 4). There is no evidence of a
lagged response in sell-side analyst earnings forecasts with respect to (panel B of
Table 4).
Having established the relative ability of to forecast both firm fundamentals and
sell-side analyst earnings revisions, we now assess whether it has any predictive value for equity
returns. Table 5 reports regression estimates of equation (5). We estimate equation (5) every
month and report averages of estimated regression coefficients. Standard errors are based on the
time series variation in estimated regression coefficients. Equation (5) is estimated three times
each month to assess the predictive content of our included explanatory variables over the next
three months. We report four panels in table 5 to correspond to the two weighting schemes
described in section 2.4.3 and the two measures: and .
Consistent with prior research we see that equity returns are (i) strongly positively
associated with )* and 0G/ , (ii) negatively correlated with the most recent stock returns,
* , (iii) negatively correlated with /0 , (iv) positively associated with IJ
(weakly so in the value-weighted regressions), (v) weakly negatively related with ,- , and (vi)
unrelated with )I . The weak association for IJ in the value weighted specification
is partly attributable to our use of a global sample of securities. Asness (2011) notes that
30
momentum is not evident in Japan, and its relation with future returns is conditional on the
inclusion of other characteristics. Indeed, when we exclude Japanese securities from our sample,
we find that IJ becomes more significant, albeit still at marginal levels. The weak
association between )I and future returns is consistent with much prior research (e.g., Fama
and French, 1992). Consistent with P4 we find a positive association between our primary
variable of interest, , and future equity returns for the next three months, but note that
(i) the significance of is marginal in the value weighted specification (panel A of
Table 5), and (ii) the significance of is limited to the first month (panels B and D of
Table 5).
To assess the economic significance of the relation between and future equity
returns, we examine portfolio level returns in table 6. As discussed in section 2.4.3, every month
we form 25 portfolios based on a conditional sort, first on and then on the respective
measure. We then compute the value weighted return for each of these 25 portfolios
over the next month. We also report a hedge return as the difference in the average portfolio
return across the extreme quintiles. Test statistics are reported based on the time series
The first row of panel A of table 6 reports the average for firms across the five
quintiles. These market values have been adjusted to 2011 dollars using a GDP
deflator to allow comparison across time. The smallest quintile contains securities with a market
capitalization of about $15 million and the largest quintile contains securities with a market
capitalization of about $2.74 billion. Clearly there is a very large difference in the economic
importance of securities across the five quintiles. Table 6 shows that, across the five
quintiles, the value weighted hedge return associated with is significant for the next
31
month, with the magnitude of the relation varying between 160 basis points per month for the
middle quintile and 87 basis points per month for the largest quintile. A stronger pattern is
observed for : the value weighted hedge return ranges between 183 basis points for
the second smallest quintile to 85 basis points per month for the largest quintile. We interpret
this as strong evidence in support of P4. While these portfolios are all buy-and-hold for a given
month, they would require rebalancing at a monthly frequency. Absent clean data on expected
transaction costs for an institutional investor in a global context, we are unable to make strong
see very significant intercepts which translate into economically and statistically significant
conditional Sharpe Ratios (see last row of table 7). These large conditional Sharpe ratios suggest
that the portfolio returns documented in table 7 cannot be explained by the set of seven risk
factors. Of course, it is always possible there is an unidentified risk factor which time varies
with our Z
/[
returns. Of the included risk factors, there is strong evidence that Z
/[
is
positively associated with \* and , and weakly negatively associated with Z3 .
Specifically, the regression coefficients suggest that the returns to a portfolio exploiting
geographic exposures tends to outperform when (i) the overall equity market is doing well, (ii)
recent ‘winners’ have been performing well, and (iii) underperform when ‘value’ firms out-
perform.
A necessary condition for to be able to forecast future firm performance and
returns is that there is cross-sectional dispersion in either (i) the 12 month ahead real GDP
32
forecasts from CE or (ii) the rolling six month country level returns we gather from MSCI.
Table 2 documents considerable variation in these forecasts across countries and time. Our
empirical analysis thus far is all cross-sectional. To identify whether there is additional
information content in the time series of our country level forecasts, we compute the across
country dispersion in the (i) 12 month ahead real GDP forecasts each month, and (ii) rolling six
month country level MSCI returns. We then associate this measure of dispersion with the
predictive ability of for revisions and returns in the following month(s). Specifically,
we correlate the % coefficients from equation (4) and equation (5) with this dispersion measure
for the next six months. Our prior is that when there is greater dispersion in beliefs across
expected country level performance, there is greater potential for information content (assuming
that forecasting skill of analysts providing forecasts is not negatively associated with this
aggregate dispersion). We find some evidence that country level dispersion is associated with
the future predictive ability of . For example, (i) the correlation between dispersion in
expected country level performance over the past six months and the % coefficient in equation
(4) averages 0.10 (0.2) for the next three months of analyst earnings (sales) revisions, and (ii) the
correlation between dispersion in expected country level performance over the past six months
and the % coefficient in equation (5) is between 0.26 and 0.12 for the next three months. These
relations are strongest for the sample of non-domestic firms and only hold for the
measure. We interpret this evidence as providing limited support for P3a and P5a.
between both measures and future analyst revisions and future stock returns. In panel
A of Table 8 we re-estimate equation (4) for various cross sectional partitions. For the sake of
33
brevity we only report the % coefficient. In panel B of Table 8 we report the hedge return using
various measures by varying the set of included regression variables and we also consider
characteristic adjusted returns as per Daniel, Grinblatt, Titman and Wermers (1997). We do not
re-estimate equation (3) across these partitions as we have no prior as to why the relation
between expected country level performance and future firm fundamentals would vary along the
dimensions we examine.
The partitions we examine are as follows: (i) firm size using both in-sample breakpoints
and NYSE breakpoints, (ii) analyst following, (iii) comparing ‘domestic’ to ‘non-domestic’ firms,
(iv) considering firm complexity within the set of ‘non-domestic’ firms using the reported
number of geographic segments as our measure of complexity, and (v) the effect of inclusion in
the MSCI indices for the stock return analyses. As outlined in P3b and P5b, we expect the
predictive ability of to vary across the size partitions and analyst following partitions,
with stronger results expected for the smaller, less followed firms. We also expect the predictive
ability of to be greater for firms that are more complex, consistent with the prior
literature on limited attention and economic linkages (see e.g., Cohen and Frazzini, 2008 and
Table 8 summarizes the results. We show the full sample results in bold at the top of each
panel for comparison with the earlier tables, and the fraction of the full sample included in each
partition in parentheses beside the partition label. For the sake of brevity we only report the %
For the analyst revision tests we find little evidence of variation in the % coefficient
across firm size and analyst following partitions for the measure. If anything, the
relation is stronger for the middle size and analyst following group. For the
34
measure, we continue to see very little evidence of a relation with analyst earnings revisions
consistent with the results in panel B of Table 4. But we do see some evidence that the relation
with analyst sales revisions is stronger for the sub-sample with the lowest analyst following.
Splitting the sample on the basis of ‘domesticity’ reveals similar economic magnitudes for
‘domestic’ and ‘non-domestic’ firms in the relation between future analyst forecast revisions of
earnings and sales for , and a stronger relation for the non-domestic firms using
. Within the sample of ‘non-domestic’ firms when we partition firms on the basis of
‘complexity’ we find a stronger relation for the complex sub-sample with the
measure, but a much weaker relation for the complex sub-sample with the measure.
Obviously the cross-sectional partitions we use are not independent, so to help control for
multiple dimensions we further partition the sample of complex ‘non-domestic’ based on firm
size and domicile. For the larger complicated non-domestic firms (5.3% of the full sample) we
find that the relation between analyst revisions of sales and earnings are more strongly associated
with both measures, however relations are not statistically significant for the
measure. While this partition covers only 5.3% of the full sample (still over 800
firms for the average month) these are the largest multi-national firms in the world comprising
between analyst earnings and sales revisions and for US domiciled firms, and a
statistically negative relation for earnings revisions and . This lack of results for US
firms is consistent with our priors for several reasons: (i) as discussed in section 3.1 we expect
greater measurement error in the geographic sales exposures for US firms, (ii) US firms are
known to be more liquid and capital market participants are more sophisticated relative to other
markets, and (iii) when limiting the sample to ever smaller numbers of firms we reduce the
35
natural variation in the respective measure making it more difficult to isolate any effect. Overall,
For the stock return analyses reported in panel B of Table 8, we find reasonably
consistent evidence of a weaker relation between both measures and future stock
returns for the largest sub-sample based on market capitalization and for firms with the greatest
analyst following. For example, across both measures we find hedge returns that are,
on average across the three measures of hedge returns, 51 (66) percent lower for the largest firms
using NYSE breakpoints (firms with highest analyst following). We further find that hedge
returns are, again on average across the three measures of hedge returns, 2.65 (1.88) times
greater for firms that are not included in the MSCI country indices within the full (‘domestic’)
sample. This result is consistent with our prediction of market segmentation where investors
trading on the basis of macroeconomic information are likely to trade baskets of securities (e.g.,
MSCI country indices) which have greater sensitivity to that macroeconomic information. We
find that hedge returns are on average significant for , and not for , when
limiting the sample to ‘non-domestic’ firms (and in that case the hedge returns are only 32
percent as large for the ‘non-domestic’ firms relative to the ‘domestic’ firms). However, we find
strong evidence that the hedge returns are on average 2.60 times greater for the complex ‘non-
domestic’ firms relative to the simple ‘non-domestic’ firms and the hedge returns for the
As discussed above, our cross-sectional partitions are not independent, so we also further
partition the sample of complex ‘non-domestic’ based on firm size and domicile. For the larger
complicated non-domestic sample we find hedge returns that are on average 2.14 times larger
than that for the smaller complicated non-domestic firms. When we split the sample of complex
36
‘non-domestic’ firms on the basis of domicile, consistent with the analyst revision results, we
find inconsistent results across the measures. For we find a negative
(positive) relation with future hedge returns for US (Non-US) firms, but the relation is only
significant for the 7-factor alpha hedge returns. For we find a consistently positive
relation with future hedge returns and while the hedge return is economically larger for the US
sample, it is statistically stronger for the non-US sample, most likely due to the difference in
sample size. As discussed above, we have priors for a weaker relation for US firms, but not a
prior for a negative return relation. Collectively, the results in panel B of Table 8 are consistent
with P5b, but the results are stronger for non-US firms.
decompose both and for sub-samples of ‘non-domestic’ firms which
have an economically meaningful exposure to foreign markets as measured by either (i) foreign
sales exceeding 20 percent of total sales, or (ii) greater than three geographic segments. For
example, Mulberry PLC is a UK listed company, and in August 2010, as described in Appendix I,
Mulberry reports sales across the following regions: (i) Europe 90%, (ii) Asia 5.1%, (iii) North
America 3.2%, and (iv) rest of the world for 1.7%. Using GDP weights, UK (Mulberry’s home
country) accounts for 11.6% of Europe, so Mulberry reports 10.44% sales from its home country
(0.116 x 0.9) and 98.56% spread across foreign countries. In this way, for every firm with
into their respective home country ( a@, ) and foreign country
a natural scale reflecting the relative importance of the home country relative to foreign countries.
37
We re-estimate equations (3) and (5) with these additive decompositions and find that both
a@, and b@?cd, are positively associated with future analyst revisions and
future stock returns. It is also important to remember the results for US firms disclosed
separately in Table 8. When using the measure there is still a robust positive
relation with future stock returns. This relation can only be explained by foreign country
We have also included both and into estimates of equations (1),
(3) and (5). Given that both measures use the same set of country exposures a possible concern
for this analysis is multicollinearity. For our sample the average cross-sectional correlation in
expected country performance across real GDP growth forecasts and rolling six month country
level stock returns is only 0.067. Further, the average firm level cross-sectional correlation
between and is 0.176. Thus, multicollinearity is not likely to be an issue
when including both measures. We find that both measures are significant across regression
equations (1), (3) and (5), consistent with the tabulated results. We also find a slight increase in
explanatory power from the inclusion of both measures, suggesting that they provide orthogonal
5. Conclusion
level forecasts. Using a large sample of publicly traded firms around the world, we show that
combining geographic segment sales disclosures and forecasts of country level performance
generates significant out of sample improvement in forecasting firm level profitability. We also
find that sell side analysts are slow to incorporate this information into their forecasts. Finally,
we find that stock prices are also slow to incorporate this information.
38
Our results suggest potential benefits to detailed contextual analysis which seeks to
identify value drivers that are external to the firm. Combining firm specific exposures to these
value drivers with a directional view on the value driver should create improvements in our
ability to understand and hopefully forecast future firm fundamentals and associated risks.
39
References
Asness, C., (2011). “Momentum in Japan: the exception that proves the rule.” The Journal of
Portfolio Management, 37 (4), 67-75.
Asparouhova, E., H. Bessembinder, and I. Kalcheva, 2010, Liquidity baises in asset pricing tests,
Journal of Financial Economics, 96:215–237.
Balakrishnan, R., T. S. Harris, and P. K. Sen. (1990). The predictive ability of geographic
segment disclosures. Journal of Accounting Research, 28, 305-325.
Bartov, E., and M. Bodnar (1994). “Firm valuation, earnings expectations, and the exchange-rate
exposure effect”, Journal of Finance, 49, 1755-85.
Bartram, S. M., and G. M. Bodnar, 2012. “Crossing the lines: The conditional relation between
exchange rate exposure and stock returns in emerging and developed markets” Forthcoming,
Journal of International Money and Finance.
Basu, S., S. Markov, and L. Shivakumar (2010). Inflation, earnings forecasts and post-earnings
announcement drift. Review of Accounting Studies, 15, 403-440.
Batchelor, R. (2001). How useful are the forecasts of intergovernmental agencies? The IMF and
OECD versus the consensus? Applied Economics, 33, 225-235.
Blume, M. and R. Stambaugh. (1983). “Biases in computed returns: An application to the size
effect”. Journal of Financial Economics, 12, 387-404.
Bradshaw, M., S. Richardson, and R. Sloan. 2001. Do analysts and auditors use information in
accruals? Journal of Accounting Research 39, 45-74.
CHEN, N., R. ROLL, and S. A. ROSS (1986). “Economic forces and the stock market.” The
Journal of Business 59 : 383-403.
Chordia, T., and L. Shivakumar (2005) Inflation illusion and post earnings announcement drift.
Journal of Accounting Research, 43, 521-556.
Cochrane, J. H., (2000). Asset Pricing, Princeton University Press.
Cochrane, J. H., (2010). Discount rates. Journal of Finance 66, 1047-1108.
Cohen, L., Frazzini, A., 2008. Economic links and predictable returns. Journal of Finance 63,
1977–2011.
Cohen, L., and D. Lou, (2012). Complicated firms. Journal of Financial Economics, 104, 383-
400.
COLLINS, D. W. (1975). “Predicting earnings with sub-entity data: Some further evidence.”
Journal of Accounting Research 14, 163-177.
40
Core, J., W. Guay and T. Rusticus (2006). Does weak governance cause weak stock returns?
An examination of firm operating performance and investors’ expectations. Journal of Finance,
56, 655-87.
Cutler, D.M., J.M. Poterba, and L.H. Summers. 1989. “What Moves Stock Prices?” Journal of
Portfolio Management, vol. 15, no. 3 (Spring):4–12.
Daniel, K., M. Grinblatt, S. Titman and R. Wermers, (1997). Measuring mutual fund
performance with characteristic based benchmarks. Journal of Finance, 52, 1035-1058.
Dimson, E., (1979) Risk measurement when shares are subject to infrequent trading. Journal of
Financial Economics 7, 197-226.
Dichev I., and W. Tang (2009). Earnings volatility and earnings predictability, Journal of
Accounting and Economics 47, 160-181.
Fairfield, P. M., R. J. Sweeney, and T. L. Yohn (1996). Accounting classification and the
predictive content of earnings. The Accounting Review, 71, 337-355.
Fairfield, P.M., Whisenant, J.S., Yohn, T.L., 2003. Accrued earnings and growth: implications
for future profitability and market mispricing. Accounting Review 78, 353–371.
Fairfield, P. M., and T. L. Yohn (2001). Using asset turnover and profit margin to forecast
changes in profitability. Review of Accounting Studies, 6, 371-385.
Fama, E., and K. French, 1997. Industry costs of equity. Journal of Financial Economics 43,
153–193.
FAMA, E., and K. FRENCH (1992). “The cross-section of expected stock returns.” Journal of
Finance 47: 427-465.
FAMA, E., and K. FRENCH (1993). “Common risk factors in the returns of stocks and bonds.”
Journal of Financial Economics 33: 3-56.
Fama, E., and K. French, 1995. Size and book-to-market factors in earnings and returns. Journal
of Finance, 50, 131-155.
Fama, E., and K. French, 2000. Forecasting profitability and earnings. Journal of Business 72,
161–175.
Fama, E., and K. French, 2008. Dissecting anomalies. Journal of Finance 63, 1653–1678.
FAMA, E. and J. MACBETH (1973). “Risk, Return and Equilibrium: Empirical Tests.” Journal
of Political Economy, 81 (3): 607-636.
41
Foster, G. (1975). “Security price revaluation implications of sub-earnings disclosure.” Journal
of Accounting Research 13, 283-292.
French, K. R., G. W. Schwert, and R. F. Stambaugh. (1987). “Expected stock returns and
volatility.” Journal of Financial Economics, 19, 3-29.
Hirhsleifer, D., and S-H. Teoh, (2003). Limited attention, information disclosure, and financial
reporting. Journal of Accounting and Economics, 36, 337-386.
Hong, H., and J. Stein (1999). A unified theory of under-reaction, momentum trading, and
overreaction in asset markets. Journal of Finance, 54, 2143-2184.
Hou, K., M. A. van Dijk and Y. Zhang. (2012). The implied cost of capital: A new approach.
Journal of Accounting and Economics, 53, 504-526.
Hughes, J., J. Liu, and W. Su. 2008. “On the Relation between Predictable Market Returns and
Predictable Analyst Forecast Errors.” Review of Accounting Studies, vol. 13, no. 2–3
(September): 266–291.
Kadan, O., L. Madureria, R. Wang, and T. Zach, 2012. “Analyst’s industry expertise”. Journal
of Accounting and Economics, forthcoming.
HUGHES, J., J. LIU, AND W. SU. 2008. “ON THE RELATION BETWEEN PREDICTABLE
MARKET RETURNS AND PREDICTABLE ANALYST FORECAST ERRORS.” REVIEW
OF ACCOUNTING STUDIES, VOL. 13, NO. 2–3 (SEPTEMBER): 266–291.
KINNEY, W. R., (1971). “Predicting earnings: Entity vs. sub-entity data.” Journal of Accounting
Research 9, 127-136.
Kirby, C., and B. Ostdiek. 2012. “It’s all in the timing: simple active portfolio strategies that
outperform naïve diversification”. Journal of Financial and Quantitative Analysis, 47, 437-467.
Korajczyk, R., R. Sadka. (2004). “Are momentum profits robust to trading costs?” Journal of
Finance, 59, 1039–1082.
Lang, L. H. P., R. Stulz and R. A. Walkling. (1991). “A test of the free cash flow hypothesis.”
Journal of Financial Economics, 29, 315-335.
42
Li, Q., M. Vassalou, and Y. Xing. (2006). “Sector investment growth rates and the cross section
of equity returns”. Journal of Business, 79, 1637-65.
Liew, J., and M. Vassalou. (2000). “Can book-to-market, size and momentum be risk factors that
predict economic growth?” Journal of Financial Economics, 57, 221-245.
Maillard, S., T. Roncalli, and J. Teiletche. 2010. “On the properties of equally weighted risk
contribution portfolios”. Journal of Portfolio Management, 36, 60-70.
Penman, S., Zhang, X., 2002. Accounting conservatism, the quality of earnings, and stock
returns. The Accounting Review 77, 237–264.
Pontiff, J., 2006. Costly arbitrage and the myth of idiosyncratic risk. Journal of Accounting and
Economics 42, 35–52.
Richardson, S., R. Sloan, M. Soliman, and I. Tuna. (2005). Accrual reliability, earnings
persistence and stock prices. Journal of Accounting and Economics, 39, 437–485.
Richardson, S. A., I. Tuna and P. Wysocki (2010). Accounting anomalies and fundamental
analysis: A review of recent research advances. Journal of Accounting and Economics 50, 410–
454.
Roberts, C. B., (1989). Forecasting earnings using geographical segment data: Some UK
evidence. Journal of International Financial Management and Accounting, 1, 130-151.
Scholes, M., and J. Williams. 1977. Estimating betas from non-synchronous trading. Journal of
Financial Economics 5, 309-327.
Silhan, P. A., (1983) “The effects of segmenting quarterly sales and margins on extrapolative
forecasts of conglomerate earnings: Extension and replication.” Journal of Accounting Research
21, 341-347.
Sloan, R., 1996. Do stock prices fully reflect information in accruals and cash flows about future
earnings? The Accounting Review 71, 289–316.
43
So, E., (2012). A new approach to predicting analyst forecast errors: Do investors overweight
analyst forecasts? Working paper, Stanford University.
Soliman, M. T. (2008) The Use of DuPont Analysis by Market Participants. The Accounting
Review 83, 823-853.
Thomas, W., ‘‘A test of the market’s mispricing of domestic and foreign earnings,’’
Journal of Accounting and Economics (December 1999), pp. 243–67.
Vassalou, M. (2003). “News related to future GDP growth as a risk factor in equity returns”,
Journal of Financial Economics, 68, 47-73.
Xie, H., 2001. The mispricing of abnormal accruals. The Accounting Review 76, 357–373.
44
Appendix I: Calculation of efghij for Mulberry Group PLC
In the fiscal year ended on March 2010, Mulberry's sales are from the following regions: (i)
Europe 90%, (ii) Asia 5.1%, (iii) North America 3.2%, and (iv) ‘Rest of the World’ 1.7%. We
use this exposure matrix to calculate for each month from August 2010 to July 2011.
For example,
for Mulberry in August 2010 is calculated as:
= % I
J
l ×
[
J
l] + % I ,
×
[
,] + % I 0
Iℎ
,
×
[
0
Iℎ
,] + % I I Iℎ o
8
×
[
I Iℎ o
8]
To compute our measures of expected performance across the geographic regions we use
Consensus Economics (CE) GDP growth forecast data. We calculate the one year head GDP
growth forecast for each country as the weighted average of the mean forecasts for the first and
second years, using (13-m)/12 as the weight for the first year forecast, where m is the forecast
month in the calendar year. For example, In March 2009, the mean US GDP growth forecasts for
2009 and 2010 are -2.82% and 1.72%, respectively. The one year ahead GDP growth forecast in
p &
March 2009 is calculated as × (−2.82%) + × 1.72% = −2.30%.
& &
For regions that comprise multiple countries, we assume that each company’s operations across
countries are directly proportional to the relative GDPs across these countries. For example, to
compute the expected performance for Europe as of August 2010 (i.e.,
[
J
l]) we:
1) Calculate the total 2009 GDP of European countries using GDP data from IMF World
Economic Outlook Databases (http://www.imf.org/external/ns/cs.aspx?id=28).
2) Calculate the GDP percentage of each of the 28 countries in Europe with CE GDP growth
forecast data.
45
Italy 0.113 United Kingdom 0.116
Latvia 0.001 UKraine 0.007
3)
[
J
l], as of August 2010, is then calculated as the sum of the
individual country level one year ahead GDP forecast, as described above, multiplied by the
GDP percentages in the above table.
46
Appendix II: Variable definitions
Variable Description
I Total assets as at the end of the fiscal year (in USD millions).
Equity market beta estimated from a rolling regression of 60 months of
)I
data requiring at least 36 months of non-missing return data.
Book-to-market ratio computed as the ratio of common equity to equity
market capitalization, both measured at the fiscal period end date for the
)*
most recent and available fiscal period prior to month t. See Figure 1 for
more details.
O
8G ln O Q , where G is Industrial Production Index at the end of month t
QTU
from the Board of Governors of the Federal Reserve System (INDPRO),
available at the St Louis Fed web site:
http://research.stlouisfed.org/fred2/
8 Change in risk premium, − X , where is the difference
between the Moody’s Seasoned BAA Corporate Bond Yield from the
Board of Governors of the Federal Reserve System (BAA) and the 10-
Year US Treasury constant maturity rate from the Board of Governors of
the Federal Reserve System (GS10). BAA and GS10 are available at the
St Louis Fed web site: http://research.stlouisfed.org/fred2/.
8* Change in term structure,* − *X, where * is the difference between
the 10-Year US Treasury constant maturity rate (GS10) and the 2-Year
US Treasury constant maturity rate (GS2), both from the Board of
Governors of the Federal Reserve System. Both GS10 and GS2 are
available at the Louis Fed web site: http://research.stlouisfed.org/fred2/
An indicator variable equal to one for firms that pay dividends and zero
/_/,5
otherwise.
/,5_7,8 Dividends per share divided by stock prices.
An indicator variable equal to one for firms that have negative earnings
/_3
before extraordinary items and zero otherwise.
The change of net operating assets, scaled by total assets, where net
operating assets are calculated as operating assets (total assets less the sum
/0
of cash and investments) minus operating liabilities (total liability minus
total debt).
An indicator variable equal to one for firms that have no foreign sales and
/
*G
zero otherwise.
Z3 Monthly mimicking global (developed market) factor portfolio return to
the value factor, obtained from Ken French’s website.
The sum product of a firm’s geographic sales exposure to a country and
the one year ahead Consensus Economics GDP growth forecast of the
country. The geographic sales data are extracted from the most recent
annual report prior to month t (ensuring at least a four month gap between
the end of the fiscal year and month t). See Section 2.3 for details and
Appendix I for an example.
47
The sum product of a firm’s geographic sales exposure to a country and
the most recent six month return from the MSCI index for that country.
The geographic sales data are extracted from the most recent annual report
prior to month t (ensuring at least a four month gap between the end of the
fiscal year and month t). See Section 2.3 for details and Appendix I for an
example.
Equity market capitalization (in USD millions).
Variable Description
\* Monthly excess (to risk free rate) global market return, obtained from Ken
French’s website.
Average global (developed market) return on the two high prior return
portfolios minus the average return on the two low prior return portfolios,
obtained from Ken French’s website.
The average monthly equity return inclusive of dividends from month t-6
IJ
to month t-1.
Earnings-to-Price ratio computed as the ratio of net income before
extraordinary items to equity market capitalization, both measured at the
0G/
fiscal period end date for the most recent and available fiscal period prior
to month t. See Figure 1 for more details.
* Monthly equity return inclusive of dividends.
Return on assets computed as the ratio of net income before extraordinary
items to average total assets.
This is the monthly revision in median consensus sell-side analyst
earnings or sales forecasts. Earnings forecast revision is calculated as
[O&P,QRS ]
5,,,F = ln [O& , where
[
12, ] is a calendar
P,QRSTU ]
48
Figure 1
Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
2010 2011
vwtj , xywtj , x_zi{{j , f|} x_x~j are all measured for the 12 months ended Dec 31, 2010.
ej f|} x~_~z}j are measured using book equity, dividends and prices as at Dec 31, 2010.
~
j is measured as at Dec 31, 2010.
* The regressions reported in table 3 are based on firm-year observations. Thus, while we are able to measure every month
we only use for the period that coincides with the end of the previous year. This is to ensure that all of the explanatory
variables are measured prior to the future profitability, F , that we are trying to forecast.
49
Figure 2
Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
2010 2011
* The regressions reported in tables 4 and 5 are based on firm-month observations. We are careful to ensure that all explanatory
variables are known to the analysts and the market at month t.
50
Table 1 Summary Statistics
51
Panel A: Country Distribution (continued)
Number of Firm-Years Percentage
Singapore 6,995 1.84
Slovakia 206 0.05
Slovenia 162 0.04
South Africa 4,682 1.23
Spain 2,049 0.54
Sweden 4,804 1.26
Switzerland 3,413 1.05
Taiwan 16,549 4.34
Thailand 6,171 1.62
Turkey 2,823 0.74
Ukraine 37 0.01
United Kingdom 25,024 6.57
United States 79,585 20.89
Venezuela 410 0.11
Total 381,030 100
52
Panel C: Industry Distribution (Fama-French 12 Industries)
Number of
Percentage
Firm-Year
Consumer Non-Durables 32,039 8.41
Consumer Durables 12,365 3.25
Manufacturing 50,021 13.13
Oil, Gas, and Coal Extraction and Products 9,964 2.62
Chemicals and Allied Products 13,401 3.52
Business Equipment 57,368 15.06
Telephone and Television Transmission 9,050 2.38
Utilities 9,544 2.50
Wholesale, Retail, and Some Services 33,787 8.87
Healthcare, Medical Equipment, and Drugs 20,139 5.29
Money and Finance 72,230 18.96
Other 61,122 16.04
Total 381,030 100
This table reports summary statistics for the sample. The sample only includes countries with
Consensus Economics GDP forecast data. The sample period is 1998-2010. The sample includes
381,030 firm-years and 4,518,347 firm-months. Panel A reports the distribution of countries of
domicile. Panel B reports firm characteristics. All variables are defined in Appendix II. Panel C
presents industry distribution. The industry classification follows the twelve primary industry
groupings identified in Fama-French (1997).
53
Table 2
Summary Statistics for Country Level Forecasts
54
Peru 4.62 1.41 2.19 0.018 0.040 0.042
Philippines 4.12 1.02 1.38 0.004 0.033 0.045
Poland 3.87 1.28 1.9 0.006 0.036 0.040
Portugal 1.17 1.57 1.12 -0.002 0.028 0.034
Romania 3.13 2.49 4.06
Russia 4.12 2.6 2.33 0.021 0.059 0.055
Singapore 4.48 2.24 1.91 0.006 0.034 0.039
Slovakia 3.97 2.16 2.5
Slovenia 3.16 1.54 0.96
South Africa 3.37 0.98 1.07 0.012 0.024 0.033
Spain 2.05 1.59 2.18 0.002 0.028 0.032
Sweden 2.45 1.26 1.23 0.005 0.037 0.047
Switzerland 1.54 0.82 0.76 -0.000 0.023 0.031
Taiwan 3.96 1.75 1.26 0.003 0.038 0.038
Thailand 4.14 1.59 1.03 0.011 0.039 0.041
Turkey 3.69 2.05 2.45
Ukraine 3.58 3.36 2.72
United Kingdom 1.85 1.18 0.84 0.000 0.020 0.021
United States 2.61 1.28 1.08 0.001 0.023 0.023
Venezuela 2.16 3.94 6.08
This table reports summary statistics for the time series distribution of two sets of country level
forecasts. The first three columns report the means, standard deviations and interquartile ranges
of monthly Consensus Economics real GDP forecasts. These forecasts are the weighted average
of the mean forecast for the first and second years such that the combined forecast always has a
twelve month horizon. The last three columns report the means, standard deviations and
interquartile ranges of country forecasts using MSCI index returns. This return forecast is based
on a rolling six month window of country level returns.
55
Table 3 Macroeconomic Information and Future Firm Performance
vwtj = + etuvwj + vwtj + ej + ~ j + xywtj + x_i{{j + x_x~j + x~_~z}j + j (1)
wtj = + etuvwj + wtj + ej + ~ j + xywtj + x_i{{j + x_x~j + x~_~z}j + j (1)
The reported regression coefficients are mean coefficients from year-industry regressions, weighting each regression by the square
root of sample size for each year-industry (assumes greater precision for those industry-year groups with greater sample size). The t-
statistics (reported in parentheses below coefficient estimates) are based on the standard errors of the coefficient estimates across the
56
year-industry regressions, adjusted for autocorrelation in the annual coefficient estimates based on an assumed AR(1) autocorrelation
structure. Standard errors are multiplied by an adjustment factor, (1 + )/(1 − ) − 2(1 − )/(1 − )& , where n is the number
of years and is the first-order autocorrelation of the annual coefficient estimates. All variables are defined in Appendix II.
57
Table 4 Macroeconomic Information and Future Analyst Forecast Revisions
v~{~i|j = + etuvwj + v~{~i|j + ej + y/j + ei|jj +xywtj + x_i{{j + j (3)
58
Sales Forecast Revision
The reported regression coefficients are mean coefficients from monthly regressions, weighting each regression by the square root of
sample size for each month. All variables are defined in Appendix II. Panel A and B (C and D) report regression results for consensus
analyst Earnings Per Share (Sales Per Share) forecast revisions.
59
Table 5 Macroeconomic Information and Future Stock Returns
vj = + etuvwj + ej + y/j + jfj + ~ j + vj + ei|jj + }ywtj + x_i{{j + j (5)
60
(1/) weighted regressions
Panel C : Consensus Economics Real GDP Forecasts (etuvwu j ) [N=2,087,585 firm-months]
Adj. R2
k=1
Coefficient 0.0074 0.0016 0.0015 0.0237 0.0018 -0.0007 -0.0274 0.0712 -0.0073 -0.0039 0.053
(t-statistic) (1.80) (2.44) (2.69) (5.12) (0.88) (-2.35) (-4.07) (3.82) (-4.63) (-4.10)
k=2
Coefficient 0.0070 0.0017 0.0013 0.0237 0.0014 -0.0005 0.0036 0.0521 -0.0063 -0.0034 0.052
(t-statistic) (1.70) (2.62) (2.30) (4.87) (0.57) (-1.86) (0.70) (3.16) (-4.24) (-3.68)
k=3
Coefficient 0.0062 0.0017 0.0015 0.0202 0.0011 -0.0005 0.0144 0.0500 -0.0065 -0.0028 0.051
(t-statistic) (1.49) (2.57) (2.55) (4.17) (0.44) (-1.60) (2.82) (3.10) (-4.28) (-3.12)
The reported regression coefficients are mean coefficients from monthly regressions. In computing averages and standard errors each
cross section is weighted by the square root of sample size given more weight to the largest cross sections. Within each cross section
security level returns are value weighted in panels A and B (where the weights are the square root of the securities market
capitalization in USD) and 1/ & weighted in panels C and D (the weights are inversely proportional to the historical volatility of
idiosyncratic returns). All other variables are as defined in Appendix II.
61
Table 6 Future Stock Returns Related to efghij Across Size Quintiles
Size Quintile
Small 2 3 4 Large
eut 14.23 52.99 149.31 485.27 3070.49
Low 0.0146 0.0055 0.0059 0.0053 0.0029
2 0.0154 0.0114 0.0109 0.0093 0.0094
MACRO 3 0.0084 0.0115 0.0069 0.0071 0.0035
Quintile 4 0.0193 0.0092 0.0078 0.0076 0.0042
High 0.0319 0.0231 0.0205 0.0181 0.0116
For each month, stocks are first sorted into five equal groups based on market capitalization
(in USD). Then, within each size group, stocks are further sorted based on . The
portfolio returns are value weighted (where the weights are computed as market capitalization,
in USD). The ‘Hedge’ return is the difference between the average portfolio returns across
extreme quintiles. The Sharpe ratio is calculated following Lewellen (2010). Return are
reported in decimal units (i.e., 0.01 is 1%). MCAP is median market capitalization in USD,
adjusted to the PPI level in September 2011.
62
Table 7
Ex Post Return Analysis
For each month, stocks are sorted into five equal groups based on . The portfolio
returns are value weighted (where the weights are market capitalization, in USD). The
Z
/[
return is the difference between the average portfolio returns across extreme
quintiles for the following month. The Sharpe ratio is calculated as the ratio of the annualized
return (as measured by the intercept) relative to the annualized standard deviation, following
Lewellen (2010). The remaining variables are defined in Appendix II.
63
Table 8
Cross sectional partitions
Panel A: Macroeconomic information and one month ahead analyst revisions
etuvwu j etuvweu
j
EPS forecast revisions Sales forecast revisions EPS forecast revisions Sales forecast revisions
Full Sample: Coefficient T-statistic Coefficient T-statistic Coefficient T-statistic Coefficient T-statistic
0.0007 (3.27) 0.0011 (17.10) 0.0291 (0.90) 0.0264 (2.33)
Partition full sample on firm size (full sample breakpoints):
Small (33.3%) 0.0007 (2.49) 0.0011 (12.25) 0.0606 (1.29) 0.0184 (1.17)
Medium (33.3%) 0.0008 (2.77) 0.0009 (10.19) -0.0121 (-0.28) 0.0292 (1.90)
Large (33.3%) 0.0006 (1.95) 0.0012 (12.77) 0.0427 (1.26) 0.0267 (1.79)
Partition full sample on firm size (NYSE breakpoints):
Small (71.0%) 0.0007 (2.56) 0.0010 (12.29) 0.0336 (0.77) 0.0225 (1.58)
Medium (18.9%) 0.0009 (2.90) 0.0010 (11.73) 0.0458 (1.22) 0.0289 (1.97)
Large (10.1%) 0.0005 (1.46) 0.0012 (9.91) 0.0190 (0.44) 0.0219 (1.35)
Partition full sample on analyst following:
Low (36.6%) 0.0005 (1.85) 0.0010 (10.60) 0.0845 (1.82) 0.0394 (2.69)
Medium (31.5%) 0.0007 (1.98) 0.0011 (11.47) -0.0315 (-0.81) 0.0177 (1.00)
High (32.9%) 0.0006 (2.12) 0.0010 (12.77) 0.0043 (0.11) 0.0160 (1.31)
Partition full sample on domesticity:
Domestic (69%) 0.0008 (3.58) 0.0010 (13.44) 0.0289 (0.86) 0.0250 (2.04)
Non-domestic (31%) 0.0009 (2.18) 0.0011 (9.92) 0.0759 (1.46) 0.0382 (2.52)
Partition non-domestic sample on complexity:
Simple (19.5%) 0.0007 (1.86) 0.0010 (9.08) 0.1224 (2.49) 0.0483 (3.13)
Complex (10.5%) 0.0012 (1.99) 0.0011 (6.25) -0.0283 (-0.33) 0.0177 (0.67)
o Partition non-domestic complicated sample on firm size:
o Small (5.3%) 0.0009 (1.19) 0.0008 (3.55) -0.0711 (-0.70) -0.0128 (-0.39)
o Large (5.3%) 0.0015 (1.91) 0.0014 (5.61) 0.0642 (0.68) 0.0378 (1.10)
o Partition non-domestic complicated sample on firm domicile:
o US (3.6%) -0.0019 (-0.73) 0.0002 (0.23) -0.7452 (-2.19) -0.1534 (-1.32)
o Non-US (6.9%) 0.0012 (1.73) 0.0011 (7.02) 0.0675 (0.75) 0.0426 (1.72)
64
Panel B: Macroeconomic information and one month ahead future stock returns
etuvwuj etuvweu
j
Raw DGTW x 7-factor Raw x DGTW x 7-factor
x
Full Sample: 0.0105 0.0096 0.0117 0.0125 0.0087 0.0113
(2.40) (2.57) (2.57) (3.36) (2.68) (3.04)
Partition full sample on firm size (full sample breakpoints):
Small (33.3%) 0.0107 0.0132 0.0135 0.0212 0.0216 0.0212
(1.74) (2.40) (2.12) (3.48) (3.76) (3.28)
Medium (33.3%) 0.0159 0.0163 0.0196 0.0184 0.0174 0.0180
(2.68) (3.23) (3.28) (4.30) (4.45) (4.06)
Large (33.3%) 0.0098 0.0072 0.0101 0.0097 0.0064 0.0093
(2.04) (1.74) (2.05) (3.07) (2.25) (2.91)
Partition full sample on firm size (NYSE breakpoints):
Small (70.3%) 0.0135 0.0145 0.0178 0.0186 0.0179 0.0181
(2.24) (2.91) (2.86) (4.26) (4.47) (3.97)
Medium (20.0%) 0.0106 0.0095 0.0132 0.0088 0.0069 0.0106
(1.82) (1.87) (2.22) (2.36) (1.95) (2.73)
Large (10.7%) 0.0087 0.0069 0.0100 0.0088 0.0060 0.0077
(2.38) (2.18) (2.68) (2.73) (2.10) (2.37)
Partition full sample on analyst following:
Low (25.0%) 0.0142 0.0134 0.0161 0.0180 0.0152 0.0206
(2.17) (2.33) (2.46) (3.47) (3.16) (3.92)
Medium (16.1%) 0.0103 0.0089 0.0145 0.0167 0.0161 0.0206
(1.53) (1.46) (2.23) (2.70) (2.62) (3.27)
High (18.7%) 0.0063 0.0051 0.0077 0.0048 0.0035 0.0046
(1.74) (1.58) (2.15) (1.65) (1.32) (1.56)
Partition full sample on MSCI index inclusion:
Constituents (7.7%) 0.0083 0.0063 0.0093 0.0040 0.0031 0.0049
(1.74) (1.90) (2.43) (1.20) (1.03) (1.45)
Non-Constituents (92.3%) 0.0114 0.0108 0.0130 0.0168 0.0126 0.0155
(2.18) (2.41) (2.35) (3.91) (3.18) (3.59)
65
Partition domestic firms on MSCI index inclusion:
Constituents (3.4%) 0.0095 0.0083 0.0096 0.0107 0.0087 0.0099
(1.69) (1.64) (1.76) (2.15) (1.87) (1.97)
Non-Constituents (67.4%) 0.0159 0.0153 0.0180 0.0215 0.0167 0.0193
(2.49) (2.76) (2.68) (4.45) (3.80) (3.98)
Partitions: The fraction of the full sample included in each cross sectional partition is shown in parentheses next to the partition label. We
partition each cross-section on the basis of firm size twice. The first approach groups firms into three equal sized groups using breakpoints for
the entire sample from the previous month. The second approach uses NYSE breakpoints from the previous month. We group firms below
(above) the NYSE size 20th (60th) percentile into the small (large) group. The remaining firms are placed into the middle group. Complex
(simple) firms are those firms with more (less) then 3 reported geographic segments. This partition is only conducted for the non-domestic
66
sample. We partition the sub-sample of firms with analyst following (62.8 percent of the full sample) into three groups based on analyst
following ‘tercile’ breakpoints from the previous month.
Panel A: We estimate regression equation (3) for the full sample and for each partition described above. For the sake of brevity we report only
the regression coefficient (with Fama-Macbeth test statistic in parentheses adjacent to coefficient) for .
Panel B: In each month all securities are sorted into five equal sized groups based on . The portfolio returns are value weighted (the
weights are market capitalization, in USD). The Z
/[
return is the difference in total returns between the average portfolio returns across
extreme quintiles for the following month. The DGTW Z
/[
return is the difference in characteristic adjusted returns between the average
portfolio returns across extreme quintiles for the following month. To compute characteristic adjusted returns we follow the procedure in Daniel,
Grinblatt, Titman and Wermers (1997). Specifically we sort all securities each month into 125 portfolios based on a conditional sorting on the
following characteristics: (i) size, (ii) B/P, and (iii) momentum. Value weighted returns are computer for each cell each month. The
characteristic adjusted return is then the difference between the security return and the average return for the characteristic portfolio that security
belongs to. The 7-factor alpha is based on regression equation (6). We report the time series average of the hedge return with the associated test
statistic below in parentheses.
67