Industry Effects on Firm and Segment Profitability Forecasting:
Do Aggregation and Diversity Matter?†
David Schröder, Birkbeck, University of London
Andrew Yim, Cass Business School, City University London
29 August 2014
Abstract. This paper analyzes the incremental advantage of industry-specific models
of profitability forecasting, over economy-wide models, for business segments and firms.
Prior research suggests that industry-specific analysis has no advantage over economywide analysis in forecasting firm profitability. This seems puzzling because many studies
in economics and strategic management have documented the importance of industry
effects in explaining firm profitability. We reconcile this apparent inconsistency by
showing that industry effects on profitability forecasting can exist at the more refined
business segment level and can yet be obscured by aggregated reporting at the firm level.
Industry-specific forecasts can be significantly more accurate in predicting profitability
for segments and undiversified firms. (JEL L25, G17, M21, M41)
Key words: Segment profitability, Earnings predictability, Earnings persistence,
Aggregation, Diversity, Industry membership
D. Schröder. Postal: Department of Economics, Mathematics and Statistics, Birkbeck,
University of London, Malet Street, London WC1E 7HX, UK. Phone: +44 20 7631-6408.
E-mail: d.schroeder@bbk.ac.uk .
A. Yim. Postal: Faculty of Finance, Cass Business School, City University London,
106 Bunhill Row, London EC1Y 8TZ, UK. Phone: +44 20 7040-0933. Fax: +44 20 70408881. E-mail: a.yim@city.ac.uk / andrew.yim@aya.yale.edu .
†
First version: June 2012. We thank Eli Amir, Art Kraft, Peter Pope, and participants of the 50th
Annual Conference of the British Accounting and Finance Association (BAFA) for their useful
feedback. All remaining errors are ours.
Industry Effects on Firm and Segment Profitability Forecasting:
Do Aggregation and Diversity Matter?
1. Introduction
The return on equity (ROE) as a measure of firm performance, or its unlevered
counterpart, the return on assets (ROA), plays an important role in firm valuation
models.1 More accurate forecasts of profitability, or of the underlying earnings, are useful
to market participants in valuing firms, contributing to the price formation in the stock
market. Prior research shows that the predictability of earnings is due to the mean
reversion of profitability (Fama and French 2000). Therefore, market participants, such as
financial analysts, can more accurately forecast earnings by incorporating the mean
reversion in profitability.
Yet, it remains an open question how to optimally exploit the mean reversion
property. In particular, if the mean reversion differs significantly across industries, an
industry-specific forecasting model should outperform an economy-wide model. A large
body of economic and strategic management research suggests that inter-industry
differences can affect firm performance systematically (e.g., Schmalensee 1985,
McGahan and Porter 1997, and Bou and Satorra 2007). Motivated by the literature,
Fairfield, Ramnath, and Yohn (2009) examine the relative forecast accuracy of mean
reverting models at the industry and economy-wide levels. Surprisingly, they find no
significant improvement in forecasting profitability with an industry-specific model,
compared to its economy-wide counterpart.
This paper proposes an intuitive reconciliation of the seemingly inconsistent findings
1
See, e.g., Frankel and Lee (1998), Gebhardt, Lee, and Swaminathan (2001), and Penman (2003).
1
of the economics and strategic management literatures and the Fairfield et al. (2009)
study. Many firms are conglomerates operating in different industries. They have various
lines of business organized into units reported as business segments. When the segments
of such diversified firms are associated with different industries, no single industry
accurately represents the whole firm. A firm-level industry-specific forecasting model
therefore may not be appropriate for forecasting the profitability of diversified firms. In
contrast, individual business segments of a firm are by definition more homogenous than
the firm itself, and often operate mainly in one industry. Hence, for a firm’s business
segments, industry-specific forecasting models should generate more accurate
profitability forecasts. In sum, we conjecture that industry effects on profitability
forecasting exist and are clearer at the segment level, but can be obscured when data are
aggregated to the firm level.
Using both firm and segment data up to 2011, this study finds solid empirical support
for the conjecture. Following the literature, we employ various versions of the persistence
model to predict future profitability. We show that compared to economy-wide
predictions, industry-specific forecasts can be significantly more accurate in predicting
profitability for individual business segments and for undiversified firms (with only one
business segment). With regard to economic significance, the forecast improvements (in
terms of absolute forecast error) for large industry sectors can be as high as around 10%
of the actual profitability to be forecast.
Yet, segment reporting standards have changed considerably in 1998, with the new
Statement of Financial Accounting Standards No. 131 (SFAS 131) superseding SFAS
2
14.2 While SFAS 131 has brought various advantages to investors and financial analysts,
this study shows that the new standard has had a mixed effect on the quality of industryspecific profitability forecasts. For undiversified firms, there is moderate evidence
suggesting that the advantage of industry-specific profitability forecasts has improved
following the introduction of SFAS 131. However, the new standard has adversely
affected the advantage of industry-specific forecasts of segment profitability. We are able
to document a significant advantage of industry-specific forecasts at the segment level
only for the pre-SFAS 131 period.
In addition to the ROE and ROA, we consider two other measures, the return on net
operating assets (RNOA) and the return on sales (ROS). The ROA, RNOA, and ROS are
common alternatives to the ROE as profitability measures. Most important, they are
among the main drivers of the ROE. According to the Du Pont analysis, what drive the
ROE are the asset-to-equity ratio (a leverage measure) and the ROA, which can be
further broken down into the asset turnover ratio and the ROS. An alternative way to
decompose the ROE traces to the RNOA and two leverage-related measures as the main
drivers, namely the financial leverage ratio and the operating spread (see Penman 2003).
Analyzing the predictability of these alternative profitability measures provides an
additional route to understanding the predictability of the ROE.
This additional route is particularly important and necessary when one uses a bottomup approach to understand the performance of a firm through the performance of its
individual segments –– the typical approach used by financial analysts. The importance
2
Under SFAS 14, firms were asked to disclose segment information according to the industry
classification of the segments. Most important, reported segment profits must conform to the US
generally accepted accounting principles (GAAP). This guarantees certain level of comparability
across firms. With the implementation of SFAS 131, firms are only required to align the segment
reporting with the internal structure and accounting. Hence, segment profit data are not as comparable
across firms as before. Due to concerns like this, some studies focus the analysis on the pre-SFAS 131
period (e.g., Hund, Monk, and Tice 2010).
3
of segment disclosures to analysts is evident in the change of the segment reporting
standard from SFAS 14 to SFAS 131. The change was partly a response to analysts’
complaints about the flexibility of the old standard that was exploited by some firms to
avoid providing segment disclosures (Botosan and Stanford 2005).
At the segment level, one cannot use the ROE and RNOA measures to assess
performance. They are either undefined or the required data for constructing the measures
are not disclosed to the public. Hence, the ROA and ROS measures become the
remaining alternatives for measuring profitability at the segment level (e.g., Berger and
Ofek 1995).
Our analysis reveals that unlike the other profitability measures, the ROS can be more
accurately forecast with industry-specific models even at the firm level, without focusing
on undiversified firms. Consistent with our conjecture, the ROS forecast improvement is
even larger when considering individual business segments and undiversified firms.
Taken together, our results show that aggregation and diversity matter in revealing the
industry effects on profitability forecasting.
Besides reconciling the seemingly inconsistent findings of prior research, this study is
also related to the growing literature that examines the usefulness of providing less
aggregated segment-level data to the public. Several studies find that segment data allows
stakeholders such as analysts, investors, and researchers to anticipate future earnings
more accurately (e.g., Ettredge et al. 2005, Berger and Hann 2003, Baldwin 1984, and
Collins 1976).
The rest of the paper is organized as follows. In the next section, we explain the
hypotheses developed from our conjecture and the research design. Section 3 describes
4
the data and the sample construction procedure. Section 4 discusses the evidence from the
firm- and the segment-level analysis, considering the effect of the SFAS 131. Concluding
remarks are given in section 5.
2. Hypotheses and Research Design
2.1 Hypotheses
Inspired by studies in the diversification and segment reporting literatures (e.g.,
Berger and Ofek 1995, Campa and Kedia 2002, Berger and Hann 2007, and Hund, Monk,
and Tice 2010), we conjecture that industry effects on profitability forecasting exist and
are clearer at the segment level but can be obscured when segment-level data capturing
the effects are aggregated to the firm level. To verify the conjecture, we examine several
implications of the conjecture.
The first hypothesis is about single- and multiple-segment firms. Multiple-segment
firms are firms that have more than one segment; single-segment firms are those have
only one segment. To the extent that single-segment firms on average are less diversified
(more homogenous) than multiple-segment firms, we can find industry effects at the firm
level for single-segment firms but not for multiple-segment firms.
H1: Industry-specific profitability forecasts are more accurate than economy-wide
forecasts for single-segment firms, but not for multiple segment firms.
The introduction of SFAS 131 in 1998 arguably has given firms less discretion in
segment aggregation. In line with this, some firms have changed the number of reported
segments from one in 1997 to more than one by 1999, suggesting that they might not be
genuinely single-segment firms prior to 1997. If many of such “change firms” indeed are
multiple-segment firms, they should look more like multiple-segment firms than single5
segment firms. In particular, “change firms” and multiple-segment firms should be quite
similar in terms of the advantage of the industry-specific profitability forecasts over their
economy-wide counterparts. Therefore, we are interested in testing the following
hypothesis.
H1a: Between change firms and multiple-segment firms, there is no difference in the
advantage of industry-specific profitability forecasts over economy-wide forecasts.
Another way to examine the effect of SFAS 131 is to compare the advantage of
industry-specific profitability forecasts over economy-wide forecasts before and after the
shift to the new standard. Before the new standard, “change firms” are part of the asreported single-segment firms. They are no longer part of that group after the introduction
of the new standard. Given that “change firms” might not be genuinely single-segment
firms, we expect the following hypothesis to hold.
H1b: For as-reported single-segment firms, the advantage of industry-specific
profitability forecasts over economy-wide forecasts is stronger after the introduction of
SFAS 131.
We also examine the advantage of industry-specific profitability forecasts at the
segment level. By definition a segment of a firm is more homogeneous in activities than
the firm itself. If as conjectured it is mainly because of aggregated reporting at the firm
level that obscures the industry effects on profitability forecasting, then we should see the
effects re-appearing at the segment level. This gives the first of our second set of
hypotheses:
6
H2: At the segment level, industry-specific profitability forecasts are more accurate
than economy-wide forecasts.
Although SFAS 131 arguably has reduced the discretion in segment aggregation,
there is less consistency in defining segment profits under the new standard (Berger and
Hann 2007). Less uniform segment profits reduce the data quality of the samples
composed of segments from different firms, increasing the difficulty in forecasting
segment profitability. Note that the within-industry samples used for estimation under the
IS approach are much smaller than the all-industry sample used under the EW approach.
Hence, the data quality problem is likely to have a greater impact on the accuracy of the
IS approach than the EW approach. We therefore expect the following hypothesis to
hold:
H2a: At the segment level, the advantage of industry-specific profitability forecasts
over economy-wide forecasts is weaker after the introduction of SFAS 131.
2.2 Research Design
We measure the advantage of industry-specific (IS) profitability forecasts over
economy-wide (EW) forecasts in terms of the absolute forecast error from the EW
analysis minus its IS counterpart. For ease of reference, we call this the forecast
improvement (of IS over EW analysis).
Among the many models that may be used to forecast profitability, the persistence
model (i.e., first-order autoregressive) is a parsimonious choice. Unlike higher-order
autoregressive models, the persistence model does not require long earnings histories and
therefore minimizes the survivor bias. The limited availability of segment-level data also
prevents us from using sophisticated models to forecast profitability at the segment level.
7
We use the standard first-order autoregressive model (the baseline specification), as
well as two augmented specifications, to forecast profitability. The IS and EW models for
the baseline specification are:
IS model: xi,t = αj,t + βj,t xi,t–1 + εi,t,
EW model: xi,t = αt + βt xi,t–1 + εi,t,
where xi,t is the profitability of firm/segment i in year t, j is the industry of the
firm/segment, and εi,t is the error term. The IS model estimates a regression for each
industry j separately, whereas the EW model pools all observations into one regression.
The model coefficients are indexed by a year subscript t because they are re-estimated
each year based on the most recent 10 years of data. The estimated coefficients from
these in-sample regressions (Step 1) are used to compute the profitability forecasts and
the forecast errors used for out-of-sample tests (Step 2). Further details of this two-step
procedure will be provided shortly.
The first augmented specification is obtained by adding to the baseline specification a
dummy variable capturing nonlinear mean reversion of profitability. Specifically, the
dummy variable Dj,i,t in the IS model for the first augmented specification is equal to 1 if
in year t – 1 the profitability of firm/segment i is below the mean profitability of all
firms/segments in the same industry j, and 0 otherwise. In the EW model in this
specification, the dummy variable Di,t is the counterpart of Dj,i,t for all firms/segments in
any industry. These dummy variables allow the mean reversion of profitability to differ
for firms/segments with above- and below-average profitability. Fama and French (2000),
among others, have documented a non-linear pattern of the mean reversion of
profitability.
8
Penman (2003) and Lundholm and Sloan (2007) suggest that predicted sales growth
is important to profitability forecasting. Therefore, the second augmented specification is
obtained by further adding the predicted sales growth of a firm/segment, PREDGSLi,t, as
follows:
IS model: xi,t = αj,t + βj,t xi,t–1 + γ j,t Dj,i,t xi,t–1 + λ j,t PREDGSLi,t + εi,t,
EW model: xi,t = αt + βt xi,t–1 + γt Di,t xi,t–1 + λt PREDGSLi,t + εi,t.
Similar to the profitability forecasts from the baseline specification, PREDGSLi,t is the
predicted value of a first-order autoregressive model of sales growth, defined as the
percentage change in sales of firm/segment i from year t – 1 to year t. The estimation of
this model is carried out for each industry j separately because Fairfield, Ramnath, and
Yohn (2009) have documented a significant industry effect on sales growth forecasting.
In the in-sample regressions, we estimate the year-indexed coefficients of the three
specifications on a rolling basis. The baseline and the first augmented specifications
require the most recent 10 years of data. For example, to estimate the coefficients for year
t like αt and βt, we use profitability data of all firms/segments from year t back to year t –
9 and their lagged values from year t – 1 back to year t – 10. The second augmented
specification requires the most recent 20 years of data because it uses the predicted sales
growth as an explanatory variable (which itself requires 10 years of data).
To obtain reasonably reliable estimates, we require a minimum of 100 observations
for each rolling regression. Some industries are excluded from the analysis owing to too
few observations. For equal-footing comparisons, we estimate the EW model using only
observations that are included to estimate the IS model.
In the out-of-sample tests, we use the estimated coefficients of the in-sample
9
regressions and the observed profitability of last year to forecast the firm/segment
profitability of the current year. For the baseline specification, this means calculating the
forecasts as follows:
IS model: EIS[xi,t]= aj,t + bj,t xi,t–1,
EW model: EEW[xi,t]= at + bt xi,t–1,
where a and b are the estimates of the model coefficients α and β. The forecasts from the
two augmented specifications are calculated analogously.
To perform an out-of-sample test on the relative accuracy of the IS and EW models,
we first calculate for each observation the absolute forecast error defined as the absolute
difference between the profitability actually observed and the profitability forecast:
AFEIS = | xi,t – EIS[xi,t] |,
AFEEW = | xi,t – EEW[xi,t] |,
where AFEIS and AFEEW are the absolute forecast errors for a firm/segment of a year
based on the IS and EW models, respectively. Next, we calculate the forecast
improvement (FI) of the IS over EW model by deducting AFEIS from AFEEW:
FI= AFEEW – AFEIS.
If IS analysis can improve the accuracy of profitability forecasting compared to EW
analysis, the FI measure should be positive on average.
3. Data and Sample Selection
In this section, we give an overview of the data used and the sample constructed,
followed by a discussion of the summary statistics.
The firm and business segment data come from the Compustat annual fundamentals
and Compustat segments databases of the Wharton Research Data Services (WRDS).
10
Business segment data are available from as early as 1976. Because the data coverage in
the initial year is not good, we use the segment data from 1977 onward.
Recall that the second augmented specification requires 20 years of data to estimate
the model coefficients. This prevents us from using the specification in the segment-level
analysis for the effect of SFAS 131 and leaves only a relatively short period for doing
out-of-sample tests for other segment-level analyses. For simplicity, we do not consider
this specification in all the segment-level analysis.
For the baseline and the first augmented specifications, only 10 years of data are
needed to estimate the model coefficients. Hence, the forecasts for the out-of-sample tests
in the segment-level analysis are available from 1987 onward. To ease the comparison
with the segment-level analysis, we also construct the forecasts in the firm-level analysis
from this year onward. Given the longer time series of firm level data, we are able to use
the second augmented specification in all the firm-level analysis. To construct the
forecasts in the firm-level analysis, the second augmented specification uses firm data
from 1967 onward, whereas the other two specifications use firm data from 1977 onward.
Panel A of table 1 gives an overview on the data included in this study.
We use the two-digit primary Standard Industry Classification (SIC) code to define
the industry to which a firm or business segment belongs.3 Observations with missing
SIC codes are excluded from the sample. To avoid distortions caused by regulated
industries, we exclude all firms and segments in the financial service and utilities sectors
(i.e., with SIC between 6000 and 7000, or between 4900 and 4950). In addition, the U.S.
3
For robustness checking, we have conducted the firm-level analysis using three other industry
classification systems: Fama-French Classification System (FF), North American Industry
Classification System (NAICS), and Global Industry Classification Standard (GICS). The results
based on the FF are very similar to those based on the SIC, sometimes even stronger. Though
qualitatively similar, the results based on the NAICS and GICS are generally weaker.
Some studies (e.g., Fairfield, Ramnath, and Yohn 2009) classify industries using the GICS codes,
which are often unavailable for segment-level data.
11
postal service (SIC 4311) and non-classifiable establishments (SIC above 9900) are
excluded.
Because firm assets might not be fully allocated at the segment level, segment assets
do not always add up to the firm assets. Similar discrepancies, though not as severe, exist
between firm sales and aggregated segment sales. Such discrepancies can lead to
measurement errors in segment ROA and ROS. To mitigate the problem, we follow
Berger and Ofek (1995) and Berger and Hann (2007) to adjust the segment assets and
sales as follows: First, we exclude segment observations with the aggregated segment
assets deviating from the firm assets by more than 25%; similarly, we exclude those with
a deviation of 5% for segment sales. Second, if a deviation is within the limits above, we
allocate the deviation proportionally to each segment based on the segment assets to firm
assets ratio (its counterpart for sales).
To construct the time series of a segment, we rely on the segment ID (SID) provided
by Compustat. Firms sometimes change the internal structure, leading to changes in the
number of disclosed segments, and possibly their SIC codes. Such a restructuring
requires firms to restate previous segment information to make them comparable across
years. We utilize the restated information in the in-sample regressions. To prevent a lookahead bias, we do not utilize the information in the out-of-sample tests.
In the firm-level analysis, we distinguish between single- and multiple-segment firms.
To do so, we match the sample of firm profitability forecast improvements with the
business segment data. As in the segment-level analysis, we exclude observations with a
25% deviation between the firm assets and aggregated segment assets (or for sales, a 5%
deviation) to alleviate the data quality concern.
12
Occasionally, some firm/segment has two observations per calendar year. We drop
identical duplicate entries. If the data of duplicate observations are diverging, e.g., due to
reasons like shortened fiscal years, we exclude them from the sample.4 To mitigate the
impact of small denominators on firm profitability measures, we exclude firm
observations with total assets, net operating assets, and sales below USD 10mn and book
value of equity below USD 1mn. For segment data, we exclude observations with total
identifiable assets and sales below USD 1mn.
To avoid the influence by outliers, observations with the absolute value of
firm/segment profitability exceeding one are excluded. To reduce the influence by
mergers and acquisitions, we remove observations with growth in operating assets, net
operating assets, book value of equity, and sales above 100%. Recall that our analysis has
an in-sample regression step and an out-of-sample test step. Before the in-sample
regressions, we further exclude observations with the profitability measure in concern
falling in the top or bottom one percentile. However, we do not apply such an extremevalue exclusion criterion before the out-of-sample tests to avoid any look-ahead bias in
the analysis.5
Panel B of table 1 summarizes the definitions of the four profitability measures
considered and the variables used to compute the measures. Panel A of table 2
summarizes the number of observations after applying each exclusion criterion described
above. For consistency, only observations with all four profitability measures available
are used in the firm-level analysis and only those with both ROA and ROS measures
4
The deletion of double observations per calendar year reduces the sample size by 4 observations
in the firm-level analysis and by 2,114 observations in the segment-level analysis.
5
All exclusion criteria are similar to those in Fairfield, Ramnath, and Yohn (2009).
13
available are used in the segment-level analysis.
Panels B and C of table 2 give an overview of the firm and segment data used to
compute the average forecast improvements reported in section 4. The firm-level analysis
uses 40,010 firm-year observations of 5,977 unique firms; the segment-level analysis is
based on 77,443 segment-year observations of 14,979 unique segments.
For firms, the ROE on average is 7.5%, while the mean ROS and ROA are slightly
higher at around 8.5%. With 14.3%, the mean RNOA is considerably higher. These
statistics are similar to those in prior studies, such as Fama and French (2000) and
Fairfield, Ramnath, and Yohn (2009). The average levels of segment profitability are
somewhat lower than their firm profitability counterparts. The mean segment ROA and
ROS are 7.95% and 6.93%, respectively.
Panel C reports for each industry the number of observations, as well as average
profitability. With 3,751 firm-year and 6,891 segment-year observations, electronic &
other electric equipment (SIC 36) constitutes the largest industry in the sample. Other
important industries are chemicals & allied products (SIC 28), industrial machinery &
equipment (SIC 35), instruments & related products (SIC 38), and business services (SIC
73).
There is substantial variation in average profitability across industries, ranging from 2.1% to +22.8%. For firms, transportation services (SIC 47) and communications (SIC
48) are the industries with the highest levels of profitability. The lowest levels come from
motion pictures (SIC 78), hotels & other lodging places (SIC 70), auto repair, services &
parking (SIC 75), and food stores (SIC 54). The highest levels of segment profitability
are from pipelines, except natural gas (SIC 46) and communications (SIC 48). Similar to
14
firm profitability, motion pictures (SIC 78) and food stores (SIC 54) exhibit the lowest
levels of segment profitability.
4. Results on Forecast Improvements of IS over EW Models
This section discusses the results of the firm- and segment-level analyses,6 followed
by an extension of the analyses to sales growth forecast improvements.
We assess the magnitude of the firm/segment profitability forecast improvement
using two notions of the mean and the median. The pooled mean and pooled median are
respectively the mean and median forecast improvements computed from all
firm/segment observations pooled together. In addition, we report the grand mean, which
is the mean of the yearly mean forecast improvements, and the grand median, which is
the median of the yearly median forecast improvements.
Along with the pooled and grand means and medians, we report the p-values for the
significance levels of the tests of these statistics. Tests of the pooled and grand means are
based on a t-test. Tests of the pooled and grand medians are based on a Wilcoxon (1945)
signed-rank test. For the pooled mean, the t-test is on the estimated coefficient of a
regression on only the intercept term, using the sample of all the forecast improvements
and with the standard errors corrected for the clustering by firm and year following
Rogers (1993). For the grand mean, the t-test is based on a similar regression using the
sample of the yearly mean forecast improvements, with the standard error adjusted
following Newey and West (1987). The Wilcoxon signed-rank tests for the pooled and
grand medians are on the median of all the forecast improvements and the median of the
yearly median improvements, respectively.
Similar to Fairfield, Ramnath, and Yohn (2009), we also report the number of
6
For brevity, we report only the results for the two augmented specifications. The results for the
baseline specification (available upon request) are similar.
15
industries (or years) in which the industry (or yearly) mean forecast improvements is
significantly positive / negative at the 10% level.
4.1 Firm-level Analysis
Table 3 shows that Fairfield, Ramnath, and Yohn’s (2009) no industry effect result for
profitability forecasting by and large holds for our out-of-sample test period (1987-2011).
Two profitability measures, ROE and RNOA, are analyzed in their study. For these two
measures, nearly all the firm profitability forecast improvements (of IS over EW
analysis) are not significantly different from zero. The only exception is the pooled
median under the second augmented specification, which is a statistic not considered in
their study. Aside from this, the no industry effect result holds regardless of the
specifications or the way the forecast improvements are computed.
Further evidence of the no industry effect result is obtained when using ROA as the
profitability measure. Including ROA in the firm-level analysis facilitates comparison
with the results from the segment-level analysis (where ROE and RNOA cannot be
computed owing to data limitations). The evidence based on ROA is similar to those for
ROE and RNOA. The only difference is the marginally significant grand median under
the second augmented specification.7
Table 3 also shows an interesting new finding: In terms of ROS, the firm profitability
forecast improvement is highly significantly positive, regardless of the specifications or
forecast improvement measures. This suggests that Fairfield, Ramnath, and Yohn’s
(2009) no industry effect result for profitability forecasting may be sensitive to the
profitability measure used. Notwithstanding this, the new finding can be completely
7
Unlike Fairfield, Ramnath, and Yohn’s (2009) original analysis, table 3 includes only firms for
which there is segment data available. In an untabulated analysis including all firms in our test period,
we are able to replicate the no industry effect result for ROE, RNOA, and ROA.
16
consistent with our conjecture that industry effects on profitability forecasting are
stronger at the segment level than at the firm level. The results of our tests to be presented
below provide evidence supporting both the conjecture and the consistency with the new
finding based on ROS.
We argue that the lack of (or weaker) industry effects at the firm level is due to
aggregated reporting that obscures the relation between profitability and industry-specific
characteristics. When the segments of a multiple-segment firm are associated with
different industries, no single industry can closely represent the whole firm. Describing a
multiple-segment firm with a primary industry ignores the relation between its
profitability and the other industries to which its segments belong. In contrast, for firms
with a single business segment, the firm-level reporting does not distort the truth – the
only segment of a single-segment firm is effectively identical to the whole firm. If the
advantage of IS over EW analysis exists at the segment level, it should also be observed
at the firm level when confining to single-segment firms. However, for multiple-segment
firms, the advantage may remain indistinguishable from zero.
To test this hypothesis (H1), we partition the forecast improvements into subsamples
for single- and multiple-segment firms. Owing to the doubt in correctly classifying firms
that changed the number of reported segments from one in 1997 to more than one by
1999, the forecast improvements of these firms are separated into a third group called
“change firms”.
The results presented in table 4 generally support the hypothesis. Under the first
augmented specification in panel A, all the forecast improvements for multiple-segment
firms are statistically indistinguishable from zero, or non-positive. As predicted, the IS
17
analysis still has no advantage over the EW analysis in forecasting firm profitability for
these firms. The results for multiple-segment firms under the second augmented
specification (panel B) are similar, though with a few exceptions (two occasions for the
pooled median and one for the grand median).
The results for single-segment firms under the two augmented specifications strongly
support the hypothesis for three of the four measures, except the ROE. Nearly all the
forecast improvements measured in terms of RNOA, ROA, and ROS are positive at the
5%, or even 1%, significance level. In terms of ROE, the forecast improvements for
single-segment firms are insignificant under the first augmented specification. However,
two of the statistics (the pooled mean and the pooled median) become significantly
positive at the 10%, or even 1%, level under the second augmented specification. Overall,
apart from the limited support from the ROE, the evidence from the other three measures
firmly support the prediction that IS analysis is useful for profitability forecasting even at
the firm level when confining to single-segment firms.
Although the forecast improvement for single-segment firms on average is not large,
the economic significance of the improvement is material for certain industries, even in
terms of the ROE. For example, one of the largest industry sectors is communications
(SIC 48). This industry sector has an ROE forecast improvement of 1.3%. Given that the
single-segment firms in this sector on average have an ROE of about 13.8%, the forecast
improvement of 1.3% means IS forecasts on average being 9.4% (= 1.3% / 13.8%) closer
to the actual ROE than EW forecasts. Other industries with material forecast
improvements for single-segment firms include chemicals & allied products (SIC 28),
railroad transportation (SIC40), and transportation services (SIC 47).
18
We also test whether the difference in the forecast improvements between single- and
multiple-segment firms is statistically significant. The results are shown in the
“Difference SS-MS” column of table 4.8 The results for the ROE continue to be an
exception. For the other three measures and for both augmented specifications, nearly all
the differences in the forecast improvements between single- and multiple-segment firms
are significantly positive at the 5%, or even 1%, level.
Following the introduction of SFAS 131, many firms have changed the number of
reported segments from one in 1997 to more than one by 1999. These “change firms”
might actually be disguised multiple-segment firms in prior to 1999. If this is true, they
should be similar to multiple-segment firms in terms of the firm profitability forecast
improvements. The “Difference MS-Change” column of the table 4 shows the results.
Consistent with our prediction H1a, almost all the differences in the forecast
improvements
between
multiple-segment
and
change
firms
are
statistically
indistinguishable from zero. The few exceptions are the grand mean and pooled median
for the ROA.
Finally, hypothesis H1b predicts that the firm profitability forecast improvement for
as-reported single-segment firms is stronger after the introduction of SFAS 131 than
before. Table 5 presents the results for this hypothesis. Given the complexity of the
hypothesis, we report only the results for the pooled and grand means. The single- and
multiple-segment firm type contained in the table are as reported, i.e., simply based on
the reported number of segments. So the “change firms” are part of the as-reported
single-segment firms before 1998 but are no longer part of them after 1998. To mitigate
8
Tests of the difference in the pooled (or grand) means are based on a t-test on the estimated
slope coefficient of a regression on the dummy for multiple-segment firms, with the standard error
corrected for the clustering by firm and year following Rogers (1993). Tests of the difference in the
pooled medians are based on an unpaired-sample Mann and Whitney (1947) U test (also called a
Wilcoxon rank-sum test). Tests of the difference in the grand medians are based on a paired-sample
Wilcoxon (1945) signed-rank test.
19
the contamination by major macroeconomic events like the aftermath of the September
11th attack, we compare only the three years after and before the introduction of SFAS
131 in 1998.
The findings provide moderate support to the hypothesis. In terms of the ROA and
ROS, the firm profitability forecast improvements for the as-reported single-segment
firms are indeed significantly stronger in the three years after than the three years before
1998. This holds for the two profitability measures regardless of the specifications. For
the other two measures, namely ROE and RNOA, the results are statistically
insignificant. However, the point estimates are still in line with the hypothesis.
Consistent with the intuition behind the hypothesis, one would expect the difference
in the forecast improvements between as-reported single- and multiple-segment firms to
be more pronounced after the introduction of SFAS 131. We examine this implication in
the “Difference” column of table 5. Unfortunately, we are unable to obtain statistically
significant results to confirm the implication. Yet, it should be noted that the statistics
reported in the “Difference” column are not simply a difference-in-difference test. It
involves three differences because the forecast improvement itself is a difference in the
absolute forecast errors between the IS and EW analysis. Given the difficulty in
empirically measuring a triple difference, the insignificant results seem unsurprising.
4.2 Segment-level Analysis
If industry effects on profitability forecasting exist at the segment level, we should
also observe forecast improvements of IS over EW analysis for segment profitability.
Table 6 shows the results for testing this hypothesis (H2). As explained in section 3, we
do not consider the second augmented specification in the segment-level analysis because
of insufficient segment data.
20
In terms of ROS, the findings provide solid support for H2. The segment profitability
forecast improvements are significantly positive, regardless of the statistics used to
measure the forecast improvements. However, in terms of ROA, only the pooled median
of the forecast improvements is significantly positive. We conclude that overall the
evidence provides moderate support for H2.
The difference between the results for ROS and ROA highlights an issue about
segment data in the post-SFAS 131 period recognized in the literature. Arguably, ROS is
more reliable than ROA as a profitability measure since sales, unlike assets, can be
correctly assigned to segments without ambiguity. The reliability of ROA decreases in
the SFAS 131 period because firms are only required to align the segment reporting with
the internal structure and accounting. Consequently, segment profits become less
comparable across firms owing to non-uniform definitions adopted by different entities
(Berger and Hann 2003 and Berger and Hann 2007). In contrast, SFAS 14 asked firms to
report segment information according to industry classification. Most important, the
segment profits reported must conform to the US generally accepted accounting
principles (GAAP), ensuring certain level of comparability across firms.
Although this problem also affects ROS, the impact is likely to be less severe because
the denominator, segment sales, can be more reliably measured than its counterpart in
ROA, the identifiable assets of a segment. Thus, the results in table 6 are consistent with
the well-recognized problem of segment profit data in the post-SFAS 131 period.
Table 7 provides a test of H2a, analyzing the difference between pre- and post-SFAS
131 forecast improvement. In line with the hypothesis, all the segment profitability
forecast improvements in the pre-SFAS 131 period are significantly positive, regardless
21
of the two profitability measures or the statistics used for measurement. For the ROA, the
significance levels of the forecast improvements vary from 10% to 1%. However, for the
ROS, which arguably is the more reliable profitability measure, the significance levels
are uniformly at 1%.
Strongly supporting H2a, the differences in the forecast improvements between the
pre- and the post-SFAS 131 period are significantly positive at the 1% level, with one
exception at the 5% level. The negative forecast improvements in the post-SFAS 131
period are consistent with the expectation that the data quality issue is likely to have a
greater impact on the accuracy of the IS approach than the EW approach.
As in the analysis for single-segment firms, communications (SIC 48) is the industry
sector with the largest forecast improvements. For instance, the segment ROS and ROA
forecast improvements are 1.0% and 0.5%, respectively. The segments in this sector on
average have an ROS of about 19% and an ROA of about 10%. Therefore, IS forecasts
on average are around 5% and 6% closer to the actual ROS and ROA, respectively, than
EW forecasts. Other industries with material forecast improvements are railroad
transportation (SIC40) and food stores (SIC54).
4.3 Sales Growth Forecasting
The second augmented specification considered in the firm-level analysis uses a
firm’s predicted sales growth (PREDGSL) as an explanatory variable. As explained in
section 2.2, this variable is itself constructed from an IS forecasting model because
Fairfield et al. (2009) has documented a significant industry effect on sales growth
forecasting. However, it is a priori unclear that such an effect also exists in our sample.
Besides, it is important to know whether our conjecture on profitability forecasting also
applies to sales growth forecasting, i.e., whether industry effects on sales growth
forecasting exist and are clearer at the segment level, but can be obscured when data are
22
aggregated to the firm level. To examine these questions, we repeat some of the firm- and
segment-level analysis on sales growth forecasting, using exactly the same empirical
approach and dataset as in the analysis on profitability forecasting.
The results presented in table 8 underline the importance of using IS models to
forecast firm and segment sales growth. Panel A of the table summarizes the results from
the firm-level analysis. Similar to Fairfield et al. (2009), we find a significant forecast
improvement of IS over EW model for the period from 1987 to 2011 for three of the four
statistics considered. This justifies the inclusion of PREDGSL as an additional
explanatory variable in forecasting firm profitability in section 4.1.
Next, we partition the sample of firms into single- and multiple-segment firms, as
well as change firms. Panel B of table 8 shows the forecast improvements for the three
subsamples. While for single-segment firms, the forecast improvement of IS over EW
model is highly significant at the 1% level, sales growth forecasts for multiple-segment
firms cannot be significantly improved using IS analysis. Furthermore, the differences in
the forecast improvements between the multiple-segment and change firms are
statistically indistinguishable from zero. The findings support the extension of H1 and
H1a to sales growth forecasting. Finally, panel C shows that at the segment level the sales
growth forecast improvement of IS over EW model is significantly positive for three of
the four statistics considered. This last result supports the extension of H2 to sales growth
forecasting.
5. Concluding Remarks
The future performance of a firm may depend on economy-wide or industry-specific
factors. Several studies in the economics and strategic management literatures have
documented the importance of industry effects in future changes in profitability and
earnings. Against the backdrop of these studies, recent research often favors industryspecific more than economy-wide forecasting models. For example, Gebhardt, Lee, and
23
Swaminathan (2001) propose a residual income model that explicitly incorporates
industry-specific paths of expected long-term profitability.9
Yet, the preference for industry-specific forecasting models does not go
unchallenged. Fairfield, Ramnath, and Yohn (2009) find that there is no significant
forecast improvement of industry-specific over economy-wide analysis in predicting firm
profitability. The finding questions the suitability of the valuation model by Gebhardt et
al. (2001), which has enjoyed great popularity in the finance literature.10
This paper proposes an intuitive reconciliation of the seemingly conflicting findings,
based on the observation that many firms have multiple business segments operating in
different industries. When segment-level data are aggregated to the firm level for external
reporting, industry effects on forecasting profitability are obscured at the firm level.
Our analysis shows that IS models are indeed significantly more accurate than EW
models in predicting segment profitability for the pre-SFAS 131 period and segment sales
growth for the whole sample period. For three of the four profitability measures
considered, we also find higher accuracy in predicting profitability at the firm level when
confining to single-segment firms, which operate in one industry only. Taken together,
the evidence strongly supports our conjecture that industry factors have an impact on
profitability and sales growth forecasting but the aggregated nature of firm-level data can
prevent the industry effects from standing out in the firm-level analysis.
Our findings on the impact of SFAS 131 on profitability forecasting highlight a
potential drawback of the new reporting standard. The new standard has brought various
9
Although firm-specific factors might be useful for forecasting, a firm-specific forecasting model
is undesirable because of the survivor bias resulting from the long-history data requirement (see Fama
and French 2000, p. 162, for more discussion). Indeed, Esplin (2012) shows that long-term growth
forecasts obtained from industry-specific prediction models are more accurate than firm-specific
forecasts.
10
The residual income model by Gebhardt et al. (2001) has been used extensively in the finance
literature to compute a firm’s implied cost of capital, see e.g., Pastor, Sinha, and Swaminathan (2008),
Chava and Purnanandam (2010), Lee, Ng, and Swaminathan (2009), and Chen, Da, and Zhao (2013).
24
benefits, such as an increase in the number of reported segments and a higher informative
value of the segment data. However, under SFAS 131 the segment data are less
comparable across firms. The increased noise in the segment data limits the usefulness of
the data in predicting profitability and hence reduces the accuracy of industry-specific
forecasting models at the segment level.
Besides the results above, we document that when profitability is measured in terms
of ROS, industry effects on profitability forecasting can be clearly seen at the firm level
even without focusing on single-segment firms. This interesting new finding strengthens
the conclusion that industry characteristics are indeed important to profitability
forecasting. The finding also serves as a support for the conventional wisdom that sales
convey valuable information about a firm’s future prospect.
The results of this study are also relevant to the accounting disclosure literature. Since
we find that segment-level data can improve a firm’s profitability forecasts, this can be
taken as evidence for the usefulness of less aggregated accounting disclosure. Yet, the
complications encountered at the segment level after the introduction of SFAS 131
highlights the importance of ensuring the comparability of the reported business segment
data.
25
References
Albrecht, W. Steve, Larry L. Lookabill, and James C. McKeown. 1977. “The Time-Series
Properties of Annual Earnings.” Journal of Accounting Research 15 (2): 226–44.
doi:10.2307/2490350.
Ali, Ashiq, April Klein, and James Rosenfeld. 1992. “Analysts’ Use of Information about
Permanent and Transitory Earnings Components in Forecasting Annual EPS.” The
Accounting Review 67 (1): 183–98.
Baginski, Stephen P., Kenneth S. Lorek, G. Lee Willinger, and Bruce C. Branson. 1999. “The
Relationship between Economic Characteristics and Alternative Annual Earnings
Persistence Measures.” The Accounting Review 74 (1): 105–20.
Baldwin, Bruce A. 1984. “Segment Earnings Disclosure and the Ability of Security Analysts to
Forecast Earnings Per Share.” The Accounting Review 59 (3): 376–89.
Bar-Yosef, Sasson, Jeffrey L. Callen, and Joshua Livnat. 1987. “Autoregressive Modeling of
Earnings-Investment Causality.” The Journal of Finance 42 (1): 11–28.
doi:10.2307/2328416.
Berger, Philip G., and Rebecca Hann. 2003. “The Impact of SFAS No. 131 on Information and
Monitoring.” Journal of Accounting Research 41 (May): 163–223. doi:10.1111/1475679X.00100.
Berger, Philip G., and Rebecca N. Hann. 2007. “Segment Profitability and the Proprietary and
Agency Costs of Disclosure.” The Accounting Review 82 (July): 869–906.
doi:10.2308/accr.2007.82.4.869.
Berger, Philip G., and Eli Ofek. 1995. “Diversification’s Effect on Firm Value.” Journal of
Financial Economics 37 (1): 39–65. doi:10.1016/0304-405X(94)00798-6.
Botosan, Christine A., and Mary Stanford. 2005. “Managers’ Motives to Withhold Segment
Disclosures and the Effect of SFAS No. 131 on Analysts’ Information Environment.” The
Accounting Review 80 (July): 751–72. doi:10.2308/accr.2005.80.3.751.
Bou, Juan Carlos, and Albert Satorra. 2007. “The Persistence of Abnormal Returns at Industry
and Firm Levels: Evidence from Spain.” Strategic Management Journal 28 (7): 707–22.
Brandon, Charles H., Jeffrey E. Jarrett, and Saleha B. Khumawala. 1983. “Revising Forecasts of
Accounting Earnings: A Comparison with the Box-Jenkins Method.” Management
Science 29 (2): 256–63.
Brown, Ld, Rl Hagerman, Pa Griffin, and Me Zmijewski. 1987. “An Evaluation of Alternative
Proxies for the Markets Assessment of Unexpected Earnings.” Journal of Accounting &
Economics 9 (2): 159–93. doi:10.1016/0165-4101(87)90004-8.
Campa, Jose Manuel, and Simi Kedia. 2002. “Explaining the Diversification Discount.” The
Journal of Finance 57 (August): 1731–62. doi:10.1111/1540-6261.00476.
Chant, Peter D. 1980. “On The Predictability of Corporate Earnings Per Share Behavior.” The
Journal of Finance 35 (1): 13–21. doi:10.2307/2327177.
Chava, Sudheer, and Amiyatosh Purnanandam. 2010. “Is Default Risk Negatively Related to
Stock Returns?” Review of Financial Studies 23 (6): 2523–59. doi:10.1093/rfs/hhp107.
Chen, Long, Zhi Da, and Xinlei Zhao. 2013. “What Drives Stock Price Movements?” Review of
Financial Studies 26 (4): 841–76. doi:10.1093/rfs/hht005.
Collins, Daniel W. 1976. “Predicting Earnings with Sub-Entity Data: Some Further Evidence.”
Journal of Accounting Research 14 (1): 163–77. doi:10.2307/2490463.
Conroy, Robert, and Robert Harris. 1987. “Consensus Forecasts of Corporate Earnings: Analysts’
Forecasts and Time Series Methods.” Management Science 33 (6): 725–38.
Cubbin, J., and P. Geroski. 1987. “The Convergence of Profits in the Long Run: Inter-Firm and
Inter-Industry Comparisons.” The Journal of Industrial Economics 35 (4): 427–42.
doi:10.2307/2098581.
26
Dechow, Patricia, Weili Ge, and Catherine Schrand. 2010. “Understanding Earnings Quality: A
Review of the Proxies, Their Determinants and Their Consequences.” Journal of
Accounting and Economics 50 (2-3): 344–401. doi:10.1016/j.jacceco.2010.09.001.
Dichev, Ilia D., and Vicki Wei Tang. 2009. “Earnings Volatility and Earnings Predictability.”
Journal
of
Accounting
and
Economics
47
(1-2):
160–81.
doi:10.1016/j.jacceco.2008.09.005.
Edwards, Edgar Owen, and Philip Wilkes Bell. 1961. The Theory and Measurement of Business
Income. Berkeley: University of California Press.
Esplin, Adam. 2012. “Industry Information and Forecasts of Long-Term Earnings Growth.”
http://business.illinois.edu/accountancy/events/forum/papers/11-12/esplin.pdf.
Ettredge, Michael L., Soo Young Kwon, David B. Smith, and Paul A. Zarowin. 2005. “The
Impact of SFAS No. 131 Business Segment Data on the Market’s Ability to Anticipate
Future
Earnings.”
The
Accounting
Review
80
(July):
773–804.
doi:10.2308/accr.2005.80.3.773.
Fairfield, Patricia M., Sundaresh Ramnath, and Teri Lombardi Yohn. 2009. “Do Industry-Level
Analyses Improve Forecasts of Financial Performance?” Journal of Accounting Research
47 (1): 147–78.
Fama, Eugene F, and Kenneth R French. 2000. “Forecasting Profitability and Earnings.” The
Journal of Business 73 (2): 161–75.
Frankel, Richard, and Charles M.C. Lee. 1998. “Accounting Valuation, Market Expectation, and
Cross-Sectional Stock Returns.” Journal of Accounting and Economics 25 (3): 283–319.
doi:10.1016/S0165-4101(98)00026-3.
Frankel, Richard, and Lubomir Litov. 2009. “Earnings Persistence.” Journal of Accounting and
Economics 47 (1-2): 182–90. doi:10.1016/j.jacceco.2008.11.008.
Gebhardt, W. R., C. M. C. Lee, and B. Swaminathan. 2001. “Toward an Implied Cost of Capital.”
Journal of Accounting Research 39 (1): 135–76. doi:10.1111/1475-679X.00007.
Glen, Jack, Kevin Lee, and Ajit Singh. 2003. “Corporate Profitability and the Dynamics of
Competition in Emerging Markets: A Time Series Analysis.” The Economic Journal 113
(491): F465–F484.
Goddard, John, Hong Liu, Philip Molyneux, and John O.S. Wilson. 2011. “The Persistence of
Bank Profit.” Journal of Banking & Finance 35 (11): 2881–90.
doi:10.1016/j.jbankfin.2011.03.015.
Hou, Kewei, Mathijs A. van Dijk, and Yinglei Zhang. 2012. “The Implied Cost of Capital: A
New Approach.” Journal of Accounting & Economics 53 (3): 504–26.
doi:10.1016/j.jacceco.2011.12.001.
Hund, John, Donald Monk, and Sheri Tice. 2010. “Uncertainty about Average Profitability and
the Diversification Discount.” Journal of Financial Economics 96 (3): 463–84.
doi:10.1016/j.jfineco.2010.02.006.
Lamont, O. A., and C. Polk. 2002. “Does Diversification Destroy Value? Evidence from the
Industry Shocks.” Journal of Financial Economics 63 (1): 51–77. doi:10.1016/S0304405X(01)00089-7.
Lee, Charles, David Ng, and Bhaskaran Swaminathan. 2009. “Testing International Asset Pricing
Models Using Implied Costs of Capital.” Journal of Financial and Quantitative Analysis
44 (2): 307–35. doi:10.1017/S0022109009090164.
Lundholm, Russell, and Richard Sloan. 2007. Equity Valuation and Analysis. 2nd ed. McGrawHill/Irwin.
Mann, H. B., and D. R. Whitney. 1947. “On a Test of Whether One of Two Random Variables Is
Stochastically Larger than the Other.” The Annals of Mathematical Statistics 18 (1): 50–
60. doi:10.1214/aoms/1177730491.
McGahan, Anita M., and Michael E. Porter. 1997. “How Much Does Industry Matter, Really?”
Strategic Management Journal 18 (July): 15–30.
27
———. 1999. “The Persistence of Shocks to Profitability.” The Review of Economics and
Statistics 81 (1): 143–53.
———. 2002. “What Do We Know about Variance in Accounting Profitability?” Management
Science 48 (7): 834–51.
Mueller, Dennis C. 1977. “The Persistence of Profits above the Norm.” Economica, New Series,
44 (176): 369–80. doi:10.2307/2553570.
Newey, Whitney K., and Kenneth D. West. 1987. “A Simple, Positive Semi-Definite,
Heteroskedasticity and Autocorrelation Consistent Covariance Matrix.” Econometrica 55
(3): 703–8. doi:10.2307/1913610.
Ohlson, James A. 1990. “A Synthesis of Security Valuation Theory and the Role of Dividends,
Cash Flows, and Earnings*.” Contemporary Accounting Research 6 (2): 648–76.
doi:10.1111/j.1911-3846.1990.tb00780.x.
Pastor, Lubos, Meenakshi Sinha, and Bhaskaran Swaminathan. 2008. “Estimating the
Intertemporal Risk-Return Tradeoff Using the Implied Cost of Capital.” Journal of
Finance 63 (6): 2859–97. doi:10.1111/j.1540-6261.2008.01415.x.
Penman, Stephen. 2003. Financial Statement Analysis and Security Valuation. 2nd ed. McGrawHill/Irwin.
Penman, Stephen H. 1983. “The Predictive Content of Earnings Forecasts and Dividends.” The
Journal of Finance 38 (4): 1181–99. doi:10.2307/2328019.
Penman, Stephen H., and Xiao-Jun Zhang. 2002. “Modeling Sustainable Earnings and P/E Ratios
with
Financial
Statement
Analysis.”
SSRN
eLibrary,
June.
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=318967.
Rogers, W. H. 1993. “Regression Standard Errors in Clustered Samples.” Stata Technical
Bulletin 13: 19–23.
Ruefli, Timothy W., and Robert R. Wiggins. 2003. “Industry, Corporate, and Segment Effects
and Business Performance: A Non-Parametric Approach.” Strategic Management
Journal 24 (9): 861–79.
Schmalensee, R. 1985. “Do Markets Differ Much.” American Economic Review 75 (3): 341–51.
Waring, Geoffrey F. 1996. “Industry Differences in the Persistence of Firm-Specific Returns.”
The American Economic Review 86 (5): 1253–65.
Watts, Ross L., and Richard W. Leftwich. 1977. “The Time Series of Annual Accounting
Earnings.” Journal of Accounting Research 15 (2): 253–71. doi:10.2307/2490352.
Wilcoxon, Frank. 1945. “Individual Comparisons by Ranking Methods.” Biometrics Bulletin 1
(6): 80–83.
28
TABLE 1
Data overview and variable definitions
Panel A: Data overview
Analysis
Specification
In-sample regressions
Out-of-sample tests
Firm-level
Baseline
1977-2010
1987-2011
First augmented
Second augmented
Baseline
1977-2010
1967-2010
1977-2010
1977-2010
1987-2011
1987-2011
1987-2011
1987-2011
Segment-level
First augmented
Panel A gives an overview on the periods of the data used in the in-sample regressions and out-of-sample tests in the firm- and
segment-level analyses for the three specifications of the forecasting model.
29
TABLE 1 (Continued)
Data overview and variable definitions
Panel B: Variable definitions
Variable name
Computation
Description
Firm-level analysis
(Compustat fundamentals annual)
Segment-level analysis
(Compustat segments)
(USD million)
NI t
Income before extraordinary items –
available for common equity
Compustat item 237
WRDS mnemonic: IBCOM
BV t
Common/ordinary shareholder’s equity
Compustat item 60
WRDS mnemonic: CEQ
OPINC t
Operating income after depreciation
Compustat item: 178
WRDS mnemonic: OIADP
Compustat item: XXX
WRDS mnemonic: OPS
TA t
Identifiable/total assets
Compustat item 6
WRDS mnemonic: AT
Compustat item: XXX
WRDS mnemonic: IAS
SALES t
Total sales
Compustat item: 12
WRDS mnemonic: SALE
WRDS mnemonic: SALES
†
Net operating assets
Common stock (60/CEQ) + preferred
stock (130/PSTK) + long-term debt
(9/DLTT) + debt in current liabilities
(34/DLC) + minority interest (38/MIB) –
cash and short-term investments (1/CHE)
ROA t
Return on assets
OPINC t /(0.5*(TA t + TA t –1))
OPINC t /(0.5*(TA t + TA t –1))
ROS t
Return on sales
OPINC t /(0.5*(SALES t + SALES t –1))
OPINC t /(0.5*(SALES t + SALES t –1))
RNOA t
Return on net operating assets
OPINC t /(0.5*(NOA t + NOA t –1))
ROE t
Return on equity
NI t /(0.5* (BV t + BV t –1))
GSL t
Sales growth
(SALES t - SALES t –1)/ SALES t –1
NOA t
†
(SALES t - SALES t –1)/ SALES t –1
If the data items for preferred stock, long-term debt, debt in current liabilities, minority interest and cash and short-term investments are not available, they are assumed to equal zero.
30
TABLE 2
Sample selection and descriptive statistics
Panel A: Sample selection
Adjustments to data sample
ROA
Observations for in-sample regressions:
Total observations, excluding utilities and financial firms/segments
Less observations with small denominators
Less observations with an absolute value larger than 1
Less observations with more than 100% growth
Less upper and lower centiles observations
Observations for out-of-sample tests, out of which:
Single-segment (as reported)
Multiple-segment (as reported)
254,248
159,188
159,150
140,882
138,066
Firm-level data
(firm-year observations)
ROS
ROE
246,176
158,737
157,494
140,153
137,351
253,419
158,381
154,808
138,352
135,586
40,010
23,369
16,641
RNOA
253,354
158,346
155,469
139,165
136,383
Segment-level data
(segment-year observations)
ROA
ROS
227,572
213,022
210,397
183,217
179,553
249,732
233,783
225,155
180,248
176,644
77,443
Panel A summarizes the sample selection procedure and the number of observations available after each filter. Besides utilities and financials, we also exclude the U.S. postal service (SIC
43), non-classifiable establishments (SIC 99) and observations without SIC code. For variable definitions, see table 1. Single-segment firms are firms that report only one segment; multiplesegment firms are those reporting more than one segment. Firms with missing segment data or where the aggregate segment data deviate substantially from firm data are excluded in the outof-sample tests. For more details, see section 3.
31
TABLE 2 (Continued)
Sample selection and descriptive statistics
Panel B: Descriptive statistics
Variable
Mean
Std. dev.
Firm-level: 5,977 firms (40,010 firm-year observations)
OPINC
297.9
1,224.7
NI
155.2
830.1
TA
3,431.0
13,376.9
SALES
3,200.2
11,790.4
BV
1,313.4
5,265.6
NOA
2,007.7
7,996.1
ROA
8.62%
7.61%
ROS
8.47%
9.04%
ROE
7.50%
15.05%
RNOA
14.30%
13.88%
Segment-level: 14,979 segments (77,443 segment-year observations)
OPINC
121.3
550.4
TA
1,279.1
5,131.1
SALES
1,408.4
6,369.9
ROA
7.95%
13.29%
ROS
6.93%
12.74%
First quartile
Median
Third quartile
5.0
1.1
101.4
124.9
50.4
63.1
4.35%
3.26%
2.40%
6.66%
27.4
11.3
363.8
430.9
173.4
226.1
8.68%
7.37%
9.81%
13.41%
139.7
66.9
1,612.6
1,675.0
691.8
994.3
13.19%
12.60%
16.02%
21.38%
0.6
32.8
42.2
2.33%
1.66%
9.8
150.5
184.0
8.58%
6.79%
61.1
664.7
751.5
14.93%
13.01%
Panel B gives an overview on the firm and segment data used to compute the average forecast improvements in the out-of-sample tests for the period from 1987
to 2011. OPINC (operating income), NI (income before extraordinary items), TA (total assets), SALES (total sales), BV (common shareholder’s equity), and
NOA (net operating assets) are reported in USD million.
32
TABLE 2 (Continued)
Sample selection and descriptive statistics
Panel C: Descriptive statistics by industry
Twodigit
SIC
01
10
12
13
14
15
16
17
20
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
Firm-level data
Description
Agricultural production-crops
Metal mining
Coal mining
Oil & gas extraction
Nonmetallic minerals
General building
Heavy construction
Special trade contractors
Food & kindred products
Textile mill products
Apparel & other textile
Lumber & wood
Furniture & fixtures
Paper & allied products
Printing & publishing
Chemicals & allied products
Petroleum & coal
Rubber & plastic products
Leather
Stone, clay & glass
Primary metal products
Fabricated metal products
Industrial machinery & equipment
Electronic & other electric
equipment
Transportation equipment
Instruments & related products
Misc. manufacturing industries
Railroad transportation
Segment-level data
Obs.
ROA
ROS
ROE
RNOA
Obs.
ROA
ROS
272
4.88%
9.10%
4.04%
7.50%
16.35%
12.13%
14.71%
10.59%
15.47%
13.52%
16.55%
18.00%
14.28%
16.51%
17.04%
12.64%
12.12%
15.04%
13.78%
12.21%
196
307
177
2,051
219
643
303
339
2,470
763
864
703
715
1,415
1,453
4,773
590
1,533
415
894
1,806
2,177
6,593
6,891
6.53%
3.81%
6.73%
6.15%
12.22%
5.36%
6.52%
6.67%
11.00%
7.61%
9.34%
8.12%
8.84%
10.66%
11.28%
10.71%
8.08%
11.21%
6.70%
9.66%
8.38%
11.14%
6.74%
6.04%
9.19%
6.41%
8.99%
12.43%
14.23%
4.10%
3.43%
3.51%
7.88%
5.31%
6.41%
5.91%
5.46%
9.58%
9.85%
9.47%
4.86%
7.72%
4.17%
8.96%
6.23%
7.84%
4.88%
4.56%
941
11
395
48
6.17%
10.12%
8.24%
7.12%
13.38%
9.55%
7.58%
5.18%
5.85%
11.28%
9.85%
7.05%
9.26%
15.64%
11.38%
15.14%
1,615
467
597
354
464
845
774
2,658
462
688
213
407
1,067
1,077
2,991
3,751
10.21%
8.50%
9.74%
6.73%
9.97%
9.17%
10.63%
10.62%
8.73%
10.61%
10.25%
8.72%
8.03%
9.50%
7.87%
6.92%
7.77%
6.37%
7.48%
5.99%
6.57%
9.19%
10.32%
11.36%
8.15%
8.02%
6.33%
9.20%
7.12%
8.06%
6.97%
6.43%
11.08%
3.98%
7.32%
4.98%
9.00%
8.93%
9.71%
10.80%
10.23%
8.36%
7.09%
9.00%
6.22%
8.48%
6.60%
4.55%
1,472
2,731
555
342
8.86%
8.98%
7.82%
7.07%
7.81%
9.28%
6.62%
17.06%
9.22%
7.25%
4.52%
9.74%
15.64%
15.09%
12.89%
11.67%
2,590
5,357
902
350
10.04%
6.80%
7.05%
7.22%
7.19%
5.93%
5.73%
16.71%
33
Panel C: Descriptive statistics by industry (Continued)
Twodigit
SIC
42
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
70
72
73
75
78
79
80
82
83
87
Total
Firm-level data
Segment-level data
Description
Obs.
ROA
ROS
ROE
RNOA
Obs.
ROA
ROS
Trucking & warehouse
Water transportation
Transportation by air
Pipelines, except natural gas
Transportation services
Communications
Electric, gas & sanitary services
Wholesale trade-durable products
Wholesale trade-nondurable
goods
Building materials
General merchandise stores
Food stores
Automotive dealers & services
Apparel & accessory stores
Furniture stores
Eating & drinking places
Miscellaneous retail
Hotels & other lodging places
Personal services
Business services
Auto repair, services & parking
Motion pictures
Amusement & recreation services
Health services
Educational services
Social services
Engineering & management
services
568
237
454
8.96%
5.99%
6.15%
6.16%
14.22%
7.15%
8.04%
5.04%
5.49%
14.87%
8.02%
11.93%
55
2,014
181
1,592
840
10.78%
10.39%
5.42%
7.79%
7.87%
9.97%
20.39%
10.21%
4.74%
4.81%
8.94%
11.99%
3.40%
6.97%
8.37%
21.16%
15.84%
8.03%
12.43%
13.96%
881
568
855
21
416
3,193
657
2,908
1,618
8.90%
6.28%
4.97%
6.72%
9.74%
9.58%
5.64%
7.36%
8.55%
5.82%
13.55%
4.96%
22.80%
8.60%
18.56%
9.10%
3.55%
4.50%
129
596
598
138
632
341
972
923
233
18
2,412
11
162
453
655
8.01%
8.80%
9.18%
7.54%
9.86%
8.09%
10.20%
8.47%
4.79%
5.11%
7.68%
6.28%
5.25%
8.74%
10.64%
3.96%
4.73%
3.26%
3.48%
4.98%
4.36%
7.64%
5.04%
9.37%
8.95%
8.13%
9.18%
5.34%
12.89%
10.45%
3.23%
8.04%
8.59%
6.29%
8.38%
5.05%
7.41%
6.25%
1.19%
5.02%
5.87%
3.02%
-2.10%
4.12%
7.76%
12.17%
14.50%
15.84%
10.21%
20.04%
15.32%
14.62%
13.98%
6.48%
8.95%
15.82%
9.23%
8.39%
12.48%
15.23%
593
8.22%
6.68%
6.25%
15.49%
267
846
777
335
1,026
577
1,689
1,641
478
193
6,426
192
537
981
1,335
173
48
1,316
7.20%
7.37%
8.71%
7.28%
10.90%
5.95%
7.78%
8.10%
5.89%
9.63%
5.46%
4.94%
3.22%
7.45%
8.67%
14.75%
12.86%
8.45%
3.68%
3.83%
2.82%
4.00%
5.21%
3.25%
5.34%
4.28%
10.13%
10.10%
5.31%
7.23%
3.88%
9.77%
7.82%
12.20%
9.97%
6.19%
40,010
8.62%
8.47%
7.50%
14.30%
77,443
7.95%
6.93%
Panel C reports the number of observations and the average firm and segment profitability by industry (in two-digit SIC).
34
TABLE 3
Firm profitability forecast improvements of industry-specific analysis over economy-wide analysis (firm-year
observations: 40,010)
In-sample estimation
AR(1) + NL dummy
Value
ROE
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
RNOA
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
ROA
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
ROS
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
p -Value
AR(1) + NL dummy + PREDGSL
Value
p -Value
0.0076%
0.0068%
-0.0120%
0.0102%
15 / 13
4/4
0.687
0.662
0.814
0.968
0.0905%
0.0929%
0.0385% ***
-0.0139%
14 / 11
9/6
0.103
0.123
<0.001
0.968
0.0095%
0.0094%
0.0087%
-0.0003%
0.512
0.450
0.101
0.696
0.0191%
0.0184%
0.0195% ***
0.0176%
12 / 5
7/4
0.378
0.381
<0.001
0.192
0.785
0.763
0.316
0.397
0.0163%
0.0168%
0.0172% ***
0.0113% *
16 / 7
6/6
0.259
0.254
<0.001
0.098
0.0568%
0.0561%
0.0535%
0.0568%
<0.001
<0.001
<0.001
<0.001
12 / 8
5/3
0.0025%
0.0025%
0.0059%
0.0040%
12 / 8
6/5
0.0373%
0.0366%
0.0248%
0.0224%
***
***
***
**
19 / 8
10 / 1
0.002
0.002
<0.001
0.016
***
***
***
***
18 / 4
15 / 1
This table summarizes the firm profitability forecast improvement of industry-specific analysis over economy-wide analysis. The firm
profitability forecast is the out-of-sample predicted value based on the in-sample estimation on a rolling basis using the most recent 10 years
of data. Three specifications are used for the in-sample estimation: (i) the baseline AR(1), i.e., the first-order autoregressive model of firm
profitability; (ii) an AR(1) model augmented with a dummy variable (NL dummy) for observations with below-average firm profitability; (iii)
an AR(1) model augmented with both the NL dummy and the predicted sales growth (PREDGSL ), where PREDGSL is the predicted value of
the first-order autoregressive model of sales growth estimated on a rolling basis using the most recent 10 years of data (for more details, see
section 2.2). For brevity, the results for the baseline AR(1) model are not reported but are available on request.
The pooled mean (median) is the mean (median) forecast improvement based on all the firm-year improvements pooled together. The grand
mean (median) is the mean (median) of the yearly mean (median) forecast improvements for all the firms in a year. For the pooled mean, the
p-values are based on the standard errors corrected for two-way clustering by firm and year following Rogers (1993). For the grand mean, the
standard errors are adjusted following Newey and West (1987). The p-vales for the pooled and grand median improvements are obtained from
a one-sample Wilcoxon (1945) signed-rank test using the pooled data and the yearly medians, respectively. “No. industries” is the number of
industries (out of 50) for which the mean improvement from using the industry-specific model is significantly positive / negative (at the 10%
significance level). “No. years” is the number of years (out of 25) that the yearly mean improvement is significantly positive / negative (at the
10% significance level).
35
TABLE 4
Firm profitability forecast improvements of industry-specific analysis over economy-wide analysis by firm type
Panel A: In-sample estimation AR(1) + NL dummy
Firm type
Observations
Single-segment
(SS) firms
19,756
Value
ROE
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
RNOA
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
ROA
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
ROS
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
p -Value
Multiple-segment
(MS) firms
13,003
Value
p -Value
Change firms
Difference
SS-MS
7,251
Value
p -Value
Value
p -Value
Value
p -Value
0.0035%
0.0034%
0.0207%
-0.0014%
0.917
0.894
0.126
0.638
0.0122%
0.0114%
-0.0023%
-0.0153%
0.743
0.686
0.805
0.619
0.011
0.005
0.010
0.510
-0.0038%
-0.0024%
0.0060%
0.0422%
0.894
0.913
0.720
0.288
0.0116%
0.0107%
-0.0051%
-0.0118%
10 / 11
2/2
0.610
0.539
0.201
0.882
0.0081%
0.0072%
-0.0258%
-0.0104%
13 / 10
3/2
0.781
0.761
0.377
0.619
-0.0040%
-0.0042%
-0.0235%
0.0050%
14 / 9
1/1
0.901
0.875
0.644
0.840
0.0397% **
0.0388% **
0.0206% ***
0.0225%
10 / 6
8/1
0.023
0.011
0.001
0.221
-0.0214%
-0.0200%
-0.0017%
0.0097%
0.322
0.258
0.581
0.757
-0.0176%
-0.0176%
-0.0076%
-0.0324%
11 / 5
4/2
0.496
0.418
0.387
0.242
0.0611% **
0.0588% ***
0.0222% ***
0.0128%
7/9
1/4
Difference
MS-Change
0.0214% **
0.044
0.0206% **
0.031
0.0216% *** <0.001
0.0212% **
0.032
10 / 8
9/0
-0.0241% *
-0.0229% **
-0.0142% ***
-0.0075%
7/9
1/8
0.062
0.040
0.002
0.192
-0.0013%
-0.0004%
0.0059%
0.0085%
13 / 6
3/3
0.935
0.976
0.788
0.657
0.0455%
0.0435%
0.0358%
0.0286%
***
***
***
***
0.001
<0.001
<0.001
0.002
-0.0228%
-0.0225% *
-0.0200% **
-0.0160%
0.156
0.068
0.037
0.150
0.0749% ***
0.0740% ***
0.0512% ***
0.0478% ***
15 / 4
12 / 0
-0.0002%
-0.0002%
0.0076%
0.0029%
12 / 6
3/2
0.988
0.986
0.651
0.677
0.0024%
0.0048%
-0.0126%
-0.0053%
14 / 6
4/4
0.911
0.811
0.551
0.493
0.0751%
0.0742%
0.0436%
0.0449%
***
***
***
***
<0.001
<0.001
<0.001
<0.001
-0.0027%
-0.0050%
0.0202%
0.0082%
0.909
0.794
0.461
0.353
<0.001
<0.001
<0.001
<0.001
36
Panel B: In-sample estimation AR(1) + NL dummy + PREDGSL
Firm type
Observations
ROE
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
RNOA
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
ROA
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
ROS
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
Single-segment
(SS) firms
19,756
Value
p -Value
0.0956% *
0.068
0.0898%
0.101
0.0417% *** <0.001
-0.0112%
0.904
11 / 9
7/2
0.0452%
0.0411%
0.0340%
0.0134%
**
**
***
*
8/5
6/0
0.044
0.049
<0.001
0.083
0.0303% **
0.017
0.0283% **
0.021
0.0278% *** <0.001
0.0133% **
0.015
12 / 8
7/0
0.0932% ***
0.0907% ***
0.0784% ***
0.0798% ***
16 / 3
16 / 0
<0.001
<0.001
<0.001
<0.001
Multiple-segment
(MS) firms
13,003
Value
p -Value
Change firms
Difference
SS-MS
7,251
Value
p -Value
Difference
MS-Change
Value
p -Value
Value
p -Value
0.0028%
-0.0053%
0.0079%
-0.0116%
0.943
0.867
0.397
0.737
0.0203%
-0.0021%
-0.0091%
-0.0084%
0.652
0.945
0.782
0.757
0.0928%
0.0951%
0.0338% ***
0.0004%
12 / 8
7/3
0.193
0.185
0.005
0.968
0.0726%
0.0972%
0.0429% **
0.0088%
15 / 7
6/0
0.175
0.121
0.013
0.313
-0.0146%
-0.0139%
0.0075%
0.0078%
10 / 7
5/4
0.640
0.624
0.455
0.677
0.0084%
0.0190%
0.0140%
0.0157%
0.760
0.516
0.337
0.288
0.0598% **
0.0550% ***
0.0265% **
0.0056%
0.014
0.008
0.025
0.174
-0.0230%
-0.0329%
-0.0065%
-0.0079%
0.416
0.173
0.758
0.696
-0.0062%
-0.0050%
0.0022%
-0.0189%
12 / 8
4/9
0.759
0.794
0.924
0.600
0.0183%
0.0264%
0.0188% *
0.0159%
12 / 3
5/3
0.372
0.264
0.061
0.201
0.0365%
0.0333%
0.0256%
0.0321%
**
**
***
**
0.016
0.015
0.000
0.023
-0.0245%
-0.0314% **
-0.0166%
-0.0347%
0.109
0.021
0.157
0.211
0.0141%
0.0139%
0.0320% ***
0.0248% **
11 / 4
4/2
0.483
0.455
0.002
0.048
0.0341%
0.0408% *
0.0249% **
0.0218%
15 / 3
8/2
0.131
0.079
0.012
0.242
0.0792%
0.0768%
0.0464%
0.0551%
***
***
***
***
<0.001
<0.001
<0.001
0.002
-0.0200%
-0.0269%
0.0071%
0.0030%
0.388
0.188
0.984
0.581
8/3
4/3
37
This table summarizes the firm profitability forecast improvement of industry-specific analysis over economy-wide analysis for three sub-samples of firms. Single-segment firms are firms that
report only one business segment; multiple-segment firms are firms that report more than one business segment. Change firms are firms that have changed the number of reported segments from
one in 1997 to more than one in 1999, suggesting that they might not be genuinely single-segment firms prior to the introduction of SFAS 131 in 1998. Owing to the doubt in correctly
classifying these firms, they are excluded from the sub-samples of single- and multiple-segment firms but form a category on their own (see section 3 for details). The three columns on the right
present the differences between the three sub-samples.
Panel A reports the out-of-sample tests when the in-sample estimation is based on an AR(1) model augmented with a dummy variable (NL dummy) for observations with below-average firm
profitability. Panel B reports the results for an AR(1) model augmented with both the NL dummy and the predicted sales growth (PREDGSL). For brevity, the results for the baseline AR(1)
model are not reported but are available on request.
See the footnote of table 3 for details on the calculation of the p-values for the mean and median forecast improvements. The p-values for the difference between the forecast improvements for
two groups of firms are calculated as follows. For the pooled mean, the p-values are based on the standard errors corrected for two-way clustering by firm and year following Rogers (1993) using
a regression on a constant and a firm-type dummy (e.g., a multi-segment firm dummy when comparing single- and multiple-segment firms). For the grand mean, the p-values are based on a
paired t-test on the yearly means. For the pooled median, the p-values are calculated using a two-group Mann and Whitney (1947) rank-sum test. For the grand median, the p-values are based on
a two-sample, paired Wilcoxon (1945) signed-rank test of the yearly medians. ***,**, and * indicate statistical significance at the 1%, 5%, and 10% level, respectively.
“No. industries” is the number of industries (out of 48 for single-segment and change firms and 50 for multiple-segment firms) for which the mean improvement from using the industry-specific
model is significantly positive / negative (at the 10% significance level). “No. years” is the number of years (out of 25) that the yearly mean improvement is significantly positive / negative (at
the 10% significance level).
38
TABLE 5
Firm profitability forecast improvements of industry-specific analysis over economy-wide analysis by firm type and
around the introduction of SFAS 131
Panel A: ROE
Firm type
Single-segment
(as reported)
Value
p -Value
Multiple-segment
(as reported)
Value
p -Value
Difference
Value
p -Value
In-sample estimation AR(1) + NL dummy
3 years before
3 years after
Change
Pooled mean
Grand mean
Pooled mean
Grand mean
Pooled mean
Grand mean
-0.0123%
-0.0129%
-0.0020%
-0.0024%
0.0103%
0.0105%
0.779
0.783
0.944
0.939
0.819
0.844
-0.0821%
-0.0857%
0.0225%
0.0202%
0.1046%
0.1059%
0.314
0.395
0.750
0.797
0.272
0.372
0.0698%
0.0728%
-0.0246%
-0.0226% ***
-0.0944%
-0.0954%
0.168
0.254
0.699
0.749
0.181
0.283
0.598
0.628
0.688
0.710
0.458
0.535
-0.1357%
-0.1377%
0.0252%
0.0220%
0.1609%
0.1597%
0.105
0.226
0.788
0.834
0.148
0.260
0.1082% ***
0.1091% *
-0.0168%
-0.0143%
-0.1250%
-0.1235%
0.007
0.069
0.828
0.867
0.101
0.205
In-sample estimation AR(1) + NL dummy + PREDGSL
3 years before
3 years after
Change
Pooled mean
Grand mean
Pooled mean
Grand mean
Pooled mean
Grand mean
-0.0276%
-0.0286%
0.0084%
0.0076%
0.0359%
0.0362%
Panel B: RNOA
Firm type
Single-segment
(as reported)
Value
p -Value
Multiple-segment
(as reported)
Value
p -Value
Difference
Value
p -Value
0.0725%
0.0704%
0.0222%
0.0245%
-0.0503%
-0.0460%
0.132
0.228
0.796
0.800
0.507
0.651
In-sample estimation AR(1) + NL dummy
3 years before
3 years after
Change
Pooled mean
Grand mean
Pooled mean
Grand mean
Pooled mean
Grand mean
-0.0416% *
-0.0424%
0.0227%
0.0236%
0.0643%
0.0661%
0.094
0.205
0.584
0.620
0.134
0.229
-0.1142% **
-0.1129%
0.0005%
-0.0008%
0.1146% *
0.1120%
0.109
0.227
0.708
0.722
0.111
0.204
-0.1611%
-0.1592%
-0.0085%
-0.0098%
0.1527%
0.1494%
0.017
0.116
0.993
0.988
0.070
0.164
In-sample estimation AR(1) + NL dummy + PREDGSL
3 years before
3 years after
Change
Pooled mean
Grand mean
Pooled mean
Grand mean
Pooled mean
Grand mean
-0.0467%
-0.0474%
0.0105%
0.0110%
0.0572%
0.0584%
*** <0.001
*
0.060
0.849
0.844
*** 0.008
*
0.067
0.1144% *** <0.001
0.1118% **
0.030
0.0190%
0.763
0.0208%
0.769
-0.0955%
0.133
-0.0910%
0.233
39
Panel C: ROA
Firm type
Single-segment
(as reported)
Value
p -Value
Multiple-segment
(as reported)
Value
p -Value
Difference
Value
p -Value
0.0411%
0.0403%
0.0512%
0.0527%
0.0101%
0.0124%
0.240
0.323
0.243
0.345
0.842
0.826
In-sample estimation AR(1) + NL dummy
3 years before
3 years after
Change
Pooled mean
Grand mean
Pooled mean
Grand mean
Pooled mean
Grand mean
-0.0187%
-0.0191%
0.0502%
0.0506%
0.0689%
0.0697%
0.204
0.284
*** 0.002
*
0.083
*** <0.001
**
0.027
-0.0598% *
-0.0593%
-0.0009%
-0.0021%
0.0588%
0.0573%
0.073
0.184
0.978
0.956
0.171
0.273
-0.0860% ***
-0.0853% *
0.0017%
0.0005%
0.0877% **
0.0858%
0.006
0.094
0.961
0.991
0.039
0.131
In-sample estimation AR(1) + NL dummy + PREDGSL
3 years before
3 years after
Change
Pooled mean
Grand mean
Pooled mean
Grand mean
Pooled mean
Grand mean
-0.0208%
-0.0211%
0.0408%
0.0410%
0.0616%
0.0621%
0.125
0.226
*** <0.001
**
0.016
*** <0.001
*** 0.010
0.0652% **
0.0642%
0.0391%
0.0405%
-0.0261%
-0.0237%
0.019
0.115
0.305
0.394
0.540
0.623
Panel D: ROS
Firm type
Single-segment
(as reported)
Value
p -Value
Multiple-segment
(as reported)
Value
p -Value
Difference
Value
p -Value
In-sample estimation AR(1) + NL dummy
3 years before
3 years after
Change
Pooled mean
Grand mean
Pooled mean
Grand mean
Pooled mean
Grand mean
0.0462%
0.0464%
0.1164%
0.1159%
0.0702%
0.0695%
*** <0.001
**
0.013
*** <0.001
**
0.018
*** <0.001
**
0.014
-0.0122%
-0.0122%
0.0074%
0.0061%
0.0196%
0.0183%
0.338
0.139
0.853
0.892
0.593
0.670
0.0584%
0.0586%
0.1090%
0.1099%
0.0506%
0.0512%
***
**
***
*
0.001
0.030
0.001
0.067
0.106
0.182
-0.0396%
-0.0381%
0.0341%
0.0330%
0.0737% *
0.0710%
0.168
0.298
0.306
0.426
0.057
0.174
0.0839% *** 0.008
0.0827%
0.108
0.0880% *** <0.001
0.0885% **
0.034
0.0040%
0.900
0.0057%
0.874
In-sample estimation AR(1) + NL dummy + PREDGSL
3 years before
3 years after
Change
Pooled mean
Grand mean
Pooled mean
Grand mean
Pooled mean
Grand mean
0.0444%
0.0446%
0.1221%
0.1214%
0.0777%
0.0768%
*** 0.003
*
0.068
*** <0.001
**
0.018
*** 0.001
**
0.021
This table compares the firm profitability forecast improvement of industry-specific analysis over economy-wide analysis before and after the
introduction of SFAS 131 in 1998. The observations of 1998 are excluded from the out-of-sample tests to account for the transition year. In this
table, the grouping of the firms into single- and multiple-segment is as reported: Firms that changed the number of reported business segments
from one to more than one following the introduction of SFAS 131 are classified as single-segment in the 3 years before (1995 to 1997) and as
multiple-segments in the 3 years after (1999 to 2001).
Panels A to D report the results for the four profitability measures considered: ROE, RNOA, ROA, and ROS. For brevity, only the results for the
pooled and grand mean forecast improvements are reported. The results for the pooled and grand median improvements and those using the
baseline AR(1) model for in-sample estimation are not reported but are available on request.
The p-values for the difference between the forecast improvements for the single- and multiple-segment firms are calculated as follows. For the
pooled mean, the p-values are based on the standard errors corrected for two-way clustering by firm and year following Rogers (1993) using a
regression on a constant and a multiple-segment firm dummy. For the grand mean, the p-values are based on a paired t-test on the yearly mean
improvements for the single- and multiple-segment firms. ***,**, and * indicate statistical significance at the 1%, 5%, and 10% level,
respectively.
40
TABLE 6
Segment profitability forecast improvements of industry-specific analysis over economy-wide
analysis (segment-year observations: 77,443)
In-sample estimation AR(1) + NL dummy
ROA
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
ROS
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
Value
p -Value
0.0214%
0.0167%
0.0096% ***
-0.0111%
0.428
0.542
<0.001
0.861
16 / 13
7 / 12
0.0643%
0.0580%
0.0573%
0.0154%
**
**
***
*
0.012
0.029
<0.001
0.083
23 / 6
10 / 2
This table summarizes the segment profitability forecast improvement of industry-specific analysis over economy-wide
analysis. The segment profitability forecast is the out-of-sample predicted value based on the in-sample estimation on a
rolling basis using the most recent 10 years of data. Two specifications are used for the in-sample estimation: (i) the
baseline AR(1), i.e., the first-order autoregressive model of segment profitability; (ii) an AR(1) model augmented with
a dummy variable (NL dummy) for observations with below-average segment profitability. For brevity, the results for
the baseline AR(1) model are not reported. They are available on request.
The pooled mean (median) is the mean (median) forecast improvement based on all the segment-year improvements
pooled together. The grand mean (median) is the mean (median) of the yearly mean (median) forecast improvements
for all the segments in a year. For the pooled mean, the p-values are based on the standard errors corrected for two-way
clustering by segment and year following Rogers (1993). For the grand mean, the standard errors are adjusted
following Newey and West (1987). The p-vales for the pooled and grand median improvements are obtained from a
one-sample Wilcoxon (1945) signed-rank test using the pooled data and the yearly medians, respectively. “ No.
industries ” is the number of industries (out of 55) for which the mean improvement from using the industry-specific
model is significantly positive / negative (at the 10% significance level). “No. years” is the number of years (out of 25)
that the yearly mean improvement is significantly positive / negative (at the 10% significance level).
41
TABLE 7
Segment profitability forecast improvements of industry-specific analysis over economy-wide analysis by
reporting regime
In-sample estimation AR(1) + NL dummy
Period
Observations
ROA
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
ROS
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
Pre-SFAS 131
(1987-1997)
39,624
Post-SFAS 131
(1999-2011)
34,860
Value
p -Value
Value
p -Value
Value
0.0965%
0.1081%
0.0705%
0.0717%
**
0.029
**
0.049
***
<0.001
*
0.091
20 / 8
7/1
-0.0602%
-0.0576%
-0.0400%
-0.0295%
***
<0.001
***
<0.001
***
<0.001
**
0.016
6 / 10
0 / 11
0.1567%
0.1657%
0.1105%
0.1012%
***
***
***
**
<0.001
0.003
<0.001
0.026
0.1423%
0.1513%
0.1319%
0.1136%
***
<0.001
***
0.004
***
<0.001
***
0.006
30 / 4
9/0
-0.0231% **
0.042
-0.0201% *
0.071
-0.0212% ***
<0.001
-0.0193%
0.101
16 / 8
0/2
0.1654%
0.1715%
0.1530%
0.1329%
***
***
***
***
<0.001
<0.001
<0.001
<0.001
Difference
p -Value
This table summarizes the segment profitability forecast improvement of industry-specific analysis over economy-wide analysis for
the pre-SFAS 131 period (1987-1997) and the post-SFAS 131 period (1999-2011), as well as the difference between the two
periods. The observations of 1998 are excluded from the out-of-sample tests to account for the transition year. The table reports the
out-of-sample tests when the in-sample estimation is based on an AR(1) model augmented with a dummy variable (NL dummy) for
observations with below-average firm profitability. For brevity, the results for the baseline AR(1) model are not reported. They are
available on request.
See the footnote of table 6 for details on the calculation of the p-values for the mean and median forecast improvements. The pvalues for the difference between the forecast improvements for the two periods are calculated as follows. For the pooled mean, the
p-values are based on the standard errors corrected for two-way clustering by segment and year following Rogers (1993) using a
regression on a constant and a post-SFAS 131 period dummy. For the grand mean, the p-values are based on a paired t-test on yearly
means. For the pooled median, the p-values are calculated using a two-group Mann and Whitney (1947) rank-sum test. For the grand
median, the p-values are based on a two-sample, paired Wilcoxon (1945) signed-rank test of the yearly medians. ***,**, and *
indicate statistical significance at the 1%, 5%, and 10% level, respectively.
“No. industries” is the number of industries (out of 54 for the pre-SFAS 131 period and 55 for the post-SFAS 131 period) for which
the mean improvement from using the industry-specific model is significantly positive / negative (at the 10% significance level).
“No. years” is the number of years (out of 11 for the pre-SFAS 131 period and 13 for the post-SFAS 131 period) that the yearly
mean improvement is significantly positive / negative (at the 10% significance level).
42
TABLE 8
Sales growth forecast improvements of industry-specific analysis over economy-wide analysis
Panel A: Firm-level analysis (firm-year observations: 40,010)
Value
p -Value
0.0589% **
0.011
0.0609% **
0.016
0.0261% *** <0.001
0.0193%
0.143
9/3
8/3
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
Panel B: Firm-level analysis by firm type
Firm type
Observations
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
Single-segment
(SS) firms
19,756
Value
0.0840%
0.0864%
0.0495%
0.0555%
p -Value
***
0.002
***
0.005
*** <0.001
**
0.028
10 / 4
9/0
Multiple-segment
(MS) firms
13,003
Value
0.0287%
0.0291%
-0.0013%
0.0046%
p -Value
0.402
0.399
0.556
0.861
4/3
5/4
Change firms
Difference
SS-MS
Difference
MS-Change
7,251
Value
p -Value
0.0449%
0.258
0.0340%
0.394
0.0269% *
0.070
-0.0165%
0.778
10 / 3
7/1
Value
p -Value
0.0553%
0.171
0.0573%
0.178
0.0508% ***
0.001
0.0509% **
0.032
Value
-0.0162%
-0.0048%
-0.0282%
0.0211%
p -Value
0.711
0.912
0.271
0.382
Panel C: Segment-level analysis (segment-year observations: 77,443)
Pooled mean
Grand mean
Pooled median
Grand median
No. industries
No. years
Value
p -Value
0.0964% ***
0.001
0.1022% ***
0.002
0.0458% *** <0.001
0.0257%
0.109
14 / 0
14 / 2
This table summarizes the sales growth forecast improvement of industry-specific analysis over economy-wide analysis. The sales growth forecast is the predicted value of the first-order
autoregressive model of sales growth estimated on a rolling basis using the most recent 10 years of data. Panel A reports the sales growth forecast improvement for all firms. Panel B reports
the sales growth forecast improvement for single- and multiple-segment firms and change firms separately, along with the differences between the groups. Panel C reports the sales growth
forecast improvement for all segments. See the footnotes of tables 3, 4, 6, and 7 for details on the calculation of the p-values. ***,**, and * indicate statistical significance at the 1%, 5%,
“No. industries” is the number of industries (out of 50 for all firms in panel A; out of 48 for single-segment and change firms and 50 for multiple-segment firms in panel B; out of 55 for all
segments in panel C) for which the mean improvement from using the industry-specific model is significantly positive / negative (at the 10% significance level). “No. years” is the number
of years (out of 25) that the yearly mean improvement is significantly positive / negative (at the 10% significance level).
43