Inter-Industry Network Structure, Information Transfers and the CrossPredictability of Earnings and Stock Returns *
Daniel Aobdia
Northwestern University, Kellogg School of Management
Judson Caskey
University of Texas Austin, McCombs School of Business
N. Bugra Ozel
UCLA, Anderson School of Management
October 2012
Inter-Industry Network Structure, Information Transfers and the CrossPredictability of Earnings and Stock Returns
Abstract
We examine the role of the inter-industry input-output network structure in the transfer
of information among firms. We focus our attention on the role of industries that serve as ‘hubs,’
or central industries, in the flow of trade across industries. Consistent with a diversification effect,
we find that firms in these central industries are more exposed to systemic risks compared to
other firms. Additionally, earnings-response-coefficients of central firms are lower than those of
other firms, consistent with investors placing less emphasis on the firm-specific information on
account of the relative importance of macroeconomic risk to central firms. Comparing central
industries to non-central industries, we also find that the stock returns and accounting
performance of central industries better predict the performance of industries linked to them.
This suggests that shocks to central industries propagate more strongly than shocks to other
industries. Our results highlight the importance of inter-industry trade flows in information
transfers.
JEL classification: D57; G14; M41.
Keywords: Information transfer; inter-industry networks; systemic risk; earnings; stock returns.
1
Introduction
The accounting literature has long recognized the role of information transfers among
firms in forming beliefs about earnings and returns. In one of the early studies in this area,
Clinch and Sinclair (1987) document intra-industry information transfers. More recent studies
document information transfers via customer/supplier links (e.g., Pandit et al. 2011) and
informativeness of individual firms’ earnings guidance about market returns (e.g., Anilowski et
al. 2007). We provide a specific explanation for why the strength of the information transfers
varies among firms that have similar characteristics and similar positions in the value chain. In
particular, we examine how the inter-industry network structure affects information transfers,
and we show that the effect is economically important.
Our analysis views industries through the lens of the network formed by their reliance on
each other for inputs or as customers. We place particular attention on ‘central’ industries that
form hubs in the network. We measure an industry’s centrality as the extent to which it has
strong direct or indirect ties to a large fraction of the economy. In particular, we use an
eigenvector measure of centrality that characterizes the network as a matrix, and captures both
direct and indirect links between industries. The measure incorporates ‘ripple effects’ whereby
an industry that serves few others may be central because those few that it serves have strong
connections with the remaining industries. In other words, the measure captures not only direct
customer and supplier relationships, but also higher-tier associations such as customers of
customers and suppliers of suppliers.
Our first set of findings provides evidence that firms in central industries (hereafter
central firms) obtain some diversification by virtue of their exposure to wide swaths of the
economy. We estimate the R2s from capital asset pricing model (CAPM) and Fama-French three-
1
factor model as a gauge of the extent to which systemic risk explains stock returns. We then
regress these R2s on explanatory factors, including centrality of the industry the firm belongs to,
firm size, analyst coverage, and trading volume. We find that central firms have higher R2s,
consistent with a diversification effect.
We also examine earnings-response-coefficients (ERCs) to assess whether investors learn
less from earnings surprises of central firms. If central firms track the overall economy more
closely, investors can place less reliance on their firm-specific earnings information. Consistent
with this, we find that central firms have lower ERCs than non-central firms. Prior studies have
shown that the stock price reaction to earnings announcements is greater for early announcing
firms than for late announcing firms (Foster 1980, 1981). 1 We find that this effect is diminished
for central firms. This is consistent with the central firms’ low ERC being due to preemption by
macroeconomic news rather than by other firms’ earnings announcements. 2
Our second set of analyses examines how centrality impacts information transfers. For
each industry, we examine the association between its monthly returns and the concurrent and
one-month-ahead returns of the industries adjacent to it. We find that this association is greater
for central industries. Similarly, we find that, compared to non-central firms, central firms’
seasonally differenced quarterly return-on-assets (ROA) have a stronger association with
adjacent firms’ both concurrent and one-quarter-ahead ROA. In other words, shocks to central
industries propagate more strongly than shocks to non-central industries. This is consistent with
the central firms having some diversification that hedges their exposure to shocks to individual
industries.
1
This is a different issue than the ‘good news early/bad news late’ phenomenon (e.g., Begley and Fischer 1998) that
examines late announcements vis-à-vis when a firm typically announces.
2
An alternative explanation for this finding is that earnings surprises are more transitory for central firms than for
non-central firms and therefore investors do not react to the surprises as strongly. We are in the process of testing
this explanation.
2
Our study relates to the literature on the gradual diffusion of information across asset
markets. Hong, Torous, and Valkanov (2007) show that stock returns of industries that associate
with macroeconomic fundamentals can predict market returns. Along the same lines, Anilowski
et al. (2007) and Bonsall et al. (2012) find that ‘bellwether’ firms’ earnings guidance predicts
market-wide returns, where they identify bellwether firms based on size and past relation with
market-wide returns, respectively. Our study adds to this literature by linking evidence of
information transfers to a specific mechanism – inter-industry trade flows. Additionally, our
study also complements prior work, such as Cohen and Frazzini (2008) and Pandit, Wasley, and
Zach (2011), who study firm-specific customer and supplier relations to identify potential
information transfers. Our study differs from these studies in two respects. First we examine the
associations from a different dimension, namely from the industry perspective. For example if a
firm has several customers, all of which are essentially from the same industry, the firm’s
performance may not associate with each of these individual customers’ performance, whereas it
will more strongly associate with the overall performance of its customers’ industry . Second,
unlike firm-specific data, which require some degree of a concentrated activity with a given
customer or supplier, the BEA input/output tables incorporate all trade-flows. For example, a
company may buy from or sell to a dispersed set of firms, none of which individually warrant
mention as a major customer or supplier at the firm level. Our results suggest that prior findings
on the cross predictability of returns across economically linked firms (Cohen and Frazzini 2008;
Menzly and Ozbas 2010 and Pandit et al. 2011) are likely stronger among non-central firms
because these firms have a less diversified set of input/output connections.
Our study is also related to the literature that examines the propagation of shocks
throughout the economy. In a recent analytical study, Acemoglu et al. (2012) examine a multi-
3
sector setting to analyze conditions under which idiosyncratic shocks can lead to aggregate
fluctuations. They show that shocks to sectors that trade with a disproportionately large number
of other sectors can be amplified into macroeconomic fluctuations. 3 While we do not
differentiate our analysis based on the origination point of the shocks, our finding that central
industries have stronger associations with systemic fluctuations is consistent with this theory.
The paper proceeds as follows. Section 2 discusses the concept of centrality as well as
our empirical predictions. Section 3 describes our data. Section 4 provides evidence that
systemic risk accounts for a relatively large portion of returns for firms in central industries.
Section 5 shows that stock market and accounting based performances of central firms provide
greater information transfer than those of non-central firms. Section 6 concludes.
2
Centrality and Empirical Predictions
2.1 The Setting
In any economy, the structure of inter-industry interactions has an impact on the
importance of an industry, and on the factors that affect that industry’s performance. Some
industries, such as retail trade or plastic production, have strong customer/supplier interactions
with several other industries and therefore are in more central positions. Others, such as tobacco
products manufacturing in the U.S., are relatively isolated.
To fix ideas, consider an economy with N industries where there is one central industry
that makes equal sized trades with all other industries. For simplicity, assume that the other N-1
industries do not trade with each other, leading to the network structure shown in Figure 1. In
this economy, as N increases, a diversification argument based on the law of large numbers
3
As Lucas (1977) and others argue, in highly disaggregated economies, idiosyncratic shocks will remain fairly
confined. While Dupor (1999) and Horvath (1998, 2000) debate on whether sectoral shocks can transfer into
aggregate fluctuations, Acemoglu et al. (2012) provide a more complete answer to this question by showing that
sectoral shocks can lead to aggregate fluctuations only in the presence of asymmetries in the roles different sectors.
Gabaix (2011) provide a similar model using firm-level shocks as a source of aggregate fluctuations.
4
implies that idiosyncratic shocks to non-central industries will have only a negligible impact on
the performance of the central industry. On the other hand, an idiosyncratic shock to the central
industry will propagate to all non-central industries and affect their performance (Acemoglu et al.
2012). Because of these forces, systemic factors play a relatively larger role in determining
central industries’ performance, as compared to the individual performances of non-central
industries.
(Insert Figure 1 about here)
In this setting, the performance of an industry will be informative about the concurrent
and/or future performance of the industries it is tied to. The predictive ability will depend on the
rate at which shocks transfer from one industry to another. If shocks transfer instantaneously,
then a shock to one industry cannot predict future shocks to another industry – it only pertains to
concurrent shocks. Even if shocks propagate with a delay, the relation between non-central
industry performance and central industry performance diminishes as the number of industries
grows. This ultimately results in a central industry’s performance predicting that of individual
non-central industries, but not vice versa.
The setting described above leads to the following two main testable predictions. First,
compared to the performance of the non-central industries, the performance of the central
industries associates more strongly with systemic fluctuations and risks. An extension of this
prediction is that investors will have better anticipated the earnings of central industries. The
reasoning follows an analogy to Ayers and Freeman (1997), who find evidence that investors
better predict the industry component of earnings than the firm-specific component. Similarly,
we expect that investors better predict the macroeconomic component of earnings than the
industry- and firm-specific components. Second, current period shocks to the performance of a
5
central industry are more strongly associated with the concurrent and/or future performance of
the industries it is linked to, than shocks to a non-central industry are associated with the
industries it is linked to.
2.2 Measure of Centrality
The economic story and predictions outlined in the previous section depend on the actual
structure of the interactions in the economy. In this section we introduce our measure of
centrality and take a closer look at the interactions in the U.S. economy. We measure centrality
using the eigenvector centrality measure, which is formally defined by Bonacich (1972). 4
Eigenvector centrality weights an industry’s ties with other industries by the importance of the
industries it is tied to. For example, industry i could be tied to industries x and y, both of which
are not tied to any other industry, and industry j could be tied to q and z, both of which are tied to
a few other industries. Ceteris paribus, industry j will have a higher eigenvector centrality than
industry i because the industries it is tied to are more important in the network than those
industry i is tied to. Formally, the eigenvector centrality for all nodes j ≠ i is calculated as follows:
=
ci
∑
1
c
λ=
j∈M ( i ) j
1
λ
∑ j Aij c j ,
(1)
where ci is the eigenvector centrality of industry i, M (i) is the set of all industries that are tied to
i, λ is a constant and Aij is the weight of the edge between i and j. 5 The term eigenvector
centrality arises because (1) can be written in matrix form as Ac = λ c , an eigenvalue equation.
The constant λ is the largest eigenvalue of the matrix A, which the Perron–Frobenius theorem
implies is the only one whose associated eigenvector c has all positive entries.
In calculation of (1) we define edge weight Aij as strength of the link between industries i
4
Variants of this measure are used in recent studies for similar purposes (see Hochberg, Ljunqvist and Lu 2007;
Ahern and Harford 2010; Acemoglu et al. 2012).
5
In the calculation of simple unweighted eigenvector centrality Aij is defined as an indicator variable that equals 1 if
i is connected to j and zero otherwise. An industry is not treated as linked to itself (Aii = 0) when writing (1) in
matrix form.
6
and j to take into account that certain links between industries are stronger than others. We
calculate the strength of the link between any given industry pair i and j using the following
undirected measure:
A=
ij
sij
s ji
s ji
1  sij

+
+
+
4  ∑ j sij ∑ i sij ∑ i s ji ∑ j s ji


,


(2)
where sij equals to the sales (in dollars) of goods and services by industry i to industry j. The first
(second) ratio measures the sales made by i to j as a percentage of i’s total sales (j’s total
purchases), gauging whether j is an important customer to i (i is an important supplier to j). The
second two ratios similarly measure whether i is an important customer to j, and j is an important
supplier to i, respectively. A smaller Aij indicates a weaker link between industries i and j
accordingly we refer the reciprocal 1 / Aij as the distance between the two industries.
We construct the eigenvector centrality measure using detailed input-output tables
published by the BEA every five years. Following Ahern and Harford (2010) and Anjos and
Fracassi (2012), we use the 1997 “Use Table” for calculation of the centrality measure. We
utilize 1997 data because 1997 is the approximate midpoint of our data period and according to
Anjos and Fracassi (2012) the flows among industries remain highly persistent over years.
We define industries using the 4-digit IO industry codes provided by the BEA. We
exclude government, special industries, value added and final users (industry definitions that
start with the letters S, V, or F). 6 Figure 2 plots the actual network structure by depicting the
connections between 123 industries. While the network in Figure 2 is far more complex than that
in Figure 1, we can conclude that the actual interactions in the U.S. economy resembles the
6
We do not include government/special industry/value added/final uses industries in the calculation of the strength
measure because of the difficulties in the interpretation of certain associations. (e.g., negative sales values, changes
in private inventories etc.) In the calculation of the strength measure, we include private consumption expenditures
to calculate total sales numbers so that our measures gauge importance relative to overall sales, rather than strictly
business-to-business sales.
7
economy depicted in Figure 1, as some industries clearly play a more central role than others
(Acemoglu et al. 2012). In Table 1, we provide a list of the most central industries.
(Insert Figure 2 and Table 1 about here)
3
Data and descriptive statistics
3.1 Data
We obtain accounting data and stock returns data from COMPUSTAT quarterly file and
CRSP daily and monthly files respectively. Analyst-related data come from I/B/E/S/. The specific
data requirements differ for returns related tests and earnings related tests and accordingly for
each analysis we start with the sample of firms from COMPUSTAT and impose additional
restrictions depending on the specification.
We require that the historical NAICS industry code be available in COMPUSTAT in
order to merge our firm-related data with our centrality measures, available at the BEA 4- digit
industry level. This restricts our sample period to the fiscal years after 1985, ending in 2011. We
exclude regulated firms (NAICS 22) and financial institutions (NAICS 52) from the sample
because their earnings and returns depend relatively more on the regulatory environment than for
other firms in the economy.
3.2 Descriptive statistics
Figure 3 plots the centrality measure by decreasing order. The figure exhibits the
common phenomenon that a small number of nodes in the network, industries in our case,
exhibit high connectivity (e.g., Newman 2010). As the figure demonstrates, while most
industries exhibit some connections with the other industries, there are relatively few ‘hub’
industries. Four industries – wholesale trade, construction, management of companies, and real
estate – exhibit the highest degree of centrality, followed by a number of other industries with
8
still relatively high eigenvector centrality. We confirm these observations in Table 2. The
centrality measure quickly decreases from its maximum of 0.266 to 0.097 at the 75th percentile
or a difference of 0.169. The distribution then levels out with 70% of the observations clustered
between the 0.036 values and the 0.097 values, or a difference of 0.060.
(Insert Figure 3 and Table 2 about here)
There are three major discontinuities in the eigenvector centrality values not counting the
two cliffs at the very beginning and very end of the distribution: the first is at 0.216, the second
is at 0.125 and the third is at 0.111. The distribution of centrality values becomes visibly smooth
beyond 0.111 (the dotted line in Figure 3), suggesting that industries below this level are
relatively more comparable in terms of centrality. Accordingly, we choose to define central
industries as those that have eigenvector centrality values greater than 0.111, which accounts for
the top 17% of the industries. This cutoff is conservative enough to ensure that all central
industries are indeed classified as central. Our results remain similar when we set the cutoff at
0.125 or when we use different cutoffs around the baseline level. 7 We do not consider 0.216 as a
cutoff point because as Figure 2 shows there are several well-connected industries that play
relatively important roles in the economy and 0.216 is an aggressive cutoff point that leads only
four industries to be classified as central.
7
In particular, our analyses remain qualitatively similar when using a cutoff of 0.12, or 12% of the distribution of
industries, and when using a cutoff of 0.10 or 21% of the industries.
9
4
Are central firms more exposed to systemic risk?
4.1 R-square regressions
Research design
As we explain in section 2.2, our first prediction is that central firms are more exposed to
systemic risks than non-central firms. In order to determine whether this prediction holds, we
estimate the ability of CAPM and Fama-French three-factor models to explain central firms’
returns. We expect that the R2s from these models will be higher for central firms than for noncentral firms.
We estimate CAPM and Fama-French three-factor models with both daily and monthly
returns data using non-overlapping five-year windows. 8 We restrict our sample to firms that are
in both CRSP and COMPUSTAT. We require firms to have at least 800 days (60 months) of
return data availability to be included in the daily (monthly) regressions. This helps us to limit
the impact of stocks with limited return information during the estimation period.
Upon estimating the CAPM and the three-factor models we test the association between
the R2s from these models and our centrality measure using the following model: 9
Rsquarei ,t= α + β1Eigenvector Centralityi ,t +β 2 Analysts Dummyi ,t
+ β3 Average Proportion Traded i ,t +β 4 Logmktcapi ,t
(3)
We predict that the coefficient on Eigenvector Centrality, β1, should be positive if central firms
are more exposed to systemic risks than non-central firms. Because the R2s in the regressions
could also depend on the information environment and the liquidity of the stocks traded, we add
three control variables (Piotroski and Roulstone 2004, Kelly 2005). 10 Analysts Dummy is equal
8
We use the following fiscal years as cutoffs for the estimation periods: 1985, 1990, 1995, 2000, 2005 and 2010.
Results are not dependent on inclusion or exclusion of the fiscal year 1985 (estimation period from 1981 to 1985) in
the sample.
9
Our approach of using R2s as the dependent variable is parallel to that used in the study of stock return
synchronicity (e.g., Morck et al. 2000 ; Piotroski and Roulstone 2004).
10
For example, the market price of a stock will remain the same if the stock is not traded during a given period,
10
to one if the stock is covered by the analysts at the end of the R2s estimation period, Average
Proportion Traded is the average trading volume as a percentage of shares outstanding during
the period, and Logmkval is the logarithm of the market capitalization of the stock at the end of
the period, equal to shares outstanding times market price at the end of the period in
COMPUSTAT. Based on Kelly (2005)’s findings that R2s from market models increase with
information availability and liquidity, we expect all coefficients on control variables to be
positive. Additional information on the computation of control variables is provided in Appendix
A. Both the dependent variable and control variables are winsorized at the 1st and 99th percentile
in order to reduce the impact of outliers in the specifications.
Results
We present our results from the estimation of (3) in Table 3. The first two columns
present the results using daily returns data. The coefficient on the Eigenvector Centrality is
positive in both CAPM and Fama-French specifications at the 1% level or better. The effect is
also economically important. For example, an increase of the centrality measure from the 25th
percentile to the 75th percentile, or 0.078, corresponds to a 0.90% increase in R2s in the CAPM
specification using daily returns, compared to a sample average R2 of 9.63%. Consistent with
Piotroski and Roulstone’s (2004) and Kelly (2005)’s findings our results indicate that firms
covered by analysts, and firms with a higher market capitalization have higher R2s, indicating
that larger firms are more exposed to systemic risks than other firms. Our analysis using monthly
returns yields results that are similar to those using daily returns. For both CAPM and FamaFrench three factor models, we find that central firms have higher R2s. Overall these findings are
consistent with our prediction that exposure to systemic risks increases with the centrality.
(Insert Table 3 about here)
which would result in a low R2.
11
4.2 Earnings Response Coefficients
Research design
In this section we examine whether investors react less to earnings announcements of
central firms compared to non-central firms. Our prediction follows from our findings in the
previous section that central firms are more exposed to systemic risks than non-central firms.
Given this association, we expect that investors can better anticipate the earnings of central
industries because investors have access to a wide variety of sources of macroeconomic news
that they can impound into prices throughout the quarter. Consequently, we predict that the
earnings response coefficients (ERCs) of central firms should be lower than ERCs of non-central
firms.
We conduct the following regression in order to investigate this hypothesis:
CARit =
α + β1 Surpriseit + β 2 Centrali + β3 Surpriseit × Centrali + β . X it + γ Surpriseit × X it (4)
where CARit is the cumulative abnormal returns for firm i and quarter t, defined over a three-day
window around the earnings announcement date reported by I/B/E/S/. We subtract the weighted
market return to the company cumulative returns; Centrali is an indicator variable equal to one
when the centrality measure is above the cutoff level of 0.111; Surpriseit is equal to the actual
earnings per share less the most recent analysts’ consensus earnings forecast, both scaled by the
stock price at the beginning of the firm’s fiscal quarter. We use the earnings per share and
earnings forecast data from I/B/E/S/ and we require that there is a lag of at least 2 days between
the most recent consensus forecast and the actual announcement date in order to avoid any
information leakage concerns. Because we use analyst forecasts to calculate the surprise, our
sample is restricted to firms that are covered by analysts. This tends to skew the sample towards
larger companies. We also require that the company fiscal quarter ends either in March, June,
September or December. This restriction is necessary because we also examine the impact of the
12
timing of announcement relative to peer firms.
We also include a vector of control variables (Xit) that are identified as determinants of
earnings response coefficients in prior research (see for example Francis and Ke, 2006, Lim and
Tan, 2008 and Aboody, Aobdia and Hughes, 2012). Xit includes the following: Tobinq, the Tobin
Q coefficient defined as the market value of assets (i.e., total liabilities plus market value of
equity at the end of the fiscal quarter) divided by the book value of total assets; StdReturns, the
90 calendar day stock returns volatility ending three days prior to the earnings announcement
date; Leverage, the firm book leverage (total debt over total debt plus book value of equity);
Logmkval, the natural logarithm of the firm fiscal quarter end market value; Loss, an indicator
variable equal to one when the actual I/B/E/S/ EPS is negative; Specialitems, an indicator
variable equal to one when special items during the quarter are above 5% of the total assets value;
Fourthquarter, an indicator variable equal to one when the fiscal quarter is the last one of the
year.
We winsorize all continuous variables at the 1st and 99th percentiles. All control variables
are interacted with the surprise in order to control for their impact on the ERCs. Because this
interaction could create some issues with interpretation of the results on the coefficient on the
surprise when the mean of the variables is not equal to zero, we normalize the mean of all
continuous explanatory control variables to zero in order to present meaningful estimates of β3,
the coefficient of interest. We predict that the coefficient β3 will be negative if more information
relevant to central firms is released outside earnings announcement date in comparison to noncentral firms.
Results
We present our results from the estimation of (4) in Table 4. Panel A presents several
13
descriptive statistics that compare central firms with non-central firms. Central firms are larger
than non-central firms as evidenced by a larger logarithm of their market capitalization. They
also have lower Tobinq, lower volatility, and higher leverage than non-central firms. 17.9% of
central firm-quarters have losses, compared to 26.8% for non-central firms. The proportion of
central firm-quarters that have large special items in their income statements as a percentage of
assets is also significantly lower than the proportion of non-central firms. All these differences
are significant at the 1% level or better when using a t-test of differences for the mean or a
Wilcoxon Mann-Whitney test for the median. Overall, these results suggest that central firms are
larger and more mature than non-central firms. This confirms the need to control for these
differences in our specifications on the ERCs. 11
(Insert Table 4 about here)
Column (1) of Table 4, Panel B presents our main results. As predicted, the interaction
coefficient on Surprise × Central loads negatively, indicating that ceteris paribus the ERCs of
central firms are lower than the ERCs of non-central firms. The results are significant at the 1%
level. Given that the coefficient on the surprise is equal to 1.056, central firms have an ERC
coefficient lower by 0.056, or approximately 5% lower than non-central firms. Column (2)
presents similar specifications to Column (1) with standard errors clustered at the quarter level.
Results on the interaction Surprise × Central remain unchanged.
In Column (3), we include a Late Announcer indicator variable for when firms announce
at the end of the announcement period. The variable takes the value one when firms announce
after the median announcement date within the quarter. We interact this variable with the
11
In untabulated results, we confirm that these differences between central firms and non-central firms hold when
incorporating the entire sample of firms from COMPUSTAT quarterly merged with CRSP, and not only the sample
restricted to companies covered by analysts.
14
Surprise and with the Central dummy. The coefficient on Surprise × Central remains negative
and significant, indicating that the ERCs of early announcing central firms are lower than those
of non-central firms. We also find a negative coefficient on the interaction Surprise × Late
Announcer, consistent with information relevant to late announcers being released by the early
announcers. However, we find that the interaction Surprise × Late Announcer × Central is
significantly positive, indicating that the reduction in ERC between early announcers and late
announcers is not as acute for central firms as it is for non-central firms. This result is consistent
with the reduction in ERCs for central firms being primarily due to pre-emption by the release of
macro-economic news and not by company-specific news. The sum Surprise × Late Announcer
+ Surprise × Late Announcer × Central remains negative and significant at 5% using an F-test of
differences, indicating that the ERC of central firms that are late announcers is still lower than
the ERC of central firms that are early announcers. The economic magnitude of the impact of
centrality is magnified for early announcers in comparison to Column (1) and (2). In particular,
the reduction in ERC for central firms represents almost 12% of the ERC of early announcing
firms. In Column (4), we confirm that the results in Column (3) continue to hold when we cluster
the standard errors at the quarter level. The results, except those of the F-tests, remain unchanged.
5
Are central firms’ returns and ROA changes more informative than those
of non-central firms?
In this section, we test our second prediction that the central industries’ performance
leads that of the non-central industries. In particular, we expect that the stock performance and
the accounting performance of central firms are informative about those of non-central firms.
15
5.1 Returns Predictability
Research design
Our analysis of the predictability of returns is similar to that in Hong, Torous and
Valkanov (2007). We test whether monthly return of an industry (hereafter source industry) has
predictive power over the future returns of the industries that it is linked to (hereafter the linked
industries) and whether the effects are enhanced when the source is a central industry. In
particular, we regress the weighted average returns of the linked industries for the month n+1 on
the returns of the source industry for month n. The linked industries are defined as the industries
that are linked to the source industry either as a supplier or a customer or both. We conduct the
following regression:
Linked Abn Returnsi ,n +1 =
α + β1Linked Abn Returnsi ,n + β 2Source Abn Returnsi ,n
+ β3Central + β 4 Central × Source Abn Returnsi ,n + γ X n
(5)
In this model, Linkedt Abn Returns is the weighted average monthly return of the abnormal
returns for each linked industry. Abnormal returns for each industry are computed using the
market value weighted raw returns of the firms comprising the industry, less the value weighted
NYSE/NASDAQ/AMEX market return. We then weight the abnormal returns of each industry
based on the strength of the link between the source industry and the linked industry, because we
anticipate that the predictive power should increase with the strength of the links. Source Abn
Returns is the market capitalization value weighted average monthly return of the source
industry less the value weighted NYSE/NASDAQ/AMEX market return; Central is an indicator
variable that equals one if the centrality of the industry is greater than 0.111 and zero otherwise.
X is a vector of control variables defined as in Hong et al. (2007) and includes the following
variables: Spread, the default spread (DSPR), defined as the difference between the Moody’s
BAA-rated and AAA-rated bond yields; Inflation, the seasonally adjusted monthly inflation,
16
measured as the growth rate of the consumer price index (CPI); Dividend rate, the one year
market dividend yield, computed as the total dividend from the CRSP market portfolio divided
by the current market level.
We predict a positive coefficient on the interaction term, Source Abn Returns × Central,
if the source industry can predict the future returns of the linked industries better when the source
is a central industry. We also predict the coefficient on Source Abn Returns be positive if source
industries can explain the returns of linked industries. A positive coefficient would be consistent
with the findings of Menzly and Ozbas (2010) who find that stocks of economically related
supplier and customer firms predict each other’s returns.
Results
We present our results from the estimation of (5) in Table 5. We start our analysis by
examining the concurrent associations. In particular, in Panel A we regress current returns of the
linked industries on the concurrent returns of the source industry in Panel A. In Column (1), we
regress current period linked industries abnormal returns on the control variables. We find that
higher spreads between AAA and BAA bonds, higher inflation rates, lower standard deviation of
market returns and lower dividend rates are positively associated with concurrent abnormal
returns. In Column (2), we add the returns of the source industries and their interactions with the
central dummy. We find that source returns are positively associated with linked returns. The
association is enhanced when the source is a more central industry, evidenced by a positive
coefficient on the interaction Source Abn Returns × Central. In particular, the coefficient in the
regression goes from 0.098 for non-central sources to 0.213 for central sources, or a 117%
increase. The R2s in the specification go from 0.03 in Column (1) to 0.13 in Column (2),
indicating that the returns of the source and their interaction with the central dummy have strong
17
incremental explanatory power in the regression. Results are unchanged when clustering the
standard errors at the year level in Column (3). Overall, these results are consistent with central
source firms having a stronger association with linked firms than non-central source firms.
(Insert Table 5 about here)
The results of the specifications using one-month-ahead returns for the linked industries
are presented in Panel B. Column (1) presents the results for baseline specifications with
inclusion of the abnormal returns for the linked industries at month n. Consistent with the
literature on momentum (see for example Moskowitz and Grinblatt, 1999), we find that the
returns of the linked industries at month n are positively associated with the returns of the linked
industries at month n+1. We introduce the returns of the source and their interactions with the
central dummy in Column (2). We find a positive coefficient on the returns of the source,
significant at the 1% level, indicating that industries predict the combined returns of their
suppliers and customers. This result is consistent with Menzly and Ozbas (2010). The interaction
Central × Source Abn Returns is positive and significant at the 5% level, indicating that central
source firms’ returns have more predictive power compared to non-central source firms returns.
The total coefficient on Source Abn Returns goes from 0.008 for non-central firms to 0.020 for
central firms, a 120% difference comparable to the increase in coefficient for concurrent returns.
Results in Column (2) are unchanged in Column (3) when standard errors are clustered at the
year level.
5.2 Earnings Predictability
Research design
Our final analysis investigates whether current accounting performance of a source
industry can explain the future accounting performance of the linked industries, and whether the
18
effects are enhanced when the source is a central industry. In particular, we regress the weighted
seasonally differenced return on assets (i.e. current quarter return on assets minus four quarter
lagged return on assets) of the linked industries for the next quarter on the seasonally differenced
return on assets of the source industry for current quarter. We conduct the following regression:
Linked ROA Changei ,n +1 =
α + β1Source ROA Changei ,n + β 2Centrali
+ β3Source ROA Changei ,n × Centrali + β 4 Linked ROA Change n
+ β5 Linked ROA Change n −1 + β 6 Linked ROA Change n − 2
(6)
+ β 7 Linked ROA Change n −3 + β8 Linked ROA Change n− 4
In order to compute Linked ROA Change at quarter n we first compute the seasonally differenced
return on assets (ROA) for each company in each industry. ROA is defined as earnings before
extraordinary items and discontinued operations (ibq in COMPUSTAT) divided by the total
assets at the end of the prior quarter. We then weight these ROA changes by the market
capitalization at the beginning of the quarter of each firm within the industry in order to obtain
industry averages of ROA changes. We then weight these ROA changes by the strength of the
link between the source industry and the linked industries in order to compute a linked industries
average ROA change, or Linked ROA Change. Similar to our analysis on returns, we anticipate
that industries with stronger links predict each other’s accounting performance better than
industries with weaker links and we consequently use the strength of the links between industries
as the weight for our analyses.
Similar to our definition of Linked ROA Change, we define Source ROA Change as the
market capitalization weighted seasonally differenced ROA of the firms comprising the source
industry. We predict that the coefficient on β1 is positive if source industries’ changes in ROA
explain future changes in ROA for linked industries. We also predict that the coefficient on β3 is
positive if the explanatory power is greater for central industries. We include lagged values of
19
Linked ROA Change in order to control for future predictability of within industry earnings
changes, following Bernard and Thomas (1990)’s findings at the firm level. In particular, we
predict that β4, β5 and β6 should be positive while β7 should be negative, due to positive
autocorrelation of earnings for the prior 3 quarters and negative autocorrelation of earnings for
the prior 4th quarter, as in Bernard and Thomas (1990).
Similar to the analysis on ERCs, and for the purpose of consistency in our regressions,
we restrict our sample to firms having fiscal quarters ending in March, June, September and
December.
Results
We present our results from the estimation of (6) in Table 6. Similar to Table 5, the
results in Panel A present specification using concurrent earnings change as the dependent
variable. Column (1) presents specifications using the control variables only. Consistent with
Bernard and Thomas (1990), we find that the first three lags of the Linked Change ROA variable
are positively associated with the current value of this variable and two of the three associations
are significant. Also consistent with Bernard and Thomas (1990) the coefficient on the fourth lag
of Linked Change ROA is negative and significant.
We add Source Change ROA in Column (2), as well as its interaction with the Central
dummy variable. We find a significant positive association between Source and Linked
concurrent year on year changes in ROA, evidenced by a positive coefficient on Source Change
ROA. We also find that this positive association is significantly enhanced when the source
industry is central. In particular, the coefficient nearly triples from 0.039 to 0.114. In Column (3)
we add source fixed effects into the analysis and confirm our findings in Column (2). The
coefficients of interest remain significantly positive and are practically unchanged compared to
20
the specifications in Column (2).
(Insert Table 6 about here)
In Panel B, we present the results of the tests where the dependent variable is one-quarter
ahead value of the Linked Change ROA. Results in Column (1), including control variables only,
are practically similar to those of Column (1) in Panel A. In Column (2), consistent with our
predictions, we find that Source Change ROA at quarter n positively predict Linked Change ROA
at quarter n+1, with a significant improvement of the predictive power when the source is a
central industry, as evidenced by a positive coefficient on the interaction Source Change ROA ×
Central. The predictive coefficient increases from 0.017 to 0.056, comparable to the results in
Panel A. Results in Column (2) remain similar when we add source fixed effects in the
specifications. The coefficient on Source Change ROA at quarter n and its interaction with
Central remain positive and significant at the 1% level and are practically unchanged.
6
Conclusion
We investigate the role of inter-industry trade flows as a source of information transfers
among firms. We hypothesize that an industry’s position in the inter-industry network is an
important determinant of that industry’s exposure to systemic shocks and the extent to which
idiosyncratic shocks to that industry affects the industries it is linked to. In particular, building on
a diversification based argument and prior theoretical findings on the propagation of industry
shocks, we predict and find that firms in central industries are more exposed to systemic risks
compared to firms in non-central industries. Additionally, we find that the earnings response
coefficients are lower for central firms, which is consistent with the investors’ ability to better
anticipate the earnings of central industries based on macroeconomic information.
We also document that shocks to a central industry’s performance, as measured by stock
21
returns and accounting earnings, propagate more strongly than shocks to a non-central industry’s
performance. In particular, we find that a central industry’s performance is more strongly
associated with the concurrent and future performance of the industries it is linked to, than the
performance of a non-central industry is associated with the industries it is linked to.
Our study adds to the literature on the gradual diffusion of information across asset
markets by identifying a specific channel – intra-industry trade flows – through which
information transfers occur. Prior research examines the cross predictability of returns across
economically linked firms, and our findings suggest that such associations are stronger among
non-central firms because these firms have a less diversified set of input/output connections. Our
paper also contributes to the literature on the propagation of idiosyncratic shocks in the economy.
Consistent with Acemoglu et al. (2012)’s prediction that central industries play a key role in
amplification of idiosyncratic shocks into aggregate shocks, our empirical evidence indicates that
the shocks to central industries propagate stronger than those to non-central industries.
22
References
Aboody, D., Aobdia, D., and J. Hughes. 2012. Seasonality in market reactions to quarterly
earnings announcements. Working Paper.
Acemoglu, D., Carvalho, V.M., Ozdaglar, A., and A. Tahbaz-Salehi. 2012. The network origins
of aggregate fluctuations. Econometrica 80(5): 1977-2016.
Ahern, K., and J. Harford. 2010. The importance of industry links in merger waves. Working
Paper.
Anilowski, C., M. Feng and D.J. Skinner, 2007. Does earnings guidance affect market returns?
The nature and information content of aggregate earnings guidance. Journal of Accounting
and Economics 44(1-2): 36-63.
Anjos, F., and C. Fracassi. 2011. Shopping for information? Diversification and the network of
industries. Working Paper.
Ayers, B., and R. Freeman. 1997. Market assessment of industry and firm earnings information.
Journal of Accounting and Economics 24(2): 205-218.
Begley, J., and P. Fischer. 1998. Is there information in an earnings announcement delay?
Review of Accounting Studies 3(4): 347-363.
Bernard, V.L. and J.K. Thomas. 1990. Evidence that stock prices do not fully reflect the
implications of current earnings for future earnings. Journal of Accounting and Economics
13(4):305-340.
Bonacich, P.B. 1972 Factoring and weighing approaches to status scores and clique
identification. Journal of Mathematical Sociology 2(1): 113–120.
Bonsall, S.,Bozanic, Z., and P.E., Fischer. 2012. What do (some) management forecasts convey
about the macroeconomy? Working Paper.
Clinch, G., and N. Sinclair. 1987. Intra-industry information releases: A recursive systems
approach. Journal of Accounting and Economics 9 (1): 89–106.
Cohen, L., and A. Frazzini. 2008. Economic links and predictable returns. Journal of Finance
63(4): 1977–2011.
Dupor, B. 1999. Aggregation and irrelevance in multi–sector models. Journal of Monetary
Economics 43(2): 391–409.
Foster, G. 1980. Externalities and financial reporting. Journal of Finance 35(2): 521-533.
Foster, G. 1981. Intra-industry information transfers associated with earnings releases. Journal of
Accounting and Economics 3(3):201-232.
23
Francis, J. and B. Ke. 2006. Non-audit fee disclosures and the market valuation of earnings
surprises. Review of Accounting Studies 11(4) : 495-523.
Gabaix, X. 2011. The granular origins of aggregate fluctuations. Econometrica 79(3): 733–772.
Hochberg, Y. Ljungqvist,A., and Y. Lu. 2007. Whom you know matters: Venture capital
networks and investment performance. Journal of Finance 62(1): 251–302.
Hong, H., Torous, W., and R. Valkanov. 2007. Do industries lead stock markets? Journal of
Financial Economics 83(2):367-396.
Horvath, M. 1998. Cyclicality and sectoral linkages: Aggregate fluctuations from sectoral
shocks. Review of Economic Dynamics 1(4): 781–808.
Horvath, M. 2000. Sectoral shocks and aggregate fluctuations. Journal of Monetary Economics
45(1): 69–106.
Kelly, P.J. 2005. Information efficiency and firm-specific return variation. Working Paper.
Lim, C., and H. Tan. 2008. Non-audit service fees and audit quality: the impact of auditor
specialization. Journal of Accounting Research 46 (1): 199-246.
Lucas, R. E. 1977. Understanding business cycles. Carnegie-Rochester Conference Series on
Public Policy 5: 7 – 29.
Menzly, L. and O. Ozbas. 2010. Market segmentation and cross-predictability of returns. Journal
of Finance 65(4):1555-1580.
Morck, R., Yeung, B., and W. Yu. 2000. The information content of stock markets: Why do
emerging markets have synchronous stock price movements? . Journal of Financial
Economics 58 (1-2):215-260.
Moskowitz, T. J., and M. Grinblatt. 1999. Do industries explain momentum? Journal of Finance
54(4): 1249–1290.
Newman, M. 2010. Networks: An introduction. Oxford University Press, Oxford, UK.
Pandit, S., C. Wasley, and T. Zach. 2011. Information externalities along the supply chain: The
economic determinants of suppliers’ stock price reaction to their customers’ earnings
announcements. Contemporary Accounting Research 28 (4): 1304-1343.
Piotroski, J.D., and D. T., Roulstone. 2004. The influence of analysts, institutional investors, and
insiders on the incorporation of market, industry and firm-specific information into stock
prices. The Accounting Review 79 (4): 119-1151.
24
Appendix A. Variable Descriptions
Variable
Analysts dummy
Average proportion
traded
CAR
Central
Dividend rate
Eigenvector
Centrality
Fourthquarter
Inflation
Late Announcer
Leverage
Logmkval
Loss
Source Abn
Returns
Source ROA
Change
Specialitems
Spread
StdReturns
Surprise
Linked Abn
Returns
Linked ROA
Change
Tobinq
Description
An indicator variable that equals one if there is a quarterly consensus EPS forecast
provided in I/B/E/S/ for the last quarter of the analysis period and zero otherwise.
Average number of shares traded as a percentage of shares outstanding. The variable
is calculated using daily (monthly) data from CRSP for daily (monthly) analysis.
Cumulative returns for a given firm's stock for the three days around the earnings
announcement date, less the value weighted NYSE/NASDAQ/AMEX market return.
An indicator variable that equals one if the centrality of the industry is greater than
0.111 and zero otherwise.
Total dividend from the CRSP market portfolio divided by the current market level.
The value of eigenvector centrality. The calculation is described in detail in Section
2.2. The measure is computed based on the 1997 BEA "Use Tables".
An indicator variable that equals one for the fourth fiscal quarter and zero otherwise.
Monthly change in the seasonally adjusted consumer price index
An indicator variable that equals one when the earnings announcement date occurs
after the median earnings announcement date during the quarter
Total debt (short term plus long term) divided by itself plus book value of equity
Logarithm of the market capitalization of the company at the end of the period (shares
outstanding x Price at the end of the fiscal quarter).
An indicator variable that equals one if the actual EPS as reported in I/B/E/S/ is
negative and zero otherwise.
Market capitalization weighted average monthly returns of the firms in the source
industry
Market capitalization weighted average of the change in ROA between quarter n and
quarter n-4 of firms in the industry. ROA is defined as earnings before extraordinary
items and discontinued operations divided by assets at the beginning of the quarter.
An indicator variable that equals one if the proportion of special items during the
quarter exceeds 5% of the value of total assets.
The difference between the yield of BAA rated and AAA rated bonds
Standard deviation of stock returns estimated over a 90 days period ending three days
prior to the earnings announcement date
Actual EPS less the most recent analyst consensus forecast, deflated by the stock price
at the beginning of the fiscal quarter. Actual and consensus are from I/B/E/S/. The
consensus is measured at least two days prior to the earnings announcement date
Weighted average of the monthly industry returns less the value weighted market
return. The weights are the strength of the link between the source and the linked
industries and the monthly industry returns are weighted using market capitalization at
the beginning of the month
Weighted average of the industry year on year changes in ROA. Weights used are the
strength of the link between source and linked firms. Industry level changes in ROA
are computed as the market capitalization weighted average of the firm level change
in ROA between quarter n and quarter n-4
Book value of total liabilities plus market value of equity at the end of the fiscal
quarter divided by the book value of total assets
25
Table 1: List of the most central industries
Rank
Name of the Industry
1
Wholesale Trade
2
Construction
3
Management of Companies/Enterprises
4
Real Estate
5
Retail Trade
6
Administrative and Support Svc.
7
Money Authorities/Credit Intermediation
8
Power Generation and Supply
9
Advertising and Related Svc.
10
Plastics/Rubber Products Mfg.
11
Motor Vehicle Body/Trailer/Parts Mfg.
12
Other Fabricated Metal Product Mfg.
13
Petroleum and Coal Products Mfg.
14
Food Mfg.
15
Oil and Gas Extraction
16
Employment Svc.
17
Semiconductor and Electronic Eq. Mfg.
18
Iron and Steel Mills
19
Management and Tech. Consulting Svc.
20
Architectural and Engineering Svc.
26
Table 2: Distribution of the centrality variable
This table presents the distribution of the eigenvector centrality measure for the industries in the U.S. The measure is
calculated using the detailed use tables of 1997 published by the Bureau of Economic Analysis (BEA). Industries
are defined using the BEA’s four-digit industry codes and the total number of industries is 123.
Variable
Eigenvector Centrality
Minimum
5th Percentile
10th Percentile
25th Percentile
50th Percentile
75th Percentile
90th Percentile
95th Percentile
Maximum
0.021
0.036
0.045
0.054
0.074
0.097
0.127
0.143
0.266
Mean
StDev
N
0.081
0.040
123
27
Table 3: Analysis of the R2s from CAPM and Fama-French three factor models
This table presents the analysis of the association between centrality and exposure to systemic risk using an ordinary least squares model. The dependent variable
is the R2 from CAPM or Fama French three factor models estimated using daily or monthly data. Estimations are conducted using five-year non-overlapping
windows, ending in 1985, 1990, 1995, 2000, 2005 and 2010. The variable of interest is the centrality measure, Eigenvector Centrality. The definitions of the
variables are available in Appendix A. Standard errors are clustered at the firm level. *, **, and *** denote significance at a two sided 10%, 5% and 1% level.
Daily R-squares
Monthly R-squares
Dependent Variable:
R-square
Eigenvector
Centrality
CAPM
0.116
Fama French
***
9.643
Analysts Dummy
0.017
0.004
***
28
0.027
***
Number
Observations
Clustering
Adjusted R-square
F-statistic
(0.087)
***
0.004
***
0.029
***
(0.082)
0.033
0.008
***
0.021
***
0.013
0.043
***
0.010
***
0.020
***
0.085
4.964
28.630
20,571
20,571
17,375
17,375
Firm level
Firm level
Firm level
Firm level
0.45
0.46
0.21
0.21
3,287.05
***
1,036.15
***
***
34.312
(41.775)
***
***
9.122
(47.466)
2,970.35
***
15.815
37.768
***
0.053
2.716
8.167
65.401
***
***
13.137
26.362
64.605
Constant
0.025
0.073
Fama French
4.164
14.298
26.313
Logmkval
***
9.494
10.696
Average Proportion Traded
0.123
CAPM
***
990.29
***
***
Table 4: Earnings response coefficients
Table 4 Panel A presents descriptive statistics for the sample of firms covered by analysts. There are 57,292
observations for central firms and 141,576 observations for non-central firms. Panel B presents the analysis of the
association between earnings response coefficients and centrality. In Panel B, cumulative abnormal return for the
three day announcement period [-1, 1] is regressed on the earnings surprise, an indicator variable for central firms
and the interaction of the central dummy with the earnings surprise. Control variables and their interactions with the
earnings surprise are also included in the model. For brevity the control variables and their interactions with the
earnings surprise variable are not reported, but they are available from the authors upon request. The definitions of
the variables are available in Appendix A. *, **, and *** denote significance at a two sided 10%, 5% and 1% level.
Panel A: Descriptive Statistics
Mean
Variable
Non Central
CAR
Surprise
Tobinq
StdReturns
Leverage
Logmkval
Loss
Specialitems
Fourthquarter
0.0006
(0.0033)
2.3845
0.0348
0.2868
6.0015
0.2678
0.0350
0.2697
Median
Central
Test of equality
Non Central
Central
0.0002
0.0002
1.7029
0.0299
0.2245
5.8589
-
0.0006
0.0001
1.4305
0.0272
0.3221
6.0881
-
0.0015
(0.0036)
1.8489
0.0319
0.3403
6.1483
0.1786
0.0246
0.2675
Mean
(2.29)
1.89
49.96
31.09
(33.62)
(16.84)
42.27
11.99
1.00
Median
**
*
***
***
***
***
***
***
(1.73)
1.47
64.28
31.40
(48.79)
(18.75)
42.09
11.98
1.00
*
***
***
***
***
***
***
Panel B: Earnings Response Coefficients (ERCs)
Dependent variable: CARs
Surprise
Central
S urprise x Central
(1)
(2)
1.056 ***
52.029
(0.001) **
(2.044)
(0.056) ***
(3.553)
(3)
1.056 ***
15.261
(0.001)
(1.577)
(0.056) ***
(2.908)
Late Announcer
Late Announcer x Central
S urprise x Late Announcer
S urprise x Late Announcer x Central
Constant
Number Observations
Clustering
Adjusted R-square
F-statistic
0.006
19.701
***
0.006
6.677
198,868
None
0.03
337.53 ***
***
198,868
Quarter
0.03
61.98 ***
1.126
51.566
(0.403)
(0.134)
(5.084)
(0.004)
(7.422)
(0.002)
(1.677)
(0.182)
(10.072)
0.122
3.742
0.006
20.546
(4)
***
***
***
*
***
***
***
198,868
None
0.03
282.09 ***
F-test: S urprise x Late Announcer Dummy + S urprise x Late Announcer Dummy x Central Dummy = 0
F-test
4.52
p-value
0.03
29
1.126
15.654
(0.303)
(0.134)
(3.258)
(0.004)
(3.338)
(0.002)
(1.565)
(0.182)
(6.991)
0.122
2.280
0.006
6.741
***
***
***
***
**
***
198,868
Quarter
0.03
56.62 ***
1.62
0.21
Table 5: Returns Predictability
This table presents the analysis of the association between centrality and the predictability of stock returns at the
industry level. The dependent variable is equal to the weighted average monthly returns of the linked industries of a
given industry i (i.e., the source industry) less the market returns. To calculate the weighted average monthly returns
of the linked industries we first calculate value weighted returns of all firms in each linked industry then weight
these returns with the strength of the link of between the industry and the source industry. Panel A presents results
for the concurrent association between the returns of the linked industries and the source and Panel B presents the
results where the dependent variable is one month ahead return of the linked industries. The definitions of the
variables are available in Appendix A. *, **, and *** denote significance at a two sided 10%, 5% and 1% level.
Panel A: Concurrent returns
Dependent Variable:
Concurrent returns linked
(1)
(2)
Returns source
0.098
53.547
1.318
0.115
19.975
0.009
23.878
0.088
2.387
(0.289)
(11.709)
(0.357)
(23.171)
(0.063)
Central dummy
Central x Returns Source
Spread
Inflation
Stdev
Dividend Rate
Constant
Number of observations
Clustering
Adjusted R-square
F-statistic
0.011
27.786
0.084
2.139
(0.419)
(16.090)
(0.432)
(26.604)
0.001
2.235
***
**
***
***
**
36,334
None
0.03
238.93 ***
(3)
***
***
***
**
***
***
36,334
None
0.13
776.47 ***
F-test: Returns source + Central x Returns S ource = 0
F-statistic
p-value
1,514.07
-
30
0.098
9.409
1.547
0.115
15.017
0.009
2.954
0.088
0.375
(0.289)
(1.175)
(0.357)
(2.333)
(0.006)
***
***
***
**
36,334
Year Level
0.13
93.11 ***
198.58
-
Table 5: Returns Predictability (continued)
Panel B: One month ahead returns
Dependent Variable:
Returns linked month n+1
Returns linked
(1)
0.069
13.255
(2)
***
0.060
10.906
0.008
3.844
0.850
0.012
1.976
0.008
20.303
0.171
4.366
(0.124)
(4.758)
(0.357)
(21.764)
(0.001)
(2.042)
Returns source
Central dummy
Central x Returns Source
Spread
Inflation
Stdev
Dividend Rate
Constant
Number of observations
Clustering
Adjusted R-square
F-statistic
0.008
20.489
0.169
4.337
(0.131)
(5.035)
(0.359)
(21.916)
(0.001)
(1.795)
***
***
***
***
*
36,334
None
0.03
198.53 ***
(3)
***
***
**
***
***
***
***
**
36,334
None
0.03
127.32 ***
F-test: Returns source + Central x Returns S ource = 0
F-statistic
p-value
11.26
0.00
31
0.060
1.753
0.008
3.341
1.268
0.012
2.193
0.008
2.620
0.171
0.736
(0.124)
(0.515)
(0.357)
(2.384)
(0.001)
(0.222)
*
***
**
**
**
36,334
Year Level
0.03
6.28 ***
10.08
0.00
Table 6: Earnings predictability
This table presents the analysis of the association between centrality and the predictability of accounting
performance at the industry level. The dependent variable is equal to the weighted average seasonally differenced
quarterly return on assets of the linked industries of a given industry i (i.e., the source industry)
The seasonally differenced quarter ROA of the linked industries is computed as the weighted average of each
industry’s average ROA change comprising the linked industries. The weights are based on the distance between the
source and the linked industry. Each industry’s average ROA is computed as the market value weighted average of
the ROA changes of the firms comprising the industry. Panel A presents results for the concurrent association
between the ROA of the linked industries and the source and Panel B presents the results where the dependent
variable is one quarter ahead ROA of the linked industries. The definitions of the variables are available in
Appendix A. *, **, and *** denote significance at a two sided 10%, 5% and 1% level.
Panel A: Concurrent ROAs
Dependent variable:
Linked Change ROA n
(1)
(2)
0.395
43.420
0.003
0.326
0.117
11.913
(0.227)
(24.812)
(0.001)
(25.843)
0.039
11.944
(0.061)
0.075
7.380
0.375
41.376
(0.002)
(0.183)
0.115
11.868
(0.221)
(24.431)
(0.001)
(23.241)
Source Change ROA
Central
Source Change ROA x Central
Linked Change ROA n-1
Linked Change ROA n-2
Linked Change ROA n-3
Linked Change ROA n-4
Constant
Number Observations
Fixed Effects
Adjusted R-square
F-statistic
11,528
None
0.20
715.16
***
***
***
***
11,528
None
0.22
458.10
***
F-test: S ource Change ROA + Central x S ource Change ROA = 0
F-statistic
p-value
32
147.19
-
(3)
***
0.039
11.954
***
***
0.076
7.409
0.370
40.683
(0.005)
(0.488)
0.112
11.536
(0.225)
(24.807)
(0.001)
(25.405)
***
***
***
***
***
***
***
***
***
***
11,528
Source Fixed Effects
0.22
521.09 ***
147.22
-
Table 6: Earnings predictability (continued)
Panel B: One quarter ahead ROAs
Dependent variable:
Linked Change ROA n+1
(1)
(2)
0.390
41.396
0.002
0.247
0.121
12.250
(0.234)
(23.459)
0.015
1.571
(0.001)
(23.337)
0.017
5.118
(0.406)
0.039
3.726
0.379
39.846
(0.002)
(0.243)
0.119
12.044
(0.233)
(23.468)
0.015
1.599
(0.001)
(21.514)
Source Change ROA
Central
Source Change ROA x Central
Linked Change ROA n
Linked Change ROA n-1
Linked Change ROA n-2
Linked Change ROA n-3
Linked Change ROA n-4
Constant
Number Observations
Fixed Effects
Adjusted R-square
F-statistic
11,528
None
0.19
546.57
***
***
***
***
11,528
None
0.20
350.34
***
F-test: S ource Change ROA + Central x S ource Change ROA = 0
F-statistic
p-value
33
31.32
-
(3)
***
0.017
5.149
***
***
0.040
3.759
0.373
39.016
(0.005)
(0.477)
0.116
11.716
(0.235)
(23.625)
0.010
1.009
(0.001)
(23.692)
***
***
***
***
***
***
***
***
***
***
11,528
Source Fixed Effects
0.19
389.54 ***
31.51
-
Figure 1: An economy where one central industry trades with N-1 other industries
N
1
2
N-1
3
4
N-2
34
Figure 2: Inter-industry network based on BEA input/output tables
This figure plots the inter-industry network based on the U.S. input-output matrix of 1997 provided by the Bureau of
Economic Analysis (BEA).Each node corresponds to an industry (based on BEA’s 4-digit industry code) and each
edge corresponds to a link between industries with strength (Aij) of 3% or more. See section 2.2 for the definition of
strength of the link between two industries.
2212
3160
3150
1120
3253
3140
4840
1130
3130
1150
8120 813A
6240
5240
6230 7130
532A 6100
2110
4860
3221
3252
1110
6220
4930
1140
2213
3240
3251
3270
2123
2300
3210
3260
4810
4820
2122
48A0
4830
3324
3359
3121
3339
2121
332B
5230
52A0
5411
5412
5414
5133
4920
3110
3333
2130
3346
3230
3259
5419
5615
336A
5132
3322
3335
3321
5120 71A0
3222
331A
3364
3353
5141
3399 3341
3344
5415
3332
3315
5111
3336
3361
8111
3345
334A
35
3122
5142
811A
331B
5321
3254
5250
5500
5418
3334 4200
3331
336B
5416
4A00
561A
5613
7220
3351
3323
5330
5620
2211
5310
5324
6210
813B
5413
32553370
3256
3352
3391
5112
5131
Figure 3: Distribution of eigenvector centrality measure
0
.05
Eigenvector Centrality
.1
.15
.2
.25
This figure plots the distribution of the eigenvector centrality measure. Eigenvector centrality (defined formally in
section 2.2) of an industry is as function of that industry’s direct or indirect ties to the other industries in the
economy as well as the strength of these ties. Larger values indicate more central industries. The dotted reference
line in the figure corresponds to 0.111, the cutoff value we use to differentiate central industries from the non-central
industries.
0
20
40
80
60
Centrality Rank
36
100
120