Corporate Governance and Equity Prices: Do Industry Adjustments Explain Results?
Stefan Lewellen Yale University
February 17, 2012∗
Abstract No. Using a wide range of common industry classification standards, I find little evidence that unexpected industry performance explains the return on governance-sorted portfolios during the 1990s. I also find little evidence that Gompers, Ishii and Metrick (2003)’s tests are “misspecified.” In contrast, the majority of tests with the strongest size and power properties in my sample yield positive industry-adjusted abnormal returns on governance-sorted portfolios during the 1990s. My results have important implications for future governance research and highlight the inherent tradeoff between industry coarseness and statistical power in calendar-time tests.
JEL: C1, C52, G11, G12, G14, G34 ∗
Comments are welcome at
[email protected]. I thank Andrew Metrick, Bill Cruise, Martijn Cremers, Quinn Curtis, Xavier Giroud, Shane Johnson, Michael Lemmon, Steve Malliaris, Holger Mueller, Jacob Thomas, Ed Vytlacil, Jackie Yen, and seminar participants at the 2010 WFA Meetings, the 2011 AFE Meetings, and Yale University for helpful comments.
1
Introduction Is unexpected industry performance responsible for the positive relationship between shareholder rights and stock returns during the 1990s? In an influential paper, Gompers, Ishii and Metrick (2003) find that firms with strong shareholder rights (“Democracies”) earned higher abnormal stock returns than firms with weak shareholder rights (“Dictatorships”) during the 1990s, even after controlling for industry membership. However, a recent paper in this journal by Johnson, Moorman and Sorescu (2009) finds that the seminal results reported by Gompers, Ishii and Metrick (2003) (hereafter GIM) are “an artifact of asset pricing model misspecification or unexpected industry performance,” and as a result, “governance quality has no reliable impact on long-term abnormal stock returns.” Given the extensive amount of recent research linking governance to equity prices and firm value, the findings documented by Johnson, Moorman and Sorescu (2009) came as a surprise to many researchers within the field of corporate governance.1 This paper attempts to resolve the debate over the relationship between industries and governance during the 1990s by expanding the set of industry classification standards used to evaluate industry-adjusted performance on GIM’s Democracies-minus-Dictatorships portfolio (hereafter the GIM governance portfolio). I begin by noting that GIM only consider a single industry classification standard in their tests, while Johnson, Moorman and Sorescu (2009) (hereafter JMS) only consider two such classification standards in their response. After reproducing the key tables in both papers using a more representative selection of industry classification standards, I find that the relationship between governance and equity prices during the 1990s is generally robust to concerns about unexpected industry performance. I reach a similar conclusion after reexamining my findings using a variety of empirical 1
Recent papers linking governance to equity prices include Cremers and Nair (2005), Core, Guay and Rusticus (2006), Masulis, Wang and Xie (2007), Bebchuk, Cohen and Ferrell (2009), Cremers, Nair and John (2009), Gompers, Ishii and Metrick (2010), Cremers and Ferrell (2011), Giroud and Mueller (2011), Gu and Hackbarth (2011), Kadyrzhanova and Rhodes-Kropf (2011), and Bebchuk, Cohen and Wang (forthcoming), among others.
2
testing methodologies, many of which are new to the literature. I also document a more general tradeoff between industry coarseness and statistical power that may be of independent interest to empirical asset pricing researchers. Taken together, my findings have important implications for future corporate governance research and shed light on the tradeoff between broad versus narrow industry definitions in the context of calendar-time asset pricing tests. Like GIM and JMS, I share the goal of measuring industry-adjusted abnormal returns on GIM’s governance portfolio in a manner that is both economically and econometrically sound. However, GIM’s Democracies and Dictatorships tend to cluster within specific industries, perhaps because of the endogenous relationship between industry membership and corporate governance. To alleviate the possibility of spurious results driven by industry membership, GIM construct industry-adjusted abnormal returns on their governance portfolio using Fama and French (1997)’s 48-industry classification scheme (hereafter FF48) and find that the relationship between governance and stock prices is not an artifact of industry clustering. However, JMS argue that the asset pricing test employed by GIM is “misspecified,” in the sense that a long/short portfolio of random firms with the same industry weights as GIM’s Democracy and Dictatorship portfolios earns a positive industry-adjusted abnormal return. To resolve the specification problems with GIM’s test, JMS identify an industry-adjusted asset pricing test involving narrowly-defined three-digit Standard Industrial Classification (SIC) industry definitions that is “well-specified” in their empirical simulations. Using this well-specified test, JMS find that long-term abnormal returns on the GIM governance portfolio are statistically zero, raising serious questions about the robustness of GIM’s original findings. Despite the gravity of their conclusions, however, JMS make no attempt to show that three-digit SIC classifications are the only industry classifications that produce well-specified tests. To mitigate this important oversight, I take the “kitchen sink” approach and examine the size and power properties of industry-adjusted asset pricing tests on the GIM gover-
3
nance portfolio using a variety of commonly-used industry classification standards. After applying the same size and power tests employed by JMS to my broader sample of industry classifications, I find that a total of 96 industry-adjusted asset pricing tests involving a wide variety of industry classifications are well-specified according to JMS’s methodology. Furthermore, the majority of the industry-adjusted asset pricing tests with the strongest size and power properties in my sample also yield statistically and economically significant industry-adjusted abnormal returns on the GIM governance portfolio. As such, I conclude that while industry adjustments do reduce the abnormal returns on the GIM governance portfolio during the 1990s, they do not reduce abnormal returns to zero. To gain additional comfort in my results, I next extend the asset pricing “precision” test developed by Metrick (1999) to incorporate industry-adjusted returns and apply this test to the various asset pricing specifications in my sample. While my precision tests do not attempt to measure size, they provide an independent assessment of statistical power. Consistent with my previous findings, I find that the most “precise” industry-adjusted asset pricing tests also yield positive and statistically significant abnormal returns on the GIM governance portfolio. Additional robustness tests validate this conclusion. Thus, while I cannot completely rule out the possibility that industry clustering is responsible for GIM’s results, I find little evidence that the relationship between governance and equity prices during the 1990s is a result of model misspecification or conditional industry return premia. Many of the industry-adjusted asset pricing tests with the best empirical properties in my sample utilize the Global Industry Classification Standard (GICS), an industry classification standard that was developed specifically for the finance profession by MSCI Barra and Standard & Poors. Interestingly, the GICS classifications considered in this paper lie in the middle of the industry granularity spectrum between more broadly-defined classifications (FF48 classifications) and more narrowly-defined classifications (three- and four-digit SIC classifications). While GICS classifications produce tests with the best size and power prop-
4
erties in my sample, many asset pricing specifications involving FF48 industry classifications also possess excellent empirical properties. In contrast, asset pricing specifications that are constructed using narrowly-defined three-digit and four-digit SIC classifications yield a much smaller number of tests with strong empirical properties. To make sense of this result, I highlight a fundamental tradeoff that exists between industry coarseness and statistical power in the context of calendar-time asset pricing tests. All else equal, smaller industries should do a better job than larger industries of grouping together firms that share similar exposures to economic fundamentals. However, as the number of firms in each industry becomes smaller, industry returns are more likely to be clouded by an idiosyncratic return component that has nothing to do with industry performance. For example, the return on an industry with only five firms is almost certain to contain idiosyncratic noise, while the return on an industry with 50 firms should contain far less idiosyncratic noise. While this idiosyncratic return component should not affect point estimates, it will reduce the statistical power of industry-adjusted abnormal return tests, making it more difficult to detect abnormal returns. Consistent with this tradeoff, I find that industries of “average” coarseness produce asset pricing tests with stronger statistical power than tests constructed using broadly-defined or narrowly-defined industries. As such, I conclude that narrowly-defined industries are not always preferable to broader industries, even if such industries do a better job of grouping together similar types of firms. Other aspects of the relationship between industries, governance, and firm value have been previously documented in the literature. Bebchuk, Cohen and Wang (forthcoming) find that industry-adjusted returns on GIM’s governance portfolio were statistically and economically significant during the period from 1990 to 2001, but were statistically zero after 2001. In contrast to Core, Guay and Rusticus (2006), Bebchuk, Cohen and Wang (forthcoming) also find that analysts and the market were surprised by the outperformance of Democracies during the 1990s, consistent with Democracies earning positive abnormal
5
returns over GIM’s sample period. In a separate study, Giroud and Mueller (2011) find that corporate governance provisions influence operating performance and long-term abnormal stock returns more acutely among firms residing in highly concentrated industries. They also document the small size of three-digit SIC industries within the context of the GIM governance sample. Furthermore, in a study examining thirty years of corporate governance data, Cremers and Ferrell (2011) run Tobin’s Q regressions and find that shareholder rights were positively correlated with firm value dating back to 1978, even after adjusting for the effects of industry. While all of these papers document a positive industry-adjusted link between governance and firm value, none of them specifically examines the sensitivity of GIM’s results to a variety of different industry classification standards. As such, the contributions of this paper represent a new addition to this important and growing literature.
1 1.1
Data Description Governance and Returns Data
I source governance data from GIM’s dataset of corporate governance provisions, which can be downloaded from Andrew Metrick’s website (http://www.som.yale.edu/faculty/am859) or through Wharton Research Data Services. GIM construct an index of shareholder rights (the “G-index”) based on corporate governance data published by RiskMetrics (formerly the Investor Responsibility Research Center, or IRRC) from 1990 to 1998. A firm’s G-index value starts at zero and is incremented by one unit for each governance provision that restricts shareholder rights. Following GIM, I refer to firms with a G-index value of five or less (out of 24) as “Democracies” and firms with a G-index of 14 or greater as “Dictatorships.” Portfolios of Democracies and Dictatorships are formed at the beginning of September 1990, July 1993, July 1995, and February 1998, corresponding to the dates on which the RiskMetrics data was first published, and G-index scores are rolled forward until the next publication date. Firms
6
with dual-class share structures are excluded from the sample. Monthly stock returns and market capitalizations are sourced from the Center for Research in Security Prices (CRSP) from September 1990 to December 1999. Where applicable, delisting returns are taken from CRSP. Market capitalizations are also lagged by one month.
1.2
Industry Classification Standards
I report test results using three different industry classification schemes: the Standard Industrial Classification (SIC) system, the 48-industry classification scheme designed by Fama and French (1997) (FF48), and the Global Industry Classification Standard (GICS).2 Since GIM obtained SIC codes from CRSP and JMS obtained SIC codes from Compustat, I obtain SIC classifications from both CRSP and Compustat.3 GICS classifications are obtained from Compustat. I also use the latest FF48 classification criteria from Kenneth French’s website. The SIC and GICS classification systems contain a number of hierarchical levels. To ensure broad coverage across a variety of industry granularity levels, I focus on the narrowest industry definitions within each hierarchy (four-digit SIC and eight-digit GICS classifications) and the second-narrowest industry definitions within the SIC and GICS hierarchies (three-digit SIC and six-digit GICS classifications). The FF48 system contains a single classification level that is defined more broadly than the other types of classifications. More information on the SIC, FF48, and GICS classification standards can be found in Bhojraj, Lee and Oler (2003) (hereafter BLO). 2
I also examined tests involving the North American Industry Classification Standard (NAICS), which is the successor to the SIC system. However, given the significant overlap between the two methodologies, I found that results using NAICS were virtually identical to results using SIC codes. Thus, for brevity, I chose not to report results using NAICS. 3 Compustat and CRSP SIC assignments differ approximately 38% of the time in my sample.
7
1.2.1
Previous Empirical Research on Industry Classification Standards
Most of the early empirical work within the industry classification literature focused solely on the SIC classification system.4 However, recent research has focused more heavily on comparing the performance of the SIC, FF48, and GICS classifications standards across various applications. BLO compare the SIC, NAICS, FF48, and GICS classification standards and find that “GICS classifications are significantly better at explaining stock return comovements,” “the GICS advantage is consistent from year to year and is most pronounced among large firms,” and “[t]he other three methods differ little from each other in most applications.” Chan, Lakonishok and Swaminathan (2007) compare GICS and FF48 classifications and find that six-digit and eight-digit GICS classifications “deliver similar levels of improvement over FF [the FF48 classification standard] on the basis of higher correlations across industry peer firms, and larger differentiation between the inside- and outside-industry correlations.” Finally, Weiner (2005) studies the performance of SIC, NAICS, FF48, and GICS in nearly a dozen different applications. He concludes that “GICS leads to lower valuation errors than the other systems. These results are robust over time.” Unfortunately, despite a substantial volume of research highlighting the strength of the GICS classification system, the overwhelming majority of researchers continue to construct industry-adjusted asset pricing tests using the SIC classification standard.5
2
Empirical Methodology and Results
2.1
Industry-adjusted Abnormal Return Tests
In this section, I augment the findings of GIM and JMS by evaluating industry-adjusted abnormal returns on the GIM portfolio using the wider sample of industry classification 4
See, e.g., Clarke (1989), Guenther and Rosman (1994), and Kahle and Walking (1996). Of the research papers in major finance, economics, and accounting journals that contain tests based on one or more industry classification systems, BLO and Weiner (2005) find that more than 90% of papers use the SIC classification system. 5
8
standards described in section 1.2. To span the range of commonly-used industry classification standards and industry granularities, I compute industry-adjusted returns on each Democracy and Dictatorship firm using a total of eight different industry classification standards: FF48 classifications formed from CRSP SIC codes, FF48 classifications formed from Compustat SIC codes, three- and four-digit SIC classifications from both CRSP and Compustat, and six- and eight-digit GICS classifications sourced from Compustat.
2.1.1
Methodology
Like GIM and JMS, I compute the value-weighted, industry-adjusted return on firm i from time t − 1 to time t as: adj Ri,t
= Ri,t −
J X
wj,t−1 Rj,t ,
(1)
j=1
where Ri,t is the total return on firm i from time t − 1 to t, the J firms used to compute the industry return share firm i’s industry classification (and may have other restrictions), and wj,t−1 is the market capitalization weight of firm j within the industry portfolio at the end of the previous period. Industry returns are assigned using the methodology employed by JMS, who stratify the GIM sample into “non-Democracies” and “non-Dictatorships.” The non-Democracy sample contains all firms in the GIM sample which are not part of GIM’s Democracy portfolio, and the non-Dictatorship sample is defined analogously. Thus, industry returns are assigned to firms in GIM’s Democracy and Dictatorship portfolios using the non-Democracy and nonDictatorship samples, respectively. For example, the industry return that is matched with a Democracy firm in industry J is computed as the value-weighted return on all non-Democracy firms in industry J. Following JMS, I assign firms to SIC classifications in this section using the DN U M variable from the legacy version of Compustat. The DN U M variable assigns firms to industries based on their current line(s) of business. As such, firms do not change DN U M 9
classifications over time. Likewise, I use the HSICCD variable in CRSP, which contains firms’ current CRSP SIC classifications. Since FF48 industries are formed from SIC classifications, the FF48 classifications in this section are also assigned based on firms’ current industries, and do not change over time. To maintain consistency, I also assign firms to GICS classifications in this section using the GIN D and GSU BIN D variables from Compustat, which contain firms’ current industry classifications. After computing industry-adjusted returns at the firm level, I next compute valueweighted, industry-adjusted returns on the Democracy and Dictatorship portfolios. Monthly industry-adjusted returns on these portfolios are computed as:
adj Rportf olio,t
=
I X
adj wi,t−1 Ri,t , portf olio ∈ {Democracies, Dictatorships},
(2)
i=1
adj where Ri,t is given by equation (1). Finally, a zero-cost portfolio that is long industry-
adjusted Democracies and short industry-adjusted Dictatorships is constructed, and returns adj on this portfolio (denoted RGIM,t ) are regressed against a set of asset pricing factors using
OLS. Like JMS, I consider three different asset pricing models in my tests. The first model, which I refer to to as the FF3 model, contains three factors – the excess return on the market portfolio (denoted M KT RFt ), a factor capturing the difference in returns between small stocks and large stocks (denoted SM Bt ), and a factor capturing the return difference between firms with high and low ratios of book value to market value (denoted HM Lt ). Additional information on the FF3 model can be found in Fama and French (1993). I also augment the FF3 model with a momentum factor that captures the return on a portfolio that buys previous winners and sells previous losers. The second model I consider, which is referred to as the PR1YR model, uses the momentum factor introduced by Carhart (1997) (denoted P R1Y Rt ). In contrast, the third model, nicknamed the UMD model, uses
10
the momentum factor constructed by Kenneth French (denoted U M Dt ).6 Thus, the regression equations for the three asset pricing models considered in this paper are:
adj F F 3 M odel : RGIM,t = α + β1 M KT RFt + β2 SM Bt + β3 HM Lt + εt ,
(3)
adj P R1Y R M odel : RGIM,t = α + β1 M KT RFt + β2 SM Bt + β3 HM Lt + β4 P R1Y Rt + εt , adj U M D M odel : RGIM,t = α + β1 M KT RFt + β2 SM Bt + β3 HM Lt + β4 U M Dt + εt .
2.1.2
Results
Table 1 contains the abnormal return estimates from these regressions. Since the regressions span eight different industry classification standards and three different asset pricing models, the table contains a total of 24 industry-adjusted abnormal return estimates. The industry-adjusted abnormal return estimates reported in Table 1 are highly significant from an economic standpoint: the mean and median abnormal return estimates in the table are 0.43% and 0.41% per month, respectively, or more than 5% per year. Furthermore, the average point estimate of 0.43% per month is virtually identical to the industry-adjusted abnormal return estimate of 0.47% per month originally reported by GIM using FF48 industry classifications constructed from CRSP SIC codes.7 While the vast majority of the point estimates in Table 1 are economically large, however, many of these point estimates are statistically insignificant (the average t-statistic is approximately 1.5). In addition, the range of industry-adjusted abnormal return estimates is extremely wide, ranging from a low of 0.24% per month to a high of 0.77% per month. Some tests are statistically significant. Table 1 shows that industry-adjusted abnormal 6
I thank Kenneth French for providing the regression factors and the FF48 industry definitions. These data can be downloaded at: http://mba.tuck.dartmouth.edu/pages/faculty/ken.french. The P R1Y R factor is constructed by hand using the methodology outlined in Carhart (1997). 7 In untabulated results, I exactly replicated GIM’s reported industry-adjusted abnormal return estimate of 0.47% per month, statistically significant at the 5% level. While there is a common perception that GIM used median industry returns in their asset pricing tests, GIM in fact used value-weighted industry returns in all of their calendar-time portfolio tests. I thank Andrew Metrick for clarifying this point.
11
tests involving six-digit GICS classifications and four-digit SIC classifications sourced from CRSP produce point estimates that are economically large (0.49% to 0.77%) and statistically significant. In contrast, I find that the industry-adjusted abnormal return test favored by JMS – a combination of three-digit SIC classifications from Compustat and the UMD asset pricing model – produces a statistically insignificant abnormal return estimate of only 0.26% per month, which is identical to the point estimate listed in their paper.
2.1.3
Discussion
Viewed in isolation, Table 1 suggests that the relationship between governance and equity prices may not be robust to industry effects. After all, despite the economic significance of the results in Table 1, only six of the 24 asset pricing specifications produce an industryadjusted abnormal return estimate that is statistically different from zero. However, three important observations are worth noting. First, the whole point of JMS’s paper is that we should not trust any of the results in Table 1 without first determining whether the asset pricing specifications listed in Table 1 produce unbiased estimates of the industry-adjusted abnormal return on GIM’s governance portfolio. Thus, before we can draw any conclusions from Table 1, we must first evaluate the size and power properties of each of the 24 tests to identify the asset pricing specifications that are both unbiased and that have the best power to detect abnormal returns. In the following section, I accomplish this task by examining the size and power properties of the 24 tests in Table 1 based on the random sampling methodology developed by JMS. Second, the t-statistics in Table 1 are generally far lower than the t-statistic reported by GIM, despite the fact that the mean abnormal monthly return reported in the table is nearly identical to the abnormal return estimate reported by GIM. For example, the combination of narrowly-defined 4-digit SIC classifications from Compustat and the PR1YR model produces a point estimate in Table 1 of 0.47% per month, which is identical to the
12
coefficient reported by GIM, but the t-statistic in Table 1 is only 1.3 (compared with GIM’s t-statistic of 2.1). Table 1 also shows that the asset pricing specification used by GIM is not statistically significant when industry returns are computed using non-Democracies and nonDictatorships, even though the monthly abnormal return estimate is nearly identical to the estimate reported by GIM (0.42% vs. 0.47%). This suggests that the use of non-Democracies and non-Dictatorships may produce tests with weak power to detect abnormal returns, which would explain the lack of stars in Table 1 despite the strong economic significance of the table’s abnormal return estimates. I return to this discussion in my robustness tests in section 3. Finally, it is worth noting that the original industry-adjusted asset pricing test employed by GIM is by no means extreme. Figure 1 provides a graphical depiction of the abnormal return estimates from Table 1 ordered from high to low. The figure shows that the asset pricing specification employed by GIM produces an industry-adjusted point estimate that lies in the middle of the abnormal return distribution from the table, while the asset pricing specification preferred by JMS lies towards the bottom of the abnormal return distribution. Thus, broadly-defined classification standards do not appear to produce the extreme point estimates that one might expect if such classifications were severely mis-measuring the true return on industries. In fact, industry-adjusted abnormal returns computed using the most granular industry classification standard (four-digit SIC classifications from CRSP) yield the highest abnormal return estimates in Table 1 and Figure 1. As such, Figure 1 appears to show that GIM employed a reasonable test in their paper, even if they did not examine the size and power properties of their preferred specification.
13
2.2 2.2.1
Size Tests Random Sampling Methodology
I next examine the statistical size of the 24 industry-adjusted asset pricing tests listed in Table 1. To measure size, I follow JMS and examine the industry-adjusted returns on portfolios of randomly selected firms that mimic the industry clustering present within GIM’s Democracy and Dictatorship portfolios. When constructing these random portfolios, however, I must be careful to avoid potential sources of bias. If Democracies unconditionally earned higher risk-adjusted returns than Dictatorships during the 1990s, random samples that include Democracies will be biased in favor of rejecting the null hypothesis of zero abnormal returns. Thus, to avoid direct use of the Democracy and Dictatorship portfolios, I follow JMS and return to the non-Democracy and non-Dictatorship samples described previously. These samples are designed to be approximately governance-neutral. As such, calendar-time regressions involving portfolios of randomly-sampled non-Democracies and non-Dictatorships that mimic the industry clustering present in the actual Democracy and Dictatorship samples should produce abnormal return estimates of close to zero if an asset pricing specification is “well-specified.” I now turn to the specifics of the random sampling methodology designed by JMS. For brevity, my description is limited to the non-Democracy sample; random samples involving non-Dictatorships are constructed similarly. First, using industry classification standard X, I assign a probability weight π to each industry within the non-Democracy sample on each of the four portfolio rebalancing dates (1990, 1993, 1995, and 1998) within GIM’s sample period. For a given industry, π is computed as the ratio of the market capitalization of that industry within the Democracy sample to the market capitalization of that industry within the non-Democracy sample. All non-Democracy firms within a given industry are assigned the same probability weight. Firms are then selected into each sample portfolio using independent Bernoulli trials with success probability π, and the sample portfolios are 14
rolled forward until the next rebalancing date. Additional details about the methodology employed by JMS can be found in Appendix A. After forming my samples, I next construct the monthly industry-adjusted return on each firm in my sample portfolios using industry classification Y . Thus, the industry classification used to form the sample portfolios need not be the same as the industry classification used to compute industry-adjusted returns. Following JMS, the industry return that is assigned to a firm in the non-Democracy sample is computed as as the value-weighted monthly return on all other non-Democracy firms that share the same industry classification as the firm of interest. Industry returns are assigned to firms in the non-Dictatorship sample along analogous lines. Value-weighted returns on the industry-adjusted non-Democracy and nonDictatorship sample portfolios are then computed using equation (2). Finally, returns on an industry-adjusted portfolio that is long non-Democracies and short non-Dictatorships are utilized as the dependent variable in the regression specifications listed in equation (3). I repeat this entire process 1,000 times for each of the 24 asset pricing tests listed in Table 1. Thus, I construct 1,000 sample portfolios and run 1,000 regressions for each asset pricing test listed in Table 1. For a given test, I then compute the fraction of 1,000 regression intercepts that reject the null hypothesis of zero abnormal returns at a given significance level, and compare this fraction with the appropriate theoretical rejection benchmark. For example, if a test is “well-specified,” we would expect approximately 25 of my 1,000 sample portfolios (2.5%) to reject the two-tailed null hypothesis of zero abnormal returns in the right tail at the 5% statistical significance level. Following Lyon, Barber and Tsai (1999) and JMS, I use a one-sided binomial test to measure the extent to which the actual rejection probabilities in 1,000 random samples exceed their associated theoretical benchmarks in both the right and left tails of the empirical distribution.
15
2.2.2
Results
Table 2 reports the results of my size tests. Tests that are “well-specified” – that is, tests that do not reject the null hypothesis too frequently at the 10%, 5%, or 1% theoretical significance levels in either tail – are shaded in gray. For brevity, the table only lists the fraction of tests that reject the null hypothesis at the 10% significance level in the right tail. However, to be “well-specified,” and thus to be highlighted in gray, tests must produce reasonable rejection probabilities at all three significance levels in both tails. For comparison purposes, the six tests examined by JMS in their paper are also denoted in the table by the symbol #. Table 2 shows that when samples are constructed to mimic the three-digit SIC clustering that is present within the Democracy and Dictatorship portfolios, asset pricing tests involving three-digit SIC-adjusted returns and the UMD asset pricing model produce well-specified tests, while tests involving FF48-adjusted returns tend to systematically over-reject the null hypothesis of zero abnormal returns in the right tail. Thus, based on the limited number of tests in their paper, JMS were correct to conclude that three-digit SIC industry adjustments were preferable to FF48 adjustments in the context of the GIM governance sample. However, JMS made no effort to show that their preferred asset pricing model was the only model that was “well-specified.” Unfortunately, after extending their tests to incorporate a more representative selection of industry classifications, I find little support for the key assertions made by JMS. First, I find that 23 of the 24 asset pricing tests listed in Table 1 have at least one well-specified test in Table 2 (the exception is the combination of six-digit GICS classifications and the FF3 model). In total, I find that 96 different tests in Table 2 are well-specified. Thus, the asset pricing model recommended by JMS is not the only model with strong econometric properties based on their testing methodology. Second, I find that all but one of the asset pricing tests that yield statistically significant industry-adjusted abnormal returns in Table 1 are also well-specified in Table 2 in at least one specification (again, the outlier is the combination of six-digit GICS and the FF3 model).
16
Thus, I find that well-specified tests do not unilaterally produce statistically insignificant industry-adjusted abnormal returns on the GIM governance portfolio. Third, the number of gray boxes in each row of Table 2 does not appear to increase or decrease systematically across the 24 asset pricing tests listed in the table. Thus, narrowly-defined classifications do not produce industry-adjusted abnormal return tests that are “better-specified” than tests involving broadly-defined classifications. Finally, Table 2 shows that JMS’s results are extremely sensitive to the assumptions made in their paper. For example, JMS argue that narrowly-defined classifications should be used to form the sample portfolios, which they form using three-digit SIC classifications. However, Table 2 shows that if JMS had constructed their random samples using four-digit SIC classifications, which are more narrowly defined than three-digit SIC classifications, their findings would be completely reversed: the combination of FF48-adjusted returns and the PR1YR model used by GIM would now be well-specified, while all models involving three- and four-digit SIC-adjusted returns would now over-reject the null hypothesis in the lef t tail and would thus be poorly-specified. Similar findings obtain when samples are constructed using (more) narrowly-defined three-digit and four-digit SIC classifications from CRSP. Thus, I find that the key assertions made by JMS do not survive even small changes to their underlying assumptions. Theoretically, the choice of the industry classification standard that is used to form the sample portfolios should only affect the number of firms in each sample portfolio and the prevalence with which certain firms appear in the random samples. This choice has no impact whatsoever on the industry-adjusted return computation that should in theory control for the effects of industry membership. Thus, we might expect tests that are well-specified when classification J is used to form the sample portfolios to also be well-specified when classification K is used to form these portfolios. Broadly speaking, Table 2 is supportive of this intuition: the gray boxes are more correlated across rows than across columns.8 8
In addition, we might expect tests to be well-specified at a higher-than-average rate when the same
17
However, Table 2 also shows that the number of well-specified tests declines precipitously when random samples are formed to mimic the three-digit SIC industry clustering present within GIM’s Democracy and Dictatorship portfolios. While there is no clear explanation for this finding, it appears that JMS selected the approach that in hindsight produced the smallest number of well-specified tests.
2.3
Power Tests
I next examine the statistical power of the asset pricing tests reported in Table 2 using the same sample portfolios that were used in the previous section. However, there is little point in evaluating the power properties of tests that are poorly-specified in Table 2. As such, I eliminate all industry classification/asset pricing model combinations that are not wellspecified in Table 2. Of the remaining tests, I then seek to identify the asset pricing models that have the best power to detect abnormal returns within the range of point estimates reported in Table 1 (about 0.2% to 0.8% per month). My testing methodology is identical to the methodology employed by Mitchell and Stafford (2000) and JMS. I begin by injecting abnormal returns ranging from 0.10% to 1.50% per month into each of my random samples. Following JMS, I add half of the stated “alpha” level to the returns on sample Democracies and subtract half of the stated “alpha” level from the returns on sample Dictatorships. I then regress the returns on these portfolios against the PR1YR, UMD, and FF3 asset pricing models. Finally, I compute the percentage of 1,000 regression intercepts for each asset pricing model that reject the null hypothesis of zero abnormal returns in the right tail at the 10% statistical significance level. Consistent with JMS’s findings, Table 3 shows that the combination of three-digit SICadjusted returns and the UMD asset pricing model has the best power to detect abnormal industry classification system that is used to construct the samples is also used to compute industry-adjusted returns. Table 2 does not support this hypothesis. However, this finding appears to be specific to the testing methodology employed by JMS and is described in more detail within Appendix C.
18
returns when the sample portfolios are formed to mimic the three-digit SIC clustering that is present within the GIM governance portfolio. However, there is no reason why samples should only be formed to mimic the three-digit SIC clustering present in the GIM governance portfolio. Hence, Table 3 lists the asset pricing model with the best power to detect abnormal returns for each of the eight industry classification standards that are used to form my sample portfolios, conditional on the asset pricing test being well-specified in Table 2 (the full test results can be found in Table A.1 of the Appendix). The table shows that a total of ten asset pricing tests have the best power to detect abnormal returns across the range of abnormal returns listed in Table 1 (0.2% to 0.8%). Asset pricing tests are shaded in gray if they also produce a positive and statistically significant industry-adjusted abnormal return estimate in Table 1. Of the 56 boxes in Table 3 (eight industry classifications used to form the sample portfolios times seven abnormal return levels), 38 such boxes, or approximately 68%, are shaded in gray. Thus, a sizable majority of well-specified tests with the best power to detect abnormal returns also produce statistically significant abnormal returns on the GIM governance portfolio in Table 1. In particular, I find that the combination of six-digit GICS-adjusted returns and the UMD asset pricing model has the best power to detect abnormal returns in 34 of the 56 boxes (61%) in Table 3, and these boxes are distributed fairly evenly across the industry classifications used to form the sample portfolios and across various levels of injected abnormal returns. Thus, from among the tests that are well-specified in Table 2, the combination of six-digit GICS-adjusted returns and the UMD asset pricing model appears to have the best power to detect abnormal returns. Importantly, Table 1 shows that the abnormal return associated with this asset pricing specification is 0.49% per month, statistically significant at the 10% level. Thus, while I cannot rule out the possibility that one of the other asset pricing tests listed in Table 3 is the “correct” test, the evidence presented in Tables 1-3 suggests that unexpected industry performance is unlikely to fully explain the
19
seminal results reported by GIM.
2.4
Precision Tests
As a second test of statistical power, I next extend the asset pricing “precison” test first developed by Metrick (1999) to a setting involving industry-adjusted returns. Consider a calendar-time regression of industry-adjusted returns on a portfolio of firms (say, Democracies) against a set of asset pricing factors. The test statistic associated with the intercept term from this regression can be written as:
tx =
αx , σx
(4)
where tx is the t statistic for the intercept term under asset pricing model x, αx is the intercept term from the regression, and σx is the standard error of the intercept term. Now suppose that each firm in the Democracy portfolio obtains a small, constant increase in its return each period. Since asset pricing factors typically contain a large number of stocks, incrementally increasing the return on each firm in the Democracy portfolio should have virtually no impact on the factor returns. Futhermore, under JMS’s methodology, each industry that is matched with Democracies contains only non-Democracies (and vice versa for Dictatorships). Thus, increasing the return on Democracies will have no impact on the industry returns that are matched with Democracies.9 As such, the increased return on each firm in the Democracy portfolio each period should be fully captured by the intercept term in the regression. The change in the test statistic associated with the increased portfolio 9
The industry return that would be matched with Dictatorships would contain non-Dictatorships, which would include Democracies. As such, increasing the return on Democracies would have a small impact on the industries matched with Dictatorships. However, JMS show in Table 2 of their paper that the nonDemocracy and non-Dictatorship portfolios are effectively governance-neutral. Thus, for simplicity, I proceed under the assumption that the industries matched with Democracies and Dictatorships would not be affected by increasing the return on all Democracies.
20
return is thus given by: ∂tx adj ∂Rport
=
1 , σx
(5)
adj where Rport is the industry-adjusted return on the Democracy portfolio. Thus, the increase
in the test statistic associated with a small, consistent change in portfolio returns is simply the inverse of the standard error when industries are constructed using non-Democracies and non-Dictatorships. I refer to the expression above as the precision of the asset pricing model in question. To compare the precision of asset pricing model x with another asset pricing model y, I simply calculate
∂tx adj ∂Rport
and
∂ty adj ∂Rport
as above. If
∂tx adj ∂Rport
>
∂ty adj , ∂Rport
model x is defined
as being more precise than model y when it comes to detecting abnormal returns. Using equation (5), I can now estimate the precision of each of the 24 asset pricing tests listed in Table 1. However, it is not clear how to estimate the precision of a long/short portfolio, since increasing the return on all firms within the portfolio will not lead to an increase in the long/short portfolio return. As such, I estimate precision separately for the Democracy and Dictatorship portfolios. I then rank each of the 24 asset pricing tests in descending order according to their estimates of precision. The same procedure is performed for Dictatorships. Finally, I compute the combined rank for each asset pricing test by summing its respective ranks within the Democracy and Dictatorship lists.10 Thus, the test with the lowest combined rank has the highest precision score. Table 4 shows that the combination of six-digit GICS industry adjustments and the FF3 asset pricing model has the lowest combined rank, and hence the highest precision score, from among the 24 tests under consideration.11 However, this combination does not produce any tests that are “well-specified” in Table 2. Thus, I turn to the test with the second 10
This is akin to equally weighting precision scores across the two sets of rankings. I could also compute the average precision of each asset pricing specification. However, given that estimates of precision vary more widely across the Dictatorship portfolios, this would be placing disproportionate weight on the Dictatorship precision estimates. 11 Out of curiosity, I also computed precision rankings using the standard errors from my long/short portfolios. One interpretation would be that I am increasing the return on Democracies, but not on Dictatorships. My results are qualitatively unchanged.
21
highest precision in Table 4, which is the combination of six-digit GICS classifications and the PR1YR model. This combination also shows up in Table 3 as having strong power properties while also being well-specified. Thus, it is the test with the highest precision score that is also well-specified. Importantly, this combination yields an industry-adjusted abnormal return that is statistically significant at the 5% level in Table 1. The combination of six-digit GICS and the UMD model that features so prominently in Table 3 has the third highest precision score out of the 24 asset pricing tests in Table 4, and this test also yields statistically significant industry-adjusted abnormal returns in Table 1. Thus, the tests with the highest precision are also tests that produce positive and statistically significant abnormal returns in Table 1. If narrowly-defined industries are the “true” industries and GIM’s results are largely driven by the performance of these industries, tests involving narrowly-defined industries should have lower standard errors than tests involving broadly-defined industries. However, Table 4 shows little support for this argument. The table shows that FF48 classifications and GICS classifications generally have strong precision rankings, while specification tests involving SIC classifications have poor precision rankings. Since precision is simply the inverse of the standard error, the results in Table 4 show that tests involving narrowly-defined industry classifications have higher standard errors than tests involving more broadly-defined classifications. This suggests that industry returns that are computed using narrowly-defined industry classifications may often contain a significant amount of noise, even if such classifications do the best job of grouping together firms with similar economic characteristics.
3
Robustness
In this section, I perform a number of robustness tests to ensure that my findings in the previous section are not being driven by specific testing assumptions or quirks in my data. I begin by describing a number of methodological changes that are made to my sample to 22
more closely align the testing assumptions used by JMS with the common standards in the literature. I then report the results of my robustness tests. For brevity, detailed descriptions of most of my tests are relegated to the Appendix.
3.1 3.1.1
Methodological Changes Current versus Historical Industry Classifications
To be consistent with JMS, firms were assigned to SIC and FF48 industries in the previous section based on the DN U M variable from Compustat. However, researchers should generally use historical industry classifications when such classifications are available. Since Compustat contains historical SIC and GICS classifications, I use Compustat’s historical classifications in the tests that follow. Unfortunately, Compustat’s historical SIC and GICS coverage is relatively poor.12 Thus, to improve coverage, I follow BLO and augment historical classification assignments with current SIC and GICS assignments from Compustat.13 This procedure ensures that I am accurately capturing historical industry assignments while also ensuring robust coverage within my sample.
3.1.2
Calendar-year versus Mid-year Industry Assignments
In the previous section, data from Compustat in year t was used to assign firms to industries from January to December of year t + 1 (following JMS). This choice was arbitrary in the previous section since firms did not change industries over time. However, the use in this section of historical industry assignments makes the choice of an assignment period relevant. 12
Compustat’s coverage is particularly poor for GICS. As noted by BLO, historical GICS codes are only available in Compustat for firms in the S&P 500, S&P Midcap 400, and S&P Smallcap 600 indices prior to 1999. Data is available for these firms beginning in 1994 with the exception of S&P 500 firms, which are available beginning in 1985. Despite this limitation, however, I am able to assign GICS classifications for 99% of firm-months in the GIM sample and 76% of firm-months in the merged CRSP/Compustat database. S&P sells a separate “GICS History” product (not available in Compustat) that provides historical coverage on more than 26,000 equities dating back to 1985. 13 I use the CRSP variables HSICCD and SICCD and the Compustat variables SIC, SICH, GIND/GSUBIND, and SPGIM, respectively, for current and historical SIC and GICS codes, and trim these variables as needed for different classification levels.
23
Importantly, one of the factors used by Compustat to assign firms to SIC classifications (and by extension, FF48 classifications) is “analyst judgment,” which likely depends in part on firms’ annual and quarterly results. Thus, I follow Fama and French (1997) and assign firms to industries from July of year t + 1 to June of year t + 2 based on year t data from Compustat. Industry classifications from CRSP are also lagged by one month.
3.1.3
Eliminating Non-Democracies and Non-Dictatorships
Following JMS, industry returns were constructed in the previous section using non-Democracies and non-Dictatorships exclusively. While this is not a common practice within the literature, it makes sense to exclude Democracies and Dictatorships from industries under the null hypothesis that Democracies outperformed Dictatorships during the 1990s. If Democracies and Dictatorships were included in industries, industry returns could contain “alpha” that would be unrelated to industry performance. If Democracies and Dictatorships are clustered within specific industries (as in the case in this setting), the “alpha” contained in industry returns will tend to reduce industry-adjusted abnormal returns on the long/short portfolio below their “fair” value. On the surface, this appears to show that the methodology employed by JMS is conservative: by excluding Democracies and Dictatorships from industries, JMS’s methodology should make it easier to identify the GIM alpha, which subsequently makes their finding of no alpha all the more striking. However, the use of non-Democracies and non-Dictatorships severely limits the size of each industry. For example, firms that are not part of the GIM governance sample are excluded by default, since they are not assigned a G-index value. As a result of these exclusions, Giroud and Mueller (2011) report that more than 80% of the firms in the merged CRSP/Compustat universe over the GIM sample period are excluded from JMS’s industry computations. Consistent with Giroud and Mueller (2011), Table A.2 of the appendix shows that industries are indeed significantly smaller when computed using non-Democracies and
24
non-Dictatorships. For example, more than 50% of three-digit SIC industries computed using JMS’s methodology contain ten or fewer firms, whereas only 13% of three-digit SIC industries contain ten or fewer firms when industries are constructed using the universe of stocks. An industry return computed using ten firms is likely to contain noise that is unrelated to industry performance. While this noise should average out across industries, it will still affect standard errors. Thus, a tradeoff exists: assuming that Democracies and Dictatorships cluster within industries, the methodology employed by JMS is likely to produce higher point estimates but also higher standard errors relative to a methodology that constructs industries using the universe of stocks. In the tests that follow, I construct industries using the universe of stocks. This approach is consistent with the null hypothesis from Table 1 that Democracies did not outperform Dictatorships over my sample period. Under a null hypothesis of zero abnormal returns, there would be no need to exclude Democracies and Dictatorships when constructing industry returns. Indeed, as shown in Table A.2, doing so would only reduce the size of each industry significantly, which could subsequently reduce the power of my tests to detect abnormal returns. Thus, although my point estimates are likely to be lower than the point estimates in the previous section, standard errors should also fall, making it easier to detect abnormal returns. Appendix B contains a simple mathematical expression of this tradeoff.
3.2
Revised Industry-adjusted Abnormal Return Tests
I begin my robustness tests by replicating the analysis from Table 1 using my revised data sample and industry construction methodology. Per the discussion above, I assign firms to industries based on historical classification assignments, which are augmented with current classifications as needed. Following Fama and French (1997)’s industry construction methodology, I assign industry classifications reported by Compustat in fiscal year t to firms from July of year t + 1 to June of year t + 2. As before, industry classifications taken from CRSP
25
are lagged by one month. I also eliminate the use of non-Democracies and non-Dictatorships by computing industry returns using all firms in the merged CRSP/Compustat database. This approach ensures that each industry contain the largest number of possible firms, which in theory should minimize the idiosyncratic noise within industry returns. Table 5 reports industry-adjusted abnormal returns on the GIM governance portfolio using my revised industry returns and data sample. As in Table 1, Table 5 shows that industryadjusted abnormal returns based on six-digit GICS and four-digit SIC (CRSP) classifications are economically and statistically significant. However, Table 5 shows that tests involving FF48 and FF48 (CRSP) classifications also yield statistically significant point estimates. Thus, after applying a more conventional set of assumptions to my industry-adjusted asset pricing tests, I find that a larger number of asset pricing specifications produce statistically significant abnormal returns on the GIM governance portfolio. As before, however, we should refrain from drawing conclusions from Table 5 without first examining the size and power properties of the 24 asset pricing specifications listed in the table. Thus, in the following sections, I re-run the size and power tests described in section 2 on my revised data sample to identify the asset pricing specification(s) with the best size and power properties in the context of my revised sample. As predicted in the previous section, I also find that point estimates in Table 5 are often substantially lower than point estimates in Table 1, particularly for tests involving narrowly defined industry classification standards. For example, the test specification based on fourdigit SIC (CRSP) classifications and the PR1YR asset pricing model yields a point estimate of 0.72% per month in Table 1 but only 0.30% in Table 5. However, while point estimates are generally lower in Table 5 than in Table 1, t-statistics are frequently higher in Table 5 than in Table 1. For example, point estimates based on three-digit SIC classifications and the PR1YR model drop from 0.39% per month to 0.25% per month, but the t-statistic actually increases from 1.24 to 1.52. This finding is consistent with the intuition described previously:
26
since the industries formed using JMS’s methodology contain fewer firms than the industries formed using the universe of stocks, the former set of industries should have higher levels of idiosyncratic noise within each industry return. All else equal, this will produce higher standard errors.
3.3
Revised Size and Power Tests
I next perform a second set of specification tests on my updated data sample using a slightly revised version of the testing methodology outlined in sections 2.2 and 2.3. As before, I would like to identify the test specifications that are “well-specified” and also possess the best power to detect abnormal returns. To accomplish this task, I eliminate all industry classification/asset pricing model combinations that are not well-specified in my size tests, which are reported in Table A.3. Of the well-specified tests, I then seek to identify the specification with the best power to detect abnormal returns within the range of alphas reported in Table 5 (about 0.1% to 0.5% per month). However, unlike the power tests in the previous section, I also standardize the beginning level of alpha to be exactly zero across all of the well-specified tests in my sample. This approach, which I believe is novel to the literature, ensures that all of the specifications are on a level playing field. Without this adjustment, tests that reject the null hypothesis in the right tail with the most frequency in my size tests would have an unfair advantage in my power tests. Additional details about the construction of my size and power tests can be found in Appendix C. Table 6 provides a summary of the power test results (the full results can be found in Table A.4). The table shows that approximately 68% of the ‘winning’ asset pricing specifications involve industry-adjusted returns computed using six-digit GICS and FF48 classifications and the PR1YR and FF3 asset pricing models. These asset pricing specifications also yield industry-adjusted abnormal returns in Table 5 that are statistically significant at the 5% level. Thus, as in the previous section, I find that many of the asset pricing specifications
27
with the best size and power properties also yield statistically significant industry-adjusted abnormal returns on the GIM governance portfolio. Consistent with Table 3, I also find that asset pricing specifications involving six-digit GICS industry adjustments are by far the best-represented specifications in the table, although here these classifications are paired with the FF3 asset pricing model rather than the UMD model. I also run a number of other size and power tests that alter the basic assumptions of JMS’s tests in various ways. To my knowledge, these methodologies are also new to the literature. As one example, I reconstruct my size tests using a methodology in which I draw firms at random from each industry without the aid of probability weights, and then simply constrain the weight on each industry within my sample portfolios to match the actual industry weights in the Democracy and Dictatorship portfolios. This methodology and the other new methodologies I consider, outlined in Appendix C, are significantly less complicated to implement than the methodology developed by JMS. Importantly, my results using these new methodologies are qualitatively unchanged from my previous results.
3.4
Revised Precision Tests
I also extend the precision tests reported in section 2.4 to a setting where industry returns are computed using the universe of stocks. This setting requires changes to the methodology described in section 2.4, since increasing the return on Democracy firm i will now also increase the return on the industry that is matched with firm i. Appendix D describes these changes, and the results of my tests are listed in Table A.5. Consistent with my results in section 2.4, I find that the combination of six-digit GICS industry adjustments and the FF3 model has the highest precision score when industries are computed using the universe of stocks. This combination is also well-represented in Table 6 as having strong size and power properties. Thus, I find that the combination of six-digit GICS adjustments and the FF3 model has the best size and power properties across all of my robustness tests. Again, the
28
abnormal return associated with this specification is statistically significant at the 5% level in Table 5. Consistent with my previous discussion, I also find that asset pricing tests are generally more precise when industries are computed using the universe of stocks.
3.5
Industry-and-DGTW-adjusted Abnormal Return Tests
The industry-adjusted abnormal return tests in Table 5 do not control for the size, bookto-market, and momentum adjustments developed by Daniel, Grinblatt, Titman and Wermers (1997) (DGTW) and later updated by Wermers (2003).14 Thus, following JMS, I also compute industry-and-DGTW-adjusted abnormal returns on the GIM governance portfolio. Every June, DGTW assign each stock in the CRSP/Compustat merged database to size, book-to-market, and momentum quintiles. DGTW then compute returns for the following twelve months on these triple-sorted portfolios. I match all firms in the CRSP/Compustat database with their respective DGTW classifications and compute the DGTW-adjusted return on each firm. I then compute value-weighted, DGTW-adjusted returns on each industry. Next, I compute the industry-and-DGTW-adjusted return on each Democracy and Dictatorship. Finally, I form value-weighted Democracy and Dictatorship portfolios and regress the returns on the corresponding industry-and-DGTW-adjusted long/short portfolio against an asset pricing model. Untabulated results show that the asset pricing specifications involving six-digit GICS classifications, FF48 classifications, and the P R1Y R and F F 3 models earn industry-and-DGTW-adjusted abnormal returns that are statistically significant, as in Table 5. Thus, all of the tests with the best size and power properties in Table 6 also have industryand-DGTW-adjusted abnormal returns that are positive and statistically significant. While I am again unable to explicitly rule out the possibility that some other asset pricing specification is the “true” specification, these results and the other results in this paper suggest that industry performance is unlikely to fully explain the positive and statistically significant 14
I thank Russ Wermers for graciously providing this data, which is available for download at http://www.smith.umd.edu/faculty/rwermers/ftpsite/Dgtw/coverpage.htm.
29
abnormal returns on governance-sorted portfolios during the 1990s.
4
Discussion: Industry Size and Statistical Power
In sections 2 and 3, I find that tests involving six-digit GICS classifications, which are of “average” coarseness, have some of the best size and power properties within my sample. In this section, I argue that my previous findings are not surprising in light of the inherent tradeoff that exists between industry size and statistical power in calendar-time settings. To understand this tradeoff, it is useful to return to the primary motivation behind industry adjustments. Industry-adjusted returns are computed for the purpose of eliminating industry-specific return premia from a firm’s realized return. Since industry returns are not directly observable, however, two potential measurement problems emerge. First, industries may be improperly defined. In this case, adjusting a firm’s stock return by its industry return may not fully remove the industry component from returns. Second, industry returns may contain an idiosyncratic (firm-specific) component that is unrelated to industry performance. Both of these potential problems can lead to noisy industry-adjusted returns. Intuitively, one can picture an objective function whose argument represents the squared “error” of industry returns. In this setting, the “error” comes from either improper industry definitions or idiosyncratic noise within industry returns. Our goal as researchers is to construct industries that mimimize this objective function. However, minimizing this objective function requires a tradeoff between narrowly-defined and broadly-defined industry classification standards. On the one hand, narrow industry definitions may do a better job than broad industry definitions of grouping together firms with similar economic characteristics, thus reducing the frequency of improper industry definitions. On the other hand, however, broad definitions should produce industry returns with less idiosyncratic noise because broad definitions are constructed using a larger number of firms.15 15
While not stated explicitly, these competing objectives are also discussed in Moskowitz and Grinblatt
30
To formalize this concept, consider a calendar-time regression of industry-adjusted returns on firm i against a set of asset pricing factors, and suppose that industry-adjusted returns can be computed using two different industry classification standards – a broadlydefined classification standard and a narrowly-defined classification standard. The t-statistics associated with the intercept terms of these regressions can be writen as:
tbroad ≡
αbroad σbroad
and tnarrow ≡
αnarrow , σnarrow
(6)
where tx is the t-statistic of the regression intercept when industry-adjusted returns are computed using classification standard x. For simplicity, suppose that the “broad” industry contains n firms and the “narrow” industry contains n − k firms. Suppose also that all firms have the same market capitalization, and that industry returns are computed using the universe of stocks. Now consider a small but consistent increase in the return on firm i each period. As in my previous precision tests, incrementally increasing one stock’s return should have virtually no impact on the factor returns. However, firm i’s increased stock return will increase the return on its industry (whether broad or narrow), since each industry is assumed to contain far fewer firms than are contained in the asset pricing factors. Thus, the change in each t-statistic due to the increase in firm i’s return can be approximated as 1 − n1 ∂tbroad = ∂Ri σbroad
1 1 − n−k ∂tnarrow = . ∂Ri σnarrow
and
(7)
A narrowly-defined classification standard will have better power to detect abnormal returns than a broadly-defined classification standard when: ∂tnarrow ∂Ri ∂tbroad ∂Ri
=
1 1− n−k σnarrow 1 1− n σbroad
>1,
(1999), Hou and Robinson (2006), and Kadyrzhanova and Rhodes-Kropf (2011).
31
(8)
or, equivalently, when: σnarrow
0, the first term on the right-hand side of (9) is always less than one and is decreasing in k. In this case, the narrowlydefined classification must produce a test with a lower standard error than the broadlydefined classification in order for the inequality in (9) to hold. Even if σbroad > σnarrow , the broadly-defined standard may still produce tests with better statistical power if the standard error associated with the narrowly-defined standard does not fall enough to offset the decline in the first term on the right-hand side of equation (9). I now extend this logic to a setting involving calendar-time portfolio returns. When industry returns are computed using the universe of stocks, an increase in the return on portfolio firm i will also increase the value-weighted return on the industry associated with firm i by a factor of wi,ind ≡
market capi market capindustry
at each point in time. Aggregating across a
portfolio of stocks, the derivative of the t statistic with respect to a change in the portfolio return at a single point in time can be written as ∂t 1−λ = , ∂Rport σ where λ ≡
PN
i=1
(10)
wi ∗ wi,ind , N is the number of firms in the portfolio, wi represents the
market capitalization weight of firm i in the portfolio, and wi,ind ≡
mktcapi PJi j=1 mktcapj
is the market
capitalization weight of firm i within its industry (which contains a total of Ji firms, including firm i). Using equation (10), I can now estimate the precision of each of the eight industry classifications under study to determine how the tradeoff documented above impacts the choice
32
between narrowly-defined and broadly-defined classifications. To estimate precision, I draw 200 firms randomly each month from the merged CRSP/Compustat database and compute value-weighted, industry-adjusted returns on my portfolio of random firms. Since GICS classifications are not widely available before 1990, my sample period runs from January 1990 to December 2011. I draw 1,000 random portfolios for each of the eight industry classifications under consideration and regress the industry-adjusted returns on my portfolios against the three asset pricing models described previously. I then compute the average precision score for each classification using the standard errors from my regressions and time-series average values of λ.16 Consistent with the tradeoff described above between industry size and statistical power, Table 7 shows that industry classifications that are of “average” coarseness have higher precision scores than industry classifications that are either broadly-defined or narrowly-defined. The table shows that standard errors generally decline as industries become more granular, indicating that narrowly-defined industries likely do a better job than broadly-defined industries of grouping together firms with similar economic characteristics. However, λ values also increase as industries become more granular, indicating that the sample firms constitute a greater and greater fraction of each industry as industries become more granular. All else equal, this will reduce the statistical power of the associated asset pricing test, making it more difficult to detect abnormal returns. Hence, the table shows that industry classifications of “average” coarseness appear to do the best job of balancing the tradeoff between improper industry definitions and the reduction in statistical power caused by idiosyncratic noise. Furthermore, given that my portfolios are completely random (and thus have nothing to do with governance), the results in Table 7 may be of interest to asset pricing researchers in a number of other contexts. 16
The λ term would disappear from (10) if sample firms were excluded from the industry returns. When this approach was evaluated in section 2.4, however, industry sizes were reduced considerably and the results mirrored the results in Table 7.
33
5
Conclusions
This paper reexamines the link between industry performance, corporate governance, and equity prices during the 1990s. Unlike previous papers, I take a “kitchen sink” approach and examine industry-adjusted returns on governance-sorted portfolios using a variety of industry classification standards that are common to the literature. Using industry-adjusted asset pricing tests that are unbiased and have strong statistical power, I find little evidence that the abnormal returns reported by GIM can be explained by conditional industry return premia or a poorly-specified asset pricing model. In contrast, the majority of tests with the strongest size and power properties in my sample yield positive industry-adjusted abnormal returns on governance-sorted portfolios during the 1990s. Thus, while I cannot explicitly rule out industry performance as a contributing factor, my findings suggest that industry adjustments alone are unlikely to explain the relationship between governance and equity prices during the 1990s first documented by GIM. I also document a more general tradeoff between industry coarseness and statistical power that may be of independent interest to empirical asset pricing researchers. Until an industry classification standard emerges in the literature that clearly dominates its competitors, researchers must continue to grapple with the inherent flaws and tradeoffs that are common to all industry classification standards. Moreover, the findings in this paper acutely highlight the perils of relying too heavily on any one particular industry classification standard. However, it is impractical for every researcher to conduct dozens of industry-adjusted size and power tests until he or she identifies the unique combination that is best suited to their data sample. As such, one of the key takeaways from this paper is that the profession needs to develop a simple, effective way of identifying the “best” industry in various settings. I leave this task to future research.
34
References Bebchuk, Lucian A., Alma Cohen and Allen Ferrell. 2009. “What Matters in Corporate Governance?” Review of Financial Studies 22:783–827. Bebchuk, Lucian A., Alma Cohen and Charles C. Y. Wang. forthcoming. “Learning and the Disappearing Association Between Governance and Returns.” Journal of Financial Economics . Bhojraj, Sanjeev, Charles M. C. Lee and Derek K. Oler. 2003. “What’s My Line? A Comparison of Industry Classification Schemes for Capital Market Research.” Journal of Accounting Research 41:745–774. Carhart, Mark M. 1997. “On Persistence in Mutual Fund Performance.” Journal of Finance 52:57–82. Chan, Louis K. C., Josef Lakonishok and Bhaskaran Swaminathan. 2007. “Industry Classifications and Return Comovement.” Financial Analysts Journal 63:56–70. Clarke, Richard N. 1989. “SICs as Delineators of Economic Markets.” Journal of Business 62:17–31. Core, John E., Wayne R. Guay and Tjomme O. Rusticus. 2006. “Does Weak Governance Cause Weak Stock Returns? An Examination of Firm Operating Performance and Investors’ Expectations.” Journal of Finance 61:655–687. Cremers, K. J. Martijn and Allen Ferrell. 2011. “Thirty Years of Corporate Governance: Firm Valuation and Stock Returns.” Yale ICF Working Paper No. 09-09. Cremers, K. J. Martijn and Vinay B. Nair. 2005. “Governance Mechanisms and Equity Prices.” Journal of Finance 60:2859–2894.
35
Cremers, K. J. Martijn, Vinay B. Nair and Kose John. 2009. “Takeovers and the CrossSection of Returns.” Review of Financial Studies 22:1409–1445. Daniel, Kent, Mark Grinblatt, Sheridan Titman and Russ Wermers. 1997. “Measuring Mutual Fund Performance with Characterstic-Based Benchmarks.” Journal of Finance 52:1035–1058. Fama, Eugene F. and Kenneth R. French. 1993. “Common Risk Factors in the Returns on Stocks and Bonds.” Journal of Financial Economics 33:3–56. Fama, Eugene F. and Kenneth R. French. 1997. “Industry Costs of Equity.” Journal of Financial Economics 43:153–193. Giroud, Xavier and Holger M. Mueller. 2011. “Corporate Governance, Product Market Competition, and Equity Prices.” Journal of Finance 66:563–600. Gompers, Paul A., Joy Ishii and Andrew Metrick. 2010. “Extreme Governance: An Analysis of Dual-Class Firms in the United States.” Review of Financial Studies 23:1051–1088. Gompers, Paul, Joy Ishii and Andrew Metrick. 2003. “Corporate Governance and Equity Prices.” Quarterly Journal of Economics 118:107–155. Gu, Lifeng and Dirk Hackbarth. 2011. “Governance and Equity Prices: Does Transparency Matter.” Working Paper, University of Illinois. Guenther, David A. and Andrew J. Rosman. 1994. “Differences between COMPUSTAT and CRSP SIC Codes and Related Effects on Research.” Journal of Accounting and Economics 18:115–128. Hou, Kewei and David T. Robinson. 2006. “Industry Concentration and Average Stock Returns.” Journal of Finance 61:1927–1956.
36
Johnson, Shane A., Theodore C. Moorman and Sorin Sorescu. 2009. “A Reexamination of Corporate Governance and Equity Prices.” Review of Financial Studies 22:4753–4786. Kadyrzhanova, Dalida and Matthew Rhodes-Kropf. 2011. “Concentrating on Governance.” Journal of Finance 66:1649–1685. Kahle, Kathleen M. and Ralph A. Walking. 1996. “The Impact of Industry Classifications on Financial Research.” Journal of Financial and Quantitative Analysis 31:309–335. Lyon, John D., Brad M. Barber and Chih-Ling Tsai. 1999. “Improved Methods for Tests of Long-Run Abnormal Stock Returns.” Journal of Finance 54:165–201. Masulis, Ronald W., Cong Wang and Fei Xie. 2007. “Corporate Governance and Acquirer Returns.” Journal of Finance 62:1851–1889. Metrick, Andrew. 1999. “Performance Evaluation with Transactions Data: The Stock Selection of Investment Newsletters.” Journal of Finance 54:1743–1775. Mitchell, Mark L. and Eric Stafford. 2000. “Managerial Decisions and Long-Term Stock Price Performance.” Journal of Business 73:287–329. Moskowitz, Tobias J. and Mark Grinblatt. 1999. “Do Industries Explain Momentum?” Journal of Finance 54:1249–1290. Weiner, Christian. 2005. “The Impact of Industry Classification Schemes on Financial Research.” SFB 649 Discussion Paper 2005-062, Humboldt University of Berlin. Wermers, Russ. 2003. “Is Money Really Smart? New Evidence on the Relation Between Mutual Fund Flows, Manager Behavior, and Performance Persistence.” Working Paper, University of Maryland.
37
Figure 1: Comparison of Monthly Industry-adjusted Abnormal Returns on the Gompers, Ishii, and Metrick (2003) Governance Portfolio
4-digit SIC (CRSP) / FF3 4-digit SIC (CRSP) / PR1YR 6-digit GICS / PR1YR
Industry Classification Standard and Asset Pricing Model
6-digit GICS / FF3 4-digit SIC (CRSP) / UMD 6-digit GICS / UMD 4-digit SIC / PR1YR 4-digit SIC / FF3
3-digit SIC (CRSP) / PR1YR FF48 / PR1YR FF48 (CRSP) / PR1YR FF48 / FF3 8-digit GICS / PR1YR
3-digit SIC (CRSP) / FF3 3-digit SIC / PR1YR
FF48 (CRSP) / FF3 3-digit SIC / FF3 8-digit GICS / FF3 3-digit SIC (CRSP) / UMD 4-digit SIC / UMD FF48 (CRSP) / UMD 3-digit SIC / UMD 8-digit GICS / UMD FF48 / UMD -0.10% 0.00% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% Monthly Abnormal Return
Notes: (1) The average monthly abnormal return estimate is indicated by the solid black line. (2) The median monthly abnormal return estimate is indicated by the dashed black line. (3) The abnormal return estimate from the test employed by Gompers, Ishii, and Metrick (2003) (FF48 (CRSP) industry classifications and the PR1YR model) is fully shaded. (4) The abnormal return estimate reported by Johnson, Moorman, and Sorescu (2009) (and replicated in Table 1) is cross-hatched.
38
Table 1 Industry-adjusted Abnormal Returns This table reports monthly abnormal return estimates from calendar-time regressions involving valueweighted, industry-adjusted returns on the long/short governance portfolio of Gompers, Ishii & Metrick (2003) (GIM). The portfolio takes a long position in firms with strong shareholder rights (Democracies) and shorts firms with weak shareholder rights (Dictatorships). The sample period is September 1990 to December 1999, and portfolio compositions are updated in September 1990, July 1993, July 1995, and February 1998. Monthly stock returns are sourced from the Center for Research in Security Prices (CRSP). Where applicable, delisting returns are used in months when a security is delisted from its primary exchange. Market capitalizations are taken from CRSP and are lagged by one month. The industry classification standards that are used to compute industry-adjusted returns on Democracy and Dictatorship firms are listed in the first column of the table. Industry definitions are sourced from Compustat, CRSP, and Kenneth French's website. Industry classifications from fiscal year t in Compustat are assigned to returns from January of year t +1 to December of year t +1. Industry classifications from CRSP are lagged by one month. The industry-adjusted return on Democracy firm i is computed as the value-weighted return on all non -Democracy firms in the GIM sample that share the same industry classification as firm i . Industryadjusted returns on Dictatorship firms are computed analogously. Industry-adjusted portfolio returns are regressed against three different asset pricing models: the three-factor model introduced by Fama and French (1993) (denoted FF3), which contains market return, size, and book-to-market factors; the FF3 model augmented with Carhart (1997)'s momentum factor (PR1YR ); and the FF3 model augmented with Kenneth French's UMD factor (UMD ). T -statistics are shown below each abnormal return estimate. *, **, and *** denote statistical significance at the 10%, 5%, and 1% level, respectively.
Industry Adjustment Method Fama-French 48 (CRSP) Fama-French 48 6-digit GICS 8-digit GICS 3-digit SIC 4-digit SIC 3-digit SIC (CRSP) 4-digit SIC (CRSP)
Asset Pricing Model PR1YR UMD FF3 0.42% 0.29% 0.39% (1.54) (1.03) (1.44) 0.44% 0.24% 0.42% (1.62) (0.86) (1.59) 0.63% ** 0.49% * 0.63% ** (2.41) (1.82) (2.47) 0.40% 0.24% 0.35% (1.55) (0.90) (1.39) 0.39% 0.26% 0.38% (1.24) (0.82) (1.25) 0.47% 0.31% 0.46% (1.33) (0.84) (1.32) 0.45% 0.31% 0.39% (1.35) (0.89) (1.19) 0.72% ** 0.57% * 0.77% ** (2.33) (1.81) (2.54)
39
Table 2 Size Tests This table examines the statistical size properties of randomly constructed portfolios containing governance-neutral firms that mimic the industry clustering present in the Gompers, Ishii & Metrick (2003) Democracy and Dictatorship portfolios. Democracies are replaced in the sample portfolios with non-Democracies, and Dictatorships are replaced with non-Dictatorships. Portfolios are designed such that the expected market capitalization within each industry in each sample portfolio matches the market capitalization of that industry in the corresponding Democracy or Dictatorship portfolio. Firms are randomly chosen for inclusion in each sample portfolio using a weighting scheme discussed in the text. The industry return assigned to each non-Democracy/non-Dictatorship firm in a given sample is computed as the value-weighted return on all other non-Democracies/non-Dictatorships in the dataset which contain the same industry classification. This industry return is then subtracted from the return on the parent firm to arrive at an industry-adjusted return. A zero-cost portfolio is then formed that takes a long position in the value-weighted, industry-adjusted sample Democracy portfolio and a short position in the analagous Dictatorship portfolio. Returns on this portfolio are then regressed against an asset pricing model. See section 2.3 and Table 1 for more details about the data and the asset pricing models employed in these tests. This procedure is employed 1,000 times, for a total of 1,000 zero-cost portfolios, and 1,000 regressions. The null hypothesis for all tests is that the regression intercept equals zero. Tests which are highlighted in gray are "well-specified" - they do not reject the null hypothesis of zero abnormal returns too frequently at any significance level in either tail relative to the appropriate theoretical rejection rates. The numbers in the table represent the fraction of tests that reject the null hypothesis at the 10% significance level in the right tail. For brevity, the other results are omitted. Tests marked with "#" can also be found in Johnson, Moorman & Sorescu (2009).
40
Table 2 (continued) Size Tests Industry classification used to compute industry-adjusted returns FF48 (CRSP)
FF48 (Compustat)
6-digit GICS
8-digit GICS
3-digit SIC
4-digit SIC
3-digit SIC (CRSP)
4-digit SIC (CRSP)
Asset Pricing Model PR1YR UMD FF3 PR1YR UMD FF3 PR1YR UMD FF3 PR1YR UMD FF3 PR1YR UMD FF3 PR1YR UMD FF3 PR1YR UMD FF3 PR1YR UMD FF3
Classification used to form portfolios that mimic the industry clustering in the Dem. / Dict. portfolios Numbers represent rejection rates at the 10% statistical significance level in the right tail FF48 FF48 6-digit 8-digit 3-digit 4-digit 3-digit 4-digit (CRSP) (Compustat) GICS GICS SIC SIC SIC (CRSP) SIC (CRSP) 2.50% 9.80% 6.20% 10.80% 10.40% 1.00% 3.00% 2.50% 1.10% 5.60% 2.60% 3.40% 5.90% 0.10% 0.70% 1.40% 4.80% 11.50% 8.60% 17.20% 11.80% 3.10% 4.90% 2.80% 5.00% 6.20% 5.80% 9.80% 15.40% # 2.90% 2.50% 3.20% 3.40% 2.60% 2.20% 2.80% 7.50% # 0.70% 1.50% 1.20% 7.10% 7.40% 7.40% 12.70% 14.40% # 3.30% 4.00% 3.70% 7.80% 9.60% 4.90% 12.80% 21.20% 6.50% 6.80% 7.80% 4.60% 6.70% 2.20% 5.10% 11.40% 2.30% 2.70% 4.90% 10.20% 12.80% 7.10% 22.00% 25.20% 9.70% 8.70% 7.50% 3.00% 4.50% 4.20% 2.90% 8.60% 0.70% 1.50% 3.00% 1.80% 2.60% 2.30% 1.10% 3.30% 0.00% 0.50% 1.40% 4.60% 6.50% 5.30% 3.20% 10.50% 1.70% 3.80% 3.00% 0.90% 3.40% 2.30% 4.40% 6.70% # 0.30% 0.70% 1.20% 0.40% 1.80% 1.10% 1.90% 4.90% # 0.00% 0.10% 0.40% 1.30% 3.40% 3.00% 4.90% 7.30% # 0.60% 1.20% 1.20% 0.60% 2.20% 2.00% 2.40% 3.70% 0.20% 0.20% 0.40% 0.20% 1.50% 1.10% 1.30% 2.10% 0.30% 0.20% 0.20% 0.60% 3.10% 2.60% 3.30% 4.10% 0.40% 0.60% 0.40% 2.60% 7.10% 3.60% 2.60% 12.10% 2.20% 3.10% 1.50% 1.50% 4.00% 1.70% 1.00% 8.50% 1.70% 1.40% 1.30% 4.20% 9.00% 4.20% 4.90% 13.60% 3.10% 4.40% 1.60% 2.00% 5.00% 2.00% 3.70% 14.50% 5.80% 2.50% 2.20% 1.30% 3.20% 1.00% 2.20% 9.00% 4.00% 1.30% 1.30% 2.30% 5.50% 2.20% 3.80% 12.90% 4.70% 3.10% 3.40% 41
Table 3 Summary of Power Test Results This table summarizes power tests involving 1,000 random sample portfolios that are formed to mimic the industry clustering present in the Gompers, Ishii & Metrick (2003) Democracy and Dictatorship portfolios. See Table 2 for a description of the sample portfolios. Each month, I add half of the alpha amount shown below to the long leg and subtract half of the alpha amount shown below from the short leg of each of the 1,000 long/short industry-adjusted portfolios. Industry-adjusted returns on these portfolios are then regressed against the PR1YR, UMD, and FF3 asset pricing models. See Table 1 for data sources and additional definitions. I perform power tests for all combinations of industry classifications and asset pricing models that produce "wellspecified" tests in Table 2. See Table 2 and the text for a description of "well-specified" tests. In each power test, I compute the fraction of 1,000 samples that reject the null hypothesis of zero alpha in the right tail using a one-sided t -test at the 10% statistical significance level. For each classification standard that is used to form the sample portfolios, the table lists the asset pricing specification with the best power to detect abnormal returns when I inject monthly abnormal returns of 0.2% to 0.8% per month (the range of abnormal returns from Table 1) into sample portfolios. Asset pricing specifications highlighted in gray yield industry-adjusted abnormal returns on the actual Gompers, Ishii & Metrick (2003) governance portfolio that are statistically significant at the 10% level or better in Table 1. Industry classification used to form portfolios that mimic the industry clustering present within the Democracy and Dictatorship portfolios FF48 (CRSP)
FF48
6-digit GICS
Industry classification used to compute industry-adjusted returns / Asset pricing model Monthly alpha injected into sample portfolios 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 6-digit 6-digit 6-digit 6-digit 6-digit 6-digit 6-digit GICS / UMD GICS / UMD GICS / UMD GICS / UMD GICS / UMD GICS / UMD GICS / UMD 8-digit GICS / PR1YR 6-digit GICS / PR1YR
8-digit GICS / PR1YR 6-digit GICS / PR1YR
FF48 (CRSP) / UMD 6-digit GICS / PR1YR
FF48 (CRSP) / UMD 6-digit GICS / PR1YR
FF48 (CRSP) / UMD 8-digit GICS / PR1YR
FF48 (CRSP) / UMD 8-digit GICS / PR1YR
FF48 (CRSP) / UMD 6-digit GICS / UMD
8-digit GICS
6-digit 6-digit 6-digit 6-digit 6-digit 6-digit 6-digit GICS / UMD GICS / UMD GICS / UMD GICS / UMD GICS / UMD GICS / UMD GICS / UMD
3-digit SIC
3-digit SIC 3-digit SIC 3-digit SIC 3-digit SIC 8-digit 8-digit 8-digit / UMD / UMD / UMD / UMD GICS / UMD GICS / UMD GICS / UMD
4-digit SIC
3-digit SIC (CRSP)
4-digit SIC (CRSP)
FF48 / PR1YR
6-digit 6-digit 6-digit 6-digit 6-digit 6-digit GICS / UMD GICS / UMD GICS / UMD GICS / UMD GICS / UMD GICS / UMD
FF48 (CRSP) / FF3
6-digit 6-digit 6-digit 6-digit 6-digit 6-digit GICS / UMD GICS / UMD GICS / UMD GICS / UMD GICS / UMD GICS / UMD
6-digit 6-digit 6-digit 6-digit 6-digit 6-digit 6-digit GICS / UMD GICS / UMD GICS / UMD GICS / UMD GICS / UMD GICS / UMD GICS / UMD
42
Table 4 Precision Tests For each asset pricing specification listed in Table 1, I regress calendar-time, value-weighted, industry-adjusted returns on the GIM Democracy and Dictatorship portfolios against the PR1YR , UMD , and FF3 asset pricing models described in the text. I perform separate regressions for the Democracy and Dictatorship portfolios. I then examine the precision of each of the 24 combinations of industry classifications and asset pricing models listed in Table 1. The variable Precision is defined as the inverse of the standard error of the intercept estimate. After computing precision, I then rank each test specification based on its precision score. I rank specifications separately for regressions involving Democracy and Dictatorship portfolio returns. Combined rank is defined as the sum of the precision ranks from the Democracy and Dictatorship columns. Hence, the test with the highest precision is the test with the lowest combined rank. Industry returns are computed using the methodology outlined in Table 1. Democracy portfolio Industry classification used to compute industry returns 6-digit GICS FF48 (CRSP) 6-digit GICS FF48 (CRSP) 6-digit GICS FF48 FF48 (CRSP) FF48 FF48 8-digit GICS 3-digit SIC 8-digit GICS 3-digit SIC (CRSP) 3-digit SIC 8-digit GICS 3-digit SIC 3-digit SIC (CRSP) 4-digit SIC (CRSP) 3-digit SIC (CRSP) 4-digit SIC (CRSP) 4-digit SIC (CRSP) 4-digit SIC 4-digit SIC 4-digit SIC
Asset Pricing Model Precision Rank FF3 5.77 1 FF3 5.73 2 PR1YR 5.61 3 PR1YR 5.58 4 UMD 5.48 5 FF3 5.40 6 UMD 5.40 7 PR1YR 5.26 8 UMD 5.12 9 FF3 5.02 10 FF3 4.91 11 PR1YR 4.88 12 FF3 4.80 13 PR1YR 4.78 14 UMD 4.76 15 UMD 4.74 16 PR1YR 4.67 17 FF3 4.58 18 UMD 4.56 19 PR1YR 4.46 20 UMD 4.34 21 FF3 4.27 22 PR1YR 4.16 23 UMD 4.09 24
Dictatorship portfolio Industry classification used to compute industry returns 8-digit GICS 8-digit GICS 6-digit GICS 8-digit GICS 6-digit GICS 6-digit GICS FF48 3-digit SIC FF48 FF48 (CRSP) 3-digit SIC FF48 FF48 (CRSP) 3-digit SIC FF48 (CRSP) 4-digit SIC (CRSP) 4-digit SIC (CRSP) 4-digit SIC (CRSP) 4-digit SIC 4-digit SIC 3-digit SIC (CRSP) 3-digit SIC (CRSP) 4-digit SIC 3-digit SIC (CRSP)
43
Asset Pricing Model Precision Rank FF3 7.36 1 PR1YR 7.17 2 FF3 7.07 3 UMD 6.97 4 PR1YR 6.89 5 UMD 6.71 6 FF3 6.71 7 FF3 6.54 8 PR1YR 6.54 9 FF3 6.51 10 PR1YR 6.39 11 UMD 6.38 12 PR1YR 6.36 13 UMD 6.15 14 UMD 6.13 15 FF3 6.12 16 PR1YR 5.98 17 UMD 5.81 18 FF3 5.59 19 PR1YR 5.47 20 FF3 5.33 21 PR1YR 5.29 22 UMD 5.25 23 UMD 5.01 24
Combined rank Industry classification used to compute industry returns 6-digit GICS 6-digit GICS 6-digit GICS 8-digit GICS FF48 (CRSP) FF48 8-digit GICS FF48 (CRSP) FF48 8-digit GICS 3-digit SIC FF48 FF48 (CRSP) 3-digit SIC 3-digit SIC 3-digit SIC (CRSP) 4-digit SIC (CRSP) 4-digit SIC (CRSP) 3-digit SIC (CRSP) 4-digit SIC (CRSP) 4-digit SIC 3-digit SIC (CRSP) 4-digit SIC 4-digit SIC
Asset Pricing Combined Model Rank FF3 4 PR1YR 8 UMD 11 FF3 11 FF3 12 FF3 13 PR1YR 14 PR1YR 17 PR1YR 17 UMD 19 FF3 19 UMD 21 UMD 22 PR1YR 25 UMD 30 FF3 34 FF3 34 PR1YR 37 PR1YR 39 UMD 39 FF3 41 UMD 43 PR1YR 43 UMD 47
Table 5 Alternative Tests of Industry-adjusted Abnormal Returns This table reports monthly abnormal return estimates from calendar-time regressions involving valueweighted, industry-adjusted returns on the long/short governance portfolio of Gompers, Ishii & Metrick (2003) (GIM). The portfolio takes a long position in firms with strong shareholder rights (Democracies) and shorts firms with weak shareholder rights (Dictatorships). The sample period is September 1990 to December 1999, and portfolio compositions are updated in September 1990, July 1993, July 1995, and February 1998. The industry classification standards that are used to compute industry-adjusted returns on Democracy and Dictatorship firms are listed in the first column of the table. A combination of historical and current industry classifications are assigned to each firm in the sample. Industry classifications from fiscal year t in Compustat are assigned to returns from July of year t +1 to June of year t +2. Industry classifications from CRSP are lagged by one month. The industry-adjusted return on Democracy firm i is computed as the value-weighted return on all firms in the merged CRSP/Compustat database that share the same industry classification as firm i . Industry-adjusted returns on Dictatorship firms are computed in an analogous fashion. See Table 1 for a description of the asset pricing models employed below and additional data definitions. Market capitalizations are lagged one month. T -statistics are shown below each abnormal return estimate. *, **, and *** denote statistical significance at the 10%, 5%, and 1% level, respectively.
Industry Adjustment Method Fama-French 48 (CRSP) Fama-French 48 6-digit GICS 8-digit GICS 3-digit SIC 4-digit SIC 3-digit SIC (CRSP) 4-digit SIC (CRSP)
Asset Pricing Model PR1YR UMD FF3 0.43% * 0.29% 0.41% (1.95) (1.26) (1.91) 0.44% ** 0.27% 0.43% (2.04) (1.20) (2.04) 0.39% ** 0.27% 0.40% (2.16) (1.47) (2.27) 0.25% 0.16% 0.24% (1.49) (0.93) (1.50) 0.25% 0.09% 0.23% (1.52) (0.56) (1.43) 0.19% 0.05% 0.16% (1.29) (0.32) (1.15) 0.31% 0.18% 0.30% (1.62) (0.93) (1.63) 0.30% ** 0.22% 0.31% (2.27) (1.59) (2.36)
44
* ** **
**
Table 6 Summary of Power Test Results based on Alternative Data Sample and Methodology This table summarizes power tests involving 1,000 random sample portfolios that are formed to mimic the industry clustering present in the Gompers, Ishii & Metrick (2003) Democracy and Dictatorship portfolios. See Table A.3 for a description of the sample portfolios. I begin by "standardizing" the abnormal return estimates on each sample portfolio by subtracting the average abnormal return estimate across all 1,000 samples from each sample's portfolio return. I then add half of the alpha amount shown below each month to the long leg and subtract half of the alpha amount shown below from the short leg of each of my 1,000 long/short industry-adjusted portfolios. I also inject abnormal returns into industry returns based on the market capitalization of the sample firms within that industry. See Appendix C of the text for more details. I then regress returns on the 1,000 sample portfolios against the PR1YR, UMD, and FF3 asset pricing models. I perform power tests for all combinations of industry classifications and asset pricing models that produce "well-specified" tests in Table A.3. For each classification standard that is used to form mimicking portfolios, the table lists the asset pricing specification with the best power to detect abnormal returns when we inject monthly abnormal returns of 0.1% to 0.5% per month into the sample portfolios. The percentages listed in the table below represent the fraction of 1,000 samples that reject the null hypothesis of zero alpha in the right tail using a one-sided t- test at the 10% statistical significance level. Asset pricing specifications highlighted in gray yield industry-adjusted abnormal returns on the actual Gompers, Ishii & Metrick (2003) governance portfolio that are statistically significant at the 10% level or better in Table 5. Industry classification used to form portfolios that Industry classification used to compute industry-adjusted mimic the industry returns / Asset pricing model clustering present within Monthly alpha injected into sample portfolios the Democracy and 0.10% 0.20% 0.30% 0.40% 0.50% Dictatorship portfolios FF48 FF48 6-digit 6-digit 6-digit FF48 (CRSP) (CRSP) / (CRSP) / GICS / FF3 GICS / FF3 GICS / FF3 PR1YR FF3 FF48
6-digit GICS
8-digit GICS
3-digit SIC
4-digit SIC
3-digit SIC (CRSP)
4-digit SIC (CRSP)
FF48 / FF3 6-digit GICS / PR1YR
6-digit 6-digit 6-digit 6-digit GICS / FF3 GICS / FF3 GICS / FF3 GICS / FF3 6-digit GICS / PR1YR
6-digit 6-digit 6-digit GICS / FF3 GICS / FF3 GICS / FF3
3-digit SIC 3-digit SIC 4-digit SIC 4-digit SIC 4-digit SIC / FF3 / FF3 / PR1YR / PR1YR / PR1YR 4-digit SIC 4-digit SIC 4-digit SIC 4-digit SIC 3-digit SIC (CRSP) / (CRSP) / (CRSP) / (CRSP) / / FF3 FF3 FF3 FF3 FF3 3-digit SIC 3-digit SIC 3-digit SIC 3-digit SIC 3-digit SIC (CRSP) / (CRSP) / (CRSP) / (CRSP) / (CRSP) / UMD UMD UMD UMD UMD 8-digit 6-digit 6-digit 6-digit 6-digit GICS / GICS / FF3 GICS / FF3 GICS / FF3 GICS / FF3 PR1YR FF48 FF48 FF48 FF48 FF48 / UMD (CRSP) / (CRSP) / (CRSP) / (CRSP) / FF3 FF3 FF3 FF3
45
Table 7 Precision Tests based on Random Samples This table reports the average precision of calendar-time asset pricing tests involving eight industry adjustment methods across 1,000 random samples drawn from the universe of stocks. The sample period is January 1990 to December 2011. I begin by drawing 200 firms at random each month from the merged CRSP/Compustat database. I then compute monthly industry-adjusted returns on each firm in the sample. I compute industry returns using all firms in the merged CRSP/Compustat database. Finally, I regress the value-weighted, industry-adjusted monthly returns on my sample portfolio against the three asset pricing models listed below. I repeat this process 1,000 times for each of the eight industry classifications listed in the first column of the table. The second column of the table reports the average number of firms in each industry across the time series. The next three columns list the average standard errors of the intercept term across 1,000 regressions. The sixth column reports the average weight of the sample firms in each industry, denoted λ. λ is defined formally in the text. Precision is defined as (1-λ)/σ. Precision scores are averaged across the three asset pricing models listed in the table below. Avg. Standard Error from 1,000 Industry-adjusted Abnormal Industry Average number Return Regressions Average Adjustment Method of firms per industry PR1YR UMD FF3 λ Fama-French 48 (CRSP) 152.2 0.18% 0.19% 0.18% 0.09 Fama-French 48 128.2 0.17% 0.18% 0.17% 0.10 6-digit GICS 82.4 0.16% 0.16% 0.16% 0.13 8-digit GICS 35.6 0.15% 0.15% 0.14% 0.21 3-digit SIC 23.1 0.15% 0.16% 0.15% 0.23 4-digit SIC 14.6 0.14% 0.14% 0.14% 0.30 3-digit SIC (CRSP) 20.1 0.16% 0.17% 0.16% 0.22 4-digit SIC (CRSP) 8.3 0.13% 0.13% 0.13% 0.35
46
Average Precision Score 501.2 520.9 549.6 544.5 516.5 513.7 491.3 511.0
Appendix to: Corporate Governance and Equity Prices: Are Results Robust to Industry Adjustments?
Stefan Lewellen Yale University
February 17, 2012
Appendix A: Description of the Random Sampling Methodology used by Johnson, Moorman & Sorescu (2009) Consider the GIM sample on a date corresponding to one of the four RiskMetrics portfolio “reset” dates. Let I denote a specific industry classification (such as a three-digit SIC code). Let N denote the set of firms in the event (Democracy or Dictatorship) portfolio with an industry code of I, and let K denote the set of firms in the non-event (non-Democracy or non-Dictatorship) portfolio with the same industry code. Finally, let wn and wk denote the market capitalizations of firms n and k, respectively, where n ∈ N and k ∈ K. Under the methodology utilized by JMS, the sample selection probability assigned to all non-event firms in industry I is computed as: PN
πI = Pn=1 K
wn
k=1 wk
.
(1)
In words, the sample selection probability for non-event firms in industry I equals the ratio of the total market capitalization in industry I within the event portfolio to the total market capitalization in industry I within the non-event portfolio. The expected market capitalization of industry I within each non-event portfolio thus matches the market capi-
1
talization of industry I within the event portfolio: E[WI ] ≡
K X
π I wk = π I
k=1
K X
PN wk =
k=1
wn Pn=1 K k=1 wk
!
K X k=1
wk =
N X
wn .
(2)
n=1
Once selection probabilities have been assigned to all firms, an i.i.d. sample of K random variables is generated from a U nif orm[0, 1] distribution. If the selection probability of firm i exceeds the randomly drawn number to which it is assigned, firm i will be included in the mimicking portfolio. This process is then repeated across all industries in the event portfolio. Selection probabilities can exceed unity using the method just described. To address this issue, JMS uniformly reduce the selection probability of all firms using a constant divisor of four. This reduces the size of each random sample by 75%, but simultaneously reduces the number of firms with π values greater than one. Using these reduced probabilities, JMS then impose a maximum π value of two and set π = 1 for all firms with a π value between one and two. Additional information about the random sampling methodology employed by JMS can be found in the Web Appendix to their paper, which is available on the Review of Financial Studies website.
Appendix B: Tradeoff Between Point Estimates and Standard Errors Null Hypothesis: α 6= 0 Consider an economy containing N firms that all reside in a single industry. All firms are assumed to have identical market capitalizations at each point in time. For now, I will assume that Democracies outperformed Dictatorships during my sample period. Specifically, I assume that each of the N firms in the economy has realized returns each period of: Ri,t = αi + δt + εi,t ,
(3)
where αi ∈ {g, −g, 0}, g is a positive constant, δt is a mean-zero industry return shock with finite variance, and εi,t ∼ i.i.d. N (0, σ 2 ) is a firm-specific return shock. Of the N firms in the sample, J ≥ 0 such firms are assumed to have an αi equal to g; these firms can be thought of as Democracies. Likewise, K ≥ 0 firms have αi = −g and can be thought of as Dictatorships.
2
Finally, the N − J − K ≥ 0 firms with αi = 0 are considered ‘governance-neutral’ firms.1 No other factors are presumed to affect realized returns. For brevity, time subscripts are dropped in what follows. I would like to determine whether a portfolio that is long Democracies and short Dictatorships earns positive industry-adjusted abnormal returns. I begin by constructing industryadjusted returns on the sample Democracies. When industry returns are computed using all firms in the universe of stocks, the industry return is simply Rind
N N 1 X J −K 1 X g+δ+ Ri = εi . = N i=1 N N i=1
(4)
Importantly, the industry return contains alpha as long as J 6= K. The industry-adjusted return on Democracy firm i is thus Ri − Rind
N J −K 1 X = 1− g + εi − εj . N N j=1
(5)
The first term of equation (5) shows that the industry-adjusted abnormal return will not be g. Thus, industry-adjusted abnormal return g – the correct value – but rather 1 − J−K N estimates will be biased when all firms in the sample are used to construct industry returns.2 Abnormal returns will be biased downward when Democracies are ‘clustered’ within industries (that is, when J > K), which is the case within the GIM governance sample. I do not consider sample Dictatorships or the long/short portfolio, as the logic is identical. Now consider a second method of constructing industry returns – specifically, the industry return methodology employed by JMS. Under the methodology employed by JMS, the industry return that is assigned to Democracies contains all firms in the sample which are not Democracies. Thus, the industry return that is matched with sample Democracies is Rind
N −J N −J −K 1 X 1 X Ri = εi . = g+δ+ N − J i=1 N −J N − J i=1
(6)
Note that the industry return now contains negative alpha as long as K > 0. The industry1
Here, ‘governance-neutral’ can have two meanings. First, these firms may have a neutral G-index rating. Second, these firms may not have a G-index rating at all; the assumption here is that firms that do not have a G-index rating are ‘governance-neutral’ on average. 2 It is important to note that this “bias” is not statistical in nature and does not influence the econometric implementation. The “bias” emerges from the manner in which I construct industries.
3
adjusted return on Democracy firm i is thus Ri − Rind
= 1+
K N −J
g + εi −
N −J 1 X εj . N − J j=1
(7)
Importantly, the industry-adjusted abnormal return in equation (7) now exceeds its correct value of g. Relative to equation (5), I would expect industry-adjusted portfolio returns computed according to equation (7) to produce higher abnormal return estimates but also higher standard errors. The increase in abnormal return estimates comes directly from a comparison of the first terms in equations (5) and (7). Standard errors are expected to be higher because a smaller number of firms are used to compute industry returns in equation (7) than equation (5). This results in a potentially higher level of idiosyncratic noise within the industryadjusted return in equation (7). A notable exception to the findings above arises when the sample firms are perfectly ‘clustered’ in specific industries – for example, when all Democracies reside in industry X and all Dictatorships reside in industry Y . Equation (7) shows that when K = 0, the methodology employed by JMS will yield the correct abnormal return estimate of g, while equation (5) shows that the industry-adjusted abnormal return will be biased downward when J > 0 and K = 0. Thus, if a data sample contains extreme levels of industry clustering, the methodology employed by JMS may be preferred to a methodology that computes industry returns using the universe of stocks. As noted by JMS, however, the industry clustering that is present within the GIM Democracy and Dictatorship portfolios is not so extreme that the portfolios are clustered within a handful of industries.
Null Hypothesis: α = 0 The previous example assumed that Democracy firms did, indeed, earn positive abnormal returns. In practice, our null hypothesis is normally that our sample portfolio earns zero abnormal returns. Under this null hypothesis, αi is assumed to equal zero for all firms in the sample, and equation (5) reduces to Ri − Rind
N 1 X εj , = εi − N j=1
4
(8)
while equation (7) reduces to Ri − Rind = εi −
N −J 1 X εj . N − J j=1
(9)
Thus, standard errors are likely to be lower when industry returns are computed using all firms in the sample (equation (8)). Since the null hypothesis in most empirical asset pricing tests is that α = 0, equations (8) and (9) imply that researchers should generally construct industries using the universe of stocks in order to minimize the amount of idiosyncratic noise contained within industry returns.
Appendix C: Size and Power Tests using Revised Sample and Industry Construction Methodology Size Tests My size tests in this appendix differ from the size tests reported in Table 2 along a number of (relatively minor) dimensions. First, while JMS form their random samples using nonDemocracies and non-Dictatorships, I form samples by drawing firms randomly from the entire universe of stocks. Second, rather than selecting firms using the probability weighting algorithm employed by JMS, I simply draw the same number of firms into each sample industry that exist in the true Democracy and Dictatorship portfolios. For example, if the Democracy portfolio contains five firms in industry I at time t, I randomly draw five firms from the universe of stocks at time t that are in industry I. Finally, the industry returns that are matched with each firm in the sample are computed using the universe of stocks. While this methodology does not attempt to directly mimic the market capitalization of each industry within the Democracy and Dictatorship portfolios, it has the virtue of being substantially easier to implement than the methodology employed by JMS. As in Table 2, I form 1,000 sample portfolios for each of the 24 combinations of industry classifications and asset pricing models listed in Table 5. I then regress value-weighted, industry-adjusted returns on each sample portfolio against an asset pricing model and evaluate the fraction of sample portfolios that reject the null hypothesis of zero abnormal returns. Table A.3 reports the results of these specification tests. As in Table 2, tests which are well-specified are shaded in gray. Table A.3 shows that at least one well-specified test exists for each of the industry clustering and industry adjustment methodologies I study. Further5
more, a number of the asset pricing tests that yield statistically significant industry-adjusted abnormal returns in Table 5 are also well specified in Table A.3. Like Table 2, Table A.3 also shows no obvious patterns related to industry granularity. However, test specifications that utilize the same industry classification standard to form sample portfolios and construct industry-adjusted returns tend to be well-specified, while tests which use different industry classification standards to form sample portfolios and construct industry-adjusted returns are generally poorly-specified. The former result accords with intuition – if samples are formed to mimic the FF48 clustering within the GIM governance portfolio, FF48-adjusted portfolio returns will likely be well-specified. The latter result highlights the danger of relying on a single industry classification standard to mimic the industry clustering present within the Democracy and Dictatorship portfolios.
Power Tests I also compute power tests to identify the specifications in Table A.3 that have the best power to detect abnormal returns. However, these power tests differ significantly from the power tests reported in Table 3. I begin by regressing the industry-adjusted returns on the 1,000 sample portfolios against the PR1YR, UMD, and FF3 asset pricing models without injecting any abnormal returns. I then compute the average abnormal return estimate across all 1,000 samples and subtract this value from each sample’s abnormal return. This standardization procedure ensures that the average abnormal return on each of the test specifications is exactly zero.3 After standardizing the abnormal return estimates across all 1,000 samples, I inject abnormal returns ranging from 0.10% to 1.50% per month into each of the sample Democracy and Dictatorship firms. Importantly, the sample Democracies and Dictatorships are also included in the industries that are used to compute industry-adjusted returns. Thus, to correctly approximate the power of each test specification, I also inject alpha into each industry return. The amount of alpha that I inject into each industry return is commensurate with the market capitalization weight of the sample Democracies (Dictatorships) within that industry.4 I then regress the returns on these portfolios against the PR1YR, UMD, and FF3 asset pricing models. Finally, I compute the percentage of sample portfolios that reject the null hypothesis of zero abnormal returns in the right tail at the 10% statistical significance 3 Even if a test specification is well-specified in Table A.3, some specifications reject the null hypothesis in the right tail more frequently than others. Without standardizing the abnormal return estimates, the test specifications that reject the null hypothesis more frequently in the right tail would gain an advantage in my power tests relative to other specifications. 4 For example, consider an industry that contains ten firms with a combined market capitalization of 100. Suppose that one of the ten firms with a market capitalization of 20 is included in the sample Democracy α 20 portfolio. The amount of alpha injected into the industry-adjusted return on this firm would be 1 − 100 2, where α is the stated alpha level.
6
level. A summary of the results of can be found in Table 6 of the paper. The full results are shown in Table A.4.
Additional Robustness Tests While the random samples in the previous section contain the same number of firms in each industry that exist within the GIM Democracy and Dictatorship portfolios, I do not attempt to match the market capitalization of each industry within the actual Democracy and Dictatorship portfolios. As such, one concern is that my sample portfolios do not adequately mimic the industry clustering that is present within the GIM Democracy and Dictatorship portfolios. I address this concern in two ways. First, I return to the sample portfolios from section 3 that were formed using the probability weighting algorithm employed by JMS. I then repeat the specification and power tests in Tables A.3 and A.4 using these sample portfolios.5 Untabulated results show that my results are qualitatively unchanged – a majority of the test specifications with the best econometric properties also yield statistically significant industry-adjusted abnormal returns on the GIM governance portfolio. I also find that test specifications involving six-digit GICS classifications have the best econometric properties among the specifications I consider. As a second robustness check, I return to the random samples used in Tables A.3 and A.4. As before, I begin by computing industry-adjusted returns on each firm in each sample. I then compute value-weighted returns on each industry in each sample. I next constrain the market capitalization weight on each industry to match the market capitalization weight of that industry within the true Democracy or Dictatorship portfolio. Using these revised weights, I then compute the portfolio return on the sample Democracy and Dictatorship portfolios. This methodology simply constrains the weight on each sample industry at each point in time to match its weight within the respective Democracy or Dictatorship portfolio at the same point in time. Untabulated results show that a majority of the test specifications with the best econometric properties also yield statistically significant industry-adjusted abnormal returns on the GIM governance portfolio. Again, specifications involving six-digit GICS classifications have the best econometric properties among the specifications I consider. 5
Unlike the tests in Tables 2 and 3, however, I compute industry returns using the universe of stocks.
7
Appendix D: Precision Tests using Revised Sample and Industry Construction Methodology In this section, I apply a slightly more formal version of the methodology introduced in section 3.4 to test the precision of the 24 asset pricing tests listed in Table 5. As in section 3.4, I consider a calendar-time regression of value-weighted, industry-adjusted portfolio returns against a set of asset pricing factors. The intercept coefficient from this regression has a t-statistic of t = ασ . When industry returns are computed using the universe of stocks (as in Table 5), an increase in the return on portfolio firm i will also increase the value-weighted market capi return on the industry associated with firm i by a factor of wi,ind ≡ market at each capindustry point in time. Aggregating across all stocks in my sample portfolio, the derivative of my t statistic with respect to a change in my portfolio return at a single point in time can be written as 1−λ ∂t = , (10) ∂Rport σ PN where λ ≡ i=1 wi ∗ wi,ind , N is the number of firms in the portfolio, wi represents the i market capitalization weight of firm i in the portfolio, and wi,ind ≡ PJmktcap is the mari j=1
mktcapj
ket capitalization weight of firm i within its industry (which contains a total of Ji firms, including firm i). Importantly, λ will change over time to reflect differences in the market capitalization weights of the portfolio firms and their respective industries. To estimate precision, I follow Metrick (1999) and simply average the values of λ across months to obtain a single estimate of λ for each test specification. Again, it is not clear how to compute a precision metric for a long/short portfolio. As such, I estimate precision separately for the Democracy and Dictatorship portfolios and form a combined ranking based on the sum of each specification’s precision ranking across the two tests. I also report λ and σ estimates separately. Table A.5 shows that six-digit GICS classifications again have the highest precision ranks, followed closely by four-digit SIC classifications. Table A.5 also shows that λ is increasing in industry granularity. In particular, narrowly defined classification standards have extremely large λ estimates. For example, four-digit SIC (CRSP) classifications have an average Democracy λ of 0.5, indicating that Democracies account for 50% of the total market capitalization within their respective 4-digit SIC (CRSP) industries. In contrast, Democracies only account for 14% of the market capitalization within FF48 industries. Table A.5 also shows that standard error estimates are generally decreasing in industry granularity. As industry granularity increases, however, standard errors do not 8
always fall enough to offset the corresponding increase in λ. For example, the test specification involving 4-digit SIC (CRSP) classifications and the UMD asset pricing model has one of the lowest standard error estimates in the Democracy column, but its extremely high λ estimate of 0.5 yields a precision metric near the very bottom of the list. As such, I find no clear relationship between industry granularity and the combined precision rankings listed in Table A.5. Two other comparisons between the precision tests in Tables 4 and A.5 are worth noting. First, the average precision in the Democracy column increases substantially between Table 4 and Table A.5 despite the fact that λ is not subtracted from the numerator in Table 4. Precisions in the Dictatorship column are similar across the two tables. Thus, I find that asset pricing tests appear to have better precision when all firms in the universe of stocks are used to form industries. Consistent with Guenther & Rosman (1994), Table A.5 also shows that SIC classifications sourced from Compustat yield industry-adjusted asset pricing tests with better precision than SIC classifications sourced from CRSP.
References Guenther, D. A. & Rosman, A. J. (1994), ‘Differences between COMPUSTAT and CRSP SIC Codes and Related Effects on Research’, Journal of Accounting and Economics 18, 115– 128. Johnson, S. A., Moorman, T. C. & Sorescu, S. (2009), ‘A Reexamination of Corporate Governance and Equity Prices’, Review of Financial Studies 22, 4753–4786. Metrick, A. (1999), ‘Performance Evaluation with Transactions Data: The Stock Selection of Investment Newsletters’, Journal of Finance 54, 1743–1775.
9
Table A.1 Power tests Please see Table 3 for a description of the power tests contained in this table. Industry classification used to form portfolios that mimic the industry clustering within the GIM (2003) Dem/Dict portfolios
FF48 (CRSP)
FF48 (Compustat)
6-digit GICS
8-digit GICS
Monthly alpha added to the portfolio 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50%
FF48 (CRSP) PR1YR UMD 10.4% 8.4% 17.0% 12.9% 24.2% 20.2% 33.8% 29.0% 44.7% 39.2% 55.0% 51.7% 62.9% 62.7% 71.2% 72.0% 79.9% 79.4% 85.0% 85.9% 90.5% 89.7% 93.0% 92.7% 95.1% 95.6% 96.7% 97.7% 98.1% 98.8% 22.2% 17.8% 29.5% 24.7% 38.7% 33.6% 48.0% 45.6% 57.5% 57.1% 65.2% 66.6% 72.8% 74.7% 79.7% 81.8% 85.8% 86.9% 90.7% 90.7% 92.8% 94.3% 95.3% 95.9% 97.3% 97.4% 98.3% 98.0% 98.5% 98.4% 19.5% 10.8% 28.8% 16.3% 37.8% 25.7% 46.6% 36.2% 56.2% 47.4% 65.0% 58.2% 73.9% 68.3% 79.8% 77.7% 86.3% 85.0% 91.4% 90.5% 94.5% 94.3% 95.8% 96.7% 97.9% 98.3% 98.3% 99.0% 99.1% 99.4% 31.0% 19.5% 42.2% 29.6% 53.7% 42.7% 65.7% 56.5% 75.8% 68.5% 82.9% 79.0% 88.4% 86.4% 92.1% 91.0% 94.1% 94.8% 96.4% 96.6% 97.7% 98.0% 98.5% 99.5% 99.3% 99.8% 99.5% 99.9% 99.7% 100.0%
FF3 14.1% 21.1% 28.3% 38.6% 48.9% 57.7% 66.2% 72.3% 80.3% 85.3% 90.0% 93.2% 94.7% 95.8% 97.8% 23.9% 32.0% 40.6% 50.0% 59.2% 66.6% 73.3% 78.9% 85.3% 90.2% 92.8% 95.4% 96.8% 98.1% 98.5% 23.1% 30.7% 39.4% 47.7% 57.0% 66.3% 73.2% 79.8% 85.3% 89.6% 93.2% 95.6% 96.9% 98.3% 98.8% 38.4% 48.8% 61.8% 71.5% 78.1% 85.1% 89.5% 92.7% 94.6% 96.3% 97.5% 98.2% 98.9% 99.4% 99.4%
FF48 (Compustat) PR1YR UMD FF3 17.2% 10.8% 19.3% 25.3% 18.1% 28.7% 34.8% 29.1% 39.0% 45.9% 40.2% 46.7% 55.6% 49.8% 56.6% 64.3% 62.2% 65.6% 71.9% 71.7% 74.8% 78.7% 79.9% 80.1% 84.9% 86.4% 85.7% 89.9% 89.9% 89.3% 92.5% 93.5% 93.0% 94.5% 95.1% 94.7% 96.4% 97.1% 96.2% 97.5% 98.6% 97.2% 98.3% 99.5% 98.5% 17.7% 12.5% 19.0% 25.4% 19.6% 28.0% 33.5% 29.1% 36.0% 42.4% 39.4% 44.1% 51.2% 50.6% 52.8% 61.0% 58.9% 62.1% 70.1% 69.5% 70.8% 76.7% 78.8% 76.4% 82.2% 84.6% 82.1% 86.7% 89.0% 85.9% 90.1% 92.5% 89.8% 93.1% 94.7% 92.2% 95.2% 96.4% 95.0% 96.4% 97.5% 96.1% 97.1% 97.9% 96.8% 19.5% 8.8% 20.6% 28.2% 14.7% 28.3% 37.8% 23.3% 38.8% 50.9% 35.1% 51.3% 60.6% 47.2% 60.2% 69.6% 59.8% 69.3% 77.9% 71.3% 77.0% 83.4% 79.4% 83.2% 89.3% 85.6% 87.8% 92.2% 91.8% 90.9% 94.9% 94.8% 93.8% 96.9% 97.2% 96.1% 98.2% 98.4% 97.0% 98.7% 99.0% 98.3% 99.4% 99.7% 98.9% 32.8% 14.5% 35.8% 45.3% 25.6% 45.8% 56.9% 39.5% 57.6% 66.9% 53.7% 67.9% 74.8% 65.9% 77.9% 84.4% 77.2% 84.0% 90.0% 85.2% 89.9% 93.7% 91.7% 93.0% 95.8% 95.1% 95.7% 96.8% 97.6% 96.8% 98.4% 98.6% 97.8% 99.2% 99.4% 98.4% 99.4% 99.9% 99.0% 99.5% 100.0% 99.5% 99.7% 100.0% 99.7%
6-digit GICS PR1YR UMD 22.9% 18.3% 31.0% 28.4% 41.3% 39.0% 53.0% 50.9% 62.4% 61.5% 71.6% 71.7% 78.0% 79.3% 84.8% 87.1% 89.9% 91.3% 93.4% 93.6% 95.1% 95.3% 96.4% 97.1% 96.8% 98.3% 97.8% 99.2% 98.6% 99.8% 26.7% 22.4% 35.1% 31.3% 44.2% 40.9% 55.6% 52.8% 63.5% 62.8% 71.9% 73.1% 79.0% 80.3% 84.6% 85.0% 87.2% 89.0% 91.3% 92.0% 93.9% 94.5% 95.8% 96.3% 96.8% 97.5% 97.6% 97.9% 98.0% 98.5% 17.3% 11.7% 26.4% 19.0% 37.7% 30.7% 50.1% 43.9% 61.1% 55.7% 70.0% 67.4% 77.7% 78.1% 84.7% 85.8% 89.8% 91.5% 93.7% 95.4% 96.4% 97.2% 97.6% 98.2% 98.2% 98.8% 98.8% 99.4% 99.2% 99.6% 40.1% 26.4% 50.7% 40.2% 62.4% 53.1% 73.5% 68.9% 82.8% 78.7% 89.2% 87.6% 93.0% 92.3% 96.0% 96.0% 97.7% 98.5% 99.1% 99.3% 99.5% 99.9% 99.8% 100.0% 99.9% 100.0% 99.9% 100.0% 99.9% 100.0%
Industry classification used to compute industry-adjusted returns 8-digit GICS 3-digit SIC 4-digit SIC PR1YR UMD FF3 PR1YR UMD FF3 PR1YR UMD 12.5% 10.8% 15.0% 5.1% 4.0% 6.6% 2.4% 2.2% 19.3% 16.0% 21.5% 9.8% 7.6% 10.6% 4.1% 3.2% 26.9% 22.4% 30.4% 14.5% 12.9% 17.1% 7.3% 5.1% 36.5% 33.6% 39.4% 21.3% 19.2% 24.7% 10.1% 8.5% 46.0% 43.8% 49.1% 31.9% 27.2% 34.2% 14.4% 14.4% 56.6% 55.6% 58.1% 42.0% 37.9% 43.6% 21.0% 20.4% 67.0% 64.7% 66.6% 53.2% 48.6% 55.0% 28.9% 27.9% 74.9% 74.1% 75.1% 63.1% 59.8% 64.5% 37.1% 35.9% 83.1% 81.7% 82.4% 72.1% 71.0% 72.9% 45.9% 44.8% 88.2% 88.8% 87.2% 79.5% 78.7% 81.3% 55.1% 54.4% 92.0% 91.5% 90.6% 85.7% 84.7% 86.4% 64.1% 62.2% 94.4% 93.8% 93.2% 89.4% 89.9% 90.1% 71.8% 70.4% 96.0% 96.4% 95.0% 93.6% 93.0% 93.5% 77.7% 76.2% 96.7% 97.9% 96.2% 95.8% 95.8% 95.4% 82.7% 82.1% 97.4% 98.5% 97.0% 97.2% 97.0% 96.8% 87.7% 86.9% 18.2% 14.0% 21.5% 10.8% 9.3% 12.1% 2.4% 2.2% 26.0% 23.4% 29.4% 17.5% 16.0% 20.4% 4.1% 3.2% 34.2% 32.6% 38.6% 25.7% 23.1% 27.4% 7.3% 5.1% 44.7% 42.1% 47.8% 34.9% 32.7% 34.8% 10.1% 8.5% 55.1% 51.9% 55.9% 42.9% 42.1% 44.9% 14.4% 14.4% 62.8% 62.3% 65.7% 52.7% 54.0% 56.1% 21.0% 20.4% 74.0% 73.0% 74.5% 63.5% 64.0% 65.7% 28.9% 27.9% 81.1% 80.2% 81.5% 71.6% 72.9% 73.9% 37.1% 35.9% 86.2% 84.8% 85.6% 79.6% 80.3% 80.3% 45.9% 44.8% 89.6% 89.2% 89.8% 85.6% 86.3% 86.1% 55.1% 54.4% 92.3% 92.2% 91.8% 89.1% 90.5% 89.8% 64.1% 62.2% 93.8% 94.8% 94.4% 92.4% 93.9% 92.8% 71.8% 70.4% 95.6% 96.5% 95.6% 95.0% 96.2% 94.9% 77.7% 76.2% 96.8% 97.3% 96.4% 97.2% 98.1% 96.4% 82.7% 82.1% 97.6% 97.5% 97.2% 98.3% 98.7% 97.8% 87.7% 86.9% 16.1% 11.7% 17.6% 10.7% 6.4% 10.5% 8.1% 5.5% 24.4% 18.0% 26.1% 16.5% 11.3% 15.2% 12.0% 9.4% 35.0% 27.5% 37.7% 23.7% 18.8% 23.1% 18.2% 12.7% 47.1% 39.4% 47.4% 33.5% 27.0% 31.2% 24.1% 18.7% 60.1% 54.5% 59.1% 44.2% 35.1% 42.9% 32.3% 26.7% 70.3% 66.0% 68.0% 55.2% 47.0% 52.5% 42.1% 35.7% 78.4% 77.2% 77.1% 65.1% 58.4% 61.8% 51.4% 45.4% 85.2% 84.0% 83.7% 74.0% 69.6% 70.5% 61.3% 55.6% 90.0% 89.0% 87.9% 82.3% 78.2% 77.9% 70.0% 64.9% 93.4% 93.6% 92.1% 87.4% 84.3% 84.0% 77.6% 72.5% 96.4% 96.6% 95.0% 91.5% 91.0% 89.3% 82.7% 79.9% 97.8% 97.8% 96.0% 95.1% 94.5% 92.5% 86.8% 86.3% 98.2% 98.6% 98.1% 97.3% 96.5% 95.4% 92.0% 90.2% 99.0% 99.5% 98.9% 98.1% 98.0% 97.1% 94.3% 93.2% 99.6% 99.5% 99.3% 98.7% 98.9% 98.2% 96.4% 95.8% 13.3% 6.5% 12.5% 17.0% 7.7% 18.5% 9.4% 4.8% 21.9% 14.0% 20.8% 25.9% 14.1% 26.0% 13.6% 9.1% 33.5% 22.8% 29.2% 36.2% 25.0% 37.1% 20.3% 14.1% 46.8% 34.6% 42.0% 47.9% 36.7% 47.3% 29.3% 21.0% 59.3% 49.0% 53.1% 60.1% 47.6% 58.4% 40.5% 28.9% 72.9% 64.8% 64.5% 70.2% 61.1% 66.5% 51.5% 40.6% 80.9% 78.6% 75.5% 78.1% 73.0% 74.9% 59.8% 52.1% 86.8% 85.6% 83.1% 84.5% 81.6% 82.8% 69.7% 61.2% 92.6% 91.7% 88.2% 89.8% 87.7% 88.5% 77.1% 71.4% 95.7% 95.7% 93.2% 94.4% 92.9% 92.8% 85.3% 80.8% 98.0% 97.4% 95.8% 96.6% 95.8% 95.6% 90.6% 86.5% 98.8% 99.0% 98.2% 97.8% 97.7% 96.6% 93.3% 90.8% 99.4% 99.6% 98.7% 98.7% 99.3% 98.1% 95.0% 94.7% 99.6% 99.8% 99.2% 99.3% 99.7% 98.6% 96.8% 96.4% 99.8% 99.9% 99.4% 99.6% 99.7% 99.5% 98.0% 97.6%
FF3 25.0% 34.8% 44.2% 53.1% 63.1% 72.5% 78.9% 84.6% 89.8% 92.6% 95.0% 96.0% 96.9% 97.6% 98.6% 28.5% 37.7% 47.5% 57.5% 65.8% 73.1% 80.1% 84.8% 88.4% 91.2% 93.7% 95.7% 97.0% 97.2% 98.0% 19.0% 26.9% 38.2% 50.9% 61.0% 68.5% 76.6% 83.3% 89.3% 92.8% 95.5% 97.0% 97.9% 98.2% 98.7% 45.6% 57.7% 69.5% 78.5% 86.9% 91.7% 95.0% 96.8% 98.6% 99.0% 99.3% 99.6% 99.6% 99.9% 99.9%
10
FF3 3.3% 4.7% 7.9% 11.0% 15.3% 22.8% 30.4% 38.9% 48.7% 56.8% 65.0% 72.2% 78.1% 83.4% 87.8% 3.3% 4.7% 7.9% 11.0% 15.3% 22.8% 30.4% 38.9% 48.7% 56.8% 65.0% 72.2% 78.1% 83.4% 87.8% 8.6% 13.5% 18.3% 24.2% 32.0% 41.6% 49.9% 59.0% 69.7% 76.4% 81.1% 87.0% 90.3% 93.2% 95.3% 10.4% 15.5% 24.2% 32.1% 42.1% 52.0% 60.9% 68.6% 76.7% 84.7% 88.9% 92.0% 94.2% 96.1% 97.5%
3-digit SIC (CRSP) PR1YR UMD FF3 10.1% 7.2% 13.3% 15.0% 11.0% 18.4% 21.4% 17.7% 24.8% 28.4% 24.1% 32.4% 36.7% 31.0% 40.4% 45.7% 40.2% 49.1% 53.9% 51.1% 57.0% 62.2% 61.2% 65.8% 71.6% 70.3% 72.3% 79.2% 78.3% 79.8% 85.1% 84.1% 85.3% 89.1% 88.8% 89.4% 91.9% 92.7% 92.1% 94.4% 95.5% 94.0% 96.6% 97.3% 95.9% 19.1% 16.2% 21.4% 26.3% 23.1% 28.3% 34.5% 30.7% 36.4% 43.1% 40.5% 45.5% 51.7% 48.6% 53.6% 58.8% 58.9% 60.6% 66.4% 66.1% 67.2% 72.6% 74.4% 72.9% 78.5% 79.7% 78.4% 82.6% 84.2% 83.3% 87.1% 87.7% 86.3% 90.0% 90.8% 89.8% 92.3% 92.4% 91.9% 93.4% 93.8% 93.2% 95.6% 95.1% 95.3% 12.4% 7.3% 13.5% 20.3% 11.0% 21.0% 26.4% 18.4% 28.1% 35.3% 25.4% 35.1% 44.7% 34.9% 44.3% 54.3% 46.8% 54.3% 63.2% 58.5% 61.9% 72.9% 68.8% 70.5% 79.2% 77.9% 77.1% 85.0% 84.4% 83.3% 88.4% 88.9% 86.9% 91.3% 91.9% 90.4% 94.1% 94.0% 92.2% 95.9% 95.8% 94.0% 96.9% 96.7% 95.4% 13.2% 5.7% 17.5% 19.7% 9.7% 24.3% 29.1% 16.9% 33.1% 38.8% 26.4% 43.0% 49.1% 35.1% 51.8% 60.3% 47.1% 62.7% 69.0% 59.1% 71.9% 77.3% 69.5% 78.5% 83.0% 77.4% 83.1% 87.8% 83.4% 87.1% 90.6% 87.6% 90.2% 92.4% 90.8% 91.6% 94.1% 93.5% 93.7% 95.7% 94.6% 95.1% 96.7% 95.4% 96.0%
4-digit SIC (CRSP) PR1YR UMD FF3 7.7% 6.1% 8.5% 10.6% 9.2% 13.3% 15.4% 12.2% 17.0% 20.8% 18.8% 23.2% 27.7% 24.9% 29.7% 34.6% 32.5% 38.4% 44.4% 40.0% 46.6% 51.7% 50.0% 53.7% 59.8% 58.2% 61.8% 66.3% 66.2% 67.4% 73.6% 72.6% 74.5% 79.6% 79.4% 79.7% 83.8% 84.8% 84.5% 87.5% 88.4% 87.6% 90.3% 91.1% 90.6% 13.5% 11.4% 14.2% 17.9% 16.8% 18.3% 24.0% 22.4% 24.3% 29.9% 29.6% 30.4% 36.6% 38.0% 38.6% 44.6% 46.4% 46.6% 52.2% 54.6% 53.0% 60.9% 63.3% 60.4% 66.9% 71.0% 66.6% 71.9% 77.0% 73.0% 78.6% 81.8% 77.9% 82.8% 85.3% 82.0% 85.9% 88.4% 85.4% 88.7% 91.3% 87.5% 91.3% 92.7% 90.5% 10.8% 7.2% 11.4% 14.9% 11.5% 15.9% 22.0% 18.7% 21.2% 29.5% 26.2% 29.8% 39.3% 34.3% 39.2% 47.4% 44.1% 47.1% 57.0% 55.4% 54.8% 65.5% 63.1% 64.5% 73.1% 71.1% 72.1% 79.0% 78.7% 78.2% 84.8% 84.8% 83.0% 89.0% 88.6% 87.8% 91.7% 91.5% 90.0% 93.6% 94.2% 92.9% 95.6% 96.0% 94.6% 12.9% 9.0% 14.3% 19.4% 14.1% 20.8% 27.3% 20.3% 25.8% 34.4% 28.5% 35.4% 44.7% 38.5% 43.0% 54.8% 47.4% 52.6% 63.9% 55.9% 60.5% 70.1% 65.5% 68.8% 76.0% 74.3% 75.2% 82.4% 80.5% 81.1% 85.8% 85.3% 85.5% 88.8% 88.4% 88.4% 91.1% 90.7% 91.0% 93.3% 92.3% 92.4% 94.3% 93.8% 94.1%
Table A.1 (Continued) Power tests Industry classification used to form portfolios that mimic the industry clustering within the GIM (2003) Dem/Dict portfolios
3-digit SIC
4-digit SIC
3-digit SIC (CRSP)
4-digit SIC (CRSP)
Monthly alpha added to the portfolio 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50%
FF48 (CRSP) PR1YR UMD 29.4% 18.8% 38.5% 27.9% 48.5% 37.4% 59.3% 49.1% 68.8% 60.8% 78.5% 72.4% 84.9% 81.9% 88.6% 87.9% 93.1% 92.3% 95.3% 94.9% 97.0% 96.7% 98.0% 98.2% 99.3% 99.3% 99.7% 99.7% 99.9% 100.0% 8.4% 4.8% 15.0% 9.8% 21.0% 17.2% 32.1% 26.7% 43.8% 36.9% 54.5% 50.4% 65.5% 62.1% 74.3% 73.1% 82.3% 81.5% 88.8% 88.6% 93.2% 93.1% 96.5% 96.4% 98.1% 97.9% 98.4% 98.7% 99.0% 99.3% 10.9% 5.5% 16.3% 9.5% 24.4% 15.1% 32.4% 23.8% 42.8% 33.4% 53.5% 43.8% 63.4% 56.1% 73.3% 66.9% 81.4% 76.4% 87.6% 85.2% 92.9% 91.7% 95.9% 95.3% 98.3% 97.0% 98.9% 98.9% 99.1% 99.2% 8.7% 5.8% 14.4% 8.6% 19.7% 12.6% 27.0% 20.0% 36.3% 29.8% 46.5% 38.7% 56.5% 48.5% 65.4% 59.3% 74.2% 68.8% 81.9% 77.4% 86.8% 84.1% 90.7% 90.6% 94.6% 93.8% 96.0% 96.7% 98.3% 98.2%
FF3 31.3% 40.3% 48.9% 60.0% 69.5% 77.7% 84.9% 88.6% 92.8% 95.2% 96.6% 98.1% 98.9% 99.4% 99.8% 11.4% 17.3% 25.8% 35.2% 46.4% 59.6% 68.5% 76.4% 82.4% 89.0% 92.8% 96.0% 97.8% 98.4% 98.9% 15.1% 21.9% 29.8% 38.0% 49.2% 57.2% 67.5% 75.9% 82.4% 88.6% 93.2% 96.1% 97.9% 98.8% 99.3% 10.0% 14.5% 20.3% 27.2% 36.1% 46.5% 56.1% 64.8% 72.9% 79.0% 85.5% 89.7% 92.8% 95.5% 96.7%
FF48 (Compustat) PR1YR UMD FF3 37.0% 24.1% 35.7% 47.4% 34.9% 46.9% 57.7% 45.6% 57.3% 67.6% 57.4% 67.4% 75.2% 67.3% 75.7% 82.0% 76.8% 81.7% 87.6% 84.5% 87.4% 90.8% 88.5% 90.7% 93.3% 93.3% 94.5% 96.7% 95.9% 96.1% 97.8% 97.8% 97.7% 99.1% 98.9% 99.0% 99.5% 99.5% 99.5% 99.7% 99.7% 99.7% 99.8% 100.0% 99.8% 15.1% 7.3% 15.5% 24.3% 14.0% 24.0% 34.2% 23.1% 34.5% 45.8% 34.5% 47.3% 56.9% 46.0% 58.0% 67.6% 58.5% 68.8% 74.9% 70.8% 76.0% 82.3% 79.6% 81.6% 87.5% 85.8% 87.3% 91.6% 90.2% 92.1% 94.2% 93.5% 94.9% 96.1% 96.1% 96.1% 97.2% 97.4% 97.6% 98.3% 98.9% 97.9% 99.0% 99.6% 99.1% 9.8% 4.9% 11.8% 15.6% 9.7% 17.6% 23.6% 17.1% 25.8% 32.9% 25.0% 34.9% 44.2% 34.8% 45.2% 54.0% 45.2% 56.0% 64.7% 57.8% 67.0% 72.7% 69.0% 75.4% 80.2% 79.6% 83.3% 88.4% 85.8% 89.0% 92.5% 90.6% 93.1% 94.9% 94.8% 95.0% 96.3% 96.0% 96.6% 97.4% 97.5% 97.9% 98.0% 98.6% 98.0% 11.6% 6.5% 10.4% 17.0% 10.7% 15.5% 24.5% 16.4% 21.9% 35.0% 24.4% 30.6% 44.7% 34.4% 40.8% 54.9% 45.5% 50.3% 64.8% 56.9% 60.8% 73.2% 66.1% 68.8% 79.7% 76.0% 75.7% 85.4% 84.6% 82.4% 90.3% 90.1% 87.7% 93.3% 93.9% 91.9% 95.8% 97.1% 95.0% 97.8% 98.2% 96.8% 98.9% 99.3% 98.2%
Industry classification used to compute industry-adjusted returns 6-digit GICS 8-digit GICS 3-digit SIC 4-digit SIC PR1YR UMD FF3 PR1YR UMD FF3 PR1YR UMD FF3 PR1YR UMD 43.4% 31.6% 47.8% 21.3% 13.6% 24.6% 22.3% 16.8% 22.0% 12.5% 11.2% 53.6% 43.2% 59.3% 31.8% 20.8% 36.1% 29.1% 24.4% 28.7% 18.8% 16.5% 65.2% 54.7% 68.0% 44.2% 29.3% 47.5% 38.9% 33.6% 37.6% 26.1% 22.5% 73.9% 66.5% 76.9% 55.4% 42.7% 56.9% 48.3% 44.9% 49.5% 33.9% 30.6% 81.0% 76.0% 83.7% 66.1% 55.2% 67.0% 60.6% 55.7% 60.7% 43.1% 38.4% 86.8% 83.3% 88.4% 75.0% 67.3% 75.3% 70.2% 65.4% 70.3% 52.1% 48.1% 91.3% 89.7% 92.6% 81.4% 77.8% 82.2% 79.1% 75.3% 77.9% 60.2% 57.9% 94.8% 93.4% 95.4% 87.6% 85.5% 86.8% 84.1% 83.0% 84.3% 67.5% 65.6% 97.2% 96.5% 97.4% 91.7% 89.5% 91.5% 89.9% 88.7% 89.5% 74.4% 72.7% 98.4% 98.3% 98.7% 94.9% 93.6% 93.9% 93.1% 92.7% 93.3% 81.2% 78.7% 99.1% 99.3% 99.1% 96.1% 96.2% 96.1% 95.7% 95.7% 96.1% 85.9% 83.7% 99.3% 99.5% 99.5% 97.2% 97.8% 96.9% 97.6% 97.6% 97.3% 89.9% 88.5% 99.8% 99.6% 99.8% 98.3% 98.6% 98.1% 98.6% 98.9% 98.5% 92.6% 91.8% 99.9% 99.8% 99.9% 99.1% 99.2% 98.8% 99.0% 99.2% 98.9% 95.4% 94.7% 99.9% 100.0% 99.9% 99.5% 99.6% 99.3% 99.5% 99.8% 99.5% 96.7% 97.0% 24.0% 12.0% 31.1% 4.2% 0.9% 6.6% 3.2% 2.1% 3.8% 3.1% 2.9% 35.7% 22.1% 43.4% 9.0% 3.4% 13.4% 6.5% 4.9% 7.8% 5.5% 4.6% 46.5% 35.5% 54.0% 16.1% 7.6% 20.5% 11.6% 9.0% 12.4% 8.9% 7.7% 57.2% 48.2% 64.8% 27.5% 14.4% 30.7% 20.8% 16.1% 22.1% 14.5% 13.0% 68.8% 59.9% 73.7% 38.9% 26.7% 42.9% 29.8% 27.8% 31.6% 20.2% 20.2% 77.4% 69.9% 82.3% 49.9% 38.2% 53.5% 40.6% 38.0% 43.1% 27.9% 27.6% 85.2% 79.2% 88.4% 62.3% 49.6% 63.6% 52.1% 47.8% 52.6% 39.6% 38.5% 89.0% 86.7% 92.9% 72.9% 62.4% 74.1% 61.4% 61.3% 65.2% 48.0% 48.7% 93.6% 91.3% 96.5% 82.0% 74.5% 84.6% 74.0% 72.9% 74.3% 60.5% 59.7% 96.7% 95.4% 98.7% 89.2% 82.4% 89.6% 82.1% 82.5% 82.1% 70.0% 69.5% 98.7% 97.9% 99.2% 94.0% 88.7% 93.3% 88.6% 88.8% 87.9% 77.5% 76.2% 99.5% 99.2% 99.6% 96.7% 93.8% 96.5% 92.3% 93.3% 92.2% 84.3% 84.0% 99.9% 99.7% 100.0% 98.0% 97.1% 98.4% 95.5% 96.1% 95.7% 89.2% 89.0% 100.0% 99.9% 100.0% 99.2% 98.8% 99.3% 96.9% 97.2% 97.4% 93.2% 92.7% 100.0% 100.0% 100.0% 99.6% 99.4% 99.6% 98.0% 98.8% 98.7% 95.5% 95.3% 21.1% 13.3% 23.7% 8.9% 4.3% 13.4% 4.0% 2.4% 4.3% 2.1% 1.7% 30.0% 21.5% 32.3% 14.0% 8.2% 18.0% 6.5% 4.1% 8.4% 3.4% 2.7% 39.7% 33.2% 42.4% 21.5% 13.5% 26.0% 10.8% 8.5% 13.0% 6.1% 4.9% 50.5% 43.1% 54.2% 30.5% 23.5% 35.0% 17.5% 12.8% 18.6% 9.3% 6.5% 61.6% 56.9% 64.6% 41.7% 34.2% 45.2% 25.3% 20.6% 25.4% 12.8% 10.6% 72.4% 66.7% 72.4% 51.3% 45.6% 55.7% 34.9% 29.0% 35.7% 18.9% 16.2% 79.8% 76.8% 81.5% 62.5% 58.4% 65.1% 45.5% 40.6% 46.2% 26.0% 23.5% 86.9% 86.2% 87.6% 73.0% 69.3% 73.7% 56.5% 51.3% 58.4% 35.2% 31.4% 92.3% 91.7% 92.5% 81.6% 80.2% 81.9% 68.8% 62.7% 69.8% 44.0% 40.2% 95.6% 96.0% 95.4% 88.5% 87.1% 88.1% 77.1% 72.8% 78.4% 54.0% 51.0% 97.1% 97.3% 96.9% 92.6% 92.3% 92.2% 84.6% 82.2% 85.4% 63.4% 59.7% 97.9% 98.8% 98.4% 95.3% 95.3% 94.4% 90.3% 89.3% 89.0% 69.7% 67.5% 98.9% 99.3% 99.1% 97.5% 98.4% 96.7% 93.5% 93.5% 93.2% 77.6% 75.8% 99.9% 99.7% 99.7% 98.8% 98.8% 98.4% 96.2% 95.3% 95.7% 82.9% 81.6% 100.0% 100.0% 99.9% 99.6% 99.2% 99.0% 97.3% 97.2% 97.4% 88.4% 86.0% 21.9% 16.6% 21.4% 11.5% 6.5% 10.7% 4.3% 3.7% 4.6% 1.7% 1.5% 30.8% 23.4% 28.0% 16.8% 12.3% 16.4% 7.6% 6.1% 6.6% 3.2% 3.5% 39.1% 33.3% 37.1% 24.2% 19.4% 24.3% 12.5% 10.4% 11.5% 6.0% 4.9% 50.3% 44.8% 46.5% 33.2% 29.2% 32.1% 19.6% 16.7% 18.0% 10.4% 8.4% 60.0% 55.0% 56.6% 42.9% 38.5% 42.1% 26.6% 24.1% 26.4% 14.0% 12.8% 69.2% 64.9% 67.0% 55.4% 49.5% 52.8% 36.4% 31.9% 34.6% 19.5% 18.6% 78.0% 75.1% 75.4% 65.4% 60.2% 63.2% 45.3% 44.9% 45.3% 28.0% 25.7% 85.4% 83.3% 83.1% 74.1% 71.4% 71.7% 57.0% 54.3% 53.9% 37.7% 34.6% 90.3% 89.8% 88.9% 83.0% 82.4% 80.2% 65.9% 65.5% 64.3% 44.4% 43.2% 94.4% 93.7% 93.2% 88.5% 88.5% 88.1% 76.2% 75.6% 73.6% 54.7% 52.6% 96.7% 96.9% 95.2% 93.3% 93.2% 91.6% 82.3% 83.8% 80.7% 64.1% 62.5% 98.1% 98.3% 97.5% 96.1% 96.1% 95.2% 87.8% 88.7% 86.9% 72.5% 69.9% 99.0% 98.9% 98.4% 97.9% 98.5% 97.4% 92.3% 93.1% 90.7% 79.2% 77.3% 99.2% 99.5% 99.2% 98.8% 99.0% 98.2% 94.7% 95.4% 93.2% 84.0% 84.6% 99.5% 99.9% 99.5% 99.4% 99.5% 99.0% 96.0% 96.5% 95.7% 88.2% 89.6%
11
FF3 13.9% 19.6% 25.8% 34.5% 43.1% 52.0% 59.9% 67.8% 74.1% 80.4% 85.6% 89.0% 91.8% 95.4% 97.3% 3.0% 5.7% 9.2% 14.8% 20.0% 30.1% 39.8% 51.1% 62.5% 71.2% 77.7% 85.3% 90.5% 93.9% 95.8% 2.3% 3.8% 6.3% 9.5% 14.6% 21.1% 29.9% 37.6% 45.7% 55.8% 62.2% 73.0% 79.3% 85.5% 89.6% 1.7% 2.9% 6.0% 9.0% 14.0% 19.7% 27.8% 35.6% 44.7% 53.7% 62.6% 70.4% 76.5% 83.3% 87.8%
3-digit SIC (CRSP) PR1YR UMD FF3 32.1% 25.3% 34.0% 39.8% 32.7% 42.9% 48.9% 41.5% 52.4% 59.9% 52.9% 64.3% 71.5% 64.8% 72.8% 79.6% 74.3% 80.0% 84.7% 82.6% 86.6% 90.6% 88.0% 91.0% 93.0% 91.9% 93.2% 95.3% 95.5% 96.4% 97.5% 97.5% 97.8% 98.6% 99.0% 98.7% 99.1% 99.3% 99.0% 99.5% 99.5% 99.3% 99.7% 99.8% 99.7% 12.2% 12.0% 14.2% 19.1% 18.9% 21.1% 27.2% 28.0% 28.6% 36.7% 40.1% 38.3% 47.8% 50.8% 51.1% 59.0% 61.2% 62.7% 68.6% 71.3% 71.5% 77.5% 80.3% 79.1% 83.7% 87.0% 84.7% 89.7% 92.0% 89.6% 93.2% 95.2% 94.1% 95.8% 97.0% 96.7% 97.7% 98.2% 98.1% 98.7% 99.3% 98.8% 99.1% 99.8% 99.3% 8.8% 6.6% 11.9% 14.1% 10.2% 17.1% 21.2% 16.0% 24.2% 28.5% 23.0% 33.9% 39.2% 32.6% 43.3% 50.7% 46.1% 54.2% 61.0% 57.3% 64.6% 71.4% 67.8% 72.0% 78.6% 79.3% 80.4% 85.9% 87.1% 85.9% 90.2% 91.4% 91.3% 94.5% 94.1% 95.1% 96.6% 97.1% 96.9% 98.0% 98.7% 97.9% 99.0% 99.4% 98.6% 7.0% 5.5% 7.3% 11.6% 8.3% 11.4% 18.1% 12.9% 17.5% 24.1% 20.8% 24.1% 31.0% 28.8% 30.6% 42.0% 37.0% 40.0% 52.0% 47.3% 49.6% 60.4% 59.3% 59.5% 70.5% 67.9% 68.0% 79.4% 76.5% 75.7% 86.4% 84.9% 83.4% 91.0% 90.6% 89.2% 94.1% 94.1% 92.8% 96.3% 97.0% 96.1% 97.7% 98.8% 97.1%
4-digit SIC (CRSP) PR1YR UMD FF3 35.1% 29.5% 34.4% 45.8% 37.5% 42.6% 55.3% 47.7% 52.7% 63.4% 57.8% 61.3% 70.4% 65.8% 68.8% 77.7% 74.6% 76.4% 83.5% 81.9% 81.3% 87.7% 86.6% 85.8% 91.0% 91.1% 90.4% 93.5% 94.9% 93.4% 95.3% 96.5% 95.8% 97.5% 97.7% 97.3% 98.5% 98.4% 97.9% 99.1% 99.2% 98.6% 99.7% 99.7% 99.5% 16.6% 16.9% 16.1% 23.4% 24.0% 23.7% 31.7% 33.6% 33.1% 42.4% 45.9% 41.9% 54.7% 57.8% 53.7% 64.4% 67.1% 63.1% 72.0% 77.2% 71.5% 79.8% 83.1% 78.8% 85.6% 88.9% 85.6% 90.2% 92.5% 90.3% 93.5% 95.4% 94.0% 95.9% 97.7% 96.3% 97.1% 98.8% 97.3% 98.0% 99.3% 98.1% 98.7% 99.5% 98.5% 8.2% 5.4% 9.8% 12.3% 8.5% 15.0% 19.2% 12.4% 20.6% 26.1% 18.6% 28.1% 34.1% 25.9% 35.8% 44.1% 34.6% 43.8% 52.9% 44.2% 53.9% 62.6% 54.8% 62.7% 69.1% 63.6% 70.4% 76.1% 71.9% 76.8% 83.7% 78.7% 82.7% 88.3% 84.7% 87.9% 92.5% 90.4% 91.8% 95.2% 93.8% 94.1% 96.9% 96.0% 96.5% 8.7% 7.3% 11.0% 13.3% 11.3% 16.7% 19.7% 15.5% 23.1% 26.6% 21.5% 30.0% 34.6% 28.0% 37.3% 42.8% 38.3% 46.6% 52.3% 46.8% 56.4% 62.8% 58.0% 66.4% 70.5% 67.1% 73.3% 77.5% 74.9% 80.0% 85.0% 82.3% 85.6% 89.6% 87.1% 90.2% 93.5% 91.3% 94.9% 95.7% 94.8% 97.2% 97.4% 96.8% 98.2%
Table A.2 Industry Size This table examines the size of the industry classifications that are matched with GIM Democracies and Dictatorships in my industry-adjusted return tests on the four portfolio rebalancing dates (September 1990, July 1993, July 1995, and February 1998). I start by creating a list of all Democracies and Dictatorships from the 1990, 1993, 1995, and 1998 RiskMetrics samples. For each Democracy and Dictatorship, I then calculate the size of the industry that is matched with that firm based on a variety of industry classification standards listed in the first column of the table. Panel A shows the cumulative percentage of industry classifications containing up to 50 firms based on the industry construction process employed by JMS. Panel B shows the cumulative percentage of industry classifications containing 0 to 50 firms based on an industry construction process that uses all firms in the merged CRSP/Compustat database to form industries. The column labeled "0 firms" lists the fraction of firms for which I am able to assign an industry but for which no other firms exist within that industry that can be used to construct industry-adjusted returns. Panel A: Distribution of the number of firms contained within each industry classification when industry returns are computed using nonDemocracies and non-Dictatorships Industry classification used to compute industry returns FF48 (CRSP) FF48 6-digit GICS 8-digit GICS 3-digit SIC 4-digit SIC 3-digit SIC (CRSP) 4-digit SIC (CRSP)
0 firms 0.0% 0.0% 0.5% 2.2% 4.9% 9.0% 7.8% 22.2%
≤ 1 firm 0.1% 1.1% 2.2% 4.2% 12.8% 21.6% 15.4% 35.8%
Percentage of industry classifications containing: ≤ 2 firms ≤ 3 firms ≤ 4 firms ≤ 5 firms ≤ 10 firms ≤ 15 firms ≤ 20 firms ≤ 50 firms 0.7% 1.8% 3.1% 3.9% 7.2% 13.3% 17.8% 61.8% 1.8% 2.0% 2.4% 2.7% 7.1% 13.7% 22.4% 63.6% 3.1% 4.2% 6.1% 7.6% 12.4% 16.6% 29.1% 83.0% 6.2% 9.0% 13.0% 17.5% 38.8% 55.9% 68.0% 91.7% 20.1% 25.9% 31.9% 37.3% 56.4% 65.6% 70.4% 94.8% 31.7% 40.9% 48.5% 54.2% 71.4% 81.2% 85.0% 99.5% 22.5% 29.3% 35.1% 41.2% 56.3% 63.3% 71.7% 86.1% 46.1% 52.5% 58.4% 62.8% 72.2% 82.0% 86.1% 92.6%
12
Table A.2 (Continued) Industry Size
Panel B: Distribution of the number of firms contained within each industry classification when industry returns are computed using all firms in the merged CRSP/Compustat database Industry classification used to compute industry returns FF48 (CRSP) FF48 6-digit GICS 8-digit GICS 3-digit SIC 4-digit SIC 3-digit SIC (CRSP) 4-digit SIC (CRSP)
0 firms 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
≤ 1 firm 0.0% 0.0% 0.0% 0.0% 0.3% 0.4% 0.5% 7.0%
Percentage of industry classifications containing: ≤ 2 firms ≤ 3 firms ≤ 4 firms ≤ 5 firms ≤ 10 firms ≤ 15 firms ≤ 20 firms ≤ 50 firms 0.0% 0.0% 0.0% 0.0% 0.3% 1.2% 1.7% 4.2% 0.0% 0.0% 0.0% 0.0% 0.1% 0.4% 1.7% 5.4% 0.0% 0.0% 0.0% 0.0% 0.1% 2.0% 3.9% 13.1% 0.0% 0.1% 0.9% 1.2% 2.4% 8.0% 12.3% 43.4% 0.9% 2.5% 3.7% 5.2% 13.2% 22.8% 33.8% 60.0% 1.3% 3.6% 6.4% 10.8% 27.8% 39.0% 50.5% 76.2% 1.3% 3.2% 5.3% 6.8% 14.7% 24.2% 30.8% 55.2% 13.3% 20.8% 25.5% 29.2% 47.1% 54.7% 59.0% 75.6%
13
Table A.3 Alternative Specification Tests Involving Long/Short Portfolios Designed to Mimic the Industry Clustering Present in the Democracy and Dictatorship Portfolios of Gompers, Ishii & Metrick (2003) This table contains the results of specification tests using randomly constructed portfolios that are designed to mimic the industry clustering present in the Gompers, Ishii & Metrick (2003) Democracy and Dictatorship portfolios. Sample portfolios are designed such that the number of firms within each sample industry matches the number of firms within that industry in the corresponding Democracy or Dictatorship portfolio. Sample firms are randomly drawn from the merged CRSP/Compustat database such that all firms in a given industry have the same probability of being included in each sample. After forming samples, I next compute industry-adjusted returns on each sample Democracy and sample Dictatorship. Industry returns are computed using all firms in the merged CRSP/Compustat database. The industry return is then subtracted from the return on the parent firm to arrive at an industry-adjusted return. The industry-adjusted returns on each firm are then value-weighted to form a return on each industry-adjusted sample Democracy or Dictatorship portfolio. A zero-cost portfolio is then formed that takes a long position in the (industry-adjusted) sample Democracy portfolio and a short position in the sample Dictatorship portfolio. Returns on this portfolio are then regressed against an asset pricing model. This procedure is employed 1,000 times, for a total of 1,000 sample Democracy portfolios, 1,000 sample Dictatorship portfolios, 1,000 zero-cost portfolios, and 1,000 regressions. The sample period is September 1990 to December 1999. Monthly stock returns are sourced from CRSP. The null hypothesis for all tests is that the regression intercept equals zero. *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels using a one-sided binomial test statistic. The table lists the percentage of 1,000 tests that reject the null hypothesis at the relevant statistical significance level. Tests which are highlighted in gray are "well-specified" - they do not reject the null hypothesis of zero abnormal returns too frequently at any significance level in either tail relative to appropriate theoretical rejection rates.
14
Table A.3 (Continued) Alternative Specification Tests Involving Long/Short Portfolios Designed to Mimic the Industry Clustering Present in the Democracy and Dictatorship Portfolios of Gompers, Ishii & Metrick (2003)
Asset Industry classification used to compute industry- Pricing Model adjusted returns PR1YR FF48 (CRSP) UMD FF3 PR1YR FF48 (Compustat) UMD FF3 PR1YR 6-digit GICS UMD FF3 PR1YR 8-digit GICS UMD FF3 PR1YR 3-digit SIC UMD FF3 PR1YR 4-digit SIC UMD FF3 PR1YR 3-digit SIC (CRSP) UMD FF3 PR1YR 4-digit SIC (CRSP) UMD FF3
Classification used to form portfolios that mimic the industry clustering in the Dem. / Dict. portfolios Numbers represent rejection rates at the 10% statistical significance level in the right tail FF48 FF48 6-digit 8-digit 3-digit 4-digit 3-digit 4-digit (CRSP) (Compustat) GICS GICS SIC SIC SIC (CRSP) SIC (CRSP) 3.9% 8.2% 11.2% 8.2% 10.6% 11.6% 8.2% 3.3% 3.9% 8.0% 7.6% 8.0% 10.4% 11.3% 7.1% 2.7% 3.9% 8.7% 9.4% 6.4% 10.8% 9.3% 7.5% 2.8% 5.9% 5.0% 10.8% 7.2% 12.2% 12.1% 14.5% 6.8% 5.8% 5.2% 6.5% 5.2% 10.0% 10.6% 11.7% 5.4% 6.3% 4.3% 9.4% 5.5% 12.0% 11.0% 13.3% 5.0% 3.6% 5.8% 4.2% 5.5% 9.3% 13.0% 7.0% 0.7% 3.3% 5.3% 4.2% 5.0% 9.3% 11.8% 8.2% 1.2% 4.6% 5.7% 5.1% 4.3% 11.2% 13.0% 5.8% 1.2% 4.3% 6.1% 3.7% 5.0% 11.1% 9.4% 5.4% 0.7% 3.9% 6.0% 4.2% 4.8% 9.9% 8.9% 5.9% 1.2% 4.4% 6.3% 3.9% 5.7% 11.2% 7.5% 5.3% 0.8% 7.7% 6.0% 6.4% 6.6% 4.8% 6.9% 5.9% 2.6% 6.7% 5.7% 5.1% 6.4% 4.0% 5.8% 5.0% 1.4% 7.2% 5.3% 5.0% 5.1% 4.5% 6.0% 6.0% 2.2% 6.3% 5.0% 5.1% 5.1% 5.0% 4.4% 4.8% 0.9% 6.2% 5.1% 4.3% 5.2% 4.8% 3.6% 4.0% 0.8% 6.4% 4.2% 4.3% 4.8% 4.8% 4.0% 5.0% 0.8% 6.7% 6.7% 5.8% 7.6% 5.4% 5.5% 5.8% 1.7% 5.8% 6.4% 5.9% 7.5% 5.5% 5.4% 5.0% 1.5% 6.4% 6.6% 5.5% 5.2% 5.3% 4.7% 5.6% 1.8% 4.7% 5.0% 6.3% 5.9% 6.1% 4.9% 5.1% 4.4% 4.6% 5.6% 5.8% 6.9% 6.8% 6.8% 5.4% 4.2% 4.6% 5.4% 5.7% 3.9% 4.8% 3.7% 4.3% 4.2%
15
Table A.4 Power Tests Please see Table 6 for a description of the power tests contained in this table. Industry classification used to form portfolios that mimic the industry clustering within the GIM (2003) Dem/Dict portfolios
FF48 (CRSP)
FF48 (Compustat)
6-digit GICS
8-digit GICS
Monthly alpha added to the portfolio 0.00% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50% 0.00% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50% 0.00% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50% 0.00% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50%
FF48 (CRSP) PR1YR UMD FF3 10.6% 11.2% 10.5% 17.1% 15.5% 16.1% 24.7% 23.5% 25.2% 34.6% 32.2% 35.8% 44.9% 43.0% 47.0% 56.7% 56.3% 57.7% 67.3% 65.7% 67.4% 77.2% 76.0% 79.3% 85.7% 84.1% 86.2% 91.6% 90.5% 91.8% 94.7% 94.4% 95.5% 97.0% 96.9% 98.3% 99.1% 98.2% 99.1% 99.6% 99.3% 99.7% 99.7% 99.7% 99.8% 100.0% 100.0% 100.0% 10.7% 10.2% 10.2% 15.5% 16.4% 16.7% 23.8% 23.7% 24.7% 35.5% 33.6% 35.7% 46.1% 43.3% 46.9% 58.1% 56.2% 57.5% 67.6% 66.2% 69.1% 77.1% 76.2% 79.8% 85.9% 82.9% 88.0% 92.4% 90.2% 93.5% 96.0% 94.4% 96.7% 97.7% 97.0% 98.0% 98.8% 98.9% 99.1% 99.2% 99.2% 99.4% 99.5% 99.5% 99.7% 99.8% 99.7% 99.9% 10.5% 10.2% 10.5% 16.9% 16.8% 16.8% 25.2% 24.1% 25.5% 34.9% 33.7% 35.0% 45.2% 43.2% 45.6% 54.7% 52.1% 57.5% 65.4% 63.3% 68.1% 75.4% 73.7% 77.9% 83.3% 82.1% 85.4% 90.6% 89.4% 91.0% 94.8% 93.8% 95.0% 97.0% 96.2% 96.9% 97.9% 97.8% 98.5% 99.1% 98.7% 99.1% 99.4% 99.6% 99.5% 99.7% 99.8% 99.6% 9.9% 10.2% 9.7% 16.2% 15.1% 14.7% 23.5% 21.9% 24.2% 33.9% 31.8% 35.3% 46.2% 44.4% 47.4% 59.1% 57.1% 60.4% 70.6% 68.3% 71.4% 78.9% 78.5% 79.8% 87.2% 86.2% 88.7% 92.9% 92.6% 94.2% 96.4% 96.1% 96.9% 98.2% 97.7% 98.3% 98.9% 98.6% 99.1% 99.5% 99.2% 99.5% 99.8% 99.8% 99.9% 99.9% 99.9% 100.0%
FF48 (Compustat) PR1YR UMD FF3 9.9% 10.0% 9.5% 15.8% 16.1% 16.0% 23.7% 23.6% 24.2% 34.2% 34.1% 34.1% 46.0% 46.9% 47.8% 58.4% 57.7% 58.5% 68.9% 67.9% 70.8% 79.7% 77.4% 79.6% 86.4% 86.0% 87.3% 91.7% 91.4% 92.2% 95.4% 95.3% 96.2% 97.5% 97.3% 98.1% 98.8% 98.5% 98.9% 99.3% 99.1% 99.4% 99.4% 99.4% 99.5% 99.7% 99.6% 99.9% 10.8% 10.4% 9.8% 17.2% 16.0% 17.3% 25.8% 23.7% 25.5% 35.8% 33.5% 35.8% 46.0% 46.2% 47.4% 59.1% 58.5% 59.8% 71.1% 69.5% 72.0% 80.6% 79.2% 82.1% 87.5% 86.4% 88.0% 92.1% 91.3% 92.3% 94.9% 94.6% 95.7% 97.9% 96.3% 97.8% 98.8% 98.6% 99.0% 99.4% 99.2% 99.5% 99.7% 99.6% 99.7% 99.9% 99.6% 99.9% 9.7% 9.5% 8.7% 15.9% 16.1% 15.7% 25.1% 22.5% 26.1% 36.5% 34.3% 36.9% 47.8% 45.9% 49.9% 59.5% 57.5% 60.9% 70.6% 66.6% 71.9% 79.3% 77.8% 80.3% 86.3% 84.9% 87.2% 91.2% 90.4% 92.2% 95.4% 94.7% 96.3% 97.6% 97.3% 98.1% 98.9% 98.7% 99.3% 99.4% 99.6% 99.4% 99.8% 99.9% 99.7% 100.0% 100.0% 99.8% 9.6% 10.1% 8.7% 15.8% 16.2% 16.0% 24.7% 24.0% 25.4% 34.7% 33.4% 35.3% 46.4% 44.1% 48.1% 60.3% 57.4% 62.7% 73.0% 70.0% 75.6% 82.9% 80.9% 83.9% 89.0% 88.4% 89.7% 93.1% 92.7% 94.2% 96.2% 95.7% 97.0% 98.1% 97.5% 98.0% 99.1% 98.9% 99.1% 99.5% 99.5% 99.7% 99.8% 99.9% 99.9% 100.0% 99.9% 99.9%
6-digit GICS PR1YR UMD 9.1% 9.3% 15.9% 15.4% 24.9% 25.1% 36.2% 37.0% 50.1% 48.8% 61.1% 60.9% 73.3% 71.5% 80.7% 79.6% 87.6% 87.1% 93.3% 92.3% 96.2% 95.7% 98.0% 97.9% 99.2% 99.6% 99.8% 100.0% 100.0% 100.0% 100.0% 100.0% 9.7% 9.1% 16.9% 15.5% 25.5% 24.6% 36.1% 35.7% 49.9% 48.0% 61.7% 59.0% 72.6% 69.5% 82.2% 81.0% 88.4% 88.0% 93.7% 93.6% 97.7% 96.8% 98.8% 98.3% 99.5% 99.5% 99.8% 99.7% 99.9% 99.9% 100.0% 100.0% 10.3% 9.5% 16.7% 16.1% 26.5% 24.7% 37.2% 35.9% 48.5% 48.3% 59.7% 58.8% 71.3% 69.3% 79.6% 78.1% 87.4% 85.8% 92.9% 91.7% 96.0% 95.4% 98.0% 97.8% 98.8% 98.7% 99.3% 99.2% 99.5% 99.4% 99.9% 99.9% 10.8% 10.1% 17.4% 16.5% 26.7% 25.7% 37.1% 36.4% 47.6% 46.0% 59.4% 58.3% 71.0% 70.3% 80.3% 78.7% 87.9% 86.6% 93.7% 91.4% 96.1% 95.9% 98.6% 97.4% 99.3% 98.9% 99.6% 99.5% 99.7% 99.6% 99.8% 99.8%
Industry classification used to compute industry-adjusted returns 8-digit GICS 3-digit SIC 4-digit SIC PR1YR UMD FF3 PR1YR UMD FF3 PR1YR UMD 9.2% 9.1% 9.5% 10.2% 8.0% 8.9% 10.5% 10.9% 14.5% 14.5% 15.6% 14.8% 12.5% 14.2% 15.5% 15.7% 23.4% 21.3% 24.4% 20.7% 20.6% 21.3% 21.8% 22.1% 32.8% 32.4% 32.8% 28.6% 28.2% 29.2% 28.7% 28.8% 44.3% 43.3% 44.5% 37.4% 38.0% 38.7% 38.4% 37.3% 55.7% 56.0% 59.5% 47.3% 46.8% 48.5% 47.7% 45.4% 68.4% 66.6% 71.1% 56.4% 55.8% 58.3% 56.7% 55.9% 79.3% 77.5% 79.5% 67.1% 64.3% 67.5% 65.5% 65.0% 86.8% 84.5% 86.6% 74.9% 73.1% 75.0% 73.9% 73.1% 91.9% 89.9% 91.8% 82.0% 80.7% 81.5% 81.6% 80.6% 95.8% 95.1% 95.9% 86.8% 86.4% 86.2% 86.6% 86.0% 97.8% 97.6% 97.8% 90.8% 90.6% 90.6% 91.7% 90.1% 98.7% 98.3% 98.9% 94.2% 93.7% 94.6% 94.3% 93.3% 99.3% 99.1% 99.5% 95.8% 96.0% 96.4% 96.5% 96.1% 99.8% 99.9% 99.8% 97.3% 97.4% 98.0% 97.9% 97.4% 99.9% 99.9% 99.9% 98.7% 98.4% 98.7% 98.9% 98.7% 9.6% 9.3% 10.0% 10.3% 9.7% 10.7% 10.9% 11.4% 16.4% 15.2% 17.0% 16.3% 15.3% 16.6% 15.6% 15.1% 24.6% 23.8% 25.0% 23.3% 21.0% 22.5% 22.8% 21.5% 34.7% 33.4% 35.9% 31.6% 29.8% 32.6% 30.5% 29.8% 45.2% 43.5% 47.1% 40.8% 38.0% 42.6% 38.5% 37.5% 56.8% 53.7% 59.5% 49.4% 47.4% 52.0% 47.6% 46.7% 67.8% 65.3% 71.2% 60.1% 58.3% 61.4% 58.4% 56.8% 76.3% 74.5% 77.3% 69.0% 66.5% 70.3% 67.8% 67.0% 83.4% 82.0% 85.2% 76.9% 75.0% 78.6% 75.9% 74.5% 90.5% 88.4% 90.1% 82.6% 82.5% 84.2% 82.4% 81.3% 94.0% 93.6% 94.2% 88.8% 87.6% 90.4% 89.0% 86.2% 96.3% 96.4% 96.6% 92.9% 91.9% 92.8% 92.9% 91.6% 98.1% 98.0% 98.3% 95.1% 94.6% 95.5% 95.6% 94.5% 99.2% 98.8% 99.5% 97.2% 96.7% 97.5% 97.0% 96.3% 99.8% 99.5% 99.6% 98.1% 97.8% 98.2% 98.4% 97.9% 99.8% 99.7% 99.8% 98.7% 98.3% 98.7% 99.1% 98.9% 11.7% 11.1% 11.4% 12.4% 12.2% 11.5% 11.7% 11.5% 17.8% 16.6% 16.6% 17.0% 16.6% 15.6% 17.0% 16.3% 25.1% 24.9% 24.4% 22.2% 21.8% 22.0% 23.0% 22.3% 34.7% 33.9% 36.4% 30.5% 28.7% 29.3% 29.6% 28.7% 46.4% 44.5% 46.4% 39.1% 36.9% 39.5% 37.3% 35.9% 56.2% 54.1% 57.7% 47.3% 45.7% 48.0% 46.8% 44.4% 66.6% 65.7% 68.4% 56.3% 55.8% 58.3% 55.0% 53.0% 76.3% 74.7% 77.5% 65.5% 63.3% 66.6% 64.7% 61.0% 84.4% 81.2% 86.0% 75.0% 71.3% 74.8% 72.0% 68.5% 89.5% 86.8% 90.2% 82.7% 79.9% 82.6% 78.6% 76.0% 93.6% 91.5% 94.4% 87.9% 86.3% 87.9% 84.6% 82.5% 96.4% 95.1% 96.8% 92.3% 91.1% 93.0% 88.9% 87.0% 98.4% 97.0% 98.4% 95.1% 94.5% 95.6% 92.7% 91.2% 99.0% 98.5% 99.2% 97.3% 96.5% 97.4% 95.1% 94.3% 99.3% 99.4% 99.4% 98.3% 97.8% 98.5% 97.2% 96.4% 99.7% 99.7% 99.6% 98.9% 99.1% 99.4% 98.2% 97.8% 9.0% 9.9% 9.0% 11.8% 10.2% 11.2% 13.1% 11.6% 13.8% 13.8% 14.0% 17.3% 14.4% 17.7% 17.6% 15.5% 20.2% 20.3% 19.5% 23.4% 20.2% 24.3% 22.5% 22.5% 28.6% 27.5% 28.8% 30.2% 28.3% 30.6% 30.8% 28.8% 38.4% 37.1% 38.5% 39.0% 37.3% 38.9% 38.9% 36.6% 48.1% 47.7% 48.7% 47.9% 45.1% 48.8% 48.9% 45.0% 59.1% 56.4% 60.2% 57.1% 53.1% 57.5% 56.7% 52.4% 68.7% 67.7% 71.7% 65.7% 61.9% 67.0% 64.4% 63.8% 78.3% 77.3% 79.2% 74.2% 69.0% 75.7% 71.9% 70.3% 84.6% 83.9% 85.2% 81.3% 78.2% 81.5% 79.2% 76.9% 88.8% 88.6% 89.8% 86.1% 84.2% 87.7% 84.7% 82.7% 93.1% 92.3% 93.5% 90.5% 89.7% 90.8% 90.4% 88.7% 95.8% 95.1% 96.4% 93.2% 92.5% 93.6% 93.2% 92.1% 97.8% 97.2% 98.3% 95.4% 94.8% 95.8% 96.3% 94.9% 98.8% 98.5% 99.0% 97.2% 96.2% 97.6% 97.9% 97.6% 99.5% 99.1% 99.4% 97.8% 97.7% 98.1% 98.6% 98.3%
FF3 8.9% 16.7% 24.8% 37.2% 50.7% 64.1% 73.3% 81.9% 88.4% 93.3% 96.5% 98.7% 99.5% 99.7% 100.0% 100.0% 10.1% 17.2% 27.0% 37.9% 50.6% 64.0% 74.1% 83.5% 90.7% 95.2% 97.6% 98.7% 99.4% 99.8% 99.9% 100.0% 9.9% 16.0% 25.8% 38.4% 49.1% 61.5% 70.9% 80.4% 88.2% 93.2% 96.9% 97.9% 98.6% 99.2% 99.6% 100.0% 10.2% 16.8% 26.2% 36.8% 48.2% 59.5% 71.5% 81.7% 88.7% 93.6% 97.3% 98.7% 99.5% 99.7% 99.8% 99.9%
16
FF3 10.4% 15.6% 22.2% 29.7% 38.7% 48.6% 58.0% 66.1% 75.3% 82.6% 87.7% 92.1% 94.9% 96.6% 98.3% 98.9% 10.6% 16.2% 23.9% 30.5% 39.3% 48.7% 58.7% 68.5% 76.6% 83.3% 90.0% 93.7% 96.5% 97.7% 98.5% 99.0% 11.9% 17.4% 24.0% 30.7% 38.6% 47.9% 57.3% 65.8% 73.7% 79.8% 85.0% 89.6% 93.5% 95.7% 96.9% 98.0% 13.1% 17.1% 22.9% 30.8% 39.8% 47.4% 56.1% 65.4% 73.1% 80.8% 86.1% 90.8% 94.2% 96.4% 98.0% 99.3%
3-digit SIC (CRSP) PR1YR UMD FF3 10.8% 10.7% 10.7% 15.5% 15.0% 15.5% 20.9% 21.0% 21.8% 28.7% 27.8% 28.8% 35.9% 35.3% 36.6% 44.1% 43.3% 45.1% 52.6% 51.3% 54.2% 61.4% 60.4% 62.6% 70.3% 69.0% 70.7% 76.0% 75.3% 77.5% 81.8% 81.4% 84.3% 87.0% 86.3% 88.4% 91.2% 90.8% 93.0% 94.4% 93.7% 95.2% 96.5% 96.0% 96.8% 97.6% 97.4% 98.0% 11.4% 10.1% 9.6% 15.8% 15.4% 15.9% 21.4% 20.6% 20.6% 27.7% 27.5% 29.2% 36.7% 35.0% 37.7% 44.9% 43.4% 46.6% 56.0% 54.6% 56.1% 65.2% 63.8% 65.7% 73.1% 71.9% 74.4% 79.8% 77.6% 80.4% 85.1% 83.9% 86.6% 89.6% 88.4% 90.2% 92.7% 92.1% 94.4% 95.1% 95.0% 96.0% 97.3% 96.8% 97.6% 98.4% 97.9% 99.1% 12.0% 11.2% 10.5% 15.5% 15.9% 14.5% 20.0% 19.6% 20.4% 27.3% 26.6% 26.5% 35.3% 33.2% 32.8% 43.5% 40.2% 41.9% 51.7% 49.5% 50.9% 60.1% 59.0% 60.0% 68.5% 66.6% 67.9% 75.6% 73.4% 75.0% 79.7% 79.0% 80.2% 84.8% 84.3% 85.5% 89.0% 88.0% 90.3% 91.9% 91.5% 93.2% 95.0% 94.5% 95.6% 97.4% 96.9% 97.8% 10.8% 10.8% 11.2% 16.6% 15.5% 14.7% 21.0% 21.2% 20.0% 28.6% 28.0% 27.0% 36.1% 35.5% 34.5% 45.5% 45.1% 43.1% 54.2% 52.4% 53.2% 62.5% 61.0% 61.4% 71.4% 69.6% 70.3% 77.8% 77.9% 77.8% 85.2% 83.7% 84.2% 89.0% 88.2% 89.1% 92.6% 91.6% 92.0% 94.7% 95.2% 94.6% 96.4% 96.7% 97.0% 98.2% 98.3% 98.2%
4-digit SIC (CRSP) PR1YR UMD FF3 9.0% 9.1% 9.2% 12.6% 12.5% 13.5% 19.0% 17.9% 19.0% 23.9% 23.5% 24.2% 30.8% 30.1% 31.8% 38.6% 38.1% 39.4% 48.8% 46.4% 49.7% 57.4% 55.0% 58.6% 66.6% 64.2% 67.8% 74.7% 72.5% 76.0% 82.1% 79.6% 83.5% 87.8% 85.6% 89.3% 92.1% 90.2% 93.3% 94.8% 94.1% 95.2% 96.6% 96.0% 96.7% 97.5% 97.2% 98.1% 8.8% 9.4% 9.6% 14.6% 14.8% 14.9% 21.4% 20.8% 20.6% 28.5% 28.6% 29.9% 38.4% 38.0% 38.4% 47.0% 47.0% 46.5% 55.7% 54.8% 57.3% 63.4% 63.5% 66.2% 71.0% 70.0% 73.0% 78.6% 76.6% 79.5% 84.1% 82.4% 84.7% 88.5% 86.9% 89.8% 92.1% 90.7% 92.8% 95.1% 94.5% 94.7% 96.9% 96.7% 96.7% 97.8% 97.5% 97.8% 10.4% 10.9% 9.5% 14.6% 15.0% 13.7% 19.5% 20.3% 19.5% 25.4% 24.7% 24.9% 31.8% 30.9% 32.7% 39.0% 37.2% 39.1% 46.4% 44.0% 47.0% 53.7% 52.1% 55.1% 62.4% 59.8% 61.6% 68.6% 66.9% 70.0% 75.2% 72.9% 76.2% 81.4% 78.5% 81.7% 86.5% 84.6% 87.2% 90.2% 88.7% 91.4% 93.3% 92.2% 94.4% 95.7% 94.7% 95.9% 10.0% 10.8% 8.8% 14.6% 15.2% 13.0% 20.2% 20.1% 19.3% 26.8% 26.7% 25.1% 32.8% 34.1% 31.5% 40.9% 41.6% 39.9% 50.0% 49.4% 48.2% 58.0% 58.8% 57.5% 65.1% 65.8% 64.1% 72.0% 70.8% 72.1% 78.6% 77.1% 78.9% 84.2% 82.6% 84.8% 87.9% 86.9% 88.5% 90.9% 90.1% 91.3% 93.5% 93.0% 93.8% 95.9% 95.6% 96.8%
Table A.4 (Continued) Power Tests Industry classification used to form portfolios that mimic the industry clustering within the GIM (2003) Dem/Dict portfolios
3-digit SIC
4-digit SIC
3-digit SIC (CRSP)
4-digit SIC (CRSP)
Monthly alpha added to the portfolio 0.00% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50% 0.00% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50% 0.00% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50% 0.00% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% 0.70% 0.80% 0.90% 1.00% 1.10% 1.20% 1.30% 1.40% 1.50%
FF48 (CRSP) PR1YR UMD 8.0% 8.0% 15.8% 15.1% 27.2% 24.6% 36.5% 35.4% 49.0% 46.7% 60.9% 58.2% 71.2% 69.9% 79.8% 78.6% 87.6% 86.3% 94.0% 91.4% 96.9% 95.8% 98.5% 97.3% 99.1% 99.1% 99.6% 99.5% 99.8% 99.7% 100.0% 99.9% 8.9% 8.9% 16.0% 14.7% 26.5% 24.8% 36.1% 34.9% 47.9% 46.7% 60.6% 58.7% 71.8% 69.7% 81.5% 80.5% 87.9% 87.6% 93.3% 91.9% 96.2% 94.8% 97.7% 97.4% 99.1% 98.6% 99.8% 99.4% 99.8% 99.9% 100.0% 100.0% 8.7% 7.9% 14.8% 14.9% 25.3% 24.1% 36.4% 35.3% 49.7% 45.9% 62.0% 58.8% 73.2% 71.2% 82.7% 79.7% 90.3% 87.6% 94.7% 93.4% 97.2% 96.5% 98.5% 97.8% 99.2% 99.1% 99.7% 99.4% 99.8% 99.9% 99.9% 100.0% 7.2% 7.1% 13.7% 13.7% 26.1% 26.2% 40.4% 39.9% 59.2% 54.3% 74.3% 70.7% 85.3% 82.9% 93.1% 90.9% 97.3% 96.6% 99.1% 98.6% 99.8% 99.7% 99.9% 99.8% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0%
FF3 8.9% 16.9% 26.8% 37.1% 49.9% 61.7% 72.0% 81.4% 89.5% 95.1% 97.7% 98.6% 99.5% 99.8% 99.8% 100.0% 9.4% 16.7% 25.9% 38.5% 50.2% 62.7% 74.7% 82.8% 89.2% 94.0% 96.6% 97.9% 99.5% 99.8% 99.9% 100.0% 8.6% 14.9% 25.8% 38.9% 51.7% 62.8% 73.5% 83.7% 90.6% 95.3% 97.4% 98.5% 99.4% 99.8% 99.9% 99.9% 7.2% 15.3% 28.3% 44.3% 62.3% 76.9% 87.8% 94.8% 98.3% 99.5% 99.7% 99.9% 100.0% 100.0% 100.0% 100.0%
FF48 (Compustat) PR1YR UMD FF3 10.6% 8.8% 10.4% 18.7% 17.0% 18.5% 27.4% 26.4% 27.8% 40.0% 36.5% 40.9% 53.2% 47.7% 54.4% 65.5% 60.1% 66.3% 75.6% 71.3% 77.8% 84.0% 80.6% 84.2% 89.9% 87.7% 90.8% 93.9% 92.4% 95.1% 97.7% 95.9% 97.3% 98.6% 98.3% 98.8% 99.2% 99.2% 99.7% 99.8% 99.7% 99.8% 99.8% 99.8% 99.8% 99.9% 99.8% 100.0% 10.1% 8.4% 9.0% 15.9% 14.3% 15.6% 24.3% 23.8% 24.1% 37.4% 33.3% 36.1% 49.7% 46.1% 51.3% 61.8% 58.7% 62.8% 74.4% 70.2% 75.9% 85.0% 81.3% 85.9% 91.1% 89.9% 92.1% 95.8% 93.7% 97.0% 98.3% 97.2% 98.2% 99.2% 99.1% 99.4% 99.7% 99.6% 99.8% 99.9% 99.8% 99.8% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 10.5% 9.7% 10.2% 16.3% 17.1% 16.8% 26.4% 26.2% 26.2% 40.1% 35.8% 39.5% 51.9% 49.1% 53.0% 61.9% 60.3% 66.1% 73.9% 71.0% 75.7% 82.2% 80.5% 83.4% 89.6% 88.9% 90.1% 94.9% 93.4% 95.7% 98.0% 97.1% 98.3% 99.0% 98.8% 99.0% 99.2% 99.3% 99.3% 99.7% 99.8% 99.6% 99.9% 99.9% 99.9% 100.0% 99.9% 99.9% 6.6% 7.7% 7.4% 13.8% 15.3% 13.8% 25.3% 27.3% 25.4% 42.5% 42.8% 43.3% 59.4% 59.1% 61.6% 74.3% 76.2% 76.4% 84.6% 85.4% 87.0% 93.8% 93.1% 94.1% 97.4% 96.9% 98.2% 99.2% 99.0% 99.4% 99.8% 99.8% 99.8% 100.0% 99.9% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0%
6-digit GICS PR1YR UMD 9.4% 9.7% 14.8% 16.0% 22.4% 22.9% 32.5% 33.3% 43.0% 44.2% 56.0% 55.5% 67.6% 65.9% 77.1% 76.4% 83.7% 83.8% 90.4% 89.6% 94.3% 93.2% 97.0% 96.6% 98.2% 98.4% 99.4% 99.7% 99.9% 99.8% 99.9% 100.0% 7.5% 8.0% 13.1% 13.1% 20.2% 22.1% 31.5% 33.2% 43.4% 44.3% 56.3% 58.3% 68.4% 68.0% 79.0% 78.0% 86.3% 86.0% 93.1% 91.8% 96.2% 95.0% 98.1% 98.3% 99.6% 99.4% 99.7% 99.6% 99.8% 99.8% 100.0% 99.9% 9.4% 8.8% 15.5% 15.2% 24.6% 24.2% 35.8% 35.1% 48.0% 46.9% 60.6% 58.1% 70.5% 68.2% 79.3% 78.1% 87.7% 86.8% 93.7% 92.8% 97.2% 96.2% 98.3% 98.4% 99.1% 99.1% 99.7% 99.5% 99.9% 99.7% 99.9% 99.9% 5.3% 5.5% 11.6% 12.0% 23.0% 21.6% 39.6% 37.0% 56.7% 54.9% 72.4% 71.0% 83.6% 82.8% 91.9% 90.9% 96.9% 96.2% 99.1% 98.8% 99.9% 99.7% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0%
Industry classification used to compute industry-adjusted returns 8-digit GICS 3-digit SIC 4-digit SIC FF3 PR1YR UMD FF3 PR1YR UMD FF3 PR1YR UMD 8.9% 6.7% 7.5% 6.9% 9.0% 8.1% 9.8% 10.4% 11.1% 15.2% 11.4% 11.1% 12.4% 14.5% 13.4% 14.9% 15.3% 15.3% 24.9% 17.8% 17.3% 18.4% 21.4% 19.8% 21.5% 20.9% 20.9% 34.7% 24.7% 24.4% 26.2% 28.8% 27.5% 28.9% 28.5% 28.2% 44.5% 34.1% 34.1% 37.5% 37.0% 35.5% 37.2% 37.9% 35.8% 58.4% 46.2% 45.9% 47.6% 46.1% 44.8% 47.7% 47.1% 44.3% 69.4% 58.2% 55.9% 58.8% 56.3% 55.0% 57.4% 54.6% 52.4% 79.2% 67.4% 65.2% 69.8% 66.3% 63.2% 66.1% 63.6% 62.1% 86.8% 76.8% 75.1% 79.7% 73.5% 73.4% 74.4% 72.1% 71.5% 91.9% 85.3% 83.5% 86.1% 82.2% 80.4% 81.8% 79.0% 78.5% 95.2% 91.1% 90.5% 92.1% 86.8% 85.7% 87.6% 84.7% 83.5% 97.3% 94.8% 94.2% 95.8% 90.6% 89.9% 91.3% 89.9% 89.2% 99.1% 97.1% 97.0% 97.3% 93.1% 93.9% 93.6% 93.0% 92.2% 99.6% 98.1% 98.0% 98.3% 95.8% 95.5% 96.2% 95.6% 94.8% 99.9% 99.0% 98.8% 99.0% 97.5% 97.2% 97.5% 97.2% 96.6% 100.0% 99.5% 99.3% 99.3% 98.1% 98.3% 98.1% 98.2% 97.7% 8.7% 7.0% 6.9% 7.0% 8.5% 8.8% 8.4% 10.8% 11.8% 13.5% 10.8% 11.6% 11.4% 12.8% 12.7% 12.8% 16.7% 16.8% 21.7% 16.7% 17.9% 17.8% 16.4% 18.1% 17.0% 21.8% 22.4% 32.7% 25.0% 25.1% 26.8% 23.1% 24.4% 22.9% 28.2% 28.5% 46.8% 35.0% 34.4% 36.5% 32.1% 31.7% 31.0% 38.2% 36.1% 59.6% 44.0% 44.7% 46.7% 39.9% 41.7% 41.2% 45.9% 44.0% 71.9% 55.5% 55.0% 56.3% 50.2% 51.2% 51.6% 54.0% 53.6% 82.0% 66.0% 64.8% 68.1% 61.1% 61.0% 62.1% 64.1% 62.1% 89.3% 75.1% 75.5% 79.1% 69.8% 69.5% 71.5% 72.6% 70.7% 94.3% 83.8% 82.3% 86.5% 77.4% 77.2% 78.3% 80.1% 78.3% 97.4% 90.2% 88.3% 91.6% 84.5% 84.8% 85.0% 85.8% 83.5% 98.8% 94.2% 93.2% 95.1% 89.8% 89.3% 90.0% 90.8% 89.2% 99.6% 96.4% 96.4% 97.3% 93.2% 92.7% 93.5% 94.8% 92.9% 99.8% 97.9% 98.0% 98.7% 95.0% 95.1% 95.7% 96.7% 96.1% 99.9% 98.7% 98.6% 99.2% 96.8% 96.6% 97.3% 98.2% 98.1% 100.0% 99.6% 99.3% 99.8% 97.8% 98.0% 98.3% 98.8% 99.0% 9.7% 10.7% 10.5% 10.2% 8.9% 9.4% 8.2% 7.9% 8.6% 15.3% 16.9% 16.4% 16.1% 12.9% 13.1% 13.1% 12.1% 12.3% 24.5% 23.3% 22.9% 25.5% 18.4% 18.6% 18.7% 17.1% 16.9% 36.1% 33.1% 31.0% 33.7% 26.1% 24.4% 25.6% 22.1% 22.5% 49.2% 42.6% 42.0% 44.0% 34.9% 33.7% 33.7% 29.8% 29.2% 61.1% 53.5% 52.2% 56.1% 43.6% 43.1% 44.2% 39.1% 38.0% 72.9% 64.4% 63.0% 65.8% 54.6% 54.3% 55.0% 48.0% 46.8% 82.1% 73.2% 71.8% 74.1% 64.3% 63.0% 63.7% 58.1% 57.5% 90.2% 81.2% 80.5% 82.6% 71.7% 70.1% 70.8% 65.9% 65.3% 94.6% 87.8% 86.2% 88.7% 77.8% 76.1% 77.9% 73.0% 72.0% 97.6% 92.7% 91.5% 92.9% 84.7% 82.2% 85.3% 80.1% 78.9% 99.0% 95.6% 95.1% 96.0% 89.9% 87.4% 90.7% 85.6% 85.4% 99.2% 97.6% 96.6% 97.7% 92.9% 92.0% 92.9% 89.6% 90.4% 99.8% 98.5% 98.6% 99.0% 95.4% 94.4% 95.4% 93.9% 92.7% 99.9% 99.0% 99.1% 99.5% 96.9% 96.0% 97.2% 96.0% 95.2% 99.9% 99.6% 99.3% 99.5% 98.0% 97.4% 98.3% 97.8% 97.3% 5.0% 5.9% 6.1% 6.2% 10.6% 12.7% 9.5% 10.4% 11.7% 10.7% 11.2% 11.1% 11.6% 17.6% 17.8% 14.3% 15.6% 16.4% 23.9% 19.0% 18.4% 19.2% 25.6% 26.2% 21.7% 20.5% 22.1% 40.4% 29.5% 27.8% 29.9% 35.9% 35.0% 31.2% 30.0% 30.0% 57.2% 41.5% 39.1% 43.5% 46.2% 46.5% 42.7% 37.6% 38.5% 73.5% 55.7% 53.2% 58.2% 57.4% 57.2% 54.1% 46.3% 47.4% 86.4% 67.8% 63.4% 71.0% 67.7% 67.5% 63.6% 55.5% 56.7% 92.9% 80.0% 75.4% 81.5% 75.8% 74.9% 71.9% 65.8% 66.2% 97.2% 87.3% 84.9% 90.1% 82.2% 81.9% 80.7% 75.9% 75.0% 99.3% 93.6% 91.5% 95.0% 88.2% 87.5% 86.1% 82.1% 82.1% 100.0% 97.1% 95.2% 97.6% 92.7% 92.9% 92.2% 88.0% 87.1% 100.0% 98.6% 97.9% 99.2% 96.6% 96.3% 94.8% 91.7% 92.7% 100.0% 99.7% 99.5% 99.7% 98.4% 97.7% 97.2% 95.0% 95.8% 100.0% 99.9% 99.8% 100.0% 99.1% 99.0% 99.4% 97.4% 97.9% 100.0% 100.0% 99.9% 100.0% 99.5% 99.7% 99.8% 98.6% 98.6% 100.0% 100.0% 100.0% 100.0% 100.0% 99.7% 99.9% 99.2% 99.3%
17
FF3 10.3% 16.1% 22.0% 28.9% 39.4% 47.1% 56.0% 64.7% 72.1% 79.3% 85.4% 90.3% 94.1% 95.9% 97.7% 98.5% 10.4% 15.8% 22.3% 29.3% 38.4% 46.2% 55.7% 64.8% 75.0% 82.0% 88.0% 91.2% 95.0% 97.0% 98.0% 99.0% 8.0% 11.6% 15.8% 23.0% 30.3% 38.7% 48.9% 58.3% 67.3% 75.5% 81.7% 87.9% 91.9% 94.8% 97.0% 97.8% 9.1% 13.6% 20.0% 27.5% 34.7% 45.8% 55.0% 64.2% 73.5% 81.1% 87.4% 91.6% 95.3% 97.4% 98.3% 99.0%
3-digit SIC (CRSP) PR1YR UMD FF3 11.8% 12.9% 11.0% 17.7% 17.7% 16.2% 22.9% 23.1% 21.8% 30.1% 31.3% 28.7% 37.0% 39.0% 37.2% 46.0% 46.5% 46.1% 54.6% 54.2% 54.2% 63.1% 61.5% 62.8% 69.1% 68.3% 69.9% 76.0% 75.3% 76.1% 82.4% 81.3% 82.9% 88.0% 86.9% 88.3% 92.7% 91.7% 92.0% 95.1% 94.6% 94.4% 96.8% 97.5% 97.3% 98.5% 98.9% 98.7% 7.1% 8.8% 6.4% 10.1% 12.0% 9.3% 14.9% 16.5% 13.3% 20.8% 22.4% 19.8% 26.9% 30.4% 26.2% 35.0% 37.7% 32.6% 43.8% 46.0% 42.8% 51.4% 55.5% 51.2% 61.2% 62.3% 59.8% 69.4% 69.8% 68.8% 76.5% 76.2% 75.7% 82.1% 83.3% 81.0% 86.9% 88.4% 87.0% 91.9% 92.3% 91.4% 94.7% 95.0% 94.7% 96.5% 96.6% 96.6% 10.9% 10.5% 11.0% 15.1% 14.0% 14.5% 18.7% 18.5% 18.7% 25.1% 24.1% 24.6% 31.5% 31.2% 31.7% 38.9% 37.6% 40.7% 49.1% 46.2% 50.3% 57.5% 55.1% 59.6% 66.2% 63.9% 67.4% 74.7% 71.6% 75.8% 80.5% 78.7% 81.6% 85.2% 83.4% 86.7% 88.6% 87.7% 89.6% 92.7% 91.7% 93.6% 94.9% 94.4% 95.4% 97.3% 97.1% 97.5% 16.3% 16.2% 15.2% 20.6% 21.6% 21.2% 27.8% 27.9% 28.4% 36.9% 36.1% 37.8% 46.2% 45.5% 47.2% 56.8% 54.7% 58.0% 65.9% 62.9% 66.6% 73.6% 70.6% 74.7% 80.5% 78.8% 81.1% 85.9% 85.0% 85.8% 89.0% 89.2% 90.7% 92.4% 93.0% 94.0% 96.4% 95.6% 96.1% 97.4% 97.1% 97.9% 98.3% 98.0% 98.6% 99.2% 98.7% 99.2%
4-digit SIC (CRSP) PR1YR UMD FF3 12.3% 12.7% 12.3% 16.7% 18.3% 17.0% 22.3% 23.9% 22.5% 29.3% 30.1% 30.4% 37.7% 38.0% 37.1% 45.6% 46.3% 45.0% 53.6% 53.9% 53.9% 61.3% 61.6% 62.2% 69.0% 68.6% 69.5% 75.5% 75.7% 76.3% 81.3% 81.7% 82.0% 86.2% 86.0% 86.4% 90.1% 89.2% 91.1% 92.7% 92.3% 93.4% 95.0% 95.3% 95.7% 96.9% 96.8% 97.6% 7.4% 8.7% 6.7% 10.7% 12.6% 10.0% 14.8% 16.2% 14.8% 20.7% 21.1% 20.3% 27.3% 27.8% 27.8% 35.2% 35.2% 36.6% 43.1% 43.6% 44.4% 50.3% 51.7% 52.0% 60.1% 60.8% 61.1% 68.5% 69.4% 69.3% 75.4% 76.3% 76.3% 81.8% 82.7% 82.5% 86.5% 88.0% 87.3% 90.4% 91.4% 90.4% 93.1% 94.0% 94.1% 95.2% 95.9% 96.3% 9.6% 10.7% 9.2% 13.1% 13.6% 12.3% 17.9% 17.0% 16.8% 21.5% 21.8% 21.4% 27.7% 26.9% 27.1% 34.2% 33.5% 35.8% 41.9% 40.3% 43.3% 49.5% 46.7% 51.6% 57.8% 55.2% 58.8% 65.2% 64.3% 65.8% 71.6% 70.8% 72.6% 78.1% 76.2% 80.1% 84.0% 82.2% 85.9% 88.6% 86.3% 89.9% 91.4% 89.6% 92.3% 94.1% 93.0% 94.9% 12.4% 11.6% 13.5% 17.5% 15.1% 17.9% 22.3% 20.7% 23.7% 28.1% 25.7% 29.1% 34.1% 31.4% 36.2% 42.4% 38.5% 44.2% 50.3% 46.8% 53.9% 60.5% 56.4% 62.0% 68.6% 65.0% 69.1% 74.5% 72.2% 75.7% 79.8% 77.6% 81.2% 84.1% 82.7% 85.3% 88.2% 86.3% 89.1% 91.0% 89.6% 92.4% 93.4% 92.6% 94.6% 96.0% 94.7% 97.1%
Table A.5 Precision Tests when Industry Returns are Computed using the Universe of Stocks For each industry classification standard listed in Table 5, I regress calendar-time, value-weighted, industry-adjusted returns on the GIM Democracy and Dictatorship portfolios against the PR1YR , UMD , and FF3 asset pricing models which are described in the text. I perform separate regressions for the Democracy and Dictatorship portfolios. I then examine the precision of each of the 24 combinations of industry classifications and asset pricing models listed in Table 5. λ is defined in Appendix D, while σ is the standard error of the regression intercept. Precision is defined as (1-λ)/σ. After computing precision, I then rank each classification based on its precision score. I rank specifications separately for regressions involving Democracy and Dictatorship portfolio returns. Combined rank is defined as the sum of the precision ranks from the Democracy and Dictatorship columns. The tests with the lowest combined rank have the highest precision. As in Table 5, industry returns are computed using all firms in the merged CRSP/Compustat database. Democracy portfolio Industry classification used to compute industry returns 4-digit SIC 4-digit SIC 4-digit SIC 6-digit GICS 6-digit GICS FF48 (CRSP) FF48 FF48 (CRSP) 3-digit SIC 6-digit GICS FF48 8-digit GICS 3-digit SIC FF48 (CRSP) 4-digit SIC (CRSP) 8-digit GICS FF48 3-digit SIC 4-digit SIC (CRSP) 3-digit SIC (CRSP) 8-digit GICS 3-digit SIC (CRSP) 4-digit SIC (CRSP) 3-digit SIC (CRSP)
Asset Pricing Model FF3 PR1YR UMD FF3 PR1YR FF3 FF3 PR1YR FF3 UMD PR1YR FF3 PR1YR UMD FF3 PR1YR UMD UMD PR1YR FF3 UMD PR1YR UMD UMD
λ 0.44 0.44 0.44 0.21 0.21 0.16 0.14 0.16 0.35 0.21 0.14 0.32 0.35 0.16 0.50 0.32 0.14 0.35 0.50 0.32 0.32 0.32 0.50 0.32
σ Precision Rank 0.065 8.53 1 0.067 8.35 2 0.068 8.15 3 0.104 7.62 4 0.107 7.43 5 0.114 7.39 6 0.119 7.25 7 0.117 7.21 8 0.090 7.20 9 0.110 7.18 10 0.122 7.06 11 0.097 7.06 12 0.092 7.02 13 0.121 6.95 14 0.073 6.88 15 0.099 6.87 16 0.126 6.86 17 0.094 6.86 18 0.075 6.71 19 0.102 6.63 20 0.103 6.63 21 0.105 6.46 22 0.078 6.46 23 0.109 6.24 24
Dictatorship portfolio Industry classification used to compute industry returns 8-digit GICS 8-digit GICS 3-digit SIC 6-digit GICS 8-digit GICS 3-digit SIC 6-digit GICS 3-digit SIC 6-digit GICS 4-digit SIC FF48 4-digit SIC FF48 FF48 4-digit SIC 4-digit SIC (CRSP) FF48 (CRSP) 4-digit SIC (CRSP) 4-digit SIC (CRSP) 3-digit SIC (CRSP) FF48 (CRSP) FF48 (CRSP) 3-digit SIC (CRSP) 3-digit SIC (CRSP)
Asset Pricing Model FF3 PR1YR FF3 FF3 UMD PR1YR PR1YR UMD UMD FF3 FF3 PR1YR PR1YR UMD UMD FF3 FF3 PR1YR UMD FF3 PR1YR UMD PR1YR UMD
18
λ 0.15 0.15 0.26 0.07 0.15 0.26 0.07 0.26 0.07 0.31 0.10 0.31 0.10 0.10 0.31 0.44 0.11 0.44 0.44 0.28 0.11 0.11 0.28 0.28
σ Precision Rank 0.119 7.15 1 0.122 6.97 2 0.108 6.87 3 0.135 6.86 4 0.125 6.82 5 0.111 6.70 6 0.139 6.69 7 0.113 6.59 8 0.141 6.57 9 0.107 6.45 10 0.139 6.44 11 0.110 6.29 12 0.143 6.28 13 0.144 6.21 14 0.112 6.16 15 0.091 6.14 16 0.148 6.03 17 0.094 5.99 18 0.094 5.98 19 0.122 5.93 20 0.152 5.87 21 0.154 5.78 22 0.125 5.78 23 0.127 5.69 24
Combined rank Industry classification used to compute industry returns 6-digit GICS 4-digit SIC 6-digit GICS 3-digit SIC 8-digit GICS 4-digit SIC FF48 8-digit GICS 4-digit SIC 6-digit GICS 3-digit SIC FF48 (CRSP) FF48 8-digit GICS 3-digit SIC FF48 (CRSP) FF48 4-digit SIC (CRSP) FF48 (CRSP) 4-digit SIC (CRSP) 3-digit SIC (CRSP) 4-digit SIC (CRSP) 3-digit SIC (CRSP) 3-digit SIC (CRSP)
Asset Pricing Combined Model Rank FF3 8 FF3 11 PR1YR 12 FF3 12 FF3 13 PR1YR 14 FF3 18 PR1YR 18 UMD 18 UMD 19 PR1YR 19 FF3 23 PR1YR 24 UMD 26 UMD 26 PR1YR 29 UMD 31 FF3 31 UMD 36 PR1YR 37 FF3 40 UMD 42 PR1YR 45 UMD 48
