Professional Documents
Culture Documents
ESG Rating Disagreement and Stock Returns
ESG Rating Disagreement and Stock Returns
To cite this article: Rajna Gibson Brandon, Philipp Krueger & Peter Steffen Schmidt (2021)
ESG Rating Disagreement and Stock Returns, Financial Analysts Journal, 77:4, 104-127, DOI:
10.1080/0015198X.2021.1963186
Using environmental, social, and There’s so much disagreement about investing, and it’s because
governance (ESG) ratings from nobody really knows.
seven different data providers for
a sample of firms in the S&P 500 —Robert J. Shiller
Index between 2010 and 2017, we
studied the relationship between Environmental, social, and governance (ESG) ratings nowadays feature
ESG rating disagreement and prominently in the financial press, regulatory and policy debates, and
stock returns. We found that stock academic studies, and they are also a hot topic in investment practice.
returns are positively related to Such ratings also increasingly shape investment decisions of institu-
ESG rating disagreement, suggest- tional investors representing trillions of dollars in assets under manage-
ing a risk premium for firms with ment (Gibson, Glossner, Krueger, Matos, and Steffen 2021; GSIA 2016;
higher ESG rating disagreement. USSIF 2020; PRI 2018). Recently, a lot of attention has been paid to
The relationship is primarily driven divergence of ratings for the same firm issued by different ESG rating
by disagreement about the envi- providers. For example, in a Wall Street Journal article, Mackintosh
ronmental dimension. We discuss (2018) pointed out that Tesla was rated highly by MSCI regarding
the practical implications of our environmental issues in 2018. In contrast, FTSE reached the opposite
findings for firms’ equity cost of conclusion, rating Tesla poorly on environmental matters. Other news
capital as well as for investment outlets, policy-oriented think tanks, and practitioner-oriented publi-
managers and asset owners who cations have made similar observations (Wigglesworth 2018; Doyle
use ESG investment strategies. 2018; Matos 2020).
104 © 2021 CFA Institute. All rights reserved. Fourth Quarter 2021
ESG Rating Disagreement and Stock Returns
equity investors who pursue sustainable investment insights for financial analysts who cover firms in
strategies. these sectors.
For the purpose of this study, we collected and stud- In our main empirical analysis, we measured ESG
ied ESG ratings from seven prominent ESG rating rating disagreement as the standard deviation of the
providers for firms in the S&P 500 Index between available ESG ratings from the seven different data
2010 and 2017. We used data from Asset4 (now providers for a given firm at a given point in time. We
Refinitiv ESG), Sustainalytics (now Morningstar), calculated the disagreement measure for the total
Inrate, Bloomberg, FTSE, KLD (now MSCI), and MSCI ESG rating and separately for the E, S, and G dimen-
IVA. We believe that our work has the most compre- sions (or “pillars”). We then related monthly stock
hensive data coverage among current papers that returns to our proxy of ESG rating disagreement
have studied the issue of ESG rating disagreement. while controlling for standard stock characteristics
For instance, Christensen et al. (forthcoming) used that are known to have predictive power in the
data from only three rating providers, and Berg et al. cross-section of stock returns (e.g., size, momentum,
(2020) did not focus on the time-series dimension. and quality).
In contrast, we built a panel dataset based on seven
different ESG ratings. Note also that the focus of the We found that stock returns are positively related to
other papers is very different from ours, in that they ESG rating disagreement. Further tests showed that
primarily examine why ESG ratings disagree. the relationship is driven mainly by disagreement
about the environmental rating. In terms of economic
We started our analysis by documenting some basic magnitude, we estimated that an interquartile range
empirical facts about ESG rating disagreement in our increase in ESG rating disagreement is associated
sample of S&P 500 firms. We found, for example, with an increase of 92 bps in the annual cost of
that the average pairwise correlation between the equity capital. Hence, ignoring differences in ESG
ESG ratings of the seven rating providers is about rating disagreement in corporate valuation could
0.45. Surprisingly, the average pairwise correlation lead to sizable mistakes when estimating the value
is lowest for the governance dimension (0.16) and of a firm’s equity: Assume, for instance, a firm has
highest for the environmental dimension (0.46). perpetual annual free cash flow to equity (FCFE) of
Our analysis of pairwise correlations between ESG $100 million per year, an expected growth rate of
ratings from different providers also highlights more the FCFE of g = 0%, and a cost of equity capital of
subtle patterns in disagreement that go against rE = 5%. An interquartile range increase in ESG rating
the common beliefs about generalized ESG rating disagreement combined with our empirical estimates
disagreement. For example, our analysis also shows would imply that the true cost of equity capital is
that correlations between ratings from some provid- 5.92%, suggesting an overestimation of the value of
ers are markedly high, with the maximum correlation the firm’s equity by about $311 million (or 18%) if the
being about 0.75. ESG rating disagreement is ignored.1
We next studied whether disagreement varies Our finding of a positive relationship between
in relation to observable firm-level financial and ESG rating disagreement and stock returns can be
accounting characteristics. We provide evidence that rationalized with a standard asset pricing argument:
disagreement tends to be more prevalent for the Higher ESG rating disagreement may be perceived
largest firms in the S&P 500 (perhaps because of the as a source of uncertainty—in the spirit of Knightian
complexity of such firms) and for firms that do not uncertainty (that is, a lack of any quantifiable
have credit ratings (potentially because the informa- knowledge about some possible occurrence)—that
tion environment for these firms is of lower quality commands an uncertainty premium. This explana-
than for other firms). In contrast, more profitable tion would clarify why uncertainty-averse inves-
firms tend to have lower ESG rating disagreement tors taking on such additional exposure wish to be
(perhaps because they are able to dedicate more compensated by higher expected returns. Consistent
resources to ESG policies and to their ESG disclo- with this conjecture, we found in standard portfolio
sures). We also found that ESG rating disagreement sorts that a portfolio that was long stocks with a high
is orthogonal to disagreement in the EPS forecasts level of disagreement and short stocks with a low
issued by analysts. We further show that rating level of disagreement generated monthly returns of
disagreement is generally more pronounced for firms about 21 bps (2.52% on an annual basis) for disagree-
that belong to the consumer durables and telecom- ment about the overall ESG and the environmental
munications industries, which provides important rating. We found similar magnitudes when adjusting
105
Volume 77 Number 4
Financial Analysts Journal | A Publication of CFA Institute
for well-known risk factors by using standard asset applies to the conclusions stemming from existing
pricing models such as the Carhart (1997) or the empirical studies conducted by academics.
Fama and French (2015) five-factor models.
Given the heightened concern that diverging ESG
Although many studies have examined the relation- ratings have recently generated in both practitioner
ship between stock returns and average ESG rat- circles and the financial press, the topic has also
ings (see, e.g., Friede, Busch, and Bassen 2015 and spurred significant academic interest. Most of the
references therein), ours is the first to systematically contemporaneous work aims primarily at explaining
test whether the second moment of ESG ratings the why—that is, what drives ESG rating disagree-
(i.e., disagreement about a firm’s ESG performance) ment. For instance, Christensen et al. (forthcoming)
has consequences for stock returns. Another impor- focused on the role of disclosure as a determinant
tant contribution of our article is to move beyond of ESG rating disagreement and found that more
simply documenting that disagreement about nonfi- disclosure leads to higher disagreement.5 In addition,
nancial information exists and instead shedding light they pointed out that the relationship between a
on whether such ESG disagreement has real conse- firm’s average ESG rating and ESG rating disagree-
quences for firms, analysts, and investors. We also ment is nonlinear. An important difference between
contribute importantly to the debate on why ESG their study and ours is that they focused primarily
rating disagreement exists by providing evidence of on explaining why disagreement exists, whereas we
the firms’ financial and accounting characteristics are more interested in examining the consequences
that correlate with ESG rating disagreement and arising from ESG rating disagreement and, specifi-
by identifying those industry sectors that are most cally, whether ESG rating disagreement has implica-
prone to ESG rating disagreement. tions for stock returns, a firm’s cost of capital, and an
investor’s performance when investing sustainably.
Overall, our empirical results should help financial
analysts, academics, institutional investors, financial Berg et al. (2020) also explored why ESG rating
advisers, policymakers, regulators, and ultimately, disagreement exists. They pursued a more granular
firms themselves to better understand that, in addi- approach, proposing a decomposition of the sources
tion to the sustainability performance as captured by of ESG rating disagreement. By subdividing the
average ESG ratings, the dispersion of these ratings ESG ratings of six providers into finer categories,
also can have an economically meaningful impact Berg et al. identified three sources of ESG rating
on stock returns and thus on firms’ equity cost divergence. First, they highlighted that raters use
of capital. different categories, which can lead to disagreement.
They referred to this source as “scope divergence.”
Second, they pointed out that ESG raters measure
Literature Review identical categories differently, which they referred
The use of ESG ratings in investment practice to as “measurement divergence.” Finally, they
increased considerably during the last two decades highlighted “weight divergence,” which results from
and has skyrocketed recently. In parallel, ESG ratings raters attaching different weights to the different
are now also commonly used in economics, manage- categories when generating an aggregated ESG rat-
ment, and finance research.2 Given the complexity ing. The authors found that most of the differences
of measuring a firm’s nonfinancial or ESG perfor- can be traced to measurement and scope divergence;
mance, the validity and convergence of these ratings weight divergence seems to play a minor role.
have been debated critically in the management In addition, Berg et al. found a “rater effect”—that is,
literature.3 Chatterji, Durand, Levine, and Touboul ratings of one provider were positively correlated
(2016), for instance, studied the convergence of cor- across different categories. The most important
porate social responsibility (CSR) ratings produced by difference between Christensen et al. and Berg et al.,
six well-established information intermediaries. They on the one hand, and our study, on the other hand, is
documented a lack of agreement among information that their studies focused mainly on explaining why
intermediaries that comes mainly from two sources— ratings disagree whereas we focus on whether there
the absence of a common theorization and lack of are consequences from ESG ratings’ disagreements
commensurability.4 These findings point out that in the form of measurable effects on stock returns.
firms’ and professional investors’ sustainable financ-
In a recent paper, Avramov, Cheng, Lioui, and Tarelli
ing and investment decisions are potentially tainted
(2021) examined the relationships among stock
by the choice of their rating providers. The same
returns, ESG ratings, and ESG rating disagreement.
Using a theoretical model that highlights the inter- We collected data from seven ESG data providers:
play between the average ESG rating and ESG rating (1) Asset4 (now Refinitiv),6 (2) Sustainalytics,7 (3)
disagreement, Avramov et al. showed that the aver- Inrate, (4) Bloomberg, (5) FTSE, (6) KLD,8 and (7)
age ESG rating is negatively associated with future MSCI Intangible Value Assessment (IVA).9 According
stock performance only for low-ESG-disagreement to a survey by Wong, Brackley, and Petroy (2019),
stocks. In contrast to Avramov et al., we focus on the most important and commonly used providers
ESG rating disagreement only. are Sustainalytics, MSCI, Bloomberg, and Asset4.
Later in the article, we report robustness analyses
Studying disagreement in ESG ratings is also in which we restricted disagreement to the ratings
reminiscent of the rich literature on heterogeneous from these four providers. In Appendix A, we provide
beliefs in financial markets (see the discussion in further information about sample selection, dataset
the subsection “Possible Theoretical Explanations”). matching, variable definitions (in Table A1), and
Many empirical studies have tested the relationship summary statistics of the variables used in this study
between dispersion in beliefs and stock returns in (in Table A2).
a variety of settings. These studies typically used
the dispersion in analyst earnings forecasts as a Table 1 displays important features of these seven
proxy for the extent to which a stock is subject to data providers. As shown in the “Origin” column,
heterogeneous beliefs. For example, Diether, Malloy, three providers are US-based (Bloomberg, KLD,
and Scherbina (2002) documented a significant and and MSCI IVA), whereas two providers have their
negative relationship between heterogeneous beliefs origins in Switzerland (Asset4 and Inrate). The other
and stock returns. Anderson, Ghysels, and Juergens two providers can be traced back to origins in the
(2005) reached the opposite conclusion. They argued Netherlands (Sustainalytics) and in the United
that disagreement about expected EPS is an addi- Kingdom (FTSE).
tional priced risk factor and provided supportive
empirical evidence for the pricing of this additional In column 2 of Table 1, we show the rating scales
source of risk. used by each provider. For example, three providers
apply a scale from 0 to 100 for their assessments;
In the subsection “Possible Theoretical Explanations,” Inrate uses a scale of 1 to 12, which is based on
we discuss another stream of the finance literature sustainability assessments ranging from D– to A+.
that would allow rationalizing the excess return for Originally, MSCI KLD did not provide a genuine scale
stocks with high ESG rating disagreement—namely, itself. However, many academic studies (e.g., Lins,
that these excess returns are a compensation for Servaes, and Tamayo 2017) have summed up KLD’s
ESG information uncertainty (see also Viale, Garcia- judgment as to “strengths” and “concerns” separately
Feijoo, and Giannetti 2014). and scaled both by the total number of strengths and
concerns available. This course of action results in a
scale of –1 to +1. Note that KLD also has strengths
Data and concerns items for norms-oriented categories
To test how stock returns are related to ESG rating related to alcohol, military, firearms, gambling,
disagreement, we constructed a representative and nuclear, and tobacco, which we decided to ignore.
homogeneous sample with available ESG ratings over
Because the rating scales differ not only in terms
the longest possible period. We faced the challenge
of their statistical support but also in terms of the
that the availability of ESG data is restricted in both
distribution across the statistical support, a simple
the cross-section and the time series. This specific
rescaling would not suffice to make the different
limitation is not unique to our setting but applies
ratings comparable. Therefore, we did the following
generally to research concerned with ESG. To use
to achieve comparability across rating providers: At
a sample as homogeneous as possible and to maxi-
each point in time, we sorted all stocks according
mize the number of available ESG ratings per firm
to the ratings of the respective providers. We then
(as well as the time-series dimension of the panel),
calculated the individual rating-specific percentile
we restricted our study to firms belonging to the
ranks and used these as adjusted scores. Using
S&P 500 and considered a sample period from 2010
ranked measures is also consistent with investment
through 2017.
practice, in which investors compare the ranked
We used financial data from CRSP and account- value of a given signal relative to the ranked values
ing data from Standard & Poor’s Compustat. of the signals for other firms. In the case of ties,
107
Volume 77 Number 4
Financial Analysts Journal | A Publication of CFA Institute
Note: The data dimensions (e.g., environmental, social, and governance) are referred to as “pillars.” The “L” here stands for “labor.”
we assigned each company the average rank. We an absolute ESG rating, whereas providers such as
normalized these ranks between 0 and 1. Sustainalytics and MSCI provide best-in-class ratings.
Also, some rating providers (e.g., Bloomberg) are
Column 4 of Table 1 shows the average number of more geared toward capturing ESG disclosure quality
sample stocks per year for which we observed an than are other providers.
ESG rating from a given data provider. Sustainalytics,
MSCI KLD, MSCI IVA, and Bloomberg had, on aver-
age, the best coverage (about 460 stocks). Inrate, Analysis
Asset4, and FTSE had the least number of stocks, on
In this main section of our article, we begin with
average, with 434, 438, and 442, respectively. The
descriptive statistics and correlations, consider the
average number of stocks for all providers is rather
financial and accounting determinants of ESG rating
high—well above 400—and we therefore considered
disagreement, and then turn to the main research
the sample to be representative for S&P 500 compa-
question: the relationship between stock returns and
nies. Note, however, that the Inrate and FTSE ratings
ESG rating disagreement. Next, we evaluate whether
were available only for a subperiod of the overall
ESG rating disagreement provide a profitable signal
sample period. Therefore, we report in column 3 the
for investing by reporting the results for portfolios
period for which data were available from a given
based on ESG rating disagreement. We then consider
provider.
the possible theoretical explanations for the relation-
The fifth column of Table 1 reports the pillar scores ship between ESG rating disagreement and stock
supplied by the providers as well as their rating returns, and finally, we discuss whether ESG rating
styles. All providers supplied a total ESG score, an disagreement affects standard equity risk measures.
environmental score, a social score, and a (corporate)
governance score. In addition, Inrate provided a Descriptive Statistics and Correlations.
labor score. Because the labor score captures a social Table 2 provides summary statistics and Pearson
topic, we used the average of the original social score correlations between the ESG ratings from the seven
and the labor score as the social score. In column 6, different data providers. We show the results for
we highlight the rating styles used by these various the total rating and the E, S, and G pillars in separate
data vendors, which may partially explain disagree- panels. The first three columns display descrip-
ments. For instance, MSCI KLD and Inrate provide tive statistics for the ranked ESG scores from the
Pearson Correlations
Data N Mean StD. EPS Asset4 Sust. Inrate Bloom. FTSE KLD
Provider (1) (2) (3) (4) (5) (6) (7) (8) (9)
A. Total pillar
Asset4 42,087 0.501 0.289
Sustainalytics 44,078 0.501 0.289 0.752
Inrate 26,037 0.501 0.284 0.233 0.303
Bloomberg 44,464 0.501 0.289 0.750 0.693 0.124
FTSE 17,220 0.501 0.288 0.568 0.614 0.267 0.586
KLD 44,951 0.501 0.288 0.524 0.559 0.292 0.477 0.488
MSCI IVA 43,775 0.501 0.289 0.396 0.434 0.318 0.303 0.266 0.439
Average correlation 0.447
B. Environmental pillar
Asset4 42,019 0.501 0.289
Sustainalytics 44,020 0.501 0.289 0.706
Inrate 26,036 0.501 0.286 0.305 0.487
Bloomberg 37,624 0.501 0.289 0.647 0.557 0.206
FTSE 17,220 0.501 0.288 0.654 0.678 0.368 0.607
KLD 44,669 0.501 0.280 0.575 0.609 0.422 0.431 0.581
MSCI IVA 43,580 0.501 0.289 0.233 0.352 0.404 0.187 0.239 0.312
Average correlation 0.455
C. Social pillar
Asset4 42,087 0.501 0.289
Sustainalytics 44,078 0.501 0.289 0.617
Inrate 26,037 0.501 0.288 0.133 0.143
Bloomberg 44,364 0.501 0.288 0.685 0.527 0.062
FTSE 17,220 0.501 0.288 0.637 0.501 0.106 0.560
KLD 44,951 0.501 0.288 0.367 0.391 0.129 0.276 0.271
MSCI IVA 43,775 0.501 0.289 0.266 0.303 0.236 0.202 0.191 0.337
Average correlation 0.330
D. Governance pillar
Asset4 42,087 0.501 0.289
Sustainalytics 44,078 0.501 0.289 0.331
Inrate 26,037 0.501 0.283 0.297 0.401
Bloomberg 44,464 0.501 0.282 0.432 0.327 0.344
FTSE 17,220 0.501 0.288 0.027 0.160 –0.029 –0.027
KLD 44,951 0.501 0.248 0.104 0.089 0.081 0.153 –0.065
MSCI IVA 43,775 0.501 0.288 0.132 0.135 0.145 0.060 0.023 0.133
Average correlation 0.155
Note: We display in the last row of each panel the average pairwise correlation between providers. “StD. EPS” = disagreement
about EPS.
109
Volume 77 Number 4
Financial Analysts Journal | A Publication of CFA Institute
different providers. The subsequent columns display performance, and GHG emissions are increasingly
the pairwise cross-correlations. We also show in the measured. In a similar spirit, nowadays firms quantify
last row of each panel the average pairwise correla- water and electricity use. Although measurement of
tion between providers, calculated as the mean of emissions is certainly not without problems (scal-
the respective pairwise cross-correlations (separately ing of emissions, missing emissions data/imputa-
for the total rating and the E, S, and G pillars).10 tion, voluntary reporting), firms at least have some
basic guidance about how to measure them (see,
Note, first, that the average pairwise correlation for instance, the Greenhouse Gas Protocol11).12
for the overall ESG ratings in Table 2 is 0.45, which In contrast, we think that when it comes to S and
is much lower than average correlations between G, there is (1) less agreement among providers on
credit ratings issued by Moody’s Investors Service what the most important issues are and (2) a worse
and Standard & Poor’s. According to Berg et al. understanding of how to quantify these issues.
(2020), correlations between those two credit-rating
providers exceed 0.99. A point worth mentioning is In Figure 1, we can see whether the average pair-
that despite the commonly held belief of generalized wise correlations between ESG ratings vary at the
ESG rating disagreement, the analysis of pairwise industry level. We plotted the average correlations
correlations in Table 2 also highlights more subtle across the seven ESG rating providers for each of
patterns in ESG disagreement, in that the pairwise the 12 Fama and French industries.13 Some industry
correlations between ratings from some providers heterogeneity appears when it comes to correlations
can be relatively high. For instance, the correlations between ESG ratings. Average correlations in the
between the total rating of Asset4 and Sustainalytics total ratings (Panel A) are lowest in the consumer
and Asset4 and Bloomberg are about 0.75, and that durables and telecommunications sectors. The low
between Sustainalytics and Bloomberg is about average pairwise correlation in the total rating for
0.69. This pattern may result from their similar rating the consumer durables industry seems to be driven
styles, as documented in Table 1. by the low pairwise correlations in the environmental
ratings (Panel B) and governance ratings (Panel D).
Next, we separately examine the average pairwise The low average correlation of the telecommunica-
correlations between providers for the E, S, and tions sector results from the low correlations of
G pillars (Panels B–D). Apart from the environmental the social ratings (Panel C). In contrast, ESG data
pillar, these average pairwise correlations are gener- providers seem to disagree the least (i.e., exhibit high
ally lower than for the total rating, which probably average correlations) as to the total rating (Panel A)
results from discrepancies in aggregation and weight- in the business equipment and manufacturing sec-
ing procedures across the three pillars. Surprisingly, tors. Another interesting observation is that rating
the average correlation is lowest for the governance providers also seem to disagree quite strongly (i.e.,
pillar (0.16) and highest for the environmental pil- correlations are low) about governance in the finance
lar (0.46), but this result can also be rationalized: sector. These findings on industry-variation in firms’
Environmental issues can be increasingly measured ESG ratings could help industry-focused financial
and quantified (e.g., water usage, greenhouse gas analysts distinguish their assessments and compari-
emissions), but the criteria applied to quantifying sons of firms, in that analysts are often confronted
governance may likely differ among rating providers. with metrics coming from various data providers.
In a similar spirit, the social rating is also likely to
require more value judgments and is thus inherently Determinants of ESG Rating Disagreement.
more subjective than the environmental pillar, sug- We do not see our main contribution as studying
gesting more disagreement among raters (i.e., lower determinants of ESG rating disagreement, but the
correlations). comprehensiveness of our sample (in terms of rating
providers covered) as well as the representativeness of
Our main argument for stating that the E rating
the sample firms (S&P 500 firms) allow us to contrib-
is more objective and measurable than the oth-
ute also to this literature. Hence, to add to the existing
ers is this: We believe that not only is there more
literature that traces the origins to disagreement in
agreement on the issues that are important in the
scopes, measurement, rating methodologies, and ESG
environmental dimension but also there are more
firms’ disclosure policies, we examined whether ESG
systematic, regulation-driven, firm-level attempts to
rating disagreement also correlates with observable
quantify these dimensions. For instance, a consensus
firm-level financial and accounting characteristics. We
now exists that greenhouse gas (GHG) emissions are
used the standard deviation of ESG ratings available
an important dimension of a firm’s environmental
Nondurables Nondurables
Consumer Durables Consumer Durables
Manufacturing Manufacturing
Energy Energy
Chemicals Chemicals
Business Equipment Business Equipment
Telecommunicaons Telecommunicaons
Ulies Ulies
Retail/Wholesale Retail/Wholesale
Health Care Health Care
Financials Financials
Other Other
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
Nondurables Nondurables
Consumer Durables Consumer Durables
Manufacturing Manufacturing
Energy Energy
Chemicals Chemicals
Business Equipment Business Equipment
Telecommunicaons Telecommunicaons
Ulies Ulies
Retail/Wholesale Retail/Wholesale
Health Care Health Care
Financials Financials
Other Other
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
Notes: Shown are average pairwise Pearson correlations between the normalized ratings of the ESG data providers for each of the
12 Fama and French industries. The vertical line represents the average correlation across all industries.
for a given firm at a given point in time as the depen- disagreement (see columns 1 and 3), as do larger
dent variable. We calculated this measure for the total firms (see columns 3 and 4).14 These results seem
rating and also separately for the E, S, and G pillars. intuitive: Profitable firms may be viewed less criti-
We explored how the levels of rating disagreement cally by ESG analysts, perhaps because they have
correlate with variables falling in one of the follow- more resources available to shape and disclose their
ing five categories: (1) balance sheet–related data, (2) ESG policies. In contrast, firms without a credit rating
industry-related data, (3) investor transparency, (4) are subject to a less transparent information environ-
valuation, and (5) price. ment, making their assessment in terms of ESG more
difficult. Also, large firms might be more diversified
We used pooled panel regressions in which the and complex and are also analyzed more thoroughly
rating disagreement measures served as dependent by ESG data providers, explaining why they exhibit
variables. We also included industry-month fixed higher rating disagreement. Finally, note that ESG
effects. Standard errors were double clustered at the rating disagreement is orthogonal to disagreement
firm and month levels. Table 3 displays the regression about EPS (StD. EPS in Table 3).
results for disagreement about the total ratings and
separate results for disagreement about the pillars. In addition, three other variables seem to affect dis-
agreement about individual pillars of the ESG rating.
Essentially, Table 3 shows that three financial For instance, tangibility plays a specific role, in that
variables play a role in explaining rating disagree- firms with more tangible assets tend to have lower
ments. First, the more profitable firms (in terms of disagreement in their environmental ratings. Again,
gross profitability) are subject to lower ESG rating this result seems intuitive, because firms with more
disagreement (see columns 1 and 2). Second, firms tangible assets are also likely to have more nega-
without a credit rating exhibit higher ESG rating tive impacts on the environment (e.g., higher GHG
111
Volume 77 Number 4
Financial Analysts Journal | A Publication of CFA Institute
Note: We measured disagreement as the standard deviation of all firm-level ratings available for a given firm at a given time.
The explanatory variables are the following: tangibility, current ratio, leverage, gross profitability (see Novy-Marx 2013), the
Herfindahl–Hirschman index (HHI) for measuring industry concentration, multisegment, no credit rating, institutional ownership,
number of analysts, dispersion of analyst forecasts of the firm’s one-year-ahead earnings forecasts (StD. EPS; see Diether et al.
2002), book-to-market ratio (see Fama and French 1995), market capitalization (see Banz 1981), momentum (see Jegadeesh and
Titman 1993), and total volatility (see Ang, Hodrick, Xing, and Zhang 2006). We included industry-month fixed effects (FE). The
t-statistics are based on double-clustered standard errors (month and firm). Bold indicates significance at the 5% level.
emissions) and thus potentially more easily measur- In column 2, we find a coefficient estimate for
able environmental ratings. Moreover, higher levels disagreement about the environmental rating of
of institutional ownership are also associated with about 1.0, which is both economically larger and
higher disagreement in the environmental rating. more significant (t-statistic = 2.38) than the estimate
Firms with high book-to-market ratios display higher for disagreement about the total rating. The coef-
disagreement in their social ratings. ficient for disagreement about the social rating (see
column 3) is positive (0.6) but not significant at con-
ESG Rating Disagreement and Stock ventional levels (t-statistic = 1.515). The coefficient
Returns. We now turn to our main research estimate for disagreement about the governance
question and examine the relationship between rating (see column 4) is small (–0.06) and insignificant
stock returns and ESG rating disagreement. As in (t-statistic = –0.136). We conclude that disagree-
the previous analyses, we used pooled panel regres- ment about the environmental rating primarily drives
sions with standard errors double clustered at the our results, with an interquartile range increase in
firm and month level. We used monthly stock returns disagreement about the environmental rating result-
as the dependent variable in the regressions. In ing in an increase of about 132 bps [(0.25 – 0.14) ×
addition to our main disagreement-related explana- 1 × 12] in a firm’s annual cost of equity capital.
tory variables, which we measured as the standard
deviations of ratings available for a given firm at Given that ratings from Inrate and FTSE were
monthly intervals and denote as Disp, we included not available for the entire sample period, we
industry-month fixed effects. We also controlled for re-estimated the regressions by using a disagree-
standard characteristics that have been found to ment measure based only on Asset4, Sustainalytics,
explain the cross-section of stock returns: market Bloomberg, MSCI KLD, and MSCI IVA. The results
capitalization (Banz 1981), book-to-market ratio are reported in Table B1 of Appendix B and continue
(Fama and French 1995), gross profitability (Novy- to show a strong positive relationship between
Marx 2013), momentum (Jegadeesh and Titman ESG rating disagreement and stock returns. In fact,
1993), dispersion of analyst forecasts of the firm’s the results in these regressions are stronger both
one-year-ahead earnings (Diether et al. 2002), the statistically and economically speaking than the
firm’s beta (Frazzini and Pedersen 2014), and total results shown in Table 4 for disagreement about the
volatility (Ang et al. 2006). Conceptually, pooled E and S pillars separately. The t-statistic for disagree-
panel regressions with industry-month fixed effects ment about the environmental rating increases to
are similar to Fama and MacBeth (1973) type regres- 3.01, and disagreement about the social rating has a
sions with industry dummies. Because our sample t-statistic of 1.72.
period is relatively short, and also because ESG rat-
The reader may wonder why the control variables
ing disagreement varies among sectors (see Figure 1),
in the return regressions of Table 4 do not turn out
controlling for differences at the industry level was
to be significant. This result might have a variety
important.
of reasons. First, the control variables are known
The coefficient estimates for the main explanatory return predictors, and there is evidence of lower
variable (Disp) and the control variables appear in post‐publication return predictability for them
Table 4. We also show t-statistics based on double- (McLean and Pontiff 2016), especially in the United
clustered standard errors. States (Jacobs and Müller 2020). Other research
also has shown that return predictability fell sharply
Table 4 reports a positive and significant coefficient after 2003 (Green, Hand, and Zhang 2017). In a
estimate for the total ESG rating disagreement proxy, similar spirit, Chordia, Subrahmanyam, and Tong
suggesting that firms with higher disagreement (2014) showed that capital market anomalies have
tend to have higher stock returns. The regression in attenuated in recent periods, coinciding with periods
column 1 shows an estimated coefficient of about 0.7 that have been accompanied by significant liquid-
for disagreement about the total ESG rating. In terms ity increases. Given the combined findings of these
of economic magnitude, consider a firm moving from studies, and noting that we examined a sample of
the first quartile (0.14) to the third quartile (0.25) highly liquid S&P 500 firms since 2010, insignificant
of the ESG rating disagreement distribution. Such control variables are perhaps not that surprising.
a move would imply an increase of about 92 bps in
annual stock returns [(0.25 – 0.14) × 0.7 × 12], an Overall, we conclude that higher stock returns for
estimate that seems plausible in terms of magnitude. firms with higher ESG rating disagreement is consis-
tent with the view that risk-averse investors perceive
113
Volume 77 Number 4
Financial Analysts Journal | A Publication of CFA Institute
Note: We measured ESG rating disagreement by the standard deviation of ratings available for a
given firm at a given point in time, Disp. Included also are industry-month fixed effects and controls
for standard characteristics that have been found to explain stock returns. The t-statistics are based
on double-clustered standard errors (month and firm). Bold indicates significance at the 5% level.
a dispersed ESG performance of a given firm as an would be driven by industry effects. As Figure 1
additional source of risk (or uncertainty) that com- shows, average ESG rating correlations, and thus dis-
mands a separate risk premium. agreement, exhibit important variation by industry,
and we wanted to ensure that the portfolio composi-
Portfolio Sorts on ESG Rating tions in the sorts were not biased by these industry
Disagreement. In the previous subsection, we differences in ESG rating disagreement. Note that in
documented a positive relationship between ESG the return regressions of Table 4, we also controlled
rating disagreement and stock returns. To evaluate for industries by including industry-month fixed
whether ESG rating disagreement provides a profit- effects, implying that our insights and results are
able signal for investing, we now report our imple- driven by differences in disagreement within, and not
mentation of portfolios sorted on the basis of ESG between, industries.
rating disagreement.
Specifically, for the portfolio sorts, we calculated
In the portfolio sorts, we used industry-adjusted ESG the industry-adjusted ESG rating disagreement
rating disagreement as the sorting variable. We did for a given firm in a given month by simply de-
so primarily to rule out the possibility that our results meaning stock-level ESG rating disagreement using
the average ESG rating disagreement in the firm’s advocated by Novy-Marx and Velikov (2016). These
Fama–French 12-industry classification in a given authors used the effective bid–ask spread measure
month. We denote this variable as Disp_adj, and of Hasbrouck (2009) to proxy for trading costs.
we sorted all stocks in the sample into five quintile Hasbrouck suggested a Bayesian Gibbs sampler
groups based on their industry-adjusted ESG rating approach to the Roll (1984) model of price dynamics.
disagreement. Table 5 shows the results. Hasbrouck showed that his estimate of the average
effective cost is comparable to estimates from high-
For each of the four rating dimensions, we report in frequency trade and quote (TAQ) data and reported
Table 5 results for the equally weighted portfolios of a 0.965 Pearson correlation of his measure with the
stocks (quintile) with the lowest industry-adjusted TAQ value. Novy-Marx and Velikov noted that this
rating disagreement (Low Disp_adj), the high- measure does not account for the price impact of
est industry-adjusted rating disagreement (High large trades, but they nevertheless considered it an
Disp_adj), and the long–short portfolio of high- and appropriate trading cost measure (for details, see
low-disagreement stocks. Returns are all expressed in Novy-Marx and Velikov, p. 108).
excess of the risk-free rate. We report mean monthly
returns, the median number of firms for the low- and To calculate the actual trading costs for each port-
high-disagreement portfolios, standard deviations, folio, we subtracted trading costs from portfolio
Sharpe ratios, and alphas computed from four differ- returns each time a given stock was introduced into a
ent factor models. We observe in Panel A of Table 5 portfolio or withdrawn from it.
that for the total rating, the equally weighted long–
short portfolio of high- and low-disagreement stocks Table 6 displays the results of the long–short strate-
generated a mean raw monthly return of 21 bps (with gies after accounting for transaction costs. Note
a t-statistic of 2.2). The factor model alphas are of that the raw returns, Sharpe ratios, and alphas are
a similar magnitude, with average monthly alphas somehow diminished by trading costs but still siz-
ranging between 23 bps and 27 bps (and t-statistics able. For example, for the total rating disagreement,
between 2.0 and 2.7). Furthermore, we observe that the monthly raw return of the long–short portfolio
a long-only strategy of high-disagreement stocks gen- of high- and low-disagreement stocks drops from
erated a raw monthly mean return of 134 bps (with a 21 bps in Table 5 to 18 bps in Table 6. The t-statistic
t-statistic of 3.5). The alphas of the long-only strategy, drops from 2.2 to 1.9. The monthly alphas range from
ranging between 14 bps and 21 bps (with t-statistics 19.5 bps to 23 bps (with t-statistics of 1.8 to 2.4). For
between 1.6 and 2.4), are somewhat lower. environmental rating disagreement, we observe sig-
nificant monthly alphas for the CAPM and the Fama–
For the environmental dimension (Panel B of Table 5), French three-factor and Carhart four-factor models.
we observe similar results. The long–short portfolio These alphas range between 17 bps and 21 bps (with
generated a mean monthly raw return of 21 bps t-statistics between 2.0 and 2.2). The long–short raw
(with a t-statistic of 2.0). The factor model alphas returns and the alphas from the Fama–French five
range between 21 bps and 25 bps (with t-statistics factor model, however, have t-statistics of about 1.6
between 2.0 and 2.5). The long-only strategy also in both cases (and mean returns of about 17 bps). As
shows a high average monthly raw return, about in the case without transaction costs, the long–short
140 bps (t-statistic of 3.5), and also high average portfolios for the social and the governance rating
monthly factor alphas, ranging from 16 bps to 21 bps disagreements display insignificant raw returns and
(with t-statistics between 1.8 and 2.3). For disagree- risk-adjusted alphas.
ment about the social and governance ratings, we
did not obtain any significant raw returns or alphas Possible Theoretical Explanations. A poten-
from the long–short strategies. Taken together, the tial way of rationalizing the positive relationship
insights from the portfolio sorts in Table 5 are con- between ESG ratings disagreement and stock returns
sistent with the evidence presented in the regression may be found in the literature on heterogeneous
analysis shown in Table 4 and emphasize the promi- beliefs in financial markets. Theoretical models of
nent role played by environmental rating disagree- heterogeneous beliefs (e.g., Atmaz and Basak 2018)
ment in explaining stock returns. provide some predictions for empirical studies. Note,
however, that these theories are based on beliefs
An important question is whether these strategies about factors that affect the returns of firms (e.g.,
are implementable in a way that survives account- consensus EPS forecasts), not factors that may or
ing for trading costs. To gain a realistic impression may not have risk and return implications, such as
as to whether they are, we followed the approach
115
Volume 77 Number 4
Financial Analysts Journal | A Publication of CFA Institute
A. Total
Low Disp_adj 1.124 94 3.809 0.295 –0.088 –0.068 –0.055 –0.074
(2.891) (–1.260) (–0.970) (–0.794) (–1.049)
High Disp_adj 1.336 94 3.769 0.355 0.144 0.164 0.211 0.165
(3.474) (1.561) (1.792) (2.425) (1.611)
H–L Disp_adj 0.212 0.942 0.225 0.232 0.232 0.267 0.239
(2.209) (2.151) (2.192) (2.664) (2.044)
B. Environmental
Low Disp_adj 1.186 93 4.057 0.292 –0.091 –0.043 –0.016 –0.036
(2.866) (–0.902) (–0.500) (–0.187) (–0.455)
High Disp_adj 1.397 93 3.906 0.358 0.163 0.183 0.212 0.178
(3.504) (1.848) (2.040) (2.274) (1.982)
H–L Disp_adj 0.211 1.026 0.205 0.254 0.226 0.228 0.213
(2.010) (2.482) (2.226) (2.322) (2.005)
C. Social
Low Disp_adj 1.283 93 3.841 0.334 0.071 0.121 0.139 0.116
(3.273) (0.865) (1.734) (1.943) (1.950)
High Disp_adj 1.330 94 3.996 0.333 0.067 0.096 0.134 0.081
(3.262) (0.658) (1.077) (1.748) (0.998)
H–L Disp_adj 0.047 0.965 0.049 –0.004 –0.024 –0.004 –0.035
(0.480) (–0.058) (–0.338) (–0.073) (–0.479)
D. Governance
Low Disp_adj 1.246 94 3.888 0.320 0.013 0.050 0.071 0.057
(3.140) (0.129) (0.535) (0.777) (0.702)
High Disp_adj 1.284 94 3.794 0.338 0.088 0.118 0.167 0.120
(3.315) (1.320) (1.871) (2.749) (2.577)
H–L Disp_adj 0.038 1.006 0.037 0.075 0.069 0.096 0.063
(0.366) (0.742) (0.680) (0.954) (0.696)
Note: Reported are mean returns (Return), the median number of observations for the high- and low-disagreement portfolios (N),
the standard deviations of returns (StD.), and the Sharpe ratio. In addition, we report alphas of the capital asset pricing model
(CAPM) and the Fama–French three-factor, Carhart four-factor, and Fama–French five-factor models. The sample includes 96
monthly time-series observations. Portfolios were formed each January with disagreement values from December of the preced-
ing year. Returns are reported in percentages. The t-statistics for the factor model alphas are based on Newey–West (1987)
standard errors with 12 lags. Bold indicates significance at the 5% level.
ESG ratings. Nonetheless, we think that discuss- In an important paper, Atmaz and Basak (2018,
ing our findings in the light of this literature will be p. 1241) argued that “. . . dispersion represents an
useful in our setting without attempting to test a additional risk for investors and therefore inves-
specific theory. tors demand a higher return to hold the stock when
Table 6. Portfolio Sorts on Industry-Adjusted ESG Rating Disagreement Adjusted for Trading
Costs, January 2010–December 2017 (t-statistics in parentheses)
A. Total
Low Disp_adj 1.119 93 3.797 0.295 –0.089 –0.068 –0.056 –0.078
(2.887) (–1.288) (–0.988) (–0.820) (–1.110)
High Disp_adj 1.330 94 3.762 0.354 0.139 0.159 0.206 0.162
(3.465) (1.500) (1.715) (2.365) (1.567)
H–L Disp_adj 0.181 0.934 0.194 0.196 0.195 0.227 0.203
(1.897) (1.890) (1.906) (2.370) (1.830)
B. Environmental
Low Disp_adj 1.178 93 4.054 0.291 –0.097 –0.050 –0.023 –0.041
(2.848) (–0.977) (–0.575) (–0.276) (–0.510)
High Disp_adj 1.379 93 3.889 0.355 0.150 0.169 0.197 0.162
(3.474) (1.775) (1.964) (2.247) (1.882)
H–L Disp_adj 0.168 1.025 0.164 0.214 0.185 0.185 0.166
(1.609) (2.155) (1.861) (1.939) (1.594)
C. Social
Low Disp_adj 1.270 93 3.835 0.331 0.059 0.109 0.126 0.102
(3.243) (0.742) (1.589) (1.758) (1.690)
High Disp_adj 1.320 93.5 3.966 0.333 0.065 0.089 0.125 0.075
(3.260) (0.681) (1.052) (1.693) (0.985)
H–L Disp_adj 0.017 0.958 0.018 –0.029 –0.055 –0.038 –0.066
(0.172) (–0.443) (–0.777) (–0.630) (–0.954)
D. Governance
Low Disp_adj 1.225 93 3.929 0.312 –0.021 0.020 0.040 0.025
(3.054) (–0.201) (0.203) (0.421) (0.297)
High Disp_adj 1.266 93 3.821 0.331 0.062 0.091 0.141 0.090
(3.247) (0.934) (1.444) (2.274) (1.916)
H–L Disp_adj 0.005 1.031 0.005 0.045 0.033 0.060 0.022
(0.044) (0.401) (0.302) (0.562) (0.227)
Note: See the notes to Table 5. We adjusted returns for trading costs according to Novy-Marx and Velikov (2016). Bold indicates
significance at the 5% level.
dispersion is higher.” In their setting, this impact analysts’ EPS forecast dispersion enhanced explana-
may have been, however, attenuated (reinforced) tory power in explaining S&P 500 excess returns
in the presence of investors’ excessive optimism after accounting for standard market risk factors.
(pessimism). Along the same lines, Anderson et al. Extended to our context, that approach would
(2005) found that adding empirical factors based on imply that ESG rating disagreement risk is priced
117
Volume 77 Number 4
Financial Analysts Journal | A Publication of CFA Institute
in expected stock returns in addition to the market Oikonomou, Sautner, Starks, and Zhou 2019; Gibson
risk exposure of stocks. In other words, our finding et al. 2020). Hence, we regressed standard risk
of a positive relationship between stock returns and measures on our ESG rating disagreement proxies
ESG rating disagreement is in line with the risk-based and a set of common control variables.15 Table 7
explanation for analysts’ EPS disagreement empiri- reflects our focus on the three risk measures—total
cally documented by Anderson et al. According to volatility, idiosyncratic volatility, and beta. Following
this risk-based explanation, more total (or environ- prior studies, we also analyzed whether downside
mental) rating disagreement implies more uncer- risk, as measured by the lower partial moment (LPM),
tainty about the ESG performance of a given firm is related to the level of ESG rating disagreement
and thus would be perceived as a separate source of (see Hoepner et al. for details on the relationship
risk that commands a risk premium if investors are between ESG policies and LPMs).
risk averse. We found supportive empirical evidence
that this explanation is correct for environmental and Table 7 shows that some risk measures are positively
total rating disagreement. So, we have thereby added related to ESG rating disagreement—in particular, to
to this stream of literature by empirically document- disagreement about the social dimension. For exam-
ing the effects of heterogeneity in beliefs about ple, total volatility, idiosyncratic volatility, and LPM
nonfinancial information on stock excess returns. all appear to be positively related to disagreement
about the social rating (see column 3). In addition, we
Another possible explanation for our findings is that found some evidence that idiosyncratic and down-
disagreement about a firm’s ESG rating is a proxy for side risk are related to total rating disagreement.
ESG uncertainty and therefore captures a specific
form of (Knightian) uncertainty. Whereas risk is asso- We believe that in light of the second theoretical
ciated with the uncertain outcome of a known prob- explanation provided in the previous subsection
ability distribution of returns, ambiguity (or Knightian (“Possible Theoretical Explanations”), to observe
uncertainty) is associated with uncertainty regarding that ESG rating disagreement is only weakly related
the probability distribution itself (see, for example, to standard risk measures is not surprising. The
Viale et al. 2014). Theoretical papers have advocated uncertainty-based explanation is consistent with the
using a two-factor model, with one factor a proxy fact that ESG rating disagreement has little power
for risk and the other factor a proxy for uncertainty in explaining standard risk measures. Therefore,
(see, in particular, Kogan and Wang 2003). Anderson, our results suggest that ESG information uncer-
Ghysels, and Juergens (2009) estimated a two-factor tainty offers explanatory power beyond traditional
model with a factor proxy for uncertainty based on risk factors.
professional forecasters’ disagreement. Viale et al.
further showed that uncertainty (or ambiguity) is Study Limitations. Our study has several limita-
priced in the cross-section of stock returns and not tions. First, the fact that we focused on the S&P 500
subsumed by standard risk factors. Hence, in the universe leaves open the question of whether our
context of our setting, a possible explanation for results can be transposed to other stock markets
the positive relationship between ESG rating dis- located outside of the United States. Also unclear
agreement and stock returns is that our ESG rating is whether and how the conclusions hold for larger
disagreement measure is a proxy for uncertainty cross-sections of stocks. Second, to encompass as
regarding ESG information—in particular, about many data providers as possible, we had to work with
environmental information. a limited time period. Thus, the power of our tests
might be an issue. This issue may explain why we
ESG Rating Disagreement and Equity found no significant effects for disagreement about
Risk. For completeness, we examine in this sub- the S and G ratings. Third, some of the data vendors
section whether ESG rating disagreement affects may have changed their rating methods during
standard equity risk measures. An examination our sample period, which would create additional
of the relationship between equity risk and ESG biases when measuring ESG rating disagreement.
disagreement is natural because prior research Finally, we focused only on stocks, but exploring
highlights a robustly negative relationship between whether our results can be transposed to fixed-
stock-level risk and the quality of a firm’s ESG rating income securities that have ESG ratings might be
(Dunn, Fitzgibbons, and Pomorski 2018; Hoepner, interesting.
Note: For the calculation of idiosyncratic volatility and stock market beta, we used the CAPM.
The t-statistics are based on double-clustered standard errors (month and firm). Bold indicates
significance at the 5% level.
119
Volume 77 Number 4
Financial Analysts Journal | A Publication of CFA Institute
Our results have important practical consequences. possible period. We faced the challenge that the
First, our analysis shows that financial analysts who availability of ESG data is restricted in both the cross-
value the equity of firms should incorporate the section and the time series. In other words, ESG
effects of ESG rating disagreement and adjust esti- data are often available only for the largest firms and
mates of firms’ equity cost of capital upward. Second, for recent years. To use a sample as homogeneous
CFOs deciding about the allocation of capital as possible and to maximize the number of avail-
expenditures should consider ESG rating disagree- able ESG ratings per firm, we restricted ourselves
ment in their capital budgeting decisions because to firms belonging to the S&P 500 and considered
such disagreement raises the investment threshold a sample period of eight years, from 2010 to 2017.
for firms subject to substantial total (and environ- See Table A1 for an overview of the variables used in
mental) rating disagreement. Third, our evidence that this study.
ESG rating disagreement varies across industries is
an important insight for financial analysts, who often
focus on specific industries. Financial Data
We used financial data from CRSP and accounting
Finally, our analysis also has important implica-
data from Standard & Poor’s Compustat. For each
tions for asset owners and investment managers
stock, we calculated idiosyncratic volatility, total
who implement responsible investment strategies.
volatility, and the stock market beta at the end of
In today’s responsible investment landscape, two
each month for up to 250 daily observations (we
strategies are currently very popular—screening and
required a minimum of 60 daily observations). We
ESG integration. If asset managers and investment
calculated market capitalization as (adjusted) total
managers following these strategies wish to optimize
shares outstanding times stock price, both at the end
financial performance while investing responsibly,
of the month. We calculated the momentum signal
they should care about ESG rating disagreement and
at time t as the continuously compounded returns
its impact on stock returns. Indeed, our results sug-
from month t – 2 to month t – 12. Book value of
gest that with positive (negative) screening, manag-
equity is the sum of shareholders’ equity, deferred
ers should buy (sell) primarily those stocks that, for
taxes, and the investment tax credit minus preferred
a given high (low) ESG rating, command the lowest
stock.16 Only firms with a positive book value were
(highest) level of ESG disagreement. This practice
selected for the sample. Following Novy-Marx
should allow positive (negative) screeners to mitigate
(2013), we calculated gross profitability as total
the adverse impact of ESG rating disagreement on
revenues minus cost of goods sold, divided by total
the expected future returns of their buy (sell) orders.
assets. In addition, we matched the dispersion in EPS
Similarly, an ESG integration strategy may fail to
forecasts for one-year-ahead earnings from IBES
deliver its financial promises if it does not search
(Diether et al. 2002).
simultaneously for those stocks that have superior
ESG ratings and embed the lowest level of ESG rating
disagreement within an industry. Indeed, controlling
for a low level of ESG rating disagreement will allow Dataset Matching
investors who integrate ESG criteria in their stock A challenge for constructing a dataset from many
selection processes to avoid a subsequent unin- subsets of data is to properly match the different
tended stock price decline. sets. We matched on the basis of three identifiers:
(1) CUSIP, (2) ISIN, and (3) company name. The
CUSIP code was available for all the providers except
Inrate.17 Because the ISIN code was available for
Appendix A. Sample Selection, Inrate, however, we extracted the CUSIP code from
the ISIN code. Note that we used only the first six
Financial Data, Dataset CUSIP characters for matching (known as the issuer
Matching, and Variable identifier). Characters seven and eight identify the
specific issue (for example, if the seventh character
Definitions is 1 and the eighth character is 0, this indicates
common equity), and the ninth character is a check
Sample Selection digit. The ISIN code was available for all providers
except MSCI KLD. For the CRSP/Compustat data,
To test our hypotheses, we constructed a represen-
we retrieved the ISIN number from the CUSIP code
tative and homogeneous sample over the longest
and the current ISO (International Organization In addition, we used a monthly frequency for our
for Standardization) country code of incorporation sample. Asset4, Sustainalytics, FTSE, and MSCI IVA
(fic).18 To perform the merge with the company already provided data at a monthly frequency; Inrate
names, we first converted the original names of the provided rating updates on a semiannual basis for
providers by using some common abbreviations to the years 2015 and 2016; and Bloomberg and MSCI
avoid rather trivial mismatches. We used the unique KLD provided data on a yearly frequency. To convert
union of all three matching procedures to compile from a semiannual or annual frequency, we simply
our sample. used the respective annual or semiannual value for
the whole period. Note that most ratings (also for
To construct the sample, we also required that at the providers with a monthly frequency) change
least three rating observations be available for each infrequently; most ratings are constant for about one
company. This choice provided us with an internally year, but some are constant for longer periods.19
consistent sample, and it was not overly restrictive.
Control variables
Tangibility Tangibility Property, plant, and equipment divided Compustat
by total assets.
Current ratio Current ratio Current assets divided by current Compustat
liabilities.
Leverage Leverage Long-term debt plus debt in current Compustat
liabilities divided by total assets.
Gross Gross profitability Revenues minus costs of goods sold Compustat
profitability divided by total assets.
HHI Herfindahl–Hirschman The HHI measures industry concentration Compustat
Index (HHI) based by using book equity and the 2-digit SIC
on book equity level.
Multisegment Multisegment Dummy variable that is 1 if the firm Compustat Segments
operates in more than one segment. Data
No credit rating Missing credit rating Dummy variable that is 1 if no credit Compustat Company S&P
rating is available. Credit Ratings
Institutional Institutional Percentage of institutional ownership. Thomson Reuters
ownership ownership Institutional (13f)
Holdings
(continued)
121
Volume 77 Number 4
Financial Analysts Journal | A Publication of CFA Institute
Risk measures
Total volatility Total volatility Standard deviation computed from CRSP
the most recent 250 daily return
observations.
Idiosyncratic Idiosyncratic volatility Total volatility minus systematic volatility CRSP
volatility (explained by the CAPM) from the most
recent 250 daily return observations.
Beta Firm’s beta Market beta computed from the most CRSP
recent 250 daily return observations.
Note: We classified the variables into five groups: ESG rating disagreement variables, additional independent variable(s), control
variables, risk measures, and downside risk measures.
Other variables
Return 51,152 0.012 0.080 –0.748 1.274 0.013 0.388 6.796
Market 51,178 0.030 0.052 0.000 0.882 0.013 4.992 38.167
capitalization
Book-to-market 49,783 0.762 2.381 0.001 108.763 0.444 20.718 621.428
ratio
Gross 49,783 0.287 0.219 –1.143 1.405 0.253 1.035 2.459
profitability
Momentum 47,736 0.150 0.305 –0.925 4.870 0.139 1.320 10.567
StD. EPS 44,310 0.822 17.904 0.000 961.040 0.060 31.516 1,147.771
Beta 51,094 1.075 0.394 –1.738 3.787 1.039 0.712 1.461
Total volatility 51,178 0.018 0.009 0.001 0.226 0.016 4.899 65.671
Tangibility 49,304 0.255 0.244 0.000 0.946 0.160 1.022 –0.102
Current ratio 42,716 1.832 1.188 0.205 11.966 1.490 2.673 11.596
Leverage 49,585 0.246 0.157 0.000 0.960 0.238 0.591 0.561
HHI 49,783 0.074 0.078 0.009 0.744 0.044 2.810 11.632
Multisegment 18,212 0.562 0.496 0.000 1.000 1.000 –0.249 –1.938
No credit rating 51,206 0.199 0.399 0.000 1.000 0.000 1.511 0.284
Institutional 50,811 0.751 0.221 0.000 1.748 0.801 –1.811 3.800
ownership
Number of 48,920 0.189 0.078 0.010 0.560 0.180 0.445 0.481
analysts
Idiosyncratic 51,094 0.014 0.008 0.001 0.226 0.012 6.994 135.876
volatility
LPM 51,094 0.012 0.006 0.001 0.103 0.011 3.054 21.447
Note: We report the number of observations (N) and the sample means and standard deviations, as well as minimums, maximums,
medians, skewness, and kurtosis. Table A1 provides a detailed description of the variables.
123
Volume 77 Number 4
Financial Analysts Journal | A Publication of CFA Institute
Note: See the notes to Table 4. Bold indicates significance at the 5% level.
Editor’s Note
Submitted 6 February 2021
Accepted 29 July 2021 by William N. Goetzmann
Notes
1. Valuation mistake = [Equity(no disagreement) – 5. In a recent study, Lopez-de-Silanes, McCahery, and
Equity(disagreement)]/Equity(disagreement) – 1 = Pudschedl (2019) found evidence that firms with good ESG
[FCFE/(rE – g)]/[FCFE/(rE + 0.0092 – g)] = [100/(0.05)]/ scores may disclose more information.
[100/(0.05 + 0.0092)] – 1 = 0.184 (or 18.4%).
6. Asset4 was acquired by Thomson Reuters in 2009, but
2. See, for instance, Hong and Kostovetsky (2012); Krüger the ESG data were made available under the old name
(2015); Lins, Servaes, and Tamayo (2017); Liang and of Asset4. After the acquisition, the name changed to
Renneboog (2017); Gibson, Krueger, and Mitali (2020); Thomson Reuters ESG Scores. Because the name Asset4
Dyck, Lins, Roth, and Wagner (2019). is widely known, however, we used the old name for
simplicity. Note that as of 2018, the ESG ratings data of
3. See, for instance, Chatterji, Levine, and Toffel (2009); Thomson Reuters are part of Refinitiv and now also known
Bouten, Cho, Michelon, and Roberts (2017); Delmas, as Refinitiv ESG.
Etzion, and Nairn-Birch (2013).
7. After acquiring about a 40% stake in Sustainalytics in
4. The concept of a common theorization refers to the idea 2017, Morningstar purchased the remaining approximately
that raters (or information intermediaries) agree on a 60% of Sustainalytics equity in 2020 (see https://bit.
common definition of CSR. Absence of commensurability ly/3oXCgxM).
captures the idea that different raters do not use the same
measures when quantifying the same feature.
8. The data from KLD originate from Kinder, Lydenberg, Philipov (2009) reported that 9,051 firms did not have a
Domini (KLD) & Co., Inc., which was acquired by credit rating. In our sample, 194 out of a total of 553 firms
RiskMetrics in 2009. In 2010, MSCI acquired RiskMetrics. did not have a credit rating for at least one month.
Eccles, Lee, and Stroehle (2019) have provided details on
the history of KLD. We refer to these data as either KLD or 15. We included industry-month fixed effects and controlled
MSCI KLD. for standard characteristics (not reported) that have been
found to explain volatility (see, for example, Dennis and
9. The MSCI IVA dataset was initially created by Innovest Strickland 2004)—namely, market capitalization (Banz
Strategic Value Advisors, which was also acquired by 1981), leverage, business segment (a dummy variable
RiskMetrics in 2009 before RiskMetrics was taken over by was 1 for multisegment firms), percentage of institutional
MSCI (see Eccles et al. 2019 for details). ownership, the ratio of mutual fund ownership to total
institutional ownership, and turnover.
10. Each provider, with the exception of Bloomberg, had a
rather constant number of observations for the differ- 16. If available, we used the redemption value as preferred
ent scores they issue. Bloomberg had substantially lower stock. Otherwise, we used the liquidating value or, if the
coverage for environmental ratings. liquidation value was also unavailable, the carrying value.
11. Available at http://www.ghgprotocol.org. 17. The MSCI KLD dataset seems to have some issues with
the CUSIP code. The codes do not always have the same
12. In addition, evidence from both academics and number of characters, and leading zeros are often trun-
practitioners indicates that environmental risk matters to cated. Therefore, we filled in leading zeros if the number
investors. Bolton and Kacperczyk (forthcoming) found a of characters was less than eight. Then, we added the
sizable risk premium for firms with high carbon emissions, self-computed check digit to the code if the eighth number
suggesting that investors do care about carbon risk. In a was not the would-be check digit if there had been an
recent survey by BlackRock (2020), 88% of the respondents additional leading zero (in that case, we added a leading
placed climate risk at the top of their portfolio concerns. zero) or the last two characters consisted of commonly
used issue codes.
13. The 12-industry classification may be found at https://
mba.tuck.dartmouth.edu/pages/faculty/ken.french/ 18. For US stocks, the ISIN number consists of the country
data_library/det_12_ind_port.html. code (first two characters), the CUSIP code (characters
3 to 11), and a check digit.
14. The reader might wonder why any S&P 500 firms would
not have a credit rating. In general, firms without a credit 19. Because the providers change their ratings at different
rating do not seem to be exceptional. For example, in a points in time, we argue that for our purposes, using a
sample of 12,312 firms, Avramov, Chordia, Jostova, and monthly frequency makes sense.
References
Anderson, E. W., E. Ghysels, and J. L. Juergens. 2005. “Do Berg, F., J. F. Koelbel, and R. Rigobon. 2020 “Aggregate
Heterogeneous Beliefs Matter for Asset Pricing?” Review Confusion: The Divergence of ESG Ratings.” Working paper.
of Financial Studies 18 (3): 875–924.
BlackRock. 2020. “BlackRock Survey Shows Acceleration of
———. 2009. “The Impact of Risk and Uncertainty on Expected Sustainable Investing.” Press release. https://bit.ly/3upXdlZ.
Returns.” Journal of Financial Economics 94 (2): 233–63.
Bolton, P., and A. Kacperczyk. Forthcoming. “Do Investors
Ang, A., R. J. Hodrick, Y. Xing, and X. Zhang. 2006. “The Cross- Care about Carbon Risk?” Journal of Financial Economics.
Section of Volatility and Expected Returns.” Journal of Finance
61 (1): 259–99. Bouten, L., C. H. Cho, G. Michelon, and R. W. Roberts. 2017.
“CSR Performance Proxies in Large-Sample Studies: ‘Umbrella
Atmaz, A., and S. Basak. 2018. “Belief Dispersion in the Stock Advocates’, Construct Clarity and the ‘Validity Police.’”
Market.” Journal of Finance 73 (3): 1225–79. Working paper.
Avramov, D., S. Cheng, A. Lioui, and A. Tarelli. 2021. Carhart, M. M. 1997. “On Persistence in Mutual Fund
“Sustainable Investing with ESG Rating Uncertainty.” Performance.” Journal of Finance 52 (1): 57–82.
Working paper.
Chatterji, A. K., R. Durand, D. I. Levine, and S. Touboul. 2016.
Avramov, D., T. Chordia, G. Jostova, and A. Philipov. 2009. “Do Ratings of Firms Converge? Implications for Managers,
“Dispersion in Analysts’ Earnings Forecasts and Credit Rating.” Investors and Strategy Researchers.” Strategic Management
Journal of Financial Economics 91 (1): 83–101. Journal 37 (8): 1597–614.
Banz, R. W. 1981. “The Relationship between Return and Chatterji, A. K., D. I. Levine, and M. W. Toffel. 2009. “How
Market Value of Common Stocks.” Journal of Financial Well Do Social Ratings Actually Measure Corporate Social
Economics 9 (1): 3–18. Responsibility?” Journal of Economics & Management Strategy
18 (1): 125–69.
125
Volume 77 Number 4
Financial Analysts Journal | A Publication of CFA Institute
Chordia, T., A. Subrahmanyam, and Q. Tong. 2014. “Have GSIA. 2016. “Global Sustainable Investment Review
Capital Market Anomalies Attenuated in the Recent Era of 2016.” Global Sustainable Investment Alliance. https://bit.
High Liquidity and Trading Activity?” Journal of Accounting ly/2U1OuoF.
and Economics 58 (1): 41–58.
Hasbrouck, J. 2009. “Trading Costs and Returns for U.S.
Christensen, D., G. Serafeim, and A. Sikochi. Forthcoming. Equities: Estimating Effective Costs from Daily Data.”
“Why Is Corporate Virtue in the Eye of the Beholder? The Journal of Finance 64 (3): 1445–77.
Case of ESG Ratings.” Journal of Accounting Research.
Hoepner, A. G., I. Oikonomou, Z. Sautner, L. T. Starks, and
Delmas, M. A., D. Etzion, and N. Nairn-Birch. 2013. X. Zhou. 2019. “ESG Shareholder Engagement and Downside
“Triangulating Environmental Performance: What Do Risk.” Working paper.
Corporate Social Responsibility Ratings Really Capture?”
Academy of Management Perspectives 27 (3): 255–67. Hong, H., and L. Kostovetsky. 2012. “Red and Blue Investing:
Values and Finance.” Journal of Financial Economics 103 (1):
Dennis, P. J., and D. Strickland. 2004. “The Determinants 1–19.
of Idiosyncratic Volatility.” Working paper.
Jacobs, H., and S. Müller. 2020. “Anomalies across the
Diether, K. B., C. J. Malloy, and A. Scherbina. 2002. Globe: Once Public, No Longer Existent?” Journal of Financial
“Differences of Opinion and the Cross Section of Stock Economics 135 (1): 213–30.
Returns.” Journal of Finance 57 (5): 2113–41.
Jegadeesh, N., and S. Titman. 1993. “Returns to Buying
Doyle, T. M. 2018. “Ratings That Don’t Rate: The Subjective Winners and Selling Losers: Implications for Stock Market
View of ESG Rating Agencies.” Technical report. American Efficiency.” Journal of Finance 48 (1): 65–91.
Council for Capital Formation. https://bit.ly/2LBwvky.
Kogan, L., and T. Wang. 2003. “A Simple Theory of Asset
Dunn, J., S. Fitzgibbons, and L. Pomorski. 2018. “Assessing Pricing under Model Uncertainty.” MIT working paper.
Risk through Environmental, Social and Governance
Exposures.” Journal of Investment Management 16 (1): 4–17. Krüger, P. 2015. “Corporate Goodness and Shareholder
Wealth.” Journal of Financial Economics 115 (2): 304–29.
Dyck, A., K. V. Lins, L. Roth, and H. F. Wagner. 2019. “Do
Institutional Investors Drive Corporate Social Responsibility? Liang, H., and L. Renneboog. 2017. “On the Foundations of
International Evidence.” Journal of Financial Economics 131 (3): Corporate Social Responsibility.” Journal of Finance 72 (2):
693–714. 853–910.
Eccles, R. G., L.-E. Lee, and J. C. Stroehle. 2019. “The Social Lins, K. V., H. Servaes, and A. Tamayo. 2017. “Social Capital,
Origins of ESG? An Analysis of Innovest and KLD.” Working Trust, and Firm Performance: The Value of Corporate Social
paper. Responsibility during the Financial Crisis.” Journal of Finance
72 (4): 1785–824.
Fama, E. F., and K. R. French. 1995. “Size and Book-to-Market
Factors in Earnings and Returns.” Journal of Finance 50 (1): Lopez-de-Silanes, F., J. A. McCahery, and P. C. Pudschedl.
131–55. 2019. “ESG Performance and Disclosure: A Cross-Country
Analysis.” ECGI working paper.
———. 2015. “A Five-Factor Asset Pricing Model.” Journal of
Financial Economics 116 (1): 1–22. Matos, P. 2020. ESG and Responsible Institutional Investing
around the World: A Critical Review. Charlottesville, VA: CFA
Fama, E. F., and J. D. MacBeth. 1973. “Risk, Return, and Institute Research Foundation (May).
Equilibrium: Empirical Tests.” Journal of Political Economy
81 (3): 607–36. Mackintosh, J. 2018. “Is Tesla or Exxon More Sustainable? It
Depends Whom You Ask.” Wall Street Journal (17 September).
Frazzini, A., and L. H. Pedersen. 2014. “Betting against Beta.” https://on.wsj.com/2MQCC4m.
Journal of Financial Economics 111 (1): 1–25.
McLean, R. D., and J. Pontiff. 2016. “Does Academic Research
Friede, G., T. Busch, and A. Bassen. 2015. “ESG and Financial Destroy Stock Return Predictability?” Journal of Finance 71 (1):
Performance: Aggregated Evidence from More Than 2000 5–32.
Empirical Studies.” Journal of Sustainable Finance & Investment
5 (4): 210–33. Newey, W. K., and K. D. West. 1987. “A Simple, Positive Semi-
Definite, Heteroskedasticity and Autocorrelation Consistent
Gibson, R., S. Glossner, P. Krueger, P. Matos, and T. Steffen. Covariance Matrix.” Econometrica 55 (3): 703–08.
2021. “Do Responsible Investors Invest Responsibly?”
University of Geneva and University of Virginia. Novy-Marx, R. 2013. “The Other Side of Value: The Gross
Profitability Premium.” Journal of Financial Economics 108 (1):
Gibson, R., P. Krueger, and S. Mitali. 2020. “The Sustainability 1–28.
Footprint of Institutional Investors.” Swiss Finance Institute
Research Paper 17-05. Novy-Marx, R., and M. Velikov. 2016. “A Taxonomy of
Anomalies and Their Trading Costs.” Review of Financial
Green, J., J. R. Hand, and X. F. Zhang. 2017. “The Characteristics Studies 29 (1): 104–47.
That Provide Independent Information about Average US
Monthly Stock Returns.” Review of Financial Studies 30 (12): PRI. 2018. “Annual Report 2018.” Principles for Responsible
4389–436. Investment. https://bit.ly/2NsU8xv.
Roll, R. 1984. “A Simple Implicit Measure of the Effective Bid- Wigglesworth, R. 2018. “Rating Agencies Using
Ask Spread in an Efficient Market.” Journal of Finance 39 (4): Green Criteria Suffer from ‘Inherent Biases.’” Financial
1127–39. Times (20 July). https://www.ft.com/content/
a5e02050-8ac6-11e8-bf9e-8771d5404543.
USSIF. 2020. “Report on U.S. Sustainable and Impact Investing
Trends 2020.” Washington, DC: Forum for Sustainable and Wong, C., A. Brackley, and E. Petroy. 2019. “Rate the
Responsible Investment. Raters 2019: Expert Views on ESG Ratings.” Sustainability.
https://www.sustainability.com/thinking/rate-raters-2019/.
Viale, A. M., L. Garcia-Feijoo, and A. Giannetti. 2014. “Safety
First, Learning under Ambiguity, and the Cross-Section of
Returns.” Review of Asset Pricing Studies 4 (1): 118–59.
127
Volume 77 Number 4