JID: FINEC

ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

Journal of Financial Economics xxx (xxxx) xxx

Contents lists available at ScienceDirect

Journal of Financial Economics

journal homepage: www.elsevier.com/locate/jfec

Asset pricing on earnings announcement days ✩

Kam Fong Chan a,∗, Terry Marsh b,c
a
Accounting and Finance, University of Western Australia, UWA Business School, 35 Stirling Highway WA, 6009, Australia
b
Quantal International Inc., Two Embarcadero Center, 8th Floor, San Francisco, CA 94111, United States
c
University of California, Berkeley, Haas School of Business, 545 Student Services Bldg, #1900 Berkeley, CA 94720, United States

a r t i c l e i n f o a b s t r a c t

Article history: Market betas have a strong and positive relation with average stock returns on a hand-
Received 15 December 2020 ful of days every year. Such unique days, defined as leading earnings announcement days
Revised 3 May 2021
(LEADs), are times when an aggregate of influential S&P 500 firms disclose quarterly earn-
Accepted 1 June 2021
ings news early in the earnings season. The positive return-to-beta relation holds for var-
Available online xxx
ious test portfolios, individual stocks, and Treasuries; and is robust to different data fre-
JEL classification: quencies and testing procedures. On days other than LEADs, the beta-return relation is
G11 flat. We conclude that waves of early earnings announcements by large firms clustered on
G12 LEADs significantly influence asset pricing.
G14
© 2021 Elsevier B.V. All rights reserved.
Keywords:
Capital asset pricing model
Earnings announcements
Security market line
Market beta

1. Introduction positive beta-return relation, but only in months of low in-

flation (Cohen et al., 2005), on macroeconomic announce-
In the capital asset pricing model (CAPM), an asset’s ment days (Savor and Wilson, 2014), when investor senti-
market risk exposure as measured by beta is linearly and ment or disagreement about the stock market’s prospects
positively related to its return expectations. Recent stud- is low (Antoniou et al., 2016; Hong and Sraer, 2016),
ies provide some new empirical evidence supporting this when borrowing constraints are slack (Jylha, 2018), when
the beta-return relation is measured using a set of
✩
key accounting variables (Penman and Zhu, 2019), when
We are indebted to Bill Schwert (the editor) and an anonymous jour-
nal referee for guidance and insightful suggestions that have significantly
investor attention or market predicted return is high
improved the paper. We thank Ray Ball, Charles Chang, David Chapman, (Da et al., 2020; Ben-Rephael et al., 2021; Hasler and
Phil Dolan, Mark Hulbert, Allan Kleidon, Stephen Penman, Kris Ramesh, Martineau, 2021), overnight while the market is closed
Laura Starks, Jing Yu and Chen Zhou for helpful suggestions. Comments (Hendershott et al., 2020), and in months after the U.S.
from seminar participants at the January 2019 seminar at Bond Univer-
midterm elections (Chan and Marsh, 2021).
sity, the February 2019 Fudan University presentation in San Francisco,
the March 2019 Research Seminar at the University of Western Australia, Our study expands on this burgeoning literature by pro-
the June 2019 Pacific Basin Finance, Economics, Accounting and Manage- viding evidence that the market beta-return relation is lin-
ment Annual Conference, Taipei, the 2nd West Coast Finance 2020 Collo- ear and strongly positive across securities when an aggre-
quium, Australia and the La Trobe 2021 Finance virtual seminar are also gate of large influential S&P 500 companies disclose cor-
appreciated. Jessie Xie provided excellent assistance in constructing the
porate earnings news early in the earnings seasons. These
abnormal institutional attention measure using data obtained from the
Bloomberg terminal. leading earnings announcement days (LEADs) also happen
∗
Corresponding author. to be times when institutional investors’ attention is high.
E-mail address: kam.chan@uwa.edu.au (K.F. Chan). LEADs occur for only a handful of adjacent days per quar-

https://doi.org/10.1016/j.jfineco.2021.06.022
0304-405X/© 2021 Elsevier B.V. All rights reserved.

Please cite this article as: K.F. Chan and T. Marsh, Asset pricing on earnings announcement days, Journal of Financial
Economics, https://doi.org/10.1016/j.jfineco.2021.06.022
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

Fig. 1. Average excess returns for beta decile portfolios. This figure plots the average daily excess returns (expressed in basis points) against market betas
for 10 value-weighted beta-sorted portfolios. Panel A plots the unconditional SML, and Panel B plots the conditional SMLs on LEADs (blue line with triangle
markers) and on other days (green line with circle markers). For each test portfolio, we use the same full-sample beta estimate for both types of day, and
we superimpose an ordinary least squares best fitted line for each type of day. The sample period covers January 2001 through December 2019. (For
interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

ter, and their timing is rather predictable. On days other beta of 1 is associated with a reduction in stock excess re-
than LEADs (i.e., “other days” or non-LEADs), the market turns of 1.95 bps. In addition, the intercept for the other-
beta-return relation is flat.1 day SML is significantly positive (a finding that is inconsis-
Fig. 1 provides the key findings that motivate tent with the CAPM), while it is not statistically significant
this study. As in Savor and Wilson (2014) and for LEAD returns.
Hendershott et al. (2020), we use a 12-month rolling We confirm that the positive beta-return relation on
regression of daily of returns to estimate betas for US LEADs is a significant and robust phenomenon. It continues
stocks. The stocks are grouped into ten value-weighted, to hold when we perform the analysis using different test
beta-sorted portfolios, and average portfolio excess returns portfolios and US Treasuries, and for individual stocks af-
are plotted against average portfolio betas in Fig. 1. Panel ter controlling for familiar firm characteristics such as size
A shows the unconditional fit of the security market line and book-to-market ratio. The results are also robust to
(SML) from 2001 to 2019. The line is mostly flat, sug- whether we test using Fama and MacBeth (1973) or panel
gesting that market betas do not explain average excess regressions, for sub-periods, and for betas estimated from
returns. This result, which is inconsistent with the CAPM, high frequency (intraday) returns. Furthermore, the re-
has been documented by Black et al. (1972), Fama and sults are not driven by strategically early, or delayed, news
French (1992, 2004), and many others. releases by LEAD announcers, nor confounded by macro
Panel B shows the SML separately for LEADs (blue line news announcements as per Savor and Wilson (2014), nor
with triangle markers) and for other days (green line with by the overnight return effect of Hendershott et al. (2020).
circle markers). The blue SML shows that on LEADs, an in- We also show that the cross-section SML finding is ac-
crease in market beta of 1 is associated with an econom- companied by an increase in the market premium à la
ically and statistically significant increase in average daily Roll (1977), whereby the average market premium on
excess market returns of 24.47 basis points (bps). The re- LEADs is seven times higher than on other days.
sult, however, is vastly different for other days: the green There are also pronounced shifts in the SML in the
SML has a slight downward slope, where an increase in hours surrounding LEADs: a negative daytime SML on the
day prior to LEADs is followed by a positive overnight SML
1 just before the first “leading” trading day. On LEADs, the
Our use of an S&P 500-based definition of “influential announcers” is
consistent with Bai et al. (2016) and Farboodi et al. (2020). They show relation between average return and market beta is lin-
that informational efficiency in prices has improved in recent years for ear and mostly positive for all but the last trading hour,
S&P 500 stocks, inter alia because they are large companies. Both studies when the SML reverts to a flat slope. On the day following
measure price informativeness by regressing future earnings on current LEADs, the SML slope is negative. Finally, there appears to
stock prices of S&P 500 firms. In this study, on the other hand, we show
that on LEADs, the entire cross-section of discount rates also changes.
be a day-of-the-week effect for the positive beta-return re-
Taken together, the results suggest that price informativeness extends to lation: it occurs predominantly when earnings announce-
that for cross-sectional discount rates for all stocks instead of just reflect- ments are disclosed on Tuesday through Thursday in the
ing future earnings expectations for S&P 500 stocks. LEAD week.

2
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

To assess the economic significance of the findings, the then-expensive leverage drops, and the size of the
we analyze a hybrid “betting-on-and-against-beta” trad- leverage-constrained segment of investors shrinks.
ing strategy that entails “betting-on-beta” (i.e., taking a The third potential explanation for “on again” CAPM
long position in high-beta stocks and a short position in risk pricing on LEADs focuses on an increase in market
low-beta stocks) on LEADs, while switching to the famil- risk on LEADs versus on other days. We show how the
iar “betting-against-beta” strategy (i.e., buy low-beta stocks average realized betas change around earnings announce-
and sell high-beta stocks) on other days. This hybrid strat- ments in a way similar to Patton and Verardo (2012), and
egy earns a Sharpe ratio equal to 0.306 when annualized. then provide evidence that the increase in average real-
This number is impressive, both in absolute terms and ized beta on LEADs is higher than that on other days.
relative to other strategies: for example, the annualized The shift (i.e., increase) in beta helps explain the posi-
Sharpe ratio generated by continuously betting-against- tive cross-sectional return-beta relation, and a higher mar-
beta over the full sample period is three times lower at ket premium on LEADs.3 We interpret the increase in av-
0.113.2 Overall, our results highlight the importance of cor- erage betas for LEADs to be due to potential spillover of
porate earnings announcements in asset pricing: market information to the cross-section of stocks in the wake of
beta exposure is priced on LEADs after all and not on other the “wave” of a quarter’s announcements made by clus-
days. In fact, one-third of the cumulative log excess mar- ters of large firms. Savor and Wilson (2016) also explore
ket return in 2001–2019 is earned on LEADs alone, even the spillover of announcement information to future aggre-
though these LEADs account for only 5% of all trading days. gate earnings growth. Two features of our results are that:
We explore three possible explanations for our re- (i) we observe the “early announcement LEAD effect” via
sults. First, our definition of LEADs entails a handful of a simple and transparent criterion for influential announc-
days when an aggregate of influential S&P 500 compa- ers, thus avoiding self-fulfilling choice and possible selec-
nies are making an early wave of earnings announcements tion bias; and (ii) LEADs are days on which cross-sectional
for the quarter. Thus, it is possible that LEAD announce- discount rates move systematically.4
ments operate as a macro-level trigger for investor atten- Our study is related to Savor and Wilson’s (2014) “tale
tion. Consistent with this interpretation, Hirshleifer and of two days” where market exposure is priced on days
Sheng (2019) explore what constitutes a “news environ- when pre-scheduled key macroeconomic news is released,
ment.” They show that on days when macro news is re- but not on other days. Chan and Marsh (2021) show
leased concomitantly with micro-level announcements, the that decreases in future economic policy uncertainty fol-
former acts as an attention trigger for the latter, thereby lowing pre-scheduled US midterm elections are likewise
driving the returns of individual (micro-level) stocks. Ben- priced. Macroeconomic news announcements, Congres-
Rephael et al. (2021) find that macro news is associated sional election-related announcements, and LEAD earnings
with micro-level risk premiums, and Da et al. (2020) re- announcements all share two common attributes: they are
port that high-beta stocks deliver higher average premi- pre-scheduled and “macro” in nature. The pre-scheduling
ums than low-beta stocks when investor attention is high. ipso facto enables better trading strategies that utilize
On the other hand, it seems that attention to earnings an- the timing at which possible insider information becomes
nouncements and associated media appearances, confer- known to the market. The LEAD announcements seem to
ence calls, and the like would, if important, already have pre-empt macro announcements insofar as the latter lose
been primed – “triggered” – at the time when the corre- their importance once LEAD announcers are taken into ac-
sponding announcement date is scheduled, typically sev- count: we find no difference between the return-beta rela-
eral days before the announcements, and mostly on the tion estimated on LEADs that happen also to overlap with
same day from year to year. macroeconomic news releases, and the return-beta relation
The second potential explanation for the LEAD effect on LEADs with no macro news.
on cross-sectional asset prices traces back to Black (1972), We structure the paper as follows. In Section 2, we
who proposes that an unconditional flat SML like we ob- describe the data and procedure to identify LEADs. In
serve over the full sample period reflects market segmen- Section 3, we analyze how stock prices behave differently
tation: leverage-constrained investors who cannot fully on LEADs and on other days, and Section 4 provides ad-
lever their portfolios are willing to pay more than pre- ditional tests. In Section 5, we discuss how our findings
dicted by the CAPM for risky stocks. This segment of could be consistent with the three potential explanations
leverage-constrained investors overweights high-risk as- outlined above. Finally, Section 6 concludes.
sets, enabled by leverage-unconstrained investors who un-
derweight them. The consequence is an equilibrium in
which the prices of high-beta assets are higher and ex- 3
Penman (1990) suggests that the earnings announcement premium is
pected returns are lower, and prices of low-beta assets a compensation for the additional cross-asset risk of information spillover
of earnings news. Penman’s (1990) work was proposed long before the
are lower and expected returns are higher, relative to the
sample period of the present study.
CAPM. This well-known segmentation model provides a 4
Andrei et al. (2020) propose that it is discount rates as seen by the
basis for the “betting-against-beta” strategy. We discuss empiricist that move when differences in investor private information are
how the SML is upward sloping on LEADs if demand for “washed away” by public announcement (prior to that, each investor only
observes the SML conditional on their private information sets and the
empiricist sees only a conglomeration of them). Our framework differs
2
Marsh and Pfleiderer (2016, exhibit 2) calibrate the betting-against- insofar as the discount rate changes as seen by the empiricist are for a
beta model of segmentation and find that it is consistent with premiums different set of stocks than the cluster of LEADers making the announce-
for low risk “value” stocks up to 60 bps when annualized. ments.

3
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

2. Data and lead definition ple has a total of 35,826 earnings announcements for con-
stituents of the S&P 500 index.6
The work by Savor and Wilson (2014) highlights the im- We then select “leading” earnings news from the pool
portance of pre-scheduled macro announcements made at of sample announcements. This procedure entails identify-
the market level in asset pricing. The present study shows ing leading earnings announcement days (LEADs) that are
that pre-scheduled corporate earnings announcements are the days when “leading news” is reported. To this end,
just as important at the aggregate level in asset pricing. To we use a method that is straightforward, transparent, and
operationalize “aggregate,” we identify a set of individual devoid of look-back bias: we select LEADs to be Tuesday
firms making corporate earnings announcements that col- through Thursday in the first week of reporting quarter q
lectively have the potential to “move” the market. We posit that has a minimum of 50 announcers (i.e., one-tenth of
that, conditional on the concomitant earnings announce- S&P 500 firms).7 We label the LEAD announcers as LEAD-
ments released by market movers, exposure to the aggre- ers. Put simply, our LEAD definition encompasses a sin-
gate market moving news is priced cross-sectionally on an- gle fraction of S&P 500 firms as announcers, thus ensur-
nouncement days. Market movers are large and influential, ing that LEADers are, by construction, large but not other-
because all else equal, uncertainty about the informative- wise specifically chosen, and their reported early earnings
ness of announcements made by smaller companies limits news is ex ante likely to be influential.8 We omit Mon-
their influence, while there is also evidence of greater in- day from our LEAD definition because firms rarely disclose
vestor attention to earnings announcements made by large earnings news on Monday (only about 5% of all S&P 500
companies (Nekrasov et al., 2021). Indeed, clusters of large firms’ earnings reports in our sample were made between
announcers in earnings week have virtually been institu- 0:00 am and 16:00 pm on Monday). As such, Monday an-
tionalized as the focus of media attention then; for ex- nouncements, if anything, are less influential. We also ex-
ample, Nasdaq lamented that: “…this week [September 4, clude Fridays from our LEADs due to general inattention
2020], there aren’t enough major companies reporting to and pre-weekend idiosyncrasies, and we provide the rele-
truly move the market.”5 vant discussion below.
Amongst the large market movers, we expect early an- Table 1 reports descriptive statistics of S&P 500 an-
nouncers to be inherently more influential because their nouncers reporting on two types of days: LEADs and other
announcements contain up-to-date information. “Strongest days. Panel A shows that there are on average 31 LEADers
signal” announcements early in the quarterly earnings sea- per day (i.e., about 93 from Tuesday through Thursday);
son are also regularly complemented by conference calls, this estimate is fourfold higher than the average num-
media attention, interviews, and general “buzz” du jour ber of other S&P 500 firms reporting on non-LEADs. In
sufficient to constitute a “macro” event for three or four Panel B, we partition the news observations into differ-
jours. In other words, early influential announcements are ent weekdays, for a total of 7138 news announcements on
plausibly signals or attention triggers. LEADs (i.e., about one-fifth of the final earnings sample)
To select early influential announcements, we begin versus 28,688 news announcements on other days. Panel
by collecting quarterly earnings announcements for all C tabulates, on a yearly basis, the total number of unique
the constituents of the S&P 500 index between Jan- announcers on LEADs and on other days, and the cross-
uary 2001 and December 2019, for a total of 4227 trad- section median of variables such as market capitalization
ing days. The 2001 starting date is motivated by stud- and analyst coverage. It is apparent from the panel that
ies such as Beaver et al. (2018, 2020), which document over the years, LEADers are, as desired, consistently larger
a considerable increase in information on earnings an- and thereby attract more attention (with a slightly higher
nouncement days post-20 0 0; a period of key account- analyst following), than other S&P 500 announcers report-
ing reforms, such as the Sarbanes-Oxley Act and Reg- ing on non-LEADs.
ulation Fair Disclosure. We extract earnings data from In Fig. 2, we plot the total number of LEADs identified
the Institutional Brokers’ Estimate System (I/B/E/S) sum- in each reporting week. The online appendix presents the
mary file, and we require that the S&P 500 announcers frequency distribution of LEADers and other S&P 500 an-
in our sample universe are listed on the NYSE, Amex or nouncers for each fiscal quarter. We can see that LEADers
Nasdaq. Following convention, we remove a small num- generally announce in week 3 or 4 in fiscal quarter one,
ber of earnings observations timestamped at 0 0:0 0:0 0 be-
cause these are likely I/B/E/S recording errors made espe-
cially in the early 20 0 0s (deHaan et al., 2015). We further 6
Because of multiple data filters, our final sample consists of 35,826
drop a handful of earnings news observations recorded news observations instead of 19 x 4 x 500 = 38,000 observations. We
on Saturdays, Sundays, and public holidays. Also, we fol- exclude, for example, Fannie Mae and Freddie Mac from the analysis be-
low Patton and Verardo (2012), Michaely et al. (2014), cause Compustat tags both stocks as traded over the counter since the
start of our sample period (even though both firms were in actuality
and deHaan et al. (2015) and assign news observations re- traded over the counter since mid 2010).
ported after the market has closed to the next trading day 7
Our LEADs are not always days with the largest number of announc-
or overnight as appropriate for our analysis. The final sam- ers, and conversely, there are typically some large announcers reporting
in advance of or following the LEAD clusters; nor do we seek to include
all the popular attention-getters du jour (e.g., FAANG stocks).
8
The online appendix tabulates 50 top frequent LEADers which include
household names such as Apple Inc., AT&T, Boeing, IBM, and Johnson &
5
https://www.nasdaq.com/articles/7- earnings- reports- to- watch- next- Johnson, as well as several lesser known companies, such as Danaher and
week- 2020- 09- 04. Nucor.

4
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

Table 1
Descriptive statistics of S&P 500 announcers.
This table reports descriptive statistics for S&P 500 announcers on LEADs and other days. Panel A reports the daily distribution
of announcers, and Panel B reports the number of announcers each day between Monday and Friday. In Panel C, Unique firms
is the number of announcers reporting in a given year, and Mkt cap, B/M and # of analysts report the respective cross-section
median of market capitalization, firms’ book-to-market ratios and number of analysts followings (all measured one quarter before
the announcement date). The sample period covers January 2001 through December 2019.

Panel A: Daily distribution of announcers Panel B: # of announcers in different weekdays

LEADs Other days LEADs Other days

Mean 31 8 Monday – 1719

Min 6 1 Tuesday 1534 5767
5 pctl 13 1 Wednesday 2123 6538
25 pctl 22 2 Thursday 3481 8387
50 pctl 29 4 Friday – 6277
75 pctl 40 9
95 pctl 59 32 Total 7138 28,688
Max 73 70

Panel C: Descriptive statistics (per year)

LEADs Other days

Unique firms Mkt cap ($ bil) B/M # of analysts Unique firms Mkt cap ($ bil) B/M # of analysts

2001 177 8.93 0.33 13 415 7.58 0.35 13

2002 167 9.63 0.35 14 423 7.36 0.39 14
2003 175 8.29 0.38 14 437 7.13 0.43 14
2004 191 10.88 0.35 16 428 9.03 0.37 14
2005 142 15.21 0.34 17 433 10.19 0.38 14
2006 182 15.11 0.31 16 450 11.36 0.36 14
2007 145 16.64 0.32 16 457 12.44 0.33 15
2008 198 11.68 0.37 14 458 9.56 0.4 14
2009 193 9.32 0.52 14 460 6.43 0.55 13
2010 189 12.07 0.43 17 463 8.91 0.46 15
2011 187 15.78 0.42 19 467 10.76 0.42 16
2012 153 14.67 0.42 19 469 11.95 0.42 17
2013 204 15.25 0.36 18 486 14.28 0.38 18
2014 207 20.04 0.32 19 488 16.77 0.33 18
2015 175 21.92 0.31 19 466 17.64 0.32 18
2016 153 20.17 0.35 19 489 17.46 0.34 18
2017 211 21.74 0.32 17 475 19.13 0.33 17
2018 191 24.26 0.29 17 481 19.92 0.33 17
2019 175 22.76 0.28 17 458 21.03 0.37 17
Avg 180 15.49 0.36 16.58 458 12.58 0.38 15.58

Fig. 2. LEADs in each quarter. In this figure, a typical year is equally partitioned into four fiscal quarters, each with 13 weeks. The number of days identified
as LEADs are reported for each week-quarter. Since each quarter has three LEADs (Tuesday, Wednesday and Thursday), we have a total of 57 LEADs for
each quarter over the 2001–2019 sample period (19 years).

5
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

and mainly in week 3 in the three remaining quarters. In 3. Relation between market betas and returns
fact, the identification of LEAD week is fairly predictable
over the years, and it is concentrated as intended in early 3.1. Beta-sorted portfolios
peak earnings weeks in January, April, July, and October.9
It is possible that the striking stock return differentials We begin by constructing beta decile portfolios using
between LEADs and other days discussed below occur be- stocks traded on the NYSE, Amex or Nasdaq from 2001 to
cause LEADers strategically time their earnings announce- 2019. We retrieve stock data from the Center for Research
ments, in which case the strategic timing would become in Security Prices (CRSP) database, and we include all com-
the leading information event. To check this possibility, mon stocks with CRSP share codes of 10 or 11.
we use a simple algorithm to project scheduled earnings In our initial set of tests, the daily pre-ranked be-
dates for quarter q in year y based on the same report- tas of individual stock j are estimated using a 12-month
ing quarter-date as in the previous year y - 1. For example, rolling regression of the stock’s daily excess returns on
if firm j reported on October 24, 2017, the algorithm pre- the market excess returns. At the start of each month, we
dicts that the firm would release its fourth quarter earn- sort the stocks into one of ten portfolios on the basis of
ings report around the same business day in the following their pre-ranking betas, and then calculate the daily value-
fiscal year. Firm j is deemed to have strategically advanced weighted returns of each portfolio over the month. Post-
its actual earnings reporting if it announced earlier than ranking portfolio betas are calculated for each day using
the expected date (i.e., October 24, 2018 in our example) a 12-month rolling regression of the daily portfolio excess
by more than five business days. Using the same criteria, returns on the daily market excess returns. These value-
firm j is deemed to strategically delay its actual news re- weighted beta decile portfolios are our main test subjects,
porting if it pushes back its announcement by more than but we also consider other test assets, and the results (re-
five business days. The algorithm shows that only 205 ported below) are robust.
(27) earnings announcements would have been advanced Panel A of Fig. 1 plots average excess returns of the
(delayed) to LEADs. With only 3% (205/6465) of the an- value-weighted beta decile portfolios on average market
nouncements having possibly been “sped up” to LEADs, betas over the 2001–2019 sample period. The slope of
and 0.4% (27/6465) delayed, we conclude that the distinct the unconditional SML is mostly flat with a coefficient
price dynamics observed on LEADs are unlikely caused by of –0.69 bps and a t-statistic of –2.38. The regression
material strategic timing.10 Also, we verify that there is intercept, on the other hand, is statistically significant
no substantial disparity between earnings beats and earn- with an estimate of 4.44 bps and t-statistic of 14.40, and
ings misses with effects that might be conflated with our the R2 is 42%. These results, which are inconsistent with
LEAD analysis: of 7138 LEAD announcements, 68.2% beat the CAPM prediction, accord with prior studies, such as
the analyst consensus (and the remaining 31.8% missed Black et al. (1972) and Fama and French (1992, 2004).
or matched analyst forecasts) versus 65.9% beats during We then partition the unconditional SML into its con-
non-LEAD announcements. In other words, the earnings ditional versions on LEADs and on other days. Panel B
beats-to-misses ratio is almost identical on both types of of Fig. 1 reports the findings. Following Savor and Wil-
days. son (2014) and Hendershott et al. (2020), we use uncondi-
In summary, LEADers in this study are large con- tional portfolio betas averaged over the full sample period
stituents of the S&P 500 index, and they are early an- to plot both conditional SMLs in the panel. The blue SML
nouncers. The aggregated early earnings news information shows a strong and positive linear relation between excess
reported by these announcers on LEADs influences and returns and market beta: on LEADs, a unit increase in beta
thus “leads” the reporting season. Finally, there is no ev- is associated with a significant increase in daily average ex-
idence suggesting that LEADers strategically and system- cess return of 24.47 bps (t-statistic = 5.14). The regression
atically advance or delay the announcements of corporate intercept has a modest negative estimate of –5.44 bps but
earnings to LEADs. the corresponding t-statistic is –1.07, rendering the neg-
ative estimate statistically insignificant. The regression R2
is 77%, suggesting that the variation in the market beta
explains a large portion of the cross-sectional excess re-
turn variation on LEADs. In sharp contrast, the slope of
9
This regularity in the timing of LEADs identifies a “quarterly week 3 the green SML is –1.95 bps (t-statistic = –7.08), indicat-
seasonality” (i.e., week 3 effect in each quarter), and a “quarterly weeks
4 and 3 seasonality” (i.e., week 4 effect in quarter one and week 3 ef-
ing a significantly negative implied market risk premium
fect in other quarters) in cross-sectional stock returns. To expand on the on other days. The regression intercept has a significantly
conjecture, the online appendix provides two separate SML plots for av- positive estimate equal to 4.94 (t-statistic = 16.82), and the
erage returns in week 3 of each quarter (the “quarterly week 3 season- R2 is 86%.
ality” effect), and average returns in week 4 in quarter one and week 3
We now provide the first piece of evidence that the
in other quarters (the “quarterly weeks 4 and 3 seasonality” effect). The
plots show that the implied market premium in the “quarterly week 3 positive beta-return relation on LEADs is a robust phe-
seasonality” effect (“quarterly weeks 4 and 3 seasonality” effect) can be nomenon. Like Savor and Wilson (2014), we group all
used as an expression for nearly 42% (65%) of the LEAD implied market stocks into 50 value-weighted, beta-sorted portfolios and
premium. We are grateful to the referee for suggesting this implication. repeat the above analysis. Fig. 3 presents the SML re-
10
We omit announcements reported in the last four quarters because
sults, with an almost identical finding. The blue SML shows
there are no matched expected-actual news announcements in the year
beyond the 2019 end date of our sample. As such, we use 6465 instead of a strongly linear and (almost) monotonically increasing
7178 news observations in analyzing possible strategic announcements. slope: on LEADs, an increase in market beta of 1 is sig-

6
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

Fig. 3. Average excess returns for 50 beta-sorted portfolios. The figure plots the average daily excess stock returns (expressed in basis points) against
market betas for 50 value-weighted, beta-sorted portfolios. Panel A plots the unconditional SML, and Panel B plots the conditional SMLs on LEADs (blue)
and on other days (green). For each test portfolio, we use the same full-sample beta estimate for both types of day, and we superimpose an ordinary least
squares best fitted line for each type of day. The sample period covers January 2001 through December 2019. (For interpretation of the references to colour
in this figure legend, the reader is referred to the web version of this article.)

nificantly associated with an increase in average excess As in Savor and Wilson (2014) and
return of 25.63 bps (t-statistic = 12.91). The intercept is Hendershott et al. (2020), we push the analysis further
negative (–7.40 bps with t-statistic = –3.45) and the re- by regressing the cross-section excess returns of portfolio
gression R2 is 78%. On other days, the beta coefficient of i for LEAD day t + 1 (xri,t+1
L ) on the prior-day portfolio
the green SML is –2.04 bps (t-statistic = –8.42), implying betas using the Fama and MacBeth (1973) regression:
a negative market risk premium on non-LEADs that lacks L
xri,t+1 = aL + bL βi,t + εi,t+1
L
(1)
economic meaning. The non-LEAD intercept is significantly
positive (5.15 bps with t-statistic = 19.74) and the regres- where β i,t is calculated from a 12-month rolling regression
sion R2 is 60%. When LEAD and non-LEAD returns are of excess returns of portfolio i on the daily market excess
pooled together, the resulting unconditional SML is largely returns. Similarly, we estimate the Fama-MacBeth regres-
flat with a beta coefficient of –0.72 bps (t-statistic = – sion for other days as:
3.12), the intercept is significantly positive (4.55 bps with
t-statistic = 18.37), and the R2 is 17%.
O
xri,t+1 = aO + bO βi,t + εi,t+1
O
(2)
If the CAPM is considered as a restriction that the in- The left-hand side of Panel A in Table 2 reports the
tercept is zero in the regression of portfolio excess re- Fama-MacBeth regression estimates and the corresponding
turns on the market excess returns (Gibbons et al., 1989), t-statistics (in parentheses). Standard errors are calculated
the insignificant intercept for LEAD returns shown in using the standard deviation of the time-series of coeffi-
Fig. 1 would be consistent with the CAPM, while the sig- cient estimates, but the significance levels barely change
nificantly positive intercept for non-LEAD returns is not. if we correct for heteroskedasticity and autocorrelation us-
For LEAD returns, the GRS F-statistic for the hypothesis ing Newey-West adjusted standard errors. The bL coeffi-
that the alphas of all the beta decile portfolios are jointly cient estimate is 22.24 bps (t-statistic = 2.19); this value
equal to zero is 1.75 (p-value = 0.07). This F-statistic is is close to the magnitude of the blue SML slope shown
marginally significant at the 10% level, but it is still much in Fig. 1. In economic terms, our finding means that on
lower than the GRS F-statistic estimated for other days LEADs, stocks with a beta that is higher by one have a
(F-statistic = 3.03 with p-value = 0.001). Likewise, the significantly higher average daily equity premium equal
GRS F-statistic for alphas of LEAD returns of the 50 beta- to 22.24 bps. In fact, this estimate is statistically indis-
sorted portfolios shown in Fig. 3 is statistically insignifi- tinguishable from the 15.86 bps average risk premium of
cant at 1.23 (p-value = 0.16) versus a statistically signif- the aggregate market portfolio realized on LEADs over the
icant F-statistic equal to 1.67 (p-value = 0.002) for non- 2001–2019 sample period (the t-test for the difference in
LEAD returns. In short, the GRS test suggests that the means gives a t-statistic of 0.51). The LEADs intercept is
CAPM captures the variation of average excess portfolio slightly negative, but it is statistically insignificant (aL = –
returns for LEADs, but the model is strongly rejected for 2.62 with t-statistic = –0.50), and the average regression
non-LEADs. R2 is 50%.

7
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

Table 2
Fama-MacBeth and panel regression results for various test portfolios.
This table reports the results from Fama-MacBeth and panel regressions of daily excess returns (expressed in
basis points) on market betas for various test portfolios. The betas used to construct the portfolios are estimated
using a 12-month rolling regression on daily excess returns. Panels A and B report the results for beta decile
portfolios (value-weighted portfolios for Panel A and equal-weighted portfolios for Panel B). Panel C reports
the results for 10 beta-sorted portfolios, 25 Fama and French size- and book-to-market sorted portfolios and 17
industry-sorted portfolios; all are value-weighted. The parenthesized t-statistics are estimated based on standard
errors calculated using standard deviations of the time-series coefficient estimates (for the Fama-MacBeth regres-
sion) and standard errors clustered by days (for the panel regression). The sample period covers January 2001
through December 2019. ∗ , ∗ ∗ and ∗ ∗ ∗ denote statistical significance at the 10%, 5%, and 1% level, respectively.

Panel A: 10 beta-sorted portfolios (value weighted)

Fama-MacBeth Pooled regression

Type of day Intercept β Avg R2 Intercept B DLEAD β x DLEAD R2

∗∗ ∗∗∗ ∗∗∗
LEADs −2.62 22.24 0.50 4.64 −1.58 −11.31 27.35 0.001
(−0.50) (2.23) (3.48) (−0.78) (−1.59) (2.63)
Other days 4.46∗ ∗ ∗ −1.60 0.51
(4.17) (−0.77)
LEADs - Other −7.08 23.84∗ ∗
(−1.44) (2.50)

Panel B: 10 beta-sorted portfolios (equal weighted)

Fama-MacBeth Pooled regression

Type of day Intercept β Avg R 2

Intercept β DLEAD β x DLEAD R2
∗∗ ∗∗∗ ∗∗ ∗∗
LEADs 7.88 13.12 0.64 9.29 −3.96 −4.50 19.50 0.001
(2.48) (1.39) (8.19) (−2.28) (−0.82) (2.50)
Other days 8.37∗ ∗ ∗ −3.27 0.61
(10.50) (−1.58)
LEADs - Other 0.49 16.39∗
(0.13) (1.73)

Panel C: 10 beta-sorted + 25 size/BM sorted + 17 industry portfolios

Fama-MacBeth Pooled regression

Type of day Intercept β Avg R2 Intercept B DLEAD β x DLEAD R2

∗∗ ∗∗∗
LEADs −3.37 20.31 0.21 2.98 0.09 −12.11 24.94 0.001
(−0.64) (2.07) (1.43) (0.05) (−1.23) (2.58)
Other days 4.38∗ ∗ ∗ −1.44 0.23
(3.73) (−0.70)
LEADs - Other −7.75 21.75∗ ∗
(−1.44) (2.29)

A different finding emerges for the other days: bO is – The right-hand side of Panel A in Table 2 reports the
1.60 and the corresponding t-statistic is –0.77, making it panel regression results, with the parenthesized t-statistics
statistically insignificant. That is, the Fama-Macbeth proce- computed using standard errors clustered by trading days.
dure suggests that the beta-return relation on other days The a and b1 coefficient estimates are 4.64 bps and –
is mostly flat. The net difference between bL and bO is 1.58 bps, respectively; these values closely match the
23.84 (t-statistic = 2.50), which indicates that the im- aO and bO estimates reported for the Fama-MacBeth re-
plied market risk premium is 23.84 bps higher on LEADs gression. The b3 coefficient is statistically significant at
than on other days. The other-day regression intercept is 27.35 bps (t-statistic = 2.63); this estimate is also close to
positive and statistically significant (aO = 4.46 with t- the net difference of 23.84 bps obtained from the Fama-
statistic = 4.17), and the average regression R2 is 51%. The MacBeth procedure. The coefficient of the DLEAD dummy
net LEADs-minus-other-days intercept is –7.08 but it is sta- variable, which captures the net “LEADs-minus-other-days”
tistically insignificant (t-statistic = –1.44). intercept, is –11.31. This number, although slightly be-
To explore further, we estimate the following panel re- low the –7.08 difference obtained using the Fama-MacBeth
gression modified from Savor and Wilson (2014): methodology, has a t-statistic of –1.59, making it statisti-
cally insignificant.
xri,t+1 = a + b1 βi,t + b2 Dt+1
LEAD
+ b3 βi,t × Dt+1
LEAD
+ εi,t+1 , (3) Panel A of Fig. 4 shows that our key result con-
where xri , t + 1 is the excess return of portfolio i on all tinues to hold for equal-weighted beta decile portfolios.
LEAD is a dichotomous variable equal to 1 on LEAD
days, Dt+1 The blue SML has a significantly positive slope on LEADs
and 0 on other days, and the b3 parameter directly cap- whereby an increase of one in market beta is associated
tures the net difference in implied market risk premium with a 15.26 bps increase in average excess returns (t-
on LEADs and on other days. statistic = 5.68). In contrast, the green SML has a signif-

8
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

Fig. 4. Average excess returns for different test portfolios. Panel A plots the average daily excess returns (expressed in basis points) against market betas
for 10 equal-weighted beta-sorted portfolios. Panel B presents analogous plots for 10 beta-sorted portfolios, 25 Fama and French size- and book-to-market
sorted portfolios and 17 industry-sorted portfolios (all are value-weighted). In each panel, we plot the conditional SMLs on LEADs (blue) and on other days
(green). For each test portfolio, we use the same full-sample beta estimate for both types of day, and we superimpose an ordinary least squares best fitted
line for each type of day. The sample period covers January 2001 through December 2019. (For interpretation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)

icantly negative slope: on other days, an increase of one the blue SML has a strongly positive slope: a unit increase
in market beta is associated with a reduction in average in beta is linearly related to an increase in average excess
excess returns of 3.62 bps (t-statistic = –8.96). When we return of 22.38 bps (t-statistic = 6.77), implying a signif-
estimate the Fama-MacBeth and panel regressions, we get icantly positive market risk premium on LEADs. The in-
similar results, as reported in Panel B of Table 2. For the tercept is negative, but it is only significant at the 10%
Fama-MacBeth test, the implied market risk premium on level (–6.34 bps with t-statistic = –1.76), and the regres-
LEADs is 16.39 bps higher than on other days (t-statistic sion R2 is 48%. On other days, the beta-return relation is
of the difference is 1.73). This finding is confirmed by the strongly negative with an implied risk premium equal to –
panel regression, where the difference between the risk 1.43 bps (t-statistic = –3.35), the SML intercept is positive
premium on LEADs and on other days is 19.50 bps (t- (4.54 bps with t-statistic = 9.78) and the regression R2 is
statistic = 2.50). nearly thrice lower (18%) relative to the LEAD’s regression
Taken together, the key results presented in this sec- R2 value.
tion suggest that the cross-section of stock prices behave We confirm the above finding through the Fama-
very differently on LEADs and on other (i.e., normal) days, MacBeth and panel regressions, with the results reported
since market betas of the same portfolios, which have in Panel C of Table 2. For the Fama-MacBeth test, the
a slightly negative (or rather flat) cross-sectional relation implied market risk premium is significantly positive for
with returns during normal days, actually explain returns LEAD returns (20.31 bps with t-statistic = 2.07) but
on LEADs. it is negative for other-day returns (–1.44 bps with t-
statistic = –0.70). The difference in implied market risk
premiums for LEADs and other days is 21.75 bps (t-
3.2. Beta, size, book-to-market ratio and industry portfolios
statistic = 2.29); this estimate is close to the LEAD-minus-
other-days implied market risk premium estimated for the
Lewellen et al. (2010) advocate expanding the set of
beta decile portfolios reported in Panel A of Table 2. The
test portfolios to include those sorted on size and book-
average R2 s are about 21%−23% for both LEAD and other-
to-market ratio (B/M), because doing so provides a higher
day returns. Using panel regression, the LEAD-minus-other-
hurdle for accepting CAPM. To this end, we add 17 in-
days implied market risk premium is equal to 24.94 bps
dustry portfolios and 25 Fama and French size- and B/M-
(t-statistic = 2.58). The coefficient on the LEAD dummy is
sorted portfolios to the 10 beta-sorted portfolios and re-
–12.11 but it is not statistically significant (t-statistic = –
peat the earlier tests. We obtain the data for size, B/M,
1.23).
and industry portfolios from Kenneth French’s online data
library (https://mba.tuck.dartmouth.edu/pages/faculty/ken.
french/data_library.html). 3.3. Treasuries
Panel B of Fig. 4 plots the average excess value-
weighted returns on all 52 portfolios against market be- Wachter and Zhu (2021) point out that pre-scheduled
tas separately on LEADs and on other days. Once again, macro announcements contain informative news about

9
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

Fig. 5. Average excess returns for Treasuries. This figure plots the average daily excess returns (expressed in basis points) against market betas for US
Treasuries with different maturities. The conditional LEAD-SMLs is in blue, and the other day-SML is in green. For each test portfolio, we use the same
full-sample beta estimate for both types of day, and we superimpose an ordinary least squares best fitted line for each type of day. Constrained by data
availability, the sample period used in the analysis covers January 2001 through December 2017. (For interpretation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)

expected inflation, and thus Treasury bonds are poten- lishes that it must if the market is mean-variance effi-
tially more exposed to scheduled news than are equities. cient. That is, if a market proxy such as the CRSP in-
Savor and Wilson (2014) provide empirical support to this dex is ex post efficient, the linear relation will tautolog-
hypothesis and show that the upward-sloping equity SML ically hold ex post with a slope equal to the ex post
observed on macroeconomic news announcement days ex- market index premium. Panel A of Table 3 provides ev-
tends to U.S. Treasury bonds. Here, we confirm this gen- idence confirming this proposition: the average premium
eral insight in the context of scheduled corporate earnings on the CRSP value-weighted market portfolio is seven
news. To this end, we retrieve US Treasury bonds with ma- times higher at 15.86 bps for LEADs versus 2.17 bps for
turities of 1, 2, 5, 7, 10, 20, and 30 years from the CRSP other days. This 13.69 bps risk premium differential is
Daily Treasury Fixed Term Indexes File. economically large and statistically significant (t-statistic
Fig. 5 plots the average excess returns on Treasuries =1.75).
against market betas (calculated using a 12-month rolling To reinforce the finding, Panel B of Table 3 reports the
regression) separately for LEADs and for other days. As cumulative log excess returns of the CRSP market port-
the figure shows, Treasury returns have a relation with folio on different types of days. Cumulatively, the market
betas that is strikingly similar to that for equities: the outperformed the risk-free asset by 0.346 on LEADs ver-
linear blue SML slope is economically large and positive sus 0.677 on the much larger number of other days. The
(38.50 bps with t-statistic = 34.89), implying a significantly 0.346 cumulative number is economically significant, since
positive risk premium on LEADs, while the magnitude of LEADs consist of only 228 out of 4779 trading days over
the intercept, although significant, is virtually zero (coeffi- the 2001–2019 sample period. In other words, of the un-
cient = 0.69 with t-statistic = 3.88). The regression R2 is conditional log sum of 1.023, one-third is earned on LEADs
99%, suggesting that variation in market betas explains al- alone even though the LEADs account for only 5% of all
most all the variation in the average excess returns of the trading days.
Treasuries on LEADs. On the other hand, the green SML has We now explore several trading implications. First, we
a significantly negative slope of –9.82 (t-statistic = –11.89). assess the performance of three alternative strategies: (i)
That is, on other (“normal”) days, a negative risk premium invest in the aggregate market portfolio on LEADs and in
estimate emerges, with an increase of one in market beta the risk-free asset on other days; (ii) invest in the risk-
associated with a reduction in Treasury returns of about free asset on LEADs and in the market portfolio on other
10 bps. days; and (iii) buy-and-hold the market portfolio between
2001 and 2019. Panel C of Table 3 reports the results. Strat-
3.4. Market premiums and trading strategies egy (i), which targets the LEADs effect, generates a daily
mean and standard deviation of 0.013% and 0.253%, re-
We have documented a strong and positive cross- spectively. These values give a daily Sharpe ratio of 0.052,
section relation between market betas and average re- which when annualized, translates to a sizable estimate of
turns for equities and Treasuries on LEADs. We now show 0.838. This number is about three times the annualized
that this cross-section finding has a counterpart at the Sharpe ratio of 0.371 generated by strategy (ii), and twice
aggregate market portfolio level, as Roll (1977) estab-

10
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

Table 3 The long-short trading strategies reported in Panel D

Market premium and trading strategies.
are not beta-neutral in the sense of Frazzini and Peder-
Panel A reports average market risk premium for three types of days:
LEADs, other days, and both types of days together, from January 2001 sen (2014). To address this issue, Panel E reports the beta-
through December 2019. Panel B reports cumulative log excess returns neutral version of the betting-on-and-against-beta strategy
earned by investing in the aggregate market portfolio. Panel C reports that encompasses the following beta-neutral betting-on-
daily summary statistics (means, standard deviations and Sharpe Ratios) beta strategy on LEADs:
for three trading strategies: (i) invest in equities on LEADs and in the
risk-free asset on other days; (ii) invest in the risk-free asset on LEADs r10 − r f r1 − r f
and in equities on other days; and (iii) buy-and-hold equities over the − ,
full sample period. Panel D reports analogous statistics for the returns β10 β1
of two long-short trading strategies: (i) “betting-against-beta” strategy
that takes a long (short) position on low (high) decile beta-sorted port- where the subscript 1 (10) denotes low-beta (high-beta)
folio over the whole sample period; and (ii) the hybrid “betting-on- portfolio, and reversing to the beta-neutral betting-against-
beta” strategy that takes a long (short) position on high (low) decile beta strategy on other days:
beta-sorted portfolio on LEADs and switch to the betting-against-beta
strategy on other days. Both strategies (i) and (ii) in Panel D, however, r1 − r f r10 − r f
are not beta neutral. Panel E reports the statistics for the analogous beta − .
neutral long-short strategies. β1 β10
Panel A: Market risk premium We use the average post-ranking portfolio betas of 0.30
LEADs Other days All days for β 1 and 1.84 for β 10 . Panel D reports the findings. The
Sharpe ratio for the beta-neutral betting-on-and-against-
Average (bps) 15.86 2.17 2.82
beta strategy is 0.043 per day, or 0.693 when annualized.
Panel B: Cumulative log excess returns
This number, although lower than the beta-non-neutral
betting-against-beta strategy over the full sample period
LEADs Other days All days
(the Sharpe ratio is 0.871 when annualized), is still eco-
Cumulative returns 0.346 0.677 1.023 nomically significant in an absolute sense.

Panel C: Investing in the aggregatedmarket portfolio

(i) (ii) (iii) 4. Additional tests

Average (%) 0.013 0.026 0.034
Std deviation (%) 0.253 1.141 1.169 4.1. Realized betas
Sharpe Ratio 0.052 0.023 0.029
So far, our results have been based on Fama-MacBeth
Panel D: Investing in long-short portfolio (not beta neutral) beta estimates where both pre-ranking stock betas and
Hybrid Betting against beta post-ranking portfolio betas are calculated using a 12-
month rolling regression with daily returns. We update
Average (%) 0.044 0.016
Std deviation (%) 2.284 2.284 the post-ranking portfolio betas every day, but some may
Sharpe Ratio 0.019 0.007 question the wisdom of synchronizing updates that involve
a rolling prior 12-month interval, albeit using daily returns,
Panel E: Investing in long-short portfolio (beta neutral) with a more dynamic time-series that involves return ex-
Hybrid Betting against beta pectations conditional on earnings announcements.
To address such concerns, we repeat the earlier tests
Average (%) 0.102 0.128
Std deviation (%) 2.375 2.375 using daily betas estimated from high-frequency (i.e., in-
Sharpe Ratio 0.043 0.054 traday) stock returns. We use Refinitiv’s Thompson Reuters
Tick Historical database to extract stock prices every
25 min between 9:45am and 4:00pm, combined with the
overnight return (calculated between 4:00pm on day t-1
the annualized Sharpe ratio of 0.468 earned from the buy- and 9:45am on day t), for a total of 16 return components
and-hold strategy (iii). per day. The 9:45am starting time is chosen to mitigate is-
To the extent that high-beta stock returns eclipse those sues related to illiquidity at the market open.
of low-beta stocks on LEADs, and low-beta stocks outper- To keep the portfolio test analysis to a manageable
form their high-beta counterparts on other days, a hybrid level, we limit ourselves to S&P 500 constituents listed be-
“betting-on-and-against-beta” trading strategy that entails tween 2001 and 2019. We end up forming realized beta
buying the high-beta (decile 10) value-weighted portfolio decile portfolios using 668 unique stocks. For the market
and selling the low-beta (decile 1) value-weighted portfo- index, we rely on the S&P 500 index exchange traded fund
lio on LEADs, and the reverse (viz. buying low-beta stocks (SPDR) with ticker symbol SPY. We estimate the daily real-
and selling high-beta stocks) on other days, should yield ized beta for stock j as:
an economically meaningful return. Panel D shows that 16
k=1 r j,t,k rm,t,k
this is indeed the case, with the betting-on-and-against- Rβ j,t = 16 2 (4)
beta strategy earning a daily Sharpe ratio of 0.019, or 0.306 k=1 rm,t,k
when annualized. This number is three times the Sharpe
ratio earned by continuously betting-against-beta over the where rj,t,k refers to the 25-minute log returns on stock j
full sample period. during the kth intraday period on trading day t and rm,t,k is

11
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

Fig. 6. Average excess returns for realized beta-sorted portfolios. This figure plots the average daily excess returns (expressed in basis points) against
market betas for 10 value-weighted portfolios sorted on realized betas estimated using Eq. (4). Panel A plots the unconditional SML, and Panel B plots
the conditional SMLs on LEADs (blue) and on other days (green). For each test portfolio, we use the same full-sample beta estimate for both types of day,
and we superimpose an ordinary least squares best fitted line for each type of day. The sample period covers January 2001 through December 2019. (For
interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

the market log return.11 This intraday beta estimation pro- market risk premium for LEADs is economically and sta-
cedure deliberately follows Patton and Verardo (2012). We tistically significant at 23.45 bps (t-statistic = 2.39) and
construct beta decile portfolios as follows: on each trading the intercept, although negative, is statistically indistin-
day, we calculate realized betas for stock j using Eq. (4) and guishable from zero (–4.26 bps with t-statistic = –0.66).
group each stock into one of ten realized beta-sorted port- In contrast, the implied market risk premium on other
folios. We then use value-weights to average the realized days has a modest positive estimate equal to 5.86 bps (t-
betas and returns of individual stocks in each portfolio to statistic = 2.98) and the regression intercept is slightly
produce daily estimates of portfolio beta and portfolio re- negative (–1.18 bps with t-statistic = –0.99). The LEAD-
turn. minus-other-day implied market risk premium is 17.59 bps
As in the earlier figures, Fig. 6 shows the unconditional with t-statistic equal to 1.76. The average regression R2 s
SML in Panel A, and the SMLs conditional on LEADs and on are 61% and 52%, respectively, for LEAD returns and for
other days, separately, in Panel B. For all the SMLs, portfo- other-day returns. Confirming the Fama-MacBeth findings,
lio realized betas are averaged over the entire 2001–2019 the panel regression shows that the difference in implied
sample period. Focusing first on Panel B, the blue SML market risk premium between LEADs and other days is
shows a strong, approximately linear and positive, relation 14.20 bps with a t-statistic of 2.21. The regression coeffi-
between average excess returns and realized beta with a cient on the realized beta in the panel regression is 1.59
slope equal to 22.04 bps (t-statistic = 15.34). This indi- but it is not statistically significant (t-statistic = 1.47).
cates a positive implied market risk premium on LEADs The qualitative similarity between the SMLs presented
just as for the portfolios shown in Fig. 1. The intercept in Fig. 1 (with betas estimated using a 12-month rolling
is slightly negative, but it is not statistically significant (– regression on daily returns) and those in Fig. 6 (with real-
3.04 bps with t-statistic = –1.55), and the regression R2 ized betas derived from high-frequency returns) is not en-
is 97%. On other days, the SML also has a positive slope, tirely surprising. To further explore the dynamics of real-
but its magnitude is relatively negligible (3.47 bps with ized betas around earnings news reported by LEADers and
t-statistic = 5.24), and the regression R2 is 77%. When non-LEADers, we use the following specification modified
both conditional SMLs are combined, the resulting un- from Patton and Verardo (2012):
conditional SML shown in Panel A has a slightly positive
slope equal to 4.35 bps (t-statistic = 6.35). The intercept Rβi,t = αi + δt + γ−10
L
DLi,t+10 + · · ·+ γ0L DLi,t + · · ·+ γ10
L L
Di,t−10
is 0.94 bps (t-statistic = 1.01) and the regression R2 is
83%.
γ−10
O
DOi,t+10 + · · ·+ γ0O DOi,t + · · ·+ γ10 Di,t−10 + εi,t ,
O O
(5)
Table 4 reports the Fama-MacBeth and panel regres-
sion results for the realized beta decile portfolios. Once where DtL s are dummy variables over the [−10, +10] event
again, using the Fama-MacBeth procedure, the implied window days when LEADers make earnings announce-
ments, DtO s are the equivalent dummy variables centered
around non-LEADs when other S&P 500 announcers re-
lease earnings reports, and α i and δ t represent firm and
11
Patton and Verardo (2012, Appendix I) show that when the sampling
frequency is sufficiently high, the realized beta converges to the noisy but
year fixed-effects, respectively, to control for differences in
unbiased estimates of the true beta.

12
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

Table 4
Fama-MacBeth and panel regression results for realized beta-sorted portfolios.
This table reports the results from Fama-MacBeth and panel regressions of daily excess returns (expressed
in basis points) on market betas for 10 value-weighted portfolios sorted based on realized betas. We form
portfolios for each trading day, with stocks sorted according to realized beta estimated on a daily basis using
Eq. (4). The parenthesized t-statistics are estimated based on standard errors calculated using standard devia-
tions of the time-series coefficient estimates (for the Fama-MacBeth regression) and standard errors clustered
by days (for the panel regression). The sample period covers January 2001 through December 2019. ∗ , ∗ ∗ and
∗∗∗
denote statistical significance at the 10%, 5%, and 1% level, respectively.

Fama-MacBeth Pooled regression

Type of day Intercept β Avg R 2

Intercept β DLEAD β x DLEAD R2
∗∗ ∗∗ ∗∗
LEADs −4.26 23.45 0.61 3.15 1.59 0.69 14.20 0.001
(−0.66) (2.39) (2.18) (1.47) (0.11) (2.21)
Other days −1.18 5.86∗ ∗ ∗ 0.52
(−0.99) (2.98)
LEADs - Other −3.08 17.59∗
(−0.55) (1.76)

Fig. 7. Changes in realized beta around earnings news days. The figure plots the estimated changes in realized betas over the [−10, +10] event window
when earnings news is released on LEADs (Panel A) and on other days (Panel B). Point estimates are marked with a solid line, and 95% confidence
intervals (calculated from standard errors that clustered using two-way firm-year technique of Petersen (2009)) are plotted with dotted lines. To facilitate
comparison, both panels have the same y-scale. The sample period covers 2001 through 2019. (For interpretation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)

betas across stocks and to capture changes in betas over Fig. 7 provides a visualization of a regression test
time. of this prediction, with the full set of results avail-
Savor and Wilson (2016) report a higher abnormal able from the authors upon request. As can be seen, on
return for earnings announcing firms than for non- LEADs realized betas constructed from 25-minute intra-
announcing firms, and that early (late) announcers deliver day pricing intervals increase significantly by 17.2% (t-
higher (lower) returns. They propose a risk-based expla- statistic = 3.16), on average. On other earnings announce-
nation: investors superimpose aggregate information on ment days, the average realized beta also increases (rel-
firms’ individual cash flow earnings news, thus generat- ative to non-announcement days), but by a more modest
ing a high conditional covariance between firm-level and magnitude of 11.6% (t-statistic = 4.34). When we consider
market-level cash flow news (Savor and Wilson, 2016, pp. both LEADs and other earnings news days together, the un-
83–84). To examine this proposition, we posit that beta in- reported coefficient on the day-0 dummy variable is 12.7%
creases with a larger magnitude on days when LEADers (t-statistic = 5.43); this number is slightly below the value
disclose early earnings news than on days when other later of 16.2% obtained by Patton and Verardo (2012) for their
announcers make theirs. As such, our prediction is that in 1996–2006 sample period.
Eq. (5), γ0L > 0, γ0O > 0, and γ0L > γ0O . To summarize, the increase in average realized beta
on LEADs is consistent with variation in market expo-

13
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

sures for stocks as a whole around the known timing The blue SMLs in both panels show that the positive re-
of LEAD announcements (Patton and Verardo, 2012), as lation between average overnight returns and market betas
we would have expected when an aggregate of influen- on LEADs is three times stronger than on other days: an
tial S&P 500 firms report early earnings news. At the increase in market beta of 1 on LEADs is associated with
same time, we observe an upward slope in the ratio of an increase in average overnight return of 12.52 bps (t-
average returns to the increase in betas on LEADs, and statistic of the slope coefficient is 8.41). The intercept is
per Roll (1977), a higher implied premium for the market negative (estimate = –6.19 bps with t-statistic = –3.96),
exposure. and the regression R2 is 90%. What sets Fig. 8 apart from
Hendershott et al. (2020) study is the daytime blue-SML:
on LEADs, an increase in beta of 1 is positively associ-
4.2. Overnight returns versus trading day returns ated with a 10.08 bps increase in average daytime re-
turns (t-statistic = 3.09). The intercept is almost zero and
We have heretofore been analyzing daily close-to-close the regression R2 is 54%. We rationalize that for daytime
stock returns. However, Hendershott et al. (2020) docu- returns, the difference in LEAD-SML and other day-SML
ment that stocks are priced such that their average excess is likely due to institutional investors closing out their
returns are linearly and positively related to CAPM betas announcement-event positions during the day following
when the market is closed for trading, but the relation the announcement (except on Fridays when traders espe-
is negative during trading hours. In this section, we ex- cially want to get “flat” well before the weekend illiquidity
tend Hendershott et al. (2020) finding with an analysis as that occurs globally, and we have excluded Friday as a can-
to whether overnight-versus-daytime asset pricing differs didate “leading” day).
when LEADers make early earnings news announcements To explore further, we run two sets of Fama-MacBeth
versus times when other announcers are making theirs. tests: first on LEADs, and then on other days. In each
Following Lou et al. (2019) and set of tests, we regress average day returns, and average
Hendershott et al. (2020), we split the close-to-close overnight returns, separately, on market betas. This test-
−to−close
return on stock j for day t (r close
jt
) into close-to- ing strategy allows the estimation of the difference in im-
open overnight return (r jt
close−to−open
) and open-to-close plied market risk premium between daytime and overnight
−to−close
SMLs on LEADs, and between daytime and overnight SMLs
day return (ropen
jt
). These components are related on other days.
as: Table 5 reports the results. Panel A shows that the av-
 
−to−close
 −to−open
 −to−close
 erage implied market risk premium during LEAD-overnight
1 + r close
jt
= 1 + r close
jt
1 + r open
jt is positive and significant with a slope coefficient equal to
(6) 12.31 bps (t-statistic = 2.32). The implied market risk pre-
mium during trading hours on LEADs is also positive with
We repeat the earlier tests on both close-to-open an estimate equal to 10.80 bps, albeit with an insignificant
overnight return and open-to-close daytime return. Pre- t-statistic of 1.21. Together, the overnight-minus-daytime
ranking betas are estimated at the start of each month risk premium difference on LEADs is 1.51 bps; this es-
by regressing close-to-close returns of stock j on close- timate is never going to be significant, either economi-
to-close market returns using a 12-month rolling regres- cally or statistically (t-statistic of the difference is 0.14). On
sion. Stocks are sorted into one of the beta decile portfo- other days, the overnight implied market risk premium is
lios, portfolio returns are value-weighted and post-ranking significantly positive (4.71 bps with t-statistic = 4.53), but
portfolio betas are estimated differently for tables and fig- the implied market risk premium during trading hours is
ures: For tables, we follow Hendershott et al. (2020) and significantly negative (–6.11 bps with t-statistic = –3.47).
regress overnight returns of portfolio i on overnight re- The difference in risk premiums between both daytime
turns of the market portfolio (constructed using value- and overnight SMLs is 10.82 with t-statistic of 5.30.
weighted overnight returns of all individual stocks) using a Overall, our finding reinforces the key insight of
12-month rolling window. The daily post-ranking betas are Hendershott et al. (2020): on “normal” days, the implied
used in the Fama-MacBeth and panel regression tests. For market risk premium is positive overnight, but it turns to
figures, we average the unconditional post-ranking beta of negative when the market opens for trading. We extend
each portfolio i over the full sample period. the literature by showing that on LEADs, high-beta stocks
Panel A of Fig. 8 plots the SMLs at times when earn higher returns, and low-beta stocks earn lower re-
the market is closed for trading (i.e., overnight SMLs) turns, as the CAPM predicts, and this finding holds for both
on LEADs and on other days. Panel B presents the daytime and overnight. Lou et al. (2019) identify a “tug of
SMLs when the market is open for trading (i.e., day- war” between day and overnight returns, and they hypoth-
time SMLs) for both types of days. Looking at the green esize that different clienteles use different specialized trad-
SMLs in both panels, we first confirm the key finding of ing strategies overnight and during the day. One obvious
Hendershott et al. (2020): for non-LEAD “normal” days, implication from our Fig. 8 is that during LEAD reporting
market betas are positively related to overnight returns weeks, Lou et al. (2019) “regular” clienteles are disrupted
(slope = 4.45 bps with t-statistic = 8.03), but they are neg- along with the patterns of information release (given the
atively associated with daytime returns (slope = –6.40 bps high concentration of after-hours releases) and the cliente-
with t-statistic = –10.13). The regression R2 s for both day- les’ trading strategies.
time and overnight SMLs are around 90%.

14
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

Fig. 8. Average excess returns for beta-sorted portfolios on day and overnight. Panel A plots the average daily excess returns estimated overnight (close-
to-open) against market betas for 10 value-weighted beta-sorted portfolios, while Panel B presents the analogous plots for excess returns estimated during
the day (open-to-close). In each panel, we compare the SMLs when earnings news is released on LEADs (blue) and on other days (green). We use the same
full-sample beta estimate for all types of day, and we superimpose an ordinary least squares best fitted line for each type of day. The sample period covers
January 2001 through December 2019. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this
article.)

Table 5
Fama-MacBeth regression results for returns on day and night.
This table reports the results for Fama-MacBeth regressions of daily excess returns on mar-
ket betas for 10 value-weighted beta-sorted portfolios on LEADs (Panel A) and on other days
(Panel B). In each panel, we report the results for excess returns during the trading day and
overnight. The parenthesized t-statistics are estimated based on standard errors calculated
using standard deviations of the time-series coefficient estimates. The sample period cov-
ers January 2001 through December 2019. ∗ , ∗ ∗ and ∗ ∗ ∗ denote statistical significance at the
10%, 5%, and 1% level, respectively.

Panel A: LEADs Panel B: Other days

Type of day Intercept β Avg R2 Intercept β Avg R2

∗ ∗∗ ∗∗∗ ∗∗∗
Overnight −5.15 12.31 0.43 −2.42 4.71 0.51
(−1.91) (2.32) (−5.67) (4.53)
Day −0.96 10.80 0.51 6.39∗ ∗ ∗ −6.11∗ ∗ ∗ 0.49
(−0.20) (1.21) (6.53) (−3.47)
Overnight - Day −4.18 1.51 −8.81∗ ∗ ∗ 10.82∗ ∗ ∗
(−0.77) (0.14) (−8.26) (5.30)

4.3. Weekday and intraday returns day after the market has closed (Panel A2) to Thursday
before the market closes (Panel D1). However, the SML is
We have explored how daytime and overnight SMLs be- negative on Monday when the market is open (Panel A1),
have distinctively on LEADs and on other days. It is worth and from Thursday evening through to Friday (Panel D2
repeating here that our analysis defines LEADs as occur- to Panel E2).12 Recognizing traders’ desire to minimize in-
ring on Tuesday through Thursday in the first week of re- ventories going into Friday night and weekends, anecdotal
porting quarter q that has a minimal aggregate of 50 S&P Wall Street evidence is that unusual pricing behavior and
500 announcers. We exclude Mondays and Fridays as pos- negative returns can accompany major announcements on
sible LEADs given the evidence of weekend-induced ef- Fridays. Johnson and So (2018), p. 219 also provide com-
fects on those days (e.g., Birru, 2018). To assess whether plementary evidence that “…[financial] intermediaries de-
daytime and overnight Mondays and Fridays are indeed à mand greater compensation for providing [earnings an-
priori different, we now examine the SMLs on weekdays nouncement risk] liquidity to sellers, relative to buyers…”
separately. The possibility that Monday and Friday market behav-
The blue lines in Fig. 9 show, for LEAD week, the re- ior is different also echoes Birru (2018), who attributes
lation between beta and average day return, and the rela-
tion between beta and average overnight return. The fig- 12
As noted earlier, firms rarely report earnings news between 0:00am
ure shows that the relation is largely positive from Mon- and 16:00pm on Monday.

15
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

Fig. 9. Average day and night returns for beta-sorted portfolios on different weekdays. The top panels plot the average daytime returns against market
betas for 10 value-weighted beta-sorted portfolios from Monday through Friday. The bottom panels present the analogous plots for average overnight
returns. In each panel, we compare the SMLs when earnings news is released on LEADs (blue) and on other days (green). LEADs occur during midweek
from Tuesday to Thursday, and the day- and night-SMLs presented in panels A1 and A2 (panels E1 and E2) are estimated on Monday prior to LEAD-Tuesday
(on Friday following LEAD-Thursday). We use the same full-sample beta estimate for all types of day, and we superimpose an ordinary least squares best
fitted line for each type of day. To facilitate comparison, the respective top (bottom) panels have the same y-scale. The sample period covers January 2001
through December 2019. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

a weekday seasonality in market anomalies to an upbeat 15:30 pm to 16:00 pm. For brevity, we aggregate the re-
mood on Fridays and downbeat blues on Mondays. That turns from 10:30 am to 15:00 pm, and present the average
said, care must be taken in contrasting our finding to aggregated return-beta relation in Panel C of Fig. 10. Two
Birru’s (2018): his results are based on day returns in- observations emerge. First, Panel A shows a sharp upward-
stead of the overnight returns that we have just analyzed, sloping SML between 9:30 am and 10:00 am.13 Second,
though we show below that our beta-return positive re- the average return-beta relation is mostly upward sloping
lation largely holds during the first half-hour of daytime throughout the day, except in the final trading hour when
trading on LEADs. DellaVigna and Pollet (2009) also an- the slope of the SML is somewhat flat between 15:00 pm
alyze Friday announcements, and they report an under- and 15:30 pm (Panel D), before turning to slightly negative
reaction of stock prices to Friday announcements along between 15:30 pm and 16:00 pm when the stock market
with lower trading volume than that accompanying an- closes (Panel E).
nouncements on other days of the week. DellaVigna and
Pollet (2009) attribute the under-reaction to investor dis- 4.4. Individual stocks
traction with non-work-related activities on Friday. As
noted earlier, a greater under-reaction for higher beta So far, the analysis of the beta-return relation has been
stocks is another way of expressing our “non” result for focused on portfolios. The portfolio-level analysis, how-
Fridays. On the other hand, deHaan et al., 38) argue ever, has a significant drawback: it could conceal a large
that “…attention to [earnings announcements] is no dif- amount of important cross-sectional information in the
ferent on Fridays as compared to Monday – Thursdays” portfolio aggregation. To check this possibility, we now
for their 2001–2011 sample period which overlaps our’s. evaluate whether market betas explain returns across in-
deHaan et al. (2015) suggest that managers attempt to hide dividual stocks on LEADs and on other days. To this end,
bad information by reporting it on Friday when investor we include all individual common stocks used in the ear-
attention is typically lower. Their proposition would sug- lier portfolio tests, except that, following convention, we
gest the presence of bad earnings news on Fridays, but we require the stocks to have minimal price per share of $2,
find that there is nothing unusual about Friday earnings and we remove stocks for which the daily return is above
announcements in our sample: the ratio of earnings beats 200%. Finally, we match Compustat’s data records (required
to misses is largely similar to that on other days. to compute the stock’s book-to-market ratio) with stock
Finally, we evaluate the intraday SML on LEADs by slic- data retrieved from CRSP. We end up with 4267 stocks.
ing the open-to-close returns into half-hourly intervals. Panel A of Table 6 reports the findings for the Fama-
Fig. 10 presents the finding. Unlike the earlier daytime SML MacBeth regression (as before, separately, for LEADs and
analysis that relies on the 12-month rolling regression, the for other days) and the panel regression of individual stock
intraday SMLs in Fig. 10 are constructed using realized be-
tas. Following Hendershott et al. (2020), we estimate the 16 13
When analyzing Panel A, it is noteworthy that the beta-return relation
returns every half-an-hour, where the first interval covers could be noisy when trading begins at the start of the day, and more so
9:30 am to 10:00 am, and the final interval extends from after an overnight announcement,
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

Fig. 10. Average intraday returns for realized beta-sorted portfolios. This figure plots the intraday average returns (expressed in basis points) against
average realized betas for 10 value-weighted beta-sorted portfolios on LEADs (blue line) and on other days (green line). The panels show the SMLs for
returns calculated for each 30-minute interval, except for Panel C which shows the SMLs estimated using aggregated returns between 10:30 and 15:00. We
form portfolios every day, with stocks sorted according to realized beta estimated every day. For each plot, we use the same full-sample beta estimate for
both types of day, and we superimpose an ordinary least squares best fitted line for each type of day. To facilitate comparison, all panels have the same
y-scale. The sample period covers January 2001 through December 2019. (For interpretation of the references to colour in this figure legend, the reader is
referred to the web version of this article.)

Table 6
Fama-MacBeth and panel regression results for individual stocks.
This table reports the results from Fama-MacBeth and panel regressions of daily excess returns (expressed in basis
points) for individual stocks. We require stocks to have a price per share of $2 and over. Panel A presents the
regression results of daily excess returns on stock market betas only, and Panels B and C present the regression
results after controlling for log market capitalization (Size), book-to-market ratios (BM) and cumulative returns
over the past 12 months (PastRet). The parenthesized t-statistics are estimated based on standard errors calculated
using standard deviations of the time-series coefficient estimates (for the Fama-MacBeth regression) and based
on standard errors clustered by days (for the panel regression). The sample period covers January 2001 through
December 2019. ∗ , ∗ ∗ and ∗ ∗ ∗ denote statistical significance at the 10%, 5%, and 1% level, respectively.

Panel A: Beta only

Fama-MacBeth Pooled regression

Type of day Intercept β Avg R 2

Intercept β DLEAD β x DLEAD R2
LEADs 0.13 23.27∗ ∗ 0.036 6.43∗ ∗ ∗ 0.54 −12.91 27.35∗ ∗ ∗ 0.001
(0.03) (2.57) (3.32) (0.27) (−1.47) (2.68)
Other days 5.71∗ ∗ ∗ 1.60 0.040
(4.93) (0.84)
LEADs - Other −5.58 21.67∗ ∗
(−1.06) (2.49)

Panel B: With firm characteristics as controls (Fama-MacBeth regression)

Type of day Intercept β Size BM PastRet (x100) Avg R2

∗∗∗ ∗∗ ∗∗∗ ∗∗∗
LEADs 18.57 29.31 −2.12 25.64 0.75 0.068
(0.94) (3.19) (−2.28) (5.39) (11.24)
Other days 40.70∗ ∗ ∗ 7.80∗ ∗ ∗ −2.64∗ ∗ ∗ 20.61∗ ∗ ∗ 0.63∗ ∗ ∗ 0.074
(4.55) (3.93) (−5.72) (16.84) (40.04)
LEADs - Other −22.13 21.51∗ ∗ 0.52 5.03 0.12
(0.55) (2.36) (0.25) (0.90) (1.63)

Panel C: With firm characteristics as controls (pooled regression)

Intercept β Size BM PastRet (x100) DLEAD β x DLEAD R2

28.61∗ ∗ ∗ 5.31∗ ∗ ∗ −2.04∗ ∗ ∗ 22.36∗ ∗ ∗ 0.50∗ ∗ ∗ −14.60∗ 28.50∗ ∗ ∗ 0.004

(3.20) (2.62) (−5.25) (8.51) (9.03) (−1.67) (2.77)

excess returns on market betas only. Quantitatively, the es- 1.60 bps (t-statistic = 0.84) on other days. That is, the im-
timates are remarkably similar to the portfolio findings re- plied risk premium on LEADs is significantly higher than
ported earlier for the value-weighted beta decile portfo- on other days by 21.67 bps (t-statistic of the difference
lios. Starting with the Fama-MacBeth regression, we can is 2.49). This finding is confirmed using the panel regres-
see that market betas are strongly and positively related sion to estimate the difference in market risk premium be-
to individual stock excess returns on LEADs, but the rela- tween both types of days: the difference is 27.35 bps (t-
tion is mostly flat on other days: the implied market risk statistic = 2.68). In fact, this LEADs-minus-other-days risk
premium is 23.27 bps (t-statistic = 2.57) on LEADs versus premium difference is identical to the estimate obtained

17
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

from the equivalent panel regression used in the portfolio conditional SML that is largely flat for (i), and that has a
analysis (see Panel A of Table 2). positive slope but with an inconsequential magnitude for
We then add three control variables typically used in (ii), suggesting that market betas do not explain average
cross-section stock expected return analysis – log market returns on days with simply “intense-by-count” earnings
capitalization (Size), book-to-market ratios (BM), and cu- announcements. To save space, we report the robustness
mulative stock returns over the past 12-months (PastRet) test results in the online appendix.
– to market beta and re-run the Fama-MacBeth and panel
regressions (see also Savor and Wilson, 2014). Panel B of 5. Explanations
Table 6 reports the Fama-MacBeth regression estimates,
and Panel C presents the panel regression findings. We 5.1. LEADers as attention-getters
can see that the beta-return relation continues to hold. It
is strongly positive on LEADs but is mostly flat on other We define LEADs as a handful of adjacent early days
days. The net difference in implied market risk premium in each quarter when prominent S&P 500 companies col-
on LEADs and on other days is 21.51 bps with a t-statistic lectively make a first wave of scheduled earnings an-
of 2.36 for the Fama-MacBeth test, and it is higher at nouncements and thus “lead” the earnings reporting sea-
28.50 bps with a t-statistic of 2.77 for the panel regres- sons. This criterion for LEADs is designed to select influ-
sion. The results for the control variables tabulated for ential announcements with minimal engineering. As noted
both types of days are largely consistent with prior stud- at the outset, such announcements are likely ipso facto
ies: size is strongly negatively related to stock excess re- to be “attention getting.” Thus, our findings intersect with
turns, and BM and past 12-month returns have a strong the growing number of studies that examines the rela-
and positive relation with average returns. tion between asset pricing and attention getting, and at-
tention getting is possibly a candidate explanation for
4.5. Robustness tests our results. For example, Ben-Rephael et al. (2017) show
that macro news, including that for large firms in aggre-
Our key finding remains remarkably consistent when gate, is associated with micro-level risk premiums, while
we partition the data into two su-period samples: the Da et al. (2020) find larger average premiums on high-
LEAD-SML has a slope coefficient of 27.46 bps (t- beta stocks. Both studies attribute the higher premiums to
statistic = 4.33) in the 20 01–20 09 sub-period, and the heightening of institutional investors’ attention on an-
22.12 bps (t-statistic = 4.57) in the 2010–2019 sub-period. nouncement day by using Bloomberg query scores as in-
The difference in the implied market risk premium be- struments for institutional attention.
tween LEAD and other days is also similar across both sub- If LEAD announcements are also attention getting, it
periods (30.64 bps in 20 01–20 09 vs. 25.51 bps in 2010– is interesting to explore how they compare with the
2019). Bloomberg query scores. To provide insight on this, we
Savor and Wilson (2014) show that stock market beta use the query measure of institutional investor attention
is strongly related to average excess returns on days when mentioned above that was developed by Ben-Rephael et al.
key scheduled macroeconomic news, such as Federal Open (2017, 2021): the measure, termed “abnormal institu-
Markets Committee (FOMC) interest rate decisions and in- tional attention (AIA),” relies on news-searching and news-
flation rate are released. Hirshleifer and Sheng (2019) sug- reading on Bloomberg terminals. We check whether the
gest that on days when such macro news is released con- AIA metric is higher on LEADs than on other days. Since
comitantly with micro-level announcements, the former Bloomberg terminal users are primarily institutional in-
could operate as an attention trigger for the latter. From vestors who have the financial resources to access the
this viewpoint, it might be that our finding is confounded terminal, a heightening of AIA on LEADs could suggest
by macroeconomic news. However, LEADs rarely coincide that institutional investors pay closer attention to spe-
with news arrivals for key macro variables. Between 2001 cific stocks on those unique LEADs than they do on other
and 2019, only 8.8% (20/228) of the total number of LEADs days.
also happen to be scheduled macro news announcement To this end, we closely follow the procedure described
days. Also, there is virtually no difference, either statisti- by Ben-Rephael et al. (2017, 2021) in constructing the daily
cally or economically, in the return-beta relation estimated AIA metric. We construct a cross-section daily AIA for all
on those LEADs that coincide with macro news release and S&P 500 constituents, and the online appendix shows that
on LEADs excluding macro news. our result is robust when we use all constituent stocks
Nor are LEADs simply over-represented by days with a in the Russell 30 0 0 index. We do not only gauge institu-
flood of announcements for constituents that are all im- tional investors’ attention to specific announcers on day
plicitly assumed to have an equally important influence t; instead, we assess the attention given to (almost) all
on investor expectations. To verify this contention, we re- the stocks, since testing asset pricing models involves test-
define LEADs as (i) consecutive weekdays in the week that ing them all. We aggregate the AIAs across the S&P 500
simply has the highest number of S&P 500 firms reporting firms to obtain sumAIA for day t. If sumAIA = 100 on day t,
in quarter q, and (ii) the top five non-consecutive days in for example, it means that Bloomberg terminal users (i.e.,
quarter q with the highest number of announcers. We then institutional investors) pay close attention to 100 firms
repeat the earlier set of asset pricing tests on the value- on that specific day. All the numbers reported in Table 7
weighted beta decile portfolios for the re-defined LEADs. should be interpreted in the same way. For brevity, we de-
These two experimental renditions of LEADs generate a tail the construction of sumAIA in the online appendix. Also

18
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

Table 7 gies. The overweighted high-risk assets are supplied by

Descriptive statistics of sumAIA.
a leverage-unconstrained group of investors who under-
This table reports descriptive statis-
tics of daily counts of AIA (sumAIA) weight them while overweighting low-risk assets. The net
separately on LEADs and on other result in normal non-LEADs times is an equilibrium in
days. Limited by data availability, which the prices of high-risk assets are higher and ex-
the sample period for the AIA mea- pected returns lower (and conversely, low-risk assets are
sure covers February 2010 through
priced lower and expected returns are higher). As a result,
December 2019. The historical data
for Bloomberg’s attention measure we observe a flat SML in “normal” non-LEAD times.
are missing from December 6, 2010 However, as LEADs approach, daily betas and announce-
to January 7, 2011, and from August ment risk increase. This reflects the potential asymmetries
17, 2011 to November 2, 2011.
in information amongst investors about impending earn-
LEADs Other days ings numbers and earnings calls, accompanying guidance,
Mean 116 81 and the like. The informational risk at announcement time
Min 32 1 for any one firm is accentuated by potential spillover of
5 pctl 90 31 earnings surprises to other related stocks in which in-
25 pctl 105 60 vestors also have positions.
50 pctl 118 80
It becomes more expensive inter alia to short on an-
75 pctl 129 100
95 pctl 143 133 nouncement night/day. Johnson and So (2018) argue that
Max 160 311 these higher selling costs occur because market makers
want to trade out of their securities inventories and be
flat on announcement night. Selling on margin (“borrow-
in the online appendix, we plot the frequency distribu- ing”) can also be expensive. Thus, borrowing-constrained
tion of the daily mean of sumAIA on both LEADs and non- investors tend to reduce their expensive levered positions,
LEADs for each reporting quarter q. Both Table 7 and the particularly on LEADs where higher information spillover is
online appendix provide exploratory evidence that institu- more likely given our selection of clusters of leading influ-
tional investors’ attention increases more on LEADs com- encers with large market cap. With the reduced segmenta-
pared to other announcing days. On average, we find that tion, the Black SML “looks” increasingly like that for “rep-
institutional investors monitor 81 firms listed in the S&P resentative” investors in a standard CAPM.
500 on non-LEADs, but this number increases 40% to 116 Regarding the CAPM-like pricing on LEADs,
firms on days when LEADers report earnings news. Roll (1977) shows that the linearity of ex post average
If the research motivation is to measure attention get- returns in beta is tautological given the ex post efficiency
ting, it is plausible that LEADs are a stronger proxy than of the market proxy. The efficiency of the market is not
the AIA metric. For example, Hirshleifer and Sheng (2019, relevant in the Black segmented equilibrium on days other
p. 3) find that: “…the attention trigger effect is more than LEADs: at those other times, neither borrowing-
pronounced among large firms and firms with high ana- constrained nor borrowing-unconstrained investors hold
lyst coverage.” It is also entirely plausible that the causal- the market, and the SML tends to be flat. The market
ity could go from asset price changes on LEADs to AIA clears with “discounts” for high-risk stocks and “premi-
queries (i.e., sharp moves in stock prices induce queries ums” for low-risk stocks (relative to the all-else-equal
rather than vice versa). Reverse causality seems even more CAPM equilibrium expected returns).
likely when LEADs and risk premium shifts are occurring We use both the CRSP and the S&P 500 indexes to
overnight while queries take place around say the begin- proxy the market. On LEADs with pre-scheduled public an-
ning of the succeeding trading day. This is one reason nouncements, the ex-post efficiency of both market prox-
why we define LEADs in terms of a clustering of firms, ies is plausibly higher: the considerable re-balancing in
but then look at the structure of cross-sectional risk pre- the market on announcement day arguably gets the av-
miums at firm micro level. Moreover, we show that the erage investor closer to efficiency (Gallagher et al., 2010;
attention-getting effect on premiums reverses the very Lee and Zhu, 2018). In the segmented investor framework,
next day. It is more plausible that a Wall Street attention in post-LEAD periods that are at the beginning of the next
span lasts at most a day (or even at most a New York earnings announcement cycle, the difficulty of leveraging
minute!) than that it lasts for a whole week as in Ben- across the now-past announcements subsides, borrowing-
Raphael et al. (2021). constrained investors again push up the demand for risky
assets with higher betas, and the SML again flattens out.
5.2. Reduced segmentation on LEADs
5.3. Market exposure and equilibrium risk premium on
We next discuss a second framework that is consis- LEADs
tent with the preceding evidence about asset pricing be-
havior on both non-LEADs and LEADs. This framework was We have provided evidence that average cross-sectional
suggested by Black (1972) and envisages a segment of in- excess returns on portfolios of equities, Treasuries, and in-
vestors who are constrained in borrowing. Frazzini and dividual stocks are positively and linearly related to mar-
Pedersen (2014), and Marsh and Pfleiderer (2016) cal- ket betas on LEADs. We have also demonstrated a higher
ibrate Black’s model in which leverage-constrained in- excess return on the aggregate market portfolio on LEADs.
vestors overweight high-risk assets in their trading strate- Taken together, these confirm Roll’s (1977) proposition that

19
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

the CRSP and SPDR market proxies that we have been us- In summary, our findings suggest that earnings an-
ing are at least approximately ex post efficient. That is, the nouncements reported on LEADs have “macro” implica-
linear return-beta relation holds tautologically ex post with tions for firm-level (micro) equity prices.
a slope equal to the ex post market premium. We now
show that the incremental increase that we reported above
for the market return premium on LEADs is also consistent References
with a simple equilibrium where market risk increases on
LEADs. Andrei, D., Cujean, J., Wilson, M., 2020. The Lost Capital Asset Pricing
Model. McGill University, Montreal, Canada Unpublished Working Pa-
The daily average realized variance (calculated as the per.
sum of squared 5-minute log returns) on the SPDR in- Antoniou, C., Doukas, J., Subrahmanyam, A., 2016. Investor sentiment, beta
dex is 1.62% on LEADs versus 1.38% on other days (i.e., and the cost of equity capital. Manag. Sci. 62, 347–367.
Bai, J., Philippon, T., Savov, A., 2016. Have financial markets become more
on average, the daily market realized variance increases by informative? J. Financ. Econ. 122, 625–654.
24 bps on LEADs). For an average investor with log utility Beaver, W., McNichols, M., Wang, Z., 2018. The information content of
(for which the one-period CAPM at issue here holds), the earnings announcements: new insights from intertemporal and cross–
sectional behavior. Rev. Account. Stud. 23, 95–135.
24 bps rise in market variance is consistent with a 24 bps Beaver, W., McNichols, M., Wang, Z., 2020. Increased market response to
rise in market premium on LEADs, which in turn is re- earnings announcements in the 21st century: an empirical investiga-
markably close to the LEAD-minus-other day implied risk tion. J. Account. Econ. 69, 101244.
Ben-Rephael, A., Carlin, B., Da, Z., Israelsen, R., 2021. Information con-
premium of 23.84 bps obtained from the Fama-MacBeth
sumption and asset pricing. J. Financ. 76, 357–394.
regression (see Table 2). In short, a LEAD announcement- Ben-Rephael, A., Da, Z., Israelsen, R., 2017. It depends on where you
risk explanation is also consistent with our results in an search: institutional investor attention and underreaction to news.
Rev. Financ. Stud. 30, 3009–3047.
equilibrium.
Birru, J., 2018. Day of the week and the cross-section of returns. J. Financ.
Econ. 130, 182–214.
Black, F., 1972. Capital market equilibrium with restricted borrowing. J.
Bus. 45, 444–455.
6. Conclusion
Black, F., Jensen, M., Scholes, M., 1972. The capital asset pricing model:
some empirical tests. In: Jensen, M. (Ed.), Studies in the Theory of
In this study, we analyze the CAPM beta-return rela- Capital Markets. Praeger Publishers Inc, New York, pp. 79–121.
Chan, K., Marsh, T., 2021. Asset prices, midterm elections and political un-
tion on days when large and influential firms disclose cor-
certainty. J. Financ. Econ. 141, 276–296.
porate earnings news announcements early in the earn- Cohen, R., Polk, C., Vuolteenaho, T., 2005. Money illusion in the stock mar-
ings seasons. We label these unique days as leading earn- ket: the Modigliani-Cohn hypothesis. Q. J. Econ. 120, 639–668.
ings announcement days (LEADs). On LEADs, market beta Da, Z., Hua, J., Hung, C.-.C., Peng, L., 2020. Market Returns and a Tale of
Two Types of Attention. University of Notre Dame, Notre Dame Un-
exposure has a strong and positive linear cross-sectional published Working Paper.
relation to average excess returns, and this relation is deHaan, E., Shevlin, T., Thornock, J., 2015. Market (in)attention and the
statistically significant and economically meaningful. This strategic scheduling and timing of earnings announcements. J. Ac-
count. Econ. 60, 36–55.
key finding extends to various test portfolios, Treasuries, DellaVigna, S., Pollet, J., 2009. Investor inattention and Friday earnings an-
and individual stocks. It holds during overnight (daytime) nouncements. J. Financ. 64, 709–749.
when the market is closed (open) for trading; and the re- Fama, E., French, K., 1992. The cross-section of expected stock returns. J.
Financ. 47, 427–465.
sult barely changes for betas estimated using a 12-month Fama, E., French, K., 2004. The capital asset pricing model: theory and
rolling regression on daily returns, or for realized betas es- evidence. J. Econ. Perspect. 18, 25–46.
timated using high-frequency (intraday) returns. Further- Fama, E., MacBeth, J., 1973. Risk, return and equilibrium: empirical tests.
J. Polit. Econ. 81, 607–636.
more, our result is not driven by strategically early or de-
Farboodi, M., Matray, A., Veldkamp, L., Venkateswaran, V., 2020. Where
layed news releases by LEAD announcers, and it is robust Has All the Data gone?. National Bureau of Economic Research Work-
to different testing procedures. On days other than LEADs, ing Paper #26927.
Frazzini, A., Pedersen, L., 2014. Betting against beta. J. Financ. Econ. 111,
the market beta-return relation is mostly flat.
1–25.
The linear upward-sloping security market line reported Gallagher, D., Looi, A., Pinnuck, M., 2010. Are active fund managers collec-
for LEADs implies a higher market risk premium on LEADs tors of private information or fast interpreters of public information?
than on other days. We confirm this proposition, and we Account. Financ. 50, 635–662.
Gibbons, M., Ross, S., Shanken, J., 1989. A test of the efficiency of a given
also show that one-third of the cumulative log excess mar- portfolio. Econometrica 57, 1121–1152.
ket return is earned on LEADs alone, even though LEADs Hasler, M., Martineau, C., 2021. Does the CAPM Predict Returns?. Univer-
account for only 5% of all trading days in our sample. We sity of Texas, Dallas Unpublished Working Paper.
Hendershott, T., Livdan, D., Rosch, D., 2020. Asset pricing: a tale of night
also propose a hybrid trading strategy that bets on beta on and day. J. Financ. Econ. 138, 635–662.
LEADs, and bets against beta on other days. The strategy Hirshleifer, D., Sheng, J., 2019. The Attention Trigger effect: Macro News
generates a Sharpe ratio equal to 0.306 when annualized, and Efficient Processing of Micro News. University of California, Irvine
Unpublished Working Paper.
an estimate that is three times the Sharpe ratio earned Hong, H., Sraer, D., 2016. Speculative betas. J. Financ. 71, 2095–2144.
from the familiar strategy that continuously bets against Johnson, T., So, E., 2018. Asymmetric trading costs prior to earnings an-
beta over the full sample period. Finally, we show that the nouncements: implications for price discovery and returns. J. Account.
Res. 56, 217–263.
average realized beta of firms announcing on LEADs in-
Jylha, P., 2018. Margin constraints and the security market line. J. Financ.
creases significantly relative to non-announcing days, and 73, 1281–1321.
that this beta increase on LEADs is also higher than the Lee, C., Zhu, C., 2018. Actively Managed Funds and Earnings News: Evi-
dence from Trade-Level Data. Stanford University, California Unpub-
rise in average realized beta of firms announcing on other
lished Working Paper.
days. Lewellen, J., Nagel, S., Shanken, J., 2010. A skeptical appraisal of asset pric-
ing tests. J. Financ. Econ. 96, 175–194.

20
JID: FINEC
ARTICLE IN PRESS [m3Gdc;July 4, 2021;10:42]

K.F. Chan and T. Marsh Journal of Financial Economics xxx (xxxx) xxx

Lou, D., Polk, C., Skouras, S., 2019. A tug of war: overnight versus intraday Penman, S., Zhu, J., 2019. An Accounting-Based Asset Pricing Model and
expected returns. J. Financ. Econ. 134, 192–213. a Fundamental factor. Columbia University, New York Unpublished
Marsh, T., Pfleiderer, P., 2016. Alpha signals, smart betas and factor model Working Paper.
alignment. J. Portf. Manag. 42, 51–66. Petersen, M., 2009. Estimating standard errors in finance panel data sets:
Michaely, R., Rubin, A., Vedrashko, A., 2014. Corporate governance and comparing approaches. Rev. Financ. Stud. 22, 435–480.
the timing of earnings announcements. Rev. Financ. 18, 2003– Roll, R., 1977. Critique of the asset pricing theory’s tests part I: on past
2044. and potential testability of the theory. J. Financ. Econ. 4, 129–176.
Nekrasov, A., Teoh, S., Wu, S., 2021. Visuals and attention to earnings news Savor, P., Wilson, M., 2014. Asset pricing: a tale of two days. J. Financ.
on Twitter. Rev. Account. Stud. (forthcoming). Econ. 113, 171–201.
Patton, A., Verardo, M., 2012. Does beta move with news? Firm-specific Savor, P., Wilson, M., 2016. Earnings announcements and systematic risk.
information flows and learning about profitability. Rev. Financ. Stud. J. Financ. 71, 83–138.
25, 2789–2839. Wachter, J., Zhu, Y., 2021. A model of two days: discrete news and asset
Penman, S., 1990. Earnings Reporting and the Return on the Stock Mar- prices. Rev. Financ. Stud. (forthcoming).
ket. University of California Berkeley, California Unpublished Working
Paper.

21