institutional trading and asset pricing · institutional trading and asset pricing bart frijns,...
TRANSCRIPT
Institutional Trading and Asset Pricing
Bart Frijns, Thanh D. Huynh, Alireza Tourani-Rad, P. Joakim Westerholm
This draft: 28 Aug 2015
Preliminary and Incomplete
Abstract
This paper examines whether the activeness of financial institutions in the market
can affect the relation between beta and average returns. We find that this relation
is strong and positive on days with high institutional trading activity (High-IT). In
contrast, on Low-IT days, the relation is negative and statistically significant. This
novel IT effect is strong and not driven by the known effect of macroeconomic news
or other alternative explanations. The evidence suggests that the trading activity of
investors could cause the Security Market Line to change slope.
IBart Frijns (E-mail: [email protected]), Thanh Huynh (Corresponding author, E-mail: [email protected]), and Alireza Tourani-Rad (E-mail: [email protected]) arefrom Auckland University of Technology, New Zealand. Joakim Westerholm (E-mail:[email protected]) is from the University of Sydney, Australia. We thank work-shop participants at McGill Global Asset Management Conference, Conference of the MultinationalFinance Society, World Finance Conference, Auckland Finance Meeting, AUT, Massey University,and Queensland University of Technology for helpful comments. We especially thank JonathanBrogaard, Andrew Karolyi, and Marios Panayides for helpful suggestions. Thanh Huynh gratefullyacknowledges the financial support from AUT Research Grant 2014 and SIRCA/AFAANZ Develop-ing Researcher Grant 2014. All errors are our own.
1. Introduction
Recent research has revisited the issue of the insignificant relation between beta
and average return. Frazzini and Pedersen (2014) suggest that the flat relation could
be attributed to leverage constraints and show that a portfolio that is long low-beta
stocks and short high-beta stocks is profitable. Savor and Wilson (2014) find that the
relation is conditional on news announcement and document a positive relation on
days with macroeconomic announcements. Motivated by DeLong, Shleifer, Summers
and Waldmann (1990) who suggest that financial institutions are less likely to be
noise traders, we hypothesize that when institutions trade intensively, noise in the
market is lower and the relation between beta and returns would be positive. In
contrast, on days when individual investors are more active, the relation would be
negative.
According to the noise trader hypothesis, when noise traders are optimistic about
high-beta stocks, they are willing to pay more for those stocks, causing them to be
overvalued. Thus, on days noise traders dominate the trading volume in the market,
we expect lower returns for high-beta stocks and, in turn, a negative relation between
beta and returns. In contrast, on days institutions are more active, the relation is
expected to be positive. Using a unique Finnish dataset that records daily trading
activities of all financial institutions, we find strong evidence for this hypothesis: on
days with high institutional trading (High-IT), beta is positively related to average
return whereas on other days, this relation is significantly negative. This finding
cannot be attributed to size, book-to-market, momentum, short-sales constraints,
liquidity effects or news announcements.
We obtain data on all stocks listed on the Nasdaq OMX Helsinki Exchange over
the period 1995-2011 from Euroclear Finland Ltd. (Euroclear). This dataset contains
full records of all trades conducted by institutions aggregated at a daily level. We
construct our measure of institutional trading (IT) activity as the fraction of total
trading volume by all institutions over total market volume, and label day t as a high
institutional trading day (High-IT) when IT on that day exceeds its average over the
past quarter. We find that, on High-IT days, an increase in beta by one is associated
1
with a significant increase in average return of about 7 basis points (bps). In contrast,
on Low-IT days, we observe a significant reduction in average daily excess return of
approximately 8 bps. These findings hold for beta-sorted portfolios, for size and
book-to-market ratio (BM) portfolios, for industry portfolios as well as for individual
stocks. The results are not driven by well-known institutional trading behaviors such
as the January and turn-of-month effects (Sikes, 2014) or by Nokia’s stock, which is
by far the dominant and most liquid stock in the Finnish market.
Finally, we also replicate our main findings using Thomson Reuters’ quarterly 13F
holdings data for the U.S. markets. Using ten beta-sorted portfolios as testing assets,
we find that the implied market risk premium is higher on High-IT quarters than
on Low-IT quarters, though the difference is weakly significant. The weak results
are primarily attributable to the fact that quarterly holdings miss out important
information in the high-frequency (daily) trade of institutions. We view this result as
strong support for our use of better Finnish daily institutional trading data as well
as a robustness test that our findings could be generalized to large markets such as
those in the U.S..
An alternative explanation for our findings is the announcement effect documented
by Savor and Wilson (2014). They show that the slope of the SML is significantly
different on days with interest rate announcements by the U.S. Federal Open Markets
Committee (FOMC) versus non-announcement days. We confirm these findings in
Finland using the European Central Bank (ECB) interest rate announcements. We
find that, after controlling for announcement days, the positive relation between beta
and average return on High-IT days remains strong and positive. Thus, our results
are not a manifestation of the announcement effect. We further examine which effect
is stronger and find that the announcement effect becomes weaker on Low-IT days,
especially when institutions buying volume is low. These findings indicate that the
IT effect is stronger.
The second alternative explanation is the leverage-constraints hypothesis. Black
(1972) and recently Frazzini and Pedersen (2014) show that, when investors are faced
with leverage constraints, they overweigh high-beta stocks, and therefore require lower
2
average returns on those stocks. Because institutions are on average less leverage-
constrained than individual investors, High-IT days may represent times when the
leverage constraint is low. A central prediction of this hypothesis is that institutions
can take advantage of the flat beta by buying low-beta stocks and sell high-beta
stocks. Our evidence is not consistent with this prediction. Regardless of the type
of days, stocks that institutions buy have a higher average beta than those that
they sell. Thus, leverage-constraints hypothesis does not seem to be the explanation.
More importantly, the pattern on Low-IT days is inconsistent with the leverage-
constraints hypothesis because, even when facing leverage constraints, investors would
not demand high-beta portfolios if they offer a risk premium lower than that of the
market.1
We next consider whether the investor disagreement hypothesis can explain the IT
effect. Hong and Sraer (2015) show that because high-beta stocks are more sensitive to
disagreement among investors about the future of the economy and because of limits
to arbitrage, high-beta stocks will be overvalued. Consequently, the relation between
beta and average return is negative on days with strong investor disagreement. On
days with weak disagreement, the SML will be upward sloping. Thus, High-IT days
could be times when investor disagreement is weak while Low-IT days could represent
strong disagreement periods.
Motivated by Dzielinski and Hasseltoft (2014), we proxy for investor disagreement
by news dispersion (defined as the daily standard deviation of firm-specific news tone
scores provided by Thomson Reuters). In line with Hong and Sraer (2015), we find
that the relation between beta and average return is significantly negative on days
with strong investor disagreement. In contrast, on weak disagreement days, the SML
is upward sloping. These results suggest that our proxy of market disagreement is
reasonable. Conditioning Low-IT days, we find that the relation between beta and
return becomes negative even when the day has weak disagreement. These findings
1This point is made clear in the model of Black (1972) (Equation 8) in which there is an upperbound on the Lagrange multiplier on the leverage constraint. This constraint is due to the fact thatno investor would demand the high-beta portfolio if it offered a risk premium lower than that of themarket, and so the Lagrange multiplier drops to zero.
3
suggest that disagreement in the market is not reflected in the trade of institutions
whose trades can affect the CAPM relation.
Finally, we examine whether the lottery preference can explain the downward
sloping SML on Low-IT days. According to this hypothesis, individual investors,
whose presence is more prevalent on Low-IT days, have a strong preference for stocks
with high skewness because they can offer more chances of large returns. Since these
stocks typically have higher betas, the strong demand for those stocks on Low-IT
days drive the SML downward. Mitton and Vorkink (2007) and Kumar (2009) show
that individual investors have a strong preference for stocks with lottery features.
Recently, Bali et al. (2015) show that this lottery preference can explain the return
on the betting-against-beta (BAB) portfolio of Frazzini and Pedersen (2014).
As suggested by Kumar (2009), we employ coskewness as a proxy for lottery stocks
and document two main findings. First, on days of low market coskewness, the SML is
upward sloping whereas on high coskewness days, the SML is downward sloping. This
result is consistent with the notion that lottery stocks are also those with high betas
and that lottery preference in the market can affect the CAPM relation. Second,
conditioned on high-coskewness days, we find that the High-IT effect is still strong
and positive, suggesting that lottery preference cannot explain our findings.
Taken together, our results are most consistent with the notion that the impact
of noise traders can cause high-beta stocks to be overvalued and therefore the SML
to be downward sloping on Low-IT days. When noise traders are not active (or
High-IT), we indeed see an upward sloping SML. Our study joins a large literature
including Barber et al. (2009), Grinblatt and Titman (1989, 1993), Daniel et al.
(1997), Grinblatt et al. (1995), and others to shed a positive light on the effect of
institutional trading. These studies generally suggest that institutions are informed
and their trades can improve intraday price discovery. Our study offers an asset
pricing perspective to this literature by linking the IT effect to the relation between
beta and average return. Our findings should not be interpreted to imply that the
performance of institutions can be explained by the CAPM as it is not the objective
4
of this paper to take side on this inconclusive debate.2 Our goal is to show that the
prediction of the CAPM relation is supported when (for whatever reason) there is
systematically High-IT activity in the market.
Though related, this study differs from Frazzini and Pedersen (2014) in three
important aspects. First, Frazzini and Pedersen (2014) also use the 13F data to test
whether institutions exploit the BAB strategy while we employ daily trading data for
all institutions. Using daily institutional trades gives us a more timely and accurate
estimate of beta. As our data covers all institutions in the Finnish market, we can
examine the effect of aggregate institutional trading on the relation between beta
and return. Furthermore, we are also able to separately study the buying and selling
sides of institutional trading, which help distinguish various competing hypotheses.
Second, the scope of our study is different. Our paper documents a novel finding
on the SML between High- and Low-IT days whereas Frazzini and Pedersen’s (2014)
focus is to explore the implication of leverage constraints through the BAB portfolio.3
Finally, we document a downward sloping SML on Low-IT days that is not consistent
with the leverage-constraints hypothesis. Our findings offer a new perspective to the
flat beta puzzle that when noise traders are less active, the CAPM is supported in
the data.
2. Data and empirical methodology
In this section, we discuss our unique data for the Finnish market that offers daily
trading records of all institutions. This gives us a comparative advantage over other
U.S.-based asset pricing studies that have to rely on quarterly institutional holdings
data. Nevertheless, we are able confirm that our conclusions do not change when
using U.S. data.
2Lewellen (2011) finds that institutions do not possess stock-selection skills and their performancecan be explained by book-to-market and momentum factors.
3Recently, Bali et al. (2015) show that the BAB phenomenon could be explained via the lotterypreference. It is not the objective of Bali et al. (2015) to explain the larger puzzle of the flat beta.
5
2.1. Data
This study employs the daily trading record of financial institutions for Finnish
stocks from Euroclear (also known as Finland’s Central Securties Depository). The
daily frequency gives this study an important advantage over prior research that
uses lower frequency data (e.g., mutual funds holdings or 13F quarterly institutional
holdings data). Moreover, the database contains trades by all institutions (identified
by a unique number of each institution aggregated at the daily level) while most
U.S. data only cover large institutions (small financial firms do not file 13Fs). These
features of our database allow us to compute a timely measure of the systematic
impact of institutional trading. The dataset has separate fields for institutional buy
and sell volumes. To trade on the Helsinki stock exchange, investors must register
with Euroclear and are given a unique account number. Our data consist of 187
stocks listed on the Nasdaq OMX Helsinki exchange between 1996 and 2011.4
We collect daily stock prices, dividends, capitalization adjustment, and the num-
ber of shares outstanding from Compustat Global.5 Book values of equity are obtained
from WorldScope. For tests that use the central bank’s interest rate announcement,
we collect scheduled days of monetary policy announcements from the European Cen-
tral Bank (ECB) website from 1999 when the ECB was officially established.6 Al-
though tests that employ ECB announcement days are limited to the sample period
between 1999 and 2011, the daily frequency of our data gives us well over 3,100 trad-
ing days, with about 370,000 of stock-days observations over this shorter period. All
returns are in euros, and excess returns are above the five-year government bond yield
obtained from Datastream. Betas are computed with respect to the value-weighted
4Grinblatt and Keloharju (2001) provide a detailed description of the Euroclear database and itsclassification between institutions and retail investors. Their data cover two years of 1995 and 1996for the 16 largest Finnish stocks. Exchange-traded funds (ETF) or index funds play a limited rolein the Finnish market because ETFs are not as popular as in the U.S.. For example, there were onlytwo ETFs listed on the Finnish market in 2006.
5Following Ince and Porter (2006) and Griffin et al. (2010), filters are applied on individual stockreturns in order to eliminate data errors. Specifically, returns that are greater than 100% in one dayis treated as missing. If daily return that is greater than 20% and then reversed immediately in thefollowing day, then returns on both days are treated as missing.
6https://www.ecb.europa.eu/press/govcdec/mopo/previous/html/index.en.html,
accessed 14 October 2014.
6
market return.7
< INSERT TABLE 1 HERE >
We form nine size- and BM- sorted portfolios in the spirit of Fama and French
(1993). Stocks are first sorted into three groups on the basis of their size (market
capitalization) at the end of June of each year. Big stocks are those in the top
30% of the market cap for the Finnish market and small stocks are those in the
bottom 30%. Independently, stocks are also sorted into three groups on the basis of
their book-to-market (BM) ratios. We use book values for the fiscal year ending in
calendar year t − 1, while market cap is for the end of December in calendar year
t−1. The nine size-BM portfolios are thus the intersections of three size and three BM
portfolios.8 We also form four beta-sorted portfolios and five industry portfolios using
Fama and French’s SIC classifications.9 Except for the beta-sorted portfolios that are
rebalanced monthly, all other portfolios are rebalanced on a yearly basis. Panel A
of Table 1 reports summary statistics for the Finnish market. The number of firms
listed on the Helsinki exchange increases from an average of 79 firms per year with
an average market capitalization (price times number of shares outstanding) of 389.9
million euros at the beginning of the sample to 140 firms per year with an average
market capitalization of 1095.5 million euros in the final years. The average fraction
of institutional trading volume (sell volume over the total market volume) increases
from 19.6% in the first five years to 37.9% in the last four years of the sample. There
is also a sharp increase in sell volume by institutions after 2008. Between 2008 and
2011, institutions’ sell fraction (over the total market volume) is 25.1%, which is
almost double the sell fraction earlier years. The number of trading accounts held by
institutions increases from 563 in the first subperiod to 722 in the third subperiod,
7We compute the market return as the value-weighted average of stock returns using the stock’slagged market capitalization. Our main results do not qualitative change when we use the MSCIindex.
8Due to the small size of the Finnish market, sorting stocks into three bins helps to maintain acertain level of diversification in each portfolio as well as the power of our tests.
9The classification is downloaded from Ken French’s website http ://www.mba.tuck.dartmouth.edu/pages/faculty/ken.french/datalibrary.html.
7
but then decreases to 656 for the last subperiod (2008-2011). Due to the substantial
rise in the trading volume in the last subsample, we ensure in one of the robustness
tests that our results are not specific to this subperiod.
Panel B reports average returns on nine value-weighted size-BM sorted portfolios.
We observe a strong size effect for growth stocks. For example, the small-growth
portfolio earns a significant average return of 10 basis points (bps) per day while the
average return on large-growth portfolio is only 6bps. The size effect is, however,
weak for value stocks. The value effect is present but in the opposite direction to the
U.S. evidence in which value stocks earn a lower average return than growth stocks.
This value effect is, nevertheless, strong for small stocks only. For large stocks, value
and growth portfolios earn similar returns. Panel C presents average returns on four
beta-sorted portfolios. The beta puzzle is apparent in the Finnish market where the
high-beta portfolio earns an average return of 3bps per day, which is lower than 5bps
average return of the low-beta portfolio. Panel D reports average returns on five
industry portfolios that are formed using industry classifications downloaded from
Ken French’s website. Only portfolios formed in Industry 2 (manufacturing, energy,
and utilities) and Industry 5 (business services and finance) earn significant average
returns over the sample period that we examine. Industry 3 (business equipment,
telephone, and television transmission), which includes Nokia, earns an average return
of 7bps, which is insignificant at the conventional level.
2.2. Empirical methodology
The main object of our analysis is the aggregate institutional trading (IT) volume
defined as IT = (buy+sell)/total market volume, where buy and sell are buy volume
and sell volume of financial institutions, respectively. We determine a day to be having
high institutional trading (High-IT day) when that day’s IT is higher than its average
over the past quarter.10 One can think of our measure as a time-series dummy variable
that takes a value of one on High-IT days and zero otherwise. There are 1598 High-
10We ensure that the results are robust to using various windows. For example, Appendix Apresents the main results for one-year windows.
8
IT days out of the total of 3,567 trading days.11 In some of the tests, we separately
examine institutions’ buying fraction (ITbuy = buy/total market volume) and selling
fraction (ITsell = sell/total market volume).
Our main empirical tests employ the two-pass Fama and MacBeth (1973) proce-
dure for the CAPM and then examine the coefficient estimate on High- and Low-IT
days.12 First, we estimate stock betas using one-year rolling regressions, adjusting
for the potential effect of non-synchronous trading by using Dimson (1979) sum beta.
Specifically, we run the regression Rt = α+β0RM,t+β1RM,t−1+β2∑4
k=2RM,t−k/3+εt,
where Rt and RM,t are excess returns on an asset and the market index, respectively.
The Dimson sum beta is then β0 + β1 + β2.13 The testing assets are four beta-sorted
portfolios, nine Fama and French size- and BM-sorted portfolios, five industry portfo-
lios, and individual stocks. Using portfolios as testing assets provides more accurate
estimates of market betas than those from individual stocks (Fama and French, 1992).
In the second-stage regression, we run the following cross-sectional regressions:
RHi,t+1 = γH0,t + γH1,tβ̂i,t (1)
RLi,t+1 = γL0,t + γL1,tβ̂i,t (2)
where β̂i,t is asset i’s stock market beta for period t estimated in the first stage;
and RHi,t+1 is the excess return on the testing asset on high institutional trading days
(High-IT or High).
As we are interested in studying the marginal effect of High-IT days on the relation
between beta and average return, we focus on the difference in the coefficient estimate
of γH1,t − γL1,t. Standard errors are computed using Newey and West’s (1987) method.
Similar to Savor and Wilson (2014), we also estimate the following pooled regres-
sion to test the difference in implied market risk premia between High-IT days and
11When we form portfolios, we lose one year of data to rank and sort stocks. Thus, the sampleperiod in tests starts in 1997.
12This methodology is similar to that of Fama and French (1992) and Savor and Wilson (2014)and is standard in the literature.
13Our results are robust to using no Dimson adjustment.
9
Low-IT days:
Ri,t+1 = γ0 + γ1β̂i,t + γ2Hight+1 + γ3β̂i,tHight+1 (3)
where Hight is a dummy variable that equals one for High-IT days and zero otherwise.
We compute clustered standard errors by date, which adjust for the cross-sectional
correlation of the residuals. The coefficient of interest is γ3, which shows the difference
in the implied risk premium on High- versus Low-IT days.
3. Results
Our empirical analysis focuses on the difference in market risk premium between
High- and Low-IT days (γH1,t − γL1,t in regressions (1) and (2)). We summarize our
findings in Figure 1, which uses portfolio returns as testing assets. We confirm that
our results are robust to individual stocks, subsample periods, the January effect, the
turn-of-month effect, and the use of U.S. 13 institutional holdings data.
3.1. Beta, book-to-market, size, and industry portfolios
Figure 1 summarizes our findings. The figure plots average excess returns on
nine portfolios (four beta-sorted and five industry portfolios) against their average
estimated betas. The upper graph shows the relation on High-IT days while the
lower graph plots the relation on Low-IT days. There is a striking difference between
the two graphs. The graph for High-IT days shows a strong, positive relation between
beta and average return. An increase in beta by 0.1 is associated with an increase in
expected return of about 1.7% per day (not reported in the graph). In stark contrast,
on Low-IT days the relation is significantly negative. An increase in beta by 0.1 is
associated with a statistically significant reduction in average daily excess return of
approximately 1.5% basis points (bps) per day, with an associated t-statistic greater
than three.14 This graph represents our novel findings that beta risk is important
14It is also interesting that the two extreme portfolios on the right-hand side of the graph exhibitcontrasting performances on High- and Low-IT days. Savor and Wilson (2014) document similarextreme performance on macroeconomic announcement days in the U.S. Although there is no reason
10
when institutions trade intensively.
Table 2 formally reports results for regressions (1) and (2) using the Fama-
MacBeth procedure. Specifically, panel A shows the results for four beta-sorted
portfolios. On High-IT days, the slope γH1,t is positive 7.3bps per day with an as-
sociated t-statistic of 2.6, significant at the 1% level. This positive slope is consistent
with the prediction of the CAPM and that beta risk is priced on High-IT days. In
contrast, the slope γL1,t on Low-IT days is −7.4bps (t-statistic = −2). The difference in
the slope between High- and Low-IT days (the bottom row) is 14.7bps per day, which
is statistically significant at the 1% level. Thus, there is a significant and negative
relation between beta and average return on Low-IT days.
The right-hand side of panel A presents the results from pooled regression (3) that
tests the difference in the implied market risk premium between High- and Low-IT
days. Consistent with the Fama-MacBeth regression results, the coefficient on the
interaction between High-IT days and beta (High × β) is positive and statistically
significant (2.8bps, with a t-statistic of 4.7 adjusted for clustering by trading day).
Again, these results show that beta is a much more important systematic risk on
High-IT days. It is also interesting to note the coefficient on High-IT dummy is
insignificant, suggesting that the difference in returns between High- and Low-IT
days is attributable to their portfolio betas, which carry a higher risk premium on
High-IT days.
As suggested by Lewellen, Nagel and Shanken (2010), it is useful to examine how
the model performs when more portfolios are included in the tests. Consequently,
panel B adds five industry portfolios to the test and brings the total of test portfolios
to nine. Consistently, we see that High-IT days exhibit a strong, positive relation
between beta and average return. The coefficient on beta is positive 7.6bps, which is
statistically significant at the 1% level. On the other hand, this coefficient is negative
to exclude those portfolios as we have already applied filters to our returns data, in unreported tests,we attempted to do so in the FM regression analysis and our main results do not qualitatively change.The slope of the security market line is still apparent when we exclude those portfolios. Finally,the results are robust to the use of individual stocks, which should span the SML more than doportfolios.
11
8bps on Low-IT days. The difference between the two days is 15.6bps per day with
an associated t-statistic of 3.2. Further, the pooled regression shows that a positive
coefficient of 18.9bps (t-statistic = 3.72) on the interaction term (High× β). These
results suggest that adding five industry portfolios to the set of testing assets does
not change our conclusions that beta risk is much more important on High-IT days.
Panel C further raises the hurdle by including nine size-BM portfolios as additional
test portfolios to panel B. The slope coefficient on beta is positive 7bps on High-IT
days (t-statistic = 2.8). On Low-IT days, the slope is again significant and negative
8.2bps. The difference in the slope on beta between High- and Low-IT samples
is 15bps with an associated t-statistic of 3.5. Consequently, we conclude that the
CAPM is supported on High-IT days, but fails on Low-IT days.15 16
< INSERT TABLE 2 HERE >
3.2. Excess returns on individual stocks
In this subsection, we employ individual stocks as testing assets and examine
whether the market risk premium is higher on High-IT days than on Low-IT days.
We report the results in Table 3.
< INSERT TABLE 3 HERE >
Panel A shows that, on High-IT days, the relation between beta and return is still
positive 3.9bps with an associated t-statistic of 1.8.17 In contrast, on Low-IT days,
15As pointed out by Savor and Wilson (2014), the difference in risk premium between High- andLow-IT days could be attributed to the difference in beta rather than in market risk premium.This concern arises because this methodology uses all data over the past year to estimate betas. Inunreported robustness tests, we follow Savor and Wilson (2014) and estimate beta in two separatesamples: High-IT days only sample and Low-IT days only sample. We then compute the difference inbetas between the two samples. Similar to the conclusion of Savor and Wilson (2014), the differenceis not economically and statistically significant, suggesting that using all data over the past yeardoes not badly affect our results. Furthermore, the results suggest that the difference in the impliedrisk premium, not betas, between High- and Low-IT days drives our findings.
16In another robustness test, we employ lagged High-IT and control for the market return inRegression (3) and find that the coefficient on β̂i,tHight+1 remains statistically significant. This isperhaps due to the persistence of the High-IT effect.
17The lower statistical significance reflects the estimation uncertainty when the testing asset isindividual stock return.
12
the slope coefficient on beta is negative 7.8bps (t-statistic = −3.2), significant at the
1% level. The difference in the slope on High- and Low-IT days (last row) is positive
11.7bps, which is statistically significant at the 1% level. In panel B, the coefficient
on the interaction term (High×β) is also significantly positive. Thus, our conclusion
that the market risk premium is higher on High-IT days than Low-IT days holds for
individual stocks.
In panels C and D, we include log(size), log(BM), and past one-year returns as
additional controls to the baseline model. On High-IT days, the coefficient on beta
is still positive while on Low-IT days, the coefficient is significantly negative. Again,
the difference in the slope on beta between two samples is positive and statistically
significant. Panel D shows that including size, BM, and past returns does not affect
the positive coefficient on the interaction term (High× β).
Another confounding effect that may affect the trade of financial institutions is
liquidity. As institutions tend to hold liquid stocks that have more efficient prices, the
significant difference in the coefficient on beta between high- and Low-IT days may be
confounded by the liquidity effect: Low-IT days may represent days with low liquidity
in the market and therefore stock prices are less efficient. We argue that liquidity
does not seem to significantly affect our results for two reasons. First, Boehmer and
Kelly (2009) show in an intraday analysis that institutions’ trade enhances market
efficiency that is beyond the effect of liquidity provision itself. Second, the effect of
liquidity, if any, would be larger for selling than for buying activity of institutions. As
we show in the next section, our results are driven by the buying, rather than selling,
side of institutional trades.
Nevertheless, we attempt to control for liquidity (proxied by turnover) in panels
E and F of Table 3. Turnover is defined as the average daily trading volume over
the past year divided by the number of shares outstanding. On each day t, we then
rank stocks based on their turnover and use the rank (Turn) as an additional control
in both Fama-MacBeth regression (panel E) and pooled regression (panel F).18 The
18As a robustness test, we also proxy for the level of liquidity of a stock by the number of oc-currences of zero returns over the past year. This liquidity proxy is in the spirit of Lesmond et al.
13
coefficient on Turn is positive but insignificant. On High-IT days, the coefficient
on beta is positive 4bps whereas on Low-IT days, this coefficient is negative 7.7bps
(t-statistic = −2.82). This leads to the average difference in the slope between High-
and Low-IT days of 12bps (t-statistic = 3.1). The pooled regression in panel F shows
consistent results.
Panels G and H repeat the regressions in panels E and F, but exclude Nokia from
the sample. Over our sample period, Nokia is the most liquid and largest firm by
market capitalization. Our findings are not driven by the dominant stock of Nokia
and that beta risk is significant on High-IT days.
3.3. Turn-of-month and January effects
Because institutions are known to pursue window-dressing behavior (Sikes, 2014),
we examine whether our results are due to their trading regularity throughout the
calendar year by excluding days that are turn-of-month from the analysis. Figure 2
contrasts the SML between high- and Low-IT days conditioned on non-turn-of-month
days. High-IT days still exhibit the positive relation between beta and average return
while Low-IT days show the negative relation. On High-IT days, an increase in beta
of the portfolio by one is associated with an rise in average return of 35bps per day
(t-statistic = 7.02). On Low-IT days, an increase in beta by one is associated with a
decrease in average return of 7bps per day (t-statistic = -3.8). These results suggest
that the turn-of-month behavior is not the explanation.
< INSERT FIGURES 2 and 3 HERE >
To complete the analysis, Figure 3 controls for the well-known January effect
(Sias and Starks, 1997) by excluding the month of January from the sample. Out-
side January, High-IT days still show strong support for the CAPM while Low-IT
days indicate the CAPM failure in explaining average returns. On High-IT days, an
(1999). Intuitively, stocks that were more illiquid over the past year should have more zero returnsthan those that were traded more frequently. Griffin et al. (2010) show that, in smaller and emerg-ing markets, this type of measure captures the illiquidity and transaction costs better than othermeasures. Our conclusions do not change.
14
increase in beta of the portfolio by one is associated with an rise in average return
of 27bps per day (t-statistic = 5.14). On Low-IT days, an increase in beta by one is
associated with a decrease in average return of 5.5bps per day (t-statistic = −2.9).
We conclude that regularities in stock returns throughout the calendar year cannot
explain our findings.
3.4. Subsample analysis: pre- and post-2008
The summary statistics in Table 1 show that there is an increase in trading volume
of institutions over the period between 2008 and 2011 that covers the period of the
Global Financial Crisis. Consequently, this subsection examines whether our results
are attributable to this later period by looking at two subsamples: pre- and post-2008.
< INSERT TABLE 4 HERE >
Table 4 reports results from Fama-MacBeth regressions and pooled regressions
that are estimated using samples before (panel A) and after 2008 (panel B). As
before, estimates are reported separately for High- and Low-IT days and testing
assets are four beta-sorted, nine size-BM, and five industry-sorted portfolios. The
general conclusion is that the difference in implied market risk premium between
High- and Low-IT days remains positive and statistically significant in all subsamples.
Before 2008, the slope on beta is positive 14bps on High-IT days with an associated
t-statistic of 2.1. On Low-IT days, the slope is negative 6bps, though statistically
insignificant. The difference in the coefficient on beta between two samples is 21bps
with a significant associated t-statistic of 2.6. The pooled regression also shows that
the coefficient on the interaction term is positive 26bps (t-statistic = 6.1).
We see a similar picture in panel B for the period between 2008 and 2011. The
relation between beta and average return is still strong and positive on High-IT days
while this relation is significantly negative on Low-IT days. The pooled regression
also shows that the implied market risk premium is much higher on High-IT days than
on Low-IT days. Consequently, the IT effect is not specific to a subsample period.
15
3.5. The use of U.S. 13F holdings data
In this robustness test, we employ Thomson Reuters 13F quarterly holdings data
for the U.S. markets and show that the High-IT effect does not seem to be specific
to the Finnish market. In each quarter, we compute aggregate change in holdings
of all 13F institutions scaled by the total trading volume between 1980 and 2014.
Similar to the main analysis, we define a quarter as High-IT when the fraction of
institutional volume over the market volume is higher than its average over the past
year. The High-IT dummy is then a monthly time-series dummy variable that takes
the value of one in months of High-IT quarters. We collect monthly market data
for all U.S. common equities (share codes of 10 and 11) from Center for Research
in Security Prices (CRSP). We estimate pre-ranking betas by regressing 60 months
of excess returns on excess market returns (adjusting for potential non-synchronous
trading using Dimson’s sum beta). We then form 10 value-weighted portfolios on the
basis of individual stocks’ monthly betas. We repeat the Fama-MacBeth regression
and pooled regression and report the results in Table 5. The primary limitation of 13F
holdings data (that motivates us to use a much comprehensive and high-frequency
Finnish dataset) is its low frequency. Since the quarterly frequency of 13F data
cannot not capture the timely trade of institutions, the power of our tests would be
significantly reduced, and inferences would have to rely on the sign and magnitude
of the coefficient rather than its statistical significance.
< INSERT TABLE 5 HERE >
Panel A of Table 5 reports the estimate from Fama-MacBeth regressions. On High-
IT quarters, the coefficient on beta is positive 27bps whereas, on Low-IT quarters,
this coefficient is negative 35bps. The difference in the coefficient on beta between
two types of day is 63bps (t-statistic = 1.8). The pooled regression shows a similar
picture that the coefficient on the interaction term (High× β) is positive 58bps with
an associated t-statistic of 2.5. Compared with the Finnish results in Table 2, the
magnitude of these coefficients is higher, and hence economically significant. These
16
findings show that the relation between beta and average returns is pronounced on
High-IT quarters in the U.S. markets.
These out-of-sample tests confirm our main conclusions and show that the High-
IT effect is not specific to the Finnish market. The results mitigate any potential
concerns of small markets and suggest that the institutional settings of the Finnish
market do not seem to be driving our findings.
4. Alternative explanations
In this section, we test whether other potential explanations of the flat beta puzzle
can explain our findings. Specifically, we investigate the role of macroeconomic an-
nouncements, short-sale constraints, leverage constraints, investor disagreement, and
lottery preference.
4.1. Macroeconomic announcements
Savor and Wilson (2014) find that the relation between beta and average return
is strong and positive on days when there is an interest rate announcement. This
relation on non-announcement days, in contrast, is downward sloping. Thus, one
possible explanation for our findings is that the IT effect could be a manifestation of
the announcement effect.
Figure 4 plots the SML on ECB scheduled monetary policy announcement days
and non-announcement days between 1999 and 2011. Similar to Savor and Wilson
(2014), we find a positive relation between beta and average returns on announce-
ment days. Non-announcement days, on the other hand, exhibit the well-documented
negative relation. Thus, the announcement effect is present in the Finnish market.
Consequently, this section examines whether the High-IT effect is driven by the an-
nouncement days. The analysis has two parts. First, we show that the announcement
effect cannot explain our findings. Second, we examine which effect is stronger by
running regressions on various combinations of High-IT and announcement days.
< INSERT FIGURE 4 HERE >
17
If the effect of institutional trading is a manifestation of the announcement effect
then we should not see the upward sloping SML on High-IT days when announcement
days are excluded from the sample. Figure 5 clearly shows that the positive relation
between beta and average returns is still present on High-IT days. Consequently, our
results cannot be explained by macroeconomic announcements.19
< INSERT FIGURE 5 HERE >
Table 6 reports regression results for various test portfolios that are estimated
on High- and Low-IT days, conditioned on those days being announcement or non-
announcement days. Panel A1 contrasts the slope coefficient on beta on days that are
both High-IT and having announcements versus all other days. (It is a time-series
dummy variable that takes the value of one if the day is High-IT and an announcement
day and zero otherwise.) On High-IT and announcement days, the coefficient on beta
is positive and statistically significant. In contrast, on other days, the relation between
beta and average return is flat. The difference in betas between the two samples is
positive and significant as shown in the last row of panel A1 as well as the interaction
term in the pooled regression.
Panel A2 considers days that are both Low-IT and having announcement ver-
sus other days. Thus, this panel is meant to study the effect of macroeconomic
announcements when the day is also Low-IT. If the announcement effect is robust
and strong, we should see a significant and positive coefficient on Low-IT and an-
nouncement days. In the Fama-MacBeth regression, although the coefficient on beta
is positive on announcement days and negative on non-announcement days, they
are both statistically insignificant and the difference between the two types of days
is also statistically insignificant (t-statistic = 1.2). However, the pooled regression
shows that the coefficient on the interaction term between announcement day dummy
and beta is positive 0.37% (t=3.4), indicating that the implied risk premium is higher
19Lucca and Moench (2014) document a drift on the day before FOMC announcement. In unre-ported robustness tests, we exclude both the ECB day and pre-ECB day. Our conclusions do notchange: the upward sloping SML is still strong on High-IT days even after excluding both days fromthe analysis. Thus, the results could not be explained by the effect of interest rate announcements.
18
on announcement days. These results suggest that the announcement effect becomes
weaker in the present of Low-IT days, though the beta risk is still important on
announcement days.
Panel A3 examines days that are High-IT and having no announcement versus
other days. This panel tests whether the effect of institutional trading is still robust
even when the day does not have macroeconomic announcement. High-IT days indeed
exhibit a positive relation between beta and average return while, on Low-IT days
this relation is significantly negative. The difference between two days is positive
25bps, which is statistically significant at the 1% level. In pooled regression results,
the coefficient on the interaction term (High × β) is also positive and statistically
significant. Consequently, these results show that the IT effect is still strong on
non-announcement days.
In short, panel A shows that the IT effect is distinct from the announcement effect
and if anything, it is even stronger.20
< INSERT TABLE 6 HERE >
Due to the conflicting results between Fama-MacBeth regression and pooled re-
gression on Low-IT-announcement days, we examine whether splitting the analysis
into IT buying and IT selling sides could offer a clearer picture. This test also shows
whether the stronger IT effect comes from buying or selling side. Panel B therefore
repeats the analysis for High-IT buying. The general conclusion from panel B is that,
when we examine the buying side of IT, the effect is stronger than the evidence of
total IT volume in panel A. The CAPM is strongly supported on High-IT buying
days, regardless of whether the day has announcement. In contrast, on Low-IT days,
the relation between beta and average return is flat even though those days also have
announcements.
20There are 85 days with High-IT and announcement while there are 92 Low-IT-announcementdays. These statistics suggest that there are enough number of observations in the two tests. Infact, the test of High-IT could be more conservative because the number of High-IT-announcementdays is fewer.
19
Panel C examines the effect of High-IT selling. We can see that when institutions
sell intensively, the relation between beta and average returns becomes insignificant
regardless of whether the day has macroeconomic announcement. On Low-IT selling
days, however, the relation is significant because days with low selling volume could
be associated with high buying volume, and therefore the low selling effect partly
picks up the effect of High-IT buying.
In short, this subsection shows that the announcement effect cannot explained
the IT effect (particularly buying). It is interesting that the effect of macroeconomic
announcements becomes insignificant when the days are Low-IT buying days. In
Appendix B, we further confirm that the IT effect cannot be explained by the U.S.
Federal Open Market Committee (FOMC) announcements.21
4.2. Short-sale constraints hypothesis
We examine whether short-sale constraints can explain our findings. This hy-
pothesis posits that High-IT days could be times when short-sale constraints are
less binding, which makes it less costly for institutions to short sell and, in turn, help
cause the stock price to be more efficiently. In fact, Nagel (2005) employs institutional
holdings as a proxy for short-sale constraints and finds that prices of stocks with low
institutional ownership do not quickly reflect cash-flow news.22 Although our data
do not allow us to identify short sales, under this hypothesis, we should observe that
the IT effect is stronger on the selling side of institutional trading. Consequently, we
separately examine the effects of High-IT buying and High-IT selling.
21Although we think that interest rate announcements are the most important macroeconomicnews, we also confirm the results in this subsection with other types of macroeconomic news aswell. Specifically, we employ the database of Thomson Reuters News Analytics (TRNA) to iden-tify macroeconomic news in Finland. Following Hendershott et al. (2015), we identify days withnews of the following topic codes as provided in the TRNA database: ECI (Economic Indicator),GVD (Government/Sovereign Debt), MCE (Macroeconomics). To be conservative, we also considerother U.S.-related macroeconomic news with the topic codes of FED (Federal Reserve) and WASH(Washington/U.S. Government news). We then repeat the exercise of this subsection. We confirmthat our conclusions do not change and the effect of High-IT is still strong even on days withoutany macroeconomic news, suggesting that macroeconomic information cannot explain our findings.These unreported results are available upon request.
22In an intraday analysis, Boehmer et al. (2008) and Boehmer and Wu (2013) find that shortselling activities tend to be more informed and enhance price discovery.
20
< INSERT TABLE 7 HERE >
Panel A of Table 7 reports regression results for High- versus Low-IT buying
days. On High-IT days, the coefficient on beta is positive 8bps with an associated
t-statistic of 3. This coefficient is even more significant than that reported in Table 2
for the total IT volume. The picture is reversed on Low-IT buying days on which the
coefficient is negative and statistically significant. The difference between two days is
17bps with an associated t-statistic above 3. The pooled regression shows a consistent
result that the coefficient on the interaction (High × β) is positive and statistically
significant. These results indicate that the IT effect is stronger on the buying side.
Panel B reports regression results for High- versus Low-IT selling days. In contrast
to the buying side, the selling side of IT exhibits a much weaker effect on the CAPM.
On High-IT selling days, the slope on beta in the Fama-MacBeth regresion is small
and statistically insignificant. The pooled regression shows a positive and significant
coefficient on the interaction term (High × β), but the magnitude is much smaller
than that for IT buying volume.
In short, Table 7 shows that the effect of IT is strong on High-IT buying days and
weak on Low-IT selling days – in consistent with the short-sale constraints hypothesis.
These findings are in fact intuitive. When institutions buy, they tend rely more on
the analysis of risk and return and could be using the standard capital asset pricing
model. On the other hand, they are less likely to sell a stock because of its risk level.23
4.3. Leverage-constraints hypothesis
We next consider the leverage-constraints hypothesis as a possible explanation of
the IT effect. Frazzini and Pedersen (2014) argue that the relation between beta
and average return is flat because investors are constrained in the leverage that they
23A related hypothesis that may explain our findings is liquidity: High-IT days may representtimes when liquidity in the market is high, and therefore stocks are priced more rationally. Thishypothesis is unlikely the explanation because we show in the previous section that our results arerobust to the control of liquidity. Moreover, if the liquidity effect drives our results, we shouldalso see stronger results on High-IT selling days because selling relies more on the liquidity in themarket. Because the effect of institutional trading is driven by the buying side, liquidity is unlikelythe explanation.
21
can take. They show that institutions that are less leverage-constrained can take
advantage of this flat beta by pursuing a profitable betting-against-beta strategy,
which buys low-beta stocks and sells high-beta stocks.
If High-IT days represent times when the overall leverage constraint is less binding
so that institutions can employ leverage to bet against beta, stocks that they buy
should have low betas while stocks they sell should have high betas. Table 8 shows
the average beta of stocks that institutions buy and sell on High- and Low-IT days.
On High-IT days, we can see that stocks that institutions buy have a higher average
beta than those that they sell and the difference is statistically significant. We also see
a similar picture on Low-IT days. Thus, the evidence does not suggest that Finnish
institutions take advantage of the betting-against-beta strategy.
< INSERT TABLE 8 HERE >
Finally, the downward sloping SML on Low-IT days is another evidence that
is inconsistent with the leverage-constraints hypothesis. Theoretically, Hong and
Sraer (2015) pointed out a limitation of the leverage-constraints hypothesis: it cannot
explain the downward sloping SML that we document on Low-IT days. This is
because investors would not buy stocks with high betas if they offer a risk premium
lower than that of the market.
4.4. Investor-disagreement hypothesis
Hong and Sraer (2015) contend that high-beta stocks are more sensitive to dis-
agreement among investors about the future of the economy. Consequently, in times
of high disagreement and together with the presence of limits to arbitrage, the op-
timists outweigh the pessimists in their trading impact on high-beta stocks. When
disagreement is sufficiently large, the expected return on high beta stock exhibits an
inverted U-shape curve. During times of weak investor disagreement, the SML will
be upward sloping. The investor disagreement hypothesis does not make a distinction
with the noise trader hypothesis and therefore our Low-IT days could also be times
when disagreement is strong.
22
Empirically, testing this hypothesis requires one to make a choice of proxy for
investor disagreement. Hong and Sraer (2015) employ the dispersion (standard devi-
ation) of analyst forecasts of the earnings-per-share and find that this measure can
yield the upward sloping SML when disagreement is low and downward sloping SML
when disagreement is high. In the context of our study that employs daily IT, this
measure is not preferable because of its low frequency and little variation in analyst
forecast over time. Consequently, we employ a new measure of aggregate news tone
dispersion that is similar in spirit to Dzielinski and Hasseltoft (2014) as a proxy of
investor disagreement.
We collect news data from Thomson Reuters News Analytics (TRNA) that sys-
tematically quantifies the tone of firm-specific news for Finnish stocks between 2003
and 2011 (2267 trading days).24 For every news item, TRNA provides the probability
that a news item has good, neutral, or negative tone (the three scores sum to one).
For each news article, we compute a unified tone score as the difference between good
and bad score.25 On each day t, we then compute news tone dispersion as the stan-
dard deviation of the tone score of all articles on that day. Finally, we define a day
to have high news dispersion if the day’s tone dispersion is greater than the average
dispersion over the past one month. As TRNA is reasonably comprehensive, approx-
imately 90% of the trading days have news and on average there are 52 news articles
per day. The average news dispersion is 0.48 per day and 1260 days are defined as
having high news dispersion (strong investor disagreement).
< INSERT FIGURE 6 HERE >
Dzielinski and Hasseltoft (2014) find that this news dispersion is strongly associ-
ated with, and can even drive, the analyst disagreement. Using this measure allows
us to construct a timely measure of investor disagreement. Figure 6 plots the SML
24This database has been employed in the literature (e.g., Hendershott et al. (2015)) to examinethe relation between news and stock returns or institutional trading behavior. Interested readersare referred to Hendershott et al. (2015) and Heston and Sinha (2014) for a detailed description ofthe database.
25We use all news items, but our results are robust to examining news with the relevance score ofone, which means that the firm’s ticker code is mentioned in the headline or title of the news article.
23
on weak- and strong-disagreement days. On weak disagreement days, the relation
between beta and average return is positive. An increase in beta by 0.1 is associated
with 5bps increase in average return, with a significant associated t-statistic at the 1%
level. In contrast, on days with strong disagreement, the SML is downward sloping:
an increase in beta by 0.1 is associated with 14bps lower average return per day, with
an associated t-statistic of 5.8. In short, with the new proxy for investor disagreement,
we are able to test and confirm the prediction of Hong and Sraer (2015).
< INSERT TABLE 9 HERE >
Table 9 tests whether the investor disagreement can explain the IT effect by run-
ning regressions on various combinations of High-IT and weak disagreement days.
In panel A, we can see that the IT effect is significant after we condition the day
on having weak disagreement, which is not surprising because both days exhibit an
upward SML. In panel B, the relation between beta and average return is flat on
Low-IT days even though the day has weak disagreement (which, as shown in Figure
6, should exhibit a significant and positive relation). This result suggests that the
IT effect is stronger than the effect of investor disagreement. Panel C shows that
the relation between beta and return is positive on days of High-IT and strong dis-
agreement. Although the relation on this day is insignificant, the difference in risk
premium between this day and others is statistically significant at the 10% level.
Further, the pooled regression shows a positive coefficient on the interaction term be-
tween beta and High-IT-strong-disagreement day, suggesting that the High-IT effect
is still stronger.
In short, the results are not consistent with the notion that High-IT days are
times of weak disagreement and Low-IT days are times of strong disagreement in the
market. We however cannot rule out the possibility that disagreement and noise are
correlated because, as for prior studies, our measure is a noisy proxy of the true market
disagreement. To the extent that our measure of market disagreement can successfully
exhibit the upward or downward slopes of the SML as shown in Figure 6, the findings
show that noise, rather than market disagreement, is apparent on Low-IT days and
24
absent on High-IT days. At the minimum the findings suggest a novel implication
that follows from the disagreement hypothesis: the degree of disagreement among
institutions, which is reflected through their trades, is an important determinant of
the relation between beta and return.
4.5. Lottery preference by individual investors on Low-IT days
Another potential explanation for the downward sloping SML on Low-IT days is
the lottery preference by individual investors as put forth by Mitton and Vorkink
(2007). According to this hypothesis, individual investors, whose trade’s impact is
more prominent on Low-IT days, are willing to sacrifice some returns to hold stocks
with lottery features such as high beta and high coskewness. In this subsection,
we test whether this hypothesis can explain the IT effect. We estimate the stock’s
coskewness (denoted as the coefficient ci) as in Harvey and Siddique (2000) and Mitton
and Vorkink (2007) by running the following quarterly regression:
Ri,t = a+ biRM,t + ciR2M,t + εi,t (4)
where Ri,t and RM,t are excess returns on the stock and the market, respectively.
The coefficient ci is the coskewness of the stock. The market coskewness is then the
simple average of individual stocks’ coskewness on each day. Day t exhibits high
coskewness if the market coskewness is higher than its average over the past quarter.
Figure 7 shows that on days of low coskewness, the SML is upward sloping while on
high-coskewness days, the relation is downward sloping. These results suggest that
lottery preference in the market can also affect the relation between beta and return.
< INSERT FIGURE 7 HERE >
Table 10 examines the IT effect conditional on days of high or low coskewness.
Panel A shows that the relation between beta and return is positive on days of High-
IT and low coskewness. This is not surprising because both types of days exhibit an
upward SML. On days of Low-IT and low coskewness, however, the relation is flat
(panel B). This result indicates that the effect of low coskewness in the market is not
25
strong enough to drive the SML upward on Low-IT days. We also observe that the
High-IT effect is strong on high coskewness days (panel C). Together, these results
suggest that the lottery preference hypothesis cannot explain our findings.26
< INSERT TABLE 10 HERE >
4.6. Endogeneity concern
One may be concerned about the endogeneity issue of the IT effect. Specifically,
the CAPM works on High-IT days either because high IT causes the CAPM to work
or because institutional traders predict that the CAPM could work on High-IT days
and therefore try to trade more. First and foremost, regardless of the direction of
the causality, our findings that the CAPM is supported on High-IT days is still novel
and interesting. Since we are interested in testing whether the CAPM works on
High-IT days, disentangling the two-way causality should not affect our conclusions.
Second, if institutions could predict the return to be higher on High-IT days, they
would trade more heavily in the direction of the return on those days. The pooled
regression in Table 2 shows that the coefficient on High-IT day dummy is statistically
insignificant, indicating that the difference in returns between High- and Low-IT days
is not significant. This result suggests that institutions are unlikely to predict the
positive relation between risk and returns on High-IT days.
Finally, if institutional traders expect the CAPM to work on High-IT days, we
should observe that the stocks they buy would have higher beta (to earn higher
returns) than those they sell. In contrast, on Low-IT days the stocks that they buy
would have lower beta than those they sell. Table 8 shows that the stocks institutions
buy have a higher average beta regardless of the type of IT days. Further, beta on
26In unreported robustness tests, we confirm that our conclusions do not change when we use thereturn on MAX portfolios of Bali, Cakici and Whitelaw (2011) as an alternative proxy for individualinvestors’ lottery preference. Specifically, stocks are ranked and split into five groups on the basis oftheir average five maximum daily returns over the past month. Portfolio 1 contains stocks with thelowest maximum daily returns while portfolio 5 consists of stocks with the highest maximum dailyreturns. The value-weighted average raw return difference between portfolio 5 and portfolio 1 is theMAX portfolio. Bali et al. (2011) show that this MAX portfolio earns a negative average returnbecause investors have a strong preference for lottery-type stocks.
26
High-IT days is lower than that on Low-IT days and the difference is statistically
significant. These findings suggest that institutions do not predict the CAPM to work
on High-IT days and then trade accordingly. Rather, the evidence points toward the
first scenario.
5. Conclusion
By employing a comprehensive dataset of daily institutional trades, this study is
one of the first to directly link the effect of institutional trading (IT) to the relation
between beta and average return. We find that this relation is strong and positive
on days when there is high institutional trading (High-IT). On Low-IT days, how-
ever, this relation is negative and statistically significant. Moreover, the difference
in market risk premium between High- and Low-IT days is positive and statistically
significant. These findings hold for various test portfolios and for individual stocks.
The results are not driven by a specific subsample period, the January effect, or the
turn-of-month effect.
Our findings are consistent with the notion that it is the trading activity of two
main trader groups that causes the relation between beta and return to change sign –
depending on who is active in the market. When individual investors are more active
(or institutions are less active), we see a downward sloping SML – consistent with the
notion that noise in the market is high on those days. Our results are unlikely to be
explained by alternative hypotheses such as the short-sale constraints hypothesis, the
liquidity effect, small market’s problems, macroeconomic announcements, leverage-
constraints, market disagreement, or lottery preference. Further, the evidence does
not seem to suggest that institutions expect the CAPM to work and therefore act
accordingly on High-IT days.
27
References
Bali, T. G., Brown, S., Murray, S. and Tang, Y. (2015) Betting against beta
or demand for lottery?. Georgetown McDonough School of Business Research
Paper No. 2408146. Available at SSRN: http://ssrn.com/abstract=2408146 or
http://dx.doi.org/10.2139/ssrn.2408146.
Bali, T. G., Cakici, N. and Whitelaw, R. (2011) Maxing out: Stocks as lotteries and
the cross-section of expected returns. Journal of Financial Economics 99: 427–446.
Barber, B. M., Odean, T. and Zhu, N. (2009) Do retail trades move markets?. Review
of Financial Studies 22: 151–186.
Black, F. (1972) Capital market equilibrium with restricted borrowing. Journal of
Business 4(3): 444–455.
Boehmer, E. and Kelly, E. K. (2009) Institutional investors and the information
efficiency of prices. Review of Financial Studies 22(9): 3563–3594.
Boehmer, E. and Wu, J. (2013) Short selling and the information efficiency of prices.
Review of Financial Studies 26(2): 287–322.
Boehmer, E., Jones, C. M. and Zhang, X. (2008) Which shorts are informed?. The
Journal of Finance 63: 491–528.
Daniel, K., Grinblatt, M., Titman, S. and Wermers, R. (1997) Measuring mutual
fund performance with characteristics-based benchmarks. The Journal of Finance
52: 1035–58.
DeLong, J. B., Shleifer, A., Summers, L. and Waldmann, R. J. (1990) Noise trader
risk in financial markets. Journal of Political Economy 98: 703–738.
Dimson, E. (1979) Risk measmeasure when shares are subject to infrequent trading.
Journal of Financial Economics 60: 1673–1712.
Dzielinski, M. and Hasseltoft, H. (2014) Why do investors disagree? The role of a
dispersed news flow. Western Finance Association 2014 Meeting.
28
Fama, E. and French, K. R. (1992) The cross section of expected stock returns.
Journal of Financial Economics 47: 427–465.
Fama, E. and French, K. R. (1993) Common risk factors in the returns on stocks and
bonds. Journal of Financial Economics 33: 3–56.
Fama, E. F. and MacBeth, J. D. (1973) Risk, return and equilibrium: Empirical
tests. Journal of Political Economy 81: 607–636.
Frazzini, A. and Pedersen, L. H. (2014) Betting against beta. Journal of Financial
Economics 111: 1–25.
Griffin, J. M., Kelly, P. J. and Nardari, F. (2010) Do market efficiency measures yield
correct inferences? A comparison of developed and emerging markets. Review of
Financial Studies 23: 3225–3277.
Grinblatt, M. and Keloharju, M. (2001) What makes investors trade?. The Journal
of Finance 56: 589–616.
Grinblatt, M. and Titman, S. (1989) Mutual fund performance: An analysis of
quarterly portfolio holdings. Journal of Business 62: 393–416.
Grinblatt, M. and Titman, S. (1993) Performance measurement without benchmarks:
An examination of mutual fund returns. Journal of Business 66: 47–68.
Grinblatt, M., Titman, S. and Wermers, R. (1995) Momentum investment strategies,
portfolio performance, and herding: A study of mutual fund behavior. American
Economic Review 85: 1088–1105.
Harvey, C. R. and Siddique, A. (2000) Conditional skewness in asset pricing. The
Journal of Finance 34: 465–477.
Hendershott, T., Livdan, D. and Schurhoff, N. (2015) Are institutions informed about
news?. Journal of Financial Economics (forthcoming).
29
Heston, S. L. and Sinha, N. R. (2014) News versus sentiment: Comparing textual
processing approaches for predicting stock returns. Robert H. Smith School Research
Paper.
Hong, H. and Sraer, D. (2015) Speculative betas. NBER Working Paper No. 18548.
Ince, O. S. and Porter, R. B. (2006) individual equity return data from thomson
datastream: Handle with care!. Journal of Financial Research 29: 463–79.
Kumar, A. (2009) Who gambles in the stock market?. The Journal of Finance
64: 1889–1933.
Lesmond, D. A., Ogden, J. P. and Trzcinka, C. A. (1999) A new estimate of transac-
tion costs. Review of Financial Studies 12(5): 1113–41.
Lewellen, J. (2011) Institutional investors and the limits of arbitrage. Journal of
Financial Economics 102: 62–80.
Lewellen, J., Nagel, S. and Shanken, J. (2010) A skeptical appraisal of asset pricing
tests. Journal of Financial Economics 96: 175 – 194.
Lucca, D. O. and Moench, E. (2014) The pre-FOMC announcement drift. The Journal
of Finance.
Mitton, T. and Vorkink, K. (2007) Equilibrium underdiversification and the prefer-
ence for skewness. Review of Financial Studies 20(4): 1255–88.
Nagel, S. (2005) Short sales, institutional investors and the cross-section of stock
returns. Journal of Financial Economics 78: 277–309.
Newey, W. K. and West, K. D. (1987) A simple, positive semi-definite, heteroskedas-
ticity and autocorrelation consistent covariance matrix. Econometrica 55: 703–708.
Savor, P. and Wilson, M. (2014) Asset pricing: A tale of two days. Journal of
Financial Economics 113: 171–201.
30
Sias, R. W. and Starks, L. T. (1997) Institutions and individuals at the turn-of-the-
year. The Journal of Finance 52(4): 1543–62.
Sikes, S. A. (2014) The turn-of-the-year efect and tax-loss-selling by institutional
investors. Journal of Accounting and Economics 57: 22–42.
31
Figure 1: Capital asset pricing model on high and low institutional tradingAverage excess returns for four beta-sorted and five industry portfolios on high and low institutionaltrading (IT) days. This figure plots average daily excess returns (in percentages) against marketbetas for four beta-sorted (value-weighted) portfolios, nine value-weighted size-BM portfolios, andfive value-weighted industry portfolios. Day t has High-IT volume (scaled by total market volume)when the IT volume at t is greater than the average IT volume over the past quarter. The firstgraph shows the relation between beta and average return on High-IT days while the second graphis on days with Low-IT. The implied ordinary least squares estimates of the securities market linefor each type of day are also plotted. The sample period is between 1997 and 2011. Individualstocks’ betas are corrected for potential asynchronous trading using the Dimson method. Thisfigure shows that the relation between beta and average return is much higher and more positiveon High-IT days (first graph) whereas, on Low-IT days (second graph), the relation is significantlynegative.
32
Figure 2: Capital asset pricing model on non-turn-of-month daysAverage excess returns for four beta-sorted, nine size-BM, and five industry portfolios on High- andLow-IT days accounting for the turn-of-month effect. This figure plots average daily excess returns(in percentages) against market betas for four beta-sorted portfolios, nine size-BM portfolios,and five industry portfolios. The first graph shows days with High-IT total volume but not theturn-of-month day. The second graph presents the relationship between excess returns and betason days with Low-IT total volume and not the turn-of-month day. The implied ordinary leastsquares estimates of the securities market line for each type of day are also plotted. The sampleperiod is between 1997 and 2011. Betas to form portfolios are corrected for potential asynchronoustrading using the Dimson method. This figure shows that, after excluding turn-of-month days, thedifference in the risk premium between High- and Low-IT days remains robust.
33
Figure 3: Capital asset pricing model on non-January daysAverage excess returns for four beta-sorted, nine size-BM, and five industry portfolios on High-and Low-IT days accounting for the January effect. This figure plots average daily excess returns(in percentages) against market betas for four beta-sorted portfolios, nine size-BM portfolios, andfive industry portfolios. The first graph shows days with High-IT volume but not in January. Thesecond graph presents the relationship between excess returns and betas on days with Low-ITvolume and not January days. The implied ordinary least squares estimates of the securities marketline for each type of day are also plotted. The sample period is between 1997 and 2011. Betas toform portfolios are corrected for potential asynchronous trading using the Dimson method. Thisfigure shows that, after excluding January, the difference in the risk premium between High- andLow-IT days remains robust.
34
Figure 4: Capital asset pricing model on announcement and non-announcement daysAverage excess returns for four beta-sorted, nine size-BM, and five industry portfolios on ECBannouncement versus non-announcement days. This figure plots average daily excess returns (inpercentages) against market betas for four beta-sorted portfolios, nine size-BM portfolios, and fiveindustry portfolios. The first graph shows the relation between beta and average return on dayswith ECB monetary policy decision. The second graph presents similar line on non-announcementdays. The implied ordinary least squares estimates of the security market line for each type of dayare also plotted. The sample period is between 1999 and 2011 (the ECB was formally established in1999). Betas to form portfolios are corrected for potential asynchronous trading using the Dimsonmethod. This figure shows that, consistent with Savor and Wilson (2014), the relation betweenbeta and average return is much more positive on monetary announcement days than on normal days.
35
Figure 5: Capital asset pricing model – Non-announcement daysAverage excess returns for four beta-sorted, nine size-BM, and five industry portfolios on High-and Low-IT days accounting for the ECB announcement effect. This figure plots average dailyexcess returns (in percentages) against market betas for four beta-sorted portfolios, nine size-BMportfolios, and five industry portfolios. The first graph shows days with High-IT volume but noECB monetary policy decision. The second graph presents the relationship between excess returnsand betas on days with Low-IT volume and no announcement. The implied ordinary least squaresestimates of the securities market line for each type of day are also plotted. The sample period isbetween 1999 and 2011 (the ECB was formally established in 1999). Betas to form portfolios arecorrected for potential asynchronous trading using the Dimson method. The purpose of this figureis to show that, even when the day has no macroeconomic announcement, the difference in the riskpremium between High- and Low-IT days remains robust. Consequently, the effect of institutionaltrading is not a manifestation of announcement effects.
36
Figure 6: Capital asset pricing model on days with high or low degree of investor disagreementAverage excess returns for four beta-sorted, nine size-BM, and five industry portfolios on high andlow investor-disagreement days accounting for the ECB announcement effect. This figure plotsaverage daily excess returns (in percentages) against market betas for four beta-sorted portfolios,nine size-BM portfolios, and five industry portfolios. A day t has strong investor disagreement ifthe cross-sectional standard deviation of news tone on day t is higher than its average over thepast month. Dzielinski and Hasseltoft (2014) show that high news dispersion is associated withstrong investor disagreement. The first graph shows days with weak investor disagreement. Thesecond graph presents the relation between excess returns and betas on days with strong investordisagreement. The implied ordinary least squares estimates of the securities market line for eachtype of day are also plotted. The sample period is daily frequency between 2003 and 2011 (the newsdataset starts in 2003). Betas to form portfolios are corrected for potential asynchronous tradingusing the Dimson method. This figure provides evidence for Hong and Sraer’s (2014) theory thatthe relation between beta and average return is negative (positive) when the aggregate investordisagreement is strong (weak).
37
Figure 7: Capital asset pricing model on days with high or low degree of coskewnessAverage excess returns for four beta-sorted, nine size-BM, and five industry portfolios on high andlow investor-disagreement days accounting for the coskewness effect. This figure plots average dailyexcess returns (in percentages) against market betas for four beta-sorted portfolios, nine size-BMportfolios, and five industry portfolios. A stock coskewness is estimated using Regression (4) andthen the overall coskewness on day t is the average of individual stocks’ coskewness. Day t exhibitshigh coskewness if the overall coskewness is higher than its average over the past quarter. The firstgraph shows days with low coskewness. The second graph presents the relation between excessreturns and betas on days with high coskewness. The implied ordinary least squares estimatesof the securities market line for each type of day are also plotted. Betas to form portfolios arecorrected for potential asynchronous trading using the Dimson method. This figure provides evi-dence for the hypothesis that when coskewness is low (high) the SML is upward (downward) sloping.
38
Table 1: Summary statistics for Finnish marketThis table reports summary statistics for the Helsinki stock exchange between 2 January 1996 and 30December 2011. “Mean Firms” are the average number of firms. “Mean ME” is the average marketcapitalization in millions of euros (price times number of shares outstanding). “Mean Fractionof Total IT Volume” is the average fraction of trading volume by financial institutions over thetotal market volume. “Mean Fraction of Sell Volume” is the average fraction of sell volume byfinancial institutions over the total market volume. “Number of Accounts” is the count of uniqueaccounts held by financial institutions. Panel B reports average returns on nine value-weightedsize-BM sorted portfolios. Panel C shows the average return on four value-weighted beta-sortedportfolios. Panel D presents the average return on five industry portfolios using Fama-French’sclassifications (downloaded from Ken French’s website), where Industry 1 is consumer-related suchas consumer durables, non-durables, wholesale and retail; Industry 2 is manufacturing; Industry3 is high-tech firms; Industry 4 is health-related services; and Industry 5 is others (e.g., mines,construction, transportation, entertainment, and finance). t-statistics computed using Newey-Weststandard errors with five lags are reported in parentheses.
Panel A: Summary statistics for the Finnish marketYear Mean Firms Mean ME Mean Fraction of Mean Fraction of Number of
(’000,000) Total IT Volume Sell Volume Accounts
1995-1999 79 389.868 0.196 0.123 5632000-2003 138 1767.843 0.169 0.104 6432004-2007 141 1335.035 0.232 0.128 7222008-2011 140 1095.511 0.379 0.251 656
Panel B: Summary statistics for nine VW size-BM sorted portfoliosLow M High tLow tM tHigh
Small 0.0010 0.0012 0.0006 (4.13) (4.43) (1.85)M 0.0006 0.0005 0.0003 (2.63) (2.73) (1.15)Big 0.0006 0.0006 0.0005 (2.15) (2.55) (1.23)
Panel C: Summary statistics for four VW beta-sorted portfoliosBeta Low 2 3 High
0.0005 0.0006 0.0005 0.0003(1.88) (3.74) (2.43) (0.90)
Panel D: Summary statistics for five VW industry portfoliosIndustry 1 2 3 4 5
0.0004 0.0006 0.0007 0.0006 0.0008(1.79) (2.43) (1.49) (1.35) (3.35)
39
Table 2: Daily excess returns on days of high and low institutional tradingThis table reports estimates from Fama-MacBeth regressions of daily excess returns on betas for various test portfolios. Estimates arecomputed for days with high institutional trading (High-IT days or High) and other days (Low-IT days or Low). Day t has High-IT volume(scaled by total market volume) when the IT volume at t is greater than the average IT volume over the past quarter. The difference inthe coefficient between high- and Low-IT days is reported in the last row of each panel. There are 1598 days with High-IT volume. Theright-hand side panel reports estimates from pooled regression of excess returns on betas, High-IT day dummy, and interaction between betaand High-IT (High*Beta). Panel A shows results for four beta-sorted (value-weighted) portfolios. Individual stocks’ betas are adjusted forpotential asynchronous trading effects using Dimson’s method. Panel B presents results for four beta-sorted and five value-weighted industryportfolios. Panel C reports results for nine value-weighted size-BM portfolios, four beta-sorted portfolios, and five industry portfolios. Thesample period is between 1997 and 2011 (we lost one year to form portfolios). t-statistics, which are computed using Newey-West standarderrors with five lags, are reported in parentheses. For pooled regressions, standard errors are corrected for clustering by trading day. Thistable shows that the difference in the coefficient on beta between High- and Low-IT days is always positive and statistically significant.
Fama-MacBeth regression Pooled regression
Type Intercept Beta Avg. R2 Intercept Beta High High*Beta Avg. R2
of day
Panel A: four beta-sorted portfoliosHigh 0.00013 0.00073 0.24 0.00068 −0.00186 0.000258 0.00283 0.53
(0.77) (2.61) (1.62) (−4.54) (0.42) (4.69)Low 0.00011 −0.00074 0.31
(0.54) (−2.00)High− Low 0.00002 0.00147
(0.08) (3.18)
Panel B: four beta-sorted and five industry portfoliosHigh 0.00011 0.00076 0.16 0.00095 −0.00120 −0.00070 0.00189 0.39
(0.65) (2.83) (3.19) −3.67 −1.51 (3.72)Low 0.00016 −0.00080 0.20
(0.50) (−1.80)High− Low −0.00005 0.00156
(−0.14) 3.20
Panel C: four beta-sorted, nine size & BM portfolios, and five industry portfoliosHigh 0.00019 0.00070 0.11 0.00060 −0.00173 0.00022 0.00296 0.37
(1.41) (2.81) (1.88) (−7.19) (0.46) (8.40)Low 0.00017 −0.00082 0.14
(0.72) (−2.26)High− Low 0.00003 0.00152
(0.12) 3.50
40
Table 3: Daily excess returns for individual stocks on days of high and low institutional tradingThis table reports estimates from Fama-MacBeth regressions of daily excess returns on betas for various test portfolios. Estimates arecomputed for days with high domestic institutional trading (High-IT days or High) and other days (Low-IT days or Low). Day t has High-ITvolume (scaled by total market volume) when the IT volume at t is greater than the average IT volume over the past quarter. The differencein the coefficient between High- and Low-IT days is reported in the last row of each panel. There are 1598 days with High-IT volume. Theright-hand side panel reports estimates from pooled regression of excess returns on betas, High-IT day dummy, and interaction between betaand High-IT days (High*Beta). Panels A and B show results for beta as the right-hand side variable in Fama-MacBeth regressions andpooled regressions, respectively. Panels C and D are similar to the first two panels, but include log(size), BM, and past one-year returns asadditional controls. Panels E and F include liquidity rank (LIQ) as an additional control. This liquidity measure is constructed based on theoccurrence of zero returns over the past year. The sample period is between 1997 and 2011 (we lost one year to form portfolios). Panels Gand H repeat the regressions in Panels E and F, but excluding Nokia, which is the largest firm on the Finnish market. t-statistics, which arecomputed using Newey-West standard errors with five lags, are reported in parentheses. For pooled regressions, standard errors are correctedfor clustering by trading day. This table shows that, for individual stocks, the coefficient on beta on High-IT days is higher than on Low-ITdays.
Panel A: beta only (Fama-MacBeth regressions)Type of day Beta Avg. R2
High 0.00039 0.02(1.80)
Low −0.00078 0.02(−3.20)
High− Low 0.00117(3.52)
Panel B: beta and interaction with High-IT (Pooled regressions)Beta High High*Beta Avg. R2
−0.00123 0.00058 0.00242 0.09(−6.59) (1.52) (8.68)
Panel C: size, BM, and past returns as controls (Fama-MacBeth regressions)Type of day Beta Size BM Past one-year Avg. R2
High 0.00044 −0.00008 −0.00009 −0.02032 0.04(1.81) (−2.59) (−3.13) (−1.18)
Low −0.00077 −0.00007 −0.00007 0.00026 0.05(−2.85) (−2.29) (−2.19) (1.43)
High− Low 0.001205 −0.00001 −0.00002 −0.020583.214 (−0.02) (−0.40) (−1.76)
41
Table 3 continued.
Panel D: size, BM, and past returns as controls (Pooled regressions)Beta Size BM Past one-year High High*Beta Avg. R2
−0.00165 −0.00016 0.00002 0.00015 0.00020 0.00280 0.09(−8.01) (−4.74) (1.26) (0.70) (0.50) (9.51)
Panel E: size, BM, past returns, and liquidity as controls (Fama-MacBeth regressions)Type of day Beta Size BM Past one-year Turn Avg. R2
High 0.00041 −0.00008 −0.00009 −0.01801 0.00413 0.04(1.88) (−2.85) (−3.15) (−1.05) (0.76)
Low −0.00077 −0.00007 −0.00007 0.00028 −0.00001 0.06(−2.82) (−2.21) (−2.19) (1.51) −0.20
High− Low 0.00118 −0.00001 −0.00002 −0.01829 0.00415(3.14) (−0.28) (−0.38) (−1.73) (0.58)
Panel F: size, BM, past returns, and liquidity as controls (Pooled regressions)Beta Size BM Past one-year Turn High High*Beta Avg. R2
−0.00168 −0.00017 0.00002 0.00014 0.00007 0.00019 0.00281 0.09(−8.03) (−4.68) (1.24) (0.68) (0.90) (0.49) (9.52)
Panel G: size, BM, past returns, and liquidity as controls (Fama-MacBeth regressions) – excluding NokiaType of day Beta Size BM Past one-year Turn Avg. R2
High 0.00039 −0.00009 −0.00010 −0.01746 0.00419 0.04(1.78) (−2.96) (−3.26) (−1.02) (0.77)
Low −0.00078 −0.00007 −0.00008 0.00028 −0.00001 0.06(−2.79) (−2.24) (−2.32) (1.53) (−0.08)
High− Low 0.001171 −0.00002 −0.00002 −0.01773 0.00419(3.10) (−0.32) (−0.31) (−1.69) (0.50)
Panel H: size, BM, past returns, and liquidity as controls (Pooled regressions) – excluding NokiaBeta Size BM Past one-year Turn High High*Beta Avg. R2
−0.00168 −0.00018 0.00001 0.00014 0.00008 0.00020 0.00280 0.09(−7.84) (−4.68) (0.33) (0.69) (0.94) (0.50) (9.26)
42
Table 4: Daily excess returns on High- and Low-IT volume: subsample analysisThis table reports estimates from Fama-MacBeth regressions of daily excess returns on betas for various test portfolios in two subsamples:before and after 2008. Estimates are computed for days with high domestic institutional trading (High-IT days or High) and other days(Low-IT days or Low). Day t has High-IT volume (scaled by total market volume) when the IT volume at t is greater than the average ITvolume over the past quarter. The difference in the coefficient between high- and Low-IT days is reported in the last row of each panel. Theright-hand side panel reports estimates from pooled regression of excess returns on betas, High-IT day dummy, and interaction between betaand High-IT (High*Beta). Test assets are four beta-sorted portfolios, nine value-weighted size-BM portfolios, and five industry portfolios.The sample period is between 1997 and 2011. Panel A uses sample before 2008 while the data in Panel B are after 2008. t-statistics, which arecomputed using Newey-West standard errors with five lags, are reported in parentheses. For pooled regressions, standard errors are correctedfor clustering by trading day. Betas are estimated using all returns data over the past one year. This table shows that the difference in thecoefficient on beta between High- and Low-IT days is always positive and statistically significant.
Fama-MacBeth regression Pooled regression
Type Intercept Beta Avg. R2 Intercept Beta High High*Beta Avg. R2
of day
Panel A: four beta-sorted, nine size-BM portfolios, and five industry portfolios, pre-2008High 0.00053 0.00143 0.26 0.00089 −0.00142 −0.0001 0.00261 0.29
(1.41) (2.09) (2.50) (−4.85) (−0.01) (6.13)Low 0.00030 −0.00064 0.141
(0.99) (−1.36)High− Low 0.00024 0.00207
(0.55) (2.67)
Panel B: four beta-sorted, nine size-BM portfolios, and five industry portfolios, post-2008High 0.00018 0.00186 0.25 0.00016 −0.00249 0.00011 0.00424 0.29
(0.317) (2.06) (0.26) (−6.31) (0.12) (7.05)Low −0.00013 −0.00124 0.15
(−0.38) (−2.48)High− Low 0.00031 0.00310
(0.51) (2.80)
43
Table 5: U.S. markets’ results: Excess returns on quarters of high and low institutional holdingsThis table reports estimates from Fama-MacBeth regressions of daily excess returns on betas forten beta-sorted portfolios using U.S. CRSP data. Estimates are computed for quarters with highinstitutional trading (High-IT or High) and other quarters (Low-IT or Low). Using data fromThomson Reuters SEC 13F holdings data, we compute aggregate change in institutional holdingsscaled by CRSP’s total market trading volume per quarter. Quarter t has Low-IT volume (scaledby total market volume) when the aggregate change in quarterly holdings at t is greater thanthe average IT volume over the past year (four quarters). Betas are estimated by running 60-month rolling window of excess returns on excess market returns (with Dimson’s adjustment). Thedifference in the coefficient between high- and Low-IT quarters is reported in the last row of eachpanel. Panel B reports estimates from pooled regression of excess returns on betas, High-IT quarterdummy, and interaction between beta and High-IT (High*Beta). The sample period is betweenQ1-1984 and Q1-2014. t-statistics, which are computed using Newey-West standard errors with fivelags, are reported in parentheses. For pooled regressions, standard errors are corrected for clusteringby month. This table shows that the difference in the coefficient on beta between High- and Low-ITquarters is positive.
Panel A: Fama-MacBeth regressionType Intercept Beta Avg. R2
of day
High 0.00040 0.002791 0.25(0.20) 1.040
Low 0.006491 −0.003520 0.27(3.56) −1.19
High− Low −0.00609 0.00631(−2.55) (1.80)
Panel B: Pooled regressionIntercept Beta High High*Beta Avg. R2
0.00979 −0.00368 −0.00491 0.00583 0.74(2.31) (−2.23) (−0.81) (2.45)
44
Table 6: Daily excess returns on various day typesThis table reports estimates from Fama-MacBeth regressions of daily excess returns on betas for various test portfolios on High- and Low-ITdays accounting for the ECB announcement effect. Day t has High-IT volume when the IT volume (scaled by market volume) on day t isgreater than its average over the past quarter. “Announcement days” are days when the ECB announced monetary policy decisions. “Yes”represents days that have the characteristics displayed in the heading whereas “No” represents days that do not have any or both of thecharacteristics displayed in the heading. For example, in panel A1, a day is a “Yes” day when it has both High-IT and announcement;otherwise, it is a “No” day. The difference in the coefficient on beta between Yes and No days is reported in the last row of each panel.The right-hand side panel reports estimates from pooled regression of excess returns on betas, Yes day dummy, and interaction between betaand Yes (Yes*Beta). Test assets are four beta-sorted portfolios, nine size-BM portfolios, and five industry portfolios. Panel A reports resultsfor IT total volume. Panels B and C break the analysis into buying and selling volume, respectively. t-statistics, which are computed usingNewey-West standard errors with five lags, are reported in parentheses. For pooled regressions, standard errors are corrected for clusteringby trading day. There are 85 days with High-IT and announcements and 92 days with Low-IT and announcement, 1,331 days with High-ITand no announcements, and 1,673 days with Low-IT and no announcements.This table shows that the IT effect is still strong on non-announcement days, suggesting that the two effects are distinct from each other.
Fama-MacBeth regression Pooled regression
Type Intercept Beta Avg. R2 Intercept Beta Yes Yes*Beta Avg. R2
of day
Panel A: IT total volumePanel A1: High-IT and announcement daysY es −0.00123 0.00555 0.31 0.00073 −0.00064 −0.00139 0.00558 0.36
(−0.73) (2.45) (3.05) (−3.56) (−0.95) (5.10)No 0.00055 −0.00055 0.26
(1.90) (−1.15)Y es−No −0.00178 0.00607
(−1.20) (2.19)Panel A2: Low-IT and announcement daysY es −0.00150 0.00287 0.31 0.00077 −0.00059 −0.00277 0.00365 0.37
(−0.94) (0.73) (3.21) (−3.29) (−1.96) (3.39)No 0.00056 −0.00046 0.26
(1.93) (−1.03)Y es−No −0.00206 0.00333
(−1.44) (1.25)Panel A3: High-IT and non-announcement daysY es 0.00063 0.00106 0.26 0.00058 −0.00151 0.00031 0.00236 0.36
(2.04) (1.84) (1.88) (−6.35) (0.65) (6.55)No 0.00040 −0.00139 0.27
(0.93) (−2.10)Y es−No 0.00023 0.00245
(0.47) (2.71)
45
Table 6 continued.
Fama-MacBeth regression Pooled regression
Type Intercept Beta Avg. R2 Intercept Beta Yes Yes*Beta Avg. R2
of day
Panel B: IT buying volume
Panel B1: High-IT buying and announcement daysY es −0.00159 0.008189 0.33 0.00076 −0.00072 −0.00265 0.00874 0.37
(−1.20) (2.25) (3.18) (−3.97) (−1.78) (7.78)No 0.00056 −0.000589 0.26
(1.93) (−1.32)Y es−No −0.00215 0.00878
(−1.43) (3.12)
Panel B2: Low-IT buying and announcement daysY es −0.00117 0.00067 0.30 0.00074 −0.00052 −0.00165 0.00096 0.36
(−0.75) (0.20) (3.08) (−2.87) (−1.19) (0.92)No 0.00055 −0.00040 0.26
(1.90) (−0.87)Y es−No −0.00172 0.00107
(−1.23) (0.407)
Panel B3: High-IT buying and non-announcement daysY es 0.00028 0.001250 0.26 0.00090 −0.00170 −0.00041 0.00276 0.37
(0.84) (2.11) (2.85) (−7.12) (−0.87) (7.66)No 0.00066 −0.00157 0.26
(1.64) (−2.48)Y es−No −0.00038 0.00282
(−0.78) (3.13)
46
Table 6 continued.
Fama-MacBeth regression Pooled regression
Type Intercept Beta Avg. R2 Intercept Beta Yes Yes*Beta Avg. R2
of day
Panel C: IT selling volume
Panel C1: High-IT selling and announcement daysY es −0.00113 0.00234 0.31 0.00072 −0.00055 −0.00110 0.00235 0.36
(−0.62) (0.90) (3.00) (−3.05) (−0.73) (2.08)No 0.000541 −0.00043 0.26
(1.90) (−0.95)Y es−No −0.00167 0.00278
(−1.09) (0.98)
Panel C2: Low-IT selling and announcement daysY es −0.00157 0.00565 0.32 0.00078 −0.00069 −0.00299 0.00651 0.36
(−1.13) (1.33) (3.26) (−3.80) (−2.17) (6.23)No 0.00056 −0.00055 0.26
(1.93) (−1.24)Y es−No −0.00213 0.00620
(−1.54) (2.39)
Panel C3: High-IT selling and non-announcement daysY es 0.00090 0.000309 0.261 0.00043 −0.00106 0.00062 0.00132 0.36
(2.53) (0.53) (1.40) (−4.47) (1.30) (3.67)No 0.00021 −0.000854 0.262
(0.56) (−1.44)Y es−No 0.000697 0.00116
(1.45) (1.29)
47
Table 7: Daily excess returns on High-IT buying and High-IT selling daysThis table reports estimates from Fama-MacBeth regressions of daily excess returns on betas for various test portfolios on high- and low-ITbuying and selling days (rather than total IT volume as in previous tables). Estimates are computed for days with high institutional trading(High-IT days or High) and other days (Low-IT days or Low). Day t is a High-IT day when IT buy (sell) volume (scaled by total marketvolume) at t is greater than its average over the past quarter. The difference in the coefficient between High- and Low-IT days is reported inthe last row of each panel. The right-hand side panel reports estimates from pooled regression of excess returns on betas, High-IT day dummy,and interaction between beta and High-IT (High*Beta). Panel A shows the estimate for buying volume (scaled by total market volume)while Panel B shows the results for selling volume (scaled by total market volume). Test assets are four beta-sorted portfolios, nine size-BMportfolios, and five industry portfolios. t-statistics, which are computed using Newey-West standard errors with five lags, are reported inparentheses. For pooled regressions, standard errors are corrected for clustering trading day. Betas are estimated using all returns data overthe past one year. This table shows that the difference in the coefficient on beta between High- and Low-IT days is driven by the buying sideof institutional trading.
Fama-MacBeth regression Pooled regression
Type Intercept Beta Avg. R2 Intercept Beta High High*Beta Avg. R2
of day
Panel A: High- versus Low-IT buying daysHigh 0.000077 0.00080 0.12 0.00095 −0.00200 −0.00056 0.00353 0.37
(0.54) (3.03) (2.99) (−8.32) (−1.19) (10.02)Low 0.00029 −0.000925 0.14
(1.32) (−2.72)High− Low −0.00021 0.00173
(−0.90) (3.99)
Panel B: High- versus Low-IT selling daysHigh 0.00036 0.000129 0.12 0.00046 −0.00084 0.00046 0.00104 0.37
(2.22) 0.500 (1.44) (−3.50) (0.97) (2.94)Low 0.00001 −0.00025 0.14
(0.05) (−0.74)High− Low 0.000349 0.00038
(1.48) (0.89)
48
Table 8: Beta on High- and Low-IT daysThis table reports the average of, IT imbalance and beta on High- and Low-IT days. Stocks are splitinto two groups on the basis of their institutional trade imbalance where the first group (Buy) haspositive average imbalance and the second group (Sell) has negative average imbalance. “IT imb”is the average imbalance (imb=(buy volume − sell volume)/total market volume) of institutions.“Beta” is the average beta across stocks in each buy or sell portfolios. High − Low in the bottompanel reports the difference in beta and news dispersion between High- and Low-IT days. The lastcolumn reports the difference in the characteristic between Buy and Sell. # denote the differencethat is statistically significant at either the 5% or 1% level. Standard errors are computed using theNewey-West method with five lags.
Buy − SellIT day Buy/Sell IT Imb Beta Beta
Low Sell −0.0504 0.6783 0.0136#
Buy 0.0408 0.6919
High Sell −0.0605 0.6453 0.0114#
Buy 0.0514 0.6567
High−Low Sell −0.0101# −0.0330#
Buy 0.0106# −0.0352#
49
Table 9: Institutional trading and investor disagreementThis table reports estimates from Fama-MacBeth regressions of daily excess returns on betas for various test portfolios on High- and Low-ITdays accounting for the market disagreement. Day t has High-IT volume when the IT volume (scaled by market volume) on day t is greaterthan its average over the past quarter. “Strong disagreement days” are days when news dispersion is higher than its average over the pastquarter. “Yes” represents days that have the characteristics displayed in the heading whereas “No” represents days that do not have any orboth of the characteristics displayed in the heading. For example, in panel A1, a day is a “Yes” day when it has both High-IT and weakdisagreement; otherwise, it is a “No” day. The difference in the coefficient on beta between Yes and No days is reported in the last row ofeach panel. The right-hand side panel reports estimates from pooled regression of excess returns on betas, Yes day dummy, and interactionbetween beta and Yes (Yes*Beta). Test assets are four beta-sorted portfolios, nine size-BM portfolios, and five industry portfolios. t-statistics,which are computed using Newey-West standard errors with five lags, are reported in parentheses. For pooled regressions, standard errorsare corrected for clustering by trading day. There are 444 days with High-IT and weak disagreement, 534 days with Low-IT and weakdisagreement, and 555 days with High-IT and strong disagreement.
Fama-MacBeth regression Pooled regression
Type Intercept Beta Avg. R2 Intercept Beta Yes Yes*Beta Avg. R2
of day
Panel A: High-IT and weak disagreement daysY es 0.000890 0.001889 0.24 0.00075 −0.00135 0.00017 0.00317 0.45
(1.98) (2.59) (2.61) (−6.99) (0.27) (7.41)No 0.000549 −0.001113 0.24
(1.41) (−2.07)Y es−No 0.000340 0.00300
(0.50) (2.74)Panel B: Low-IT and weak disagreement daysY es 0.000646 −0.001305 0.24 0.00088 −0.00058 −0.00046 -0.00052 0.45
0.779 −1.13 (3.00) (−2.99) (−0.76) (-1.26)No 0.000608 −0.000270 0.24
2.044 −0.61Y es−No 0.000037 -0.001034
(0.60) (-1.00)Panel C: High-IT and strong disagreement daysY es 0.000680 0.000716 0.24 0.00071 −0.00107 0.00030 0.00138 0.45
(1.39) (1.03) (2.38) (−5.31) (0.51) (3.53)No 0.000596 −0.000924 0.24
(1.47) (−1.68)Y es−No 0.000083 0.001640
(0.13) (1.67)
50
Table 10: Institutional trading and coskewnessThis table reports estimates from Fama-MacBeth regressions of daily excess returns on betas for various test portfolios on High- and Low-ITdays accounting for the coskewness effect. Day t has High-IT volume when the IT volume (scaled by market volume) on day t is greater thanits average over the past quarter. A stock coskewness is estimated using Regression (4) and then the overall coskewness on day t is the averageof individual stocks’ coskewness. Day t exhibits high coskewness if the overall coskewness is higher than its average over the past quarter.“Yes” represents days that have the characteristics displayed in the heading whereas “No” represents days that do not have any or both ofthe characteristics displayed in the heading. For example, in panel A1, a day is a “Yes” day when it has both High-IT and low coskewness;otherwise, it is a “No” day. The difference in the coefficient on beta between Yes and No days is reported in the last row of each panel.The right-hand side panel reports estimates from pooled regression of excess returns on betas, Yes day dummy, and interaction between betaand Yes (Yes*Beta). Test assets are four beta-sorted portfolios, nine size-BM portfolios, and five industry portfolios. t-statistics, which arecomputed using Newey-West standard errors with five lags, are reported in parentheses. For pooled regressions, standard errors are correctedfor clustering by trading day. There are 682 days with High-IT and low coskewness, 811 days with Low-IT and low coskewness, and 852 dayswith High-IT and high coskewness.
Fama-MacBeth regression Pooled regression
Type Intercept Beta Avg. R2 Intercept Beta Yes Yes*Beta Avg. R2
of day
Panel A: High-IT and low coskewness daysY es 0.000321 0.001544 0.25 0.00063 −0.00071 0.00015 0.00182 0.37
(0.75) (2.06) (2.42) (−3.63) (0.26) (4.26)No 0.000549 −0.001113 0.26
(1.30) (−1.22)Y es−No −0.000109 0.002150
(−0.18) (2.00)Panel B: Low-IT and low coskewness daysY es −0.000173 0.000165 0.27 0.00083 −0.00042 −0.00080 0.00039 0.37
(−0.25) (0.16) (3.11) (−2.12) (−1.47) (0.95)No 0.000592 −0.000281 0.25
(2.05) (−0.62)Y es−No −0.000765 0.000447
(−1.38) (0.44)Panel C: High-IT and high coskewness daysY es 0.000545 0.001531 0.25 0.00059 −0.00092 0.00015 0.00236 0.37
(1.232) (1.97) (2.22) −4.56) 0.29) (5.83)No 0.000364 −0.000744 0.26
(1.059) (−1.41)Y es−No 0.000181 0.002275
(0.333) (2.28)
51
Appendices
A. Capital asset pricing model between High- and Low-IT days (one-yearwindow)
In the main text, we report results from High- and How-IT days that are definedon the basis of the past quarter. In this appendix, we use the window of one year asthe basis to determine a High- or Low-IT day. Specifically, a day has High-IT whenthe fraction of institutional trading volume over the market volume is greater thanits average over the past year.
Figure B.1 plots the security market line for four beta-sorted, nine size-BM, andfive industry portfolios. The relation between beta and average return is still muchhigher and positive on High-IT days. An increase in beta by 0.5 is associated with astatistically significant increase in daily excess return of 9.7bps per day. In contrast,this relation is significantly negative on Low-IT days (second graph). An increasein beta by 0.5 leads to a reduction in average excess return of 3.1bps per day (t-statistic = -3.1). We thus conclude that our results are not sensitive to the choiceof window length to define High- or Low-IT days. In untabulated results, we alsorepeat the estimation for the level institutional volume (without scaling by marketvolume). The difference in implied market risk premium between High- and Low-ITdays remains significant.
B. High-IT vs. FOMC announcements
This section examines whether the IT effect in Finland could be driven by days ofscheduled FOMC announcements in the U.S.. First, Figure B.2 establishes the effectof FOMC announcements on the relation between beta and returns in the Finnishmarket. On days when the FOMC announces interest rate, the SML is upwardsloping whereas on non-announcement days, the SML is downward sloping. We thentest whether the IT effect in Finland could be driven by FOMC announcement days.Replicating the methodology in the main text for Figure 5, we examine the SML ondays of High-IT and non-FOMC announcement. We find that the High-IT effect inFinland is still strong even though the day does not have an announcement. Theseresults confirm that our IT effects are distinct from the announcement effect.
52
Figure B.1: Capital asset pricing model on High- and Low-IT volume daysAverage excess returns for four beta-sorted, nine size-BM, and five industry portfolios on Highand Low-IT volume. This figure plots average daily excess returns (in percentages) against marketbetas for three beta-sorted portfolios, nine size-BM portfolios, and five industry portfolios. Thefirst graph shows days with high-IT volume (days with the fraction of IT volume greater than theaverage past one-year volume). The second graph presents the relationship between returns andbetas on low-IT volume days. The implied ordinary least squares estimates of the securities marketline for each type of day are also plotted. The sample period is between 1996 and 2011 (we lost oneyear to form portfolios). Betas to form portfolios are corrected for potential asynchronous tradingusing the Dimson method. This figure shows that the relation between beta and average return ismuch higher and more positive on High-IT days (first graph) than that on Low-IT days (secondgraph).
53
Figure B.2: Capital asset pricing model on FOMC announcement and non-announcement daysAverage excess returns for four beta-sorted, nine size-BM, and five industry portfolios on FOMCscheduled announcement versus non-announcement days. This figure plots average daily excessreturns (in percentages) against market betas for four beta-sorted portfolios, nine size-BMportfolios, and five industry portfolios. The first graph shows the relation between beta andaverage return on days with FOMC monetary policy decision. The second graph presents similarline on non-announcement days. The implied ordinary least squares estimates of the securitymarket line for each type of day are also plotted. The sample period is between 1997 and 2011.Betas to form portfolios are corrected for potential asynchronous trading using the Dimsonmethod. This figure shows that, consistent with Savor and Wilson (2014), the relation betweenbeta and average return is much more positive on monetary announcement days than on normal days.
54
Figure B.3: Capital asset pricing model – Non-FOMC-announcement daysAverage excess returns for four beta-sorted, nine size-BM, and five industry portfolios on High-and Low-IT days accounting for the FOMC announcement effect. This figure plots average dailyexcess returns (in percentages) against market betas for four beta-sorted portfolios, nine size-BMportfolios, and five industry portfolios. The first graph shows days with High-IT volume butno FOMC monetary policy decision. The second graph presents the relationship between excessreturns and betas on days with Low-IT volume and no announcement. The implied ordinary leastsquares estimates of the securities market line for each type of day are also plotted. The sampleperiod is between 1997 and 2011. Betas to form portfolios are corrected for potential asynchronoustrading using the Dimson method. The purpose of this figure is to show that, even when theday has no macroeconomic announcement, the difference in the risk premium between High- andLow-IT days remains robust. Consequently, the effect of institutional trading is not a manifestationof announcement effects.
55