Intra-day momentum∗

Oleg Komarov
Imperial College London†

First draft: January 25, 2017

This version: February 7, 2017

∗
I would like to thank Walter Distaso for helpful suggestions. All errors remain my own.
†
Imperial College Business School
South Kensington Campus, London SW7 2AZ (UK)
Email: [email protected]
 Intra-day momentum

Abstract

There is limited evidence of intraday predictability both in the cross-section of US

stock returns (see Heston et al., 2010) and in the time-series of the aggregate stock

market (see Gao et al., 2015). I find that statistical time-series predictability does not

imply economic profitability, whereas cross-sectional sorts on past performance see

stocks, which lost or won the most in the morning, earn in the last half-hour of trading

about 15.6 and 19.4% in annualized terms, and well above the rest of the cross-section.

The effect is fundamentally different from Heston et al. (2010) and is robust to

stock characteristics, the day-of-week effect, variations in the formation and holding

periods (afternoon), but exhibits some dependence on the sample period, suggesting

that specific market mechanisms or frictions play a relevant role on intraday price

formation.

Keywords: intraday predictability, cross-section, time-series, momentum

JEL Classification: G12, G17

ii
1 Introduction

Recent but limited evidence indicates the existence of intraday predictability both in the
cross-section of US stock returns and in the time-series of the aggregate stock market. I
reconsider the time-series dimension using all common US stocks from 1993 until 2010
and, building on this, I present the cross-sectional dimension with new and complimentary
evidence. Consistent with prior literature, I divide the trading session in 13 half-hours
and focus entirely on patterns in returns within the day. The impact of the overnight
return is considered separately. I find that statistical time-series predictability does not
imply economic profitability, whereas cross-sectional sorts on past performance see stocks,
which lost or won the most in the morning, earn above the rest of the cross-section in
the afternoon, and especially during the last half-hour of trading. The effect shows some
dependence on the interval of the holding period and the horizon of analysis, suggesting
that specific market mechanisms or frictions play a relevant role in price formation.
While the literature has examined intraday returns at the unconditional level, e.g.
providing a simple breakdown of averages by half-hours or hours, studies about intraday
predictability in returns are limited to two contributions.1 On one hand, Heston et al.
(2010) (henceforth HKS) apply the well-known “cross-sectional” momentum by Jegadeesh
and Titman (1993) to a high-frequency setting. The authors analyse return periodicity in
the cross-section and find patterns of continuation at half-hour intervals that are multiple
of one day and last up to 40 trading days. In other words, any half-hour today is e.g.
positively predicted by the same half-hour yesterday. On the other hand, Gao et al. (2015)
investigate the intraday time-series dimension of momentum. The effect, first documented
at the monthly frequency and across several asset classes by Moskowitz et al. (2012),
concerns individual securities and anticipates continuation of past performance. Gao et al.
(2015) concentrate on the time-series predictability of the S&P 500 ETF and find that the
1
A partial list documenting intraday average returns includes Wood et al. (1985), Harris (1986) and
Jain and Joh (1988).

1
first half-hour return predicts the return of the last half-hour.
My work differs from the previous two in several aspects: Gao et al. (2015) use an
aggregate index while I look at the whole cross-section which allows me to identify richer
evidence on time-series predictability. Additionally, I disentangle the overnight return from
the first half-hour and analyse its impact in a separate instance. With respect to HKS,
who have data from 2001 until 2005, my sample period is longer and covers the years from
1993 until 2010. Most important, my work is complimentary to HKS because I focus on
specific half-hours of a single day. Formation periods do not extend back than the previous
close and are generally confined to the open of the current trading session. In contrast,
HKS study the lagged impact of intraday returns up to the previous 40 days and report
an effect averaged by lag across all intervals. They do not focus on a specific interval, and
even their table IV, which looks at a time-of-day decomposition, reports the average return
to the winners-minus-losers strategy, a composite estimate, and takes into consideration
lags which are multiples of one day.
My contribution fills a gap in the existing literature by arguing that although time-
series predictability can exist statistically, there is no coherent economic evidence and
the observed results are the manifestation of a cross-sectional pattern, distinct from the
evidence in HKS. To establish the claim, I first look at which half-hours are statistically
predicted by past half-hours of the same day. I use the methodology of pooled regressions
with clustered standard errors by Moskowitz et al. (2012). In line with Gao et al. (2015), I
find that the last interval is predicted by the first and, in a more flexible specification, by
the larger interval from 9:30 until 12:00. I name this pattern last.2 Additionally, I observe
that the second half of the trading day, namely from 13:30 until 15:30 is negatively related
to the morning period from 9:30 to 13:00. I name this pattern afternoon.
Secondly, to gauge the periodicity shown by the time-series regressions, I sort stocks in
2
I will interchangeably refer to the set of formation and holding periods with pattern, configuration or
specification.

2
two groups, those with positive and those with negative returns in the morning, and form
equal-weighted portfolios. The return of these portfolios is calculated in the afternoon
according to the two specifications summarized above. While last anticipates continua-
tion and afternoon reversal, the performance of the two portfolios is inconsistent with both
specifications. The evolution of the strategies over-time clearly shows that stocks, indepen-
dent of their past performance, appreciate in the second half of the trading day. Moreover,
morning losers exhibit higher returns than their winning counterparts, hence suggesting
the existence of systematic differences in the cross-section. Additionally, a winners-minus-
losers strategy exhibits a clear structural break around the beginning of 2001. While the
date coincides with decimalization of stock quotes, I do not investigate a causal link within
the scope of this research.
Motivated by the inconsistencies in the behaviour of time-series strategies, I explore
differences in returns from a cross-sectional perspective for both the last and afternoon
configurations. I sort stocks into ten groups based on their performance during the intraday
formation period and form ten equal-weighted portfolios. The return to those portfolios
is calculated over the afternoon and last half-hour. I find a U-shaped pattern in average
returns where stocks that lost or gained the most in the morning, earn higher and positive
returns in the second half of the trading day, and especially during the last half-hour.
The U-shaped pattern in average returns to portfolios sorted on past performance is
robust to stock characteristics (I consider size, traded volume, illiquidity, tick size, volatility
and skewness), the day-of-week effect, variations to formation and holding periods, but
exhibits some dependence on the sample period. Namely, under afternoon, returns to cross-
sectional portfolios before the decimalization in 2001 are systematically different from the
returns after that date. On the other hand, the U-shaped pattern in returns under the last
configuration appears more resilient. Nonetheless, the general observation is that morning
winners are more profitable than morning losers before 2001 but these circumstances are
reversed afterwards, and the change is radical under afternoon.

3
 I also assess the impact of including the overnight component into the formation period
and I find that returns to the extreme winners and losers are amplified in a way which
makes reversal a profitable strategy. That is, stocks that lost the most in the morning,
inclusive of the overnight return, earn even higher returns in the afternoon, while stocks
that won the most, now gain less. The outcome is independent of the holding period, i.e.
persists under last and afternoon, and raises the economic opportunity to go long past
losers and to short past winners.

The remainder of the study is organized as follows: section 2 describes the data and presents
preliminary results on intraday unconditional average returns. Section 3 examines time-
series predictability and implements market timing strategies. Section 4 moves the focus
from the time-series dimension to the cross-sectional one, documents the U-shaped pattern
in average returns of portfolios formed on morning performance and tests the robustness of
this pattern. Finally, Section 5 concludes and suggests directions for further research.

2 Data and preliminary results

This section describes the sample and presents preliminary results on the unconditional
average returns by half-hour.

The sample includes all US equities that belong to both the Center for Research in
Security Prices (CRSP) and the Trades And Quotes (TAQ) databases, with a coverage
that extends to the New York Stock Exchange (NYSE), the American Stock Exchange
(AMEX or currently NYSE MKT), the NYSE Arca exchange and NASDAQ’s National
Market System (NMS, limited to OTC trades by TAQ).
I only keep common stocks, i.e. series with CRSP Share Type Code of 10 or 11, and
exclude observations that belong to microcaps in order to mitigate the spurious effects
induced by microstructure issues. That is, holding period returns do not include obser-

4
vations from stocks which, on the previous day, either had a price below $5 or a market
capitalization in the lowest New York Stock Exchange (NYSE) decile. Moreover, to allevi-
ate the impact of stale prices, I require stocks to be sufficiently liquid. Specifically, I only
keep the stock-date pairs that had at least 79 observations on the preceding day, which is
equivalent to a security being traded on average every 5 minutes during the 9:30 to 16:00
session. The resulting sample has a total of 8924 equities with an average of more than
1800 stocks per day, and covers the period from January 1993, first date of availability of
the TAQ, until May 2010.3
High-frequency data from TAQ are cleaned of irregular and misreported trades following
the usual rules from Barndorff-Nielsen et al. (2009) and Bollerslev et al. (2016), and are
assigned permnos from CRSP. The link between the two datasets is established through
the historical cusip and matches on average 98% of the daily TAQ observations.4 Intraday
prices are used to calculate half-hour returns by splitting the 9:30 to 16:00 trading day into
13 non-overlapping intervals.5 I keep the overnight return separate from the first interval
and I avoid backfilling the first price of the day to the opening of the session. In addition
to returns, I also calculate the realized volatility over each half-hour of interest using prices
sampled at 5-minute intervals.

2.1 Unconditional moments of intraday returns

This section explores patterns in the unconditional average and standard deviation of half-
hour returns and compares the evidence to existing literature. For instance, Wood et al.
(1985) and Jain and Joh (1988) find that returns are higher at the beginning and end of
the trading day, with a similar pattern applying to volatility. However, Harris (1986) shows
3
Days with partial trading times, either due to recurring festivities like Christmas’ Eve, or due to major
disruptions like the power outage of 2003/08/15, are excluded from the analysis.
4
For a detailed and general description of the linking procedure between TAQ and CRSP, the cleaning
and sampling of high-frequency data, refer to Komarov (2016).
5
All intervals are of the lb ≤ x < ub type except for the last that half-hour which also includes the
upper bound.

5
that the initial positive returns are driven by the previous close-to-open appreciation, and
that generally, returns are negative right after the open.6 Moreover, while the author does
not address differences in standard deviation, I provide such decomposition.

Figure 1 shows in the top plot the cross-sectional dispersion in intraday returns for the
13 half hours and the overnight interval. For each stock and sub-interval, I calculate the
average return across days. Then, for each half-hour, the 25th, 50th and 75th percentiles
of the cross-sectional distribution are respectively denoted in the plot by the lower bar, the
marker and the upper bar. The bottom graph plots the cross-sectional dispersion in the
standard deviation, calculated across days, for the half-hour and overnight returns. All
values are annualized.
The pattern in unconditional returns is consistent with the results from Harris (1986).
The first two half-hours record a median annualized return of -9.2% and -5.9%, with the
lower bars stretching much further than the upper ones, indicating a general trend of
negative returns at the beginning of the trading day. The last half-hour exhibits the
opposite behaviour, with a value of 8.9% and a much longer upper bar than the negative
one, showing overall very high returns near close-of-day. The intermediate half-hours, i.e.
from 10:30 until 15:30, show absence of strong trends and exhibit contained and symmetric
dispersion with median values that range from -1.4% to 2.5% circa. The overnight return
exhibits the highest annualized value, at 20.1%, and the biggest dispersion, with the upper
bar stretching four times as high as the median and the lower bar going negative but close
to zero.
In related works, Kelly and Clark (2011) document a risk premium for the overnight
return at about 23% whereas Lou et al. (2016) find that return to momentum strategies
are mostly realized overnight, which they estimate at about 11%. Meanwhile, the intraday
return is either strongly negative or relatively negative in both studies, thus confirming the
6
Henceforth, I will refer interchangeably to previous close-to-open return as overnight.

6
Figure 1: cross-sectional dispersion of annualized averages (top) and standard deviations (bottom) of
half-hour and overnight returns. The trading day is partitioned in 13 half-hours and for every stock the
time-series average and standard deviation of returns are calculated for each sub-period and the overnight
time. Then, the lower bar represent the 25th percentile of the cross-sectional dispersion for a given half-
hour, while the marker (x) and the upper bar, respectively delimit the median and the 75th percentile.
The time on the x-axis marks the start of a half-hour. Values are expressed in percentage. The sample
consists of all common NYSE stocks excluding microcaps, i.e. stocks with a price below 5$ or market
capitalization in the bottom decile, and covers the period from January 1993 to May 2010.

magnitude and shape of the results presented by figure 1. Additionally, the reported time-
series averages are not driven by a specific period of time. Figure A1 plots cross-sectional
averages over time and shows that returns at beginning of the day are usually lower than
those at the end, and that the overnight return is the highest.
The lower plot in figure 1, instead of averages, shows the cross-sectional dispersion in
standard deviations. The intraday volatility draws a u-shaped pattern, with annualized
median standard deviations of e.g. 27.4%, 12.5% and 17.6% for the first, middle and last
half hour. The close-to-open period exhibits similar median and dispersion to the first half-
hour. Overall, the differences among each sub-interval are less drastic if compared to the
distribution of the averages. Nonetheless, the cross-sectional variability is still significant,
with the difference between the 90th and 10th percentiles exceeding 40% volatility.

7
To summarize, the sample includes US common stocks matched in both the TAQ and
CRSP databases and excludes microcaps, i.e. those stocks with a price below $5 or a
market capitalization in the bottom NYSE decile. Results on unconditional average returns
by half-hour see negative returns in the first two intervals, positive and extremely positive
returns in the last and overnight intervals respectively. The evidence is consistent with
prior literature.

3 Time series momentum

In this section I build on the pattern in unconditional intraday returns from section 2 by
examining intraday time-series predictability of equity returns. After identifying two main
patterns, I explore their economic significance by implementing intraday timing strategies.

3.1 Intraday predictability

To examine intraday predictability I use pooled panel regressions with standard errors
clustered by time as in Moskowitz et al. (2012).

Methodology. I split the trading day, which starts at 9:30 and ends at 16:00, in 13 disjoint
half-hours and denote the intervals with h = 1, . . . , 13. I then regress the return rh,t
∗
of
half-hour h in day t on same-day returns of all previous half-hours except the contiguous
interval h − 1:7

h−2
∗
= a + b1 r1,t
∗
+ . . . + bh−2 rh−2,t
∗
+ eh,t = a + ∗
+ eh,t . (1)
X
rh,t bk rh,t
k=1

Returns are scaled by ex-ante volatility, i.e. rh,t

∗
= rh,t /σh,t−1 , where the ex-ante variance
7
The inclusion of the contiguous interval h − 1 among the explanatory variables does not affect the
main results, but introduces the spurious effect of the bid-ask bounce described by Roll (1984). In other
words, large and negative t-statistics are limited to the contiguous interval, while everything else remains
unchanged. Results are reported in table A1.

8
 2
σh,t−1 is calculated as the exponential average of daily realized variances RVh,t for the
half-hour of interest:8
∞
2
= α(1 − α)i RVh,t−1−i
X
σh,t−1
i=0

and α is chosen such that 83% of the weight is given to the first 60 days, i.e. α = 2/(60+1).
The use of standardized returns is justified by the evident cross-sectional differences in
volatility plotted in figure 1. Nonetheless, the specification in equation (1) is robust to the
use of unscaled returns rh,t , as sustained by results in table A1.

Results. Table 1 reports t-statistics and scaled coefficients (by 100) of time-series predicta-
bility regressions. Each row represents a pooled panel regression described by equation (1),
where the explained half-hour is indicated by the row header and the predicting intervals
are specified in the columns. For example, the first row regresses the return realized over
the 15:30 – 16:00 interval onto the half-hour returns that go from 9:30 until 15:00. The
second row regresses the 15:00 – 15:30 interval onto half-hours from 9:30 until 14:30, and
so forth. The dependent variable from the last row is the return over the 10:30 – 11:00
interval.
Results are laid out to allow the column-wise inspection of the impact of a single half-
hour on several future intervals. At the same time, the multivariate regression favours the
interpretation of a predictive interval composed by several half-hours. Nevertheless, esti-
mates are robust to the univariate equivalent of equation (1) and are reported in table A1.
Similarly, rows also can be pooled together into an interval that spans several half-hours,
granted an appropriate cutoff is selected for the predictive period. In this fashion, I identify
two main patterns from the t-statistics in table 1:

• from the first row, the 15:30 – 16:00 return is positively predicted by the interval from
9:30 to 12:00. I will commonly refer to this combination of predictive and predicted
8
An overview of the literature on realized measures is given in Komarov (2016). For this application,
daily realized variances are calculated for each half-hour as the sum of squared returns sampled at the
5-minute frequency.

9
Table 1: t-statistics and coefficients from time-series predictability regressions. Each row is a pooled
panel regression of the return from the half-hour denoted by the row-header onto a constant and all, but
the contiguous, half-hour returns of the same day that precede it. Returns are scaled by ex-ante volatility,
which is measured by the exponential average of the 5-minute realized volatilities calculated over the
respective half-hour. The centre of mass of the exponential average is fixed at 60 days. The t-statistics are
computed from standard errors clustered by time at the daily level. Darker colours highlight statistical
significance at the 1% level and lighter colours that at the 5%. The sample consists of all common NYSE
stocks excluding microcaps, i.e. stocks with a price below 5$ or market capitalization in the bottom decile,
and covers the period from January 1993 to May 2010.

t-statistics
c 09:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30
15:30 3.46 2.24 1.92 1.77 0.78 2.44 1.12 -1.32 -0.10 -0.01 0.89 1.47
15:00 2.00 -2.04 -1.13 -2.70 -0.87 -0.32 -2.05 -2.05 0.33 -0.58 0.91
14:30 2.58 -0.04 -0.92 1.39 0.97 0.64 0.67 0.69 2.48 0.81
14:00 -0.87 -2.03 -0.76 -0.72 -0.49 -0.71 -0.33 -0.28 0.83
13:30 0.11 -0.11 0.29 -1.39 -2.12 -1.31 -0.72 -1.00
13:00 1.76 1.55 0.09 0.36 0.21 -1.67 -0.78
12:30 2.47 0.68 0.22 0.69 -1.02 -0.69
12:00 0.97 0.86 1.16 1.32 1.35
11:30 0.10 0.75 -1.32 1.82
11:00 -0.96 1.02 -1.76
10:30 -0.10 -2.08
coefficients * 100
c 09:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30
15:30 2.43 0.74 0.72 0.86 0.38 1.24 0.57 -0.75 -0.07 0.00 0.64 1.00
15:00 1.45 -0.74 -0.44 -1.46 -0.48 -0.18 -0.95 -1.28 0.24 -0.35 0.71
14:30 1.75 -0.01 -0.31 0.59 0.44 0.30 0.27 0.31 1.21 0.36
14:00 -0.60 -0.61 -0.29 -0.32 -0.21 -0.36 -0.14 -0.16 0.50
13:30 0.07 -0.03 0.09 -0.56 -0.89 -0.64 -0.32 -0.56
13:00 1.07 0.40 0.03 0.16 0.09 -0.75 -0.37
12:30 1.39 0.18 0.07 0.27 -0.42 -0.32
12:00 0.55 0.23 0.34 0.53 0.57
11:30 0.06 0.21 -0.47 0.75
11:00 -0.57 0.29 -0.61
10:30 -0.06 -0.58

10
 interval as the last specification;

• the return from 13:30 to 15:30 is negatively predicted by the interval from 9:30 until
13:00. I will refer to this combination as the afternoon specification.

In particular, the predictive regression for the last half-hour lists t-statistics of 2.24 and
2.44 for the (explanatory) half-hours starting respectively at 9:30 and 11:30, and exhibits
statistical significance at the 10% level in the 10:00 and 10:30 periods. However, the highest
impact is captured by the intercept with a scaled coefficient of 2.43 basis points (bp) per
day and a t-statistic of 3.46. The afternoon combines predicted sub-intervals from 13:30
to 15:30 and loads throughout the morning more sparsely than last, but exhibits stronger
statistical significance (all t-statistics are below -2). For instance, the half-hours starting
at 9:30, 10:30, 12:00 and 12:30 negatively predict the 15:00 – 15:30 interval. The 9:30
period is also relevant for the prediction of the 14:00 – 14:30 interval, while the half-hour
starting at 11:00 exhibits a t-statistic of -2.12 for the 13:30 – 14:00 period. As with last,
the intercept is significant and has a positive sign for two of the predicted sub-intervals.
While the choice of the predictive and predicted intervals is to some extent discretionary,
the focus is on identifying a general and preferably robust regularity in returns rather than
on finding the optimal timing strategy.9 In this spirit, the last and afternoon configuration
aim to capture the tendency suggested by the t-statistics of morning returns predicting
reversal in afternoon returns, but with a strong exception for the last half-hour, charac-
terized instead by continuation. Nevertheless, section 3.2 covers alternative specifications
and compares results with Gao et al. (2015).
Finally, for completeness, R2 from the regressions of table 1 are included in the first
column of table A2. Values range from 0.063% for the prediction of the half-hour starting
at 15:00, to 0.003% for the half-hour at 10:30. In comparison, e.g. Campbell (1991)
documents a R2 of 2.5% for monthly regressions from 1926 in equities. Considering that to
9
In the context of a timing strategy, the predictive interval will constitute the formation period while
the predicted interval will be the holding period.

11
longer holding periods correspond larger R2 (e.g Cochrane, 2008; Fama and French, 1988),
the values of my analysis remain reasonable.

3.2 Timing strategies

Next, I investigate the profitability of trading strategies based on the timing patterns
defined by last and afternoon in the previous section. For each stock and day, I calculate
the return of the intraday formation period. For last, it is the return through half-hours 1
to 5, i.e. from 9:30 until 12:00, and for afternoon it is the return through half-hours 1 to 7,
i.e. from 9:30 until 13:00. Then, I sort stocks into those which had positive and negative
returns in the morning and form two equal-weighted portfolios. For the afternoon strategy
I hold positions through half-hours 9 to 12, i.e. from 13:30 until 15:30, while for the last
strategy, stocks are kept for the last half-hour only, i.e. from 15:30 until close at 16:00.
Mathematically this can be summarized as:

PN win
rtwin = 1
r9:12,i if r1:7,i > 0


win i=1

afternoon :  N
lose
rtlose = r9:12,i if r1:7,i < 0
 1 PN
lose i=1

N

(2)

PN win
rtwin = 1
r13,i if r1:5,i > 0


win i=1

last : N
lose
rtlose = r13,i if r1:5,i < 0

 1 PN
lose i=1

N

where N ∗ is the number of stocks with positive or negative returns in the formation period
and rstart:end is short-hand for intraday return of day t through half-hours h = start, . . . , end.
The winners-minus-losers and long-only portfolios are simply the difference and the sum
of the win and lose portfolios:

rtWML = rtwin − rtlose

(3)
rtlong = rtwin + rtlose .

12
 Table 2 reports descriptive statistics of daily returns earned by the equal-weighted
portfolios of morning losers (lose) and winners (win), the winners-minus-losers (WML)
timing strategy and the long-only portfolio (win plus lose) from equation (3). Among
other statistics, I list the average return (avgret) and standard deviation (std), the Newey
and West (1987) standard error of the average with default-bandwidth (se) and its pvalue
(pval), the annualized return, standard deviation and Sharpe Ratio (annret, annstd and
SR), skewness and kurtosis (skew and kurt), and the time-series average of the cross-
sectional averages of returns during the formation period (avgsignal).10 Returns and
related measures are expressed in percentage. The table summarizes results for the last
and afternoon specifications respectively on the left and right sides.
First, the average daily return of the lose portfolio under the last specification is in-
consistent with the predictability results from table 1. Instead of continuing the negative
performance, morning losers earn an average of 3.8 bp per day. Second, the reversal pre-
dicted with the afternoon pattern is also not supported by the profitability of the morning
winners and losers. The former group, instead of the anticipated negative performance,
exhibits a positive return of 2.5 bp per day, and the latter group records a statistically
insignificant daily value of 0.7 bp. In terms of WML returns, the last and afternoon re-
spectively record an average of -0.1 and 1.8 bp, hence, denying continuation in the former
case and reversal in the latter. Finally, the long-only variant suggests that, no matter the
performance recorded in the morning, stocks appreciate in the second half of the day, and
especially so in the last half-hour. The evidence strongly suggests absence of predictability
and confirm the unconditional pattern in intraday returns shown in figure 1.
The summary statistics of equal-weighted portfolios do not offer any insight on the
evolution of the strategies over-time. For this purpose, figure 2 plots the monthly cumulated
returns earned by equal-weighted portfolios of morning winners and losers, of the winners-
10
Daily returns are annualized by a factor of 252 and daily standard deviations by the square root of
252.

13
Table 2: descriptive statistics of daily returns earned by the equal-weighted portfolio of morning losers
(lose), winners (win), the winners-minus-losers (WML) timing strategy and the long-only variant (win
plus lose). Specifically, stocks are sorted every day into winners and losers based on the performance over
the intraday formation period, and the return of equal-weighted portfolios is calculated for the holding
period. The table is organized in two panels. On the left is the last timing specification, with formation
period from 9:30 until 12:00 and holding period from 15:30 to 16:00, and on the right is the afternoon
timing specification, with formation period from 9:30 until 13:00 and holding period from 13:30 to 15:30.
The table lists in this order: the average return (avgret), standard deviation (std), the default-bandwidth
Newey-West standard error of the average (se) and its pvalue (pval), the annualized return, standard
deviation and Sharpe Ratio (annret, annstd and SR), the annualized downside deviation with return
threshold at 0 (downstd), the minimum, median and maximum returns (minret, medret and maxret),
skewness and kurtosis (skew and kurt), the maximum drawdown return, its length in months and the
number of months that the price series took to recover from it (mdd, mddlen and reclen), the sortino ratio
(sortino), and the time-series average of the cross-sectional averages of returns during the formation period
(avgsignal). Asterisks in the reclen value indicate truncation from end-of-sample period. The average
signal for WML and long-only is omitted. Returns and related measures are expressed in percentage. The
sample consists of all common NYSE stocks excluding microcaps, i.e. stocks with a price below 5$ or
market capitalization in the bottom decile, and covers the period from January 1993 to May 2010.

last afternoon
win lose WML long win lose WML long
avgret 0.037 0.038 -0.001 0.075 0.025 0.007 0.018 0.032
std 0.354 0.365 0.106 0.712 0.569 0.611 0.227 1.159
se 0.005 0.005 0.002 0.010 0.007 0.008 0.004 0.014
pval 0.000 0.000 0.680 0.000 0.001 0.326 0.000 0.022
annret 9.318 9.527 -0.209 18.846 6.304 1.861 4.443 8.164
annstd 5.621 5.798 1.675 11.297 9.036 9.706 3.600 18.406
SR 1.658 1.643 -0.125 1.668 0.698 0.192 1.234 0.444
downstd 3.756 3.802 1.252 7.498 6.147 6.423 2.808 12.413
minret -3.260 -3.269 -1.077 -6.358 -3.671 -3.957 -7.457 -7.628
medret 0.040 0.035 -0.001 0.080 0.050 0.013 0.014 0.065
maxret 3.346 3.511 1.259 6.857 7.003 8.380 1.937 13.348
skew -0.106 0.044 -0.784 -0.078 0.615 1.392 -8.757 0.836
kurt 20.019 19.580 19.776 19.973 15.357 22.737 282.823 16.194
mdd 14.529 9.253 33.958 23.132 13.620 53.671 54.608 30.398
mddlen 637 380 3620 637 1107 2828 3506 1121
reclen 477 28 9* 46 728 3073 4* 714
sortino 2.481 2.506 -0.167 2.514 1.025 0.290 1.582 0.658
avgsignal 0.016 -0.016 - - 0.017 -0.017 - -

14
minus-losers (WML) trading strategy and the long-only variant.

Figure 2: monthly cumulated returns earned by equal-weighted portfolios of morning winners (cross-
marked) and morning losers (circle-marked), the winners-minus-losers (WML) trading strategy (triangle-
marked) and the long-only variant (winners plus losers, no markers). Shaded areas denote NBER recessi-
ons. Specifically, stocks are sorted every day into winners and losers based on the performance over the
intraday formation period, and the return of equal-weighted portfolios is calculated for the holding period.
In the top plot is the last timing specification, with formation period from 9:30 until 12:00 and holding
period from 15:30 to 16:00, and in the bottom plot is the afternoon timing specification, with formation
period from 9:30 until 13:00 and holding period from 13:30 to 15:30. The sample consists of all common
NYSE stocks excluding microcaps, i.e. stocks with a price below 5$ or market capitalization in the bottom
decile, and covers the period from January 1993 to May 2010.

The portfolio with the morning losers from the last strategy constantly appreciates over-
time, strengthening the conclusions drawn from table 2 and again rejecting any pattern
of performance continuation. For the afternoon specification, the win and lose portfolios
display momentum until 2001 but reversal afterwards, the latter being consistent with
the predictability results. However, rather than a convergence to the expected behaviour,
the afternoon strategy is now mimicking last, where morning losers appreciate faster than
the winners. The absence of predictability is also confirmed by the low average rate of
correct predictions under afternoon, which is equal to 45.1% for the whole sample and
to a slightly higher 48.0% after 2001. Similar estimates come from last, with averages

15
of 43.7% and 49.1% for the whole sample and for the period after 2001. Predictability
aside, figure 2 suggests the presence of a structural break in January 2001, which coincides
with the introduction of decimalization in stock prices.11 While decimalization had an
undeniably strong economic impact on financial markets, it is outside the scope of this
analysis to discuss the microeconomic foundations that might support (or not) intraday
predictability.12 Notwithstanding, the patterns in the t-statistics from table 1 only really
appear after 2001, as the sub-period analysis of predictability suggests in table A1.

Alternative specifications. In related work, Gao et al. (2015) find that the return of the
first half-hour predicts the return of the last half-hour. Their work differs in a few aspects:
first, their analysis is based on the SP500 ETF only, second, they include the overnight
return in the definition of the first half-hour return and finally, they cover the sample
period from 1993 until 2013.13 Even though I am able to closely match their predictability
results, using the first half-hour as the sole predictive interval still produces inconsistent
results in terms of win and lose portfolios. Moreover, the inclusion of the overnight return,
while marginal in terms of predictability (see table A1), produces an unexpected reversal
effect with average returns for the lose, win and WML portfolios of 1.6 (3.7), 5.7 (3.8) and
-4.2 (-0.1) bp per day (values in parenthesis are from table 2). The impact of the overnight
return is analysed in more detail in section 4.4.
Additionally, I implement other strategies and report the correlations with the last and
11
The decimalization was launched as a pilot in September 2000 following orders by SEC (2000b) and
SEC (2000a), and was fully implemented by NYSE in January 2001. NASDAQ and other regional markets
followed at the end of March 2001.
12
For an analysis of the impact of the decimalization on market quality, see Bessembinder (2003), while
for a more general overview see SEC (2012). In general, tick-size reduction plays a fundamental role in
liquidity provision and its effects have been analysed by Goldstein and A. Kavajecz (2000) after prices, in
1997, went from being quoted in eights to sixteenths. From another perspective, Angel (1997) and Schultz
(2000) give evidence that (wider) relative tick-sizes act as incentive for the market dealer and that strategic
splits can promote trading in the stock. Hence the connection between optimal tick sizes and liquidity.
In this respect, Ball and Chordia (2001) show that the optimal spread for large stocks was bound by the
tick-size prior to decimalization.
13
They also exclude days with less than 500 trades which effectively sets the beginning of their sample
period to 1997.

16
afternoon in table 8. For instance, I exclude the first half-hour from the formation pe-
riod of the last specification, setting it to the interval from 10:30 until 12:00. Next, since
the second-last half-hour, i.e. the one starting at 15:00, has the most of the significant
t-statistics among the whole afternoon, I set it aside in a separate specification (the for-
mation period remains from 9:30 to 13:00). I also check the relevance of the first half-hour
in predicting the second-last half-hour by excluding it from the formation period. The
evidence from the correlations in table 8 is examined in more detail in the next section,
but confirms that alternative specifications are only marginally different from last and
afternoon.

To summarize, I investigate the existence of intraday predictability in returns and im-

plement timing strategies based on the last and afternoon patterns. The former predicts
continuation of morning performance in the last half-hour, while the latter predicts rever-
sal of morning performance in the afternoon (excluding the last interval). However, the
economic performance of the trading strategies is inconsistent with the results from the
predictability regressions. In fact, I observe a general tendency of stocks to earn positive
returns in the afternoon, especially in the last half-hour, and morning losers to appreciate
faster than morning winners. The evidence is consistent with the pattern of unconditional
returns from figure 1 and hints at a cross-sectional phenomenon rather than a time-series
one. Additionally, the evolution of the timing strategies uncover a structural break at the
beginning of 2001 which coincides with the price decimalization of US stocks.

4 Cross-sectional patterns

Section 3 does not find economic evidence supporting intraday predictability. However,
the systematic appreciation that stocks exhibit in the second half of the day, with past
winners earning higher returns than past losers, raises the question whether there are
significant and systematic cross-sectional differences. I address this question in this section

17
and sort stocks on their performance and on several stock characteristics. I find that stocks
which either performed very poorly or extremely well in the morning, keep earning in the
afternoon more than others. I also investigate how the uncovered pattern behaves before
and after the decimalization in 2001, whether it is exclusive to a particular day of the week
and how it is affected by the inclusion of the overnight return in the formation signal.

Methodology. I use the last and afternoon specifications from section 3 to be consis-
tent with the time-series analysis, and explore variations to those in section 4.4. As in
section 3.2, for each stock and day I calculate the return of the intraday formation period.
For last, it is the return from 9:30 until 12:00, and for afternoon it is the return from
9:30 until 13:00. Then, I sort stocks into deciles based on their performance and form
ten equal-weighted portfolios. Finally, the performance of the ten portfolios is measured
from 13:30 until 15:30 according to the afternoon timing, and over the last half-hour of the
trading day as defined by the holding period in last.

Results. Table 3 reports descriptive statistics of daily returns (akin to table 2) earned
by decile portfolios formed on past performance and by an equal-weighted portfolio of all
stocks. Portfolios are ordered from lose to win, i.e. in ascending order of the formation
signal. In particular, the table lists the average return (avgret) and its Newey and West
(1987) standard error with default-bandwidth (se), the annualized return, standard devia-
tion and Sharpe Ratio (annret, annstd and SR), skewness and kurtosis (skew and kurt),
and the time-series average of the cross-sectional averages of returns during the formation
period (avgsignal). Returns and related measures are expressed in percentage. The table
summarizes results for the last and afternoon specifications respectively in the top and
bottom panels.
First, the annualized average return of 9.4% earned by the portfolio of all stocks under
the last definition is in line with the median value of 8.9% provided in section 2.1 by
the unconditional average return of the last half-hour. Second and most importantly, the

18
average returns of the ten portfolios form a U-shaped pattern, i.e. stocks that lost or gained
the most in the morning, tend to earn higher and positive returns in the second half of the
trading day, and especially during the last half-hour. For instance, under last the lose and
win portfolios respectively earned an impressive 6.2 and 7.7 bp per day, which translate
to 15.6 and 19.4% in annualized terms. On the other hand, most of the intermediate
portfolios, with stocks belonging to deciles 2 through 9, recorded average returns below 3
bp and never higher than 4 bp per day. It is also worth pointing out that standard errors
create a similar U-shape, and while the lose and win portfolios exhibit larger dispersion
in average returns than the intermediate deciles, the second moments are not big enough
to offset the superior performance. In fact, not only the win and lose portfolios have the
highest Sharpe Ratios, they also have positive skewness, unlike portfolios 3 through 7, and
the lowest values of kurtosis. In other words, stocks belonging to the two extreme deciles,
earn positive returns during the last half-hour more consistently than those from the other
deciles.
In regard to afternoon, the U-shaped pattern in average returns uncovered by last is
slightly weaker. The lose and win portfolios still have the highest values at 6.2 and 2.2 bp
but the latter return is much lower than the 7.7 bp from the last equivalent. This indicates
that morning losers keep earning superior returns in the afternoon, i.e. from 13:30 until
15:30, but past winners are closer to stocks from intermediate deciles. Most indicative of
this difference in the cross-sectional behaviour are skewness and kurtosis: portfolios exhibit
monotonically decreasing values in both moments. For example, skewness goes from 3.88
to -0.05 from lose-to-win of the cross-sectional deciles. Similarly, kurtosis starts at 78.43
on the first decile and ends at 10.86 on the last decile, with the first (last) five portfolios
having higher (lower) estimates than those shown under last. In short, the upside potential
of morning winners is somewhat more limited during the interval from 13:30 to 15:30 than
in the last half-hour of trading.

19
Table 3: descriptive statistics of daily returns earned by equal-weighted decile portfolios formed on past
performance. Every day, stocks are sorted into ten groups based on their performance during the intraday
formation period. Then, ten equal-weighted portfolios are formed and their return is calculated over the
holding period. Additionally, the return for an overall equal-weighted portfolio is reported in the last
column. The table is organized in two panels. In the top panel is the last timing specification, with
formation period from 9:30 until 12:00 and holding period from 15:30 to 16:00, and in the bottom panel
is the afternoon timing specification, with formation period from 9:30 until 13:00 and holding period
from 13:30 to 15:30. The table lists in this order: the average return (avgret) and the default-bandwidth
Newey-West standard error of the average (se), the annualized return, standard deviation and Sharpe Ratio
(annret, annstd and SR), the minimum, median and maximum returns (minret, medret and maxret),
skewness and kurtosis (skew and kurt), and the time-series average of the cross-sectional averages of returns
during the formation period (avgsignal). Returns and related measures are expressed in percentage. The
sample consists of all common NYSE stocks excluding microcaps, i.e. stocks with a price below 5$ or
market capitalization in the bottom decile, and covers the period from January 1993 to May 2010.

last
lose 2 3 4 5 6 7 8 9 win all

Avgret 0.062 0.040 0.030 0.028 0.026 0.025 0.024 0.026 0.033 0.077 0.037
Se 0.007 0.006 0.005 0.005 0.005 0.005 0.004 0.005 0.005 0.007 0.005
Annret 15.625 10.085 7.553 6.984 6.496 6.245 5.949 6.628 8.234 19.413 9.423
Annstd 7.473 6.454 5.819 5.562 5.402 5.352 5.334 5.441 5.830 7.068 5.691
SR 2.091 1.563 1.298 1.256 1.203 1.167 1.115 1.218 1.412 2.747 1.656
Minret -3.500 -3.441 -3.359 -3.290 -3.431 -3.092 -3.007 -3.074 -3.047 -3.744 -3.203
Medret 0.050 0.038 0.033 0.028 0.032 0.029 0.026 0.028 0.033 0.074 0.039
Maxret 4.349 4.349 3.603 3.245 3.174 3.424 3.444 3.388 3.771 6.036 3.468
Skew 0.648 0.513 -0.153 -0.185 -0.212 -0.238 -0.136 0.068 0.045 0.455 -0.045
Kurt 15.579 20.457 21.167 21.487 20.928 21.451 20.989 20.739 18.335 19.623 20.039
Avgsignal -0.038 -0.018 -0.011 -0.007 -0.003 0.001 0.005 0.009 0.016 0.038 0.000
afternoon
lose 2 3 4 5 6 7 8 9 win all

Avgret 0.062 0.009 0.000 0.000 0.003 0.009 0.012 0.020 0.021 0.022 0.016
Se 0.011 0.009 0.008 0.007 0.007 0.007 0.007 0.007 0.008 0.010 0.007
Annret 15.620 2.146 0.106 0.027 0.828 2.328 2.972 4.992 5.182 5.666 4.046
Annstd 14.333 11.328 10.199 9.063 8.604 8.320 8.346 8.618 9.509 11.689 9.454
SR 1.090 0.189 0.010 0.003 0.096 0.280 0.356 0.579 0.545 0.485 0.428
Minret -5.361 -5.238 -4.050 -3.921 -3.810 -3.891 -3.587 -3.679 -4.062 -5.382 -3.901
Medret 0.048 0.016 0.005 0.013 0.003 0.023 0.020 0.034 0.044 0.072 0.034
Maxret 20.770 16.698 13.892 9.293 6.764 6.250 6.988 6.928 7.036 7.524 7.558
Skew 3.884 3.667 3.445 1.986 1.371 1.049 1.085 0.795 0.519 -0.046 1.251
Kurt 78.426 82.276 67.662 33.455 22.733 18.962 20.011 17.176 14.073 10.862 21.230
Avgsignal -0.041 -0.019 -0.012 -0.007 -0.003 0.001 0.005 0.010 0.018 0.042 0.000

20
4.1 Stock characteristics

The previous section uncovered the U-shaped pattern in the cross-section of average re-
turns, with morning winners and losers earning more in the afternoon, and especially during
the last half-hour, than stocks from intermediate deciles. In this section, I explore if this
pattern is driven by cross-sectional differences in stock characteristic like e.g. size.

Selecting stock features. Among the numerous stock characteristics I pick a classical
one like size (size), others which are strongly related to intraday activity like illiquidity
(illiq), volume (vol) and tick-size (tick), a feature emerging from recent results, i.e. the
realized skewness (skew), and for a purely descriptive perspective the realized volatility
(std).14 I use the logarithm of market capitalization from the previous month to proxy for
size, the logarithm of the previous-month illiquidity measure of Amihud (2002) for illiq,
the logarithm of the average daily volume for vol, and the average of the relative tick-size
for tick.15 For skew, I take the average of the daily realized skewness calculated as in
Amaya et al. (2015) and std is defined as the average of daily realized volatilities. Averages
are taken over the past quarter and rolled on a daily basis. Finally, realized measures are
calculated with intraday returns sampled at the 5-minute frequency.
Since many of the proposed characteristics are related to each other, only relevant stock
attributes are considered for the double sorts in order to avoid redundancy in the results.
14
Banz (1981) finds that small stocks outperform bigger stocks. Amihud (2002) suggests that an increase
in expected illiquidity should be compensated with higher returns. Similarly, Amihud and Mendelson
(1986) show that stock returns increase in the bid-ask spread, and hence in the relative tick-size which sets
the lower boundary for the spread. Admati and Pfleiderer (1988) propose a theoretical framework where
liquidity and informed traders will trade during periods of high volume. Amaya et al. (2015) find stock
returns with lower realized skewness outperform those with higher skewness.
15
The Amihud illiquidity measure is defined as:
252
1 X |rt−n |
illiqt = .
252 n=0 volt−n pt−n

The absolute tick size was fixed at 1/8 of a dollar until June 1997 when it was reduced to a 1/16 and
finally to a 1/100 in April 2001. The tick is then the size of the tick divided by the price from the previous
close.

21
For this purpose, table 4 displays the time-series average of cross-sectional correlations be-
tween stock characteristics. The table has two types of correlations, with Spearman’s rank
coefficients appearing above the main diagonal and the conventional Pearson’s coefficients
below the main diagonal. Specifically, at each point in time, I calculate the cross-sectional
correlation between the values of two characteristics. This produces a time-series of coef-
ficients for each pairing which I then average and summarize in the table.

Table 4: time-series average of cross-sectional correlations between stock characteristics. Spearman’s

rank correlations appear above the main diagonal and Pearson’s coefficients are below the main diagonal.
Definition of characteristics: size is the logarithm of the previous-month market capitalization; vol is the
logarithm of the average daily volume; skew is the average of the daily realized skewness from Amaya
et al. (2015); illiq is the logarithm of the previous-month illiquidity measure from Amihud (2002); tick is
the average of the relative tick-size; std is the average of the daily realized volatility. Average values are
computed over the past quarter and rolling on a daily basis. Realized measure are calculated with intraday
returns sampled at the 5-minute frequency. Colours highlight coefficients above 0.5 (green) and below -0.5
(orange). The sample consists of all common NYSE stocks excluding microcaps, i.e. stocks with a price
below 5$ or market capitalization in the bottom decile, and covers the period from January 1993 to May
2010.

size vol skew illiq tick std

size 0.55 0.16 -0.87 -0.60 -0.67

vol 0.59 0.04 -0.65 -0.09 -0.12
skew 0.14 0.05 -0.11 -0.09 -0.12
illiq -0.88 -0.62 -0.12 0.63 0.63
tick -0.45 -0.14 -0.11 0.50 0.62
std -0.58 -0.11 -0.14 0.57 0.59

From the first row, size is positively correlated to vol, with a coefficient of 0.55, and
is negatively correlated to illiq, tick and std, with values of -0.87, -0.60 and -0.67, corre-
spondingly. This type of strong affinity is expected since small stocks tend to be illiquid,
hence sporting less trading volume, to be more volatile and have wider relative tick-sizes in
order to attract market dealer’s interest (see Angel, 1997; Schultz, 2000). The implications
just outlined between market capitalization and the other features naturally translates into
the negative dependence between illiq and vol, with a Spearman correlation of -0.65, and
positive pairwise associations between illiq, tick and std, with values that fare in the low

22
0.60s.16 The skew remains orthogonal to all other characteristics since it preserves the sign,
in contrast with std, and because it describes past returns rather than company qualities.
From the analysis of correlations, I choose to report results from the double sorts for
size, since it is easier to interpret than illiq and has longer-standing contributions in the
literature, for skew, because of its orthogonality to the other features, and for std, which
diversifies on the evidence. On the other hand, I do not report results from the double
sorts on attributes related to trading activity since they do not add qualitative information
on top of that provided by the sort on market capitalization.

Double sorts. Every day, stocks are sorted into five groups based on their performance
during the intraday formation period. Stocks are also allocated independently to three
groups sorted on either size, std or skew. The intersection of the two sorts produces 15
groups which will form equal-weighted portfolios. The return of the portfolios is calculated
over the holding period. As explained at the beginning of this section, formation and
holding periods are defined by the last and afternoon configurations.
Table 5 collects average annualized returns earned by equal-weighted portfolios formed
on past performance and a stock attribute, and reports t-statistics in parentheses. Each
combination yields 15 values arranged into three-row by five-column matrices. Columns
are sorted from lose to win, i.e. in ascending order of the return realized in the formation
period, and rows are sorted in ascending order of the stock characteristic, e.g. for the size
characteristic into small, medium and large stocks. The table summarizes results for the
last specification on the left and for afternoon on the right side.
A general inspection of the table shows that the U-shaped pattern from the cross-
sectional sort on performance (see table 3), also persists after controlling for stock attribu-
tes. For example, let us consider the last specification: the shape of average returns across
the small, medium and large size groups is clearly parabolic, i.e. extreme morning winners
16
The negative dependence between volume and illiquidity arises mechanically since the former appears
in the denominator of the Amihud measure.

23
and losers earn more than the intermediate portfolios. More specifically, small stocks from
the first (lose), third and fifth (win) groups, have an annualized return of 19.1, 12.6 and
21.5%, on average. As a side note, the same findings extend to vol, illiq and tick, but are
omitted for conciseness.17
A minor exception to the usual U-pattern, which I will explore in more detail in the next
section, arises for the low-volatility group. Returns decrease monotonically as morning
performance improves. In particular, while morning losers yield 7.9% in the afternoon,
morning winners average at 1.1% only. Otherwise, the medium- and high-volatility groups
conform with the previous results. This description suggests that the less volatile stocks
can recover from a drawdown with a sudden surge, but are not likely to jump up if they
were already earning a profit.
The U-pattern aside, size, std and the unreported features are still responsible for cross-
sectional variation in returns. Considering market capitalization again, e.g. the lose group
exhibits decreasing returns of 19.1, 9.6 and 7.4%, as we look at small, medium and then
large stocks. However, the influence of skew on the cross-section is less pronounced, but
at the same time the U-shape is clearly outlined. For instance, the maximum difference
in returns between the low- and high-skewness groups is of only 2.1% (win column). In
contrast, size sets a distance of 16.3% between small and large stocks (win column), and
std gets an even bigger 18.8% between the low and high portfolios (win column).
Most of the conclusions drawn for last also apply to the afternoon specification, with
two caveats. First, returns are generally lower, consistently with the unconditional results
by half-hour (section 2), and, as observed in table 3, the parabolic pattern is weaker.
For example, low-skewness stocks from the first (lose), third and fifth (win) performance
groups, exhibit estimates of 7.0, -0.2 and 3.2% respectively, sign that the pattern is not as
well defined as under last, where the profit of the win group is on par with that of the losers.
Second, the impact of features on average returns is not always coherent across performance
17
Results for double sorts on the excluded characteristics are available upon request.

24
groups, which is probably also due to lower statistical significance as t-statistics are often
smaller than 2 in absolute terms. Let us consider size: smaller stocks are more lucrative
than bigger ones in the first and last portfolios but this scenario is reversed for the second
and third group. Similarly, average returns are inversely proportional in the std for all
but the win portfolio which behaves just the opposite. On the other hand, the effect of
skew is largely consistent across groups formed on morning returns but it is proportionally
increasing in the values of skewness, which is in opposition with the influence shown under
last.

25
Table 5: average annualized returns earned by equal-weighted portfolios formed on past performance and
stock characteristics. Every day, stocks are sorted into five groups based on their performance during the
intraday formation period. Stocks are allocated independently to three groups sorted on either size, std or
skew. The intersection of the two sorts produces 15 equal-weighted portfolios and their return is calculated
over the holding period. Definition of characteristics: size is the logarithm of the previous-month market
capitalization, std is the average of the daily realized volatility and skew is the average of the daily realized
skewness from Amaya et al. (2015). Average values are computed over the past quarter and rolling on
a daily basis. Realized measure are calculated with intraday returns sampled at the 5-minute frequency.
In parentheses are t-statistics calculated from the default-bandwidth Newey-West standard errors of the
average. The table is organized in two panels. In the left panel is the last timing specification, with
formation period from 9:30 until 12:00 and holding period from 15:30 to 16:00, and in the right panel is
the afternoon timing specification, with formation period from 9:30 until 13:00 and holding period from
13:30 to 15:30. Returns are expressed in percentage. The sample consists of all common NYSE stocks
excluding microcaps, i.e. stocks with a price below 5$ or market capitalization in the bottom decile, and
covers the period from January 1993 to May 2010.

last afternoon
lose 2 3 4 win lose 2 3 4 win
size
small 19.1 13.1 12.6 13.3 21.5 10.5 -3.4 1.2 3.9 6.8
[10.7] [8.3] [8.4] [8.9] [12.0] [4.0] [-1.6] [0.6] [1.9] [2.8]
medium 9.6 6.3 5.5 5.6 10.8 8.8 1.0 0.9 3.2 3.0
[6.4] [5.2] [4.9] [5.0] [7.5] [3.5] [0.5] [0.5] [1.9] [1.3]
large 7.4 4.5 2.7 1.9 5.3 7.5 2.8 2.3 3.9 4.3
[4.7] [4.1] [2.7] [1.8] [3.9] [2.8] [1.5] [1.5] [2.4] [2.0]
std
low 7.9 3.8 2.0 0.7 1.1 17.2 5.2 3.0 3.1 -0.1
[6.0] [3.8] [2.1] [0.8] [1.0] [8.1] [3.4] [2.2] [2.2] [-0.1]
medium 9.4 6.3 5.4 5.2 8.1 11.1 0.9 1.9 3.0 4.2
[6.5] [4.9] [4.5] [4.4] [5.8] [4.5] [0.5] [1.1] [1.7] [2.0]
high 16.4 13.1 13.4 13.8 19.9 4.2 -5.7 -3.8 1.9 4.9
[8.9] [7.7] [8.2] [8.6] [10.7] [1.5] [-2.3] [-1.6] [0.8] [1.8]
skew
low 12.8 7.8 6.5 6.4 13.6 7.0 -0.4 -0.2 3.9 3.2
[7.8] [6.1] [5.6] [5.4] [9.0] [2.8] [-0.2] [-0.1] [2.2] [1.4]
medium 11.5 6.7 6.0 6.4 12.2 8.4 0.4 1.3 2.0 5.0
[7.4] [5.4] [5.3] [5.6] [8.1] [3.1] [0.2] [0.8] [1.2] [2.2]
high 11.1 6.3 4.9 4.8 10.1 9.4 0.8 2.1 3.9 5.0
[7.1] [5.2] [4.5] [4.3] [7.1] [3.7] [0.4] [1.3] [2.3] [2.3]

26
4.2 Sub-period analysis

In section section 3.2 I pinpoint a potential structural break at the beginning of 2001. Both
figure 2 and the robustness by sub-period on the predictability analysis covered in table A1,
highlight a change in the behaviour of morning winners and losers after the suggested date.
Moreover, the timing coincides with the introduction of the decimalization of US quotes,
which bore deep microstructural implications on day-trading (for a general overview see
Bessembinder, 2003). In this section, I split the horizon at the end of April 2001 in two sub-
periods, and form double-sorts on each interval. Implications for the simple cross-sectional
sort on past performance are easily deduced.18
Table 6 reports the sub-period analysis of average annualized returns earned by equal-
weighted portfolios independently formed on past performance and a stock attribute. Met-
hodology and organization of results mimic table 5. Additionally, the upper panel presents
estimates for the period up to April 2001 and the lower panel presents estimates for the
subsequent interval. The t-statistics of the averages are listed in table A4.
Two main points arise from a general overview of the estimates. First, the parabolic
pattern in average profits is the strongest before April 2001 and overall under last. While
this is generally true across all stock attributes, small stocks generate the highest observed
returns with e.g. the first (lose), third and fifth (win) groups listing annualized values
of 20.7, 17.4 and 34.4% respectively. Instead, afternoon’s profile takes the shape of a
right-smirk with the first three groups usually yielding negative profits. As an illustration,
low-volatility stocks from the first group generate one of the worst performances with an
annualized value of -10.7%. In summary, unconditional on the timing specification, the
first sub-period favours morning winners.
Second, the period following April 2001 sees morning losers come ahead. Under last,
18
Figure A2 plots monthly cumulated returns earned by equal-weighted portfolios formed on past per-
formance. The structural break is particularly visible for the afternoon specification on the win and lose
portfolios.

27
while returns are generally lower than during the first sub-period, the U-shaped pattern
morphs into a left-smirk. However, this transition comes from morning losers losing upside
potential as e.g. high-volatility stocks settle onto a new low of 9.2% from a previous value
of 31.9% whereas the annualized profit of morning losers, from the same stock category,
is essentially unchanged at 16.3%. The transition is even more pronounced under the
afternoon specification. If earlier, the shape of average annualized profits was that of
a right-smirk, returns are now monotonically decreasing with the win portfolios yielding
between -3.5 and 1.5% (the sole positive estimate).
To conclude, while the U-shaped pattern is strong under last and before April 2001,
it becomes a left-smirk when sorting stocks by std in the second sub-period. Additional
investigations on the reason of this behaviour are warranted. Moreover, the (weaker)
parabolic profile under afternoon in table 5 is the combination (average) of the right-smirk
and decreasing monotonic patterns from the two estimation intervals.

28
Table 6: sub-period analysis of average annualized returns earned by equal-weighted portfolios indepen-
dently formed on past performance and stock characteristics. The sorts are formed as in table 5. The
sample period is split at the end of April 2001. Definition of characteristics: size is the logarithm of the
previous-month market capitalization, std is the average of the daily realized volatility and skew is the
average of the daily realized skewness from Amaya et al. (2015). Average values are computed over the past
quarter and rolling on a daily basis. Realized measure are calculated with intraday returns sampled at the
5-minute frequency. The table is organized in two panels. In the left panel is the last timing specification,
with formation period from 9:30 until 12:00 and holding period from 15:30 to 16:00, and in the right panel
is the afternoon timing specification, with formation period from 9:30 until 13:00 and holding period from
13:30 to 15:30. Returns are expressed in percentage. The sample consists of all common NYSE stocks
excluding microcaps, i.e. stocks with a price below 5$ or market capitalization in the bottom decile, and
covers the period from January 1993 to May 2010.

last afternoon
lose 2 3 4 win lose 2 3 4 win

Pre April 2001

size
small 20.7 17.1 17.4 18.6 34.4 -1.7 -15.9 -4.8 1.0 12.6
medium 12.5 8.8 7.9 8.5 18.3 -3.7 -9.0 -5.2 2.4 10.0
large 9.2 5.8 3.2 2.6 8.5 -0.1 -3.1 -1.9 3.5 9.1
std
low 12.4 6.0 3.0 1.0 1.5 12.5 -0.8 -1.1 2.7 1.8
medium 11.4 8.4 7.8 8.1 14.7 -2.6 -9.9 -3.0 2.3 11.8
high 16.5 15.6 17.2 19.1 31.9 -10.7 -17.8 -14.3 -2.6 11.0
skew
low 15.1 10.6 8.8 8.7 22.5 -6.3 -10.8 -8.2 2.4 8.0
medium 13.8 9.3 8.2 9.5 19.8 -3.4 -8.6 -3.9 0.0 12.0
high 12.3 7.9 6.2 6.4 15.2 -3.3 -7.9 -2.1 3.1 11.6
Post April 2001
size
small 17.6 9.5 8.4 8.4 9.6 21.7 8.1 6.6 6.5 1.5
medium 6.9 3.9 3.2 2.9 4.0 20.3 10.2 6.4 3.9 -3.5
large 5.7 3.4 2.3 1.2 2.3 14.5 8.1 6.2 4.3 -0.1
std
low 3.8 1.9 1.1 0.5 0.7 21.4 10.5 6.7 3.4 -1.9
medium 7.7 4.3 3.4 2.6 2.3 23.3 10.6 6.3 3.7 -2.6
high 16.3 10.8 10.1 9.0 9.2 17.6 5.1 5.5 6.0 -0.5
skew
low 10.8 5.2 4.5 4.4 5.6 18.8 9.0 7.1 5.3 -1.0
medium 9.4 4.4 4.0 3.6 5.4 18.9 8.5 6.0 3.9 -1.2
high 10.0 4.8 3.7 3.4 5.6 20.7 8.6 5.8 4.7 -0.9

29
4.3 Day of the week

French (1980) finds that returns are lower on Monday than on other days of the week.
While I have investigated the impact of stock characteristics and the period of analysis
on stocks sorted by morning performance, the substantial difference documented across
days of the week might constitute a relevant driver of the patterns observed so far. In this
section, I investigate the difference in returns by day of the week and by sub-period and
find that the U-shaped pattern is not caused by a specific day.
Table 7 reports average annualized returns earned by equal-weighted portfolios formed
on past performance and grouped by day of the week. Every day, stocks are sorted into
five groups based on their performance during the intraday formation period (see last and
afternoon specifications). Then, five equal-weighted portfolios are formed and their return
is calculated over the holding period and reported for each day of the week. The analysis
is carried out on the whole horizon and on the two sub-periods created by splitting the
sample at the end of April 2001. The t-statistics are listed in table A5.
The estimates on the whole sample are consistent with the weekend effect. Under last,
Monday returns are economically lower with e.g. the lose portfolio recording a mere 5.5%
while the same group on Friday scores 19.2%, in annualized terms. For the afternoon
specification, the phenomenon is even stronger since the equal-weighted portfolios yield
negative estimates between -1.5 and -9.1%, while on the other days, performance is always
positive. The pattern in average profits is parabolic on all days during the last half-hour and
for afternoon from Wednesday onwards, Monday being the biggest exception with negative
but increasing returns. This is in line with the findings from the previous sections.
The estimates for the days of the week combined with the sub-period analysis are
consistent with the observations from section 4.2. For example, average returns for the
afternoon configuration change in shape from a right-smirk to a monotonically decreasing
pattern. These results from the previous section simply combine with the Monday effect.

30
Table 7: average annualized returns earned by equal-weighted portfolios formed on past performance and
grouped by day of the week. Every day, stocks are sorted into five groups based on their performance
during the intraday formation period. Then, five equal-weighted portfolios are formed and their return is
calculated over the holding period and reported for each day of the week. The analysis is carried out on
the whole horizon and on the two sub-periods created by splitting the sample at the end of April 2001. The
table is organized in two panels. In the left panel is the last timing specification, with formation period
from 9:30 until 12:00 and holding period from 15:30 to 16:00, and in the right panel is the afternoon timing
specification, with formation period from 9:30 until 13:00 and holding period from 13:30 to 15:30. Returns
are expressed in percentage. The sample consists of all common NYSE stocks excluding microcaps, i.e.
stocks with a price below 5$ or market capitalization in the bottom decile, and covers the period from
January 1993 to May 2010.

last afternoon
lose 2 3 4 win lose 2 3 4 win

Whole sample
Mon 5.5 3.6 3.2 3.7 10.6 -7.8 -9.1 -4.5 -1.5 -2.5
Tue 10.2 6.2 5.8 6.4 13.6 9.8 0.8 1.8 4.2 3.5
Wed 14.4 7.6 5.7 6.2 12.7 12.4 1.5 3.3 7.6 12.4
Thu 14.8 8.5 7.9 7.8 15.7 9.4 2.3 1.8 4.6 7.1
Fry 19.2 10.6 8.4 8.0 16.4 19.7 4.5 3.9 4.4 5.9
Pre April 2001
Mon 1.4 3.6 3.3 5.2 17.5 -18.3 -15.3 -6.2 1.4 8.1
Tue 11.4 7.3 6.9 8.3 19.4 0.4 -9.9 -6.3 -1.1 3.8
Wed 20.0 12.7 9.7 10.3 22.6 -0.9 -8.9 -2.6 7.5 20.6
Thu 17.6 10.3 9.4 10.2 24.3 0.5 -3.9 -2.6 5.0 15.1
Fry 24.8 14.8 12.2 11.3 27.2 4.8 -8.3 -2.8 2.0 11.4
Post April 2001
Mon 9.2 3.7 3.1 2.4 4.3 1.9 -3.3 -3.1 -4.2 -12.2
Tue 9.0 5.1 4.7 4.7 8.1 18.5 10.8 9.4 9.2 3.2
Wed 9.3 2.9 2.1 2.3 3.6 24.7 11.2 8.8 7.6 4.8
Thu 12.1 6.9 6.4 5.7 7.8 17.6 8.0 5.8 4.2 -0.3
Fry 14.1 6.8 5.0 4.9 6.4 33.3 16.3 10.1 6.5 0.9

31
As for the whole period, Mondays report lower returns across all performance groups under
afternoon, and for the win group during the first sub-period and for the lose group after
April 2001 under last.

The evidence does not suggest that the U-shaped pattern concentrates on a particular
day of the week.

4.4 Overnight return and alternative specifications

In this section, I consider alternatives to the last and afternoon specifications introduced in
section 3.2. First, I assess how the inclusion in the signal of the overnight return affects the
performance of the cross-sectional equal-weighted portfolios. Second, I introduce variations
in the formation and holding periods by e.g. excluding the first half-hour from the signal,
or by investing only during the second-last half-hour.

Overnight return. Rogalski (1984) and Harris (1986) give an early account about the
relevance of the overnight return in the context of the weekend effect. More recently, Kelly
and Clark (2011) estimate the magnitude of investing during trading hours against the
non-trading periods and find the latter case to be overwhelmingly more profitable than
the former, and Lou et al. (2016) find similar evidence in an application to momentum
strategies. Moreover, the unconditional breakdown of the day into half-hours and the
overnight component in figure 1, confirms these findings. Hence, it is natural to ask how
the cross-sectional sorts are affected when last and afternoon compound the overnight
return into their definition of morning signals.
Figure 3 plots monthly cumulated returns earned by equal-weighted quintile portfolios
formed on past performance with (dashed line) or without the overnight return (solid line).
For clarity, I only show the bottom and top quintiles. The last and afternoon configurations
are respectively displayed in the top and bottom axes.
The impact of the overnight return on the cross-sectional sorts is unequivocal and

32
 1 1
1 1
1 1
1
5 5 5 5 5
1 5 5 1 1
5 1 1 1
1 1 1
1 5 5 5 5 5 5 5
1 1 5
1 5 5 5
1 5 1
5
1 1 1 1 5
1
5 5 5 5 5

1
1
1 1 1
5 5 5 1
5 5 5 1 5 5 5 5 5
5 5 5 5 5 1
1 5 5 1
5 5 5
5 1 1
5 1 1 1
1
5 1
1 1
1 1 1 1
1 1
1 1

Figure 3: monthly cumulated returns earned by equal-weighted quintile portfolios formed on past perfor-
mance including the overnight return (dashed lines) and without it (solid lines). Only the bottom and top
quintiles are displayed. Shaded areas denote NBER recessions. Specifically, stocks are sorted every day
into five equally populated groups based on the performance over the intraday formation period plus the
eventual overnight component. Then, the return of equal-weighted portfolios is calculated for the holding
period. In the top plot is the last timing specification, with formation period from 9:30 until 12:00 and
holding period from 15:30 to 16:00, and in the bottom plot is the afternoon timing specification, with
formation period from 9:30 until 13:00 and holding period from 13:30 to 15:30. The sample consists of all
common NYSE stocks excluding microcaps, i.e. stocks with a price below 5$ or market capitalization in
the bottom decile, and covers the period from January 1993 to May 2010.

independent of the specification: the returns of the equal-weighted portfolios are amplified
in a way which makes trading on reversal profitable. The lose portfolios (first quintiles)
average at 5.1 and 3.5 bp per day on the last and afternoon specifications without the
overnight component (solid lines) but improve by about 50% to 8.6 and 5.3 bp when the
close-to-open return is included in the signals (dashed lines). The win portfolios (fifth
quintiles), on the other hand, average at 5.4 and 2.1 bp per day on the two intraday
specifications and worsen on average by 45% to 2.6 and 1.3 bp after the inclusion of
the overnight return. Hence, the inclusion of the overnight return in the signal creates
the opportunity for a cross-sectional reversal where a losers-minus-winners strategy would
yield between 6-7 bp per day, whether positions are held from 1:30 to 15:30 or only during

33
the last half-hour of trading.

Table 8: correlations between timing strategies of winners-minus-losers (WML) and cross-sectional quin-
tile portfolios based on alternative intraday specifications. Alternative holding and formation periods
are defined in the table, where last and afternoon correspond to point A) and E). The vwap variations
define buy and sell prices as 5-minute volume-weighted averages. WML portfolios are formed for each
specification as in table 2, and pairwise correlations between different strategies are reported under the
time-series panel. Likewise, five cross-sectional portfolios are formed for each specification as in table 3,
and five correlation matrices are calculated between alternative specifications of the same quintile. Then,
the average correlation is displayed under the cross-sectional panel. Coefficients above 0.45 are highlighted
in green. The sample consists of all common NYSE stocks excluding microcaps, i.e. stocks with a price
below 5$ or market capitalization in the bottom decile, and covers the period from January 1993 to May
2010.

signal hpr
A) 9:30 - 12:00 15:30 - 16:00
B) 9:30 - 12:00 15:30 - 16:00 vwap
C) 10:30 - 13:00 15:30 - 16:00
D) 9:30 - 10:00 15:30 - 16:00
E) 9:30 - 13:00 13:30 - 15:30
F) 9:30 - 13:00 13:30 - 15:30 vwap
G) 9:30 - 13:00 15:00 - 15:30
H) 10:30 - 13:00 15:00 - 15:30
time-series
A) B) C) D) E) F) G) H)
A) 1.00
B) 0.87 1.00
C) 0.53 0.46 1.00
D) 0.66 0.57 0.20 1.00
E) 0.18 0.18 0.13 0.13 1.00
F) 0.21 0.20 0.16 0.14 0.94 1.00
G) 0.15 0.17 0.10 0.11 0.65 0.64 1.00
H) 0.10 0.11 0.06 0.05 0.46 0.46 0.63 1.00
cross-sectional
A) B) C) D) E) F) G) H)
A) 1.00
B) 0.96 1.00
C) 0.96 0.93 1.00
D) 0.98 0.94 0.96 1.00
E) 0.21 0.17 0.21 0.21 1.00
F) 0.20 0.17 0.20 0.20 0.98 1.00
G) 0.23 0.19 0.23 0.23 0.63 0.61 1.00
H) 0.23 0.18 0.22 0.23 0.62 0.60 0.98 1.00

Alternative formation and holding periods. This section builds on the predictability

34
patterns from section 3 and defines additional intraday specifications as variations to the
commonly used last and afternoon. To recall, the former defines the formation period
from 9:30 until 12:00 and the holding period from 15:30 to 16:00. The latter defines the
formation period from 9:30 until 13:00 and the holding period from 13:30 to 15:30.
In modifications to last, the formation period is adjusted to the interval from 10:30 to
13:00, or to only include the first half-hour as in Gao et al. (2015), while the holding period
is kept unchanged in both variants. As an alternative to afternoon, its holding period
is shortened to the second-last half-hour, and in one case the formation period remains
unchanged while in the other I use the 10:30 to 13:00 interval (excludes first half-hour).
Table 8 reports correlations between timing strategies of winners-minus-losers (WML)
and cross-sectional quintile portfolios based on alternative intraday specifications. WML
portfolios are formed for each specification as in table 2, and pairwise correlations between
different strategies are reported under the time-series panel. Likewise, five cross-sectional
portfolios are formed for each specification as in table 3, and five correlation matrices
are calculated between alternative specifications of the same quintile. Then, the average
correlation across quintiles is displayed under the cross-sectional panel.
Conclusions are similar for both time-series and cross-sectional portfolios. As expected,
high coefficients cluster around the last and afternoon specifications showing that alterna-
tive intraday strategies are only marginally different from the base cases. This is especially
true for the cross-sectional sorts, where all variations of the last configuration correlate
above 0.90 among each other, while estimates from alternatives to afternoon all score
above 0.60. On the other hand, (cross) combinations of last and afternoon variants record
values in the low 0.20 or below.

4.5 Cost analysis

This section assesses the impact of costs on the gross average returns of cross-sectional
portfolios like those reported in table 3.

35
Many anomalies disappear after considering transaction costs, especially when turnover
is above 50% (see Novy-Marx and Velikov, 2015). In the context of intraday strategies,
the turnover is 100% since positions are closed within the day, which poses the question
of the economic feasibility of the previously documented U-pattern in returns. I use a
cost-minimizing approach similar to Novy-Marx and Velikov (2015) and show that the
average return of quintile portfolios sorted by past performance is generally negative after
taking into account costs (consistent with Heston et al., 2010) with an exception. The
5% of stocks with the lowest ex-ante costs exhibit statistically positive returns under the
afternoon specification and after the decimalization in 2001.
Table 9 reports daily average returns, gross and net of costs, earned by equal-weighted
quintile portfolios formed on past performance and an equal-weighted portfolio of all stocks.
Estimates are reported for all stocks and for the group with the lowest 5% bid-ask spread
relative to the midpoint on the previous close (ba spread). The table also reports the
standard-bandwidth Newey-West standard error of the average net return,the averages of
the estimated ba spread (cost) and the stock count per group. The analysis is applied for
the whole sample and for the period after April 2001.
By considering the whole sample period (first row), the average bid-ask spread for the
whole cross-section is about 79 bp and varies between 67 and 82 bp for portfolios formed
on past performance. On the other hand, the highest average return is earned by the win
portfolio of the last specification and amounts to 4.8 bp only. Hence, roundtrip costs are
much higher than average returns and this is also true if we restrict the analysis to the
5% of the cheapest stocks, i.e. those with the lowest ex-ante bid-ask spread. In the latter
case, the average cost is much lower and stable at around 11 bp, but the best performance,
again by the win portfolio of the last specification, is less than half the ba spread at 3.8 bp
per day.
The situation changes a bit if we only consider the post-decimalization period, i.e.
after April 2001. Consistent with literature (e.g. Bessembinder, 2003), bid-ask spreads are

36
Table 9: daily average returns, gross and net of costs, earned by equal-weighted quintile portfolios formed
on past performance and an equal-weighted portfolio of all stocks. Estimates are reported for all stocks and
for the group with the lowest 5% bid-ask spread relative to the midpoint from previous close (ba spread).
The table reports averages of gross portfolio returns, the estimated ba spread (cost), portfolio returns net
of costs and its standard-bandwidth Newey-West standard errors in parentheses, and the count of stocks
per group. The analysis is applied for the whole sample and for the period after April 2001. The table
is organized in two panels. In the left panel is the last timing specification, with formation period from
9:30 until 12:00 and holding period from 15:30 to 16:00, and in the right panel is the afternoon timing
specification, with formation period from 9:30 until 13:00 and holding period from 13:30 to 15:30. Returns
are expressed in percentage. The sample consists of all common NYSE stocks excluding microcaps, i.e.
stocks with a price below 5$ or market capitalization in the bottom decile, and covers the period from
January 1993 to May 2010.

last afternoon
lose 2 3 4 win all lose 2 3 4 win all

Whole Sample
ba spread all
gross ret. 0.048 0.029 0.024 0.024 0.051 0.037 0.029 -0.003 0.002 0.013 0.019 0.016
cost 0.821 0.686 0.672 0.709 0.808 0.785 0.822 0.693 0.680 0.706 0.806 0.785
net ret. -0.773 -0.657 -0.648 -0.685 -0.757 -0.747 -0.793 -0.696 -0.677 -0.693 -0.787 -0.769
[0.029] [0.023] [0.022] [0.024] [0.027] [0.027] [0.032] [0.026] [0.025] [0.025] [0.028] [0.028]
count 331 325 327 342 333 1658 331 327 328 339 333 1658
ba spread low 5%
gross ret. 0.008 0.015 0.013 0.012 0.038 0.037 -0.006 -0.016 0.001 0.021 0.036 0.016
cost 0.105 0.104 0.104 0.105 0.104 0.108 0.105 0.104 0.104 0.105 0.104 0.108
net ret. -0.097 -0.089 -0.090 -0.093 -0.067 -0.070 -0.111 -0.120 -0.103 -0.083 -0.068 -0.092
[0.011] [0.009] [0.008] [0.008] [0.010] [0.008] [0.015] [0.012] [0.011] [0.011] [0.013] [0.010]
count 15 17 17 17 15 81 15 17 17 17 15 81
Post April 2001
ba spread all
gross ret. 0.040 0.019 0.016 0.014 0.021 0.024 0.075 0.034 0.025 0.018 -0.005 0.031
cost 0.303 0.273 0.265 0.270 0.302 0.289 0.304 0.272 0.265 0.269 0.302 0.289
net ret. -0.264 -0.253 -0.250 -0.255 -0.281 -0.265 -0.229 -0.238 -0.241 -0.252 -0.307 -0.259
[0.016] [0.015] [0.015] [0.015] [0.016] [0.015] [0.018] [0.018] [0.017] [0.017] [0.019] [0.018]
count 419 419 420 420 420 2097 419 419 419 420 420 2097
ba spread low 5%
gross ret. 0.017 0.013 0.008 0.005 0.010 0.024 0.057 0.025 0.022 0.021 0.003 0.031
cost 0.012 0.013 0.013 0.013 0.012 0.012 0.012 0.013 0.014 0.013 0.012 0.012
net ret. 0.006 0.000 -0.005 -0.008 -0.002 0.011 0.045 0.012 0.009 0.008 -0.009 0.018
[0.009] [0.008] [0.007] [0.007] [0.008] [0.007] [0.015] [0.012] [0.011] [0.011] [0.014] [0.011]
count 18 21 22 22 19 102 18 21 22 22 18 102

reportedly lower during this period when compared to the whole sample, i.e. estimates for
the whole cross-section are at 29 bp against the 79 bp over the whole horizon. Yet, only
the 5% of the cheapest stocks show some positive net returns. For example, let us consider
the last configuration: the lose portfolio and the average of all stocks earn respectively

37
0.6 and 1.1 bp, but the estimates are statistically not significant. Surprisingly, the lose
portfolio of the afternoon specification reports a statistically significant net return of 4.5
bp. Yet, this result should be seen as an exception and the low number of stocks in the
group, 18 on average, warrants caution in the interpretation.

5 Conclusion and further research

Literature on intraday predictability in returns is limited to a cross-sectional contribution

by Heston et al. (2010) and a time-series investigation by Gao et al. (2015). Motivated by
this scarcity of evidence I provide new and complimentary findings by using US common
stocks data from 1993 until 2010. I divide the trading day in half-hours and study how
returns in the afternoon are predicted by those in the morning. I find similar patterns
of statistical predictability to those in Gao et al. (2015), i.e. morning returns anticipate
continuation in the last half-hour, and I extend them with additional evidence of reversal
in the afternoon. However, the economic profitability of timing strategies based on such
periodicities turns out to be inconsistent with the statistical estimates. Moreover, the
performance of these strategies over time suggests the existence of systematic differences
in the cross-section.
By looking at the average return of ten cross-sectional portfolios formed on morning
performance, I uncover a U-shaped pattern in average returns where stocks that lost or
gained the most in the morning, earn higher and positive returns in the second half of the
trading day, and especially during the last half-hour. For example, stocks that ranked in
the bottom and top decile, respectively earn in the last half-hour 6.2 and 7.7 bp per day
or about 15.6 and 19.4% in annualized terms. The finding is fundamentally different from
the evidence documented by Heston et al. (2010), and is robust to: stock characteristics (I
consider size, traded volume, illiquidity, tick size, volatility and skewness), the day-of-week
effect, and variations to formation and holding periods. However, the pattern shows some

38
dependence on the period of analysis. Specifically, morning winners are more profitable
than morning losers before the decimalization in 2001, but these circumstances are reversed
afterwards.
What is causing the cross-sectional U-shaped pattern in average returns during the last
half-hour of trading remains an open question. I conjecture two interrelated effects. First,
institutional trading by e.g. indexed funds might induce significant positive price pressure,
and second, trading mechanisms, such as the closing auction, and/or limits to speculative
short sales and excessive depreciations establish a floor on the losses and leave the upside
potential untouched. This latter effect would then be responsible for the differences ob-
served in the average returns of the cross-section. There is established literature for most
of the elements in the conjecture. For example, the last minutes of trading account for
most of the daily return (Cushing and Madhavan, 2000) and this phenomenon is consis-
tent with institutional investors strategically timing their activity during periods of high
volume (Admati and Pfleiderer, 1988), i.e. at the open and close of the session. Moreover,
Edelen and Warner (2001) document that institutional order imbalances bear impact on
contemporaneous returns. Additionally, Amihud and Mendelson (1987) document a sig-
nificant effect of trading mechanisms, like the opening auction, on the stock price. More
research is needed to establish whether there is a positive price impact by institutional
investors during the last half-hour of trading and whether market frictions and/or the clo-
sing auction play a role in the superior appreciation of morning losers. The decimalization
of stock quotes in 2001 might prove to be relevant to the latter. Understanding the impact
of decimalization on the U-shaped cross-sectional pattern might also improve statistical
predictability by allowing the inclusion of regime switches, as Pesaran and Timmermann
(1995) suggest.

39
Appendices

Appendix A Additional tables and figures

1994
20
1998

2002 0

2006 −20

2010
Ron 9:30 10:30 11:30 12:30 13:30 14:30 15:30
80
1994

1998 40
2002
0
2006

2010 −40
Ron 9:30 10:30 11:30 12:30 13:30 14:30 15:30

Figure A1: annualized cross-sectional average of half-hour returns over time. The trading day is par-
titioned in 13 half-hours and, at each point in time, the cross-sectional average of returns is calculated
for each sub-period and the overnight time. The top plot focuses on intraday averages by dropping the
overnight return estimate. The time on the x-axis marks the start of a half-hour. Values are expressed
in percentage and darker colours correspond to lower returns. The sample consists of all common NYSE
stocks excluding microcaps, i.e. stocks with a price below 5$ or market capitalization in the bottom decile,
and covers the period from January 1993 to May 2010.

40
 5 5 5 5 5
5 1
5 1 1
5 1 1
1
1
5
5 1
5 1
5 1
5 1 1
1
5 1

5 5 1
5 5 5 5 5 5
5 5 1
5 1
5 5 1 1
1
1
5 1 1 1
1 1
1 1

Figure A2: monthly cumulated returns earned by equal-weighted quintile portfolios formed on past
performance. Shaded areas denote NBER recessions. Specifically, stocks are sorted every day into five
equally populated groups based on the performance over the intraday formation period. Then, the return
of equal-weighted portfolios is calculated for the holding period. For clarity, only the first (lose) and the
fifth (win) quintiles are marked. In the top plot is the last timing specification, with formation period
from 9:30 until 12:00 and holding period from 15:30 to 16:00, and in the bottom plot is the afternoon
timing specification, with formation period from 9:30 until 13:00 and holding period from 13:30 to 15:30.
The sample consists of all common NYSE stocks excluding microcaps, i.e. stocks with a price below 5$ or
market capitalization in the bottom decile, and covers the period from January 1993 to May 2010.

41
Table A1: t-statistics from time-series predictability regressions. Each panel assesses the robustness of
the base case (see table 1) with respect to single changes in methodology or in the period of analysis.
In this order: A) univariate equivalent of the equation (1); B) regressions with non-standardized returns;
C) include all previous half-hour in the RHS of the regression; D) include the overnight return as a
separate regressor; E) excludes period of the financial crisis from December 2007 until May 2009; F) uses
sample until April 2001 (pre-decimalization); G) uses sample after April 2001 (post-decimalization); H)
uses periods from the beginning until end of 2000, then from May 2002 until June 2007 and from start of
2009 until May 2010 (bull markets); I) uses periods from 2000 until May 2002 and from June 2007 until
end of 2009 (bear markets).

A) univariate
c 09:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30
15:30 2.17 1.76 1.71 0.75 2.42 1.17 -1.36 -0.03 -0.06 0.83 1.46
15:00 -2.04 -0.97 -2.66 -0.82 -0.29 -1.95 -1.97 0.47 -0.63 0.96
14:30 0.00 -1.00 1.42 0.94 0.61 0.62 0.51 2.43 0.64
14:00 -2.01 -0.70 -0.68 -0.47 -0.71 -0.32 -0.31 0.84
13:30 -0.12 0.39 -1.37 -2.10 -1.30 -0.66 -0.96
13:00 1.53 0.06 0.33 0.21 -1.66 -0.74
12:30 0.66 0.19 0.69 -1.03 -0.68
12:00 0.82 1.07 1.26 1.32
11:30 0.76 -1.43 1.88
11:00 1.05 -1.77
10:30 -2.08
B) no standardization
c 09:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30
15:30 6.10 1.47 1.80 2.28 1.93 3.52 1.08 -1.42 0.99 1.04 2.00 2.02
15:00 1.55 -0.09 0.93 -1.75 1.37 0.17 -0.48 -1.34 0.67 0.45 1.26
14:30 2.54 1.87 -0.69 1.10 -0.15 -0.63 -0.19 1.09 0.94 -1.75
14:00 -0.57 -0.67 0.32 -1.15 0.01 -1.38 -1.84 -1.47 -1.03
13:30 0.59 -0.90 -0.32 -1.91 -2.26 -1.29 -1.24 -2.06
13:00 2.07 1.88 1.20 -0.02 -0.13 -1.48 -1.66
12:30 1.76 1.95 0.08 0.78 -1.15 -0.12
12:00 0.32 1.57 0.66 -0.07 -0.55
11:30 0.20 0.77 -1.92 -0.46
11:00 -1.74 0.41 -3.87
10:30 0.06 -1.34 continues...

42
...continues from table A1
C) no skip
c 09:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30 15:30
15:30 3.53 2.05 1.81 1.53 0.71 2.37 0.99 -1.44 -0.07 -0.05 0.92 1.03 -1.35
15:00 2.08 -2.04 -1.16 -2.64 -0.84 -0.30 -2.02 -2.03 0.40 -0.56 0.78 -3.85
14:30 2.55 -0.09 -0.94 1.36 0.95 0.62 0.66 0.68 2.49 0.59 -3.44
14:00 -0.86 -2.02 -0.75 -0.77 -0.56 -0.75 -0.36 -0.33 0.41 -4.23
13:30 0.23 -0.01 0.30 -1.34 -2.10 -1.41 -0.80 -1.55 -1.89
13:00 1.88 1.59 0.10 0.40 0.16 -1.73 -1.28 -8.96
12:30 2.50 0.71 0.26 0.75 -0.96 -0.93 -3.08
12:00 0.97 0.87 1.12 1.37 1.32 -5.80
11:30 0.10 0.76 -1.33 1.79 -0.89
11:00 -0.96 0.97 -1.99 -4.37
10:30 -0.20 -2.29 -2.43
10:00 -2.61 -5.07
D) overnight
c Ron 09:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30
15:30 2.59 0.53 2.89 2.24 2.13 0.90 2.19 2.04 -0.93 -0.61 -0.18 0.89 1.66
15:00 0.96 0.41 -1.35 -0.61 -2.65 -0.12 -0.06 -1.47 -1.28 1.34 0.42 1.41
14:30 1.88 -0.24 -0.29 -0.75 1.32 1.34 1.05 0.27 0.10 1.92 0.68
14:00 -1.26 -0.84 -0.49 -0.54 -1.16 -0.54 0.04 0.08 -0.34 0.20
13:30 -0.34 0.48 -0.24 -0.14 -1.23 -1.80 -1.27 -0.28 -1.80
13:00 1.53 1.21 3.17 0.63 0.23 0.22 -2.15 -1.46
12:30 1.16 -0.14 -0.96 -0.92 -0.29 -2.27 -0.89
12:00 0.20 0.87 0.16 1.24 0.46 -1.04
11:30 0.72 2.56 0.50 -1.02 0.61
11:00 -0.92 2.10 2.25 -1.60
10:30 0.63 -1.10 -0.60
10:00 -1.56 -3.34
E) exclude crisis 2007-2009
c 09:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30
15:30 4.02 3.17 1.77 1.35 0.82 2.21 -0.06 -0.16 -0.53 -0.19 1.72 0.81
15:00 1.64 -2.38 -1.85 -1.09 -1.12 0.64 -2.29 -0.58 -0.52 -0.67 2.47
14:30 1.50 -0.98 -1.55 0.54 0.98 -0.74 -0.56 1.09 2.19 1.21
14:00 -0.77 -2.76 -1.21 0.39 -0.63 -0.72 0.98 0.00 0.86
13:30 0.95 -0.41 0.38 -0.97 -1.61 -0.84 -0.48 -1.05
13:00 1.70 1.35 0.43 0.91 -0.60 -0.77 -0.64
12:30 3.17 -0.71 0.79 -0.05 -1.57 -0.66
12:00 1.39 0.27 1.08 1.69 1.78
11:30 0.31 1.14 -1.07 1.82
11:00 -0.25 0.88 -1.02
10:30 -0.23 -3.04 continues...

43
...continues from table A1
F) pre-decimalization
c 09:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30
15:30 6.10 -1.47 0.06 -0.49 0.00 1.08 -0.73 -0.28 -1.07 -0.80 0.54 0.50
15:00 0.65 -1.96 -0.19 -0.28 2.12 -0.37 0.08 0.46 1.98 1.66 3.05
14:30 -0.11 -1.62 0.04 -0.27 -0.32 -0.26 -0.44 2.10 2.69 -0.68
14:00 -1.32 -0.43 1.58 -0.31 0.91 -0.74 1.21 0.05 -0.66
13:30 0.46 -1.41 -0.82 -1.42 -1.06 -0.75 -0.94 -1.37
13:00 1.13 1.72 1.05 1.21 -0.12 -1.42 0.81
12:30 -0.55 0.88 2.40 1.55 1.45 2.06
12:00 -1.20 2.44 2.00 2.44 3.60
11:30 -1.39 2.09 -0.12 3.01
11:00 -1.59 2.53 1.23
10:30 0.58 2.64
G) post-decimalization
c 09:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30
15:30 1.66 2.86 2.05 1.86 0.80 2.25 1.38 -1.30 0.08 0.10 0.78 1.41
15:00 1.83 -1.63 -1.18 -2.73 -1.47 -0.22 -2.12 -2.28 -0.07 -1.23 0.15
14:30 2.55 0.62 -1.03 1.61 1.10 0.69 0.82 -0.02 1.83 1.10
14:00 -0.53 -1.93 -1.33 -0.58 -0.67 -0.40 -0.67 -0.25 1.13
13:30 -0.09 0.31 0.44 -0.93 -1.94 -1.07 -0.56 -0.45
13:00 1.34 1.02 -0.29 0.03 0.27 -1.35 -0.98
12:30 2.60 0.44 -0.75 0.25 -1.44 -1.67
12:00 1.28 0.20 0.47 0.82 0.67
11:30 0.55 0.11 -1.41 1.03
11:00 -0.39 0.42 -2.46
10:30 -0.27 -3.20 continues...

44
 ...continues from table A1
H) Bull markets
c 09:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30
15:30 3.53 3.55 1.48 0.77 1.35 3.74 2.18 1.30 -1.72 0.21 1.21 0.98
15:00 0.98 -1.89 -0.36 0.11 -1.64 0.22 -2.40 0.15 0.10 -1.33 -0.37
14:30 1.96 0.17 -1.51 1.38 1.08 -0.22 -0.62 0.93 1.46 1.79
14:00 -0.16 -2.12 -1.05 0.91 -0.07 -0.34 1.20 0.93 1.11
13:30 0.70 -0.71 -0.41 -0.55 -0.84 -0.74 0.54 1.40
13:00 1.03 0.70 1.01 0.91 0.47 -0.57 -1.00
12:30 3.14 -1.19 0.17 -0.14 -1.13 -1.52
12:00 2.93 0.40 1.40 1.77 1.34
11:30 -0.17 2.03 0.02 2.19
11:00 0.77 0.92 -0.84
10:30 0.91 -3.64
I) Bear markets
c 09:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30
15:30 1.29 -0.19 1.06 1.22 -0.14 -0.02 0.25 -3.21 1.33 0.17 -0.32 1.62
15:00 1.47 -0.33 -0.04 -2.85 0.07 -1.19 -0.46 -2.88 0.72 -0.01 0.72
14:30 2.06 1.05 0.17 1.13 0.14 0.87 1.35 -0.62 1.46 -0.96
14:00 -1.27 -0.44 0.11 -2.32 -0.59 -0.82 -2.30 -0.74 -0.09
13:30 -1.30 0.16 0.39 -1.48 -1.90 -1.44 -1.31 -1.78
13:00 1.51 1.32 -0.59 -0.19 0.54 -1.43 0.06
12:30 0.39 2.49 -0.12 1.28 -0.07 0.55
12:00 -2.00 0.89 0.85 0.31 -0.07
11:30 0.28 -0.84 -1.72 0.28
11:00 -2.86 0.36 -2.84
10:30 -0.80 1.25 end

Table A2: R-squared coefficient from predictability regressions. The first column contains coefficients
from the base-case regressions of table 1 and the remaining column represent variants to the base-case as
outlined in table A1, with the exception of variant A), which is not reported for compactness.

B) C) D) E) F) G) H) I)
15:30 0.051 0.093 0.447 0.112 0.039 0.006 0.134 0.083 0.303
15:00 0.063 0.040 0.199 0.081 0.048 0.062 0.144 0.032 0.529
14:30 0.025 0.037 0.094 0.030 0.023 0.023 0.047 0.036 0.092
14:00 0.011 0.044 0.175 0.009 0.016 0.013 0.032 0.023 0.146
13:30 0.017 0.097 0.390 0.039 0.010 0.018 0.024 0.009 0.159
13:00 0.009 0.018 0.316 0.048 0.006 0.010 0.014 0.009 0.031
12:30 0.004 0.009 0.229 0.031 0.007 0.033 0.021 0.011 0.043
12:00 0.007 0.004 0.070 0.011 0.009 0.017 0.005 0.011 0.011
11:30 0.010 0.007 0.012 0.026 0.009 0.022 0.011 0.015 0.033
11:00 0.006 0.019 0.037 0.033 0.002 0.006 0.021 0.003 0.052
10:30 0.003 0.005 0.176 0.007 0.007 0.006 0.016 0.016 0.006
10:00 - - 0.023 0.053 - - - - -

45
Table A3: average counts of constituents from portfolios formed independently on past performance and
a stock characteristic. Complements table 5.

last afternoon
lose 2 3 4 win lose 2 3 4 win
size
small 156 100 93 105 148 157 101 93 103 149
medium 117 120 120 128 119 117 121 120 127 119
large 89 134 142 143 96 88 134 144 142 95
std
low 68 131 149 143 79 68 131 150 142 79
medium 114 114 109 118 116 114 115 109 117 116
high 161 90 79 92 149 161 91 79 91 149
skew
bottom 123 110 107 113 118 123 110 107 112 118
medium 113 112 113 119 113 113 113 114 118 113
top 106 114 117 122 112 106 114 117 121 112

46
Table A4: t-statistics for sub-period analysis of average annualized returns earned by equal-weighted
portfolios independently formed on past performance and stock characteristics. Complements table 6.

last afternoon
lose 2 3 4 win lose 2 3 4 win

Pre April 2001

size
small 8.6 8.5 8.8 10.0 15.6 -0.4 -5.3 -1.5 0.4 3.7
medium 6.0 6.4 6.7 7.0 9.9 -1.0 -3.6 -2.7 1.2 3.4
large 3.8 4.0 2.6 2.0 4.2 -0.0 -1.1 -1.0 1.8 3.2
std
low 6.5 5.2 2.9 0.9 1.2 3.7 -0.4 -0.7 1.6 0.9
medium 5.8 5.6 5.6 5.9 8.1 -0.7 -3.9 -1.4 1.1 4.2
high 6.3 6.6 7.4 8.8 12.9 -2.6 -5.1 -4.1 -0.8 2.7
skew
low 6.3 6.7 6.4 6.3 11.5 -1.7 -4.2 -3.8 1.1 2.5
medium 6.5 6.3 6.3 7.0 10.2 -0.8 -3.4 -2.0 0.0 4.0
high 5.7 5.4 5.0 5.0 7.9 -0.8 -3.1 -1.1 1.6 3.9
Post April 2001
size
small 7.1 4.1 3.9 3.8 3.9 6.0 2.7 2.3 2.1 0.4
medium 3.3 2.1 1.8 1.6 1.9 6.0 3.5 2.4 1.4 -1.1
large 3.0 2.0 1.5 0.8 1.3 4.4 3.1 2.4 1.6 -0.0
std
low 2.2 1.2 0.7 0.3 0.4 7.9 4.7 3.0 1.5 -0.7
medium 3.7 2.2 1.8 1.4 1.1 7.2 3.6 2.3 1.3 -0.8
high 6.5 4.6 4.5 4.1 3.7 4.5 1.5 1.6 1.7 -0.1
skew
low 4.9 2.7 2.5 2.3 2.7 5.5 3.2 2.6 1.9 -0.3
medium 4.3 2.3 2.2 2.1 2.6 5.5 3.0 2.2 1.4 -0.4
high 4.6 2.6 2.1 1.9 2.8 6.2 3.1 2.2 1.7 -0.3

47
Table A5: t-statistics for average annualized returns earned by equal-weighted portfolios formed on past
performance and grouped by day of the week. Complements table 7.

last afternoon
lose 2 3 4 win lose 2 3 4 win

Whole sample
Mon 1.3 1.2 1.2 1.3 2.9 -1.2 -1.8 -1.1 -0.4 -0.5
Tue 3.0 2.2 2.1 2.3 4.0 1.3 0.2 0.4 1.0 0.6
Wed 3.0 1.8 1.5 1.6 2.8 1.8 0.3 0.7 1.6 2.0
Thu 3.9 3.0 3.0 3.0 4.7 1.7 0.5 0.5 1.2 1.5
Fry 5.9 4.5 3.8 3.6 5.8 3.9 1.1 1.0 1.1 1.3
Pre April 2001
Mon 0.2 0.9 1.1 1.5 3.2 -1.8 -2.2 -1.1 0.3 1.1
Tue 2.7 2.5 2.7 3.1 5.3 0.0 -1.2 -1.2 -0.2 0.5
Wed 4.2 4.0 3.8 4.0 6.3 -0.1 -1.4 -0.5 1.4 2.3
Thu 3.3 2.8 3.0 3.1 5.3 0.1 -0.7 -0.6 1.1 2.5
Fry 5.9 5.3 5.0 4.7 8.6 0.8 -1.9 -0.7 0.5 2.0
Post April 2001
Mon 1.6 0.8 0.7 0.5 0.9 0.2 -0.5 -0.5 -0.7 -1.6
Tue 1.7 1.1 1.0 1.0 1.5 2.1 1.5 1.4 1.3 0.4
Wed 1.2 0.4 0.3 0.3 0.5 2.5 1.5 1.3 1.1 0.6
Thu 2.4 1.7 1.7 1.4 1.7 2.1 1.2 0.9 0.6 0.0
Fry 2.9 1.8 1.4 1.3 1.5 4.4 2.5 1.7 1.1 0.1

48
References
Admati, Anat R. and Paul Pfleiderer (1988). “A Theory of Intraday Patterns: Volume and
Price Variability”. In: The Review of Financial Studies 1.1, pp. 3–40.
Amaya, Diego et al. (2015). “Does realized skewness predict the cross-section of equity
returns?” In: Journal of Financial Economics 118.1, pp. 135–167.
Amihud, Yakov (2002). “Illiquidity and stock returns: cross-section and time-series effects”.
In: Journal of Financial Markets 5.1, pp. 31–56.
Amihud, Yakov and Haim Mendelson (1986). “Asset pricing and the bid-ask spread”. In:
Journal of Financial Economics 17.2, pp. 223–249.
— (1987). “Trading Mechanisms and Stock Returns: An Empirical Investigation”. In: The
Journal of Finance 42.3, pp. 533–553.
Angel, James J. (1997). “Tick Size, Share Prices, and Stock Splits”. In: The Journal of
Finance 52.2, pp. 655–681.
Ball, Clifford A. and Tarun Chordia (2001). “True Spreads and Equilibrium Prices”. In:
The Journal of Finance 56.5, pp. 1801–1835.
Banz, Rolf W. (1981). “The relationship between return and market value of common
stocks”. In: Journal of Financial Economics 9.1, pp. 3–18.
Barndorff-Nielsen, Ole E. et al. (2009). “Realized kernels in practice: Trades and quotes”.
In: Econometrics Journal 12.3, pp. 1–32.
Bessembinder, Hendrik (2003). “Trade Execution Costs and Market Quality after Decima-
lization”. In: The Journal of Financial and Quantitative Analysis 38.4, pp. 747–777.
Bollerslev, Tim et al. (2016). “Roughing up beta: Continuous versus discontinuous betas
and the cross section of expected stock returns”. In: Journal of Financial Economics
120.3, pp. 464–490.
Campbell, John Y. (1991). “A Variance Decomposition for Stock Returns”. In: The Eco-
nomic Journal 101.405, pp. 157–179.
Cochrane, John H. (2008). “The Dog That Did Not Bark: A Defense of Return Predicta-
bility”. In: Review of Financial Studies 21.4, pp. 1533–1575.
Cushing, David and Ananth Madhavan (2000). “Stock returns and trading at the close”.
In: Journal of Financial Markets 3.1, pp. 45–67.
Edelen, Roger M and Jerold B Warner (2001). “Aggregate price effects of institutional
trading: a study of mutual fund flow and market returns”. In: Journal of Financial
Economics 59.2, pp. 195–220.
Fama, Eugene F. and Kenneth R. French (1988). “Dividend yields and expected stock
returns”. In: Journal of Financial Economics 22.1, pp. 3–25.

49
French, Kenneth R. (1980). “Stock returns and the weekend effect”. In: Journal of Financial
Economics 8.1, pp. 55–69.
Gao, Lei et al. (2015). “Intraday momentum: The first half-hour return predicts the last
half-hour return”. Working paper.
Goldstein, Michael A. and Kenneth A. Kavajecz (2000). “Eighths, sixteenths, and mar-
ket depth: changes in tick size and liquidity provision on the NYSE”. In: Journal of
Financial Economics 56.1, pp. 125–149.
Harris, Lawrence (1986). “A transaction data study of weekly and intradaily patterns in
stock returns”. In: Journal of Financial Economics 16.1, pp. 99–117.
Heston, Steven L. et al. (2010). “Intraday Patterns in the Cross-section of Stock Returns”.
In: The Journal of Finance 65.4, pp. 1369–1407.
Jain, Prem C. and Gun-Ho Joh (1988). “The Dependence between Hourly Prices and
Trading Volume”. In: The Journal of Financial and Quantitative Analysis 23.3, pp. 269–
283.
Jegadeesh, Narasimhan and Sheridan Titman (1993). “Returns to Buying Winners and
Selling Losers: Implications for Stock Market Efficiency”. In: The Journal of Finance
48.1, pp. 65–91.
Kelly, Michael A. and Steven P. Clark (2011). “Returns in trading versus non-trading hours:
The difference is day and night”. In: Journal of Asset Management 12.2, pp. 132–145.
Komarov, Oleg (2016). “Low-Frequency Investment with High-Frequency Measures: Is it
Profitable?” Working paper.
Lou, Dong et al. (2016). “A Tug of War: Overnight Versus Intraday Expected Returns”.
Working paper.
Moskowitz, Tobias J. et al. (2012). “Time series momentum”. In: Journal of Financial
Economics 104.2, pp. 228–250.
Newey, Whitney K. and Kenneth D. West (1987). “A Simple, Positive Semi-Definite, He-
teroskedasticity and Autocorrelation Consistent Covariance Matrix”. In: Econometrica
55.3, pp. 703–708.
Novy-Marx, Robert and Mihail Velikov (2015). “A Taxonomy of Anomalies and Their
Trading Costs”. In: The Review of Financial Studies 29.1, pp. 104–147.
Pesaran, M. Hashem and Allan Timmermann (1995). “Predictability of Stock Returns:
Robustness and Economic Significance”. In: The Journal of Finance 50.4, pp. 1201–
1228.
Rogalski, Richard J. (1984). “New Findings Regarding Day-of-the-Week Returns over Tra-
ding and Non-Trading Periods: A Note”. In: The Journal of Finance 39.5, pp. 1603–
1614.

50
Roll, Richard (1984). “A Simple Implicit Measure of the Effective Bid-Ask Spread in an
Efficient Market”. In: The Journal of Finance 39.4, pp. 1127–1139.
Schultz, Paul (2000). “Stock Splits, Tick Size, and Sponsorship”. In: The Journal of Finance
55.1, pp. 429–450.
SEC (2000a). Order Directing the Exchanges and NASD To Submit a Decimalization Im-
plementation Plan. Commission Order and Notice Release No. 34-42360. Securities and
Exchange Commission.
— (2000b). Order Directing the Exchanges and NASD to Submit a Phase-In Plan to Imple-
ment Decimal Pricing In Equity Securities and Options. Commission Order and Notice
Release No. 34-42914. Securities and Exchange Commission.
— (2012). Report to Congress on Decimalization. JOBS Act Section 106. Securities and
Exchange Commission.
Wood, Robert A. et al. (1985). “An Investigation of Transactions Data for NYSE Stocks”.
In: The Journal of Finance 40.3, pp. 723–739.

51