Intraday Momentum: The First Half-Hour

Return Predicts the Last Half-Hour Return

Lei Gao
Iowa State University
and
Yufeng Han
University of Colorado at Denver
and
Guofu Zhou
Washington University in St. Louis∗

First Draft: March, 2014

Current Version: May, 2014

∗ We are grateful to Campbell Harvey, Matthew Ringgenberg, Ronnie Sadka, Robert Stambaugh,
and seminar participants at Washington University in St. Louis for very insightful and helpful
comments. Correspondence: Guofu Zhou, Olin School of Business, Washington University, St.
Louis, MO 63130; e-mail: zhou@wustl.edu, phone: 314-935-6384.

Electronic copy available at: http://ssrn.com/abstract=2440866

 Intraday Momentum: The First Half-Hour
Return Predicts the Last Half-Hour Return

Abstract

In this paper, using intra data from January 4, 1999 to December 31, 2012, we doc-
ument an intraday momentum pattern that the first half-hour return on the market
predicts the market return in the last half-hour. The predictability is both statisti-
cally and economically significant, and is stronger on high volatile days, recession days
and some macroeconomic news release days. We interpret the trading behavior of
daytraders and informed traders as the economic driving forces behind the intraday
momentum.

JEL Classification: G11, G14

Keywords: Predictability, Intra-day, Economic value

Electronic copy available at: http://ssrn.com/abstract=2440866

1. Introduction

Since the seminal work of Jegadeesh and Titman (1993), it is well-known that winners of
the past 6 months or a year tend to continue to be winners, and the losers tend to continue
to be losers. Griffin, Ji, and Martin (2003) show that such momentum is common in global
stock markets. Recently, Moskowitz, Ooi, and Pedersen (2012) and Asness, Moskowitz, and
Pedersen (2013) provide strong evidence that momentum is pervasive across asset classes
such as bond and currencies. However, to our knowledge, all momentum studies are confined
at the monthly frequency exception a couple on the weekly returns. The open question is
whether there is intraday momentum. This question is of interest not only for examining
robustness of momentum strategies, but also for understanding intraday market efficiency
and the role played by daytraders including in particular the high-frequency traders.

In this paper, we find strong evidence of market intraday momentum. The first half-
hour return on the market, as represented by the actively traded S&P 500 ETF, predicts
significantly the last half-hour return with an R2 of 2.0%, matching or exceeding the typical
predictive R2 ’s at the monthly frequency (see, e.g., Rapach and Zhou, 2013). If the first
half-hour return is combined with the twelfth half-hour return (the half-hour prior to the
last half-hour), the R2 increases further to 3.5%. In addition, we find that the predictability
rises generally with volatility. When the first half-hour return volatility is high, the R2
increases to 4.8% for the two combined predictors. Moreover, we find that the predictability
is stronger during recessions and on days for some major macroeconomic news releases.

On out-of-sample (OOS) predictability, the R2 is as high as 1.8% and 2.7%, with the use
of the single first half-hour return predictor and of the combined predictors, respectively.
Like in the in-sample results, the degree of OOS predictability is greater than that typically
found at the monthly frequency. In terms of economic significance, the predictability from
the two sets of predictors generates certainty equivalent annual returns of 7.23% and 7.61%,
respectively, for a mean-variance investor. In terms of marketing timing, the economic value
is also significant. Therefore, the intraday momentum is both statistically and economically
significant.1

What drives the intraday momentum? While there is a lack of theory at this time, we
1
Our paper here focuses on the market intraday momentum, while leaving the study of cross-section of
stocks for future research.

1
provide two explanations. The first is based on the trading behavior of daytraders. Most
major economic news, such as GDP and CPI, are released prior to 8:30 am Eastern time,
one hour before the stock market starts trading. There are in addition various overnight
news. Hence, when the first half-hour return is up substantially, it is likely due to some
good economic news. In responding to the price move, many of the daytraders may go short
to provide liquidity to the market. But they will almost surely unwind to go flat before
market closes. The findings of Shefrin and Statman (1985), Odean (1998), Locke and Mann
(2000), Coval and Shumway (2005), and Haigh and List (2005) all suggest that daytraders
may be subject to the disposition effect — they may be more reluctant to unwind losing
positions than winning ones. Thus, as many of them may wait to unwind in the last half-
hour, their trading is likely to cause higher prices. Our empirical evidence seems consistent
with this explanation. On a day when the first half-hour return is up substantially, the
twelfth half-hour return is on average slightly positive, making those who procrastinate to
unwind during this period to wait to do so in the last half-hour, making more unwinding at
the end. Moreover, the opening price the following day is on average lower, suggesting that
there is an adjustment of the price from the last half-hour buying pressure.

Our second explanation is based on the strategic trading of informed traders. It is a

well-known empirical fact that the trading volume has a U-shaped pattern. Heavy trading
occurs in the beginning and at the end of the trading day, and light trading happens in the
middle of the day (see, e.g., Jain and Joh, 1988). This is particularly true for the trading
activity on the S&P 500 ETF. Admati and Pfleiderer (1988) show theoretically that informed
traders will time their trades to high trading volume periods, or during the first and last
half-hours in our context. With a different preference specification, Hora (2006) also shows
that an optimal trading strategy is to trade rapidly at the beginning and at the end of the
trading horizon, and trade more slowly in the middle of the day. Therefore, given good
economic news in the first half-hour, the informed trader are likely to bid up the asset price
substantially. Then, in the last half-hour, their heavy buying is likely to continue to push
the price up further. Both of the above explanations explain the intraday momentum that
the market first half-hour return predicts the last half-hour return.

Our paper is related to the literature on intraday asset prices. Much of the existing
studies have been focused on trading activity and volatility (see, e.g., Chordia, Roll, and
Subrahmanyam, 2011; Corwin and Schultz, 2012). Heston, Korajczyk, and Sadka (2010)

2
seems the only study that is closely related to ours. They find a striking intraday pattern
that returns on individual stocks tend to continue at half-hour intervals across trading days,
and that this pattern can last up to 40 days. In contrast to their study, we analyze intraday
momentum, the predictability of the first half-hour return on the last half-hour return on
the same day.

Our paper is also related to the literature on price discovery. Barclay and Warner (1993),
Chakravarty (2001) and Boehmer and Wu (2013) study trading and traders of different types’
contribution to price discovery during the trading day and in longer horizon. Our study seems
to suggest that the price discovery process can take a day for the information integration,
resulting in the intraday momentum.

The rest of the paper is organized as follows. Section 2 provides a description of the data.
Section 3 documents the intraday momentum both in-sample and out-of-sample, in addi-
tion to their properties over volatility regimes. Section 4 provides an economic evaluation.
Section 5 investigates its behavior over macroeconomic regimes and news announcements.
Section 6 examine the robustness and Section 7 concludes.

2. Data

The intraday trading prices of actively traded SPDR S&P 500 ETF Trust (ticker SPY) are
taken from Trade and Quote database (TAQ) to compute the half-hour returns used in this
paper. The sample period is from January 4, 1999 to December 31, 2012. In our robustness
study, we also use the NASDAQ proxy, the PowerShares QQQ Trust (ticker QQQ, was
QQQQ before March 22, 2011). The sample period for the latter is from March 10, 1999 to
December 31, 2012. In our analysis of major news releases, we obtain the historical release
dates of the GDP estimate from Bureau of Economic Analysis, the historical release dates
of the CPI from Bureau of Labor Statistics, and the historical release dates of the Federal
Open Market Committee (FOMC) from the Federal Reserve.2

To examine the intraday return predictability, we calculate half-hour (30 minutes) returns
2
The website for historical GDP release is bea.gov/newsreleases/relsarchivegdp.htm, for Bureau of La-
bor Statistics announcements is www.bls.gov/bls/archived sched.htm, and for FOMC minutes release is
www.federalreserve.gov/monetarypolicy/fomccalendars.htm.

3
from 9:30 am to 4:00 pm Eastern time, a total of 13, from
pj
rj = − 1, j = 1, · · · , 13, (1)
pj−1
where pj are the prices at each cut off point of half-hour from 1 to 13, and pj−1 are the prices
at the previous cut off point of half-hours. Note that p0 is the previous trading day’s price
at the 13th half-hour (4:00 pm). That is, we use the previous trading day’s close price as
the starting price when calculating the first half-hour return on date t, so that the return
captures the impact of information since the previous trading day’s closing time. To assess
the impact of return volatility on return predictability, we also compute the volatility of the
first half-hour return as follows. We first calculate the returns minute by minute within the
first half-hour, and then estimate the standard deviation using the 30 one-minute returns.

3. Intraday Momentum

In this section, we run first predictive regressions to show intraday momentum, and examine
next the impact of volatility. Then we investigate out-of-sample performance. Final, we
provide two intuitive explanations.

3.1. Predictive regressions

Consider first the simple predictive regression of the last half-hour return on the first half-
hour return,

r13,t = α + βr1,t + ϵt , t = 1, · · · , T, (2)

where r13,t and r1,t are the last half-hour return and the first half-hour return on day t,
respectively, and T is the sample size or the total number of trading days.

The second column of Table 1 reports the results. The first half-hour return positively
predicts the last half-hour return with a slope of 0.077 that is statistically significant at
the 1% level. The R2 is 2%. This magnitude of R2 is impressive since almost all existing
predictors have lower R2 ’s (see, e.g., Rapach and Zhou, 2013).

Since the twelfth half-hour, the half-hour prior to the last half-hour, may affect the last
half-hour return too if there is any information carry over. The third column of Table 1

4
examines this issue. It does predicts the last half-hour return at the 1% level and has an
R2 of 1.5%. However, as shown next, its strong predictability largely comes from the most
recent financial crisis period. On the other hand, the predictability of the first half-hour
return is always significant.

Since both r1 and r12 predicts r13 individually, it is of interest to examine wether they can
predict r13 jointly. The fourth column of Table 1 reports the results. Surprisingly, the slopes
are little changed from their individual regression values. Moreover, the R2 , 3.5%, is equal
to the sum of the individual R2 ’s. The evidence suggests that r1 and r12 are complimentary
in forecasting the last half-hour return.

The standard monthly momentum strategy is known to have performed poorly during
the recent financial crisis, though it still beats the market. How well the intraday momentum
performs in this period, from January 1, 2006 to December 2009, is an interesting question.
Panel B of Table 1 reports the results. The predictive power of r1 becomes in fact stronger,
with a larger slope of 0.137 and a greater R2 of 4.3%. Moreover, the combined two predictors
yield an amazingly large R2 of 7.1%, rarely seen anywhere else. It may be noted that the
predictive power of r1 and r12 are complimentary during the crisis period too.

Since the performance during the crisis period is so remarkable, it is then a logical
question to what extent it affects the results of the whole sample period. Panel C of Table 1
addresses this question. Excluding those crisis days, the performance clearly becomes much
weaker. Although r12 is no longer significant, r1 remains a powerful predictor of r13 with
a sizable R2 of 0.8%, comparable to many good predictors at the monthly frequency. The
combined predictors yield a higher R2 of 1.3%. Therefore, though like many strategies the
predictability is not stable due to the financial crisis, there is no doubt for the validity of the
intraday momentum through the entire sample period.

3.2. Volatility

Since financial crisis is characterized by high volatility, earlier results during the crisis period
is a special case of how the intraday momentum performs under high volatility. In general,
we can examine the impact of volatility by sorting all the trading days into three equal
groups according to the first half-hour volatility, low, medium and high. For brevity, we
consider the case of joint predictors of r1 and r12 only.

5
 Table 2 reports the results. The predictability appears an increasing function of volatility.
When the volatility is low, the predictability is minimal with an R2 of 0.5%. When the
volatility is at intermediate level, the R2 rises only slightly to 0.6%. This magnitude of R2
is still economically significant. However, when the volatility is high, the R2 increases more
than nine times to as high as 4.8%.

Overall, the intraday momentum seems highly related to volatility. The higher the volatil-
ity, the greater the predictability. This appears consistent with the theoretical model of
Zhang (2006) that the greater the uncertainty, the greater the persistent of a trend. In our
context, the greater the likelihood of the first half-hour trend (up or down) carry over to the
last half-hour.

3.3. Out-of-sample recursive regression

Our previous intraday momentum analysis is based on the entire sample (in-sample) esti-
mation. While in-sample is econometrically more efficient if the regressions are stable over
time. But the financial crisis clearly destabilizes the estimation. At the monthly frequency, ?
find that many macroeconomic predictors suffer stability problems, and their predictability
is largely vanished once the predictive regressions are estimated recursively out-of-sample
(OOS). That is, to forecast return at any time t, only data up to time t − 1 is used.

To test whether intraday momentum exists OOS, we run recursive regressions, starting
the regression using the first year (1999) returns, and then progressively adding one more
year of returns each time. Following Campbell and Thompson (2008), Rapach, Strauss, and
Zhou (2010), and Neely, Rapach, Tu, and Zhou (2013), we use the OOS R2 to measure the
out-of-sample predictability, which is defined as,
∑T
(r13,t − r̂13,t )2
OOS R = 1 − ∑t=1
2
, (3)
t=1 (r13,t − r̄13,t )
T 2

where r̂13,t is the forecasted last half-hour return from the predictive regression estimated
through period t − 1, and r̄13,t is the historical average return estimated through period t − 1.
A positive OOS R2 indicates that the forecast beats the simple historical average.

Table 3 reports the results. When we use the first half-hour return alone, the OOS R2 is
1.76%. When we use the twelfth half-hour return alone, the OOS R2 is 1.01%. When we use
both of them, the OOS R2 achieves the highest value of 2.74%. The OOS R2 ’s are matching

6
or exceeding those at the monthly frequency. As shown by Campbell and Thompson (2008)
for monthly returns and confirmed later here, these levels of OOS R2 are of substantial
economic significance.

3.4. Explanations

Statistically, both the in- and out-of-sample analysis clearly provides strong evidence on
the intraday momentum. From an economic point of view, an interesting question is what
economic forces that drive it. We provide two intuitive explanations.

Our first explanation is based on the trading behavior of daytraders. On a day when
the first half-hour return is up substantially, which might be due to overnight or morning
news, some traders may expect price reversion and go short. Since they will almost surely
unwind to go flat before market closes, and some of them may wait to unwind in the last
half-hour. Due to the disposition effect (see, e.g., Shefrin and Statman, 1985; Odean, 1998;
Locke and Mann, 2000; Coval and Shumway, 2005; Haigh and List, 2005), they may be more
reluctant to unwind losing positions than winning ones. On the other hand, on the days
with substantial rise in price, the twelfth half-hour return is on average slightly positive,
making those who plan to unwind during this period to wait to do so in the last half-hour.
Therefore, there are likely even more unwinding the loser positions than usual in the last
half-hours. Collectively, their buying is likely to push the last half-hour return higher than
otherwise. Indeed, the opening price on the following day is on average lower, suggesting an
adjustment of the price from the last half-hour buying pressure.

Our second explanation is based on the strategic trading of informed traders. Admati
and Pfleiderer (1988) show theoretically that informed traders will time their trades to high
trading volume periods, and, with a different preference specification, Hora (2006) also shows
that an optimal trading strategy is to trade rapidly at the beginning and at the end of the
trading horizon, and trade more slowly in the middle of the day. Figure 1 plots the average
trading volume of the S&P 500 ETF every half hour. Both the first and the last half-hours
have trading volume close to 15 millions shares, but the middle of the day has only about
5 million shares. The plot has a perfect U-shape, consistent with earlier findings about
intraday trading activity (see, e.g., Jain and Joh, 1988). Now, based on the theories, given
good economic news the informed traders are likely to trade more actively in the first half

7
hour and thus bid up the price substantially. In the last half hour, their heavy buying is
likely to continue to push the price up further.

Both of the above explanations explain the intraday momentum that the market first
half-hour return predicts the last half-hour return. Clearly, our explanations are limited
in their scope. Future research on developing rigorous theories for understanding fully the
economic forces is called for.

4. Economic Significance

In this section, we explore the economic significance of intraday momentum. We first use the
first half-hour and twelfth half-hour returns as timing signals either individually or combined
to examine the performance relative to a passive strategy that always holds the market (SPY)
during the last half hour, and then use the predicted returns to assess the certainty equivalent
utility gains for a mean-variance investor.

4.1. Market timing

We use the first and twelfth half-hour returns as timing signals to time the market. Specif-
ically, we will take a long position of the market at the beginning of the last half hour if
the timing signal is positive, and take a short position otherwise. It is worth noting that
the position (long or short) is closed at the market close each trading day. Mathematically,
the marketing timing strategy based on signal r1 on day t will have a return in the last half
hour,
{ r , if r1 > 0;
13
η(r1 ) = (4)
−r13 , if r1 ≤ 0.
The formula is similar for signal r12 . In addition, we also consider the market timing strategy
based on both r1 and r12 . In this case, we buy only if both signals are positive, and sell
when both are negative. Otherwise, we stay out of the market. Mathematically, the return
is computed from
{ r13 , if r1 > 0 & r12 > 0;
η(r1 , r12 ) = −r13 , if r1 ≤ 0 & r12 ≤ 0; (5)
0, otherwise.

Panel A of Table 4 reports the summary statistics of returns generated from the three
timing strategies. When using the first half-hour return as the signal to trade in the last

8
half-hour, the average return is 6.34% on an annual basis. At a first glance, this does not
seem too large. To gauge the performance, we report two benchmark returns. The first is an
‘Always Long’ strategy that we always take a long position in the market at the beginning
of the last half hour and close it at the closing of the market. The first row of Panel B of
Table 4 shows that the annualized average return of this strategy is only −0.47%. Hence,
the timing strategy η(r1 ) outperforms this passive strategy substantially.

The second benchmark is the buy-and-hold strategy of the market that we simply take a
long position of the market from the beginning of the sample, and hold it till the end of the
whole sample period. The results are reported on the second row of Panel B. The average
return is only 2.66% per year, which is also less than the average return delivered by the
timing strategy, η(r1 ). Hence, it is remarkable considering that we are only in the market
for an half hour each trading day instead of six and half hours each day or all the time.

Of course, we have to factor the risk into consideration. The standard deviation is 6.28%
per annum for the timing strategy η(r1 ), and as a result the Sharpe ratio is 1.01. In contrast,
the ‘Always Long’ strategy has a comparable standard deviation of 6.30%, but a negative
Sharpe ratio of −0.07. The long-term buy-and-hold strategy has a much higher standard
deviation of 20.99%, and a much lower Sharpe ratio of 0.13. Note that the timing strategy
η(r1 ) also enjoys a large positive skewness of 1.01 and large kurtosis of 16.47, suggesting that
it often delivers large positive returns.

We also measure annual cumulative returns as another measure of performance. The

average annual cumulative return is 6.97% for the timing strategy η(r1 ) and −0.53% for the
‘Always Long’ strategy, and 2.34% for the buy-and-hold strategy. It is even striking if we
measure the cumulative return over the entire sample period. The timing strategy η(r1 )
delivers 134.57%, whereas the other two benchmark strategies deliver −8.90% and 6.48%,
respectively.

Finally we report the success rate which is the percentage of trading days of positive
returns.3 The success rate of the ‘Always Long’ strategy is 50%, suggesting that the uncon-
ditional probability for the last half-hour returns is 50 to 50. This is also the success rate
of the buy-and-hold strategy for the entire period. However, the success rate of the timing
strategy η(r1 ) is 53.92%, greater than 50%.
3
For the timing strategy using both r1 and r12 as signal, the calculation of the success rate also includes
trading days when the market yields negative returns but the strategy is out of the market.

9
 Using the twelfth half-hour return as the timing signal yields similar but weaker results.
The average return is about 2.22% per annum, the Sharpe ratio is 0.352, skewness is 0.60,
kurtosis is 16.56, and success rate is 50.54%. Overall, it still has a higher Sharpe ratio and
greater annual cumulative return than the two benchmarks.

Combining the two signals, r1 and r12 , delivers an improved performance over using the
twelfth half-hour return. However, the performance is slightly weaker than using just the first
half-hour return signal. For example, the average return is now 4.52% vs. 6.34% per annum.
But the success rate is much higher, with an impressive value of 76.88%. This means that
combining both r1 and r12 does improve the percentage of being right substantially. Then,
why does higher success rate yield lower average returns? The reason is that, when combining
the two signals, we take the long or short position only when both of them are positive or
negative, which reduces substantially the number of days we are in the market.4

4.2. Mean-variance portfolios

In contrast with using only the signs for timing strategies, we in this subsection use both
the signs and magnitudes of the predictors to forecast the future returns. Then we apply
this expected return to form the optimal portfolio for a mean-variance investor who allocates
funds between the market (SPY) and the risk-free asset (the Treasury T-bill).

The mean-variance efficient portfolio weights are given as

r̂13,t+1
wt = 2
, (6)
γ σ̂13,t+1

where r̂13,t+1 is the forecasted last half-hour return on day t + 1 conditional on information
available at t, σ̂13,t+1 is the standard deviation of the last half-hour return, both of which are
estimated from the recursive regression, and the relative risk aversion coefficient, γ, is set
to be 5. To be more realistic, we impose the portfolio constraint that weights on the risky
asset must be between −0.5 and 1.5, meaning that the investor is allowed to borrow or short
50% on margin. This will limit the potential economic gains from the usual unconstrained
weights.5
4
If we exclude the no trading days with zero returns, the strategy performs the best as expected, with an
annualized average return of 9.15%, a standard deviation of 6.42%, thus a Sharpe ratio of 1.43, a comparable
skewness of 1.54 and kurtosis of 19.28.
5
Unreported results, available upon request, show that the performance of the unrestricted portfolios is
much stronger.

10
 Over the out-of-sample period, the certainty equivalent rate of return of the realized
utility is
1
CER = µ̂p − γ σ̂p2 , (7)
2
where µ̂p and σ̂p are computed based on the realized portfolio returns.

The results are reported in Table 5. Using the historical averages to predict future returns
yields an average return of 0.11% per annum, a Sharpe ratio of 0.03, and a CER of 0.10%
per annum. In sharp contrast, using the first half-hour returns to forecast the last half-hour
returns yields an average returns of 7.27% per annum and a standard deviation of 6.18%
per annum. So the portfolio yields a Sharpe ratio of 1.18 with large positive skewness and
kurtosis. The CER is 7.23% per annum. Weaker performance is observed when using the
twelfth half-hour returns to forecast the last half-hour returns, but the Sharpe ratio is still
much higher than that of the benchmark. Finally, with both the first and twelfth half-hour
returns used to forecast the last half-hour returns, the portfolio delivers an average return of
7.65% per annum, a Sharpe ratio of 1.15, and a CER of 7.61% per annum. Note that unlike
the case with market timing, the performance of using both predictors is slightly better than
using the first half-hour return alone because we are now are always in the market. It is just
that the allocation varies daily.

5. Macroeconomic Events

In this section, we explore the relation between the intraday momentum and macroeconomy.
We examine its performance first over business cycles, and then on macroeconomic news
releases.

5.1. Business Cycles

With the NBER dates for expansions and recessions, we can divide all the trading days into
these two types. The performances of the intraday momentum are reported in Table 6.

During the expansions, only the first half-hour return can predict the last half-hour
return, and the predictability, though significant, is relatively weaker, with an R2 of 1.3%.
However, during recessions, both predictors are highly significant and R2 increases about 5
times to 7.6%.

11
5.2. News Releases

Previously, we have found that the intraday momentum is stronger with higher volatility.
One possibly source of volatility may be the release of major economic news. It is hence of
interest to examine this empirically.

While there are many regular news releases, we here focus on three whose release dates
are easily collected and they represent different times of the day. The first two are the major
macro variables, the gross domestic product (GDP) and the consumer price index (CPI).
Both of them are released monthly on pre-specified dates at 8:30 am before the market
opens, like most other macroeconomic news. The third is the minutes of Federal Open
Market Committee (FOMC), which is released regularly at about 2:00 pm. We analyze the
impact of news release by dividing all the trading days into two groups: days with news
release, and days with no news release.

Table 7 report the performances of the intraday momentum on the two groups of days.
On days with no GDP news, the R2 is 3.6%, slightly higher than 3.5%, the unconditional
value or the R2 on all trading days. On days with GDP releases, the R2 is lowered to
2.2%, that is, the intraday momentum seems weaker. It may simply indicates that the
market incorporates the GDP news fairly fully in the first half-hour, and hence not much
continuation of the prices. This seems consistent with the fact that the average volatility of
the first half-hour on GDP release days is only at the intermediate level.

However, the market reaction to the CPI news is different. On days with no CPI news
releases, the R2 is 3.3%, slightly lower than 3.5%. But on days with CPI releases, the R2
is more than doubled to 7.6%. This seems to suggest there is a carry over effect of the CPI
news.

The most astonishing result is on the release of the FOMC minutes. While the no release
days have only an R2 of 3.2%, the R2 increases enormously to 23.4%. There are two ways
to see why this result is astonishing. First, the size of the R2 is large by any standards,
and exceeds by far almost all predictors at the usual monthly frequency. Second, the market
participants seems to anticipate correctly in the first half-hour what message the Fed is going
to send out to the market. After the news release, it seems that there is a strong reaction
of the market to continue the trend after the news confirmation of the correctness of the
anticipation.

12
 Will the larger R2 s on the news release days imply greater economic gains? To answer
this question, we examine the performances of the earlier market timing strategies on days
with news release, and days with no news release. Table 8 reports the results. For the GDP
news, the profits on release days is about twice greater. This is interesting as the profits are
made on relatively lower out-of-sample R2 days. This suggests that, as some studies found,
statistical significant is not always the same as economic significance. For the CPI, the gains
are even greater, to about 3 times. The greatest economic gains occur on the release days of
the FOMC minutes. The annualized average return reaches a high value of 22.14%. This is
about 4 times greater than the no news lease days. Overall, the performance of the intraday
momentum is much stronger economically on the days with the three macro news releases
(clearly it may or may not be stronger on other macro news releases).

6. Robustness

In this section, we examine the robustness of the intraday momentum along two dimensions.
First, we examine how the economic value measure may vary for various mean-variance
portfolio. Then, we apply the same strategy to NASDAQ market to see whether there exists
similar intraday momentum.

6.1. mean-variance portfolio

In Table 9, we examine the robustness of the out-of-sample mean-variance portfolio perfor-

mance by varying the relative risk aversion coefficient, γ, and/or imposing different restric-
tions on portfolio weights. We use the portfolio formed from forecasting the last half-hour
return using both returns. In Panel A, we keep γ constant at five and change the port-
folio weight restrictions. The first alternative restriction is no short sell and no borrowing
(ψ2 : 0 ≤ w ≤ 1.0), which is more restrictive than the one used in Table 5. Nor surprisingly,
the performance is weaker with an average return of 3.82% per annum but a Sharpe ratio
of 0.90. The Sharpe ratio is only slightly lower because of lower volatility of the portfolio.
Relaxing the restriction by allowing shorting (ψ3 : −1.0 ≤ w ≤ 1.0) increases the average
return but also volatility. In this case, the average return is around 7.55% per annum, and
CER 7.48%. Finally, we allow both shorting and borrowing (ψ4 : −1.0 ≤ w ≤ 2.0), which
delivers much higher return (11.31% per annum), Sharpe ratio (1.20), and CER (11.22% per

13
annum).

In Panel B, we further reduce the relative risk aversion coefficient to two, and then
impose various portfolio weight restrictions, and in Panel C, we further increase the relative
risk aversion coefficient to 10. The results overall are very similar to each other and to the
previous case where γ is five. Of course, when no restriction is imposed, the average return
and standard deviation are indeed very different for different γ, and the lower γ is, the higher
the average return and standard deviation are. But the Sharpe ratio is the same because
they are all on the same efficient frontier. Imposing portfolio restrictions, on the other hand,
makes γ more or less irrelevant, and the portfolio performance is very close.

6.2. NASDAQ Market

To assess whether there exists similar intraday momentum in the NASDAQ market vs the
large cap stocks as represented by the S&P500, we consider, for brevity, only the economic
measure of significance.

Table 10 reports the results based on trading the NASDAQ 100 ETF (QQQ). The results
are qualitatively similar to those in Table 5. For example, the certainty equivalent return of
using r1 is 8.74%, compared with 7.23% for the S&P500. However, while r12 has predictive
value for the S&P500, it does not have any economic value for the NASDAQ. Either using r1
or using both r1 and r12 , the economic values are quite close to those earlier results. In short,
the intraday momentum is almost as strong for the NASDAQ market as for the S&P500.

7. Conclusions

Extending to the well-known momentum that winners of past six months or a year tend to
be winners and the losers tend to be losers (Jegadeesh and Titman, 1993), we, in this paper,
document that the first half-hour return on the market predicts the market return in the last
half-hour. The intraday predictability is statistically significant both in- and out-of-sample.
In terms of market timing or asset allocation, the economic gains of using the predictability
are substantial too. In addition, we find that the intraday momentum is stronger on high
volatile days, recession days, and some macroeconomic news release days. We interpret the
trading behavior of daytraders and informed traders as the economic driving forces behind

14
the intraday momentum.

There are a number of open issues on intraday momentum. For examples, while this paper
studies the intraday momentum at the market level, it is unknown at the cross-section. In
addition, while Griffin et al. (2003), Moskowitz et al. (2012) and Asness et al. (2013) show
momentum with monthly data hold internationally and across asset classes such as bond
and currencies, it is unknown whether there are similar empirical patterns for the intraday
data. There are important topics for future research.

15
References
Admati, Anat R. and Paul Pfleiderer, 1988, A theory of intraday patterns: volume and price
variability, Review of Financial Studies 1, 3–40.

Asness, Clifford S., Tobias J. Moskowitz, and Lasse Heje Pedersen, 2013, Value and momen-
tum everywhere, Journal of Finance 68, 929–985.

Barclay, Michale J. and Jerold B. Warner, 1993, Stealth trading and volatility which trades
move prices?, Journal of Financial Economics 34, 281–305.

Boehmer, Ekkehart and Juan Wu, 2013, Short selling and the price discovery process, Review
of Financial Studies 26, 287–322.

Campbell, John Y. and Samuel B. Thompson, 2008, Predicting excess stock returns out
of sample: Can anything beat the historical average?, Review of Financial Studies 21,
1509–1531.

Chakravarty, Sugato, 2001, Stealth-trading: Which traders’ trades move stock prices?, Jour-
nal of Financial Economics 61, 289–307.

Chordia, Tarun, Richard Roll, and Avanidhar Subrahmanyam, 2011, Recent trends in trad-
ing activity and market quality, Journal of Financial Economics 101, 243 – 263.

Corwin, Shane A. and Paul Schultz, 2012, A simple way to estimate bid-ask spreads from
daily high and low prices, The Journal of Finance 67, 719–760.

Coval, Joshua D. and Tyler Shumway, 2005, Do behavioral biases affect prices?, The Journal
of Finance 60, 1–34.

Griffin, John M., Xiuqing Ji, and J.Spencer Martin, 2003, Momentum investing and business
cycle risk: Evidence from pole to pole, Journal of Finance 58, 2515–2547.

Haigh, Michael S. and John A. List, 2005, Do professional traders exhibit myopic loss aver-
sion? an experimental analysis, The Journal of Finance 60, 523–534.

Heston, Steven L., Robert A. Korajczyk, and Ronnie Sadka, 2010, Intraday patterns in the
cross-section of stock returns, The Journal of Finance 65, 1369–1407.

16
Hora, Merrell, 2006, Tactical liquidity trading and intraday volume, Working paper, Credit
Suisse Group.

Jain, Prem C. and Gun-Ho Joh, 1988, The dependence between hourly prices and trading
volume, Journal of Financial and Quantitative Analysis 23, 269–283.

Jegadeesh, Narasimhan and Sheridan Titman, 1993, Returns to buying winners and selling
losers: implications for stock market efficiency, Journal of Finance 48, 65–91.

Locke, Peter R. and StevenC. Mann, 2000, Do professional traders exhibit loss realization
aversion?, Working paper, Texas Christian University.

Moskowitz, Tobias J., Yao Hua Ooi, and Lasse Heje Pedersen, 2012, Time series momentum,
Journal of Financial Economics 104, 228–250.

Neely, Christopher J., David E. Rapach, Jun Tu, and Guofu Zhou, 2013, Forecasting the
equity risk premium: The role of technical indicators, Management Science forthcoming.

Newey, Whitney K. and Kenneth D. West, 1987, A simple, positive semi-definite, het-
eroskedasticity and autocorrelation consistent covariance matrix, Econometrica 55, 703–
708.

Odean, Terrance, 1998, Are investors reluctant to realize their losses?, The Journal of Fi-
nance 53, 1775–1798.

Rapach, David and Guofu Zhou, 2013, Chapter 6 - forecasting stock returns, in Graham
Elliott and Allan Timmermann, eds.: Handbook of Economic Forecasting (Elsevier, ).

Rapach, David E., Jack K. Strauss, and Guofu Zhou, 2010, Out-of-sample equity premium
prediction: Combination forecasts and links to the real economy, Review of Financial
Studies 23, 821–862.

Shefrin, Hersh and Meir Statman, 1985, The disposition to sell winners too early and ride
losers too long: Theory and evidence, The Journal of Finance 40, 777–790.

Zhang, X. Frank, 2006, Information uncertainty and stock returns, Journal of Finance 61,
105–137.

17
 Average 30 Minute Trading Volume
1.5e+07
1.0e+07
Volume
5.0e+06
0

1 2 3 4 5 6 7 8 9 10 11 12 13

Figure 1: Average 30 minute trading volume of SPY.

For every 30 minute from 9:30 am to 4:00 pm Eastern time, Figure 1 shows the average trading
volume for SPY by the period of the day from January 4, 1999 to December 31, 2012. Each 30
minute period is labeled from one to thirteen sequentially.

18
 Table 1: Predictability of the Final Half-Hour Returns
This table reports the results of regressing the final half-hour returns (r13 ) on the first half-hour return (r1 ) and the twelfth half-hour
return (r12 ) of the day. Panel A, B, and C are the results for three periods: the whole sample period, financial crisis period from January
2, 2007 to December 31, 2009, and period excluding the financial crisis, respectively. Newey and West (1987) robust t-statistics are in
parentheses and significance at the 1%, 5% , or 10% level is given by an ***, an ** or an *, respectively. The sample period is from
January 4, 1999 to December 31, 2012.

Predictor r1 r12 r1 and r12 r1 r12 r1 and r12 r1 r12 r1 and r12

19
Panel A: Whole Sample Period Panel B: Financial Crisis (2007 - 2009) Panel C: Excluding Financial Crisis
Intercept(×10000) -0.267 -0.277 -0.356 -0.225 -1.120 -0.680 -0.154 -0.074 -0.171
(-0.45) (-0.46) (-0.59) (-0.12) (-0.59) (-0.36) (-0.27) (-0.13) (-0.30)
βr1 0.077∗∗∗ 0.076∗∗∗ 0.137∗∗∗ 0.133∗∗∗ 0.044∗∗∗ 0.044∗∗∗
(4.15) (4.29) (3.15) (3.26) (3.07) (3.13)
βr12 0.138∗∗∗ 0.137∗∗∗ 0.211∗∗ 0.202∗∗ 0.073 0.074
(2.66) (2.71) (2.06) (2.13) (1.86) (1.91)
R2 (%) 2.0 1.5 3.5 4.3 3.1 7.1 0.8 0.4 1.3
 Table 2: The Impact of Volatility
This table reports the regression results of regressing the final half-hour return (r13 ) on the first half-hour return (r1 ) and the twelfth
half-hour return (r12 ), under different levels of volatility. The first half-hour volatility is estimated using 1-minute returns within the
first half hour period and volatilities over all trading days are ranked into three levels: low, medium and high. Newey and West (1987)
robust t-statistics are in parentheses and significance at the 1%, 5% , or 10% level is given by an ***, an ** or an *, respectively. The
sample period is from January 4, 1999 to December 31, 2012.

Volatility Low Medium High

20
Intercept -0.000∗ -0.000 0.000
(-1.69) (-0.86) (0.54)
βr1 0.026 0.047∗∗ 0.083∗∗∗
(1.03) (2.37) (4.03)
βr12 0.078∗ 0.066 0.163∗∗
(1.65) (1.50) (2.38)
R2 (%) 0.5 0.6 4.8
 Table 3: Out-of-Sample Predictability
This table examines the out-of-sample predictability of the final half-hour return (r13 ) by the first half-hour return (r1 ) and the twelfth
half-hour return (r12 ) based on recursive estimations. The window of the estimation initially uses the first year (1999) and progressively
includes one more month of returns. The out-of-sample predictability is measured by the out-of-sample R-squared (OOS R2 ),
∑T
2 (r13,t − r̂13,t )2
OOS R = 1 − ∑t=1
T
,
t=1 (r13,t − r̄13,t )2

where r̂13,t is the forecasted last half-hour return from the predictive regression estimated through period t − 1, and r̄13,t is the historical
average return of the last half-hour estimated through period t − 1. Newey and West (1987) robust t-statistics are in parentheses and
significance at the 1%, 5% , or 10% level is given by an ***, an ** or an *, respectively. The sample period is from January 4, 1999 to

21
December 31, 2012.

r1 r12 r1 and r12

βr1 0.06∗∗∗ 0.06∗∗∗

(33.9) (34.4)
βr12 0.07∗∗∗ 0.07∗∗∗
(17.1) (16.7)

OOS R2 (%) 1.76 1.01 2.74

 Table 4: Market Timing
This table reports the economic value of timing the last half-hour market return using the first half-hour return, or the twelfth half-hour
return or both. We use the sign of the first (twelfth) half-hour return as the timing signal - when the first (twelfth) half-hour return is
positive (negative), we take a long (short) position in the market. When both returns are used, we only trade when both returns have
the same sign - long when both are positive and short when both are negative.
{ r , if r > 0;
13 1
η(r1 ) =
−r13 , if r1 ≤ 0.
{ r , if r > 0;
13 12
η(r12 ) =
−r13 , if r12 ≤ 0.

{ r13 , if r1 > 0 & r12 > 0;

η(r1 , r12 ) = −r13 , if r1 ≤ 0 & r12 ≤ 0;
0, otherwise,
The benchmark (‘Always Long’) is to invest in the market for the last half hour each trading day; ‘Buy-and-Hold’ is to buy and hold the
market on a daily basis. Newey and West (1987) robust t-statistics are in parentheses and significance at the 1%, 5% , or 10% level is
given by an ***, an ** or an *, respectively. The sample period is from January 4, 1999 to December 31, 2012.

22
Annual
Timing Signal Avg Ret(%) Std Dev(%) SRatio Skewness Kurtosis Cum Ret(%) Success
Cum Ret(%)
Panel A: Market Timing
r1 6.34∗∗∗ 6.28 1.01 1.01 16.47 6.97 134.57 53.92
(3.76)
r12 2.22 6.30 0.35 0.60 16.56 2.35 32.39 50.54
(1.31)
r1 and r12 4.52∗∗∗ 4.52 1.00 2.38 36.36 4.88 85.09 76.88
(3.74)
Panel B: Benchmark
Always Long -0.47 6.30 -0.07 -0.22 16.59 -0.53 -8.90 50.00
(-0.28)
Buy-and-Hold 2.66 20.99 0.13 0.07 6.89 2.34 6.48
(0.47)
 Table 5: Mean-Variance Portfolio Performance
This table reports the economic value of recursively predicting the final half-hour market return
using the first half-hour return, or the twelfth half-hour return or both. We use the predicted returns
to form mean-variance optimal portfolio for a mean-variance investor with a relative risk aversion
of five. We report summary statistics such as average return (Avg Ret) in percentage, standard
deviation (Std Dev) in percentage, Sharpe Ratio (SRatio), Skewness, and Kurtosis. Also reported
are the certainty equivalent rate of returns in percentage, CER. r̄13 uses the recursively estimated
average returns of the last half hour returns (r13 ) to form optimal mean-variance portfolio. Newey
and West (1987) robust t-statistics are in parentheses and significance at the 1%, 5% , or 10% level
is given by an ***, an ** or an *, respectively. The sample period is from January 4, 1999 to
December 31, 2012.

Predictor Avg Ret(%) Std Dev(%) SRatio Skewness Kurtosis CER(%)

r̄13 0.11 3.06 0.03 0.34 19.01 0.10

(0.12)
βr1 7.27∗∗∗ 6.18 1.18 1.90 39.94 7.23
(4.23)
βr12 2.57 6.44 0.40 0.86 61.46 2.53
(1.43)
β1 r1 + β2 r12 7.65∗∗∗ 6.68 1.15 1.08 48.16 7.61
(4.12)

23
 Table 6: The mpact of Business Cycle
This table examines the predictability of the final half-hour return (r13 ) by the first half-hour return (r1 ) and the twelfth half-hour return
(r12 ) in different stages of the business cycle. The expansion and recession periods are defined by the NBER. Newey and West (1987)
robust t-statistics are in parentheses and significance at the 1%, 5% , or 10% level is given by an ***, an ** or an *, respectively. The
sample period is from January 4, 1999 to December 31, 2012.

Business Cycle Expansion Recession

24
Intercept -0.00 0.00
(-1.41) (1.11)
βr1 0.05∗∗∗ 0.12∗∗∗
(3.53) (3.01)
βr12 0.05 0.23∗∗
(1.21) (2.42)
R2 (%) 1.3 7.6
 Table 7: The Impact of Macro News Release on Predictive Regression
This table contrasts the results of regressing the final half-hour returns (r13 ) on the first and twelfth
half-hour returns of the day (r1 and r12 ) when there are macro news releases with the regression
results when there are no macro news releases. The first half-hour return (r1 ) is calculated from
the close price of the previous trading day to the first half hour (10:00 am Eastern Time). GDP:
monthly GDP estimate release at 8:30 am Eastern Time; CPI: monthly release of CPI at 8:30 am
Eastern Time; FOMC: Federal Open Market Committee minutes release at 2:00 pm Eastern Time.
Newey and West (1987) robust t-statistics are in parentheses and significance at the 1%, 5% , or
10% level is given by an ***, an ** or an *, respectively. The sample period is from January 4,
1999 to December 31, 2012.

No Release Release No Release Release No Release Release

GDP CPI FOMC

Intercept -0.00 0.00 -0.00 0.00 -0.00 -0.00
(-0.75) (0.39) (-0.74) (0.77) (-0.41) (-1.29)
βr1 0.08∗∗∗ 0.10∗∗ 0.07∗∗∗ 0.13∗∗ 0.07∗∗∗ 0.19∗∗∗
(4.11) (2.06) (4.07) (2.04) (4.11) (3.04)
βr12 0.14∗∗∗ -0.01 0.14∗∗∗ 0.16 0.13∗∗∗ 0.54∗∗∗
(2.75) (-0.05) (2.68) (0.92) (2.58) (2.69)
R2 (%) 3.6 2.2 3.3 7.6 3.2 23.4

25
 Table 8: The Impact of Macro News Release on Timing Performance
This table reports the profitability of timing the final half-hour market return using the first half-
hour return, contrasting the days with certain macro news release with the days with no macro
news release. We use the sign of the first half-hour return as the timing signal - when the first
half-hour return is positive (negative), we take a long (short) position in the market. We report
summary statistics such as average return (Avg Ret) in percentage, standard deviation (Std Dev) in
percentage, Sharpe Ratio (SRatio), Skewness, and Kurtosis. GDP: monthly GDP estimate release
at 8:30 am Eastern Time; CPI: monthly release of CPI at 8:30 am Eastern Time; FOMC: Federal
Open Market Committee minutes release at 2:00 pm Eastern Time. Newey and West (1987) robust
t-statistics are in parentheses and significance at the 1%, 5% , or 10% level is given by an ***, an
** or an *, respectively. The sample period is from January 4, 1999 to December 31, 2012.

Macro News Avg Ret(%) Std Dev(%) SRatio Skewness Kurtosis

Non-Release GDP 6.02∗∗∗ 6.30 0.96 1.03 16.95

(3.48)
Release GDP 13.12∗ 5.92 2.22 0.54 3.56
(1.74)
Non-Release CPI 5.78∗∗∗ 6.29 0.92 1.00 16.94
(3.35)
Release CPI 18.29∗∗ 6.05 3.02 1.34 5.19
(2.37)
Non-Release FOMC 5.82∗∗∗ 6.29 0.93 0.99 16.71
(3.39)
Release FOMC 22.14∗∗ 6.01 3.68 1.55 8.82
(2.44)

26
 Table 9: Robustness of Out-of-Sample Mean-Variance Portfolio Performance
This table reports the out-of-sample performance of different combinations of the relative risk
aversion coefficient, γ, and portfolio weight restrictions, ψi , i = 1, · · · , 4. The recursive regression
uses both the first half-hour return and the twelfth half-hour return as described in Table 5. We
report summary statistics such as average return (Avg Ret) in percentage, standard deviation (Std
Dev) in percentage, Sharpe Ratio (SRatio), Skewness, Kurtosis and the certainty equivalent rate
of returns in percentage (CER). Newey and West (1987) robust t-statistics are in parentheses and
significance at the 1%, 5% , or 10% level is given by an ***, an ** or an *, respectively. The sample
period is from January 4, 1999 to December 31, 2012.

Variable Avg Ret(%) Std Dev(%) SRatio Skewness Kurtosis CER(%)

Panel A: γ = 5
ψ2 : 0 ≤ w ≤ 1.0 3.82∗∗∗ 4.24 0.90 1.29 59.16 3.80
(3.23)
ψ3 : −1.0 ≤ w ≤ 1.0 7.58∗∗∗ 6.14 1.23 0.85 18.80 7.54
(4.43)
ψ4 : −1.0 ≤ w ≤ 2.0 11.31∗∗∗ 9.41 1.20 0.98 39.03 11.22
(4.31)
Panel B: γ = 2
ψ1 : −0.5 ≤ w ≤ 1.5 7.68∗∗∗ 6.79 1.13 1.05 44.99 7.57
(4.06)
ψ2 : 0 ≤ w ≤ 1.0 3.77∗∗∗ 4.29 0.88 1.25 56.29 3.73
(3.16)
ψ3 : −1.0 ≤ w ≤ 1.0 7.69∗∗∗ 6.23 1.23 0.81 17.69 7.59
(4.43)
ψ4 : −1.0 ≤ w ≤ 2.0 11.52∗∗∗ 9.63 1.20 0.93 35.61 11.29
(4.30)
Panel C: γ = 10
ψ1 : −0.5 ≤ w ≤ 1.5 7.39∗∗∗ 6.48 1.14 1.20 54.29 7.37
(4.09)
ψ2 : 0 ≤ w ≤ 1.0 3.77∗∗∗ 4.15 0.91 1.35 64.20 3.76
(3.26)
ψ3 : −1.0 ≤ w ≤ 1.0 7.56∗∗∗ 5.98 1.26 0.92 20.83 7.54
(4.54)
ψ4 : −1.0 ≤ w ≤ 2.0 10.87∗∗∗ 8.99 1.21 1.12 46.45 10.83
(4.34)

27
 Table 10: Out-of-Sample Porftolio Performance using QQQ
This table reports the economic value of recursively predicting the last half-hour returns using Nasdaq 100 ETF as the underlying
asset. As a benchmark, we use the recursively estimated average returns to predict the last half-hour returns. We form mean-variance
r̂ 2
2
efficient strategy using both predicted returns. The weights are given as wt = γ σ̂13,t+1 . The variance σ̂13,t+1 is estimated in the recursive
13,t+1
estimation, and γ is set to be five. The portfolio weights are restricted between -0.5 and 1.5. We report summary statistics such as
average return (Avg Ret) in percentage, standard deviation (Std Dev) in percentage, Sharpe Ratio (SRatio), Skewness, and Kurtosis.
Also reported are the certainty equivalent rate of returns in percentage, CER = µ̂p − 12 γ σ̂p2 . Newey and West (1987) robust t-statistics
are in parentheses and significance at the 1%, 5% , or 10% level is given by an ***, an ** or an *, respectively. The sample period is
from March 10, 1999 to December 31, 2012.

28
Predictor Avg Ret(%) Std Dev(%) SRatio Skewness Kurtosis CER(%)

r̄13 0.82 4.30 0.19 -0.53 16.49 0.80

(0.69)
βr1 8.83∗∗∗ 9.20 0.96 1.45 37.40 8.74
(3.45)
βr12 -0.59 8.81 -0.07 -0.12 23.57 -0.67
(-0.24)
β1 r1 + β2 r12 7.37∗∗∗ 9.24 0.80 -0.12 19.53 7.28
(2.86)