SSRN Id3850494

Chapter 59
Technical Analysis in the Stock

Market: A Review
Yufeng Han
Belk College of Business, University of North Carolina-Charlotte, Charlotte, North Carolina,
USA
yhan15@uncc.edu
Yang Liu
School of Economics and Management, Tsinghua University, Beijing, China
liuy9.15@sem.tsinghua.edu.cn
Guofu Zhou
Olin School of Business, Washington University in St. Louis, St. Louis, Missouri, USA
Zhou, Guofu ¡zhou@wustl.edu¿
Yingzi Zhu
School of Economics and Management, Tsinghua University, Beijing, China
zhuyz@sem.tsinghua.edu.cn
CHAPTER OUTLINE:
59.1 Introduction
59.2 Time-series Predictability
59.2.1 Market
59.2.1.1 Empirical studies
59.2.1.2 The theory
59.2.1.3 Brock, Lakonishok, and LeBaron (1992): An update
59.2.2 Portfolios and Other Assets
59.2.2.1 Related studies
59.2.2.2 Han, Yang and Zhou (2013): An update
59.3 Cross-section Predictability
59.3.1 Fama-MacBeth Regressions
59.3.1.1 Han, Zhou and Zhu (2016): An update
Electronic copy available at: https://ssrn.com/abstract=3850494

59.3.2 Machine Learning Methods
59.4 Conclusion
Abstract
Technical analysis is the study for forecasting future asset prices with past data. In this
survey, we review and extend studies on not only the time-series predictive power of technical
indicators on the aggregated stock market and various portfolios, but also the cross-sectional
predictability with various firm characteristics. While we focus on reviewing major academic
research on using traditional technical indicators, but also discuss briefly recent studies that
apply machine learning approaches, such as Lasso, neural network and genetic programming,
to forecast returns both in the time-series and on the cross-section.
Keywords
Technical Analysis, Machine Learning, Genetic Programming, Cross-sectional Returns, Pre-
dictability
59.1 Introduction
Why does technical analysis matter? It is about forecasting future asset prices using the
past data, and clearly contradicts with market efficiency theory. In an ideal world in which all
investors have the same information and the same rational expectations, the asset prices must
be a random walk and must be unpredictable by any means. However, the real world is much
different. First, no all investors have the same information on stocks. More importantly, even
if they do have the same information, not everyone processes the information the same way,
and some can view the information as good news and some can view as bad. As a result, how
the market collective reacts to the information is never known an ante for sure by anyone,
and there is no rational or all agreed equilibrium prices in practice. Therefore, investors have
to observe the price path carefully to see how the market price adjusts over time relative
to the information and related to their own expectations, resulting in the relevance of past
data to future prices, yielding a certain degree of predictability.
There are many other reasons why the path matters (see, Han, Zhou and Zhu (2016) for
a review). A simple example of path-dependent asset prices is that investors require greater
returns after a big market crash. Another example is that the market is not perfectly liquid,
and so larger traders must take time in days or even months to build or unwind their large
positions, revealing possibly up or down trends in the market.
Technical indicators have a long history of use by practitioners to predict stock returns.
Smidt (1965) surveys amateur traders and finds that over half of them use charts to identify
trends. Schwager (1993, 1995) and Lo and Hasanhodzic (2009) show that many top traders
and fund managers also rely heavily on technical analysis. Moreover, Covel (2005), citing ex-
amples of successful hedge funds, advocates the use of technical analysis exclusively without
exploiting any fundamental information on the market. Today, with increasing computing

power and data availability, technical indicators are popular and almost all brokerages pro-
vide chart analysis. High-frequency trading, algorithmic trading, and systematic trading
in general are widespread (see, e.g., Harvey, 2021), which are data-driven and rule-based,
sophisticated forms of technical analysis.
In this article, we provide a review on technical analysis. We will focus on the stock
market, while briefly discussing other asset classes. We will first survey major advances on
time-series predictability, and then on cross-section predictability.
On technical indicators, we focus exclusively on the use of moving averages because they
are the most widely used tools for capturing a trend, and trend-following is the major trading
strategy in practice. Moreover, they are also the ones on which we have ample evidence.
Since technical analysis is in general the study for forecasting prices via past data, it includes
machine learning approaches, such as Lasso, neural network and genetic programming, as
special cases. We also discuss the latter as they are getting increasing attention and becoming
more and more useful in applications.
On time-series predictability, we extend the market predictability study of Brock, Lakon-
ishok, and LeBaron (1992) based on price moving averages to recent years, and find that the
early predictability disappears almost completely. Following Schwert (2003) and McLean
and Pontiff (2016), we explain this as publication effect as many profitable patterns of the
stock market tend to disappear, reverse, or attenuate after their publications. Another rea-
son is that the market as a whole is known as very difficult to forecast to begin with, as there
are so many factors that can drive the ups and downs unless some other markets such as
foreign exchanges of which the predictability is greater. However, using more sophisticated
machine learning tools, such as Lasso and a new C-Lasso, Rapach and Zhou (2020) find that
the market can still be profitably predicted with the moving averages in conjunction with
macro predictor, even in recent years.
59.2 Time-series Predictability
There are both time-series predictability and cross-section predictability. We discuss the
former in this section, and the later in the next section.
59.2.1 Market
In this subsection, we focus on time-series predictability of the aggregate stock market

such as an index.
59.2.1.1 Empirical studies
Earlier studies on the usefulness of technical analysis concentrate on the market for its
data availability. Cowles (1933) is perhaps the first empirical study on the profitability of
professional forecasters who presumably use technical analysis and other means. But he

finds that they cannot beat the buy-and-hold strategy. Fama and Blume (1966) examine
a number of filter rules introduced by Alexander (1961) in the securities in the Dow Jones
Industrial Average (DJIA) during 1956 through 1962, and conclude that the trading rules
cannot beat the buy-and-hold strategy either. Other popular technical indicators include
the moving averages (Cootner, 1962; Van, Horne and Parker, 1967, 1968; James, 1968; Dale
and Workman, 1980) and relative strength (Levy, 1967; Jensen and Benington, 1970). The
studies also find that technical trading rules fail to generate profitable performance. In short,
the majority of earlier studies are skeptical about the usefulness of the technical analysis in
the stock market.
With the availability of cheaper computing power and the development of electronic
database, later studies generally have improved in terms of the testing procedure with more
data and more elaborate strategies. Utilizing the DJIA from 1897 to 1986, Brock, Lakon-
ishok, and LeBaron (1992) examine two of the simplest and the most popular technical
trading rules – moving average and trading range break. In contrast to earlier studies, they
find that all 26 technical strategies in their study generally exhibit strong profitability. Using
the same technical rules, Bessembinder and Chan (1995) also find significant predictability
in forecasting index return for a group of Asian stock markets, including Malaysia, Thailand,
and Taiwan. Kwon and Kish (2002) extend the work of Brock, Lakonishok, and LeBaron
(1992) by including trading volume moving averages indicators. Their results confirm that
technical trading rules indeed add value, as compared to a buy-and-hold strategy.
However, there is concern of data snooping bias. Sullivan, Timmermann, and White
(1999) find, using the White (2000) reality check bootstrap methodology, weak profitability
over the 10 year out-of-sample period, suggesting that efficiency of the stock markets has
improved. Our replications of Brock, Lakonishok, and LeBaron (1992) also show diminishing
profits. In general, technical strategies are more robust for equity indexes in emerging
markets (Bessembinder and Chan, 1995; Ito, 1999; Ratner and Leal, 1999), while that in
developed markets are much weaker or have shrink over time (Mills, 1997; Bessembinder and
Chan, 1998).
On the other hand, Neely, Rapach, Tu, and Zhou (2014) show that technical indicators,
including moving average, momentum and volume-based indicators, can provide complemen-
tary information beyond 14 macroeconomic variables collected by Welch and Goyal (2008)
to forecast the equity risk premium. They also show that the predictability of technical
indicators becomes stronger during recessions. Since the frequency of trend breaks happen
more often in recent years, Garg, Goulding, Harvey and Mazzoleni (2020) study the impact
of trend breaks on the performance of standard trend-following strategies across several as-
sets and asset classes, and construct dynamic multi-asset trend-following portfolios, with
performance more than doubling the standard strategies over the last decade.
While most studies are based on linear models, a growing body of recent research combines
traditional technical indicators with non-linear and advanced machine learning techniques,
such as genetic programming and neural networks. Genetic programming (Koza, 1992) is a
supervised machine learning method based on the principle of Darwinian natural evolution.
When applied to technical trading rules, the building blocks of genetic trading rules consist of

various functions of past prices, trading volume, numerical and logical constants, and logical
functions. Allen and Karjalainens (1999) is among the first to apply genetic programming
to examine the profitability of technical trading rules in the stock market. They use the
combination of the moving averages of the past prices to find profitable trading rules to beat
the S&P 500 index, but unsuccessfully. Neely (2003) also finds no evidence that genetic
programming can generate trading rules for the S&P 500 index to beat the buy-and-hold
strategy, while Neely, Weller, and Dittmar (1997) and Neely and Weller (2001) show that ge-
netic trading rules produce superior performance in the foreign exchange markets. Recently,
however, Brogaard and Zareei (2019), with modified genetic programming algorithms, are
able to identify stronger time-series predictability of the S&P 500 index. Berutich, Lopez,
Luna, and Quintana (2016) also present a robust genetic programming approach to deter-
mine the potential buy and sell conditions, and the resulting method yields robust solutions
able to withstand extreme market conditions.
Neural network, based on biological neurosystems, is another commonly used machine
learning approach in generating technical trading rules in the stock market. Gencay (1998a)
uses a simple feed forward neural network model to explore the non-linear technical trading
rules in the DJIA data over 1964 to 1988. The resulting trading rules are shown to outperform
easily the buy-and-hold strategy in terms of both return and Sharpe ratio. Utilizing the
DJIA data over a longer period from 1897 to 1988, Gencay (1998b) further shows strong
and robust evidence of the nonlinear predictability of the trading rules. Gencay and Stengos
(1998) further find that incorporating a 10-day volume indicator into the neural network
model as an additional indicator help to improve the forecast accuracy on the DJIA return.
Other advanced techniques, such as reinforcement learning and support vector machine,
are also utilized to produce technical trading strategies. Lee (2009) develops a prediction
model based on support vector machine (SVM) to predict the trend of the stock market.
Tan, Quek, and Cheng (2011) use the reinforcement learning (RL) technique, based on
the momentum and moving average indicators, to determine the best time to go long and
short. Meanwhile, there are also many studies utilizing the above techniques simultaneously
to improve the profitability of various technical trading strategies. For example, Kim and
Han (2000) propose a hybrid neural network model along with genetic algorithms to predict
the stock price index, and show that the combination of the two techniques outperforms
the conventional models. Chavarnakul and Enke (2009) develop an intelligent hybrid stock
trading system that utilizes neural networks, fuzzy logic, and genetic algorithms to prove
the predictability of the volume adjusted moving average indicator.
Recall that our definition of technical analysis is more general. The predictors are not
limited to past price and volume information, but allowing for past data of any variables. In
the academic literature, there are many economic variables that are well motivated to have
predictive power on the stock market.
Welch and Goyal (2008) find that, while numerous economic variables have in-sample
predictive ability for the market premium, they fail to deliver consistent out-of-sample fore-
casting gains relative to the simple historical average. Campbell and Thompson (2008) argue
a few predictors can have out-of-sample predictive power, but it is difficult to know ex ante

which of the predictors will do well in the future. Without forcing such a choice, Rapach,
Strauss and Zhou (2010) show that the market is predictable by using a forecast combina-
tion method (averaging all the forecasts), thus providing the first convincing out-of-sample
evidence on market predictability. Subsequently, there are many studies that find market
predictability. For examples, Li, Ng and Swaminathan (2013) and Kelly and Pruitt (2013)
and Rapach, Ringgenberg and Zhou (2016) find, respectively, that the aggregated implied
cost of capital, book-market ratio and short interest are powerful predictors of the market,
while Huang, Jiang, Tu and Zhou (2015), Jiang, Lee, Martin and Zhou (2019) and Chen,
Tang, Yao and Zhou (2020) find investor sentiment, manager sentiment and employee senti-
ment can predict the market too. Koijen and Van Nieuwerburgh (2011), Rapach and Zhou
(2013) and Timmermann (2018) provide reviews of the literature.
Recently, Gu, Kelly, and Xiu (2020) apply a comprehensive set of popular machine learn-
ing tools to both macro economic and firm specific predictors in a panel set-up for forecasting
firm returns, and then obtain a significant market forecast by summing the component fore-
casts. Instead of using panel models, Dong, Li, Rapach and Zhou (2021) show, in traditional
time-series regressions, that mispricing is related to predictability, and so about 100 asset
pricing anomalies become a large set of new time-series predictors on the market beyond the
usual dozen macro predictors.
The least absolute shrinkage and selection operator (Tibshirani 1996, LASSO) is one of
the most popular machine learning techniques in finance. Rapach, Strauss and Zhou (2013)
seems the first to apply it for forecasting returns, and they find that US market leads the
world. There does not appear any major LASSO applications until Chinco, Clark-Joseph
and Ye (2019) who use the LASSO to predict individual stock returns one minute ahead.
Han, He, Rapach and Zhou (2020) propose a C-Lasso approach that can be applied more
effectively to combining information from a large number of forecasts. Rapach and Zhou
(2020) apply it to both technical indicators and macro variables and show that the market
can be profitably predicted.
Since machine learning tools are getting popular and new methods are coming out fast,
their application to forecasting the stock market is almost sure to increase substantially over
time and may be found effects on the market prices too.
59.2.1.2 The theory
A short survey of theoretical models that justify the use of technical analysis is provided
by Han, Zhou and Zhu (2016). Here we focus on the most widely used technical indicators,
the moving averages (MAs), which are the foundation of trend-following.
Zhu and Zhou (2009) seems the first to provide a theoretical basis for the MAs. In a
partial equilibrium model for a small investor, the MAs are fast learning methods about
the underlying true model of asset dynamics. Under uncertainty about predictability or
uncertainty about the true parameters, the MA learning can add value to an asset allocation
problem in reasonable sample sizes. In contrast, sophisticated econometric methods, though
asymptotically optimal, underperform the simple MAs due to not enough data. As an

extension, Zhou, Zhu and Qiang (2012) provide an optimal asset allocation strategy using
the MAs, and show it makes a significant economic difference empirically.
Han, Zhou and Zhu (2016) propose further an equilibrium model in which there are
informed traders and technical traders. In the presence of noise traders, they show that,
even in equilibrium, the technical traders can survive in the long-run and the MAs have
predictive power on asset prices. However, the fraction of technical traders can matter.
When the fraction is small, MAs indicate trend-following, but when the fraction is large,
they predict counter-trends. In contrast to Han, Zhou and Zhu (2016) where the technical
traders are assumed exogenously, Detzel, Liu, Strauss, Zhou and Zhu (2021) propose a novel
equilibrium model in which technical analysis can arise endogenously via rational learning.
They document that ratios of prices to their moving averages forecast daily Bitcoin returns
in- and out-of-sample, and similar results hold for small-cap, young-firm, and low-analyst-
coverage stocks as well as NASDAQ stocks during the dotcom era.
A limitation of the theoretical model of Han, Zhou and Zhu (2016) is that it justifies the
use of one MA as a predictor. In practice, multiple MAs may be used at the same time.
For simplicity, we consider two MA predictors. We show in what follows that the main
conclusion can be extended to allowing for two MAs as predictors.
Consider now three types of investors: the informed, the short-term and long-term techni-
cal traders. Let w1 and w2 be the fraction of the short- and long-term technical traders. The
informed investors observe the dividend Dt , mean growth rate of dividend πt , the price as
well as all history of the variables, while they do not directly observe the supply of asset, the
fluctuation of which is due to noise trader behavior. Formally, F i (t) = {Dτ , Pτ , πτ : τ ≤ t}
is the informed investors’ information set at time t. On the other hand, technical traders
only observe dividend and price, and do not directly observe the mean grow rates of the
dividends. The two types of technical investors use
Z t
Ait ≡ exp [−αi (t − s)]Ps ds, (59.1)
−∞
with i = 1, 2 and α1 > α2 > 0, to infer information. A1t corresponds to the short-term signal
and A2t the long-term signal. Formally, Fi u (t) = {1, Dt , Pt , Ait }, i = 1, 2 is the information
sets of the technical traders of type i at time t. The result is summarized as follows.
Theorem: In an economy defined in above Assumptions, there exists a stationary rational
expectations equilibrium. The equilibrium price function has the following linear form:
Pt = p0 + p1 Dt + p2 πt + p3 θt + p4 A1t + p5 A2t , (59.2)
where p0 , p1 , p2 , p3 , p4 and p5 are constants determined only by model parameters.

Proof: Similar to Equations (A.1) and (A.2) of Han, Zhou and Zhu (2016), the informed
investors’ optimization problem is
Z ∞
−ρs−c(s)
max
i i
E − e ds|Ft s.t. dW i = (rW i − ci )dt + η i dQ.
i
(59.3)
η ,c t

Let J i (W i , Dt , πt , θt , A1t , A2t ; t) be the value function, then it satisfies the HJB equation

1 i iT i2
0 = max −e−ρt−c + JW (rW i − ci + η i eiQ Ψ) + σQ σQ η JW W + η i σQ i iT
σΨ JW Ψ
c,η 2

i T 1 i iT
−ρJ + (eΨ Ψ) JΨ + σΨ JΨΨ σΨ . (59.4)
2
The optimal demand for stock is given by
η i = f i Ψi = f0i + f1i Dt + f2i πt + f3i θt + f4i A1t + f5i A2t , (59.5)
where f0i , f1i , f2i , f3i , f4i and f5i are constants.
To solve the technical traders’ optimization problem, for i = 1, 2, let Wiu be the wealth
of a technical trader i, ηiu be the holding of stock and cui be the consumption. Then the
optimization problem is
Z ∞
−ρs−ci (s)
max
u u
E − e u
ds|Fit s.t. dWiu = (rWiu − cui )dt + ηiu dQui . (59.6)
ηi ,ci t
Let J u (Wiu , Ψui ; t) be the value function, then it solves the following HJB equation,

u 1 u uT u2
0 = max −e−ρt−ci + JW u u u u u u
u (rWi − ci + ηi eQ Ψi ) + σ σ η JWiu Wiu + ηiu σiQ
u uT u
σΨ JW u Ψu
u u
ci ,ηi i 2 Q Q

u u u T u 1 u u uT
−ρJ + (eΨ Ψ ) JΨ + σΨ JΨu Ψu σΨ . (59.7)
2
The technical traders’ optimal demand for stock is given by
ηiu = f0iu + f1iu Dt + f2iu Pt + f3iu Ait , (59.8)
where f0iu , f1iu , f2iu and f3iu are constants for i = 1, 2.
We have now Equations (59.5) and (59.8), the demands of stock by informed and technical
investors, the market clearing condition requires
(1 − w1 − w2 )η i + w1 η1u + w2 η2u = 1 + θt ,
or equivalently,
2
X
(1−w1 −w2 )[f0i +f1i Dt +f2i πt +f3i θt +f4i A1t +f5i A2t ]+ wj [f0ju +f1ju Dt +f2ju Pt +f3ju Ajt ] = 1+θt .
j=1
(59.9)

Substitute Pt in (59.2) into above, and by matching coefficients of state variables, we obtain
 P2
i ju ju
(1 − w1 − w2 )f0 + Pj=1 wj (f0 + f2 p0 ) = 1,

(1 − w1 − w2 )f1i + 2j=1 wj (f1ju + f2ju p1 ) = 0,





(1 − w − w )f i + P2 w f ju p = 0,

1 2 2 j 2 2
i
Pj=12 ju (59.10)

 (1 − w1 − w )f
2 3 + j=1 w f
j 2 p 3 = 1,
P2
wj f2ju p4 = 0,

i 1u
(1 − w − w )f + w f +



 1 2 4 1 3 j=1
(1 − w − w )f i + w f 2u + P2 w f ju p = 0,

1 2 5 2 3 j=1 j 2 5
The solution to Equation (59.10) determines the coefficients p0 , p1 , p2 , p3 , p4 and p5 for the
price function of (59.2). It is easy to show that a unique solution exists under general
conditions. This implies that the Proposition holds. Q.E.D.
If all the investors are informed, i.e., w = 0, there is an explicit solution to the problem
with the parameters in Equation (59.2) given as
h 2 2
i
σD σπ
p0 = Φ + p∗0 = r(r+ααDπ)(r+α
π̄
π)
− (r+αD )2
+ (r+αD )2 (r+απ )2
, (59.11)
p1 = p∗D = 1
r+αD
, (59.12)
p2 = p∗π = 1
(r+αD )(r+απ )
, (59.13)
with p3 < 0 and p4 = p5 = 0. Q.E.D.
59.2.1.3 Brock, Lakonishok, and LeBaron (1992): An update
Since Brock, Lakonishok, and LeBaron (1992) is one of the major studies to provide
empirical support to technical analysis, it is of interest to examine how the strategies perfor-
mance after publication. The original sample period is from January 2, 1897 to December
31, 1986, and in this review, we extend the sample period to December 31, 2020.
We consider three trading rules: variable-length moving average (VMA), fixed-length
moving average (FMA), and trading range break-out (TRB). To construct these strategies,
we first calculate the moving average signals. The moving average (MA) at day t of lag L is
defines as
Pt−L−1 + Pt−L−2 + · · · + Pt−1 + Pt
At,L = , (59.14)
L
where Pt is the market index level (DJIA) at day t.
The first rule, variable-length moving average (VMA), generates buy (sell) signals when
the short moving average is above (below) the long moving average by an amount larger
than the band. With a band of zero, this method classifies all trading days into either buys
or sells. Mathematically, the returns on VMA strategy are
At−1,Lshort
n M KT ,
t if At−1,Llong
≥ 1 + b;
R̃V M A,t = (59.15)
−M KTt , otherwise,

where M KTt is the return on the market index at day t, Llong (Lshort ) is the lag window for
the long (short) moving average signal, and b is the band.
The second rule, fixed-length moving average (FMA) differs from VMA in the sense that
it puts emphasis on the crossing of the moving averages. In specific, FMA trading rule
initials a buy (sell) signal when the short moving average cuts the long moving average from
below (above). Returns during the next ten days are recorded, and other signals during this
period are ignored. As for the final rule, trading range break-out (TRB), a buy signal is
initiated if the price penetrates the resistance level, which is defined as the local maximum
price. Meanwhile, a sell signal is generated if the price goes below the support level, which
is the local minimum. Define M AXt,L and M INt,L the maximum and minimum price over
the past L days include day t, respectively.
M AXt,L = max{Pt , Pt−1 , · · · , Pt−L+1 }, (59.16)
M INt,L = min{Pt , Pt−1 , · · · , Pt−L+1 }. (59.17)

Then, the returns on TRB strategy are
Pt−1
n M KTt , if M AXt−1,L
≥ 1 + b;
R̃T RB,t = Pt−1 (59.18)
−M KTt , if M INt−1,L
≤ 1 − b.
While numerous variations of the above rules with different parameters are used in prac-
tice, we follow Brock, Lakonishok, and LeBaron (1992) to focus on the simplest and the
most popular ones. For VMA and FMA, the windows for the short-long moving average
are: 1-50, 1-150, 5-150, 1-200, and 2-200. For TRB, the resistance (support) level is defined
as the maximum (minimum) price over the past 50, 150, and 200 days. In addition, all the
rules are implemented with and without a one percent band.
Table 59.1 reports the summary statistics for the entire series and five subsamples for
the daily and 10-day returns on the DJIA. Returns are defined as log differences of the
index level. Panel A shows that the index yields the highest average daily return of 0.033%
during the most recent subperiod from 1987 to 2020. Meanwhile, the lowest skewness (-1.58)
and the greatest kurtosis (41.43) are also observed in this subperiod, indicating that the
DJIA return represents a strong non-asymmetric and leptokurtic feature. Panel B reports
the statistics for 10-day nonoverlapping returns. It is interesting to note that the resulting
kurtosis is much lower than that for the daily returns for all subperiods.
Table 59.2 reports the performance of the 10 variable-length moving average (VMA)
trading rules in the full sample and five subperiods. The second column specifies the trading
rules. For instance, (1, 50, 0.01) indicates that the short period is one day, the long period
is 50 days, and the band is 1%. “N (Buy)” and “N (Sell)” are the number of the buy and sell
signals. It is expected to see that the trading rules with zero band generally initiates more
trading signals. The last column reports the difference between the buy and sell returns.
Panel A shows that during the full sample over 1897 to 2020, all these differences are positive
and the resulting t-stats are highly significant. The column of “Buy” and “Sell” shows that
the average return for buy and sell signals is positive and negative, respectively, for each

of the ten trading rules. Averaging across the rules, the last row shows that the mean buy
return is 0.038%, while the mean sell return is -0.01%.
Panel B shows the performance in several subperiods. To save space, we only report the
results for (1, 150, 0) rule. During each of the first four subperiods, the average return on
buy (sell) signals is positive (negative), and the difference between the buy and sell returns
is positive and significant, which is consistent with the results of Brock, Lakonishok, and
LeBaron (1992). However, during the most recent subperiod from 1987 to 2020, the average
return on buy signals decreases almost by half to only 0.021%. Worse still, the average
return on sell signals turns positive (0.064%). As a result, the buy-sell difference is negative,
which indicates that the trading rule does not continue to generate a profitable return in
this subsequent period.
Table 59.3 provides the results for other VMA trading strategies over 1987 to 2020. The
results are similar. All of the ten VMA trading rules generate positive returns on the sell
signals, and the resulting buy-sell differences are negative. This result is in stark contrast to
the strong profitability of technical trading rules over the early period examined by Brock,
Lakonishok, and LeBaron (1992).
Table 59.4 shows the performance of the fixed-length moving average (FMA) rules. Panel
A shows that over the full sample from 1897 to 2020, the sell returns are negative, and the
buy-sell differences are always positive for all the 10 FMA rules. In sharp contrast to Panel
A, Panel B shows that FMA rules have experienced a huge decrease in profitability during
the post-1987 subperiod. 9 out of the 10 rules yield positive returns on sell signals, and
7 rules generate negative buy-sell returns. On average, the buy-sell difference is negative
(-0.029 %).
Table 59.5 reports the performance of the trading range break (TRB) rules. The results
are similar to that of the two moving average rules: although TRB rules show some prof-
itability over the full sample, it fails to generate positive buy-sell return in the post-1987
subperiod.
In short, we confirm Brock, Lakonishok, and LeBaron’s (1992) result that the technical
trading rules exhibit strong profitability during 1897 to 1986. However, our updated results
indicate that it is unlikely that the same strategies can earn any profits during the post-1987
period. The publication effect may be one of the main causes. Today, it seems that different
strategies are required to make abnormal profits in the changing world.
59.2.2 Portfolios and Other Assets
There are many studies on the use of technical analysis in other markets besides the stock
market. Taylor and Allen (1992) show that currency market seems the next largest place
where technical analysis is widely used. However, as shown by Hsu and Taylor (2014), the
predictive power, like that in the stock market, tends to decrease over time.
In general, Fung and Hsieh (2001) find that trend-following trading is of great importance
for explaining hedge fund returns. Burghardt and Walls (2011) show that a simple mechani-
10

cal trading rule based on the MAs can yield favorable returns in trading futures contracts and
the return correlation with the managed futures index exceeds 70%. Olszweski and Zhou
(2014) show that combining both technicals and macro/fundamentals offers a significant
improvement in risk-adjusted returns.
59.2.2.1 Related studies
Recently, Moskowitz, Ooi and Pedersen (2012) provide evidence on momentum across
asset classes, that is, the past values have predictive power. However, Huang, Li, Wang
and Zhou (2020) argue that, while asset classes may have predictability, the predictability
is not simply fixed at the 12-month horizon with the past 12-month return as the sole
predictor. They show that asset-by-asset time-series regressions reveal little evidence of 12-
month momentum, both in- and out-of-sample. From an investment perspective, a strategy
of using the 12-month momentum has similar performance with a strategy that is based on
historical sample mean and does not require predictability. In other words, while Huang, Li,
Wang and Zhou (2020) do not rule out the predictive power of past returns on future values
for a wide range of assets, their study merely points out that the predictability can be much
more complex than using just the past 12-month return.
Going beyond technical indicators, Filippou, Rapach, Taylor and Zhou (2020) show that
the currency market is predictable with country characteristics, global variables, and their
interactions, and the predictability yields sizable carry trade profits. For the corporate bond
market, Guo, Lin, Wu, and Zhou (2020) provide the first predictability evidence across bond
rating along with a survey of the literature.
In summary, technical analysis appears useful not only in the stock market, but also
valuable across asset classes. However, as it is in the stock market, the predictability is
small and tends to decline over time.
59.2.2.2 Han, Yang and Zhou (2013): An update
Instead of forecasting the stock index which can have too many factors affecting its returns,
Han, Yang and Zhou (2013) provide perhaps the first application of a moving average timing
strategy to portfolios, which are sorted by volatility or other characteristics of the stocks.
We replicate this study below to see whether the publication effct is strong enough to weaken
its performance too.
Their study is focused on portfolios sorted by volatility as stock volatility is a simple
proxy of information uncertainty. The more uncertain the future information about a stock
is, the more volatile the stock price is. The volatility sorted portfolios are also of interest
from the theoretical perspectives about technical analysis. Rational models, such as Brown
and Jennings (1989), show that rational investors can gain from forming expectations based
on historical prices and this gain is an increasing function of the volatility of the asset.
For their major results, they apply the moving average (MA) timing strategy to the CRSP
NYSE/Amex volatility decile portfolios constructed based on the NYSE/Amex stocks sorted
11

into ten groups (deciles) by their annual standard deviations estimated using the daily returns
within the prior year. The portfolios are rebalanced each year at the end of the previous
year. The original sample period for the volatility decile portfolios is from July 1, 1963 to
December 31, 2009 to coincide with the Fama-French factors. In this review, we extend the
sample period to September 30, 2020.
Denote by Rjt (j = 1, . . . , 10) the returns on the volatility decile portfolios, and by Pjt
(j = 1, . . . , 10) the corresponding portfolio prices (index levels). The moving average (MA)
at day t of lag L is defined as
Pjt−L−1 + Pjt−L−2 + · · · + Pjt−1 + Pjt
Ajt,L = , (59.19)
L
which is the average price of the past L days include day t. The paper mainly considers
10-day moving averages but also examines 20-, 50-, 100- and 200-day moving averages for
robustness.
On each trading day t, if the last closing price Pjt−1 is above the MA price Ajt−1,L , the
MA strategy will invest in the decile portfolio j for the trading day t, otherwise it will invest
in the 30-day Treasury bill. So the MA provides an investment timing signal with a lag of
one day. The idea of the MA is that an investor should hold an asset when the asset price
is on an uninterrupted up trend, which may be due to a host of known and unknown factors
to the investor. However, when the trend is broken, new factors may come into play and the
investor should then sell the asset. Mathematically, the returns on the MA timing strategy
are
Rjt , if Pjt−1 > Ajt−1,L ;
n
R̃jt,L = (59.20)
rf t , otherwise,
where Rjt is the return on the j-th volatility decile portfolio on day t, and rf t is the return
on the risk-free asset, the 30-day Treasury bill.
They define the differences in return between the MA timing strategy and a buy-and-
hold strategy, R̃jt,L − Rjt , as returns on the MA portfolios (MAPs), which measure the
performance of the MA timing strategy relative to the buy-and-hold strategy. With the 10
decile portfolios, we thus obtain 10 MAPs,
MAPjt,L = R̃jt,L − Rjt , j = 1, . . . , 10. (59.21)
A MAP can also be interpreted as a zero-cost arbitrage portfolio that takes a long position
in the MA timing portfolio and a short position in the underlying volatility decile portfolio.
The abnormal performance of the MAPs indicates the profitability of the MA investment
timing strategy.
Table 59.6 reports the updated results of the original Table 1 of the paper. It shows the
basic characteristics of the returns on the decile portfolios, Rjt , the returns on the 10-day
MA timing portfolios, R̃jt,10 , and the returns on the corresponding MAPs, MAPjt,10 . The
updated results are very similar to the old results even with additional 12 year’s daily data,
which demonstrates the robustness of the MA timing strategy.
Panel A provides the average return, the standard deviation, the skewness, and the Sharpe
ratio of the buy-and-hold strategy across the ten volatility deciles. The average returns
12

are more or less an increasing function of the deciles, ranging from 10.17% per annum for
the lowest decile to 38.13% per annum for the highest decile. The last row in the table
provides the difference between the highest and the lowest deciles. For the decile portfolios,
the difference is 27.96% per annum, both statistically and economically highly significant.
Similarly, the MA timing portfolios, reported in Panel B, also have returns varying (strictly)
positively with the deciles, ranging from 17.72% to 53.14% per annum.1 In addition, the
returns on the MA timing portfolios not only are larger than those on the decile portfolios,
but also enjoy substantially smaller standard deviations. For example, the standard deviation
is 4.21% versus 7.35% for the lowest decile, and 15.06% versus 21.59% for the highest decile.
In general, the MA timing strategy yields only about 65% volatility of the decile portfolios.
As a result, the Sharpe ratios are several times higher for the MA timing portfolios than
for the volatility decile portfolios. Furthermore, while the volatility decile portfolios display
negative skewness (except for the highest volatility decile), the MA timing strategy yields
either much smaller negative skewness or positive skewness across the volatility deciles.
Panel C reports the results for the MAPs. The returns increase monotonically from 7.55%
to 15.42% per annum across the deciles (except for the highest volatility decile). While the
standard deviations are much smaller than those of Rjt in the corresponding deciles, they
are not much different from those of R̃jt,L . However, the skewness of the MAPs across all
deciles is positive and large.
Table 59.7 is similar to Table 2 of the paper but uses the up-to-date Fama-French five-
factor (FF5) model (Fama and French, 2015) instead of the three-factor model (Fama and
French, 1993). It reports the results of the daily FF5 regressions of the MAPs formed with
10-day MA timing strategy. Again, the results are very similar. The alphas or risk-adjusted
returns are even greater than the unadjusted ones ranging from 9.36% to 20.70% per annum.
The alphas also increase monotonically from the lowest volatility decile to higher volatility
deciles, except that the highest decile yields a slightly lower alpha than the ninth decile.
Nevertheless, the highest volatility decile still generates an alpha that is about twice as large
as that generated by the lowest decile (19.49 versus 9.36). The difference is reported in the
last row, which is substantial and highly significant.
Table 59.8 reports the daily FF5 regression results using only samples after the publica-
tion, i.e., from January 4, 2010 to September 30, 2020. The results are even stronger than
the ones reported in the paper for the sample period from July 1, 1963 to December 31,
2009, and the ones reported in Table 59.7 for the extended sample period. For example, the
spread between the highest and lowest deciles is 13.38%, while it is 10.13% in the extended
sample period.
In summary, the performance of the MA timing on volatility portfolio seems very robust,
and even stronger after publication. Hence, the MA timing is an example for which there
is most likely no publication effect. One possibility might be that it is much more costly to
construct and trade the volatility portfolio than trading the market index.
1 To put this in perspective, the equal-weighted NYSE/Amex index has an average return of 19.22% per annum, and a
standard deviation of 17.14% per annum.
13

59.3 Cross-section Predictability
Cross-section predictability is about the relative performance of assets, and hence the
econometric tools will be different from time-series regressions and the like. Instead, it relies
on cross-section regressions, panel models and their extensions.
59.3.1 Fama-MacBeth Regressions
The cross-section predictive power of a firm characteristic can be assessed by sorting the
variable across firms, or by running Fama and MacBeth’s (1973) regressions, in which asset
returns are regressed on firm specific variables across firms. The latter is the most used
method for examining the cross-section predictive power of more than one firm characteris-
tics. Haugen and Baker (1996) provide excellent illustration of implementing the latter.
The size and the book-to-market are the earlier well known predictors, and the momen-
tum effect of Jegadeesh and Titman (1993) is the next most well known. Jagannathan,
Schaumburg and Zhou (2010), Nagel (2013), and Lewellen (2015) review more predictive
variables and related studies. Recent research, however, makes use of more complex models
and machine learning tools, to be discussed in the next subsection.
Han, Zhou and Zhu (2016) is the first to study the predictive power of technical indicators
in the cross-section. Based purely on the MAs over various horizons, they find that the
resulting trend factor earns a high Sharpe ratio, beating almost all known fundamental
factors. Because of large individual trading volume in China, Liu, Zhou and Zhu (2021a)
propose a new trend factor accounting for the trading volume, and find that it improves the
usual trend factor substantially and explains almost all the know anomalies in the Chinese
stock market. Avramov, Kaplanski and Subrahmanyam (2021) propose a novel moving
average distance as a predictor for the cross-section of stock returns. They find that the
resulting spread earns annualized value-weighted alphas around 9%, and the predictability
goes beyond momentum, 52-week highs, profitability, and other prominent anomalies.
59.3.1.1 Han, Zhou and Zhu (2016): An update
It is of interest to examine how the trend factor performs after its publication. To con-
struct the trend factor, the paper first calculates the MA prices on the last trading day of
each month t, which is defined as
t t t t
Pj,d−L+1 + Pj,d−L+2 + · · · + Pj,d−1 + Pj,d
Ajt,L = , (59.22)
L
t
where Pjd is the closing price for stock j on the last trading day d of month t, and L is the
lag length. Then, the moving average prices are normalized by the closing price on the last
trading day of the month,
Ajt,L
Ãjt,L = t . (59.23)
Pjd
14

To predict the monthly expected stock returns cross-sectionally, a two-step procedure is
employed. In the first step, a cross-section regression of stock returns on observed normalized
MA signals is run in each month t to obtain the time-series of the coefficients on the signals,
X
rj,t = β0,t + βi,t Ãjt−1,Li + j,t , j = 1, · · · , n, (59.24)
i
where
rj,t = rate of return on stock j in month t,
Ãjt−1,Li = trend signal at the end of month t − 1 on stock j with lag Li ,
βi,t = coefficient of the trend signal with lag Li in month t,
β0,t = intercept in month t,
and n is the number of stocks. It should be noted that only information in month t − 1 or
prior is used above to regress returns in month t. For the signals in the above regressions,
the paper considers MAs of lag lengths 3-, 5-, 10-, 20-, 50-, 100-, 200-, 400-, 600-, 800-, and
1,000-days. These MA signals roughly indicate the daily, weekly, monthly, quarterly, 1-year,
2-year, 3-year, 4-year, and 5-year price trends of the underlying stock.
Then, in the second step, as in Haugen and Baker (1996), we estimate the expected return
for month t + 1 from X
Et [rj,t+1 ] = Et [βi,t+1 ]Ãjt,Li , (59.25)
i
where Et [rj,t+1 ] is our forecasted expected return on stock j for month t + 1, and Et [βi,t+1 ]
is the estimated expected coefficient of the trend signal with lag Li , and is give by
12
1 X
Et [βi,t+1 ] = βi,t+1−m , (59.26)
12 m=1
which is the average of the estimated loadings on the trend signals over the past 12 months.
Finally, to construct the trend factor, stocks are sorted into five portfolios by their fore-
casted expected returns. The portfolios are equal-weighted and rebalanced every month.
The return difference between the quintile portfolio of the highest expected returns and the
quintile portfolio of the lowest is defined as the return on the trend factor. Intuitively, the
trend factor buys stocks that are forecasted to yield the highest expected returns (Buy High)
and shorts stocks that are forecasted to yield the lowest expected returns (Sell Low).
Table 59.9 reproduces the original Table 1 in the paper with the sample extended to
December 2019 and replaces the momentum factor with Fama-French UMD factor. It also
includes the trend factor constructed using Fama-French approach (Table 10 in the paper).
The trend factor is constructed by excluding stocks with prices below $5 (price filter) and
stocks that are in the smallest decile sorted with NYSE breakpoints (size filter) at the end
of each month. It reports the summary statistics of the trend factor, short-term reversal
factor (SREV), momentum factor (UMD), long-term reversal factor (LREV), as well as the
Fama-French three factors, eg, Market, small minus big (SMB), and high minus low (HML).
15

The average monthly return of the trend factor from June 1930 through December 2019 is
1.56% per month (compared to 1.63% in the paper), more than doubling the average return
of any of the other factors including SREV, whose average return is the highest among the
other factors but is only 0.77% per month. The standard deviation of the trend factor is
unchanged from data up to December 2014 (3.42% versus 3.45%), lower than that of any
other factor except SMB. As a result, the Sharpe ratio of the trend factor is much higher than
those of the other factors. For example, the trend factor has a Sharpe ratio of 0.46, whereas
the next highest Sharpe ratio is only 0.23 generated by SREV. The trend factor constructed
using Fama-French approach (Trend(FF)) yields a higher average return of 1.22% per month
and lower standard deviation of 2.90%, and thus delivers a higher Sharpe ratio of 0.42. As
expected, the momentum factor has a very large negative skewness (−3.02) and very large
kurtosis (27.4). In contrast, the trend factor has a large positive skewness (1.44 and 1.61)
and large kurtosis (11.3 and 11.2), indicating a fat right tail, and great chances for large
positive returns.
Table 59.10 reports the same summary statistics of these factors after publication of
the paper, i.e., from January 2015 to December 2019. In this short five-year period, the
trend factor does not perform as well as it used to. It does only slightly better than the
momentum factor, but worse than the short-term reversal factor. In contrast, the market
performs remarkable well during the last five years, delivering 0.89% a month on average.
What drives the relatively underperformance of the trend factor? The first may be the
publication effect emphasized by Schwert (2003), and found true for many anomalies by
McLean and Pontiff (2016). While the publication effect may play a role here, there seem
to exist other fundamental causes. The long-term reversal factor is particularly bad, losing
−0.72% a month. The SMB and HML factors lose money significantly over the years too,
whereas the market outperforms all other factors substantially, a very rare phenomenon. The
trend now comes down from the usual number one spot to the third. Although it may have
something to do with the poor performance of the long-term reversal factor, the impact is
likely small since the trend factor captures trends over all horizons. On the other hand, while
many anomalies are weakened after publication, do some come back in this extraordinary
market environment? Note that our update period here is after the publication of McLean
and Pontiff (2016) whose data ends in 2013. These are open questions of future research.
59.3.2 Machine Learning Methods
Machine learning methods, in contrast to the Fama-MacBeth regressions, can incorpo-

rate more predictors and can examine nonlinear relations. Freyberger, Neuhierl and Weber
(2020), Gu, Kelly and Xiu (2020) and Kozak, Nagel and Santosh (2020) are recent examples
of using firm characteristics to predict the cross-section of stock returns, which can easily
adapted to using the technical indicators of Han, Zhou and Zhu (2016) or using these in
conjunction of the firm characteristics. Typically, such studies assess the economic value of
forecasting by examining the performance of a spread portfolio similar to the Fama-MacBeth
regression case except that now the estimated expected returns are computed from the new
methods rather than from the Fama-MacBeth regression. Liu, Zhou and Zhu (2021b) use a
16

genetic programming approach to maximize the Sharpe ratio of the spread portfolio in esti-
mating the model. As it turns out, this economically motivated objective helps to improve
substantially the performance.
The above machine learning methods rely on panel data. Han, He, Rapach and Zhou
(2020) propose a C-Lasso approach that allows for missing and adding gradually available
new characteristics. This may be particularly useful for applying to high-frequency technical
indicators which are available only in recent decades. It will be of interest to apply all these
machine learning methods to more pure technical indicators, those depend on price and
trading volume only.
59.4 Conclusion
This article reviews primarily the use of technical analysis in the stock market, though
we do discuss briefly applications in other asset classes. The empirical evidence shows that
technical analysis is useful and has significant predictive power on stock returns both in the
time-series and in the cross-section.
However, there are three important caveats. First, the degree of predictability is typi-
cally small, and has to be well managed to be exploited profitably. Although cross-section
predictability is stronger than time-series predictability, the trading cost of the former is
much higher. Second, the predictability is time-varying, and tends to decrease over time.
A particular profitable predictability strategy often subjects to a publication effect (Schw-
ert, 2003 and McLean and Pontiff, 2016) that, its superior performance often disappears,
reverses, or attenuates after its publication. We have updated three major studies with data
after publication, and find that two of them have weaker results than before publication.
Third, all trading rules, technical strategies included, are subject to a tournament effect. No
matter what the state of art machine learning tools or artificial intelligence packages all the
traders are using, about half of them will fail to beat the market return, while the others
will probably outperform.
17

References
Alexander, S. S., 1961. Price movements in speculative markets: Trends or random walks,
Industrial Management Review 2, 7–26.
Allen, F., Karjalainen, R., 1999. Using genetic algorithms to find technical trading rules.
Journal of Financial Economics 51, 245–271.
Avramov, D., Kaplanski, G., Subrahmanyam, A., 2021, Moving Average Distance as a
Predictor of Equity Returns. Review of Financial Economics, Forthcoming, Available
at SSRN: https://ssrn.com/abstract=3111334.
Berutich, J.M., Lopez, F., Luna, F., Quintana, D., 2016. Robust technical trading strategies
using GP for algorithmic portfolio selection. Expert Systems with Applications 46, 307–
315.
Bessembinder, H., Chan, K., 1995. The profitability of technical trading rules in the Asian
stock markets. Pacific-Basin Finance Journal 3, 257–284.
Bessembinder, H., Chan, K., 1998. Market efficiency and the returns to technical analysis.
Financial Management 27, 5–17.
Brock, W., Lakonishock, J., LeBaron, B., 1992. Simple technical trading rules and the
stochastic properties of stock returns. Journal of Finance 47, 1731–1764.
Brown, D. P., and Jennings, R. H., 1989. On technical analysis. Review of Financial Studies
2, 527–551.
Brogaard, J., Zareei A., 2019. Machine learning and the stock market. Working paper,
Available at SSRN 3233119.
Burghardt, G., Walls, B., 2011. Managed futures for institutional investors: Analysis and
portfolio construction. John Wiley & Sons.
Campbell, J.Y., Thompson, S.B., 2008. Predicting excess stock returns out of sample: Can
anything beat the historical average?. Review of Financial Studies 21, 1509–1531.
Chavarnakul, T. and Enke, D., 2009. A hybrid stock trading system for intelligent technical
analysis-based equivolume charting. Neurocomputing 72, 3517–3528.
Chen, J., Tang, G., Yao, J., Zhou, G., 2019. Investor attention and stock returns. Journal of
Financial and Quantitative Analysis, Forthcoming, Available at SSRN: https://ssrn.com/abstract=3
Chinco, A., Clark-Joseph, A.D., Ye, M., 2019. Sparse signals in the cross-section of returns.
The Journal of Finance 74 , 449–492.
Cootner, P.H., 1962. Stock prices: Ramdom vs. systematic changes. Industrial Manage-
ment Review 3, 24–45.
Covel, M., 2005. Trend Following: How Great Traders Make Millions in Up or Down
Markets. Prentice-Hall, New York.
18

Cowles, A., 1933. Can stock market forecasters forecast? Econometrica 1, 309–324.
Dale, C., Workman, R., 1980. The arc sine law and the treasury bill futures market.
Financial Analysts Journal 36, 71–74.
Detzel, A., Liu, H., Strauss, J., Zhou, G. and Zhu, Y., 2021. Learning and predictability via
technical analysis: Evidence from bitcoin and stocks with hard-to-value fundamentals.
Financial Management 50, 107–137.
Dong, Xi and Li, Yan and Rapach, David and Zhou, Guofu, Anomalies and the Expected
Market Return (February 24, 2021). Baruch College Zicklin School of Business Research
Paper No. 2020-02-02, Available at SSRN: https://ssrn.com/abstract=3562774.
Fama, E.F., Blume, M.E., 1966. Filter rules and stock-market trading. The Journal of
Business 39, 226–241.
Fama, E.F., French, K.R., 1993. Common risk factors in the returns on stocks and bonds.
Journal of financial economics 33, 3–56.
Fama, E.F., French, K.R., 2015. A five-factor asset pricing model. Journal of Financial
Economics 116, 1–22.
Fama, E.F., MacBeth, J.D., 1973. Risk, return, and equilibrium: Empirical tests. Journal
of political economy 81, 607–636.
Filippou, I., Rapach, D., Taylor, M.P., Zhou, G., Exchange Rate Prediction with Ma-
chine Learning and a Smart Carry Portfolio (September 10, 2020). Available at SSRN:
https://ssrn.com/abstract=3455713.
Freyberger, J., Neuhierl, A. and Weber, M., 2020. Dissecting Characteristics Nonparamet-
rically. Review of Financial Studies 33, 2326–2377.
Fung, W., Hsieh, D.A., 2001. The risk in hedge fund strategies: Theory and evidence from
trend followers. Review of Financial Studies 14, 313–341.
Garg, A., Goulding, C.L., Harvey, C.R. and Mazzoleni, M., 2020. Breaking Bad Trends.
Available at SSRN: https://ssrn.com/abstract=3594888.
Gencay, R., 1998a. Optimization of technical trading strategies and the profitability in
security markets. Economics Letters 59, 249–254.
Gencay, R., 1998b. The predictability of security returns with simple technical trading
rules. Journal of Empirical Finance 5, 347–359.
Gencay, R., Stengos, T., 1998. Moving average rules, volume and the predictability of
security returns with feedforward networks. Journal of Forecasting 17, 401–414.
Gu, S., Kelly, B., Xiu, D., 2020. Empirical asset pricing via machine learning, Review of
Financial Studies 33, 2223–2273.
19

Guo, X., Lin, H., Wu, C., Zhou, G., Extracting Information from Corporate Bond Yields: A
Machine Learning Approach (December 2, 2020). Available at SSRN: https://ssrn.com/abstract=37
Han, Y., He, A., Rapach, D., Zhou, G., 2020. Firm characteristics and expected stock
returns. Available at SSRN 3185335.
Han, Y., Yang, K., Zhou, G., 2013. A new anomaly: The cross-sectional profitability of
technical analysis. Journal of Financial and Quantitative Analysis, 1433–1461.
Han, Y., Zhou, G., Zhu, Y. 2016. A trend factor: Any economic gains from using informa-
tion over investment horizons?. Journal of Financial Economics 1222, 352–375.
Harvey, C., 2021, Why is systemic investing important, journal of systemic investing 1,i–iv.
Haugen, R.A., Baker, N.L., 1996. Commonality in the determinants of expected stock
returns. Journal of financial economics 41, 401–439.
Hsu, P.H., Taylor, M.P., Forty Years, Thirty Currencies and 21,000 Trading Rules: A
Large-Scale, Data-Snooping Robust Analysis of Technical Trading in the Foreign Ex-
change Market (June 2014). CEPR Discussion Paper No. DP10018, Available at SSRN:
https://ssrn.com/abstract=2501490.
Huang, D., Jiang, F., Tu, J.. Zhou, G., 2015. Investor sentiment aligned: A powerful
predictor of stock returns. Review of Financial Studies 28, 791–837.
Huang, D., Li, J., Wang, L., Zhou, G., 2020. Time series momentum: Is it there?. Journal
of Financial Economics 135, 774–794.
Ito, A., 1999. Profits on technical trading rules and time-varying expected returns: evidence
from Pacific-Basin equity markets. Pacific-Basin Finance Journal 7, 283–330.
James, F.E., 1968. Monthly Moving Averages–An Effective Investment Tool?. Journal of
financial and quantitative analysis, 315–326.
Jegadeesh, N., Titman, S., 1993. Returns to buying winners and selling losers: implications
for stock market efficiency. Journal of Finance 48, 65–91.
Jagannathan, R., Schaumburg, E., Zhou, G., 2010. Cross-sectional asset pricing tests.
Annual Review of Financial Economics 2, 49–74.
Jensen, M.C., Benington, G.A., 1970. Random walks and technical theories: some addi-
tional evidence. Journal of Finance 25, 469–482.
Jiang, F., Lee, J., Martin, X., Zhou, G., 2019. Manager sentiment and stock returns.
Journal of Financial Economics 132, 126–149.
Kelly, B. and Pruitt, S., 2013. Market expectations in the cross-section of present values.
Journal of Finance, 1721–1756.
Kim, K.J., Han, I., 2000. Genetic algorithms approach to feature discretization in artificial
neural networks for the prediction of stock price index. Expert systems with Applica-
tions 19, 125–132.
20

Koijen, R.S., Van Nieuwerburgh, S., 2011. Predictability of returns and cash flows. Annual
Review of Financial Economics 3, 467–491.
Koza, J.R., 1992. Genetic Programming: On the Programming of Computers by Means of
Natural Selection. MIT Press, Cambridge.
Kozak, S., Nagel, S., Santosh, S., 2020. Shrinking the cross-section. Journal of Financial
Economics 135, 271–292.
Kwon, K.Y., Kish, R.J., 2002. A comparative study of technical trading strategies and
return predictability: an extension of Brock, Lakonishok, and LeBaron (1992) using
NYSE and NASDAQ indices. The Quarterly Review of Economics and Finance 42,
611–631.
Lee, M.C., 2009. Using support vector machine with a hybrid feature selection method to
the stock trend prediction. Expert Systems with Applications 36, 10896–10904.
Levy, R.A., 1967. Relative strength as a criterion for investment selection. Journal of
Finance 22, 595–610.
Lewellen, J. 2015. The Cross-section of Expected Stock Returns. Critical Finance Review
4, 1–44.
Li, Y., Ng, D.T., Swaminathan, B., 2013. Predicting market returns using aggregate implied
cost of capital. Journal of Financial Economics 110, 419–436.
Liu, Y., Zhou, G., Zhu, Y., 2020, Maximizing the Sharpe Ratio: A Genetic Programming
Approach. Available at SSRN 3726609.
Liu, Y., Zhou, G., Zhu, Y., 2021a, Trend Factor in China: The Role of Large Individual
Trading. Available at SSRN: https://ssrn.com/abstract=3402038.
Liu, Y., Zhou, G., Zhu, Y., 2021b, Maximizing the Sharpe Ratio: A Genetic Programming
Approach. Available at SSRN: https://ssrn.com/abstract=3726609.
Lo, A.W., Hasanhodzic, J., 2009. The Heretics of Finance: Conversations with Leading
Practitioners of Technical Analysis. Bloomberg Press.
Mills, T.C., 1997. Technical analysis and the London Stock Exchange: testing trading rules
using the FT30. International Journal of Finance and Economics 2, 319–331.
McLean, R.D., Pontiff, J., 2016. Does academic research destroy stock return predictabil-
ity?. The Journal of Finance 71, 5–32.
Moskowitz, T.J., Ooi, Y.H., Pedersen, L.H., 2012. Time series momentum. Journal of
financial economics 104, 228–250.
Nagel, S., 2013. Empirical cross-sectional asset pricing. Annual Review of Financial Eco-
nomics 5, 167–199.
21

Neely, C.J., 2003. Risk-adjusted, ex ante, optimal technical trading rules in equity markets.
International Review of Economics and Finance 12, 69–87.
Neely, C.J., Rapach, D.E., Tu, J., Zhou, G., 2014. Forecasting the equity risk premium:
the role of technical indicators. Management science 60, 1772–1791.
Neely, C.J., Weller, P.A., 2001. Technical analysis and central bank intervention. Journal
of International Money and Finance 20, 949–970.
Neely, C., Weller, P., Dittmar, R., 1997. Is technical analysis in the foreign exchange market
profitable? A genetic programming approach. Journal of financial and Quantitative
Analysis, 32, 405–426.
Newey, W.K., West, K. D., 1987. A simple, positive semi-definite, heteroskedasticity and
autocorrelation consistent covariance matrix. Econometrica 55, 703–708.
Olszweski, F., Zhou, G., 2014. Strategy diversification: Combining momentum and carry
strategies within a foreign exchange portfolio. Journal of Derivatives & Hedge Funds
19, 311–320.
Rapach, D.E., Ringgenberg, M.C., Zhou, G., 2016. Short interest and aggregate stock
returns. Journal of Financial Economics 121, 46–65.
Rapach, D.E., Strauss, J.K., Zhou, G., 2010. Out-of-sample equity premium prediction:
Combination forecasts and links to the real economy. Review of Financial Studies 23,
821–862.
Rapach, D.E., Strauss, J.K., Zhou, G., 2013. International stock return predictability:
What is the role of the United States? Journal of Finance 68, 1633–1662.
Rapach, D., Zhou, G., 2013. Forecasting stock returns. Handbook of economic forecasting
2, 328–383, Elsevier.
Rapach, D.E., Zhou, G., 2020. Time-series and Cross-sectional Stock Return Forecast-
ing: New Machine Learning Methods. Machine Learning for Asset Management: New
Developments and Financial Applications, pp.1–33.
Ratner, M., Leal, R.P.C., 1999. Tests of technical trading strategies in the emerging equity
markets of Latin America and Asia. Journal of Banking and Finance 23, 1887–1905.
Schwager, J., 1993. Market Wizards: Interviews with Top Traders. Collins, New York.
Schwager, J., 1995. The New Market Wizards: Conversations with America’s Top Traders,
Wiley, New York.
Schwert, G.W., 2003. Anomalies and market efficiency. In: Constantinides, G.M., Harris,
M., Stulz, R.M. (Eds.), Handbook of the Economics of Finance, 1. Elsevier, Amsterdam,
Netherlands, pp. 939–974. chap. 15.
Smidt, S., 1965. Amateur Speculators. Ithaca, NY: Graduate School of Business and Public
Administration, Cornell University.
22

Sullivan, R., Timmermann, A., White, H., 1999. Data-snooping, technical trading rule
performance, and the bootstrap. Journal of Finance 53, 1647–1691.
Tan, Z., Quek, C., Cheng, P.Y., 2011. Stock trading with cycles: A financial application of
ANFIS and reinforcement learning. Expert Systems with Applications 38, 4741–4755.
Taylor, M.P., Allen, H., 1992. The use of technical analysis in the foreign exchange market.
Journal of international Money and Finance 11, 304–314.
Tibshirani, R., 1996. Regression shrinkage and selection via the lasso. Journal of the Royal
Statistical Society: Series B (Methodological) 58, 267–288.
Timmermann, A., 2018. Forecasting Methods in Finance. Annual Review of Financial
Economics 10, 449–479.
Van Horne, J.C., Parker, G.G.C., 1967. The random-walk theory: an empirical test. Fi-
nancial Analysts Journal 23, 87–92.
Van Horne, J.C., Parker, G.G.C., 1968. Technical trading rules: a comment. Financial
Analysts Journal 24, 128–132.
Welch, I., Goyal, A., 2008. A comprehensive look at the empirical performance of equity
premium prediction. The Review of Financial Studies 21, 1455–1508.
White, H., 2000. A reality check for data snooping. Econometrica 68, 1097–1126.
Zhou, G., Zhu, Y., Qiang, S., 2012. Asset allocation: can technical analysis add value?.
International Journal of Portfolio Analysis and Management 1, 43–58.
Zhu, Y., Zhou, G., 2009. Technical analysis: An asset allocation perspective on the use of
moving averages. Journal of Financial Economics 92, 519–544.
23

Table 59.1
Summary Statistics for Daily and 10-Day Returns
Results are presented for the full sample over 1897 to 2020 and five nonoverlapping subperiods. Returns
are measured as log differences of the level of the Dow index. 10-day returns are based on nonoverlapping
10-day periods.
Panel A: Daily Returns

Full Sample 1897-1914 1915-1938 1939-1962 1962-1986 1987-2020
N 33583 5272 7165 6570 6283 8545
Mean 0.00021 0.00011 0.00015 0.00022 0.00015 0.00033
Std. 0.01093 0.01001 0.01463 0.00742 0.00849 0.01167
Skew -0.56880 -0.51406 0.01834 -0.82081 0.23110 -1.57504
Kurtosis 24.27 9.15 12.79 12.69 5.66 41.43
Panel B: 10-Day Returns
Mean 0.00206 0.00124 0.00142 0.00219 0.00161 0.00324
Std. 0.03549 0.03403 0.04749 0.02729 0.02982 0.03210
Skew -1.04321 -0.17634 -0.83248 -1.34787 -0.10958 -1.47806
Kurtosis 14.18 4.55 8.81 12.94 4.44 11.82
24

Table 59.2
Performances of the Variable-Length Moving Average (VMA) Rules: 1897-2020
Results for daily data from 1897-2020. Rules are identified as (short, long, band) where short and long
are the short and long moving averages respectively, and band is the percentage difference that is needed
to generate a signal. “N (Buy)” and “N (Sell)” are the number of buy and sell signals reported during the
sample. Numbers in parentheses are standard t-ratios testing the difference of the mean buy and mean sell
from the unconditional 1-day mean, and buy-sell from zero. “Buy > 0” and “Sell > 0” are the fraction of buy
and sell returns greater than zero. The last row reports averages across all 10 rules. Results for subperiods
are given in Panel B.
Panel A: Full Sample

Period Test N (Buy) N (Sell) Buy Sell Buy>0 Sell>0 Buy-Sell
1897-2020 (1,50,0) 20046 13488 0.00042 -0.00012 0.53662 0.48880 0.00054
(2.24) (-2.92) (4.47)
(1,50,0.01) 16452 10245 0.00050 -0.00016 0.53805 0.49175 0.00066
(2.24) (-3.07) (4.60)
(1,150,0) 21184 12250 0.00036 -0.00007 0.53573 0.48955 0.00043
(1.65) (-2.37) (3.48)
(1,150,0.01) 19338 10486 0.00038 -0.00013 0.53801 0.49199 0.00051
(1.73) (-2.60) (3.75)
(5,150,0) 21200 12234 0.00033 -0.00001 0.53552 0.48913 0.00035
(1.32) (-1.90) (2.79)
(5,150,0.01) 19245 10438 0.00036 -0.00006 0.53666 0.49272 0.00041
(1.40) (-2.10) (3.03)
(1,200,0) 21707 11677 0.00037 -0.00012 0.53577 0.49165 0.00049
(1.79) (-2.70) (3.88)
(1,200,0.01) 20178 10302 0.00037 -0.00013 0.53717 0.49534 0.00050
(1.68) (-2.63) (3.73)
(2,200,0) 21722 11662 0.00035 -0.00008 0.53508 0.49082 0.00044
(1.59) (-2.41) (3.47)
(2,200,0.01) 20150 10287 0.00035 -0.00008 0.53593 0.49344 0.00043
(1.45) (-2.25) (3.21)
Average 0.00038 -0.00010 0.00048
Panel B: Subperiods
1897-1914 (1,150,0) 2925 2198 0.00039 -0.00029 0.53231 0.49727 0.00068
(1.26) (-1.52) (2.41)
1915-1938 (1,150,0) 4217 2948 0.00053 -0.00041 0.55229 0.50068 0.00094
(1.36) (-1.73) (2.67)
1939-1962 (1,150,0) 4242 2328 0.00036 -0.00004 0.54267 0.47509 0.00040
(0.96) (-1.43) (2.07)
1962-1986 (1,150,0) 3672 2611 0.00041 -0.00022 0.52206 0.51972 0.00063
(1.49) (-1.87) (2.91)
1987-2020 (1,150,0) 6193 2352 0.00021 0.00064 0.52931 0.45323 -0.00043
(0.61) (1.16) (-1.53)
25

Table 59.3
Performances of the Variable-Length Moving Average (VMA) Rules: 1987-2020
Results for daily data from 1987-2020. Rules are identified as (short, long, band) where short and long
are the short and long moving averages respectively, and band is the percentage difference that is needed
to generate a signal. “N (Buy)” and “N (Sell)” are the number of buy and sell signals reported during the
sample. Numbers in parentheses are standard t-ratios testing the difference of the mean buy and mean sell
from the unconditional 1-day mean, and buy-sell from zero. “Buy > 0” and “Sell > 0” are the fraction of
buy and sell returns greater than zero. The last row reports averages across all 10 rules.
Period Test N (Buy) N (Sell) Buy Sell Buy>0 Sell>0 Buy-Sell

1987-2020 (1,50,0) 5666 2879 0.00025 0.00048 0.52983 0.45780 -0.00024
(0.40) (0.63) (-0.89)
(1,50,0.01) 4659 2076 0.00015 0.00052 0.52350 0.45568 -0.00038
(0.50) (0.85) (-1.17)
(1,150,0) 6193 2352 0.00021 0.00064 0.52931 0.45323 -0.00043
(0.61) (1.16) (-1.53)
(1,150,0.01) 5661 1874 0.00020 0.00063 0.53030 0.45411 -0.00043
(0.51) (1.05) (-1.35)
(5,150,0) 6218 2327 0.00021 0.00064 0.53040 0.45552 -0.00043
(0.60) (1.13) (-1.50)
(5,150,0.01) 5634 1837 0.00022 0.00064 0.53088 0.46053 -0.00042
(-0.49) (1.02) (-1.30)
(1,200,0) 6398 2147 0.00028 0.00047 0.53329 0.46344 -0.00019
(0.25) (0.52) (-0.67)
(1,200,0.01) 5949 1771 0.00025 0.00074 0.53169 0.46132 -0.00049
(0.55) (1.22) (-1.54)
(2,200,0) 6402 2143 0.00026 0.00053 0.53249 0.46104 -0.00027
(0.36) (0.73) (-0.94)
(2,200,0.01) 5938 1762 0.00025 0.00070 0.53116 0.46254 -0.00045
(0.50) (1.11) (-1.40)
Average 0.00023 0.00060 -0.00037
26

Table 59.4
Performances of the Fixed-Length Moving Average (FMA) Rules
Cumulative returns are reported for fixed 10-days periods after signals. Rules are identified as (short, long,
band) where short and long are the short and long moving averages respectively, and band is the percentage
difference that is needed to generate a signal. The last row reports averages across all 10 rules. The sample
period in Panel A is from 1897 to 2020. The sample period in Panel B is from 1987 to 2020.
Test N (Buy) N (Sell) Buy Sell Buy>0 Sell>0 Buy-Sell

Panel A: Sample period over 1897-2020
(1,50,0) 452 500 0.0029 -0.0017 0.6150 0.5080 0.0047
(0.51) (-2.37) (2.02)
(1,50,0.01) 411 411 0.0043 -0.0024 0.6253 0.5012 0.0067
(1.28) (-2.53) (2.70)
(1,150,0) 232 271 0.0046 -0.0002 0.5690 0.5609 0.0048
(1.07) (-1.07) (1.52)
(1,150,0.01) 225 206 0.0045 -0.0025 0.6267 0.5728 0.0070
(1.03) (-1.83) (2.04)
(5,150,0) 195 201 0.0061 -0.0011 0.6103 0.5920 0.0072
(1.58) (-1.25) (2.01)
(5,150,0.01) 185 176 0.0048 -0.0030 0.6162 0.5455 0.0078
(1.05) (-1.88) (2.09)
(1,200,0) 174 219 0.0043 -0.0010 0.6034 0.5434 0.0053
(0.81) (-1.28) (1.47)
(1,200,0.01) 173 172 0.0045 -0.0041 0.6243 0.5291 0.0087
(0.91) (-2.29) (2.27)
(2,200,0) 160 203 0.0039 -0.0024 0.6125 0.5517 0.0062
(0.63) (-1.78) (1.66)
(2,200,0.01) 167 159 0.0006 -0.0047 0.5389 0.5220 0.0052
(0.53) (-2.38) (1.33)
Average 0.0040 -0.0023 0.0064
Panel B: Sample period over 1987-2020
(1,50,0) 110 153 0.0020 0.0042 0.6909 0.6078 -0.0022
(0.41) (0.36) (-0.55)
(1,50,0.01) 115 108 0.0004 0.0059 0.6087 0.6389 -0.0055
(0.94) (0.85) (-1.27)
(1,150,0) 75 84 0.0003 0.0016 0.5200 0.5357 -0.0013
(0.79) (0.48) (-0.25)
(1,150,0.01) 64 59 -0.0021 0.0060 0.5625 0.6610 -0.0081
(-1.33) (0.66) (-1.40)
(5,150,0) 62 61 0.0028 -0.0022 0.5806 0.6066 0.0050
(0.09) (-1.32) (0.87)
(5,150,0.01) 57 51 0.0009 0.0003 0.5088 0.5490 0.0006
(0.55) (0.65) (0.10)
(1,200,0) 60 64 0.0026 0.0029 0.5667 0.5469 -0.0003
(0.14) (0.08) (-0.05)
(1,200,0.01) 47 51 -0.0012 0.0074 0.5745 0.6863 -0.0086
(-0.95) (0.93) (-1.33)
(2,200,0) 51 62 0.0025 0.0020 0.5882 0.5806 0.0005
(0.16) (0.30) (0.09)
(2,200,0.01) 52 45 -0.0025 0.0065 0.5000 0.7111 -0.0090
(-1.28) (0.69) (-1.38)
Average 0.0006 0.0035 -0.0029
27

Table 59.5
Performances of the Trading Range Break (TRB) Rules
Cumulative returns are reported for fixed 10-days periods after signals. Rules are identified as (short, long, band)
where short and long are the short and long moving averages respectively, and band is the percentage difference that
is needed to generate a signal. The last row reports averages across all 10 rules. The sample period in Panel A is
from 1897 to 2020. The sample period in Panel B is from 1987 to 2020.
Test N (Buy) N (Sell) Buy Sell Buy>0 Sell>0 Buy-Sell

Panel A: Sample period over 1897-2020
(1,50,0) 1031 521 0.0039 0.0012 0.5790 0.5413 0.0027
(1.65) (-0.56) (1.43)
(1,50,0.01) 333 315 0.0073 -0.0005 0.6156 0.5460 0.0079
(2.69) (-1.29) (2.82)
(1,150,0) 761 255 0.0037 0.0005 0.5782 0.5098 0.0032
(1.23) (-0.71) (1.24)
(1,150,0.01) 219 169 0.0073 -0.0008 0.6256 0.4970 0.0081
(2.17) (-1.03) (2.21)
(1,200,0) 699 215 0.0037 0.0006 0.5808 0.5209 0.0031
(1.18) (-0.60) (1.11)
(1,200,0.01) 202 149 0.0066 -0.0001 0.6040 0.5101 0.0067
(1.82) (-0.73) (1.75)
Average 0.0054 0.0002 0.0053
Panel B: Sample period over 1987-2020
(1,50,0) 299 104 0.0011 0.0062 0.5819 0.5385 -0.0051
(1.15) (0.93) (-1.41)
(1,50,0.01) 83 63 0.0045 0.0030 0.5542 0.5873 0.0015
(0.35) (0.07) (0.28)
(1,150,0) 240 37 0.0010 0.0204 0.5875 0.6216 -0.0195
(1.08) (3.25) (-3.43)
(1,150,0.01) 58 26 0.0035 0.0131 0.5345 0.5769 -0.0096
(0.07) (1.57) (-1.27)
(1,200,0) 224 28 0.0016 0.0208 0.5893 0.6786 -0.0191
(0.73) (2.89) (-2.98)
(1,200,0.01) 53 24 0.0041 0.0229 0.5472 0.7083 -0.0188
(0.20) (3.00) (-2.38)
Average 0.0026 0.0144 -0.0118
28

Table 59.6
Summary Statistics
We calculate the 10-day moving average prices each day using the last 10 days’ closing prices including the current closing
price, and compare the moving average price with the current price as the timing signal. If the current price is above the
moving average price, it is an in-the-market signal, and we will invest in the decile portfolios for the next trading day; otherwise
it is an out-of-the-market signal, and we will invest in the 30-day risk-free Treasury Bill for the next trading day. We use
the 10 NYSE/Amex volatility decile portfolios as the investment assets. We report the average excess return (Avg Ret), the
standard deviation (Std Dev), and the skewness (Skew) for the buy-and-hold benchmark decile portfolios (Panel A), the moving
average timing decile portfolios (Panel B), and the moving average portfolios (MAPs) that are the differences between the MA
timing portfolios and the buy-and-hold portfolios (Panel C). The results are annualized and in percentage. We also report
the annualized Sharpe ratio (SRatio) for the buy-and-hold portfolios and the moving average timing portfolios, and report the
success rate for the MAPs. The sample period is from July 1, 1963 to September 30, 2020. t-statistics are in parentheses and
significance at the 1% and 5% levels is given by an ** and an *, respectively.
Rank Avg Ret Std Dev Skew SRatio Avg Ret Std Dev Skew SRatio Avg Ret Std Dev Skew Success
Panel A Panel B Panel C

Volatility Decile Portfolios MA(10) Timing Portfolios MAP
Low 10.17∗∗ 7.35 -2.49 0.77 17.72∗∗ 4.21 1.33 3.15 7.55∗∗ 5.96 4.81 0.61
(10.46) (31.81) (9.58)
2 12.13∗∗ 10.31 -1.70 0.74 18.92∗∗ 6.11 0.55 2.37 6.79∗∗ 8.27 3.25 0.58
(8.89) (23.40) (6.21)
3 13.34∗∗ 12.23 -1.39 0.72 19.44∗∗ 7.52 -0.16 1.99 6.11∗∗ 9.62 2.54 0.57
(8.25) (19.56) (4.80)
4 13.97∗∗ 13.92 -1.00 0.68 21.33∗∗ 8.68 -0.38 1.94 7.36∗∗ 10.85 1.68 0.57
(7.59) (18.58) (5.12)
5 14.26∗∗ 15.52 -0.86 0.63 22.57∗∗ 9.75 -0.28 1.86 8.31∗∗ 12.03 1.47 0.57
(6.95) (17.51) (5.22)
6 15.08∗∗ 16.56 -0.71 0.64 23.99∗∗ 10.52 -0.17 1.86 8.92∗∗ 12.75 1.25 0.57
(6.88) (17.25) (5.29)
7 14.80∗∗ 17.85 -0.63 0.58 24.73∗∗ 11.34 -0.00 1.79 9.94∗∗ 13.74 1.18 0.57
(6.27) (16.49) (5.47)
8 14.40∗∗ 19.22 -0.48 0.52 27.95∗∗ 12.31 0.29 1.91 13.54∗∗ 14.68 1.05 0.58
(5.67) (17.17) (6.97)
9 17.13∗∗ 20.24 -0.41 0.62 32.55∗∗ 13.21 0.39 2.13 15.42∗∗ 15.23 1.01 0.57
(6.40) (18.63) (7.65)
High 38.13∗∗ 21.59 0.07 1.56 53.14∗∗ 15.06 1.42 3.23 15.01∗∗ 15.30 0.72 0.60
(13.35) (26.68) (7.42)
High - Low 27.96∗∗ 18.46 0.37 0.77 35.42∗∗ 13.99 1.40 3.15 7.46∗∗ 13.06 0.40 0.61
(11.45) (19.14) (4.32)
29

Table 59.7
Fama-French Alphas
The table reports the alphas, betas and the adjusted R-squares of the regressions of the MAPs formed from the 10-day MA
timing strategy on the Fama-French five factors. The alphas are annualized and in percentage. Newey and West (1987) robust
t-statistics are in parentheses and significance at the 1% and 5% levels is given by an ** and an *, respectively. The sample
period is from July 1, 1963 to September 30, 2020.
Rank α(%) βmkt βsmb βhml βrmw βcma Adj. R2 (%)
Low 9.36∗∗ -0.19∗∗ -0.05∗ -0.04 -0.08∗∗ -0.02 24.51

(10.50) (-8.10) (-2.01) (-1.19) (-4.37) (-0.57)
2 10.10∗∗ -0.35∗∗ -0.11∗∗ -0.09∗ -0.11∗∗ -0.04 42.87
(9.31) (-11.76) (-4.13) (-2.23) (-4.72) (-1.36)
3 10.29∗∗ -0.44∗∗ -0.16∗∗ -0.09∗ -0.15∗∗ -0.07∗ 49.52
(9.11) (-14.41) (-6.19) (-2.21) (-6.04) (-2.31)
4 12.20∗∗ -0.51∗∗ -0.22∗∗ -0.10∗∗ -0.16∗∗ -0.08∗ 52.91
(10.39) (-16.79) (-8.12) (-2.53) (-5.92) (-2.26)
5 13.64∗∗ -0.57∗∗ -0.28∗∗ -0.14∗∗ -0.16∗∗ -0.03 55.26
(10.98) (-19.14) (-10.71) (-3.01) (-5.84) (-0.95)
6 14.47∗∗ -0.60∗∗ -0.34∗∗ -0.12∗∗ -0.16∗∗ -0.03 55.28
(11.07) (-21.04) (-11.97) (-2.73) (-5.64) (-0.87)
7 15.59∗∗ -0.64∗∗ -0.39∗∗ -0.11∗ -0.12∗∗ -0.01 55.62
(11.31) (-22.66) (-13.29) (-2.43) (-4.02) (-0.37)
8 19.25∗∗ -0.67∗∗ -0.44∗∗ -0.07 -0.09∗∗ -0.01 54.28
(13.19) (-23.88) (-13.92) (-1.48) (-2.88) (-0.27)
9 20.70∗∗ -0.66∗∗ -0.49∗∗ -0.03 -0.02 0.02 51.68
(13.35) (-23.21) (-15.21) (-0.67) (-0.73) (0.51)
High 19.49∗∗ -0.57∗∗ -0.51∗∗ -0.05 0.06 0.03 42.05
(11.53) (-18.99) (-14.12) (-0.97) (1.82) (0.69)
High - Low 10.13∗∗ -0.38∗∗ -0.45∗∗ -0.01 0.14∗∗ 0.05
(6.42) (-20.56) (-15.95) (-0.14) (4.48) (1.31)
30

Table 59.8
Fama-French Alphas After Publication
The table reports the alphas, betas and the adjusted R-squares of the regressions of the MAPs formed from the 10-day MA
timing strategy on the Fama-French five factors. The alphas are annualized and in percentage. Newey and West (1987) robust
t-statistics are in parentheses and significance at the 1% and 5% levels is given by an ** and an *, respectively. The sample
period is after the sample period used in the paper, from January 4, 2010 to September 30, 2020.
Rank α(%) βmkt βsmb βhml βrmw βcma Adj. R2 (%)
Low 5.91∗ -0.18∗∗ -0.00 -0.10 -0.08 0.09 18.53

(2.24) (-3.98) (-0.04) (-1.07) (-1.52) (0.79)
2 4.54 -0.42∗∗ 0.03 -0.14 -0.12 0.08 44.34
(1.36) (-7.03) (0.57) (-1.61) (-1.81) (0.71)
3 2.36 -0.54∗∗ 0.01 -0.13 -0.16∗ 0.04 53.98
(0.71) (-8.68) (0.24) (-1.74) (-2.30) (0.34)
4 4.82 -0.61∗∗ -0.04 -0.16∗ -0.14 0.08 58.80
(1.47) (-10.73) (-0.72) (-2.05) (-1.91) (0.64)
5 4.64 -0.67∗∗ -0.11 -0.15 -0.16∗ 0.06 61.16
(1.37) (-12.55) (-1.86) (-1.83) (-2.05) (0.54)
6 2.78 -0.67∗∗ -0.17∗∗ -0.13 -0.14 0.07 59.62
(0.78) (-12.66) (-2.99) (-1.70) (-1.78) (0.68)
7 5.24 -0.70∗∗ -0.23∗∗ -0.12 -0.09 0.09 60.17
(1.41) (-15.03) (-4.08) (-1.68) (-1.11) (0.84)
8 8.41∗ -0.69∗∗ -0.28∗∗ -0.08 0.02 -0.02 56.59
(2.11) (-15.08) (-4.42) (-1.09) (0.29) (-0.12)
9 9.65∗ -0.68∗∗ -0.36∗∗ -0.05 0.08 -0.00 53.75
(2.32) (-14.13) (-5.14) (-0.61) (1.05) (-0.02)
High 19.30∗∗ -0.63∗∗ -0.37∗∗ -0.05 0.24∗∗ -0.06 44.57
(3.84) (-12.06) (-4.83) (-0.67) (2.75) (-0.39)
High - Low 13.38∗∗ -0.44∗∗ -0.36∗∗ 0.04 0.32∗∗ -0.15
(2.75) (-11.96) (-6.99) (0.63) (4.31) (-1.49)
31

Table 59.9
The Trend Factor and Other Factors: Summary Statistics
This table reports the summary statistics for the trend factor (Trend), another version of the trend factor following
Fama-French approach (Trend(FF)), the short-term reversal factor (SREV), the umdentum factor (umd), the long-
term reversal factor (LREV), and the Fama-French three factors including the market portfolio (Market), SMB and
HML factors. For each factor, we report sample mean in percentage, sample standard deviation in percentage, Sharpe
ratio, skewness, and excess kurtosis. The t-statistics are in parentheses and significance at the 1%, 5%, and 10%
levels is given by an ***, an **, and an *, respectively. The sample period is from June, 1930 through December,
2019.
Factor Mean(%) Std Dev(%) Sharpe Ratio Skewness Kurtosis
Trend 1.56 3.42 0.46 1.44 11.3

Trend(FF) 1.22 2.90 0.42 1.61 11.2
SREV 0.77 3.43 0.23 1.00 8.38
UMD 0.61 4.74 0.13 -3.02 27.4
LREV 0.28 3.44 0.08 2.92 25.1
Market 0.64 5.31 0.12 0.25 8.13
SMB 0.24 3.19 0.08 2.01 20.0
HML 0.36 3.52 0.10 2.19 19.2
32

Table 59.10
The Trend Factor and Other Factors After Publication: Summary Statistics
This table reports the summary statistics for the trend factor (Trend), another version of the trend factor following
Fama-French approach (Trend(FF)), the short-term reversal factor (SREV), the umdentum factor (umd), the long-
term reversal factor (LREV), and the Fama-French three factors including the market portfolio (Market), SMB and
HML factors. For each factor, we report sample mean in percentage, sample standard deviation in percentage, Sharpe
ratio, skewness, and excess kurtosis. The t-statistics are in parentheses and significance at the 1%, 5%, and 10%
levels is given by an ***, an **, and an *, respectively. The sample period starts after the sample period in the paper,
from January, 2015 through December, 2019.
Factor Mean(%) Std Dev(%) Sharpe Ratio Skewness Kurtosis
Trend 0.24 2.58 0.09 0.07 3.21

Trend(FF) 0.24 2.22 0.11 -0.51 2.01
SREV 0.47 2.26 0.21 0.40 1.15
UMD 0.22 3.81 0.06 0.12 0.14
LREV -0.72 2.25 -0.32 0.36 -0.32
Market 0.89 3.59 0.25 -0.59 0.98
SMB -0.14 2.39 -0.06 0.30 -0.35
HML -0.36 2.62 -0.14 0.96 1.36
33

SSRN Id3850494

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

SSRN Id3850494

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SSRN Id3850494

Uploaded by

Copyright:

Available Formats

Chapter 59

Technical Analysis in the Stock

Electronic copy available at: https://ssrn.com/abstract=3850494

Electronic copy available at: https://ssrn.com/abstract=3850494

59.2 Time-series Predictability

In this subsection, we focus on time-series predictability of the aggregate stock market

59.2.1.1 Empirical studies

Electronic copy available at: https://ssrn.com/abstract=3850494

Electronic copy available at: https://ssrn.com/abstract=3850494

Electronic copy available at: https://ssrn.com/abstract=3850494

59.2.1.2 The theory

Electronic copy available at: https://ssrn.com/abstract=3850494

Pt = p0 + p1 Dt + p2 πt + p3 θt + p4 A1t + p5 A2t , (59.2)

where p0 , p1 , p2 , p3 , p4 and p5 are constants determined only by model parameters.

Electronic copy available at: https://ssrn.com/abstract=3850494

The optimal demand for stock is given by

η i = f i Ψi = f0i + f1i Dt + f2i πt + f3i θt + f4i A1t + f5i A2t , (59.5)

The technical traders’ optimal demand for stock is given by

ηiu = f0iu + f1iu Dt + f2iu Pt + f3iu Ait , (59.8)

where f0iu , f1iu , f2iu and f3iu are constants for i = 1, 2.

Electronic copy available at: https://ssrn.com/abstract=3850494

with p3 < 0 and p4 = p5 = 0. Q.E.D.

59.2.1.3 Brock, Lakonishok, and LeBaron (1992): An update

Electronic copy available at: https://ssrn.com/abstract=3850494

M AXt,L = max{Pt , Pt−1 , · · · , Pt−L+1 }, (59.16)

M INt,L = min{Pt , Pt−1 , · · · , Pt−L+1 }. (59.17)

Electronic copy available at: https://ssrn.com/abstract=3850494

59.2.2 Portfolios and Other Assets

Electronic copy available at: https://ssrn.com/abstract=3850494

59.2.2.1 Related studies

59.2.2.2 Han, Yang and Zhou (2013): An update

Electronic copy available at: https://ssrn.com/abstract=3850494

Electronic copy available at: https://ssrn.com/abstract=3850494

standard deviation of 17.14% per annum.

Electronic copy available at: https://ssrn.com/abstract=3850494

59.3.1 Fama-MacBeth Regressions

59.3.1.1 Han, Zhou and Zhu (2016): An update

Electronic copy available at: https://ssrn.com/abstract=3850494

Electronic copy available at: https://ssrn.com/abstract=3850494

59.3.2 Machine Learning Methods

Machine learning methods, in contrast to the Fama-MacBeth regressions, can incorpo-

Electronic copy available at: https://ssrn.com/abstract=3850494

Electronic copy available at: https://ssrn.com/abstract=3850494

Electronic copy available at: https://ssrn.com/abstract=3850494

Electronic copy available at: https://ssrn.com/abstract=3850494

Electronic copy available at: https://ssrn.com/abstract=3850494

Electronic copy available at: https://ssrn.com/abstract=3850494

Electronic copy available at: https://ssrn.com/abstract=3850494

Electronic copy available at: https://ssrn.com/abstract=3850494

Panel A: Daily Returns

Electronic copy available at: https://ssrn.com/abstract=3850494

Panel A: Full Sample

Electronic copy available at: https://ssrn.com/abstract=3850494

Period Test N (Buy) N (Sell) Buy Sell Buy>0 Sell>0 Buy-Sell

Electronic copy available at: https://ssrn.com/abstract=3850494

Test N (Buy) N (Sell) Buy Sell Buy>0 Sell>0 Buy-Sell

Electronic copy available at: https://ssrn.com/abstract=3850494

Test N (Buy) N (Sell) Buy Sell Buy>0 Sell>0 Buy-Sell