Cryptocurrency Factor Portfolios: Performance, Decomposition and

Pricing Models

Weihao Han, wh587@[Link], School of Management, University of Bath, Claverton

Down, Bath, BA2 7AY, UK.

David Newton, [Link]@[Link], School of Management, University of Bath,

Claverton Down, Bath, BA2 7AY, UK, Tel: +44(0) 1225 386890

Emmanouil Platanakis, [Link]@[Link], School of Management, University of Bath,

Claverton Down, Bath, BA2 7AY, UK. Tel: +44(0) 1274 235311

Charles Sutcliffe, [Link]@[Link], ICMA Centre, Henley Business School,

University of Reading, PO Box 242, Reading RG6 6BA, UK, Tel +44(0) 118 378 6117

Xiaoxia Ye, [Link]@[Link], Management School, University of Liverpool,

Chatham Street, Liverpool L69 7ZH, UK, Tel +44(0) 151 795 7638

This draft: 28 January 2021

Abstract

We employ a non-parametric technique, almost stochastic dominance, and find that portfolios
of cryptocurrencies based on nine factors dominate the S&P500, US 10-year T-bonds, US T-
bills and a cryptocurrency index over longer investment horizons. After decomposing those
nine long-short factor portfolios, we notice that their dominance relative to above four
benchmarks and equity portfolios based on size, momentum and book-to-market, is mainly
attributable to their long legs. A three-factor cryptocurrency model (market, size, and
momentum) has insignificant alphas for five of the dominant cryptocurrency portfolios,
indicating that their dominance is due to a risk premium. We then add the combinations of four
mispricing factors and cryptocurrency fundamental factors to the three-factor model. The
alphas of the other four dominant cryptocurrency portfolios remain significant, indicating that
their dominance is due to mispricing.

Keywords: Asset Pricing, Cryptocurrencies, Almost Stochastic Dominance, Risk Premium,

Mispricing

JEL Classification: G11, G12

1
1. Introduction

Cryptocurrencies are built on blockchain technology (Abadi and Brunnermeier, 2018;

Biais et al., 2019) which permits transactions without central supervision, and the market has
experienced a surge in the number of cryptocurrencies. Since the most famous, Bitcoin,
appeared in 2009, more than 50 million investors have traded cryptocurrencies on more than
100 global exchanges, and over 100,000 companies worldwide accept payment in bitcoins and
bitcoin debit cards (Easley et al., 2019; Makarov and Schoar, 2020). Companies at an early
stage can raise money from initial coin offerings, rather than being financed through venture
capital. As cryptocurrencies attract attention, whether they have investment value becomes an
important issue, and both academics and practitioners are examining their properties for uses
other than speculation. If we treat cryptocurrencies as a financial asset, it is vital to understand
their characteristics in performance and pricing for at least three reasons. First, to find an
appropriate performance metric for comparing cryptocurrencies with other financial classes.
Second, to find factors that can be used to form outperforming cryptocurrency portfolios. Third,
to determine whether any outperformance by cryptocurrency portfolios is caused by a risk
premium or by mispricing.

An extensive academic finance literature on cryptocurrencies has been developed in

recent years, especially in the areas of of ethics, initial coin offerings (ICOs) and analysis of
several of the most popular coins (e.g. Ciaian et al., 2016; Foley et al., 2019; Gan et al., 2020).
There is only limited literature that explores the factors that influence returns and even
lessfocusing on the performance of cryptocurrency factor portfolios. For instance, Liu et al.
(2020) and Borri (2019) study the relation between popular cryptocurrencies, finding only
exposure to cryptocurrency-related factors. JUst a few of the most popular cryptocurrencies
account for a vast majority of the capitalization of the whole cryptocurrency market. Their
characteristics are well documented from the asset pricing side but further investigation is still
needed to evaluate the performance, decomposition and pricing models of cryptocurrency
anomalies. Hence, in this paper we examine questions associated with previous research: 1)
Are cryptocurrency portfolios profitable compared to traditional financial asset classes? We
evaluate the performance of portfolios of cryptocurrencies factors relative to conventional
financial asset classes such as stocks, bonds, treasury bills, and a cryptocurrency index. 2) Is
there a proper metric to compare the performances of different underlying assets? Because the
empirical distributions of cryptocurrency returns are highly non-normal, this has limited the
number of researchers who have examined the risk factors and analysed cryptocurrencies'

2
performances. 3) If outperformance of cryptocurrency exists, do long leg or short leg of
cryptocurrency factor portfolios contribute to outperformance? 4) Does any outperformance of
cryptocurrency factor portfolios come from risk premium or mispricing? 5) Can outperforming
cryptocurrency factor portfolios be accurately explained by adding mispriced factors and
cryptocurrency fundamental factors to a coin market three-factor model?

To fill this gap, we examine the performance of cryptocurrency portfolios based on

different risk factors by conducting almost stochastic dominance (ASD) proposed by Leshno
and Levy (2002) on the 400 largest cryptocurrencies, which account for over 80% of total
market capitalization. Almost stochastic dominance plays a vital role in comparing the
performance of different asset classes that are highly skewed and leptokurtic, as the
conventional mean-variance approach fails to measure the returns correctly (Bali et al., 2013).
On the other hand, it would also eliminate the impact of investors’ extreme preferences, where
mean-variance cannot deliver an accurate measure. Furthermore, to form the factor portfolios,
we rely on each cryptocurrency’s open price, high price, low price, close price, volume and
market capitalization, which comprise the only public information for each cryptocurrency.
Inspired by Fama and French (1993), Carhart (1997), Liu et al. (2019) and Feng et al. (2020),
we divide the available factors information into four categories: size, momentum, volume and
volatility. We sub-divide the four large categories into 27 factor portfolios (see Table 3) to
study whether cryptocurrencies share similarities with stocks regarding anomalies and examine
each factor's relative performance against selected benchmarks. Subsequently, we decompose
the long-short portfolios to detect which legs contribute to outperformance, and then we
improve the existing coin market three-factor model developed by Liu et al. (2019) via
incorporating mispriced factors and cryptocurrency fundamental factors (electricity and
computing power).

This paper is the first to utilize the almost stochastic dominance (ASD) approach of
Leshno and Levy (2002) to examine cryptocurrency factor portfolios’ relative performance.
ASD is a non-parametric method which compares two uncertain prospects by maximizing
expected utility, and does not require any assumption about the return distribution. Due to
cryptocurrencies’ highly skewed returns distributions, standard performance metrics such as
the mean-variance approach and Sharpe ratio cannot provide precise measure since these two
approaches require the assumption of normality (Farinelli et al., 2008; Bali et al., 2013).
Moreover, there exists an extreme utility function (pathological preference) of investors who
are indifferent to a small amount of income and a large amount of income. Although this type

3
of investor is a minority, standard metrics also fail to estimate accurately (Leshno and Levy,
2002; Bali et al., 2009). To avoid such problems, we employ almost first-order stochastic
dominance (AFSD) and almost second-order stochastic dominance (ASSD) to compare the
relative performance of each factor portfolio against S&P 500, US T-bonds, US T-bills and our
cryptocurrency index at 4-week, 13-week, 26-week, 52-week and 78-week investment
horizons. Our longest horizon is 1.5 years because cryptocurrencies have not existed for as
long as equities and bonds, and because investors are unwilling to hold cryptocurrencies for
relatively long horizons due to their high volatility. We obtain nine dominant portfolios against
our four benchmarks in the sense of AFSD and ASSD.

Even though numerous studies (e.g. Stambaugh et al., 2012; Daniel and Moskowitz,
2016; Chu et al., 2020) argue that the importance of short legs is not as prominent as long legs,
and their impacts are not symmetric as short legs face more frictions, we aim to clarify whether
long leg or short leg or both legs contribute to outperformance among nine dominant factors
for cost saving. For instance, a short sale may incur higher fees than buying an asset due to the
frictions of short selling. Our paper is the first to decompose long-short portfolios into their
long and short components in order to examine the individual contribution. Using AFSD and
ASSD tests, we find that dominance is mostly attributable to the long-legs. For the long-only
portfolios, unlike the long-short portfolios, there is no AFSD and ASSD dominance for
horizons of 52-weeks and below. We conduct the same test against three equity benchmarks –
equity portfolios based on size, momentum and BE/ME - since these factors have been widely
used. We find that the dominant portfolios against equity portfolios based on size, momentum
and BE/ME are to the same as those for the portfolios above. We conclude that the superior
returns generated by long-short strategies are mainly contributed by their long legs. Although
such portfolios could achieve AFSD and ASSD dominance via the long-only strategy, their
violation areas are larger than those of the long-short portfolios, implying that the long-only
strategy might not be optimal. Hence, it is likely that the long-short strategy is preferable to the
long-only strategy and can boost performance, which is consistent with Stambaugh et al. (2012)
and Edelen et al. (2016) for the case of equity factors.

Finally, we examine whether the variation in returns of the nine dominant factor
portfolios can be captured by both the original and the improved coin market three-factor model
to determine whether the outperformance comes from risk premium or mispricing. Even though
hundreds of predictive cross-sectional asset pricing anomalies have been explored by finance
academics, the debate on whether abnormal equity returns come from risk premium or

4
mispricing never ends (Harvey et al., 2015; Chu et al., 2020). For instance, an inverse relation
caused by risk premium between book-to-market portfolio and future growth rates have been
documented by Fama and French (1995) and Penman (1996), whereas Lakonishok et al. (1994)
and La Porta et al. (1997) propose that return on value/growth stocks is due to mispricing as
investors have overly optimistic (pessimistic) expectations for future. The seminal literature
focuses on stocks and hardly any research in this area explores cryptocurrencies, adding to our
motivation. We find that a coin market three-factor model explains cross-sectional returns to a
varying extent. Four dominant factor portfolios have positive and statistically significant alphas,
and relatively small R2 values, indicating their outperformance results from mispricing. A
three-factor cryptocurrency model (market, size, and momentum) has insignificant alphas for
the remaining five dominant cryptocurrency portfolios, indicating that their dominance is due
to a risk premium. We then add four mispricing factors (subcategories of momentum, volume
and volatility) and two cryptocurrency fundamental factors (electricity and computing power)
to our three-factor model. The alphas of the four dominant cryptocurrency portfolios remain
significant, while their R2 remain low, indicating that their dominance is due to mispricing.

This paper is organized as follows. Section 2 discuss the related literature. Section 3
describes the data and summary statistics. Section 4 presents the methodology of the ASD
approach. Section 5 describes the summary statistics of factor portfolios constructed in Section
4. Section 6 evaluates the empirical results of the ASD, decomposes long-short portfolios, and
examines the cross-sectional return on dominant factor portfolios using a coin market three-
factor model. Section 7 concludes the paper.

2. Related Literature

Although we are not aware of any other existing literature that focuses on
cryptocurrency factor portfolios' performance, our paper is linked to various literature from the
perspective of almost stochastic dominance, decomposition of long-short portfolios and pricing
models.

2.1. Risk Factors Related to Traditional Assets

Financial economists have studied cross-sectional variations in stock and bond returns
to find risk factors that can be used to create trading strategies. Fama and French (1993)
document that the size factor (SMB) and the book-to-market factor (HML) can significantly

5
explain the cross-sectional variation of stock returns. Subsequently, Carhart (1997) improved
their model by adding the momentum factor (MOM), and found that it can explain cross-
sectional returns on equity mutual funds. Bai et al. (2019) analyze the risk factors for corporate
bonds by assuming that the downside risk is a strong indicator of future bond returns. Zhang
(2005) and Lettau and Wachter (2007) evaluate the value premium of equities, suggesting that
statistical models can capture portfolios' value premium. Stambaugh and Yuan (2017) argue
that return anomalies can be captured by a model that incorporates mispricing factors, and risk
factors such as the size factor have a risk premium twice the usual estimate. Neely et al. (2014)
study the predictive power of technical indicators on the equity risk premium and find that
technical indicators can provide in-sample and out-of-sample forecasts for equities. Likewise,
Rapach et al. (2009) identifies the predictive power of combining individual forecasts of the
equity premium, since combining forecasts generates economically and statistically significant
gains both in and out-of-sample. The studies above emphasise the importance of studying risk
factors which could explain the cross-sectional variation in asset returns and provide predictive
power. However, few studies focus on cryptocurrencies, most of them evaluating the relation
between several popular coins and different asset classes (equity, bond and commodity etc.).
The relatively small sample size cannot detect cryptocurrencies’ anomalies, and may incur
sample bias. Hence, we are motivated to develop appropriate factors capturing the cross-
sectional variation of cryptocurrency returns, taking into account the risk premium and
mispricing of cryptocurrencies.

Our work relates to the literature that evaluates the factor anomalies, and a vast amount
of literature documents whether an anomaly is caused by risk premium or mispricing.
Keloharju et al. (2020) show that the seasonalities of stocks can be balanced out by seasonal
reversal, demonstrating that seasonalities contribute to temporary mispricing. Ali et al. (2003)
propose that book-to-market (B/M) has a greater effect on stocks with higher idiosyncratic
volatility and higher transaction cost due to market-mispricing. Wang (2019) establish that
zero-investment strategy that longs the portfolios with the lowest cash conversion cycle (CCC)
and shorts portfolios with the highest CCC can generate 7% return annually due to mispricing.
In this paper, we seek to diagnose 27 cryptocurrency factor portfolios to find whether anomalies
can be generated in both long term and short term, and we use coin market pricing model to
scrutinise which legs of factor portfolios contribute to outperformance. For robustness, we use
the methos of Stambaugh and Yuan (2017) to test dominant portfolios with an adjusted three-
factor model incorporating mispricing factors.

6
2.2. Almost Stochastic Dominance

Our paper is related to the literature that examines financial assets' performance by
using almost stochastic dominance. Bali et al. (2009) analyze US stocks' performance versus
bonds using almost stochastic dominance (ASD), and provide firm evidence that stocks
dominate bonds by ASD at a long horizon after adjusting for pathological preferences.
Similarly, Levy and Levy (2019) employ first-degree stochastic dominance with a riskless asset
(FSDR) to evaluate the performance of stocks and bonds; finding that stocks tend to dominate
bonds by FSDR for any investment horizon longer than three years. Similarly, Post (2003)
evaluates the stochastic dominance of portfolios constructed from a list of assets and argues
that benchmark portfolios based on market capitalization and book-to-market ratios are
remarkably efficient compared to the Fama and French market portfolio. Second-degree
stochastic dominance (SSD) is also applied by Board and Sutcliffe (1994) to rank a series of
portfolios with or without short sales, and they provide robust results thorough the SSD
approach at the 1% level. To examine the performance of hedge funds, Bali et al. (2013)
employ almost stochastic dominance on hedge funds, and propose that long-short equity hedge
and emerging markets hedge strategies dominate the US equity market. These two papers study
the performance of common financial assets. The above literature evaluates stocks’, bonds’
and hedge funds’ performance using a non-parametric approach, and underlying assets’
superior performance has been detected. However, these studies do not mention how to build
such dominant portfolios, even if investors knew that the performance of stocks, bonds and
hedge funds is superior, investors are still left with the problem of assets selection. In contrast,
we construct cryptocurrency factor portfolios and study their performance through almost
stochastic dominance, to both avoid influence of special return distributions and clarify whether
cryptocurrencies are worthy of investing.

2.3. Long/Short Legs of Zero-Investment Portfolios

Our research is also related to papers that evaluate whether cryptocurrency factor
portfolios' outperformance is attributed to their long legs. Extensive research has employed the
long-short strategy in studying the cross-section of factor portfolios. For instance, Israel and
Moskowitz (2013) examine the effect of a long position and short position on overall stock
portfolio performance. They find that outperformance of size factor, value factor, and
momentum factor is attributed to a long position of portfolios. Similarly, Blitz et al. (2019)
decompose Fama-French style equity portfolios into long- and short-leg portfolios, they find

7
that most added return comes from long-leg portfolios and long-leg portfolios are more
diversified than that of short-leg portfolios. A number of papers study long-short strategy (e.g.
Frazzini and Pedersen, 2014; Barroso and Santa-Clara, 2015; Daniel and Moskowitz, 2016).
Nevertheless, most of previous literature concentrates on restricted asset classes such as
equities, and the research on cryptocurrency regarding the decomposition of long-short
portfolios is rare. For this reason, we examine the performance of long and short legs of each
factor portfolio relative to four widely used benchmarks (S&P500, US 10-year T-bonds, US T-
bills and a cryptocurrency index) and equity anomalies (size, momentum and book-to-market
ratio), to gain a solid insight into cryptocurrencies.

2.4. Main Difference Between Current Literature on Cryptocurrency

Unlike equities and bonds that have been well evaluated, studies on cryptocurrencies
are an emerging field that needs to be [Link] et al. (2019) as pioneers, study the cross-
sectional returns of a large number of cryptocurrencies from the aspect of asset pricing.
Specifically, they first form several effective risk factors with significant alphas, and they find
such significant factors can be well explained by a coin three-factor model, where the model
shares similarities with an equity pricing model. Similarly, Prior to Liu et al. (2019), Liu et al.
(2020) examine the relationship between risk and returns for three mainstream cryptocurrencies
(e.g. Bitcoin, Ripple and Ethereum). They demonstrate which exposures cryptocurrencies have,
and narrow the scope of potential factors that may influence returns of cryptocurrencies to aid
further studies. Particularly, they first test the exposures of cryptocurrencies to stocks (Fama-
French factors), major currencies (Australian Dollar, Canadian Dollar, Euro etc.), precious
metals (gold, silver and platinum) and macroeconomic factors (industrial production growth
and personal income growth etc.), they claim that little or no evidence shows cryptocurrencies
have exposures to the factors above. Thus, they evaluate the exposures of cryptocurrencies only
associated with cryptocurrency market such as cryptocurrency momentum, proxy for investor
attention, proxy for price-to-fundamental value and cost of mining, they conclude that factors
have certain predictive power in certain horizons. This implies the direction of further study of
cryptocurrencies.

Our research has some similarity to Liu et al. (2019) but with important differences.
They find risk factors that capture variations in the cross-sectional returns of factor portfolios,
and develop a cryptocurrency three-factor model. In contrast, our study focuses on factor
portfolio performance. Because cryptocurrencies have non-normal returns distributions, we

8
evaluate the investment value of cryptocurrencies by comparing their performance with those
of different asset classes. Our paper also differs from Liu et al. (2020). They study the
relationship between cryptocurrencies and a list of potential factors, aiming to exploit those
with predictive power. In contrast, our paper focuses on which factors generate high excess
returns. We also use a non-parametric approach, which allows for the highly non-normal
distributions of cryptocurrency returns. We also disaggregate the long-short portfolios to
examine whether outperformance comes from long or the short legs, and analyze whether the
dominance of some factors is due to mispricing or a risk premium. Thus, we study
cryptocurrencies mainly from a performance, rather than an asset pricing, perspective.

In summary, we test cryptocurrency portfolios using the non-parametric approach of

almost stochastic dominance, and we then show that the dominance of some factor portfolios
comes mainly from the long-leg portfolios due to either a risk premium or mispricing.

9
3. Data

The data of cryptocurrencies are collected from [Link], which is a leading

source of cryptocurrencies price, volume and market capitalization. [Link]
aggregates information from over 200 major cryptocurrencies’ exchanges data on opening
price, high price, low price, close price, volume and market capitalization for most
cryptocurrencies. Moreover, a cryptocurrency must meet certain criteria to be listed on
exchange website, such as trading on a public exchange with an application programming
interface which can show closing prices and non-zero trading volume during the previous 24
hours. Another reason for choosing [Link] is that it includes both defunct and
active cryptocurrencies which might mitigate survivorship bias. Also some influential papers
use this website as source of data (e.g. Detzel et al., 2020; Liu et al., 2020). We chose only
cryptocurrencies with market capitalization larger than 1 million dollars. Because data on
trading volume became available in the last week of 2013, we use cryptocurrencies date from
the beginning of 2014. We select data for the top 400 cryptocurrencies based on their market
capitalizations in USD over the period 1st January 2014 to 31st December 2019. The market
capitalization of these 400 cryptocurrencies accounts for 82% percent of the whole market
capitalization. We use the data of S&P 500, 10-year T-bond from CRSP and risk-free rate from
Kenneth French’s website.

To gain better insights into the investment value of cryptocurrencies, we calculate the
market-capitalization-weighted return of all 400 cryptocurrencies. First, we calculate the log
daily return across the sample period and allocate the weights to each asset every day, then get
the summation of the value weighted return on each day to construct the daily market index
return. The summary statistics of each asset are in Table 1. According to Table 1, it is clear that
the coin market weekly return of 0.0103 is an order of magnitude higher than the return on the
S&P 500 of 0.0019, return on T-bill of 0.0002 and proceeds of T-bond of 0.0007, but the risk
of coin market is also much larger than those of stocks, bonds and the risk-free rate. The overall
coin market return is positively skewed, similar to T-bills and T-bonds, whereas the skewness
of returns on the S&P 500 is negative. The kurtosis of coin market and S&P 500 are both above
the normal distribution of 3, which indicates the distributions of coin market and S&P 500 are
leptokurtic. Although risk-adjusted returns could compare different asset classes such as
Sharpe ratio (Sharpe, 1966), it might not be efficient with cryptocurrencies because of the high
skewed distribution of returns. Thus, it is difficult to compare different asset classes due to
their distinctive properties of returns by conducting the standard measure. This paper

10
determines upon finding an appropriate mean to examine the investment value of
cryptocurrencies.

Table 1 Descriptive Statistics

Mean Median Std Skewness Kurtosis J-B test P-value

Coin Market (CMKT) 0.0103 0.0086 0.1054 0.1109 4.5026 2677 0.001
Bitcoin Return 0.0071 0.0040 0.1046 0.0299 4.0146 2761 0.001
Ripple Return 0.0063 -0.0067 0.1822 1.9627 11.1288 1479 0.001
Ethereum Return 0.0166 0.0100 0.1957 -0.4957 12.4147 3465 0.001
S&P 500 0.0019 0.0030 0.0174 -0.8320 5.1659 893.7 0.001
One-month T-bill 0.0002 0.0001 0.0002 0.6166 1.7923 97.81 0.001
Ten-year T-bond 0.0007 0.0000 0.0079 0.3104 3.4265 1872 0.001

Notes. This table presents descriptive statistics of weekly returns (not in percentage) on Coin market
return including three popular cryptocurrencies such as Bitcoin, Ripple and Ethereum for data period
2014 to 2019. As well as the S&P 500 index, T-bill and T-bond are collected to be set as control
groups. Specifically, coin market is the value-weighted average returns of 400 coins through sample
period. This table reports the mean, median, standard deviation skewness and kurtosis to identify the
distributions of corresponding assets. The Jarque-Bera (JB) test statistics and corresponding p-value
indicate the shape is different from normal distribution.

Three popular cryptocurrencies are chosen to present a portrait of individual investors

who invest in mainstream virtual currencies in the cryptocurrency market. Among the three
popular cryptocurrencies, Ripple possesses the lowest weekly return of 0.0063 with a standard
deviation of 0.1822, and Ethereum has the highest weekly return of 0.0166 with a standard
deviation of 0.1957. Bitcoin, the most popular and famous cryptocurrency, has a weekly return
of 0.0071, which is smaller than that of Ethereum, along with half its standard deviation. The
kurtosis of Bitcoin is the most similar to that of a normal distribution, which is also the lowest
kurtosis among the three coins. In contrast, comparing coin market return to three popular
cryptocurrencies shows similar dispersion to Bitcoin but with higher weekly return of 0.0103
and the kurtosis of coin market return is slightly higher than that of Bitcoin but significantly
lower than those of Ripple and Ethereum.

Compared with traditional assets, such as stocks, investors might gain larger returns by
holding cryptocurrencies over long horizons. Figure 1 illustrates the cumulative returns on the
Cryptocurrency market index and S&P 500 stock index, respectively. The red line is the
cumulative return on coin market. Although it experienced the sharp drop over the period 2014
to 2016, it started to recover in 2016 and achieved exceptional return in the longer period.
Conversely, the blue line represents the cumulative return on the stock market, which is steadily
increasing over time but the growth rate of that is an order of magnitude lower than the coin

11
market return. Thus, the attractive returns on coin market might provide an alternative
investment opportunity for investors.

Figure 1 Cumulative return of Coin Market against S&P 500

Notes. This figure reports the cumulative return of Cryptocurrencies against S&P 500 over the period
from 2014 to 2019. It is clear that the Cryptocurrencies is more profitable compared with the stock
market.

4. Methodology

Because of the inapplicability of standard metrics such as mean-variance and normal

stochastic dominance, we apply almost stochastic dominance to test the performance of
cryptocurrencies against benchmarks: S&P 500, 10-year T-bond and 30-day T-bill. Since the
distributions of cryptocurrencies are skewed, the usual performance metrics that are based on
assumption of normal distribution are not appropriate to measure performance and capture the
generic properties of returns on cryptocurrencies. Almost stochastic dominance does not
require a parametric specification of investors’ preferences and assumption of a normal
distribution.

12
4.1. Inability of Mean Variance Approach

According to Bali et al. (2013), a classical mean-variance investment decision rule

might be inefficient to identify the preference between two portfolios when all investors or
almost all investors would choose one portfolio over another portfolio. To illustrate, consider
two portfolios, H and L:

H: 𝜇 = 100%, 𝜎 = 5.1%

L: 𝜇 = 1%, 𝜎 = 5.0%

where H and L represents portfolios with high and low expected returns, respectively. The 𝜇
and 𝜇 are returns on portfolio H and L, respectively; the standard deviations of portfolios H
and L are 𝜎 and 𝜎 , respectively. If portfolio H dominates portfolio L by the mean-variance
approach, then the condition that 𝜇 > 𝜇 and 𝜎 < 𝜎 must be met. In this example, the
expected return on portfolio H is dramatically higher than that of portfolio L (100 times) but
with a slightly higher standard deviation (0.1 percent). However, in reality, most investors
would choose portfolio H over portfolio L, because the decrease in expected utility of slightly
higher risk is much less than the increase in expected utility of much higher expected return.
Therefore, although most investors would choose portfolio H over L, yet the mean-variance
approach fails to identify the outperforming portfolio.

4.2. Stochastic Dominance

In comparison with the mean-variance approach, stochastic dominance provides an

alternative perspective to compare the performance of two assets. Recall the definition of first-
order stochastic dominance (FSD) and second-order stochastic dominance (SSD) in the
following (Leshno and Levy, 2002; Bali et al., 2013).

First-Order Stochastic Dominance. Suppose there are two hypothetical risky portfolios,
H and L, and the cumulative distribution of H and L are denoted by 𝐹 and 𝐹 , respectively.
Portfolio H dominates portfolio L by first-order stochastic dominance (H FSD L) if 𝐹 (𝑟) ≤
𝐹 (𝑟) for all return values r and a strict inequality holds for at least some r. Specifically, H FSD
L if and only if 𝐸 𝑢(𝑟) ≥ 𝐸 𝑢(𝑟) for all 𝑢 ∈ 𝒰 , where 𝒰 is the set of all nondecreasing
differentiable real-valued functions. To get better understanding of FSD, Figure 2 might
provide an intuitive insight into FSD through graphical explanation.

13
 Figure 2 First-Order Stochastic Dominance

Notes. This figure reports the pattern that Portfolio H dominates Portfolio L by FSD.

According to Figure 2 which represents H FSD L, it is clear that portfolio H (red solid
line) is plotted below the portfolio L (blue dashed line) for every point of return, which
represents portfolio H would gain a higher return than that of portfolio L for any given level of
probability. Hence, portfolio H reveals superiority over portfolio L since with portfolio H is
always possible to achieve higher return that results in higher preference to portfolio L.
However, FSD is difficult to achieve since the situation in reality is more complicated than at
the theoretical level. Therefore, the birth of second-order stochastic dominance (SSD) is
important, which enables the condition of domination less harsh than FSD and more effectively
to use.

Second-Order Stochastic Dominance. Suppose there are two hypothetical risky

portfolios, H and L, H represents high expected portfolio and L represents low expected return.
The cumulative distribution of H and L are denoted by 𝐹 and 𝐹 , respectively. Portfolio H
dominates portfolio L by second-order stochastic dominance (H SSD L) if ∫ [𝐹 (𝑠) −
𝐹 (𝑠)]𝑑𝑠 ≥ 0 for all return values r and a strict inequality holds for at least some r. Particularly,
H SSD L if and only if 𝐸 𝑢(𝑟) ≥ 𝐸 𝑢(𝑟) for all 𝑢 ∈ 𝒰 , where 𝒰 is the set of all
nondecreasing real-valued functions such that 𝑢 ≤ 0. To get better understanding of SSD,
Figure 3 explains the second-order stochastic dominance by utilizing graphical explanation.
Figure 3 illustrates the fact that H (red solid line) dominates L (blue dashed line) by SSD. To

14
illustrate, there exist an area that portfolio H plots above portfolio L, which is denoted by b and
falls into the range of [2,3].

Figure 3 Second-Order Stochastic Dominance

Notes. This figure reports the pattern that Portfolio H dominates Portfolio L by SSD.

Because this violation area that break the FSD rule that portfolio H dominates portfolio
L if portfolio H plots below the portfolio L for every given return r, so FSD does not exist due
to the reason above. Moreover, from the perspective of the FSD proposition, the 𝐹 (𝑟) is larger
than 𝐹 (𝑟) in the interval of [2,3], which also indicates nonexistence of FSD in such a case.
However, there might exist SSD according to the SSD proposition, because SSD examines the
dominance regard to the area that enclosed by cumulative distributions of two portfolios, which
is more tolerant than that of FSD that only compares the value of cumulative distributions.
According to Figure 2, portfolio H dominates portfolio L when return is up to 2, which makes
area a is larger than zero by SSD condition by SSD proposition ∫ [𝐹 (𝑠) − 𝐹 (𝑠)]𝑑𝑠 ≥ 0.
Furthermore, portfolio H cannot dominate portfolio L when return falls into interval [2,3] since
the area b enclosed by H and L is smaller than zero by SSD proposition ∫ [𝐹 (𝑠) −
𝐹 (𝑠)]𝑑𝑠 ≥ 0. The portfolio H would dominate L again after x larger than 4 since portfolio H
plots portfolio L. In other words, if portfolio H SSD L, such condition 𝒂 ≥ 𝒃 must be held to
ensure that ∫ [𝐹 (𝑠) − 𝐹 (𝑠)]𝑑𝑠 ≥ 0. Additionally, if the first point of portfolio H prior to
that of portfolio L (the minimum value of portfolio H is smaller than that of portfolio L) that
produces negative area by the SSD formula, then no matter how big the positive area after their

15
first intersection, there would not be SSD because it violates the condition that
∫ [𝐹 (𝑠) − 𝐹 (𝑠)]𝑑𝑠 ≥ 0 in the interval. Nevertheless, portfolio H might fail to dominate L
if area of b is slightly bigger than a no matter how big the positive area in the following part,
and this is the reason why this paper conducts almost stochastic dominance, to avoid such an
economically irrational selection.

The key difference between FSD and SSD is the assumption of an investor’s utility
function (Daskalaki et al., 2017). FSD allows that investors may be risk averters or risk lovers
because it only requires investors prefer more to less (mathematically, 𝑢 > 0). SSD assumes
investors are risk averters, which has a concave utility function and prefer certainty to gambling
(mathematically, 𝑢 > 0 and 𝑢 < 0 ). As most investors are risk averse, SSD plays a
significant role in measuring performance of assets for risk averters. In following section, we
discuss the almost stochastic dominance (ASD) for first-order and second-order respectively.
Intuitively, ASD is also stochastic dominance but with a looser assumption, it assumes a part
of utility function caused by extreme (unusual) utility function can be ‘ignored’ only if the
violation area is small enough.

4.3. Almost Stochastic Dominance

Although most investors would prefer one asset to another asset in the real world,
stochastic rules cannot explain such preference due to some extreme utility functions that only
violate a small portion of these rules. Almost stochastic dominance (ASD) can provide a more
realistic situation for solving this sort of problem. To illustrate the necessity of almost
stochastic dominance, we begin with a hypothetical cash flow table as table 2 shows.

Table 2 demonstrates the returns on portfolio H and L with corresponding probabilities.

Portfolio H dominates L by FSD, and the cumulative distribution of H should plot strictly
below the cumulative distribution of L by FSD condition, which indicates that the return on
portfolio H should always be higher than that of portfolio L.

Table 2 Cash Flow for Portfolio H and L

Statement Probability (%) L (£) H (£)
Low 1 2 1
Medium 2 4 4
High 97 6 100k

Notes. This table presents the cash flows with

probabilities for three statements (low,
medium and high).

16
 However, in this example, portfolio H plots below portfolio L for most possible regions
except the statement of Low where portfolio L generates £2 and portfolio H earns £1. Hence,
portfolio H fails to dominate portfolio L in such situation, although for most investors would
choose portfolio H over L since H offers a greater amount of money than L at the same
probability level. The inability of standard stochastic dominance also reveals a pathological
utility function that only a few investors feel indifferently on receiving £6 or £100k, which is
economically irrelevant or irrational (Bali et al., 2013).

4.3.1. Almost First-Order Stochastic Dominance (AFSD)

Almost First-Order Stochastic Dominance (AFSD) might solve the problem of a

pathological utility function by excluding a few extreme utility functions and examining
whether the small violation of FSD can be ‘ignored’ (Leshno and Levy, 2002). Additionally,
to see whether AFSD exists, it is necessary that one distribution should be ‘close to’ to specific
distribution that dominates another distribution in the traditional sense of FSD. In other words,
if two cumulative distributions are close to FSD, then AFSD should be used rather than ASSD
(Leshno and Levy, 2002).

To interpret the concept of almost stochastic dominance, we define the violation area
that interfere the success of FSD. When considering whether portfolio H (red solid line)
dominates L (blue dashed line) by AFSD, the area that cumulative distribution of H is above
the cumulative distribution of L is called the violation area (denoted by M in Figure 4), which
results in the failure of FSD.

According to (Bali et al., 2013), the violation area M could be defined as

∫ [𝐹 (𝑠) − 𝐹 (𝑠)]𝑑𝑠, where the FSD violation range is given by

𝑅 (𝐹 , 𝐹 ) = {𝑠 ∈ (𝑟 , 𝑟 ): 𝐹 (𝑠) < 𝐹 (𝑠)} (1)

The empirical violation area is defined as

∫ [𝐹 (𝑠) − 𝐹 (𝑠)]𝑑𝑠
𝜀 = (2)
∫ |𝐹 (𝑠) − 𝐹 (𝑠)|𝑑𝑠

where 𝐹 and 𝐹 have a finite support [min, max]. Equation (1) measures the violation range
where the probability of H is larger than L (range [1,2] in Figure 4).

17
 Figure 4 Almost First-Order Stochastic Dominance

Notes. This figure reports the cumulative distribution of high return portfolio (H) and low return
portfolio (L). Because of the violation area M, H fails to dominate the by FSD, SSD or Mean-Variance
approach. There are some extreme utility functions that assigns large weight to area M and a small or
zero weight to area N. However, most of investors would choose H over L, which indicates the
existence of AFSD if the violation area is small enough.

Moreover, according to Equation (2), 𝜀 is defined as the area above [1,2] (area M in Figure
4) divided by the total absolute area enclosed between 𝐹 and 𝐹 (area 𝑀 + 𝑁 in Figure4). It
is clear that FSD may exist if 𝜀 = 0, which implies no violation area at all. However, for 𝜀 >
0, although H fails to dominate L by FSD, AFSD may exist if 𝜀 is small enough to be ‘ignored’.
Levy et al. (2010) conducted the empirical test to find the minimum tolerance and suggests that
the violation area could be ‘ignored’ if 𝜀 is smaller or equal than 5.9%, which implies AFSD
may exist if the violation area is smaller than the minimum accepted loss of investors.

To examine whether AFSD is applicable, we evaluate cumulative distributions of

cryptocurrency portfolios against benchmarks such as stock, bond and risk-free rate for short
and long horizon. The minimum returns on cryptocurrency portfolios being always smaller
than the benchmarks in the sample period, which indicates the CDF of portfolios always starts
prior to benchmarks. As Leshno and Levy (2002) emphasize, the precondition to test AFSD is
the cumulative distribution should be in traditional sense of FSD. In other words, firstly,
ignoring the violation area and evaluating whether the rest part of CDF has a traditional sense
of FSD; secondly, continue the test if the CDF meets the precondition.

Figure 5 gives details for understanding AFSD, which exhibits the empirical CDF
distribution for a portfolio against three benchmarks. Specifically, the red lines represent the

18
CDF of the portfolio and blue lines are the benchmarks. According to Figure 5, the portfolio
cannot dominate the benchmark by FSD due to the violation area that CDF of portfolio plots
above that of benchmarks. However, the portfolio may dominate the benchmarks by AFSD if
∗
the violation area is small enough. Moreover, the set of preference 𝒰 is defined as

1
𝒰 ∗ (𝜀 ) = {𝑢 ∈ 𝒰 ; 𝑢 (𝑠) ≤ Inf{𝑢 (𝑠)} −1 ,
𝜀 (3)
∀ s ∈ (min, max)},

Figure 5 Cumulative Distribution of Real Data

Notes. This figure reports the CDF of a portfolio against benchmarks. In this figure, red line represents
the CDF of a portfolio and blue line represents stock, bond and risk-free rate. It is clear that H cannot
dominate L by FSD, SSD or MV approach since there exist a violation area. L cannot dominate H due
to the fact H has a higher expected return.

4.3.2. Almost Second-Order Stochastic Dominance (ASSD)

Similar to AFSD, there might exist almost second-order stochastic dominance (ASSD)
when the cumulative distribution is close to traditional sense of SSD. In this paper, we first
define the violation area in ASSD.

Figure 6 illustrates a case that H (red solid line) has a higher mean than L (blue dashed
line) but due to the negative area C, H fail to dominate L by SSD. Specifically, because the
integrated area between 𝐹 and 𝐹 become negative at the beginning of area C, which violates
the SSD condition ∫ [𝐹 (𝑠) − 𝐹 (𝑠)]𝑑𝑠 ≥ 0 for all given 𝑠. The range of SSD violation area
(C in this paper) could be defined as

𝑅 (𝐹 , 𝐹 ) = {𝑠 ∈ 𝑅 (𝐹 , 𝐹 ): [ 𝐹 (𝑠) − 𝐹 (𝑠)]𝑑𝑠 < 0} (4)

19
Also, the empirical ASSD violation area 𝜀 may be defined by

∫ [𝐹 (𝑠) − 𝐹 (𝑠)]𝑑𝑠
𝜀 = (5)
∫ |𝐹 (𝑠) − 𝐹 (𝑠)|𝑑𝑠

where 𝜀 is defined as the violation area (area C in Figure 6) divided by the total absolute area
enclosed between 𝐹 and 𝐹 . Similar to AFSD, Levy et al. (2010) suggests that the threshold
∗
value 𝜀 of ASSD is 3.2%, which reveals the minimum tolerance for most investors. Hence,
∗
if 𝜀 is smaller than or equal to 𝜀 of 3.2%, ASSD may exist. Additionally, the set of
∗
preference 𝒰 is defined by

1
𝒰 ∗ (𝜀 ) = {𝑢 ∈ 𝒰 ; −𝑢 (𝑠) ≤ Inf{−𝑢 (𝑠)} −1 ,
𝜀 (6)
∀ s ∈ (min, max)},

where 𝒰 excludes the utility function of pathological preference and for all 𝑢 ∈ 𝒰 ∗ , H
dominates L by ASSD if and only if 𝜀 ≤ 𝜀 ∗ .

Figure 6 Almost Second-Order Stochastic Dominance

Notes. This figure demonstrates a case that portfolio H has a higher mean than that of L but there
is no SSD of H over L due to there exist a negative area C, which makes the SSD condition
failed. Nevertheless, if area C is relatively small, ASSD may exist.

Hence, in this paper, the violation area for ASSD would be same as that of AFSD but
needs to be compared with a lower critical value of 3.2% because ASSD/SSD focuses on risk
averters, which is developed by Levy et al (2010).

20
4.4. Portfolio Formation

To alleviate the anomalies of returns on cryptocurrencies and capture the generic

properties of cryptocurrencies, we build zero-investment long-short based on a list of
established factors complied with Liu et al. (2019) and Feng et al. (2020). We select factors
that can be established by only using the available information such as opening price, closing
price, high price, low price, trading volume and market capitalization.

The construction of Zero-Investment long-short portfolios

Category Factor Used Definition

Size MARCAP Log last day market capitalization in the portfolio formation week
Size LPRC Log last day price in the portfolio formation week
Size MAXPRC The maximum price of portfolio formation week
Size Age The number of existent weeks that listed on [Link]
Momentum MOM1 One-week momentum
Momentum MOM2 Two-week momentum
Momentum MOM3 Three-week momentum
Momentum MOM4 Four-week momentum
Momentum MOM8 Eight-week momentum
Momentum RMOM1 One-week risk-adjusted momentum based on Sharpe ratio
Momentum RMOM2 Two-week risk-adjusted momentum based on Sharpe ratio
Momentum RMOM3 Three-week risk-adjusted momentum based on Sharpe ratio
Momentum RMOM4 Four-week risk-adjusted momentum based on Sharpe ratio
Momentum RMOM8 Eight-week risk-adjusted momentum based on Sharpe ratio
Volume VOL Log average daily volume in the portfolio formation week
Volume VOLPRC Log average daily volume times price in the portfolio formation week
Volume VOLSCALE Log average daily volume times price then divided by market capitalization
in the portfolio formation week
Volatility RETVOL The standard deviation of daily returns in the portfolio formation week
Volatility RETSKEW The skewness of daily returns in the portfolio formation week
Volatility RETKURT The kurtosis of daily returns in the portfolio formation week
Volatility MAXRET The maximum daily return of the portfolio formation week
Volatility STDPRCVOL Log standard deviation of dollar volume in the portfolio formation week
Volatility MEANABS The mean absolute daily return divided by dollar volume in the portfolio
formation week
Volatility BETA The regression coefficient of 𝛽 in 𝑅 − 𝑅 = 𝛼 + 𝛽 𝑀𝐾𝑇 + 𝜀 . The
model is estimated by using the daily return of previous 365 days before
formation week
Volatility BETA^2 Beta squared
The idiosyncratic volatility is measured as the standard deviation of the
Volatility IDIOVOL residual after estimating 𝑅 − 𝑅 = 𝛼 + 𝛽 𝑀𝐾𝑇 + 𝜀 . The model is
estimated by using the daily return of previous 365 days before formation
week.
Volatility DELAY The improvement of 𝑅 in
𝑅 −𝑅 =𝛼 +𝛽 𝑀𝐾𝑇 + 𝛽 𝑀𝐾𝑇 +
𝛽 𝑀𝐾𝑇 + 𝜀 compared to regression that only uses
𝑀𝐾𝑇, where 𝑀𝐾𝑇 and 𝑀𝐾𝑇 are lagged one- and two-day
market index return. The model is estimated by using the daily
return of previous 365 days before formation week.

Notes. This table reports the construction of each portfolio based on specific factors. For instance, the procedure
of establish a portfolio based on MARCAP is that sort each cryptocurrency by market cap into quintiles in
formation week, then track the return of each portfolio in the week that follows. All portfolios are rebalanced
weekly. MKT refers to coin market return that is value-weighted return of all selected cryptocurrencies.

21
 Table 3 shows the constructions of different zero-investment long-short portfolios
based on related factors. To grasp a comprehensive understanding of the investment value of
cryptocurrencies, we identify four main aspects which are size, momentum, volume and
volatility to form the corresponding portfolios to test. In addition, the main four aspects can be
further divided into 18 meaningful factor-portfolios to be examined for stochastic dominance
against different assets’ benchmarks such as stock, bond and risk-free rate.

To form a portfolio, we sort individual cryptocurrency into quintiles in ascending order

according to corresponding factors in each week, then track the return of each portfolio in the
following week. Five quintile portfolios are acquired by repeating this procedure for each factor
mentioned in Table 3. Subsequently, excess mean return over risk-free rate of each quintile
portfolio is calculated for each factor portfolio, and we also compute the excess return of long-
short portfolios based on the difference between fifth quintile and first quintile, which
represents the long-short strategy. For instance, suppose there are 𝑁 periodic returns (weekly
in this paper) on cryptocurrencies for data period from 2014 to 2019 denoted as 𝐶 , , where 𝑖
represents the 𝑖 cryptocurrency and 𝑖 ∈ [1, 𝑁]; 𝑗 indicates the specific weekly time point and
𝑗 ∈ [2014,2019]. 𝐶 , is a matrix of 𝑗 rows (number of weeks) and 𝑖 columns (number of
cryptocurrencies). Additionally, we sort 𝐶 , in ascending order by a given factor in Table 3 at
every time j to acquire a new return matrix denoted as 𝐶 ∗, based on the give factor. Then we
divide 𝐶 ∗, into quintiles to form the five portfolios and denote the first quintile portfolio as
𝐶[∗ , ], , the second quintile portfolio as 𝐶[∗ , ], ⋯ and the fifth quintile portfolio
𝐶[∗ , ], , where the 1𝑠𝑡, 2𝑛𝑑 ⋯ 5𝑡ℎ represent quintiles. After five quintile return series are
established, multiply their corresponding weight and sum them up to obtain five value-weighed
quintile portfolios 𝑃 , where 𝑚 indicates the 𝑚 portfolio. Moreover, the mean excess return
of 𝑃 is net of risk-free rate. The following paragraphs in this section will explain the detailed
factor-portfolio and their mean excess returns.

4.4.1. Size-Related Portfolios

The size-factor-portfolios are constructed based on market capitalization, last price,

maximum price and age factors. To illustrate, this paper sorts individual cryptocurrency into
quintile portfolios based on given factors in each week, then tracks the return in the following

22
week and repeats this procedure to establish quintile portfolios for each given size-related
factor. Moreover, we estimate the excess return of each portfolio over risk-free rate and the
excess return of zero-investment long-short investment which is represented by the difference
between fifth quintile portfolio and first quintile portfolio (e.g. subtracting first quintile from
fifth quintile).

Table 4 Portfolio Returns Based on Size

Quintiles
1 2 3 4 5 5-1

MARCAP Low High

Mean 0.0298 0.0121 -0.0049 -0.0033 -0.0018 -0.0316

LPRC Low High

Mean 0.0323 0.0093 0.0324 0.0102 0.0095 -0.0228

MAXPRC Low High

Mean 0.0176 0.0053 0.0149 -0.0051 -0.0057 -0.0233

Age Low High

Mean 0.0080 0.0092 0.0038 -0.0014 0.0028 -0.0052

Note. This table presents the weekly mean excess returns on quintile
portfolios for market capitalization, last price, maximum price and age
factors. The mean excess returns are defined as the excess value-
weighted expected returns, and ‘5-1’ represents the long-short strategy.

Table 4 demonstrates the expected excess return for each quintile portfolio that is
created on given factors. Fifth quintile portfolio minus first quintile portfolio represents the
mimic long-short strategy - the negative sign before the value is not important because a
positive return could be gained by conducting opposite action. For instance, the negative sign
before the difference between the fifth and first quintile portfolios suggests that investors might
loss return of 0.0316 per week if investors long the portfolio with the largest market
capitalization and short the portfolio with the smallest market capitalization; however, an
investor could gain a return of 0.0316 per week by taking an opposite action such as long the
portfolio with the smallest market capitalization and short the portfolio with the largest market
capitalization, which is similar to the size effect of stocks.

The portfolio of MARCAP generates the highest absolute return of 0.0316 on long-
short strategy, the portfolio of AGE has an absolute return of 0.0052 on long-short strategy
which is the worst performing of the portfolios in the size group. Moreover, the absolute returns
on portfolios based on LPRC and MAXPRC are similar, they are 0.0228 and 0.0233,
respectively. As noted above, the negative sign could be eliminated by conducting the opposite

23
action which is to long first quintile portfolio and short fifth quintile portfolio. Moreover,
compared to the mean return of summary statistics in Table 1, it is clear that the long-short
portfolios have much higher expected returns than that of individual cryptocurrency and other
benchmarks. Hence, these portfolios are used to test the investment value of cryptocurrencies.

4.4.2. Momentum-Related Portfolios

The momentum-factor-portfolios are formed based on one-, two-, three-, four-, eight-
week momentum and risk-adjusted momentum (Sharpe ratio) factors. Specifically, we sort
individual cryptocurrencies into quintile portfolios based on given factors in each week, then
trace the return in the following week and repeat this procedure to form the quintile portfolios
for each given momentum-related factor. In addition, we estimate the excess return of each
portfolio over the risk-free rate and excess return of zero-investment long-short portfolio which
is defined as the difference between fifth quintile portfolio and first quintile portfolio (e.g.
subtracting first quintile portfolio from the fifth quintile portfolio). In terms of momentum
portfolios, the low quintile portfolio indicates the badly performed portfolio (e.g. first quintile)
and the high quintile portfolio represents the winning portfolios (e.g. fifth quintile).

Table 5 reports the expected excess return for each quintile portfolio that is created on
given momentum factors. The fifth quintile portfolio minus the first quintile portfolio illustrates
the procedure of going long the winning portfolio and short the losing portfolio. The highest
expected excess return of 0.0367 on momentum-portfolio is established by using a three-week
momentum factor, whereas the portfolio of eight-week momentum provides lowest expected
return of 0.0149. Hence, expected excess return of portfolios shows an almost monotonically
increasing pattern for one-week momentum factor to three-week momentum factor, and it starts
to drop from three-week momentum factor to eight-week momentum factor. The long-short
strategy for momentum-portfolio applied in this paper exhibits similar effect to that of stock.

Conversely, for risk-adjusted momentum portfolios, RMOM1 has the highest return of
0.0341 and the RMOM8 has the worst performance of 0.0202. In contrast to the almost
monotonic increase pattern of momentum portfolios, the risk-adjusted momentum portfolios
exhibit an almost monotonic decreasing pattern as the horizon becomes longer (one-week to
eight-week horizon). The return on RMOM8 is 1 percent less than that of RMOM1 per week,
which shows the risk-adjusted momentum strategy could best perform in relatively short period
(one-week to four-week horizons).

24
 Table 5 Portfolio Returns Based on Momentum
Quintiles
1 2 3 4 5 5-1

MOM1 Low High

Mean 0.0019 -0.0084 0.0035 0.0197 0.0365 0.0346

MOM2 Low High

Mean 0.0054 -0.0065 0.0006 0.0284 0.0355 0.0300

MOM3 Low High

Mean 0.0031 -0.0016 0.0104 0.0128 0.0398 0.0367

MOM4 Low High

Mean 0.0047 -0.0073 0.0053 0.0146 0.0309 0.0262

MOM8 Low High

Mean 0.0060 0.0067 0.0093 0.0219 0.0209 0.0149

RMOM1 Low High

Mean 0.0012 0.0025 0.0069 0.0231 0.0353 0.0341

RMOM2 Low High

Mean -0.0035 0.0088 0.0072 0.0157 0.0280 0.0315

RMOM3 Low High

Mean 0.0052 0.0006 0.0074 0.0083 0.0296 0.0244

RMOM4 Low High

Mean 0.0013 -0.0035 0.0127 0.0140 0.0288 0.0275

RMOM8 Low High

Mean 0.0009 0.0088 0.0161 0.0157 0.0211 0.0202

Note. This table presents the weekly mean excess returns on quintile
portfolios for one-, two-, three-, four- and eight-week momentum (risk-
adjusted momentum) factors. The mean excess returns are defined as the
excess value-weighted expected returns, and ‘5-1’ represents the long-
short strategy.

4.4.3. Volume-Related Portfolios

The volume-factor-portfolios are constructed based on volume, volume times price and
scaled volume times volume factors. We sort individual cryptocurrency into quintile portfolios
based on given factors in each week, then track the return in the following week and repeats
this procedure to establish quintile portfolios for each volume-related factor. We estimate the
excess return of each portfolio over risk-free rate and the excess return of zero-investment long-
short portfolio which is represented by the difference between fifth quintile portfolio and first
quintile portfolio (e.g. subtracting first quintile from fifth quintile).

25
 Table 6 Portfolio Returns Based on Volume
Quintiles
1 2 3 4 5 5-1

VOL Low High

Mean 0.0398 0.0460 0.0281 0.0104 0.0117 -0.0281

VOLPRC Low High

Mean 0.0390 0.0389 0.0258 0.0080 0.0100 -0.0291

VOLSCALE Low High

Mean 0.0248 0.0245 0.0171 0.0150 0.0091 -0.0157

Note. This table presents the weekly mean excess returns on quintile
portfolios for volume, volume times price and scaled volume times price
factors. The mean excess returns are defined as the excess value-weighted
expected returns, and ‘5-1’ represents the long-short strategy.

Table 6 presents the expected excess return for each quintile portfolio that is created on
given factors. Fifth quintile portfolio minus first quintile portfolio represents the mimic long-
short strategy, the negative sign before the return on long-short portfolio could be turned into
positive by conducting long first quintile portfolio and short fifth quintile portfolio. In respect
of portfolios based on volume, the highest absolute excess return is 0.0291 which is generated
by portfolio based on volume times price, whereas the portfolio of scaled volume times price
delivers the lowest absolute return of 0.0157. It is clear that for every factor in volume
portfolios, the first quintile portfolio always generates the highest return among all quintiles,
which discloses the relationship between the volume and mean return.

4.4.4. Volatility-Related Portfolios

The volatility-factor-portfolios are constructed based on standard deviation of return,

skewness of return, kurtosis of return, maximum return, log standard deviation of dollar volume
and mean absolute daily return scaled by dollar volume factors. As an illustration, we sort
individual cryptocurrencies into quintile portfolios based on given factors in each week, then
track the return in the following week and repeat this procedure to establish quintile portfolios
for given volatility-related factors. In addition, we estimate the excess return of each portfolio
over risk-free rate, and the excess return of zero-investment long-short investment which is
represented by the difference between the fifth quintile portfolio and the first quintile portfolio
(e.g. subtracting first quintile from the fifth quintile).

26
 Table 7 Portfolio Returns Based on Volatility
Quintiles
1 2 3 4 5 5-1

RETVOL Low High

Mean 0.0075 0.0025 0.0159 0.0121 0.0350 0.0275

RETSKEW Low High

Mean 0.0060 -0.0003 0.0145 0.0039 0.0181 0.0121

RETKURT Low High

Mean 0.0098 0.0046 0.0104 0.0046 0.0116 0.0018

MAXRET Low High

Mean 0.0068 0.0099 0.0190 0.0231 0.0367 0.0299

STDPRCVOL Low High

Mean 0.0414 0.0294 0.0266 0.0115 0.0080 -0.0334

MEANABS Low High

Mean 0.0075 0.0181 0.0227 0.0408 0.0373 0.0298

BETA Low High

Mean 0.0113 0.0230 0.0205 0.0206 0.0052 -0.0061

BETA2 Low High

Mean 0.0113 0.0231 0.0205 0.0206 0.0052 -0.0061

IDIOVOL Low High

Mean 0.0120 0.0160 0.0148 0.0236 0.0371 0.0251

DELAY Low High

Mean 0.0157 0.0036 0.0118 0.0205 0.0219 0.0062

Note. This table presents the weekly mean excess returns on quintile
portfolios for RETVOL, RETSKEW, RETKURT, MAXRET,
STDPRCVOL, MEANABS, BETA, BETA^2, IDIOVOL, DELAY. The
mean excess returns are defined as the excess value-weighted expected
returns, and ‘5-1’ represents the long-short strategy.

Table 7 shows the expected return for each quintile portfolio that is created on given
factors. The fifth quintile portfolio minus first quintile portfolio represents the mimic long-
short strategy. Particularly, portfolio based on log standard deviation of dollar volume could
achieve the highest absolute mean return of 0.0334, and portfolio based on BETA and BETA^2
deliver the lowest return of 0.0061. The low return on BETA and BETA^2 is counterintuitive
since the portfolio with higher beta (systematic risk) should be compensated with higher return,
yet this phenomenon cannot be found in our work. This might be because that we constructed
the coin market index using the value-weighted average method, Bitcoin occupied the largest
share of it among all the cryptocurrencies, so the coin market index is influenced by Bitcoin
significantly; whereas other cryptocurrencies might not have such a property and have only a

27
little relationship with market index, hence beta might not be an appropriate factor to rank the
cryptocurrencies to form the portfolios. Moreover, the returns on quintile portfolios of
maximum return shows conspicuous increasing pattern, and the returns on quintile portfolios
for log standard deviation of return exhibits distinct decreasing trend from first quintile to fifth
quintile.

5. Summary Statistics of Factor Portfolios

This section illustrates the summary statistics of factor portfolios based on the factors
in table 3. The large skewness and kurtosis indicate that the distributions of factor portfolios
heavily deviate from normality, which motivates us to find an alternative method to measure
the performance of factor portfolios.

Table 8 Descriptive Statistics

Mean Median Std Skewness Kurtosis J-B test P value
MARCAP 0.0316 0.0253 0.1517 2.0147 21.7848 4813.75 0.001
LPRC 0.0242 0.0138 0.1421 2.1506 16.3488 2565.19 0.001
MAXPRC 0.0233 0.0129 0.1435 2.0957 15.8456 2381.11 0.001
AGE -0.0052 -0.0034 0.1438 -0.5203 8.4189 397.09 0.001
MOM1 0.0346 0.0093 0.2307 1.8286 12.9793 1468.51 0.001
MOM2 0.0300 0.0119 0.2348 0.7013 10.7627 808.95 0.001
MOM3 0.0367 0.0093 0.2363 0.3991 12.1101 1083.71 0.001
MOM4 0.0262 0.0151 0.2149 1.1057 8.9437 519.49 0.001
MOM8 0.0149 0.0014 0.2062 1.0367 10.5786 787.11 0.001
RMOM1 0.0341 0.0154 0.2000 1.6654 14.5713 1890.91 0.001
RMOM2 0.0315 0.0200 0.1875 1.4430 16.5212 2485.00 0.001
RMOM3 0.0244 0.0202 0.2166 -0.2546 14.7155 1781.94 0.001
RMOM4 0.0275 0.0197 0.1791 0.6379 9.3132 535.83 0.001
RMOM8 0.0202 0.0063 0.1762 1.5710 15.7319 2192.65 0.001
VOL 0.0281 0.0223 0.2161 0.8855 22.1340 4815.59 0.001
VOLPRC 0.0291 0.0077 0.1765 2.1719 15.2141 2191.69 0.001
VOLSCALE 0.0157 -0.0026 0.1580 2.5540 15.2384 2293.63 0.001
RETVOL 0.0275 -0.0032 0.2469 0.8554 14.0169 1621.08 0.001
RETSKEW 0.0121 0.0020 0.1707 0.1087 7.8823 311.49 0.001
RETKURT 0.0018 0.0025 0.1677 0.0427 7.7197 290.60 0.001
MAXRET -0.0299 -0.0075 0.2615 0.1757 17.8606 2881.71 0.001
STDPRCVOL 0.0334 0.0054 0.1811 2.3954 15.0932 2206.61 0.001
MEANABS 0.0298 0.0107 0.1809 2.3847 15.6372 2379.40 0.001
BETA -0.0061 -0.0031 0.1449 -0.4172 6.4648 138.12 0.001
BETA^2 -0.0061 -0.0031 0.1449 -0.4176 6.4635 138.05 0.001
IDIOVOL 0.0251 0.0053 0.1787 3.0674 23.8671 5144.64 0.001
DELAY 0.0062 -0.0013 0.1433 0.5956 7.6729 252.90 0.001

This table reports the summary statistics such as mean, median, stand deviation, skewness, kurtosis,
J-B test and corresponding p values.

Table 8 shows the mean, median, standard deviation, skewness, kurtosis, J-B test and
their p values for total 27 factor portfolios. For instance, the portfolio of MOM3 has the highest

28
mean weekly return of 0.367, whereas the portfolio of MAXRET has the lowest average
weekly return of -0.0299. Moreover, the portfolio of MAXRET is the most volatile portfolio
with standard deviation of 0.2615, the portfolio of LPR is the steadiest portfolio with a standard
deviation of 0.1421. Recalling the mean return of S&P is 0.0019 and the standard deviation is
0.0174 in Table 1, investors might compare the performance between factor portfolios and S&P
500 by using some simple risk-adjusted metrics. However, it is necessary to examine the return
distribution before conducting performance comparison because the assumption of normality
is vital. As demonstrated in Table 8, there are dramatic departures from normality in the return
distribution of factor portfolios. To illustrate, the empirical return distributions are highly
skewed and peaked around the mean. we conduct the Jarque-Bera (J-B) test to analyze whether
factor portfolios are normally distributed, the extremely large J-B statistics for every factor
portfolio reject the null hypothesis of normal distribution.

The considerable J-B statistics indicate the non-normality of factor portfolios, which
does not allow us to study the factor portfolios via widely used methods based on an assumption
of normal distribution. Hence, we are motivated to develop another methodology that not based
on assumption of normality to investigate the factor portfolios in the aspect of performance.

6. Empirical Analysis

In this section, we analyze the empirical result of AFSD, ASSD and decomposition of
zero-investment for 4-week, 13-week, 26-week, 52-week and 78-week horizon portfolios
based on factors listed in Table 3. The formation for n-week long-horizon is as follows: sum
the value from the first position of return series to 1 + (𝑛 − 1) position to get the first n-
week value, then sum the value from the second position of return series to 2 + (𝑛 − 1)
position to get the second n-week value, after that we can iteratively get a series of n-week
return. We employ this method to construct different horizon portfolios for 4-week, 13-week,
26-week, 52-week and 78-week portfolio to examine the investment value of cryptocurrencies.

6.1. AFSD

To test whether AFSD and ASSD exist, we need to calculate the ratio of violation area
(M in Figure 4) over total enclosed area (M + N in figure 4) for each portfolio and compare
these ratios 𝜀 with critical value of AFSD (5.9%) and ASSD (3.2%). We define 𝐴 as the area
between the cumulative distributions when the cumulative distribution of a portfolio plots

29
above the cumulative distribution of benchmarks (e.g. S&P 500, T-Bill and T-Bond). Likewise,
we define 𝐴 as the area between the cumulative distributions when the cumulative distribution

of a benchmark plots above the portfolio. The measure of 𝜀 for AFSD is 𝜀 = by

conducting empirical test. If a portfolio has a 𝜀 that is smaller than critical value of AFSD
(ASSD), then we conclude that this portfolio dominates the AFSD (ASSD). The violation area
is positively related to 𝜀 , so we can horizontally compare the violation area among the
portfolios if they are against same benchmark (e.g. S&P 500, T-Bill and T-Bond) through the
value of 𝜀 .

Table 9 demonstrates AFSD empirical results of 𝜀 values for each factor portfolio
compared with S&P 500 index, T-Bill, T-Bonds and cryptocurrency index for 4-week to 78-
week investment horizons. Among four panels, it is clear that almost every 𝜀 of portfolios is
monotonically decreasing as investment horizon become longer (4-week to 78-week horizon)
except portfolios based on age, kurtosis of returns, beta and beta squared. As 𝜀 is positively
related to violation area, the increasing 𝜀 indicates the growing violation area, which reveals
that these factors might not be efficient factors to construct profitable portfolios since these
their factor portfolios neither dominate the benchmarks nor have long horizon investment value.

6.1.1. AFSD against S&P 500

Panel A of Table 9 reports the empirical 𝜀 values for each factor portfolio compared to
S&P 500 index for 4-week to 78-week investment horizons. Among 4-week and 13-week
horizons, no portfolios show dominance over the S&P 500 until the holding period is extended
to a 26-week horizon. Specially, 𝜀 is directly determined by the size of violation area, the
violation area is mainly determined by extreme negative values of two cumulative distribution
functions (CDF). Hence, relatively large 𝜀 indicates that the negative returns of portfolio
cannot be absorbed enough by positive returns after extending holding period from 4-week
horizon to 13-week horizon. In this paper, 10 portfolios out of 27 portfolios exhibit AFSD at
26-week horizon as their 𝜀 values are less than 5.9%, they are portfolios of LPRC, MAXPRC,
MOM1, MOM2, MOM3, ROMO1, RMOM2, RMOM3, RMOM4 and STDPRCVOL,
respectively. Specifically, the portfolio of LPRC has the smallest empirical 𝜀 of 1.43%, which
is the best of total six dominant portfolios. In contrast, the portfolio of RMOM1 has the largest
empirical 𝜀 of 5.05%, which is the worst of the total ten dominant portfolios. All ten portfolios’
empirical values are smaller than 5.9% of the AFSD critical value, which reveals that although
portfolios of cryptocurrencies based on factors above have much higher volatility than S&P

30
500, yet their exceptional return could compensate the corresponding risk. Additionally, there
are 17 portfolios out of 27 portfolios AFSD against benchmarks at 52-week horizon. To
illustrate, apart from the ten portfolios that dominate the S&P 500 at both 26-week and 52-
week horizon, seven new portfolios join the dominant group. They are Portfolios of MARCAP,
VOLPRC, VOLSCALE, RETVOL, RETSKEW, STDPRCVOL, MEANABS and IDIOVOL,
respectively. Among 17 dominant portfolios at 52-week horizon, 10 portfolios (portfolios of
LPRC, MAXPRC, MOM1, MOM2, MOM3, RMOM1, RMOM2, RMOM3, RMOM4,
STDPRCVOL) not only dominate S&P 500 at 26-week horizon but also dominate S&P 500 at
52-horizons, and the remaining dominant portfolios have no such properties. However, 7 new
dominant portfolios emerged at 52-week horizon disclose that negative returns are absorbed by
positive returns after extending the investment horizon, which might be a proof of investment
value for cryptocurrencies. In this case, 𝜀 equals 0 for portfolio of MAXPRC and RMOM2 are
the smallest violation among 17 outperformed portfolios, indicating FSD since there does not
exist violation area. Nevertheless, Portfolio of RETVOL has the largest 𝜀 of 5.44%, which is
very close to critical value of AFSD, revealing that this portfolio is the worst performed
portfolio in 13 outperformed portfolios. The dominant portfolios at 52-week horizon behave
better as the investment horizon extends to 78 weeks. The portfolio of VOL becomes a
dominant portfolio against S&P 500 and be the member of the dominant portfolios group,
which makes the total 18 portfolios dominate the stock index. Moreover, the portfolios with
tiny 𝜀 values start to dominate the S&P 500 in the sense of FSD (see example such as LPRC,
MOM1, RMOM1, RMOM2 and RMOM3), which reveals that the extreme negative return
might be offset by significant positive return in long-term horizon.

6.1.2. AFSD against T-Bond

Panel B of Table 9 reports empirical 𝜀 values for each factor portfolio compared to ten-
year T-Bond for 4-week to 78-week investment horizons. We find the same dominant
portfolios as in Panel A at 26-week to 78-week horizon after conducting the empirical AFSD
test. For instance, portfolio of LPRC still possesses the smallest 𝜀 value of 1.52%, indicating
the best performance among six dominant portfolios at 26-week horizon. Moreover, portfolio
of RMOM1 with a 𝜀 of 5.23% is the worst in ten dominant portfolios. As investment horizon
extends to 52 weeks, we also get the same outperformed portfolio as in Panel A (17 out of 27
portfolios). The portfolio of RMOM2 remains the best performer with 𝜀 of 0, and the portfolio
of RETVOL is the worst dominant portfolio with 𝜀 of 5.15%. Moreover, as the movement of
returns on T-bond is steadier than stock index, portfolio of MAXPRC does not show FSD as

31
in Panel A against S&P 500 index. As holding period lengthens to 78-week horizon, the
dominant portfolios are the same as those in Panel A but with smaller 𝜀 value. In particular,
portfolio base on MARCAP, MOM2, RMOM4, VOLPRC, STDPRCVOL becomes FSD
against the T-bond. Hence, this phenomenon also provides the evidence that portfolios based
on cryptocurrency factors tend to dominate more of the benchmark with less volatility (T-bond)
compared to more volatile benchmark (S&P 500).

6.1.3. AFSD against T-Bill and Cryptocurrency Index

Panel C of Table 9 demonstrates empirical 𝜀 values for each factor portfolio compared
to one-month T-Bill for 4-weel to 78-week investment horizons. For 26-week horizon, we get
the same dominant portfolios as in Panel A, but the best performing portfolio is that based on
MOM1 with 𝜀 equals 2.16% rather than the portfolio based on LPRC that is the best dominant
portfolio in Panel A and B. However, the portfolio of RMOM1 with a 𝜀 of 4.98% remains the
worst performer in Panel A and B. Hence, we might infer that portfolio of cryptocurrencies
based on certain factors might behave similar compared to stock index and bond index. The
outperformance of portfolio based on MOM1 in Panel C indicates that the extreme negative
value of MOM1 might be small in order to dominate risk-free rate by smallest violation area at
26-week horizon. For 52-week horizon, the portfolio of RMOM2 with the smallest 𝜀 of 0 is
the best performed among 17 outperformed same as in Panel A and Panel B, and portfolio on
RETVOL with 𝜀 of 5.4% represents the worst performed dominant portfolio same as in Panel
A and Panel B. In regard to 78-week horizon, the dominant portfolios in Panel A and B still
show dominance against T-bill except portfolio based on STDPRCVOL deteriorates from 𝜀
from 0 to 0.06%. Additionally, we found the 𝜀 value of MAXRET is -1 for S&P 500, T-bond
and T-bill, indicating that those benchmarks dominate the portfolio of MAXRET. However,
given that such a portfolio dominated by benchmarks is in a minority, we can still conclude
that the dominant 52-week and 78-week factor portfolios are behaving the same against S&P
500, T-bond and T-bill, which might confirm the investment value of cryptocurrencies.

Panel D of Table 9 illustrates the 𝜀 values of factor portfolios against cryptocurrency

index (CMKT in Table 1). Unlike benchmarks of traditional assets, the AFSD first appears at
4-week horizon. Specifically, portfolios based on LPRC, MAXPRC and RMOM4 dominate
the cryptocurrency index. Nevertheless, the persistence of the performance of such portfolios
cannot last long, since none of the three portfolios above dominate cryptocurrency market in
long horizons (52-week and 78-week horizon). We also find that the portfolios with a 𝜀 value

32
of -1 or very large value (AGE, RETKURT, MAXRET, BETA and BTEA^2) that close to 1
do not dominate the traditional benchmarks either in any horizons, which performs cross
validation to previous test. In addition, most portfolios based on momentum and risk-adjusted
momentum dominate the cryptocurrency market in the sense of FSD, as they previously
dominated other traditional benchmarks.

To sum up, we test the AFSD of 27 factor portfolios against four benchmarks such as
S&P500, T-bond, T-bill and cryptocurrency index for 4-week to 78-week horizons. We find
there are ten portfolios dominating every benchmark at 52-week and 78-week horizons, they
are portfolios based on MOM1, MOM2, MOM3, RMOM1, RMOM2, RMOM3, VOLPRC,
RETVOL, STDPRCVOL and MEANABS. The dominant portfolios are classified as
momentum and volatility category, which proposes that portfolios constructed on such factors
might generate high excess return compared to other factor portfolios.

6.2. ASSD

In addition to AFSD in Table 10, we conduct ASSD for each portfolio at 4-week to
78-week horizon in Table 10. The key difference between AFSD and ASSD is the critical value,
ASSD has a lower critical value of 3.2% than that of AFSD 5.9%, which implies ASSD has
less tolerance of loss than that of AFSD. Similar to AFSD, the ASSD first appears at 26-week
horizon against stock, bond and risk-free rate, and the ASSD against cryptocurrency index first
appears at 4-week horizon.

6.2.1. ASSD against S&P 500

Panel A of Table 10 documents empirical 𝜀 values for each factor portfolio compared
to S&P 500 for 4-week to 52-week investment horizons. From 4-week to 13-week horizon,
none of portfolios dominate S&P 500 index. For 26-week horizon, only 5 out of 27 portfolios
dominate benchmark in the sense of ASSD. These are portfolios of LPRC, MAXPRC, MOM1,
MOM3 and RMOM2, respectively. Among the dominant portfolios, the portfolio of LPRC is
the best performing with 𝜀 1.43% and the portfolio of ROMO2 performs worst with 𝜀 2.95%.
As the critical value decreases, the number of dominant portfolios in ASSD drops by five, yet
the best performing portfolio is still the portfolio of LPRC which has the smallest violation
area. As the investment horizon extends to 52 weeks, the number of dominant portfolios falls
from 17 in AFSD to 14 in ASSD. The portfolios of MAXPRC and RMOM2 have 𝜀 of 0 which
indicates ASD has no violation area, and is the best performed portfolio among 14 dominant
portfolios. Among the remaining 12 dominant portfolios, two portfolios (MARCAP, LPRC)

33
belong to size-related factor, seven portfolios (MOM-1,2,3 and RMOM-1,2,3,4) come from
momentum-related factor, only one portfolio (VOLPRC) from volume-related factor and three
portfolios come from volatility-related factor. Hence, we conclude that volume-related factor
might not be an effective factor to construct portfolio of cryptocurrencies. As the investment
horizon extends to 78-week horizon, 16 out of 27 portfolios exhibit ASSD against the S&P 500,
showing similar outcome to AFSD. One interesting observation is that market risk factor
(BETA and BETA^2) shows no impact on abnormal positive return since they cannot dominate
the S&P 500 in terms of any investment horizon. This result apparently conflicts with both the
CAPM model and the Fama-French three-factor model which assume a market factor is an
important risk factor to explain returns. However, cryptocurrencies have existed for only a
relatively short compared to traditional assets, the insufficient dataset might not provide a
comprehensive insight into cryptocurrencies.

6.2.2. ASSD against T-Bond

Panel B in Table 10 documents empirical 𝜀 values for each factor portfolio compared
to T-bond for 4-week to 52-week investment horizons. There are no portfolios that dominate
the T-bond until 26-week, and the dominant portfolios are the same as those in Panel A except
for portfolio of RMOM2. For a 26-week horizon, the portfolio of LPRC (last price) with the
smallest 𝜀 value of 1.52% performs best among four dominant portfolios, whereas the
portfolio of MOM3 (three-week momentum) performs badly with a 𝜀 of 2.81%. For a 52-week
horizon, the portfolio of RMOM2 (risk-adjusted momentum) has the smallest 𝜀 value of 0 and
the poorly performing portfolio of VOLPRC has the largest 𝜀 value of 0.98%. For a 78-week
horizon, dominant portfolios are same as those in Panel A except the 𝜀 values are generally
smaller than those in Panel A. This might be because T-bond is less variant than S&P 500 and
results in a smaller violation area.

6.2.3. ASSD against T-Bill and Cryptocurrency Index

Panel C in table 10 reports the 𝜀 values for each factor portfolio compared to treasury
bill for 4-week to 52-week investment horizons. The dominance first appears at 26-week
horizon, but a difference is that only four portfolios (LPRC, MAXPRC, MOM1 and RMOM2)
dominate the Treasure Bill compared to five dominant portfolios in Panel A. Another
noteworthy point is that the portfolio of MOM1 with a 𝜀 of 2.16% becomes the best
performing portfolio, and the portfolio of RMOM2 with a 𝜀 of 2.95% is the worst performing
portfolio among the three outperforming portfolios. Moreover, the portfolios behave similarly

34
to dominant portfolios in Panel A, B and C, providing evidence that portfolios will dominate
three traditional benchmarks in long horizon if portfolios dominate one of three benchmarks.

Panel D in table 10 demonstrates the 𝜀 value for each portfolio against cryptocurrency
index. Similar to Panel D in Table 9, ASSD of LPRC and MAXPRC first occurs at 4-week
horizon. However, the first dominant portfolios fail to continue extraordinary performance as
investment horizon becomes longer, which is identical with the result in AFSD. Since the
dominant portfolios outperformed three benchmarks at same time, the investment value of
cryptocurrencies based on certain factors might have been identified. There are several
portfolios such as MOM1, MOM3, MOM4, VOLPRC, STDPRCVOL and MEANABS
dominate cryptocurrency index for 13-week to 78-week horizon.

To conclude, we test the ASSD of 27 factor portfolios against four benchmarks such as
S&P500, T-bond, T-bill and cryptocurrency index for 4-week to 78-week horizons. There are
nine out of 27 portfolios dominate four benchmarks at the same time. They are portfolio based
on MOM1, MOM2, MOM3, RMOM1, RMOM2, RMOM3, VOLPRC, STDPRCVOL and
MEANABS. Compared to ten dominant portfolios in the sense of AFSD, portfolio based on
RETVOL failed to ASSD benchmarks. Therefore, we have nine portfolios show AFSD and
ASSD in total.

6.3. Decompose of Long-Short Portfolios

6.3.1 Decompose the Long/Short Portfolios against Four Indices

So far, this paper has considered the investment values of cryptocurrencies by

conducting AFSD and ASSD on long-short factor portfolios. To investigate which counterpart
(long/short) contributes most to ASD, we decompose the all the long-short portfolios into long
only and short only parts and continue AFSD and ASSD. For instance, we list the mean return
on quintile portfolios for each factor from Table 4 to Table 7, and we assume 5-1 as the normal
action. If the 5-1(long fifth quintile and short first quintile) portfolio is negative, then we
employ the opposite action that is 1-5 (long first quintile and short fifth quintile) to make our
portfolio is profitable. We also conduct ASD against some stock benchmarks based on factors
from Carhart four-factor model (Carhart, 1997) and Fama-French three-factor model (Fama
and French, 1993) to examine the performance of cryptocurrency portfolios. The data of stock
portfolios based on corresponding factors is collected from Kenneth French’s website.

35
Table 9 Almost First-Order Stochastic Dominance

Panel D: Portfolios against Cryptocurrency

Panel A: Portfolios against S&P 500 Index Panel B: Portfolios against Ten-year T-Bond Panel C: Portfolios against One-month T-Bill
Index

4- 13- 26- 52- 78- 4- 13- 26- 52- 78- 4- 13- 26- 52- 78- 4- 13- 26- 52- 78-
Portfolios
week week week week week week week week week week week week week week week week week week week week
MARCAP 0.2283 0.1382 0.0648 0.0075* 0.0006* 0.2354 0.1349 0.0634 0.0079* 0.0000* 0.2416 0.1406 0.0678 0.0084* 0.0000* 0.0620 0.0835 0.0511* 0.0633 0.0299*
LPRC 0.2723 0.1307 0.0140* 0.0003* 0.0000* 0.2777 0.1339 0.0152* 0.0009* 0.0000* 0.2836 0.1429 0.0239* 0.0014* 0.0000* 0.0203 0.0519* 0.1262 0.1304 0.1029
*
MAXPRC 0.2896 0.1363 0.0170* 0.0000* 0.0000* 0.2932 0.1378 0.0178* 0.0005* 0.0000* 0.2983 0.1481 0.0267* 0.0009* 0.0000* 0.0208 0.0430* 0.1379 0.1571 0.1462
*
AGE 0.5706 0.5998 0.6374 0.6763 0.6955 0.5551 0.5770 0.6077 0.6263 0.6049 0.5483 0.5649 0.5905 0.6028 0.5823 0.9281 -1.0000 -1.0000 0.9777 -1.0000
MOM1 0.2740 0.1180 0.0160* 0.0033* 0.0000* 0.2771 0.1196 0.0170* 0.0037* 0.0000* 0.2810 0.1267 0.0216* 0.0042* 0.0000* 0.1121 0.0075* 0.0000* 0.0000* 0.0000*
MOM2 0.3224 0.1721 0.0370* 0.0005* 0.0001* 0.3230 0.1714 0.0370* 0.0008* 0.0000* 0.3252 0.1761 0.0425* 0.0011* 0.0000* 0.2425 0.0586* 0.0031* 0.0000* 0.0000*
MOM3 0.2522 0.0799 0.0280* 0.0024* 0.0014* 0.2565 0.0836 0.0280* 0.0029* 0.0009* 0.2608 0.0905 0.0330* 0.0041* 0.0015* 0.1043 0.0064* 0.0000* 0.0000* 0.0000*
MOM4 0.3127 0.1946 0.1522 0.0983 0.0848 0.3134 0.1913 0.1445 0.0883 0.0728 0.3162 0.1956 0.1487 0.0920 0.0756 0.1842 0.0076* 0.0251* 0.0106* 0.0754
MOM8 0.4176 0.3363 0.2924 0.2016 0.1267 0.4124 0.3262 0.2772 0.1850 0.1035 0.4119 0.3257 0.2771 0.1861 0.1074 0.4417 0.3390 0.4625 0.6606 0.8756
RMOM1 0.2655 0.1278 0.0505* 0.0028* 0.0000* 0.2736 0.1334 0.0520* 0.0038* 0.0000* 0.2695 0.1297 0.0498* 0.0033* 0.0000* 0.1398 0.1007 0.0210* 0.0000* 0.0000*
RMOM2 0.2860 0.1415 0.0295* 0.0000* 0.0000* 0.2923 0.1482 0.0340* 0.0000* 0.0000* 0.2887 0.1428 0.0295* 0.0000* 0.0000* 0.1567 0.0737 0.0001* 0.0053* 0.0000*
RMOM3 0.3071 0.1328 0.0372* 0.0023* 0.0000* 0.3116 0.1401 0.0400* 0.0028* 0.0000* 0.3088 0.1348 0.0369* 0.0026* 0.0000* 0.2199 0.0924 0.0189* 0.0085* 0.0000*
RMOM4 0.2713 0.1014 0.0374* 0.0025* 0.0001* 0.2802 0.1126 0.0440* 0.0035* 0.0000* 0.2754 0.1044 0.0364* 0.0029* 0.0000* 0.0527 0.0000* 0.0000* 0.0959 0.1928
*
RMOM8 0.3616 0.2503 0.2381 0.1922 0.1791 0.3612 0.2484 0.2282 0.1787 0.1475 0.3598 0.2454 0.2263 0.1773 0.1501 0.2924 0.1318 0.2291 0.5353 0.9591
VOL 0.2694 0.1638 0.1176 0.0800 0.0501* 0.2733 0.1639 0.1130 0.0758 0.0443 0.2777 0.1691 0.1177 0.0781 0.0466* 0.1232 0.0410* 0.0039* 0.0000* 0.0195*
VOLPRC 0.2887 0.1449 0.0659 0.0093* 0.0004* 0.2907 0.1446 0.0634 0.0098* 0.0000* 0.2946 0.1514 0.0701 0.0110* 0.0000* 0.0846 0.0000* 0.0000* 0.0001* 0.0000*
VOLSCALE 0.3747 0.2982 0.2056 0.0505* 0.0461* 0.3708 0.2888 0.1896 0.0428* 0.0356* 0.3719 0.2898 0.1929 0.0481* 0.0395* 0.3311 0.2464 0.2665 0.1612 0.0298*
RETVOL 0.3739 0.2733 0.1649 0.0544* 0.0093* 0.3718 0.2681 0.1578 0.0515* 0.0076* 0.3725 0.2698 0.1615 0.0540* 0.0090* 0.3492 0.2066 0.0764* 0.0295* 0.0172*
RETSKEW 0.3785 0.2792 0.1393 0.0377* 0.0050* 0.3745 0.2699 0.1249 0.0343* 0.0029* 0.3756 0.2727 0.1341 0.0389* 0.0050* 0.2505 0.2968 0.4244 0.3926 0.4629
RETKURT 0.5031 0.5258 0.5789 0.5963 0.6017 0.4919 0.5061 0.5498 0.5490 0.5411 0.4880 0.4974 0.5353 0.5314 0.5276 0.7608 0.9802 -1.0000 -1.0000 -1.0000
MAXRET 0.6728 0.7898 0.9144 0.9903 -1.0000 0.6614 0.7786 0.9046 0.9867 -1.0000 0.6548 0.7702 0.8941 0.9818 -1.0000 0.8577 0.0000* 0.0000* 0.0001* -1.0000
STDPRCVO 0.2655 0.1453 0.0399* 0.0032* 0.0005* 0.2686 0.1449 0.0392* 0.0037* 0.0000* 0.2730 0.1507 0.0451* 0.0045* 0.0006* 0.0624 0.0055* 0.0000* 0.0000* 0.0000*
L
MEANABS 0.2812 0.1559 0.0676 0.0014* 0.0000* 0.2838 0.1554 0.0652 0.0019* 0.0000* 0.2800 0.1613 0.0714 0.0024* 0.0000* 0.0647 0.0063* 0.0000* 0.0000* 0.0000*
BETA 0.5991 0.6467 0.6718 0.7197 0.7956 0.5751 0.6121 0.6365 0.6654 0.7398 0.5830 0.6250 0.6483 0.6836 0.7216 0.9390 -1.0000 -1.0000 -1.0000 -1.0000
BETA^2 0.5988 0.6463 0.6713 0.7190 0.7947 0.5749 0.6117 0.6330 0.6646 0.7387 0.5827 0.6245 0.6478 0.6828 0.7206 0.9389 -1.0000 -1.0000 -1.0000 -1.0000
IDIOVOL 0.3113 0.1693 0.0730 0.0037* 0.0004* 0.3152 0.1740 0.0774 0.0058* 0.0000* 0.3117 0.1665 0.0683 0.0045* 0.0000* 0.1080 0.0053* 0.0194* 0.1641 0.1159
DELAY 0.4756 0.4498 0.4215 0.3445 0.2852 0.4633 0.4304 0.3954 0.3104 0.2350 0.4662 0.4351 0.4010 0.3147 0.2345 0.6746 0.8440 0.8893 0.9251 0.9902

Notes. This table presents the empirical estimates of 𝜀 for 4-week to 78-week investment horizon. To complete the stochastic dominance, this paper defines 𝐴 as the area between the cumulative distributions when the cumulative distribution of
a portfolio plots above the cumulative distribution of benchmarks (e.g. S&P 500, T-Bond, T-Bill and Cryptocurrency Index). Likewise, this paper defines 𝐴 as the area between the cumulative distributions when the cumulative
distribution of a benchmark plots above the long-short portfolio. The measure of 𝜀 for AFSD is 𝜀 = by conducting empirical test. Moreover, the critical value for AFSD is 𝜀 ∗ = 5.9%, if the 𝜀 for any portfolio is less than the
critical value, then the result indicates the outperformance of that portfolio. This paper uses * to label portfolios which is less than the critical value. The 𝜀 value of -1 indicates that benchmarks dominate corresponding portfolio.

36
Table 10 Almost Second-Order Stochastic Dominance

Panel D: Portfolios against Cryptocurrency

Panel A: Portfolios against S&P 500 Index Panel B: Portfolios against Ten-year T-Bond Panel C: Portfolios against One-month T-Bill
Index
4- 13- 26- 52- 78- 4- 13- 26- 52- 78- 4- 13- 26- 52- 78- 4- 13- 26- 52- 78-
Portfolios
week week week week week week week week week week week week week week week week week week week week
MARCAP 0.2283 0.1382 0.0648 0.0075* 0.0006* 0.2354 0.1349 0.0634 0.0079* 0.0000* 0.2416 0.1406 0.0678 0.0084* 0.0000* 0.0620 0.0835 0.0511 0.0633 0.0299*
LPRC 0.2723 0.1307 0.0140* 0.0003* 0.0000* 0.2777 0.1339 0.0152* 0.0009* 0.0000* 0.2836 0.1429 0.0239* 0.0014* 0.0000* 0.0203 0.0519 0.1262 0.1304 0.1029
*
MAXPRC 0.2896 0.1363 0.0170* 0.0000* 0.0000* 0.2932 0.1378 0.0178* 0.0005* 0.0000* 0.2983 0.1481 0.0267* 0.0009* 0.0000* 0.0208 0.0430 0.1379 0.1571 0.1462
*
AGE 0.5706 0.5998 0.6374 0.6763 0.6955 0.5551 0.5770 0.6077 0.6263 0.6049 0.5483 0.5649 0.5905 0.6028 0.5823 0.9281 -1.0000 -1.0000 0.9777 -1.0000
MOM1 0.2740 0.1180 0.0160* 0.0033* 0.0000* 0.2771 0.1196 0.0170* 0.0037* 0.0000* 0.2810 0.1267 0.0216* 0.0042* 0.0000* 0.1121 0.0075* 0.0000* 0.0000* 0.0000*
MOM2 0.3224 0.1721 0.0370 0.0005* 0.0001* 0.3230 0.1714 0.0370 0.0008* 0.0000* 0.3252 0.1761 0.0425 0.0011* 0.0000* 0.2425 0.0586 0.0031* 0.0000* 0.0000*
MOM3 0.2522 0.0799 0.0280* 0.0024* 0.0014* 0.2565 0.0836 0.0280* 0.0029* 0.0009* 0.2608 0.0905 0.0330 0.0041* 0.0015* 0.1043 0.0064* 0.0000* 0.0000* 0.0000*
MOM4 0.3127 0.1946 0.1522 0.0983 0.0848 0.3134 0.1913 0.1445 0.0883 0.0728 0.3162 0.1956 0.1487 0.0920 0.0756 0.1842 0.0076* 0.0251* 0.0106* 0.0754
MOM8 0.4176 0.3363 0.2924 0.2016 0.1267 0.4124 0.3262 0.2772 0.1850 0.1035 0.4119 0.3257 0.2771 0.1861 0.1074 0.4417 0.3390 0.4625 0.6606 0.8756
RMOM1 0.2655 0.1278 0.0505 0.0028* 0.0000* 0.2736 0.1334 0.0520 0.0038* 0.0000* 0.2695 0.1297 0.0498 0.0033* 0.0000* 0.1398 0.1007 0.0210* 0.0000* 0.0000*
RMOM2 0.2860 0.1415 0.0295* 0.0000* 0.0000* 0.2923 0.1482 0.0340 0.0000* 0.0000* 0.2887 0.1428 0.0295* 0.0000* 0.0000* 0.1567 0.0737 0.0001* 0.0053* 0.0000*
RMOM3 0.3071 0.1328 0.0372 0.0023* 0.0000* 0.3116 0.1401 0.0400 0.0028* 0.0000* 0.3088 0.1348 0.0369 0.0026* 0.0000* 0.2199 0.0924 0.0189* 0.0085* 0.0000*
RMOM4 0.2713 0.1014 0.0374 0.0025* 0.0001* 0.2802 0.1126 0.0440 0.0035* 0.0000* 0.2754 0.1044 0.0364 0.0029* 0.0000* 0.0527 0.0000* 0.0000* 0.0959 0.1928
RMOM8 0.3616 0.2503 0.2381 0.1922 0.1791 0.3612 0.2484 0.2282 0.1787 0.1475 0.3598 0.2454 0.2263 0.1773 0.1501 0.2924 0.1318 0.2291 0.5353 0.9591
VOL 0.2694 0.1638 0.1176 0.0800 0.0501 0.2733 0.1639 0.1130 0.0758 0.0443 0.2777 0.1691 0.1177 0.0781 0.0466 0.1232 0.0410 0.0039* 0.0000* 0.0195*
VOLPRC 0.2887 0.1449 0.0659 0.0093* 0.0004* 0.2907 0.1446 0.0634 0.0098* 0.0000* 0.2946 0.1514 0.0701 0.0110* 0.0000* 0.0846 0.0000* 0.0000* 0.0001* 0.0000*
VOLSCALE 0.3747 0.2982 0.2056 0.0505 0.0461 0.3708 0.2888 0.1896 0.0428 0.0356 0.3719 0.2898 0.1929 0.0481 0.0395 0.3311 0.2464 0.2665 0.1612 0.0298*
RETVOL 0.3739 0.2733 0.1649 0.0544 0.0093* 0.3718 0.2681 0.1578 0.0515 0.0076* 0.3725 0.2698 0.1615 0.0540 0.0090* 0.3492 0.2066 0.0764 0.0295* 0.0172*
RETSKEW 0.3785 0.2792 0.1393 0.0377 0.0050* 0.3745 0.2699 0.1249 0.0343 0.0029* 0.3756 0.2727 0.1341 0.0389 0.0050* 0.2505 0.2968 0.4244 0.3926 0.4629
RETKURT 0.5031 0.5258 0.5789 0.5963 0.6017 0.4919 0.5061 0.5498 0.5490 0.5411 0.4880 0.4974 0.5353 0.5314 0.5276 0.7608 0.9802 -1.0000 -1.0000 -1.0000
MAXRET 0.6728 0.7898 0.9144 0.9903 -1.0000 0.6614 0.7786 0.9046 0.9867 -1.0000 0.6548 0.7702 0.8941 0.9818 -1.0000 0.8577 0.0000* 0.0000* 0.0001* -1.0000
STDPRCVO 0.2655 0.1453 0.0399 0.0032* 0.0005* 0.2686 0.1449 0.0392 0.0037* 0.0000* 0.2730 0.1507 0.0451 0.0045* 0.0006* 0.0624 0.0055* 0.0000* 0.0000* 0.0000*
L
MEANABS 0.2812 0.1559 0.0676 0.0014* 0.0000* 0.2838 0.1554 0.0652 0.0019* 0.0000* 0.2800 0.1613 0.0714 0.0024* 0.0000* 0.0647 0.0063 0.0000* 0.0000* 0.0000*
BETA 0.5991 0.6467 0.6718 0.7197 0.7956 0.5751 0.6121 0.6365 0.6654 0.7398 0.5830 0.6250 0.6483 0.6836 0.7216 0.9390 -1.0000 -1.0000 -1.0000 -1.0000
BETA^2 0.5988 0.6463 0.6713 0.7190 0.7947 0.5749 0.6117 0.6330 0.6646 0.7387 0.5827 0.6245 0.6478 0.6828 0.7206 0.9389 -1.0000 -1.0000 -1.0000 -1.0000
IDIOVOL 0.3113 0.1693 0.0730 0.0037* 0.0004* 0.3152 0.1740 0.0774 0.0058* 0.0000* 0.3117 0.1665 0.0683 0.0045* 0.0000* 0.1080 0.0053* 0.0194* 0.1641 0.1159
DELAY 0.4756 0.4498 0.4215 0.3445 0.2852 0.4633 0.4304 0.3954 0.3104 0.2350 0.4662 0.4351 0.4010 0.3147 0.2345 0.6746 0.8440 0.8893 0.9251 0.9902

Notes. This table presents the empirical estimates of 𝜀 for 4-week to 78-week investment horizon. To complete the stochastic dominance, this paper defines 𝐴 as the area between the cumulative distributions when the cumulative
distribution of a portfolio plots above the cumulative distribution of benchmarks (e.g. S&P 500, T-Bond, T-Bill and Cryptocurrency Index). Likewise, this paper defines 𝐴 as the area between the cumulative distributions when
the cumulative distribution of a benchmark plots above the long-short portfolio. The measure of 𝜀 for ASD is 𝜀 = 𝜀 = by conducting empirical test. Moreover, the critical value for AFSD is 𝜀 ∗ = 3.2%, if the 𝜀 for any
portfolio is less than the critical value, then the result indicates the outperformance of that portfolio. This paper uses * to label portfolios which is less than the critical value. The 𝜀 value of -1 indicates that benchmarks dominate
corresponding portfolio.

37
6.3.2 AFSD of Long-Leg Portfolios against S&P 500

Table 12 reports the empirical 𝜀 value of each long leg of portfolio in the sense of
AFSD and ASSD against four benchmarks. We start with AFSD then ASSD. Panel A of Table
12 illustrate empirical 𝜀 for the long part of each portfolio against S&P 500. Unlike the
dominance for the long-short portfolios in Table 9, none of any portfolios shows dominance
against the S&P 500 at a 26-week horizon. As the investment horizon extends to 52 weeks, the
phenomenon that total 13 portfolios out of 27 portfolios AFSD the S&P 500 appears. Among
these 13 dominant portfolios, the portfolio of VOL has the smallest 𝜀 value of 0.92%, whereas
the portfolio of RETVOL performs worst with an 𝜀 value of 5.08%. Although the best
performed long-short portfolios are portfolio of MAXPRC and RMOM2 with 𝜀 values of 0 at
52-week horizon, none of each long leg portfolio dominate the stock index. Furthermore, long
leg portfolios of MARCAP (𝜀 equals 7.96%) and LPRC (𝜀 equals 18.48%) are also absent
from the dominant group without long-short strategy. Compared to the result in Table 9 of
long-short strategy, we find that the dominance of portfolios based on size factor might be
susceptible to the choice of trading strategy. For 78-week horizon, the number of dominant
portfolios against all four benchmarks increased to 21, and the non-dominant portfolios are
only LPRC, MAXPRC, AGE, RETKURT, BETA and BETA^2, respectively. This evidence
shows that nearly 77.8% (21 out of 27 portfolios) of factor portfolios could dominate the S&P
500 in long term, which justifies the fact that cryptocurrencies can generate excess high return
as long as long holding period.

6.3.3 AFSD of Long-Leg Portfolios against T-Bond

Panel B of Table 12 reports the empirical 𝜀 for the long leg of each portfolio against
T-bond. The performance of portfolios is similar to Panel A of Table 12, none of the portfolios
exhibiting dominance until a 52-week horizon. However, 15 portfolios out of 27 portfolios
AFSD the T-bond at the 52-week horizon, which has two more dominant portfolios compared
to dominant portfolios in Panel A. Moreover, the best performing portfolio is still VOL with 𝜀
0.9% and the poorest portfolios 5.9%, which just meet the condition of AFSD. As the dominant
portfolio is increased by two (MOM8 and RMOM2), we conclude that portfolios tend to behave
relatively better when compared to a steadier benchmark and show more dominance. For a 78-
week horizon, the same portfolios as in Panel A dominate the T-bond; all 𝜀 values of portfolios
in Panel B are smaller than those in Panel A except for portfolios of RMOM1, RMOM3, VOL
and STDPRCVOL that have zero 𝜀 values in both Panel A and Penal B.

38
6.3.4 AFSD of Long-Leg portfolios against T-Bill

Panel C of Table 12 reports the AFSD for long leg of portfolios against T-bill. There
are no dominant portfolios for 4-week to 26-week horizon, which is the same as Panel A and
Panel B. In Panel C of Table 12, 13 same portfolios as in Panel A dominate T-bill in the sense
of AFSD. Among the dominant portfolios, portfolio of VOL performed best with a 𝜀 value
equals 0.1%, whereas portfolio of RETVOL performed poorest with a 𝜀 value 5.07%. In
contrast, recalling there are 17 dominant long-short portfolios at 52-week horizon against S&P
500 in Table 9, and 13 long-only portfolios dominate S&P 500 at 52-week horizon in Table 12,
the number of dominant portfolios drops by four. This indicates that the excess return could be
increased if investors conduct appropriate investment strategies. As investment horizon
extends to 78 weeks, the dominant portfolios against T-bill are same as portfolios in Panel A
and B despite the dominant portfolios are various at 52-week horizon for each panel.

6.3.5 AFSD of Long-Leg portfolios against Cryptocurrency Index

Panel D of Table 12 represents the AFSD for long-leg portfolios against cryptocurrency
index. Unlike the other three panels, AFSD first appears in 13-week, and the dominant
portfolios at 52-week and 78-week horizons are the same. The factor portfolios and
cryptocurrency index are identical type of asset, which ensures their comparability. Therefore,
on one hand, we conclude that the performance of factor portfolios is superior to every four
selected indices, and even perform better when compare to cryptocurrency index. On the other
hand, we assert that factor portfolios can alleviate the risk of high return since most long-leg
factor portfolios dominate index of identical and different types of assets.

6.3.6 AFSD of Short-Leg Portfolios against Four Benchmarks

Table 13 reports the AFSD for short leg of portfolios against different benchmarks.
Among Panel A, B and C in Table 13, there exist no dominant portfolio at any investment
horizon. Moreover, the 𝜀 value for each portfolio increases as the investment horizon become
longer, which illustrates that portfolios only perform poorly if investors long the portfolios
labelled as short portfolio. We first examine the 52-week and 78-week investment horizon
because there would exist no more dominance in shorter period if no significant dominance
can be found in 52-week and 78-week horizon. Specifically, Panel A of Table 13 demonstrates
the 𝜀 value for each portfolio labelled as short portfolio. At 52-week horizon, the portfolio of
LPRC has the smallest 𝜀 value of 50%, which is almost 10 times of the critical value of 5.9%.
Even the best performed portfolio cannot meet the condition, therefore no short portfolios
39
dominate the S&P 500 in our case. At 78-week horizon, the extremely large 𝜀 value does not
improve as time goes by, even portfolios of VOL and STDPRCVOL have a 𝜀 over 90%,
indicating that the S&P 500 benchmark dominate the factor portfolios over 90% of the time.
This phenomenon not only exist when against stock, bond and risk-free rate but also occurs in
terms of cryptocurrency index.

Furthermore, Panel B of Table 13 reports the 𝜀 value of each portfolio labelled as short
against T-bond. The best performed portfolio is still the portfolio of LPRC with a 𝜀 value of
46.8%, which cannot dominate the benchmark either. Therefore, no portfolios could dominate
the T-bond as the best portfolio even fail to show dominance. Panel C of Table 13 shows the
same performance as in Panel A and B, the portfolio of LPRC remains the best position with
𝜀 value of 45.9%, while the portfolio of IDIOVOL has the worst performance with 𝜀 value of
88.2%, which indicates nearly 90 percent of all the return cannot dominate benchmark.
Moreover, there is no ASSD for short labelled portfolios since no portfolios in this case have
𝜀 values less than ASSD critical value of 3.2%. Panel D of Table 13 demonstrates 𝜀 values of
the short-leg portfolios and most of factor portfolios are dominated by the cryptocurrency
benchmarks, as most of their 𝜀 values are -1. For instance, there are eight portfolios are
dominated by cryptocurrency index at 4-week horizon, and all of the portfolios are dominated
by benchmark as investment horizon extends to 78 weeks. It is clear that Short-leg portfolios
not only cannot gain excess return but also experience significant loss conversely. Therefore,
we conclude that long the short leg of portfolio cannot gain excess return, and the short-leg
portfolios should be used with long-leg portfolios to earn high return.

6.3.7 ASSD of Long-Leg Portfolios versus Four Benchmarks

On the other hand, some long leg of portfolios in Table 12 also dominate benchmarks
at 52-week and 78-week horizon in the sense of ASSD. In Panel A of Table 12, there are six
portfolios ASSD the S&P 500, the portfolio of VOL is the best portfolio with 𝜀 value of 0.9%,
whereas portfolio of MEANABS performed worst with 𝜀 value equals 3.1%. In contrast to the
13 ASSD long-short portfolios in Table 10, the number of ASSD long-only portfolios falls
significantly. This might prove that appropriate long-short strategy could enhance the
performance of portfolios dramatically. Furthermore, the same phenomenon of dominance
exists in Panel B and C with the same portfolios, and the best performed portfolio is still
portfolio of VOL and the worst performed portfolio is portfolio of MEANABS. For 78-week
horizon, the number of dominant portfolios raised to 16, which is the similar for Panel B and

40
C. In regard to ASSD, Panel D of Table 12 behaves differently from Panel A, B and C as their
asset classes are variant. Although the criteria of ASSD is stricter than AFSD, the number of
dominant portfolios of AFSD is consistent with that of ASSD, which reveals the investment
value of factor portfolios again. Therefore, there are five long-leg portfolios out of nine
dominant long-short portfolios dominate four benchmarks in the sense of AFSD and ASSD,
respectively. Hence, we suggest the extraordinary performance of cryptocurrency portfolios
results from risk premium as their long counterpart dominate the benchmarks, which share
similarities with equities.

To summarize, the decomposition of long and short portfolio against four benchmarks
illustrates that the power of dominance of long-short portfolios is mostly contributed by the
long leg of the portfolios, since none of the short leg of portfolios dominate benchmarks at any
investment horizons. Moreover, we document that the dominance is mainly from risk premium
on each factor. However, the decreased number of dominant long-only portfolios compared to
the amount of dominant long-short portfolio also indicates the necessity of conducting the long-
short strategy as this grant investors more probability to gain the benefits.

6.3.8 Decompose the Long/Short Portfolios against Stock Factor Portfolios

We have so far provided evidence that cryptocurrency portfolios based on factors could
dominate S&P 500, T-bond, T-bill and cryptocurrency index in long horizon, and the
dominance is mainly contributed by long-leg portfolios. In this section, we examine the
performance of long-leg portfolios against stock factor portfolios based on size, momentum
and book-to-market equity (BE/ME) factors, because Fama and French (1992) document that
size and BE/ME factors are two vital factors that could capture the variation in average stock
returns. Meanwhile, Jegadeesh and Titman (1993) propose the strategy that long the winning
stocks and short the losing stocks disclose momentum factor is also noteworthy for portfolio
selection and construction. We select portfolios with highest top 20% return for each factor to
ensure that selected stock benchmarks are profitable and attractive.

6.3.9 ASD for Long-Leg portfolios versus Stock Factor Portfolios

Panel A of Table 14 reports the 𝜀 values for each portfolio against stock portfolio based
on size. We only focus on the portfolio dominating both in the sense of AFSD and ASSD as
we want to compare them to nine dominant portfolios in last section. Hence, there are six
portfolios dominate the benchmark at 52-week horizon. Among these portfolios, five portfolios
of MOM2, RMOM1, RMOM3, STDPRCVOL and MEANABS belong to the group of nine

41
Table 12 AFSD and ASSD for Long Leg of Portfolios against S&P500, T-Bond, T-bill and Cryptocurrency Market

Panel D: Portfolios against Cryptocurrency

Panel A: Portfolios against S&P 500 Index Panel B: Portfolios against Ten-year T-Bond Panel C: Portfolios against One-month T-Bill
Index
4- 13- 26- 52- 78- 4- 13- 26- 52- 78- 4- 13- 26- 52- 78- 4- 13- 26- 52- 78-
Portfolios
week week week week week week week week week week week week week week week week week week week week
MARCAP 0.300 0.218 0.170 0.080 0.017** 0.301 0.216 0.164 0.075 0.014** 0.304 0.219 0.167 0.078 0.016** 0.183 0.123 0.095 0.010** 0.003**
LPRC 0.397 0.337 0.281 0.185 0.114 0.392 0.327 0.267 0.172 0.102 0.393 0.327 0.266 0.174 0.104 0.392 0.361 0.324 0.163 0.140
MAXPRC 0.402 0.346 0.291 0.197 0.123 0.397 0.336 0.278 0.185 0.111 0.397 0.335 0.277 0.186 0.113 0.404 0.381 0.350 0.188 0.164
AGE 0.464 0.432 0.375 0.268 0.163 0.457 0.421 0.359 0.246 0.139 0.455 0.417 0.356 0.245 0.141 0.535 0.583 0.603 0.713 0.949
MOM1 0.298 0.181 0.096 0.036* 0.005** 0.299 0.179 0.093 0.035* 0.003** 0.302 0.183 0.097 0.036* 0.005** 0.199 0.048* 0.007** 0.000** 0.000**
MOM2 0.302 0.162 0.083 0.030** 0.003** 0.303 0.161 0.080 0.029** 0.002** 0.306 0.166 0.085 0.031** 0.003** 0.207 0.039* 0.003** 0.000** 0.000**
MOM3 0.287 0.156 0.096 0.046* 0.016** 0.288 0.155 0.093 0.044* 0.013** 0.291 0.160 0.097 0.046* 0.015** 0.180 0.038* 0.026** 0.004** 0.000**
MOM4 0.323 0.189 0.114 0.065 0.033* 0.323 0.187 0.110 0.062 0.028** 0.325 0.190 0.114 0.064 0.030** 0.248 0.058* 0.027** 0.007** 0.001**
MOM8 0.368 0.240 0.149 0.062 0.032* 0.366 0.235 0.141 0.059* 0.026** 0.367 0.238 0.146 0.061 0.029** 0.326 0.087 0.013** 0.001** 0.000**
RMOM1 0.300 0.165 0.090 0.029** 0.000** 0.301 0.163 0.087 0.029** 0.000** 0.304 0.168 0.092 0.030** 0.000** 0.195 0.050* 0.003** 0.000** 0.000**
RMOM2 0.338 0.223 0.146 0.060 0.020** 0.337 0.219 0.141 0.058* 0.017** 0.339 0.222 0.144 0.060 0.019** 0.268 0.127 0.048* 0.001** 0.002**
RMOM3 0.302 0.176 0.086 0.028* 0.000** 0.303 0.175 0.083 0.027** 0.000** 0.306 0.179 0.088 0.028** 0.000** 0.189 0.050* 0.007** 0.000** 0.000**
RMOM4 0.303 0.152 0.093 0.048* 0.006** 0.304 0.152 0.090 0.046* 0.004** 0.308 0.158 0.095 0.048* 0.006** 0.174 0.019** 0.011** 0.000** 0.000**
RMOM8 0.356 0.239 0.150 0.080 0.020** 0.355 0.234 0.142 0.076 0.014** 0.356 0.237 0.147 0.078 0.018** 0.290 0.064 0.049* 0.010** 0.001**
VOL 0.267 0.146 0.087 0.009** 0.000** 0.270 0.146 0.084 0.009** 0.000** 0.274 0.150 0.088 0.010** 0.000** 0.160 0.025** 0.005** 0.000** 0.000**
VOLPRC 0.279 0.165 0.115 0.043* 0.003** 0.282 0.164 0.111 0.042* 0.002** 0.285 0.168 0.115 0.043* 0.003** 0.174 0.063 0.036* 0.002** 0.000**
VOLSCALE 0.342 0.251 0.185 0.094 0.043* 0.341 0.246 0.177 0.090 0.037* 0.343 0.248 0.180 0.092 0.039* 0.265 0.139 0.112 0.019** 0.003**
RETVOL 0.348 0.249 0.144 0.051* 0.001** 0.347 0.245 0.139 0.049* 0.000** 0.348 0.247 0.142 0.051* 0.000** 0.313 0.185 0.068 0.006** 0.000**
RETSKEW 0.362 0.269 0.200 0.096 0.044* 0.360 0.263 0.192 0.090 0.039* 0.361 0.265 0.194 0.093 0.041* 0.286 0.165 0.112 0.007** 0.005**
RETKURT 0.421 0.350 0.286 0.172 0.123 0.415 0.340 0.271 0.155 0.102 0.415 0.339 0.271 0.157 0.105 0.455 0.386 0.376 0.369 0.379
MAXRET 0.337 0.228 0.121 0.038* 0.003** 0.337 0.226 0.271 0.037* 0.002** 0.338 0.228 0.121 0.038* 0.003** 0.294 0.167 0.051* 0.003** 0.000**
STDPRCVO 0.274 0.155 0.089 0.030** 0.000** 0.277 0.155 0.271 0.029** 0.000** 0.280 0.159 0.091 0.031** 0.000** 0.172 0.061 0.017** 0.000** 0.000**
L
MEANABS 0.289 0.165 0.096 0.031** 0.002** 0.291 0.164 0.271 0.030** 0.001** 0.294 0.168 0.097 0.031** 0.002** 0.187 0.055* 0.006** 0.000** 0.000**
BETA 0.403 0.335 0.272 0.180 0.105 0.396 0.323 0.271 0.163 0.085 0.396 0.322 0.255 0.165 0.087 0.447 0.400 0.353 0.315 0.260
BETA^2 0.404 0.335 0.272 0.180 0.105 0.397 0.323 0.271 0.163 0.085 0.397 0.322 0.255 0.165 0.088 0.448 0.401 0.354 0.317 0.262
IDIOVOL 0.269 0.147 0.081 0.034* 0.008** 0.272 0.147 0.271 0.033* 0.005** 0.276 0.152 0.083 0.034* 0.007** 0.130 0.018** 0.006** 0.000** 0.000**
DELAY 0.321 0.215 0.133 0.076 0.050* 0.322 0.211 0.271 0.070 0.040* 0.325 0.215 0.130 0.073 0.043* 0.119 0.049* 0.017** 0.007** 0.000**

Notes. This table reports the 𝜀 of AFSD and ASSD for each long portfolio against S&P 500, bond, T-bill and cryptocurrency index. The portfolio with * means AFSD
against the benchmark and portfolio with ** represents both AFSD and ASSD. 𝜀 value equals -1 indicates that benchmark dominates corresponding portfolio.

42
Table 13 AFSD and ASSD for Short Leg of Portfolios against S&P500, T-Bond, T-bill and Cryptocurrency Market
Panel D: Portfolios against Cryptocurrency
Panel A: Portfolios against S&P 500 Index Panel B: Portfolios against Ten-year T-Bond Panel C: Portfolios against One-month T-Bill
Index
4- 13- 26- 52- 78- 4- 13- 26- 52- 78- 4- 13- 26- 52- 78- 4- 13- 26- 52- 78-
Portfolios
week week week week week week week week week week week week week week week week week week week week
MARCAP 0.509 0.531 0.565 0.625 0.699 0.497 0.512 0.541 0.595 0.664 0.492 0.503 0.530 0.583 0.654 0.854 0.952 -1.000 -1.000 -1.000
LPRC 0.464 0.465 0.471 0.500 0.554 0.455 0.449 0.448 0.468 0.512 0.452 0.443 0.440 0.459 0.504 0.658 0.845 0.990 -1.000 -1.000
MAXPRC 0.468 0.471 0.481 0.522 0.583 0.459 0.454 0.459 0.492 0.544 0.456 0.449 0.451 0.482 0.536 0.677 0.879 1.000 -1.000 -1.000
AGE 0.563 0.616 0.666 0.758 0.847 0.548 0.594 0.643 0.733 0.819 0.541 0.582 0.629 0.718 0.808 0.986 -1.000 -1.000 -1.000 -1.000
MOM1 0.540 0.571 0.614 0.684 0.734 0.528 0.553 0.590 0.642 0.682 0.523 0.544 0.577 0.622 0.667 0.781 0.986 -1.000 -1.000 -1.000
MOM2 0.568 0.599 0.615 0.676 0.769 0.555 0.582 0.596 0.649 0.735 0.549 0.573 0.585 0.636 0.724 0.863 0.960 0.998 -1.000 -1.000
MOM3 0.535 0.542 0.555 0.627 0.744 0.523 0.525 0.534 0.594 0.697 0.519 0.517 0.524 0.581 0.683 0.803 0.948 0.996 -1.000 -1.000
MOM4 0.568 0.620 0.675 0.813 0.897 0.554 0.603 0.655 0.789 0.872 0.548 0.593 0.642 0.773 0.861 0.868 1.000 -1.000 -1.000 -1.000
MOM8 0.589 0.661 0.717 0.813 0.917 0.575 0.642 0.697 0.793 0.898 0.569 0.631 0.683 0.779 0.890 0.886 -1.000 -1.000 -1.000 -1.000
RMOM1 0.538 0.581 0.596 0.639 0.743 0.525 0.560 0.571 0.601 0.690 0.520 0.549 0.558 0.585 0.674 0.853 0.958 -1.000 -1.000 -1.000
RMOM2 0.491 0.487 0.499 0.535 0.619 0.480 0.470 0.477 0.507 0.582 0.476 0.463 0.468 0.498 0.574 0.789 0.955 0.965 1.000 -1.000
RMOM3 0.559 0.556 0.559 0.591 0.683 0.544 0.538 0.535 0.560 0.641 0.538 0.529 0.524 0.548 0.630 0.916 0.961 -1.000 -1.000 -1.000
RMOM4 0.545 0.587 0.642 0.732 0.832 0.530 0.567 0.615 0.700 0.795 0.525 0.556 0.599 0.683 0.781 0.900 0.997 -1.000 -1.000 -1.000
RMOM8 0.548 0.590 0.628 0.700 0.814 0.535 0.573 0.611 0.680 0.791 0.529 0.564 0.600 0.669 0.783 0.877 0.956 0.992 -1.000 -1.000
VOL 0.648 0.716 0.757 0.824 0.910 0.628 0.695 0.738 0.806 0.891 0.618 0.681 0.724 0.794 0.884 -1.000 -1.000 -1.000 -1.000 -1.000
VOLPRC 0.632 0.694 0.732 0.800 0.895 0.612 0.673 0.713 0.781 0.873 0.603 0.660 0.699 0.769 0.865 -1.000 -1.000 -1.000 -1.000 -1.000
VOLSCALE 0.642 0.711 0.754 0.822 0.930 0.619 0.687 0.732 0.801 0.906 0.607 0.671 0.716 0.788 0.895 -1.000 -1.000 -1.000 -1.000 -1.000
RETVOL 0.627 0.697 0.750 0.805 0.924 0.605 0.672 0.726 0.782 0.904 0.594 0.657 0.709 0.767 0.894 -1.000 -1.000 -1.000 -1.000 -1.000
RETSKEW 0.589 0.650 0.722 0.813 0.882 0.573 0.629 0.701 0.793 0.857 0.565 0.617 0.687 0.778 0.847 0.964 -1.000 -1.000 -1.000 -1.000
RETKURT 0.601 0.683 0.792 0.906 0.957 0.587 0.666 0.774 0.889 0.944 0.581 0.655 0.759 0.873 0.936 0.879 0.981 -1.000 -1.000 -1.000
MAXRET 0.610 0.645 0.685 0.748 0.846 0.590 0.622 0.660 0.718 0.807 0.581 0.608 0.644 0.701 0.793 -1.000 -1.000 -1.000 -1.000 -1.000
STDPRCVO 0.627 0.693 0.740 0.808 0.913 0.605 0.669 0.717 0.786 0.886 0.594 0.654 0.701 0.771 0.875 -1.000 -1.000 -1.000 -1.000 -1.000
L
MEANABS 0.629 0.700 0.752 0.816 0.932 0.606 0.674 0.727 0.793 0.903 0.594 0.658 0.709 0.777 0.890 -1.000 -1.000 -1.000 -1.000 -1.000
BETA 0.582 0.651 0.713 0.837 0.917 0.567 0.630 0.691 0.811 0.895 0.560 0.617 0.675 0.793 0.884 0.947 -1.000 -1.000 -1.000 -1.000
BETA^2 0.582 0.651 0.713 0.837 0.917 0.567 0.630 0.691 0.811 0.895 0.560 0.617 0.675 0.793 0.884 0.947 -1.000 -1.000 -1.000 -1.000
IDIOVOL 0.703 0.792 0.843 0.908 0.981 0.675 0.765 0.823 0.894 0.962 0.660 0.745 0.806 0.882 0.951 -1.000 -1.000 -1.000 -1.000 -1.000
DELAY 0.676 0.742 0.799 0.882 0.940 0.658 0.724 0.782 0.869 0.926 0.648 0.711 0.768 0.858 0.919 0.999 -1.000 -1.000 -1.000 -1.000

Notes. This table reports the 𝜀 of AFSD and ASSD for each short portfolio against S&P 500, bond, T-bill and cryptocurrency index. The portfolio with * means AFSD
against the benchmark and portfolio with ** represents both AFSD and ASSD. 𝜀 value equals -1 indicates that benchmark dominates corresponding portfolio.

43
dominant portfolios. Moreover, portfolio based on RMOM3 is the best performed portfolio
with the smallest 𝜀 value of 2.86%, whereas portfolio of MEANABS is the worst portfolio
with 𝜀 value of 3.16%. For a 78-week horizon, there are 16 portfolios that outperform the
benchmark, and each portfolio of dominant portfolios beats the benchmark. Among the nine
dominant portfolios, portfolio based on STDPRCVOL has the best performance with 𝜀 value
of 0.03%.

Panel B of Table 14 demonstrates the 𝜀 values for each portfolio against stock portfolio
based on momentum, and the dominant portfolios are same as Panel A. Moreover, the best
performing portfolios at 52-week and 78-week holding periods are still portfolios based on
RMOM3 and STDPRCVOL, respectively. Panel C of Table 14 reports slightly different results
from Panel A and B. Specifically, there are six out of nine dominant portfolios dominating the
BE/ME portfolios. In addition to 5 dominant portfolios in Panel A and Panel B, portfolio based
on MOM1 also dominate BE/ME benchmark at 52-week horizon. For 78-week horizon, every
portfolio of nine dominant portfolio dominates the BE/ME benchmark, and the portfolio of
RMOM3 and STDPRCVOL are best performed among nine dominant portfolios including
MOM1-3, RMOM1-3, VOLPRC, STDPRCVOL and MEANABS.

In summary, the decomposition of long and short portfolio illustrates that the power of
dominance of long-short portfolios is mostly contributed by the long leg of the portfolios, since
none of the short leg of portfolios dominate benchmarks at any investment horizons. In the
long run (52-week and 78-week holding periods), we find five long-leg portfolios that not only
outperform the four benchmarks (S&P500, T-bond, T-bill and cryptocurrency index) but also
dominate the three stock factor benchmarks (size, momentum and BE/ME), and they are
portfolio based on MOM2, RMOM1, RMOM3, STDPRCVOL and MEANABS. However, the
decreased number of dominant long-only portfolios at 52-week horizon compared to the
amount of dominant long-short portfolio also indicates the necessity of conducting the long-
short strategy as this grant investors more probability to gain the benefits.

6.4 Regression Analysis

We have discovered nine outperforming portfolios by conducting almost stochastic

dominance. In this section, we decide to investigate whether cross-sectional return of nine
dominant portfolios can be explained by applying the coin market three factor model of Liu et
al. (2019).
44
 We start with construction of three factors. The three-factor model is inspired by the
Fama-French three factor model, and the three factors are cryptocurrency market, size and
momentum. The formation of cryptocurrency market excess is discussed in Section 2. For size
factor, we sort the coins into three size groups by market capitalization: bottom 30 percent
(small group, S), middle 40 percent (middle group, M) and top 30 percent (big group, B). We
then form value-weighted portfolios for each of three groups, and the cryptocurrency size factor
(CSMB) is the return on small group minus that on big group. We construct the momentum
factor by using three-week momentum as the return on long-short three-week momentum is
the highest among the momentum factors (See Table 5). In particular, we sort the coins into
three three-week momentum groups: bottom 30 percent, middle 40 percent and top 30 percent,
then we form value-weighted portfolios for each of the three three-week momentum groups.
The cryptocurrency momentum factor (CMOM) is the difference between return on top 30
percent portfolios and return on bottom 30 percent portfolios. The coin market three factor
model is shown below.

𝑅 −𝑅 = 𝛼 +𝛽 𝐶𝑀𝐾𝑇 + 𝛽 𝐶𝑆𝑀𝐵 + 𝛽 𝐶𝑀𝑂𝑀 + 𝜖

where 𝑅 is the return on factor portfolios, 𝑅 is risk-free rate, CMKT is the cryptocurrency
market excess return, CSMB is the cryptocurrency size factor and CMOM is the
cryptocurrency momentum factor.

Table 15 reports regression results for nine dominant portfolios from previous section.
We find that R-squared values of momentum portfolios are commonly higher than other types
of portfolios, which provides more explanatory power. For instance, portfolio of MOM3 has
the highest 𝑅 of 0.5544, whereas portfolio of MEANABS that is the most explained portfolio
in other categories, only has the 𝑅 of 0.0954. The portfolio of MOM1 is an outlier with 𝑅
0.024, while the other momentum portfolios have a relatively high 𝑅 range from 0.2574 to
0.5544, indicating that portfolio based on momentum factors could be well explained by coin
market three factor models. Furthermore, the values of 𝛼 for MOM1, VOLPRC,
STDPRCVOL and MEANABS are positive and significant at 1 percent level, suggesting that
the four portfolios might be mispriced. Combined with relatively small 𝑅 that ranges from
0.0254 for portfolio of MOM1 to 0.0954 for portfolio of MEANABS, we suggest that the return
anomalies on portfolio of MOM1, VOLPRC, STDPRCVOL and MEANABS might be caused
by mispricing.

45
Table 14 Long Leg of Portfolios Against Momentum, Size and BE/ME Portfolios
Panel A: Portfolios against Size Portfolios Panel B: Portfolios against Momentum Portfolios Panel C: Portfolios against BE/ME Portfolios
Portfolios 4-week 13-week 26-week 52-week 78-week 4-week 13-week 26-week 52-week 78-week 4-week 13-week 26-week 52-week 78-week
MARCAP 0.3019 0.2209 0.1728 0.0821 0.0173** 0.2995 0.2176 0.1705 0.0822 0.0177** 0.2873 0.2054 0.1578 0.0681 0.0128**
LPRC 0.3998 0.3420 0.2860 0.1901 0.1181 0.3981 0.3390 0.2843 0.1900 0.1189 0.3875 0.3240 0.2630 0.1655 0.1004
MAXPRC 0.4045 0.3507 0.2963 0.2028 0.1271 0.4030 0.3479 0.2947 0.2026 0.1279 0.3926 0.3334 0.2745 0.1789 0.1104
AGE 0.4666 0.4367 0.3811 0.2766 0.1708 0.4660 0.4348 0.3802 0.2763 0.1721 0.4586 0.4240 0.3612 0.2432 0.1390
MOM1 0.3004 0.1838 0.0982 0.0370* 0.0057** 0.2982 0.1803 0.0954 0.0371* 0.0060** 0.2875 0.1671 0.0842 0.0297** 0.0018**
MOM2 0.3041 0.1650 0.0856 0.0312** 0.0034** 0.3019 0.1613 0.0827 0.0313** 0.0037** 0.2913 0.1483 0.0715 0.0246** 0.0010**
MOM3 0.2889 0.1583 0.0979 0.0468* 0.0171** 0.2866 0.1549 0.0953 0.0469* 0.0174** 0.2759 0.1429 0.0853 0.0397* 0.0113**
MOM4 0.3250 0.1920 0.1170 0.0670 0.0351* 0.3230 0.1887 0.1143 0.0670 0.0356* 0.3130 0.1761 0.1009 0.0549* 0.0251**
MOM8 0.3708 0.2444 0.1538 0.0644 0.0350* 0.3690 0.2406 0.1504 0.0645 0.0356* 0.3589 0.2249 0.1294 0.0513* 0.0222**
RMOM1 0.3017 0.1676 0.0926 0.0300** 0.0006** 0.2994 0.1636 0.0897 0.0301** 0.0007** 0.2885 0.1496 0.0781 0.0241** 0.0000**
RMOM2 0.3401 0.2259 0.1489 0.0623 0.0208** 0.3381 0.2224 0.1462 0.0624 0.0212** 0.3277 0.2088 0.1326 0.0513* 0.0157**
RMOM3 0.3041 0.1788 0.0886 0.0286** 0.0004** 0.3018 0.1751 0.0856 0.0288** 0.0005** 0.2902 0.1617 0.0742 0.0224** 0.0000**
RMOM4 0.3054 0.1552 0.0955 0.0495* 0.0073** 0.3029 0.1512 0.0923 0.0497* 0.0077** 0.2905 0.1376 0.0805 0.0408* 0.0017**
RMOM8 0.3590 0.2429 0.1539 0.0833 0.0224** 0.3570 0.2391 0.1503 0.0834 0.0231** 0.3459 0.2236 0.1323 0.0689 0.0103**
VOL 0.2692 0.1484 0.0894 0.0095** 0.0001** 0.2669 0.1449 0.0869 0.0097** 0.0002** 0.2563 0.1331 0.0749 0.0071** 0.0000**
VOLPRC 0.2811 0.1672 0.1168 0.0444* 0.0034** 0.2789 0.1639 0.1145 0.0444* 0.0036** 0.2684 0.1530 0.1041 0.0371* 0.0011**
VOLSCALE 0.3450 0.2546 0.1890 0.0969 0.0457* 0.3429 0.2514 0.1868 0.0970 0.0462* 0.3317 0.2379 0.1698 0.0830 0.0344*
RETVOL 0.3494 0.2515 0.1464 0.0523* 0.0014** 0.3481 0.2490 0.1441 0.0524* 0.0016** 0.3407 0.2389 0.1324 0.0436* 0.0004**
RETSKEW 0.3649 0.2729 0.2039 0.0994 0.0464* 0.3628 0.2693 0.2014 0.0994 0.0470* 0.3505 0.2543 0.1841 0.0819 0.0361*
RETKURT 0.4240 0.3548 0.2919 0.1796 0.1305 0.4226 0.3517 0.2899 0.1789 0.1315 0.4121 0.3370 0.2667 0.1447 0.0998
MAXRET 0.3388 0.2305 0.1228 0.0392* 0.0031** 0.3373 0.2280 0.1205 0.0393* 0.0034** 0.3296 0.2185 0.1104 0.0324* 0.0014**
STDPRCVOL 0.2759 0.1577 0.0912 0.0310** 0.0003** 0.2738 0.1545 0.0888 0.0310** 0.0004** 0.2635 0.1439 0.0798 0.0251** 0.0000**
MEANABS 0.2906 0.1672 0.0987 0.0316** 0.0020** 0.2885 0.1638 0.0960 0.0317** 0.0022** 0.2778 0.1523 0.0838 0.0252** 0.0004**
BETA 0.4076 0.3407 0.2781 0.1865 0.1120 0.4055 0.3372 0.2761 0.1860 0.1128 0.3909 0.3189 0.2505 0.1543 0.0814
BETA^2 0.4079 0.3410 0.2785 0.1869 0.1123 0.4058 0.3376 0.2765 0.1864 0.1131 0.3913 0.3193 0.2509 0.1547 0.0817
IDIOVOL 0.2708 0.1492 0.0834 0.0349* 0.0085** 0.2682 0.1454 0.0803 0.0350* 0.0088** 0.2560 0.1326 0.0692 0.0271** 0.0036**
DELAY 0.3245 0.2194 0.1373 0.0794 0.0540* 0.3215 0.2146 0.1334 0.0793 0.0547* 0.3062 0.1959 0.1108 0.0595 0.0363*

Notes. This table reports the 𝜀 of AFSD and ASSD for each long-leg portfolio against size, momentum and BE/ME portfolios. The portfolio with * means AFSD against
the benchmark and portfolio with ** represents both AFSD and ASSD.

46
 Table 15 Coin Market Three-Factor model

𝛼 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0382*** 0.2673** 0.0419 -0.1453** 0.0254

T-Statistics 2.8837 2.155 0.4079 -1.9837
MOM2 0.0034 0.1396 0.1567* 0.7839*** 0.3629
T-Statistics 0.3079 1.3641 1.846 12.9671
MOM3 0.0015 0.1287 -0.2448*** 0.9433*** 0.5544
T-Statistics 0.1637 1.4957 -3.4327 18.5655
RMOM1 0.0143 0.049 0.1367* 0.5544*** 0.2574
T-Statistics 1.4547 0.5311 1.7874 10.1778
RMOM2 0.0107 0.1624** 0.1799*** 0.6055*** 0.351
T-Statistics 1.2195 1.9701 2.6311 12.4335
RMOM3 -0.0074 0.218*** -0.1754*** 0.8369*** 0.5266
T-Statistics -0.8558 2.6825 -2.603 17.4358
VOLPRC 0.033*** 0.2943*** 0.2829*** -0.1181** 0.0795
T-Statistics 3.3419 3.1836 3.6904 -2.1635
STDPRCVOL 0.0352*** 0.3454*** 0.3069*** -0.0619 0.0817
T-Statistics 3.4765 3.6467 3.9077 -1.1059
MEANABS 0.032*** 0.341*** 0.355*** -0.0594 0.0954
T-Statistics 3.1883 3.6307 4.5591 -1.0713

Notes. This table reports the results of regression analysis for nine dominant portfolios. *, **, ***
represent significance at 10%, 5%, 1% level, respectively.

In regard to cryptocurrency market factor, all non-momentum portfolios have

significant exposures to cryptocurrency market factor, and only half of momentum-
based portfolio (MOM1, RMOM2 and RMOM3) have significant relatively small
exposures compared to non-momentum strategies. In other words, the cryptocurrency
market factor can capture the cross-sectional return for non-momentum strategies. For
the coin market size factor, nearly all dominant portfolios have statistically significant
loadings but portfolio of MOM1. Moreover, size factor might have the reverse effect
on portfolio of MOM3 and RMOM3 as their exposures are less than zero. For the coin
market momentum factor, cross-sectional return of portfolios based on momentum
strategies could be well captured as all their factor loadings are statistically significant,
but other portfolios based on volatility (STDPRCVOL and MEANABS) cannot be
explained appropriately as their momentum factor loadings are not significant.

To conclude, we test whether the dominant portfolio can be explained by a coin

market three-factor model and find that the model could explain the cross-sectional
return to a certain extent. The return anomalies for portfolio of MOM1, VOLPRC,

47
STDPRCVOL and MEANABS are possibly associated with mispricing. Additionally,
momentum factor provides stronger explanatory power concerning momentum strategy,
market factor could capture the cross-sectional return of non-momentum strategy
portfolios and size factor could explain both momentum and non-momentum strategy.

6.4.1 Enhance Performance of Model by Incorporating Mispricing Factors

The previous section illustrates the explanatory power of a coin market three-
factor model when examining nine dominant portfolios, and there exist four factor
portfolios that cannot be addressed well by coin market three-factor model. Moreover,
inspired by Stambaugh and Yuan (2017) who claim that an asset pricing model might
increase its explanatory power by incorporating mispricing factors, we decide to find
factors that can explain the anomalies by incorporating unexplained portfolios in last
section. Specifically, we add the potential combinations of four unexplained factor
portfolios (MOM1, VOLPRC, STDPRCVOL and MEANABS) and two crypto
fundamental factors (Electricity and computer power) to existing coin market three-
factor model and examine the performance after incorporating mispricing factors. This
procedure creates 28 adjusted models, which can be found in an appendix.

We construct mispricing factors based on their categories. For instance, four

mispricing portfolios belong to three large categories: MOM1 of momentum factors,
VOLPRC of volume factors, STDPRCVOL and MEANABS of volatility factors.
Because there are two volatility factors, we apply equally weighted average rankings
with respect to STDPRCVOL and MEANABS, in order to form a mispricing factor that
captures the comprehensive effect of volatility. For MOM1 factor, we sort the coins
into three one-week momentum groups: bottom 30 percent, middle 40 percent and top
30 percent, then we form value-weighted portfolios for each of the three three-week
momentum groups. The mispricing cryptocurrency momentum factor (CMOM1) is the
difference between return on top 30 percent portfolios and return on bottom 30 percent
portfolios. Similarly, to construct VOLPRC, we sort the coins into three VOLPRC
groups (We refer this to volume factor): bottom 30 percent, middle 40 percent and top
30 percent, then we form value-weighted portfolios for each of the three volume groups.
The mispricing cryptocurrency volume factor (CVOL) is the difference between return
on top 30 percent portfolios and return on bottom 30 percent portfolios. Moreover, to
construct a composite mispricing factor for volatility, we apply equally weighted

48
average to rankings with respect to two anomalies, then form the mispricing volatility
factors based on averaged rankings. We sort the coins into three volatility groups,
bottom 30 percent, middle 40 percent and top 30 percent, then we form value-weighted
portfolios for each of the three volume groups. The mispricing cryptocurrency volatility
factor (CSTD) is the difference between return on top 30 percent portfolios and return
on bottom 30 percent portfolios.

We first try evaluating the performance of coin market three-factor model

incorporating mispricing momentum (CMOM1). The adjusted coin market three-factor
model is shown below after incorporating CMOM1.

𝑅 −𝑅 =𝛼 +𝛽 𝐶𝑀𝐾𝑇 + 𝛽 𝐶𝑆𝑀𝐵 + 𝛽 𝐶𝑀𝑂𝑀 + 𝛽 𝐶𝑀𝑂𝑀1

+𝜖

where 𝑅 , 𝑅 , 𝐶𝑀𝐾𝑇 , 𝐶𝑆𝑀𝐵 , 𝐶𝑀𝑂𝑀 have been illustrated above, 𝐶𝑀𝑂𝑀1 is the
mispricing momentum factor.

Table 16 reports the regression results of adjusted coin market three-factor

model. We find that the R-squared for each portfolio does not improve significantly,
indicating that CMOM1 cannot provide explanatory power to coin market three-factor
model. For example, the values of alpha for mispriced portfolios are still significant and
their R-squared are quite small, even the biggest R-squared of them is less than 0.1. We
also find that the adjusted model could capture more cross-sectional variation for
portfolios based on momentum (e.g. MOM2, MOM3, RMOM1, RMOM2) as their
exposures are statistically significant at 1% level. However, we cannot suggest that the
adjusted model improve the performance of coin market three factor model, since
dominant portfolios based on other factors still failed to be explained by incorporating
CMOM1 factor.

Next, we test the adjusted coin market three-factor model incorporating CVOL
factor. The adjusted coin market three-factor model is shown below.

𝑅 −𝑅 =𝛼 +𝛽 𝐶𝑀𝐾𝑇 + 𝛽 𝐶𝑆𝑀𝐵 + 𝛽 𝐶𝑀𝑂𝑀 + 𝛽 𝐶𝑉𝑂𝐿 + 𝜖

where 𝑅 , 𝑅 , 𝐶𝑀𝐾𝑇 , 𝐶𝑆𝑀𝐵 , 𝐶𝑀𝑂𝑀 have been illustrated above, 𝐶𝑉𝑂𝐿 is the
mispricing volume factor.

49
 Table 17 reports regression results for coin market three-factor model
incorporating volume factor. We find that mispriced portfolios of VOLPRC,
STDPRCVOL and MEANABS can be well explained by adjusted model under
conditions of not decreasing other portfolios’ explanatory power. To illustrate,
compared to R-squared of coin market three-factor model for VOLPRC, STRPRCVOL
and MEANABS which ranges from 0.0759 to 0.0954, the R-squared of adjusted model
for them are 0.5553, 0.5963 and 0.6032, respectively. More importantly, their
exposures to CVOL factor are all statistically significant at 1% level, their factor
loadings of CSMB become insignificant, which might suggest that CSMB might not
the most important factor to capture the variation among volume and volatility
portfolios. In other words, we may argue that trading volume is an important factor
driving the cryptocurrencies trading and development. On the other hand, Momentum
factor still works well for portfolios based on momentum, and their R-squared does not
exhibit considerable fluctuation.

Table 16 Three-Factor Model Incorporating CMOM1

𝛼 𝛽 𝛽 𝛽 𝛽 R-squared
MOM1 0.0385*** 0.2632*** 0.0429 -0.1184 -0.0423 0.0262
T-Statistics 2.9007 2.1151 0.4165 -1.3009 -0.5001
MOM2 0.0007 0.1736* 0.1488* 0.5586*** 0.3537*** 0.4164
T-Statistics 0.0684 1.7657 1.8291 7.7685 5.2927
MOM3 0.0003 0.1435* -0.2482*** 0.8451*** 0.1542*** 0.5644
T-Statistics 0.0384 1.6808 -3.5141 13.5377 2.6570
RMOM1 0.0095 0.1103 0.1226** 0.1481*** 0.6380*** 0.5069
T-Statistics 1.1868 1.4625 1.9637 2.6846 12.4442
RMOM2 0.0094 0.1797*** 0.1759*** 0.4909*** 0.1799*** 0.3727
T-Statistics 1.0818 2.2090 2.6124 8.2503 3.2540
RMOM3 -0.0078 0.2228*** -0.1765*** 0.8049*** 0.0503 0.5279
T-Statistics -0.8979 2.7353 -2.6183 13.5117 0.9078
VOLPRC 0.0333*** 0.2898*** 0.2839*** -0.0878 -0.0476 0.0812
T-Statistics 3.3718 3.1252 3.7009 -1.2955 -0.7552
STDPRCVOL 0.0356*** 0.3393*** 0.3083*** -0.0211 -0.0640 0.0846
T-Statistics 3.5198 3.5739 3.9250 -0.3038 -0.9925
MEANABS 0.0321*** 0.3396*** 0.3553*** -0.0507 -0.0136 0.0955
T-Statistics 3.1899 3.6034 4.5551 -0.7362 -0.2128

Notes. This table reports the results of regression analysis for nine dominant portfolios. *, **, ***
represent significance at 10%, 5%, 1% level, respectively.

50
 However, alpha for portfolio of MOM1, VOLPRC, STDPRCVOL and
MEANABS are still statistically significant at 1% level, indicating that there still exist
anomalies that cannot be captured by adjusted model.

Finally, we test the adjusted coin market three-factor model by adding

mispricing volatility factor (CSTD) to it. The adjusted model is shown below.

𝑅 −𝑅 =𝛼 +𝛽 𝐶𝑀𝐾𝑇 + 𝛽 𝐶𝑆𝑀𝐵 + 𝛽 𝐶𝑀𝑂𝑀 + 𝛽 𝐶𝑆𝑇𝐷 + 𝜖

where 𝑅 , 𝑅 , 𝐶𝑀𝐾𝑇 , 𝐶𝑆𝑀𝐵 , 𝐶𝑀𝑂𝑀 have been illustrated above, 𝐶𝑆𝑇𝐷 is the
mispricing volatility factor.

Table 17 Three-Factor Model Incorporating CVOL

𝛼 𝛽 𝛽 𝛽 𝛽 R-squared
MOM1 0.0380*** 0.2388** -0.0079 -0.1325** -0.1683 0.0320
T-Statistics 2.8767 1.9046 -0.0732 -1.7993 -1.4404
MOM2 0.0032 0.1186 0.1199 0.7934*** -0.1242 0.3664
T-Statistics 0.2976 1.1455 1.3397 13.0415 -1.2866
MOM3 0.0016 0.1394 -0.2259*** 0.9384*** 0.0637 0.5553
T-Statistics 0.1700 1.5997 -3.0000 18.3244 0.7840
RMOM1 0.0144 0.0580 0.1525** 0.5504*** 0.0533 0.2583
T-Statistics 1.4582 0.6204 1.8874 10.0197 0.6119
RMOM2 0.0107 0.1614* 0.1780** 0.6059*** -0.0061 0.3510
T-Statistics 1.2169 1.9299 2.4639 12.3325 -0.0787
RMOM3 -0.0073 0.2425*** -0.1325 0.8259*** 0.1446** 0.5321
T-Statistics -0.8438 2.9587 -1.8715 17.1536 1.8936
VOLPRC 0.0319*** 0.1005 -0.0566 -0.0312 -1.1455*** 0.5553
T-Statistics 4.8743 1.6179 -1.0551 -0.8543 -19.7913
STDPRCVOL 0.0340*** 0.1457** -0.0430 0.0277 -1.1805*** 0.5963
T-Statistics 5.1143 2.3064 -0.7870 0.7477 -20.0543
MEANABS 0.0308*** 0.1350** -0.0057 0.0330 -1.2169*** 0.6032
T-Statistics 4.9416 2.2817 -0.1114 0.9483 -22.0672

Notes. This table reports the results of regression analysis for nine dominant portfolios. *, **,*** represent
significance at 10%, 5%, 1% level, respectively.

Table 18 presents the regression results for coin market three-factor model
incorporating mispricing volatility factor (CSTD). Compared to adjusted coin market
three-factor model in Table 17, the degree of fitting deteriorates heavily as indicating
by R-squared for portfolios of VOLPRC, STDPRCVOL and MEANABS. The R-
squared for them drops significant from average 0.5 to average 0.1. Moreover, the alpha
for portfolio of MOM3 becomes significant whereas the alpha for portfolio of

51
MEANABS turns into insignificant. Although the exposures to CSTD are statistically
significant for VOLPRC, STDPRCVOL and MEANABS, yet it cannot prove the
efficiency of this mispricing factor since the adjusted model lacks explanatory power
for these three portfolios. Furthermore, the anomaly of MOM1 still cannot be captured
by developed factor because its R-squared is always small (average 0.03) and its alpha
is always statistically significant at 1% level. Therefore, further research needs to be
done to discover such anomalies.

Table 18 Three-Factor Model Incorporating CSTD

𝛼 𝛽 𝛽 𝛽 𝛽 R-squared

MOM1 0.0463*** 0.3351*** 0.0581 -0.1449** -0.0827 0.0319

T-Statistics 3.2194 2.5278 0.5628 -1.9810 -1.4314
MOM2 -0.0144 -0.0080 0.1215 0.7830*** 0.1799*** 0.3924
T-Statistics -1.2380 -0.0746 1.4545 13.2396 3.8513
MOM3 -0.0162* -0.0181 -0.2798*** 0.9423*** 0.1789*** 0.5832
T-Statistics -1.6694 -0.2031 -4.0258 19.1460 4.6016
RMOM1 0.0164 0.0663 0.1408* 0.5545 -0.0211 0.2580
T-Statistics 1.5280 0.6705 1.8280 10.1672 -0.4898
RMOM2 0.0031 0.0991 0.1648** 0.6051*** 0.0772** 0.3595
T-Statistics 0.3257 1.1282 2.4079 12.4867 2.0161
RMOM3 -0.0092 0.2033** -0.1789*** 0.8368*** 0.0180 0.5270
T-Statistics -0.9723 2.3329 -2.6358 17.4116 0.4731
VOLPRC 0.0186* 0.1747* 0.2544*** -0.1189** 0.1458*** 0.1137
T-Statistics 1.7584 1.7960 3.3563 -2.2159 3.4379
STDPRCVOL 0.0224** 0.2393** 0.2816*** -0.0626 0.1294*** 0.1073
T-Statistics 2.0564 2.3887 3.6090 -1.1324 2.9644
MEANABS 0.0168 0.2151** 0.3250*** -0.0602 0.1533*** 0.1315
T-Statistics 1.5685 2.1800 4.2274 -1.1067 3.5657

Notes. This table reports the results of regression analysis for nine dominant portfolios. *, **, ***
represent significance at 10%, 5%, 1% level, respectively.

In summary, apart from testing the explanatory power of coin market three
factor model, we also evaluate adjusted three-factor model by incorporating anomalies
of COMO1, CVOL and CSTD independently. We find that adjusted three-factor model
incorporating CSTD could enhance the performance of coin market three-factor model,
especially for portfolios based on volume factors and volatility factors. Such evidence
illustrates that volume factor might drive the cryptocurrencies trading and development.
The other adjusted models fail to improve the coin market three-factor model as no

52
considerable improvement is observed. Moreover, the portfolio of MOM1 cannot be
either explained by coin-market three-factor model or adjusted three-factor model,
which needs further research.

7 Conclusions

This paper employs non-parametric almost stochastic dominance to identify

portfolios of cryptocurrencies based on 27 different cryptocurrency factors to
investigate whether they can generate superior returns benchmarked against US equities,
US Treasury bills, US Treasury bonds and a cryptocurrency index over a relatively long
horizon (52 and 78 weeks). We find nine dominant factor portfolios, in the sense of
almost first degree stochastic dominance (AFSD) and almost second degree stochastic
dominance (ASSD, which are based on momentum (1 to 3 weeks), risk momentum (1
to 3 weeks), daily turnover, the standard deviation of daily turnover, and the mean
absolute return divided by daily turnover. Benchmarking these nine dominant factor
portfolios against equity portfolios based on size, momentum and book-to-market, we
find that the long-only strategy contributes more to performance than the short-only
strategy. However, we suggest that the performance of long-short portfolios is generally
preferable to that of long-only factor portfolios. We then test whether a coin market
three-factor model can capture the variation of cross-sectional returns, and find that this
model has limited success in explaining returns of the nine dominant factor portfolios.
For five factor portfolios their alphas are statistically insignificant, and adjusted R2
values reasonably high, so we conclude that the dominance of these five cryptocurrency
factors can be explained by a risk premium. To investigate further, we added four
mispricing factors (subdivision of momentum, volume and volatility) and two
cryptocurrency fundamental factors (electricity and computing power) to our three
factor model. Even with the addition of these mispricing factors, the alphas of the other
four factor portfolios remained significant, and their adjusted R2 values stayed low,
indicating that their dominance may be due to mispricing, rather than a risk premium.

Our findings should improve investors’ understanding of cryptocurrencies, and

support a factor-based approach to analysing the cryptocurrency market. However,
although we use 400 cryptocurrencies representing over 80% of cryptocurrency market
capitalization, our empirical results might be changed by the inclusion of the remaining

53
20% of market capitalization. In addition, due to the high volatility of cryptocurrencies,
different downside risk metrics might provide conflicting outcomes. Hence, further
study is needed.

54
 References

Abadi, J. and Brunnermeier, M., 2018. Blockchain Economics. (0898-2937).

Ali, A., Hwang, L.-S. and Trombley, M.A., 2003. Arbitrage risk and the book-to-market
anomaly. Journal of Financial Economics, 69(2), pp. 355-373.

Bai, J., Bali, T.G. and Wen, Q., 2019. Common risk factors in the cross-section of corporate
bond returns. Journal of Financial Economics, 131(3), pp. 619-642.

Bali, T.G., Brown, S.J. and Demirtas, K.O., 2013. Do hedge funds outperform stocks and bonds?
Management Science, 59(8), pp. 1887-1903.

Bali, T.G., Demirtas, K.O., Levy, H. and Wolf, A., 2009. Bonds versus stocks: Investors’ age
and risk taking. Journal of Monetary Economics, 56(6), pp. 817-830.

Barroso, P. and Santa-Clara, P., 2015. Momentum has its moments. Journal of Financial
Economics, 116(1), pp. 111-120.

Biais, B., Bisiere, C., Bouvard, M. and Casamatta, C., 2019. The blockchain folk theorem.
Review of Financial Studies, 32(5), pp. 1662-1715.

Bianchi, D., Guidolin, M. and Pedio, M., 2020. Dissecting Time-Varying Risk Exposures in
Cryptocurrency Markets. BAFFI CAREFIN Centre Research Paper, (2020-143).

Blitz, D., Baltussen, G. and van Vliet, P., 2019. When Equity Factors Drop Their Shorts.
Available at SSRN 3493305.

Board, J.L. and Sutcliffe, C.M., 1994. Estimation methods in portfolio selection and the
effectiveness of short sales restrictions: UK evidence. Management Science, 40(4), pp. 516-
534.

Borri, N., 2019. Conditional tail-risk in cryptocurrency markets. Journal of Empirical Finance,
50, pp. 1-19.

Carhart, M.M., 1997. On persistence in mutual fund performance. Journal of Finance, 52(1),
pp. 57-82.

Chu, Y., Hirshleifer, D. and Ma, L., 2020. The Causal Effect of Limits to Arbitrage on Asset
Pricing Anomalies. Journal of Finance, 75(5), pp. 2631-2672.

Ciaian, P., Rajcaniova, M. and Kancs, d.A., 2016. The economics of BitCoin price formation.
Applied Economics, 48(19), pp. 1799-1815.

Daniel, K. and Moskowitz, T.J., 2016. Momentum crashes. Journal of Financial Economics,
122(2), pp. 221-247.

Daskalaki, C., Skiadopoulos, G. and Topaloglou, N., 2017. Diversification benefits of

commodities: A stochastic dominance efficiency approach. Journal of Empirical Finance, 44,
pp. 250-269.

55
Detzel, A., Liu, H., Strauss, J., Zhou, G. and Zhu, Y., 2020. Learning and predictability via
technical analysis: Evidence from Bitcoin and stocks with hard‐to‐value fundamentals.
Financial Management.

Easley, D., O'Hara, M. and Basu, S., 2019. From mining to markets: The evolution of bitcoin
transaction fees. Journal of Financial Economics, 134(1), pp. 91-109.

Edelen, R.M., Ince, O.S. and Kadlec, G.B., 2016. Institutional investors and stock return
anomalies. Journal of Financial Economics, 119(3), pp. 472-488.

Fama, E.F. and French, K.R., 1992. The cross‐section of expected stock returns. Journal of
Finance, 47(2), pp. 427-465.

Fama, E.F. and French, K.R., 1993. Common risk factors in the returns on stocks and bonds.
Journal of Financial Economics, 33(1), pp. 3-56.

Fama, E.F. and French, K.R., 1995. Size and Book-to-Market Factors in Earnings and Returns.
Journal of Finance, 50(1), pp. 131-155.

Farinelli, S., Ferreira, M., Rossello, D., Thoeny, M. and Tibiletti, L., 2008. Beyond Sharpe
ratio: Optimal asset allocation using different performance ratios. Journal of Banking &
Finance, 32(10), pp. 2057-2063.

Feng, G., Giglio, S. and Xiu, D., 2020. Taming the factor zoo: A test of new factors. Journal
of Finance, 75(3), pp. 1327-1370.

Foley, S., Karlsen, J.R. and Putniņš, T.J., 2019. Sex, Drugs, and Bitcoin: How Much Illegal
Activity Is Financed through Cryptocurrencies? Review of Financial Studies, 32(5), pp. 1798-
1853.

Frazzini, A. and Pedersen, L.H., 2014. Betting against beta. Journal of Financial Economics,
111(1), pp. 1-25.

Gan, J., Tsoukalas, G. and Netessine, S., 2020. Initial Coin Offerings, Speculation, and Asset
Tokenization. Management Science, 0(0), p. null.

Harvey, C.R., Liu, Y. and Zhu, H., 2015. … and the Cross-Section of Expected Returns. Review
of Financial Studies, 29(1), pp. 5-68.

Israel, R. and Moskowitz, T.J., 2013. The role of shorting, firm size, and time on market
anomalies. Journal of Financial Economics, 108(2), pp. 275-301.

Jegadeesh, N. and Titman, S., 1993. Returns to buying winners and selling losers: Implications
for stock market efficiency. Journal of Finance, 48(1), pp. 65-91.

Keloharju, M., Linnainmaa, J.T. and Nyberg, P., 2020. Are return seasonalities due to risk or
mispricing? Journal of Financial Economics.

La Porta, R., Lakonishok, J., Shleifer, A. and Vishny, R., 1997. Good News for Value Stocks:
Further Evidence on Market Efficiency. Journal of Finance, 52(2), pp. 859-874.

56
Lakonishok, J., Shleifer, A. and Vishny, R.W., 1994. Contrarian Investment, Extrapolation,
and Risk. Journal of Finance, 49(5), pp. 1541-1578.

Leshno, M. and Levy, H., 2002. Preferred by “all” and preferred by “most” decision makers:
Almost stochastic dominance. Management Science, 48(8), pp. 1074-1085.

Lettau, M. and Wachter, J.A., 2007. Why is long‐horizon equity less risky? A duration‐based
explanation of the value premium. Journal of Finance, 62(1), pp. 55-92.

Levy, H., Leshno, M. and Leibovitch, B., 2010. Economically relevant preferences for all
observed epsilon. Annals of Operations Research, 176(1), pp. 153-178.

Levy, H. and Levy, M., 2019. Stocks versus Bonds and the Investment Horizon. Available at
SSRN 3458828.

Li, J., Li, N., Peng, J., Cui, H. and Wu, Z., 2019. Energy consumption of cryptocurrency mining:
A study of electricity consumption in mining cryptocurrencies. Energy, 168, pp. 160-168.

Liu, Y., Tsyvinski, A. and Goldstein, I., 2020. Risks and Returns of Cryptocurrency. Review
of Financial Studies.

Liu, Y., Tsyvinski, A. and Wu, X., 2019. Common risk factors in cryptocurrency. (0898-2937).

Makarov, I. and Schoar, A., 2020. Trading and arbitrage in cryptocurrency markets. Journal of
Financial Economics, 135(2), pp. 293-319.

Neely, C.J., Rapach, D.E., Tu, J. and Zhou, G., 2014. Forecasting the equity risk premium: the
role of technical indicators. Management Science, 60(7), pp. 1772-1791.

Penman, S.H., 1996. The Articulation of Price-Earnings Ratios and Market-to-Book Ratios and
the Evaluation of Growth. Journal of Accounting Research, 34(2), pp. 235-259.

Post, T., 2003. Empirical tests for stochastic dominance efficiency. Journal of Finance, 58(5),
pp. 1905-1931.

Rapach, D.E., Strauss, J.K. and Zhou, G., 2009. Out-of-Sample Equity Premium Prediction:
Combination Forecasts and Links to the Real Economy. Review of Financial Studies, 23(2),
pp. 821-862.

Sharpe, W.F., 1966. Mutual fund performance. Journal of Business, 39(1), pp. 119-138.

Stambaugh, R.F., Yu, J. and Yuan, Y., 2012. The short of it: Investor sentiment and anomalies.
Journal of Financial Economics, 104(2), pp. 288-302.

Stambaugh, R.F. and Yuan, Y., 2017. Mispricing factors. Review of Financial Studies, 30(4),
pp. 1270-1315.

Wang, B., 2019. The cash conversion cycle spread. Journal of Financial Economics, 133(2),
pp. 472-497.

Zhang, L., 2005. The value premium. Journal of Finance, 60(1), pp. 67-103.

57
 Supplementary Appendix
A1. Simulation Test for ASD Critical Values

We discussed the critical values of 𝜀 and 𝜀 in Section 4.3, which are the maximum
ratio of violation area to enclosed area of portfolio with high return (H) and portfolio with low
return (L). According to Levy et al. (2010), they exploit experiments based on 400 subjects’
choices to clarify the economically relevant set of preferences and quantify the critical epsilon
values of ASD (𝜀 ∗ = 5.9%, 𝜀 ∗ = 3.2%) that avoid the paradoxical choices. Moreover, the
outcomes of Levy et al. (2010) are robust and not sensitive to both magnitude of asset returns
and different asset classes. Therefore, we suggest that the critical values of 𝜀 and 𝜀 employed
in our empirical analysis are robust criteria to reflect almost all investors’ choices when
investors are facing the similar scenarios. Nevertheless, one may be concerned about the
reliability of critical values of Levy et al. (2010) because they completed the test ten years ago.
To alleviate such potential concerns, we conduct analysis of critical values of 𝜀 and 𝜀 based
on randomization techniques (Bali et al., 2013) to ensure the reliability of critical values for
our study, and we focus on the p-values associated with critical values used in our paper.

We select monthly S&P 500 index return as a proxy for stock return over the time
period from January 1926 to December 2019 and generate the distribution of 𝜀 and 𝜀 with
repeated samples. Particularly, in each time of simulation, two series of 1000 monthly return
observations are picked (with replacement) from the S&P 500 index, we further calculate and
record the empirical 𝜀 and 𝜀 values. This procedure is repeated 3000 times, which generate
3000 pairs of 𝜀 and 𝜀 values. The null hypothesis is that two return series do not dominate
each other in the sense of AFSD or ASSD. Mathematically, for AFSD, the null hypothesis is
𝐻 ∶ 𝜀 ≥ 5.9%; for ASSD, the null hypothesis is 𝐻 ∶ 𝜀 ≥ 3.2%.

Table A.1 reports the descriptive statistics for distribution of 𝜀 and 𝜀 based on
randomization techniques. Specifically, the minimum, maximum, mean and 1, 3, 5, 10, 25, 50,
75, 90, 95, 99 percentiles of generated 𝜀 and 𝜀 values. We examine the p-values of each
distribution to clarify whether the null hypothesis can be rejected. For 𝜀 , the third percentile
of distribution of 𝜀 is 6.99%, which is greater than 5.9%, indicating that estimated 𝜀 values
reported in Table A.1 have p-values lower than 3%. This result firmly rejected the null
hypothesis of 𝜀 at significance level of 3%. Similarly, from the perspective of 𝜀 , the 10th
percentile of distribution of 𝜀 is 5.89%, we can reject the null hypothesis of 𝜀 at significance
level of 10%.

58
 Table A.1: Distribution of 𝜀 and 𝜀 from Randomization
Min 1% 3% 5% 10% 25% 50% 75% 90% 95% 99% Max Mean
𝜀 0.0017 0.0326 0.0699 0.1047 0.1709 0.3791 0.6498 0.8513 0.9619 0.9880 1.0000 1.0000 0.6059
𝜀 0.0000 0.0010 0.0108 0.0256 0.0589 0.1813 0.3965 0.6650 0.8610 0.9374 0.9945 1.0000 0.4302
Notes. This table reports the statistics for distribution of 𝜀 and 𝜀 based on randomization techniques.

To sum up, we secure the results of Levy et al. (2010) by conducting ASD on randomly
generated data, which provide robustness about our empirical results of cryptocurrency factors.

A2. Incorporating Mispricing and Fundamental Factors into Coin Market Model

In this section, we aim to construct every potential combination of mispricing factors

and fundamental factors (electricity and computer power), we then incorporate these potential
factors into coin market three factor model developed by Liu et al. (2019) to evaluate whether
the performance of adjust model can be improved (Stambaugh and Yuan, 2017). We consider
the two most important fundamental factors for cryptocurrencies: electricity and computer
power, since these two factors are essential requirements to mine cryptocurrencies (Bianchi et
al., 2020; Liu et al., 2020). We select two proxies for electricity, they are value-weighted stock
return of US-listed electricity firms and value-weighted stock return of China-listed electricity
firms, respectively. According to Li et al. (2019), We include China-listed electricity firms as
China is assumed to have the vastest mining operations and electricity supply is location
specific. Moreover, we choose one with the highest average correlation coefficients of each
dominant factor. From the perspective of computer power, we consider the stock returns of
primary GPU manufacturers such as Nvidia and TSMC as our two proxies for computer power
(Liu et al., 2020). Likewise, we still choose one with the highest average correlation
coefficients of each dominant factor as our proxy for computer power.

A2.1. Correlation between Fundamental Factors and Dominant Factors

The data for US-listed electricity firms, China-listed electricity firms, Nvidia and
TSMC are collected from DataStream over the period from 2014 to 2019. We first examine the
correlation between fundamental factors and nine dominant factors, choose one from each
fundamental group with higher correlation to continue the combinations of potential models.
The correlation coefficients between each fundamental factor and each dominant factor is
reported in Table A.2.

Panel A of Table A.2 illustrates the correlation between US-listed electricity firms and
nine dominant factors, MOM1 has the highest coefficient of 0.1520, whereas RMOM3 is

59
almost not related to US-listed electricity as indicated by coefficient of 0.003, and the average
correlation coefficient is 0.0580. Panel B of Table A.2 demonstrates the correlation between
China-listed electricity firms and dominant factors. We can observe that none of dominant
factors are strongly related to China-listed electricity firms and the average correlation
coefficient is 0.0237. Therefore, we choose value-weighted returns of US-listed firms as our
proxy for electricity.

Table A.2: Correlation between US, China Electricity Index and Nine Factors
Panel A: Correlation between 9 Factors and US Electricity Index
Factor MOM1 MOM2 MOM3 RMOM1 RMOM2 RMOM3 VOLPRC STDPRCVOL MEANABS AVG
Corr 0.1520 -0.0614 -0.0319 -0.0792 -0.0533 0.0030 -0.0727 -0.0290 -0.0392 0.0580
Panel B: Correlation between 9 Factors and China Electricity Index
Factor MOM1 MOM2 MOM3 RMOM1 RMOM2 RMOM3 VOLPRC STDPRCVOL MEANABS AVG

Corr -0.0192 -0.0191 0.0325 0.0328 0.0078 0.0504 -0.0115 0.0270 -0.0129 0.0237
Notes. This table reports the correlation between US, China electricity and dominant factors.

We evaluate the suitability of two computer power in the same manner of electricity.
Table A.3 reports the correlation between Nvidia, TSMC and nine dominant factors.

Table A.3: Correlation between AMD, TSMC and Nine Factors

Panel A: Correlation between 9 Factors and AMD
Factor MOM1 MOM2 MOM3 RMOM1 RMOM2 RMOM3 VOLPRC STDPRCVOL MEANABS AVG

Corr 0.0102 -0.0350 0.0314 0.0161 -0.0254 -0.0237 0.0681 0.0860 0.0485 0.0383
Panel B: Correlation between 9 Factors and TSMC
Factor MOM1 MOM2 MOM3 RMOM1 RMOM2 RMOM3 VOLPRC STDPRCVOL MEANABS AVG

Corr -0.0712 0.0123 0.0600 0.0660 0.0365 0.0447 -0.0163 0.0441 -0.0178 0.0410
Notes. This table reports the correlation between AMD, TSMC and dominant factors.

Panel A of Table A.3 shows the correlation between nine factors and AMD, none of
dominant factors heavily correlate to stock return of AMD, and the average correlation
coefficient is 0.0383. Moreover, Panel B of Table A.3 indicates the similar results in Panel A
but with a slightly higher average correlation coefficient of 0.041. Although the results are
quite lower than what we expected, it is still consistent with Liu and Tsyvinski’s finding (2020).
Therefore, we choose TSMC as our proxy for computer power.

A2.2. Potential Combinations for Adjusted Models

After we determine to use stock returns of US-listed electricity firms as a proxy for
electricity and consider stock return of TSMC as a proxy for computer power (Graphics in
table), we list all adjusted models incorporating combinations of fundamental factors and
mispricing factors, which are 28 models in total. Table A.4 lists all combination of these factors,

60
where volatility category combines two volatility factors into one factor by using average
technique. The first column represents the specific model number, and the first row illustrates
coin market three-factor model, mispricing factors and selected fundamental factors. The value
‘Y’ means that corresponding model include this factor to form an adjusted model.

Table A.4: Models of Potential Combinations

𝐶𝑜𝑖𝑛 3 − 𝐹𝑎𝑐𝑡𝑜𝑟 𝑀𝑜𝑑𝑒𝑙 𝐶𝑀𝑂𝑀1 𝐶𝑉𝑂𝐿 𝐶𝑆𝑇𝐷 𝐸𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦 𝐺𝑟𝑎𝑝ℎ𝑖𝑐𝑠
1 Y Y
2 Y Y
3 Y Y
4 Y Y Y
5 Y Y Y
6 Y Y Y
7 Y Y Y Y
8 Y Y Y
9 Y Y Y
10 Y Y Y
11 Y Y Y Y
12 Y Y Y Y
13 Y Y Y Y
14 Y Y Y Y Y
15 Y Y Y
16 Y Y Y
17 Y Y Y
18 Y Y Y Y
19 Y Y Y Y
20 Y Y Y Y
21 Y Y Y Y Y
22 Y Y Y Y
23 Y Y Y Y
24 Y Y Y Y
25 Y Y Y Y Y
26 Y Y Y Y Y
27 Y Y Y Y Y
28 Y Y Y Y Y Y
Notes. This table reports the 28 adjusted model based on mispricing factors and determined fundamental factors.

To get a better understanding of Table A.4, we explain the terms in this paragraph. Coin three-
factor model represents a model consists of coin market (𝐶𝑀𝐾𝑇), size factor (𝐶𝑆𝑀𝐵) and
momentum factor (𝐶𝑀𝑂𝑀), which is developed by Liu et al. (2019) and discussed in Section
6.4. Moreover, one-week momentum (𝐶𝑀𝑂𝑀1), volume factor (𝐶𝑉𝑂𝐿) and volatility factor
(𝐶𝑆𝑇𝐷) are mispricing factors with significant alphas and relative low R-squared that cannot
be explained by origin three factor model. We choose value-weighted returns of US-listed
electricity firms as a proxy for electricity (𝐸𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦 in Table A.4) and stock return of TSMC
as a proxy for computer power (𝐺𝑟𝑎𝑝ℎ𝑖𝑐𝑠 in Table A.4).

61
 The construction of each mispricing factor follows the Fama and French (1993) method.
To illustrate, For 𝐶𝑀𝑂𝑀1 factor, we sort the coins into three one-week momentum groups:
bottom 30 percent, middle 40 percent and top 30 percent, then we form value-weighted
portfolios for each of the three three-week momentum groups. The mispricing cryptocurrency
momentum factor (𝐶𝑀𝑂𝑀1) is the difference between return on top 30 percent portfolios and
return on bottom 30 percent portfolios. Similarly, to construct 𝐶𝑉𝑂𝐿, we sort the coins into
three volume groups (We refer this to volume factor): bottom 30 percent, middle 40 percent
and top 30 percent, then we form value-weighted portfolios for each of the three volume groups.
The mispricing cryptocurrency volume factor (𝐶𝑉𝑂𝐿) is the difference between return on top
30 percent portfolios and return on bottom 30 percent portfolios. Moreover, to construct a
composite mispricing factor for volatility (𝐶𝑆𝑇𝐷 ), we apply equally weighted average to
rankings with respect to two anomalies, then form the mispricing volatility factors based on
averaged rankings. We sort the coins into three volatility groups, bottom 30 percent, middle 40
percent and top 30 percent, then we form value-weighted portfolios for each of the three volume
groups. The mispricing cryptocurrency volatility factor ( 𝐶𝑆𝑇𝐷) is the difference between
return on top 30 percent portfolios and return on bottom 30 percent portfolios.

Among 28 models, each model is formed by selected column(s) at each row, and the
symbol ‘Y’ indicates that corresponding factor (column) is included in a model. For example,
Model 1 is comprised of two parts: coin market three-factor model and one mispricing factor
𝐶𝑀𝑂𝑀1, mathematical expression tends to provide a clearer picture:

𝑅 −𝑅 =𝛼 +𝛽 𝐶𝑀𝐾𝑇 + 𝛽 𝐶𝑆𝑀𝐵 + 𝛽 𝐶𝑀𝑂𝑀 + 𝛽 𝐶𝑀𝑂𝑀1 + 𝜖

where 𝑅 , 𝑅 , 𝐶𝑀𝐾𝑇, 𝐶𝑆𝑀𝐵, 𝐶𝑀𝑂𝑀 have been illustrated above, 𝐶𝑀𝑂𝑀1 is the mispricing
momentum factor.

The rest models are constructed in the same manner.

A2.3. Empirical Results for 28 Models

In this section, we examine whether the returns of nine dominant factors can be
explained better by 28 adjusted models, and we choose the best fit one as our improved model.
The first row contains coefficients of regression analysis such as alpha, beta for each factor and
adjusted R-squared. Moreover, the first column represents each one of nine dominant factors,
and its t-statistics which is included in parenthesis.

62
Table A.5: Model 1

𝛼 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0385*** 0.2632*** 0.0429 -0.1184 -0.0423 0.0262

T-Statistics (2.9007) (2.1151) (0.4165) (-1.3009) (-0.5001)
MOM2 0.0007 0.1736* 0.1488* 0.5586*** 0.3537*** 0.4164
T-Statistics (0.0684) (1.7657) (1.8291) (7.7685) (5.2927)
MOM3 0.0003 0.1435* -0.2482*** 0.8451*** 0.1542*** 0.5644
T-Statistics (0.0384) (1.6808) (-3.5141) (13.5377) (2.657)
RMOM1 0.0095 0.1103 0.1226** 0.1481*** 0.6380*** 0.5069
T-Statistics (1.1868) (1.4625) (1.9637) (2.6846) (12.4442)
RMOM2 0.0094 0.1797** 0.1759*** 0.4909*** 0.1799*** 0.3727
T-Statistics (1.0818) (2.209) (2.6124) (8.2503) (3.254)
RMOM3 -0.0078 0.2228*** -0.1765*** 0.8049*** 0.0503 0.5279
T-Statistics (-0.8979) (2.7353) (-2.6183) (13.5117) (0.9078)
VOLPRC 0.0333*** 0.2898*** 0.2839*** -0.0878 -0.0476 0.0812
T-Statistics (3.3718) (3.1252) (3.7009) (-1.2955) (-0.7552)
STDPRCVOL 0.0356*** 0.3393*** 0.3083*** -0.0211 -0.064 0.0846
T-Statistics (3.5198) (3.5739) (3.925) (-0.3038) (-0.9925)
MEANABS 0.0321*** 0.3396*** 0.3553*** -0.0507 -0.0136 0.0955
T-Statistics (3.1899) (3.6034) (4.5551) (-0.7362) (-0.2128)
Average 0.2973
Notes. This table reports the empirical results of Model 1. *, **, *** represent significance at 10%, 5%, 1%
level, respectively. The average adjusted R-squared of Model 1 is 0.2973. Four mispricing factors still cannot
be well explained due to their statistically significant alpha values.

63
Table A.6: Model 2

𝛼 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0380*** 0.2388** -0.0079 -0.1325** -0.1683 0.032

T-Statistics (2.8767) (1.9046) (-0.0732) (-1.7993) (-1.4404)
MOM2 0.0032 0.1186 0.1199 0.7934*** -0.1242 0.3664
T-Statistics (0.2976) (1.1455) (1.3397) (13.0415) (-1.2866)
MOM3 0.0016 0.1394 -0.2259*** 0.9384*** 0.0637 0.5553
T-Statistics (0.17) (1.5997) (-3) (18.3244) (0.784)
RMOM1 0.0144 0.058 0.1525** 0.5504*** 0.0533 0.2583
T-Statistics (1.4582) (0.6204) (1.8874) (10.0197) (0.6119)
RMOM2 0.0107 0.1614* 0.1780** 0.6059*** -0.0061 0.351
T-Statistics (1.2169) (1.9299) (2.4639) (12.3325) (-0.0787)
RMOM3 -0.0073 0.2425*** -0.1325 0.8259*** 0.1446** 0.5321
T-Statistics (-0.8438) (2.9587) (-1.8715) (17.1536) (1.8936)
VOLPRC 0.0319*** 0.1005 -0.0566 -0.0312 -1.1455*** 0.5553
T-Statistics (4.8743) (1.6179) (-1.0551) (-0.8543) (-19.7913)
STDPRCVOL 0.0340*** 0.1457** -0.043 0.0277 -1.1805*** 0.5963
T-Statistics (5.1143) (2.3064) (-0.787) (0.7477) (-20.0543)
MEANABS 0.0308*** 0.1350** -0.0057 0.033 -1.2169*** 0.6032
T-Statistics (4.9416) (2.2817) (-0.1114) (0.9483) (-22.0672)
Average 0.4278
Notes. This table reports the empirical results of Model 2. *, **, *** represent significance at 10%, 5%, 1%
level, respectively. The average adjusted R-squared of Model 2 is 0.4278. Four mispricing factors still cannot
be well explained due to their statistically significant alpha values.

64
Table A.7: Model 3

𝛼 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0463*** 0.3351*** 0.0581 -0.1449** -0.0827 0.0319

T-Statistics (3.2194) (2.5278) (0.5628) (-1.981) (-1.4314)
MOM2 -0.0144 -0.008 0.1215 0.7830*** 0.1799*** 0.3924
T-Statistics (-1.238) (-0.0746) (1.4545) (13.2396) (3.8513)
MOM3 -0.0162* -0.0181 -0.2798*** 0.9423*** 0.1789*** 0.5832
T-Statistics (-1.6694) (-0.2031) (-4.0258) (19.146) (4.6016)
RMOM1 0.0164 0.0663 0.1408* 0.5545 -0.0211 0.258
T-Statistics (1.528) (0.6705) (1.828) (10.1672) (-0.4898)
RMOM2 0.0031 0.0991 0.1648** 0.6051*** 0.0772** 0.3595
T-Statistics (0.3257) (1.1282) (2.4079) (12.4867) (2.0161)
RMOM3 -0.0092 0.2033** -0.1789*** 0.8368*** 0.018 0.527
T-Statistics (-0.9723) (2.3329) (-2.6358) (17.4116) (0.4731)
VOLPRC 0.0186* 0.1747* 0.2544*** -0.1189** 0.1458*** 0.1137
T-Statistics (1.7584) (1.796) (3.3563) (-2.2159) (3.4379)
STDPRCVOL 0.0224** 0.2393** 0.2816*** -0.0626 0.1294*** 0.1073
T-Statistics (2.0564) (2.3887) (3.609) (-1.1324) (2.9644)
MEANABS 0.0168 0.2151** 0.3250*** -0.0602 0.1533*** 0.1315
T-Statistics (1.5685) (2.18) (4.2274) (-1.1067) (3.5657)
Average 0.2783
Notes. This table reports the empirical results of Model 3. *, **, *** represent significance at 10%, 5%, 1%
level, respectively. The average adjusted R-squared of Model 3 is 0.2783. Three mispricing factors still cannot
be well explained due to their statistically significant alpha values. However, MOM3 becomes statistically
significant.

65
Table A.8: Model 4

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0383*** 0.2352* -0.0067 -0.1073 -0.0398 -0.1672 0.0168

T-Statistics (2.8923) (1.8697) (-0.0619) (-1.1768) (-0.471) (-1.4284)
MOM2 0.0006 0.1511 0.1089 0.5676*** 0.3558*** -0.1345 0.4109
T-Statistics (0.0551) (1.5202) (1.2708) (7.8782) (5.3316) (-1.4542)
MOM3 0.0004 0.1534* -0.2307*** 0.8412*** 0.1533*** 0.0592 0.5581
T-Statistics (0.0452) (1.7739) (-3.0912) (13.4158) (2.6391) (0.7362)
RMOM1 0.0096 0.1162 0.1329** 0.1458*** 0.6375*** 0.0348 0.4993
T-Statistics (1.1898) (1.5193) (2.0149) (2.6298) (12.4158) (0.4898)
RMOM2 0.0094 0.1778** 0.1725** 0.4916*** 0.1801*** -0.0114 0.3625
T-Statistics (1.0787) (2.1553) (2.4238) (8.2199) (3.2512) (-0.1479)
RMOM3 -0.0077 0.2469*** -0.134* 0.7954*** 0.0481 0.1432* 0.5256
T-Statistics (-0.8842) (3.0054) (-1.891) (13.3583) (0.8721) (1.8743)
VOLPRC 0.0321*** 0.0977 -0.0557 -0.0119 -0.0303 -1.1446*** 0.5904
T-Statistics (4.8996) (1.5692) (-1.0367) (-0.2643) (-0.7251) (-19.7564)
STDPRCVOL 0.0344*** 0.1415** -0.0415 0.0571 -0.0462 -1.1791*** 0.5983
T-Statistics (5.1621) (2.236) (-0.761) (1.2451) (-1.0888) (-20.0332)
MEANABS 0.0308*** 0.1354** -0.0058 0.0300 0.0047 -1.217*** 0.6452
T-Statistics (4.9223) (2.281) (-0.114) (0.6962) (0.1186) (-22.0293)
Average 0.4674
Notes. This table reports the empirical results of Model 4. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model 4
is 0.4674. Four mispricing factors still cannot be well explained due to their statistically significant alpha values.

66
Table A.9: Model 5

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0463*** 0.3306** 0.0582 -0.1284 -0.0259 -0.0802 0.0163

T-Statistics (3.2108) (2.4743) (0.5628) (-1.4084) (-0.3037) (-1.3727)
MOM2 -0.0138 0.0485 0.1204 0.5772*** 0.3233*** 0.1489*** 0.4269
T-Statistics (-1.2255) (0.4662) (1.4941) (8.1271) (4.8648) (3.2707)
MOM3 -0.0159* 0.0029 -0.2802*** 0.866*** 0.1199** 0.1674*** 0.5825
T-Statistics (-1.6537) (0.0319) (-4.054) (14.2152) (2.1041) (4.287)
RMOM1 0.0177** 0.1809** 0.1386** 0.1376** 0.6552*** -0.084** 0.5080
T-Statistics (2.0392) (2.247) (2.2246) (2.5056) (12.7496) (-2.3851)
RMOM2 0.0034 0.1284 0.1642** 0.4985*** 0.1674*** 0.0611 0.3677
T-Statistics (0.3655) (1.4714) (2.431) (8.373) (3.0059) (1.6013)
RMOM3 -0.0091 0.2116** -0.1791*** 0.8066*** 0.0475 0.0134 0.5204
T-Statistics (-0.9618) (2.412) (-2.637) (13.477) (0.8486) (0.3494)
VOLPRC 0.0184* 0.1609 0.2546*** -0.0687 -0.0789 0.1533*** 0.1039
T-Statistics (1.7451) (1.6453) (3.363) (-1.0293) (-1.2641) (3.5844)
STDPRCVOL 0.0222** 0.2231** 0.2819*** -0.0038 -0.0923 0.1382*** 0.0988
T-Statistics (2.0429) (2.2175) (3.6192) (-0.0558) (-1.4369) (3.1415)
MEANABS 0.0167 0.2071** 0.3251*** -0.0310 -0.0459 0.1577*** 0.1187
T-Statistics (1.5587) (2.0839) (4.2261) (-0.4579) (-0.7232) (3.6291)
Average 0.3048
Notes. This table reports the empirical results of Model 5. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model 5
is 0.3048. Three mispricing factors still cannot be well explained due to their statistically significant alpha values. Furthermore, alpha values of MOM3 and RMOM1 become
statistically significant.

67
Table A.10: Model 6

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0478*** 0.3147** 0.0015 -0.1294* -0.2021* -0.0995* 0.0254

T-Statistics (3.3259) (2.3714) (0.0139) (-1.7621) (-1.7102) (-1.7026)
MOM2 -0.0139 -0.0146 0.1033 0.7879*** -0.0649 0.1745*** 0.3834
T-Statistics (-1.1943) (-0.1352) (1.1765) (13.2107) (-0.6763) (3.6798)
MOM3 -0.0171* -0.0052 -0.2439*** 0.9325*** 0.1280 0.1895*** 0.5800
T-Statistics (-1.7668) (-0.0579) (-3.3499) (18.8526) (1.6093) (4.8182)
RMOM1 0.0161 0.0711 0.1541* 0.5509*** 0.0475 -0.0172 0.2465
T-Statistics (1.4915) (0.7153) (1.9025) (10.0125) (0.5365) (-0.3922)
RMOM2 0.0030 0.1012 0.1706** 0.6035*** 0.0207 0.0789** 0.3491
T-Statistics (0.309) (1.1456) (2.3692) (12.341) (0.2626) (2.0286)
RMOM3 -0.0103 0.219** -0.1354* 0.825*** 0.1551** 0.0308 0.5255
T-Statistics (-1.0948) (2.5149) (-1.9092) (17.1185) (2.0008) (0.8046)
VOLPRC 0.0268*** 0.0607 -0.0616 -0.0328 -1.1278*** 0.0521* 0.5940
T-Statistics (3.7617) (0.9238) (-1.15) (-0.9019) (-19.2793) (1.8032)
STDPRCVOL 0.0309*** 0.121* -0.0460 0.0267 -1.1695*** 0.0323 0.5983
T-Statistics (4.25) (1.8051) (-0.8424) (0.7206) (-19.5922) (1.0954)
MEANABS 0.0255*** 0.0939 -0.0108 0.0313 -1.1986*** 0.0539* 0.6496
T-Statistics (3.7667) (1.502) (-0.212) (0.9039) (-21.5259) (1.9562)
Average 0.4391
Notes. This table reports the empirical results of Model 6. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model 6
is 0.4391. Four mispricing factors still cannot be well explained due to their statistically significant alpha values. Furthermore, alpha value of MOM3 becomes statistically
significant.

68
Table A.11: Model 7

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0478*** 0.3114** 0.0019 -0.1172 -0.0193 -0.2009* -0.0975 0.0223

T-Statistics (3.317) (2.3297) (0.0176) (-1.2865) (-0.2271) (-1.6955) (-1.6488)
MOM2 -0.0131 0.0404 0.0964 0.582*** 0.3261*** -0.0855 0.1415*** 0.4266
T-Statistics (-1.1673) (0.3864) (1.1387) (8.1706) (4.9007) (-0.9237) (3.0615)
MOM3 -0.0168* 0.0144 -0.2464*** 0.8593*** 0.116** 0.1207 0.1778*** 0.5843
T-Statistics (-1.7462) (0.1607) (-3.4004) (14.0987) (2.0372) (1.5232) (4.4946)
RMOM1 0.0177** 0.1814** 0.1403** 0.1373** 0.655*** 0.0060 -0.0834** 0.5064
T-Statistics (2.0272) (2.2423) (2.1397) (2.4889) (12.7118) (0.0837) (-2.3311)
RMOM2 0.0034 0.1294 0.167** 0.4979*** 0.1671*** 0.0101 0.0620 0.3657
T-Statistics (0.3564) (1.4747) (2.35) (8.3283) (2.992) (0.1295) (1.5971)
RMOM3 -0.0102 0.2261*** -0.1363* 0.7981*** 0.0425 0.1524** 0.0265 0.5248
T-Statistics (-1.0829) (2.5805) (-1.9202) (13.3627) (0.7624) (1.9626) (0.6844)
VOLPRC 0.0267*** 0.0536 -0.0607 -0.0062 -0.0421 -1.1251*** 0.0564* 0.5940
T-Statistics (3.7468) (0.811) (-1.1333) (-0.1372) (-1.0017) (-19.2138) (1.9299)
STDPRCVOL 0.0307*** 0.1119* -0.0449 0.0610 -0.0542 -1.166*** 0.0378 0.5991
T-Statistics (4.2354) (1.661) (-0.8222) (1.3275) (-1.2631) (-19.5335) (1.2688)
MEANABS 0.0255*** 0.0928 -0.0107 0.0355 -0.0067 -1.1982*** 0.0545* 0.6485
T-Statistics (3.7579) (1.473) (-0.2089) (0.8273) (-0.1672) (-21.4617) (1.9566)
Average 0.4746
Notes. This table reports the empirical results of Model 7. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model 7
is 0.4746. Four mispricing factors still cannot be well explained due to their statistically significant alpha values. Furthermore, alpha values of MOM3 and RMOM1 become
statistically significant.

69
Table A.12: Model 8

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0344*** 0.2863** 0.0405 -0.1241 -0.0233 1.9823*** 0.0340

T-Statistics (2.6044) (2.3191) (0.3976) (-1.3782) (-0.2773) (2.7393)
MOM2 0.0012 0.1712* 0.1491* 0.5593*** 0.3517*** -0.2123 0.4071
T-Statistics (0.1092) (1.7341) (1.8296) (7.764) (5.2368) (-0.3669)
MOM3 0.0001 0.145* -0.2484*** 0.8447*** 0.1554*** 0.1337 0.5574
T-Statistics (0.0081) (1.6925) (-3.5109) (13.5072) (2.6659) (0.266)
RMOM1 0.0098 0.1087 0.1227** 0.1485*** 0.6366*** -0.1425 0.4990
T-Statistics (1.2137) (1.4351) (1.9635) (2.6874) (12.3569) (-0.3211)
RMOM2 0.0097 0.178** 0.1761*** 0.4913*** 0.1785*** -0.1476 0.3626
T-Statistics (1.108) (2.1796) (2.6111) (8.2431) (3.2126) (-0.3083)
RMOM3 -0.0088 0.2287*** -0.1771*** 0.8034*** 0.0551 0.5077 0.5219
T-Statistics (-1.0118) (2.8018) (-2.6277) (13.486) (0.9925) (1.0608)
VOLPRC 0.0349*** 0.2809*** 0.2848*** -0.0856 -0.0549 -0.7590 0.0721
T-Statistics (3.5126) (3.0277) (3.7183) (-1.2646) (-0.8691) (-1.3952)
STDPRCVOL 0.0362*** 0.3358*** 0.3087*** -0.0202 -0.0669 -0.2975 0.0705
T-Statistics (3.5533) (3.5251) (3.9248) (-0.2909) (-1.032) (-0.5326)
MEANABS 0.0328*** 0.3353*** 0.3557*** -0.0497 -0.0172 -0.3739 0.0820
T-Statistics (3.2427) (3.5459) (4.5566) (-0.7196) (-0.2677) (-0.6744)
Average 0.2896
Notes. This table reports the empirical results of Model 8. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model 8
is 0.2896. Four mispricing factors still cannot be well explained due to their statistically significant alpha values. Only MOM1 has exposure to electricity.

70
Table A.13: Model 9

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.034*** 0.2599** -0.0104 -0.126* -0.1701 2.0046*** 0.0406

T-Statistics (2.5887) (2.0922) (-0.0973) (-1.7282) (-1.4715) (2.7891)
MOM2 0.0042 0.1138 0.1204 0.7919*** -0.1238 -0.4586 0.3572
T-Statistics (0.3785) (1.096) (1.3451) (13.0012) (-1.2816) (-0.7637)
MOM3 0.0015 0.1397 -0.226*** 0.9385*** 0.0637 0.0208 0.5480
T-Statistics (0.1643) (1.5967) (-2.9953) (18.2863) (0.7824) (0.0411)
RMOM1 0.0156 0.0517 0.1532* 0.5484*** 0.0538 -0.5977 0.2492
T-Statistics (1.5699) (0.5521) (1.8973) (9.9825) (0.6181) (-1.1035)
RMOM2 0.0113 0.1585* 0.1784** 0.605*** -0.0059 -0.2746 0.3411
T-Statistics (1.2697) (1.8896) (2.4658) (12.2941) (-0.0756) (-0.5659)
RMOM3 -0.0082 0.2474*** -0.1331* 0.8274*** 0.1442* 0.4636 0.5259
T-Statistics (-0.9446) (3.0125) (-1.8794) (17.1748) (1.8881) (0.976)
VOLPRC 0.0333*** 0.0933 -0.0558 -0.0334 -1.1449*** -0.6811* 0.5945
T-Statistics (5.0723) (1.5058) (-1.0437) (-0.9188) (-19.8649) (-1.9004)
STDPRCVOL 0.0345*** 0.1435** -0.0427 0.0271 -1.1803*** -0.2099 0.5971
T-Statistics (5.1406) (2.2646) (-0.7813) (0.728) (-20.0288) (-0.5727)
MEANABS 0.0315*** 0.1316** -0.0053 0.0319 -1.2166*** -0.3204 0.6462
T-Statistics (5.0128) (2.22) (-0.1035) (0.9175) (-22.0572) (-0.9341)
Average 0.4333
Notes. This table reports the empirical results of Model 9. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model 9
is 0.4333. Four mispricing factors still cannot be well explained due to their statistically significant alpha values. Only MOM1 and VOLPRC have exposure to electricity.

71
Table A.14: Model 10

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0421*** 0.3537*** 0.0555 -0.1386* -0.0795 1.9784*** 0.0399

T-Statistics (2.9386) (2.6923) (0.5435) (-1.9139) (-1.3899) (2.751)
MOM2 -0.0135 -0.0119 0.1220 0.7816*** 0.1792*** -0.4165 0.3835
T-Statistics (-1.1536) (-0.1108) (1.4598) (13.1997) (3.8329) (-0.708)
MOM3 -0.0163* -0.0175 -0.2799*** 0.9425*** 0.179*** 0.0692 0.5764
T-Statistics (-1.6722) (-0.1953) (-4.0205) (19.1102) (4.5961) (0.1412)
RMOM1 0.0177 0.0607 0.1416* 0.5526*** -0.0221 -0.6016 0.2489
T-Statistics (1.6393) (0.613) (1.8388) (10.1307) (-0.5127) (-1.1103)
RMOM2 0.0037 0.0967 0.1651** 0.6042*** 0.0768** -0.2550 0.3496
T-Statistics (0.3804) (1.0983) (2.4098) (12.4489) (2.0025) (-0.5288)
RMOM3 -0.0102 0.2077** -0.1795*** 0.8383*** 0.0187 0.4733 0.5208
T-Statistics (-1.0733) (2.3805) (-2.6447) (17.4337) (0.4933) (0.9909)
VOLPRC 0.02* 0.1684* 0.2553*** -0.1211** 0.1446*** -0.6826 0.1040
T-Statistics (1.888) (1.7298) (3.3714) (-2.2577) (3.4145) (-1.281)
STDPRCVOL 0.0228** 0.2373** 0.2819*** -0.0633 0.129*** -0.2165 0.0931
T-Statistics (2.0841) (2.3621) (3.6074) (-1.1428) (2.9516) (-0.3938)
MEANABS 0.0175 0.2121** 0.3254*** -0.0613 0.1528*** -0.3222 0.1182
T-Statistics (1.6218) (2.1443) (4.2282) (-1.1239) (3.549) (-0.5948)
Average 0.2705
Notes. This table reports the empirical results of Model 10. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
10 is 0.2705. Three mispricing factors still cannot be well explained due to their statistically significant alpha values, MOM3 becomes significant as well. Only MOM1 has
exposure to electricity.

72
Table A.15: Model 11

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0342*** 0.2579** -0.0098 -0.1129 -0.0207 -0.1695 1.9899*** 0.0376

T-Statistics (2.5945) (2.068) (-0.0911) (-1.2515) (-0.2464) (-1.4636) (2.7549)
MOM2 0.0010 0.1487 0.1093 0.5681*** 0.3538*** -0.1342 -0.2063 0.4092
T-Statistics (0.0949) (1.4911) (1.2726) (7.873) (5.2761) (-1.4495) (-0.3571)
MOM3 0.0001 0.1549* -0.2309*** 0.8408*** 0.1545*** 0.0591 0.1310 0.5567
T-Statistics (0.0155) (1.7846) (-3.089) (13.386) (2.6476) (0.7332) (0.2605)
RMOM1 0.0099 0.1145 0.1331** 0.1462*** 0.6361*** 0.0350 -0.1441 0.4978
T-Statistics (1.217) (1.4922) (2.0152) (2.6326) (12.3281) (0.4914) (-0.3242)
RMOM2 0.0097 0.1761** 0.1727** 0.492*** 0.1787*** -0.0112 -0.1471 0.3606
T-Statistics (1.1048) (2.1271) (2.4232) (8.2125) (3.2098) (-0.1455) (-0.3068)
RMOM3 -0.0087 0.2526*** -0.1348* 0.794*** 0.0529 0.1427* 0.5013 0.5258
T-Statistics (-0.9972) (3.069) (-1.9021) (13.3334) (0.9563) (1.867) (1.0517)
VOLPRC 0.0336*** 0.0896 -0.0546 -0.0099 -0.0371 -1.1438*** -0.7076** 0.5942
T-Statistics (5.1128) (1.4429) (-1.0211) (-0.2209) (-0.8889) (-19.8344) (-1.9668)
STDPRCVOL 0.0349*** 0.1387** -0.0412 0.0578 -0.0486 -1.1788*** -0.2445 0.5975
T-Statistics (5.1995) (2.1851) (-0.7534) (1.2587) (-1.1392) (-20.0096) (-0.6653)
MEANABS 0.0314*** 0.1318** -0.0053 0.0309 0.0017 -1.2167*** -0.3192 0.6450
T-Statistics (4.9942) (2.2142) (-0.1044) (0.7168) (0.0415) (-22.0169) (-0.9259)
Average 0.4694
Notes. This table reports the empirical results of Model 11. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
11 is 0.4694. Four mispricing factors still cannot be well explained due to their statistically significant alpha values. Only MOM1 and VOLPRC have exposures to electricity.

73
Table A.16: Model 12

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0421*** 0.3523*** 0.0556 -0.1339 -0.0073 -0.0788 1.9734*** 0.0367

T-Statistics (2.9335) (2.6599) (0.5429) (-1.4844) (-0.0858) (-1.3623) (2.7306)
MOM2 -0.0134 0.0464 0.1207 0.5778*** 0.3214*** 0.1488*** -0.1954 0.4252
T-Statistics (-1.18) (0.4441) (1.4951) (8.121) (4.8144) (3.2628) (-0.343)
MOM3 -0.0163* 0.0045 -0.2804*** 0.8656*** 0.1214** 0.1675*** 0.1527 0.5812
T-Statistics (-1.6752) (0.0506) (-4.0508) (14.1835) (2.1193) (4.2833) (0.3124)
RMOM1 0.018** 0.1792** 0.1388** 0.138** 0.6537*** -0.0841** -0.1521 0.5066
T-Statistics (2.0615) (2.2189) (2.2245) (2.5091) (12.662) (-2.3847) (-0.3452)
RMOM2 0.0037 0.1268 0.1644** 0.4989*** 0.1661*** 0.0610 -0.1407 0.3658
T-Statistics (0.3944) (1.4488) (2.43) (8.365) (2.9679) (1.5962) (-0.2946)
RMOM3 -0.0102 0.2172** -0.1797*** 0.8051*** 0.0523 0.0138 0.5093 0.5206
T-Statistics (-1.0702) (2.472) (-2.6474) (13.4524) (0.9316) (0.3591) (1.0625)
VOLPRC 0.02* 0.1528 0.2556*** -0.0666 -0.0859 0.1528*** -0.7417 0.1066
T-Statistics (1.8865) (1.5614) (3.3811) (-0.9993) (-1.3739) (3.5771) (-1.3894)
STDPRCVOL 0.0228** 0.22** 0.2823*** -0.0030 -0.0950 0.138*** -0.2818 0.0966
T-Statistics (2.0835) (2.1801) (3.6195) (-0.0442) (-1.4717) (3.133) (-0.5118)
MEANABS 0.0175 0.2032** 0.3256*** -0.0300 -0.0492 0.1575*** -0.3560 0.1171
T-Statistics (1.6184) (2.0389) (4.2281) (-0.4426) (-0.7729) (3.6196) (-0.6548)
Average 0.3063
Notes. This table reports the empirical results of Model 12. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
12 is 0.3063. Four mispricing factors still cannot be well explained due to their statistically significant alpha values. Only MOM1 and VOLPRC have exposures to electricity.

74
Table A.17: Model 13

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0436*** 0.3332** -0.0013 -0.1231* -0.2027* -0.0963* 1.9809*** 0.0461

T-Statistics (3.0457) (2.5345) (-0.0118) (-1.6925) (-1.7344) (-1.6657) (2.7635)
MOM2 -0.0131 -0.0184 0.1039 0.7866*** -0.0647 0.1739*** -0.4157 0.3823
T-Statistics (-1.1107) (-0.1708) (1.1821) (13.1708) (-0.6743) (3.662) (-0.706)
MOM3 -0.0172* -0.0045 -0.244*** 0.9327*** 0.1280 0.1896*** 0.0676 0.5786
T-Statistics (-1.7686) (-0.0507) (-3.3456) (18.8172) (1.6064) (4.8122) (0.1383)
RMOM1 0.0174 0.0655 0.1549* 0.549*** 0.0477 -0.0181 -0.6022 0.2471
T-Statistics (1.6026) (0.6582) (1.9135) (9.9761) (0.539) (-0.4144) (-1.1101)
RMOM2 0.0035 0.0988 0.1709** 0.6027*** 0.0207 0.0785** -0.2552 0.3476
T-Statistics (0.3636) (1.1158) (2.3712) (12.3034) (0.2633) (2.0153) (-0.5285)
RMOM3 -0.0113 0.2233** -0.1361* 0.8265*** 0.1549** 0.0316 0.4714 0.5254
T-Statistics (-1.1949) (2.5621) (-1.9183) (17.1409) (1.9987) (0.8241) (0.9917)
VOLPRC 0.0282*** 0.0545 -0.0606 -0.0350 -1.1276*** 0.0511* -0.6685* 0.5973
T-Statistics (3.9562) (0.8312) (-1.1372) (-0.9645) (-19.3544) (1.773) (-1.8714)
STDPRCVOL 0.0313*** 0.1191* -0.0457 0.0261 -1.1694*** 0.0320 -0.2020 0.5974
T-Statistics (4.2798) (1.7727) (-0.8362) (0.7019) (-19.5687) (1.083) (-0.5512)
MEANABS 0.0262*** 0.0911 -0.0104 0.0303 -1.1985*** 0.0534* -0.3072 0.6494
T-Statistics (3.8403) (1.4538) (-0.2035) (0.8746) (-21.5173) (1.9373) (-0.8995)
Average 0.4413
Notes. This table reports the empirical results of Model 13. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
13 is 0.4413. Four mispricing factors still cannot be well explained due to their statistically significant alpha values, MOM3 has a significant alpha as well. Only MOM1 and
VOLPRC have exposures to electricity.

75
Table A.18: Model 14

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0436*** 0.3331** -0.0012 -0.1227 -0.0006 -0.2027* -0.0962 1.9805*** 0.0430
T-Statistics (3.0406) (2.5138) (-0.0116) (-1.3607) (-0.0068) (-1.7294) (-1.6445) (2.7494)
MOM2 -0.0127 0.0383 0.0967 0.5825*** 0.3243*** -0.0854 0.1414*** -0.1924 0.4250
T-Statistics (-1.123) (0.3651) (1.1406) (8.1642) (4.8503) (-0.9204) (3.0543) (-0.3377)
MOM3 -0.0171* 0.0160 -0.2466*** 0.8589*** 0.1174** 0.1206 0.1779*** 0.1484 0.5830
T-Statistics (-1.766) (0.1782) (-3.3984) (14.0675) (2.052) (1.5192) (4.4902) (0.3043)
RMOM1 0.018** 0.1798** 0.1405** 0.1377** 0.6535*** 0.0061 -0.0835** -0.1523 0.5050
T-Statistics (2.0495) (2.2147) (2.1402) (2.4923) (12.6241) (0.0856) (-2.3304) (-0.3451)
RMOM2 0.0037 0.1278 0.1673** 0.4983*** 0.1658*** 0.0102 0.0619 -0.1411 0.3638
T-Statistics (0.3853) (1.4524) (2.3495) (8.3203) (2.954) (0.131) (1.5923) (-0.2949)
RMOM3 -0.0113 0.2316*** -0.1371* 0.7967*** 0.0473 0.1519* 0.0269 0.5040 0.5250
T-Statistics (-1.1894) (2.6392) (-1.9318) (13.3386) (0.8453) (1.9569) (0.6929) (1.0562)
VOLPRC 0.0282*** 0.0459 -0.0596 -0.0042 -0.0488 -1.1245*** 0.056* -0.7021** 0.5978
T-Statistics (3.9517) (0.6969) (-1.1175) (-0.0946) (-1.1614) (-19.2919) (1.9233) (-1.9602)
STDPRCVOL 0.0312*** 0.1093 -0.0445 0.0616 -0.0565 -1.1658*** 0.0377 -0.2408 0.5983
T-Statistics (4.2772) (1.6175) (-0.8143) (1.3405) (-1.3107) (-19.5111) (1.2623) (-0.6559)
MEANABS 0.0262*** 0.0894 -0.0102 0.0364 -0.0097 -1.1979*** 0.0543* -0.3139 0.6483
T-Statistics (3.8329) (1.4157) (-0.199) (0.847) (-0.2404) (-21.4503) (1.9487) (-0.9146)
Average 0.4766
Notes. This table reports the empirical results of Model 14. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
14 is 0.4766. Four mispricing factors still cannot be well explained due to their statistically significant alpha values, MOM3 and RMOM1 have a significant alpha as well.
Only MOM1 and VOLPRC have exposures to electricity.

76
Table A.19: Model 15

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0398*** 0.262** 0.0408 -0.1114 -0.0434 -0.4057 0.0134

T-Statistics (2.9819) (2.1052) (0.396) (-1.2202) (-0.5137) (-0.9846)
MOM2 0.0016 0.1728* 0.1474* 0.5634*** 0.353*** -0.2743 0.4082
T-Statistics (0.1491) (1.7565) (1.8105) (7.8068) (5.278) (-0.8422)
MOM3 0.0005 0.1434* -0.2484*** 0.8457*** 0.1541*** -0.0377 0.5573
T-Statistics (0.0509) (1.6768) (-3.5105) (13.4853) (2.6507) (-0.133)
RMOM1 0.0093 0.1106 0.1231** 0.1465*** 0.6383*** 0.0936 0.4991
T-Statistics (1.1436) (1.4641) (1.9683) (2.6436) (12.4307) (0.3741)
RMOM2 0.0096 0.1795** 0.1755*** 0.4921*** 0.1797*** -0.0729 0.3626
T-Statistics (1.1012) (2.2029) (2.6023) (8.2339) (3.245) (-0.2703)
RMOM3 -0.0074 0.2225*** -0.1772*** 0.8071*** 0.0499 -0.1277 0.5205
T-Statistics (-0.8471) (2.7271) (-2.6242) (13.4904) (0.9) (-0.4729)
VOLPRC 0.0333*** 0.2898*** 0.2841*** -0.0883 -0.0475 0.0254 0.0662
T-Statistics (3.3427) (3.1208) (3.6958) (-1.2959) (-0.7528) (0.0825)
STDPRCVOL 0.0346*** 0.3403*** 0.3101*** -0.0271 -0.0631 0.3469 0.0733
T-Statistics (3.3985) (3.5858) (3.9483) (-0.3889) (-0.9775) (1.1041)
MEANABS 0.0321*** 0.3396*** 0.3552*** -0.0507 -0.0136 -0.0051 0.0807
T-Statistics (3.1715) (3.5972) (4.5464) (-0.7314) (-0.2127) (-0.0163)
Average 0.2868
Notes. This table reports the empirical results of Model 15. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
15 is 0.2868. Four mispricing factors still cannot be well explained due to their statistically significant alpha values. No factors have exposures to graphics.

77
Table A.20: Model 16

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0393*** 0.2375* -0.0105 -0.1261* -0.1697 -0.4096 0.0193

T-Statistics (2.9588) (1.8939) (-0.0967) (-1.7055) (-1.4519) (-0.9973)
MOM2 0.0042 0.1176 0.1180 0.7981*** -0.1252 -0.3032 0.3577
T-Statistics (0.3815) (1.1356) (1.3179) (13.0654) (-1.2965) (-0.8935)
MOM3 0.0017 0.1393 -0.2262*** 0.9391*** 0.0635 -0.0454 0.5480
T-Statistics (0.1842) (1.5954) (-2.9981) (18.2394) (0.7808) (-0.1589)
RMOM1 0.0142 0.0582 0.1528* 0.5496*** 0.0535 0.0527 0.2462
T-Statistics (1.4328) (0.6212) (1.888) (9.9508) (0.6129) (0.1719)
RMOM2 0.0110 0.1611* 0.1775** 0.6073*** -0.0064 -0.0853 0.3406
T-Statistics (1.2392) (1.9236) (2.4523) (12.2945) (-0.0822) (-0.3109)
RMOM3 -0.0069 0.2421*** -0.1333* 0.8279*** 0.1442* -0.1252 0.5248
T-Statistics (-0.7943) (2.9498) (-1.8795) (17.107) (1.8857) (-0.4656)
VOLPRC 0.032*** 0.1004 -0.0567 -0.0309 -1.1456*** -0.0181 0.5897
T-Statistics (4.8526) (1.6142) (-1.0552) (-0.842) (-19.7589) (-0.0888)
STDPRCVOL 0.0331*** 0.1467** -0.0411 0.0230 -1.1795*** 0.3031 0.5995
T-Statistics (4.9602) (2.3262) (-0.7538) (0.6184) (-20.0737) (1.4675)
MEANABS 0.031*** 0.1348** -0.0060 0.0338 -1.2171*** -0.0538 0.6453
T-Statistics (4.9381) (2.2752) (-0.1177) (0.9674) (-22.0357) (-0.277)
Average 0.4301
Notes. This table reports the empirical results of Model 16. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
16 is 0.4301. Four mispricing factors still cannot be well explained due to their statistically significant alpha values. No factors have exposures to graphics.

78
Table A.21: Model 17

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0478*** 0.3354** 0.0562 -0.1384* -0.0843 -0.4194 0.0194

T-Statistics (3.3047) (2.5296) (0.5446) (-1.8854) (-1.4593) (-1.0207)
MOM2 -0.0135 -0.0079 0.1203 0.787*** 0.1789*** -0.2628 0.3837
T-Statistics (-1.1532) (-0.0733) (1.4394) (13.2502) (3.8256) (-0.7903)
MOM3 -0.0161* -0.0181 -0.2799*** 0.9425*** 0.1788*** -0.0127 0.5764
T-Statistics (-1.654) (-0.2027) (-4.0194) (19.0471) (4.591) (-0.0459)
RMOM1 0.0162 0.0663 0.141* 0.5538*** -0.0209 0.0464 0.2459
T-Statistics (1.5032) (0.6692) (1.8275) (10.0998) (-0.4846) (0.1513)
RMOM2 0.0034 0.0991 0.1645** 0.6061*** 0.0769** -0.0699 0.3491
T-Statistics (0.349) (1.1269) (2.3993) (12.443) (2.0052) (-0.2562)
RMOM3 -0.0088 0.2033** -0.1795*** 0.8388*** 0.0175 -0.1276 0.5196
T-Statistics (-0.9197) (2.3307) (-2.6405) (17.3651) (0.4592) (-0.472)
VOLPRC 0.0184* 0.1747* 0.2546*** -0.1198** 0.146*** 0.0574 0.0993
T-Statistics (1.7283) (1.7928) (3.3538) (-2.2205) (3.4365) (0.19)
STDPRCVOL 0.0211* 0.2391** 0.2833*** -0.0684 0.1309*** 0.3770 0.0971
T-Statistics (1.9279) (2.3886) (3.6328) (-1.2342) (2.9993) (1.2152)
MEANABS 0.0167 0.2151** 0.3251*** -0.0607 0.1534*** 0.0261 0.1172
T-Statistics (1.5499) (2.1763) (4.2214) (-1.1082) (3.5609) (0.0852)
Average 0.2675
Notes. This table reports the empirical results of Model 17. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
17 is 0.2675. Three mispricing factors still cannot be well explained due to their statistically significant alpha values, MOM3 has a significant alpha value as well. No factors
have exposures to graphics.

79
Table A.22: Model 18

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0396*** 0.2337* -0.0092 -0.1001 -0.0409 -0.1685 -0.4124 0.0169

T-Statistics (2.9753) (1.8582) (-0.0851) (-1.0944) (-0.4846) (-1.4397) (-1.0025)
MOM2 0.0014 0.1501 0.1072 0.5724*** 0.355*** -0.1354 -0.2797 0.4104
T-Statistics (0.1375) (1.5096) (1.2499) (7.9181) (5.3171) (-1.4632) (-0.8602)
MOM3 0.0005 0.1533* -0.2309*** 0.8418*** 0.1532*** 0.0591 -0.0353 0.5567
T-Statistics (0.0568) (1.7695) (-3.0883) (13.3626) (2.6329) (0.7336) (-0.1246)
RMOM1 0.0093 0.1165 0.1335** 0.1441*** 0.6377*** 0.0352 0.0950 0.4978
T-Statistics (1.1462) (1.5214) (2.0203) (2.5883) (12.4023) (0.4934) (0.3792)
RMOM2 0.0096 0.1776** 0.1721** 0.4929*** 0.1799*** -0.0116 -0.0734 0.3605
T-Statistics (1.0981) (2.1488) (2.4131) (8.2034) (3.2423) (-0.1507) (-0.2716)
RMOM3 -0.0073 0.2464*** -0.1347* 0.7975*** 0.0478 0.1428* -0.1220 0.5244
T-Statistics (-0.8354) (2.9962) (-1.8986) (13.3353) (0.8647) (1.8666) (-0.4537)
VOLPRC 0.0322*** 0.0976 -0.0558 -0.0116 -0.0304 -1.1447*** -0.0201 0.5890
T-Statistics (4.8785) (1.5654) (-1.037) (-0.2553) (-0.7252) (-19.7242) (-0.0986)
STDPRCVOL 0.0335*** 0.1425** -0.0397 0.0519 -0.0454 -1.1782*** 0.3001 0.5997
T-Statistics (5.0079) (2.2566) (-0.7285) (1.1295) (-1.0709) (-20.052) (1.4532)
MEANABS 0.031*** 0.1353** -0.0062 0.0309 0.0046 -1.2172*** -0.0535 0.6441
T-Statistics (4.9186) (2.2743) (-0.1201) (0.7145) (0.1146) (-21.9976) (-0.275)
Average 0.4666
Notes. This table reports the empirical results of Model 18. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
18 is 0.4666. Four mispricing factors still cannot be well explained due to their statistically significant alpha values. No factors have exposures to graphics.

80
Table A.23: Model 19

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0478*** 0.3307** 0.0563 -0.1213 -0.0268 -0.0818 -0.4207 0.0165

T-Statistics (3.2962) (2.475) (0.5446) (-1.3273) (-0.3139) (-1.399) (-1.0223)
MOM2 -0.0129 0.0486 0.1193 0.5814*** 0.3228*** 0.148*** -0.2471 0.4261
T-Statistics (-1.1427) (0.4663) (1.4792) (8.1566) (4.8536) (3.2474) (-0.7703)
MOM3 -0.0159 0.0029 -0.2802*** 0.8661*** 0.1199** 0.1674*** -0.0069 0.5811
T-Statistics (-1.6405) (0.0319) (-4.0472) (14.1532) (2.1003) (4.2778) (-0.0251)
RMOM1 0.0174** 0.1809** 0.139** 0.1363** 0.6553*** -0.0837** 0.0782 0.5066
T-Statistics (1.9952) (2.2435) (2.2265) (2.4709) (12.7333) (-2.3726) (0.3149)
RMOM2 0.0037 0.1284 0.1639** 0.4995*** 0.1673*** 0.0609 -0.0617 0.3658
T-Statistics (0.3858) (1.4692) (2.4228) (8.3534) (2.9988) (1.5923) (-0.2294)
RMOM3 -0.0087 0.2116** -0.1796*** 0.8087*** 0.0473 0.0129 -0.1253 0.5191
T-Statistics (-0.9101) (2.4091) (-2.6414) (13.4561) (0.8428) (0.3367) (-0.4632)
VOLPRC 0.0182* 0.1609 0.2549*** -0.0696 -0.0788 0.1535*** 0.0536 0.1010
T-Statistics (1.7163) (1.6426) (3.3603) (-1.0381) (-1.2603) (3.5822) (0.1775)
STDPRCVOL 0.0209* 0.2231** 0.2836*** -0.0101 -0.0915 0.1396*** 0.3725 0.1001
T-Statistics (1.9157) (2.2185) (3.6426) (-0.1466) (-1.4258) (3.1741) (1.2029)
MEANABS 0.0166 0.2071** 0.3252*** -0.0314 -0.0458 0.1578*** 0.0239 0.1159
T-Statistics (1.5409) (2.0805) (4.2199) (-0.4617) (-0.7212) (3.624) (0.0779)
Average 0.3036
Notes. This table reports the empirical results of Model 19. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
19 is 0.3036. Three mispricing factors still cannot be well explained due to their statistically significant alpha values, RMOM1 has a significant alpha value as well. No
factors have exposures to graphics.

81
Table A.24: Model 20

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0493*** 0.3147** -0.0010 -0.1226* -0.2041* -0.1013* -0.4310 0.0257

T-Statistics (3.4147) (2.3719) (-0.0092) (-1.6631) (-1.7277) (-1.7339) (-1.0524)
MOM2 -0.0130 -0.0146 0.1018 0.7921*** -0.0662 0.1734*** -0.2666 0.3826
T-Statistics (-1.1079) (-0.135) (1.1579) (13.2223) (-0.6891) (3.6518) (-0.8009)
MOM3 -0.0171* -0.0052 -0.244*** 0.9326*** 0.1280 0.1895*** -0.0054 0.5786
T-Statistics (-1.7532) (-0.0578) (-3.344) (18.7506) (1.6061) (4.8075) (-0.0196)
RMOM1 0.0159 0.0711 0.1544* 0.5501*** 0.0477 -0.0170 0.0492 0.2441
T-Statistics (1.4658) (0.7141) (1.9025) (9.9438) (0.5382) (-0.3866) (0.16)
RMOM2 0.0032 0.1012 0.1702** 0.6046*** 0.0203 0.0786** -0.0687 0.3471
T-Statistics (0.332) (1.1438) (2.3594) (12.2965) (0.258) (2.017) (-0.2515)
RMOM3 -0.0099 0.219** -0.1361* 0.8269*** 0.1545** 0.0303 -0.1188 0.5242
T-Statistics (-1.044) (2.5117) (-1.9159) (17.0685) (1.9906) (0.7899) (-0.4413)
VOLPRC 0.0268*** 0.0607 -0.0616 -0.0327 -1.1278*** 0.0521* -0.0071 0.5927
T-Statistics (3.7403) (0.9223) (-1.1486) (-0.8939) (-19.2456) (1.7984) (-0.0349)
STDPRCVOL 0.0298*** 0.121* -0.0442 0.0218 -1.168*** 0.0337 0.3102 0.5999
T-Statistics (4.088) (1.8087) (-0.8109) (0.5873) (-19.605) (1.1424) (1.502)
MEANABS 0.0257*** 0.0939 -0.0110 0.0320 -1.1988*** 0.0537* -0.0424 0.6485
T-Statistics (3.764) (1.4997) (-0.2164) (0.9182) (-21.493) (1.9456) (-0.2195)
Average 0.4382
Notes. This table reports the empirical results of Model 20. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
20 is 0.4382. Four mispricing factors still cannot be well explained due to their statistically significant alpha values, MOM3 also has a significant alpha. No factors have
exposures to graphics.

82
Table A.25: Model 21

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0493*** 0.3113** -0.0006 -0.1099 -0.0202 -0.2029* -0.0993* -0.4319 0.0227
T-Statistics (3.4059) (2.3292) (-0.0053) (-1.2025) (-0.2369) (-1.7125) (-1.6784) (-1.0529)
MOM2 -0.0123 0.0403 0.0950 0.5863*** 0.3256*** -0.0867 0.1405*** -0.2519 0.4259
T-Statistics (-1.0827) (0.3854) (1.1206) (8.2016) (4.8901) (-0.9356) (3.036) (-0.785)
MOM3 -0.0168* 0.0144 -0.2464*** 0.8593*** 0.116** 0.1207 0.1778*** -0.0002 0.5829
T-Statistics (-1.7346) (0.1604) (-3.394) (14.0343) (2.0337) (1.5205) (4.4854) (-0.0007)
RMOM1 0.0174** 0.1815** 0.1408** 0.1359** 0.6551*** 0.0064 -0.0831** 0.0785 0.5050
T-Statistics (1.9827) (2.2393) (2.1429) (2.4538) (12.6954) (0.0886) (-2.3177) (0.3158)
RMOM2 0.0036 0.1293 0.1667** 0.499*** 0.167*** 0.0098 0.0617 -0.0612 0.3637
T-Statistics (0.3767) (1.4722) (2.3408) (8.3083) (2.9851) (0.1257) (1.5875) (-0.227)
RMOM3 -0.0098 0.2261** -0.137* 0.8001*** 0.0423 0.1519* 0.0261 -0.1169 0.5235
T-Statistics (-1.0329) (2.5767) (-1.9266) (13.3389) (0.7573) (1.9528) (0.6709) (-0.434)
VOLPRC 0.0267*** 0.0536 -0.0607 -0.0060 -0.0422 -1.1252*** 0.0564* -0.0090 0.5927
T-Statistics (3.7265) (0.8096) (-1.1321) (-0.1332) (-1.0005) (-19.1805) (1.9247) (-0.0442)
STDPRCVOL 0.0297*** 0.112* -0.0431 0.0557 -0.0536 -1.1646*** 0.0391 0.3078 0.6007
T-Statistics (4.0745) (1.6656) (-0.7911) (1.2122) (-1.2517) (-19.5464) (1.3133) (1.4917)
MEANABS 0.0257*** 0.0928 -0.0109 0.0363 -0.0068 -1.1984*** 0.0544* -0.0427 0.6474
T-Statistics (3.7553) (1.4705) (-0.2133) (0.8404) (-0.169) (-21.4288) (1.9465) (-0.2207)
Average 0.4738
Notes. This table reports the empirical results of Model 21. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
21 is 0.4738. Four mispricing factors still cannot be well explained due to their statistically significant alpha values, MOM3 and RMOM1 also has a significant alpha. No
factors have exposures to graphics.

83
Table A.26: Model 22

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0359*** 0.2864** 0.0373 -0.1144 -0.0234 2.1431*** -0.5908 0.0374

T-Statistics (2.7147) (2.3241) (0.3663) (-1.2689) (-0.2794) (2.9315) (-1.4345)
MOM2 0.0018 0.1712* 0.1477* 0.5636*** 0.3516*** -0.1410 -0.2621 0.4064
T-Statistics (0.1719) (1.7336) (1.8104) (7.797) (5.2327) (-0.2406) (-0.794)
MOM3 0.0002 0.1451* -0.2487*** 0.8455*** 0.1554*** 0.1474 -0.0504 0.5560
T-Statistics (0.022) (1.6899) (-3.5083) (13.4608) (2.6615) (0.2894) (-0.1756)
RMOM1 0.0096 0.1087 0.1233** 0.1467*** 0.6367*** -0.1721 0.1084 0.4977
T-Statistics (1.1742) (1.4329) (1.9699) (2.6441) (12.3408) (-0.3825) (0.4278)
RMOM2 0.0098 0.178** 0.1757*** 0.4923*** 0.1785*** -0.1309 -0.0616 0.3606
T-Statistics (1.1207) (2.1763) (2.6014) (8.2239) (3.2074) (-0.2697) (-0.2254)
RMOM3 -0.0084 0.2288*** -0.1781*** 0.8063*** 0.0551 0.5556 -0.1757 0.5210
T-Statistics (-0.9565) (2.7996) (-2.6389) (13.4833) (0.9908) (1.1459) (-0.6432)
VOLPRC 0.0347*** 0.2809*** 0.2854*** -0.0872 -0.0548 -0.7844 0.0931 0.0693
T-Statistics (3.4724) (3.023) (3.7185) (-1.2816) (-0.8675) (-1.4227) (0.2999)
STDPRCVOL 0.0353*** 0.3358*** 0.3108*** -0.0265 -0.0668 -0.4014 0.3816 0.0718
T-Statistics (3.4493) (3.5268) (3.9532) (-0.3804) (-1.0314) (-0.7106) (1.1992)
MEANABS 0.0328*** 0.3353*** 0.3559*** -0.0501 -0.0172 -0.3815 0.0278 0.0791
T-Statistics (3.2202) (3.54) (4.55) (-0.723) (-0.2671) (-0.6789) (0.088)
Average 0.2888
Notes. This table reports the empirical results of Model 22. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
22 is 0.2888. Four mispricing factors still cannot be well explained due to their statistically significant alpha values. Only MOM1 has exposure to electricity. No factors have
exposures to graphics.

84
Table A.27: Model 23

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0356*** 0.2597** -0.0143 -0.1161 -0.1722 2.1677*** -0.5984 0.0441

T-Statistics (2.701) (2.0943) (-0.134) (-1.5882) (-1.4924) (2.9855) (-1.4579)
MOM2 0.0048 0.1137 0.1187 0.7963*** -0.1247 -0.3851 -0.2696 0.3564
T-Statistics (0.4394) (1.0944) (1.3242) (13.0096) (-1.2905) (-0.6333) (-0.7844)
MOM3 0.0016 0.1396 -0.2263*** 0.9393*** 0.0635 0.0340 -0.0484 0.5466
T-Statistics (0.1768) (1.5939) (-2.9938) (18.1934) (0.779) (0.0663) (-0.1669)
RMOM1 0.0153 0.0517 0.1539* 0.5466*** 0.0542 -0.6270 0.1073 0.2470
T-Statistics (1.5352) (0.5517) (1.9026) (9.8926) (0.6215) (-1.1421) (0.3459)
RMOM2 0.0114 0.1585* 0.178** 0.6061*** -0.0061 -0.2574 -0.0629 0.3390
T-Statistics (1.2817) (1.8864) (2.4555) (12.2427) (-0.0783) (-0.5234) (-0.2262)
RMOM3 -0.0078 0.2473*** -0.1342* 0.8303*** 0.1436* 0.5098 -0.1696 0.5250
T-Statistics (-0.8912) (3.0088) (-1.8926) (17.1411) (1.8783) (1.0594) (-0.6234)
VOLPRC 0.0332*** 0.0933 -0.0555 -0.0341 -1.1448*** -0.6927* 0.0422 0.5932
T-Statistics (5.032) (1.5036) (-1.0366) (-0.9325) (-19.8295) (-1.9065) (0.2055)
STDPRCVOL 0.0336*** 0.1436** -0.0405 0.0216 -1.1791*** -0.2995 0.3292 0.5991
T-Statistics (5.0118) (2.272) (-0.7435) (0.5801) (-20.0559) (-0.8096) (1.5739)
MEANABS 0.0315*** 0.1316** -0.0055 0.0324 -1.2167*** -0.3132 -0.0265 0.6451
T-Statistics (4.9996) (2.2163) (-0.1067) (0.9246) (-22.0217) (-0.9007) (-0.1349)
Average 0.4328
Notes. This table reports the empirical results of Model 23. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
23 is 0.4328. Four mispricing factors still cannot be well explained due to their statistically significant alpha values. Only MOM1 and VOLPRC has exposure to electricity
No factors have exposures to graphics.

85
Table A.28: Model 24

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0439*** 0.3555*** 0.0526 -0.1287* -0.0816 2.1427*** -0.6053 0.0435

T-Statistics (3.0563) (2.7115) (0.5159) (-1.7732) (-1.4288) (2.9498) (-1.474)
MOM2 -0.0128 -0.0112 0.1209 0.7854*** 0.1784*** -0.3535 -0.2321 0.3824
T-Statistics (-1.0915) (-0.104) (1.4449) (13.1956) (3.8112) (-0.5933) (-0.6891)
MOM3 -0.0163* -0.0174 -0.28*** 0.9428*** 0.1789*** 0.0744 -0.0192 0.5750
T-Statistics (-1.6581) (-0.1943) (-4.0145) (19.0038) (4.5855) (0.1498) (-0.0683)
RMOM1 0.0174 0.0604 0.1421* 0.551*** -0.0217 -0.6290 0.1010 0.2467
T-Statistics (1.6043) (0.6089) (1.8421) (10.0425) (-0.5037) (-1.1454) (0.3254)
RMOM2 0.0038 0.0969 0.1649** 0.605*** 0.0766** -0.2417 -0.0489 0.3475
T-Statistics (0.3932) (1.0982) (2.4021) (12.3924) (1.9943) (-0.4945) (-0.1769)
RMOM3 -0.0097 0.2082** -0.1803*** 0.8411*** 0.0181 0.5202 -0.1727 0.5198
T-Statistics (-1.0162) (2.3841) (-2.6539) (17.4004) (0.477) (1.0751) (-0.6315)
VOLPRC 0.0197* 0.168* 0.2558*** -0.1231** 0.1451*** -0.7150 0.1194 0.1015
T-Statistics (1.8465) (1.7236) (3.3737) (-2.2811) (3.4185) (-1.3241) (0.3912)
STDPRCVOL 0.0217** 0.236** 0.2838*** -0.0699 0.1304*** -0.3266 0.4053 0.0951
T-Statistics (1.9723) (2.3522) (3.6357) (-1.2583) (2.986) (-0.5874) (1.2897)
MEANABS 0.0173 0.212** 0.3257*** -0.0622 0.153*** -0.3372 0.0554 0.1154
T-Statistics (1.5989) (2.1391) (4.2242) (-1.1338) (3.5467) (-0.6142) (0.1784)
Average 0.2697
Notes. This table reports the empirical results of Model 24. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
24 is 0.2697. Three mispricing factors still cannot be well explained due to their statistically significant alpha values, MOM3 has a significant alpha as well. Only MOM1
has a exposure to electricity. No factors have exposures to graphics.

86
Table A.29: Model 25

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0358*** 0.2577** -0.0137 -0.1029 -0.0208 -0.1716 2.1529*** -0.5985 0.0412
T-Statistics (2.7066) (2.07) (-0.1277) (-1.1396) (-0.2481) (-1.4845) (2.9506) (-1.4559)
MOM2 0.0017 0.1486 0.1075 0.5726*** 0.3537*** -0.1352 -0.1333 -0.2682 0.4085
T-Statistics (0.1591) (1.4892) (1.2511) (7.9077) (5.2725) (-1.4588) (-0.2278) (-0.8138)
MOM3 0.0003 0.1549* -0.2312*** 0.8416*** 0.1545*** 0.0589 0.1440 -0.0477 0.5553
T-Statistics (0.0286) (1.7815) (-3.0873) (13.3386) (2.6432) (0.7299) (0.2825) (-0.1663)
RMOM1 0.0096 0.1146 0.1339** 0.1444*** 0.6361*** 0.0354 -0.1741 0.1100 0.4964
T-Statistics (1.1771) (1.4908) (2.0227) (2.5886) (12.312) (0.4961) (-0.3864) (0.4335)
RMOM2 0.0098 0.1761** 0.1723** 0.4931*** 0.1787*** -0.0114 -0.1302 -0.0621 0.3586
T-Statistics (1.1176) (2.1235) (2.413) (8.1932) (3.2046) (-0.1481) (-0.2679) (-0.2269)
RMOM3 -0.0082 0.2525*** -0.1359* 0.7968*** 0.0529 0.1421* 0.5475 -0.1693 0.5248
T-Statistics (-0.9437) (3.0651) (-1.9152) (13.3287) (0.9548) (1.8571) (1.1337) (-0.6224)
VOLPRC 0.0335*** 0.0896 -0.0543 -0.0106 -0.0371 -1.1437*** -0.7191** 0.0421 0.5929
T-Statistics (5.0724) (1.4409) (-1.014) (-0.2355) (-0.8873) (-19.799) (-1.9719) (0.2047)
STDPRCVOL 0.0341*** 0.1388** -0.0390 0.0523 -0.0485 -1.1777*** -0.3341 0.3290 0.5995
T-Statistics (5.0708) (2.1923) (-0.7155) (1.1386) (-1.1406) (-20.0368) (-0.9004) (1.5737)
MEANABS 0.0315*** 0.1318** -0.0055 0.0313 0.0017 -1.2168*** -0.3120 -0.0265 0.6439
T-Statistics (4.9811) (2.2105) (-0.1075) (0.7238) (0.0413) (-21.9813) (-0.8928) (-0.1346)
Average 0.4690
Notes. This table reports the empirical results of Model 25. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
25 is 0.4690. Four mispricing factors still cannot be well explained due to their statistically significant alpha values. Only MOM1 and VOLPRC has exposure to electricity
No factors have exposures to graphics.

87
Table A.30: Model 26

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0439*** 0.3542*** 0.0527 -0.1243 -0.0070 -0.0809 2.1379*** -0.6052 0.0404
T-Statistics (3.051) (2.6793) (0.5153) (-1.376) (-0.0825) (-1.4012) (2.9289) (-1.4714)
MOM2 -0.0127 0.0471 0.1195 0.5816*** 0.3216*** 0.1479*** -0.1314 -0.2358 0.4243
T-Statistics (-1.1149) (0.4509) (1.4796) (8.1459) (4.8123) (3.2411) (-0.2276) (-0.7251)
MOM3 -0.0162* 0.0046 -0.2805*** 0.8659*** 0.1214** 0.1674*** 0.1583 -0.0206 0.5799
T-Statistics (-1.6606) (0.0512) (-4.0448) (14.1279) (2.116) (4.2731) (0.3194) (-0.0737)
RMOM1 0.0178** 0.1789** 0.1393** 0.1365** 0.6537*** -0.0837** -0.1775 0.0935 0.5052
T-Statistics (2.0207) (2.212) (2.2281) (2.4717) (12.6431) (-2.3713) (-0.3975) (0.3716)
RMOM2 0.0039 0.1270 0.1642** 0.4997*** 0.1661*** 0.0608 -0.1269 -0.0508 0.3638
T-Statistics (0.4078) (1.4483) (2.4221) (8.343) (2.9636) (1.5885) (-0.2622) (-0.1862)
RMOM3 -0.0097 0.2177** -0.1806*** 0.8079*** 0.0524 0.0132 0.5564 -0.1733 0.5196
T-Statistics (-1.013) (2.4756) (-2.6566) (13.4496) (0.9322) (0.3428) (1.146) (-0.6335)
VOLPRC 0.0196* 0.1524 0.2562*** -0.0685 -0.0860 0.1532*** -0.7744 0.1204 0.1041
T-Statistics (1.8447) (1.5553) (3.3834) (-1.0241) (-1.3729) (3.5809) (-1.4316) (0.395)
STDPRCVOL 0.0216** 0.2188** 0.2842*** -0.0095 -0.0952 0.1395*** -0.3923 0.4064 0.0986
T-Statistics (1.9713) (2.1697) (3.6479) (-0.1387) (-1.4764) (3.1679) (-0.7048) (1.2956)
MEANABS 0.0173 0.203** 0.3259*** -0.0309 -0.0493 0.1577*** -0.3712 0.0559 0.1143
T-Statistics (1.5953) (2.0338) (4.2241) (-0.4538) (-0.7721) (3.6172) (-0.6736) (0.1801)
Average 0.3056
Notes. This table reports the empirical results of Model 26. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
26 is 0.3056. Three mispricing factors still cannot be well explained due to their statistically significant alpha values, MOM3 and RMOM1 has a significant alpha as well.
Only MOM1 has a exposure to electricity. No factors have exposures to graphics.

88
Table A.31: Model 27

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0454*** 0.3348** -0.0051 -0.1128 -0.2057* -0.0987* 2.1486*** -0.6175 0.0501
T-Statistics (3.1677) (2.5518) (-0.0475) (-1.5472) (-1.7634) (-1.7099) (2.9681) (-1.5088)
MOM2 -0.0124 -0.0178 0.1024 0.7905*** -0.0659 0.1729*** -0.3516 -0.2360 0.3813
T-Statistics (-1.0472) (-0.165) (1.1642) (13.1674) (-0.6856) (3.6383) (-0.5896) (-0.7)
MOM3 -0.0172* -0.0045 -0.2441*** 0.9329*** 0.1280 0.1896*** 0.0707 -0.0115 0.5772
T-Statistics (-1.7561) (-0.0503) (-3.3402) (18.7074) (1.6028) (4.8014) (0.1428) (-0.0412)
RMOM1 0.0170 0.0652 0.1556* 0.5472*** 0.0482 -0.0177 -0.6304 0.1039 0.2449
T-Statistics (1.5667) (0.6545) (1.9181) (9.8863) (0.5438) (-0.4045) (-1.1466) (0.3342)
RMOM2 0.0036 0.0989 0.1706** 0.6035*** 0.0205 0.0783** -0.2423 -0.0477 0.3455
T-Statistics (0.3762) (1.1154) (2.3627) (12.2458) (0.2599) (2.0066) (-0.495) (-0.1722)
RMOM3 -0.0109 0.2238** -0.1371* 0.8292*** 0.1542** 0.0310 0.5158 -0.1635 0.5244
T-Statistics (-1.1392) (2.5641) (-1.93) (17.1038) (1.9861) (0.8065) (1.0711) (-0.6007)
VOLPRC 0.0281*** 0.0543 -0.0603 -0.0358 -1.1273*** 0.0513* -0.6827* 0.0522 0.5960
T-Statistics (3.9149) (0.8278) (-1.1291) (-0.9827) (-19.3174) (1.7766) (-1.8855) (0.2549)
STDPRCVOL 0.0303*** 0.1183* -0.0437 0.0205 -1.1678*** 0.0333 -0.2931 0.3356 0.5994
T-Statistics (4.1414) (1.7644) (-0.8002) (0.5501) (-19.5892) (1.1293) (-0.7924) (1.6049)
MEANABS 0.0262*** 0.0911 -0.0105 0.0306 -1.1986*** 0.0533* -0.3028 -0.0162 0.6482
T-Statistics (3.8275) (1.4521) (-0.2051) (0.877) (-21.4804) (1.9311) (-0.8747) (-0.0825)
Average 0.4408
Notes. This table reports the empirical results of Model 27. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
27 is 0.4408. Four mispricing factors still cannot be well explained due to their statistically significant alpha values, MOM3 becomes significant. Only MOM1 and VOLPRC
has exposure to electricity No factors have exposures to graphics.

89
Table A.32: Model 28

𝛼 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 𝛽 Adj R-squared

MOM1 0.0454*** 0.3347** -0.0051 -0.1126 -0.0002 -0.2057* -0.0987* 2.1484*** -0.6175 0.0470
T-Statistics (3.1624) (2.5315) (-0.0474) (-1.2484) (-0.0021) (-1.7586) (-1.6889) (2.9535) (-1.5063)
MOM2 -0.0120 0.0389 0.0952 0.5864*** 0.3244*** -0.0865 0.1405*** -0.1269 -0.2410 0.4241
T-Statistics (-1.0562) (0.371) (1.1219) (8.1905) (4.849) (-0.9322) (3.0304) (-0.2199) (-0.7408)
MOM3 -0.0171* 0.0160 -0.2467*** 0.8591*** 0.1174** 0.1205 0.1778*** 0.1520 -0.0133 0.5816
T-Statistics (-1.7529) (0.1783) (-3.393) (14.0093) (2.0488) (1.5156) (4.4798) (0.3076) (-0.0479)
RMOM1 0.0177** 0.1795** 0.1411** 0.1362** 0.6535*** 0.0066 -0.0832** -0.1778 0.0939 0.5036
T-Statistics (2.0082) (2.2083) (2.1454) (2.4544) (12.6049) (0.0918) (-2.3159) (-0.3976) (0.3725)
RMOM2 0.0038 0.1279 0.1669** 0.4991*** 0.1658*** 0.0100 0.0617 -0.1274 -0.0502 0.3617
T-Statistics (0.3987) (1.4515) (2.3407) (8.2978) (2.9498) (0.1276) (1.5841) (-0.2628) (-0.1837)
RMOM3 -0.0108 0.2321*** -0.1382* 0.7994*** 0.0474 0.1511* 0.0262 0.5486 -0.1643 0.5240
T-Statistics (-1.1335) (2.6413) (-1.9435) (13.3327) (0.8463) (1.9442) (0.6753) (1.1351) (-0.603)
VOLPRC 0.028*** 0.0458 -0.0592 -0.0051 -0.0488 -1.1242*** 0.0562* -0.7165** 0.0529 0.5965
T-Statistics (3.9101) (0.6937) (-1.1093) (-0.1133) (-1.1604) (-19.2549) (1.9267) (-1.9737) (0.2587)
STDPRCVOL 0.0303*** 0.1083 -0.0424 0.0561 -0.0567 -1.1642*** 0.0390 -0.3323 0.3365 0.6004
T-Statistics (4.1385) (1.6083) (-0.778) (1.2211) (-1.3192) (-19.5319) (1.3096) (-0.8967) (1.611)
MEANABS 0.0262*** 0.0894 -0.0103 0.0367 -0.0097 -1.198*** 0.0543* -0.3095 -0.0160 0.6471
T-Statistics (3.82) (1.414) (-0.2005) (0.8494) (-0.2398) (-21.4134) (1.9426) (-0.8898) (-0.0816)
Average 0.4762
Notes. This table reports the empirical results of Model 28. *, **, *** represent significance at 10%, 5%, 1% level, respectively. The average adjusted R-squared of Model
28 is 0.4762. Four mispricing factors still cannot be well explained due to their statistically significant alpha values, MOM3 and RMOM1 become significant. Only MOM1
and VOLPRC has exposure to electricity No factors have exposures to graphics.

90
 Table A.33: Summary of 28 Models
Number of Number of
Significant Significant
Model Average Adj R-squared Number of Significant 𝛼
exposures to exposures to
𝐸𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦 𝐺𝑟𝑎𝑝ℎ𝑖𝑐𝑠
1 0.2973 5 - -
2 0.4278 4 - -
3 0.2783 4 - -
4 0.4674 4 - -
5 0.3048 5 - -
6 0.4391 5 - -
7 0.4746 6 - -
8 0.2896 4 1 -
9 0.4333 4 2 -
10 0.2705 4 1 -
11* 0.4694 4 2 -
12 0.3063 5 1 -
13 0.4413 5 2 -
14 0.4766 6 2 -
15 0.2868 4 - 0
16 0.4301 4 - 0
17 0.2675 4 - 0
18 0.4666 4 - 0
19 0.3036 4 - 0
20 0.4382 5 - 0
21 0.4738 6 - 0
22 0.2888 4 1 0
23 0.4328 4 2 0
24 0.2697 4 1 0
25 0.4690 4 2 0
26 0.3056 5 1 0
27 0.4408 5 2 0
28 0.4762 6 2 0
Notes. This table summarizes the average adjusted R-squared, number of significant 𝛼 (out of 9), number of
significant exposures to 𝐸𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦 and 𝐺𝑟𝑎𝑝ℎ𝑖𝑐𝑠 (out of 9) for each of our 28 adjusted model, respectively.
The – symbol indicates that no corresponding independent factor exists in this model. The model number with
* symbol illustrates the best adjusted model among 28 models.

We test whether nine dominant factors can be better explained by conducting

all the combinations of three different mispricing categories (momentum, volume and
volatility) and two cryptocurrency fundamental factors (electricity and computer power)
as additional independent factors that is added to origin three factor model of Liu et al.
(2019), we establish that the explanatory power of new adjusted model No.11, which
incorporates mispriced momentum and volume factor as well as electricity factor,
provides the best explanatory power as it has the highest adjusted R-squared and

91
minimum numbers of significant 𝛼 values for nine dominant factors. To illustrate, we
set the selection criteria as follow: we first determine the adjusted model with minimum
numbers of significant 𝛼 values, because increasing numbers of significant 𝛼 values
indicate deterioration as appearance of more abnormal returns which cannot be captured
by model. Second, we pick the one with highest adjusted R-squared, since adjusted R-
squared measures the proportion of the variation that can be captured by corresponding
model, and adjusted R-squared also penalize for adding any uncorrelated independent
factors that do not contribution to the fitness of the model, which mitigates the risk of
overfitting the model when we construct the 28 models. According to Table A.33,
Model No. 11 has the highest adjusted R-squared of 0.4694 and has four mispricing
factors, we therefore assume this model is the best adjusted model in our case.

Table A.33 also demonstrate the numbers of significant exposures to

𝐸𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦 and 𝐺𝑟𝑎𝑝ℎ𝑖𝑐𝑠 (proxy for computer power) as shown in the last two
columns. We propose that cryptocurrencies may have no correlation with computer
power, since none of 28 models have a significant exposure to this factor. Similarly,
from the perspective of 𝐸𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦, there are only one to two dominant factors are
exposed to 𝐸𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦 , indicating a weekly relationship with cryptocurrencies.
Although this finding is complied with Liu et al. (2020), further study is needed to dive
deeper into this area.

92