Economic Modelling 86 (2020) 299–305

Contents lists available at ScienceDirect

Economic Modelling
journal homepage: www.journals.elsevier.com/economic-modelling

Common risk factors in the returns on cryptocurrencies

Weiyi Liu a, Xuan Liang b, *, Guowei Cui c
a
School of Finance, Capital University of Economics and Business, Beijing, 100070, China
b
Research School of Finance, Actuarial Studies and Statistics, The Australian National University, Acton ACT 2601, Australia
c
School of Economics, Huazhong University of Science and Technology, Wuhan, 430074, China

A R T I C L E I N F O A B S T R A C T

JEL classification: This paper identifies three common risk factors in the returns on cryptocurrencies, which are related to crypto-
G12 currency market return, market capitalization (size) and momentum of cryptocurrencies. Investigating a collec-
Keywords: tion of 78 cryptocurrencies, we find that there are anomalous returns that decrease with size and increase with
Cryptocurrency return momentum, and the momentum effect is more significant in small cryptocurrencies. Moreover, Fama-
Market return Macbeth regressions show the size and momentum combine to capture the cross-sectional variation in average
Size
cryptocurrency returns. In the tests of the three-factor model, we find most cryptocurrencies and their portfolios
Momentum
Factor model
have significant exposures to the proposed three factors with insignificant intercepts, demonstrating that the three
factors explain average cryptocurrency returns very well.

1. Introduction anomalies and consider them to be potential factors in asset pricing

models. Nevertheless, Fama and French (2018) argued these various
Since its inception in 2009 as an open-source digital currency, Bitcoin anomalies may produce similar results to the combinations of the six
has inspired and provoked the release of a large number of crypto- factors mentioned above. Compared with the abundant researches in
currencies based on blockchain technology. The cryptocurrencies are stock markets, the study on asset pricing for cryptocurrencies is still very
shown to be extremely volatile with large average returns (Brauneis and limited and not well developed.
Mestel, 2018) in contrast to the traditional currencies. Thus, they are Although there are several similarities between the stocks and cryp-
mostly regarded as a new class of assets. In consideration of the huge and tocurrencies on the empirical facts, such as leptokurtosis (Chan et al.,
rapid growth of the cryptocurrency market, developing an appropriate 2017), heteroscedasticity (Gkillas and Katsiampa, 2018) and
asset pricing model on cryptocurrencies is an important and emergent long-memory (Phillip et al., 2019), cryptocurrency assets are funda-
complement to modern financial economics and investment theory. mentally different from stocks, as the effective stock factors mentioned
Finding common risk factors is always an open challenge in applying above are mostly based on the traditional financial theories that suggest
the arbitrage pricing theory, in which the expected return of a financial stock prices are determined by the present discounted value of funda-
asset can be modeled as a linear function of various factors or theoretical mentals (Miller and Modigliani, 1961; Campbell and Shiller, 1988),
market indices. There has been a great amount of literature focusing on while these fundamentals are not directly related to cryptocurrencies.
looking for effective factors for stock returns. For example, the market Gregoriou (2019) regresses the ten most heavily traded cryptocurrencies
excess return is the first factor that forms the CAPM model (Sharpe, to the Fama French stock factors and obtains significant positive alphas,
1964); the size and value factors are proposed in the Fama and French indicating that the abnormal returns on cryptocurrencies cannot be
(1993) three-factor model; the momentum factor is further added in the explained by stock markets. Liu and Tsyvinski (2018) further show
Carhart (1997) four-factor model; the profitability and investment fac- cryptocurrencies have no exposure to most common stock markets and
tors instead of the momentum factor are constructed in the Fama and macroeconomic factors or the returns of currencies and commodities.
French (2015) five-factor model. The above six factors are the most Bhambhwani et al. (2019) use the data of the five most prominent
widely adopted ones in real applications, although there are still many cryptocurrencies and also show there is little relationship between the
other factors being proposed in the hope of adding more contribution to cryptocurrency returns and stock factors, while they posit that crypto-
the explanation of average returns. Harvey et al. (2016) catalogue 316 currency prices are related to the computing power and the adoption of

* Corresponding author.
E-mail addresses: [email protected] (W. Liu), [email protected] (X. Liang), [email protected] (G. Cui).

https://doi.org/10.1016/j.econmod.2019.09.035
Received 8 July 2019; Received in revised form 22 September 2019; Accepted 25 September 2019
Available online 26 September 2019
0264-9993/© 2019 Elsevier B.V. All rights reserved.
W. Liu et al. Economic Modelling 86 (2020) 299–305

their respective blockchains. In addition, Corbet et al. (2018) also find From the double sorting results, we further find the momentum effect
evidence that the dynamics of cryptocurrencies are relatively isolated to is more significant for small cryptocurrencies, which makes a gap of
a variety of other financial assets, indicating that it is practically 50.64 percent per week with a t-statistic of 4.42 for buying a small winner
impossible to construct effective cryptocurrency factors based on the portfolio and selling a small loser portfolio. The SML and WML factors in
external information from other types of financial markets. this paper are constructed through the double sorting results, and then
In the light of the research above, this paper focuses on the factors their correlation can be reduced to 0.66. Implementing the time-series
that are specific to cryptocurrency markets. We note Sovbetov (2018) regressions to the 78 cryptocurrencies on the three factors, it is shown
find some factors, such as market beta, trading volume, and volatility, are from the significance of the intercepts that 93.99% cryptocurrencies in
significantly related to the prices of the most common five crypto- our sample period can be well explained through our three-factor model.
currencies by Autoregressive Distributed Lag cointegration framework. When we turn to the size-momentum portfolios, they generally have
Unlike their approach, the method we use is inspired by Fama and French significant exposures to the three factors with insignificant intercepts,
(1993, 2012, 2015) to construct the cryptocurrency factors, which is again demonstrating our three-factor model performs well in explaining
quite different from directly using the Fama-French stock factors that can the common variation in cryptocurrency returns.
be obtained from Kenneth French’s website. Since there is no financial The rest of paper is organized as follows. Section 2 describes the data
data supporting the intrinsic value of cryptocurrencies, we finally obtain used in our analyses and the anomaly variables we use to predict returns.
three factors, market return, SMB (small minus big), and WML (winner Section 3 studies the average returns from sorts and cross-sectional re-
minus loser), that can be considered to analogically construct crypto- gressions on the anomaly variables. In Section 4, our three-factor pricing
currency factors, based on a collection of all the well-recorded crypto- model is defined and the results of a set of time-series regression tests are
currencies (78 complete time series described in Section 2), instead of discussed. Section 5 discusses the robustness check. Section 6 concludes
only several prominent ones, so that the proposed factors can be well the paper.
defined and the split-sample portfolio analysis become possible.
Anomalies and factors are treated different in this paper, where the 2. Data and variables
anomalies are variables which differ from cryptocurrencies, such as
market beta, size and momentum, while the corresponding factors We obtain daily cryptocurrency prices, dollar volume, and market
(market return, SML, and WML) are common for cryptocurrencies and capitalization from https://coinmarketcap.com/. The sample period is
their loadings determine cryptocurrency return attributions. The poten- from 07-Aug-15 to 31-Dec-18 with total 1243 trading days, where 07-
tial economic intuitions behind the three factors can be similar with those Aug-15 is the starting date of Ethereum trading. Similar starting date is
in stock markets. The market factor can be defined via the value- also used in Liu (2019). We use cryptocurrencies with a complete time
weighted cryptocurrency portfolio returns, which is based on the series at the end of 2018, leaving a data set consisting of 78 crypto-
consideration that the cryptocurrencies may have some co-movements currencies. The risk-free rate is defined by the one-month U.S. Treasury
that formulate systematic risk across cryptocurrencies, similar to those bill rate from the Center of Research in Security Prices.
described in the CAPM model; The SML factor suggests that small cryp- We examine the patterns of average returns in a weekly horizon. The
tocurrencies have higher average returns mainly because they are less weekly returns and corresponding anomaly variables are converted from
liquid and the small coin holders require higher returns for accepting the daily data of each cryptocurrency (from Tuesday close price to
liquidity risk (Amihud and Mendelson, 1986; Liu, 2006). The WML factor Tuesday close price in next week), which finally leads to 178-week ob-
assumes buying past winners and selling past losers realize abnormal servations. We also carried out the robustness check of daily data in
returns, which might be caused by the delayed price reactions to Section 5 and the main findings remain the same.
cryptomarket-related information, as those also discussed in Jegadeesh Our asset pricing tests are based on two different methods. The first
and Titman (1993), Carhart (1997), and Grobys and Sapkota (2019). method is conducted by sorting the anomaly variables and the cross-
As anomaly variables are known to cause problems for the factor sectional regression of Fama and MacBeth (1973). The second method
models, it is reasonable to analyze the anomalies before the factors. We uses the time-series regression approach on the factor models as Fama
use two steps to identify anomalies. First, we sort the cryptocurrency and French (1993, 2012, 2015), where the common risk factors are
returns on anomaly variables, i.e. market beta, size, and momentum. usually formed by excess returns that are related to the anomaly vari-
When sorting on market beta, the resulting portfolio return differences ables. In this section, we introduce the anomaly variables of crypto-
are not statistically significant. However, when sorting on size and mo- currencies, while we leave the factor definitions and factor models to be
mentum, it shows significant anomalous returns that decrease with size discussed in Section 4.
and increase with return momentum, respectively. The long-short port- We measure the anomaly variables used to forecast returns in a
folio produces an average weekly return of 35.87 percent with a t-statistic rolling-window scheme. In other words, we use available information at
of 3.83 for the lowest level of size to the highest level of size and 36.26 time t to forecast the returns in t to tþ1, where the time tþ1 means one
percent with a t-statistic of 3.89 for the highest level of momentum to the week after time t in this paper. Here we introduce three variables that
lowest level of momentum. We also confirm the significant negative could be potential anomalies.
relation between size and future returns as well as the significant positive The first one is market beta. We define the cryptocurrency market
relation between momentum and future returns by using Fama-MacBeth beta as the slope in the regression of a cryptocurrency’s return on the
regression. market return, where the market return is the return on the value-
The above results have some interactions with the findings in Liu weighted cryptocurrency market portfolio. We use 52 week data
et al. (2019). Their paper also considered some other (nearly 1 year) preceding time t to estimate the betas in a rolling-window.
cryptomarket-related factors such as volatility and trading volume, while Then, the estimated betas can be treated as proxies for the real ones at
they are not so significant or generally explained by the size and mo- time t, although might suffer from estimation errors, to forecast the
mentum factors. Unlike their factors that are defined directly by anom- returns in t to tþ1 in the sorting and cross-sectional regression analyses,
alies or via the one-dimensional sorting results, our factors are defined as we will see these results in Section 3.
different from anomalies and are based on the double sorting results, The second one is market capitalization (size). The definition of the
following the methods by Fama and French (2012, 2015), so that the size variable is much easier than that of market betas, since it can be
factor correlation can be reduced and additional results are further precisely measured for each individual cryptocurrency. We define the
investigated. We then use the time-series regression approach to study market capitalization as the natural log of price times circulating supply
how cryptocurrencies and their portfolios are priced through the pro- at time t to be another anomaly variable.
posed three-factor model. The third one is momentum. We use one-year momentum in

300
W. Liu et al. Economic Modelling 86 (2020) 299–305

Table 1
Average weekly percent excess returns for cryptocurrency portfolios formed on one-dimensional sorts on beta, size and momentum: 07-Aug-15 to 31-Dec-18, 178 weeks.
At the end of each Tuesday, cryptocurrencies are allocated to six groups based on quintiles (each group has 13 cryptocurrencies), using beta, size and momentum,
respectively. In the sort time t, beta means the market beta estimated by the data from t-52 to t; size is the market cap at time t; momentum is the cumulative return from
t-52 to t. Then, the equal-weighted and value-weighted returns are calculated from t to tþ1. Alpha is the intercept obtained from regression of the portfolio return on
market return.
Equal-weighted Value-weighted

Panel A: Sorted on beta

Low 2 3 4 5 High Low 2 3 4 5 High
Return 15.75 5.82 7.18 6.42 9.43 26.95 2.36 4.65 3.23 3.35 2.75 3.68 (1.84)
(2.61) (3.38) (3.51) (3.67) (4.40) (3.55) (2.18) (2.35) (2.03) (1.82) (1.71)
Alpha 11.86 3.37 4.49 3.95 6.63 25.54. 0.91 2.69 0.95 0.95 0.70 0.82 (0.59)
(2.04) (2.78) (2.90) (3.19) (4.07) (3.32) (1.13) (1.45) (0.86) (0.68) (0.57)
Panel B: Sorted on size
Low 2 3 4 5 High Low 2 3 4 5 High
Return 38.67 15.22 6.72 4.87 3.27 2.80 (1.76) 24.38 14.48 6.61 4.84 2.55 2.12 (1.63)
(3.94) (6.23) (3.46) (2.91) (2.26) (6.57) (6.48) (3.43) (2.86) (1.95)
Alpha 35.40 12.82 3.98 2.49 0.89 0.28 (0.29) 21.43 12.10 3.89 2.46 0.06 0.01
(3.59) (5.95) (2.89) (2.12) (1.12) (6.21) (6.32) (2.86) (2.05) (0.07) (0.72)
Panel C: Sorted on momentum
Low 2 3 4 5 High Low 2 3 4 5 High
Return 4.42 4.43 6.20 6.96 8.86 40.68 3.01 2.99 3.10 3.95 2.89 7.88 (3.24)
(2.92) (2.98) (3.64) (3.48) (3.76) (4.19) (1.86) (2.05) (1.72) (2.05) (1.61)
Alpha 2.02 2.12 3.61 4.29 5.98 37.84 0.95 0.58 0.70 1.32 0.35 5.91 (2.62)
(2.18) (2.38) (3.30) (2.87) (3.18) (3.87) (0.76) (0.72) (0.52) (0.94) (0.28)

cryptocurrency returns, i.e., the cumulative return from t-52 to t for each returns for portfolios formed on market beta, size and momentum, in
cryptocurrency. The reason for choosing one-year horizon to define Panel A, B and C, respectively. The two kinds of weighted average returns
momentum can be attributed to the study of Jegadeesh and Titman both show a decreasing pattern on size and an increasing pattern on
(1993), Carhart (1997), and Grobys and Sapkota (2019). momentum. The return for cryptocurrencies with the lowest level of size
is 38.67 percent for equal-weighted portfolios and 24.38 percent for the
3. Cross-section of expected returns value-weighted portfolios, while the return for cryptocurrencies with the
highest level of size is 2.80 percent for equal-weighted and 2.12 percent
In this section, we first report the characteristics of the portfolios for value-weighted portfolios. Similarly, the return for cryptocurrencies
formed by sorting cryptocurrencies into groups based on the anomaly with the highest level of momentum is 40.68 percent for equal-weighted
variables introduced in Section 2. Then the interaction between any two portfolios and 7.88 percent for the value-weighted portfolios, while the
of the anomaly variables using double sorts are studied. Finally, we carry return for cryptocurrencies with the lowest level of momentum is 4.42
out Fama-MacBeth regressions to analyze the cross-section of crypto- percent for value-weighted and 3.01 percent for value-weighted portfo-
currency returns on the anomaly variables. lios. However, there is no obvious relationship between market beta and
average return. It seems a U shape for the equal-weighted case, and
3.1. One-dimensional sorts relatively flat for the case of value-weighted case. The insignificant t-
statistic values for most of market beta portfolios in the value-weighted
Each Tuesday, 78 cryptocurrencies are ranked into 6 groups with case indicate that it lacks strong explanatory power for subsequent
equal group size of 13 cryptocurrencies according to their market value, returns.
size, and momentum, respectively. Then, we construct equal-weighted We also assess the empirical relation between cryptocurrency returns
and value-weighted portfolios by the returns in the following week. and anomaly variables by adjusting for standard measures of risk from
Notice that the anomaly variables should fall behind the average returns Table 1. The alphas presented in table are the intercepts from regressions
with lag 1 to make sure the information we use is known. of the portfolio returns on the cryptocurrency market return. Note that
Table 1 reports the equal-weighted and value-weighted weekly the alphas are large and statistically significant for the lowest level of size

Table 2
Average weekly percent excess returns for cryptocurrency portfolios formed on two-pass sorts on size and momentum, size and beta, momentum and beta: 07-Aug-15 to
31-Dec-18, 178 weeks. At the end of each Tuesday, cryptocurrencies are allocated to three groups based on tertiles of the first variable in the first step, and then each
group is further divided into three subgroups based on tertiles of the second variable. The definitions of beta, size and momentum are the same as Table 1.
Equal-weighted Value-weighted

Panel A: Size-Momentum portfolios

Loser Neutral Winner Loser Neutral Winner
Small 9.17 (4.88) 15.51 (5.04) 59.81 (3.91) Small 8.86 (4.57) 11.24 (4.35) 39.07 (7.94)
Neutral 4.30 (2.77) 5.75 (3.01) 7.53 (3.51) Neutral 4.01 (2.72) 5.55 (2.96) 6.89 (3.14)
Big 2.97 (2.05) 3.18 (2.12) 2.80 (1.62) Big 3.45 (2.15) 2.71 (1.60) 3.16 (1.73)
Panel B: Size-Beta portfolios
Low Neutral High Low Neutral High
Small 22.56 (2.61) 16.03 (6.07) 44.16 (3.58) Small 9.27 (4.55) 14.68 (5.88) 25.54 (6.69)
Neutral 6.08 (3.60) 4.63 (2.65) 6.77 (3.03) Neutral 6.06 (3.60) 4.03 (2.41) 6.16 (2.74)
Big 2.37 (1.81) 3.06 (1.98) 3.76 (2.04) Big 2.54 (2.13) 3.86 (2.59) 1.82 (1.29)
Panel C: Momentum-Beta portfolios
Low Neutral High Low Neutral High
Winner 21.86 (2.54) 13.13 (5.35) 41.14 (3.35) Winner 2.63 (1.54) 4.89 (2.39) 8.53 (3.43)
Neutral 6.00 (3.69) 7.45 (3.32) 6.24 (3.30) Neutral 3.08 (1.93) 2.72 (1.33) 4.04 (2.00)
Loser 4.19 (2.89) 4.48 (2.89) 4.61 (2.92) Loser 4.51 (2.36) 2.47 (1.82) 2.56 (1.67)

301
W. Liu et al. Economic Modelling 86 (2020) 299–305

Table 3
Average coefficients from week-by-week cross-sectional regressions of cryptocurrency returns on beta, size and momentum: 07-Aug-15 to 31-Dec-18, 178 weeks. The
average slope is the time-series average of the weekly regression slopes, and the t-statistic is the average slope divided by its time-series standard error. The variables are
the same as Table 1.
(1) (2) (3) (4) (5) (6) (7)

Intercept 10.94 (3.96) 76.13 (3.63) 0.81 (0.25) 66.11 (3.90) 3.82 (1.05) 26.53 (1.87) 24.08 (2.02)
Market beta 0.05 (0.02) 0.14 (0.07) 1.46 (0.75) 1.38 (0.70)
Log(size) 4.01 (3.51) 3.48 (3.50) 1.64 (2.05) 1.66 (2.17)
Momentum 0.017 (3.46) 0.020 (4.86) 0.015 (3.36) 0.017 (5.39)

R2 0.044 0.035 0.109 0.077 0.193 0.132 0.214

Adjusted R2 0.031 0.022 0.097 0.052 0.171 0.108 0.182

as well as the highest level of momentum, while there are no significant cross-sectional regression models, where the last column represents the
alphas for any of the market beta portfolios in the value-weighted case. full model (1). The first column presents the results of the regression of
Hence, we may still conclude that the size and momentum are reliable cryptocurrency return on market beta. The coefficient associated with
indicators of subsequent one-week cryptocurrency returns in cross- market beta is 0.05 with a t-statistic of 0.02. This confirms that there
section aspect. does not seem to be a significant relation between cryptocurrency returns
and market betas. The second and third columns confirm the relation
3.2. Double sorts between the cryptocurrency return and size and momentum, respec-
tively. In column 2, the coefficient associated with log size is 4.01 with
We further analyze the interaction between size and momentum, size a t-statistic of 3.51. Similarly, in column 3, the coefficient on mo-
and beta, as well as momentum and beta, respectively. Taking the pair of mentum is 0.017 with a t-statistic of 3.46. In column 4–6, we report
size and momentum as an example, the double sort is conducted as fol- regression results to any two of the anomaly variables. We can observe
lows: Cryptocurrencies are first ranked into 3 groups on size, namely that the coefficients of beta and log size have some changes after adding
small, neutral, and big portfolios. Then, each group is further divided into the momentum term, but the corresponding statistical significance re-
3 subgroups (winner, neutral and loser) on momentum. Finally, we can mains unchanged. In the last column, we use all the anomaly variables
obtain the 3  3 equal-weighted and value-weighted portfolios from two- simultaneously. The coefficients on log size and momentum are again
pass sorts of cryptocurrencies on size and momentum. significantly negative and positive, respectively. The results displayed in
Panel A of Table 2 shows weighted average returns for portfolios Table 3 are consistent with the results in Tables 1 and 2.
formed on size and momentum. We can observe that, in each column of
Panel A, average returns largely decrease from small cryptocurrencies to 4. The three-factor model
big cryptocurrencies, showing a significant size effect. In each row of
Panel A, the momentum effect of cryptocurrencies is also quite clear for In this section, we first introduce the definitions of the factors and the
small and neutral groups of cryptocurrencies, while the big crypto- design of the factor model. Then the proposed factors are used to
currencies is the only exception. According to the results of Panel A, the construct three-factor model to explain the average returns of crypto-
equal-weighted premium of 56.84 percent and value-weighted premium currencies. Finally, we test the performance of the factor model on the
of 35.62 percent can be earned by buying small winner cryptocurrencies cryptocurrency portfolios formed on size and momentum.
and selling big loser cryptocurrencies.
Panel B and Panel C illustrate similar information. In each column of 4.1. Factor definitions
Panel B, average returns fall from small cryptocurrencies to big crypto-
currencies, similar to Panel A. In each row of Panel C, the momentum This paper focuses on three factors that are related to the anomaly
effect also exhibits in different level of market beta. Comparing the re- variables mentioned in Section 3. The first factor is the market excess
sults of Panel B and Panel C, we can find that the size effect is more return, which has been described before. Here we introduce the con-
consistent than momentum effect in the double sorting portfolios. In each struction process of size and momentum factors.
column of Panel B and Panel C, we can find the structure of market beta is As we have stated in Section 3, we use two-pass sorts instead of in-
not obvious for any of size-beta portfolios or momentum-beta portfolios, dependent sorts to construct the 3  3 size and momentum portfolios.
which confirms the findings in Table 1 that the market beta does not have That is mainly because we have in total 78 cryptocurrencies, and the two-
explanatory power for average returns in cross-section aspect. pass sorts can make sure each subgroup has nearly the same number of
cryptocurrencies, while the independent sorts define the portfolios by
3.3. Fama-MacBeth regressions intersections which might lead to imbalance on the number of each
subgroup. Moreover, we use 3  3 instead of 4  4 or 5  5 so that each
To further evaluate the relation between future returns and anomaly subgroup has a certain number of cryptocurrencies. Notice that the two-
variables, we carry out various cross-sectional regressions using the pass sort results may be different if we change the order of the first sort
method proposed in Fama and MacBech (1973). Each week t, we and the second sort, i.e., small winner (SW) is usually distinct from
compute the anomaly variables for cryptocurrency i and estimate the winner small (WS), same for the others. Therefore, we define our SMB
following cross-sectional regression: and WML as follows:

ri;tþ1 ¼ γ 0;t þ γ 1;t βi;t þ γ 2;t logðsizeit Þ þ γ 3;t momentumi;t þ εi;tþ1 ; (1) SMBW ¼ SW  BW; SMBN ¼ SN  BN; SMBL ¼ SL  BL; (2)

where ri;tþ1 is the weekly return (in percent) of the i-th cryptocurrency for WMLS ¼ WS  LS; WMLN ¼ WN  LN; WMLB ¼ WB  LB; (3)
week t þ 1, βi;t is the market beta of the i-th cryptocurrency estimated via
52 week data (nearly 1 year) preceding time t, logðsizeit Þ is the log of SMB ¼ ðSMBW þ SMBN þ SMBL Þ=3; (4)
market cap of the i-th cryptocurrency at time t, and momentumi;t is the
lagged return in the past one year of the i-th cryptocurrency. WML ¼ ðWMLS þ WMLN þ WMLB Þ=3; (5)
Table 3 reports the time series average of the coefficients for seven

302
W. Liu et al. Economic Modelling 86 (2020) 299–305

Table 4
Summary statistics for weekly factor percent returns: 07-Aug-15 to 31-Dec-18, 178 weeks. At the end of each Tuesday, cryptocurrencies are assigned to 3  3 subgroups
through a two-pass sorts described in Table 2. In each subgroup, the return of the portfolio is based on value-weighted. SMBW is the average of return on the portfolios of
small winner cryptocurrencies minus the average return on the portfolios of big winner cryptocurrencies, SMBN and SMBL are the same but for portfolios of neutral and
loser cryptocurrencies. WMLS, WMLN, and WMLB are defined in the same way. SMB is the average of SMBW, SMBN, and SMBL. WML is the average of WMLS, WMLN, and
WMLB. Panel A, Panel C and Panel D show average weekly returns, the standard deviations and t-statistics of the average returns. Panel B reports the correlation matrix
of the three factors.
Panel A Panel B

RM -RF SMB WML RM -RF SMB WML

Mean 2.12 16.61 10.94 RM -RF 1.0000

Std dev. 12.27 21.93 22.62 SMB 0.0027 1.0000 0.6606
t-statistic 1.96 8.57 5.47 WML 0.0074 0.6606 1.0000
Panel C Panel D
SMBW SMBN SMBL WMLS WMLN WMLB
Mean 35.91 8.52 5.41 Mean 30.21 2.88 0.29
Std dev. 54.72 25.50 20.07 Std dev. 56.02 18.22 22.84
t-statistic 7.42 3.78 3.05 t-statistic 6.10 1.79 0.14

4.2. Explain average returns of cryptocurrencies

Table 5
Using three factors in regressions to explain average returns on three biggest,
We now use the returns of 78 cryptocurrencies to test the explanatory
three medium, and three smallest cryptocurrencies: 07-Aug-15 to 31-Dec-18, 178
weeks. RM -RF is the value-weighted return on the cryptocurrency market port-
ability of the three-factor model. If an asset pricing model completely
folio; SMB (small minus big) is the size factor of cryptocurrencies; WML (winner captures expected returns, the intercept is indistinguishable from zero.
minus loser) is the momentum factor of cryptocurrencies. The detailed factor To save space, we only exhibit the cases of three biggest, three medium,
definitions are described in Table 4. and three smallest cryptocurrencies in Table 5.
Int RM -RF SMB WML R2 Adj.
Panel A of Table 5 reports the estimate results for three biggest
R2 cryptocurrencies: Bitcoin, Ethereum, and XRP. We can see the intercepts
of these three cryptocurrencies are all insignificant, demonstrating that
Panel A: 3 biggest cryptocurrencies
Bitcoin 0.03 0.91 0.05 0.06 0.90 0.89 the three-factor model explains the three cryptocurrencies returns very
(0.07) (32.79) (2.25) (3.16) well. The exposures of these big cryptocurrencies are mainly on the
Ethereum 0.26 1.13 0.03 0.11 0.56 0.55 market excess return factor, while Bitcoin has significant exposures to all
(0.19) (12.40) (0.50) (1.62) of the three factors with adjusted R2 at 0.89, which indicates that the
XRP 5.44 1.32 0.10 0.06 0.25 0.23
(1.66) (6.26) (0.63) (0.36)
three-factor model fits the returns of Bitcoin very well.
Panel B: 3 medium cryptocurrencies Panel B of Table 5 shows the results of three medium cryptocurren-
BlackCoin 0.84 1.06 0.13 0.04 0.37 0.36 cies: BlackCoin, Potcoin, and MonetaryUnit. From the t-statistics of the
(0.43) (8.47) (1.41) (0.48) three cryptocurrencies we can conclude that the three-factor model also
Potcoin 2.49 1.04 0.26 0.33 0.18 0.16
explains their average returns. The BlackCoin and Potcoin have signifi-
(0.73) (4.73) (1.60) (2.05)
MonetaryUnit 7.51 0.94 0.71 0.65 0.05 0.03 cant exposures to the market factor, while the MonetaryUnit has signif-
(0.93) (1.81) (1.82) (1.71) icant exposure to the size and momentum factor. Unlike the biggest
Panel C: 3 smallest cryptocurrencies cryptocurrencies, the exposures of the medium cryptocurrencies are
Bata 0.15 1.34 0.76 0.32 0.18 0.16 diversified.
(0.03) (4.04) (3.09) (1.33)
TEKcoin 27.82 0.82 0.22 0.66 0.07 0.05
Panel C of Table 5 provides the results for three smallest
(3.13) (1.43) (0.52) (1.60)
AnarchistsPrime 27.82 2.29 21.29 11.90 0.07 0.05
(0.19) (0.24) (2.98) (1.72) Table 6
Tests of three-factor model for weekly value-weighted returns on cryptocurrency
portfolios from 3  3 sorts on size and momentum: 07-Aug-15 to 31-Dec-18, 178
where the return of each subgroup is value-weighted so that the defined weeks. The regressions use the three-factor model to explain the excess return
factors can be more robust to the outliers. Finally, our factor asset-pricing cryptocurrency portfolios formed on size and momentum. Panel A shows three-
factor intercepts and their t-statistics. Panel B, Panel C, and Panel D report the
model for cryptocurrencies can be written as
slopes for RM -RF, SMB, and WML, and t-statistics for these coefficients. R(t)-
Ri ðtÞ  RF ðtÞ ¼ ai þ bi ½RM ðtÞ  RF ðtÞ þ si SMBðtÞ þ wi WMLðtÞ þ ei ðtÞ: (6) RF(t) ¼ a þ b[RM(t)- RF(t)] þsSMB(t)þwWML(t)þe(t).
Size\Momentum Loser Neutral Winner Loser Neutral Winner
Table 4 shows summary statistics for factor returns. From the results
in Panel A, we can observe the three factors all have significant positive Panel A: a t(a)
Small 1.68 1.24 3.15 1.10 0.55 1.44
average returns, where the SMB enjoys the highest return that is 16.61 Neutral 0.87 1.83 2.49 0.79 1.25 1.23
percent per week with a t-statistic of 8.57. WML is also very significant at Big 1.44 0.55 1.59 1.01 0.40 1.04
10.94 percent per week with a t-statistic of 5.47. The market excess re- Panel B: b t(b)
turn is only 2.12 percent per week with a t-statistic of 1.96. Panel B shows Small 1.12 1.21 0.93 11.42 8.28 6.59
Neutral 1.05 1.27 1.15 14.75 13.40 8.86
the correlation matrix of the three factors. We can find that SMB and
Big 0.99 1.19 1.08 10.65 13.40 10.95
WML are nearly uncorrelated to the market excess return with correlation Panel C: s t(s)
0.0027 and 0.0074 respectively, while there is a certain degree of cor- Small 0.62 0.93 1.16 8.46 8.59 11.04
relation between SMB and WML at 0.6606. Panel C and Panel D exhibit Neutral 0.27 0.33 0.28 5.04 4.66 2.85
three versions of SMB and WML defined in (2) and (3). The results pre- Big 0.18 0.10 0.37 2.60 1.48 5.03
Panel D: w t(w)
sented in Panel C and Panel D are consistent with the double sorting Small 0.50 0.51 1.34 7.09 4.83 13.16
results in Table 2. Neutral 0.16 0.07 0.21 3.18 1.05 2.27
Big 0.28 0.11 0.50 4.20 1.78 6.96

303
W. Liu et al. Economic Modelling 86 (2020) 299–305

Table 7 the positive or negative value of the intercepts for different portfolios.
Interprets from nested factor models in regressions to explain weekly value- Nevertheless, the t-statistics demonstrate the intercepts are all econom-
weighted returns on cryptocurrency portfolios formed on size and momentum: ically and statistically close to zero. The highest one is the small winner
07-Aug-15 to 31-Dec-18, 178 weeks. Panel A shows one-factor intercepts and portfolio, whose intercept is 3.15 with a t-statistic of 1.44, still far less
their t-statistics produced by the cryptocurrency market return. Panel B and Panel than the rejection threshold. In short, the three-factor model works well
C show two-factor intercepts and their t-statistics, with adding SMB and WML to
for the size-momentum portfolios.
the model of Panel A, respectively. Panel D provides the intercepts and their t-
Panel B to Panel D of Table 6 show the slopes and t-statistics for the
statistics of the proposed three-factor model.
three factors. It is observed that the market factor slopes are always close
Size\Momentum Loser Neutral Winner Loser Neutral Winner
to 1 and are all strongly significant. The SMB slopes are strongly positive
Panel A: R(t)-RF(t) ¼ a þ b[RM(t)- RF(t)]þe(t) for small cryptocurrencies and slightly negative for big cryptocurrencies.
a t(a) The WML slopes are significantly positive for winner cryptocurrencies
Small 6.47 8.68 37.04 4.21 3.83 7.56
and also significantly negative for loser cryptocurrencies. From the t-
Neutral 1.78 2.85 4.43 1.85 2.20 2.42
Big 1.35 0.17 0.86 1.11 0.15 0.60 statistics of the three slopes, we observe that all the 9 portfolios have
Panel B: R(t)-RF(t) ¼ a þ b[RM(t)- RF(t)]þsSMB(t)þe(t) significant exposures to the three factors, so that these three factors play a
a t(a) very important role in pricing cryptocurrencies.
Small 1.88 1.03 3.61 1.10 0.42 1.77
Table 7 shows matrices of the intercepts and their t-statistics for
Neutral 0.80 1.80 2.57 0.70 1.23 1.26
Big 1.56 0.55 1.39 1.02 0.39 0.77
nested models. Panel A is the CAPM type model for cryptocurrencies. The
Panel C: R(t)-RF(t) ¼ a þ b[RM(t)- RF(t)]þwWML(t)þe(t) small winner portfolio’s intercept is still very large at 37.04 with a t-
a t(a) statistic of 7.56. Obviously, it lacks of power to explain the average
Small 7.64 7.72 14.30 4.52 3.07 5.26 returns for either small cryptocurrencies or winner cryptocurrencies.
Neutral 1.70 1.33 0.17 1.59 0.95 0.09
Panel B and Panel C add the SMB and WML factors to the model of Panel
Big 3.17 0.39 1.96 2.43 0.32 1.33
Panel D: R(t)-RF(t) ¼ a þ b[RM(t)- RF(t)] þsSMB(t)þwWML(t)þe(t) A separately. While both of them have improved the explanatory power,
a t(a) the t-statistics reduced much more for adding SMB factor. As a conse-
Small 1.68 1.24 3.15 1.04 0.55 1.44 quence, only small winner cryptocurrencies cannot be explained by the
Neutral 0.87 1.83 2.49 0.79 1.25 1.23
model of Panel B, while most of the small cryptocurrencies cannot be
Big 1.44 0.51 1.59 1.01 0.36 1.04
explained by the model of Panel C. Panel D is the full three-factor model,
which is exactly the same results we presented in Panel A of Table 6.
cryptocurrencies: Bata, TEKcoin, and AnarchistsPrime. We can find that Thus, we can conclude the three-factor model performs well in explaining
Bata and AnarchistsPrime are well priced by the three factor model, and the common variation in cryptocurrency returns.
the exposures of these small cryptocurrencies tend to be located on the
size and momentum factors. However, we note that the intercepts of 5. Robustness analysis
TEKcoin is significantly bounded away from zero and there are no sig-
nificant exposures to any of the three factors, which means the three- In this section, we examine the results under a daily scheme as a
factor model fails to explain the returns of TEKcoin. robustness check kindly suggested by a referee. The raw daily data is
Beyond the above reported 9 cryptocurrencies, we have implemented from 07-Aug-15 to 31-Dec-18 with total 1243 trading days. While
the regressions to all the 78 cryptocurrencies and obtained the corre- monthly or weekly data are more frequently adopted for analyzing the
sponding intercepts and t-statistics. Among the 78 cryptocurrencies, long-short strategies in real applications, using daily frequency data
there are only 5 of them rejecting the null hypothesis of intercept being would significantly enlarge the sample size, as the development of
zero at 0.05 significant level. In other words, 93.99% cryptocurrencies in cryptocurrency markets is still in its infancy.
our sample period can be well explained through our three-factor model. Panel A of Table 8 shows the average daily percent excess returns for
cryptocurrency portfolios via one-dimensional sorts on market beta, size,
4.3. Asset pricing tests for size-momentum portfolios and momentum, respectively. Similar to the results in Table 1, the
average returns show a decreasing pattern on size and an increasing
To access the model performance, we examine the regressions on the pattern on momentum, while there is no obvious pattern on market beta.
portfolios from 3  3 sorts on size and momentum. The regression details It can be observed that the returns of the long-short zero-investment
are reported in Table 6, including the intercepts and pertinent slopes of strategies are significant for size and momentum, but not significant for
the three-factor model. market beta. Panel B of Table 8 shows the average slopes from day-by-
Panel A of Table 6 shows intercepts from the three-factor regressions day cross-sectional regressions under the Fama-MacBeth method pre-
for the 3  3 size-momentum portfolios. There is no obvious pattern of sented in equation (1). It can be found that no matter the whole model or

Table 8
One-dimensional sorts and Fama-MacBeth regressions using daily data: 07-Aug-15 to 31-Dec-18, 1243 days. Pane A reports the average daily percent excess returns for
cryptocurrency portfolios formed on beta, size, and momentum, respectively. Panel B shows the average slopes from day-by-day cross-sectional regressions of cryp-
tocurrency returns on beta, size and momentum.
Low 2 3 4 5 High High-Low

Panel A: One-dimensional Sorts

Beta 3.22 (3.49) 1.02 (4.88) 1.94 (1.79) 1.16 (4.70) 1.08 (4.69) 1.94 (5.90) 1.26 (-1.32)
Size 6.38 (4.47) ‘1.88 (6.75) 0.68 (3.09) 0.57 (2.81) 0.45 (2.67) 0.41 (1.99) 5.97 (-4.18)
Momentum 0.76 (3.31) 0.63 (3.39) 0.89 (4.18) 1.18 (4.52) 1.15 (4.26) 5.75 (4.94) 4.98 (3.51)
Panel B: Fama-MacBeth Regressions
(1) (2) (3) (4) (5) (6) (7)
Intercept 2.85 (2.27) 12.73 (3.64) 0.08 (0.15) 12.67 (3.50) 1.10 (0.70) 6.08 (2.46) 7.59 (2.46)
Market beta 1.25 (0.93) 1.76 (1.28) 1.21 (0.72) 1.35 (0.81)
Log(size) 0.69 (3.52) 0.59 (3.28) 0.36 (2.59) 0.38 (2.86)
Momentum 0.0023 (3.11) 0.0025 (3.24) 0.0018 (2.85) 0.019 (3.42)
R2 0.042 0.018 0.075 0.059 0.115 0.091 0.130
Adjusted R2 0.029 0.005 0.063 0.034 0.091 0.066 0.094

304
W. Liu et al. Economic Modelling 86 (2020) 299–305

Table 9 potential candidates in factor models for cryptocurrencies. Our analysis

Double sorts and factor model regressions using daily data: 07-Aug-15 to 31-Dec- could serve as a benchmark for identifying the additional factors.
18, 1243 days. Panel A reports the average daily percent excess returns for
cryptocurrency portfolios formed on two-pass sorts on size and momentum. Acknowledgement
Panel B provides the estimation results of the three-factor model in regressions
for daily value-weighted returns on cryptocurrency portfolios formed on size and
We would like to thank the editor and two anonymous referees for the
momentum.
very helpful comments received. We also appreciate Dr. Tao Zou from the
Size\Momentum Loser Neutral Winner Loser Neutral Winner Australian National University for his suggestions during our research.
Panel A: Double sorts This work is supported by grants from the National Natural Science
μ t(μ) Foundation of China (No. 71601132) and Foundation for the Excellent
Small 1.26 2.62 9.05 3.82 5.63 6.32
Talents of Beijing (No. 2016000020124G083). Data used in the paper are
Neutral 0.66 0.93 0.25 3.10 3.54 2.03
Big 0.46 0.47 0.36 2.55 2.04 1.55 available upon request from authors. We have no competing interests.
Panel B: Factor model regressions Unfortunately, our friend Guowei Cui passed away on Jun 21, 2019. He
a t(a) was an extraordinary researcher and a generous friend with a warm
Small 0.12 0.59 0.15 0.81 1.43 0.75 heart. He will be greatly missed.
Neutral 0.19 0.24 0.07 1.53 1.32 0.35
Big 0.96 0.51 0.30 0.81 1.41 1.85
b t(b) References
Small 1.09 1.03 0.92 31.55 18.32 20.11
Neutral 1.01 0.99 1.07 34.75 23.44 22.07 Amihud, Y., Mendelson, H., 1986. Asset pricing and the bid-ask spread. J. Financ. Econ.
Big 0.90 1.12 1.02 50.91 23.22 27.31 17 (2), 223–249.
s t(s) Bhambhwani, S., Delikouras, S., Korniotis, G., 2019. Do fundamentals drive
Small 0.45 0.71 1.15 17.88 17.11 34.20 cryptocurrency prices? SSRN Working. Paper NO. 3342842.
Brauneis, A., Mestel, R., 2018. Price discovery of cryptocurrencies: Bitcoin and beyond.
Neutral 0.12 0.15 0.01 5.72 4.87 0.28
Econ. Lett. 165, 58–61.
Big 0.09 0.32 0.47 7.04 8.94 17.20
Campbell, J.Y., Shiller, R.J., 1988. Stock prices, earnings, and expected dividends.
w t(w)
J. Financ. 43 (3), 661–767.
Small 0.38 0.41 1.48 14.53 9.58 42.21 Carhart, M.M., 1997. On persistence in mutual fund performance. J. Financ. 52 (1),
Neutral 0.14 0.01 0.29 6.30 0.28 7.71 57–82.
Big 0.13 0.24 0.58 9.56 6.37 20.18 Chan, S., Chu, J., Nadarajah, S., Osterrieder, J., 2017. A statistical analysis of
cryptocurrencies. Risk and Financial Management 10 (12), 1–23.
Corbet, S., Meegan, A., Larkin, C., Lucey, B., Yarovaya, L., 2018. Exploring the dynamic
the nested models, they all demonstrate the significance of size and relationships between cryptocurrencies and other financial assets. Econ. Lett. 165,
28–34.
momentum and the insignificance of market beta, which are consistent Fama, E.F., MacBeth, J., 1973. Risk, return and equilibrium: empirical tests. J. Political
with the findings in Table 3. Econ. 81, 607–636.
Panel A of Table 9 presents average daily excess returns for 3  3 Fama, E.F., French, K.R., 1993. Common risk factors in the returns on stocks and bonds.
J. Financ. Econ. 33 (1), 3–56.
portfolios formed on two-pass double sorts on size and momentum. From
Fama, E.F., French, K.R., 2012. Size, value, and momentum in international stock returns.
each column of Panel A, we can find obvious size effects in winner, J. Financ. Econ. 105 (3), 457–472.
neutral, and loser groups. On the other hand, for the rows of Panel A, it Fama, E.F., French, K.R., 2015. A five-factor asset pricing model. J. Financ. Econ. 116 (1),
reveals significant momentum effect especially for small cryptocurren- 1–22.
Fama, E.F., French, K.R., 2018. Choosing factors. J. Financ. Econ. 128 (2), 234–252.
cies. These findings are quite similar to that in Table 2. Panel B of Table 9 Gregoriou, A., 2019. Cryptocurrencies and asset pricing. Appl. Econ. Lett. 26 (12),
reports the estimation results by regressing the 3  3 portfolios on the 995–998.
three factors, respectively. From the t-statistics presented in Table 9, we Gkillas, K., Katsiampa, P., 2018. An application of extreme value theory to
cryptocurrencies. Econ. Lett. 164, 109–111.
can find that the portfolios generally have significant exposures to the Grobys, K., Sapkota, N., 2019. Cryptocurrencies and Momentum. Economics Letters
proposed three factors, while the regression intercepts are all insignifi- forthcoming.
cant, which again shows the three-factor model works well for the size- Harvey, C., Liu, Y., 2016. ...and the cross-section of expected returns. Rev. Financ. Stud.
29, 5–68.
momentum portfolios. Jegadeesh, N., Titman, S., 1993. Returns to buying winners and selling losers:
implications for stock market efficiency. J. Financ. 48, 35–91.
6. Conclusion Liu, W., 2006. A liquidity-augmented capital asset pricing model. J. Financ. Econ. 82 (3),
631–671.
Liu, W., 2019. Portfolio diversification across cryptocurrencies. Financ. Res. Lett. 29,
In this paper, three common factors specific to cryptocurrency mar- 200–205.
ket, i.e. the market factor and factors related to size and momentum, are Liu, Y., Tsyvinski, A., 2018. Risks and Returns of Cryptocurrency. NBER Working Paper
NO. 24877.
proposed to study the average returns of cryptocurrencies. To evaluate
Liu, Y., Tsyvinski, A., Wu, X., 2019. Common Risk Factors in Cryptocurrencies. NBER
the factor performance, we use both the methods of cross-sectional an- Working Paper NO. 25882.
alyses for the anomaly variables and time-series regressions for the risk Miller, M., Modigliani, F., 1961. Dividend policy, growth, and the valuation of shares.
factors. The results show the size effect and momentum effect are strong J. Bus. 34, 411–433.
Phillip, A., Chan, J., Peris, S., 2018. On long memory effects in the volatility measure of
in cryptocurrency markets, and the proposed three-factor model has cryptocurrencies. Financ. Res. Lett. 28, 95–100.
satisfied explanation ability on cryptocurrency returns. Sharpe, W.F., 1964. Capital asset prices: a theory of market equilibrium under conditions
On the other hand, the complexities of the cryptocurrency markets of risk. J. Financ. 19 (3), 425–442.
Sovbetov, Y., 2018. Factors influencing cryptocurrency prices: evidence from Bitcoin,
are far from fully explored. Although our model has well explained a Ethereum, dash, litcoin, and monero. Journal of Economics and Financial Analysis 2
large amount of cryptocurrencies’ returns, many new patterns beyond (2), 1–27.
those from stock markets could be discovered in future and become

305