SSRN Id4222864

Trend-based Forecast of Cryptocurrency Returns*
Xilong Tan†, Yubo Tao‡
April 16, 2023
Abstract
Cryptocurrencies are widely known for their limited publicly available information,
making it challenging to predict market returns. Technical analysis has emerged as an essential
tool in this context, but its effectiveness in the cryptocurrency market remains an open question.
Using data from nearly 3,000 cryptocurrencies at daily, weekly, and monthly horizons from
2013 to 2022, we systematically re-examine the efficacy of trend-based technical indicators in
predicting cryptocurrency market returns and find that price-based signals are more effective
in predicting short-term horizons, while volume-based signals are more powerful in predicting
long-term horizons. Further analysis shows that machine learning techniques can significantly
improve the performance of technical indicators, and technical indicators based on different
information respond differently to the COVID-19 outbreak. These results provide direct
evidence that volume imparts information to technical analysis independently of price.
JEL Classification: G12, G14, G17.
Keywords: Cryptocurrency, Return predictability, Technical analysis, Investment horizon,

Machine Learning, COVID-19.
* We express our gratitude to Sushanta Mallick (Co-editor) and two anonymous referees for providing their valuable
comments and suggestions, which have significantly enhanced the quality of this paper. We would like to thank Dashan
Huang and Guoshi Tong for sharing their codes to implement sSUFF, and the participants of The Second International
Conference on Digital Economy for their helpful discussions and feedback. Yubo Tao would like to acknowledge the
financial support provided by the Start-up Research Grant (No. SRG2022-00016-FSS) from the University of Macau.
† Department of Economics, Faculty of Social Sciences, University of Macau. Email:
ben.tanxilong@connect.um.edu.mo.
‡ Corresponding author. Department of Economics, Faculty of Social Sciences, and Asia-Pacific Academy of
Economics and Management (APAEM), University of Macau. Email: yubotao@um.edu.mo.
Electronic copy available at: https://ssrn.com/abstract=4222864

1 Introduction
Trend-based trading strategies (e.g., moving average and momentum) have been widely employed
by practitioners from every financial market. However, the effectiveness of these strategies is
under debate in academia. Very recently, Jiang, Kelly, and Xiu (2022) revisited the trend-based
predictability in the stock market using computer vision and motivated researchers and investors
to rethink technical analysis. In this paper, we systematically examine the efficacy of trend-based
technical indicators in the cryptocurrency (crypto for short, hereafter) market over various horizons
in the hope of shedding new light on the issue.
We performed three empirical tests to demonstrate the predictive power of trend-based

technical indicators on crypto market returns. Firstly, we tested the predictability of individual
technical indicators and machine learning methods using the full sample over varying horizons.
Secondly, to address potential overfitting concerns by Welch and Goyal (2008), we employed
an out-of-sample assessment to evaluate how effectively technical indicators can predict future
performance. Lastly, we conducted an asset allocation exercise for a mean-variance investor and
calculated the certainty equivalent return improvements and Sharpe ratios to assess the economic
benefits of the out-of-sample forecasts.
Three novel findings emerge from these empirical analyses. Firstly, we demonstrate that
volume-based technical indicators operate primarily on a monthly frequency, whereas price-based
indicators are effective at daily and weekly frequencies in predicting crypto market returns. This
finding provides direct evidence in support of Blume, Easley, and O’hara (1994) assertion that
volume imparts information to technical analysis independently of price. Secondly, we show that
machine learning methods, particularly scaled sufficient forecasting, can significantly enhance the
predictive performance of technical indicators on in-sample and out-of-sample tests across various
horizons. Lastly, we find that technical analysis can more effectively forecast the market behavior
of cryptos with higher market capitalization. This finding is consistent with empirical evidence
suggesting that positive feedback trading is prevalent in top cryptos (da Gama Silva et al., 2019;

King and Koutmos, 2021) and such trading activity can establish price patterns that technical
indicators can detect (Neely et al., 2014).
Our research offers several significant contributions to the empirical understanding of the
cryptocurrency market. To the best of our knowledge, this article is the first to conduct a
systematic analysis of the effectiveness of technical analysis across different sample frequencies
in the crypto market. Moreover, our study breaks new ground by investigating the predictive
power of technical indicators using advanced machine learning methods beyond PCA, including
PLS, sPCA, and sSUFF. These sophisticated techniques significantly enhance the performance
of technical indicators in all scenarios, thereby strengthening our findings. Lastly, our study’s
subsample analysis of COVID-19 and market sub-indices complements the growing literature on
the COVID-19 impact on the crypto market and deepens our knowledge of the crypto market
structure.
Our work is closely linked to the burgeoning literature on predicting crypto performance.
In time series analysis, scholars have demonstrated that several predictors, such as economic
fundamentals, trading volume (Balcilar et al., 2017; Bouri et al., 2019), policy uncertainty (Demir
et al., 2018; Colon et al., 2021), investor attention (Shen, Urquhart, and Wang, 2019; Lin, 2021),
and sentiment (Anastasiou, Ballis, and Drakos, 2021; Guégan and Renault, 2021), can effectively
forecast future Bitcoin or crypto market returns. In the cross-section, researchers have found
that various factors, such as network effect (Liu, Tsyvinski, and Wu, 2021), downside risk
(Zhang et al., 2021), and seasonality (Kaiser, 2019), can be employed to form portfolios that
generate significantly positive future returns. Our paper provides evidence of time-series return
predictability at the market level from a technical analysis perspective.
Our study also contributes to the research on the effectiveness of technical analysis. Previous
studies have demonstrated the practicality of technical analysis across various asset types,
including the stock market (Zhu and Zhou, 2009; Neely et al., 2014), futures and commodity
market (Park and Irwin, 2010; Yin, Yang, and Su, 2017), currency market (Abbey and Doukas,
2012; Neely and Weller, 2012), and crypto market (Corbet et al., 2019; Gerritsen et al., 2020;

Detzel et al., 2021). Unlike prior research that focuses on a single frequency, our article evaluates
the effectiveness of technical indicators at multiple frequencies and synthesizes technical analysis
using the latest machine learning techniques.
The rest of the paper is structured as follows. Section 2 describes the data, variables, and
forecasting techniques used in our study. Section 3 presents and discusses our empirical findings.
Finally, Section 4 concludes the paper.
2 Data, Variables and Forecasting Methods
In this section, we describe how we construct the crypto market index and introduce the three types
of technical indicators used in our analyses. We then briefly overview the forecasting methods
employed for testing the return predictability of the crypto market.
2.1 Data and Variables
We obtained cryptocurrency data from Coinmarketcap.com and constructed a value-weighted

crypto market index at a daily frequency using all available coins. Our initial data collection
included 5,909 coins with their daily statistics on open, close, high, and low prices, trading volume,
and market capitalization from January 1, 2013 to May 28, 2022. To ensure data quality and
comprehensive market coverage, we excluded cryptos with a market capitalization smaller than
US$500,000 and slightly broadened our market coverage compared to Liu, Tsyvinski, and Wu
(2022), who only included coins with market capitalizations over US$1,000,000. Additionally, we
limited our sample to cryptos with complete price, trading volume, and capitalization statistics,
resulting in a narrower sample size of 2,963 cryptocurrencies for the entire study period, starting
from the beginning of 2014.
The yearly characteristics of the coins are presented in Panel A of Table 1. It is evident that the
number of coins has been steadily increasing over the sampling period, rising from 125 in 2014 to

1,951 in May 2022. However, an intriguing finding from the descriptive statistics is that the crypto
market size is starting to shrink in 2022, despite 2021 being the year when the market achieved the
highest number of coins satisfying our filters and the highest daily average trading volume of over
160 million dollars. This could suggest a possible slowdown or correction in the cryptocurrency
market in 2022, potentially due to the market becoming over-saturated with a large number of
coins or a shift in investor sentiment. However, further research and analysis would be required to
confirm the underlying causes and implications for the cryptocurrency market going forward.
[Insert Table 1 about here]
The excess market return for cryptocurrencies is calculated by subtracting the return on the
crypto market level index from the risk-free rate, as measured by the yield on 1-month T-bills. We
consider crypto excess market returns at three data frequencies: daily, weekly, and monthly, where
the weekly and monthly returns are cumulatively calculated using the daily returns at the respective
frequency.1 Panel B of Table 1 reports the return characteristics of the market indices and major
cryptos under three data frequencies, respectively. In particular, we separately constructed two
sub-indices using the top 10 largest cryptos (Mega) in the market cap and the rest of the cryptos
(ExMega). It shows that the Mega index returns track the features of the market index returns
very well at all frequencies, while ExMega index returns are generally higher and more volatile.
In addition, the sizable annualized Sharpe ratios indicate that cryptocurrencies, therefore as an
alternative asset category, could cater to the investment needs of the investors at different trading
frequencies (see Brauneis and Mestel, 2019; Nagy and Benedek, 2021, for example). All the
summary statistics at each frequency are consistent in magnitude with those reported in Liu and
Tsyvinski (2021).
To examine the performance of technical indicators in forecasting the returns on the cryptocur-
rency market. we follow Neely et al. (2014); Detzel et al. (2021), among many others, to construct
1 Sincethe cryptocurrency market operates 24/7, we use the Sunday-Sunday definition for identifying a trading
week. In case the readers/practitioners are interested in the results on buy-and-hold returns (weekly/monthly
rebalancing), we report the results in Appendix B.

24 technical indicators at each trading frequency based on three major trend-following strategies:
moving average (MA), momentum (MOM), and “on-balance” volume (VOL). In particular, MA
is a commonly used indicator that shows the average price of an asset over a specified period of
time. Trend-following traders use moving averages to smooth out the price data, making it easier
to identify trends and enter trades in the same direction as the trend. MOM is a technical indicator
that shows the rate of change in the price of an asset over a given period of time. It measures the
strength of the trend and can be used to identify potential trend reversals. VOL is calculated by
adding the volume of trades on up days and subtracting the volume of trades on down days, which
can also be used to identify the strength of the trend and potential trend reversals. In combination,
MA, MOM, and VOL can provide trend-following traders with a comprehensive view of the
strength and direction of the trend in an asset. By using these indicators in combination, traders can
reduce the likelihood of false signals and increase the probability of identifying profitable trading
opportunities.
For MA strategy, we define a buy signal if the short-term moving average exceeds or is equal
to the long-term, and a sell signal vice versa:


1 if MAs,t ≥ MAl,t ,

Si,t = (1)

0
 if MAs,t < MAl,t .
with
1 j−1
MA j,t = ∑ Pt−i for j = s, l, (2)
j i=0
where Pt is the index value of the cryptocurrency market at time t, and s (l) represents short (long)
MA (s < l). Intuitively, the short MA is at high sensitivity to recent price movement than the long
MA, the MA rule intuitively recognizes shifts in price patterns for cryptocurrencies. For clarity,
we denote the MA indicator with MA lengths s and l by MA(s, l). Following the convention of
technical analysis, we choose s = 5, 10, 30 and l = 90, 180, 360 for daily frequency, s = 1, 2, 4 and
l = 12, 26, 52 for weekly frequency, and s = 1, 2, 3 and l = 6, 9, 12 for monthly frequency.

The momentum strategy is based on comparing the current crypto market index level with
its historical level. If the present market index is greater than its m-period-ahead price level, the
market is bullish, a buy signal is generated due to “positive” momentum indicating a continuing
trend of earning higher yields in the future. Mathematically, we write


1 if Pt ≥ Pt−m ,

Si,t = (3)

0
 if Pt < Pt−m .
We consider both short- and long-term momentum strategies in this study. That is, we set m = 5,
10, 30, 90, 180, 360 at daily frequency, m = 1, 2, 4, 12, 26, 52 at weekly frequency, m = 1, 2, 3, 6,
9, 12 at monthly frequency, respectively, to include momentum horizons from one day to one year.
Lastly, we consider the volume-based trading strategies complementary to the price-following

strategies. Following Neely et al. (2014), we generate the trading signals using the “on-balance”
volume, which is defined as
t
OBVt = ∑ VOLk Dk , (4)
k=1
where VOLk is the aggregated trading volume at time k and Dk is the dummy variable, an 1 if Pk
is higher or equal to Pk−1 and if smaller a −1 otherwise. Thus the signal derived by “on-balance”
volume is 
if MAOBV ≥ MAOBV

1
l,t ,

s,t
Si,t = (5)
MAOBV < MAOBV

0
 if s,t l,t ,
where
1 j−1
MAOBV
j,t = ∑ OBVt−i for j = s, l. (6)
j i=0
Similar to the moving average strategy, we denote the volume-based signals VOL(s, l) with MA
lengths s and l and consider the same short and long horizon parameter settings as in MA strategies
under each trading frequency.

2.2 Forecasting Methods
This paper examines the joint predictive power of trend-based trading signals using several cutting-
edge machine learning techniques, including principal component analysis (PCA), partial least
squares (PLS), scaled principal component analysis (sPCA), scaled sufficient forecasting (sSUFF),
least absolute shrinkage and selection operator (Lasso), and elastic net (ENet).
The initial four machine learning methods - PCA, PLS, sPCA, and sSUFF - are considered to be
techniques for reducing dimensionality. These methods aim to extract common latent factors from
predictors through averaging, in order to reduce noise and enhance the signal. They also enable
the decorrelation of predictors that are strongly dependent on each other. While PCA compresses
data into principal components based on the covariance structure among predictors, it fails to
consider the connection between predictors and future returns. This means that the ultimate goal
of forecasting returns is not integrated into the dimension reduction phase.
PLS is an improvement on PCA as it considers both the influence from the predictors and
the target. As shown by Kelly and Pruitt (2013, 2015), PLS is a special case of the three-pass
regression. The algorithm starts by computing the univariate return prediction coefficient for each
predictor using OLS, which stands for the degree to which the returns are sensitive to different
predictors. By taking the average of all predictors into one component with weights corresponding
to the first-stage regression coefficient, PLS put the greatest weight on the strongest predictors and
the smallest weight on the weakest. This allows PLS to exploit the predictor covariation with the
forecast objective directly.
To overcome the deficiencies of PCA, Huang et al. (2022b) also suggested a modified version of
PCA, designated sPCA, which includes useful information from the target in the factor-extracting
technique. Under the sPCA framework, each predictor is scaled according to its predictive slope
or t-statistics on the target variable, resulting in a panel of scaled predictors. The common factors
are then determined by using standard PCA to the scaled predictors. Notably, sPCA is typically
more effective than PCA in the presence of weak components since it gives diminishing weights

to irrelevant predictors.
In addition to linear factor models, Huang et al. (2022a) extends the same scaling idea to
the sufficient forecasting method (SUFF) by Fan, Xue, and Yao (2017), which is designed for
estimating a nonlinear predictive relationship with high-dimensional predictors. Similarly, the
efficacy of sSUFF is focused on outperforming SUFF in the presence of weak factors, while the
presence of strong factors is unlikely to result in adverse effects. In practice, we employ both
linear and nonlinear sSUFF in our predictive model. The linear sSUFF (sSUFFl ) is conducted by
directly regressing algorithm-generated predictive indices on future returns, while the nonlinear
sSUFF employs a fitted nonlinear combination of predictive indices using local linear regression
and then regresses the fitted value on the target.
Apart from the dimension reduction methods, we also consider the two most commonly
used variable selection methods: Lasso by Tibshirani (1996), and ENet by Zou and Hastie
(2005). Both models can be classified as penalized linear regression models, where a penalty
term (or regularization) is introduced to the OLS objective function to avoid over-fitting issues.
In particular, the Lasso imposes an L1 parameter penalization while the ENet imposes both L1
and L2 penalization to the linear regression model. Mechanically, the penalization can produce
suboptimal forecasts and stabilize the model’s out-of-sample performance when predictors are
highly correlated.
3 Empirical Results
In this section, we investigate the predictive power of technical indicators for crypto market returns.
Initially, we test the baseline in-sample forecasting performance of the individual indicators and
the common components derived by different machine learning algorithms at varying frequencies.
After that, we look at how well the prediction worked during out-of-sample periods. Lastly, we
look at the economic value of the out-of-sample performance from an asset allocation standpoint
and investigate the profitability from investing in a smaller basket of cryptocurrencies using sub-

indices.
3.1 In-sample Forecasts
We begin by evaluating the in-sample performance of crypto market return prediction. Firstly, we
examine the usual univariate predictive regression model at three distinct time frequencies:
Rt+1 = αi + βi Si,t + εi,t+1 , (7)
where Rt+1 is the excess return of the crypto market index over the risk-free rate at time t + 1; Si,t
is one of the 24 trading signals at the respective trading frequency. The regression uses data from
January 1, 2015 to May 28, 2022, and adopts Newey and West (1987)’s robust variance-covariance
estimator to adjust the standard error.
[Insert Table 2 about here]
The Panel A of Table 2 reports the in-sample regression slopes, Newey-West t-statistics, and
R2 s in predicting the market excess returns with individual technical indicators under each trading
frequency. Evidently, all the technical indicators positively predict the market returns, and the in-
sample R2 s increase as the trading frequency decreases from daily to monthly. This finding accords
with the graphical evidence shown in Figure 1, showing that index price trajectories are rougher at
higher frequencies as high-frequency tradings are more likely to inject the market microstructure
noise into the crypto prices.
[Insert Figure 1 about here]
We then test the joint return predictability of technical indicators. To avoid overfitting, we
employ several state-of-the-art machine learning methods, namely, PCA, PLS, sPCA, and sSUFF,
to reduce the predictors’ dimensionality while preserving the predictive signals. Following Neely

et al. (2014), with each forecasting method, we construct a single factor to predict the crypto market
returns. We also evaluate Lasso and elastic net to see how effectively they forecast market excess
returns.
Panel B of Table 2 summarizes the F-statistics and the R2 s of each forecasting method.2 Three
findings are in order. First, the F-statistics are all significant at 5% significance level, indicating
that all machine learning approaches have solid prediction ability for the whole data period of 2015
to 2022. Second, the single factor models all achieve sizable R2 s, especially the nonlinear sSUFF
method gains the highest in-sample R2 , which is several times more than the largest one that an
individual technical indicator can achieve. For example, the sSUFF method in daily predictive
regression achieves an in-sample R2 of 3.4%, which is almost 8 times more than MOM(30). At
monthly frequency, the R2 exhibited from the in-sample test of sSUFF can be to the size of 24.54%,
which is also more than 2 times the R2 generated by VOL(3,9). Lastly, the variable selection
methods also achieve impressive results in predicting market returns. Specifically, the Lasso
gains an in-sample R2 of 0.72%, 4.28%, and 12.89% at daily, weekly, and monthly frequency,
respectively. These numbers are substantially larger than the largest R2 that an individual predictor
or a single factor3 can achieve, indicating that technical forecasting using multiple indicators could
be much more effective than relying on a single one.
3.2 Out-of-sample Forecasts
In-sample analysis has been shown to be susceptible to over-fitting and sample-size biases, such
as the Stambaugh bias and the look-ahead problem (see Welch and Goyal, 2008, among others).
Therefore, to address these issues, we also evaluate the out-of-sample forecasting performance
of the technical indicators. Following the approach of Welch and Goyal (2008), Kelly and
Pruitt (2013), and others, we implement a recursive predictive regression model based on various
2 We report the F-statistics to reflect the joint significance of the predictor because the Lasso and ENet include
more than one variable after shrinkage and selection.
3 Note that sSUFF is a multi-factor model while sSUFF is a single-factor model.
l
10

technical indicators. Specifically, we estimate the model as:
Rbt+1 = α
bt + βbt S1:t;t (8)
bt and βbt are derived by regressing {Rs+1 }t−1

where α s=1 on a constant and a technical indicator via
OLS, using {S1:t;s }t−1

s=1 as the regressors. We also test the out-of-sample forecasting ability of the
previously described machine learning algorithms, using the period from January 1, 2018 to May
28, 2022 as the out-of-sample assessment period. To ensure economic rationality, we follow the
concept of Campbell and Thompson (2008) and set the predicted return to zero whenever a negative
return forecast is made.
To evaluate the performance of the predictions in out-of-sample periods, we use two statistics:
the R2OS statistic from Campbell and Thompson (2008) and the MSFE-adjusted statistic from Clark
and West (2007). The R2OS measures the proportionate decrease in mean squared forecast error
(MSFE) for the predictive regression prediction compared to the average historical baseline:
2
T −1
∑t=p Rt+1 − Rbt+1
R2OS = 1 − 2
. (9)
T −1
∑t=p (Rt+1 − R̄t+1 )
Here, Rbt is the forecast based on each technical indicator, and R̄t represents the historical average
baseline based on the constant expected return model (Rt+1 = α + εt+1 ).4
[Insert Table 3 about here.]
The out-of-sample predictive regression results using individual technical indicators are
summarized in Panel A of Table 3. Interestingly, the results reveal a striking pattern in the
relationship between the strategy type, formation horizons, and trading frequencies. First of
4 Itcould also be of interest to apply the in-sample and out-of-sample Sharpe ratios (see, e.g., Barillas et al., 2020;
Kan, Wang, and Zheng, 2022) as the evaluation criteria which take into account the fat-tailed feature of financial data,
estimation risk, and the transaction cost. However, it is beyond the current scope of the paper and will be left for future
studies.
11

all, it shows that the price-following strategies (MA and MOM) are more effective at daily and
weekly frequencies than at monthly frequencies, especially when the strategies are formed using
a relatively short-horizon signal. For example, at daily frequency, the R2OS s are only positive
for MA indicators with long leg horizon of fewer than 360 days, and MA(5,90) achieves the
largest out-of-sample R2 of 0.21% is significant at 5% significance level. Similarly, the momentum
strategies that gain positive R2OS are all with a formation period of less than one year. These results
reinforce the findings in the cryptos literature where strong return predictability is mostly found at
high frequency, i.e., 5-minute, hourly, and daily, using the price-based predictors (e.g., Aslan and
Sensoy, 2020; Akyildirim, Goncu, and Sensoy, 2021; Wen et al., 2022, etc.).
In contrast to price-following indicators, volume-based trading signals demonstrate an excel-

lent out-of-sample predictive ability for market-level returns at monthly frequencies. However, at
daily and weekly frequencies, they are inadequate. This finding is consistent with Gerritsen et al.
(2020), where the authors found that on-balance volume trading strategies could not beat buy-and-
hold benchmarks at the daily frequency. Additionally, our novel finding on the dichotomy between
price- and volume-following strategies supports Blume, Easley, and O’hara (1994), where they
constructed a theoretical model showing that volume provides information separately from price.
The authors claim that in early times, traders may acquire information indicating a significant price
discrepancy relative to the genuine value but take conservative positions due to uncertainty about
the underlying value. In later times, as prices approach their true values, traders become more
confident and take larger positions to capitalize on even minor price differences. Therefore, the
theory suggests that volume delays in reflecting the timely information carried by the price. Our
empirical results confirm this theory by demonstrating that volume-based indicators perform much
better in forecasting monthly returns than daily returns.
Following in-sample analysis, we also check the out-of-sample return predictability using
machine learning methods that aggregate all the information provided by the technical indicators.
The results are presented in Panel B of Table 3. Similar to the in-sample results, the common
predictors extracted by PCA, PLS, sPCA, and sSUFF all perform well in terms of generating
positive and sizable out-of-sample R2 s. Particularly, the nonlinear sSUFF method preserves its
12

outstanding in-sample performance in the out-of-sample analysis. It generates the largest R2OS at all
three sample frequencies with daily 0.30%, weekly 2.70%, and monthly 9.19%. We also checked
the performance of Lasso and elastic net. However, their performances are not as impressive as
in the in-sample analysis and are unstable across different sample frequencies. This is reasonable
because, as mentioned in Gu, Kelly, and Xiu (2020), penalized linear regressions may only provide
inferior predictions when predictors are strongly correlated, and so may not perform as capably as
predictor-averaging-based machine learning approaches, such as PCA and PLS.
Knowing that the COVID-19 outbreak may substantially impact cryptos’ price volatility, we
separately evaluate the out-of-sample performance of each machine learning method under pre-
and post-COVID subsamples using March 2020 when WHO proclaimed COVID-19 a worldwide
pandemic as the cutoff. It shows that the predictability for daily and weekly returns is primarily
concentrated in the post-COVID period (i.e., R2OS,Post is larger than R2OS,Pre ). By contrast, the
predictability is concentrated in the pre-COVID period for monthly returns (see, e.g., sSUFF and
Lasso). Mechanically, price-based indicators drive the daily return predictability, and volume-
based indicators drive the monthly return predictability. Therefore, the pre- and post-COVID
R2OS s result from the fact that when the market becomes more unstable, the timely information
in price becomes more valuable and elevates the predictive power of price-based indicators, while
information delay in volume will depreciate the forecasting ability of volume-based indicators.
We are also aware that the level of market efficiency, specifically in its weak form, can
fluctuate over time and can impact the predictability of returns. Cui et al. (2023) asserted that
the portfolio construction in cryptos market should explicitly take in account the substantial tail
risk that investors confront. Therefore, we conducted further analysis to assess the out-of-sample
R2 (R2OS,down ) during market downturns, defined as the period when returns fall in the bottom
decile (i.e., left tail). Our findings show that R2OS,down values are significantly positive and large,
indicating that return predictability is substantial during market downturns. It is important to note
that this predictability may differ from that observed solely in the post-COVID period.
13

To further investigate how market size affects return predictability, we reconduct the out-of-
sample analysis on the two market sub-indices, where the first index is constructed using the top
10 largest cryptos (Mega) and the latter index excludes the 10 largest cryptocurrencies from its
calculations (ExMega), and report the results in Table 4. Evidently, the R2OS s for the Mega index are
all greater than those for the ExMega index, indicating the cryptos with large market capitalization
are more predictable using technical analysis. This finding is in line with the empirical discovery
that positive feedback trading is prevalent among top cryptocurrencies. (see da Gama Silva et al.,
2019; King and Koutmos, 2021) and such trades may establish price patterns that can be captured
by technical indicators (Neely et al., 2014).
3.3 Asset Allocation Analysis
Having observed the substantial predictive power of technical indicators on cryptocurrency

returns, we now seek to measure the economic value that risk-averse investors could realize by
incorporating this knowledge into their asset allocation decisions. To this end, we follow Campbell
and Thompson (2008) and Ferreira and Santa-Clara (2011) to compute the certainty equivalent
return (CER) gain and the Sharpe ratio, assuming a mean-variance investor who employs out-of-
sample return forecasts to allocate between the crypto market index and risk-free bills. At the
end of each period t, the investor determines the optimal proportion of the portfolio to allocate to
cryptocurrency for the next period as
1b rt+1
wt = 2
, (10)
γ σbt+1
where γ represents the degree of risk aversion, b

rt+1 is a forecast of the crypto market excess return,
and σ 2 is a forecast of its variance. The investor then allocates share 1 − w to risk-free bills, and
bt+1 t
14

the realized portfolio return in period t + 1 is
p f
Rt+1 = wt rt+1 + rt+1 , (11)
f
where rt+1 is the risk-free return. We estimate the variance of the crypto market return using a
rolling window of past returns covering the preceding five years, as recommended by Campbell
and Thompson (2008), Neely et al. (2014), and Guo et al. (2022), among others. We also restrict
wt ∈ [0, 1] to exclude short sales and leverage.
The portfolio’s CER for such investors can be expressed as:
1 2
bp − γ σ
CERp = µ b , (12)
2 p
where µ bp2 is the

b p is the mean of the investor’s portfolio throughout the out-of-sample period, and σ
variance. We calculate the CER gain as the difference in CER between the return forecasts derived
by each ML technique and the historical average. We also investigate whether the portfolio’s
annualized Sharpe ratios based on predictive regression are significantly higher than those of the
investment portfolio that uses the historical average. To further examine the robustness of the asset
allocation exercise, we adjust the risk aversion parameter (γ) to 5 and consider the cases when
the bid-ask spread is counted in trading or not.5 Since we do not have access to the quote data,
we adopt the efficient estimator proposed by Ardia, Guidotti, and Kroencke (2022) to estimate the
bid-ask spread using daily open, high, low, and close price data.6

5 Recent Forbes news mentioned that many US crypto exchanges are embracing zero-fee trading. Binance. US’s
fees were already among the smallest, at 10 bps, matching that of FTX (10 bps for makers and 20 for takers) before
the zero-cost policy was implemented. https://www.forbes.com/sites/javierpaz/2022/07/21/binance-out-on-a-no-fee-
bitcoin-limb-limits-volume-decline-but-muddies-its-profit-prospects/?sh=50ae987563ff
6 There are many alternative methods for estimating bid-ask spreads using low-frequency information, e.g., Roll
(1984); Hasbrouck (2009); Corwin and Schultz (2012); Abdi and Ranaldo (2017), to name a few. However, most of
the methods are designed for stocks only Hasbrouck (2009); Corwin and Schultz (2012); Abdi and Ranaldo (2017).
15

Consistent with the out-of-sample analysis, the Panel A of Table 5 shows that all machine
learning techniques achieve significant and positive CER gains for market index when their R2OS
values are statistically significant, as reported in Table 3. Specifically, the sSUFF algorithm leads
to a monthly CER gain of 7.27%, indicating that investors with a risk aversion of 5 may accept a
yearly portfolio management fee of 7.27% as an opportunity cost for adopting sSUFF’s monthly
predictive regression forecast instead of using the historical average, assuming a transaction cost
of half effective spreads. Furthermore, investment portfolios with positive CER gains all exhibit
significant Sharpe ratios.7
Considering that the market index has included many illiquid small-cap cryptos, it would
be more practically relevant to check the asset allocation results by only using liquid cryptos.
Therefore, we reconduct the analysis with the Mega index and report the results in the Panel B
of Table 5. It shows that our main findings in Panel A are preserved. Specifically, the sSUFF
algorithm leads to a post-transaction-cost monthly CER gain of 7.18% which is of a similar
magnitude to that in Panel A.
Overall, the asset allocation exercise reinforces the in-sample and out-of-sample test results,
indicating that technical indicators can offer considerable economic value for risk-averse investors.
4 Conclusion
This article provides empirical evidence on the predictability of crypto market returns using trend-
based technical indicators at daily, weekly, and monthly frequencies. Our results show that price-
based indicators have statistically and economically significant predictive power for daily and
weekly returns, while volume-based indicators have strong predictive power for monthly returns.
Additionally, we employ several machine learning methods for market return prediction and find
that sSUFF consistently outperforms individual technical indicators and other methods both in
7 As one of the referees has suggested, it would be of interest to compare the machine learning methods with the
1/N simple averaging strategy (Naı̈ve), we report the corresponding asset allocation result in Appendix A. It shows
that the machine learning methods consistently beat the Naı̈ve strategy in the out-of-sample test and asset allocation
exercise.
16

and out of sample across all trading frequencies. Finally, we conduct a sub-index analysis and
conclude that the predictability of crypto market returns is mainly attributable to the leading
cryptocurrencies. Overall, our findings suggest that technical indicators can provide valuable
information for investors in the crypto market, and machine learning methods can further enhance
their predictive power.
References
Abbey, B. S., Doukas, J. A., 2012. Is technical analysis profitable for individual currency traders?
Journal of Portfolio Management 39, 142–150.
Abdi, F., Ranaldo, A., 2017. A simple estimation of bid-ask spreads from daily close, high, and
low prices. Review of Financial Studies 30, 4437–4480.
Akyildirim, E., Goncu, A., Sensoy, A., 2021. Prediction of cryptocurrency returns using machine
learning. Annals of Operations Research 297, 3–36.
Anastasiou, D., Ballis, A., Drakos, K., 2021. Cryptocurrencies’ price crash risk and crisis
sentiment. Finance Research Letters 42, 101928.
Ardia, D., Guidotti, E., Kroencke, T. A., 2022. Efficient estimation of bid-ask spreads from open,
high, low, and close prices. Available at SSRN 3892335 .
Aslan, A., Sensoy, A., 2020. Intraday efficiency-frequency nexus in the cryptocurrency markets.
Finance Research Letters 35, 101298.
Balcilar, M., Bouri, E., Gupta, R., Roubaud, D., 2017. Can volume predict bitcoin returns and
volatility? A quantiles-based approach. Economic Modelling 64, 74–81.
Barillas, F., Kan, R., Robotti, C., Shanken, J., 2020. Model comparison with Sharpe ratios. Journal
of Financial and Quantitative Analysis 55, 1840–1874.
17

Blume, L., Easley, D., O’hara, M., 1994. Market statistics and technical analysis: The role of
volume. Journal of Finance 49, 153–181.
Bouri, E., Lau, C. K. M., Lucey, B., Roubaud, D., 2019. Trading volume and the predictability of
return and volatility in the cryptocurrency market. Finance Research Letters 29, 340–346.
Brauneis, A., Mestel, R., 2019. Cryptocurrency-portfolios in a mean-variance framework. Finance

Research Letters 28, 259–264.
Campbell, J. Y., Thompson, S. B., 2008. Predicting excess stock returns out of sample: Can
anything beat the historical average? Review of Financial Studies 21, 1509–1531.
Clark, T. E., West, K. D., 2007. Approximately normal tests for equal predictive accuracy in nested
models. Journal of Econometrics 138, 291–311.
Colon, F., Kim, C., Kim, H., Kim, W., 2021. The effect of political and economic uncertainty on
the cryptocurrency market. Finance Research Letters 39, 101621.
Corbet, S., Eraslan, V., Lucey, B., Sensoy, A., 2019. The effectiveness of technical trading rules in
cryptocurrency markets. Finance Research Letters 31, 32–37.
Corwin, S. A., Schultz, P., 2012. A simple way to estimate bid-ask spreads from daily high and
low prices. Journal of Finance 67, 719–760.
Cui, T., Ding, S., Jin, H., Zhang, Y., 2023. Portfolio constructions in cryptocurrency market: A
cvar-based deep reinforcement learning approach. Economic Modelling 119, 106078.
da Gama Silva, P. V. J., Neto, A. F., Klotzle, M. C., pinto, A. C. F., Gomes, L. L., 2019. Does the
cryptocurrency market exhibits feedback trading? Economics Bulletin 39, 2830–2838.
Demir, E., Gozgor, G., Lau, C. K. M., Vigne, S. A., 2018. Does economic policy uncertainty
predict the bitcoin returns? an empirical investigation. Finance Research Letters 26, 145–149.
18

Detzel, A., Liu, H., Strauss, J., Zhou, G., Zhu, Y., 2021. Learning and predictability via
technical analysis: Evidence from bitcoin and stocks with hard-to-value fundamentals. Financial
Management 50, 107–137.
Fan, J., Xue, L., Yao, J., 2017. Sufficient forecasting using factor models. Journal of Econometrics
201, 292–306.
Ferreira, M. A., Santa-Clara, P., 2011. Forecasting stock market returns: The sum of the parts is
more than the whole. Journal of Financial Economics 100, 514–537.
Gerritsen, D. F., Bouri, E., Ramezanifar, E., Roubaud, D., 2020. The profitability of technical
trading rules in the bitcoin market. Finance Research Letters 34, 101263.
Gu, S., Kelly, B., Xiu, D., 2020. Empirical asset pricing via machine learning. Review of Financial
Studies 33, 2223–2273.
Guégan, D., Renault, T., 2021. Does investor sentiment on social media provide robust information
for bitcoin returns predictability? Finance Research Letters 38, 101494.
Guo, L., Peng, L., Tao, Y., Tu, J., 2022. Joint news, attention spillover, and market returns.
Available at SSRN 2927561.
Hasbrouck, J., 2009. Trading costs and returns for US equities: Estimating effective costs from
daily data. Journal of Finance 64, 1445–1477.
Huang, D., Jiang, F., Li, K., Tong, G., Zhou, G., 2022a. Are bond returns predictable with real-time
macro data? Working Paper, Singapore Management University .
Huang, D., Jiang, F., Li, K., Tong, G., Zhou, G., 2022b. Scaled PCA: A new approach to dimension
reduction. Management Science 68, 1678–1695.
Jiang, J., Kelly, B. T., Xiu, D., 2022. (Re-)Imag(in)ing price trends. Journal of Finance
forthcoming.
Kaiser, L., 2019. Seasonality in cryptocurrencies. Finance Research Letters 31.
19

Kan, R., Wang, X., Zheng, X., 2022. In-sample and out-of-sample Sharpe ratios of multi-factor
asset pricing models. Available at SSRN 3454628 .
Kelly, B., Pruitt, S., 2013. Market expectations in the cross-section of present values. Journal of
Finance 68, 1721–1756.
Kelly, B., Pruitt, S., 2015. The three-pass regression filter: A new approach to forecasting using
many predictors. Journal of Econometrics 186, 294–316.
King, T., Koutmos, D., 2021. Herding and feedback trading in cryptocurrency markets. Annals of
Operations Research 300, 79–96.
Lin, Z.-Y., 2021. Investor attention and cryptocurrency performance. Finance Research Letters 40,
101702.
Liu, Y., Tsyvinski, A., 2021. Risks and returns of cryptocurrency. Review of Financial Studies 34,
2689–2727.
Liu, Y., Tsyvinski, A., Wu, X., 2021. Accounting for cryptocurrency value. Available at SSRN
3951514 .
Liu, Y., Tsyvinski, A., Wu, X., 2022. Common risk factors in cryptocurrency. Journal of Finance
77, 1133–1177.
Nagy, B. Z., Benedek, B., 2021. Higher co-moments and adjusted sharpe ratios for
cryptocurrencies. Finance Research Letters 39, 101543.
Neely, C. J., Rapach, D. E., Tu, J., Zhou, G., 2014. Forecasting the equity risk premium: the role
of technical indicators. Management Science 60, 1772–1791.
Neely, C. J., Weller, P. A., 2012. Technical analysis in the foreign exchange market. Handbook of
Exchange Rates pp. 343–373.
Newey, W. K., West, K. D., 1987. A simple, positive semi-definite, heteroskedasticity and
autocorrelation consistent covariance matrix. Econometrica 55, 703–708.
20

Park, C.-H., Irwin, S. H., 2010. A reality check on technical trading rule profits in the us futures
markets. Journal of Futures Markets 30, 633–659.
Roll, R., 1984. A simple implicit measure of the effective bid-ask spread in an efficient market.
Journal of Finance 39, 1127–1139.
Shen, D., Urquhart, A., Wang, P., 2019. Does twitter predict bitcoin? Economics Letters 174,
118–122.
Tibshirani, R., 1996. Regression shrinkage and selection via the lasso. Journal of the Royal
Statistical Society: Series B 58, 267–288.
Welch, I., Goyal, A., 2008. A comprehensive look at the empirical performance of equity premium
prediction. Review of Financial Studies 21, 1455–1508.
Wen, Z., Bouri, E., Xu, Y., Zhao, Y., 2022. Intraday return predictability in the cryptocurrency
markets: momentum, reversal, or both. North American Journal of Economics and Finance p.
101733.
Yin, L., Yang, Q., Su, Z., 2017. Predictability of structural co-movement in commodity prices: the
role of technical indicators. Quantitative Finance 17, 795–812.
Zhang, W., Li, Y., Xiong, X., Wang, P., 2021. Downside risk and the cross-section of
cryptocurrency returns. Journal of Banking & Finance 133, 106246.
Zhu, Y., Zhou, G., 2009. Technical analysis: An asset allocation perspective on the use of moving
averages. Journal of Financial Economics 92, 519–544.
Zou, H., Hastie, T., 2005. Regularization and variable selection via the elastic net. Journal of the
Royal Statistical Society: Series B 67, 301–320.
21

Daily Market Indice Weekly Market Indice
8000 8000
7000 7000
6000 6000
5000 5000
Value of Investment
Value of Investment
4000 4000
3000 3000
2000 2000
1000 1000
0 0
-1000 -1000
01
01
53
52
52
52
52
52
51
01
01
01
01
01
01
01
01
01
14
15
15
16
17
18
19
20
21
14
15
16
17
18
19
20
21
22
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
Time Time
(a) Daily (b) Weekly

Monthly Market Indice
8000
7000
6000
5000
Value of Investment
4000
3000
2000
1000
-1000
01
01
01
01
01
01
01
01
15
16
17
18
19
20
21
22
20
20
20
20
20
20
20
20
Time
(c) Monthly
Figure 1. Market Indices at Various Frequencies. From January 1, 2014 through May 28,
2022, three market indices for cryptocurrencies are rebalanced on a daily, weekly, and monthly
basis, respectively, using all cryptocurrencies available on the market with market capitalizations
exceeding $500,000. Index return is a value-weighted series based on the coin’s last period market
capitalization, set the same initial value of 100 to facilitate comparison.
22

Table 1 Summary Statistics
This table reports the summary statistics of the cryptocurrencies. Panel A summarizes the number of
cryptocurrencies, mean and median of market capitalization and trading volumes on yearly basis from 2014
to 2022. The trading volumes are reported at three frequencies, i.e., daily, weekly, and monthly, where the
weekly and monthly trading volumes are the cumulative trading volumes of the daily volumes within that
week and month, respectively. Panel B reports the return characteristics (i.e., mean, standard deviation,
Sharpe ratio) of three crypto indices and three major crypto coins at various frequencies. The market index
is the value-weighted average price of the cryptos that satisfies our filters. The (ExMega) Mega index
is constructed using the (cryptos excluding) top 10 cryptos in market capitalization. The crypto indices,
Bitcoin, and Ripple data range from January 1, 2014 to May 28, 2022. The Ethereum data ranges from
August 8, 2015 to May 28, 2022.
Panel A: Characteristics by Year

Market Capitalization Trading Volume (in thousands)
(in millions) Daily Weekly Monthly
Year Number Mean Median Mean Median Mean Median Mean Median
2014 125 214.26 2.49 1,021.16 21.22 7,148.18 160.22 30,918.06 688.82
2015 85 117.16 1.79 1,031.81 7.42 7,168.48 58.36 31,704.77 262.30
2016 170 136.94 1.97 1,547.03 10.02 10,845.55 82.45 47,218.73 406.89
2017 639 470.41 7.64 20,839.31 88.79 147,501.21 783.40 669,635.82 4,657.34
2018 1,253 402.80 7.98 23,583.89 96.81 165,145.09 739.02 721,196.58 3,372.51
2019 1,268 252.19 4.19 64,525.27 97.96 450,466.92 759.33 1,970,449.03 3,478.90
2020 1,532 335.89 5.18 122,334.99 172.71 863,851.92 1,361.77 3,780,990.27 6,494.10
2021 2,261 1,247.71 13.21 167,330.72 506.08 1,169,166.09 3,945.81 5,148,487.36 18,857.40
2022 1,951 1,092.57 12.13 102,477.33 441.39 709,718.32 3,273.96 2,977,134.74 14,561.13
Full 2,963 702.72 7.43 101,015.29 203.07 707,409.32 1,581.51 3,084,395.06 7,584.37
Panel B: Market Indices and Major Cryptocurrency Returns

Daily Weekly Monthly
Index/Coin Mean Stdev. SR Mean Stdev. SR Mean Stdev. SR
Market Index 0.002 0.039 0.815 0.013 0.109 0.835 0.065 0.267 0.842
Mega Index 0.002 0.039 0.786 0.012 0.108 0.807 0.064 0.259 0.858
ExMega Index 0.002 0.048 0.762 0.017 0.141 0.875 0.089 0.403 0.768
Bitcoin 0.002 0.042 1.131 0.014 0.105 0.950 0.061 0.230 0.916
Ethereum 0.005 0.062 1.408 0.035 0.194 1.315 0.167 0.505 1.147
XRP 0.004 0.079 0.952 0.040 0.373 0.775 0.233 1.122 0.719
23

Table 2 In-sample Forecast of Cryptocurrency Market Returns
This table provides in-sample estimation results for the predictive regression
Rt+1 = α + β Xt + εt+1 ,
where Rt+1 denotes the excess market return (in %) at various frequencies and Xt is one of the return
predictors at the respective data frequency. In Panel A, the return predictors are 24 technical indicators from
three categories: moving average (MA), momentum (MOM), and volume (VOL), each of which generates a
dummy trading signal, with a 1 representing buying and a 0 representing selling. The names of the indicators
included in brackets indicate the short and long horizons for respective trading frequencies. In Panel B, the
return predictors are the common factors extracted from the 24 technical indicators using various machine
learning methods, namely, the PCA, PLS, sPCA by Huang et al. (2022b), the scaled sufficient forecasting
(sSUFF) by Huang et al. (2022a) with sSUFFl being the linear forecast and sSUFF being the nonlinear
forecast using local linear regression, the least absolute shrinkage and selection operator (Lasso), and the
elastic net (ENet). The sample ranges from January 1, 2015 to May 28, 2022. The t-statistics reported are
based on the Newey-West standard errors with 4 lags and R2 s are in percentages (%). ***, ** and * denote
statistical significance at the 1%, 5%, and 10% levels, respectively.
Panel A: Individual Technical Indicators

Predictor β t-stat R2 Predictor β t-stat R2 Predictor β t-stat R2
MA(5,90) 0.19*** 2.57 0.23 MA(1,12) 1.33** 2.16 1.46 MA(1,6) 4.97** 1.68 3.35
MA(5,180) 0.18*** 2.47 0.22 MA(1,26) 1.36*** 2.36 1.54 MA(1,9) 4.53 1.61 2.79
MA(5,360) 0.16** 2.26 0.17 MA(1,52) 1.29*** 2.35 1.38 MA(1,12) 3.51 1.14 1.67
MA(10,90) 0.15** 2.04 0.15 MA(2,12) 1.17** 1.90 1.13 MA(2,6) 4.74 1.64 3.05
MA(10,180) 0.17** 2.30 0.18 MA(2,26) 1.33** 2.31 1.47 MA(2,9) 3.88 1.29 2.05
MA(10,360) 0.15** 2.13 0.15 MA(2,52) 1.27** 2.31 1.34 MA(2,12) 3.21 1.04 1.40
MA(30,90) 0.16** 2.14 0.16 MA(4,12) 0.98** 1.65 0.80 MA(3,6) 7.29*** 2.92 7.20
MA(30,180) 0.18*** 2.49 0.21 MA(4,26) 1.21** 2.01 1.20 MA(3,9) 5.70** 2.02 4.41
MA(30,360) 0.13** 1.84 0.11 MA(4,52) 0.92* 1.56 0.69 MA(3,12) 4.12 1.37 2.31
MOM(5) 0.12* 1.57 0.09 MOM(1) 0.87** 1.81 0.63 MOM(1) 4.16 1.60 2.34
MOM(10) 0.23*** 3.08 0.33 MOM(2) 1.70*** 3.16 2.38 MOM(2) 5.93** 2.15 4.78
MOM(30) 0.26*** 3.37 0.43 MOM(4) 1.54*** 3.06 1.95 MOM(3) 4.78 1.58 3.09
MOM(90) 0.13** 1.72 0.11 MOM(12) 1.39*** 2.34 1.59 MOM(6) 3.39 1.13 1.56
MOM(180) 0.22*** 3.00 0.31 MOM(26) 1.17** 2.05 1.13 MOM(9) 3.41 1.25 1.58
MOM(360) 0.07 0.98 0.03 MOM(52) 0.36 0.66 0.11 MOM(12) 2.69 0.98 0.98
VOL(5,90) 0.14** 1.86 0.13 VOL(1,12) 1.75*** 3.03 2.52 VOL(1,6) 8.47*** 3.23 9.74
VOL(5,180) 0.16** 2.00 0.17 VOL(1,26) 1.31** 2.09 1.43 VOL(1,9) 8.38*** 2.95 9.52
VOL(5,360) 0.14** 1.82 0.13 VOL(1,52) 1.49*** 2.54 1.85 VOL(1,12) 7.97*** 2.76 8.62
VOL(10,90) 0.19** 2.30 0.22 VOL(2,12) 1.40*** 2.53 1.62 VOL(2,6) 6.99*** 2.54 6.63
VOL(10,180) 0.17** 2.11 0.18 VOL(2,26) 0.92* 1.44 0.71 VOL(2,9) 8.55*** 3.06 9.91
VOL(10,360) 0.12* 1.48 0.09 VOL(2,52) 1.56*** 2.72 2.01 VOL(2,12) 7.93*** 2.65 8.52
VOL(30,90) 0.04 0.50 0.01 VOL(4,12) 1.13** 1.86 1.05 VOL(3,6) 8.85*** 3.37 10.62
VOL(30,180) 0.10 1.26 0.06 VOL(4,26) 1.06** 1.76 0.92 VOL(3,9) 9.29*** 3.40 11.71
VOL(30,360) 0.05 0.65 0.02 VOL(4,52) 1.06** 1.74 0.94 VOL(3,12) 7.86*** 2.57 8.37
24

Table 2 (Cont’d) In-sample Forecast of Cryptocurrency Market Returns
Rt+1 = α + β Xt + εt+1 ,
learning methods, namely, the simple averages (Naı̈ve), PCA, PLS, sPCA by Huang et al. (2022b), the
scaled sufficient forecasting (sSUFF) by Huang et al. (2022a) with sSUFFl being the linear forecast and
sSUFF being the nonlinear forecast using local linear regression, the least absolute shrinkage and selection
operator (Lasso), and the elastic net (ENet). The sample ranges from January 1, 2015 to May 28, 2022. The
R2 s are reported in percentages (%). ***, ** and * denote statistical significance at the 1%, 5%, and 10%
levels, respectively.
Panel B: Machine Learning Methods

Predictor F-statistic R2 F-statistic R2 F-statistic R2
Naı̈ve 8.12*** 0.30 9.81*** 2.49 6.89** 7.42
PCA 7.06*** 0.26 9.17*** 2.33 6.75** 7.28
PLS 9.80*** 0.36 10.70*** 2.71 8.37*** 8.87
sPCA 8.03*** 0.30 9.47*** 2.41 7.70*** 8.22
sSUFFl 7.06*** 0.26 9.17*** 2.33 6.76** 7.29
sSUFF 95.02*** 3.40 156.72*** 28.53 30.45*** 24.54
Lasso 19.55*** 0.72 17.16*** 4.28 12.72*** 12.89
ENet 18.23*** 0.67 17.03*** 4.25 11.81*** 12.07
25

Table 3 Out-of-sample Forecast of Cryptocurrency Market Returns
This table provides out-of-sample estimation results by recursively estimating the predictive regression.
In Panel A, the return predictors are 24 technical indicators from three categories: moving average (MA),
momentum (MOM), and volume (VOL), each of which generates a dummy trading signal, with a 1
representing buying and a 0 representing selling. The names of the indicators included in brackets indicate
the short and long horizons for respective trading frequencies. In Panel B, the return predictors are the
common factors extracted from the 24 technical indicators using principal component analysis (PCA), partial
least squares (PLS), scaled PCA (sPCA) by Huang et al. (2022b), and scaled sufficient forecasting (sSUFF)
by Huang et al. (2022a) with sSUFFl being the linear forecast and sSUFF being the nonlinear forecast using
local linear regression, respective. Two shrinkage and variable selection methods, i.e., the least absolute
shrinkage and selection operator (Lasso) and the elastic net (ENet), are also considered for comparison. All
R2 s are in percentages (%) and CW denotes the Clark and West (2007) test statistic. R2OS,Pre and R2OS,Post
denote the out-of-sample R2 with the pre- and post-COVID samples, respectively. The sample is split in
March 2020 when WHO declared COVID-19 a global pandemic. R2OS,down is the out-of-sample R2 evaluated
using the sample when the returns are in the bottom decile. The full sample ranges from January 1, 2015 to
May 28, 2022. The out-of-sample period is from January 1, 2018 to May 28, 2022. ***, ** and * denote

Predictors R2OS CW Predictors R2OS CW Predictors R2OS CW
MA(5,90) 0.21** 1.98 MA(1,12) 1.34** 1.92 MA(1,6) −1.35 0.51
MA(5,180) 0.09* 1.40 MA(1,26) 0.52* 1.30 MA(1,9) −2.88 0.13
MA(5,360) −0.06 0.61 MA(1,52) 0.06 1.00 MA(1,12) −7.27 −0.29
MA(10,90) 0.04 1.06 MA(2,12) 0.71* 1.39 MA(2,6) −1.15 0.57
MA(10,180) 0.01 1.03 MA(2,26) 0.46 1.26 MA(2,9) −3.32 0.18
MA(10,360) −0.08 0.50 MA(2,52) −0.03 0.86 MA(2,12) −7.59 −0.43
MA(30,90) 0.13* 1.57 MA(4,12) 0.17 0.86 MA(3,6) 7.64** 1.81
MA(30,180) 0.11* 1.52 MA(4,26) 0.46 1.16 MA(3,9) −1.17 0.61
MA(30,360) −0.11 0.19 MA(4,52) −0.85 0.05 MA(3,12) −5.84 −0.15
MOM(5) −0.02 0.78 MOM(1) −0.18 0.43 MOM(1) −0.22 0.58
MOM(10) 0.32*** 2.38 MOM(2) 1.86** 2.21 MOM(2) 2.15 1.25
MOM(30) 0.35** 2.24 MOM(4) 1.59*** 2.36 MOM(3) −1.02 0.58
MOM(90) 0.05 1.07 MOM(12) 1.16* 1.63 MOM(6) −4.67 −0.11
MOM(180) 0.24** 2.07 MOM(26) 0.36 1.08 MOM(9) −7.65 −0.52
MOM(360) −0.20 −0.98 MOM(52) −1.77 −1.87 MOM(12) −8.63 −1.24
VOL(5,90) 0.10* 1.31 VOL(1,12) 2.86*** 2.80 VOL(1,6) 12.13*** 2.99
VOL(5,180) 0.04 0.92 VOL(1,26) 1.37** 1.85 VOL(1,9) 8.85** 1.96
VOL(5,360) 0.00 0.62 VOL(1,52) 1.39** 1.73 VOL(1,12) 6.89** 1.82
VOL(10,90) 0.17* 1.53 VOL(2,12) 1.58** 2.27 VOL(2,6) 5.38** 1.82
VOL(10,180) 0.06 1.04 VOL(2,26) 0.29 0.80 VOL(2,9) 10.79*** 2.37
VOL(10,360) −0.10 0.04 VOL(2,52) 1.17* 1.52 VOL(2,12) 6.52** 1.71
VOL(30,90) −0.21 −0.79 VOL(4,12) 0.89* 1.48 VOL(3,6) 11.43*** 2.51
VOL(30,180) −0.18 −0.25 VOL(4,26) 0.33 0.87 VOL(3,9) 12.14*** 2.38
VOL(30,360) −0.25 −1.75 VOL(4,52) −0.71 0.15 VOL(3,12) 6.20** 1.68
26

Table 3 (Cont’d) Out-of-sample Forecast of Cryptocurrency Market Returns
This table provides out-of-sample estimation results for recursively estimating the predictive regression. In Panel A, the return predictors are
24 technical indicators from three categories: moving average (MA), momentum (MOM), and volume (VOL), each of which generates a dummy
trading signal, with a 1 representing buying and a 0 representing selling. The names of the indicators included in brackets indicate the short and
long horizons for respective trading frequencies. In Panel B, the return predictors are the common factors extracted from the 24 technical indicators
using simple averages (Naı̈ve), principal component analysis (PCA), partial least squares (PLS), scaled PCA (sPCA) by Huang et al. (2022b), and
scaled sufficient forecasting (sSUFF) by Huang et al. (2022a) with sSUFFl being the linear forecast and sSUFF being the nonlinear forecast using
local linear regression, respective. Two shrinkage and variable selection methods, i.e., the least absolute shrinkage and selection operator (Lasso) and
the elastic net (ENet), are also considered for comparison. All R2 s are in percentages (%) and CW denotes the Clark and West (2007) test statistic.
R2OS,Pre and R2OS,Post denote the out-of-sample R2 with the pre- and post-COVID samples, respectively. The sample is split in March 2020 when WHO
declared COVID-19 a global pandemic. R2OS,down is the out-of-sample R2 evaluated using the sample when the returns are in the bottom decile. The
full sample ranges from January 1, 2015 to May 28, 2022. The out-of-sample period is from January 1, 2018 to May 28, 2022. ***, ** and * denote

27
Predictor R2OS CW R2OS,Pre R2OS,Post R2OS,down R2OS CW R2OS,Pre R2OS,Post R2OS,down R2OS CW R2OS,Pre R2OS,Post R2OS,down
Naı̈ve 0.13* 1.59 0.12 0.15* 0.66** 1.32** 2.10 0.79 2.06** 1.53* 5.24* 1.37 4.62 6.18 11.46**
PCA 0.20** 1.91 0.06 0.33** 0.99** 1.93*** 2.33 1.10 3.08** 2.56* 4.66* 1.34 3.34 6.66 13.53***
PLS 0.25** 2.11 0.10 0.39** 1.04** 1.94*** 2.38 1.22* 2.94** 3.13** 5.66* 1.54 4.05 8.10* 15.13***
sPCA 0.19** 1.87 0.02 0.34** 0.86** 1.66** 2.16 0.85 2.79** 2.69* 5.61* 1.52 4.01 8.02 15.54***
sSUFFl 0.19** 1.85 0.06 0.31** 0.98** 1.89** 2.30 1.05 3.07** 2.67* 5.45* 1.49 4.23 7.28 12.78***
sSUFF 0.30** 2.12 0.25* 0.35** 1.40*** 2.70*** 2.46 1.24 4.72*** 4.04** 9.19** 1.88 9.89* 8.14 16.59**
Lasso 0.24** 2.06 0.20 0.27* 1.14*** 1.23** 1.71 0.96 1.62* 2.28* 8.83** 2.10 9.57* 7.71* 17.26*
ENet 0.02 0.95 −0.15 0.18 1.01*** 0.69* 1.43 −0.01 1.68* 1.85* 6.08** 1.74 6.34* 5.69 12.22*
Table 4 Out-of-sample Forecast of Sub-Indices
This table provides out-of-sample estimation results on crypto market sub-indices by recursively
estimating the predictive regression. Mega Index is a value-weighted index based on the top 10 coins
ranked by the previous period’s market capitalization, ExMega Index uses all the cryptos in the market
excluding the top 10 largest cryptos. The return predictors are the common factors extracted from the
24 technical indicators using simple averages (Naı̈ve), principal component analysis (PCA), partial least
squares (PLS), scaled PCA (sPCA) by Huang et al. (2022b) and scaled sufficient forecasting (sSUFF) by
Huang et al. (2022a) with sSUFFl being the linear forecast and sSUFF being the nonlinear forecast using
local linear regression, respectively. Two shrinkage and variable selection methods, i.e., least absolute
shrinkage and selection operator (Lasso) and elastic net (ENet) are also considered for comparison. All R2OS s
are in percentages (%) and CW denotes the Clark and West (2007) test statistic. The in-sample estimation
period is from January 1, 2015 to December 31, 2017 and the out-of-sample evaluation period is January 1,
2018 through May 28, 2022. ***, ** and * denote statistical significance at the 1%, 5%, and 10% levels,
respectively.
Mega Index ExMega Index

Daily Weekly Monthly Daily Weekly Monthly
Predictors R2OS CW R2OS CW R2OS CW R2OS CW R2OS CW R2OS CW
Naı̈ve 0.19** 1.94 0.76* 1.35 4.16 1.24 0.08 1.22 0.89 1.27 3.53 1.05
PCA 0.20** 1.86 0.88* 1.47 3.14 1.21 0.11* 1.49 0.65* 1.33 2.44 1.17
PLS 0.26** 2.08 0.88* 1.48 4.02* 1.40 0.06* 1.48 0.39* 1.35 −1.19 0.92
sPCA 0.18** 1.80 0.77* 1.40 4.06* 1.39 0.05* 1.31 0.17 1.23 −0.69 0.95
sSUFFl 0.18** 1.81 1.39** 1.76 5.82* 1.60 0.12* 1.50 0.61* 1.32 0.41 1.03
sSUFF 0.14* 1.60 2.66*** 2.64 11.56* 1.64 −0.04** 1.79 0.44* 1.39 5.22** 1.65
Lasso 0.36*** 2.67 0.75 1.23 10.84*** 2.56 −0.10 0.45 1.30** 1.78 −8.24 0.25
ENet 0.23** 2.07 0.08 0.83 2.84* 1.39 −0.02 0.71 0.14 1.07 −6.63 0.38
28

Table 5 Asset Allocation Performance
This table reports the annualized CER gains (in %) and annualized Sharpe ratios for a mean-variance
investor, who allocates assets between the cryptocurrency market and risk-free bills using the out-of-sample
forecasts in daily, weekly, and monthly frequency. The investor’s risk aversion is set to five. We consider
two scenarios: zero transaction cost and a dynamic transaction cost of half-spread per transaction. The
daily bid-ask spread is estimated using Ardia, Guidotti, and Kroencke (2022), and the weekly and monthly
bid-ask spread is the daily average of the corresponding week and month, respectively. Panel A reports
the results using the market index and Panel B presents the results using the Mega index. The in-sample
estimation period is from January 1, 2015 to December 31, 2017, and the out-of-sample evaluation period
is from January 1, 2018 through May 28, 2022.
Panel A: Market Index

Predictors CER0 SR0 CERbas SRbas CER0 SR0 CERbas SRbas CER0 SR0 CERbas SRbas
Naı̈ve 4.12 0.21 3.43 0.17 5.14 0.36 4.91 0.34 4.21 0.52 4.14 0.51
PCA 5.21 0.33 4.81 0.31 8.19 0.53 7.81 0.51 4.82 0.58 4.69 0.57
PLS 7.97 0.45 6.52 0.40 8.10 0.56 7.42 0.53 5.77 0.64 5.60 0.62
sPCA 4.46 0.35 3.97 0.33 6.29 0.49 5.86 0.47 5.78 0.63 5.62 0.62
sSUFFl 4.80 0.32 4.29 0.29 8.27 0.53 7.84 0.51 4.82 0.62 4.67 0.60
sSUFF 14.50 0.62 4.08 0.29 9.28 0.65 6.84 0.57 7.47 0.74 7.27 0.73
Lasso 10.87 0.44 3.41 0.19 6.72 0.48 4.88 0.41 6.14 0.66 5.91 0.64
ENet 1.40 0.13 −5.65 −0.11 3.17 0.37 1.51 0.31 4.26 0.56 4.08 0.55
Panel B: Mega Index

Naı̈ve 5.56 0.27 5.32 0.26 1.82 0.26 1.62 0.25 3.93 0.53 3.87 0.52
PCA 3.09 0.35 2.94 0.34 1.71 0.37 1.35 0.35 4.21 0.58 4.11 0.57
PLS 6.36 0.48 5.98 0.46 1.01 0.39 0.44 0.37 5.22 0.64 5.09 0.63
sPCA 2.43 0.36 2.25 0.35 −0.05 0.36 −0.43 0.34 5.30 0.63 5.18 0.62
sSUFFl 2.51 0.33 2.34 0.33 5.23 0.47 4.72 0.45 6.56 0.67 6.43 0.66
sSUFF 8.76 0.57 5.52 0.48 12.06 0.57 10.24 0.51 7.40 0.74 7.18 0.73
Lasso 16.02 0.64 13.76 0.57 2.49 0.41 1.06 0.37 8.92 0.80 8.67 0.79
ENet 10.90 0.48 8.75 0.41 −0.48 0.28 −1.71 0.23 3.37 0.56 3.16 0.54
29

Appendices
A Additional Asset Allocation Results
Table A1 Asset Allocation Performance with Naı̈ve Strategy as Benchmark

investor, who allocates assets between the crypto market index and risk-free bills using the out-of-sample
two scenarios: zero transaction cost and a dynamic transaction cost of half-spread per transaction. The daily
bid-ask spread is estimated using Ardia, Guidotti, and Kroencke (2022), and the weekly and monthly bid-ask
spread is the daily average of the corresponding week and month, respectively. The in-sample estimation
period is from January 1, 2015 to December 31, 2017, and the out-of-sample evaluation period is from
January 1, 2018 through May 28, 2022.

PCA 1.10 0.33 1.22 0.32 3.06 0.53 2.91 0.51 0.60 0.58 0.55 0.57
PLS 3.85 0.45 3.51 0.43 2.96 0.56 2.51 0.53 1.56 0.64 1.46 0.62
sPCA 0.34 0.35 0.43 0.34 1.15 0.49 0.95 0.47 1.57 0.63 1.48 0.62
sSUFFl 0.68 0.32 0.76 0.30 3.13 0.53 2.93 0.51 0.60 0.62 0.53 0.60
sSUFF 10.38 0.62 5.82 0.47 4.14 0.65 1.94 0.57 3.26 0.74 3.13 0.73
Lasso 6.75 0.44 3.56 0.32 1.58 0.48 −0.03 0.41 1.93 0.66 1.77 0.64
ENet −2.72 0.13 −5.71 0.01 −1.97 0.37 −3.40 0.31 0.05 0.56 −0.06 0.55
30

B Results on Buy-and-hold Returns
Table B1 Summary Statistics

This table reports the summary statistics of the cryptocurrencies. Panel A summarizes the number of
cryptocurrencies, mean and median of market capitalization, and trading volumes on yearly basis from
2014 to 2022. The trading volumes are reported at three frequencies, i.e., daily, weekly, and monthly, where
the weekly and monthly trading volumes are the cumulative trading volumes of the daily volumes within
that week and month, respectively. Panel B reports the return characteristics (i.e., mean, standard deviation,
Sharpe ratio) of three crypto indices and three major crypto coins at various frequencies. The market index
is the value-weighted average price of the cryptos that satisfies our filters. The (ExMega) Mega index
is constructed using the (cryptos excluding) top 10 cryptos in market capitalization. The crypto indices,
Bitcoin, and Ripple data range from January 1, 2014 to May 28, 2022. The Ethereum data ranges from
August 8, 2015 to May 28, 2022.
Panel A: Characteristics by Year

Market Capitalization Trading Volume (in thousands)
(in millions) Daily Weekly Monthly
Year Number Mean Median Mean Median Mean Median Mean Median
2014 125 214.26 2.49 1,021.16 21.22 7,010.81 146.21 29,046.53 456.32
2015 85 117.16 1.79 1,031.81 7.42 7,198.66 56.44 31,464.03 222.47
2016 170 136.94 1.97 1,547.03 10.02 10,944.64 80.60 49,019.65 373.79
2017 639 470.41 7.64 20,839.31 88.79 150,390.66 805.56 732,329.80 5,087.00
2018 1,253 402.80 7.98 23,583.89 96.81 166,267.74 721.97 741,680.21 2,940.46
2019 1,268 252.19 4.19 64,525.27 97.96 450,091.76 737.69 1,979,036.33 3,107.67
2020 1,532 335.89 5.18 122,334.99 172.71 867,795.31 1,351.90 3,873,111.79 6,005.66
2021 2,261 1,247.71 13.21 167,330.72 506.08 1,173,810.60 3,888.18 5,283,453.15 17,876.90
2022 1,951 1,092.57 12.13 102,477.33 441.39 706,541.02 3,191.42 2,928,584.63 12,797.64
Full 2,963 702.72 7.43 101,015.29 203.07 709,828.50 1,555.71 3,147,280.94 6,938.25
Panel B: Market Indices and Major Cryptocurrency Returns

Index/Coin Mean Stdev. SR Mean Stdev. SR Mean Stdev. SR
Market Index 0.002 0.039 0.815 0.014 0.107 0.934 0.062 0.272 0.795
Mega Index 0.002 0.039 0.786 0.013 0.105 0.892 0.060 0.263 0.795
ExMega Index 0.002 0.048 0.762 0.019 0.151 0.908 0.095 0.415 0.797
Bitcoin 0.002 0.042 1.131 0.014 0.106 0.963 0.063 0.238 0.911
Ethereum 0.005 0.062 1.408 0.035 0.182 1.376 0.178 0.524 1.177
Ripple 0.004 0.079 0.952 0.032 0.276 0.844 0.266 1.270 0.725
31

Table B2 In-sample Forecast of Cryptocurrency Market Returns
Rt+1 = α + β Xt + εt+1 ,
learning methods, namely, the PCA, PLS, sPCA by Huang et al. (2022b), the scaled sufficient forecasting
(sSUFF) by Huang et al. (2022a) with sSUFFl being the linear forecast and sSUFF being the nonlinear
forecast using local linear regression, the least absolute shrinkage and selection operator (Lasso), and the
elastic net (ENet). The sample ranges from January 1, 2015 to May 28, 2022. The t-statistics reported are
based on the Newey-West standard errors with 4 lags and R2 s are in percentages (%). ***, ** and * denote

Predictor β t-stat R2 Predictor β t-stat R2 Predictor β t-stat R2
MA(5,90) 0.19*** 2.57 0.23 MA(1,12) 1.46*** 2.45 1.85 MA(1,6) 5.41** 1.75 3.77
MA(5,180) 0.18*** 2.47 0.22 MA(1,26) 1.16** 1.92 1.18 MA(1,9) 4.54* 1.51 2.65
MA(5,360) 0.16** 2.26 0.17 MA(1,52) 1.16** 2.11 1.17 MA(1,12) 4.89** 1.70 3.08
MA(10,90) 0.15** 2.04 0.15 MA(2,12) 1.47*** 2.46 1.87 MA(2,6) 5.28** 1.75 3.59
MA(10,180) 0.17** 2.30 0.18 MA(2,26) 1.24** 2.22 1.34 MA(2,9) 4.79* 1.62 2.95
MA(10,360) 0.15** 2.13 0.15 MA(2,52) 1.07** 1.96 1.01 MA(2,12) 4.55* 1.57 2.67
MA(30,90) 0.16** 2.14 0.16 MA(4,12) 1.05** 1.76 0.97 MA(3,6) 8.00*** 3.09 8.24
MA(30,180) 0.18*** 2.49 0.21 MA(4,26) 1.00* 1.63 0.87 MA(3,9) 5.57** 1.92 3.99
MA(30,360) 0.13** 1.84 0.11 MA(4,52) 0.94** 1.67 0.77 MA(3,12) 4.36* 1.47 2.45
MOM(5) 0.12* 1.57 0.09 MOM(1) 0.81* 1.49 0.57 MOM(1) 3.87* 1.43 1.93
MOM(10) 0.23*** 3.08 0.33 MOM(2) 1.32*** 2.61 1.52 MOM(2) 3.71* 1.47 1.77
MOM(30) 0.26*** 3.37 0.43 MOM(4) 1.33*** 2.44 1.53 MOM(3) 5.84** 1.93 4.39
MOM(90) 0.13** 1.72 0.11 MOM(12) 1.06** 1.87 0.98 MOM(6) 4.26* 1.45 2.33
MOM(180) 0.22*** 3.00 0.31 MOM(26) 1.07** 1.87 1.00 MOM(9) 2.90 1.01 1.08
MOM(360) 0.07 0.98 0.03 MOM(52) 0.14 0.24 0.02 MOM(12) 1.08 0.39 0.15
VOL(5,90) 0.14** 1.86 0.13 VOL(1,12) 0.55 0.84 0.27 VOL(1,6) 6.10** 2.25 4.79
VOL(5,180) 0.16** 2.00 0.17 VOL(1,26) 1.08** 1.80 1.01 VOL(1,9) 6.82*** 2.36 5.99
VOL(5,360) 0.14** 1.82 0.13 VOL(1,52) 0.63 1.07 0.34 VOL(1,12) 6.68** 2.27 5.74
VOL(10,90) 0.19** 2.30 0.22 VOL(2,12) 0.85* 1.47 0.63 VOL(2,6) 6.18** 2.09 4.91
VOL(10,180) 0.17** 2.11 0.18 VOL(2,26) 1.08** 1.93 1.01 VOL(2,9) 8.55*** 3.01 9.42
VOL(10,360) 0.12* 1.48 0.09 VOL(2,52) 0.34 0.54 0.10 VOL(2,12) 8.08*** 3.01 8.41
VOL(30,90) 0.04 0.50 0.01 VOL(4,12) 0.37 0.58 0.12 VOL(3,6) 6.36** 2.13 5.21
VOL(30,180) 0.10 1.26 0.06 VOL(4,26) 1.16** 2.07 1.18 VOL(3,9) 7.72*** 3.00 7.68
VOL(30,360) 0.05 0.65 0.02 VOL(4,52) 0.71* 1.32 0.45 VOL(3,12) 7.37*** 2.77 6.99
32

Table B2 (Cont’d) In-sample Forecast of Cryptocurrency Market Returns
Rt+1 = α + β Xt + εt+1 ,
learning methods, namely, the simple averages (Naı̈ve), PCA, PLS, sPCA by Huang et al. (2022b), the
sSUFF being the nonlinear forecast using local linear regression, the least absolute shrinkage and selection
operator (Lasso), and the elastic net (ENet). The sample ranges from January 1, 2015 to May 28, 2022. The
R2 s are reported in percentages (%). ***, ** and * denote statistical significance at the 1%, 5%, and 10%

Predictor F-statistic R2 F-statistic R2 F-statistic R2
Naı̈ve 8.12*** 0.30 6.95*** 1.78 5.57** 6.08
PCA 7.06*** 0.26 6.54** 1.68 5.57** 6.08
PLS 9.80*** 0.36 8.09*** 2.06 6.32** 6.85
sPCA 8.03*** 0.30 7.49*** 1.91 6.03** 6.55
sSUFFl 7.06*** 0.26 6.73*** 1.72 5.58** 6.09
sSUFF 95.02*** 3.40 75.08*** 16.22 38.20*** 30.20
Lasso 19.55*** 0.72 11.52*** 2.91 10.11*** 10.52
ENet 18.23*** 0.67 11.15*** 2.82 9.31*** 9.76
33

Table B3 Out-of-sample Forecast of Cryptocurrency Market Returns
This table provides out-of-sample estimation results by recursively estimating the predictive regression.
In Panel A, the return predictors are 24 technical indicators from three categories: moving average
(MA), momentum (MOM), and volume (VOL), each of which generates a dummy trading signal, with
a 1 representing buying and a 0 representing selling. The names of the indicators included in brackets
indicate the short and long horizons for respective trading frequencies. In Panel B, the return predictors
are the common factors extracted from the 24 technical indicators using simple averages (Naı̈ve), principal
component analysis (PCA), partial least squares (PLS), scaled PCA (sPCA) by Huang et al. (2022b), and
sSUFF being the nonlinear forecast using local linear regression, respective. Two shrinkage and variable
selection methods, i.e., the least absolute shrinkage and selection operator (Lasso) and the elastic net
(ENet), are also considered for comparison. All R2 s are in percentages (%) and CW denotes the Clark
and West (2007) test statistic. R2OS,Pre and R2OS,Post denote the out-of-sample R2 with the pre- and post-
COVID samples, respectively. The sample is split in March 2020 when WHO declared COVID-19 a global
pandemic. R2OS,down is the out-of-sample R2 evaluated using the sample when the returns are in the bottom
decile. The full sample ranges from January 1, 2015 to May 28, 2022. The out-of-sample period is from
January 1, 2018 to May 28, 2022. ***, ** and * denote statistical significance at the 1%, 5%, and 10%

Predictors R2OS CW Predictors R2OS CW Predictors R2OS CW
MA(5,90) 0.21** 1.98 MA(1,12) 1.13* 1.58 MA(1,6) 0.35 0.80
MA(5,180) 0.09* 1.40 MA(1,26) −0.45 0.69 MA(1,9) −1.97 0.41
MA(5,360) −0.06 0.61 MA(1,52) −0.10 0.86 MA(1,12) −2.65 0.45
MA(10,90) 0.04 1.06 MA(2,12) 1.01* 1.50 MA(2,6) 0.54 0.91
MA(10,180) 0.01 1.03 MA(2,26) 0.53* 1.31 MA(2,9) −2.90 0.43
MA(10,360) −0.08 0.50 MA(2,52) −0.41 0.60 MA(2,12) −3.80 0.27
MA(30,90) 0.13* 1.57 MA(4,12) 0.78* 1.33 MA(3,6) 8.23** 1.96
MA(30,180) 0.11* 1.52 MA(4,26) 0.14 0.88 MA(3,9) −1.03 0.72
MA(30,360) −0.11 0.19 MA(4,52) −0.74 0.32 MA(3,12) −4.17 0.18
MOM(5) 0.02 0.78 MOM(1) −0.70 0.28 MOM(1) 0.93 0.92
MOM(10) 0.32*** 2.38 MOM(2) 0.08 1.25 MOM(2) 0.90 1.00
MOM(30) 0.35** 2.24 MOM(4) 0.52 1.26 MOM(3) 2.61 1.19
MOM(90) 0.05 1.07 MOM(12) 0.40 1.18 MOM(6) −1.52 0.46
MOM(180) 0.24** 2.07 MOM(26) −0.01 0.87 MOM(9) −6.51 −0.32
MOM(360) −0.20 −0.98 MOM(52) −1.87 −1.41 MOM(12) −8.24 −1.76
VOL(5,90) 0.10* 1.31 VOL(1,12) −0.57 −0.28 VOL(1,6) 4.80** 1.81
VOL(5,180) 0.04 0.92 VOL(1,26) −0.33 0.45 VOL(1,9) 0.89 0.96
VOL(5,360) 0.00 0.62 VOL(1,52) −1.47 −0.51 VOL(1,12) 1.19 0.97
VOL(10,90) 0.17* 1.53 VOL(2,12) −0.28 0.41 VOL(2,6) 3.64* 1.35
VOL(10,180) 0.06 1.04 VOL(2,26) −0.19 0.76 VOL(2,9) 9.16** 1.99
VOL(10,360) −0.10 0.04 VOL(2,52) −1.97 −1.35 VOL(2,12) 7.28** 1.81
VOL(30,90) −0.21 −0.79 VOL(4,12) −0.92 −0.56 VOL(3,6) 3.43* 1.32
VOL(30,180) −0.18 −0.25 VOL(4,26) −0.12 0.84 VOL(3,9) 5.91** 1.79
VOL(30,360) −0.25 −1.75 VOL(4,52) −0.98 −0.08 VOL(3,12) 4.34* 1.56
34

Table B3 (Cont’d) Out-of-sample Forecast of Cryptocurrency Market Returns
This table provides out-of-sample estimation results for recursively estimating the predictive regression.
In Panel A, the return predictors are 24 technical indicators from three categories: moving average
(MA), momentum (MOM), and volume (VOL), each of which generates a dummy trading signal, with
a 1 representing buying and a 0 representing selling. The names of the indicators included in brackets
indicate the short and long horizons for respective trading frequencies. In Panel B, the return predictors
are the common factors extracted from the 24 technical indicators using simple averages (Naı̈ve), principal
component analysis (PCA), partial least squares (PLS), scaled PCA (sPCA) by Huang et al. (2022b), and
sSUFF being the nonlinear forecast using local linear regression, respective. Two shrinkage and variable
selection methods, i.e., the least absolute shrinkage and selection operator (Lasso) and the elastic net
(ENet), are also considered for comparison. All R2 s are in percentages (%) and CW denotes the Clark
and West (2007) test statistic. R2OS,Pre and R2OS,Post denote the out-of-sample R2 with the pre- and post-
COVID samples, respectively. The sample is split in March 2020 when WHO declared COVID-19 a global
pandemic. R2OS,down is the out-of-sample R2 evaluated using the sample when the returns are in the bottom
decile. The full sample ranges from January 1, 2015 to May 28, 2022. The out-of-sample period is from
January 1, 2018 to May 28, 2022. ***, ** and * denote statistical significance at the 1%, 5%, and 10%

Predictor R2OS CW R2OS,Pre R2OS,Post R2OS CW R2OS,Pre R2OS,Post R2OS CW R2OS,Pre R2OS,Post
Naı̈ve 0.13* 1.59 0.12 0.15* 0.43 0.93 0.39 0.47 3.44 1.16 3.82 2.70
PCA 0.20** 1.91 0.06 0.33** 0.41 1.10 0.02 0.88 3.25 1.20 4.01 1.76
PLS 0.25** 2.11 0.10 0.39** 0.27 1.08 −0.07 0.68 2.86 1.17 3.40 1.80
sPCA 0.19** 1.87 0.02 0.34** 0.30 1.09 −0.11 0.81 2.82 1.17 3.26 1.95
sSUFFl 0.19** 1.85 0.06 0.31** 0.33 1.07 −0.04 0.79 3.40 1.20 3.90 2.43
sSUFF 0.30** 2.12 0.25* 0.35** 1.55** 1.79 1.89 1.14* 4.70* 1.37 4.77 4.57
Lasso 0.24** 2.06 0.20 0.27* −1.50 −0.72 −2.29 −0.54 3.71* 1.33 7.38* −3.47
ENet 0.02 0.95 −0.15 0.18 −0.49 0.29 −0.12 −0.93 3.03 1.17 3.60 1.90
35

Table B4 Out-of-sample Forecast of Sub-Indices
This table provides out-of-sample estimation results on crypto market sub-indices by recursively
estimating the predictive regression. Mega Index is a value-weighted index based on the top 10 coins
ranked by the previous period’s market capitalization, ExMega Index uses all the cryptos in the market
excluding the top 10 largest cryptos. The return predictors are the common factors extracted from the
24 technical indicators using simple averages (Naı̈ve), principal component analysis (PCA), partial least
squares (PLS), scaled PCA (sPCA) by Huang et al. (2022b) and scaled sufficient forecasting (sSUFF) by
Huang et al. (2022a) with sSUFFl being the linear forecast and sSUFF being the nonlinear forecast using
local linear regression, respectively. Two shrinkage and variable selection methods, i.e., least absolute
shrinkage and selection operator (Lasso) and elastic net (ENet) are also considered for comparison. All R2OS s
are in percentages (%) and CW denotes the Clark and West (2007) test statistic. The in-sample estimation
period is from January 1, 2015 to December 31, 2017 and the out-of-sample evaluation period is from
January 1, 2018 through May 28, 2022. ***, ** and * denote statistical significance at the 1%, 5%, and 10%
Mega Index ExMega Index

Daily Weekly Monthly Daily Weekly Monthly
Predictors R2OS CW R2OS CW R2OS CW R2OS CW R2OS CW R2OS CW
Naı̈ve 0.19** 1.94 0.77 1.25 3.51 1.18 0.08 1.22 0.31 0.94 3.13 1.26
PCA 0.20** 1.86 0.75* 1.29 3.63 1.27 0.11* 1.49 −0.53 0.78 3.48* 1.42
PLS 0.26** 2.08 0.60 1.22 2.99 1.19 0.06* 1.48 −0.97 0.69 2.57* 1.28
sPCA 0.18** 1.80 0.54 1.18 2.92 1.18 0.05* 1.31 −0.81 0.68 2.37 1.27
sSUFFl 0.18** 1.81 0.87* 1.41 4.22* 1.36 0.12* 1.50 −0.41 0.86 3.05* 1.37
sSUFF 0.14* 1.60 1.00* 1.61 4.46* 1.40 −0.04** 1.79 0.51* 1.44 −5.82 0.33
Lasso 0.36*** 2.67 −0.70 0.25 1.16 0.73 −0.10 0.45 −0.41 0.81 1.03 1.01
ENet 0.23** 2.07 −0.79 0.19 2.98 1.12 −0.02 0.71 −0.94 0.50 1.16 0.91
36

Table B5 Asset Allocation Performance
investor, who allocates assets between the cryptocurrency market and risk-free bills using the out-of-sample
two scenarios: zero transaction cost and a dynamic transaction cost of half-spread per transaction. The daily
bid-ask spread is estimated using Ardia, Guidotti, and Kroencke (2022), and the weekly and monthly bid-ask
spread is the daily average of the corresponding week and month, respectively. The in-sample estimation
period is from January 1, 2015 to December 31, 2017, and the out-of-sample evaluation period is from
January 1, 2018 through May 28, 2022.

Naı̈ve 4.12 0.21 3.43 0.17 1.80 0.31 1.50 0.29 −0.53 0.57 −1.04 0.51
PCA 5.21 0.33 4.81 0.31 1.09 0.40 0.53 0.37 −0.84 0.60 −1.81 0.53
PLS 7.97 0.45 6.52 0.40 1.39 0.41 0.53 0.38 −1.35 0.60 −2.43 0.52
sPCA 4.46 0.35 3.97 0.33 0.31 0.40 −0.27 0.37 −1.26 0.60 −2.29 0.53
sSUFFl 4.80 0.32 4.29 0.29 0.47 0.38 −0.15 0.36 −0.59 0.60 −1.54 0.52
sSUFF 14.50 0.62 4.08 0.29 12.60 0.69 10.16 0.60 2.29 0.63 0.97 0.53
Lasso 10.87 0.44 3.41 0.19 −7.48 0.09 −8.97 0.03 1.35 0.64 −0.50 0.53
ENet 1.40 0.13 −5.65 −0.11 −2.33 0.23 −3.66 0.18 −1.15 0.56 −2.10 0.48
37

SSRN Id4222864

Uploaded by

Copyright:

Available Formats

SSRN Id4222864

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SSRN Id4222864

Uploaded by

Copyright:

Available Formats

Trend-based Forecast of Cryptocurrency Returns*

Xilong Tan†, Yubo Tao‡

April 16, 2023

JEL Classification: G12, G14, G17.

Keywords: Cryptocurrency, Return predictability, Technical analysis, Investment horizon,

Economics and Management (APAEM), University of Macau. Email: yubotao@um.edu.mo.

Electronic copy available at: https://ssrn.com/abstract=4222864

We performed three empirical tests to demonstrate the predictive power of trend-based

Electronic copy available at: https://ssrn.com/abstract=4222864

Electronic copy available at: https://ssrn.com/abstract=4222864

2 Data, Variables and Forecasting Methods

2.1 Data and Variables

We obtained cryptocurrency data from Coinmarketcap.com and constructed a value-weighted

Electronic copy available at: https://ssrn.com/abstract=4222864

[Insert Table 1 about here]

Electronic copy available at: https://ssrn.com/abstract=4222864

Electronic copy available at: https://ssrn.com/abstract=4222864

Lastly, we consider the volume-based trading strategies complementary to the price-following

Electronic copy available at: https://ssrn.com/abstract=4222864

Electronic copy available at: https://ssrn.com/abstract=4222864

Electronic copy available at: https://ssrn.com/abstract=4222864

3.1 In-sample Forecasts

Rt+1 = αi + βi Si,t + εi,t+1 , (7)

[Insert Table 2 about here]

[Insert Figure 1 about here]

Electronic copy available at: https://ssrn.com/abstract=4222864

3.2 Out-of-sample Forecasts

Electronic copy available at: https://ssrn.com/abstract=4222864

bt and βbt are derived by regressing {Rs+1 }t−1

OLS, using {S1:t;s }t−1

[Insert Table 3 about here.]

Electronic copy available at: https://ssrn.com/abstract=4222864

In contrast to price-following indicators, volume-based trading signals demonstrate an excel-

Electronic copy available at: https://ssrn.com/abstract=4222864

Electronic copy available at: https://ssrn.com/abstract=4222864

3.3 Asset Allocation Analysis

Having observed the substantial predictive power of technical indicators on cryptocurrency

where γ represents the degree of risk aversion, b

Electronic copy available at: https://ssrn.com/abstract=4222864

The portfolio’s CER for such investors can be expressed as:

where µ bp2 is the

[Insert Table 5 about here.]

Electronic copy available at: https://ssrn.com/abstract=4222864

Electronic copy available at: https://ssrn.com/abstract=4222864

Electronic copy available at: https://ssrn.com/abstract=4222864

Brauneis, A., Mestel, R., 2019. Cryptocurrency-portfolios in a mean-variance framework. Finance

Electronic copy available at: https://ssrn.com/abstract=4222864

Kaiser, L., 2019. Seasonality in cryptocurrencies. Finance Research Letters 31.

Electronic copy available at: https://ssrn.com/abstract=4222864

Electronic copy available at: https://ssrn.com/abstract=4222864

Electronic copy available at: https://ssrn.com/abstract=4222864

(a) Daily (b) Weekly

Electronic copy available at: https://ssrn.com/abstract=4222864

Panel A: Characteristics by Year

Panel B: Market Indices and Major Cryptocurrency Returns

Electronic copy available at: https://ssrn.com/abstract=4222864

Panel A: Individual Technical Indicators

Electronic copy available at: https://ssrn.com/abstract=4222864

Panel B: Machine Learning Methods