Abstract
This paper comprehensively examines the performance of a host of popular variables to predict Bitcoin returns. We show that time-series momentum, economic policy uncertainty, and financial uncertainty outperform other predictors in all in-sample, out-of-sample, and asset allocation tests. Bitcoin returns have no exposure to common stock and bond market factors but rather are affected by Bitcoin-specific and external uncertainty factors.
Acknowledgement
We thank Chris Adcock (the Editor), two anonymous referees, Fergal Carton, Andrew Detzel, Gerhard Kling, Athanasios Sakkas, Chris Stanley, and Andrew Urquhart for their insightful comments and suggestions. We are also grateful for helpful comments and suggestions from conference participants at Financial Inclusion and Fintech SOAS University of London, 2019 and Cryptocurrency Research Conference, University of Southampton, 2019.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Notes
1 Stambaugh (Citation1999) indicates that coefficients in such predictive regressions as Equation (1) exhibit a finite sample bias, and a normal t-test could be misleading when the predictors are highly persistent.
2 The forecast evaluation period begins on 29 May 2014, or the 901st observation in our sample.
3 The forecast evaluation period for the out-of-sample estimation also begins on 29 May 2014.
4 A utility gain of 2% or more in the predictive model is typically considered economically significant (Rapach and Zhou Citation2013).
6 We standardize each predictor to have a standard deviation of one and report returns in percentage in Table .
7 Our result is consistent with Kim et al. (Citation2016), demonstrating that the scaled time-series momentum delivers a large, significant alpha for a diversified portfolio of international futures contracts.
8 While the of BD is negative at the 21- and 28-day horizons, the Clark and West (Citation2007) test statistic is significant. This is similar to Neely et al. (Citation2014) which shows that certain macroeconomic predictors have negative
values and significant test statistics.
9 The magnitude can be different when the relative risk aversion coefficient varies. However, the patterns remain qualitatively similar. For example, the CER gains of BD decrease when the relative risk aversion coefficient increases.
10 See data.bitcoinity.org for details.
11 We obtain similar results when we use 0.1% and 0.3% as transaction costs. Detzel et al. (Citation2020) suggest that the transaction costs of Bitcoin range from 0.1% to 0.3%.
12 See Table of Elliott and Müller (Citation2006) for the critical values.