Machine learning research has gained momentum—also in finance. Consequently, initialmachine-learning-based statistical arbitrage strategies have emerged in the U.S. equities marketsin the academic literature, see e.g., Takeuchi and Lee (2013); Moritz and Zimmermann (2014);Krauss et al. (2017). With our paper, we pose the question how such a statistical arbitrage approachwould fare in the cryptocurrency space on minute-binned data. Specifically, we train a randomforest on lagged returns of 40 cryptocurrency coins, with the objective to predict whether a coinoutperforms the cross-sectional median of all 40 coins over the subsequent 120 min. We buy the coinswith the top-3 predictions and short-sell the coins with the flop-3 predictions, only to reverse thepositions after 120 min. During the out-of-sample period of our backtest, ranging from 18 June 2018to 17 September 2018, and after more than 100,000 trades, we find statistically and economicallysignificant returns of 7.1 bps per day, after transaction costs of 15 bps per half-turn. While this findingposes a challenge to the semi-strong from of market efficiency, we critically discuss it in light of limitsto arbitrage, focusing on total volume constraints of the presented intraday-strategy.
展开▼
