Machine Learning for Bitcoin/Ethereum Daily Trading

Can machine learning usefully inform daily crypto-asset trading? In their August 2021 paper entitled “Boosting Cryptocurrency Return Prediction”, Ilias Filippou, David Rapach and Christoffer Thimsen apply the XGBoost algorithm to decision trees to generate next-day forecasts of bitcoin and Ethereum excess returns. They consider 39 potential predictor inputs, including: valuation ratios (network value-to-transactions, addresses-to-network value and fee-to-price); return volatilities over the past one, two or three months; deviations of prices from moving averages for four lookback intervals; cumulative excess returns (momentum) for three lookback intervals; and, sentiment indicators based on Google Trends searches, Reddit comments or Factiva articles. They further explore the relative importance of individual predictors. Their benchmark predictors are inception-to-date mean past returns. Using daily prices from CoinMetrics for bitcoin since since mid-July 2010 and for Ethereum since early August 2015, both through January 2021, along with the contemporaneous risk-free rate and data needed to calculate non-price input predictors, they find that:

(more…)