Sources of Cryptocurrency Return Predictability: Econometric and Machine Learning Evidence Across Market States

Abstract

This study examines whether nonlinear machine-learning methods improve cross-sectional cryptocurrency return prediction beyond parsimonious linear benchmarks. Using a dynamic point-in-time cryptocurrency universe, we construct weekly characteristics capturing size, past returns, liquidity, risk, and return distributions and compare OLS, Elastic Net, Random Forest, XGBoost, a neural network, and forecast combinations under strictly out-of-sample evaluation. Cryptocurrency returns remain difficult to predict in levels, and greater model complexity does not generate monotonic forecasting gains. Random Forest delivers the strongest rank prediction, while predictive importance is concentrated in risk, return-distribution, and past-return characteristics. Importantly, the relative forecast-loss performance of nonlinear models improves as the probability of a high-volatility market state increases, although this improvement does not translate into comparably stronger cross-sectional rankings. Forecast-sorted portfolio returns are also statistically imprecise. Overall, the evidence indicates that cryptocurrency return predictability is characterized by a concentrated but time-varying information structure, for which model flexibility has conditional rather than universal value.