Predicting cryptocurrency returns with machine learning: Evidence from high-dimensional factor modeling

We investigate the predictability of cryptocurrency returns using a comprehensive set of macroeconomic and cryptocurrency-specific factors and a set of 12 machine learning models. To enhance interpretability, we employ SHAP analysis to quantify the marginal contribution of each factor to model outputs. We further assess the economic value of predictive signals by constructing long-short and long-only portfolios. Empirically, tree-based methods, particularly random forests, deliver the highest predictive accuracy and outperform neural network and linear benchmarks, with predictability substantially stronger than that documented in equity markets. Across models, the market-to-realized-value ratio, new addresses, and active addresses consistently emerge as the most influential predictors, with higher values associated with higher expected returns. Portfolio results show that neural network-based strategies achieve the highest cumulative performance, indicating meaningful investment gains. Overall, our findings demonstrate the value of machine learning for return forecasting in the cryptocurrency market and provide practical insights for investors and financial analysts operating in highly volatile and evolving cryptocurrency environments.