Sources of Cryptocurrency Return Predictability: Econometric and Machine Learning Evidence Across Market States (Frontiers in Business and AI)

Summary

• Predicting cryptocurrency return levels out-of-sample is highly challenging, with every model producing a negative out-of-sample R², where Random Forest performs best at -0.29% and OLS records -1.28%.p.16
• Cross-sectional asset ranking provides stronger signals than level accuracy, with OLS yielding the highest Pearson information coefficient at 0.092 and Random Forest delivering the highest Spearman rank correlation at 0.116.p.17
• Economic evaluation shows Random Forest generating a weekly high-minus-low portfolio return spread of 0.70%, though it remains statistically imprecise with an HAC t-statistic of 1.18.p.17
• SHAP feature importance analysis reveals that predictive signals are concentrated in risk variables (37.8% for Random Forest), return-distribution characteristics (25.3%), and past returns (24.7%).p.18
• A two-state Markov-switching model identifies a low-volatility state with mean realized market volatility of 3.7% and a high-volatility state averaging 8.1% volatility.p.19

Key insights

1. All evaluated econometric and machine learning models produce negative out-of-sample R² metrics for one-week-ahead return level predictions, ranging from -0.29% for Random Forest to -18.76% for a neural network.p.16
2. Evaluating models on asset rankings reveals that Random Forest achieves the strongest Spearman information coefficient at 0.116 despite its negative return-level R².p.17
3. High-minus-low long-short portfolio spreads based on Random Forest forecasts reach a weekly return of 0.70%, but fail to reach statistical significance with an HAC t-statistic of 1.18.p.17
4. The high-volatility market regime exhibits high persistence in a two-state Markov-switching framework, with an estimated transition probability of 0.912.p.19
5. Nonlinear and combination model forecast losses decrease relative to OLS as the probability of a high-volatility market state increases.p.20
6. Point-in-time market concentration fluctuates significantly over time, with the top cryptocurrency's capitalization share dropping from 78% in 2020-2021 to 53% in 2022 before rising to 68% in 2025.p.14