Introduction We test whether classical market microstructure features can predict short-term cryptocurrency returns once data leakage and trading costs are properly accounted for. Methods Using over three million minute-level observations from six major cryptocurrencies on Binance spot and perpetual futures markets, we evaluate nine microstructure measures through a pipeline that includes hierarchical modelling, stability selection, gradient boosting with SHAP, meta-learning, and purged walk-forward cross-validation benchmarked against naive forecasters. Results All features are stably selected at minute frequency, with range-based spread proxies and realised volatility the most robust. However, gradient-boosted models overfit severely under proper leakage controls, and no strategy survives realistic exchange fees. Models trained on one cryptocurrency do not transfer to others, although they transfer well between the spot and futures venues of the same asset. Discussion The results show that microstructure signals carry genuine but weak information content that is useful for understanding market quality but not exploitable at standard retail fee levels.
Edson Pindza (Thu,) studied this question.