When Do Regime Signals Work? HMM Uncertainty and Trade Execution Across Asset Classes

Abstract

HMM-based confidence scores are natural candidates for filtering regime-aware trade execution, yet their empirical predictive validity has not been tested across real assets and evaluation horizons. We conduct the first systematic cross-asset study of three HMM uncertainty signals, composite confidence, raw posterior entropy, and regime persistence, across eight asset classes at daily, 3-day, 5-day, 10-day, and 21-day horizons. At daily resolution, signals are largely uninformative: no individual asset achieves a significant, correctly-signed result that survives Bonferroni correction across 24 per-asset tests (α Bonferroni ≈ 0.002). Signals emerge clearly at 3-day aggregation and strengthen with horizon, with raw entropy achieving Spearman ρ =-0.411 (p < 0.001, bootstrap 95% CI [-0.555,-0.347]) for IWM at 10 days. The temporal threshold broadly aligns with empirical mean regime durations, though the relationship is not strictly linear across assets. Raw entropy outperforms composite confidence for equity indices; stay probability better predicts execution edge for cryptocurrency assets. All results use in-sample HMM fitting; reported effect sizes should be treated as upper bounds on out-of-sample predictive power. Results have direct implications for uncertainty metric design in regime-aware execution systems.

Keywords: Hidden Markov Model, Regime Detection, Optimal Trade Execution, Market Microstructure, Uncertainty Quantification, Temporal Aggregation, Posterior Entropy, TWAP, Financial Machine Learning, Regime-Aware Strategies

JEL Classification: G14, G12, C22, C58, C63, G17