ML Feature Importance in Trading Models

Feature importance tells you what the model leans on. If atm_iv and oi_buildup dominate, the model is reading sentiment (good). If a leaky feature (e.g., same-bar close) dominates, the model is cheating (bad). Always audit: plot importance, check for leakage, confirm economically sensible. A model that ranks 'max_pain_distance' and 'vix_z' high is interpretable; one ranking a random ID is not. Interpretability is the moat -- publish the importance, not just the PnL.