When Alpha Breaks: Two-Level Uncertainty for Safe Deployment of Cross-Sectional Stock Rankers

Researchers developed a two-level uncertainty framework for AI stock ranking models that struggled during 2024's AI thematic rally and sector rotation. The approach uses regime-trust gates to decide when to trade and epistemic uncertainty caps to manage tail risk, improving risk-adjusted performance.

• →LightGBM stock ranking models failed during 2024's AI sector rotation, highlighting the need for uncertainty-aware deployment strategies.
• →Direct Epistemic Uncertainty Prediction (DEUP) was adapted for ranking by predicting rank displacement relative to point-in-time baselines.
• →Epistemic uncertainty shows strong correlation (0.6) with signal strength, making inverse-uncertainty sizing counterproductive for performance.
• →Two-level policy uses regime-trust gates (0.72-0.75 AUROC) to decide trading activity and tail-risk caps for position sizing.
• →The framework improves risk-adjusted returns by treating uncertainty as a tail-risk guard rather than continuous sizing denominator.