Two-Level Uncertainty for Safe Stock Rankers

🔑 Enhanced Key Takeaways

• •The paper was authored by Ursina Sanderink and initially submitted to arXiv on February 23, 2026[4].
• •AI Stock Forecaster generates rankings at 20d, 60d, and 90d horizons specifically over a dynamic universe of AI-exposed U.S. equities[1].
• •Tier 2 uncertainty estimators using per-stock LightGBM quantile regression on features like vol_20d, adv_20d, market_vol_21d, vix_percentile_252d, mom_1m, and sector_enc succeeded at 60d horizons (ρ=0.317) but failed at 20d (ρ=0.024)[1].
• •The two-level policy with G(t) ≥ 0.2 threshold, volatility sizing on active dates, and epistemic tail-risk capping achieved superior metrics including Sharpe 3.121, Sortino 5.410, max drawdown -18.1%, and Calmar 6.07 compared to baselines[1].

🛠️ Technical Deep Dive

• •DEUP adaptation for ranking predicts rank displacement relative to a point-in-time (PIT-safe) baseline to define epistemic uncertainty signal êhat[2].
• •Median correlation ρ(êhat, |score|) = 0.616 across 1,865 dates indicates structural coupling between epistemic uncertainty and signal strength[1][2].
• •Regime-trust gate G(t) uses AUROC of 0.72 overall and 0.75 in final evaluation for deciding whether to trade[2].
• •Tiered uncertainty estimators include Tier 2: walk-forward LightGBM quantile models predicting 25th and 75th percentiles of rank displacement from features (vol_20d, adv_20d, market_vol_21d, vix_percentile_252d, mom_1m, sector_enc)[1].
• •Operational policy: trade only if G(t) ≥ 0.2, apply volatility sizing on active dates, cap top epistemic tail-risk[2].

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline