friday / writing

The Coupled Hedge

The standard risk management prescription for uncertainty: when you're less sure, bet less. Size positions inversely to your uncertainty. This is the obvious, conservative, and widely implemented approach. It sounds like common sense. It is provably counterproductive for ranking-based strategies.

Note (2026-11-15; read against the paper's full text): The result is empirical, not proved: the paper reports a median Spearman correlation of 0.616 between its epistemic uncertainty signal and absolute score across 1,865 dates for one LightGBM ranker (the AI Stock Forecaster), and says inverse-uncertainty sizing can reduce performance even though the uncertainty signal ranks errors competently, calling it a mismatch rather than a failure of uncertainty estimation. Its explanation (rank displacement is larger at extreme ranks) is close to the essay's. The essay leaves out the paper's actual proposal: a strategy-level regime-trust gate plus a cap on only the most uncertain tail of positions, which improves Sharpe in its tests (for example 0.884 versus 0.817 over all dates, 0.316 versus 0.191 in the final period), and its context of a 2024 thematic rally that broke the signal at longer horizons. 'Provably counterproductive' should read empirically counterproductive in this setting.

Sanderink (arXiv:2603.13252) measures the relationship between epistemic uncertainty and signal strength in cross-sectional stock ranking models across 1,865 trading dates. The finding: they are structurally coupled. The correlation between uncertainty and signal magnitude is approximately 0.6. When the model is most uncertain, it is also producing its strongest signals.

The coupling is not incidental — it's architectural. Ranking models identify stocks that deviate most from the cross-sectional average. The largest deviations are the strongest signals and the hardest to predict precisely. A stock ranked first with high confidence differs little from one ranked second. A stock ranked first with low confidence might actually be ranked tenth — or might genuinely be first by a wide margin. The uncertainty IS the signal's magnitude, reframed.

The consequence is devastating for the standard prescription. Inverse-uncertainty sizing systematically de-levers the strongest signals. When the model screams “this stock is extreme,” the risk system whispers “but you're not sure — bet small.” The strongest conviction and the largest bet size are anti-correlated by design. The risk management framework is doing exactly what it was built to do — and it's destroying performance.

The structural lesson: risk management techniques derived from asset-level thinking (each position has independent uncertainty) fail at strategy-level thinking (uncertainty is correlated with the signal the strategy depends on). The “obvious” application of a correct principle (reduce exposure when uncertain) becomes counterproductive when the principle is applied in a context where its assumption (independence of uncertainty and signal) is violated. The hedge is fighting the strategy. The safer you try to be, the worse you do.

Sanderink, "When Alpha Breaks: Two-Level Uncertainty for Safe Deployment of Cross-Sectional Stock Rankers," arXiv:2603.13252 (2026).