Resolve the overshoot–shrinkage discrepancy in posterior sizing

Determine why the posterior-sizing shrinkage applied to slow-reversion windows runs against the observed overshoot finding, and establish how this discrepancy should affect the use of the sizing rule for statistical-arbitrage positions.

Background

The paper finds that, contrary to its preregistered prediction, a 60-day window labeled as diverging produces higher realized profit under the contrarian rule than a window labeled as converging. This overshoot effect is replicated on a disjoint period and survives the reported controls.

The posterior-sizing rule nevertheless shrinks positions for slow-reversion windows, which are associated with the diverging regime. The paper explicitly identifies the resulting tension between the empirical overshoot pattern and the model-based shrinkage mechanism as unresolved.

References

The sizing rule eq:sizing does not use it, and the shrinkage it applies to slow-reversion windows runs against this finding, which is left open.

— Amortized Neural Filtering for Calibrated Posteriors in Statistical Arbitrage  (2610.25802 - Chang, 19 Oct 2026) in Section 4.1, subsection “Results,” paragraph “Overshoot”