The best of our 31 candidate rules scored 52.9%. We rejected it — because of that.

In the spring of 2026 we wrote and tested 31 candidate signal rules. The best one faded extreme speculative positioning, and scored 52.9% accuracy. We rejected it.

Not despite being the best of 31. Because of it.

This article explains the mechanism that makes this paradox necessary: the hypothesis counter we keep publicly, the threshold that hardens with every attempt, and the counting trap that inflated our own samples — in short, why any backtest presented without a count of its attempts should leave you cold.