A statistically gated ML deployment in NASDAQ equities — where the honest null result is the entire finding.
Most quant ML papers report in-sample alpha that vanishes out-of-sample. The selection bias is structural: thousands of model–feature combinations are searched, the winner is reported, and the search itself is never priced into the significance test.
Deployment is governed by a statistical gate: a strategy only goes live in a fold if its out-of-sample edge clears a pre-registered significance threshold after multiple-testing correction. The gate is evaluated on a strict walk-forward schedule across 1,512 out-of-sample trading days, so no future information ever leaks backward.
The gate opened 0 of 12 walk-forward folds. No statistically defensible alpha survived honest out-of-sample testing — and that is the contribution. A discipline that publishes its nulls as loudly as its wins is the only one that can be trusted when the gate does open.
All 12 walk-forward folds — every gate stayed closed