Frozen walk-forward combined scores, 59 symbols, REGION_A_CFG + chop_scale 0. Null = permute combined within each date across names, re-simulate. OOS >= 2023-01-01, N perms = 20.
Rule: REJECT if placebo p>=0.05.
- observed OOS Sharpe 1.47 (full 0.97)
- null mean 0.79, p95 1.19
- empirical placebo p = 0.000
- temporal null (k-bar shift) OOS Sharpes: -63:1.40, -42:0.45, -21:0.72, 21:1.12, 42:0.66, 63:1.22 -> temporal p = 0.000
Verdict: SIGNAL
- Cross-sectional p tests whether the name assignment matters; the temporal p tests whether the timing matters. Both small => real cross-sectional ML selection at the right times.
- Caveat: this tests the name selection only; the simulate engine, stops and sizing are held fixed across re-simulations.