Main _ADJ panel (500 names, 2006-2026). Observed = full-window net Sharpe at 10bps; null = within-date score permutation, N=40. Rule: REJECT if placebo p>=0.05; else SIGNAL.
| Candidate | Observed SR | null mean | placebo p | corr vs BASE | verdict |
|---|---|---|---|---|---|
| BASE xsmom126 | 0.88 | 0.70 | 0.000 | nan | SIGNAL |
| H1 vol-confirm-mom | 0.90 | 0.70 | 0.000 | 1.00 | SIGNAL |
| H2 liq-shock-rev | 0.53 | 0.52 | 0.375 | 0.38 | REJECT |
| H5 vol-compression | 0.64 | 0.61 | 0.325 | 0.68 | REJECT |
| H6 beta-hedged-mom | 0.89 | 0.69 | 0.000 | 0.95 | SIGNAL |
| H7 overnight-mom | 0.88 | 0.72 | 0.000 | 0.91 | SIGNAL |
| H8 intraday-mom | 0.87 | 0.71 | 0.000 | 0.96 | SIGNAL |
| B0 st-reversal-21 | -0.23 | -0.75 | 0.000 | -0.04 | SIGNAL |
Size calibration (pure-noise signals)
K=6 noise signals -> permutation p: 0.33, 0.25, 0.40, 0.93, 0.65, 0.23; fraction p<0.05 = 0% (target ~5-17% at K=6).
Reading
- SIGNAL = the cross-sectional ranking beats random names from the same eligible set. REJECT = the ranking is noise.
- corr vs BASE near 1 with a low placebo p means 'real but a momentum clone' (the edge is the exposure, not the selection).
- A well-calibrated test should NOT reject noise: if the calibration fraction is far above ~5-17%, the gate is miscalibrated.