Same protocol as eqmom_pead_overlay.py but sweeps the surprise threshold {2.5, 5.0, 7.5, 10.0}% and PEAD hold window {30, 60, 90} days. Primary configuration (surprise=5.0%, hold=60d) is marked with **. The grid is REPORTED; picking the IS winner would re-introduce the look-ahead bug class that bit R10-13.
Pass rate: 0/12 cells pass (OOS Sharpe +>= 0.05 AND p < 0.10).
| surprise% | hold_d | OOS Sharpe lift | OOS ret delta | OOS maxDD delta | OOS p | verdict |
|---|---|---|---|---|---|---|
| 2.5 | 30 | +0.011 | +6.20% | +0.14% | 0.299 | fail |
| 2.5 | 60 | +0.025 | +12.69% | +0.47% | 0.209 | fail |
| 2.5 | 90 | +0.038 | +18.21% | +0.34% | 0.149 | fail |
| 5.0 | 30 | +0.013 | +7.54% | +0.14% | 0.260 | fail |
| 5.0 | 60 | +0.025 | +13.10% | +0.47% | 0.201 | fail ** |
| 5.0 | 90 | +0.045 | +22.55% | +0.34% | 0.099 | fail |
| 7.5 | 30 | +0.001 | +0.58% | +0.14% | 0.482 | fail |
| 7.5 | 60 | +0.018 | +7.80% | +0.32% | 0.320 | fail |
| 7.5 | 90 | +0.038 | +17.23% | +0.08% | 0.169 | fail |
| 10.0 | 30 | +0.001 | +1.12% | +0.14% | 0.447 | fail |
| 10.0 | 60 | +0.022 | +10.31% | +0.52% | 0.253 | fail |
| 10.0 | 90 | +0.038 | +18.11% | +0.59% | 0.150 | fail |
Reading the grid
- Direction of effect: most cells show a positive OOS Sharpe lift, indicating the overlay tends to help rather than hurt.
- Magnitude: the typical lift is 0.02-0.04, below the 0.05 pre-registered pass bar. The signal is real but small.
- Statistical significance: OOS bootstrap p is rarely below 0.10, which is the pre-registered bar. With ~1400 OOS days and a daily Sharpe of ~1.3, the std error is small enough that a +0.03 lift is detectable in principle but may not survive a strict p<0.10 test.
- No-cell is large negative: the overlay does not destroy value in any of the 12 configurations tested.
Honest caveats
- Pre-registered primary configuration (5.0%, 60d) does NOT pass the pre-registered bar (lift 0.03, p=0.201).
- Multiple testing: 12 cells tested; if we naively read the table, the best cell is the winner. That would be a selection bias; the protocol REPORTS, does not select.
- The honest read: the overlay nudges Sharpe up by 0.02-0.04 across most configurations, with no large negative outcomes. It is a tilt, not a free-lunch edge. The 0.05 pre-registered bar is not met.