Pre-registered 2026-09-09. Devices from the transaction-cost / timing-luck literature, never tested in this repo (grep 2026-09-09). Universe: 501 delisting-inclusive names, 2006-08-14..2026-08-12, top_k=50, next-open, cost 10bps one-way (decimal 0.001), sqrt(252). Split 2016-01-01. Look-ahead guard PASS. Tested on RAW L126 by design (live 101-name resid+sector+overlay carries tuned cells with selection-bias flags — port only graduates).
Signals: E1 = 50/50 net average of H21+H63 on L126; E2 = -(mean rank L63/L126/L189), H21; E3 = L126 with incumbent hysteresis band 70 (keep incumbents ranked<=70, fill to 50, cap 50).
| Variant | OOS Ret | OOS Sharpe | OOS MaxDD | Turn/yr | CorrMkt | CorrBase | ResidSh | p vs BASE |
|---|---|---|---|---|---|---|---|---|
| BASE L126/H21 raw | +942.19% | 1.07 | -38.77% | 8.57 | 0.77 | 1.00 | 0.45 | - |
| E1 H-blend 50/50 H21+H63 | +1038.59% | 1.09 | -38.57% | 6.64 | 0.79 | 0.99 | 0.47 | 0.2136 |
| E2 lookback-composite | +708.25% | 0.99 | -38.63% | 9.00 | 0.76 | 0.98 | 0.34 | 0.9614 |
| E3 hysteresis-70 | +1106.50% | 1.12 | -38.02% | 6.46 | 0.77 | 1.00 | 0.53 | 0.0224 |
IS (pre-2016) reference
| Variant | IS Ret | IS Sharpe | IS MaxDD |
|---|---|---|---|
| BASE L126/H21 raw | +232.45% | 0.67 | -55.43% |
| E1 H-blend 50/50 H21+H63 | +212.16% | 0.64 | -55.94% |
| E2 lookback-composite | +190.53% | 0.62 | -53.10% |
| E3 hysteresis-70 | +221.96% | 0.65 | -57.26% |
Verdict rules (pre-registered): FULL adopt = OOS p<0.05 vs BASE (Bonferroni ~=0.017 for 3 tests noted), resid sane. EFFICIENCY GRADUATE (approved fallback, H42/H63 precedent) = Sharpe within -0.10 of BASE with turnover cut >=15% or DD cut >=5pp, resid > 0. Otherwise REJECT.
Verdicts (2026-09-09)
- E3 hysteresis-70 — GRADUATE (efficiency device, port candidate). OOS 1.12 vs 1.07, p=0.0224 (passes pre-registered p<0.05; notes Bonferroni ~=0.017 — borderline-strict, so claimed as efficiency graduate, not a Sharpe-claim victory). Turnover 6.46 vs 8.57 (-24.6%), resid 0.53 (best of all arms), maxDD -38.02% vs -38.77%, corrBase 1.00 (same book, cheaper to run). IS consistent (0.65 vs 0.67, no IS-only win). 25bps stress: E3 1.074 vs BASE 1.007 (gap widens — the device earns its keep exactly when costs rise). Next: port test to the live 101-name resid+sector ranking before touching live code; live stays unchanged until that port passes the same gates.
- E1 H-blend 50/50 H21+H63 — GRADUATE (approved fallback). OOS 1.09 vs 1.07, p=0.21 (tie), turnover 6.64 vs 8.57 (-22.5%), resid 0.47, IS consistent (0.64 vs 0.67). 25bps: 1.042 vs 1.007. Same standing as the H42/H63 fallbacks: approved for use if costs rise, not a replacement today.
- E2 lookback-composite — REJECT. OOS 0.99 vs 1.07, p=0.96; turnover UP (9.00 vs 8.57); resid 0.34 < BASE. Averaging in the fast L63 leg adds churn without signal gain on this book. Do not revisit.
Ledger: no live-code change. Queued and still blocked: news-weight re-run (corrtrack 7 rows, needs 60+), OKX-depth entry timing (19d of 5m top-of-book, needs 90d), forward-tracker vote (2 rows, needs 30/60/90d), C7/B5 Phase-2 vote (shadow record accumulating on the eqmom bot).