Forward window: 2026-08-01 .. 2026-08-31 (31 days, 0 AVAX-gate60 long days). SPLIT moved to 2026-07-31 so this is a strict holdout: the data here was not used in the original study's selection or pre-registration. Pre-registered thresholds unchanged: look=168h, carry|z|>=1.5, basis|z|>=2.0, multipliers {0.5, 1.0, 1.25}.
Forward OOS variants
| Variant | Days | LongDays | Ret | Sharpe | PF | MaxDD |
|---|---|---|---|---|---|---|
| baseline (gate60, no overlay) | 31 | 0 | +0.00% | 0.00 | inf | +0.00% |
| primary overlay (0.5/1.0/1.25) | 31 | 0 | +0.00% | 0.00 | inf | +0.00% |
Forward bucket distribution (gate60-long days only)
| Bucket | Days | Share |
|---|
Bootstrap p(forward overlay beats baseline): 1.000
Original study (for context)
Original OOS 2025-01..2026-07-31: Sharpe -0.64 -> -0.28, p=0.000, maxDD delta -13.57%, verdict=PASS (see reports/carry_basis_overlay_avax.md).
Verdict
- Forward OOS baseline Sharpe: 0.00, ret +0.00%, maxDD +0.00%
- Forward OOS overlay Sharpe: 0.00, ret +0.00%, maxDD +0.00%
- Bootstrap p(forward): 1.000
- Result: INERT (0 of 0 long days fire 0.5x/1.25x; nothing to confirm)
Honest caveats
- 31 days is a small sample. The std error on a daily-Sharpe at 31 obs is roughly sqrt(252/31) ~ 2.85x the long-run std. A single lucky or unlucky day can move the verdict; the bootstrap p accounts for this but the absolute Sharpe has wide confidence intervals.
- Inert signal risk: if the bucket_active_frac is below 5%, the verdict is INERT and the overlay did not have a chance to confirm or refute itself.
- Forward verdict is NOT a promotion gate by itself. It is one more data point. The original study's pre-registered verdict (PASS) is the primary gate. A clean forward confirmation is a positive signal; a marginal or INERT forward is a yellow flag, not a veto.