Backtest results · Generated → gated → certified
Strong candidates, honestly gated — and a clean floor that survives.
The search produced books above 1.6 Sharpe. Adversarial gating then deflated most of it — and suspended what didn't survive rather than adjusting it. What remains: a conservative, pre-registered floor that clears its bar, and a stronger candidate under live forward validation. All figures are annualized, hypothetical/backtested, before vs after conservative modeled costs. No live-capital track record is claimed.
The ladder — from raw generation to certified floor
Candidate META book (26.1% CAGR) before modeled costs — the ceiling the search produced.
A look-ahead leak was found in the dominant sleeve; the headline was suspended, not adjusted.
Defensible after-cost range — and execution cost is measured near-zero, so the book holds near the top. Registered holdout 1.57.
The stronger candidate is under live paper test; execution cost is measured near-zero.
Backtest detail — before vs after modeled implementation cost
| Book | Sharpe (before → after cost) | CAGR | Max drawdown | Status |
|---|---|---|---|---|
| META — combined candidate | 1.66 → 1.31–1.43 | 26.1% → 20–23% | −29.6% | candidate · forward-validating |
| Registered forward anchor | net of 2× stressed cost · live-lane target |
"Before cost" is the raw / low-cost backtest; "after cost" deducts conservative modeled trading, spread, and borrow — and execution cost is measured near-zero against 291 M auction prints, so the after-cost range holds near the top. META is the strongest candidate but not yet certified — part of it inherits a suspended sleeve, and a 13-year sample cannot establish Sharpe ≥ 1.30; the forward record settles it.
Why the ceiling isn't the headline
Deflated Sharpe gating
Candidates are deflated against the honest ~11,300-look trial burden, not a narrowed denominator. The bar is DSR ≥ 0.90 — and the program suspends what doesn't clear it.
Sealed holdout
The cleanest out-of-sample point is a blind, era-sealed re-screen — selection happens before the holdout is ever seen.
Cost is measured, not assumed
Execution cost — the usual killer of paper Sharpe — is reconciled against 291 M official auction prints: a 0.0 bp median. See Execution & cost.