Retractions
A book about not fooling yourself must publish its own corrections first. Three headline results in this project's history were later retracted. Each is listed here with what was claimed, what the bug or flaw was, and the corrected number. The full write-ups live in the book and in the experiment record.
1. StrengthSpread and the pre-RCA book — the lookahead retraction
- What was claimed
- A family of trained strategies with backtests running from 42 to 747 pips/day across sixteen deployed strategies — and, in the same family but caught before deployment, StrengthSpread at an implausible +203 pips/day out-of-sample, the number that triggered the root-cause analysis.
- What the bug was
- Lookahead bias. A
merge_asoftimestamp-alignment bug leaked up to 55 minutes of future data into StrengthSpread's features, and a related timeframe-downsample bug in the shared indicator pipeline leaked future bars into every strategy trained on it. The apparatus was reading tomorrow's newspaper. - The corrected number
- All sixteen deployed strategies were net-negative live: ≈55,000 pips retracted across 4,869 live trades. StrengthSpread, which never went live, scored −14.18 pips/day on its causal re-run. The best of sixteen causal retrains managed +0.11 pips/day — zero causal edge. See experiments #12, #13 and #20.
2. The first-touch H4 fade — a bar-grid artifact
- What was claimed
- Fading the first touch of a fresh H4 swing level on thin tick-volume earned +9.49 pips/trade net out-of-sample (win rate 54%, Monte-Carlo p = 0.018, positive in all three calendar thirds) — at the time, the search's first statistically significant and stationary contrarian edge.
- What the bug was
- A bar-grid (timezone) artifact. The study's H4 bars were resampled on a UTC-midnight grid; the broker's actual H4 bars run on the New York 17:00 grid. The "edge" existed only on bars the market never printed.
- The corrected number
- On the correct NY-17:00 H4 grid the signal scores −7.31 pips/trade against a coin-flip control's +3.52 — roughly 11 pips worse than random, 0 of 4 gates passed. See experiments #68 and #72.
3. The regime-gated mean-reversion survivor — retracted at the premise
- What was claimed
- The book's high-water mark: a daily mean-reversion strategy fired only in calm, choppy regimes, at +13.0 pips/trade with daily-clustered t = 2.25 and block-bootstrap p = 0.025 — real but marginal, flagged for a paper test.
- What the flaw was
- Not a code bug this time — the premise. The Coda's field-scan re-run, under a different gate, re-measured the underlying daily mean-reversion signal far smaller gross (+1.1 pips/trade) and negative net of the daily rollover financing cost, which consumes the gross edge several times over.
- The corrected number
- +1.1 pips/trade gross; negative net of rollover. The paper-test flag was closed at the premise, and nothing from this family is running. See experiment #58.
Every other verdict in the record stands as published — including the forty-five negatives, which need no retracting.