What we got wrong
Research tool · not investment advice. Full disclaimer →
New here? How to read
Every number on this page names its source — start with “How to read”.
Bottom line7 claims we published and then took back, each with what it cost us and the rule it produced. Two of them are corrections of corrections — places where the fix itself had to be withdrawn.
Why this page exists. These are published to keep the fence honest, not because the method lacks value: a record that only ever catches somebody else's errors is a marketing document. Everything here was found by our own later work, and each entry ends in the standing rule it produced — which is the part that compounds.
This is the retraction record: claims that were ours, in our words, that we later destroyed. It is deliberately not the same list as the approaches we have recorded as dead — those live on Strategy validation, verbatim, so a proposal that resembles one is answered before it is run.
Withdrawn we published it, then took it backFalsified we tested it and it failedPattern the same mistake, found across several findingsCorrected twice we corrected it, and the correction needed correcting
WithdrawnJuly 2026
We reported that our corporate-action table was about 30% incomplete — that the record needed roughly 1,746 entries and we held 1,224 — and recommended paying to rebuild the feed.
- What we found
- The 1,224 entries were the complete exchange record. Nothing was missing. The shortfall had been inferred from an assumption and then tested against that assumption rather than against the exchange.
- What changed
- The funding recommendation was withdrawn — there was nothing to fund. An outside reviewer had been briefed on the 30% figure as fact; removing the bad premise made its verdict stronger, not weaker. The separate price-adjustment repair from the same work was real and still stands.
- The rule
- Before calling a dataset incomplete, query the primary source and diff it.
PatternJuly 2026
Eight separate findings, published across one stretch of work, each confident and each argued from evidence.
- What we found
- All eight failed the same way: assert a fact, then test against the assertion instead of against a source. The eighth was the claim that the database was incomplete — the one directly above.
- What changed
- The last of them would have cost real money to repair a database that was already correct.
- The rule
- Before citing a detector's hit count, compute its false-positive rate.
Falsified2026
Momentum ranks future returns — a gross return-to-volatility of 1.29 on a monthly top-25 book.
- What we found
- Net of realistic cost (about 36% a year on roughly 100% monthly turnover) that 1.29 becomes about 0.09, with a negative compound growth rate and a 69% worst drawdown. Buying and holding the broad index scores 0.89.
- What changed
- The headline number was a flat-cost illusion. Momentum stays a selection lens; it is not sold as a fundable book.
- The rule
- Any fundable form must be low-turnover — and is then defensive, not alpha.
WithdrawnJuly 2026
Our recorded champion book scored 1.32 — the best thing we had built.
- What we found
- That 1.32 was flat-cost only. Charged the cost of actually participating in the market at size, it nets 0.52 at ₹25 crore, 0.17 at ₹50 crore and −0.30 at ₹100 crore. It beats the index at no size at all.
- What changed
- The champion was demoted to a descriptive overlay in the same session it was measured.
- The rule
- A cost model that ignores your own size is not a cost model.
FalsifiedJuly 2026
Prices drift after an earnings surprise, and we can trade it — the strongest cohort showed a 60-day excess of +7.62%.
- What we found
- The drift is real descriptively and every tradeable wrapper fails. Trailing 0.10, within-season 0.06 (the pre-registered cell, tested last), hedged −0.58, against 0.85 for simply holding the benchmark.
- What changed
- It ships as a descriptive event lens and is never sold as a book.
- The rule
- “The effect is real” and “the effect is tradeable” are different claims, and the second one has to be earned separately.
Corrected twiceAugust 2026
A fourteen-pattern composite explains which companies avoid catastrophic loss.
- What we found
- The risk separation is real and rebuilds from stored inputs with zero mismatches across 43,412 observations — but the attribution to fourteen patterns is false. Four of the patterns are hard-coded constants carrying no information at all (25.8% of the weight); 41.6% of the weight units are constants; and the top tier has never once fired in 43,412 observations.
- What changed
- Then the correction needed correcting. We had gone on to claim a two-variable screen beat the composite. That was withdrawn in the same session: the confidence interval on the ratio includes 1.0 at every threshold, so the two are indistinguishable. The finding now rests on the weaker and correct ground — the composite has never demonstrated value over two ratios in EITHER direction.
- The rule
- You cannot sell as skill a difference you cannot show.
Corrected twiceAugust 2026
Companies whose profits cover their interest at least three times over blow up far less often — a relative risk of 0.36. A free filter.
- What we found
- They also stop doubling almost exactly as much. Hold volatility fixed within date and both tails compress to the same number: 0.668 on the downside, 0.683 on the upside. The asymmetry that made it look free was a volatility artefact. Inside a volatility bucket it carries no evidence about the direction of the next twelve months, only about the width.
- What changed
- And the rule this entry itself published was wrong. It had said to carry both tails — but the pair it named was a crude upside against a controlled downside, which is the same one-sided rendering in new clothes. The matched number existed nowhere; two independent seats computed it and agreed to six digits.
- The rule
- A one-sided rendering of a two-sided result is the defect, not the number. Quote matched controls or quote neither.
Every entry is drawn from the strategy ledger kept alongside the research, and is checked against it on each build — an entry cannot outlive the record it summarises. Download as CSV.