The point-in-time correction: we revised our published numbers down (2026-08-11)
Public, expurgated summary of internal audit 200. The full audit contains implementation internals (exact features, thresholds, file paths) and stays internal; every figure below is reproduced from the same deterministic run (seed 42).
What happened
An external reviewer asked a fair question: could our backtest be consuming information that did not exist at signal time? We re-audited the entire published construction under strict point-in-time discipline. The answer was yes, in three places:
- Market capitalisation vintage. The eligibility screen and one scoring input used TODAY's market cap to judge filings from 2022. A company that later collapsed was silently screened out of the past; a company that later grew was silently kept in.
- Currency conversion dating. Market caps were converted to EUR at the most recent exchange rate instead of the rate of each filing's publication date.
- Ranking window. Each month's filings were ranked against the whole calendar month, so a filing published on the 3rd was selected using filings from the 28th. (Measured effect: small, but structural.)
A fourth finding was a data-coverage artefact: one scoring input was only populated on a quarter of the corpus, and its mere PRESENCE correlated with outcomes in a way a live scorer could not have known.
The numbers, before and after
Same corpus, same monthly book of 510 picks over 51 months (2022 to 2026 Q1), top-10 per month, held T+90, net of a 0.6% round-trip cost. Only the information set changes.
| Metric | Previously published | Point-in-time strict |
|---|---|---|
| Sharpe (annualized) | 1.73 | 0.82 |
| Deflated Sharpe | 1.24 | 0.46 |
| CAGR (net) | +60.8% | +27.0% |
| 10,000 EUR became | 75,317 EUR | 27,620 EUR |
| Max drawdown | -24.5% | -27.4% |
| Win rate (T+90) | 65.1% | 60.2% |
Re-measuring the old construction on today's corpus gives 1.59 (ordinary data maturation explains 0.14 of the gap vs 1.73); the remaining drop to 0.82 is information that did not exist at signal time.
The weak window stays weak: 2023-2024 is approximately flat (Sharpe 0.02) and calendar 2024 is a losing year (-12.5%). The strong 2025-2026 window remains regime-inflated and is never presented as a base rate.
What survives
After every correction, the ranking still selects better than chance: the point-in-time book beats a buy-everything version of its own eligible pool (Sharpe -0.38 on the same window) by more than a full Sharpe point, and the selection delta is positive on all three sub-windows. The product's claim, a transparent selection screen, stands. The old headline level does not.
The decision
We replaced the published family with the lower, point-in-time honest one, on every public surface. Nothing about the live ranking changed; what changed is that the published evidence now describes only what a live scorer could actually have computed on each filing's publication date.
We publish revisions like this one deliberately. A backtest you cannot trust downward will never be trusted upward.