Confirmed Dutch filing corrections: the published backtest re-run (2026-10-13)
Public summary of audit 260, following the date correction in audit 258. The full audit names tables, scripts and internal tools and stays internal. The table now includes the confirmed direction and duplicate repairs measured in audit 260, using the same frozen price and date inputs. No strategy rule, threshold or selection logic changed: this is a correction of how the published backtest is measured.
What was wrong
The Dutch regulator's register of managers' transactions gives the date of each trade, not the date the filing was published. Our backtest used that trade date as the publication date, so it entered every Dutch filing the day after the trade, before a reader could have seen it. European rules give a manager up to three business days to notify a trade.
What we did
Every Dutch register filing is now dated at that legal deadline, the trade date plus three business days. Each stored return was moved to the new entry day using the same daily prices on both dates. Every figure the site and the API publish was then re-run, with the accounting of the previous corrections: a third of the capital per monthly book, entry the day after publication, exit after 90 days, dividends on both sides, 0.6% round-trip cost.
The corpus of this run also includes the data corrections of the previous days (duplicate companies merged, subscriptions and allotments no longer read as purchases, prices that could not be right set aside).
A first version over-stated the result
A first version of this re-run, online for a few hours on 13 October, showed 14.5% a year. It had also removed about 2,500 older Dutch filings whose prices we cannot rebuild at the new date. Removing them is not part of the date correction, and it turned out to be the most favourable of every treatment we tested: kept in the data, the ranking makes 12.2%; even if all of them were counted as total losses, it would make 10.4%. The figures below keep them.
Confirmed direction and duplicate repairs
A further check found purchases created from sales and repeated declarations in the Dutch register. We corrected 87 transaction directions and set aside 27 confirmed duplicates (4 rows belong to both groups). Of the affected declarations, 23 were in the backtest decision period, 20 in 2025 and 3 in 2026. We rebuilt the scores and all published proof families with those corrections, keeping prices, entry dates, selection rules and transaction costs unchanged.
This moves the served ranking from 12.2% to 12.5% a year. It is a data correction, not a new strategy. The broader experimental removal of same-day mirrored transactions is not applied.
The numbers, before and after
Served construction, 2022-01 to 2026 Q1 decisions, 416 executed entries, EUR portfolio, total returns, 0.6% round-trip cost; final holdings close through June 2026.
| Figure | Before (9 October) | After |
|---|---|---|
| Annual return of the ranking | 11.2% | 12.5% |
| Sharpe, with 95% confidence interval | 1.04 (0.14 to 2.07) | 1.17 (0.28 to 2.33) |
| Worst drawdown | -8.7% | -9.4% |
| 10,000 EUR becomes | 16,095 EUR | 16,950 EUR |
| Share of entered picks that gained | 63.3% | 66.1% |
| STOXX Europe 600, same dates | 8.9% a year | 9.3% a year |
| Ranking minus the index, points a year (95% interval) | +1.9 (-4.0 to +7.6) | +2.7 (-3.0 to +8.5) |
| All eligible insider purchases | 0.3% a year | 0.1% a year |
| Ranking at 1.5% / 2.5% round-trip cost | 8.2% / 5.0% | 9.5% / 6.3% |
Of the earlier change from 11.2% to 12.2%, about 0.7 point came from the data corrections of the previous days and 0.3 from the new dates. The confirmed direction and duplicate repairs add another 0.3 point. Applied to the exact data of 9 October, the new dates move the ranking from 11.2% to 11.5%.
How to read it
- Moving the Dutch filings three days later lowers their own returns, by about 0.6 point per filing on average, as expected when a look-ahead is removed. The ranked book moves by a few tenths because a few of its picks change.
- The difference with the index has a 95% interval that contains zero. Corrected for the number of configurations tried since the project began (991), the deflated Sharpe stays negative at -0.58. The site's verdict does not change: no demonstrated edge over the index.
- About a fifth of the picks of the backtest are Dutch filings, and they carry much of its advantage over the index. The live paper portfolio now tracks that share.
Caveats
- The legal deadline stands in for the publication date, which the register does not give.
- The correction applies to the backtest. The stored filings keep their register dates for now.
- Dividends are gross, as in the previous correction; the index figure is gross too.
- A backtest is not a live record.