How the strategy evolved
From the first backtest to today's proof. Every number comes from the single source of truth, never varnished. The ship gates, the known biases and each version's deflated Sharpe are published as-is, including the August 2026 point-in-time correction that deliberately revised the published figures down.
What the ranking returned, next to the index
Average gap with the index: +2.7 points a year, 95% confidence interval -3.0 to +8.5 points. Without the ranking, all eligible insider purchases returned 0.1%/yr.
Archived discovery cohort. These figures are not the return of the active V15 ranking.
- The deflated Sharpe (Bailey-Lopez de Prado) of this discovery cohort drops below zero after search-bias correction. The active V15 ranking publishes no validated return. The historical composite construction remains archived for transparency, not as proof of the active product.
- The Sharpe confidence interval is wide and straddles zero. The out-of-sample window is short and single-regime.
- The cross-sectional Sharpe (dispersion of individual trades) must never be cited alone. Only the portfolio Sharpe is comparable to other strategies.
- May 2026
First scoring backtest
We score eligible insider purchases, then test whether higher scores predict higher forward returns over a 90-day horizon. The early read: selection beats the eligible baseline.
Read the auditT+90Holding window391,106Universe size - May 2026
Walk-forward validation
Train on the past, test on the strictly-later window, roll forward. The discipline that separates a real edge from a curve-fit. The filtered subset wins more often than the raw universe.
Read the audit51.5%Subset win-rate46.3%Universe win-rate - May 2026
Going multi-market
The signal is re-scored across dozens of regulators worldwide, not just one. Coverage explodes, but so does the noise: some venues add edge, others destroy it.
Read the audit - May 2026
The honest-tape decision
The legacy pooled test was negative on the largest venue, so the May 2026 version used a European research universe. The current model supersedes that decision with market-local ranking and evidence-based market weights.
Read the auditSupersededStatus - May 2026
Ship decision
The candidate that cleared every gate goes live: top picks per month, net of fees. The returns published at the time were later withdrawn (audits 200 and 231).
Read the audit0.6%Round-trip cost - June 2026
Independent integrity audit
A read-only quant review pressure-tests the whole pipeline for look-ahead bias, survivorship, and multiple-testing inflation. It is the kind of scrutiny most strategies never publish.
Read the audit< 0Deflated Sharpe3,864Subset size - June 2026
Re-bake and a retired edge
A signal that once looked spectacular collapsed on the full live cohort, so we retired it, on purpose, even though it lowered the headline. Removing it actually lifted the out-of-sample result.
Read the audit56.3%Direct win-rate50.5%Retired-signal win-rate - June 2026
The current, honest proof
After a data-mapping fix removed contaminated rows, every headline was re-baked downward. This is where the archived discovery proof landed: a positive in-sample track record, a thin out-of-sample window, and a deflated Sharpe below zero. The current V15 ranking does not inherit this return.
Read the audit51.5%Filtered win-rate+2.6%Mean T+90 return+/-1.6ptsWin-rate CI - July 2026
Stress-testing an outside strategy
A hedge-fund friend proposed a US micro-cap biotech recipe. Its point-in-time backtest failed. That result applies to the tested recipe, not to every US insider purchase. The current model ranks each market locally and publishes the losing test in full.
Read the auditFAILVerdictLegacy pooled testScope - July 2026
Capital-feasible rerun: hold
A fresh daily-NAV rerun includes capital limits, overlapping positions, costs and risk caps. Risk-adjusted evidence remains too weak, so V15 keeps no return or allocation claim.
Read the auditHOLDVerdictNONEV15 return claim - August 2026
The point-in-time correction
An external review found the published backtest consumed information unavailable at signal time (today-dated market caps, latest FX, within-month ranking). We re-measured under strict point-in-time discipline and revised the headline down. Its returns were restated again in September (audit 231).
Read the auditRevised downHeadline - September 2026
Returns restated: portfolio accounting
Overlapping 90-day books had been compounded monthly, as if each held the whole capital. The corrected backtest splits the capital across the open books and is shown next to the STOXX Europe 600 on the same dates. The ranking beats the pool of all insider purchases, not the index.
Read the auditRetractedFormer figures - October 2026
Dividends counted on both sides
The index was measured with dividends and the ranking without. Both are now total returns. The ranking is level with the STOXX Europe 600 on the same dates, with no demonstrated edge over the index.
Read the auditTotal returnReturn basis - October 2026
Dutch filings dated at the legal deadline
The Dutch regulator's register gives the trade date, not the publication date, and the backtest had entered those filings the day after the trade. They are now entered after the three business days the rules allow for the notification, and every published figure was re-run. Still no demonstrated edge over the index.
Read the audit+12.5%/yrCorrected backtest+9.3%/yrSTOXX Europe 600, same dates0.28 to 2.33Sharpe CI95