The State of V13.5 Before the Audit
V13.5_stack was the production scorer for several months. Its logic: a composite score combining insider role (CEO +2.5, CFO +2.5, BOARD +0.5), signal type (cluster +2), pctMcap (x1.4), capitalization (small cap +1), and a sector multiplier based on European sector ETFs.
On the initial cohort (BUY-only, 60k declarations, partial universe), OOS performance was satisfactory: Sharpe 1.21, DSR 0.42, CAGR +45.8%, maxDD -14.1%, hit rate 57.1%. These figures were published.
The Alert: Cohort Expansion
In May 2026, we launched a major dataset expansion:
- Integration of SEC filings (XNAS, US market) for the first time at scale.
- History extension: removal of the pubDate floor, giving access to full historical depth.
- Result: dataset grew from ~60k to 293,566 rows (BUY + SELL sign-flipped).
First warning sign: on this expanded 293k cohort, V13.5_stack posts Sharpe 0.98, CI95 [-0.90, +3.13], DSR 0.12. This is the worst of all candidates tested.
Audit 101: The XNAS Diagnostic
Audit 101 (see dedicated article) isolated the XNAS cohort: 1,229 BUY OOS rows, raw mean r90 +5.13%. The V13.5 picker achieves Sharpe 0.01 on this subset. It actively selects the worst buckets (CEO negative on XNAS, clusters negative on XNAS, mid-high pctMcap negative on XNAS).
The score function was calibrated on EU declarations. The same signals that proxy conviction in Europe invert their sign on the US tape. The picker destroys value on XNAS rather than creating it.
The EU_strict Bake-off
With audit 101 established, audit 102 tested a universe restriction: keep only the 8 European markets where signals match the calibration.
| Variant | Scorer | Sharpe | CI 95 | CAGR% | MaxDD% | Win% | DSR |
|---|---|---|---|---|---|---|---|
| EU_strict | V13.3s_sector_only | 1.21 | [-0.60, 3.71] | 35.0 | -10.6 | 55.0 | 0.29 |
| EU_strict | V13.5_stack | 1.18 | [-0.63, 3.71] | 34.1 | -10.6 | 55.0 | 0.25 |
| EU_strict | V14e_tuned_light | 0.53 | [-0.46, 4.20] | 17.1 | -27.0 | 48.7 | -0.01 |
The universe restriction alone (moving from 293k to 38k EU_strict) improves V13.5 from Sharpe 0.98 to 1.18. Adding V14e multipliers (entity +0.10, family +0.20) pushes it to 0.53.
The Decision: Replace V13.5 in Production
Audit 103 compared 6 candidates and selected V14e_tuned_light on EU_strict:
- Best Sharpe (0.53) of all tested combinations.
- Best DSR (-0.01), clearly above the +0.0 luck floor by a meaningful margin.
- CI95 Lo = -0.46: the tightest of all candidates.
- MaxDD = -10.6%: the lowest in the V13.5+ family (EU restriction + proxy multiplier shaved DD by ~3pp).
- win rate 48.7% vs 52.1% on the full cohort.
- V14e multipliers survive the cohort change: valid on BUY-only 60k (Sharpe 1.31) AND on EU_strict 38k (Sharpe 0.53). Same coefficients across two independent cohorts.
The editorial decision: publish all audits (100, 101, 102, 103), update figures on the site, and do not downplay the V13.5 degradation.
What We Do Not Know Yet
Three honest uncertainties:
1. The OOS window is short. 14 months (Jan 2025 -- May 2026) is a directional signal, not proof. A Sharpe of 0.53 over 14 months can be luck.
2. The family cohort is still thin. 79 family declarations (spouse + child + parent) out of 595k. The family multiplier (+0.20) is calibrated on too few data points to be statistically isolated. The majority of V14e lift comes from the entity multiplier (44k controlled rows).
3. XNAS may have a subset edge. Audit 101 suggests surgical filters (drop CEO + CHAIR + cluster on XNAS, or filter pctMcap < 0.05). We preferred full exclusion for the immediate ship. A V15 "US-specialist" will be researched independently.
Note: the results above are validated on the EU_strict universe (8 European markets), after exclusion of XNAS following audit 101 (signal inversion on the US tape).