81 · Director-vs-CEO divergence detector (2026-05-19)

Verdict

Kept V13.1g baseline. No V13.4 produced.

Director-vs-CEO divergence signal failed the anti-overfit gate. Pattern A (directors buy while exec sells) showed a strong in-sample lift (+23.5pp win90) on 2021-2024 priced rows but collapsed to flat (+2.5pp) in the held-out 2025-2026 OOS window, well below the required +10pp threshold. Pattern B (exec buy while directors sell) OOS sample was too small (N=15) to use as a negative filter without risk of regret.

Method

scripts/_divergence_v13_4/detect.ts walks every priced declaration with a backtest result (n=112 194, 2021-01-01 → 2026-05-19) and for each row counts, in a ±30 day rolling company window:

exec buys (CEO + CFO)
exec sells
board buys (CHAIRMAN + BOARD)
board sells

Roles resolve via the same regex used in scripts/_v13_bakeoff/bake.ts (PDG, CEO, CFO, DAF, Chairman, Administrateur, …).

Patterns:

Pattern A board buys ≥ 2 AND exec sells ≥ 1
Pattern B exec buys ≥ 1 AND board sells ≥ 2
Consensus ≥ 70 percent of window flow in one direction with zero counter-direction signal (skipped from analysis)

Universe split:

TRAIN 2021-01-01 → 2024-12-31, n=86 586 priced rows, 9 118 buys
OOS 2025-01-01 → 2026-05-19, n=25 608 priced rows, 2 850 buys

Returns measured from pubDate+1 (retail entry).

Results

TRAIN 2021-2024

Bucket	n	win90 percent	median90 percent	mean90 percent	win365 percent	median365 percent
baseline (all buys)	9 118	43.6	-1.91	-1.30	44.7	-3.92
Pattern A (buys only)	404	67.1	+2.16	+1.82	57.7	+7.76
Pattern B (buys only)	56	55.4	+4.33	+0.20	41.1	-9.87
Pattern A (all rows)	840	60.7	+2.16	-1.57	53.6	+4.29
Pattern B (all rows)	220	55.0	+0.11	-1.39	19.1	-8.61
Consensus (all rows)	7 623	38.6	-3.35	-3.00	41.3	-6.67

TRAIN headline: Pattern A buys-only win90 = 67.1 percent vs baseline 43.6 percent (+23.5pp). Median 90d return flips sign from -1.91 percent to +2.16 percent. Pattern A "all rows" 365d median +4.29 percent vs baseline -3.92 percent (+8.21pp).

Consensus rows underperform baseline by 5pp on win90, confirming the "everyone agrees" bucket is the bad cohort, as expected.

OOS 2025-2026 (held-out)

Bucket	n	win90 percent	median90 percent	mean90 percent	win365 (n)	median365 percent
baseline (all buys)	2 850	45.6	-1.12	+1.85	35.9 (n=853)	-10.74
Pattern A (buys only)	54	48.1	-0.26	-1.55	29.6 (n=27)	-23.99
Pattern B (buys only)	15	100.0	+23.85	+24.65	100.0 (n=15)	+66.20

Pattern A OOS lift over baseline: +2.5pp win90 (well below the +10pp threshold), median90 -0.26 percent vs baseline -1.12 percent (negligible), 365d median collapses to -24 percent vs baseline -10.7 percent.

Pattern B OOS sample N=15, too small. The 100 percent win rate is a sample-size artifact, not a tradable edge. Most of the rows happen to be RSU vesting events ("Acquisition d'actions gratuites") which co-occur with quarterly executive ladder sells and have no contrarian content.

Counts

Bucket	TRAIN	OOS
Pattern A	840	75
Pattern B	220	29
Consensus	7 623	(not measured)

Why the signal failed

RSU/award contamination. Many "board buys" tagged Pattern A are "Acquisition d'actions gratuites" (free-share grants triggered by compensation calendars), not discretionary purchases. They cluster on the same dates as the CEO's quarterly disposal of vested stock. There is no behavioral information in this co-occurrence. Filtering nat string to discretionary BUYs would shrink Pattern A by an estimated 40-60 percent and almost certainly kill the remaining edge.
Same-day duplication on widely-held companies. A single Dassault Systemes board-grant day puts 30+ rows into Pattern A at once, inflating the in-sample win rate because every row inherits the same pricing path. The OOS period happens to have fewer such bulk-grant events, so the inflated TRAIN lift evaporates.
Cohort overlap with cluster bonus. Pattern A by construction requires 3+ insiders in 30 days, a near-superset of the existing isCluster and clusterParticipantCount ≥ 5 features already in V13.1g. The marginal lift after controlling for cluster width is close to zero.

Sample rows (5 per pattern, one per company)

Pattern A (board-buy while exec-sells, ±30d window):

pubDate	company	market	row role	row tx	ret90 percent	ret365 percent
2021-12-08	dassault-systemes	XPAR	OTHER	Cession	-21.9	-33.3
2021-12-08	totalenergies-se	XPAR	OTHER	Cession	+2.2	+18.6
2024-03-21	publicis-groupe	XPAR	CFO	Acquisition	+2.5	-7.7
2021-12-08	lair-liquide	XPAR	BOARD	Acquisition	-8.1	-3.0
2023-10-11	vinci	XPAR	OTHER	Acquisition	+11.3	+4.3

Pattern B (exec-buy while board-sells, ±30d window):

pubDate	company	market	row role	row tx	ret90 percent	ret365 percent
2022-05-23	dassault-systemes	XPAR	CHAIR	Cession	+5.0	+5.8
2024-05-30	totalenergies-se	XPAR	OTHER	Acq actions gratuites	-3.0	-14.9
2021-12-08	safran	XPAR	BOARD	Conversion auto	-2.5	+12.8
2024-09-17	publicis-groupe	XPAR	CEO	Cession	+5.0	-20.5
2021-12-08	nextedia	XPAR	BOARD	Cession	-24.1	-28.4

Decision rule applied

Per mission spec Phase 3:

Pattern A OOS win90 lift = +2.5pp, threshold +10pp. Reject.
Pattern B OOS n=15. Below the minimum sample for a soft negative filter (would need ≥ 60 to deflate Sharpe by less than the noise floor). Reject as negative filter.
Both patterns: no schema change, no scoring change.

V13.1g_stacked remains canonical: Sharpe 0.70, DSR +0.33, CAGR +25.4 percent on the same OOS window (see doc 78).

Follow-ups (not blocking)

Re-run after transactionNature cleanup to exclude RSU / free-share / conversion natures from the Pattern A counter. If the residual edge survives at OOS +10pp on N ≥ 100, reopen.
Try a tighter window (±14 days) and a stricter "must-be-discretionary" filter on the exec-sell side (filter out 10b5-1 plan disposals once that flag exists in the schema).
Cross-check vs clusterParticipantCount ≥ 5: if Pattern A is substantially a re-skin of wide-cluster, drop it. Pre-test: approximately 60-70 percent overlap by row count expected.

Artifacts

scripts/_divergence_v13_4/detect.ts (detection + stats, no DB writes)
scripts/_divergence_v13_4/stats.json (raw numbers backing this doc)