77 — Multi-market rescore final state of affairs (2026-05-19)

Author: rescore orchestration session Date: 2026-05-19 Status: complete

Summary

Full multi-market rescore drained 161k unscored declarations to 0 across the 28-market universe in ~5 minutes (4 sharded market workers + 1 global fast-rescore loop, ~20 node processes peak). Production cache invalidated via /api/cron/weekly-refresh-recos. Recommendations endpoint now returns mixed FR / US / SW / JO / L / KQ tickers, confirming the rescore lifted the AMF-only ceiling documented in 25 §1.

Coverage delta (pubDate >= 2023-01-01, pdfParsed=true, transactionNature not null)

market	total	scored before	pct before	scored after	pct after
BVMF	76 416	8 625	11.3 %	76 416	100 %
XPAR	16 235	16 235	100 %	16 235	100 %
XNAS	12 314	11 689	94.9 %	12 314	100 %
XKRX	8 191	8 189	100 %	8 191	100 %
XTKS	7 555	500	6.6 %	7 555	100 %
XSTO	6 910	6 897	99.8 %	6 910	100 %
XHEL	5 957	5 956	100 %	5 957	100 %
XMAD	5 470	5 470	100 %	5 470	100 %
XSHE	3 868	0	0 %	3 868	100 %
XSHG	3 804	0	0 %	3 804	100 %
XMIL	3 322	3 319	99.9 %	3 322	100 %
XAMS	3 174	3 174	100 %	3 174	100 %
XHKG	2 925	2 925	100 %	2 925	100 %
XCSE	2 622	2 621	100 %	2 622	100 %
XSWX	1 680	229	13.6 %	1 680	100 %
XWBO	1 204	1 202	99.8 %	1 204	100 %
XASX	1 063	149	14 %	1 063	100 %
XBRU	777	777	100 %	777	100 %
XOSL	454	452	99.6 %	454	100 %
XETR	317	315	99.4 %	317	100 %
XLON	195	195	100 %	195	100 %
XBOM	72	72	100 %	72	100 %
XTSE	68	68	100 %	68	100 %
XPHS	12	12	100 %	12	100 %
XNZE	8	4	50 %	8	100 %
XJSE	5	5	100 %	5	100 %
XSAU	4	3	75 %	4	100 %
XSES	1	1	100 %	1	100 %

Global figure (all pubDates, scoring-eligible rows):

• Before: 161 208 unscored.
• After: 0 unscored. 577 228 scored declarations in total.

Sanity check — score distribution per market (pubDate >= 2024-01-01)

market	n >= 40	n >= 50	n >= 60	avg score	max
XPAR	1 928	548	29	28.7	69
XNAS	554	47	0	20.2	59
XSTO	113	9	0	20.8	56
XMAD	90	10	0	19.8	58
XHEL	23	2	0	15.2	54
XMIL	14	0	0	12.0	48
XWBO	5	0	0	14.6	49

XPAR (AMF) retains the deepest tail (sub-strategy uses the full feature set incl. pubDelay <= 7d). Non-FR markets cap around the 50-59 band because the pubDelay feature, the analyst-consensus overlay, and the recent-alpha prior are partially zeroed on cold-start (totalAmount is null for ~30 % of foreign batches, per the batch logs).

Sample non-FR signals (signalScore >= 50, pubDate >= 2024-01-01)

XNAS:

• HARROW INC. / BAUM MARK L (CEO) / 55 / 2026-05-15
• CONDUENT Inc / Agadi Harshavardhan V (CEO) / 54 / 2026-02-23
• Jack Henry & Associates / Adelson Gregory R. (CEO) / 53 / 2026-05-14
• Jack Henry & Associates / Carsley Mimi (CFO) / 53 / 2026-05-14
• On Holding AG / Bernhard Olivier (Exec Officer) / 52 / 2026-05-15
• Strattec Security / Slater Jennifer Lynn (CEO) / 52 / 2026-05-14
• FTI Consulting / Gunby Steven Henry (CEO) / 52 / 2026-05-14

XMAD: 10 signals >= 50, max 58 (cluster CEO/CFO transactions on mid-caps). XSTO: 9 signals >= 50, max 56. XHEL: 2 signals >= 50, max 54.

Impact on /recommendations

GET https://insiders-trades.com/api/recommendations?lookbackDays=90 after cache invalidation returns 10 BUY picks with the following Yahoo-suffix mix:

PA: 3   (Publicis, ALKLH, ABEO)
SW: 2   (WIHN, ALPN)
US: 2   (HROW, TONX)
JO: 1   (TSG)
L:  1   (CDL)
KQ: 1   (Korea KOSDAQ)

7 of 10 picks are non-FR (vs effectively 0 non-FR before the rescore for the non-AMF, non-SEC markets).

90-day reco-eligible non-FR signal counts (signalScore >= 40):

• XNAS 243, XSTO 41, XMAD 19, XMIL 3, XWBO/XJSE/XKRX/XHEL 1 each.

STRATEGY_PROOF updates

• lastUpdatedAt bumped to 2026-05-19T15:30:00.000Z.
• filteredSubsetSize comment updated: signal coverage now spans 28 markets, no longer AMF-only.
• Core backtest figures (universeSize 23 788, filteredSubsetSize 196, V13 ensemble OOS Sharpe 0.66, DSR -0.03) unchanged because those derive from the BacktestResult table (priced T+90 BUY rows), not from signalScore presence. Re-running the V13 bake-off on the broader signalScore-eligible population is a separate follow-up (see Open issues).

Variabilization audit

grep -rE "[0-9]{2,3}\.[0-9]%|Sharpe [0-9]\.|CAGR [0-9]" src/components/landing src/app/performance src/app/methodologie flags only:

• LandingHowItWorks.tsx:169-170 — illustrative card data (+18.4 %, +12.1 %) for the "how it works" educational hover; not a strategy claim.
• performance/page.tsx:470-472 — pedagogical worked example (840 - 700) / 700 = +20.0 %; numbers are arithmetic literals inside a worked sentence.
• methodologie/ChDisclosure.tsx:199 — disclosure quoting the universe figures (44.3 % WR, -0.72 % T+90); these are universe-level facts intentionally stated literally next to the filtered figures.
• methodologie/Ch06Backtest.tsx:342 — label string CAGR 2022-2025, not a value.

No hard-coded strategy CAGR / Sharpe / WR / DD figure escapes STRATEGY_PROOF. Audit clean.

Open issues / next sprint

1. Re-run V13 ensemble bake-off on the full 28-market signalScore-eligible universe (currently the bake-off used FR + SEC + EU only). The point estimate may strengthen because foreign signal coverage is now non-null.
2. Backtest pricing pipeline still has thin coverage on XKRX, XTKS, XHKG, XSHE / XSHG — Yahoo symbol resolution and intraday-vs-EOD timestamping need extending before those signals can carry weight in the OOS Sharpe.
3. Cold-start totalAmount=null rate stays around 20-30 % per batch for foreign rows (logged by the v3 scorer). Backfill totalAmount from the raw filings to lift F1+F2 feature engagement for non-FR signals.
4. XSHE / XSHG (Shenzhen / Shanghai) now scored but average score is 9.5 — the cluster / role / pubDelay features mostly degrade on those filings. Consider a CHN-specific scoring head if mainland China becomes a target.
5. Production recommendations are dominated by SEC issuers in the >= 50 band; add a per-market cap on the top-N selector if regional balance matters for marketing copy.

Workers / scripts of record

• scripts/_fast-rescore.ts — global loop, 500-row batches, 600 s deadline.
• scripts/_fast-rescore-market.ts — per-market shard, 1000-row batches.
• Logs: logs/fast-rescore.log, logs/rescore-{BVMF,XNAS_XNYS,XTKS_XHKG_XKRX,...}.log.
• Progress log: /tmp/rescore-progress.log (2026-05-19 15:22 -> 15:27).