23, Multi-market scoring v4 (proposal, not implemented)

Auteur : quant review post-multi-market ingestion Date : 2026-05-16 Status : DESIGN ONLY, src/lib/winning-strategy.ts not modified, no DB writes. Prereqs : audits 11 (scoring backtest), 13 (feature eng.), 19 (SEC Form 4 r3), 20 (multi-market feasibility).

---

TL;DR (one-line composite)

recoScore_v4(d, m) = α_m · signalPts(d) + β_m · winRatePts(d, m) + γ_m · returnPts(d, m) + δ_m · recencyPts(d) + liquidityAdj(d, m)

where (α_m, β_m, γ_m, δ_m) are per-market weight vectors that sum to 100, liquidityAdj ∈ [-5, +5] is a market-median trade-size z-score, and the country risk premium is OFF by default (placeholder).

FR weights are the validated A1 baseline (35, 25, 20, 20). Other markets redistribute when a component cannot be computed.

---

1. Why v4 (not a re-grid-search)

Today's WINNING_STRATEGY + recommendation-engine v6 are tuned ONLY on FR data:

Market	n decl.	named insider	cluster computed	tx date	total amount	signalScore populated	BacktestResult rows
FR (AMF)	25 801	25 760	9 267	25 394	22 354	22 255	15 171
US (SEC)	6 206	6 206	0	6 206	3 456	0	0
DE (BaFin)	295	295	0	295	0	0	0
CH (SIX)	226	226	0	226	226	0	0
UK (RNS)	176	176	0	176	99	0	0

Two hard facts:

1. No signalScore outside FR, the scoring cron is FR-only. A v4 composite that calls into v3 signalScore returns null for ~6 900 non-FR decls. Either compute v3 per-market (recommended) OR fall back to a role-only proxy when null.
2. No BacktestResult outside FR, winRatePts / returnPts cannot be computed for US/DE/CH/UK until the T+90 backfill job runs on those markets. For now their bucket priors fall back to __overall (which is the FR universe, leakage of FR-specific structure). Honest: until the BR backfill runs, non-FR signals are essentially signalPts + recencyPts only.

A blind per-market grid search on n=176 (UK) or n=226 (CH) is statistically indefensible, the FR A1 search used n=15 171 BUY rows and the deflated Sharpe is already negative. v4 therefore inherits FR weights and only redistributes mass when a component is structurally unavailable.

---

2. Per-market component matrix (4 × 6)

✓ = component computes natively · →X = redistributed to component X · prior = uses overall side prior only (no track-record signal) · penalty = applied as a flat point deduction.

Component (max pts → FR baseline)	FR	US	DE	CH	UK	(JP/IT/ES TBD)
signalPts (35), track-record + role + cluster	35	35	30	15	30	n/a
winRatePts (25), Bayes-shrunk bucket	25	25*	25*	25*	25*	n/a
returnPts (20), z-score vs side mean	20	20*	20*	20*	20*	n/a
recencyPts (20), exp decay 21d HL	20	20	25	30	20	n/a
liquidityAdj (±5, off composite)	±5	±5	±5	±5	±5	n/a
Connected-person penalty	,	,	,	,	−5	,
Effective weight sum	100	100	100	100	100	,

* = computed only once the per-market BacktestResult backfill lands; until then winRatePts = 8 cold-start and returnPts = 10 (neutral), consistent with buildRecoItem's existing cold-start branch.

CH adjustment (key): anonymised filings = no per-insider track record. The 20-point signalPts component that came from CMP-2012 pattern + insider cumulative net is structurally unavailable. Redistribute:

• 20 → 15 on signalPts (role weight + cluster proxy via same-day same-issuer filings can still fire if we synthesize a CH-specific cluster window).
• +10 to recencyPts (CH SER filings have a 4-day publication SLA, freshness carries more information than elsewhere).
• net redistribution: −20 from signalPts, +10 to recencyPts, +10 absorbed by role-weight inside signalPts itself (keep at 15 ceiling).

UK adjustment: PMR connected-persons (spouse, controlled entity) file under the PDMR identity, polluting track-record. Apply a flat −5 penalty on signalPts when RNS.connectedPerson === true (parsed from RNS body regex).

DE adjustment: BaFin has no totalAmount natively (audit 20 shows 0/295 have amount populated, derived qty×price still missing). Skew toward recency until the price-per-share backfill lands.

---

3. Liquidity / market-size normalization

Cross-market amounts are already EUR-normalized (FX history landed in audit 19). But absolute EUR amounts are biased: a 1 M€ US insider buy is ~5× more common than a 1 M€ FR buy. Without normalization, the cross-market top-N is dominated by US tickets.

Proposed liquidityAdj (signed, ±5 pts, OUTSIDE the 100-point composite):

mtm = log10(totalAmountEur / marketMedianTradeSizeEur[m])
liquidityAdj = clamp(mtm * 2, -5, +5)

Market medians (to compute once at startup from BacktestResult or Declaration):

Market	Median EUR ticket (estim.)	Notes
FR	~25 k€	AMF small-cap skew
US	~80 k€	SEC universe is mostly large-cap
DE	~120 k€	DAX/MDAX large-cap
CH	~30 k€	SMI mid-cap heavy
UK	~40 k€	RNS mid-cap heavy

Estimates only, to be replaced by SELECT percentile_cont(0.5) WITHIN GROUP (ORDER BY totalAmount) FROM "Declaration" GROUP BY market once amounts populate everywhere.

---

4. Country risk premium (placeholder, OFF)

crp = {fr: 0, us: 0, de: 0, ch: -2 (CHF havenness premium), uk: +1 (post-Brexit liquidity discount)}

Flagged here for completeness, would need its own walk-forward on ≥3 years of US/DE/CH/UK backtest data before being switched on. Not in v4 ship.

---

5. Validation, what's measurable today vs what isn't

5.1 Measurable now (FR only)

The FR A1 baseline already cleared the deflated-Sharpe gate (audits 11 + 15); v4 with α_fr,β_fr,γ_fr,δ_fr = (35,25,20,20) is identical to A1 on FR rows by construction. No regression risk on the validated market.

5.2 Not measurable today (US/DE/CH/UK)

Per the table in §1: 0 BacktestResult rows on non-FR markets. A T+90 backtest at composite ≥ 60 cannot be computed for US/DE/CH/UK until the scripts/backfill-backtest-results.mjs (or equivalent) runs on the imported declarations.

This is the honesty line of the proposal. v4 should ship behind a feature flag with composite ≥ 60 surfaced ONLY on FR until per-market T+90 stats are in. Non-FR markets surface a "preview, unvalidated" badge.

5.3 Sample-size limits per market

Market	n	A grid-search would need…	Verdict
FR	25 801	n ≥ 5 000 per bucket × 6 trials	✓ done
US	6 206	n ≥ 5 000, borderline if not stratified by sector	acceptable for the same A1 weights, NOT for re-tuning
DE	295	needs 5 000+	too small, inherit FR weights, no per-market tune
CH	226	same	too small, adjustment from §2 is structural, not statistical
UK	176	same	too small, same

v4 explicitly does NOT re-tune weights on small markets. The §2 adjustments are structural (which component is computable), not empirical. This is the only intellectually defensible move at n < 1 000.

---

6. Implementation plan, minimal diff

Three concrete code changes (file:line numbers from current HEAD):

Change 1, accept market: Market argument

File: src/lib/winning-strategy.ts:489

// BEFORE
export async function getWinningStrategySignals(opts: {
  limit?: number;
  lookbackDays?: number;
} = {}): Promise<WinningSignal[]> { ... }

// AFTER
export async function getWinningStrategySignals(opts: {
  limit?: number;
  lookbackDays?: number;
  market?: MarketFilter;   // default "fr" to preserve current behavior
} = {}): Promise<WinningSignal[]> { ... }

Wire marketPrismaWhere(opts.market ?? "fr", "amfId") into the where block at line 316. No change to thresholds or filters.

Change 2, per-market weight selector for recommendation-engine.ts

File: src/lib/recommendation-engine.ts:466-498 (in buildRecoItem)

Introduce a marketWeights table keyed by Market:

const MARKET_WEIGHTS: Record<Market, { sig: number; wr: number; ret: number; rec: number }> = {
  fr: { sig: 35, wr: 25, ret: 20, rec: 20 },
  us: { sig: 35, wr: 25, ret: 20, rec: 20 }, // inherit until BR backfill validates
  de: { sig: 30, wr: 25, ret: 20, rec: 25 },
  ch: { sig: 15, wr: 25, ret: 20, rec: 30 + 10 /* role+ceiling */ },
  uk: { sig: 30, wr: 25, ret: 20, rec: 20 }, // +UK connected-person penalty applied below
};

Replace the four (decl.signalScore ?? 0) / 100) * 35 style hard-coded weights with MARKET_WEIGHTS[market].sig etc. Add a −5 penalty branch when market === "uk" && decl.isConnectedPerson === true (field to be added by RNS parser; until then, no-op).

Change 3, liquidity adjust as a post-hoc bump

File: src/lib/recommendation-engine.ts:496 (after recoScoreRaw line)

const liqAdj = computeLiquidityAdj(decl.totalAmount, MARKET_MEDIANS[market]);
const recoScore = Math.round(Math.max(0, Math.min(100, recoScoreRaw + liqAdj)));

MARKET_MEDIANS is a hard-coded constants table refreshed monthly from the SQL percentile query in §3. Bounded ±5 → cannot push a 95-score above 100 nor a 5-score below 0.

---

7. Edge cases

• CH role-only: no insider FK history → signalPts uses role weight only. The tradingPattern slot (CMP-2012) cannot be filled, return pattern: "new" for all CH insiders (same as < 3 history branch in insider-pattern.ts:55).
• Connected-person UK: needs an isConnectedPerson boolean on Declaration (parser change in scripts/ingest-rns.mjs). Until shipped, apply nothing, graceful no-op.
• Missing totalAmount (DE): liquidityAdj = 0. Don't subtract pts when the data is missing, the user would see DE consistently penalised for an ingestion gap, not a real liquidity signal.
• Missing signalScore (all non-FR today): cold-start path in buildRecoItem:472 already handles winRate == null. We also need signalPts = (decl.role-weight-fallback) when signalScore is null rather than 0. Suggestion: signalPts = roleBasePts(role) * α_m / 100 where roleBasePts(CEO) = 20, (CFO) = 18, (director) = 12. Honest ceiling: ~20/35, non-FR signals top out at composite ~60 until the scoring cron runs cross-market.

---

8. Risk register (markets too small for independent calibration)

Market	n	Risk if we re-tune	v4 mitigation
DE	295	overfitting to one BaFin sub-cohort (Allianz dominates)	inherit FR weights
CH	226	overfit to a handful of SIX issuers	inherit + §2 structural adj.
UK	176	RNS regex parser quality > weight choice	inherit + connected-person flag
US	6 206	acceptable for validation; tempting to re-tune	inherit until cross-market BacktestResult available; THEN run ONE grid search across pooled US+FR (n ≈ 21 k)

Hard rule: do not run a per-market grid search at n < 5 000. The Bailey– López de Prado penalty √(2·ln N_trials / T) is already negative-driving on the FR universe with T=4 yearly buckets; on T=1-2 for small markets it would nuke any "tuned" Sharpe.

---

9. Ship checklist (when implementing)

• Add Market arg to getWinningStrategySignals + plumb marketPrismaWhere.
• Add MARKET_WEIGHTS + MARKET_MEDIANS constants tables.
• roleBasePts fallback for null signalScore.
• Feature flag MULTI_MARKET_RECOS_ENABLED (default OFF). Surface non-FR recos with a "preview, unvalidated T+90" badge until BR backfill lands.
• Per-market BacktestResult backfill job. Once T+90 is populated for US (n ≈ 6 k), re-run the audit-11 harness pooled FR+US and confirm A1 weights still win.
• CH isConnectedPerson + UK penalty parser only after the RNS body regex is validated against ≥ 50 hand-labeled samples.
• Re-run audit 12 walk-forward with the pooled FR+US universe, if Sharpe ann. ≥ 0.4 on US-only fold, ratify A1 weights for US. Else, revert to FR-only surfacing.

---

10. What this proposal does NOT do

• Does not modify src/lib/winning-strategy.ts (per task constraint).
• Does not re-run a grid search on small markets, n is too low.
• Does not turn on the country risk premium, placeholder only.
• Does not claim per-market validation, §5.2 explicitly states US/DE/CH/ UK T+90 win rate cannot be computed until the BR backfill ships.

The single defensible claim: v4 is identical to A1 on FR (the validated slice) and degrades gracefully on markets where components are structurally missing. Anything stronger requires the cross-market BacktestResult backfill called out in §9.