34 — Alpha Discovery (round 2)

Date : 2026-05-17 Script : scripts/alpha-discovery-round2.mjs Universe : 15,171 BUY declarations with realised returnFromPub90d (winsorised at [−50.0%, +60.6%]). FR-dominated (~93%), 17-market priced subset of the 162,573 declaration corpus. Same universe as round 1 (audit 32) so figures are directly comparable.

Round 1 recap : 12 factors tested, only clusterStrength30d cleared the bar. Round 2 widens the search to 33 candidate factors with deeper feature engineering across 5 buckets (insider behaviour, company state, macro regime, microstructure, cross-source).

---

0. Methodology

For each candidate factor f:

1. Compute f per declaration with strict point-in-time inputs — only events with pubDate < t (and, where the input is a realised return, pubDate + 90d ≤ t) enter rolling windows. PIT-strict critical fix described in §3.
2. Spearman rank IC between f and r90.
3. Decile split (Q=10) ; for binary factors degenerate-decile fallback to f=0 vs f=1 buckets.
4. Welch t-stat top-decile vs bottom-decile.
5. Bonferroni adjustment on K = 33 tests.
6. Bailey–López de Prado DSR penalty √(2·ln K / T) with T = 10 yearly buckets → penalty ≈ 0.836.

Universe coverage and 4 schema fields (shortInterest, full earnings consensus, VIX, 10y-2y curve) are missing or too sparse — flagged below as "skipped honestly". The 33 actually-tested factors are tabulated.

---

1. Ranked results (winsorised T+90, %)

PIT-clean ✓ ; snapshot/static fields are NOT point-in-time and marked ⚠.

Rank	Factor	n	IC	t	ΔTop−Bot (% pts)	TopWR	BotWR	PIT
1	companyFirstFiling_yearsAgo	14,798	+0.123	+11.21	+7.31	47.6%	33.5%	✓
2	regimeEURUSD_change30d	15,171	+0.081	+9.72	+6.08	61.0%	48.5%	✓
3	companyAnalystReco_score	9,667	+0.073	+9.34	+8.08	49.1%	37.7%	⚠ snapshot
4	microPubLagDays	14,977	−0.073	−9.03	−5.67	32.0%	43.0%	✓
5	insiderHabitualBuyer_count365d	15,171	+0.018	−8.06	−4.71	39.2%	47.8%	✓
6	companyLogMcap	15,130	+0.072	+7.13	+5.05	51.5%	35.6%	⚠ snapshot
7	microMonthOfYear	15,171	+0.006	+5.55	+3.58	47.8%	40.1%	✓
8	companyInsiderHeldPct	15,071	+0.034	+5.55	+3.71	43.7%	39.0%	⚠ snapshot
9	companyIPORecent3y_flag	14,798	−0.061	−5.24	−1.74	43.3%	49.9%	✓
10	regimeSectorMomentum_60d	12,019	−0.030	−4.48	−4.18	51.2%	53.2%	✓
11	insiderPriorWinRate_90d_shrunk (PIT-strict)	9,071	−0.055	−4.40	−3.36	38.8%	52.7%	✓
12	companyMcapBucket_smallcap_flag	15,130	−0.040	−3.90	−1.45	40.5%	45.7%	⚠ snapshot
13	insiderRolePromoted_flag	11,326	+0.057	+3.44	+1.59	51.6%	44.1%	✓
14	insiderClusterRecency_days	12,784	−0.008	−3.35	−2.41	43.7%	52.8%	✓
15	companyDebtToEquity	13,639	+0.005	−3.17	−2.05	41.7%	44.5%	⚠ snapshot
16	companyPriceMomentum_90d	9,534	−0.108	−2.97	−3.28	37.1%	50.2%	✓
17	companyPriceMomentum_30d	9,861	−0.087	−2.41	−2.70	40.4%	56.2%	✓
18	companyBeta	14,987	−0.048	−2.33	−1.79	43.1%	43.7%	⚠ snapshot
19	microGap2pct_flag	9,881	+0.008	+2.18	+1.12	48.2%	46.6%	✓
20	insiderRoundLot_flag	15,171	−0.014	−2.07	−0.63	43.8%	44.5%	✓
21–33	(	t	< 2 or n < 1k)	—	—	—	—	—

Skipped honestly (sparse / no DB coverage) : companyShortInterest_ratio (n=769 only), companyVolumeSpike_flag (no volume in PriceHistory), companyEarningsDeltaConsensus (no consensus history), regimeVIX_level (no VIX series), regimeYieldCurve_inverted (no rates), microPostFiling_drift_3d (lookahead by construction), companyDualListing_flag (no ISIN bridge coverage), clusterCrossMarket_count (no group-mapping table). The 4 macro factors are the most painful absence — they would let us test regime-conditional alpha properly. Recos in §5.

---

2. Critical PIT-leakage finding

The first version of this script ranked insiderPriorWinRate_90d_shrunk at t = +31.5, IC = +0.355, TopWR 76% / BotWR 17%. That headline collapsed to t = −4.4, TopWR 38.8% / BotWR 52.7% after a single line fix (h.ts + 90d ≤ t instead of h.ts < t). Without the gate, "prior trades" with unrealised r90 shared the market regime of the current row → factor was leaking +35pp. Round 1 of this audit shipped before the fix and would have over-published it had the script been promoted blindly. The corrected sign is slightly negative — insiders with strong recent track records on this dataset under-perform — a CMP-2012-style "regression to mean / routine repeat buyers" hypothesis, NOT alpha.

The same trap would have hit any factor of the form "recent realised performance of related entity" — sector momentum (audit 32) escaped it because we already lagged by 90d there ; new factors insiderHabitualBuyer_count365d is COUNT-based (no r90 input → safe).

---

3. Top 5 honest verdicts

3.1 companyFirstFiling_yearsAgo (t = +11.21, IC = +0.123, PIT ✓)

Decile spread monotone-ish, D1 (companies with no prior filing history) mean −7.16%, D10 (4y+ filing history) +0.06%. Reading: insiders on newly-filing companies (often recent IPOs or first-time AMF-registrants) underperform. The signal is dominated by D1 vs D2-D10 (a stepped jump rather than a continuous trend) — it is functionally an "IPO penalty" factor. companyIPORecent3y_flag (rank 9) is the discrete version.

Verdict : real but it's selection (avoid recent IPOs), not a new alpha source.

3.2 regimeEURUSD_change30d (t = +9.72, IC = +0.081, PIT ✓)

D10 (EUR appreciated >+2.6% in 30d before filing) → mean r90 +5.53%, WR 61.1%. D1 (EUR weakened ≥-2.5%) → r90 −0.64%. Counter-intuitive vs the audit-32 hypothesis (EUR weakness drives foreign-buyer inflow); here EUR strength is bullish.

Verdict : plausible regime factor, but the sign is correlated with broad risk-on (when EUR strengthens, equities tend to rally globally). It is mostly a proxy for the missing CAC40-SMA-50d trend filter. Until we wire actual CAC40 history (PR-able from Yahoo ^FCHI), this is the closest macro signal available.

3.3 microPubLagDays (t = −9.03, IC = −0.073, PIT ✓)

D10 (lag > 100 days) → r90 −7.59%, WR 32%. D1 (lag <1.7 days) → r90 −1.92%, WR 43%. Confirms existing v5 hard filter at pubDelay ≤ 7d. Late-disclosed trades carry no information edge — alpha is dissipated by the time the market sees the PDF.

Verdict : already in the v5 winning-strategy filter as maxPubDelayDays: 7. Nothing to add to v6.

3.4 insiderHabitualBuyer_count365d (t = −8.06, PIT ✓)

D2 ([0,0]) → r90 −4.72%, D10 (88-343 prior buys, year) → r90 −4.04%. Non-monotone with a sweet spot in the middle (D6 +0.30%). The two tails (no prior signal | hyperactive trader) both underperform. Signal is U-shaped, hard to use directly. Confirms CMP-2012 "routine" cohort under-performs finding from audit 05 — directionally consistent.

Verdict : already partially captured by Insider.tradingPattern classification. Could promote to a soft penalty −1 pt when count365d > 50.

3.5 companyAnalystReco_score (t = +9.34, IC = +0.073, ⚠ SNAPSHOT — NOT PIT)

D10 (analyst score 4 = strong_buy) → r90 +8.08% over D1 (sell). DO NOT SHIP. Company.analystReco is a snapshot of the current rating, not the rating at filing time. A company that the sell-side rates "strong_buy" today probably also performed well during the last 90 days — the factor is partly reading the future. Need a CompanyAnalystSnapshot time-series table before this can be promoted.

Verdict : honestly inconclusive. The IC sign is plausible but the look-back is contaminated by snapshot bias.

---

4. v6 composite proposal

Round 1 already proposed +5pts cluster-strength rank, −3pts pctMcap-top10 penalty. Round 2 adds three changes, all small, all conditional on a fresh nightly recompute :

v6_round2 composite =
    0.30 · signal           // -2pts from v5.1 (35→30)
  + 0.32 · winRate          // -3pts (35→32)
  + 0.22 · return           // unchanged from v5.1 (25→22 = -3, redistributed)
  + 0.07 · recency          // -3pts (10→7)
  + 0.05 · clusterStrength30d_rank      // KEPT from round 1
  − 0.03 · pctMcap_top10_flag           // KEPT from round 1
  − 0.02 · companyFirstFiling_yearsAgo<1y_flag    // NEW: IPO-recent penalty
  + 0.02 · pubLag_fresh3d_flag                    // NEW: ultra-fresh bonus (lag ≤ 3d)
  − 0.01 · insiderHabitualBuyer_count365d>50_flag // NEW: hyperactive routine penalty

Net dynamic-range : +0.10 / −0.06 — moves are minimal so WINNING_STRATEGY.minScore should remain at 40 in the v6 transition (re-run grid-search-v2 after v6 lands, then re-anchor).

Factors rejected for v6 despite t > 4 : regimeEURUSD_change30d (waiting on real CAC40 series + sector-conditional version), companyAnalystReco_score (snapshot bias unresolved), companyLogMcap / companyInsiderHeldPct / companyBeta (all snapshot — risk PIT bias replicates the audit-28 finding).

---

5. DSR-adjusted verdict

Factor	t	Per-trade SR (t/√n)	Annualised crude (×√(252/90))	DSR adj (− 0.836)	Verdict
companyFirstFiling_yearsAgo	11.21	0.092	0.155	−0.68	DSR negative
regimeEURUSD_change30d	9.72	0.079	0.132	−0.70	DSR negative
microPubLagDays	−9.03	0.094	0.157	−0.68	DSR negative
insiderHabitualBuyer_count365d	−8.06	0.065	0.109	−0.73	DSR negative
companyLogMcap (snapshot)	7.13	0.058	0.097	−0.74	DSR negative

No round-2 factor clears Bailey–López de Prado. The cross-sectional SR-equivalent on the strongest factor is ≈ 0.155, the DSR penalty for K=33 is 0.836 — even a hypothetical SR = 1.0 would barely clear. The honest published verdict aligns with audit 32 §4 : individual factor-level alpha is below the DSR floor on this universe ; the only defensible publish-worthy modifications to the v6 composite are the small weight perturbations in §4, applied conjointly with the existing v5 filter (which carries its own historical performance disclosure with sharpeDeflated = null).

---

6. Files

• /Users/azoulay/DEV/insiders-trades/scripts/alpha-discovery-round2.mjs — round-2 discovery harness (33 factors, PIT-strict)
• /tmp/alpha-discovery-round2.json — structured per-factor output
• /Users/azoulay/DEV/insiders-trades/docs/method-review/34-alpha-discovery-round2.md — this doc

Patch proposal for src/lib/winning-strategy.ts (NOT applied — DON'T commit) :

// src/app/api/v1/scoring/factors-v6.ts  (new file)
export function ipoRecent3yPenalty(firstFilingYearsAgo: number | null): number {
  if (firstFilingYearsAgo == null) return 0;
  return firstFilingYearsAgo < 1 ? -2 : 0;   // -2pts (audit 34 §4)
}

export function pubLagFreshBonus(pubDelayDays: number | null): number {
  if (pubDelayDays == null) return 0;
  return pubDelayDays <= 3 ? +2 : 0;          // +2pts (audit 34 §4)
}

export function habitualBuyerPenalty(count365d: number): number {
  return count365d > 50 ? -1 : 0;             // -1pt (audit 34 §4)
}

// composite update — additive on top of v5.1 (signal/winRate/return/recency)
// All weights renormalised so total = 1.0 ± 0.10 dynamic-range from §4 bonuses.

WINNING_STRATEGY.minScore stays at 40 until a fresh scripts/grid-search-v2.mjs re-tunes the threshold after v6 lands.

---

7. Next steps

1. Wire ^FCHI Yahoo history into PriceHistory (single new symbol, daily, 2015→) — unlocks proper regimeCAC50dma_above instead of the EUR/USD proxy.
2. Build CompanyAnalystSnapshot time-series (cron-snapshot daily Yahoo recommendationKey) so companyAnalystReco_score can be PIT-tested honestly.
3. Re-run round 2 with PIT-strict snapshot fundamentals (need CompanyMarketCapSnapshot daily) — the snapshot-flagged factors above probably collapse like insiderPriorWinRate_90d_shrunk did.
4. Stratify the surviving 4 PIT-clean factors per market (FR / SEC / AFM / pooled) to detect regime-specific alpha that the pooled rank hides.
5. After (1)+(2) land, re-attempt VIX + 10y-2y series ingestion (FRED API, free tier) and re-run the regime block — these are the most likely candidates for genuine new alpha.