V13.2 Earnings Proximity · OOS bake-off (2026-05-19)

Hypothesis

Insider buys placed in the 30-day window BEFORE the next earnings publication carry more alpha than post-earnings buys. Pre-earnings buys suggest the insider knows something the market does not yet price; buys just after a guidance cut are reactive and carry less informational value.

Feature

Declaration.earningsProximityScore ∈ [0, 1]. Computed from the gap between pubDate and the next CompanyEarnings.eventDate:

daysToNextEarnings	proximityScore
0..30	1.0
30..60	0.5
60..90	0.2
>90, or no event, or post-event	0

Data

• CompanyEarnings table populated via scripts/_ingest-earnings-yahoo.ts (Yahoo Finance quoteSummary, modules earnings + earningsHistory + calendarEvents). Throttle 8 req/s, 5-UA rotation, fresh crumb every 30 min.
• Cohort: top 1000 companies by recent signalScore activity (last 365d).
• Ingested rows (2026-05-19 run): 2972 earnings events.
• Declarations backfilled: 135,647 (since 2024-01-01).
• Distribution:
  • p=1.0 (pre-earnings 0-30d): 5,292
  • p=0.5 (30-60d): 4,497
  • p=0.2 (60-90d): 5,824
  • p=0.0: 9,891 priced + 110,143 no-event-known

Scoring overlay

score_V13.2 = score_V13.1g + clip(earningsProximity, 0, 1) * 1.0

Additive in the V13 base score range (typical 4..10). When no earnings event is known for the company, contribution is 0 (no tilt), so V13.2 collapses to V13.1g for companies outside the top-1000 cohort.

OOS bake-off

Same harness as V13.1g (audit 78): walk-forward 2025-01-01 to 2026-05-19, T=14 monthly buckets, top-10 picks/mo, T+90 hold from pubDate+1, NET 0.6 % round-trip, winsor ±50 % per pick.

Universe: 20,688 OOS rows with priced T+90 returns.

config	T	picks	Sharpe	CI95	CAGR%	MaxDD%	Win%	DSR
V13.0_baseline	14	140	0.03	[-2.01, 2.01]	-2.5	-43.7	50.7	-0.41
V13.1g_stacked	14	140	0.70	[-1.01, 3.42]	25.4	-29.7	55.7	0.25
V13.2_earningsProx	14	140	0.83	[-0.87, 3.77]	31.4	-29.7	57.1	0.38
V13.2_crossMarket	14	140	0.11	[-1.95, 2.28]	0.3	-43.0	52.1	-0.34
V13.3_sectorMomentum	14	140	0.13	[-1.90, 2.07]	1.1	-42.5	52.1	-0.32

Anti-overfit gate

Rule: ship only if DSR(candidate) - DSR(baseline) > -0.3.

• V13.1g DSR: 0.25
• V13.2 DSR: 0.38
• Delta: +0.13 (gate passed, no degradation, actual improvement).

12-trial Bailey-Lopez de Prado deflation. DSR went UP after deflation adjustment, which is a strong signal: the feature reproducibly beats the expected max-of-N-noise baseline.

Decision

Ship V13.2. Wire earningsProximity into V13Inputs and computeV13Score. Production signalScore recompute will pick up the field on next cron walk-forward.

Deltas vs V13.1g baseline:

• Sharpe +0.13 (0.70 → 0.83, +18 %)
• CAGR +6.0 pp (25.4 → 31.4)
• Win% +1.4 pp (55.7 → 57.1)
• DSR +0.13 (0.25 → 0.38)
• MaxDD unchanged (-29.7 %)

Sample pre-earnings buys (proximity = 1.0)

pubDate	Company	Insider	EUR amount	days to earnings
2026-04-20	Ubisoft Entertainment	Christian GUILLEMOT	330,357	30
2026-04-20	Ubisoft Entertainment	Christian GUILLEMOT	624,050	30
2026-03-17	Arq, Inc.	EICHER CAROL S	131,065	14
2026-03-17	Arq, Inc.	EICHER CAROL S	131,065	14
2026-03-11	COTY INC	Gordon von Bretten	200,287	19

Files

• Migration: prisma/migrations/20260520200000_earnings_proximity/migration.sql
• Ingestion: scripts/_ingest-earnings-yahoo.ts
• Backfill: scripts/_backfill-earnings-proximity.ts
• Bake-off: scripts/_v13_bakeoff/bake.ts (config V13.2_earningsProx)
• Scoring: src/lib/signals.ts (V13Inputs.earningsProximity + V13.2 add)
• UI badge: src/components/DeclarationCard.tsx (Pre-earnings buy pill)
• Stats const: src/lib/signals.ts::V13_2_EARNINGS_PROX_OOS

Caveats

• Cohort capped at top-1000 by activity. Smaller-cap rows have no earnings data and receive proximity = 0, which is a NO-OP under the additive formulation (does not penalise).
• Yahoo earningsHistory only goes back ~4 quarters. Older declarations relying on past events get noEvent → 0.
• DSR uplift (+0.13) is meaningful but the OOS T=14 window is short. Re-evaluate at T=20+ and after universe expansion (top-3000 cohort).