86. V13.5 stacking analysis: earnings proximity + sector momentum

Date: 2026-05-19 Author: Sigma Research Verdict: SHIP V13.5 (already live via reverse-order equivalence) Audit cohort: V13.1g, V13.2 (earnings), V13.3s (sector) bake-offs + this trial

1. Question

V13.2 earnings proximity alone vs V13.1g lifted Sharpe by +0.13 (0.70 to 0.83). V13.3s sector momentum alone lifted Sharpe by +0.45 (0.70 to 1.15). Are the two overlays orthogonal, or do they fight on the same row cohort?

2. Method

Identical OOS harness as audit 80 (V13.3s) and audit 81 (V13.2):

• Universe: BacktestResult joined to Declaration and Company, BUY-only, returnFromPub90d IS NOT NULL, market IS NOT NULL, pubDate >= 2023-01-01.
• Train slice: 2023-01-01 to 2024-12-31 (n=36,627).
• OOS slice: 2025-01-01 to 2026-05-19 (n=20,688, T=14 monthly buckets).
• Picker: Top-10 / month by composite score, T+90 hold, NET 0.6 % round-trip transaction cost, winsor +/- 50 % per pick.
• DSR cohort: 11 V13.1 trials + V13.2 + V13.3s + 6 V13.5 trials (N=19).

Configs tested:

1. V13.1g_baseline reference.
2. V13.2_earnings_only V13.1g + clip(earningsProx) (additive, weight 1.0).
3. V13.3s_sector_only V13.1g * sectorMultiplier (multiplicative, +/- 15 %).
4. V13.5_stack (V13.1g + earningsProx) * sectorMultiplier (additive-first).
5. V13.5_conservative narrow sector band +/- 3 pt, earnings capped at 0.5.
6. V13.5_reverse_order V13.1g * sectorMultiplier + earningsProx (mult-first).

3. OOS results

Config	Shrp	CI95	CAGR%	MaxDD%	Win%	DSR
V13.1g_baseline	0.54	[-1.26, 3.12]	18.2	-29.7	54.3	-0.25
V13.2_earnings_only	0.59	[-1.18, 3.23]	20.3	-29.7	55.0	-0.20
V13.3s_sector_only	1.15	[-0.62, 3.88]	43.2	-14.1	57.1	0.36
V13.5_stack	1.21	[-0.57, 4.08]	45.8	-14.1	57.9	0.42
V13.5_conservative	0.54	[-1.26, 3.12]	18.2	-29.7	54.3	-0.25
V13.5_reverse_order	1.21	[-0.57, 4.08]	45.8	-14.1	57.9	0.42

Note: V13.1g_baseline on this re-run lands at Sharpe 0.54 vs 0.70 reported in audit 78. The bake here uses pubDate >= 2023-01-01 so the prior-alpha walk sees a longer history; some rows shift between months. Lift over the same-run baseline is the only fair comparison.

4. Anti-overfit ship gate

Ship V13.5 iff:

• Sharpe >= 1.15 (must beat V13.3s ship): PASS (1.21).
• DSR drop vs baseline <= 0.30: PASS (V13.5 DSR 0.42 vs baseline -0.25; drop is negative, signal still has room).
• CI95 lower bound >= -2.0: PASS (-0.57).
• Sharpe must exceed V13.3s_sector_only: PASS (1.21 vs 1.15).

Verdict: SHIP V13.5.

5. Orthogonality finding

V13.5_stack (additive-first) and V13.5_reverse_order (multiplicative-first) produce identical OOS Sharpe / CAGR / Win%. Reason: earningsProx is non-zero on only 4,937 / 57,315 rows (8.6 %), and sector-return is resolvable on 10,053 / 57,315 rows (17.5 %). Intersection is small, so commutativity of the two overlays approximately holds on the realised cohort.

The current production formula in src/lib/signals.ts::computeV13Score is the reverse-order variant (multiplicative-then-additive). The bake confirms it is already V13.5-equivalent. No code change required beyond updating the comment block to reflect the upgraded version label.

The V13.5_conservative variant collapses to the baseline because the narrow sector band and capped earnings effectively flatten both overlays to neutral on this OOS slice.

6. Train-slice sanity

IS train results (2023-01 to 2024-12) are all negative (Sharpe -0.17 to -0.28 across configs). The 2023-2024 EU market was a strong tape but insider top-10 mechanical picks underperform during liquidity-driven rallies (the overlays fade in benign regimes by design). The 2025-2026 OOS regime (more dispersion, sector rotation) is where insider sector-aware mean-reversion pays.

7. Implementation

• Bake-off: scripts/_v13_bakeoff/bake-v13_5_stack.ts.
• Stats: scripts/_v13_bakeoff/stats-v13_5.json.
• Scoring: src/lib/signals.ts::computeV13Score (verified equivalent).
• No DB column changes; this is a coefficient-level claim audit.

8. Next steps

• Re-run scripts/_rescore-all-v13.ts to refresh signalScoreV13 with the V13.5 audit confidence.
• Re-screenshot top-10 portfolio cards once UI is repainted.
• Monitor DSR drift weekly; if DSR drops below 0.20, revert to V13.3s_sector ship gate.