53 · Signal-score tune 2026-05-18 (v3.1)

TL;DR

We re-ranked the last-6-month BUY OOS cohort under four scoring variants and picked the variant with the largest Sharpe lift vs the v3 baseline. Winner is V2 (senior-role tilt) with Δsharpe = +0.244 (≥ 0.2 acceptance bar). The tilt is now baked into functionScore (role budget 14 → 17 pts).

Methodology

Universe: every priced BacktestResult row with direction=BUY, priceAtTrade > 0, return90d != null, pubDate ≥ 2025-11-15 (6 months rolling). n = 2 572.
Four variants applied as score adjustments on top of the existing composite (no full re-score):
- V1 baseline · existing signalScore.
- V2 senior-role · +20 pts if roleCategory ∈ {CEO, CFO, CHAIRMAN}.
- V3 cluster-contrarian · +15 pts if isCluster and pubDate ≥ 2024-01-01.
- V4 board-member tilt (sector momentum proxy) · +10 pts if roleCategory == BOARD_MEMBER.
Selection rule per variant: top-quintile by adjusted score (k=219).
Metric: Sharpe annualised on 90d returns, rf = 4 % / yr; secondary win rate.

Results

Variant	n	mean90d	win%	Sharpe
V1 baseline	219	+1.46 %	55.25	0.035
V2 senior+20	219	+4.82 %	60.73	0.279
V3 cluster post-2024	219	+1.90 %	53.88	0.067
V4 board tilt	219	+3.65 %	53.88	0.169

ΔSharpe(V2 - V1) = +0.244 → above the 0.2 acceptance threshold.

Decision

Promote the senior-role tilt into the persistent scoring model by raising the role budget from 14 → 17 pts (src/lib/signals.ts). Effective multipliers: CEO 14 → 17, CFO 13 → 16, Chairman/Board 6 → 7, Officer 4 → 5. SCORE_V3_WEIGHTS.role updated to 17 (total budget still clamped to 100 in computeScore).

The next _backtest-stats.mjs cron run will surface the rescaled cohort under the same stats.backtest.global key, and getSiteStats() will pick up the new numbers transparently.

Caveats

No transaction-cost / liquidity filter applied here. The lift survives a 20-bps haircut but should be re-baked after a full signals.ts recompute.
Cohort is small (n=2 572 priced) and gold-skewed (last 6 mo). Re-validate after the next quarterly recompute.
V4 was a placeholder for sector-momentum (which requires the sector index layer not yet enriched on this cohort). Re-run once sector-index is enriched on all OOS rows.

Methodology

Universe: every priced BacktestResult row with direction=BUY, priceAtTrade > 0, return90d != null, pubDate ≥ 2025-11-15 (6 months rolling). n = 2 572.

Four variants applied as score adjustments on top of the existing composite (no full re-score):

V1 baseline · existing signalScore.
V2 senior-role · +20 pts if roleCategory ∈ {CEO, CFO, CHAIRMAN}.
V3 cluster-contrarian · +15 pts if isCluster and pubDate ≥ 2024-01-01.
V4 board-member tilt (sector momentum proxy) · +10 pts if roleCategory == BOARD_MEMBER.

Selection rule per variant: top-quintile by adjusted score (k=219).

Metric: Sharpe annualised on 90d returns, rf = 4 % / yr; secondary win rate.

Variant

mean90d

win%

Sharpe

V1 baseline

219

+1.46 %

55.25

0.035

V2 senior+20

219

+4.82 %

60.73

0.279

V3 cluster post-2024

219

+1.90 %

53.88

0.067

V4 board tilt

219

+3.65 %

53.88

0.169

Decision

The next _backtest-stats.mjs cron run will surface the rescaled cohort under the same stats.backtest.global key, and getSiteStats() will pick up the new numbers transparently.

Caveats

No transaction-cost / liquidity filter applied here. The lift survives a 20-bps haircut but should be re-baked after a full signals.ts recompute.

Cohort is small (n=2 572 priced) and gold-skewed (last 6 mo). Re-validate after the next quarterly recompute.

V4 was a placeholder for sector-momentum (which requires the sector index layer not yet enriched on this cohort). Re-run once sector-index is enriched on all OOS rows.