43 · Scoring rerun on enriched dataset (2026-05-17)

Walk-forward backtest on 162 k Declarations · 23,788 BUY rows with r90. Train window 2015-2024 (rolling 24-month priors, rebuilt quarterly). Test window 2025-01-01 to 2026-05-17.

Methodology

Universe: every Declaration with a populated BacktestResult.returnFromPub90d and a parseable BUY direction.
Scoring:
- v5.1 (production) — composite score (sig / wr / ret / rec), market-aware weights from MARKET_WEIGHTS table.
- v6 (candidate) — v5.1 + role-quality multiplier (CEO/CFO +5 pct), cluster premium (+10 pct), pub-delay freshness penalty (>14 days x 0.85).
Portfolio: weekly top-10, equal-weight, hold 90 days. Monthly portfolio returns = arithmetic mean of picks within month.
Stats: Sharpe / Sortino annualised from monthly returns (x sqrt 12). Calmar = annualised return / |maxDD|. Max DD on the monthly NAV. Hit rate = pct of picks with r90 > 0. Turnover = avg fraction of top-N replaced week-on-week.
DSR (Bailey-Lopez de Prado): SR_def = SR - sqrt(2 ln(N)/T) with N = 9,420 effective trials (anchored on v5.1 grid 314 weights x alpha-r3 30 features) and T = monthly obs.
Bootstrap CI95: 1000 resamples on monthly returns, seed 42, annualised by x sqrt 12.

v5.1 (production) walk-forward

Market	Picks	Sharpe_ann	Sortino	Calmar	maxDD %	HitRate %	Turnover %	DSR	CI95
FR	595	1.685	3.928	9.649	-19.9	52.4	99.8	0.648	[0.385, 3.076]
US	133	1.367	1.761	4.673	-18.1	54.1	98.5	0.014	[-0.831, 4.160]
NL	484	1.840	3.594	7.371	-6.4	58.1	100.0	0.736	[0.305, 3.523]
BE	225	0.836	0.967	1.233	-35.0	60.0	92.6	-0.233	[-0.714, 2.820]
COMBINED	662	1.319	2.015	4.090	-17.9	52.0	99.6	0.311	[-0.076, 2.785]

v6 (candidate) walk-forward

Market	Picks	Sharpe_ann	Sortino	Calmar	maxDD %	HitRate %	Turnover %	DSR	CI95
FR	595	1.246	2.473	5.469	-25.0	49.6	99.6	0.209	[-0.319, 2.588]
US	133	1.334	1.705	4.490	-18.1	54.1	98.5	-0.019	[-0.932, 4.098]
NL	484	1.840	3.594	7.371	-6.4	58.1	100.0	0.736	[0.305, 3.523]
BE	225	0.836	0.967	1.233	-35.0	60.0	92.6	-0.233	[-0.714, 2.820]
COMBINED	662	0.981	1.567	2.370	-27.4	47.3	99.8	-0.028	[-0.634, 2.480]

A/B summary v5.1 vs v6

Market	dSharpe	dSortino	dCalmar	dMaxDD pp	dHitRate pp	dDSR
FR	-0.439	-1.455	-4.180	-5.08	-2.86	-0.439
US	-0.033	-0.056	-0.183	0.00	0.00	-0.033
NL	0.000	0.000	0.000	0.00	0.00	0.000
BE	0.000	0.000	0.000	0.00	0.00	0.000
COMBINED	-0.338	-0.448	-1.720	-9.46	-4.68	-0.339

MARKET_WEIGHTS — refresh proposal

The pooled FR+SEC winner (30 / 35 / 25 / 10) was calibrated when AFM=4 330 / SEC=2 391. SEC is now 126 607 declarations and AFM is 6 392. AMF (FR) BUY at T+90 was mean -0.87 pct (negative). We re-grid-searched a small per-market neighbourhood (sig in {25,30,35,40}, wr in {30,35,40}, ret in {15,20,25}, rec in {5,10,15}) and kept the v5.1 baseline as the floor.

Deltas (only markets where the local optimum beats current by Sharpe_ann > 0.05 AND DSR remains best-in-class):

Market	Current	Proposed	Notes
FR	30/35/25/10	(kept) 30/35/25/10	neighbour search did not clear honesty threshold (dSharpe > 0.05)
US	30/35/25/10	(kept) 30/35/25/10	neighbour search did not clear honesty threshold (dSharpe > 0.05)
NL	30/35/25/10	(kept) 30/35/25/10	neighbour search did not clear honesty threshold (dSharpe > 0.05)
BE	30/35/25/10	(kept) 30/35/25/10	neighbour search did not clear honesty threshold (dSharpe > 0.05)

Decision: keep MARKET_WEIGHTS unchanged. The dataset growth confirms the v5.1 plateau result from doc 11; per-market neighbour deltas remain inside noise after DSR.

Recommendation: v5.1 vs v6

v6 must beat v5.1 on DSR-deflated Sharpe AND on at least 3 of {Sortino, Calmar, hit rate, max DD} on the COMBINED portfolio.

COMBINED:

v5.1: Sharpe_ann = 1.319 · DSR = 0.311 · Sortino = 2.015 · Calmar = 4.090 · HitRate = 52.0 pct · MaxDD = -17.9 pct
v6: Sharpe_ann = 0.981 · DSR = -0.028 · Sortino = 1.567 · Calmar = 2.370 · HitRate = 47.3 pct · MaxDD = -27.4 pct

Recommendation: keep v5.1. DSR delta = -0.339. Secondary wins = 0/4 (need >= 3 AND positive DSR delta). v6 archived behind USE_SCORING_V6 flag (default off).

Reproducibility

node --env-file=.env.local --import tsx scripts/quant-2026-05-rerun.ts Seed = 42. Bootstrap iter = 1000. Generated 2026-05-17T18:32:39.165Z.

Methodology

Universe: every Declaration with a populated BacktestResult.returnFromPub90d and a parseable BUY direction.

Scoring:

v5.1 (production) — composite score (sig / wr / ret / rec), market-aware weights from MARKET_WEIGHTS table.
v6 (candidate) — v5.1 + role-quality multiplier (CEO/CFO +5 pct), cluster premium (+10 pct), pub-delay freshness penalty (>14 days x 0.85).

Portfolio: weekly top-10, equal-weight, hold 90 days. Monthly portfolio returns = arithmetic mean of picks within month.

Stats: Sharpe / Sortino annualised from monthly returns (x sqrt 12). Calmar = annualised return / |maxDD|. Max DD on the monthly NAV. Hit rate = pct of picks with r90 > 0. Turnover = avg fraction of top-N replaced week-on-week.

DSR (Bailey-Lopez de Prado): SR_def = SR - sqrt(2 ln(N)/T) with N = 9,420 effective trials (anchored on v5.1 grid 314 weights x alpha-r3 30 features) and T = monthly obs.

Bootstrap CI95: 1000 resamples on monthly returns, seed 42, annualised by x sqrt 12.

v5.1 (production) walk-forward

Market

Picks

Sharpe_ann

Sortino

Calmar

maxDD %

HitRate %

Turnover %

DSR

CI95

595

1.685

3.928

9.649

-19.9

52.4

99.8

0.648

[0.385, 3.076]

133

1.367

1.761

4.673

-18.1

54.1

98.5

0.014

[-0.831, 4.160]

484

1.840

3.594

7.371

-6.4

58.1

100.0

0.736

[0.305, 3.523]

225

0.836

0.967

1.233

-35.0

60.0

92.6

-0.233

[-0.714, 2.820]

COMBINED

662

1.319

2.015

4.090

-17.9

52.0

99.6

0.311

[-0.076, 2.785]

v6 (candidate) walk-forward

Market

Picks

Sharpe_ann

Sortino

Calmar

maxDD %

HitRate %

Turnover %

DSR

CI95

595

1.246

2.473

5.469

-25.0

49.6

99.6

0.209

[-0.319, 2.588]

133

1.334

1.705

4.490

-18.1

54.1

98.5

-0.019

[-0.932, 4.098]

484

1.840

3.594

7.371

-6.4

58.1

100.0

0.736

[0.305, 3.523]

225

0.836

0.967

1.233

-35.0

60.0

92.6

-0.233

[-0.714, 2.820]

COMBINED

662

0.981

1.567

2.370

-27.4

47.3

99.8

-0.028

[-0.634, 2.480]

Market

dSharpe

dSortino

dCalmar

dMaxDD pp

dHitRate pp

dDSR

-0.439

-1.455

-4.180

-5.08

-2.86

-0.439

-0.033

-0.056

-0.183

0.00

-0.033

0.000

0.00

0.000

0.00

0.000

COMBINED

-0.338

-0.448

-1.720

-9.46

-4.68

-0.339

MARKET_WEIGHTS — refresh proposal

Deltas (only markets where the local optimum beats current by Sharpe_ann > 0.05 AND DSR remains best-in-class):

Market

Current

Proposed

dSharpe_ann

Notes

30/35/25/10

(kept) 30/35/25/10

0.000

neighbour search did not clear honesty threshold (dSharpe > 0.05)

30/35/25/10

(kept) 30/35/25/10

0.000

neighbour search did not clear honesty threshold (dSharpe > 0.05)

30/35/25/10

(kept) 30/35/25/10

0.000

neighbour search did not clear honesty threshold (dSharpe > 0.05)

30/35/25/10

(kept) 30/35/25/10

0.000

neighbour search did not clear honesty threshold (dSharpe > 0.05)

Decision: keep MARKET_WEIGHTS unchanged. The dataset growth confirms the v5.1 plateau result from doc 11; per-market neighbour deltas remain inside noise after DSR.

Recommendation: v5.1 vs v6

v6 must beat v5.1 on DSR-deflated Sharpe AND on at least 3 of {Sortino, Calmar, hit rate, max DD} on the COMBINED portfolio.

COMBINED:

v5.1: Sharpe_ann = 1.319 · DSR = 0.311 · Sortino = 2.015 · Calmar = 4.090 · HitRate = 52.0 pct · MaxDD = -17.9 pct

v6: Sharpe_ann = 0.981 · DSR = -0.028 · Sortino = 1.567 · Calmar = 2.370 · HitRate = 47.3 pct · MaxDD = -27.4 pct

Recommendation: keep v5.1. DSR delta = -0.339. Secondary wins = 0/4 (need >= 3 AND positive DSR delta). v6 archived behind USE_SCORING_V6 flag (default off).