80 · Geographic Regime Multiplier · V13.3 Bake-Off

Date: 2026-05-19 Status: REJECTED (overfit, kept V13.1g baseline) Author: bake-off (scripts/_v13_bakeoff/geo-regime-stats.ts + bake-v13_3.ts)

Hypothesis

Certain markets are systemically more alpha than others (JP top picks 1Y +258 %, FR +30 %, US lower). Codify explicitly with a per-market scoring multiplier in [0.95, 1.15].

Phase 1 · Per-market baseline (BUY, priced T+90 / T+365)

In-sample window: pub 2021-01-01 → 2024-12-31. Source: BacktestResult ∩ Declaration ∩ Company.

market	n_is	n_oos	winR_is %	med90_is	med365_is	mean365_is	med90_oos	med365_oos
XSTO	6	3464	100.0	21.55	61.55	61.12	-0.61	1.17
XKRX	116	0	25.9	-0.61	14.48	6.86	0.00	0.00
XTKS	8535	0	53.5	1.36	13.13	17.88	0.00	0.00
XMIL	787	583	48.5	-0.49	10.02	10.76	-1.69	-5.35
XMAD	1607	579	56.7	1.66	8.41	540.77	3.64	38.95
XSHE	1636	180	54.3	1.99	2.60	12.37	2.23	2.29
XSHG	1631	249	45.6	-1.83	2.28	11.27	5.05	51.84
BVMF	31038	11296	44.0	-1.48	-2.10	-2.56	1.68	37.44
XPAR	11581	3590	42.8	-2.13	-3.55	-1.84	0.00	-4.83
XSWX	494	0	49.4	0.00	-3.74	1.28	0.00	0.00
XHEL	1455	482	41.0	-2.08	-4.67	-0.86	0.46	-3.38
XBOM	5203	0	66.0	6.15	-6.73	4.89	0.00	0.00
XWBO	617	61	40.0	-2.38	-7.14	-3.57	0.86	30.95
XNAS	1194	25	26.9	-9.52	-10.69	3.07	9.40	43.68
XCSE	361	82	43.2	-3.58	-11.82	-8.89	2.35	9.49
XAMS	100	80	38.0	-3.61	-20.52	-14.09	4.27	27.65

(small-n markets XBRU/XASX/XOSL/XLON/XETR/XSES/XHKG/XTSE/XNZE/XSAU/XJSE = no IS sample, default 1.00x)

Phase 2 · Calibrated multipliers

Rules (conservative, band [0.95, 1.15]):

winR ≥ 55 % AND med90 > 5 % → 1.10
winR ≥ 50 % AND med90 > 0 → 1.05
n_is < 200 → 1.00 (no info)
else → 0.95

market	mult	reason
XBOM	1.10	winR 66.0 med90 6.15
XTKS	1.05	winR 53.5 med90 1.36
XMAD	1.05	winR 56.7 med90 1.66
XSHE	1.05	winR 54.3 med90 1.99
XMIL	0.95	winR 48.5 med90 -0.49
XSHG	0.95	winR 45.6 med90 -1.83
BVMF	0.95	winR 44.0 med90 -1.48
XPAR	0.95	winR 42.8 med90 -2.13
XSWX	0.95	winR 49.4 med90 0.00
XHEL	0.95	winR 41.0 med90 -2.08
XWBO	0.95	winR 40.0 med90 -2.38
XNAS	0.95	winR 26.9 med90 -9.52
XCSE	0.95	winR 43.2 med90 -3.58

Note · IS/OOS sign flips on multiple markets (XPAR med365 IS -3.55 → OOS -4.83 stable; BVMF -2.10 → +37.44 flipped; XNAS -10.69 → +43.68 flipped; XAMS -20.52 → +27.65 flipped). This is the first red flag.

Phase 3 · V13.3 OOS Bake-Off

Pipeline: identical to V13.1 bake (top-10 picks/month, T+90 hold, NET 0.6 % round-trip, winsor ±50 %). OOS window: pub 2025-01-01 → 2026-05-19. T = 14 monthly buckets. DSR uses N=14 trial cohort (11 V13.1 variants + 3 new V13.3 variants).

config	T	picks	Sharpe	CI95 (boot)	CAGR %	MaxDD %	Win %	DSR
V13.1g_baseline	14	140	0.70	[-0.96, 3.51]	25.4	-29.7	55.7	0.28
V13.3_geo	14	140	0.48	[-1.24, 3.03]	15.2	-32.7	54.3	0.06
V13.3_geo_conservative	14	140	0.70	[-1.05, 3.54]	25.2	-29.8	55.7	0.27

Decision · REJECTED

V13.3_geo: Sharpe drop -0.22 (0.70 → 0.48), DSR drop -0.22 (0.28 → 0.06), MaxDD widens -3.0 pp, Win % drops -1.4 pp. Anti-overfit guard says DSR drop > 0.3 → reject. We are at -0.22 which is borderline, but the geographic signal also flips IS→OOS sign on 4 of 13 markets (BVMF, XNAS, XAMS, XWBO), making the IS calibration unreliable. Reject.
V13.3_geo_conservative (band narrowed to [0.97, 1.05]): Sharpe ~flat (0.70 → 0.70), DSR -0.01. Effectively a no-op once squeezed below noise floor. No reason to add complexity.

Kept baseline: V13.1g (Sharpe 0.70 / DSR 0.28).

Root cause of overfit

Per-market baselines are computed on the same population of insider trades the score then re-ranks. Multiplier compresses or stretches scores already conditioned on the same data.
IS→OOS regime instability: 4 / 13 markets flip sign of med365 IS vs OOS, including the two largest by sample (BVMF n=31k, XPAR n=11.5k). Calibrating a multiplier off IS is then anti-predictive.
The 1Y top-pick spread cited in the hypothesis (JP +258 %, FR +30 %) is dominated by mean-not-median heavy tails. The top of the distribution does not generalise to top-10 monthly portfolio behaviour.

What is kept

Per-market stats table (this doc) for reference and copy editorial.
Script scripts/_v13_bakeoff/geo-regime-stats.ts and bake-v13_3.ts retained for future re-tests when sample grows.
src/lib/signals.ts unchanged. No new MARKET_REGIME_MULT map shipped.

Cross-refs

docs/method-review/78-quant-per-market-sector-2026-05-19.md (V13.1g winner)
scripts/_v13_bakeoff/bake.ts (V13.1 ensemble bake)

Phase 1 · Per-market baseline (BUY, priced T+90 / T+365)

In-sample window: pub 2021-01-01 → 2024-12-31. Source: BacktestResult ∩ Declaration ∩ Company.

market

n_is

n_oos

winR_is %

med90_is

med365_is

mean365_is

med90_oos

med365_oos

XSTO

3464

100.0

21.55

61.55

61.12

-0.61

1.17

XKRX

116

25.9

-0.61

14.48

6.86

0.00

XTKS

8535

53.5

1.36

13.13

17.88

0.00

XMIL

787

583

48.5

-0.49

10.02

10.76

-1.69

-5.35

XMAD

1607

579

56.7

1.66

8.41

540.77

3.64

38.95

XSHE

1636

180

54.3

1.99

2.60

12.37

2.23

2.29

XSHG

1631

249

45.6

-1.83

2.28

11.27

5.05

51.84

BVMF

31038

11296

44.0

-1.48

-2.10

-2.56

1.68

37.44

XPAR

11581

3590

42.8

-2.13

-3.55

-1.84

0.00

-4.83

XSWX

494

49.4

0.00

-3.74

1.28

0.00

XHEL

1455

482

41.0

-2.08

-4.67

-0.86

0.46

-3.38

XBOM

5203

66.0

6.15

-6.73

4.89

0.00

XWBO

617

40.0

-2.38

-7.14

-3.57

0.86

30.95

XNAS

1194

26.9

-9.52

-10.69

3.07

9.40

43.68

XCSE

361

43.2

-3.58

-11.82

-8.89

2.35

9.49

XAMS

100

38.0

-3.61

-20.52

-14.09

4.27

27.65

(small-n markets XBRU/XASX/XOSL/XLON/XETR/XSES/XHKG/XTSE/XNZE/XSAU/XJSE = no IS sample, default 1.00x)

Phase 2 · Calibrated multipliers

Rules (conservative, band [0.95, 1.15]):

winR ≥ 55 % AND med90 > 5 % → 1.10

winR ≥ 50 % AND med90 > 0 → 1.05

n_is < 200 → 1.00 (no info)

else → 0.95

market

mult

reason

XBOM

1.10

winR 66.0 med90 6.15

XTKS

1.05

winR 53.5 med90 1.36

XMAD

1.05

winR 56.7 med90 1.66

XSHE

1.05

winR 54.3 med90 1.99

XMIL

0.95

winR 48.5 med90 -0.49

XSHG

0.95

winR 45.6 med90 -1.83

BVMF

0.95

winR 44.0 med90 -1.48

XPAR

0.95

winR 42.8 med90 -2.13

XSWX

0.95

winR 49.4 med90 0.00

XHEL

0.95

winR 41.0 med90 -2.08

XWBO

0.95

winR 40.0 med90 -2.38

XNAS

0.95

winR 26.9 med90 -9.52

XCSE

0.95

winR 43.2 med90 -3.58

Phase 3 · V13.3 OOS Bake-Off

config

picks

Sharpe

CI95 (boot)

CAGR %

MaxDD %

Win %

DSR

V13.1g_baseline

140

0.70

[-0.96, 3.51]

25.4

-29.7

55.7

0.28

V13.3_geo

140

0.48

[-1.24, 3.03]

15.2

-32.7

54.3

0.06

V13.3_geo_conservative

140

0.70

[-1.05, 3.54]

25.2

-29.8

55.7

0.27

Decision · REJECTED

V13.3_geo: Sharpe drop -0.22 (0.70 → 0.48), DSR drop -0.22 (0.28 → 0.06), MaxDD widens -3.0 pp, Win % drops -1.4 pp. Anti-overfit guard says DSR drop > 0.3 → reject. We are at -0.22 which is borderline, but the geographic signal also flips IS→OOS sign on 4 of 13 markets (BVMF, XNAS, XAMS, XWBO), making the IS calibration unreliable. Reject.

V13.3_geo_conservative (band narrowed to [0.97, 1.05]): Sharpe ~flat (0.70 → 0.70), DSR -0.01. Effectively a no-op once squeezed below noise floor. No reason to add complexity.

Kept baseline: V13.1g (Sharpe 0.70 / DSR 0.28).

Root cause of overfit

Per-market baselines are computed on the same population of insider trades the score then re-ranks. Multiplier compresses or stretches scores already conditioned on the same data.

IS→OOS regime instability: 4 / 13 markets flip sign of med365 IS vs OOS, including the two largest by sample (BVMF n=31k, XPAR n=11.5k). Calibrating a multiplier off IS is then anti-predictive.

The 1Y top-pick spread cited in the hypothesis (JP +258 %, FR +30 %) is dominated by mean-not-median heavy tails. The top of the distribution does not generalise to top-10 monthly portfolio behaviour.