observatory: predictability, biases, signal decay

← lab · model run 18 · ran 7/14/2026

What this is: A one-off study — its numbers below are exactly what the run reported.

note	skill_captured_vs_market: 1.0 = as sharp as the closing line, 0 = no better than the base rate. Biases marked EDGE are stable disagreements that PAY; MODEL_DEFECT are stable disagreements that COST (fix the model there). Signals go weak_or_decaying when any era's correlation flips sign.
model run	18

signal decay

n	signal	status	target	corr overall	corr by window
33,804	match.ref_cards_avg	alive	cards	0.241	0.2554, 0.2884, 0.2158, 0.1996
44,630	match.style_clash_corners	alive	corners	0.0816	0.086, 0.0701, 0.0865, 0.0693
137,822	match.mismatch	alive	goal_margin	0.1932	0.1518, 0.2123, 0.2149, 0.2031
131,137	team.eye_test_v1	alive	team_goals	0.1856	0.1861, 0.1892, 0.1873, 0.1789
275,644	team.elo	alive	team_goals	0.2001	0.1853, 0.2167, 0.2086, 0.1955

n	market	verdict	mean gap	selection	price band	following gap pays
578	Over/Under 2.5 goals	MODEL_DEFECT	0.0661	Under	priced 2.60–4.50	-0.0185
326	Over/Under 2.5 goals	MODEL_DEFECT	-0.0495	Under	priced 1.00–1.80	-0.0221
1,190	Over/Under 2.5 goals	MODEL_DEFECT	-0.044	Over	priced 1.00–1.80	-0.0095
1,428	Over/Under 2.5 goals	MODEL_DEFECT	0.0213	Over	even	-0.0129
834	Match result	MODEL_DEFECT	0.0186	home side win	priced 4.50+	-0.0072
1,331	Match result	MODEL_DEFECT	0.0146	home side win	priced 2.60–4.50	-0.0103
945	Match result	MODEL_DEFECT	-0.0139	away side win	even	-0.0082
1,628	Match result	noise	0.0124	home side win	even	-0.0124
1,491	Match result	MODEL_DEFECT	0.0083	Draw	priced 4.50+	-0.015
3,829	Match result	noise	-0.0079	Draw	priced 2.60–4.50	-0.0094
1,816	Match result	noise	-0.0054	away side win	priced 2.60–4.50	-0.0123
1,527	Match result	noise	-0.0051	home side win	priced 1.00–1.80	-0.0128
685	Match result	noise	-0.0044	away side win	priced 1.00–1.80	-0.0257
1,747	Over/Under 2.5 goals	noise	-0.0026	Under	even	-0.0065
1,874	Match result	noise	-0.0019	away side win	priced 4.50+	-0.0108

n	question	by window	brier base	brier model	brier market	skill captured vs market
15,960	1x2	0.19467, 0.1883, 0.19234, 0.1965	0.2222	0.1929	0.1885	0.871
5,320	ou_goals_2_5	0.24364, 0.24554, 0.24695	0.25	0.2457	0.2402	0.437

AUC	predictability: 0.50 = coin flip, ~0.70 = ceiling for sports
Importance	how much the model leans on this factor (permutation importance)
Direction	sign of the raw correlation with the outcome
Survives all eras	effect points the same way in every historical era

Spec · the reproducible recipe

{
  "kind": "observatory",
  "model_run": 18
}