revalidate roster: consensus price + train/test holdout

← lab · n live 30 · n scored 5 · n skipped 25 · ran 7/9/2026

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

note	market-shape re-score at consensus opening price, train<2022 / test>=2022. Model-gated rules scored on price band+selection only (gate ignored) — a SHAPE-NEGATIVE verdict is NOT a kill (e.g. wc_ko_unders is 5/6 in forward paper). NOT-REPLAYABLE strategies must be judged on their forward paper record (paper_bet.status in won/lost), not this proxy.
n live	30
n scored	5
n skipped	25

scored

sel	band	code	comps	market	test n	train n
A	1, 2.6	away_fav_continental_v1	SA, FL1, PPL	Match result	1,331	4,396
—	2.2, 3.5	mlb_dog_variance_v1	all	Match winner (2-way)	0	0
D	1, 1000	ref_cards_draw_v1	all	Match result	12,649	55,726
under	1, 1000	wc_ko_unders_v1	all	Over/Under 2.5 goals	12,649	10,313
—	4, 10	wc_upset_hunter_v1	all	Match result	8,893	32,825

not replayable

code	reason
elo_value_mlb_v1	model-only edge rule: no market shape to test
drought_fade_v1	model-only edge rule: no market shape to test
steam_confusion_v1	model-only edge rule: no market shape to test
hot_hand_fade_v1	model-only edge rule: no market shape to test
chaos_parlay_v1	composite: depends on its leg strategies
wc_sim_value_v1	model-only edge rule: no market shape to test
steam_chaser_v1	steam: needs our snapshot-move history, not price levels
portfolio_v1	composite: depends on its leg strategies
pitcher_edge_v1	model-only edge rule: no market shape to test
prop_scorer_v1	soft/prop market: ~no odds history (forward-capture only)
cards_over_ref_v1	soft/prop market: ~no odds history (forward-capture only)
corners_style_over_v1	soft/prop market: ~no odds history (forward-capture only)
wc_sim_value_kor_v1	model-only edge rule: no market shape to test
mls_sim_value_v1	model-only edge rule: no market shape to test
dutch_cover_v1	model 'sim' output not reconstructable from consensus odds
dd_live_v1	in-play: no historical live odds (synthetic only)
simple_parlay_v1	composite: depends on its leg strategies
prop_scorer_v2	soft/prop market: ~no odds history (forward-capture only)
prop_soa_v1	soft/prop market: ~no odds history (forward-capture only)
soa_duo_parlay_v1	soft/prop market: ~no odds history (forward-capture only)
fade_goal_live_v1	in-play: no historical live odds (synthetic only)
mad_game_live_v1	in-play: no historical live odds (synthetic only)
goal_bets_v1	model 'goal_matrix' output not reconstructable from consensus odds
goal_bets_ht_v2	in-play: no historical live odds (synthetic only)
goal_bets_late_over_v3	in-play: no historical live odds (synthetic only)

Reading the columnswhat each number actually means

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": "revalidate",
  "cutoff": "2022-07-01",
  "min_test": 50,
  "min_train": 100
}