Edge framework

Strategy

Latest evaluation: line-move-v1b
2026-07-24T23:07:48+00:00

Active candidates

CandidateMarketStatus
open_fav_v1H2H opening favouritecandidate / not promoted
rapm_timing_v1H2H / line team-list timingcandidate / missing CLV
line_move_v1Line movement directionlive PAPER / awaiting 20 settled CLV reads

Frozen areas: generic H2H feature churn, divergence signal, broad Elo remixes. Nothing here is a proven edge or a real-money instruction.

Canonical promotion status

DO NOT PROMOTE — paper evidence only; no real-money activation is implied.

Market-only baseline evidence: insufficient_evidence.

GateStatus
sufficient_betsBLOCKED
roi_gt_5pctBLOCKED
positive_mean_and_median_clvBLOCKED
market_baseline_improvementBLOCKED
home_away_robustBLOCKED
concentration_limitPASS
leakage_checksPASS
sufficient_seasonsBLOCKED
evaluated_policy_matches_livePASS

Current edge thesis

The exploitable inefficiency is the OPENING price. Three independent instruments agree: line_ridge_core's +4.4% ROI exists only at opening lines; open-priced models' disagreements with the open are confirmed by the close ~54–55% of the time (z ≥ 2.7); and closing-featured diagnostics settle negative. The close is efficient — the open is not. Every strategy below is a way of getting to the soft price before the market hardens.

Language rule: everything below is a candidate / hypothesis / paper trade — nothing is a proven edge until it clears the promotion gates out of sample.

Staking policies

PolicyRuleStatus
Flat Every Game
flat_every_game
$1 on the max-EV side of every fixture. Benchmark, not a strategy. benchmark
Edge Capped
edge_capped
$1–2 when raw model EV ≥ 5%. Retained for comparison; raw EV in a 2%-vig market is mostly model error. legacy comparison
Shrunk ¼-Kelly
shrunk_kelly
Quarter-Kelly on the market-blended probability (λ fit walk-forward), EV clamped to [2%, 10%]. The reference policy: bets rarely, sizes honestly. reference

Promotion gates (paper → real consideration)

GateBar
Sample≥ 100 settled paper bets
ROI> +5%
CLVPositive average AND median CLV
ConcentrationNo round > 30% of profit
Seasons≥ 2 full seasons for watch status

Canonical values: data/promotion_gates.py. Negative-CLV models get zero vote weight regardless of ROI.

Strategies under paper observation

StrategyMarketTriggerAction
movement_total_2pt_v1 Totals |latest − open| ≥ 2.0 pts, ≥2 bookmakers Paper-trade the steam side
movement_spread_2pt_v1 Spreads |latest − open| ≥ 2.0 pts, ≥2 bookmakers Paper-trade the steam side
movement_h2h_3pct_v1 H2H No-vig probability shift ≥ 3% Paper-trade the rising side
rapm_timing_v1 H2H / Line RAPM lineup delta at Tuesday team-list drop vs opening price Paper-trade with CLV collection — the leakage-free test of the team-news-timing hypothesis
line_move_v1 Line Tuesday directional confidence ≥ 0.60 at the real opening handicap PAPER GO at the open; WATCH below threshold; settle line CLV at close

Latest model-build gates (line-move-v1b)

GateDescriptionStatusDetail
LM1 directional accuracy >= 53% pass accuracy=0.667789
LM2 directional accuracy >= majority-class base rate + 2% pass accuracy=0.667789; base_rate=0.516274; required=0.536274
LM3 z >= 2 against majority-class base rate pass z_vs_base=9.0501
LM4 n >= 500 non-zero line moves pass n=891
LM5 2023+ era slice is broken out report_only n_2023_plus=216