Edge Roadmap — July 2026

Status: review-owned. Updated after each strategy-shift phase review.
North star: a strategy that clears the promotion gates out of sample —
≥100 settled paper trades, ROI > +5%, positive mean AND median CLV — placed at
opening prices. Everything on this roadmap either moves toward that or gets cut.

Where we stand (evidence, not vibes)

Claim	Status	Evidence
The OPENING price is soft	Established	line_ridge_core +4.4% ROI exists only at open (CI −4.4/+19); open-priced models' edge signals confirmed by the close at 54–55% (CLA z = 2.7–3.4); close-featured diagnostics settle negative
The closing price is efficient	Established	Every model priced/settled at close is long-run negative
H2H market beatable with team-level features	Rejected so far	λ ≈ 0.02–0.10 across the bench — the market does ≥90% of the predicting
ss_strength_v1 > Elo	Established (small)	G1 pass, best raw ECE on bench; λ 0.055 — a better component, not an edge
Compound-Poisson scoring > ridge	Rejected (as built)	G3/G4 fail both runs; one iteration remains before retirement
Lineup/RAPM channel adds signal	SIGNAL CONFIRMED, GATE NOT CLEARED	2026-07-10: two bugs fixed (date-key join + lowercase team-key mismatch). With real deltas: G7 PASS λ=0.141 (lineup carries market-beating information); G6 FAIL by 0.5 milli-log-loss (0.6266 vs 0.6261 ablation). CLA z=5.22 — market moves toward RAPM picks. See phase3-rapm-v3 / -decay-v2 / phase4-stats-prior-v2 reports.
Team-list timing edge (Tuesday window)	Untested hypothesis	The one mechanism consistent with everything above; needs live paper trades + CLV
Line-move direction predictable from Tuesday features	CONFIRMED + positive backtested CLV (2026-07-25)	line-move-v1b: full 66.8% vs market_form_only 59.0% vs open_line_only 53.5%. Lineup channel adds 7.74pp over market structure, paired z=4.18 (mean-reversion-only REJECTED — open line alone barely beats base). CLV leg: mean +1.47 line points, 95% CI [1.08,1.87] excludes zero, 8 seasons. Backtested, not live. See review-2026-07-25.md
Lineup channel moves the LINE (Law 3 for line market)	CONFIRMED (2026-07-25)	Market-only ablation isolates it: 7.74pp lift over market+form, paired z=4.18. Single-feature ablations were masked by multicollinearity. First gate-clearing confirmation of the lineup channel
Favourites underpriced at open (models confirm)	Established (2026-07-12)	CLA segmentation: backing the favourite = 62.8–65.0% agreement (z 5.5–6.9); backing dogs = 41–49% (anti-signal); alive in 2023+ era. Defines `open_fav_v1`. See docs/edge-review-2026-07-12.md
`open_fav_v1` backtest (both legs)	CLA strong, ROI unproven — live CLV adjudicates	2026-07-16: OF1 PASS (CLA 0.737 / 0.690 on 2023+, 307 eligible); ROI report-only: flat +1.2% CI [−7.2,+9.1], edge_capped −1.1% CI [−9.9,+7.9]. Filter carries the signal (G2/G6 fail individually). First 2 paper trades seeded R20. See results-open-fav-v1-roi + review-2026-07-16
Underdog business-logic overlay	Tested and REJECTED (4 of 6 segments)	2026-07-16 pre-registered diagnostic: no segment clears CLA ≥0.55 ∧ z ≥3 ∧ n ≥100 (best: away dog 0.637 z=2.62). origin_week + rain segments UNEVALUABLE (data gaps) — one re-run allowed after backfill, then closed
H2H↔line cross-market divergence	Tested and KILLED	Line stopped correcting toward H2H-implied margins after 2018 (59.5% pre-2018 → 37–43% since). Archived; revisit only with exchange data

Instrument state: leakage guards test-enforced; CLA in every report;
bootstrap CIs require ≥5 seasons AND ≥100 bets; results contract (schema v3)
round-trips between sessions; UI surfaces models/strategy/ops/DQ/lineage.

Re-sequenced 2026-07-12 (docs/edge-review-2026-07-12.md is the detail)

Priority order now: N1 open-fav-v1 backtest verdict → N2 live evidence
quota every round (both strategies) → everything else. cpois RETIRED;
phase3-rapm-v4 demoted to idle rounds; H2H feature engineering FROZEN
pending a new information source; Betfair exchange data priced in September
(N4 — the only route to a materially better backtest).

Horizon 0 — superseded (kept for history: the week of R19)

DONE ✓ Re-run RAPM with fixed join — phase3-rapm-v3 committed 2026-07-10.
Two bugs fixed: kickoff-date join and lowercase team-key mismatch. First real
deltas confirmed (abs sum 6731 across 2018–2026). Results: G7 PASS λ=0.141,
G6 FAIL 0.6266 vs 0.6261 (ablation). CLA z=5.22.
DONE ✓ Decay and stats-prior re-runs — phase3-rapm-decay-v2 and
phase4-stats-prior-v2. Same pattern: G7 PASS, G6 FAIL by 0.5 milli-log-loss.
Keep the live loop fed: R19 close snapshots per kickoff
(schedule_close_snapshots.py), CLV update post-round, rapm_timing_v1
paper trades at team-list drop. The Jobs page shows if any of this slips.

Horizon 1 — next 3–4 rounds

Line market first. The only established soft spot is opening lines, and
the strongest model is line_ridge_core. Two cheap upgrades, gated on MAE +
open-line ROI CI: (a) feed ss_strength_v1 margin + variance into the
cover-probability (replace the pooled residual σ with the state-space
variance per match); (b) add the now-actually-populated lineup deltas as
features. Report as --phase line-v2.
CLA-gate every new feature family: if adding a family doesn't lift CLA
or log-loss vs its ablation control, it doesn't ship. No exceptions —
this is what kept the fake RAPM result out of production twice.
Accumulate rapm_timing_v1 evidence — target ≥40 trades by end of
H1 with CLV populated; interim read at 40 (direction only, no promotion).

Horizon 2 — by end of season 2026

Promotion decision on the line channel: line_ridge(-v2) at open with
shrunk-Kelly staking — does the season's paper sample clear the gates?
That's the first candidate with a realistic shot.
RAPM verdict: with real deltas, either the lineup channel earns λ ≥ 0.10
somewhere (season-grain OR timing-grain) or the whole family is archived
with the evidence attached.
cpois final iteration (fed by ss strengths, negative-binomial tries) —
pass G3/G4 or retire.
2027 pre-season: whatever survived gets a pre-registered live plan;
open-price capture automation (the Monday open snapshot is the single most
valuable data point we collect) hardened before R1.

Review 2026-07-11 — reading phase3-rapm-v3 correctly

The G6/G7 split is not a contradiction; the two gates measure different things
and the split tells us what to build next:

λ 0.141 vs control 0.103 and CLA 0.583 (988 signals, z=5.22) vs 0.550
(733): the lineup delta makes the model disagree with the open MORE OFTEN
and be confirmed by the close MORE OFTEN — the strongest closing-line
agreement of any actionable model. Direction is right. This is information
the opening market does not have.
G6 fail by 0.0005 log-loss with slightly worse ECE (0.019 vs 0.016):
the delta's MAGNITUDE is noisy — directionally right, poorly scaled, so
probability quality doesn't improve even though betting-relevant signal does.
Consequence (two work items, pre-registered):
1. Magnitude calibration: fit the delta through a monotone/clipped
transform inside the walk-forward (e.g. per-season standardization +
winsorize at ±2σ before the logit). Gate: G6 vs the same control, plus
ECE must not degrade. Phase label phase3-rapm-v4.
2. Timing channel unchanged as the decisive test — CLA 0.583 at
season-grain deltas is exactly the precondition the Tuesday-window
hypothesis needed. rapm_timing_v1 trades are now the highest-value
evidence in the program.
Shrunk-Kelly ROIs on these models (39–48 bets) are small-sample noise and
correctly not significant — do not quote them.
UI note: model cards fixed this round (deterministic report pick by content
timestamp, cross-report metric fallback so no track ever renders blank,
market-grouped ordering H2H → Line → Total, primary-model fallback metrics).

Standing guardrails

Nothing skips the results contract; failed gates get pushed.
No gate edits in result commits; ROI never quoted without its CI.
Closing data: diagnostic labels only, enforced by tests.
The catalogue (/models), strategy page, jobs page, glossary, and this
roadmap update together — a roadmap the UI contradicts is worse than none.