MATCHDAY RESEARCH
Matchday tests every proposed model feature against a real out-of-sample baseline before it's allowed anywhere near a live pick — no feature ships just because it sounds sophisticated. As of 2026-08-02, 9 feature experiments have a recorded decision.
Kept, in capped production (1): Elo calibration (nfl) — Approved for a capped, reversible historical pilot: production_weight 0.10, max_shift_points 3.0, allow_pick_flip false.
Still collecting evidence (2): player availability / injury signal (nfl) — Blocked on acquiring a licensed availability source and enough frozen pregame history before any effect can be trained. | run-strength model (Pythagorean strength + run differential per game) (mlb) — Frozen model attaches a prospective shadow probability to eligible fixtures via the append-only forecast ledger; production_weight is still 0 pending the prospective promotion gate (>=500 games across >=30 date blocks, log-loss + Brier improvement, paired interval upper bound below zero, named manual review).
Rejected (3): learned nflverse play-by-play challenger (nfl) — Remains zero-weight shadow research; the negative result is kept as a permanent part of the model record rather than tuned away. | quarterback availability/performance signal (nfl) — Kept only as an uncertainty/provenance receipt, not an accepted production feature. A future timestamped pregame-availability feed is required before the actual-starter version can be tested honestly. | advanced-feature residual model (nfl) — Remains rejected for production.
Not yet run (3): box-score-derived efficiency/shooting/possession/tempo/context features (basketball) — Zero-weight research path pending a licensed NBA/NCAAM data source per Phase 15's premium-data justification process. | opponent-adjusted PPA/success/explosiveness/talent/context features (ncaaf) — Zero-weight research path pending a bundled/licensed multi-season export. | xG strength / shot quality / possession-territory / pressure / set-piece / lineup-continuity features (soccer) — Zero-weight, attribution-bound research path.
Full methodology, evaluation numbers, and citations for every row above are public in this repository's docs/experiments.json and docs/PREDICTION_RESEARCH_ROADMAP.md — nothing here is summarized from a private dataset.
Matchday — pregame predictions, market context and postgame grading.