The existing neutral-field rating prevents a single noisy opener from producing a fake transformation.
SCORES
All scores ↗The 4M model room
SHOW
THE WORK.
PIPELINE · 4M INTELLIGENCE 1.5
A useful model should be explainable, versioned and judged against results it did not already know. CollegeFootballData supplies the football evidence; SportsDataIO can supply verified availability when a real-data license is active. The calculation and editorial judgment remain ours.
The starting position
A prior that can move—but cannot panic.
Every refresh creates a new database snapshot and must clear roster, coach, schedule and game-pricing coverage gates before it can be published. No historical edition or pregame forecast is silently overwritten.
Recency-weighted game PPA, EPA, success rate, explosiveness and havoc deepen multiple opponent-adjusted and predictive views.
Every rostered player receives a role probability, production grade and replacement baseline. Quarterback rooms receive explicit starter probabilities; OL, receivers, defensive front and secondary are graded as separate position groups.
CORE, adjusted EPA, success rate, explosiveness and special teams earn more influence as 2026 snaps accumulate.
Ratings, unit matchups and pace create a neutral-field margin. Base home field, capacity, altitude, relative travel, body clock, rest and verified weather then form a visible point-by-point bridge to the final independent number.
The independent 4M line remains football-only. A separately labeled market-calibrated line shrinks uncertainty toward consensus. The action threshold determines whether the difference is large enough to consider after vig and data risk.
Every schedule, conference race, playoff selection and elimination round is simulated when a new edition is created.
Inputs, source health, market context and output are stored together so every public claim can be reconstructed.
Game-level PPA, success, explosiveness, finishing drives, line yards and pace are accumulated only after each game. Separate margin, total and win models are tested on 2020–2025; a challenger is promoted only when held-out accuracy improves without damaging the other targets.
The generic home advantage is learned, then venue and conference residuals are added only after enough historical games exist. Capacity, altitude, travel, body clock and rest remain transparent matchup adjustments.
Out, doubtful, questionable and probable designations are multiplied by each player’s modeled point value. Scrambled trial feeds are blocked from public numbers, and every active adjustment keeps its source.
Daily operating rhythm
Update the data daily. Move the rating only when the evidence earns it.
A timestamped edition keeps the public board honest and prevents silent revisions after a result.
Open the public model audit ↗- Reconcile
Final scores, play-by-play, advanced box scores, participation and corrected game records.
- Refresh
Opponent-adjusted team metrics, roster transactions, injuries, depth charts and schedule data.
- Challenge
Retrain after the full slate, compare the challenger with the active champion and preserve the incumbent unless every promotion gate clears.
- Freeze the matchup
Refresh market consensus, weather and availability, then preserve the complete official prediction before kickoff.
The public scorecard
Accuracy has more than one definition.
Average absolute error between projected and actual scoring margin.
Whether a stated 60% outcome occurs near 60% over a meaningful sample.
Rewards accurate probabilities and punishes unjustified confidence.
Whether published positions consistently beat the eventual consensus close.
Tracked at the exact posted number and price, never a later favorable line.
Every weekly release archived so results cannot be rewritten after kickoff.