How it works

Games. Every team gets an Elo rating plus opponent-adjusted efficiency (EPA per play) ratings for offence and defence, weighted toward recent form, and every QB gets his own rating so a backup starting is priced in. A ridge regression projects the margin (and ridge plus gradient-boosted trees the total), then the win probability is blended with the betting market. The blend is calibrated on out-of-sample predictions so it never over-trusts the model.

Bet ladder. A different goal from the rest of the site: the best chance of multiplying a stake by a target amount, three weeks running, rolling each return into the next bet. For every game we price each alternate spread and moneyline by how often teams at that closing spread have cleared it since 2002 (weighted to recent seasons). That captures how margins bunch at 3 and 7, which is why moving a small underdog from +2.5 to +8.5 is worth more than six points suggest. Main-line player props are candidates too, using the props model's probability blended with the bookmaker's (its probabilities are well calibrated on unseen seasons, but whether it beats the book is still being measured). An optimiser then picks the accumulator with the highest chance of winning at each path's payout (about 3x, 5x and 7x), at most one leg per game, since legs from the same game move together. The safer paths keep the safest legs of the riskier one. These chances were calibrated walk-forward: legs given 70–80% won 74.5% in 2024–25. Even so, three wins in a row is unlikely, so only stake what you're happy to lose.

Spread and total views. Each card shows the model's spread and total next to Vegas, for information. They are not picks: over 2012–2025 the model's biggest disagreements with the closing line won about half the time, short of the 52.4% needed to break even. Only angles that have tested positive are marked as picks, such as the Under in 15+ mph wind.

Honest expectations. On the held-out 2024–2025 seasons, which no tuning ever touched, the blend picked 68.0% of winners against 68.2% for the Vegas favourite, with near-identical probability accuracy. That is about the ceiling for NFL prediction: the market is very efficient. Spread and total views showed no reliable edge, so they are shown as information, not picks. The CLV record tracks whether that changes live.

What we tested and rejected. Market-implied team ratings, garbage-time-filtered EPA, special-teams EPA, travel distance and body-clock kickoff times, referee crew tendencies, QB experience, Next Gen Stats (separation, cushion, YAC over expected), player-specific outcome volatility, a starting-QB flag for QB props, kicker-prop drivers (coach 4th-down aggressiveness, drive stall rates) and short-favourite moneyline fades were all tested on 2012–2023 and either failed the held-out check or added nothing. Features only go in when they improve unseen seasons.

Player props. Projections combine usage (targets, carries, snap share, share of team volume), efficiency, game context from the betting line, the opponent's defence against that position, usage freed up by teammates ruled Out or Doubtful on the official injury report (and who is next in line to inherit it), and who is starting at QB. Absences are defined exactly as they are known before kickoff, in training and live alike. Over/under probabilities come from the real spread of NFL outcomes around each projection. Every prop with a sportsbook line is graded, and the model's probability is blended with the book's using a weight fitted on those results (halfway until enough are graded), so it trusts itself exactly as much as its record justifies. Props are withheld when a teammate's status is uncertain.

Closing line value (CLV). Each lean is saved with the line at first publish (Wednesday). If the line later moves toward our side before kickoff, that is positive CLV: the betting market came round to our view. Consistently beating the closing line is the standard professional test of a real edge, and it shows up within weeks, long before win/loss records stop being noise.

Weather. Kickoff forecasts (temperature and wind) feed the totals model for outdoor games. With 15+ mph forecast wind the Under is marked as a tested pick: on 2012–2023 closing lines it hit 57.5% in such games.

Segments. Situations where research suggests the market may be beatable (windy games, backups stepping into a role after a starter is ruled Out) are tracked separately on the Record page, so the evidence decides what we trust.

Trends are current streaks ranked by how rare they are league-wide. They are context, not predictions.

Schedule. Predictions publish Wednesday night (with prop lines for Thursday's game), then refresh Saturday with final injury reports, prop lines and the bet ladder. Games that have kicked off are locked. Discord gets the short list: the ladder, Strong and Lock winners, and tested edges.