Predictive Models

Every team carries ratings that update after each game: scoring margin, and the success rate of its offense and of the offenses it faces (the share of plays that gain expected points), each judged against the strength of the opponent. Recent games count more than old ones, and each season starts from a regressed version of last season’s ratings.

Quarterbacks matter most, so each one has their own rating from EPA per dropback across their career. When a team’s starter is better or worse than the quarterback play it has been getting (an injury, a benching, a return), the forecast moves accordingly. The starter is the one nflverse lists, about a week ahead; before that, the QB1 on the team’s depth chart for games up to six weeks away, which forecast better than assuming nothing changes. The model also counts the snap share of players ruled out on the injury report, and estimates home-field advantage from recent seasons rather than assuming a fixed edge (it has shrunk).

Offensive linemen count through the injury report like everyone else, and a line's quality is already part of the team's EPA and success-rate ratings. Separate line features (starting linemen ruled out, pass protection and run blocking against the opponent's front, for the winner and for total points) were tested and didn't improve the forecasts, so they aren't used.

A logistic regression turns those differences into a win probability; the projected margin is the margin that probability implies (σ ≈ 12.2 points), so the favourite and the projected score always agree. Totals come from a separate model that adds pace, quarterback quality, wind and cold.

Only information available before kickoff is used: ratings come from earlier games, and the regression weights were fit on 2006–2025, so this season’s numbers are genuine pregame forecasts. Playoff odds come from simulating the rest of the season 10,000 times.