How the Model Works
Every number on this site traces back to something on this page. No black box, no vibes: if a claim can't be checked, it doesn't go on the site. That includes admitting where the model is weaker than the betting market, which happens on average (see Track Record).
Win Probability
The pregame model is a regularized logistic regression, not a deep or boosted model: with only a few hundred games worth of training data per era of the league, a more complex model would just memorize noise instead of learning anything real. Every prediction is calibrated (isotonic regression) against held-out games, because for a public track record, an honest 65% matters more than a confident-sounding 90%.
Because it's a linear model on standardized inputs, each factor's push on a prediction is exact arithmetic (coefficient times value), not an approximation. The reasons shown on a game page for why the model leans one way are the model's actual, literal reasons, not a plausible-sounding guess reconstructed after the fact.
Confidence Bands
A win probability is never shown as raw as "64.7%" without also saying what kind of number that is. The bands:
What Goes Into a Prediction
- Offense and defense strength, each team's own EPA per play over its last 8 games
- Starting quarterback form, weighted separately from the rest of the offense
- Rest advantage, short/long week, and how many time zones the away team is crossing
- Travel distance, altitude, indoor vs. outdoor, and wind
- Divisional matchup and neutral-site games
Deliberately not in the model: the Vegas line itself. It's tracked separately, purely as a benchmark to check the model against, never fed in as an input.
The Vegas Line
The closing spread is converted to an implied win probability with a normal approximation (a standard deviation of 13.5 points, the long-run spread of NFL margins), the same conversion behind every gap shown on the Value Picks page. This is a benchmark, not an input: the honest goal is to see how close the model gets to a line that already bakes in injury news, sharp money, and information this site simply doesn't have.
Predicted Scores
Two more regressions, trained separately from the win-probability model, predict the final margin and total points. Both are held to the same standard: their error is measured on games they never saw, and shown next to the market's own error on the same games.
Fourth-Down Decisions
Every 4th-down call in Biggest Decisions is priced in expected points, not a rule of thumb: going for it, kicking a field goal, and punting are each valued as p(success) × the value of succeeding, plus p(failure) × the cost of handing the ball back, and every branch already accounts for the opponent's next possession, not just the immediate play.
Field position after a kickoff, a missed field goal, or a punt isn't hardcoded: it's measured fresh from play-by-play data each season, since the NFL has changed kickoff rules twice in recent years and a stale constant would quietly go wrong.
What This Site Doesn't Do
- No player props or individual stat predictions
- No live, in-game win probability: every number here is set before kickoff and stays put
- Playoff tiebreakers are simplified to point differential only, not the NFL's full multi-step procedure
- A probability is not a certainty; see Track Record for how often the stated numbers actually held up