How our NFL model works

The basic unit is the drive, not the play. For each drive we know where it started, how far it got and how it ended. We treat a drive like a survival problem: for every yard of field it covers, how likely is it to end there? Touchdowns, and drives that run out the clock at the end of a half, count as drives that survived.

Drives can end short of the end zone in two competing ways, and each has its own model:

• Stalling in a punt or a field-goal attempt. Teams' ratings on this model are their drive efficiency.
• Turning the ball over: interceptions, fumbles, turnovers on downs, safeties and blocked kicks. Teams' ratings on this model are their turnover propensity.

Both are Bayesian models with an offense rating and a defense rating for every team, partially pooled toward league average, so a team needs consistent evidence to earn an extreme rating. The field is split into zones, each with its own baseline chance of a drive ending. The models also account for home field, the two-minute drill, and games that are out of hand (a margin of more than 22 points), when teams play differently.

Ratings use the current season only: the models are refit from scratch every week on that season's games so far. Each rating is published as a median with 50% and 94% intervals, and zero is league average. Three smaller models fill in the rest of a drive: whether a stalled drive punts or tries a field goal, whether the kick is good, and how much clock the drive uses (drive time).

See the current drive ratings.

To predict a game we play it out with those ratings, one drive at a time, using a Monte Carlo simulation:

• A coin flip decides who receives; the other team receives to start the second half.
• Kickoffs follow that season's kickoff rules, and possession changes after every drive.
• Each drive moves through the field zones until it ends. If it ends, the two drive models decide whether it was a turnover or a stall; a stall becomes a punt or a field-goal attempt, which is made or missed. A drive that survives to the goal line is a touchdown, followed by an extra point or a two-point try.
• The drive-time model runs the clock, which decides how many possessions each team gets.
• Games tied after regulation go to a 10-minute overtime under that season's regular-season rules, and can end in a tie.

Each simulated game draws a fresh set of team ratings from the models' uncertainty, so the spread of outcomes reflects both game-day randomness and how sure we are about each team. We run each game 2,000 times. The home team's win probability is the share of those games it wins outright, and the predicted score is the average and median of the simulated scores.

A 70% win probability is not a promise. It says that in games like this one, the favorite should win about 7 times in 10, so a 70% favorite should lose about 3 games in 10. One upset says little about the model; the test is whether, across many games, our 70% calls win about 70% of the time. That property is called calibration, and the track record below measures it, along with the Brier score.

Power rankings. Each week every team is simulated against the same imaginary opponent whose ratings are exactly league average, with home field assigned at random. Teams are ranked by their simulated win rate in those games, so the ranking measures strength without the schedule. Current power rankings.

Projected wins. Games already played keep their real results; every remaining regular-season game is simulated, and the standings show each team's average and median projected wins and the middle 50% range. Each team's remaining schedule is simulated separately, not as one joint league season, so we do not publish playoff, division or Super Bowl probabilities.

Play-by-play, drive, schedule and score data come from nflverse, used under the Creative Commons Attribution 4.0 license. We process and model that data; nflverse does not endorse this site. Seasons before 2026 on this site were built from Pro Football Reference data by an earlier version of our pipeline. The ratings, simulations and predictions are our own.

We update once a week during the season, after the week's last game (normally Monday night) and before the next week's first game (normally Thursday). The update loads the new results, refits every model on the season's games so far, re-runs the simulations and publishes new ratings, rankings, projected wins and predictions for the coming week.

Predictions start in week 4 of each season. The models use only the current season, so they need three weeks of games before they can rate teams; weeks 1 to 3 show schedules and scores only.

• No injuries, quarterback changes or roster moves. Ratings come from how a team's drives have gone, so a change of starter shows up only as later games are played.
• No weather, travel or rest. Apart from home field, game conditions are not modeled.
• One season at a time. Ratings ignore earlier seasons, which keeps them current but makes them noisy in the first weeks.
• A simplified game. The simulator plays drives, not plays. It does not produce defensive or return touchdowns or safeties, and punts and touchdowns use no clock beyond the drive-time model.
• Some drives are left out of training: overtime drives, drives that start with 45 seconds or less left in a half, and drives that open with a kneel-down.
• Projected wins simulate each team separately, so they are not a joint league simulation and cannot give playoff odds.

How our home-team win probabilities have done on every completed game that has one stored, regular season and playoffs, computed live from our database (latest game data updated October 6, 2026). The pick is the team we gave more than a 50% chance. The Brier score is the average squared gap between the forecast and the result: lower is better, and always saying 50% scores 0.250.

Games counts completed, non-tied games with a forecast. 2 tied games are left out. 4 games forecast at exactly 50% have no pick but are in the Brier score. Weeks 1 to 3 of each season, and a few other games, have no stored forecast and are not counted.

Calibration

Games grouped by the favorite's win probability. If the model is well calibrated, the favorite's actual win rate in each row should be close to the average forecast.

This record covers straight-up winners only. Nothing on this page is betting advice; see our disclaimer.