Click a team for its schedule, win distribution and seed outlook. Rating = roughly how many points better than an average FBS team on a neutral field; the model's point spread is the rating difference × 1.09 (see About).
The bracket
The announced field and seeds. Completed games show final scores; the rest show each team's chance to win that game, from the simulations.
Most likely team at each seed
Where they'll play
Chance each team appears in each CFP bowl. Announced sites are used once known; otherwise seeds 1–3 are modelled as taking the closest open quarterfinal bowl in seed order (No. 4 gets the one left), with no conference tie-ins — an approximation of the committee's placement.
Predict any FBS matchup
Uses current ratings, home-field advantage and rating uncertainty.
Win-probability calibration
Out-of-sample games, grouped by predicted home-win chance. Points on the diagonal = perfectly calibrated.
Season by season (holdout)
Each game predicted using only games played before that week. Games counts every game, bowls and playoff included; the Model and Vegas columns compare the regular-season games that had a line.
Committee model — leave-one-season-out
How well the committee model reproduces each final CFP ranking when that season is held out of training.
What this is
A computer model that tries to answer one question: who will make the College Football Playoff?
It has no opinions, doesn't watch film and doesn't read the polls. It learns from this season's scores, starting from a preseason guess built from past seasons, returning players and recruiting (step 2 below). From those it rates every FBS team, then plays out the rest of the season many times on a computer and counts what happens. When it says a team has a 30% playoff chance, that team made the field in about 30 of every 100 simulated seasons. Nothing on this site is a prediction that something will happen; it is a count of how often it happened in the simulations.
Scores and betting lines come from CollegeFootballData.com and the page is rebuilt automatically a few times a day, so the numbers move every time a game ends.
How it works, in five steps
-
Rate every team
Every FBS team gets a rating: how many points better (or worse) it would be than an average FBS team on a neutral field. A rating of +20 means nearly three touchdowns better than average; a rating of −10 means about a touchdown and a field goal worse. (Strictly, ratings are in the model's own units; step 3 turns them into points, currently ×1.09.)
The ratings are a statistical fit to margins of victory, not just wins and losses. Beating a good team by 20 says more than beating a bad team by 3, and a close loss to a great team isn't held against you much. Home-field advantage, worth about 3.9 points this season, is taken out of every result. Games against FCS opponents count, with FCS teams treated as one group of roughly similar (weak) teams, so running up the score on one proves little.
Each rating comes with an error bar. In September the model is much less sure of a team than it will be in November, and it keeps track of that.
-
Start from a preseason guess
Before kickoff there are no scores to go on, so each team starts from an educated guess built from how it rated in each of the last two seasons, how much of last year's production is back, and how well it has recruited over the last four years. That starting point counts for less with every game played, but it never vanishes entirely: one season of scores is noisy enough that last year's results still carry real information in November.
-
Turn two ratings into a point spread and a win chance
Rating difference, plus home field for the home team, times a scale factor gives the model's point spread for any game, just like a betting line. The scale factor (now 1.09; bigger early in the season) is fit on past seasons so that predicted margins come out the right size on average; it is close to 1, so the rating difference is roughly the spread. On any given Saturday the actual margin scatters around that spread by about ±14 points, because football is random. That is why a team favored by a touchdown wins only about 70% of the time, and a 3-point favorite about 59%. The Matchup predictor tab does exactly this for any two teams.
-
Play out the season, thousands of times
The simulation doesn't just pencil in the favorites. In each simulated season it first draws every team's "true" strength from inside its error bar, so a team the model is unsure about is sometimes much better and sometimes much worse than its rating. Then it plays every remaining game with the randomness from step 3, works out the conference standings with tiebreakers, picks the title-game participants and plays those games too. Do that many times and you get, for every team, how often it won its conference, how often it made the playoff, and which seed it got.
-
Guess what the committee does, then play the bracket
At the end of each simulated season a stand-in for the selection committee ranks the teams. It was taught by the real committee's final rankings from every season since 2016 and rewards what that committee has rewarded: wins, and especially wins over good teams (strength of record); the team's rating; and playing in a power conference. A conference title helps a little on top of the win it took to earn it. Random noise stands in for the committee's judgment calls, so a bubble team is sometimes in and sometimes out.
The bracket is then played out game by game with the same randomness, which is where the national title odds and the "Where they'll play" table come from.
Reading the numbers
- Rating
- Roughly how many points better than an average FBS team on a neutral field. The model's neutral-field point spread is the difference between two ratings times a scale factor (now 1.09), so it is close to the plain difference.
- Spread
- The model's predicted margin, written like a betting line: "Ohio State −7" means it expects Ohio State to win by about 7.
- Win %
- The chance of winning one particular game.
- Playoff %
- The share of simulated seasons in which the team made the 12-team field.
- Conference %
- The share of simulated seasons in which the team won its conference title.
- Title %
- The share of simulated seasons in which the team won the national championship.
- Projected wins
- The average number of wins across the simulations, counting the regular season and the conference title game (bowls and the playoff aren't included). A team at 9.4 wins is usually a 9- or 10-win team, not always.
- Strength of record
- How a team's record compares with what a top-25-caliber team would be expected to post against the same schedule. Going 8–4 against a brutal schedule can beat 10–2 against a soft one.
- Vegas
- The betting market's line for the game. The Games tab flags games where the model and the market differ by 3 points or more.
- Why this pick?
- Under each game still to play on the Games tab: how its spread is built. It starts from the two teams' preseason ratings, then adds what each game played this season changed. A game counts for how far it beat or fell short of what the preseason ratings expected, not for who won, so a team expected to win by 25 that wins by 7 moves the other way. "Everyone else's results" are games the two teams didn't play in: they change how good the teams' opponents look, and home field. The parts add up to the model's spread, and the curve below them shows how far the final margin can land from it.
Common questions
My team didn't play. Why did its odds change? Everything is connected. When an opponent your team beat in September wins big, that win looks better and your rating rises. When a conference rival wins, your path to the title game gets harder.
My team won. Why did its odds go down? The model rates on margin. A narrow win over a weak team can lower a rating, because the model expected more. A rival's result the same day can also outweigh your own.
Why does the model disagree with Vegas? The betting market knows things the model doesn't: injuries, suspensions, quarterback changes, weather, and the opinions of everyone with money on the game. The model only knows scores, plus its preseason starting point (past seasons, returning production, recruiting). Disagreements of a few points are normal.
Can I bet on this? Please don't. In games where it disagrees with the Las Vegas line by 3 points or more, the model picked the side that covered the line only about 49% of the time in seasons it was never tuned on, which is a coin flip. It is an independent rating, not a betting edge.
Why 12 teams, byes and these bowls? Those are the real playoff rules for this season, described in step 5 above and on the Bracket & bowls tab.
How good is it?
The Model accuracy tab has the details, including a chart showing that when the model says 70%, the favorite wins about 70% of the time.
What it doesn't know
- Injuries, suspensions, quarterback changes, coaching changes, weather. None of it. Beyond its preseason starting point, the model only learns about a team from its scores, so a star's injury shows up only once it costs the team points on the field.
- Conference tiebreakers are simplified to the rules most leagues share: head-to-head among the tied teams (when they all played each other, or one of them beat all the rest), then record against the conference opponents they all played, then, in the Power 4, their opponents' conference win percentage, then the model's rating in place of each league's own computer or strength-of-record formula.
- Bowl placement is approximated until the sites are announced: the top seeds are assumed to take the closest open quarterfinal bowl.
- The committee stand-in only knows what the real committee has done since 2016. It can't anticipate a new emphasis, an "eye test", or politics.
- The season is long. In September the error bars are wide and almost every decent team has a path. The odds sharpen week by week.
Under the hood
Ratings. Each team has a strength rating fit by Bayesian regression on game margins, adjusted for opponent and home field (FCS opponents share a hierarchical rating). The posterior gives both a rating and its uncertainty.
Preseason priors. Before any games, a team's rating starts from a regression fit on every past season since 2015 (except 2020): last season's rating, the season before, returning production and four-year recruiting. As games are played, results take over.
Simulation. Each simulated season draws every team's true strength from its posterior, plays the remaining schedule, determines conference title-game participants (division winners or top two by conference record; ties go to head-to-head, record against common conference opponents and, in the Power 4, opponents' conference win percentage, then rating) and plays those games.
Committee. A ranking model fit on the committee's actual final rankings (every season since 2016) scores teams on rating, wins, losses, strength of record, conference titles and power-conference membership, with noise for committee judgement. The field and seeds follow this season's rules (step 5 above), then the bracket is played out.