Who wins the 12-team playoff?

Links go to Amazon. CFB Natty may earn a commission from qualifying purchases, at no extra cost to you.

How it works

1. Power ratings. Every completed game this season is fit with a margin-of-victory model (blowouts capped at 28 points, home field worth 2.5). Each team starts from a baseline built on last season's results, which this season's games gradually override. That steadies early-season noise and reflects a program's talent level. A rating of +10 means 10 points better than an average FBS team on a neutral field.
2. Remaining games. When a game has a posted betting line, the prediction leans 90% on the line, since betting markets already price in injuries, matchups and coaching. Games without a line use the rating gap. Each simulated season also gives every team a little random up-or-down swing in strength, because ratings aren't perfect.
3. Conference championships. The top two teams in each conference standings (or division winners) meet in a title game.
4. Committee selection. Teams are ranked on strength, wins and losses, schedule, conference titles and how they're currently ranked: once the committee releases its rankings, those count 60% and the newer AP poll 40%. Like the real committee, a team is ranked ahead of a close rival it beat head-to-head. The five highest-ranked conference champions get automatic bids, seven at-large teams fill the field, and seeds 1–12 go in ranking order.
5. The bracket. Seeds 5–12 play first-round games on the higher seed's campus. Seeds 1–4 get byes to the quarterfinals, then it's single elimination to the title.
6. Tested against real results. Before going live, each piece was checked by replaying this season week by week, predicting each week's games using only earlier results. The last-season baseline and the betting lines both made predictions measurably more accurate. Weighting recent games more did not, so it isn't used.
7. Injuries. Injured starters lower a team's strength only for games before their expected return. A player back in December counts again for conference title games and the playoff. The AI analyst researches injury news twice a week, and the site owner can add details it misses.
8. What we tested and left out. Teams don't reliably bounce back after a loss or let down after a big win. Across about 2,000 games over two seasons, teams coming off a blowout did slightly better than our ratings expected, but no better than the betting line, which already carries most of the weight for the next game.
9. Vegas comparison. National championship odds from DraftKings are converted to percentages (with the bookmaker's built-in margin removed) and shown next to ours. The headline champion averages the two, since betting markets often see talent and matchups that results alone miss.
10. Predicted scores. Each team gets offense and defense ratings from the points it scores and allows. A game's predicted score combines those with the expected margin, and leans on the betting over/under when one is posted.
11. AI analyst. Before each update, Claude reviews the model's numbers, searches current news (injuries, quarterback changes, the committee's rankings) and can move a team's rating by up to 4 points, with a reason shown for each change. Turn it off above to see the pure model.