How It Works
How the ratings, simulations, committee model and weekly picks are built, and how they have scored on past seasons.
—
—
How the odds are made
- Rate every team. Each team gets a rating in points from this season’s games, adjusted for opponent and for where the game was played (home field is worth — points). Point margin counts, and so does per-play efficiency.
- Start from the roster, not the name. Before any games, a team’s rating comes from last season’s production of the players on this season’s roster, wherever they played, plus the roster’s recruiting talent. That starting point counts about as much as — games of results, so it fades as the season goes on.
- Play out the rest of the season. The remaining schedule and conference championship games are simulated — times, with uncertainty in each team’s true strength.
- Rank each simulated season the way the committee does. A model fit on every final CFP ranking since the playoff began turns each simulated season’s records, schedules, quality wins and conference titles into a ranking. Given the real full-season results, it matches — of the final top 4 on average.
- Count. A team’s chance of finishing #1 is the share of simulated seasons in which the committee model puts it there.
What the committee model learned about the committee: a home loss costs about — more than a road loss, power-conference teams get a large edge, and a preseason #1 gets about — extra win’s worth of credit over an unranked team, even with identical results. The model includes these habits because its job is to predict the committee, not to judge it.
How the weekly picks work
- Same ratings, one game at a time. After each week, every team’s rating (in points better than an average FBS team) predicts each of next week’s games: the rating gap, plus — points for the home team, gives the predicted margin. The chance to win comes from how far real games land from predictions.
- Published Sunday night, then locked. Picks go up after Sunday’s AP poll and never change. Results fill in as games are played.
- Graded against Vegas. The benchmark is the consensus closing betting line: the toughest fair comparison there is, since it moves on news all week while the pick stands from Sunday.
- Reconstructed early weeks. Weeks 1–7 were picked when this page launched, using only results known before each week. Rosters come from the current data pull, so those weeks are labeled and counted separately from live picks.
The three playoff formats
| 12-team (2026 rules) | 24-team: CFP plan | 24-team: All Champions | |
|---|---|---|---|
| Automatic bids | SEC, Big Ten, ACC and Big 12 champions, plus the highest-ranked Group of Six champion | SEC, Big Ten, ACC and Big 12 champions, plus the highest-ranked Group of Six champion | All ten conference champions |
| At-large | 6 (7 if Notre Dame isn’t protected) | 19 | 14 |
| Notre Dame | In if it finishes in the top 12 | No special treatment | No special treatment |
| Byes | Seeds 1–4 | Seeds 1–8 | Seeds 1–8 |
| Seeding | Straight by projected ranking; first-round games on the higher seed’s campus; the team the model favors advances | ||
Compared with a Colley-style ranking
The Colley Matrix ranks teams from wins and losses, adjusted for the opponents they came against, and is the base of zero-based rankings such as Peach Pundit’s. The two answer different questions.
| Colley-style ranking | This model | |
|---|---|---|
| Question | Who has earned what so far? | Where will each team finish in the committee’s final ranking? |
| Inputs | This season’s wins and losses | Margin, efficiency, venue, roster production and talent, the remaining schedule, the committee’s habits |
| Point margin | Ignored, so running up the score earns nothing | Used: it predicts future games about — better than wins and losses alone |
| Home vs. road | Treated the same | Adjusted for venue |
| Early season | Five or six results carry everything | A roster starting point steadies it, then fades |
| Games left to play | Not considered | Simulated |
| Output | One ordered list | A probability for every position |
| Can it be checked? | No: “deserved” has no answer key | Yes: scored against final rankings in seasons it never saw |
| Reputation | None: everyone starts at zero | Reproduces the committee’s, because that is what it predicts |
Where this model is stronger
- It can be graded. It predicts something that happens, so every weight was fit on past seasons and scored on seasons it never saw.
- It uses more of what the games say. Margin and venue carry real information that wins and losses drop.
- It is steadier early. With five or six games, a win/loss rating swings on one result. The roster starting point damps that.
- It looks ahead. Two teams with the same record can face very different remaining schedules.
- It shows how close teams are. A list puts a clear favorite and a long shot one spot apart. Probabilities show the gap.
Where a Colley-style ranking is stronger
- It measures merit cleanly. No preseason information and no committee habits are built in.
- Anyone can check the math. It is a few lines of algebra anyone can rerun from the scores. This model has many more parts to audit.
- It doesn’t reward running up the score.
- It needs only scores. Nothing breaks if a data source changes.
- It doesn’t inherit bias. This model reproduces the committee’s treatment of Group of Five teams and its lean toward preseason favorites. An unbeaten James Madison sits low here for that reason.
- Its order means one thing. Here a team can move without playing, and the order is expected finish, not “best team.”
Track record, 2015–2025
Average chance each method gave, at that point in the season, to the teams that actually finished #1, in the top 4 and in the top 12 of the final CFP ranking. Seasons 2015–2025, without 2020; every season was predicted without its own data. Higher is better. “Betting-market ratings” are team ratings implied by point spreads, run through the same simulation and committee model.
What to keep in mind
- The track record covers — seasons. Small gaps between methods are within normal season-to-season noise.
- The committee model learned mostly from the four-team playoff era. If the committee treats Group of Five champions differently in the 12-team era, the model will be slow to catch it.
- Odds come from simulation, so a team’s chance can move by about a point between runs with no new games.
- Offensive linemen and defensive backs enter only through recruiting talent. Box scores don’t measure them.
- Conference title-game tiebreakers are approximated by conference record, then overall record.