How MMX-F rates college football

One number, one question

MMX-F answers a single question about every FBS team: how many points better than an average FBS team is it, on a neutral field, right now. A team at +21 is expected to beat a perfectly average opponent by three touchdowns on neutral grass. Zero is average. Negative is below it. The scale runs roughly from +25 down to -25, and FCS opponents are pinned at -24 for schedule math and nothing else.

That is the entire output. Everything else on the football pages, from the quadrant tags to the 10,000-simulation playoff odds, is built on top of it.

How the rating is fit

The model is a ridge regression on game margins. Every final FBS-vs-FBS game becomes one equation:

capped margin = home rating - away rating + home-field advantage

Three details do most of the work.

1. Margins are capped at 28 points. A 63-7 win is worth exactly as much as a 35-7 win, so running up the score buys nothing.
2. Every team is solved at once, with the ratings constrained to sum to zero. Beating a good team by 10 moves a rating more than beating a bad team by 10, because the opponent's rating sits on the other side of the same equation.
3. Home-field advantage is fit, not assumed. It is a single number estimated from the same games. The current fit puts it at 2.91 points, and it is zero for neutral-site games, conference championships included.

Ridge regression adds a small pull toward zero for every team, which keeps early-season ratings from swinging wildly on one or two results. The strength of that pull (alpha) was tuned on the 2025 season: fit weeks 1 through 12, then score the model's margin predictions on the games that followed. The winner was a light touch, alpha = 1.

What it deliberately ignores

No polls. No preseason brand. No recruiting rankings. No eye test. The rating sees scores, sites, and opponents, and nothing else. That is a feature. The committee model, which is a different tool with a different job, is the one that gets to care about reputation. MMX-F only cares about who beat whom, by how much, and where.

The preseason prior, and how fast it disappears

In week 1 there is nothing to fit, so every returning team starts at 85 percent of its final 2025 rating, and programs new to FBS start at -10. The prior enters the regression as pseudo-games against an average opponent: eight of them before the first kickoff, one fewer for every game a team actually plays. After eight games the prior is gone and the rating is 100 percent this season's results.

In the first week of the 2026 season the model had exactly eight real games to work with, so the preseason ratings were still doing most of the talking. By mid-October they will be doing none of it.

How it did on 2025

Fit on the 809 FBS games played through week 12 of 2025, the model's predicted margins for the held-out games that followed missed by 12.7 points on average.

Context matters for that number. Game outcomes are noisy: the win-probability formula below assumes a standard deviation of 13.5 points around the expected margin, and that noise alone implies an average miss near 11 points even for a flawless rating. A 12.7-point miss on unseen games leaves some room to improve, which is why a play-quality layer (v1) is on the roadmap, but it sits in the neighborhood of what the noise allows. It is not a claim of clairvoyance. It is a claim that the ratings are honest.

From rating to win probability

A rating gap becomes a win probability through the standard normal distribution:

P(home wins) = Phi((home rating - away rating + home-field advantage) / 13.5)

Two examples. A +10 team hosting a +3 team has a 9.9-point edge after home field, which works out to about a 77 percent chance to win. Move the same game to a neutral site and the edge drops to 7 points, about 70 percent. Nothing in the model ever says a team will win. It says how often it would, over many replays of the same matchup.

Quadrants: the schedule, graded

Basketball settled its resume arguments years ago with quadrant records. Football never had the vocabulary, so we defined it. Every game lands in a quadrant based on where it is played and how good the opponent is right now, using the opponent's current MMX-F rank among the 138 FBS teams.

Every FCS opponent is Q4, full stop. Conference championship games use the neutral band. The cutoffs are proportionally tighter than basketball's because a 12-game season makes Q1 chances rare, and a Q1 win in football should be genuinely precious.

Quadrants re-rate every week using current ranks. A September win over a team that collapses quietly degrades from Q1 to Q2; beat a sleeper who turns out to be great and an old win gets promoted. That movement is deliberate, and it is content in its own right.

Strength of schedule and strength of record

Strength of schedule (SOS) is the average MMX-F rating of a team's opponents, with FCS opponents counted at -24. It says how hard the road is.

Strength of record (SOR) says how well a team has traveled it. It is the team's actual wins minus the wins a reference team rated +18 (a generic top-10 team) would be expected to collect against the exact same schedule, at the same sites. A positive SOR means the team has outperformed what a top-10 team would typically do with that slate. SOR needs played games, so it appears once the season is under way.

The 10,000-simulation playoff odds

Every night the model plays the rest of the season 10,000 times with a fixed random seed, so the same inputs always give the same odds. Each replay:

1. resolves every remaining game with the MMX-F win probability above;
2. sends the top two teams in each conference by conference winning percentage to the title game, breaking ties by head-to-head where it settles the matter and by rating otherwise (a simplification of the real tiebreakers, and labeled as one);
3. plays the championship games on a neutral field;
4. ranks the finished season and selects the field under the 2026 rules: the ACC, Big 12, Big Ten, and SEC champions get in regardless of ranking, the highest-ranked team from the other six conferences gets in, Notre Dame gets in if it finishes ranked 12th or better, and the remaining spots go to the highest-ranked teams left; seeds run in rank order, and seeds 1 through 4 get byes.

Until the committee model activates in the early weeks of the season, the simulation ranks teams by MMX-F itself. Once it is live, each simulated season is ranked the way the committee model expects the committee to rank it.

Probabilities are shown as whole percents. Anything between zero and one percent shows as "under 1 percent", never as "0%", because "impossible" is not something a simulation gets to say.

#	Team	Playoff	Bye	Avg wins	Likely seed
1	IndianaBig Ten	93%	66%	11.1	1
2	Notre DameFBS Independents	93%	45%	10.4	3
3	Ohio StateBig Ten	90%	64%	9.9	1
4	Texas TechBig 12	90%	11%	11.6	6
5	MiamiACC	82%	37%	10.9	4
6	OregonBig Ten	71%	43%	9.7	1
7	UtahBig 12	65%	42%	10.2	2
8	GeorgiaSEC	61%	35%	9.1	2
9	VanderbiltSEC	40%	4%	8.8	7
10	Ole MissSEC	39%	18%	8.7	3
11	Texas A&MSEC	39%	13%	9.0	4
12	South FloridaAmerican	39%	0%	10.1	12

What comes next

Two things are on the roadmap. A play-quality layer (v1) that blends the points-based rating with opponent-adjusted efficiency numbers, shipping only if it beats v0 on the same 2025 test. And the committee model, whose entire purpose is predicting the humans rather than measuring the teams. Its twelve years of backtests are published in full, misses included.