Method
MRI is a computer rating system. Ben ran it as an Excel workbook for college football and college basketball; the basketball workbooks stop at 2019–20. This is the same idea rebuilt, and carried forward through the seasons in between.
The rating
Every game becomes one equation —
margin = rating(home) − rating(away) + home court
— and the whole season is solved at once. A team's rating depends on its opponents' ratings, which depend on theirs, all the way down, so strength of schedule is not a separate statistic bolted on afterwards. It is what the solution is made of.
- Margin is compressed. A 30-point win is worth more than a 20-point win and not very much more, through a curve that is near-identity inside a normal result and flattens past it. Nothing is gained by running up a score.
- Ratings are pulled toward a preseason estimate built from last season's rating and who is on the roster, so November means something. The pull fades on its own as games accumulate, and it is gone well before conference play.
- Home court is estimated from the games rather than assumed, and neutral sites are excluded. It sits at 3.0 points here, and has run between 2.7 and 3.3 across the six rebuilt seasons.
- Non-Division-I opponents are rated individually rather than pooled, so a November exhibition against a good one is not the same result as a November exhibition against a bad one.
Two numbers
Power answers how good a team is, in points against an average Division I team, so the gap between two teams is a predicted spread. Résumé answers what a team has earned, measured as wins above what an average team would have managed against the same schedule at the same venues. They are different questions and they are published separately.
Preseason priors
A season starts from a prior, and basketball's was last season's rating pulled 35% of the way to average for everyone. Rosters turn over faster in college basketball than anywhere else, so that treats a team that lost its whole rotation like one that kept it. The prior is now a regression on last season's rating and what is known about the roster: returning production (the share of last season's win shares still on the roster), incoming production (what the newcomers did last season somewhere else, which is the transfer portal) and the freshman class.
Rosters for a coming season are posted late. Before they are, the prior uses only what last season and the draft already imply: how many of last year's minutes belong to players in their fourth year or later or headed to the NBA draft, and the freshman class. Rosters for 2026–27 are posted for 365 of 365 teams, so this season's preseason ratings use the roster. The preseason rankings show where each team starts.
Scored on how good teams turned out to be, leaving each season out in turn, a season's final rating is missed by 5.1 points under the old rule, 4.6 with the before-rosters version and 4.2 with a roster. Part of the gap to the old rule is only its weight on last season being tuned for something else; the fair comparison is a last-season-only refit at 4.7. The game-by-game test below is the one to trust.
Every game of 2014–2026 (2021, the COVID season, chained but not scored), predicted from only the earlier games of its own season, in slices of the season (the average miss, in points, of each prior; the last column is what a roster gains over the old rule). Each prior is chained properly, starting every season from the previous season's final ratings under the same method, and the new priors' coefficients are fitted leaving the predicted season out.
| Season so far | Games | Old | Before rosters | With roster | Gain |
|---|---|---|---|---|---|
| 0%-5% | 3,626 | 13.91 | 13.89 | 13.75 | +0.16 ±0.06 |
| 5%-10% | 3,629 | 12.60 | 12.46 | 12.31 | +0.29 ±0.08 |
| 10%-20% | 7,260 | 10.54 | 10.42 | 10.29 | +0.25 ±0.05 |
| 20%-40% | 14,520 | 10.46 | 10.36 | 10.30 | +0.16 ±0.03 |
| 40%-100% | 43,567 | 9.13 | 9.07 | 9.06 | +0.07 ±0.01 |
The gain is real and modest, and it fades. Basketball plays about 6,000 games among 365 teams, so by the time a fifth of the season is gone the games have taught the model most of what the prior was guessing. Football, with a fraction of the games, keeps the benefit longer.
Two adjustments, found by testing and not by argument. The new prior keeps the division's level where the old one had it, because a level that is a point off lands on every game against a non-Division I opponent, which is most of November. And its spread is pulled in to 70%: at full spread it is worse than the old prior, at that setting it is better in every part of the season, and at 50% it starts to give the gain back.
What is and is not established
The original formula was ported and checked against the surviving workbooks: it reproduces every published rating in 2012–13, 2017–18, 2018–19 and 2019–20 to floating-point tolerance. That is the part that is proven.
Ratings since then are a chain: 2019–20 primes 2020–21, which primes the next, through to now. Walking forward within each season — fit on the games played, score the games that come next — the model calls about 70% of games correctly with a mean margin error near 9.3 points.
Its settings were searched over 80 combinations and the search found nothing that beat the starting guesses on seasons it had not seen. That is reported here because it is true: at this many games per team the knobs stop mattering, and a tuning exercise that claims a win it did not get is worth less than one that admits it.
What it cannot do
A margin error near 9.3 points is about where college basketball closing spreads sit. As with football, that is the honest signal that this should not be expected to beat a closing line. There is no betting page for basketball and there will not be one unless a backtest earns it.