Every figure below was computed the way the model works on game day: from games played before tip-off, last season's ratings and each team's roster. Lines are ESPN's closing line (DraftKings first). Division I games only; the 2022 season is the warm-up and 2026 was held out when the model was tuned. Rosters are kept from 2026-27 on, so past seasons use the players who appeared for the team that year, which flatters the early-season record a little.
Straight talk: the model has not beaten the closing line. Its average miss on the margin is about a fifth of a point worse than the market's and on totals about four tenths worse, and its spread and over/under picks have won about half of their bets. It is a way to see where it differs from the market, not a betting system.
Season by season
| Season | Games | Winner: model | market | Margin miss: model | market | Total miss: model | market | ATS, edge 3+ | O/U, edge 3+ |
|---|---|---|---|---|---|---|---|---|---|
| 2022-23 | 5721 | 72% | 72.8% | 8.898 | 8.806 | 13.147 | 12.991 | 491-439 52.8% | 858-783 52.3% |
| 2023-24 | 5726 | 72.6% | 72.9% | 8.901 | 8.842 | 13.17 | 12.895 | 377-319 54.2% | 985-972 50.3% |
| 2024-25 | 5764 | 72.8% | 73.5% | 8.931 | 8.79 | 12.99 | 12.664 | 341-324 51.3% | 950-1029 48% |
| 2025-26 | 5752 | 72.3% | 72.8% | 9.024 | 8.938 | 13.736 | 13.371 | 409-367 52.7% | 1221-1213 50.2% |
| 2026-27 so far | 3 | 66.7% | 33.3% | 5.327 | 8.5 | 22.62 | 21.167 | 1-0 100% | 1-0 100% |
By kind of game, 2025-26
| Games | Margin miss: model | market | ATS, edge 3+ | O/U, edge 3+ | |
|---|---|---|---|---|---|
| conference games | 3660 | 8.744 | 8.735 | 149-111 57.3% | 712-721 49.7% |
| non-conference | 1987 | 9.536 | 9.306 | 250-247 50.3% | 478-464 50.7% |
| power v power | 1026 | 9.097 | 9.109 | 84-55 60.4% | 240-242 49.8% |
| power v mid-major | 655 | 9.756 | 9.544 | 105-103 50.5% | 174-131 57% |
| mid-major v mid-major | 4071 | 8.888 | 8.798 | 220-209 51.3% | 807-840 49% |
| neutral site | 700 | 9.054 | 9.081 | 76-48 61.3% | 182-214 46% |
| conf tournaments | 300 | 8.073 | 8.1 | 13-12 52% | 73-98 42.7% |
| NCAA tournament | 67 | 9.184 | 9.351 | 7-4 63.6% | 24-24 50% |
| NIT / other post | 38 | 8.925 | 8.553 | 3-5 37.5% | 7-4 63.6% |
| Nov-Dec | 2208 | 9.425 | 9.214 | 267-262 50.5% | 527-510 50.8% |
| Jan-Feb | 2918 | 8.78 | 8.772 | 111-78 58.7% | 540-545 49.8% |
| first 6 games of a team | 1183 | 9.656 | 9.317 | 164-176 48.2% | 291-287 50.3% |
1st half
The 1st-half model is fitted from the full-game model and past halves. ESPN posts no 1st-half lines for college games, so it is graded on the half-time score only.
| Season | Games | Winner | Margin miss | Half of full game | Total miss | Half of full game |
|---|---|---|---|---|---|---|
| 2022-23 | 5715 | 66.7% | 7.275 | 7.311 | 7.833 | 8.638 |
| 2023-24 | 5726 | 66.9% | 7.172 | 7.19 | 7.858 | 8.744 |
| 2024-25 | 5764 | 66.7% | 7.356 | 7.341 | 7.958 | 8.936 |
| 2025-26 | 5752 | 67.5% | 7.357 | 7.354 | 8.205 | 9.305 |
| 2026-27 | 3 | 100% | 3.037 | 2.943 | 6.633 | 8.477 |
Research log
What we tested, and what came of it. Nothing marked "Not used" changes a number on this site.
This year's roster in the preseason ratingIn the modelNov 2, 2026
Does knowing who is on the roster, transfers included, make the early-season numbers more accurate?
Yes. 30% of a team's preseason rating now comes from its roster at last season's player ratings (tuned on 2023-25). On the held-out 2025-26 season the average miss on the margin in a team's first six games fell from 9.86 to 9.64 points; the closing line's was 9.32, so it still trails the market. Past seasons used the players who appeared for each team, which flatters the result a little; real rosters are kept from 2026-27 on.
3-point luck adjustmentIn the modelOct 22, 2026
Does taking the noise out of 3-point shooting, the way Torvik and KenPom do, make the ratings more accurate?
Yes, a little. A team's opponents' 3-point percentage is mostly chance, so 35% of each side's 3-point over- or under-performance against the league rate is removed from the score before the ratings are fitted (the share was tuned on 2023-25). On the held-out 2025-26 season the average miss on the margin fell from 9.14 to 9.08 points against the closing line's 8.94, and on the total from 13.81 to 13.76 against 13.37. Spread picks at 1.5 or more points of edge went to 52.5%, which is at the break-even line, not over it, and the earlier seasons sit at 50-53%. The closing line is still ahead.
Four Factors and Elo on top of the ratingsNot usedOct 22, 2026
Do shooting, turnover, rebounding and free-throw rates, or an Elo rating, add anything once the model has a rating for each team?
No. Each team's season-to-date Four Factors (both sides of the ball) and a points-scale Elo were fitted as corrections to the model line, season by season, with each season predicted from the others. Every weight came out near zero and the margin miss did not move. This matches the published finding that the factors carry no information a rating does not already have.
Home court per teamNot usedOct 22, 2026
Is a team's own home-court edge worth modelling, instead of one figure for the whole division?
No. Giving every team its own home-court term, pulled toward the league figure, changed the margin miss by 0.002 points. Published studies find the same: a team's home edge cannot be told apart from noise in one season of games. Altitude and travel distance could not be tested, since the site has no venue coordinates.
A preseason prior that fades as games are playedNot usedOct 22, 2026
Should last season's rating count for more in November and fade through the season, as the ESPN BPI and KenPom preseason ratings do?
No. A prior that starts strong (worth 10, 20 or 40 games) and decays toward the usual weight was worse at every setting, by 0.01 to 0.17 points, and worst early in the season, which is where it was meant to help. Rosters turn over too much in college for last season to deserve more weight than about five games.
Pace times efficiencyNot usedOct 22, 2026
Is a tempo-free model (points per possession times possessions, the KenPom form) better than points per game?
Not overall. Fitting efficiency and pace separately from the team box scores made the total miss 0.03 points better at best and the margin miss 0.02 to 0.10 points worse. Two models for one gain of three hundredths of a point is not worth it.
Win probability curveNot usedOct 22, 2026
Is the win probability shown on game cards well calibrated, and would a fitted curve do better?
It is calibrated as it stands. The normal curve on the model line (spread of misses about 11.4 points) scores the same as a logistic curve fitted to past seasons and slightly better than one that widens with the game total; in every 10% band the teams won within 2-3 points of the stated chance. The closing line's own curve is a little sharper (Brier 0.174 against 0.177). Nothing changed.
The opening line as an inputNot usedOct 22, 2026
For information only: if the model were allowed to see the opening line, could the blend beat the closing line?
No. A blend of the model line and ESPN's opener, fitted on earlier seasons, got the margin miss to 9.05 points on 2025-26, still behind the close at 8.94. The model never uses a market line, so this changes nothing on the site; it is noted because it shows how close to the ceiling the closing line sits.
AP rank as a preseason hint and damped blowoutsIn the modelOct 22, 2026
Does a team's AP ranking in its first game help the prior, and should 40-point margins count in full?
Both help a little, about 0.01 points of margin miss each. A ranked team's prior rating gets a bump that scales with its rank (up to 8 points for the top), and the part of a margin beyond 40 points counts at 30%, since garbage time says little about either team. Both are in the model.
Returning minutes and player availabilityNot usedOct 22, 2026
Does knowing how much of last season's rotation came back, or who is missing tonight, improve the prediction?
Neither did. Scaling how much of last season's rating carries over by the share of returning minutes changed nothing, because so few minutes return (Duke brought back about a quarter). A player layer that watches who was missing from the previous game's box score was worth 0.01 points even when it was told who actually sat out. College basketball has no public injury report, so there is nothing better to feed it.
How it works
- Every Division I team has an offence and a defence rating, refitted each game day from the season's results, so beating a strong team counts for more than beating a weak one.
- At the start of a season each team begins from last season's rating, pulled toward its conference's average because rosters turn over so much in college basketball, then adjusted for who is on this year's roster: returning players and transfers at last season's box-score numbers, newcomers rated on their program.
- Home court is worth about three points and nothing on a neutral floor. The NCAA tournament and conference tournaments are modelled the same way.
- It tells you where it differs from the market (Model edge), in points. A big edge is a reason to look, not a bet. College basketball has no public injury report, so the model cannot see who is missing.