How the college basketball ratings work

Teams play at very different speeds, so points per game mixes how good a team is with how fast it plays. Every rating here is per 100 possessions instead. Possessions are estimated from the box score as FGA − OREB + TOV + 0.475 × FTA. The 0.475 is the long-standing coefficient for trips to the free-throw line that end a possession (two-shot fouls end one, and-ones do not); using 0.5 would bias every number the same way.

• AdjO — points a team would score per 100 possessions against an average Division I defence on a neutral floor.
• AdjD — points it would allow per 100 possessions against an average offence. Lower is better, so it is ranked from lowest.
• AdjEM — AdjO minus AdjD: how many points per 100 possessions better than an average team. This is the ranking number.
• AdjT — possessions per game, corrected for the pace of the opponents a team has played. A game's pace is set by both teams, so a team that meets slow opponents looks slower than it is.

Every team-game is one observation: the points a team scored per 100 possessions against one opponent, at home, away or neutral. The model treats that as the league average, plus the team's offence, plus the opponent's defence, plus a home-court term — and solves for every team's offence and defence at once. The solve alternates: offences given the current defences, then defences given the offences, then the home advantage, until nothing moves by more than a thousandth of a point.

Home advantage is fitted, not assumed. It is roughly three points per 100 possessions in college basketball but moves from year to year, and a wrong constant leaks into the ratings of teams with unbalanced home and away schedules. The fitted value is printed under the ratings table.

Shrinkage. Each team is pulled toward the league average by the equivalent of five games of average results. In November a team may have played three games, two against overmatched opponents, and an unshrunk fit would hand it an extreme rating. At five games a team is pulled halfway back; by thirty games the pull is about a seventh of the estimate. All games count equally — weighting recent games more heavily was tested on past tournaments and made the ratings worse.

The raw numbers (unadjusted points per 100) are kept beside the adjusted ones. The ratings table's Δ SoS column shows how many places a team moves once its schedule is accounted for — the size of the correction, which is the reason to adjust at all.

The ratings are not one end-of-season fit. They are refitted every week on only the games played up to that date, and each fit is stored with its date. That gives a time machine: the ratings as they stood on any Monday of the season, which is what a team page's trajectory chart draws.

It is also what keeps a projection honest. A projection only ever uses a snapshot dated before the game. A snapshot dated on or after a game has already seen its result, so a “projection” from it would be describing a score it was told. Every projected number on the scoreboard, game pages and team pages carries the date of the snapshot it came from, and games before the season's first snapshot get no number at all.

For two teams, the projected margin is the difference in AdjEM, plus the home advantage when one side is at home, multiplied by the expected number of possessions (the average of the two adjusted tempos) and divided by 100. That is the only formula the scoreboard, the game page, the team page and the compare page use.

The stats pages also show Dean Oliver's four factors for and against — effective field-goal percentage, turnover rate, offensive rebound rate and free-throw rate. They are computed from season totals of the box scores and are not adjusted for opponents; they explain why a team rates where it does rather than replacing the rating.

These are measurements of what has happened, adjusted for who it happened against. They are not graded in public on this site, and nothing here says how they compare with any other rating system.

See the ratings table, the explorer or the four-factor leaderboards.