How CBB Stats Numbers Hold Up
What each CBB Stats number measures, what kind of data it draws on, how it held up when tested, and where it falls short.
FEDERER team ratings
What it measures: team quality, read from how possessions play out across a game rather than from the final score alone.
What it draws on: play-by-play from Division I games.
How well it holds up: Not yet validated.
Known limits: non-Division-I teams are excluded, and ratings are refreshed on a schedule, so they trail live games.
Transfer value
What it measures: an in-house estimate of how much a transfer-portal player is worth to a roster, for sorting the portal.
What it draws on: player ratings and portal context.
How well it holds up: Not yet validated.
Known limits: it is a sorting aid, not a prediction of how a player will perform at a new school.
Next-season projections
What it measures: how a player's stats are expected to change next season.
What it draws on: past seasons of Division I box-score stats and team context.
How well it holds up: a version of the model trained only on seasons through 2024-25 was tested on the 2025-26 season, which it had not seen. Depending on the stat, it explained between 15% and 47% of the year-to-year change (audit of 2026-08-24). The model the site uses is then refit on every season, including 2025-26.
Known limits: shooting percentages on few attempts are mostly noise, so their projections are loose. A stat that fails the test is hidden rather than shown; defensive rating is hidden for this reason.
Portal-entry probability
What it measures: the chance that a player enters the transfer portal after a season.
What it draws on: season stats, team context and past portal entries.
How well it holds up: on the held-out 2026 cycle it ranked players who entered above players who did not enter 70.6% of the time (95% interval 69.2% to 72.1%), and its predicted probabilities sat within 2.5 percentage points of the actual rates on average.
Known limits: it is weakest for seniors (64.8%). It cannot see private reasons such as playing-time promises, money or personal circumstances. A season the model was built on is scored in hindsight and is flagged as such.
Depth charts
What it measures: a projected rotation for next season, five position slots per team, ordered by minute share.
What it draws on: this season's stats and roster information.
How well it holds up: Not yet validated.
Known limits: seniors, portal players and players who have left college are excluded, so it is a roster projection and not a record of who played.
Luck-adjusted shooting
What it measures: shooting percentages with the luck of a small sample pulled back toward what a player's attempts support.
What it draws on: box-score shooting totals across a player's seasons.
How well it holds up: in a run of 2026-08-29 over 81,753 player-seasons from 17 seasons, the adjusted percentages predicted the next season's shooting with 22% to 44% less error than the raw percentages. The adjustment's settings were fitted on these same seasons, so this is not a held-out test.
Known limits: it helps most for three-point and mid-range shooting, where a season is mostly noise, and least for free throws and shots at the rim. Players with few attempts move the most.
Derived box-score stats
What it measures: points, minutes, rebounds, assists and turnovers, as totals, per game and per 40 minutes, which the database does not store directly.
What it draws on: rate stats and season totals from box scores.
How well it holds up: Not yet validated on held-out data. A spot check on nine players found derived points per game within 0.1 of ESPN's reported points per game for seven of them.
Known limits: the offensive and defensive rebound split is an approximation. Steals and blocks are not derived, because the data to do so does not exist in our collections.