How the College Basketball Model Works
NCAA men's Division I basketball. A research board: it shows what the model thinks and how that has held up, and it makes no picks.
The model
Every team carries an offense, a defense and a pace rating, updated one game at a time and only from results that were known before the game being forecast. The ratings feed a simulation of each game's final score, 10,000 times, including overtime. The spread, the total and the win probability are all read from that one set of simulated scores, so they cannot disagree with each other.
The forecast is team-level: it does not yet adjust for rosters, injuries or minutes. Games against teams outside Division I carry wider uncertainty, and every card says so.
The market line
Beside our numbers we show a reference line: the middle of the sportsbook lines collected before the forecast. No book is named and no price is shown. With no price there is no market win probability, so the win row shows ours alone.
What research status means
Before running the model on last season we wrote down the tests it had to pass. Until it passes them, it publishes as research: no picks, no EV and no claim beyond what the record shows. The backtest page has every test with its numbers, and the record scores every frozen forecast this season, including early line value: whether the line moves toward our number after we post, shown whether that is good or bad.
When we freeze
Two forecasts a day, each frozen and never edited afterwards: the night before tip at 05:30 UTC, and in the morning at 14:30 UTC. The board shows the latest freeze; the record scores both.
What the board may claim
- Status
- research
- Picks
- No
- EV
- No
- Roster adjustments
- No
- Injury adjustments
- No
Gates
- G-B0: Did not pass5 of 6 criteria passed; failed: interval coverage 50/80/95, margin and total. Registered 2026-11-07
- G-B1c: Registered, not yet runRegistered 2026-11-11; scoring block 2026-12-01 to 2027-03-08; binding run after 2027-03-09