College Basketball Power Rankings

				Last	Next

					Last	Next

Projected field

% = chance to make itACC projected automatic bidOpening Round

Title odds

Team	Seed	R32	S16	E8	F4	Final	Title

Bubble

Team	Conf	Rating	SOR	At-large

Bids by conference

Expected NCAA teams

Today's games

— —

— · — · — in conference

Rating

—

— · — in 7 days

Proj. W–L

—–—

80%: —–— wins

Conference

—–—

Title — · Tourney —

NCAA

—

Seed — · Title —

Preseason

—

KenPom — · Torvik — /100 poss.

SOS · SOR

— · —

Wins

Tournament path

NCAA seed

Rating

Games left

— left · — exp. W

HomeNeutralAway5-game avg

Results

Date		Opponent	Result	Score	Pregame	Δ

Schedule

Date	ET		Opponent	Win	Spread

How the model works

Ratings. Published in points against an average D1 team: 0 is average, and on a neutral court a +10 team is favoured by 10 points over an average one. A game's spread is the difference of the two ratings plus home court and the other context terms. The engine runs in Elo underneath; one point is — Elo.
Priors. With 360-odd teams there are no Vegas season lines to start from, so each team opens from last season's final rating pulled toward its conference's level, preseason efficiency margins from KenPom and Bart Torvik, or a blend; a team new to D1 starts from a fitted level. Best in testing: —.
Updates. After every game between two D1 teams both sides move by K × a margin-of-victory multiplier × (result − expected). Margins are capped, so blowouts stop counting past the cap, and an autocorrelation term shrinks the multiplier when a favourite wins big. Games against non-D1 opponents count in the record but move no ratings.
Game context. Home court shifts a game's expected result (not at neutral sites). Each other factor was tested by leave-one-season-out fits and stays in only if it lowers held-out log loss. In: —. Out: —.
Forecast. The rest of the schedule is played — times with ratings updating game by game, then every conference tournament from its own bracket, the NCAA selection (automatic bids, then at-large teams by rating and strength of record), S-curve seeding with the Opening Round (the 76-team field from 2027), and the bracket.

Priors tested

Each way of setting the preseason ratings, scored on the same games; Δ is against the best.

Prior	Seasons	Games	Log loss	Δ vs best	Note

Backtest

Pregame win probabilities against results, games between D1 teams. Lower log loss and Brier are better; the best on each split is bold.

Model	Split	Seasons	Games	Log loss	Brier	Accuracy

Calibration

Test split: predicted home win chance against how often the home team won. Dot size is the number of games.

Core parameters

Fitted on past seasons; rating-scale values in points.

Parameter	Value	Unit	Note

Factors tested

Change in log loss per game when the factor is added; negative helps. Held-out: averaged over leave-one-season-out fits on the training seasons, and it alone decides Kept. Better in: how many held-out seasons improved. Test: the same comparison on the test seasons, which played no part in the choice.

Factor	Value	Unit	Status	Held-out Δ	Better in	Test Δ