Reigning champions

All-time scorecard

Season by season

Log losspunishes overconfident models while rewarding well-calibrated probabilities. Lower is better (0.693 is the log loss of a model that picks at random) Accuracyshare of games the projected favorite won Tourney pointsBracket score (each first-round game is worth 10 points, doubling each round) if each model filled out a bracket by picking its favored team in each round (1,920 max). Expected pointsaverage number of tourney points a bracket filled out using a single simulation based on the model's odds at the start of the tournament would earn.

Composite averages the three models. Torvik's odds are log5 of his Selection Sunday Barthag, fixed before the tournament like the models' (2010 on; his ratings archive starts then). Chalk picks the better seed and counts as 99.99% sure for evaluation purposes.