Simulations
Every rating system reduced to the same thing, a win probability for any two teams, then made to play this bracket 200,000 times. Ours is the one we ship. The rest are outside opinions, and the public is who you're actually up against in a pool.
The scorecard
How well each system called the real tournament games (log loss, lower is better). Haslametrics and ESPN BPI are left out on purpose: we can only get their ratings as they stand today, which for a finished season already knows how it went.
Who wins
Model vs. the public
Edge = how much more often the model has a team here than the public's brackets do. Positive means the field is under-picking them, which is where pool value comes from.
Each system's bracket
Where these numbers come from
Each system publishes a team rating, not matchup odds. They're put on one scale the same way: P(a beats b) = σ(k · (ra − rb) · tempo factor), with one slope k per system fitted on real tournament games, so no system gets an unfair sharpness advantage. None of them is an input to our model.