"It is difficult to make predictions, especially about the future."
These are the forecasts for 2025. For the previous year's forecast, see the March Madness 2024 page.
Forecasts are made with two types of models: Bayesian matrix factorizations following Salakhutdinov and Mnih (2008) and Elo-type models. The factorization models are referred to as 'PMF.' The 'Baseline' PMF models are estimated without hyperpriors, but '+Hyperpriors' models include conjugate hyperpriors for the unconditional mean and variance of latent offense/defense vectors. The Elo-type models are referred to as 'Aug. Elo' and the '+Margin of Victory' version models the margin of victory jointly with the binary game outcome.
The estimated models produce the posterior odds of team 'x' beating team 'y.' These are presented graphically under 'Matchups.' The 'Bracket' produced by each model is generated by choosing the team favored by the posterior odds.
PMF (Baseline): [Matchups], [Bracket]
PMF (+Hyperpriors): [Matchups], [Bracket]
Aug. Elo (Baseline): [Matchups], [Bracket]
Aug. Elo (+Margin of Victory): [Matchups], [Bracket]
PMF (Baseline): [Matchups], [Bracket]
PMF (+Hyperpriors): [Matchups], [Bracket]
Aug. Elo (Baseline): [Matchups], [Bracket]
Aug. Elo (+Margin of Victory): [Matchups], [Bracket]