Mark E. Glickman, JQAS Editor-in-Chief

The March 2015 special issue (volume 11, issue 1) of the Journal of Quantitative Analysis in Sports (JQAS) consists of five articles on prediction models and methods for the NCAA men’s basketball tournament. The solicitation for the articles in this issue originated from the “March Machine Learning Mania” Kaggle prediction contest held last year in which contestants made probability predictions for all the team matchups in the 2014 NCAA tournament. The articles in this issue feature innovative modeling methods and strategies adopted by the authors.

“Building an NCAA Men’s Basketball Predictive Model and Quantifying Its Success” by Michael Lopez and Gregory Matthews, the ultimate winners of the 2014 Kaggle competition, demonstrated that a fairly simple model can result in strong predictive accuracy. Their approach consisted of constructing an ensemble model based on a weighted average of two logistic regressions. Given their top performance in the Kaggle contest, they examined the role of luck through a simulation study.

“A Mixture-of-Modelers Approach to Forecasting NCAA Tournament Outcomes” by Luke Bornn et al. also fit an ensemble model to predict game outcomes, though theirs consisted of three logistic regressions, a stochastic gradient boosted tree model, and a neural network model. Recognizing that their performance statistics to predict previous years’ NCAA tournament games incorporated data from the tournaments themselves, they provided details on decontaminating these variables.

“Nearest-Neighbor Matchup Effects: Accounting for Team Matchups for Predicting March Madness” by Andrew Hoegh et al. used a linear model for score differences, but included novel game-specific adjustments to account for non-transitivity of team strengths. Specifically, they included a model term that accounted for the over- or under-performance by each team in a matchup through the outcomes against teams with the similar background performance statistics.

“A Generative Model for Predicting Outcomes in College Basketball” by Francisco Ruiz and Fernando Perez-Cruz is the Editor’s Choice article for the issue, and is available for free download for 12 months after the issue is published. The authors developed a model in which the final scores of a game follow separate Poisson distributions as a function of both offensive and defensive parameters. Their model has strong connections to a commonly used model to estimate soccer team strengths. They fit their model using a variational inference algorithm to approximate the posterior distribution.

Finally, “A New Approach to Bracket Prediction in the NCAA Men’s Basketball Tournament Based on a Dual-Proportion Likelihood” by Ajay Gupta submitted his manuscript independently of the Kaggle competition. His article focused more on traditional bracket predictions. Gupta developed a “dual-proportion” model in which the likelihood function weighted conference tournament games more heavily than regular season games, and accounted for individual games resulting in blowouts.

With the exception of the Editor’s Choice article, all articles will be freely accessible for download through April 15, after which they are available on a subscription basis. They are available at the JQAS website, along with the journal’s aims, scope, and manuscript submission instructions.