Predicting discrete outcomes with the maximum score estimator: the case of the NCAA men's basketball tournament

Seedings as a predictor of winning in the men's NCAA basketball tournament have recently been examined by Boulier and Stekler, who estimate a probit model to establish a relationship between seedings and the probability of winning. This study discusses the merits of a maximum score estimator for the prediction of discrete outcomes. Unlike the probit model, the maximum score estimator maximizes the number of correct predictions. The maximum score estimator is applied to updated data on the men's NCAA basketball tournament. The score estimator has better in-sample performance than the probit model used by BS. When out-of-sample predictions are examined using a series of rolling or recursive regressions, the maximum score estimator performs slightly better than the probit/maximum likelihood models. These results illustrate the potential advantages of using the maximum score estimator when predicting discrete outcomes.