March Madness, Quantile Regression Bracketology, and the Hayek Hypothesis

A quantile regression variant of the classical paired comparison model of mean ratings is proposed. The model is estimated using data for the regular 2004-2005 U.S. college basketball season, and evaluated based on predictive performance for the 2005 National Collegiate Athletic Association (NCAA) basketball tournament. Rather than basing predictions entirely on conditional mean estimates produced by classical least-squares paired comparison methods, the proposed methods produce predictive densities that can be used to evaluate point spread and over/under gambling opportunities. Mildly favorable betting opportunities are revealed. More generally, the proposed methods offer a flexible approach to conditional density forecasting for a broad class of applications. An electronic supplement containing predictive densities for point totals and team offensive and defensive ratings is available from the JBES website.