Bradley-Terry Modeling with Multiple Game Outcomes with Applications to College Hockey

Abstract

The Bradley-Terry model has previously been used in both Bayesian and frequentist interpretations to evaluate the strengths of sports teams based on win-loss game results. It has also been extended to handle additional possible results such as ties. We implement a generalization which includes multiple possible outcomes such as wins or losses in regulation, overtime, or shootouts. A natural application is to ice hockey competitions such as international matches, European professional leagues, and NCAA hockey, all of which use a zero-sum point system which values overtime and shootout wins as 1/3 of a win, and overtime and shootout losses as 1/3 of a win. We incorporate this into the probability model, and evaluate the posterior distributions for the associated strength parameters using techniques such as Gaussian expansion about maximum a posteriori estimates, and Hamiltonian Monte Carlo.