Predicting the outcomes of National Football League games

Rankings have predictive value for determining the outcomes of basketball games and tennis matches. Rankings, based on power scores, are also available for NFL teams. This paper evaluates power scores as predictors of the outcomes of NFL games for 1994–2000 seasons. The evaluation involves a comparison of forecasts generated from probit regressions based on power scores published in The New York Times with those of a naive model, the betting market, and the opinions of the sports editor of The New York Times. We conclude that the betting market is the best predictor followed by the probit predictions based on power scores. We analyze the editor's predictions and find that his predictions were comparable to a bootstrapping model of his forecasts but were inferior to those based on power scores and even worse than naive forecasts.