Testing predictive performance of binary choice models

Binary choice models occur frequently in economic modeling. A measure of the predictive performance of binary choice models that is often reported is the hit rate of a model. This paper develops a test for the outperformance of a predictor for binary outcomes over a naive prediction method, which predicts the outcome that is most often observed. This is done for a general class of prediction models, including the well known Probit and Logit models. In many cases the test is easy to compute. The test is then applied and compared to a general test of Pesaran and Timmermann (1992) for dependence between predictors and realizations.