In this paper, the kernel density estimator is applied to choose the explanatory variables for credit risk models. For doing it, given each considered explanatory variable, its ability on discriminating between healthy and bankrupt companies is measured using its one-dimensional kernel density estimate. All considered explanatory variables are then separated into groups according to the magnitude of their discriminant abilities. The prediction performance of both credit risk models, the modified discriminant analysis model (Welch, 1939) and the logistic model (Ohlson, 1980), is compared on each group of explanatory variables. Empirical studies demonstrate that the two credit risk models using explanatory variables of better discriminant ability have better prediction performance, in the sense of yielding smaller out-of-sample error rates.