MINIMIZING ERROR COSTS IN DISCRETE PREDICTION MODELS

Cutoff point determination is one of the major problems in the binary prediction studies. In most of the previous studies, the cutoff point was selected to minimize the total number of errors. Because of the different misclassification costs of default errors and non-default errors, the decision rules obtained in these studies might not be the most advantageous. This paper presents a more generally applicable procedure to determine the optimal cutoff point to minimize the total error cost by Bayes rules. Normal and beta distributions are considered in details. This procedure can be used in all discrete prediction models, such as discriminant analysis, logistic analysis and neural networks. The procedure is illustrated with the example of sovereign debt service capacity prediction.