Predicting Default Probability of Bank’s Corporate Clients in the Czech Republic. Comparison of Generalized Additive Models and Support Vector Machine Approaches

This article evaluates predictive properties of alternative classification methods using real (anonymized) 2011–2016 data on the probability of default of bank’s corporate clients in the Czech Republic. Specifically, logistic regression, generalized additive model and support vector machines are used and their efficiency is evaluated with respect to a binary dependent variable representing corporate default status. Confusion matrices, ROC curves and area under the curve metrics are used to compare model predictions. Overall, support vector machines exhibit the best classification characteristics. However, some prediction properties (overall accuracy) are inferior to a generalized additive model.