Modelling sovereign credit ratings: The accuracy of models in a heterogeneous sample

The accuracy of sovereign credit ratings renewed interest toward sovereign credit ratings in the aftermath of the 2008 financial crisis. The controversy over the accuracies encouraged internal credit scoring systems to reduce reliance on sovereign credit ratings. By employing classification and regression trees (CART), multilayer perceptron (MLP), support vector machines (SVM), Bayes Net, and Naïve Bayes; we explore the prediction performance of several artificial intelligence (AI) techniques in predicting sovereign credit ratings in a heterogeneous sample. The results suggest that AI classifiers outperform the conventional statistical technique in terms of accurate prediction. According to within one notch and two notches accurate prediction measure, the prediction performances of the AI classifiers exceed 90% accuracy whereas the performance of the conventional statistical method is around 70%. The results further reveal that the prediction performance of the models declines around the threshold rating that is located between investment grade and speculative grade which is not necessarily the result of inadequacy of the models. Rather, this is potentially due to CRAs' cautious behaviour toward those countries around threshold rating which can be interpreted as the certification price of upgrading to investment grade.