This article investigates the predictive performance of several machine learning models for sovereign bond credit ratings. Given that previous methods suffer from certain deficiencies, this work compares ensemble and non-parametric learners, including XGBoost, AdaBoost, Random Forest and K-nearest neighbours, with benchmark classifiers previously used in the sovereign-rating literature through a repeated grouped 10-fold cross-validation framework. It deploys two different voting schemes to formulate the final ensemble predictions of the trained models. The analysis employs sovereign credit ratings and certain determinants by making use of data from the Big Three international credit rating agencies: Standard & Poor’s, Moody’s, and Fitch Ratings and for 157 different countries, both developed and developing ones, spanning the period 1962–2024. The results show that macroeconomic and political indicators contain substantial predictive information for sovereign credit-rating classification. They also indicate that more flexible classifiers, especially XGBoost and Random Forest, perform strongly under the grouped validation framework, although the performance gap relative to the strongest benchmark models is not uniform across all evaluation settings.
Apergis et al. (Mon,) studied this question.