Sovereign credit rating classification: evaluating the accuracy and driving factors using machine learning techniques

Purpose This paper explores the effectiveness of machine learning algorithms in predicting sovereign credit ratings, benchmarking their performance against traditional linear and panel regression models. Design/methodology/approach The analysis, which incorporates macroeconomic, political, and institutional variables as predictors, utilizes data from 41 countries between 2000 and 2018. Supervised learning techniques, including Random Forest, Neural Networks, XGBoost, and K-Nearest Neighbors (KNN), are employed in conjunction with feature selection methods such as Lasso and Ridge. Findings The results demonstrate that machine learning models, particularly Random Forest, deliver superior predictive accuracy, outperforming even fixed-effects panel models. Random Forest accurately classified slightly over 40% of the ratings, compared to just over 30% for the second-best model. Originality/value These findings underscore machine learning approaches’ flexibility and predictive strength, which operate with fewer assumptions and effectively capture complex interactions and nonlinear relationships among variables. Additionally, the study reaffirms the central role of institutional quality and political stability in determining sovereign credit ratings, contributing to the expanding use of computational tools in economics, particularly for classification and forecasting tasks.