Abstract/Description
Sovereign credit rating provided by credit rating agencies like S&P, Moody’s and Fitch has always been of great importance, but these agencies have been facing criticism on their rating methodology which create subjectivity and potential biases. This harms the image of developing economies in international markets. This research is responsible to study subjectivity, its extent and evaluate the ML methodologies for better assessment of credit ratings through objective and transparent framework. The balanced panel data consisting of ten diverse countries from 2000 to 2023 is used. The study uses factors responsible for explaining credit worthiness of a country. These factors are divided into 5 WDI (World Development Indicators) and 5 WGI (World Governance Indicators) as independent variables. The dependent variable is considered as historical credit ratings of country. Nine distinct ML algorithms were tested starting from Linear Regression (baseline), KNN, SVR, Decision Tree, Random Forest, ANN, CatBoost, XGBoost and Light GBM. Results of these algorithms were evaluated by performance metrics such as R2, Adjusted R2, RAE, RMSE, MSE, Precision, Recall and F1 score. Results of these algorithms showed that nonlinear algorithm models outperformed the traditional models. Among all models tested, Light GBM emerged as the most robust model achieving highest adjusted R2 (0.7675) and lowest prediction errors such as (RMSE 0.4263), that shows model ability to capture largest proportion of variance explained by predictors with less predictive risk. Feature importance analysis provides insights with the best paramount indicators. Those indicators are labeled as Feature “CG” (Core economic strength) and I (Institutional Quality). This study explores the potential of ML algorithms to create objective rating framework that mitigate subjectivity in rating and to assess sovereign risk in the global financial landscape.
Keywords
Sovereign Credit Ratings, Machine Learning, Subjectivity, Credit Risk Assessment, Economic Indicators, Institutional Quality, LightGBM, Transparency, Developing Economies
Track
Finance
Session Number/Theme
Finance - Session II
Session Chair
Dr. Noureen Ayaz
Start Date/Time
14-6-2025 10:55 AM
End Date/Time
14-6-2025 12:35 PM
Recommended Citation
Siddiqui, A., Jahanzeb, H., Sohail, K., & Imran, M. (2025). Analysis of subjectivity on sovereign credit ratings through machine learning. IBA SBS 4th International Conference 2025. Retrieved from https://ir.iba.edu.pk/sbsic/2025/program/78
Included in
Business Analytics Commons, Finance and Financial Management Commons, Technology and Innovation Commons
Analysis of subjectivity on sovereign credit ratings through machine learning
Sovereign credit rating provided by credit rating agencies like S&P, Moody’s and Fitch has always been of great importance, but these agencies have been facing criticism on their rating methodology which create subjectivity and potential biases. This harms the image of developing economies in international markets. This research is responsible to study subjectivity, its extent and evaluate the ML methodologies for better assessment of credit ratings through objective and transparent framework. The balanced panel data consisting of ten diverse countries from 2000 to 2023 is used. The study uses factors responsible for explaining credit worthiness of a country. These factors are divided into 5 WDI (World Development Indicators) and 5 WGI (World Governance Indicators) as independent variables. The dependent variable is considered as historical credit ratings of country. Nine distinct ML algorithms were tested starting from Linear Regression (baseline), KNN, SVR, Decision Tree, Random Forest, ANN, CatBoost, XGBoost and Light GBM. Results of these algorithms were evaluated by performance metrics such as R2, Adjusted R2, RAE, RMSE, MSE, Precision, Recall and F1 score. Results of these algorithms showed that nonlinear algorithm models outperformed the traditional models. Among all models tested, Light GBM emerged as the most robust model achieving highest adjusted R2 (0.7675) and lowest prediction errors such as (RMSE 0.4263), that shows model ability to capture largest proportion of variance explained by predictors with less predictive risk. Feature importance analysis provides insights with the best paramount indicators. Those indicators are labeled as Feature “CG” (Core economic strength) and I (Institutional Quality). This study explores the potential of ML algorithms to create objective rating framework that mitigate subjectivity in rating and to assess sovereign risk in the global financial landscape.
