Enhancing Credit Rating Prediction with Machine Learning and Explainable AI: Insights into Feature-Level Drivers of Creditworthiness

Abstract

The purpose of the study is to examine how machine learning models can improve the prediction of a company’s credit rating with the help of financial ratios. By comparing traditional statistical methods with more advanced machine learning algorithms such as Random Forest, Gradient Boosting, and Artificial Neural Network, the prediction accuracy is analyzed on different levels of aggregation, from detailed classifications (16 classes) to broader groupings (two classes). On the 16-class scale Random Forest achives 47% accuracy and an F1-score of 0.47, more than double the 21% accuracy achieved by logistic regression. When the ratings are grouped into six classes, accuracy rises to 71% for Random Forest and 68% for Gradient Boosting, versus 51% for the logistic baseline. In the practically important binary split between investment-grade and speculative-grade issuers, Gradient Boosting reaches 89% accuracy and Random Forest 88%, while logistic regression achieves 80%. To improve the transparency and usability of the models in practice, methods in explainable artificial intelligence (XAI) are also applied, specifically SHAP and LIME. These tools enable deeper insights into what financial factors influences the decision of the models. The results shows that financial ratios linked to debt and profitability, such as Total Debt/EBITDA, Dividend Yield, and Interest Coverage Ratios, have the greatest impact on credit rating predictions. The study contributes insights both methodological and practical, since it demonstrates how advanced models can be efficiently used in credit risk management, while highlighting the importance of explainability of models to increase reliability towards stakeholder and regulatory authorities.