Evaluating the Efficiency of Explainable Artificial Intelligence Methods in Determining the Importance of Variables in Predictive Models

The prediction of non-performing loan recovery is one of the main challenges in the banking system. Delays in the timely repayment of loans increase credit risk for banks and undermine their financial health. This study aims to design an accurate, interpretable, and AI-based model to assess the importance of variables influencing the prediction of receivables collection over a 30 to 90-day period. The research methodology is analytical-applied in nature. Real credit and banking data from 750,000 individual customers were utilized, and advanced machine learning algorithms (such as Random Forest and XGBoost), along with explainable artificial intelligence (XAI) methods such as SHAP and LIME, were employed. The results indicated that the algorithms were able to predict delinquent contracts with high accuracy, and SHAP successfully identified variables such as the number of negative months and the average overdue debt in the last three months as the most influential features. The use of explainable artificial intelligence not only preserves the accuracy of predictive models but also enables banking analysts to make decisions based on transparent data interpretation—an element directly contributing to the enhancement of risk management strategies.