Using Explainable Machine Learning to Predict Loan Risk in Consumer Finance

The high default rate of consumer finance loans poses a significant challenge in predicting credit risk. This study presents a framework for predicting loan risk by leveraging the strengths of machine learning models. Compared to logistic regression models, LightGBM and CatBoost models demonstrate superior predictive performance. Furthermore, model fusion techniques contribute to improved performance over individual approaches. To improve the interpretability of the predictions, we use Shapley Additive Explanations (SHAP), which clarify the impact of each feature on the model's outputs. Key risk factors identified include gender, days past due, income type, late payment percentage, contract start date, total loan payments, and outstanding amount. By combining the strong predictive capabilities of LightGBM and CatBoost with the transparency provided by SHAP, this study aims to build trust in machine learning applications within the consumer finance sector. Our findings offer valuable recommendations for improving credit risk management practices, ultimately contributing to more reliable and transparent decision-making processes in consumer finance loan assessments.