Explainable Credit Default Prediction Using a Hybrid LSTM–XGBoost Model with SHAP Interpretability

The accurate prediction of credit default risk remains a significant challenge for financial institutions operating within increasingly complex data environments. This study pro-poses a hybrid Long Short-Term Memory (LSTM) and eXtreme Gradient Boosting (XGBoost) model that integrates deep learning and ensemble machine learning techniques to enhance predictive performance while preserving interpretability. The LSTM component effectively captures temporal patterns in borrower behavior, and its output is utilized as a meta-feature within the XGBoost framework. The model is evaluated using a bench-mark credit dataset and is compared with conventional machine learning approaches. The results indicate that the proposed hybrid model outperforms standalone models across key evaluation metrics, achieving high accuracy, F1-score, and ROC–AUC. To enhance transparency, Shapley Additive Explanations (SHAP) are employed to analyse feature contributions and directional effects. The findings reveal that repayment behavior, particularly recent delinquency, serves as the most influential predictor of default risk, followed by indicators of financial capacity. The feature derived from the LSTM demonstrates the strongest overall impact, thereby confirming the significance of temporal dependencies in credit risk prediction. This study illustrates that the integration of deep learning with ensemble techniques establishes a robust and interpretable framework for credit risk assessment, thereby providing practical value for enhancing financial decision-making and risk management.