Abstract
Credit risk prediction is a vital task in financial services, ensuring that institutions can manage their lending risks effectively. This study investigates the effectiveness of deep learning (DL) models for credit risk prediction, with a focus on addressing the challenge of class imbalance and the black box nature of these models using the Synthetic Minority Over-sampling Technique-Edited Nearest Neighbor (SMOTE-ENN) resampling method and Shapley Additive Explanations (SHAP), respectively. The study compares the performance of various DL architectures, including Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), Gated Recurrent Units (GRU), and Graph Neural Networks (GNN), on two real-world datasets: the Australian and German credit datasets. The findings reveal that the GRU model, enhanced with SMOTE-ENN resampling, outperforms other models in terms of accuracy, sensitivity, and specificity. The superior performance of the GRU-SMOTE-ENN model demonstrates its potential as a robust deep learning technique for financial institutions to enhance credit risk assessment. Additionally, the study demonstrates how the integration of SHAP values significantly improves the interpretability of deep learning models, making them more transparent and trustworthy for stakeholders.