Hybrid CNN-transformer architecture for personal credit risk prediction with comparative insights into model explainability

Abstract

Accurately predicting personal credit delinquency is challenging because credit data are high-dimensional, highly imbalanced, and behaviorally heterogeneous. This study proposes a hybrid deep learning framework to enhance the accuracy of personal credit risk prediction, particularly in high-dimensional and imbalanced financial datasets. Recognizing the limitations of traditional statistical models and shallow machine learning algorithms in capturing complex variable interactions and temporal patterns, we develop a novel model that combines a one-dimensional convolutional neural network with a Transformer architecture. This hybrid design enables both local feature extraction and global dependency modeling, which are essential for understanding borrower behavior across multiple credit indicators. Utilizing real-world credit transaction data from the Korea Credit Bureau, comprising 164,549 records, the proposed model outperformed baseline algorithms in recall and ROC-AUC metrics. To enhance interpretability, we employ explainable artificial intelligence techniques—SHAP and LIME—to analyze the contribution of key financial variables. We describe heterogeneous explanation patterns through clustering to characterize distinct borrower risk profiles. Given the scale of bureau-level credit records and the need for timely portfolio monitoring, the proposed framework is intended to be deployed with high-performance computing (HPC) resources to support credit risk scoring. This supercomputing capability helps meet operational latency requirements in financial institutions by enabling rapid model inference on high-dimensional credit data for early warning and credit decision support. Overall, the proposed framework improves delinquency detection while offering interpretable evidence that can support risk monitoring and credit decision processes.