생성형 AI의 RAG를 활용한 개인 신용대출 부도예측에 대한 연구

As big data and AI become increasingly important in the financial industry, their application in credit risk management has emerged as a key issue. Default prediction is essential for enhancing financial institutions' asset soundness and loan screening efficiency, and traditional logistic regression models have evolved into machine learning and deep learning techniques to improve accuracy. Recently, financial institutions have been exploring the application of generative AI and Large Language Model (LLM) in loan screening, but concerns have been raised that LLMs may not fully capture the detailed information specific to the financial domain. As a result, Retrieval-Augmented Generation (RAG) technology, which enhances information retrieval and prediction accuracy, has gained attention. This study investigates RAG-based default prediction models and RAG-LLM hybrid models for personal credit loan default prediction. Experimental results showed that their predictive performance was lower than traditional logistic regression and XGBoost models, likely due to the inability to fully incorporate the complexity of credit-related variables. Moreover, the RAG-LLM hybrid model also exhibited lower predictive accuracy than existing methods, highlighting the limitations of LLMs in specialized financial domains. On the other hand, the RAG-based model demonstrated stable predictive performance in small-scale data environments, particularly for thin-filer borrowers, suggesting its potential as an alternative approach. These findings provide important insights for financial institutions, helping to reduce trial and error in the adoption of generative AI for credit risk assessment.