DEEP LEARNING-BASED CREDIT RISK MODELING: ADDRESSING DATA IMBALANCE AND INVARIANCE

Credit risk modeling plays a crucial role in financial decision-making, helping lenders assess the likelihood of default and optimize lending strategies. Traditional credit risk assessment models, including logistic regression and decision trees, often struggle with data imbalance and invariance issues, leading to biased risk predictions and reduced generalization. The rapid advancement of deep learning (DL) techniques has introduced more sophisticated models capable of learning complex credit risk patterns. However, most DL-based credit scoring models still suffer from class imbalance in default prediction and fail to maintain fairness and stability across different demographic and economic conditions. This study proposes a DL-based credit risk modeling framework designed to address data imbalance through advanced resampling techniques and generative modeling, while also incorporating adversarial learning to improve model invariance across diverse borrower segments. The proposed framework utilizes autoencoders, generative adversarial networks (GANs), and cost-sensitive learning techniques to enhance risk assessment accuracy while reducing bias. Additionally, domain adaptation techniques are introduced to ensure that the model remains robust across different financial environments. Experiments on real-world credit datasets demonstrate that the proposed framework significantly improves credit risk prediction accuracy, enhances model fairness, and reduces sensitivity to class imbalance compared to traditional credit scoring approaches. The findings highlight the importance of integrating data-centric augmentation techniques with fairness-aware deep learning to improve the reliability of credit risk modeling in modern financial applications.