Financial Prediction Using DeepFM: Loan Repayment with Attention and Hybrid Loss

This paper presents a novel DeepFM-based transfer learning model designed to predict loan repayment behavior for underbanked populations, utilizing alternative data sources such as telecommunication and transaction information. The model incorporates an attention mechanism to enhance feature learning and employs a hybrid loss function combining Focal Loss and Cross-Entropy Loss to address data imbalance. Experimental results show that the proposed model outperforms traditional methods, including Logistic Regression, Random Forest, and LightGBM, achieving higher AUC-ROC, Precision, Recall, and F1-Score. From an audit and compliance perspective, this model improves the accuracy of loan portfolio risk assessment, ensuring it aligns with regulatory requirements and reducing the risk of non-compliance. This approach helps in encouraging financial inclusion as well as managing important compliance and audit goals.