A Comparative Study of TabTransformer and Temporal Fusion Transformer for Loan Approval Prediction Using Static and Temporal Financial Features

Abstract: In the banking and financial services industry, loan approval is an important process that helps minimize credit risk and increase lending efficiency. Traditional machine-learning models can struggle to represent complex interactions among customer financial features and changes in financial activity over time. This paper presents a comparative study of two transformer-based deep-learning models - TabTransformer and Temporal Fusion Transformer (TFT) - for intelligent loan approval prediction. TabTransformer learns meaningful representations from structured customer data, whereas TFT is trained using twelve months of sequential financial data to model long-term financial patterns. For a fair comparison, both models are trained and tested using the same enhanced loan dataset, preprocessing pipeline and experimental setting. Performance is assessed using accuracy, precision, recall, F1-score, ROC-AUC and confusion- matrix analysis. The results show that TFT achieves 94.75% accuracy compared with 93.13% for TabTransformer, indicating that temporal financial information provides additional predictive value for automated loan application decision-making.

Keywords: Loan Approval Prediction, Tab Transformer, Temporal Fusion Transformer, Deep Learning, Credit Risk Assessment, Financial Data Analytics, Transformer Networks, Sequential Learning, Artificial Intelligence.