IMPROVED CREDIT RISK ASSESSMENT WITH A SELFATTENTION-INTEGRATED HYBRID LSTM FRAMEWORK

Abstract

This extended research introduces a Self-Attention–enhanced Hybrid LSTM model for credit scoring prediction, designed to overcome the limitations of traditional machine learning and basic deep learning models. The integration of a self-attention mechanism allows the network to selectively focus on the most influential temporal and behavioral features within high-dimensional financial datasets, improving contextual understanding and reducing the impact of irrelevant patterns. In addition to temporal modeling, the framework incorporates advanced preprocessing techniques such as Min-Max normalization, SMOTE for imbalanced data handling, RFE for feature selection, and PCA for dimensionality reduction, ensuring higher data quality and stability. To enable practical implementation, the proposed model is deployed through a Flask-based web application that supports real-time credit score prediction by allowing financial institutions to upload user data files and receive instant risk assessments. Experimental analysis demonstrates that the SelfAttention Hybrid LSTM outperforms baseline models, achieving higher accuracy, better generalization across heterogeneous datasets, and improved robustness in dynamic financial environments. This extended framework offers a scalable, efficient, and intelligent solution for modern credit risk assessment.