DEEP LEARNING-BASED CREDIT RISK PREDICTION USING SELFATTENTION DRIVEN HYBRID LSTM MODEL

Abstract

In order to mitigate the shortcomings of deep learning and machine learning, this extended paper offers a Self-Attention based Hybrid LSTM credit score prediction model. In order to have the most contextual information and minimize the amount of irrelevant patterns, the network emphasizes self-attention that concentrates on purchasing the most significant temporal and behavioral variables on high-dimensional financial data. The enhancement of data quality and stability is achieved by such techniques as temporal modeling, Min-Max normalization, to balance the data, SMOTE to select the features and PCA to reduce the dimensionality. The technique is feasible because a financial company can input data file containing customer details and get real-time information on risk assessment to forecast a credit score through an internet application on Flask. Based on the experimental findings, the Self-Attention Hybrid LSTM is better than the base models in their accuracy, data generalisation, and financial power. This is the improved process of assessing the credit risk which has become intelligent, scalable and efficient.