Enhanced Bank Lending Decisions: A Deep Learning Strategy with Feature Selection Based on Lightweight Parrot Bi-LSTM

Bank loan decisions are primarily based on creditworthiness assessments that estimate the likelihood of repayment. This concentration could lead to the exclusion of applicants with short credit histories or unusual sources of income, increasing financial inequality by reducing the amount of credit available to otherwise dependable consumers. In the paper, Enhanced Bank Lending Decisions (EBLD): A deep strategy with feature selection based on lightweight parrot Bi-LSTM (LPBiLSTM) is discussed. The input data are gathered from a standard database and the data sets are named as credit-risk, credit-risk analysis, and credit-risk classification. Also, a range-controlled minority oversampling technique is enabled to address the class imbalance and generate synthetic samples to enhance accuracy. A lightweight bidirectional LSTM optimized using the Parrot Optimization Algorithm (POA) provides efficient prediction of borrower behavior. A deep-spiking Kepler neural network, utilizing the Kepler Optimization Algorithm (KOA), improves real-time lending decision-making and risk assessment. The proposed method achieves the performance metrics across different data sets. For the credit-risk data set, it achieves an accuracy of 98.27%, precision of 97.60%, recall of 98.25%, [Formula: see text]1-score of 98.13%, and ROC AUC score of 97.55%. In the credit-risk analysis data set, it achieves an accuracy of 99.69%, precision of 99.14%, recall of 95.23%, [Formula: see text]1-score of 97.15%, and ROC AUC score of 98.60%. However, the credit-risk classification data set achieves an accuracy of 98.81%, precision of 99.43%, recall of 98.80%, [Formula: see text]1-score of 98.18%, and ROC AUC score of 99.96%.