Digital Financial Risk Early Warning Using Particle Swarm Optimization and Bidirectional Long Short-Term Memory

In modern times, the risk of loan defaults has increased with the rise in digital lending, resulting in the need for an accurate and timely financial risk early warning system. However, existing approaches, such as the Light Gradient Booting Machine (LightGBM)-based model, performed well with tabular financial data but lacked modeling sequential behavioral patterns and depended heavily on manual hyperparameter tuning, which affected its generalization. To overcome these challenges, a Particle Swarm Optimization-based Bidirectional Long Short-Term Memory (PSO-BiLSTM) is proposed for early prediction of digital financial risk. This research begins with the collection of detailed borrower demographics, repayment records, credit histories, and scores from Home Credit Default Risk. Next, preprocessing is performed using K Nearest Neighbor (KNN) imputation, label encoding, one-hot encoding, and min max normalization. Next, the preprocessed data are fed to Bidirectional Long Short-Term Memory (BiLSTM) to capture temporal behavior of borrower from both past and future scenarios, while Particle Swarm Optimization (PSO) is employed to automate hyper parameter tuning to maximize predictive performance. Therefore, the proposed PSO-BiLSTM achieved the highest accuracy (0.945), precision (0.948), strong recall (0.941) and high F1-score (0.944) with enhanced performance.