Temporal Dependency Modeling in Loan Default Prediction with Hybrid LSTM-GRU Architecture

Abstract

This paper addresses the challenge of sequence modeling in loan default prediction by proposing a deep learning classification model that integrates Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks. The model adopts a cascaded structure that combines LSTM's strength in capturing long-term dependencies with GRU's advantage in computational efficiency. This design improves the accuracy of identifying potential risk behavior in financial time series data. The Kaggle public loan dataset was used for model training. Categorical features were encoded using label encoding, and numerical features were standardized. Sensitivity experiments were conducted under different hyperparameter settings. The comparison of learning rates and optimizers confirmed the model's training stability and superior performance. Furthermore, a comprehensive evaluation was conducted using accuracy, precision, and recall, along with visual tools such as training loss curves, accuracy curves, and confusion matrices. The results demonstrate that the proposed method achieves significant advantages in the loan default classification task.