Kolmogorov–Arnold networks-based GRU and LSTM for loan default early prediction

This study addresses a key challenge in time series anomaly detection: the prediction of loan defaults several months in advance. Using real-world monthly repayment data, the study aims to provide financial institutions with a reliable basis for proactive intervention. Existing studies often neglect early prediction capabilities and rely on out-of-sample rather than out-of-time testing, limiting the evaluation of model adaptability to evolving financial conditions and constraining the effective use of historical data. To bridge this gap, this study introduces Kolmogorov–Arnold Network (KAN)-enhanced Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) models, denoted as GRU-KAN and LSTM-KAN, along with a “blank interval” concept to simulate early prediction scenarios. Evaluations compare these models against baseline methods, including LSTM, GRU, LSTM-Attention, and LSTM-Transformer, across varied resampling strategies, feature window lengths, blank interval durations, sample sizes, and multi-year and cross-quarter test sets. Ablation studies isolate the individual contributions of each model component. The results of the experiments indicate that the proposed models consistently outperform the baselines, with random undersampling and feature windows of 12 to 18 months producing better performance. GRU-KAN achieves optimal results for blank intervals of 1 to 8 months, while LSTM-KAN excels for intervals of 9 to 12 months.