Article: Corporate loan risk prediction model based on improved LightGBM Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.78 pp.52 - 78 Abstract: Corporate loan risk prediction is challenged by high-dimensional features, nonlinear relationships and severe class imbalance. To improve prediction accuracy and robustness, an improved LightGBM-based corporate loan risk prediction model is proposed. The framework integrates D-B-SMOTE for minority sample balancing, a feature importance fusion selection mechanism for adaptive feature screening, and a hybrid loss function combining focal loss and Huber loss to enhance sensitivity to default samples and reduce noise interference. In addition, an SSA-GWO hybrid optimisation strategy is introduced for hyperparameter tuning, and a distributed parallel training architecture is constructed to improve scalability. Experimental results show that the proposed model achieves an accuracy of 0.893, recall of 0.822, F1-score of 0.837 and AUC of 0.886, demonstrating strong robustness and generalisation ability in enterprise loan risk prediction. Inderscience Publishers - linking academia, business and industry through research

Title: Corporate loan risk prediction model based on improved LightGBM

Authors: Xueyi Li

Addresses: School of Management, Northwest Normal University, Lanzhou, 730070, China

Abstract: Corporate loan risk prediction is challenged by high-dimensional features, nonlinear relationships and severe class imbalance. To improve prediction accuracy and robustness, an improved LightGBM-based corporate loan risk prediction model is proposed. The framework integrates D-B-SMOTE for minority sample balancing, a feature importance fusion selection mechanism for adaptive feature screening, and a hybrid loss function combining focal loss and Huber loss to enhance sensitivity to default samples and reduce noise interference. In addition, an SSA-GWO hybrid optimisation strategy is introduced for hyperparameter tuning, and a distributed parallel training architecture is constructed to improve scalability. Experimental results show that the proposed model achieves an accuracy of 0.893, recall of 0.822, F1-score of 0.837 and AUC of 0.886, demonstrating strong robustness and generalisation ability in enterprise loan risk prediction.

Keywords: corporate loans; feature fusion screening; LightGBM; risk prediction; sparrow search algorithm; SSA.

DOI: 10.1504/IJICT.2026.155026

International Journal of Information and Communication Technology, 2026 Vol.27 No.78, pp.52 - 78

Received: 16 Jan 2026
Accepted: 02 Apr 2026
Published online: 22 Jul 2026 *

Title: Corporate loan risk prediction model based on improved LightGBM

Keep up-to-date