A novel hybrid credit scoring model: integrating RNN and XGBoost for improved creditworthiness assessment

Abstract

Credit scoring is essential for financial management, enabling lenders to assess creditworthiness. Traditional methods often struggle with complex data, leading to suboptimal results. We propose a hybrid credit scoring model combining Recurrent Neural Networks (RNNs) and XGBoost ensemble techniques, enhanced by Randomised Averaging. RNNs process sequential data, capturing long-term dependencies in credit histories, while XGBoost handles structured data patterns. Transfer learning from the Diane dataset adds diverse credit market insights. Randomised Averaging improves predictive performance and reduces overfitting. Experiments on datasets from Australia, Germany, Japan, and Diane show the model's superior accuracy, precision, recall, F1 score, and Brier score, outperforming existing methods and traditional credit scoring. This innovative approach effectively addresses the challenges of low accuracy and high training time in creditworthiness assessment, highlighting its robustness and practical applicability in various financial contexts.