关键词: 债券违约预测, 交叉熵损失, 不平衡样本, 概念漂移, SHAP解释

Abstract: Bond default prediction faces multitude challenges, including severely imbalanced samples, concept drift, and the difficulty of identifying hard-to-classify samples. Existing models, both basic ones and improved ones to address individual issues, often fall short of meeting these complex demands. To address this, this paper proposes a concise composite loss function based on cross-entropy loss, termed as TW-focal loss, which adjusts the loss weights of different samples by incorporating tailored improvement factors, enabling the model to effectively learn from default samples, new samples and hard-to-classify samples. Using publicly issued credit bond data in China from 2014 to 2022, and adopting XGBoost as the experimental model, the empirical results show that TW-focal loss effectively controls the Type I error rate while reducing the Type II error rate. Compared to the standard cross-entropy loss, the performance evaluation index Gmean has increased by 46.4%, and by 12.9% compared to the weighted cross-entropy loss that focused on imbalance. Additionally, using the SHAP interpretation method, this paper analyzes the distribution of feature importance and the partial dependency curves under different loss functions. The results reveal that the model can control the recognition of default samples by altering the impact degree and range of features. This paper provides an effective attempt to complete the design of bond default prediction models and to explore the discriminative logic of the models.

Key words: bond default prediction, cross-entropy loss, imbalanced samples, concept drift, SHAP interpretation