非平衡样本下基于生成对抗网络过抽样技术的公司债券违约风险预测研究  

Imbalanced corporate bond default modeling using generative adversarial networks oversampling techniques

作　　者：姚潇 李可 余乐安[4] YAO Xiao;LI Ke;YU Le'an(Business School,Central University of Finance and Economics,Beijing 100081,China;Center of Statistical Research,School of Statistics,Southwestern University of Finance and Economics,Chengdu 611130,China;Collaborative Innovation Center of Financial Security,Southwestern University of Finance and Economics,Chengdu 611130,China;Business School,Sichuan University,Chengdu 610064,China)

机构地区：[1]中央财经大学商学院,北京100081 [2]西南财经大学统计学院统计研究中心,成都611130 [3]西南财经大学金融安全协同创新中心,成都611130 [4]四川大学商学院,成都610064

出　　处：《系统工程理论与实践》2022年第10期2617-2634,共18页Systems Engineering-Theory & Practice

基　　金：国家自然科学基金(71901230);西南财经大学数据科学与商业智能联合实验室资助。

摘　　要：本文基于我国在公开市场发行过信用类债券的违约数据,利用了基于Wasserstein距离的生成对抗网络模型和SMOTE抽样技术对违约样本进行过抽样以提高非平衡样本下违约风险模型的预测效果.为检验过抽样技术对分类模型的改进效果,实证分析对不同的重抽样样本类别比例下分类模型的预测结果进行比较.研究结果表明过抽样技术能够显著地分类模型的预测精度,而且预测效果随着样本类别比例达到平衡而不断提高.和经典的SMOTE抽样技术相比,基于Wasserstein距离的生成对抗网络过抽样技术不仅可以提高分类模型的AUC指标,同时还能显著地改进F1得分.研究结果表明通过生成对抗网络对少数类样本进行过抽样能够显著地提升机器学习算法对债券违约风险的预测效果,为研究非平衡样本下的债券违约风险预测提供一种新的解决思路.Based on the data of corporate bond issuers in Chinese market,this study applies oversampling techniques including Wasserstein generative adversarial networks(WGAN)and SMOTE to the imbalanced sample to improve the performance of bond default prediction.To explore the effect of oversampling techniques on classification models,the predictive outputs with difference imbalanced ratios are reported in the experimental results.It finds that the classification performance is significantly improved with the application of oversampling techniques,and the improvement is further enhanced when the sample distribution becomes more balanced.Compared to the classical SMOTE technique,both AUC and F1 score can be improved by WGAN.Overall,the experimental results demonstrate that the predictive performance of bond default models can be effectively boosted by generating artificial minority samples based on WGAN combined with the application of machine learning algorithms,which provides new insights into the bond default risk prediction of imbalanced samples.

关 键 词：债券违约风险 生成对抗网络 非平衡样本分类 过抽样技术 

分 类 号：F830.5[经济管理—金融学]