主管:中国科学院
主办:中国优选法统筹法与经济数学研究会
   中国科学院科技战略咨询研究院

中国管理科学 ›› 2026, Vol. 34 ›› Issue (5): 57-71.doi: 10.16381/j.cnki.issn1003-207x.2024.1256cstr: 32146.14.j.cnki.issn1003-207x.2024.1256

• • 上一篇    下一篇

降维能够帮助深度神经网络更好地学习信用风险吗?——基于因子增强可解释学习模型的研究

王璞1, 李鲲鹏1(), 苏立2   

  1. 1.首都经济贸易大学国际经济管理学院,北京 100070
    2.中国人民大学应用经济学院,北京 100872
  • 收稿日期:2024-07-24 修回日期:2024-09-17 出版日期:2026-05-25 发布日期:2026-04-21
  • 通讯作者: 李鲲鹏 E-mail:kunpenglithu@126.com
  • 基金资助:
    国家自然科学基金项目(72225006);北京市属高等学校高水平科研创新团队建设支持计划(BPHR20220119);科技部重点研发项目(2023YFA1009200)

Can Dimensionality Reduction Enhance Deep Neural Networks in Learning Credit Risk? ——A Study Based on Factor-Augmented Explainable Learning Models

Pu Wang1, Kunpeng Li1(), Li Su2   

  1. 1.International School of Economics and Management,Capital University of Economics and Business,Beijing 100070,China
    2.School of Applied Economics,Renmin University of China,Beijing 100872,China
  • Received:2024-07-24 Revised:2024-09-17 Online:2026-05-25 Published:2026-04-21
  • Contact: Kunpeng Li E-mail:kunpenglithu@126.com

摘要:

本文将大数据降维技术和深度学习的逼近手段相结合,提出了一种新的可解释深度学习方法-因子增强的加性神经网络模型,并依据该方法建立信用风险度量模型,结合我国债券市场公司数据进行实证分析,同时与目前流行的可解释机器学习方法进行比较。研究发现:(1)本文提出的模型在各项评估指标上优于外部评级、线性回归,以及目前流行的可解释机器学习方法,具备更好的泛化能力;(2)价格指标可以动态反映企业风险溢价,在信用风险度量方面最为重要且时效性较高,提高了本文模型动态捕捉风险的能力;(3)本文提出的方法也佐证了在评级过程和方法中引入降维方法、深度学习方法提高评级质量必要性。本文提出的方法是对可解释深度学习、企业信用风险度量等相关研究的有益补充,为信用风险防范和信用评级提供了一种有现实价值的信用风险度量方法。

关键词: 信用风险, 可解释深度学习, FAAN

Abstract:

The challenge of accurately assessing credit risk has grown with the complexity of financial markets and the availability of big data. Traditional models often fall short of capturing dynamic risk factors. The problem is addressed by proposing the Factor-Enhanced Additive Neural Network (FAAN) model, combining big data dimensionality reduction techniques with deep learning approximation to improve accuracy and interpretability in credit risk measurement.The main research question is designing an interpretable deep learning model that outperforms existing methods in measuring corporate credit risk. The FAAN model tackles this by integrating factor-based dimensionality reduction to remove noise and deep learning to capture complex non-linear relationships. It is more effective than traditional models like linear regression and popular interpretable machine learning methods (e.g., GAM, EBM).Empirical analysis using real-world data shows that the FAAN model: (1) consistently outperforms external credit ratings and other models across various metrics, demonstrating superior generalization; (2) highlights price indicators as crucial dynamic factors that enhance credit risk detection; and (3) confirms that combining dimensionality reduction with deep learning significantly improves credit rating quality.It contributes to research on interpretable deep learning and corporate credit risk, offering a powerful tool for dynamic risk assessment and credit rating practices, advancing both academic understanding and practical applications.

Key words: credit risk, interpretable deep learning, FAAN

中图分类号: