基于泰勒展开的BPNN-TaylorLoss非均衡样本违约预测模型  

Default prediction model of BPNN-TaylorLoss based on Taylor expansion

作　　者：杨莲 刘文秀 张志鹏[3] 石宝峰[4,5] YANG Lian;LIU Wenxiu;ZHANG Zhipeng;SHI Baofeng(College of Economics and Management,Shandong Agricultural University,Taian 271000,China;Business School,Nanjing Normal University,Nanjing 210000,China;Antai College of Economics and Management,Shanghai Jiao Tong University,Shanghai 200000,China;College of Economics and Management,Northwest A&F University,Yangling 712100,China;Research Center on Credit and Big Data Analytics,Northwest A&F University,Yangling 712100,China)

机构地区：[1]山东农业大学经济管理学院,泰安271000 [2]南京师范大学商学院,南京210000 [3]上海交通大学安泰经济与管理学院,上海200000 [4]西北农林科技大学经济管理学院,杨凌712100 [5]西北农林科技大学信用大数据应用研究中心,杨凌712100

出　　处：《系统工程理论与实践》2024年第12期4045-4063,共19页Systems Engineering-Theory & Practice

基　　金：国家自然科学基金(72303139,71873103,72173096,72401189);国家社会科学基金重大项目(23&ZD175,21&ZD115);中国博士后科学基金面上项目(2023M732204);中和农信星空计划(K4030218167);西北农林科技大学仲英青年学者项目(2021-04)。

摘　　要：针对现有违约预测模型对非均衡样本适用性不强、对具有不同信贷特征数据集可扩展性较弱的现状,利用泰勒展开原理将交叉熵函数转化为多项式的线性组合,并在第1项多项式系数中添加扰动因子ε,构建可以根据不同非均衡信贷数据特点进行灵活调整的BPNNTaylorLoss违约预测模型;利用4个真实信贷数据、7种对比模型、5个模型评价准则验证模型性能.研究表明:所提模型有助于降低违约客户误判给金融机构带来的损失以及预防非违约样本误判导致的优质客户流失;所提模型在多数信贷数据集中表现出了较为稳健的违约预测性能,具有较好的模型可扩展性.本文的创新与特色:利用扰动因子ε对标准交叉熵函数泰勒展开式进行修正,构建BPNN-TaylorLoss非均衡样本违约预测模型,实现只需对与扰动因子ε相关的1个超参数进行微调,即可改变现有评价模型对非均衡信贷数据集适用性不强,以及对具有不同信贷特征数据集可扩展性较弱的现状.本研究为非均衡样本违约风险预测提供了新的研究视角.The existing credit evaluation models have weak applicability to imbalanced samples and weak scalability to datasets with different credit characteristics.In this study,the principle of Taylor expansion is used to transform the Cross Entropy function into a linear combination ofpolynomials, and the perturbation factor ε is added to the first polynomial coefficient. Based onthis, the BPNN-TaylorLoss default prediction model is constructed. We use 4 real credit data,7 comparative models, and 5 model evaluation criteria to verify the performance of the model.The results show that the proposed model helps to reduce the losses caused by misjudgmentof defaulting customers to financial institutions and prevent the loss of high-quality customerscaused by misjudgment of non-defaulting samples. The proposed model exhibits robust defaultprediction performance in most credit data sets, thus exhibiting good model scalability. TheInnovation of this study is we use the perturbation factor ε to modify the Taylor expansion ofthe Cross Entropy function, and constructs the BPNN-TaylorLoss default prediction model. Itcan change the current situation that the existing credit evaluation models are not suitable forunbalanced data sets and have weak scalability for data sets with different credit characteristicsby adjusting only one hyperparameter ε. This study provides a new research perspective forcredit risk assessment of imbalanced samples.

关 键 词：违约预测 非均衡 泰勒展开 交叉熵 

分 类 号：F830.56[经济管理—金融学] N945.16[自然科学总论—系统科学]