基于最优指标组合的企业信用风险预测  

Credit Risk Prediction of Company Based on Optimal Feature Set

作　　者：周颖[1] 苏小婷 ZHOUYing;SU Xiaoting(School of Economics and Management,Dalian University of Technology,Dalian 116024,Liaoning,China)

机构地区：[1]大连理工大学经济管理学院,辽宁大连116024

出　　处：《系统管理学报》2021年第5期817-838,共22页Journal of Systems & Management

基　　金：国家自然科学基金重点项目(71731003);国家自然科学基金面上项目(72071026,72173096,71873103,71971051,71971034);国家自然科学基金青年科学基金资助项目(71901055,71903019);国家自然科学基金地区科学基金资助项目(72161033);国家社会科学基金重大项目(18ZDA095)。

摘　　要：信用风险预测是指构建企业历史数据与违约状态之间的对应关系,根据现在的数据对企业在未来是否会发生违约做出预判。将近邻成分分析引入信用风险领域进行指标组合遴选,以违约预测精度AUC最大反推最优的指标组合。利用随机欠采样方法,以违约预测精度G-mean最大为标准反推违约客户与非违约客户的最佳比例,确定最优训练样本。采用t-m(m=1,2,3,4,5)年的指标数据x_(t-m)和t年的企业违约状态y_(t),利用最优指标组合和最优训练样本建立了基于线性支持向量机的信用风险预测模型,达到了运用t年的指标数据x_(t)预测第t+m年企业违约状态y_(t+m)的效果。实证结果表明,本研究的违约预测精度高于非线性SVM、LR、DT、KNN和LDA等典型的大数据预测模型。研究发现:每股收益EPS-扣除/稀释、货币供应量M_(0)(亿元)和货币供应量M_(1)(亿元)3个指标对企业未来1-3年的短期违约状态具有关键影响;当日总市值/负债总计、每股EBITDA和固定资产周转率3个指标对企业未来4-5年的长期违约状态具有关键影响;经营活动产生的现金流量净额/经营活动净收益和审计意见类型2个指标,不论对于企业未来1-3年的短期、还是未来4-5年的长期违约状态,均有关键影响。Credit risk prediction is the process that predicts the default status by using feature data of a tested company based on the default prediction model constructed on the training dataset.The contribution of this paper is three-folds.First,it introduces the neighborhood components analysis mothed to select feature set for the credit risk prediction for the first time.It infers reversely the optimal feature set by maximizing the default prediction accuracy AUC.Next,it determines the best training sample with the optimal ratio of the number of default customers to that of non-default ones by maximizing the default prediction comprehensive accuracy G-mean.Finally,it constructs the L-SVM model by using the feature data x_(t-m)(m=1,2,3,4,5)and default status y_(t),based on the optimal feature set and training sample determined,to achieve the prediction of default status y_(t+m) of(t+m)-th year using the data x_(t) of t-th year.The empirical evidence shows that the accuracy of the proposed model is higher than that of typical models including nonlinear SVM,LR,DT,KNN,and LDA.Besides,the three features including"EPS-deduct/dilution","money supply M_(0)(0.1 billion RMB)",and"money supply M_(1)(0.1 billion RMB)"have an important impact on the short-term prediction within 3 years.The three features including"total market value/total liabilities","EBITDA per share",and"fixed asset turnover"have an important impact on the long-term prediction of 4 to 5 years.The two features of"net cash flow from operating activities/net income from operating activities"and"audit opinion type"have an important impact either on the default prediction within 5 years.

关 键 词：违约预测 最优指标组合 样本比例 近邻成分分析 线性支持向量机 

分 类 号：F832.4[经济管理—金融学]