基于最优信用特征组合的违约判别模型——以中国A股上市公司为例

违约判别是信用风险评估的一种方式，提高违约判别精度一直是学界和业界重点关注的问题.本文从最优信用特征组合而不是最优指标组合的角度建立违约判别模型，提高违约判别精度.本文的创新有三个方面：一是以信息值最大为目标建立优化模型，将指标数据划分成能最大区分违约状态的多个信用特征.二是采用弹性网回归对信用特征进行遴选，反推违约判别误差最小的最优信用特征组合.三是以组间离散度与组内离散度之比最大为目标，构建数学规划，反推一组权重，得到线性判别方程.本文基于2000-2017年共2169家中国A股上市公司的数据进行实证，研究表明经过特征划分的线性判别分析、K近邻、支持向量机等模型的精度整体高于没有经过特征划分的模型精度.

Default discriminant study based on optimal credit feature set: A case study of China A-share listed companies

Default discriminant is a method of credit risk assessment. Improving the accuracy of default discriminant has always been a major concern of academics and industry. In this study, we establish a default discriminant model not with optimal indicator set but with optimal credit feature set in order to improve the performance. There are three innovations in this study. First, we divide the indicators into multiple credit features which can distinguish the default state by maximizing the information values. Secondly, we select the optimal credit feature set by using elastic net regression with an objective to minimize the error. Thirdly, to maximize the ratio of the dispersion between groups and the dispersion within the group, we construct a mathematical programming and obtain the optimal weight of the linear discriminant analysis. Based on the data of 2169 China A-share listed companies from 2000 to 2017, this study shows that the performance of models, namely linear discriminant analysis, K-nearest neighbor and support vector machine, with credit feature division are better than those without credit feature division.