基于Twin-SVR的公司违约风险预测研究

Research on Corporate Default Risk Prediction Based on Twin-SVR

本文以发行短期融资券的公司为研究对象,首先基于信用利差构建出公司违约风险变量,然后运用Twin-SVR模型对公司违约风险展开预测,最后对影响公司违约风险Twin-SVR模型预测性能的特征指标解释能力进行了探讨。实证结果表明:在Twin-SVR模型的构建过程中,发现RBF核函数展示出了比Polynomial、Linear和Sigmoid核函数更加卓越的预测性能;与传统的SVR、BPNN以及Logistic公司违约风险预测模型相比,Twin-SVR模型不仅在整体行业中具有最优的预测性能,而且在分行业中具有优异的泛化性能;对于影响Twin-SVR模型预测性能的前两位特征指标解释能力来说,在整体行业和有色金属行业中前两位最具解释能力的指标为信用评级和是否为国有企业,在煤炭和钢铁行业中为信用评级和净资产收益率。

This paper,taking companies that issue short-term financing bond as the objects of research,establishes a suitable method based on the credit spread to analyze the corporate default risk,then applies Twin-SVR( Twin Support Vector Regression) to predict the corporate default risk and finally explores the ability to interpret the characteristic indicators that affect the predictive performance of Twin-SVR default risk model. The empirical results are as follows: During the construction of the Twin-SVR model,we find that the RBF kernel function exhibits better prediction performance than the Polynomial,Linear,and Sigmoid kernel functions;compared with the traditional SVR,BPNN and Logistic corporate default risk prediction models,the Twin-SVR model not only has the best prediction performance and stability performance in the overall industry,but also has excellent generalization performance in the sub-industry;for the interpretation ability of the top two characteristic indicators that affect the prediction performance of Twin-SVR model,we find that the most explanatory two indicators in the overall industry and non-ferrous metal industry are credit rating and SOE/non-SOE background,and in the coal and steel industry,the top two indicators are credit rating and ROE.