基于类别不平衡的企业信用风险违约测度探索——以制造业上市公司为例

针对所获取的类别不平衡的深沪A股制造业上市公司财务数据,为了预测制造业上市公司信用违约情况,提出基于欠采样改进的Lasso-Logistic模型;首先通过计算WOE和IV值,剔除风险识别能力和稳定性较差的变量,接着从"数据"层面对现有的Lasso-Logistic模型进行批量欠采样处理,最后结合"算法"层面对Lasso-Logistic子模型的预测概率进行简单平均集成来研究模型的改进效果;结果表明,从模型整体效果的测度指标AUC值和区分度指标KS值来看,基于欠采样改进的带有变量筛选能力的Batch-US-LLR模型能有效提升企业信用风险违约测度的效果,对完善企业风险预警机制,提升违约风险识别能力具有可行性和有效性。

Exploration on the Measurement of Enterprise Credit Risk Default Based on Class Imbalance:Take Listed Manufacturing Companies as an Example

In view of financial data obtained from Shanghai-Shenzhen A share listed companies of manufacturing industry with class imbalance, in order to predict the credit default of the listed companies of manufacturing industry, Lasso-Logistic model based on under-sampling improvement is proposed. Firstly, by calculating WOE and IV values, the variables with poor risk identification ability and poor stability are eliminated, then, from "data" level, the existing Lasso-Logistic model is processed for batch under-sampling, finally, the improved effect of the model is studied by simply mean integration of the prediction probability of Lasso-Logistic sub model based on "algorithm" level. Results show that from the perspective of the model holistic effect measurement indicator AUC value and distinguishing degree indicator KS value, Batch-US-LLR model with variables screening ability based on under-sampling improvement can effectively improve the effect of enterprise credit risk default measurement and have the feasibility and validity for perfecting early-warning mechanism for enterprise risk and promoting default risk identification ability.