基于混合生存分析的动态信用评分方法

Dynamic credit scoring method based on mixture survival analysis

当今国内外经济形势复杂多变,不确定因素增多,金融市场中信用风险的动态性显著增强,动态信用风险评价成为金融机构迫切需要解决的问题.为此,本文提出了一种基于混合生存分析的动态信用评分方法.该方法由三部分组成:首先,构建基于混合生存分析的动态信用评分模型,包括违约状态判别模型和违约时间估计模型,用于预测评价对象"是否违约"以及"何时违约";其次,利用面板数据构建多重生存状态向量,用于表征信用特征和生存时间的动态关联;最后,基于生成的多重生存状态向量,利用期望最大算法迭代估计模型参数.实验研究表明该方法的预测效果显著优于基于单分类、基于集成学习以及基于生存分析的信用评分方法.

Currently,the domestic and international economic situation has become more complex and changeable,with increasing uncertainties,and the dynamics of credit risk in the financial market has been significantly enhanced.Dynamic credit risk evaluation has become an urgent problem that financial institutions need to solve.To this end,this paper proposes a mixture survival analysis-based dynamic credit scoring method,which consists of three parts.First,constructing mixture survival analysis-based dynamic credit scoring model,including default status discrimination model and default time estimation method,to predict "whether default" and "when to default" for evaluation objects.Then,generating multiple survival status vectors using panel data to characterize the dynamic correlations between credit features and survival time.Finally,based on generated multiple survival status vectors,iteratively estimating model parameters using the expectation maximum algorithm.Experimental research shows that the predictive performance of the proposed method is significantly superior to the single classification-based,ensemble learning-based,and survival analysis-based credit scoring methods.