个人信用风险评分的贝叶斯有偏连接模型研究

Research on Bayesian Skewed Link Models in Credit Scoring

经典Logistic回归模型与Probit回归模型的连接函数都是固定的对称连接函数,数据的不平衡性使这两个对称连接模型的参数估计偏差和均方误差显著上升,预测效果也会下降。在潜变量模型的基础上,将非对称连接函数的思想引入到信用风险评分中,采用贝叶斯估计和Gibbs抽样对有偏连接模型中的参数进行估计。实证结果表明:两类有偏连接模型对对称连接模型的改造是成功的,并兼备其系数可解释的优点。

The link function of the Logistic and Probit regression model are fixed symmetric link functions. The characteristic of imbalance of data leads to a substantial increase in the bias and mean squared error of the parameter estimates as well as the predicted probabilities. Based on latent variable models, asymmetrical link functions are used to credit rating. We estimate parameters of models based on Bayesian and Gibbs sample algorithm. The study shows that skewed link models are more successful than symmetrical models. In addition, the parameters of skewed link models are also explicable.