基于显著区分两类客户的小型建筑企业信用评价模型研究

本文以中国某区域性商业银行185个小型建筑企业的贷款客户为样本,将熵值法权重、CRITIC法权重和方差齐性检验法权重进行组合,通过构建非线性目标规划函数反推出单一赋权方法的组合系数,构建了显著区分违约和非违约客户的小型建筑企业信用评价模型.通过ROC曲线原理,对不同赋权模型的结果进行违约判别能力的检验.本文的创新与特色:一是通过组合赋权得到的信用得分的组内平方和越小、组间平方和越大、那么违约与非违约客户差异越显著的思路建立非线性目标规划模型,通过目标函数最大反推出单一赋权权重的组合系数,保证组合权重的大小能够显著区分客户的违约状态.解决了信用风险评价中组合权重的大小必须对违约状态有显著鉴别能力的难题,避免了现有研究的信用评分模型由于忽略指标权重区分违约状态的能力、导致出现越是可能违约的客户、信用得分反而越高的不合理现象,开拓了信用风险评价指标赋权的新思路.二是根据违约样本均值偏离全部样本均值程度越大、这个指标区分违约状态能力越强、权重越大的思路对指标进行赋权,通过方差齐性检验F值刻画指标的权重,使指标权重的大小反映指标鉴别违约状态能力的大小,改变了现有研究的指标客观赋权方法与违约状态鉴别能力无关的弊端.实证结果表明,与单一赋权结果相比,组合赋权模型的灵敏度和特异度分别为83.33%和95.95%,对客户的违约判别能力更强.

Credit evaluation model of small construction enterprises based on discriminating between two types of customers

This paper takes 185 loan customers of small and medium construction enterprises in a Chinese regional commercial bank as the sample. It makes a combination of entropy weight, CRITIC method weight and homogeneity of variance weight, by constructing nonlinear goal programming function, the combination coefficient of single empowerment method is deduced. It establishes a credit evaluation model of small construction enterprises that distinguish between default and non default customers. Through the principle of ROC curve, it tests the default judgment ability of different weighting model results. The contributions of this paper are in two aspects. First, it creates the non-linear goal programming function by minimizing the sum of squares between types, and maximizing the sum of squares within each type. The smaller, the smaller the difference between credit scores for customers of the same type. The larger, the larger the difference in average credit scores between types. The smaller, and the larger, the larger the value for the goal programming function. As such, we maximize the value for the goal programming function, to ensure we maximize the difference in credit scores between good and bad customers and are able to clearly differentiate between good and bad customers. Greater mean deviation of default sample from the whole sample lead to bigger deviation from non-default sample as well, and the indicator can easily distinguish default and non-default sample. According to this rule, we assign a larger weight to the index able to identify default state by F value, different form existing index system, which ignores that ability. The results show that the combination weighting model shows better effect than the single weighting model in the capacity of default judgment. Combination weighting model's sensitivity and specificity are 83.33% and 95.95%, and both better than the single weighting model.