【作者】 孙静；

【导师】 李敏强；

【作者基本信息】 天津大学 ， 管理科学与工程， 2013， 博士

【摘要】 小微企业是民生银行三大核心客户群体之一，民生银行已经成为全球最大的小微企业金融服务提供商。在“高信用风险、高人工成本”的小微企业贷款业务中，如何降低不良贷款率是小微金融能否持续发展的关键之一。降低不良贷款率的核心在于如何从众多小微企业中甄选出优质客户。本文以客观数据为依据，基于定量分析方法，得到结论，对小微金融的发展提供借鉴和帮助。根据民生银行TJ分行商贷通申请材料，整理出24个指标。以授信额度为因变量，其它变量为自变量。采用最小角回归、广义线性模型弹性网和分组Lasso三种方法计算。结果发现Lasso变量选择功能最强，弹性网次之，而分组Lasso较弱。综合几种方法结果，企业从事行业对于银行贷款授信额度影响最大，一般建筑、交通运输设备容易获得贷款，而从事金属、日用品的企业较难获得更多贷款。处于事业巅峰期的中年企业家，不管是离异还是没有离异都较易获得更多的银行授信额度，而处于创业期的青年未婚者不太容易从银行获得贷款。根据Lasso回归分析结果表明，企业从事行业对于小微贷款的影响力最大，这点与民生银行推行的“打造专业化支行”战略不谋而合。比较成功的例子是民生银行QZ分行和FZ分行。在分析传统版小微业务存在的缺陷后，提出提升版小微金融基本特征。从优化业务结构措施、担保细则、小微业务风险对策、内部制度建设、售后管理方法五个方面提出改进小微金融的具体措施和条件。Copula实际是一个多元分布函数，主要用于独立同分布多元时间序列建模。本文首次将Coupla方法用于综合评价。根据Lasso回归结果分析，最终选择婚姻状况、已有额度、资产合计、从事行业II四个指标作为申请人贷款能力综合评价指标体系。对于每个指标单独拟合一元分布，把累积分布函数看成标准化函数，标准化数值服从[0,1]区间的均匀分布。采用基于经验Copula的拟合优度检验从5种常见的Copula，正态Copula、t Copula、Gumbel Copula、Frank Copula和ClaytonCopula中选择最优的t Copula，极大似然估计其参数。最后将Copula分布函数值当成综合评价值，给出申请人贷款能力排名。排名结果符合实际授信额度。根据民生银行汇总数据，为了比较各分行存贷效率，选择了存款和贷款指标。以累加DEA模型、自助DEA模型、对偶DEA模型、超效率DEA模型计算各分行效率，各种模型计算结果几乎一致。效率最低的是CC分行，其次是WH分行。TY、JN、TJ、KM、ZZ和NC效率在0.8到0.9之间。ST分行效率接近1，是所有没有达到技术有效分行中效率最高的。最后以SFA计算效率，给出各分行最终排名。更多还原

【Abstract】 Small and micro businesses is one of the three core customers in Minsheng Bank.Minsheng Bank has become the world’s largest small and micro financial serviceproviders. High credit risk and high labor costs are the main characters in small andmicro business. How to reduce the non-performing loan ratio is the key for sustainabledevelopment. A guarantee is to select high quality customers from numerous applicant.Based on applicant data, conclusion was gained to provide reference and help of to thedevelopment of small and micro finance by quantitative analysis.24indicators were complied from application materials of Commercial Sense inMinsheng Bank TJ Branch. Credit indicator was taken as the dependent variable andother variables as independent variables. Three algorithms, Least Angle Regression,Elastic Network and Group Lasso were chosen. The results found that Lasso had thestrongest capacity of variable selection, Elastic Network was the middle, and GroupLasso was the weakest. In a whole, enterprises industry had the greatest efect on credit.Industries of general construction and transportation equipment were easy to get loans.Meanwhile, enterprises engaged in metal and daily necessities were difcult to get moreloans.The Lasso regression showed that engaged industry of small and micro enterpriseshas the greatest influence for loan. It happened to have the same view with the strategyof”building professional branch” in Minsheng Bank. Successful examples were FZbranch and QZ branch. The imperfection of traditional small and micro business wasanalyzed and upgrading version characteristics were put forward. Details of structureoptimization measures, guarantee conditions, small and micro business risk counter-measure, internal system construction and after-sales management came up to improvesmall and micro finance.Copulas is actually a multivariate distribution function. It is mainly used for mul-tivariate time series modeling of independent identically distributed case. We originallyused Coupla as comprehensive evaluation. According to the experience, four indica-tors, marital status, existing lines, total assets and engaged industry were chosen asevaluation indexes system for loan applicant ability. Each index’s distribution was fit-ted separately and the cumulative distribution function was taken as standardization. Standardized value obeyed uniform distribution of [0,1] interval. Based on a goodness-of-fit test from experience copula, five copula, Normal Copula, t Copula, Gumbel Cop-ula, Frank Copulas and Clayton Copula, were taken as candidates and the optimum tCopula was t Copula. Parameters was estimated by.Maximum likelihood Estimates.Finally Copula distribution function was taken as evaluation value which gave capacityranking of applicant lending. The order conformed to the actual credit ranking results.According to Minsheng bank summary data, saving and loan indicators were se-lected for efciency evaluation. Accumulate DEA, Bootstrap DEA, Dual DEA, SuperDEA were applied to calculate efciencies of each branch. The results of four modelwere almost the same. CG branch had the lowest efciency. and WH branch followed.The efciencies of TY, KM, JN, TJ, ZZ and NC branches distributed between0.8and0.9. ST branch’s efciency was close to one and had the highest efcient in all thetechnical inefciencies. Finally the ranking of each branch was computed by SFA.更多还原