大数据环境下的微观信用评价理论与方法研究

Theories and methods of credit evaluation for individuals and enterprises will be built by analyzing multidimensional mass data such as financial information, business reputation, and network data. A credit evaluation framework in big data environment which is DISC methodology will be built in this study, including Data processing (D), Indicator system (I), Scoring (S), Credit ratings (C). By converting a variety of data into structured data, the regularity of default will be revealed. By revealing the regularity between customer’s single indicator and his or her default state, a series of single indicator selection models will be built. An optimization model will be built by studying the relationship between several indicators and the default state, and the indicator system with the greatest default identification capability will be chosen. According to the relationship between the identification capability and the weight vector, the optimal weight vector with maximum identification capability will be determined, and a series of credit evaluation models will be built. We will build a variety of credit rating models which meet the default pyramid principle "the higher the credit rating, the lower the loss given default". The key indicator feature screening model will be built by mining the relationship between the key feature of the key indicator and the loss given default, and the key feature of the key indicator with the biggest loss given default will also be determined. This study is devoted to the theories and methods of microscopic credit evaluation in big data environment, and through empirical research, the credit evaluation system of farmers, merchants and enterprises in China will be built.

通过解析财务信息、商业信誉、网络数据等多维海量数据，构建企业、个人等微观信用评价理论与方法。本研究构建大数据环境下的微观信用评价体系—DISC方法论，即数据处理(D)→指标体系的构建(I)→评价方程的建立(S)→信用等级的划分(C)。通过种类繁多的大数据的结构化处理挖掘违约规律。通过指标数据与违约状态的联系建立一系列单个指标遴选模型。通过指标组合数据与违约状态的函数关系，遴选违约鉴别能力最大的指标体系。基于评价结果的违约鉴别能力与权重向量的函数关系，反推违约鉴别能力最大的最优权重向量，建立一系列信用评价模型。根据“信用等级越高、违约损失率越低”的违约金字塔原理构建多种信用等级划分模型。通过挖掘关键指标的关键特征与违约损失率的关系，建立关键特征的甄别模型，确定哪一种特征的客户违约损失最大。本研究致力于大数据环境下微观信用评价理论与方法研究，并在实证中建立我国农户、商户和企业信用评价体系。

本研究通过挖掘财务、非财务和宏观环境等海量指标数据与违约状态的规律性关系，构建了大数据环境下的微观信用评价理论与方法。.基于科学问题的研究内容有六：一是最优指标组合的遴选。通过指标组合与违约状态的函数关系，遴选违约鉴别能力最强的指标体系。二是指标最优权重向量的确定。通过违约鉴别能力与权重向量的非线性关系，确定最优指标权重向量，确保指标加权评价结果能够显著区分违约状态。三是最优违约判别临界点的确定。通过违约预测误差最小，反推最优违约判别临界点，保证违约判别临界点能够有效区分违约与非违约两类客户。四是最优信用等级的确定。通过“信用等级越高、违约损失率越低”的违约金字塔原理，构建信用等级划分模型，保证了划分的等级满足信用等级越高、违约损失率越低的评级标准。五是最优预测模型的遴选。通过不同方法和样本的对比检验，遴选违约预测精度最高的模型。六是信用评价非平衡样本处理。通过扩充少数类违约样本等方式生成平衡样本，提高模型对少数类违约客户的预测精度。.基于评价对象的研究内容有五：一是上市公司违约预测。二是小企业违约判别。三是公司债违约预测。四是农户贷款违约判别。五是个体商户违约判别。基于方法论的研究内容有三：一是统计学习方法。二是机器学习方法。三是深度学习方法。.理论成果方面，共发表和录用230篇学术论文。其中，UT-Dallas24期刊论文1篇，FT50期刊论文1篇，AJG(ABS)4*期刊论文3篇，AJG(ABS)3*期刊论文19篇，国家自然基金委管理科学部认定A类期刊论文86篇。高被引论文7篇。完成学术专著6部。实践成果方面，落地上市公司智能风险管控等成果转化3项。人才培养方面，项目成员晋升正教授3人，培养博士后1名，博士生9名，硕士生30名。.综上，本项目已达成预期研究目标。全部研究内容均有若干论文发表，在有关难点和问题上取得了突破，拓展了大数据环境下微观信用评价理论与方法的研究及应用。