基于BP神经网络的中小企业信用评级

发布时间：2018-05-03 20:33

本文选题：中小企业 + 信用评级　；参考：《江西财经大学》2016年硕士论文

【摘要】：中小企业的发展关系着国家经济社会的发展,我国企业绝大部分是中小企业,其在城镇人口就业、出口贸易、技术创新等方面都发挥着重要作用。目前其主要融资来源是商业银行,但是商业银行对于中小企业融资还存在很大的担忧。究其原因,一方面是中小企业的财务数据等信息不透明;一方面是由于信用评级体系和方法的限制,商业银行无法对中小企业进行有效的信用评估,导致商业银行惜贷。因而,完善中小企业信用评级体系也就变得尤为重要。企业的信用评级有助于商业银行评估中小企业信用风险,帮助优秀的中小企业接受银行更多的政策、资金倾斜;同时,对于经营管理不够好的企业,起着警示督促作用,具有显着的实践意义。在指标体系选取中,结合了定性和定量指标,在完备中小企业信用评级体系,提供了一定的参考;把BP神经网络用于中小企业信用评级,丰富和更新了商业银行信用评级方法,具有重要的学术价值。本文从中小企业融资难作为切入点,让融资问题和信用评级进行良好的对接,为中小企业融资问题解决提供突破口。对于中小企业信用评级指标体系,结合了定性和定量指标,并将定性指标定量化表示,减少了指标选取中的主观人为因素。评分模型运用了 BP神经网络的方法,以50个中小板企业为样本,进行对信用评分,然后也对构建的的神经网络模型进行检验,由此同时,也利用了线性回归的方法对上述样本线性拟合,比较两种模型的结果,分析它们结果的原因,进一步阐述了神经网络在信用评级邻域中具有的重大优势。首先,本文对中小企业面临的融资困境的原因进行阐述,而导致这一现象的原因之一就是银行对于企业信用的担忧,提出了解决融资难的一个突破口,就是完善商业银行对中小企业的信用评级体系。然后,本文参考了中国人民银行,组织协会,建设银行,标准普尔以及穆迪公司的企业信用评级指标,但是由于我国中小企业在信息公开比较少,也存在财务指标数据的不真实的情况,结合此特点,文本建立了中小企业的信用评级指标体系,偿债能力、盈利能力、营运能力、成长能力、经营者及员工素质、创新能力六个一级指标,二级指标有16个。在二级指标的筛选中,通过对前人研究成果的搜集。整理,综合筛选出16个指标,指标体系结合了定性和定量指标,也有财务指标和非财务指标的体现,并且也对领导者管理水平、员工素质、创新能力三个定性指标,进行定量化处理,用可量化的行政管理人员比重、大专以上人员比重、科研技术人员比重来表示。建立起的指标体系比较适用于中小企业,为银行评估企业信用提供一定参考。在指标体系构建完成后,需要选取一个评分模型,传统的信用评级方法有5C、5P、5W,这些方法对于不能量化的因素带有主观不确定性,并且对于专家人员的数量和质量要求高,需要不断更新专家库;统计模型法在信用评级方面的运用,有其简单易操作的特点,但是模型不能反映一个动态的过程;层次分析法用于信用评级在判断相对重要性时存在较大的主观性。然而,在阐述BP神经网络的理论过程中,发现它能够避免在权值确定等方面时的主观性,也能够很好处理非线性问题,具有很强的学习能力,并且适用于中小企业信用评级。经过上述指标和模型的选取,接下来,进行基于此模型的实证分析,并和线性回归模型结果进行对比分析。根据此问题的需要和神经网络理论,BP神经网络采用了 16-8-1的拓扑结构,即设置了只有一层隐含层的网络结构,其中输入层有16个神经元节点(16个二级指标),隐含层有8个神经元,一个输出值就是企业的信用评分。以50个中小板企业样本数据为例,采用样本剖分法,40个样本用于训练BP神经网络,10组数据用于检验网络结构。以绝对误差小于0.05作为容忍范围,训练样本的预测评分准确率高达92.5%,检验样本的准确率为80%。同时,也用相同的数据,对上述样本进行多元线性拟合,因变量为期望信用评分,自变量为16个指标节点,拟合结果的残差却要大的多,拟合样本的准确率只有42.5%,显示出线性回归在中小企业信用评级中的巨大弊端。实证的结果表明,BP神经网络在中小企业信用评级中体现出了巨大优势。最后,本文的结论是:(1)中小企业的融资难问题主要原因是信息不对称,商业银行为了规避风险,对中小企业息贷,因此加强企业信息公开,商业银行加强对中小企业信用评级体系建设,为融资难问题破解助力。(2)中小企业信用指标应同时结合定性指标和定量指标、财务指标和非财务指标的选取,对定性指标定量化,减少在指标中的主观性。(3)信用评级的统计模型法受制于变量数据正态分布的假设,而财务指标数据一般是不会服从正态分布的,在实证的对比中,也看出了统计模型并不适用于中小企业信用评级。(4)同时,在实证中,也证明了,BP神经网络在信用评级中的优势:一是它具有良好的自适应能力,在确定各指标体系权重中,不需要人为来确定,而是根据数据反复训练学习,来确定和调整输入和输出之间的关系,能够弱化主观因素的存在;二是处理非线性的能力。用线性回归模型得出的评分残差比较大,而BP神经网络的残差比较小,体现出处理非线性问题的能力。三是BP网络具有很好的动态评价效果。
[Abstract]:The development of small and medium-sized enterprises is related to the development of the national economy and society. Most of our enterprises are small and medium-sized enterprises, which play an important role in urban population employment, export trade and technological innovation. At present, the main source of financing is commercial banks, but the commercial banks still have great worries about the financing of small and medium-sized enterprises. On the one hand, the financial data of small and medium-sized enterprises are not transparent. On the one hand, because of the limitation of the credit rating system and methods, commercial banks can not carry out effective credit evaluation to small and medium-sized enterprises and lead to the credit crunch of commercial banks. Therefore, it becomes particularly important to improve the credit rating system of small and medium-sized enterprises. It helps the commercial banks to assess the credit risk of small and medium-sized enterprises, help the outstanding small and medium-sized enterprises to accept more policies and fund the banks, and at the same time, it has a warning and supervision role for enterprises which are not good enough in management. In the selection of the index system, it combines qualitative and quantitative indicators to complete the credit of small and medium-sized enterprises. The rating system provides a certain reference; it is of important academic value to use the BP neural network in the credit rating of small and medium-sized enterprises and to enrich and update the credit rating methods of commercial banks. This paper makes a good connection between the financing problem and credit rating from the financing difficulty of small and medium-sized enterprises, and proposes to solve the financing problems of small and medium-sized enterprises. For the credit rating system, the credit rating index system of small and medium-sized enterprises, combining qualitative and quantitative indicators, and quantifying qualitative indicators, reduces the subjective factors in the selection of indicators. The scoring model uses the BP neural network method, takes 50 small and medium-sized board enterprises as samples, carries on the credit score, and then also constructs the nerve. The network model is tested, and at the same time, linear regression is also used to fit the above samples linearly, compare the results of the two models, analyze the causes of their results, and further elaborate the important advantages of the neural network in the credit rating neighborhood. First, the reasons for the financing difficulties faced by the SMEs are explained. One of the reasons for this phenomenon is that the bank is worried about the credit of the enterprise, and puts forward a breakthrough to solve the difficulty of financing. It is to perfect the credit rating system of commercial banks to small and medium-sized enterprises. Then, this article refers to the credit evaluation of the people's Bank of China, the organization association, the Construction Bank, the standard & Poor's and the Moodie company. But because the small and medium enterprises in our country have little information disclosure, there is also an untrue situation of financial index data. In combination with this characteristic, the text establishes the credit rating index system of small and medium-sized enterprises, the solvency, profitability, operation ability, growth ability, the quality of operators and employees, and the six first grade index of innovation ability and two level. There are 16 indicators. In the screening of the two level indicators, through the collection of previous research results, the comprehensive screening of 16 indicators, the index system combines the qualitative and quantitative indicators, also the embodiment of financial and non-financial indicators, and also the leadership management level, staff quality, innovation ability three qualitative indicators, the quantitative department. According to the proportion of quantifiable administrative staff, the proportion of college and above personnel, the proportion of scientific and technical personnel, the established index system is more suitable for small and medium enterprises, and provides a reference for the bank to evaluate the enterprise credit. After the completion of the index system, a scoring model should be selected, and the traditional credit rating method is 5C, 5P, 5W, these methods have subjective uncertainty for the factors that can not be quantified, and the requirements for the quantity and quality of the experts are high, and the expert library needs to be updated continuously. The application of the statistical model in credit rating has its simple and easy to operate characteristics, but the model can not reflect a dynamic process; the analytic hierarchy process is used for credit. Rating has great subjectivity in judging the relative importance. However, in the theory of BP neural network, it is found that it can avoid subjectivism in the determination of weights and so on. It also can handle nonlinear problems well, has strong learning ability and is suitable for credit rating of small and medium-sized enterprises. According to the need of the problem and the neural network theory, the BP neural network adopts the topology of 16-8-1, that is, a network structure with only one layer of hidden layer is set up, in which there are 16 neuron nodes in the input layer (16 two). There are 8 neurons in the hidden layer. One output value is the credit score of the enterprise. Taking the sample data of 50 small and medium sized enterprises as an example, the sample dissection method is used, 40 samples are used to train the BP neural network, and the 10 sets of data are used to test the network structure. The absolute error is less than 0.05 as tolerance range, and the accuracy of the training sample is predicted. Up to 92.5%, the accuracy of the test sample is 80%. at the same time, and the same data is also used for multivariate linear fitting of the above samples. Because the variable is the expected credit score, the independent variable is 16 index nodes, the residual error of the fitting result is much larger, the accuracy rate of the fitting sample is only 42.5%, showing the linear regression in the credit rating of the small and medium-sized enterprises. The empirical results show that the BP neural network has shown great advantages in the credit rating of small and medium enterprises. Finally, the conclusion of this paper is: (1) the main reason for the financing difficulty of SMEs is information asymmetry, the commercial banks have to avoid the risks and interest loans to small and medium-sized enterprises, so the information disclosure of enterprises is strengthened and the commercial banks are strengthened. The construction of credit rating system for small and medium enterprises can help solve the problem of financing difficulties. (2) the credit index of SMEs should be combined with qualitative and quantitative indicators, the selection of financial and non-financial indicators, the quantitative index of the qualitative indicators and the reduction of subjectivity in the indicators. (3) the statistical model method of credit rating is subject to the normal state of variable data. The distribution hypothesis, and the financial index data is generally not subordinate to the normal distribution, in the empirical comparison, it is also found that the statistical model does not apply to the credit rating of small and medium enterprises. (4) at the same time, in the empirical, it is also proved that the BP neural network has the advantage in the credit rating: first, it has good adaptive ability and determines the index body. The weight of the system does not need to be determined artificially, but is trained and learned repeatedly according to the data to determine and adjust the relationship between the input and output, and can weaken the existence of the subjective factors; two is the ability to deal with the nonlinearity. The residual error of the score is relatively large with the linear regression model, and the residual difference of the BP neural network is small, reflecting the non line processing. The ability of sexual problems. Three, the BP network has a good dynamic evaluation effect.

【学位授予单位】：江西财经大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP183;F276.3;F270

【参考文献】