Evaluation of Credit Risk of Listed Companies Based on BP Neural Network
1. Introduction Credit risk management is an eternal theme for commercial banks. Especially after the Southeast Asian financial crisis, it has attracted the attention and attention of all countries. Both banks and intermediaries urgently need to use new technologies and methods to make more accurate evaluations and decisions on the credit status of enterprises. At present, the commonly used statistical methods are discriminant analysis, LogisTIc regression, principal component analysis, cluster analysis, etc. The statistical models established by these methods can indeed provide a more scientific analysis of credit ratings, but these statistical models There are the following defects: First, all the variables adopted by all models, whether they are financial indicators or not, come from the subjective selection of researchers. It is difficult to assert that no important variables are omitted. Second, the basic assumption of linear multivariate discriminant analysis for variables is a multivariate normal distribution, but most empirical data mostly violate this assumption, and also assume that the covariance matrix between different groups is equal, violating the two samples from two separate groups. intuition. Third, the linear probability model cannot meet the definition requirement of "probability", and has inherent defects. Since the late 1980s, Western developed countries have introduced artificial intelligence into the banking industry to assist banks in making loan decisions. Among them, artificial neural networks in particular have shown great advantages and potential in corporate financial analysis. In our country, whether it is statistical methods or neural network technology to study credit risk, it is still in its infancy. Yang Baoan, Li Hai (2001) used BP neural network to conduct early warning research on corporate financial crisis; Liu Bingxiang, Sheng Zhaohan (2002) used rough neural network to analyze corporate financial crisis; Zhu Shunquan used principal components for the same sample data respectively Analysis and fuzzy comprehensive evaluation methods have studied the financial status of listed companies. Based on the sample data in Zhu Shunquan's article, this article will use the neural network method to further discuss the company's credit rating.
Second, the choice of samples and the selection of indicators The source of sample data is as described above, and will be consistent with Zhu Shunquan's article for comparison. The sample data is selected from 20 companies and 15 evaluation index values ​​selected from the financial data table disclosed by China Securities Journal on April 4, 2000. The evaluation indicators are as follows: the main business profit rate (X1), the return on net assets (X2), the return on total assets (X3), the current ratio (X4), the quick ratio (X5), the total asset turnover ratio (X X6), inventory turnover rate (X7), accounts receivable turnover rate (X8), fixed asset turnover rate (X9), shareholder equity turnover rate (X10), operating cash flow to net profit ratio (X11), main business Revenue cash flow (X12), net profit growth rate (X13), long-term debt ratio (X14), shareholder equity ratio (X15). Since each index in the index system has different dimensions, which brings many difficulties to the evaluation, it is necessary to transform the evaluation indexes of different dimensions into a dimensionless standardized index through appropriate transformation. The data after standardization is shown in Table 1.
3. Credit risk evaluation model based on BP algorithm BP algorithm is a supervised learning algorithm, which uses the minimum mean square error and gradient descent method to achieve the correction of network connection weights and bias weights. The basic idea of ​​its learning is: first set the connection weight and bias weight between each unit to a small random number, then select a training sample, and calculate the error gradient of the sample. This involves two processes: one is the forward process, which passes the input value through each unit until the output unit obtains the output of the network; the other is the reverse process: the difference between the actual output value and the expected output value The error is gradually returned to the input layer through the output layer, and the connection weight and the offset weight are adjusted until the error between the actual output value of the sample and the expected output value is less than the predetermined value.
A BP network with three layers is now established to solve this classification problem. The input variables of the network are composed of 15 indicators that can reflect the solvency, profitability, and development capabilities of the enterprise. Correspondingly, the input layer in the neural network structure requires 15 nodes. For the output layer, we take 3 nodes and use the output values ​​(1, 0, 0), (0, 1, 0), (0, 0, 1) to represent "good credit", "general credit", and "" "Credit difference" three credit levels. The hidden layer nodes should generally satisfy 2n> m, where n is the number of hidden layer nodes and m is the number of training samples. Since the number of samples in this article will be taken as 15 (the remaining 5 samples are used for simulation), n may be taken as 5, where there are 5 nodes in the hidden layer. The structure of the neural network is shown in Figure 1.
According to the network structure of Figure 1, we can establish the following credit risk evaluation model based on BP algorithm:
Expressed as a vector: Y = g {V [g (WX) + B1] + B2}
Using the neural network toolbox technology in Matlab, when we set the maximum training steps to 8000, the error index to 0.02, and the learning rate to 0.01, the results shown in Table 1 were obtained. Continue to use the W, V, B1, B2 values ​​obtained by the study to simulate the remaining five companies, and the results shown in Table 2 can be obtained. It can be clearly seen from Table 2 that the results of the neural network simulation and Zhu Wenzhong's principal component analysis are roughly consistent.
Fourth, the conclusion first, the enterprise credit risk evaluation model based on neural network, using the neural network system's fault tolerance, parallel processing ability, anti-interference ability and ability to deal with nonlinear problems, does not need to subjectively determine the weight of each indicator, And the experimental results show that the BP neural network technology method is effective in the application of corporate credit rating. Second, the use of BP neural network for corporate credit risk rating also has certain limitations. It requires companies to have no unfair and unfair Transaction behavior, the company's financial data must be true and accurate. In addition, the selection criteria of the input variables in the credit evaluation need to be further explored. In addition to the factors of the company's financial technology, the input variables should also consider the factors of the industry and the overall economic base.
Geekvape Vape,Geek Bar Disposable Vape,Geek Bar Vape Kit,Geek Bar Vape Pod
TSVAPE Wholesale/OEM/ODM , https://www.tsecigarette.com