Classical Credit Analysis Approach to Credit Analysis

Thus, in the credit decision-making process, the credit manager's expertise, subjective judgment, and the weight of certain key elements to be considered are the most important determinants.

Under the expert system approach, the vast majority of banks focus on the "5c's" of the borrower, i.e., character and reputation (character), qualifications and ability (capacity), capital or cash, collateral, operating conditions or business cycle (collateral), and the creditworthiness of the borrower (collateral), and the creditworthiness of the borrower (collateral). The creditworthiness of a bank is based on its character, qualifications and ability, capital or cash, collateral, operating conditions or business cycle. Some banks also summarize the content of credit analysis as "5w" or "5p". The "5w" refers to the borrower (who), the purpose of borrowing (why), the repayment period (when), the collateral (what), how to repay (how); "5p" refers to the personal factors (personal), the purpose factors (purpose), the repayment factors (payment); "5p" refers to the personal factors (personal), the purpose factors (purpose), the repayment factors (payment). The "5p" refers to personal factor (personal), purpose factor (purpose), repayment factor (payment), protection factor (protection), perspective factor (perspective). The defect of this method is that it is too subjective and can only be used as an auxiliary credit analysis tool. Loan rating classification model is the financial institutions in the United States Office of the Comptroller of the Currency (oc) the earliest development of the rating system on the basis of the expansion of oc on the loan portfolio is divided into normal, concern, subprime, doubtful, loss and so on 5 categories, and requires a different proportion of different loans to withdraw a different percentage of the loss reserves in order to make up for the loss of the loan.

In China, before 1998, the commercial banks loan classification method has been used by the Ministry of Finance, "Financial System of Financial and Insurance Enterprises," the loan is divided into four categories of normal, overdue, sluggish, doubtful, and the latter three categories collectively referred to as non-performing loans, referred to as "one overdue and two doubtful method". This method underestimated the non-performing loans because it did not include the high-risk loans that still pay interest and have not yet been rolled over. 1998 China started to reform the loan classification based on the international supervisory experience, and classified the loans into five categories of normal, concern, substandard, doubtful, and loss according to the degree of risk, i.e., the five-tier classification method. 2003 China Banking Regulatory Commission (CBRC) issued a document deciding that from January 1, 2004 onwards. In December 2003, the CBRC issued a document deciding that since January 1, 2004, all financial institutions operating credit business in China would formally implement the five-level classification system for loans. The credit scoring method is to assign certain weights to a number of indicators reflecting the borrower's economic status or affecting the borrower's credit status, and to obtain a comprehensive credit score or default probability value through certain specific methods, and to compare it with the benchmark value to decide whether to grant a loan as well as the pricing of the loan, which is represented by the z-scoring model.

The z-score model is a multivariate model based on financial ratios proposed by Altman in 1968. The model uses multivariate discriminant analysis to derive a z discriminant function by analyzing a set of variables to minimize within-group differences while maximizing between-group differences, in which alternative variables are selected or discarded based on statistical criteria. A distinction is made between insolvent and non-insolvent firms based on the magnitude of the z-value as compared to a measure of z. In 1995, for unlisted firms, Altman modified the z-model to obtain the z′scoring model. Altman, Haldeman, and Narayannan extended the original z-scoring model in 1977 to create a second-generation zeta credit risk model. model. The model can effectively classify firms that are about to go bankrupt 5 years before bankruptcy, with an accuracy of more than 90% 1 year before bankruptcy and more than 70% 5 years before bankruptcy. The new model is not only applicable to the manufacturing industry, but its effectiveness is also applicable to the retail industry. In both models, the zeta classification accuracy is higher than the z-score model, especially the prediction accuracy of the longer period before bankruptcy is relatively high. Due to the simplicity of the method, low cost and good results, the above methods are widely used.

It is worth noting that the mathematical and scientific methods in the construction of this kind of model, since the synthesis, there are mainly the following kinds:

1. Discriminant analysis (discriminant analysis)

Discriminant analysis (DA for short) is based on the observation of a number of statistical features. Classification of objective things to determine the category of things. It is characterized by having mastered a number of samples of each category in history, summarizing the regularity of classification, and establishing discriminant formulas. When encountering new things, as long as according to the summarized discrimination formula, you can determine the category to which things belong.

The key to da lies in the establishment of the discriminant function. At present, statistics to establish the discriminant function commonly used methods: one is unknown overall distribution, according to the individual to the overall distance to discriminate between the distance discriminant function; two is known under the premise of the overall distribution of the average probability of misclassification of the minimum classification of the discriminant function, also known as the distance discriminant function, usually referred to as the Bayesian (bayes) discriminant function; the third is the overall distribution of the unknown overall distribution or the overall distribution of the function of the unknown overall distribution of the premise according to Fischer's (bayes) discriminate function. The optimal linear discriminant function according to the Fisher criterion under the premise of the unknown overall distribution or the unknown overall distribution function.

2. Multivariate discriminant analysis (multivariate discriminant analysis)

Multivariate discriminant analysis (MDA) is the most used statistical method in countries other than the United States. The multivariate linear discriminant analysis method can be specified as general discriminant analysis (without considering variable screening) and stepwise discriminant analysis of quantitative information (considering variable screening). However, there are three main assumptions in applying multivariate discriminant analysis (MDA): that the variable data are normally distributed; that the covariances are the same across groups; and that the mean vector, covariance matrix, a priori probability, and cost of misspecification are known for each group.

The shortcoming of this method is that it must be based on a large number of reliable historical statistics, which is a difficult prerequisite in developing countries such as China.

3. Logit Analysis Discriminant Method

The essential difference between logit analysis and discriminant analysis is that the former does not require to satisfy the normal distribution or equal variance, thus eliminating the limitations of the MDA model's assumption of normal distribution. The model mainly uses logistic functions.

The problem with this model is that when the sample points are completely separated, the maximum likelihood estimates of the model parameters may not exist, and the validity of the model is questionable, so that its discriminant correctness is not higher than that of the discriminant analysis method in the case of normality. In addition the method is more sensitive to the discrimination of intermediate regions, leading to the instability of the discriminant results.

4. Neural network analysis method (artificial neural network, referred to as ANN)

Neural network analysis method is a kind of processing method with highly parallel computing ability, self-learning ability and fault tolerance ability developed by applying mathematical methods from the research results of neuropsychology and cognitive science. It can effectively solve the problem of non-normally distributed, non-linear credit assessment, the result of which is between 0 and 1, which is the probability of default under the measure of credit risk. The advantage of the neural network analysis method applied to credit risk assessment is that it has no strict assumption limitations and has the ability to deal with nonlinear problems.Altman, Marco and Varetto (1994) applied neural network analysis in the prediction of the financial crisis of the Italian company; Coats and Fant (1993) Trippi used neural network analysis to predict the financial crisis of the U.S. company and the bank, respectively. Coats and Fant (1993) Trippi used neural network analysis to predict the financial crisis of U.S. companies and banks respectively, and achieved better results. However, to get a better neural network structure, human random debugging is needed, which requires a lot of manpower and time, coupled with the fact that the conclusions of the method have no statistical theoretical basis and are not strong in explanatory, so the application is greatly restricted.

5. Cluster analysis (cluster analysis)

Cluster analysis (cluster analysis) belongs to non-parametric statistical methods. It is used in credit risk analysis to categorize borrowers based on their distance in the sample space calculated from their indicators. One of the main advantages of this method is that it does not require a specific distribution of the population; it is possible to use nominal scales, order scales for the variables, so the method can be used for quantitative research and also to analyze the properties of reality that can not be accurately expressed in numerical terms. This applies well to the requirements of credit risk analysis to categorize data information that does not obey certain distributional characteristics according to quantitative indicators (profitability ratio, quick ratio, etc.) and qualitative indicators (management level, credit rating, etc.). For example, Lundy used this method to process the typical credit application data of consumer loan applicants and their age, occupation, marital status, and living conditions into six categories and return scores to each category, which not only categorizes the borrowers effectively, but also helps the commercial banks to determine their lending strategies.

6. k-Nearest Neighbor Discriminant (k-Nearest Neighbor)

k-Nearest Neighbor Discriminant (k-NND) is a group of k samples selected from the samples with the shortest distance from the identified vectors under the concept of a certain distance in accordance with a number of quantitative variables, which is suitable for the initial distribution and the data collection range of the fewer limitations and reduces the requirements of the content in the form of a functional expression. In addition, knn approximates the sample distribution by dividing the variables into any number of decision intervals within the sample as a whole. tametal uses it for credit risk analysis, taking the martens distance, and classifying the sample with 19 variables selected from the perspective of liquidity, profitability, and quality of capital, and the classification results are not as accurate as those of lda, lg, and neural networks. The reason is that under the same sample capacity, if there is a specific parametric model for a specific problem and it is possible to find out, the non-parametric method is not as efficient as the parametric model.

7. Hierarchical analysis method (AHP)

This method emphasizes the role of human thinking and judgment in the decision-making process, through a certain mode of decision-making thinking process standardization, which is applicable to the combination of qualitative and quantitative factors, especially the qualitative factors play a leading role in the problem, the comprehensive evaluation of corporate credit rating is such a qualitative factors play a leading role in the problem.

Basic step of AHP method is: to establish a hierarchical structure, the hierarchical structure, the hierarchical structure, the hierarchical structure, the hierarchical structure and the hierarchical structure. The steps are: to establish a hierarchical structure, to construct a judgment matrix, to find the largest characteristic root of the matrix and its corresponding eigenvector, to determine the weights, and to carry out the consistency test.

8. Other methods

In addition, there are many other methods: probit method, factor-logistic method, fuzzy mathematical methods, chaos and mutation level method, gray correlation entropy, principal component analysis comprehensive scoring method, principal component analysis and the ideal point of the combination of methods, the original ant colony algorithm, the data envelopment discriminant method and so on. . The application of these methods will be discussed later in the empirical section.