Research On The Default Discriminant Model Based On Clustering Balanced Sample Processing

Enterprise default judgment is to construct the regular functional relationship between the enterprise's financial indicators,non-financial indicators,macro indicators and other indicator data and the default state to determine the enterprise's default state.Its purpose is to reduce the credit risk faced by banks and other financial institutions when they provide loans to enterprises.This study establishes a discriminant model for the problem of judging the default of small companies.It takes the small business data of a Chinese commercial bank as an example to conduct an empirical study and analyzes the micro-case analysis of the company's cash flow and operating capabilities.And put forward corresponding suggestions.This research includes five chapters: The first chapter is the introduction.The second chapter is the construction of the default judgment model,the third chapter is an empirical study based on the data of Chinese small enterprises,the fourth chapter is the application of the model in the Huadong CNC company,and the fifth chapter is the conclusion.The research focus of this study is twofold: One is the processing of unbalanced sample data.No matter what method or model,there is a strong dependence on samples,which is also a common feature of data-driven.Therefore,if the sample is not processed well,any method used to process the model will produce large errors.The second is the voting method of multiple discriminant models.In the default judgment model based on multiple models,the weight of each model is different,and the judgment result is different,or even completely opposite.The default discrimination model trained by samples with similar characteristics has high discrimination accuracy.Objectively,there must be a set of weight vectors to improve the accuracy of the default judgment model after multiple judgment models vote.In this study,after clustering the samples,a k-class discriminant model was established,and the discriminant results of new customers were obtained by weighted average of the discriminant results of k models.The main innovations of this study: First,when dealing with unbalanced samples,the k-means method is used to cluster customers and then the samples in each category are balanced(SMOTE)to ensure that similar sample treatments A similarly balanced sample.Training the model with similar balanced samples ensures the one-to-one correspondence between the balanced samples and the model,which helps to improve the accuracy of a single model.The second is in the voting method of multiple models.The processing method of this article is to calculate the Euclidean distance from the new customer to the cluster center of each type of sample.The smaller the Euclidean distance,the greater the weight of the model trained based on this type of sample,and vice versa.The smaller.It is ensured that the model trained with similar balanced samples has a large weight,and the discrimination accuracy is improved.The empirical research of small enterprises shows that the accuracy of the proposed model is higher than that of five big data models such as support vector machine,decision tree and k-nearest neighbor.The study found that in the indicator system for judging the default of Chinese small enterprises,25 indicators such as "Engel coefficient","net cash ratio of operating activities",and "capital immobilization ratio" are the key indicators for judging the default of Chinese small enterprises.The five indicators of "Engel coefficient","net cash ratio of operating activities","capital immobilization ratio","EBITDA margin" and "shareholder equity ratio" are the most important.Although the number of these five indicators only accounts for 20 of the total number of indicators %,but their importance accounts for 54.4% of all indicators.