Empirical Study On Credit Risk Of Listed Manufacturing Companies Based On Unbalanced Samples

Manufacturing industry is an important part of China's industrial structure and the foundation of China's real economy,with a good development momen-tum.However,in recent years,the related credit default disputes have gradu-ally increased,and the credit risk problems have become increasingly prominent.These problems may lead to a slowdown in the growth of the real economy,af-fecting the actual interests of investors and adversely affecting China's economic development.Therefore,it is necessary to evaluate the credit risk of listed manu-facturing companies.In addition,because only a small number of companies will have credit risk,resulting in highly unbalanced samples,it is necessary to adopt reasonable methods to solve the classification problem under unbalanced data.In view of the high imbalance of corporate credit risk data,this paper adopts four methods to classify unbalanced samples based on data and algorithm,determines the best model to evaluate corporate credit risk and the best index to influence classification through comparative analysis.The main research contents are as follows:Firstly,financial and non-financial indicators are introduced to reflect the com-pany's situation more comprehensively.Data preprocessing is completed through the operations of index digitization,splitting training test sets,missing data pro-cessing,and eliminating indicators with low contribution.Secondly,at the data and algorithm levels,four methods are adopted to solve the classification prob-lem of unbalanced samples.First,the training set is resampled(combined sam-pling,oversampling and undersampling)and then put into the random forest model to ensure that the number of positive and negative samples participating in the training is similar.The second is to change the cost of misjudging positive and negative samples by adjusting the training weight of random forest model.Thirdly,the majority samples equal to the minority samples were randomly se-lected and repeated 200 times to construct multiple random forest models,and the results of each model were integrated by voting.Fourth,it is transformed into semi-supervised anomaly detection,that is,detecting a small number of outliers in the data set.The first method is data level,and the last three methods are algorit,hm level.Finally,based on sample imbalance,AUPR and F1 values are selected as the main evaluation indexes of the model,and the performances of the four methods are compared on independent test sets,so as to determine the best model of corporate credit risk evaluation and the best index that affects the classification effect.The results show that the integrated model has the highest AUPR value,and has a high recall rate while ensuring the precision,so it is determined as the final prediction model.The feature selection results of the model show that "return on assets","net profit growth rate","sustainable growth rate","asset-liability ratio","operating profit rate" and"return on net assets" are the best indicators that affect.the classification effect of the model,reflecting that the growth ability,profitability and solvency of listed companies have a great impact on the com-pany's credit risk status.To sum up,this paper compares four methods to classify unbalanced samples,determines the best model to predict the credit risk of the company,and puts for-ward reasonable suggestions for the healthy operation of the company according to the results of the model.