Default Discriminant Model Based On The Best Equilibrium Sample Ratio And Optimal Index Combination

The construction of the index system is very important when conducting the discriminatory study of default,because the use of an index system that does not have the ability to discriminate against customers is used to make the default judgment.No matter what method is used to construct the discriminant equation,the accuracy of the discriminant results will not be high.In the process of constructing the system of default discriminating indicators,due to factors such as collinearity between the indicators,the single-indicator's default discriminating ability is strong,and the overall default discriminatory ability of the index system composed of multiple indexes with strong default discriminating ability is not necessarily strong.Corporate bonds that do not breach the“good company”are more likely to obtain loans and their corporate bonds are also easier to purchase.Therefore,in the sample of actual loans or corporate bond purchases,there are very few“bad customers”who default on the contract,and there are many“good customers”who are non-defaulting.This is the well-known“non-balanced sample”problem.If the unbalanced sample cannot be processed,the prediction accuracy of the established model will not be high.The typical treatment of unbalanced samples is to"over-sample"to fit a small number of samples or"under-sample"to extract a large number of samples.In this way,a sample set with a relatively balanced ratio of good and bad customers is constructed.So the question is,what is the optimal ratio of the two types of samples,so that the accuracy of the model can be the highest?Innovation and characteristics:First,In the 2~m combinations of m indicators,the NN was used to measure the accuracy of each combination of indicators,and the one with the highest discrimination accuracy was selected as the optimal indicator combination to establish the model to ensure the highest discriminant accuracy of the model.Second,by constructing a ratio ?between a number of good customers and the total number of bad customers;by means of ratio ratios such as ?=1,2,…,5,there are corresponding ? paired samples;The one with the highest discriminant accuracy of the NN model with multiple different indicator combinations for each of the ? paired samples can be used to backstep the optimal indicator combination of this paired sample.From this pair of ? paired samples,we can obtain the ? index combination and its corresponding neural network model.Through the ? neural network model to predict the training sample and the test sample,we can get the accuracy of the matching ratio.By comparing the accuracy of a plurality of proportion ratios,the best ratio ?*with the highest discrimination accuracy can be obtained.Third,an empirical study of a sample of 1,231non-industrial small enterprises in a commercial bank in China shows that the optimal ratio of good customers to the total number of poor customers ?*=2:1.