Research On Default Risk Prediction Based On Convolutional Neural Network

Default risk prediction is to infer the future default status and default probability of an enterprise based on the index data,environmental data and default status of the enterprise in the past and present.Default prediction is very important for stock investment,bond investment,bank loan and commercial credit decision.In recent years,with the rapid rise of deep learning,convolutional neural network is being widely used in various fields of identification problems,in some fields has higher identification accuracy than traditional methods.However,few achievements have been made in the field of enterprise default risk prediction.Among the reasons,some scholars may analyze that convolutional neural network is more suitable for processing images,rather than general financial index data.Therefore,this study tries to use the convolutional neural network model of the processed enterprise data to predict the default risk of enterprises.Part 1 is the introduction,part 2 is the literature review,part 3 is the sample mapping and model construction,part 4 is the empirical analysis of default prediction model,and part 5 is the conclusion.Main work of this study: this paper constructs a pseudo-image data processing based on the standard of the shortest double Euclidean distance,and combines it with the default risk prediction model of the convolutional neural network optimized by parameters.By applying the default risk prediction model in this paper to small enterprises in China,it is verified that the model has a good ability to predict default risk.This paper takes small enterprises as the research object,which has certain practical significance,and can provide a reference for Banks,regulatory authorities,bond and stock market investors to make decisions.The main conclusions of this study are as follows: first,the default risk prediction model constructed in this paper improves the identification effect of defaulting customers,and the overall accuracy is high.The default risk prediction model constructed in this paper is a representative convolutional neural network model in deep learning,with low structural complexity and good prediction effect,which is obviously better than LDA,logistic regression,naive bayes and support vector machine(SVM)models in the discrimination of default customers.The second is that the prediction effect of the model is better after the mapping.Moreover,after optimizing the parameters of this model,this study finds the optimal parameters that are more suitable for processing the data of Chinese small enterprises.After a series of robustness tests,the good prediction ability of this model is more determined.Main innovations and characteristics: first,on the basis of the existing research,a quasi-image arrangement standard with the shortest double euclide distance is proposed,which provides new ideas and ideas for the data processing of the convolutional neural network model,and the empirical results show that the quasi-image arrangement standard with the shortest double euclide distance is effective.Compared with traditional machine learning models such as logistic regression,naive bayes,support vector machines and LDA,the model in this study has higher accuracy in default prediction.Second,based on the classical lenet-5 convolutional neural network model,through a series of robustness tests,this study optimized the model parameters with the highest accuracy,such as the number of convolutional neural network structure layers,learning rate,convolution kernel depth and size,and critical point of model default discrimination,and effectively improved the model prediction ability.