Research On Financial Crisis Prediction Model For Listed Companies Based On Convolutional Neural Network

Most of the traditional financial crisis prediction models for listed companies have strict requirements on the data distribution and multiple collinearity of prediction variables.When there are too many prediction variables and multiple collinearity,traditional research uses principal component analysis and other methods to reduct data dimension,which reduces the presentation of the company's overall financial situation.The layers of the BP Neural Network model are fully connected.When there are more input features,the number of parameters expands sharply,and the limitation of the number of research samples makes it difficult to converge.As a deep learning algorithm,Convolutional Neural Network(CNN)has no assumption limit on prediction variables and their collinearity.CNN adopts local connection between layers to process more dimensional variables.The weight sharing method reduces the number of training parameters,so that it can converge better under a limited number of samples.The deep neural network will gradually extract the variables of higher importance,which achieves the combination of variable selection and modeling.In view of CNN's powerful pattern recognition capability,this paper adopts comparative research mtehod and machine learning method,defines the relevant concepts of financial crisis prediction models for listed companies on the basis of summarizing previous literature,and introduces Logistic Regression and BP Neural Network prediction model,as well as CNN related theories and common application areas.A CNN prediction model based on 253 imaged prediction variables was constructed,and compared with Logistic Regression and BP Neural Network models based on 27 prediction variables(represented by 11 principal components).On the data sets of 304 financial crisis samples and 912 financial normal samples,the CNN model achieved higher classification accuracy and recall rate of financial crisis companies than the traditional models.The result shows that the convolutional neural network,which was originally used for image classification,is also suitable for classifying companies with different financial conditions.Compared with traditional prediction models,CNN model can analyze more prediction variables,process more data at a larger level,and has a stronger ability to extract the correlationship between variables.The contribution of this paper is applying the convolutional neural network to financial crisis prediction,and proposing a MOC(Minimization of Overall Correlation)method for imaged prediction variables.The prediction variables include financial indicators,audit opinions,equity structure,and macroeconomics indicators.A better classification effect is achieved on 1216 samples.CNN model has great reference value for regulators and investors to identify financial crisis companies in advance.The disadvantage is that the interpretability of model parameters needs to be strengthened.