Use The MLP Neural Network Model To Identify The Scheme Of Financial Fraud In Listed Companies

Financial fraud has always been one of the most abhorrent phenomena in the capital market.Therefore,how to establish an effective scheme to predict financial fraud is of great practical significance.Based on this demand,this article attempts to use neural network models and other methods to establish a scheme that can effectively predict financial fraud.This article first sorts out the concept of financial fraud and related research on motivations of financial fraud at home and abroad,and then summarizes some researches on financial fraud identification models at home and abroad.Then the concept and nature of the indicators of the financial statements are discussed,and the 26 financial indicators used in this article are determined from the seven aspects of listed companies.After that,we used the "CSMAR" database to filter out all the companies and the year of violations notified by the official agency for financial violations in 1999-2018,and then selected the data for financial fraud from this data.After obtaining the list data,and using the Join Quant Quantized Transaction platform,under the strict selection principle of non-financial fraud companies,we obtained a list of non-financial fraud samples and 26 financial indicators of all samples.After cleaning the sample data,a total of 1046 non-financial fraud sample data and 733 financial fraud sample data were obtained.After obtaining valid sample data,in order to reduce the correlation between various indicators and improve the quality of the data,first use Matlab to reduce the original data by using the PCA method to obtain new data with 15 new indicators.Compared with the original data,the quality of the processed data has been greatly improved.In addition,according to general experience,the data is divided into a training set and a test set according to a ratio of 8: 2.After obtaining the training and test data,first establish the initial model parameters through a scientific method.In the initial model: the activation function is set to the currently popular Re LU function;the number of hidden layers is set to 1;the number of hidden layer nodes is set to 5;and the SGD is used to optimize the BP algorithm.The results of the initial model show that the accuracy of the model under the current parameters is not high: the precision rate on the test set is 81.48%,the recall rate is 14.97%,the F1 Score is 25.29%,and the comprehensive accuracy is 63.48%,Therefore,the model must be optimized.The first is to optimize the BP algorithm.By comparing the results of models written by different BP optimization algorithms,it is found that after using the L-BFGS algorithm,the model indicators have improved qualitatively.The second is to optimize the number of nodes in the hidden layer.Through continuous trial and error,it is found that the model index performs best when the number of nodes in the hidden layer is set to 6.The final optimized model results have greatly improved compared to the initial model: the precision rate in the test set reached 93.28%,the recall rate reached 85.03%,the F1 Score was 88.97%,and the comprehensive accuracy was 91.29%.Compared with other recognition models in other literatures,the recognition rate of the final model in this paper is relatively high.In short,the financial fraud identification scheme established in this article has strong feasibility in practice,whether it is from data acquisition,data processing,or model construction and solution results.