Research On Financial Report Fraud Identification And Prediction Model Of Listed Companies Based On Stacking

With the continuous development of the social economy,the exposure of listed companies in the domestic and foreign financial reporting fraud cases is not uncommon.The disclosure of such incidents not only makes the authenticity of the listed company's financial report questionable,but also the public's confidence in the capital market is seriously The destruction also caused the public to doubt the validity of the work functions of the listed company's regulators and auditors.Therefore,how to quickly and effectively identify and predict the financial reporting fraud of listed companies is of great importance to the majority of small and medium investors,audit institutions and regulatory agencies.This paper studies the data analysis methods of financial fraud in listed companies,mainly focusing on the existing financial reporting fraud identification methods.Most of them are based on whether a listed company has fraudulently constructed a two-category single model for fraud identification,or identification of multiple individual classification models.Performance comparison,the method is too singular,and based on the integrated learning algorithm for the identification and prediction of financial reporting fraud behavior of listed companies.Focusing on the resolution of financial reporting fraud identification and prediction issues,the work of this paper is as follows:(1)This paper sorts out the related literatures on financial report fraud theory,means,characteristics and identification research,clarifies the research and development process of financial report fraud,and summarizes the technical means used in this study,the integrated learning algorithm,combined with the integrated learning algorithm.The principle is based on Stacking's financial report fraud identification and prediction model.(2)In the process of building the model,the financial report fraud recognition and prediction model is constructed based on neural network technology,support vector machine technology and Stacking integrated learning technology respectively.BP neural network and support vector machine are selected as the primary classifier of integrated learning;Logistic regression is used.The model serves as a secondary classifier for integrated learning.(3)Based on the new fraud diamond theory,from the five aspects of fraud motivation,opportunity,rationalization,ability and corporate governance mechanism as the theoretical basis of screening feature indicators,construct the index system of financial report fraud identification,and initially select 42 financial indicators and non-As the initial characteristic index set of the experimental design,the financial indicators are extracted based on the independent sample T test and the self-encoder,and 10 characteristic indicators with significant difference are selected at the P=0.05 significance level.The advantages and disadvantages of the two methods of extracting feature indicators are as follows: the self-encoder is also set to extract 10 feature indicators,and then the training of BP neural network,support vector machine and Stacking integrated learning model is carried out with 10 feature indicators and 42 initial feature indicators respectively.To examine the effect of model recognition based on different combinations of feature indicators.(4)Based on the test of significance,since the encoders feature extraction method and the characteristics of the initial indicators to build the BP neural network,a single classifier and support vector machine(SVM)integrated classifier,and comparing the above three cases to build a single classifier and the model of integrated classifier performance,and the extracted features based on different ways to construct the model of integrated classifier performance.(5)According to the recognition performance of different models,the model results with the highest recognition accuracy and the best comprehensive recognition performance are output to predict the possibility of financial report fraud of listed companies.The main innovation work of this paper includes:(1)Based on the theory of new diamond for fraud in recent years,the latest research results of the model feature index of the filter,namely pressure losses from fraud,fraud opportunities,rationalization(excuses)for fraud,cheating capability and the corporate governance mechanism five aspects of selection of characteristic indexes,compared with previous research,enrich the classification of indicators;And on the index selection,no copy of previous studies have considered indicators,but combining Wang Zhihuan edited book "financial management",the analysis of financial indicators and the finance ministry make no.41 "enterprise financial rules" for the division of the company's financial indicators selection and calculation some innovative,such as long-term debt and the ratio of working capital,interest tax,depreciation and amortisation/total liabilities predecessors did not use the indexes such as,in its list of models for evaluation.In addition,in terms of the extraction method of model feature index,this paper innovatively applied the tool of selfencoder to extract feature index,and achieved better model recognition effect.(2)Existing financial reporting fraud identification study,mostly based on whether fraud of listed companies to build a single model of two classification is used to identify the fraud,or a variety of model of a single classification recognition performance comparison,few scholars will be a single classification model based on integrated learning algorithm to fusion modeling for listed companies financial reporting fraud identification and prediction research.In this paper,the classic financial report fraud recognition algorithm BP neural network and support vector machine were fused with the help of Logistic regression algorithm,and an integrated classifier was constructed to obtain the financial report fraud recognition model with strong predictive ability.(3)In this paper,Logistic regression model is used as the secondary classifier of the integrated model.This algorithm has a strong advantage that the model result can not only predict the "category",but also the probability of something happening,which is helpful for the task that needs to make decision with the help of probability.In existing studies,few scholars have studied the possibility of corporate fraud.In my opinion,it is too simple to draw a conclusion on whether a company is cheating or not.The prediction of the possibility of fraud by the model is of great significance to both regulatory agencies and auditors.Regulators can detect whether a listed company has an inkling of fraud in advance,so as to carry out targeted management and control on the listed company,and strictly examine the company that is predicted to have a high possibility of fraud,so as to prevent the fraud from happening,which is conducive to the effective play of regulatory functions of regulators.With the help of the possibility of fraud,audit institutions can make new arrangements for the audit work,and audit the companies with high probability of fraud can be more rigorous and detailed,so as to identify whether there is fraud,and thus publish the right type of audit opinions.