Research On Financial Fraud Detection Based On Machine Learning From The Perspective Of Signaling Theory: Evidence From Listed Companies In China

With the continuous development of China's capital market,the total number of listed companies is increasing.In order to protect the interests of investors,listed companies are required to disclose financial information regularly.But in the process of information disclosure,financial fraud by listed companies for various purposes has occurred from time to time.This not only brings losses to all stakeholders,but also severely disrupts the development order of country's capital market.Therefore,how to accurately and timely detect the financial fraud of listed companies has become a difficult task.Although scholars at home and abroad have paid great attention to this task,there are still some shortcomings in the selection of financial fraud detection theory,the construction of financial fraud indicators and the research of financial fraud detection methods.Thus,it is of important theoretical and practical significance to construct a more complete and effective financial fraud detection model for listed companies.Based on an overview of the literature in the field of financial fraud detection,this paper selects the signaling theory as the financial fraud detection theory,constructs financial fraud indicators from the perspective of signaling theory,and further performs machine learning algorithms to obtain final detection results.In order to construct more comprehensive financial fraud indicators,this paper selects signaling theory as guidance and picks financial ratio signals,non-financial ratio signals,social structural signals and social emotional signals as financial fraud indicators by combining the behaviors of fraudulent companies and investors with core elements of signaling theory.The financial ratio signals reflecting companies' financial performance and non-financial ratio signals reflecting corporate governance structure,internal control,and external auditing are from the annual financial reports.The social structural signals and social emotional signals are mined from three financial social platforms.So a complete financial fraud detection signals system suitable for China's listed companies has been established.In order to test the rationality of the financial fraud detection signals system,this paper further uses the normal distribution test(K-S test)and non-parametric test(Mann-Whitney U test)to analyze the difference of signals between fraudulent samples and non-fraudulent samples.And signals in the system are gradually put into the single classifiers algorithms SVM,CART and ensemble learning algorithms RF,Adaboost to verify the effectiveness of the selected financial fraud detection signals system.At the same time,this paper uses three different sample matching rates to test the impact on the detection of financial fraud by machine learning algorithms.The experimental results show that the signaling theory does have practical guidance for detection of financial fraud in listed companies,because the selected financial fraud detection signals system according to the signaling theory can indeed detect financial fraud more effectively than only financial indicators or/and non-financial indicators.At the same time,the results also confirmed that sample matching rate indeed will affect the effect of financial fraud detection,and the ensemble learning algorithms perform stably better than single classifiers algorithms.The contributions of this paper are as follows:(1)Discuss the practical guidance value of signaling theory to the financial fraud detection in listed companies.Previous studies mostly lack an effective theory to guide the selection of financial fraud indicators,and a few studies use a certain theory but the selected theory makes the indicators incomplete.Based on this situation,the signaling theory selected in this paper not only enriches the application scenarios of itself,but also makes the selection process of indicators more reasonable and comprehensive.(2)Construct a financial fraud detection signals system,which can get better classification results than only using traditional financial and non-financial indicators.And the financial fraud detection signals system is constructed by extracting financial and non-financial ratio signals from annual financial reports and mining social structural and emotional signals from financial social platforms.(3)Test the effect of different sample matching ratios in detecting financial fraud.In the past,scholars used inconsistent sample matching ratios when matching fraudulent samples to non-fraudulent samples.So this paper uses three different sets of samples with different sample matching ratios to test the influence on the same machine learning algorithms,and suggests that a ratio of 1:1 may be more suitable for financial fraud detection research.