Financial Fraud Detection Of Listed Companies By Combining Financial Indicators And Texts

Recently,although China’s regulatory system for the securities market has become more and more perfect,but the financial fraud of listed companies has not disappeared with the improvement of the system,and various financial fraud cases have emerged.The act will not only cause the loss of interests of statement users,but also impact the capital environment of the whole market.In the face of the difficult problem of financial fraud,how to build an accurate and efficient financial fraud identification model has become one of the most important concerns of financial statement users and market regulators.Existing research shows that the samples with financial fraud can be effectively identified based on financial indicators and financial texts.In this thesis,we take the financial statement data of listed companies and the text of annual reports of listed companies as the research objects,and build a financial fraud identification model combining financial indicators and financial texts,and the main research contents are as follows.(1)Most existing studies have built machine learning models to discern financial fraud based on financial metrics,but these studies usually use only current year financial metrics and ignore the role of historical data.If a company’s operating income indicator shows a rapid and stable growth trend during the operating period,this is often an early warning sign of financial fraud,because the indicator fluctuates up and down as a normal operation of the company.In this regard,this thesis will construct a year-over-year matrix of financial indicators that contains both historical financial indicator time series information and the magnitude of changes in financial indicators between adjacent years.The Bi LSTM model with attention mechanism is then used to extract deeper financial indicator features from the year-over-year financial indicator matrix to identify financial frauds.(2)Recently,some scholars have mined financial text information for financial fraud identification,in which most of them extracted sentiment features from financial texts and then combined them with machine learning models to identify frauds.Subsequently,some studies have combined financial indicators with financial texts for fraud identification and achieved better results.However,all these methods focus on the study of financial text information,not only do they not deeply explore the financial indicators,but also do not deeply explore the data fusion methods.In response,this thesis extracts deep-level features from historical financial indicators on the one hand,and uses pre-trained language models for textual characterization of financial annual reports on the other.Finally,multiple feature fusion methods are proposed to discern financial frauds.Firstly,the section of "Discussion and Analysis of Operations" is extracted from the text of the company’s annual report.Secondly,it is input into Ro BERTa model for text characterization,aiming to obtain the textual representation of financial text by using the powerful characterization ability of the pre-trained model.Finally,the mined financial indicator features and financial text representations are fused to identify financial frauds.The final experimental results show that the results of the financial fraud identification model combining financial indicators and text constructed in this thesis have good results.