Research On Financial Fraud Warning Model Based On Deep Learning

Currently,China’s overall economy has been constantly transitioning from highspeed growth to high-quality development,and adapting to the changing economic stage has become an inevitable trend for the efficient development of Chinese enterprises.However,the incessancy financial fraud of listed companies not only injures the fame and profit of the companies,but also undermines the interests and trust of investors,and even more seriously,disrupts the fairness and justice of the market,affecting the stability and development of the entire capital market.Therefore,it is necessary to identify the risks of financial fraud of listed companies,and achieving more effective and accurate identification is crucial.Thesis constructs a deep learning-based hybrid model to effectively identify financial fraud behavior.Thesis object is the A-share listed companies with financial fraud behavior from 2010 to 2020(excluding the financial industry)in China.It combines financial indicators(mainly including debt repayment capabilities,profitability,development capacity,cash flow analysis,and operating ability);non-financial indicators(mainly including management compensation,external organizations and shareholders’ holding ratios,and internal executive information);text indicators(mainly including annual report text tone,management discussion and analysis,anti-fraud situation,and corporate social responsibility).The rolling time series method is used to make each indicator have time series features,and deep learning technology is used for modeling and prediction.The experimental results show that the deep learning model proposed in thesis can effectively identify financial fraud behavior,with an accuracy rate of up to 89.43%.Moreover,the identification results of the data feature combination of "financial indicators & non-financial indicators & text indicators" are better than those of single financial indicators and the combination of "financial indicators & non-financial indicators",and this conclusion also applies to the prediction results.At the same time,thesis also discusses the problem of insufficient samples in deep learning model research,which may lead to the exclusion of data from other time periods,thereby affecting the accuracy of research results and causing bias.Secondly,the lack of non-financial and unstructured indicators may limit the performance of the model in identifying financial fraud.Finally,the inability to distinguish specific fraud types may limit the in-depth understanding and countermeasures of financial fraud behavior.In conclusion,thesis summarizes the research findings,and believes that the deep learning-based financial fraud identification model constructed in this article can provide a feasible method for financial regulatory agencies and investors to identify potential financial fraud behavior and protect the stability of investors and the market.However,the results of the model cannot be regarded as absolute truth and should be used as a reference for decision-making.