Financial statements make users understand a company's financial status and operating performance. However, some companies issue false financial statements to hide the company's actual operating conditions for the sake of profit. Therefore, it is particularly important to prevent fraud in financial statements. The research samples of is adopted from the Taiwan Economic Journal (TEJ), and based on the two types of compensation cases, the false financial report and the false public statement, announced by the Securities Investor and Futures Trader Protection Center of the Consortium Corporation, The published prosecutions and verdicts of major securities crimes are used as research samples. The sample data is the listed and OTC companies in Taiwan from 2001 to 2021, and 17 financial variables and 8 non-financial variables were used. This study uses decision tree CART, decision tree CHAID and neural network to initially screen important variables, and then uses Convolutional Neural Network (CNN) and support vector machine (SVM) to establish an effective financial statement fraud detection model, in which the decision tree CART combined with convolutional neural network has the best prediction ability, and its test group has the highest accuracy rate of 90.21%, and the error rate of type 1 and type 2 is 2.86 %.
