Predicting fraud in financial statements using supervised methods: An analytical comparison

The current era is known as the "age of information," and the capital market is built on information as the economy's primary engine. The system of financial statements of corporations, which is the most significant source of information used in the capital market, produces an information system called accounting. Fraud and manipulation in these financial statements raise corporate risk, erode investor confidence, and cast doubt on the objectivity of accounting experts. Owing to the significance of fraud, this study aims to offer a way to foretell the likelihood of fraud in the financial statements of businesses admitted to the Tehran Stock Exchange between 2014 and 2021. 180 enterprises listed on the stock exchange make up the statistical sample (532 years of companies - suspected fraud years and 908 years - of non-fraudulent companies). According to the independent auditor's assessment, the existence of dormant assets and items, the doubting of the assumption of continuity of activity, the presence of tax discrepancies with other tax areas, and the dearth of adequate performance tax reserves led to the selection of the companies suspected of fraud. 96 financial ratios have been compiled by examining the theoretical foundations and research. In this research, the supervised methods of support vector machine, K-nearest neighbor, Bayesian network, neural network, decision tree, logistic regression, random forest and the hybrid method (bagging) have been used. The results of the research showed that the performance evaluation criteria of precision, accuracy, sensitivity, and F-Measure and efficiency (ROC) and the accuracy result of the confusion matrix in the combined method (bagging) were 72.45, 61.21, 64.74, 62.93, 73.50, and 72.45 percent, respectively, which indicates the better performance and greater ability of this method to predict the possibility of fraud in financial statements compared to other proposed methods.