Fraud Prediction in Financial Statements through Comparative Analysis of Data Mining Methods

Fraud increases business risks and costs, creates investor distrust, and questions the professional competence and credibility of accounting. Hence, this study aims to employ data mining methods for fraud risk prediction at the companies listed in the Tehran Stock Exchange within the 2014–21 period. For this purpose, 96 financial ratios were collected by reviewing theoretical foundations and research literature. The proposed classifiers such as the k-nearest neighbors algorithm, Bayesian network, support vector machine, and bagging classifier were adopted for fraud prediction. The performance of all classifiers were evaluated relatively poor . Therefore, financial ratios were reduced to enhance the proposed classifiers through the particle swarm optimization algorithm. In fact, 11 effective financial ratios were extracted with a precision of 72.92% and a prediction accuracy validity of 84,82 %. The extracted ratios were then reevaluated by the proposed classifiers for fraud prediction. According to the reevaluation results, all of the proposed methods improved with the extracted financial ratios. The research results indicated that the bagging classifier yielded the highest precision and accuracy, i.e., 84.28% and 76.85%, respectively, and the lowest prediction error, i.e., 23.15%. It was also 87% efficient in fraud prediction.