The efficiency of statistical and machine learning models in fraud financial statement

The existence and persistence of fraud in financial statements can have adverse impact on the sustainable development of the capital markets as well as the financial health of companies. Using conventional audit procedures which is applied to prevent and detect fraudulent financial statements, auditors fail to cope with emerging accounting frauds. This can be due to many reasons, such as the lack of the required data mining knowledge, the complexity and infrequency of financial frauds, and the auditors without much experience. Accordingly, due to importance of identifying fraud in capital market, different types of statistical and machine learning based models were examined to establish a rigorous and effective model to detect financial statements fraud in this study. For this purpose, 20 variables in the form of the pentagonal fraud with emphasis on the structure of internal controls (pressure, opportunity, justification, capability, arrogance and internal control structure) were used from 166 manufacturing companies listed on Tehran stock exchange over the period 2009-2018. Based on the statistical indices obtained, machine learning based models exhibited higher predictive ability and accuracy than statistical based models in predicting financial statement fraud. The results also showed that C5, CHAID and C&R decision tree models were highly accurate in prediction of fraudulent datapresented in fnancial statement. Accordingly, the efficacy of combination of CHAID, C5 and C&R decision tree algorithms which had the highest accuracy in prediction of fraudulent financial reporting was examined. The high accuracy of 92.61% of the combination of these algorithms in fraud prediction shows that data mining methods based on machine learning models and especially their combination can be used successfully in fnancial statement fraud prediction.