Data Mining Techniques and Forecasting Financial Statement Fraud

The purpose of this study is to compare neural network, decision tree, nearest neighbor and support vector machine data mining techniques in predicting fraudulent and non-fraudulent financial statements. The research method is descriptive-applied and time domain from 2008 to 2018. In this study, financial ratios for two fraudulent and non-fraudulent samples and data mining methods were analyzed. Statistical hypotheses of normality, homogeneity and linearity test for financial ratios of fraudulent and non-fraudulent samples were tested. The normality hypothesis was tested using Kolmogorov-Smirnov test and Shapiro Wilk test. Then Pearson correlation coefficient for the existence of the model for financial ratios and elimination of correlated independent variables was tested. Next, data mining methods are used to test them in predicting financial statement fraud and distinguishing fraudulent and non-fraudulent financial statements. In general, the results show that data mining methods are effective in differentiating fraudulent and non-fraudulent financial statements. The neural network method had a correct prediction of 69.4%, decision tree 65.4%, nearest neighbor 64.4% and support vector machine 78%.