A hybrid model based on machine learning and genetic algorithm for detecting fraud in financial statements

Abstract

Financial statement fraud has increasingly become a serious problem for business, government, and investors. In fact, this threatens the reliability of capital markets, corporate heads, and even the audit profession. Auditors in particular face their apparent inability to detect large-scale fraud, and there are various ways to identify this problem. In order to identify this problem, the majority of the proposed methods are based on existing algorithms and have only attempted to identify human or simple data mining methods that have high overhead and are also costly. The data mining methods presented so far have had high computational overhead or low accuracy. The purpose of this study is to present a model in which an improved ID3 decision tree with a support vector machine is used as a hybrid approach and also to improve the performance and accuracy, genetic algorithm and multilayer perceptron neural networks are applied. More efficient feature selection has been used to reduce computational overhead. The tree proposed in the proposed method has the lowest depth possible and therefore has high velocity and low computational overhead. For this purpose, the financial statements of 151 listed companies in Tehran Stock Exchange during 2014-2015 were surveyed and 125 financial ratios were extracted using ANOVA test, 23 fraud related ratios were selected as model input data. The proposed model has a high accuracy of about 80% of prediction accuracy compared to similar models.