Detecting Financial Statement Fraud Using Machine-Learning Methods

Financial statement fraud raises substantial concerns for regulators worldwide, and regulators face severe challenges in detecting and addressing the increased incidence of this type of fraud. This chapter compares three popular machine-learning approaches based on 35 firm-level financial and linguistic features derived from annual reports. Using hand-collected financial statement fraud data in China, we aim to compare different machine-learning models and select the most accurate fraud detection model to improve fraud detection ability. In particular, we aim to assess the predictive performance of the least absolute shrink-age and selection operator (LASSO); random forest and bagging; and support vector machine (SVM) models, and compare the results with the logistic regression method. The findings suggest that the LASSO method outperforms relative to other two methods. This chapter contributes to the literature by selecting both financial and linguistic fraud predictors and contributing to an under-researched area by employing different machine-learning algorithms to detect fraud in financial statements.