Using data-driven methods to detect financial statement fraud in the real scenario

Abstract

This study seeks to explore the potential of data-driven methods for developing a financial statement fraud prediction model. We emphasize that building a fraud prediction model that can be used to detect fraud in real-world applications should receive attention from researchers. However, the severe class imbalance issue and the complex nature of fraudulent activities make it a rather challenging task. To address these problems, we apply the combinations of different sampling techniques and tree-based ensemble classifiers to an extensive set of raw financial statement data. The results show that the models using an extensive set of raw financial data, undersampling techniques and boosting tree classifiers are superior in fraud detection. Moreover, several features without a priori knowledge are identified to be important for fraud prediction models by feature importance evaluation. Accordingly, this study provides a methodological guide for designing fraud prediction models for real-world applications and serves as a preliminary step of the knowledge discovery process to complement fraud detection knowledge systems.