Synergizing domain knowledge and machine learning: Intelligent early fraud detection enhanced by earnings management analysis

Abstract

This paper addresses firm fraud detection by synergizing domain knowledge with machine learning through heuristic and explainable feature construction. Unlike traditional approaches that focus on algorithmic improvements, our method introduces a set of features based on earnings management analysis, providing factors influencing firm fraudulent behavior. Empirical results using a firm-year dataset from China demonstrate better classification accuracy of fraud detection compared to machine learning models with raw financial statement features alone. Additionally, the results remain robust with different false positive rates, observation periods, and firm groups. This domain knowledge-enhanced machine learning method, with alternative features for fraud detection, leads to more transparent regulation and the potential for similar counterfeit detection applications in China and beyond.