Key Considerations to be Applied While Leveraging Machine Learning for Financial Statement Fraud Detection: A Review

Financial statement fraud (FSF) is a challenging issue in capital markets and severely affects their overall health and stability. The effective prediction of FSF has become an urgent need for investors. In recent years, scholars have developed several machine learning-based FSF prediction models. This study conducted a systematic review of such models to facilitate an understanding of the latest developments in this field. First, sample and data preprocessing were analyzed, focusing on key aspects such as data sources, splitting training and testing sets, imbalanced samples, cross-period FSF, and handling of zero and missing values. Second, existing research on FSF prediction models was reviewed considering structured and unstructured data. The existing studies exhibited two significant characteristics: expansion if data from structured to unstructured formats and the evolution of methodologies from traditional machine learning to deep learning approaches. Third, the effectiveness of FSF prediction models was evaluated. Indiscriminately pursuing higher recall rates is not advisable. Rather, the effectiveness of the model in terms of predicting FSF must be scientific assessed. Finally, the challenges and opportunities in current research were summarized. The core challenges were identified as the development of robust prediction models and the incorporation of unstructured data into these models. Moreover, leveraging diverse data, deep learning, and large language models can significantly enhance the performance of prediction models. Furthermore, to advance research in this field, this study advocates the construction of a shared and open FSF database.