Borderline Cases: Evaluating Resampling Techniques for Extreme Imbalance in Financial Statement Fraud Detection

Machine learning is a novel technique that is increasingly used in diverse application areas. The usefulness of machine learning models in real applications has been severely impacted though by the improper consideration of challenges that accompany its use leading to overly optimistic results. One of these challenges is imbalanced data. Machine learning has been reported in several academic studies as a promising data analysis technique for financial statement fraud detection. Our results illustrate that the usage of machine learning models in accounting research often suffers from ignoring class imbalance and using improper performance metrics. We investigate which kind of resampling techniques provide the best results in the context of using machine learning techniques for financial statement fraud detection. We illustrate in computational experiments how they affect the performance of logistic regression models under extreme class imbalance (<1% fraud rate) as necessary for detecting financial statement fraud. The dataset used consists of 90,006 U.S. firm-years on which six resampling methods were compared using recall, precision, F₂ score, and Area Under the Precision–Recall Curve (AUPRC). Our results showed that neglecting the class imbalance and relying on unsuitable performance metrics in this setting leads to misguiding performance results. We further found that the Borderline-SMOTE resampling method significantly outperformed the baseline model achieving the best balance of recall (77.3%) and AUPRC (0.71). The results show that boundary-focused resampling can enable interpretable models such as logistic regression to successfully detect rare financial statement fraud cases without producing excessive false positives