Resumen

Fraud and error are two underlying sources of misstated financial statements. Modern machine learning techniques provide a potential direction to distinguish the two factors in such statements. In this paper, a thorough evaluation is conducted evaluation on how the off-the-shelf machine learning tools perform for fraud/error classification. In particular, the task is treated as a standard binary classification problem; i.e., mapping from an input vector of financial indices to a class label which is either error or fraud. With a real dataset of financial restatements, this study empirically evaluates and analyzes five state-of-the-art classifiers, including logistic regression, artificial neural network, support vector machines, decision trees, and bagging. There are several important observations from the experimental results. First, it is observed that bagging performs the best among these commonly used general purpose machine learning tools. Second, the results show that the underlying relationship from the statement indices to the fraud/error decision is likely to be non-linear. Third, it is very challenging to distinguish error from fraud, and general machine learning approaches, though perform better than pure chance, leave much room for improvement. The results suggest that more advanced or task-specific solutions are needed for fraud/error classification.