Is the Use of Raw Accounting Data and Self-Supervised Neural-Networks the Next Frontier of Financial Statement Fraud Detection?

Financial statement fraud has been estimated to do more than one trillion Euros of damage each year. Despite modern computing power enabling ever more sophisticated data analytics models, an article published in The Accounting Review found that well-known models for such fraud are still not economically viable in practice. Over the past decade, my research team have made advancements in the detection of financial statement fraud using ensembles of statistical learning models and variable selection theory. The question now is what will provide the next step in the fight against financial statement fraud? One study found that a flexible model was able to extract more information from raw data, instead of the more commonly used financial ratios chosen by experts. However, there have been published criticisms of this research, and further investigation is needed to confirm or refute those findings. Additionally, self-supervised neural-network models have shown promising results in other related fields but have yet to be applied to such fraud detection. Thus, in our most recent work, we propose and evaluate a self-supervised neural-network model within a comparative framework comprising well-known published benchmark models and a large dataset spanning 15 years. We find that with suitable machine learning; raw accounting data can outperform expert-chosen financial ratios. On the other hand, except for very specific circumstances, self-supervised neural-networks fail to outperform tree-based RUSBoost on real-world financial statement fraud data despite their conceptual suitability to the detection task. Finally, there are also circumstances where the best outcome is achieved with a combination of raw data and expert-chosen ratios. This presentation will go deeper into the methodology and results from this comparative analysis, revealing insights for industry and academic research stakeholders.