Is fraud detection feasible without training data? Testing an expert-based approach

Abstract

We aim to derive a fraud detection approach applicable to conditions where historical fraud data is absent, inadequate, or outdated for making predictions. To this end, we propose a new approach to fraud detection based on expert opinion, enabling tailored tools for various conditions of economic/institutional environments. For this, we determined the relative importance of common fraud indicators based on a widely used model in the literature. We then used this information to formulate a scoring alternative to conventional versions, which uses either the original coefficients or the coefficients obtained from training the model. Finally, these scoring alternatives were compared by their detection performances. The design of this research demanded a multifaceted dataset consisting of expert opinions, financial statement data of non-financial companies in the Istanbul Stock Exchange, and local regulatory authority’s notifications on fraudulent companies. The analysis of the detection performances indicates that the proposed alternative scoring method poses a feasible alternative with competitive performance and fewer data requirements. This research’s approach sidesteps the training data requirement and provides financial analysts, auditors, and regulatory bodies a versatile classifier for various use cases regarding financial data, such as detecting fraudulent financial activity, as demonstrated in this study.

Keywords

• financial statement fraud
• fraud detection
• beneish model
• probit regression
• expert knowledge
• best-worst method