Using Interpretable Machine Learning for Accounting Fraud Detection – A Multi-User Perspective

Machine learning models are increasingly used to identify accounting manipulations based on disclosed information. Still, most approaches focus on accuracy, which at the same time leads to a high number of false-positive predictions that practically hinder their application. The paper analyzes the need for interpretable machine learning techniques from the perspective of primary users such as statutory auditors, enforcement institutions, or investors with respect to their specific legal and organizational frameworks. From this, requirements for additional explanations are derived, which in turn serve as indicators for plausibility checks or as starting points for investigations, thus improving the manageability of predictions and promoting the implementation of the models.