Disentangling the black box around CEO and financial information-based accounting fraud detection: machine learning-based evidence from publicly listed U.S. firms

This study investigates the predictive power of CEO characteristics on accounting fraud utilizing a machine learning approach. Grounded in upper echelons theory, we show the predictive value of widely neglected CEO characteristics for machine learning-based accounting fraud detection in isolation and as part of a novel combination with raw financial data items. We employ five machine learning models well-established in the accounting fraud literature. Diverging from prior studies, we introduce novel model-agnostic techniques to the accounting fraud literature, opening further the black box around the predictive power of individual accounting fraud predictors. Specifically, we assess CEO predictors concerning their feature importance, functional association, marginal predictive power, and feature interactions. We find the isolated CEO and combined CEO and financial data models to outperform a no-skill benchmark and isolated approaches by large margins. Nonlinear models such as Random Forest and Extreme Gradient Boosting predominantly outperform linear ones, suggesting a more complex relationship between CEO characteristics, financial data, and accounting fraud. Further, we find CEO Network Size and CEO Age to contribute second and third strongest towards the best model’s predictive power, closely followed by CEO Duality. Our results indicate U-shaped, L-shaped, and weak L-shaped associations for CEO Age, CEO Network Size, CEO Tenure, and accounting fraud, consistent with our superior nonlinear models. Lastly, our empirical evidence suggests that older CEOs who are not simultaneously serving as chairman and CEOs with an extensive network and high inventory are more likely to be associated with accounting fraud.