Despite its severe and lasting social and financial ramifications, corporate financial crime remains one of the most understudied crime types, as it is often hindered by two challenges. First, its multidisciplinary nature requires both financial and criminological expertise among others to conduct proper investigations. Second, corporate crime data is fraught with constraints such as high dimensionality, complex interactions, and nonlinear functional forms that are ill-suited for classical statistical modeling. The lack of research coupled with the limited resources in corporate crime enforcement represent a great impediment to the advancement of fraud interventions. This dissertation seeks to overcome these specific challenges by unifying cross-disciplinary financial fraud research under a risk factor framework, and by leveraging recent advancements in artificial intelligence. The goal is to examine whether two machine learning algorithms-random forest and neural network-can be used to enhance corporate fraud risk detection/prediction beyond more commonly employed analytical techniques. Findings from the analysis showed that the random forest algorithm outperformed logistic regression and a naive classifier in a 1:1 matched sample. The neural network performed better than a naive classifier but slightly worse than logistic regression. Feature selection improved the algorithms' predictive accuracy and ability to distinguish between classes even further. Despite promising results from the 1:1 matched sample, both machine learning algorithms struggled with a heavily imbalanced 1: many dataset, which represents a more realistic setting. With the implementation of an oversampling strategy and feature selection, the algorithms improved substantially in identifying the rare fraud cases, and showed promise of improvement with further research on imbalanced classification. Feature importance from the random forest classifier identified risk factors that are consistent with findings from prior studies. Measures of financial distress ranked lower in importance than measures of financial health, suggesting future research can build on prior findings on corporate strain to examine specific mechanisms. The analysis also identified auditor independence as a key concept of guardianship and opportunity structure that warrants further study. Findings from this research also have important methodological implications for corporate crime studies-namely, the need to improve measurements of organizational-level fraud risk factors.
Read
