Improved models to detect fraud in financial statements

Studies have estimated the median loss from a single financial statement fraud scheme to be at least one million US dollars and the annual cost of financial statement fraud could exceed 1.2 trillion US dollars worldwide. Many business decisions rely on the accuracy of financial statements, but resources are not available to comprehensively investigate all of them.

Detection of this type of fraud is difficult. Consequently, there is a need for better decision aids such as those developed in this research.

Standard parametric regression-based techniques, particularly logistic regression, have been extensively studied for detecting financial statement fraud. More investigation is needed into non-parametric techniques such as decision trees and ensemble techniques that combine multiple models. Consequently, 34 different models have been compared over a range of ratios of the cost of failing to detect fraud relative to the cost of falsely alleging it, as these costs differ among stakeholders. Some models are the same as those used in prior studies, some are modifications of previously used models, and entirely new ones have also been developed. A large number of potential explanatory variables are also investigated in order to study which are the most useful to detection models. Empirical support has been found for both financial and non-financial explanatory variables, including new variables.

New models developed in this research outperform extant ones on holdout data. Using these models, financial statements can be automatically classified as either fraudulent or legitimate, as well as be ranked according to their likelihood of being fraudulent. This information can then be used to improve early detection, which would mitigate fraud’s cost and help deter its future occurrence.