Title: Detecting Fraudulent Firms in External Audits Using Artificial Neural Networks: An Experimental Study on the UCI Audit Risk Dataset
Authors: Wesam H. Ashour, Samy S. Abu-Naser
Volume: 10
Issue: 7
Pages: 124-129
Publication Date: 2026/07/28
Abstract:
External auditors face the menace of financial statement fraud, making it one of their biggest risks. Common methods of audit sample would not be suited for the complexities of today's vast amounts of financial data. This research study will evaluate if artificial neural networks (ANNs) can classify firms as fraudulent or non-fraudulent using the UCI Machine Learning Repository's Audit Risk dataset. This dataset has a total of 776 records of firm data with 27 attributes collected by an external audit office. During the data preparation stages for this project there was an issue with data leakage; the following attributes: Risk_A , Risk_F, Inherent_Risk, Control_Risk, Detection_Risk and Audit_Risk were found to be mathematically deterministic functions of the outcome variable Risk. Consequently, these attributes were removed prior to modeling to ensure that the resulting classification task would be valid and generalizable. Normalization of the remaining 16 attributes was completed using min-max scaling and utilized to develop a feedforward neural network with a single hidden layer of five neurons established in the JustNN program. The backpropagation training utilized a learning rate of 0.6 with a momentum value of 0.8 with training automatically stopping when the average training error was equal to or less than 0.01.