Predicting tax fraud using supervised machine learning approach

Abstract

With the advancement in technology, the tax base in Rwanda has become broader, and as a result, tax fraud is growing. Depending on the dataset used, fraud detection experts and researchers have used different methods to identify questionable cases. This paper aims to predict features of tax fraud using the most robust supervised machine-learning model. This research provides a context where a fraud expert can use a machine-learning model, and an implemented model offers instant feedback to the fraud expert. We evaluate supervised machine learning models such as Artificial Neural Network, Logistic Regression, Decision Tree, Random Forest, GaussianNB and XGBoost. Based on different evaluation metrics, Artificial Neural Network was the most robust model for predicting tax fraud. Findings reveal that the time of business that indicates the difference in time from when a business started and the time it was audited, the domestic businesses, taxpayers who import and export goods, those with no losses, those whose businesses are located in the eastern province, and those registered on withholding and Value Added Tax types are more susceptible to tax fraud. This study is among the few to evaluate the effectiveness of multiple supervised machine-learning models for identifying tax fraud factors on an accurate data set with numerous tax types. The evidence generated in the current study will serve as a valuable tool for both tax policymakers and auditors, as well as for enhancing awareness of more robust methods for predicting tax fraud.