Tax fraud prediction using machine learning models in Kenya

Abstract

With the rapid technological advancements in Kenya, the tax base has expanded, resulting in an increase in tax fraud cases. Detecting and preventing tax fraud has become a priority for tax agencies to maximize revenue and ensure compliance. This study focused on developing a robust tax fraud prediction model using machine learning techniques. Our approach involved training and evaluating multiple models, including Logistic regression, which was deployed as the baseline model for prediction. Key features such as age, business turnover, total turnover, and total financing expenses were identified and engineered to enhance predictive accuracy. The Random Forest model demonstrated superior performance in identifying fraudulent transactions, achieving notable precision and recall rates of 0.96 and 0.77 respectively. Additionally, exploratory data analysis (EDA) revealed significant patterns that contributed to the understanding of tax fraud behavior. This study highlights the effectiveness of machine learning in accurately detecting tax fraud and provides insights into the most influential factors driving fraudulent activities. Our findings support the application of predictive models for improving fraud detection efficiency in tax systems. KEY WORDS: Tax Fraud, Fraud Prediction, Machine Learning Algorithms, Classification.