Comparative Evaluation of Random Forest and XGBoost for Fraud Detection in Mobile Money Payment Systems using SMOTE

Mobile Money Payment Systems (MMPS) have emerged as a rapidly growing tool in the fintech space across Sub-Saharan Africa, transforming the financial landscape of the continent. By introducing mobile money technology, MMPS provide millions of previously unbanked individuals the opportunity to access basic financial services. According to the World Bank Global Findex (2021), mobile money account ownership in Sub-Saharan Africa grew from 12% to 33% between 2014 and 2022, representing nearly 330 million active users. Furthermore, 11 of the world's economies where MMPS dominate are located in Sub-Saharan Africa(World Bank, 2022). However, this rapid expansion has been accompanied by growing challenges regarding fraud. The volume of transactions increased, with the industry processing over $1 trillion in transactions by 2023, yet suffered approximately $1.5 billion in fraud losses in 2022 alone (Adongo, 2025). While machine learning can automate MMPS fraud detection, extreme data imbalance creates significant technical challenge. Class imbalance is a fundamental problem in fraud detection, as machine learning models often struggle to identify rare fraudulent activities when trained on predominantly legitimate data. To address this issue, this study employs the Synthetic Minority Oversampling Technique (SMOTE) to balance the dataset and compares the fraud detection performance of two ensemble learning algorithms: Random Forest and XGBoost. Despite these algorithms being well-researched in other fraud contexts (Hajek, Abedin and Sivarajah, 2023; Albalawi and Dardouri, 2025; Ikermane, Mohy-eddine and Rachidi, 2025), their comparative performance specifically on MMPS data with SMOTE applied remains understudied.