A Comparative Evaluation of Machine Learning Algorithms for Anomaly Detection in Mobile Money Transactions

Abstract

Nigeria's mobile money ecosystem has expanded rapidly, with mobile money operators processing tens of trillions of naira annually and serving over 220 million subscribers; however, financial institutions lost ₦52.26 billion to fraud in 2024 alone, and automated fraud detection approaches remain largely unexplored specifically within the Nigerian mobile money context. This study presents a comparative evaluation of four machine learning algorithms Logistic Regression, Decision Tree, Random Forest, and XGBoost for anomaly detection in mobile money transactions, with the aim of identifying a viable deployment candidate for Nigerian fintech platforms. The four algorithms were trained and tested on the PaySim synthetic mobile money dataset under identical experimental conditions, using an 80/20 stratified train-test split combined with a hybrid resampling strategy that undersampled the majority class before applying SMOTE to the minority fraud class. Model performance was evaluated using Precision, Recall, F1-Score, AUC-ROC, and Inference Speed. XGBoost achieved the highest F1-Score of 35.70% and AUC-ROC of 99.98%, with all tree-based models recording Recall above 99% and all four algorithms demonstrating real-time viable inference speeds. Precision remained low across the tree-based models, attributed to the synthetic nature of the PaySim dataset and the undersampling strategy applied during training. XGBoost is recommended as the primary deployment candidate for Nigerian fintech platforms, with precision improvement through access to real transaction data and threshold optimization identified as immediate next steps.