<p>This study addresses the critical issue of fraud detection in mobile money transactions (MMTs) in developing countries. It evaluates the performance of various machine learning classifiers, including Naïve Bayes, Gradient Boosting Decision Trees, Random Forest, and XGBoost, along with a Feedforward Neural Network (FNN) for deep learning. The research focuses on tackling the challenges posed by imbalanced datasets through the application of SMOTE and other resampling techniques. Findings indicate that XGBoost, particularly when combined with SMOTE, achieves the best overall performance, providing a balanced trade-off between precision, recall, and F1-score. XGBoost also demonstrates efficiency in terms of computational time and storage, where time metrics were calculated by measuring the model’s execution time on the test dataset, and storage metrics were assessed by tracking memory consumption during model training and prediction. This analysis reveals that XGBoost, with its optimized computational performance, is a promising candidate for real-time fraud detection applications. Random Forest, while effective, faces overfitting issues and is less computationally efficient. Other models, including Gaussian Naïve Bayes and FNN, show varied results, with some limitations in handling imbalanced data. Additionally, this study explores the integration of fraud detection models into mobile money platforms, emphasizing the need for model adaptation, computational efficiency, real-time performance, and data privacy. By evaluating the feasibility of deploying these models in real-world environments, the research highlights key considerations for financial institutions and policymakers looking to enhance the security and sustainability of mobile money ecosystems. These findings are crucial for improving user trust and mitigating fraudulent activities in developing regions.</p>