The rapid digital transformation of financial systems has increased the risk of fraud in mobile payment ecosystems. This paper analyzes fraudulent behavior in the PaySim mobile-money dataset using feature engineering and supervised classification. We trained and compared Logistic Regression, K-Nearest Neighbors (KNN), Decision Tree, and Random Forest classifiers using stratified 80:20 splitting and class-weighting to counter extreme class imbalance. For the test set, Decision Tree achieved the best overall balance between precision and recall (Precision = 0.6835, Recall = 0.9696, F1 = 0.8018, ROC-AUC = 0.9845). Random Forest produced very high recall (0.9838) and ROC-AUC (0.9990) but low precision (0.1576), resulting in many false positives. These results indicate ensemble and tree-based methods can detect most fraud events in this dataset, but there is a trade-off between minimizing missed fraud (false negatives) and limiting false alarms for legitimate users. We recommend using precision–recall analysis, threshold tuning, and cost-sensitive methods in operational settings to control that trade-off.