<![CDATA[Digital payment fraud in Indonesia has grown alongside the dramatic expansion of mobile money services, creating a detection problem that conventional rule-based systems are increasingly ill-equipped to handle. This paper examines whether a soft-voting ensemble of Random Forest, Gradient Boosting, and XGBoost can offer a more effective solution. The model was trained on the PaySim synthetic dataset, consisting of 6.36 million mobile money transactions in which fraudulent cases account for just 0.129 percent of all records. SMOTE was used exclusively on the training data to address the extreme class imbalance before model fitting. Five-fold cross-validated Grid Search determined the hyperparameter configuration for each constituent model. On the held-out test set, the ensemble achieved 94.7 percent precision, 91.3 percent recall, 93.0 percent F1-score, and 0.987 AUC-ROC figures that consistently exceeded those of any single algorithm. Examining feature contributions revealed that the sender balance difference and transaction amount carried the most discriminative weight, a finding that aligns with known fraud behavior in mobile payment datasets. A local streaming latency test across 5,000 consecutive transactions produced an average response time of 147.3 ms, with the 99th percentile remaining below the 200 ms operational threshold. Taken together, the results indicate that the ensemble approach is not only statistically superior but also practically deployable within the real-time constraints of digital banking environments.]]>
