Empirical Examination of Machine Learning Model Performance for Fraud Detection in Nigerian Mobile Banking Transactions

Mobile banking fraud detection requires predictive models that can identify anomalous transaction patterns in real time while minimising false positives that disrupt legitimate customer activity, and the performance of machine learning approaches on Nigerian mobile banking transaction data has not been empirically compared. This study empirically examines and compares the performance of multiple machine learning algorithms for fraud detection using a representative Nigerian mobile banking transaction dataset. A dataset of 500,000 labelled mobile banking transactions (2 percent fraudulent) from a Nigerian commercial bank was used, with features including transaction amount, time of day, geolocation deviation, device identifier, and transaction frequency patterns. Five algorithms were trained and evaluated: logistic regression, decision tree, random forest, gradient boosting (XGBoost), and a long short-term memory neural network. Performance metrics included precision, recall, F1-score, area under the ROC curve, and computational efficiency. Class imbalance was addressed using SMOTE oversampling. Available fraud detection literature from Nigerian mobile banking identifies transaction velocity anomalies and unusual geolocation deviation as the most discriminative fraud features. The Supervised Learning Framework and the Fraud Triangle Theory provide the analytical basis. Findings indicate that XGBoost achieved the highest AUC of 0.978 and best F1-score of 0.913, outperforming the LSTM model at lower computational cost. The study fills an empirical comparison gap for ML fraud detection models calibrated to Nigerian transaction patterns. Recommendations address real-time model deployment architecture and quarterly retraining schedules. Keywords: fraud detection, machine learning, mobile banking, XGBoost, Nigeria.