Pattern Recognition for Fraud Detection in Mobile Money Transactions in Nigeria using a Stacked Ensemble Technique

Abstract

Mobile money has made it easier for people in Nigeria to store, send, and receive fund using their mobile phones. This makes financial services more accessible to both rural and urban communities. Although this growth has improved financial inclusion, it has also created opportunities for fraud where existing detection systems struggles due to high false-positive rates that disrupt legitimate transactions. Our study explored a stacked ensemble machine learning model aimed at improving fraud detection while reducing false positives. We used the PaySim dataset from Kaggle, which originally contained 6,362,620 transactions. Random undersampling was used to handle class balance in the dataset, resulting in 13,140 records for model development. Exploratory analysis identified common fraud patterns, including transaction amount, frequency, and unusual balance changes. XGBoost, Random Forest, and Logistic Regression serve as base models, with LightGBM as the meta-learner. We evaluated performance using precision, recall, F1-score, false-positive rate (FPR), and AUC metrics. The model achieved a recall of 99.51% and an AUC of 99.4%, outperforming individual base models. Hyperparameter tuning reduced the FPR by 19%, from 0.94 to 0.76, reducing the misclassification of legitimate transactions. These findings revealed that a stacked ensemble approach detects fraud more effectively and reduces false positives and could be extended to other areas of financial fraud across Nigeria’s financial ecosystem.