ENSEMBLE-BASED FRAUD DETECTION IN NIGERIANBANKING: A SYNTHETIC DATA BENCHMARK ANDCOST-SENSITIVE ANALYSIS

Abstract

 Fraud detection in Nigerian banking is critically hampered by a
scarcity of authentic transaction data for model development, a challenge exacerbated by the rapid growth of digital payments. To address this foundational data gap, this study introduces a novel, high-fidelity synthetic benchmark dataset of 1,000,000 financial transactions, meticulously calibrated to reflect the fraud patterns reported by the Nigeria Inter-Bank Settlement System (NIBSS). Using this dataset, we develop a comprehensive analytical framework to evaluate the economic efficacy of advanced machine learning models. The results demonstrate a substantial potential for fraud loss reduction: optimized Random Forest models achieved a 69.1% decrease in simulated fraud-related costs
(from ₦81.8M to ₦25.2M) while maintaining perfect precision. Alternatively, XGBoost delivered superior recall (74.6%) with an F1-score of 0.854, providing a strategic option for institutions prioritizing fraud detection rates. A SHAP analysis identified transaction amount and associated behavioral features as the strongest fraud indicators and highlighted Web and Mobile channels as requiring enhanced monitoring. This paper makes three principal contributions: the first publicly available, NIBSS-calibrated fraud detection dataset for Nigeria, addressing a pivotal data scarcity in African financial research; empirically validated evidence that ensemble methods, combined with threshold optimization, can reduce fraud costs by up to 69%; and actionable implementation
guidelines for Nigerian banks operating within existing regulatory compliance frameworks. The synthetic data methodology offers a replicable and privacy preserving blueprint for other emerging markets facing similar constraints ondata access and availability.