Graph modelling of transaction networks for fraud detection and systemic risk assessment in Nigerian banks

Abstract

This study employs a graph modelling approach for the analysis of fraudulent transactions and the assessment of systemic risk in the banking system of Nigeria. Five different approaches of graph modelling; the transaction network analysis, community detection, ego networks, anomaly detection, and multilayer graphs, are utilized for the evaluation of the data set of 1,000 transactions over a period of 41 days. Betweenness centrality of the nodes was identified, where the centrality values range from 0.169 to 0.195, indicating the potential for the identification of liquidity hubs of the system, which are critical for the assessment of systemic risk. In addition, community detection using the greedy modularity approach identified clusters of users displaying high levels of cohesion, indicating the existence of fraud rings. Ego networks identified high levels of behavioural heterogeneity, where some users displayed high transaction frequency but low fraud levels, while other users displayed high fraud levels despite low transaction frequency. Anomaly detection identified high levels of fraud subgraphs and cyclic patterns, where K-cycle detection identified 326 cycles, where the fraud ratio increases from 38% to 58% for cycles of length 2, 3, 4, and 5, respectively. Temporal modularity analysis at seven-time window intervals (Q scores: 0.6428-0.7329) enables the analysis of network evolution and fraud concentration patterns. Multilayer modelling with transaction amount, fraud rates, and frequency/KYC tiers enables the identification of behavioural asymmetries. At the optimal fraud ratio threshold of ρ ≥ 0.5, the K-cycle detection framework yields a true accuracy of 61.0%, precision of 37.4%, recall of 47.9%, and an F1-score of 42.0%, These figures compare favourably with the single-feature baseline, which achieves 58.0% accuracy, 29.9% precision, 31.6% recall, and an F1-score of 30.7% on the same cycle set. The results show the viability of using graph methods for fraud detection and risk management in emerging markets' FinTech systems.