Multinomial Statistical Modeling for Predictive Detection and Security of Suspicious Transactions in a Digital Wallet

Abstract

Context and Justification: As financial services undergo accelerated digitalization, the expansion of electronic transactions within digital wallets increases vulnerabilities to fraud, anomalous behavior, and sophisticated cyberattacks. While this transformation strengthens financial inclusion, it simultaneously exposes systems to risks linked to massive data flows, automated payment mechanisms, and adaptive malicious techniques. In this context, predictive detection of abnormal transactional behaviors becomes a critical component for enhancing cybersecurity, resilience, and digital trust. Problem Statement: The central challenge is to design a model that is both high-performing and explainable, ensuring compliance with ethical and regulatory standards while enabling automatic detection of suspicious activities without compromising algorithmic transparency. Methodology: The study uses 1000 real transaction records generated by 100 users of a mobile operator’s digital wallet service in Côte d’Ivoire, each performing ten typical daily operations (payments, withdrawals, deposits, transfers). All transactions were fully anonymized prior to analysis to ensure confidentiality and adherence to data protection and cybersecurity regulations. Each transaction includes quantitative and categorical features describing monetary behavior (amount, frequency, failure rate, inactivity period) and contextual attributes (location, device type, network). Categorical variables were encoded, and continuous variables normalized for comparability across users and cities. Results: Based on probabilistic evaluation, the Multinomial Logistic Regression (MLR) model achieved strong intrinsic performance, with a weighted F1-score of 0.83, precision of 0.84, recall of 0.82, AUC-PR of 0.89, and a low log-loss of 0.42. However, hard-label evaluation reveals a macro F1-score of 0.58, with differentiated class-level performance (F1 = 0.42 for normal, 0.67 for suspicious, 0.65 for fraud), indicating a conservative decision profile aimed at minimizing false negatives. Compared with more complex models such as Random Forest or XGBoost, MLR offers a well-balanced compromise between detection capability and interpretability, ensuring full decision traceability required for auditing and regulatory compliance. Conclusion: This research confirms the relevance of MLR as an explainable, robust, and computationally efficient model for multi-class predictive detection in digital wallets. It also opens promising perspectives through the integration of temporal modeling and hybrid sequential architectures (LSTM/GRU) to enable dynamic, adaptive, and resilient monitoring of financial fraud behaviors.