TEMPORAL AND CHANNEL-SPECIFIC PATTERNS INNIGERIAN FRAUD: INTERPRETABLE MACHINE LEARNINGON A LARGE-SCALE SYNTHETIC DATASET

Abstract

An examination of temporal and channel-specific fraud patterns
is conducted using a large-scale synthetic dataset (1 million records) calibrated to Nigerian (NIBSS 2023) fraud distributions. Comparative evaluation of Logistic Regression, Random Forest, and XGBoost models, supported by SHAP interpretability, reveals that the Web (0.34%) and Mobile (0.33%) channels has highest risk. January (0.53%) and 01:00 (0.36%) are identified as peak fraud periods. Analysis confirms that there exist negligible linear correlation between temporal features and fraud, validating the need for non-linear ensemble approaches. The study concludes by proposing an interpretable, channel-aware framework for real-time risk scoring applicable to emerging markets.