An Original Investigation into Temporal Data Shift and Model Degradation in Nigerian Financial Services AI Systems

Machine learning models deployed in financial services experience performance degradation over time as data distributions shift, yet the specific temporal patterns and drivers of this degradation in Nigerian financial contexts, characterised by macroeconomic volatility and rapid digital adoption shifts, have not been theoretically or empirically investigated. This study conducted an original theoretical and empirical investigation into temporal data shift and model degradation in Nigerian financial services AI systems. A model monitoring methodology combining theoretical analysis and empirical longitudinal study was employed, drawing on deployment logs from credit scoring, fraud detection, and customer churn models at two financial institutions in Lagos and Abuja, accessed under research agreements covering 36 months of post-deployment monitoring data. The theoretical analysis identified three shift types specific to Nigerian financial AI: macroeconomic shock shift (discrete performance breaks during currency devaluations and fuel subsidy removal events), digital adoption shift (gradual distribution change as mobile banking demographics expanded), and regulatory shock shift (abrupt feature distribution changes following cashless policy implementation in 2023). Empirical degradation curves confirmed that AUC declined at an average rate of 0.012 per month without retraining and experienced discrete breaks at macroeconomic shock events (average AUC drop 0.09). An original Temporal Shift Typology for Nigerian Financial AI classifies shift events and specifies detection triggers for each type. Adaptive retraining schedules significantly outperformed calendar-based retraining. Expert review confirmed the theoretical and empirical originality.

Keywords: model degradation, temporal data shift, Nigerian financial AI, drift detection, continuous learning