Concept Drift Detection in Insurance Customer Analytics: An Empirical Longitudinal Study of Structural Breaks in Conversion Rate Models (2022–2025)[Détection de dérive des modèles prédictifs en CRM assurantiel : une étude empirique longitudinale des ruptures structurelles des modèles de taux de concrétisation (2022–2025)]

Predictive models for customer behaviour in insurance CRM are typically calibrated on stable historical data and deployed without continuous monitoring mechanisms. This paper provides the first longitudinal empirical documentation of concept drift in an insurance CRM context over a 48-month observation window (January 2022 – December 2025). Drawing on monthly conversion rate data across four acquisition channels, two customer segments and two product lines in a French affinity insurer, we apply Page-Hinkley and 2σ control chart methods to detect structural breaks. Our central finding is that drift becomes detectable as early as June 2024 — 12 to 18 months before the organisational reaction. Drift intensity is segment-dependent: policyholders are more severely affected than prospects, and the digital channel exhibits persistent drift across 22 of 24 post-baseline months. ADWIN does not detect this drift regardless of parameterisation, demonstrating that gradual drift in low-frequency insurance data requires different algorithmic tools.