View source for Definition:Model drift

📉 '''Model drift''' occurs when a [[Definition:Predictive model | predictive model]] used in insurance — whether for [[Definition:Underwriting | underwriting]], [[Definition:Pricing | pricing]], [[Definition:Fraud detection | fraud detection]], or [[Definition:Claims management | claims]] triage — gradually loses accuracy because the statistical relationships it was trained on no longer reflect current reality. Insurance is inherently tied to shifting conditions: emerging risks, regulatory changes, economic cycles, and evolving customer behavior all alter the data landscape. When the input data or the underlying target variable distribution changes meaningfully after a model has been deployed, the model's outputs begin to diverge from actual outcomes, a phenomenon that can quietly erode an insurer's [[Definition:Loss ratio (L/R) | loss ratio]] or customer segmentation quality.

🔧 Drift typically takes two forms. Data drift, sometimes called covariate shift, happens when the characteristics of incoming [[Definition:Risk | risks]] change — for instance, a [[Definition:Commercial lines | commercial lines]] portfolio that once skewed toward retail tenants now includes a growing share of dark kitchens with different [[Definition:Hazard | hazard]] profiles. Concept drift, by contrast, occurs when the relationship between inputs and the target outcome itself shifts — a [[Definition:Catastrophe model | catastrophe model]] calibrated on historical [[Definition:Hurricane | hurricane]] frequencies may understate risk if [[Definition:Climate risk | climate change]] is altering storm patterns. Insurers and [[Definition:Insurtech | insurtechs]] combat drift through continuous monitoring dashboards, scheduled model recalibration, and champion-challenger frameworks that pit the production model against retrained alternatives on live data.

🛡️ Ignoring model drift can have serious financial and regulatory consequences. A [[Definition:Rate | rating]] model that systematically under-prices a segment will attract [[Definition:Adverse selection | adverse selection]], inflating [[Definition:Incurred loss | incurred losses]] before portfolio managers notice the trend. Conversely, over-pricing caused by stale models drives away profitable [[Definition:Policyholder | policyholders]], shrinking the book. Regulators and [[Definition:Rating agency | rating agencies]] are increasingly asking insurers to demonstrate governance over their [[Definition:Artificial intelligence (AI) | AI]] and [[Definition:Machine learning (ML) | machine learning]] pipelines, including how they detect and remediate drift — making robust [[Definition:Model risk | model risk]] management not just a technical best practice but a compliance imperative.

'''Related concepts:'''
{{Div col|colwidth=20em}}
* [[Definition:Model risk]]
* [[Definition:Predictive model]]
* [[Definition:Machine learning (ML)]]
* [[Definition:Adverse selection]]
* [[Definition:Catastrophe model]]
* [[Definition:Model validation]]
{{Div col end}}