Glossary term
Model Drift
Subject: Risk Management, Operational and Model Risk
Also known as: Concept drift, Data drift
Model drift is the gradual loss of a model's accuracy because the data or relationships it was built on have changed. Predictions that once worked begin to miss, often without any visible error.
Explanation
A model learns patterns from past data. If customer behaviour, markets, regulation or data collection change, those patterns stop holding. Two forms are usually separated. Data drift means the inputs shift, for example a new customer group with different incomes. Concept drift means the link between inputs and outcome changes, for example fraud tactics that evade old signals. Drift is common in credit scoring, fraud detection, pricing and machine learning systems of every kind. Teams detect it by comparing live inputs with the training data, tracking performance against actual outcomes and setting thresholds that trigger a review. The usual responses are recalibration, retraining or retirement of the model. Approval dates say little; the monitoring pack shows whether the model is still fit to use.
Example
A lender trained its credit scoring model before a long period of low interest rates. When rates rise sharply, borrowers behave differently and the share of defaults among high-scoring applicants climbs from 1% to 3%. Monitoring flags the shift, the model risk team rules it out of date and the model is retrained with recent data.
Not to be confused with
- Model error
- A model error is a flaw present from the start. Drift is a model decaying after launch because conditions changed.