Model drift
Also called model decay, data drift, concept drift
Model drift is the decline in a model's performance after deployment because the data it sees or the relationships it learned have changed. Data drift means inputs shift away from the training data. Concept drift means the link between inputs and outcomes changes. Both make predictions less reliable over time.
How it works
Model owners set a performance baseline at validation, then track live results against it. Statistical tests compare the distribution of incoming data with training data, and outcome analysis compares predictions with what actually happened. When a metric breaches a set threshold, the model is escalated for review, recalibrated, retrained or withdrawn.
Drift is a core model risk. A fraud or credit model trained before a change in customer behaviour can miss risk or reject good customers. US supervisory guidance SR 26-2 and the PRA's SS1/23 expect ongoing monitoring that detects deterioration early. Monitoring results feed revalidation and are reported to model risk committees.
The rules
Superseded SR 11-7 and SR 21-8 and sets expectations for model validation and ongoing performance monitoring scaled to model materiality.
Banking organisations supervised by the Federal Reserve, OCC and FDIC, mainly those with over $30 billion in assetsSets five principles for model risk management, including ongoing monitoring of model performance; took effect on 17 May 2024 and was updated in April 2026.
UK banks, building societies and PRA-designated investment firms with internal model approvalCommon questions
What is the difference between data drift and concept drift?
Data drift is a change in the inputs a model receives, such as a new customer mix. Concept drift is a change in how inputs relate to outcomes, such as fraudsters changing tactics. Both reduce accuracy.
How often should models be checked for drift?
Rules do not set a single frequency. Supervisory guidance ties monitoring intensity to model materiality, so high-impact credit, fraud and capital models are usually tracked continuously or monthly, and lower-risk models less often.
Does drift affect AI models more than traditional models?
Machine learning models can drift faster because they rely on complex patterns in large datasets that change with behaviour. Traditional statistical models also drift, which is why monitoring applies to every model in the inventory.
