Model drift is the degradation of a machine learning model's predictive performance over time as the statistical properties of the input data or the relationship between inputs and the target variable change relative to the training distribution.
In practicePractitioners distinguish data drift, where the input distribution shifts, from concept drift, where the underlying relationship between inputs and outputs changes. Both can degrade accuracy, calibration and fairness, often without any change to the model itself. Drift monitoring is a core control under SR 11-7 and the EU AI Act post-market monitoring regime. A common confusion is to rely solely on accuracy on recent labelled data, which arrives too late: leading indicators on input features are usually needed to detect drift early.
A bank using a fraud detection model observes that transaction patterns have shifted after a product launch and triggers a drift alert, prompting recalibration of decision thresholds and a scheduled retraining run before fraud losses materialise.
This definition is maintained by Moweb partners and used in live client engagements. For how Model drift applies to your estate, or to challenge a working definition, speak to a partner.