Data drift

Data drift is a change over time in the statistical properties of the data a deployed model receives, compared with the data it was trained on. If customer demographics shift, a new product line launches, or an upstream system changes how it logs a field, the input distribution the model now sees may no longer resemble its training data, even though the underlying relationship between inputs and outcome has not changed.

This differs from concept drift, where the relationship between inputs and the target itself changes, for example customer behavior genuinely shifting after a market change; data drift is a change in the inputs alone. Both can degrade accuracy, but they call for different fixes: data drift may be addressed by retraining on more representative data or fixing an upstream pipeline issue, while concept drift requires the model to learn a new relationship.

Because a model gives no automatic warning that its assumptions no longer hold, teams detect drift with statistical tests that compare recent input distributions against the training baseline, as part of model monitoring. Left undetected, data drift is one of the most common causes of a model that quietly loses accuracy after deployment, which is why drift detection is a standard part of production MLOps.