Concept drift

Concept drift is a change over time in the true relationship between a model's inputs and the target it is predicting, so that the pattern the model learned during training no longer holds. A fraud model trained before a new scam technique emerged, or a demand-forecasting model trained before a competitor entered the market, can both keep receiving familiar-looking inputs while the relationship those inputs have to the outcome has genuinely shifted.

This differs from data drift, where the inputs themselves change but the underlying relationship stays the same; concept drift is a change in the relationship itself, which is often harder to detect from input data alone and usually only shows up once prediction errors start rising. Drift can be gradual, a slow trend, or sudden, a step change tied to an external event.

Because concept drift directly erodes accuracy rather than just changing input patterns, it is tracked by monitoring live prediction performance against actual outcomes as part of model monitoring, not just by monitoring input statistics. The standard response is scheduled or triggered retraining on recent training data; in fast-changing environments some systems use online learning that updates continuously rather than retraining in batches.