model drift
Model drift describes a decline in a machine learning model's usefulness when the data or relationships it encounters in production change from those represented during training. The term is often used broadly: data drift means the distribution of model inputs has shifted, while concept drift means the relationship between inputs and the outcome the model is meant to predict has changed. These changes can arise gradually, such as customer preferences evolving, or suddenly, following a policy or market disruption. Teams monitor relevant input and outcome measures to detect possible drift, investigate whether model performance has actually degraded, and decide whether recalibration, retraining, or a different model is appropriate. Drift signals are useful warnings, but they do not by themselves establish the cause or prove that a model needs retraining.