Model Drift
also: data drift · concept drift
A deployed model getting worse over time as the real-world data diverges from what it was trained on.
Model drift is the gradual degradation of a deployed model’s accuracy as live data drifts away from its training distribution. Two flavors: data drift (the input distribution shifts) and concept drift (the relationship between inputs and target changes). Unlike code, a model can silently rot without any deploy.
Worked example: a demand-forecasting model trained pre-pandemic degrades as buying behavior shifts; nothing in the code changed, but its predictions drift off because the world it learned no longer matches the world it serves. Gotcha: you cannot catch drift without monitoring — track input-feature distributions and, where you get delayed ground truth, live accuracy — because by the time users complain the model has been wrong for a while; the fix is scheduled retraining or a trigger when drift metrics cross a threshold.