Model drift is the gradual decline in a model's accuracy when the real-world data it sees in production starts to differ from the data it was trained on.
Model drift happens when the joint distribution of inputs and outputs changes over time. Data drift means input features change. Concept drift means the relationship between inputs and outputs changes. Both erode performance unless the model is monitored and retrained.
That silent decay is the main reason deployed models need operational care. A recommendation model trained on holiday shopping behavior will drift when the catalog and buyer intent shift to back-to-school season. A fraud model drifts when attackers change tactics. Drift is not a bug; it is the expected behavior of static models in dynamic environments.
Think of it like this. Think of a navigation app trained on summer road conditions. When winter arrives, closures, ice, and construction change the map. The app still works, but its suggestions become slower and riskier unless it learns the new patterns.
Monitoring systems compare production feature distributions, prediction distributions, and labels against training baselines using statistical tests such as KS tests, population stability index, or embedding distance. When drift exceeds a threshold, alerting triggers investigation and retraining pipelines may activate.
"If accuracy is fine, there is no drift." Accuracy can mask drift if labels lag or if performance is averaged across mixed populations. "One threshold fits all features." Different features drift at different rates. "Retraining fixes everything." Retraining on drifted labels without root-cause analysis can bake the new bias deeper.
Monitoring adds operational cost and false-positive noise, but missing drift causes slow silent failures. Automated retraining reduces manual load but risks feedback loops if labels or upstream data are contaminated.