Data Drift

What is data drift?

A model trains on past data, but reality doesn't stand still. Data drift is what happens when the distribution of incoming production data slowly wanders away from the data you trained on. Say a model predicts "what people buy on weekends" using 2019 data; a few years later people behave differently, but the model still thinks the world is the same.

What types are there?

Concept drift
Not the data changing, but the relationship between input and output. The definition of a "good email" shifts as spam tactics evolve.
Covariate shift
The input distribution changes — say the age mix of users shifts — while the underlying rule stays the same.
Label drift
The distribution of the target itself changes, like the sales share of a product category moving around.

Why does it happen?

User behavior changes, seasons change, policy changes, competitors change — all of it makes "now" look different from "training time". The closer a model sits to the real world, the more drift hits it.

How do you spot and handle it?

Keep comparing distributions
Use statistics to regularly compare production vs. training data, and alarm when the gap widens.
Retrain in time
Once drift is confirmed, retrain on fresh data so the model catches up.
Add monitoring
Fold drift detection into model monitoring and make it an automatic early-warning system.

Bottom line: data drift is "the world changed and the model didn't". Detecting and retraining for it is what keeps AI reliable long-term.