Data Drift
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
A shift in the distribution of a model's input data after deployment, away from what it was trained on, that quietly degrades prediction accuracy.
What is Data Drift?
Data drift is a shift in the statistical distribution of a model’s input data after deployment, away from the distribution it was trained on. The model itself is unchanged; the world it describes has moved, and its predictions quietly lose accuracy.
Drift is why a deployed model is never finished. Production systems need monitoring that compares live inputs against the training distribution and flags when the gap grows too wide.
How Data Drift Happens
- Behavior changes: customers, markets, or fraudsters change their patterns, so yesterday’s signals stop meaning what they meant.
- Upstream changes: a data source alters its format, units, or collection method, and the model receives inputs it never saw in training.
- Seasonal and cyclical shifts: holidays, weather, and economic cycles move the data in ways a short training window never captured.
A related failure, concept drift, goes one step further: not just the inputs but the relationship between inputs and outcome changes. Both are detected the same way, by watching production data over time.
Data Drift vs Data Leakage
Data drift is a production-time failure; data leakage is a training-time failure. A drifting model was built correctly and worked at launch, but the world moved away from its training data. A leaky model was broken before launch: it learned from information it will never have at prediction time. The names sound alike; the failure modes, symptoms, and fixes have nothing in common.
| Data Drift | Data Leakage | |
|---|---|---|
| When it happens | After deployment | During training |
| Root cause | The input distribution shifts away from the training data | Information unavailable at prediction time enters the training data |
| First symptom | Production accuracy decays, often gradually | Test metrics look too good; production performance collapses |
| Remedy | Monitoring, then retraining on recent data | Strict train/test separation, feature review |
Example of Data Drift
A grocery chain runs a demand-forecasting model trained on years of stable purchasing history. In March 2020, pandemic lockdowns change shopping behavior in days: bulk buying of staples, collapse in prepared foods, whole categories swinging wildly.
The model keeps forecasting the old world, so stores overstock some shelves and run empty on others. Retraining on recent data helps only once enough post-shift history accumulates. The lasting fix is drift monitoring, so the next shift is detected in the input data before it shows up as losses.
Related AI terms: Data Leakage · Data Quality · Model Evaluation · Robustness
Did you like the Data Drift gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About