AI glossary · Advanced · MLOps & LLMOps · Data & evaluation
Data drift
Data drift is when the characteristics of the data a model actually receives gradually diverge from the data it was trained on. As a result, a model that was accurate at first quietly becomes less accurate over time.
About this term
- Level
- Advanced · Specialist terms that need the inner workings
- Category
- Data, ops & platforms
- Fields
- MLOps & LLMOps, Data & evaluation
- Added
- 2026-10-08
How it works
A model learns the patterns in its training data. But after a service launches, the distribution of incoming data can change, for example because the user base shifts, the seasons change or sensors are replaced. A change in what the input data looks like is sometimes called data drift in the narrow sense, as distinct from concept drift, where the relationship between inputs and correct answers itself changes. The two are also often grouped together as 'model drift'.
Data drift and concept drift
Data driftThe input changes
- Incoming data looks different
- e.g., changes in user base or season
- Detected relatively quickly from input statistics
Concept driftThe relationship changes
- The relationship between inputs and correct answers changes
- The same input now has a different correct answer
- Checked with real performance metrics
공통 Both are cases where real use differs from training, so predictions quietly become wrong
Why it matters
Drift is hard to notice because it happens without error messages. The system keeps running normally, but recommendations become off-target or fraud detection misses more cases. So models in production are monitored by continuously tracking input data statistics and predictions and raising alerts when they differ substantially from training time. When an alert fires, the model is retrained on new data or fixed.
Things to know
Not every change is a problem. If performance holds even when the distribution shifts slightly, there is no need to retrain, so it is best to judge together with actual performance metrics. In tasks where correct answers are confirmed late (such as whether a loan is repaid), performance drops are hard to measure right away, so changes in the input distribution are checked first. The same problem can arise in LLM services when the types of user questions change, so LLM observability tools are sometimes used to track question patterns.
Example
A demand forecasting model for an online store can miss badly when purchase patterns change suddenly, for example around holidays or with shifting trends. So the operations team compares the distribution of key inputs on a dashboard against the distribution at training time, and retrains the model when the gap exceeds a threshold. The same concern, a gap between the distribution seen in training and the one seen in real use, also comes up in generative model research. The 'In-Distribution Forcing' paper featured on this site covers a method for keeping the values a model handles within the distribution it saw during training when it generates long videos.