What is model drift?
Model drift is the degradation of a model's performance over time as the conditions it was evaluated under change. The model itself is unchanged, and what has moved is the data arriving at it or the policy its answers are judged against. Performance declines without any error being raised.
Also called: Model decay, Performance drift
Why nothing in the system reports that it has stopped being correct
No alert exists for a model that is answering fluently and wrongly. Latency is normal and error rates are normal; the service is available while the answers have quietly stopped matching a policy that changed in March. Drift is discovered by one of two routes, either a scheduled evaluation against a fixed set or a complaint from someone who noticed. The second route is considerably more expensive, and it arrives with a volume of affected decisions already attached.
- Drift produces wrong answers without producing any technical failure.
- Source documents change under a retrieval system that has no view of versions.
- The interval between onset and discovery determines the volume of affected decisions.
- Only a fixed evaluation set separates drift from a change in the questions asked.
How model drift works
- A baseline is established on a fixed set before the deployment opens.
- The same set is re-evaluated on a stated cadence rather than on suspicion.
- Input characteristics are monitored for movement away from the baseline distribution.
- Changes to source documents and policy are treated as drift triggers.
- A material decline halts the affected path and raises a named owner.
- Each evaluation is retained per version, so the onset date can be located.
How model drift is evaluated
Drift is detected by comparison against a fixed reference rather than against last week's traffic. Monitor the following continuously and review them on a cadence:
- accuracy on the golden dataset by evaluation date
- movement in the distribution of incoming requests
- proportion of retrieved documents superseded since indexing
- complaint and correction rate from the people receiving outputs
- elapsed time between a decline beginning and its detection
Limitations, risks and common failure modes
- Evaluation is performed at acceptance and then never scheduled again.
- The golden dataset is updated alongside the policy, which hides the decline.
- Monitoring watches availability and latency while accuracy goes unmeasured.
- A model version is upgraded by the vendor without the enterprise being told.