Which is a sensible way to decide when to retrain a model?
Options
- Every time any library has a new version
- Whenever a developer feels like it
- On a schedule that fits how fast the data changes, plus triggers from monitoring such as drift or a quality drop (Correct option)
- Never: a good model does not need updates
Answer
On a schedule that fits how fast the data changes, plus triggers from monitoring such as drift or a quality drop
Why
Retraining costs compute and carries risk, so it should be a decision based on evidence, not on habit.
Start with a schedule that matches the pace of change (daily for ads, monthly for a stable process), and add triggers: drift alerts, a drop in quality once labels arrive, or new data volume.
Every retrained model still goes through the same evaluation gate before replacing the current one.
Remember
Retrain by schedule plus evidence from monitoring.