Suppose an inventory prediction system depends on external data, and that data stops arriving. The primary model may be healthy but unable to make a useful prediction. A fallback is a backup response or implementation that lets the system keep making decisions when its primary approach fails, is unavailable, or produces an unacceptable answer. section 13.4
Choose a fallback that remains useful under the failure you expect:
- Keep a simpler baseline alongside the main model. It can return an answer in the required format when the primary prediction fails. section 8.1
- Use a secondary model trained on different data or built with a different algorithm. One design relies on stable features first and uses a wider feature set to correct its output when appropriate. section 13.4
- Use rules for simple cases or when a stable, predictable response matters more than a model prediction. Decide in advance which inputs or conditions select each response. section 13.4
Inventory predictions with several backups
An inventory system can first fall back to a model using only significant features, then to a time-series model less dependent on external features, and finally to sales resembling the previous week’s, adjusted for expected events and holidays. Data-drift and quality alarms trigger the switch between layers. section 13.5.1
A fallback is not just an error message: it should provide a usable answer when normal prediction cannot. Plan both the alternative and the condition that activates it. Keeping a simple baseline available can make that plan easier to maintain. section 13.4 section 8.1
