The decision
Retraining costs compute, risk, and review. The retraining trigger is the policy that decides when a refresh is justified, balancing freshness against cost and stability.
Trigger types
- Scheduled retrains on a fixed cadence such as weekly. Simple but blind to actual need.
- Performance based retrains when a tracked metric drops below a threshold. Targeted but needs labels.
- Drift based retrains when input or prediction drift crosses a bound. Early but can fire on harmless drift.
- Volume based retrains after enough new labeled data accumulates.
Designing a good policy
Most teams combine triggers with guards. A drift alarm proposes retraining, a performance check confirms it is warranted, and a cooldown prevents thrashing from retraining too often.
Avoiding two failure modes
- Retraining too rarely lets the model decay and lose accuracy.
- Retraining too often burns resources, adds risk, and can chase noise into a worse model.
The goal is to retrain when it actually improves the model, confirmed by the evaluation gate.
Key idea
A retraining trigger is a policy combining schedule, performance, drift, and data volume signals, with guards, to retrain only when a refresh genuinely helps.