What triggers should decide when to retrain a production model?

Retraining can be triggered by schedule, performance, or data signals. Scheduled retraining (e.g., weekly) is simple and predictable but may retrain unnecessarily or too late. Performance-based triggers fire when a monitored metric (accuracy, AUC, business KPI) drops below a threshold once ground-truth labels arrive, retraining only when quality actually degrades. Drift-based triggers fire when input distributions shift significantly (PSI or KS above a bound), useful when labels are delayed since drift is a leading indicator. Volume-based triggers retrain after enough new labeled data accumulates. In practice teams combine these: monitor drift and performance continuously, and gate automatic retraining behind an evaluation step that promotes the new model only if it beats the incumbent on a holdout set. Also consider retrain cost, label latency, and stability, since retraining too aggressively can add noise and instability.