Research Reveals MLOps Retraining Schedules Fail Due to Distribution Shock, Not Forgetting

Research fitting the Ebbinghaus forgetting curve to 555,000 real fraud transactions achieved R² = −0.31, proving calendar-based retraining is ineffective. The analysis reveals models experience distribution shock rather than gradual forgetting, invalidating conventional ML ops wisdom. A shock-detection approach is proposed as a practical alternative for production systems, challenging the industry standard of scheduled model retraining.