Calendar-based MLOps retraining schedules are fundamentally flawed because production models don't decay gradually — they fail in sudden, unpredictable shocks. When researchers fitted an exponential forgetting curve to 555,000 fraud transactions, the result was R² = −0.31, meaning it performed worse than predicting the mean. This finding explains why standard decay models consistently misfire in production environments.

Most MLOps platforms assume models degrade smoothly over time, borrowing the Ebbinghaus forgetting curve from 19th-century psychology. The assumption is that feature distributions shift gradually, allowing teams to schedule retraining based on estimated half-lives. However, domains like fraud detection, content recommendation, and supply chain forecasting experience sudden performance drops when new patterns emerge overnight — regulatory changes, competitor exits, or policy shifts create discontinuities rather than gradual decay.

Instead of calendar-based schedules, teams should run a simple diagnostic on their weekly metrics. If R² ≥ 0.4, smooth regime retraining works. If R² < 0.4, episodic regime requires shock detection. The 26-week simulation on Kaggle's fraud dataset revealed two major performance shocks where recall dropped from 0.9375 to 0.7500 and later to 0.7429, with rapid recoveries following each disruption. This seismic pattern proves that production ML systems need real-time anomaly detection, not half-life estimates.