A retraining schedule is the rule that decides when a model is refitted and on which data: on a calendar (every week, on a rolling window), on a detector’s alarm, continuously (an online estimator), or by a combination; it is part of the model and is validated with it.
| schedule (least squares on the last 500 observations unless stated) | prequential | refits |
|---|---|---|
| never (fitted on the first 1 000 steps) | 0 | |
| every 250 steps | 2.62% | 75 |
| on a Page–Hinkley alarm (refit on the data since the alarm) | 2 | |
| recursive least squares, | 3.24% | every step |
| the truth | 4.62% |
ml_online.schedules.