Retrain on a schedule when the data changes steadily and retraining is cheap, or on a trigger when measured or estimated performance falls below a threshold. Drift alone is a weak trigger, because not all drift hurts the model.
Card 1 of 228. Answer: Retrain on a schedule when the data changes steadily and retraining is cheap, or on a trigger when measured or estimated performance falls below a threshold. Drift alone is a weak trigger, because not all drift hurts the model.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
The options are a fixed schedule (daily, weekly, monthly), a trigger from monitoring, or continual updates as new data arrives. To pick a schedule, measure how fast the model decays: train on data up to a cut-off, then score each later month separately. The shape of that curve says how often retraining pays.
Example: backtesting a demand forecast shows a mean absolute percentage error of 12% in the first month after training, 14% in the second and 19% by the fourth. A monthly retrain keeps error near 12 to 14% for the cost of one job a month; weekly retraining would buy little more. A model whose error is flat for a year can be retrained on a trigger instead.
Follow-ups: which window to train on (a sliding window forgets old patterns, an expanding one keeps them); how label delay limits how fresh the training data can be; and how to swap safely, by validating the new model against the current one on recent data before promotion, the champion-challenger setup. Retraining also can't fix a broken pipeline, so rule that out first.
