AIF-C01 Practice Question 72: Forecast Accuracy Decline

AWS defines it directly: *data drift occurs when the statistical properties of model inputs change over time, causing discrepancies between training data and production data*, and this *can happen gradually due to evolving user behaviors, market conditions, or abrupt changes like economic shifts*. When drift sets in, *the model is essentially making predictions on data distributions it wasn't trained on*. The Machine Learning Lens lists *assuming that a model trained on historical data will remain accurate indefinitely* as an anti-pattern; the remedy is drift monitoring with retraining triggered when performance crosses a defined threshold (option C).

- A is wrong because overfitting is a property of *training*, visible as a gap between training and validation performance before the model ever ships. A deployed model's weights are frozen; it cannot become more overfitted while sitting still. - B is wrong because a context window is a limit on how much text a *foundation model* can read in one prompt. A tabular demand forecaster has no context window, and exceeding a limit would throw an error rather than cause a slow slide in accuracy. - D is wrong because hyperparameters are settings chosen before training: learning rate, tree depth, number of epochs. They are baked into the finished model and do not decay. Nothing about a static artifact degrades on its own; what changed is the world, not the file.

Reference: MLPERF06-BP03 Evaluate data drift