Which practice helps detect model drift over time?

• A. Monitoring model performance over time for signs of drift; using drift tests and statistical checks.
• B. Increasing the size of the training algorithm.
• C. Reducing the dataset to a small sample.
• D. Ignoring changes in data distribution.

Answer: (A)

Monitoring how a model performs over time is essential to catching drift early. Model drift happens when the data the model sees in operation—the inputs or their relationship to the target—changes from what it was trained on. By continuously tracking performance metrics like accuracy, AUC, calibration, and error rates, you get a signal if the model’s effectiveness starts to fall.

Using drift tests and statistical checks strengthens that signal. You compare current data distributions and feature behavior to the training data, using tests such as distribution comparisons or drift detectors that alert when changes exceed thresholds. This combination lets you distinguish normal variability from meaningful shifts that warrant action.

When drift is detected, you can respond by gathering new labeled data, retraining or updating the model, or adjusting features and thresholds to align with the new data landscape. This approach directly addresses changes over time rather than ignoring them or assuming the model will stay valid.

Increasing the model size doesn’t measure or reveal drift; it changes capacity. Reducing the dataset reduces visibility into current data patterns. Ignoring data distribution changes guarantees drift will go unnoticed.