Question

You deployed a Vertex AI online model that predicts loan default risk. Ground-truth outcomes (whether a customer actually defaults) only become known 6-12 months after prediction, so labels are heavily delayed. Business stakeholders want early warning if the model's behavior degrades in production. Which monitoring approach gives you the earliest actionable signal without waiting for labels?

Accepted Answer

Enable prediction (input feature) drift detection comparing the live serving feature distributions against the training baseline distribution.

Answer

Compute accuracy, precision, and recall on a rolling weekly window as predictions arrive.

Answer

Wait for the delayed labels and back-test the model quarterly using the confusion matrix.

Answer

Configure endpoint autoscaling alerts on replica CPU utilization and request latency.