Question

You deployed a loan-default prediction model on a Vertex AI endpoint. Actual default outcomes (ground-truth labels) only become known 6 to 12 months after a prediction is made. Business stakeholders want early warning if the model's real-world quality begins to degrade, well before those delayed labels arrive. Which monitoring approach best addresses this constraint?

Accepted Answer

Configure input feature drift and prediction distribution drift monitoring on the endpoint so shifts can be detected immediately, and reserve delayed-label performance evaluation for confirmation once outcomes mature.

Answer

Wait for the 6-to-12-month ground-truth labels to arrive, then compute accuracy and AUC to determine whether the model has degraded.

Answer

Disable monitoring entirely because meaningful performance metrics cannot be produced until labels are available.

Answer

Retrain the model on a fixed weekly schedule regardless of any signals, assuming frequent retraining eliminates the need for monitoring.