Question

A retail company deploys a demand-forecasting model to a Vertex AI endpoint. Offline evaluation showed strong accuracy, but production predictions are consistently biased low. Investigation reveals that during training, the 'avg_weekly_sales' feature was computed as a 7-day rolling average from a batch pipeline, while the online serving path computes it from the most recent single day's sales pulled from a real-time database. What is the MOST likely cause of the degraded production performance?

Accepted Answer

Training-serving skew, because the feature is engineered differently in the training and serving pipelines.

Answer

Concept drift, because the underlying relationship between features and demand has changed since training.

Answer

Label drift, because the distribution of the target demand variable has shifted in production.

Answer

Data leakage, because future sales information was included in the training features.