Favorita: Vertex AI for ML
Take the forecast-ready Favorita tables and train, evaluate, and serve a sales-forecasting model on Google Vertex AI — a managed path from clean data to predictions you can actually call from production.
Primary Outcome
A trained, deployed sales-forecasting model on Vertex AI with a callable prediction endpoint.
Solution
A Vertex AI pipeline that trains on BigQuery features, evaluates and registers the model, and deploys a callable prediction endpoint — with one-command retraining.
Deliverables
A training pipeline that reads features from BigQuery and trains on Vertex AI
Model evaluation with forecasting metrics (RMSE, MAE, and error by store and item family)
A registered model version in the Vertex Model Registry
A deployed prediction endpoint for online or batch inference
Reproducible config so retraining is a single command
Strategic Context
A model that lives in a notebook is a science project. A model that retrains on a schedule and serves predictions through an endpoint is a business asset. Vertex AI removes most of the undifferentiated infrastructure work between those two states.
Technical Architecture
A model that lives in a notebook is a science project. A model that retrains on a schedule and serves predictions through an endpoint is a business asset. Vertex AI removes most of the undifferentiated infrastructure work between those two states.
Problem Statement
Forecasting models that work in a notebook rarely make it to production. There's no managed way to retrain them, version them, or serve predictions that the rest of the business can call.
Links
What's Included
Vertex training scripts, the feature-extraction queries against BigQuery, evaluation code, model-registry registration, and endpoint deployment configuration.
FAQs
Do I need GCP for this?
Yes — it's built on Vertex AI and BigQuery. The modeling logic transfers to other clouds, but the orchestration is GCP-native.
AutoML or custom training?
The template shows the custom-training path so you control the model, but the registry and serving steps work with AutoML too.
Tech Stack
Tool 1
Tool 4
Tool 4
Tool 3
Tool 2
Tool 4
Primary Outcome
A trained, deployed sales-forecasting model on Vertex AI with a callable prediction endpoint.
Problem Statement
Forecasting models that work in a notebook rarely make it to production. There's no managed way to retrain them, version them, or serve predictions that the rest of the business can call.
Solution
A Vertex AI pipeline that trains on BigQuery features, evaluates and registers the model, and deploys a callable prediction endpoint — with one-command retraining.
Links
Deliverables
A training pipeline that reads features from BigQuery and trains on Vertex AI
Model evaluation with forecasting metrics (RMSE, MAE, and error by store and item family)
A registered model version in the Vertex Model Registry
A deployed prediction endpoint for online or batch inference
Reproducible config so retraining is a single command
What's Included
Vertex training scripts, the feature-extraction queries against BigQuery, evaluation code, model-registry registration, and endpoint deployment configuration.
Strategic Context
A model that lives in a notebook is a science project. A model that retrains on a schedule and serves predictions through an endpoint is a business asset. Vertex AI removes most of the undifferentiated infrastructure work between those two states.
FAQs
Do I need GCP for this?
Yes — it's built on Vertex AI and BigQuery. The modeling logic transfers to other clouds, but the orchestration is GCP-native.
AutoML or custom training?
The template shows the custom-training path so you control the model, but the registry and serving steps work with AutoML too.
Technical Architecture
A model that lives in a notebook is a science project. A model that retrains on a schedule and serves predictions through an endpoint is a business asset. Vertex AI removes most of the undifferentiated infrastructure work between those two states.
Tech Stack
Tool 1
Tool 4
Tool 4
Tool 3
Tool 2
Tool 4