Favorita: End-to-End Forecasting Framework
The whole stack in one repo: dbt on BigQuery for transforms, Vertex AI and XGBoost for models, Prefect for orchestration, and MLflow for tracking — wired together on the Corporación Favorita grocery data. Clone it and study how the pieces actually fit.
Primary Outcome
A working reference for how a forecasting system fits together end to end — data, models, orchestration, and tracking in one repo.
Solution
One coherent repo showing how data transforms, model training, orchestration, and tracking fit together — and section templates to adopt each layer independently.
Deliverables
A dbt project on BigQuery (staging, intermediate, marts, BQML models)
Vertex AI training for XGBoost, random forest, and time-series models
Prefect flows that run dbt and Vertex steps on a schedule
MLflow experiment tracking across model runs
Docker, Makefile, and GitHub Actions for a reproducible workflow
Strategic Context
Most teams assemble a forecasting stack one bolted-on tool at a time and end up with something nobody can explain. Seeing a coherent end-to-end example first saves months of architectural backtracking. This is the map; the section templates are the territory.
Technical Architecture
Most teams assemble a forecasting stack one bolted-on tool at a time and end up with something nobody can explain. Seeing a coherent end-to-end example first saves months of architectural backtracking. This is the map; the section templates are the territory.
Problem Statement
Teams bolt forecasting tools together one at a time and end up with a stack nobody fully understands or can hand off.
Links
What's Included
The full repository: dbt/ for transforms and BQML, vertex/ for training jobs and KFP pipelines, orchestration/ for Prefect flows, MLflow tracking utilities, scripts/ for ingestion, and root Makefile, Dockerfile, docker-compose.yml, and CI workflows.
FAQs
Do I have to adopt the whole thing?
No. The layers are decoupled — take just the dbt project, or just the Vertex models. The section templates exist for exactly that.
Is this production-ready as-is?
It's production-shaped, not turnkey. Treat it as a reference architecture you adapt to your data and cloud setup.
Tech Stack
Tool 1
Tool 4
Tool 4
Tool 3
Tool 2
Tool 4
Primary Outcome
A working reference for how a forecasting system fits together end to end — data, models, orchestration, and tracking in one repo.
Problem Statement
Teams bolt forecasting tools together one at a time and end up with a stack nobody fully understands or can hand off.
Solution
One coherent repo showing how data transforms, model training, orchestration, and tracking fit together — and section templates to adopt each layer independently.
Links
Deliverables
A dbt project on BigQuery (staging, intermediate, marts, BQML models)
Vertex AI training for XGBoost, random forest, and time-series models
Prefect flows that run dbt and Vertex steps on a schedule
MLflow experiment tracking across model runs
Docker, Makefile, and GitHub Actions for a reproducible workflow
What's Included
The full repository: dbt/ for transforms and BQML, vertex/ for training jobs and KFP pipelines, orchestration/ for Prefect flows, MLflow tracking utilities, scripts/ for ingestion, and root Makefile, Dockerfile, docker-compose.yml, and CI workflows.
Strategic Context
Most teams assemble a forecasting stack one bolted-on tool at a time and end up with something nobody can explain. Seeing a coherent end-to-end example first saves months of architectural backtracking. This is the map; the section templates are the territory.
FAQs
Do I have to adopt the whole thing?
No. The layers are decoupled — take just the dbt project, or just the Vertex models. The section templates exist for exactly that.
Is this production-ready as-is?
It's production-shaped, not turnkey. Treat it as a reference architecture you adapt to your data and cloud setup.
Technical Architecture
Most teams assemble a forecasting stack one bolted-on tool at a time and end up with something nobody can explain. Seeing a coherent end-to-end example first saves months of architectural backtracking. This is the map; the section templates are the territory.
Tech Stack
Tool 1
Tool 4
Tool 4
Tool 3
Tool 2
Tool 4