Favorita: XGBoost on Vertex AI

The XGBoost model itself: train, tune with Optuna, and predict — written to run as a Vertex AI job. The interpretable tabular baseline, packaged for managed training instead of a notebook.

Primary Outcome

A tuned XGBoost forecasting model that trains and predicts as a managed Vertex AI job.

Solution

Train, Optuna-tune, and predict scripts that run the XGBoost model as a tracked, reproducible Vertex AI job.

Deliverables

  • train_xgboost.py — training entry point with feature handling

  • optimize_xgboost.py — Optuna hyperparameter search

  • predict_xgboost.py — batch prediction with feature importances

  • Integration with the Vertex job submission and MLflow tracking layers

Strategic Context

For tabular, store-item forecasting, XGBoost remains hard to beat on the accuracy-per-effort curve. The value here isn't the algorithm — it's having it tuned, tracked, and running on managed infrastructure so results are reproducible rather than artisanal.

Technical Architecture

For tabular, store-item forecasting, XGBoost remains hard to beat on the accuracy-per-effort curve. The value here isn't the algorithm — it's having it tuned, tracked, and running on managed infrastructure so results are reproducible rather than artisanal.

Problem Statement

XGBoost forecasting models stay stuck in notebooks with hand-tuned parameters, so results can't be reproduced or run on managed infrastructure.

Links

What's Included

The vertex/models/xgboost directory: train_xgboost.py, optimize_xgboost.py, and predict_xgboost.py, designed to run via the Vertex job submission layer and log to MLflow.

FAQs

Why Optuna?

It searches the hyperparameter space far more efficiently than grid search, which matters when each trial runs on managed infra.

Can I run it without Vertex?

The scripts are standard Python — you can run them locally and skip the submission layer for development.

Tech Stack

Tool 1

Tool 4

Tool 4

Tool 3

Tool 2

Tool 4

Primary Outcome

A tuned XGBoost forecasting model that trains and predicts as a managed Vertex AI job.

Problem Statement

XGBoost forecasting models stay stuck in notebooks with hand-tuned parameters, so results can't be reproduced or run on managed infrastructure.

Solution

Train, Optuna-tune, and predict scripts that run the XGBoost model as a tracked, reproducible Vertex AI job.

Links

Deliverables

  • train_xgboost.py — training entry point with feature handling

  • optimize_xgboost.py — Optuna hyperparameter search

  • predict_xgboost.py — batch prediction with feature importances

  • Integration with the Vertex job submission and MLflow tracking layers

What's Included

The vertex/models/xgboost directory: train_xgboost.py, optimize_xgboost.py, and predict_xgboost.py, designed to run via the Vertex job submission layer and log to MLflow.

Strategic Context

For tabular, store-item forecasting, XGBoost remains hard to beat on the accuracy-per-effort curve. The value here isn't the algorithm — it's having it tuned, tracked, and running on managed infrastructure so results are reproducible rather than artisanal.

FAQs

Why Optuna?

It searches the hyperparameter space far more efficiently than grid search, which matters when each trial runs on managed infra.

Can I run it without Vertex?

The scripts are standard Python — you can run them locally and skip the submission layer for development.

Technical Architecture

For tabular, store-item forecasting, XGBoost remains hard to beat on the accuracy-per-effort curve. The value here isn't the algorithm — it's having it tuned, tracked, and running on managed infrastructure so results are reproducible rather than artisanal.

Tech Stack

Tool 1

Tool 4

Tool 4

Tool 3

Tool 2

Tool 4