Demand Forecasting PoC

Problem

A large US retailer weighing an alpha-stage automl-based demand-forecasting model before committing to it.

What was built

• Analysis of ~600 GB of sales history (BigQuery, three years, apparel).
• Feature engineering plus training and validation code around the vendor AutoML model.
• A benchmark against ARMA / ARIMA / ARIMAX and a summary report for the client.

Outcome

The simple ARMA baseline outperformed the alpha-stage vendor model — a data-backed technology assessment delivered before any rollout decision.

Challenges and Lessons

• Communication is key to project success — even small message counts.
• Each query against a large dataset has a real cost — a FinOps lesson; use BigQuery’s row-preview instead of a full query when it’s sufficient.
• Feature engineering often matters more than model choice.
• Simpler solutions (ARIMAX-class models) can be more accurate and cheaper than very complex ones (AutoML).
• Before constructing models, one should deeply understand the data and its characteristics.