Build a SKU-Location Demand Forecasting Pipeline with BigQuery ML

This tutorial describes a practical pattern for building a scalable SKU-location forecasting pipeline with BigQuery ML while keeping operational validation in the loop.

1. Shape the demand series

Transform transactions into consistent SKU-location time series. Resolve missing periods, stock-out effects, returns and calendar issues before training.

2. Define a baseline

Create a simple benchmark forecast and keep it alongside the machine-learning output. Scale does not remove the need for a reference model.

3. Train and forecast

Use a time-series model appropriate to the data and produce both point forecasts and uncertainty information where available.

4. Evaluate by cohort

Group series by history length, intermittency, volatility and commercial importance. Promote models only on cohorts where they outperform the baseline.

5. Materialize planner-ready outputs

Store forecast, horizon, confidence information, model version and validation status together so downstream planning systems can distinguish experimental from approved outputs.

6. Automate refresh and monitoring

Schedule retraining and track freshness, bias and cohort degradation. Revalidation should be part of the pipeline rather than a one-time project gate.

Failure controls

• Minimum-history rule.
• Intermittent-demand flag.
• Holdout evaluation.
• Bias monitoring.
• Data freshness control.