Define the forecast contract
Specify the target, series, frequency, forecast origin, horizon, and when the result is needed. For example: every Monday at 06:00, forecast daily unit demand for each store for the next seven days. Record which inputs are known at the origin. A planned promotion may be known in advance; actual future sales are not. Account for data arrival delays and revisions when reconstructing historical forecasts.
A forecast is produced Monday for Tuesday through Monday. Identify its origin, horizon, and one feature that may be unavailable for the entire horizon. Explain why using actual future weather in a backtest can make the result optimistic.
Compare with an appropriate baseline
A useful baseline might repeat the latest observation or the value from the same weekday in the previous week. Choose a baseline that fits the business cadence and data availability. For actual values 10, 14, and 12 and forecasts 11, 10, and 15, absolute errors are 1, 4, and 3; mean absolute error (MAE) is 8/3, approximately 2.67 units. Compare models on the same origins, horizons, and series, and use a cost-sensitive measure when overforecasting and underforecasting have different consequences.
A competing forecast has absolute errors 2, 2, and 2. Its MAE is 2.0. Which forecast has lower MAE, and why might the business still care about whether each error was an overestimate or an underestimate?
Backtest without seeing the future
Use historical cutoffs: at each origin, train or update using only information available then, generate the required horizon, and compare with later realized values. Repeat over multiple origins to cover relevant seasons and operational conditions. Preserve a final recent period for a last assessment after choosing features and methods. Aggregate carefully across series and horizons; a good average can hide poor performance on critical stores or long horizons.
For weekly retraining and a seven-day forecast, propose three historical forecast origins. For each, write the latest permissible training timestamp and the subsequent evaluation window. Identify what must be known at each origin.
Interpret diagnostics and uncertainty
BigQuery ML supports ARIMA_PLUS forecasting and provides model diagnostics through ML.ARIMA_EVALUATE. Those diagnostics, including candidate-fit information such as AIC, are not a substitute for comparing forecasts with held-out future observations. ML.EVALUATE can assess forecasting accuracy with new data for supported models. Inspect prediction intervals and empirical coverage across horizons, while recognizing that a nominal interval is not a guarantee for every future period or series.
A model selects a candidate with a favorable AIC but performs worse than a seasonal baseline on recent holdouts. Explain why the holdout comparison matters for deployment, and list one possible cause of the mismatch.
What is the strongest way to estimate how a weekly forecast will perform in production?
Keep exploring
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.