Forecast demand on sample data

Scenario

A fictional retailer holds two years of weekly orders for 40 product lines in the sample set. Leadership asks: how much will each line sell over the next eight weeks, and what drives the number? One product line has a promotion in the sample window that the data barely hints at — your job includes noticing what the data cannot know.

Image: sample order history chart with the promotion window marked.

Definition of done

• One numeric forecast per product line over the eight-week horizon.
• Drivers readable per line, evidence traceable per figure.
• The promotion-affected line flagged as uncertain rather than confidently wrong.

Steps

1. Connect the sample order source and confirm two full years synced — seasonality needs history, and a partial load silently becomes a wrong forecast.
2. Shape the slice. Product and weekly-sales entities, sold-over-time relationships, identifiers on SKU plus week. Materialize and confirm no gaps: missing weeks are the most common demand-forecast killer.

Image: materialized weekly series with one gap week flagged before it can poison the forecast.

3. Describe the decision. "Forecast units per product line for the next eight weeks with drivers." Check the proposal's horizon and audience before anything runs.
4. Dry-run and interrogate the drivers. Seasonality, trend, recent level — do they make business sense per line? Find the promotion line: inputs hint at it weakly, and the honest output is wide uncertainty, not a sharp wrong number.

Image: driver breakdown per line with the promotion line's weak signals and wide band called out.

5. Commit and defend one line end to end. Walk a single figure from forecast through fact set and mapping to source weeks.
6. Record the outcome when the window closes. The sample ships actuals — compare, score, and note where uncertainty was the right call. Measured accuracy beats assumed accuracy, even on samples.

Image: forecast versus actuals for three lines, including the promotion line with its uncertainty band.

What good looks like

• Stable lines forecast tightly, volatile lines honestly wide.
• The promotion line flagged uncertain with recorded reasons.
• Actuals compared, accuracy stated, lessons captured as feedback.

Next

• Same flow, your data: Get your first forecast.
• Kinds, modes, budgets: Forecasts.
• Closing loops as habit: Forecasts overview.