Case study
Demand forecasting case study: weekly supply-planning forecasts for a packaged food maker
ResultWeekly product and distribution-centre forecasts were connected to supply planning, with comparable accuracy reports and planner overrides.
- Client
- A US packaged food maker (anonymised)
- Industry
- Consumer goods (packaged food)
- Service
- AI demand forecasting: a backtest, shadow forecasting and live forecasts by category
- Type
- Weekly forecasts by product and distribution centre, fed into supply planning
- Status
- Delivered
We built a demand forecasting workflow for a packaged food maker selling through retail chains and distributors. The work connected order history, prices and promotion information to weekly forecasts by product and distribution centre. Forecasts feed supply planning, while planners retain the ability to review and override them.
The starting point
Planners brought together forecasts in spreadsheets using historical shipments and account managers' promotion estimates. Promotion information was spread across spreadsheets and email, making a consistent view difficult to maintain. Stock shortages could make shipments look like weaker demand, and new products had no sales history of their own.
What we delivered
We reconstructed the promotion calendar and aligned it with order, shipment and price history. A shared forecasting model uses calendar, price and promotion features across product series. New-product forecasts draw on comparable launches. Product and distribution-centre forecasts roll up to category and company totals, with forecast ranges passed into supply planning.
Rollout and handover
We backtested against both the planners' forecasts and a seasonal baseline using the same forecast horizons. Shadow forecasting let the planning team compare outputs before enabling categories in the live workflow. The operating process includes accuracy reporting, model comparison and checks on additional input signals.
Where people remain in control
| Term | Agreed definition |
|---|---|
| Metric | WAPE, with bias beside it so a steady lean to over-forecasting shows |
| Grain | Product by distribution centre by week |
| Horizons | The production lead times agreed with the planning team |
| Baselines | The same week last year first, then the planners' forecast |
| Short-shipped weeks | Left out of the error and listed, since a week the company couldn't ship in full understates what customers wanted |
| Promotion weeks | Scored on their own, so a strong promotion forecast can't cover for weak ordinary weeks |
| Planner overrides | Kept, and scored against the forecast each one replaced |
| Switching a category | Once the model beats the planners' forecast for the agreed number of cycles in a row, signed off by the demand planning lead |
Project outcome
The delivery connected forecast generation, evaluation and supply planning in one repeatable process. Planners can inspect the model output and record overrides. Accuracy is reviewed by category and horizon, so promotion effects and systematic bias can be examined separately from overall error.
How performance is evaluated
| Measure | How it is measured | Comparison point |
|---|---|---|
| WAPE at two, four and eight weeks, by category | Absolute errors summed, over actual units summed, on the same weeks | The planners' forecast, and the same week last year |
| Bias | Signed errors summed, over actual units summed | The planners' forecast |
| Error in promotion weeks | WAPE on promoted weeks only | The planners' forecast on the same weeks |
| What overrides add | The change in WAPE from each override, by planner and category | From the first shadow week |
| Short shipments | Order lines not shipped in full, in the categories switched | The year before each switch |
| Stock written off past its date | Value written off, in the categories switched | The year before each switch |
Technologies and tools
- Order, shipment and price history from the ERP
- The rebuilt promotion calendar
- Gradient-boosted trees across every series
- A weekly forecast file into supply planning
- An accuracy report every cycle
- The company's own cloud account
More case studies
All case studies- Case studyLapse risk and customer lifetime value for a specialty retailerLoyalty-member lapse scores, value bands and explanations were delivered to the campaign tool, with offers controlled by the CRM team.
- Case studyField stations that identify birds by sound, with set-up cut to under 10 minutesAverage station set-up: from nearly 25 to 30 minutes to under 10
Bring us the forecast your planning team relies on
Tell us the level you order at, your lead times and how today's forecast is prepared. An engineer replies within two business days with how a backtest would compare it with a simple seasonal forecast and what history it would need.