Forecasting demand for 2,400 SKUs

The situation

Carter Foods runs three plants and ships 2,400 active SKUs. Planning worked from last year’s volumes plus the planner’s judgement, in a spreadsheet rebuilt each Monday.

It was not a bad process. It was an undocumented one that depended on two people, and it could not react to a promotion the commercial team had agreed on Friday.

Constraints

Planners had to stay in control. An earlier attempt at automated replenishment had been abandoned because overriding it was slower than ignoring it, so the override path mattered more than the model.

What we did

We built one demand history across retail, foodservice and wholesale, with returns and promotional lift handled consistently for the first time. Forecasting runs per SKU and region and retrains weekly.

The override sits in the planning screen the team already used. A planner changes a number, the reason is recorded, and that reason feeds the next evaluation round.

What surprised everyone

The largest single gain came from fixing how promotions were recorded, not from the model. Three months of history had been attributing lift to the wrong week.

Results

Measure	Before	After
Mean absolute percentage error	31%	12%
Stockouts per quarter	184	96
Planning hours per week	60	22
Models in production	0	3