The challenge
The CFO had identified a structural commercial problem: the operator was simultaneously over-ordering to protect the 15-minute promise on perishables and writing off the resulting expired stock at month-end. The weekly forecast cycle that drove replenishment decisions was the root cause — it was structurally incapable of responding to the hour-by-hour demand variation that a fast-fulfilment perishables business operates on. Both numbers needed to move together, not traded against each other.
Discovery
Six weeks of dark-store observation identified the specific inputs the existing forecast was missing. Demand varied hour-by-hour in response to signals the forecast cycle could not incorporate: weather patterns, hyper-local events, and marketing pushes the growth team had launched that same morning. The forecast model itself was well-constructed; it was responding to the wrong inputs — historical patterns rather than the real-time signals that actually drove demand.
Our approach
Forecast model redesign: built an hourly per-SKU per-store demand forecast model ingesting weather, local events, marketing actions, and the operator's promotional calendar
Replenishment programme redesign: replaced weekly replenishment runs with a continuous reorder programme respecting supplier minimum order quantities and lead times
Shared operating view: built a waste-and-stockout console at city operations level — the same data the CFO and the city operations manager see and act on
Continuous improvement programme: established a daily backtest quantifying forecast error per SKU and feeding the worst-performing categories into the next training cycle
The Outcome
The engagement delivered an hourly forecast model coupled with a continuous replenishment programme and a single waste-and-stockout view shared across operations and finance. The operator stopped managing stockouts and waste as competing objectives and started moving both together.
The results
“We had been trading stockouts against waste for three years. The hourly forecast was the first thing that let us move both numbers at once.”
