Turning 40 hour demand forecasts into 3 minute predictions

The Challenge

The planning team spent roughly 40 hours each week building demand forecasts by hand across 11 facilities. Stockouts were hitting 130 SKU-locations per week, each one costing $140 in client penalties, adding up to $18,200 a week. The operations lead knew exactly what system they needed, but no one on the team had the ML experience to build it.

What We Built

Three engineers joined the planning team within six days, sat in on daily standups, and started building. The system pulls from point-of-sale feeds, seasonal patterns, and transit times to forecast SKU-level demand at all 11 warehouses, 72 hours out. A second module compares current stock positions against those forecasts and generates transfer orders between facilities before gaps open. Full forecast cycle: under 3 minutes, refreshed every 4 hours.facilities before anything runs short. A full forecast cycle runs in under 3 minutes and refreshes every 4 hours.

Results

Stockouts went from 130 per week to fewer than 15. Penalty charges fell from $18,200/week to under $2,100, saving $64,000 in the first month. Thirty-five analyst hours per week moved from manual forecasting to carrier negotiations and new client onboarding. The full engagement, three engineers over 9 weeks, cost less than one month of penalties.