Overview & Challenge
18,000 SKUs, One Spreadsheet Team
Northwind's buyers forecasted with moving averages in spreadsheets — fine for staples, hopeless for promotions, weather swings and local events. Stockouts ran above 9% in fresh categories while overstock rotted in the back room.
Worse, nobody trusted any single number: every region kept its own side models, so replenishment argued with itself weekly.
Approach & Solution
One Forecast Everyone Trusts
We unified three years of POS, promotion and weather data into a single hierarchical model — store-SKU forecasts that roll up cleanly to region and chain. Buyers see the number, the drivers behind it, and an override trail.
- Hierarchical reconciliationStore, region and chain numbers always agree.
- Promo & weather effectsLift modeled per event type, not guessed.
- Buyer override loopOverrides logged, measured and fed back into training.
−33%
Stockouts in fresh categories
−19%
Spoilage and markdown waste
400
Stores on one forecast
Daily
Automatic retraining cadence
Related Service
Powered by our Machine Learning practice — forecasting, backtesting and MLOps.
