The problem
Forecasts were assembled monthly in a chain of spreadsheets owned by one analyst. Accuracy was unmeasured, and nobody could explain a number to a store manager who disagreed with it.
What we did
- Replaced the spreadsheet chain with contracted, tested pipelines
- Established a baseline model first, so every later change had something to beat
- Backtested against three years of history before anything went live
- Shipped explanations alongside forecasts so store managers could challenge them
