Demand forecasting across 180 stores
A forecasting and replenishment system that cut stockouts while reducing held inventory.
- Fewer stockouts
- 31%
- Less held inventory
- 12%
- Stores live
- 180
The challenge
Replenishment ran on a spreadsheet model maintained by two people, using a single national average that ignored local seasonality. Stores compensated by over-ordering, which tied up capital and still left gaps on high-turnover lines.
Our approach
Built an evaluation harness against three years of history before building any model, so improvements could be measured rather than argued about.
Started with a well-tuned statistical baseline, which turned out to beat the spreadsheet by enough to justify shipping before any machine learning was involved.
Added per-store and per-category effects only where the evaluation showed a real gain, keeping the model explainable to the buying team.
Set up drift monitoring so quality degradation surfaces as an alert rather than as a stockout.
The outcome
Stockouts fell 31% while held inventory dropped 12% — the two usually move in opposite directions. The buying team reviews forecasts rather than producing them, and the original spreadsheet has been retired.
Working on something similar?
If this looks like your situation, we can probably tell you within one call whether the same approach applies.