'/%3e%3ccircle%20cx='98'%20cy='41'%20r='9'%20fill='%2322D3EE'/%3e%3c/svg%3e){kind=small}
'/%3e%3cellipse%20cx='520'%20cy='260'%20rx='360'%20ry='132'%20stroke='%238B93B0'%20stroke-opacity='0.18'%20stroke-width='1'%20transform='rotate(-16%20520%20260)'/%3e%3ccircle%20cx='520'%20cy='260'%20r='46'%20fill='%23111A38'%20stroke='%2322D3EE'%20stroke-opacity='0.35'/%3e%3ccircle%20cx='920'%20cy='80'%20r='6'%20fill='%2322D3EE'/%3e%3c/svg%3e)
Demand forecasts that buyers actually trust
Challenge
Replenishment ran on a legacy moving average that under-called promotions and seasonal peaks, so category buyers overrode it by hand and stopped trusting the number.
Approach
We built store- and SKU-level forecasts with gradient-boosted models (LightGBM) over lagged demand, price, promotion, and calendar features, and produced quantile outputs (P10/P50/P90) so buyers could set safety stock against a service-level target rather than a single point estimate. Rolling-origin backtesting measured accuracy the way the business feels it — one period ahead, repeatedly — before anything reached replenishment.
Architecture
Feature store (sales · price · promo · calendar) → LightGBM quantile models → rolling-origin backtest → replenishment feed
Python · LightGBM · Pandas · Airflow — quantile loss (P10/P50/P90), rolling-origin backtesting
Results
- Roughly 18% lower forecast error than the moving-average baseline*
- Quantile outputs let buyers set safety stock by service-level target, not guesswork
- Backtested over 24 months of history before rollout
* Illustrative figure from a representative engagement, shown to convey typical scope. Replaced with client-verified numbers before publication.