Sample use case · Retail & supply chain
Demand forecasting & allocation
Weekly store-level forecasts that decide what to ship, where, before stockouts happen.
LightGBM quantile ensemble
gradient-boosted trees per store-SKU, Prophet-style seasonality & holiday features
Allocation engine
draft replenishment orders, reasoning attached, one-click approval
The workhorse is a gradient-boosted quantile ensemble (LightGBM) — fast to train per store-SKU, explainable, and honest about uncertainty. Where history is deep enough, a Temporal Fusion Transformer takes over the longest horizons.
The problem
A multi-brand consumer goods group allocated inventory to stores on gut feel and last year's numbers. Fast movers stocked out mid-season while slow stores sat on cash tied up in excess.
What we build
- A forecasting stack built on LightGBM quantile ensembles — gradient-boosted trees per store-SKU with Prophet-style seasonality, weather and promotion features — producing weekly P10/P50/P90 demand curves, with a Temporal Fusion Transformer for the longest horizons.
- An allocation engine that turns forecasts into draft replenishment orders — each with the reasoning attached, waiting for one-click approval.
- A scoreboard that grades every past week's allocation against what actually sold, so the model earns trust with evidence.
Proven before you commit
- Backtested on two years of sales history before touching a single live order.
- Piloted on one brand and eight stores; the planner approved every order manually until the hit rate held above target for six straight weeks.
The outcome
-38%
stockout events in pilot stores
-17%
excess inventory value
82%
forecast hit rate at handover
A representative engagement — numbers illustrate the shape of the result, not a named client reference.