Problem
An online retailer with thousands of SKUs loses money two ways: stockouts (missed sales) and overstock (cash tied up). Most forecasting demos stop at a prediction. The useful question is what to reorder, when, and how much.
Method
A bake-off of Prophet, SARIMA and gradient boosting on Olist order data, evaluated on a time-ordered 60-day holdout, with no shuffling, because shuffling leaks the future into the past. SARIMA won. On top of the forecast sits a recommendation layer: reorder points at a 95% service level, a push list of products trending up, and markdown candidates trending down.
Evaluation
The recommendations were backtested against what actually happened: −82% stockout days and −84% lost sales, at +59% stock held. The last number is the honest cost. The model buys availability with inventory, and whether that is worth it depends on margin against holding cost.
The decision it drives
A weekly reorder list a buyer can act on, with the service-level trade-off stated rather than hidden.