Concept build: a demonstration project, not client work
A demand forecasting platform for a distributor. It forecasts weekly demand per product and warehouse, then drafts the purchase orders.
Move across the chart to read any week. Further out, the interval widens: the model is honest about what it doesn't know.
Try it
Move the horizon and change the confidence level. The further ahead the forecast looks, the wider its range becomes: the buyer sees how sure the model is, not only its best guess.
A distributor stocks thousands of products across several warehouses and plans purchasing with spreadsheets and gut feel. Fast sellers run out while slow lines sit on shelves, tying up cash. Seasons, promotions and supplier lead times all matter, and no one can track them by hand. When a senior buyer retires, their instincts retire too.
In Cratecast, sales, open orders, promotions and lead times load into PostgreSQL each night. A Python pipeline trains XGBoost models on product and warehouse features. Each forecast comes as a range: low, likely and high. The reorder step sizes safety stock from that spread and drafts purchase orders. Buyers adjust and approve them, and every override is saved with a reason.
XGBoost has to beat a seasonal naive forecast on held-out weeks. Complexity has to earn its place.
Models are scored by replaying past seasons week by week, using only data available at the time.
Every adjusted order carries a reason code, reviewed at retraining. Buyer know-how ends up in the next model.
How a request moves through the system.
01Nightly jobs load sales history, stock and promotions
02Features are built per product, warehouse and week
03Models retrain and must pass a backtest first
04Forecast ranges set safety stock and draft orders
05Buyers adjust, approve and note their reasons
Next concept build
FieldkeelField servicesDescribe your version of the problem. We reply with questions and a proposed approach.