Forecast error down 40%, stockouts down to zero

Demand for the manufacturer's specialty-grade products swings with individual industrial clients' own production schedules — a single large customer pausing one production line could look, in the sales data, identical to genuine falling demand, and the forecast had no way to tell the difference.

A single static safety-stock formula — average demand multiplied by a fixed lead-time buffer — was applied uniformly across a catalogue running from high-volume commodity-grade material to low-volume specialty grades with long, variable supplier lead times. The same rule was wrong in both directions at once.

The result was two problems happening simultaneously in different parts of the same warehouse: specialty-grade stockouts that held up a client’s own production line, and commodity-grade overstock quietly tying up warehouse space and working capital that could have funded something else.

Before any model was built, a data-quality pass across two years of sales history caught a problem common to distribution forecasting generally: a handful of one-time liquidation sales and duplicate SKU codes were quietly teaching the existing forecast to expect demand that would never recur.

A probabilistic forecasting model replaced the single-point average forecast with a full range of likely demand per SKU, feeding directly into a dynamic safety-stock calculation instead of one static buffer rule applied to the entire catalogue.

The model draws on the ERP's sales and shipment history, open purchase orders, and supplier lead-time records together rather than sales history alone, so a supplier delay already known to the business is priced into the reorder point before it becomes a shortage.

A weekly exception report flags only the SKUs where actual demand is drifting outside the forecast’s confidence band, so planners investigate a handful of genuine anomalies each week instead of re-reviewing the entire catalogue by hand.

Forecast error fell by 40% within two quarters of go-live, the plant has not recorded a specialty-grade stockout since, and the working capital previously tied up in excess commodity-grade stock dropped by a quarter — freed up for the business to redeploy elsewhere.