ML Demand Forecasting Eliminating $4M in Annual Stockouts

A national retail group with 340 stores was losing $4M annually to stockouts driven by spreadsheet-based demand planning that couldn't incorporate weather, trends, or competitor signals. We deployed ML forecasting models trained on 5 years of sales data plus external signals — improving forecast accuracy from 61% to 94% and connecting predictions directly to ERP replenishment workflows.

The retailer operates 340 stores across 8 regions, with a product catalogue of 15,000 SKUs across apparel, homewares, and electronics. Demand planning had been handled by a central team of 6 planners using Excel-based models. The models performed reasonably well for steady-state products but consistently failed on seasonal items, new launches, and products sensitive to external factors like weather.

Stockouts were concentrated in three categories: seasonal products (where demand spikes were consistently underestimated), new product launches (where no historical data existed), and regionally variable products (where national averages masked local demand patterns). The planning team knew the models were insufficient but lacked the tools to improve them.

We built a multi-model forecasting system: a base ARIMA model for stable SKUs, a gradient boosting model incorporating external signals for volatile SKUs, and a cold-start model using category analogues for new product launches. All three models feed a unified ensemble layer that selects the best forecast per SKU. Predictions are published directly to the client's SAP ERP via API, triggering automated replenishment orders.

Twelve months post-deployment, forecast accuracy improved from 61% to 94% on the SKUs covered by the system. Stockout-related revenue loss fell by $4M. The planning team's exception management time fell from 70% to 25% of capacity, allowing them to focus on strategic range planning.