Demand forecasting and inventory optimisation across 320 stores

Context

A retail chain with 320 stores across 28 provinces and roughly 12,000 active SKUs. Ordering was decided store by store from experience, producing two problems at once: slow-moving stock in some lines and stockouts in the best sellers.

The challenge

Sales data lived in the POS system but could not be joined with inventory or the promotion calendar. Nobody measured forecast error because there had never been a formal forecast.

What BigAI delivered

The first phase consolidated POS, warehouse and promotion data through BigAI DataHub. A two-tier forecasting model followed: a base model for trend and seasonality, plus a second model handling promotional lift.

Crucially, the model does not override store managers. It proposes an order with a confidence range; managers can adjust, and every adjustment is logged and fed back into retraining.

Results

After three months in production, forecast error fell from 31% to 12%. What mattered more to the board was the working capital released from slow-moving stock.

Results

Swipe to see the full table

What convinced the board was not model accuracy. It was the working capital released after the first quarter.

Technology used