Business problem

A national retailer was over-ordering slow movers and stocking out of best sellers, tying up working capital while missing sales during promotional peaks.

Dataset

4 years of daily sales for 1,200 SKUs across 48 stores, joined with promotion calendars, weather and public holiday data (~26M rows).

Methodology

• Built a reconciled hierarchy at SKU, store and region level.
• Engineered lag, rolling-window, promotion and holiday features.
• Trained gradient-boosted models with time-series cross-validation.
• Deployed a weekly batch pipeline with drift monitoring.

Key findings

• Promotion depth explained 38% of forecast variance for top SKUs.
• Weather features mattered only for seasonal categories.
• MAPE improved from 21% to 11.4% across the portfolio.

Recommendations

• Shift replenishment to a weekly forecast-driven cycle.
• Cap promotion depth on SKUs with high cannibalisation scores.
• Retrain quarterly and monitor drift on the top 200 SKUs.