Demand Forecasting in E-commerce: A Measured Drop in Stock Costs

Problem

Ordering decisions at a large e-commerce operation rested on category managers’ intuition: some shelves overflowed and tied up capital while other SKUs ran out and lost sales. Forecasts existed — one-size-fits-all, in hand-updated spreadsheets.

Approach

• Forecasting at SKU × warehouse level, with seasonality, campaign calendar, price changes and holidays as features.
• A model mix by product lifecycle: gradient boosting for established SKUs, similarity-based cold start for new ones.
• Interval forecasts rather than point estimates; order suggestions derived from the interval against a service-level target.
• A controlled rollout: suggestions first ran in shadow mode against category decisions, then went live in stages.

Result

In period comparison, stock costs fell 30% and stock-out incidents dropped. In the operations manager’s words: “We make decisions with data now, not intuition.” The system was handed over with drift monitoring and a monthly retraining loop.