Store-level demand forecasting for a grocery chain, with honest backtesting

FreshCart runs 38 supermarkets supplied from one distribution centre. Store managers order fresh and packaged goods by gut feel, two weeks ahead. The result: about 4.5% of fresh produce is thrown away, and on busy weekends the top 200 products run out in about 1 store in 6. Finance estimates the waste alone at 2.3% of fresh revenue.

The supply-chain head wants a forecast for each store and product family, 14 days ahead, refreshed weekly. He has been burned before: a consultant showed a model with "97% accuracy" that fell apart in its first month because it had been tested on the same period it was trained on.

You will use a public retail dataset with daily store sales, promotions, holidays and oil prices (the Kaggle "Store Sales — Time Series Forecasting" data is a good fit), treat its stores as FreshCart's, and build: - a seasonal-naive baseline (same day last week); - a gradient-boosted model with calendar, promotion and lag features; - a rolling-origin backtest with at least 6 folds of 14 days each; - a simple order rule that turns the forecast into a suggested order with a safety buffer.

The target is a weighted MAPE at least 15% lower than the baseline across the backtest, without any fold being worse than the baseline by more than 5%. Done looks like a re-runnable pipeline, a backtest report with fold-by-fold numbers, and a one-page note for store managers explaining how to read the forecast.