Practice — Forecasting at Scale: Global, Probabilistic & Hierarchical Models (6 questions)

A grocery chain wants one daily demand forecast for each of 400,000 SKU-store pairs, 14 days ahead, refreshed every night. You have two years of daily sales, a planned-promotion table (each promo is entered about 3 weeks ahead), a price table, a product catalogue (category, brand, pack size), a store table (format, region) and a holiday calendar. About 12% of the pairs have fewer than 60 days of history, and sales per pair range from 0.2 to 900 units a day.

1. Why one global model rather than 400,000 per-series models? Name two situations where you would still keep a local model.
2. List the feature families you would build and say which ones need extra care to avoid leakage. Give the minimum lag you allow, and why.
3. How do you handle the 900-versus-0.2 scale spread? Say what you divide by, over which window, and what you check afterwards.
4. A colleague wants to add the SKU-store id as a categorical feature. Do you? What do you do for a pair with no history?

You have to produce 28-day-ahead daily forecasts for 150,000 series with LightGBM. The nightly batch has 3 hours on a 32-core machine. A teammate proposes a recursive one-step model; another proposes 28 direct models; a third proposes one model with every lag at least 28 days back.

1. Describe what each strategy trains on and what it feeds into the features at test time.
2. Give the main failure mode of each.
3. Which would you try first, and what experiment decides? What would change your mind?
4. How do you report horizon-wise accuracy honestly, and what pitfall is hiding in a plot of error against horizon?