Design Perishable Grocery Replenishment Ranges

Design a system that recommends a replenishment quantity range for perishable grocery products at each store. Products can spoil, demand varies by day, deliveries have lead times and case-pack constraints, and both stockouts and waste are costly. Explain the data, forecasting target, decision rule, offline and online evaluation, serving architecture, and monitoring. The output must be a range with an interpretable recommended point, not a single unexplained forecast. ### Clarifying Questions to Ask - Is the decision made per store-SKU-day, and how often can it be revised? - What are shelf life, lead time, minimum order, case pack, storage, and supplier constraints? - How are lost sales and censored demand represented when inventory reaches zero? - What relative costs or service-level targets define stockout versus waste trade-offs? ### Part 1 — Data and demand distribution Define training labels, features, leakage controls, and a probabilistic forecast over the relevant inventory horizon. #### What This Part Should Cover - Sales, inventory, stockout, waste, price, promotion, calendar, weather, lead-time, and substitution signals. - Separation of observed sales from latent demand under stockouts. - Rolling time splits and store-SKU cold-start behavior. - Quantiles or a full distribution rather than only a point estimate. ### Part 2 — Replenishment decision and range Convert the forecast into a feasible order recommendation and lower-upper range. #### What This Part Should Cover - Inventory position, demand through lead time and shelf life, and existing aging stock. - A cost or service-level quantile tied to underage and overage costs. - Case-pack, capacity, minimum-order, and expiry constraints. - An uncertainty range that has a clear operational interpretation. ### Part 3 — Evaluation, serving, and monitoring Describe a safe rollout and system architecture. #### What This Part Should Cover - Forecast calibration plus decision metrics such as waste, availability, lost sales, and margin. - Backtesting or simulation limitations and a randomized or phased field test. - Batch feature computation, decision service, overrides, fallbacks, and audit logs. - Drift, override, calibration, and constraint-violation monitoring. ```hint Optimize the decision, not forecast error alone Two models with similar prediction error can create different waste and stockout costs when their uncertainty is calibrated differently. ``` ### What a Strong Answer Covers - A coherent path from censored demand data to a probabilistic decision. - Explicit perishability and inventory constraints. - A range whose endpoints are defined, calibrated, and actionable. - Evaluation at both forecast and business-decision levels. ### Follow-up Questions - How would you model substitution when one item stocks out? - What fallback would you use for a newly launched product? - How would a supplier's variable lead time enter the range? - When should a store manager override the recommendation?