Forecast Food Stocking Needs Under Waste and Stockout Costs

A retailer wants to predict how much food to stock at each location for the next replenishment cycle. Too little inventory causes stockouts and lost sales; too much creates holding cost and spoilage. Walk through the problem from business definition to a modeling and evaluation plan. Explain the prediction target, features, treatment of censored demand, model choice, validation design, and how forecasts become stocking decisions. ### Constraints & Assumptions - The decision is made separately for product-location pairs. - Historical sales can be lower than true demand when inventory is unavailable. - Promotions, prices, holidays, and replenishment lead times can change. - The goal is a decision with asymmetric costs, not the lowest generic prediction error. ### Clarifying Questions to Ask - What is the order cadence, supplier lead time, shelf life, and minimum order quantity? - How are stockouts, waste, substitutions, and lost sales observed? - Which products are stable staples versus intermittent or newly launched items? - What service level or financial cost defines understocking and overstocking? ```hint Match the horizon to the decision Forecast demand over the period that new inventory must cover, then subtract usable inventory already available or arriving. ``` ### What a Strong Answer Covers - A precise grain, horizon, label, and point-in-time feature set. - Demand censoring, seasonality, promotions, and cold starts. - Time-based validation and baselines at relevant product-location segments. - Probabilistic forecasts or quantiles tied to asymmetric decision costs. - Waste, availability, margin, and lost-sales evaluation rather than RMSE alone. ### Follow-up Questions - How would you estimate demand when a product was stocked out for half the day? - What baseline would you use for a new store? - How would substitution between similar products affect labels and evaluation? - Why might a forecast with lower RMSE produce worse stocking decisions?