Reduce Fresh-Food Waste with an End-to-End ML Approach

A supermarket has rising waste in fresh-food categories while some stores still experience stockouts. Design an end-to-end machine learning approach, from data collection and cleaning through model selection, evaluation, and deployment. Explain how you would determine whether the main opportunity is better demand prediction, better ordering policy, or an operational process change. ### Constraints & Assumptions - Food has finite shelf life and inventory is tracked by product, store, and day. - Recorded sales may understate demand during stockouts. - Waste can result from forecasting, ordering constraints, handling, or data errors. - The system should support a business decision, not only predict a label. ### Clarifying Questions to Ask - Which categories, stores, and waste reasons drive the increase? - How accurate are on-hand inventory, expiry, markdown, and disposal records? - Who makes the order, when, and under which supplier and case-pack constraints? - What are the costs of waste, stockout, labor, and lost customer trust? ```hint Diagnose the funnel before modeling Decompose purchased units into sold, remaining, transferred, marked down, and discarded units; an inventory reconciliation gap may be more important than model choice. ``` ### What a Strong Answer Covers - Data lineage and inventory-flow reconciliation. - Latent demand, shelf life, promotions, lead time, and product substitution. - Time-aware baselines and probabilistic or quantile forecasting. - A constrained order decision with asymmetric costs. - Offline simulation limits, field testing, monitoring, and human overrides. ### Follow-up Questions - What would convince you not to deploy a model? - How would you distinguish spoilage from unrecorded shrink? - Which metrics prevent waste reduction from simply causing more stockouts? - How would a sudden supplier delay affect recommendations already issued?