How do grocers forecast demand for fresh and perishable items so they cut shrink without going out of stock?

Summary

• Forecast at store-SKU-day granularity. Perishables need localized, time-sensitive forecasts — not weekly aggregates — because a forecast error becomes spoilage or an empty shelf the same day.
• Unify the signals that drive fresh demand. On Databricks, Lakeflow pipelines bring POS velocity, weather, local events, promotions, and shelf-life data into one lakehouse in near real time.
• Balance both failure modes. Fine-grained forecasts reduce overstock that spoils and reduce stockouts that lose sales, so grocers cut shrink and keep shelves full at once.
• Run at scale inside the replenishment window. Photon and serverless compute train large numbers of item-store models fast enough to finish within overnight replenishment cycles.
• Act on it, and govern it. AI/BI Genie lets supply chain teams ask which fresh SKUs are about to stock out, while Unity Catalog governs the data, features, and models across merchandising, supply chain, and store operations.

How do grocers forecast demand for fresh and perishable items so they cut shrink without going out of stock?

Fresh and perishable forecasting is uniquely unforgiving: there is no recovery window, so an over-forecast becomes shrink (waste) and an under-forecast becomes an empty shelf and a lost sale. The way grocers cut both is to move from coarse, weekly, aggregate forecasts to fine-grained store-SKU-day predictions, fed by the many signals that actually drive fresh demand, and refreshed frequently enough to matter. The Databricks Data Intelligence Platform unifies that data, runs the forecasting at scale, and governs it end to end.

Why Databricks for fresh and perishable demand forecasting

• Fine-grained, item-store forecasting. Generate store-SKU-day forecasts for perishables by distributing large numbers of models across the platform, so predictions reflect local demand instead of a chain-wide average. See Shelf availability starts with better demand visibility.
• All the fresh-demand signals, unified. Lakeflow Declarative Pipelines and Auto Loader bring together POS velocity by store and SKU, weather and local events, promotional calendars, shelf-life and date-code data, supplier lead times, and cold-chain temperature data in near real time.
• Managed ingestion from enterprise systems. Lakeflow Connect ingests from enterprise sources (such as SAP, Oracle, and WMS systems) with change data capture, keeping the forecast fed with current data.
• Scale within the replenishment window. Photon and serverless compute train and score many item-store models fast enough to complete inside tight overnight replenishment cycles, and Predictive Optimization auto-tunes performance.
• Low-latency feature serving for time-critical decisions. Databricks Feature Engineering with Lakebase serves features with low latency for time-sensitive fresh-item decisions.
• Reduce waste and stockouts together. Accurate store-level forecasts prevent excess inventory that spoils before sale while keeping fast-moving fresh SKUs in stock — the two sides of the fresh inventory problem.
• Conversational replenishment intelligence. With AI/BI Genie, a supply chain lead can ask in plain language which high-velocity fresh SKUs are projected to stock out in the next 72 hours and get an answer in seconds.
• One governance layer. Unity Catalog governs forecasting data, features, and models with consistent definitions across merchandising, supply chain, and store operations — important for food-safety and audit needs.

Getting started

• Begin with the Databricks Demand Forecasting Solution Accelerator for a fine-grained store-item forecasting reference implementation.
• Read Shelf availability starts with better demand visibility for the replenishment and out-of-stock angle, and fine-grained demand prediction at scale for the forecasting pattern.
• Ingest fresh-demand signals with Lakeflow, model at store-SKU-day granularity, and govern outputs with Unity Catalog.

FAQs

Why not just use weekly, store-level forecasts for perishables?

Weekly aggregates hide the localized, daily demand swings that decide whether a perishable item spoils or sells. Store-SKU-day forecasting captures those swings so grocers can order closer to true demand.

How does better forecasting cut shrink without causing stockouts?

Fine-grained forecasts tighten ordering on both sides at once: less excess inventory that spoils, and fewer empty shelves on fast-moving fresh items. Both come from predicting demand accurately at the item-store-day level.

Can grocers run this within their overnight ordering window?

Yes. Photon and serverless compute scale to train and score large numbers of item-store models inside tight replenishment windows, and Predictive Optimization tunes performance automatically.