Journal / Kirill Naumov
AI Demand Forecasting in Retail: The Forecast Is Only the First Link
Connect retail demand predictions to replenishment, promotions, availability and human review without mistaking model accuracy for customer value.
By Kirill Naumov ·
A forecast does not put the product on the shelf. It can help a retailer anticipate demand, but the result still depends on whether the signal reaches purchasing, allocation, replenishment and store execution in time. A model can be more accurate and customers can still face an empty shelf if an order arrives late, stock is allocated elsewhere or the inventory record is wrong.
Connect the forecast to an operational decision
Start with the decision to improve: how much to order, where to send stock, when to replenish or which products need human review. Set the unit and horizon clearly—a SKU by store for the next few days is a different problem from category demand for a season. Name the team that can act on the prediction and the system where the decision happens.
Map the handoffs. Which forecasts become order suggestions? Who can override them? How are supplier lead times, minimum quantities, shelf capacity, warehouse stock and delivery calendars handled? If the forecast stops at a dashboard, the hard part is still waiting for someone else.
Sales history has blind spots
Past sales are not the same as unconstrained demand. When an item is unavailable, sales can fall because customers could not buy it. Promotions, price changes, weather, holidays, local events, substitutions and new products also change the pattern. Keep those events visible and label exceptional periods rather than asking the model to treat every week as interchangeable.
Compare the existing approach with the proposed model on periods where the answers are known, then review results by product type, store and forecast horizon. Track bias as well as average error: being consistently short on one category can be more damaging than a larger but balanced miss elsewhere.
Learn from retail examples without turning them into promises
Google Cloud’s published [Cainz case study](https://cloud.google.com/customers/cainz) describes a Japanese home-improvement retailer that expanded demand forecasting across 209 stores and changed its architecture as the scope grew. The company says its earlier moving-average approach struggled with seasonal products and short-term trends. This illustrates how scale and product mix affect model choice; it does not guarantee the same outcome elsewhere.
Google Cloud also publishes [OTTO’s forecasting case](https://cloud.google.com/customers/otto) and [Tchibo’s case](https://cloud.google.com/customers/tchibogmbh). These are vendor-published customer stories, so read their results in that context. The transferable themes are broader than any one tool: combine product, sales, promotion and logistics information; keep analysts involved; connect predictions to planning decisions.
Measure availability, not just model error
Track forecast error and bias, but also outcomes that matter to the business: on-shelf availability, defensible lost-sales estimates, waste or markdowns, inventory held, urgent transfers and time spent correcting suggestions. Watch the trade-offs by category. Cutting excess stock in one range while increasing missed sales in another is not a single clean efficiency win.
Agree the comparison before the pilot. Use similar stores or periods, account for promotions and seasonal changes, and document which recommendations were accepted or overridden. A better forecast shows that an input improved; it does not by itself prove that customers found the product or the business earned more.
Make exceptions easy to see
Managers need to know when a recommendation reflects a normal pattern and when a new product, local event or missing data makes it fragile. Show a short reason, the relevant signals and a way to flag a correction. Repeated overrides may reveal a local constraint the model does not know—or a data problem upstream.
A sound first project is bounded: one category, one planning decision, a working handoff and a measurement plan. Expand only when the prediction changes an action, the action improves availability or another agreed outcome, and the team knows what to do when the model is uncertain.