Demand forecasting and replenishment

• Demand is forecast every night for each product, store and channel, using sales history, promotions, prices, weather forecasts, holidays and local events.
• Forecasts come with a range, not just a number, so a buyer can see how sure the model is, and order more cautiously where it isn't.
• Suggested orders and store allocations are calculated from the forecast, stock in stores and warehouses, lead times, pack sizes, shelf space and shelf life.
• Buyers and merchandisers review and approve the suggestions, working through an exceptions list instead of every line.
• Fresh lines get earlier warnings: short-dated stock is flagged for markdown or redistribution before its use-by date, within rules the store teams set.
• Every forecast, suggestion, approval and override is logged with the reason, so the team can see what the model got wrong and why.

• GCAGroceries Supply Code of Practice: forecasting
  →

  This applies only to the 14 grocers the CMA has designated. A retailer must compensate a supplier for its own forecasting errors unless it prepared the forecast in good faith, with due care and after consulting the supplier. Share forecasts with suppliers, capture their input, and log the model version and data behind each one so due care can be shown, with extra care for promotions and seasonal lines.

• FSAFood safety and use-by dates
  →

  Food can't be sold or given away after its use-by date. Markdown and redistribution suggestions use the date on the stock record and stop before use-by. Store teams decide what is marked down, redistributed or thrown away.

• DefraFood waste measurement and Simpler Recycling
  →

  Since 31 March 2025, businesses in England with ten or more employees have to separate food waste for collection. Public reporting of food waste by large businesses is voluntary at the time of writing, and the government has said it is reconsidering a mandate. Measure waste by product and store in a consistent way now, so the numbers stand up either way.

• ICOUK GDPR
  →

  Forecasts work on aggregated sales, so they usually involve no personal data. If loyalty-card data is used to understand demand, it is aggregated before it reaches the model, with a DPIA and a clear retention period, and providers are contractually barred from training their models on it.

• CommercialSupplier agreements and EDI
  →

  Suggested orders become purchase orders only after a buyer approves them, and are sent through the existing EDI or supplier portal. The AI never changes supplier terms, cancels an order or commits spend on its own.

On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →

• Buyers and merchandisers involved in setting the rules: which lines need review, what counts as an exception, and how much they can change without a reason.
• Override reasons captured in a short list (local event, supplier issue, known error), so the model learns from them and the team can see where it is weak.
• A weekly accuracy review by category with the commercial team, not just the data team.
• Promotions agreed early enough to forecast them. Late promotions are flagged as higher risk.
• For grocers covered by GSCOP: forecasts shared with suppliers, their input recorded, and the evidence kept.
• A named owner for each model, with change control, and a rule that nothing changes in the four weeks before Christmas.

1. 3–4 weeks

   Discovery

   Check two to three years of sales, stock and promotion data, measure today's forecast error and waste, and pick a pilot category with fresh and ambient lines.

2. Gate · must be true to passData quality understood, baseline agreed, pilot category and stores chosen, buyers involved.

3. 10–12 weeks

   Pilot

   Forecasts and suggestions for one category in 20–30 stores, in shadow mode first: buyers order as normal, and the two are compared week by week.

4. Gate · must be true to passError and bias better than today's method, no fall in availability, buyers willing to use it.

5. 6–8 weeks

   Production

   Suggestions go live for the pilot category across all stores and channels, with monitoring, support and a rollback plan.

6. Gate · must be true to passNightly run reliable, monitoring and support in place, buyers trained, rollback rehearsed.

7. Ongoing

   Scale

   Add categories one at a time, then use the forecast for warehouse labour planning and supplier collaboration.

8. Gate · must be true to passEach new category passes the same pilot gate.

R ResponsibleA AccountableC ConsultedI Informed· not involved

Where I'd fit

1. 1

   Goes wrong

   The model is trained on sales, not demand, so every out-of-stock day looks like low demand and it orders less next time.

   How the design avoids it

   Correct for days with no stock or short shelves before training, using stock records from the EPOS and WMS.

2. 2

   Goes wrong

   Promotions are added to the system late, so the forecast never sees them.

   How the design avoids it

   A promotions calendar fed into the run, with late changes flagged as a risk on the exceptions list.

3. 3

   Goes wrong

   Buyers don't trust it and re-key their own orders, so nothing improves.

   How the design avoids it

   Shadow mode first, accuracy shared openly every week, and buyers setting the exception rules.

4. 4

   Goes wrong

   A missing store feed shows as zero sales, and the next order is cut.

   How the design avoids it

   Data checks before every run, and missing stores forecast from their last good data.

5. 5

   Goes wrong

   The data science is good, but the orders never reach the ERP properly.

   How the design avoids it

   Plan the ERP and EDI integration in discovery, not after the model is built.

1. 2026–2027
   The Groceries Code Adjudicator's focus on forecasting

   The GCA's 2025/26 annual report keeps forecasting among its priorities, and it raised forecasting with retailers' Code Compliance Officers in 2026. Sponsorship of the GCA moved to Defra on 1 July 2026. If you are a designated grocer, build supplier input and evidence of due care into the process now.

2. Open
   A decision on mandatory food waste reporting

   Defra consulted in 2022 and has said it is reconsidering a mandate for large food businesses. Nothing is confirmed. Measure waste consistently by product and store now, so reporting is a report rather than a project.

3. 1 April 2029
   Mandatory B2B e-invoicing on a Peppol model

   HMRC published the roadmap in October 2026. Supplier invoices will need to match purchase orders automatically. Keep purchase order data clean and structured, and plan Peppol alongside your existing EDI.

4. Late 2026–2028
   The Cyber Security and Resilience Bill

   Medium and large managed service providers come into regulation, with faster incident reporting. If your data platform or forecasting service is run by a supplier, expect new duties for them and updated contracts for you.