Key takeaways
- Forecast where the order is placed: per product, per store, per day, over the lead time plus the time to the next order.
- Sales are not demand. Days with an empty shelf, negative stock records and markdowns all hide part of it.
- At store level, simple methods such as exponential smoothing hold their own against machine learning; slow sellers need a method built for intermittent demand.
- Measure the error in units over the order's horizon, with bias beside it, and compare it with a naive rule. Do not use MAPE on slow sellers.
- Promotions, holidays and weather move demand more than any method choice, so put them in as known events.
Demand forecasting in a store chain is less about predicting the year and more about predicting tomorrow, product by product and store by store. The order a store sends this afternoon is a forecast of what its customers will buy before the next delivery, whether anyone wrote it down or not. Too high, and cash sits on the shelf or goes in the bin; too low, and the shelf is empty when the customer arrives. This guide covers what to forecast, which methods hold up at that level, and how to tell whether the forecast is any good.
1. What a chain should forecast
A forecast is useful at the level where a decision is made. For store replenishment that is one product in one store, day by day, because the order, the delivery and the empty shelf all happen there. A chain total is fine for the budget, but it averages away exactly what the order needs: one store selling out while another holds three weeks of the same line.
Large grocers forecast at this level as a matter of course. Albert Heijn forecasts demand for every product in every store for the next 50 days10, and Tesco says it uses AI to optimise store orders and reduction levels11. Sainsbury's moved its food forecasting to a machine-learning platform and, together with work with its suppliers, raised food availability by 190 basis points over four years9. A smaller chain does not need their systems to work at the same level; it needs the same unit of account.
2. Sales are not demand
Every store forecast starts from till data, and till data records what was sold, not what customers wanted. Three things open a gap between the two. A product that ran out sold nothing for the rest of the day, however many people looked for it; the guide to censored demand shows how this drags the forecast down. A stock record that is wrong makes a store look empty when it is not, or full when it is empty. And reductions sell units that would not have sold at full price: in one study of 1,705 products across 676 stores of a European grocer, the retailer's forecast underestimated demand on discounted days for about 92% of products5.
So before choosing a method, clean the history: mark the days a shelf was empty, flag negative or frozen stock, and keep promotion and markdown days apart from ordinary ones. Researchers who classified 31,006 product series of a three-shop retailer put it plainly: however stock-outs are detected, they have to be treated in the model3.
3. Methods, from simple to machine learning
- Naive and seasonal naive. Tomorrow sells what today sold, or next Wednesday sells what last Wednesday did. It is the benchmark every other method has to beat1.
- Moving average. The mean of the last few weeks. Simple, but every day in the window counts the same and the day before it counts nothing, which makes the result jump when an unusual day drops out1.
- Exponential smoothing. A weighted average in which recent days count most and older ones fade gradually, with factors for the day of the week and public holidays. It is the workhorse of store forecasting.
- Methods for intermittent demand. When most days sell nothing, they forecast how often a sale happens and how big it is, separately. Croston's method and its successors such as TSB belong here; ordinary exponential smoothing does not1.
- Machine learning. Gradient-boosted trees and similar models learn from many products and stores at once, and can take promotions, prices and weather as inputs.
Machine learning wins clearly on the chain total and much less in each store. In the M5 competition on Walmart data, the winning model beat exponential smoothing by 22.4% overall but by only 3.4% at the product and store level, where the data is noisy and intermittent. When Petropoulos and colleagues repeated the test over a full year of rolling forecasts for 28,298 product and store series, exponential smoothing came out slightly more accurate than a replica of the winning machine-learning ensemble, and cheaper to run2. For a chain of a few stores, a well-set simple method with clean data beats an elaborate model on dirty data.
4. Most lines sell less than one a day
The textbook picture of a smooth sales curve rarely fits a store. In the three-shop retailer above, about 78% of series were intermittent even when counted by the week3. In 46,841 daily series from 446 stores, the average series had 492 days without a sale out of 7274. A forecast for such a line is a fraction, 0.3 a day, and the stock that covers it is a few units with a lot of uncertainty around them.
Here is what that looks like in a real order. Yammy Way, a candy store chain in Romania that uses marql, has a store in Brașov whose order for 17 November 2026 held 17 lines; marql marked 11 of them as running out before the next truck. These are some of them, as marql showed them that day:
| Product | Sells a day | Stock lasts | What marql shows |
|---|---|---|---|
| Toxic Waste Blue Raspberry Chew Bar 20 g | Sells a day1.5 | Stock lasts2.7 days | What marql showsRuns out before the truck |
| Toxic Waste Apple Chew Bar 20 g | Sells a day1 | Stock lasts1 day | What marql showsRuns out before the truck |
| Sour Patch Blue Raspberry 102 g | Sells a day0.4 | Stock lasts2.6 days | What marql showsRuns out before the truck |
| Hello Kitty Gummi Treats Fruit 88 g | Sells a day0.3 | Stock lasts6.2 days | What marql showsRuns out before the truck; check first, little data |
| Johny Bee 3D Gummy Cup Cake 18 g | Sells a day0.7 | Stock lasts0 days | What marql showsStock is −3 in the ERP, read as an empty shelf; check before sending |
| Amos Peelerz Raspberry 65 g | Sells a day0.1 | Stock lasts7.5 days | What marql showsRuns out before the truck |
Three things in this order are typical of a small chain. Most lines sell well under two units a day, so the forecast is a rate, not a count. Some stock records are negative, which is impossible on a shelf and has to be checked by a person before the order goes. And the lead time is not yet known, so the horizon is a default rather than a fact. The Yammy Way and Pralio story describes the chain.
5. How far ahead to forecast
An order has to last until the next one arrives, so the horizon that matters is the supplier's lead time plus the time until the next order. For a store that orders every Tuesday from a supplier who delivers in three days, that is ten days. Forecasting further is planning, forecasting less leaves a gap. The lead time itself varies, and that variation needs stock of its own: the reorder point and safety stock guides show how the forecast and its uncertainty become a quantity.
6. How to measure forecast accuracy
Measure the error in units, summed over the horizon of the order, and put two numbers side by side. WAPE, the total absolute error divided by total sales, shows how far off the forecast is. Bias, the total error with its sign, shows whether it is too high or too low on average1. A forecast with low WAPE and a steady negative bias still empties shelves.
Avoid MAPE for store forecasts. It cannot be calculated on a day with no sales, which for many store lines is most days, and minimising it pushes forecasts too low. Even absolute-error measures mislead on slow sellers: for a product that usually sells nothing, the forecast with the lowest absolute error is zero, which would never restock it1. Three habits keep the measure honest: compare with a naive rule on the same days, leave out days with an empty shelf, because their sales were capped by stock, and watch bias at least as closely as error.
7. Promotions, holidays and weather
Most of the large forecast errors come from events, not from the method. Promotions move volume sharply, and the mechanic matters: in eight Swedish supermarkets, a two-for offer sold 19.5% more units than the equivalent single-unit discount6. Holidays are worse: German grocers turned over 50% more in the week before Christmas 2021 than in an average week, and some reported forecast accuracy of only 35% in holiday seasons7. Weather moves whole categories: UK food store volumes fell 5.4% in May 2025 and rose 0.7% in June, when supermarkets pointed to the warm weather8.
Treat these as known events rather than noise. Enter promotions with their dates and mechanics, keep a calendar of public holidays for every country you trade in, and let a person override the forecast for what the history has not seen, such as a local festival or roadworks outside the store. The guide to AI replenishment covers which of these decisions to leave to the algorithm and which to the buyer. Better forecasts also reduce waste on fresh lines: Albert Heijn credits its store forecasts and dynamic markdowns together with saving over 1.5 million kilograms of food waste in 202410; the food waste guide covers the rest.
A note on the evidence: three of the studies cited here are preprints not yet peer-reviewed, the retailers' figures are their own and combine forecasting with other work, and the holiday accuracy figures come from interviews without a common measure.
How marql shows it in the store order
In marql, the AI Inventory Manager prepares each store's order from the sales and stock it reads from the POS and ERP a chain already uses. Every line shows how much the product sells a day in that store, how many days the stock on the shelf will last and whether it runs out before the next delivery. Each order says what it is based on, including the supplier's lead time, and marks it as a default when the lead time is not known yet. Lines with little data, an old or missing stock reading or a negative stock record are listed first, so someone looks at the shelf before the order goes. marql calculates the quantities; a person checks them, adjusts them and sends the order.
Sources
- Kolassa, S., Rostami-Tabar, B., Siemsen, E. (2023). Demand Forecasting for Executives and Professionals. Chapman & Hall/CRC. doi:10.1201/9781003399599. dfep.netlify.app
- Petropoulos, F., Grushka-Cockayne, Y., Siemsen, E., Spiliotis, E. (2025). Wielding Occam's razor: Fast and frugal retail forecasting. Journal of the Operational Research Society 76(8), 1564-1583. doi:10.1080/01605682.2024.2421339. arxiv.org
- Svetunkov, I., Sroginis, A. (2025-11-13). Why do zeroes happen? A model-based approach for demand classification. arXiv:2504.05894 (preprint). arxiv.org
- Hobor, L., Brčić, M., Polutnik, L., Kapetanović, A. (2026-03-11). Comparative Analysis of Modern Machine Learning Models for Retail Sales Forecasting. arXiv:2506.05941 (preprint). arxiv.org
- Winkelmann, D., Elbracht, T., Brenker, J., Gerzen, A. (2026-02-04). Discounted Sales of Expiring Perishables: Challenges for Demand Forecasting in Grocery Retail Practice. arXiv:2602.04464 (preprint). arxiv.org
- Gravert, C., Mormann, M. (2025). In-store food promotions increase sales as well as household food waste. Scientific Reports 15, 18035. doi:10.1038/s41598-025-01800-x. doi.org
- Obermair, A., Holzapfel, A., Kuhn, H. (2023). Operational planning for public holidays in grocery retailing: managing the grocery retail rush. Operations Management Research 16, 931-948. doi:10.1007/s12063-022-00342-z. doi.org
- Office for National Statistics (2025-07-25). Retail sales, Great Britain: June 2025. ons.gov.uk
- J Sainsbury plc (2025). Annual Report and Financial Statements 2025, strategy overview. sainsburys.co.uk
- Ahold Delhaize (2025). Annual Report 2024. aholddelhaize.com
- Tesco PLC (2026-05-15). Sustainability Report 2026. tescoplc.com
TopicsDemand forecasting, Replenishment, Forecast accuracy, Store orders
All articles