AI demand forecasting
Also called sales forecasting, inventory forecasting, machine learning forecasting, demand planning
What forecasting competitions show about AI demand forecasting, how much sales history a small business needs, and where it goes wrong.
AI forecasting has won on huge retail datasets; for a small catalogue, a well-run simple method is the bar any tool must beat on your own data. In the M5 competition, built on about five years of Walmart sales across 30,490 product-store series, the winning machine-learning entries beat the best benchmarks by more than 20%. On the same data, though, a simple statistical method (exponential smoothing) matched a close copy of the winning method for single products in single stores, and ran faster.
At a glance
| Status | In pilots |
|---|---|
| What it does | Predicts how many units of each product you will sell over the coming weeks or months, from your sales history and, where available, promotions and holidays. |
| Typical cost | From US$119.99 per month for one Shopify forecasting and purchasing app (vendor list price, checked Oct 2026); otherwise varies by vendor. |
| Data it needs | Daily or weekly unit sales per product, ideally two full years or more so a yearly seasonal pattern can be estimated, plus stockout dates and a promotions calendar. |
| Setup effort | Medium: connecting the software is quick; cleaning product codes and sales history takes the time. |
| Evidence | Moderate. Large open forecasting competitions and peer-reviewed studies on retailer data; one small-firm pilot study, funded by a software firm. |
What is AI demand forecasting?
Demand forecasting estimates how many units of each product customers will buy in a future period. The number feeds purchasing, production and cash planning. In many small businesses it is done by eye or in a spreadsheet: last year's sales, nudged up or down by a feeling about this year.
AI demand forecasting replaces that with a model that learns patterns from sales history. The label covers two broad families. Machine learning models, such as gradient-boosted trees, learn from many products at once and can take in promotions, holidays and prices. Statistical methods, such as exponential smoothing, fit each product's own history and are cheap to run.
The label on the box matters less than the result. The question for an owner is whether a tool's forecasts beat a simple method on your products, over a period it has not already seen.
How does it work?
The tool pulls sales history from your shop, point-of-sale or accounting system, usually by product and by day or week. It fits one or more models, tests them on recent months it held back, and produces a forecast for each product with a range around it. Better tools then turn the forecast into a suggested order quantity using supplier lead times and the stock you hold.
A key advantage of machine learning is cross-learning: one model learns from thousands of related products at once. In a review of Kaggle forecasting competitions, every winner used such global models, information on promotions and holidays was usually highly useful, and weather and economic data added little. That is why scale matters. A shop with 80 products and patchy history gives a model far less to learn from than a retailer with tens of thousands of series.
What it needs to work
Forecasting only works if recorded sales reflect real demand and each product can be followed over time. That means usable history per product, joined to stock and promotion records.
- Unit sales by product (SKU) by day or week, exported from your shop, point-of-sale or accounting system: two full years before trusting yearly seasonal patterns; one small-firm study left out products with less than a year of history
- Stable product codes, with a mapping of old to new SKUs where codes changed and variants (size, colour) linked to a parent product
- Stockout history: the dates each product had zero stock, so lost sales are not read as low demand
- A calendar of promotions, price changes and public holidays, on the same dates as the sales data
- Supplier lead times and minimum order quantities, so forecasts can become order suggestions
- A recent stretch of history held back to test the tool against a simple benchmark before you rely on it
Most use cases in this guide succeed or fail here, on the data underneath them, before any AI is involved.
What does the evidence show?
The best evidence comes from open forecasting competitions run by academics. A 2018 study of 1,045 monthly series found ten machine learning methods were less accurate than statistical methods at every forecast horizon, and cost more to run. In the M4 competition of 100,000 series, all six pure machine learning entries performed poorly; the winner was a hybrid of statistical and machine learning methods.
The M5 competition reversed that on retail data. Built on 42,840 Walmart sales series, 30,490 of them for single products in single stores and the rest adding up to total sales, it saw the winning machine learning entries beat the best benchmarks by more than 20%. Nearly all of the top 50 used LightGBM, a popular machine-learning method. Two caveats matter. The data held roughly five years of history from ten stores, and a follow-up study found exponential smoothing matched a close copy of the winning LightGBM method at product-store level while running faster.
The only small-firm study we found is a 2024 preprint, funded by a software firm, that piloted automated forecasting at five companies; its recommended model significantly beat a naive baseline at four.
Machine learning forecasting vs a simple statistical method
| Compared on | Machine learning forecasting | Simple statistical method |
|---|---|---|
| Where it has won | Large retail datasets with thousands of related products and promotions data | The earlier M3 and M4 competition datasets |
| History needed | More is better; learns across products | Copes with shorter histories; standard methods need two seasonal cycles to estimate seasonality |
| Computing cost | Substantial at scale | Lower and faster to rerun |
| Explaining a forecast | Harder to say why a number moved | Easier: level, trend and season |
| Best for | Large catalogues, several years of data, frequent promotions | Small catalogues, shorter histories, steady sellers |
Status and evidence
Competition evidence is strong that machine learning beats simple methods on large retail datasets. But those datasets hold about five years of sales across tens of thousands of series, which few small firms have, and on the same data a simple method, exponential smoothing, matched it at product-store level. The only small-firm study we found is a vendor-funded preprint covering five companies, so we rate this pilots with moderate evidence.
Where does it go wrong?
- Sales are not demand. When a product was out of stock, recorded sales show zero and the model learns that nobody wanted it. The small-firm pilot flagged exactly this gap.
- Too little history. Standard methods for splitting out a seasonal pattern need at least two full cycles of data, so two years for a yearly pattern; workarounds rest on extra assumptions. Short histories favour very simple models.
- Messy product codes. In the same pilot, real business data had outliers and product codes that changed without notes, and strong deep learning results seen on a hand-picked dataset could not be reproduced.
- Lumpy demand. About 73% of the M5 series were classed as intermittent, with many periods of no sales. Slow sellers defeat precise weekly forecasts by any method.
- Gut-feel overrides. In a study of more than 60,000 forecasts at four companies, small manual adjustments often made forecasts worse, and upward adjustments were much less likely to help than downward ones.
Numbers it should move
Ask the demo business
Northgate Supply is a fictional wholesale distributor of catering and packaging supplies: 28 staff, R46.6m in sales over the last year. Its data is generated for this demo, to 30 September 2026.
We invoiced R3.98m in September 2026, and R46.6m over the last 12 months. July to September came in 1.2% below April to June. The biggest month was December 2025, at R5.66m, on festive-season trade. Disposables is the largest category, at 22% of sales.
How this was worked out
Invoiced sales excluding VAT, after discounts, per calendar month, and each product category’s share of the 12-month total.
Where the data usually lives
Where the records this use case needs are usually kept in a small business. It is not a list of recommended tools.
Frequently asked questions
How much sales history do I need for AI demand forecasting?
Two full years if you want the tool to learn a yearly seasonal pattern: standard methods for splitting it out need at least two full years of data. Beyond that there is no magic minimum; a standard forecasting textbook calls fixed rules of thumb misleading and notes that short histories favour simple models. One small-firm study excluded any product with less than a year of history. For new products, use judgement or the history of a similar product until data builds up.
Is AI forecasting more accurate than my spreadsheet?
Sometimes. On large retail data, machine learning beat the best benchmarks by more than 20%; on older competition data, simple statistical methods won. A spreadsheet that applies a sensible method is a fair benchmark. One built on gut feel is weaker, since small manual adjustments often reduce accuracy. Ask any vendor to show its accuracy against a simple method on your own past sales.
How much does demand forecasting software cost?
Prices vary by vendor and sales channel. One forecasting and purchasing app aimed at small, single-warehouse Shopify merchants lists US$119.99 per month (checked Oct 2026). Budget too for the staff time needed to clean product codes and stock records before you start, because the software can only learn from the history you give it.
Do I need a developer or data scientist to use AI forecasting?
Not for off-the-shelf tools that connect to common shop and inventory systems. You do need someone who knows the data: which products were renamed, when stock ran out and when promotions ran. In one small-firm pilot, real business data proved hard to bring up to the quality of a hand-picked test set. A custom model built from scratch does need specialist skills.
Why does my forecast keep getting slow-moving products wrong?
Slow sellers have many days with no sales, which makes any forecast noisy. In the M5 retail data, about 73% of series were classed as intermittent. For these products, aim for a sensible stock level and reorder point rather than a precise weekly number, and judge the tool on whether you ran out or overstocked, not on percentage error.
More in stock, supply & fulfilment
Sources
- 1M5 accuracy competition: Results, findings, and conclusionsInternational Journal of Forecasting (via IDEAS/RePEc), 2022Study
- 2Exploring the representativeness of the M5 competition dataarXiv (Theodorou, Wang, Kang, Spiliotis, Makridakis, Assimakopoulos), 2021Study
- 3Transfer learning for hierarchical forecasting: Reducing computational efforts of M5 winning methodsKU Leuven Lirias (Wellens, Udenio, Boute), 2021Study
- 4Wielding Occam's razor: Fast and frugal retail forecastingarXiv (Petropoulos, Grushka-Cockayne, Siemsen, Spiliotis), 2023Study
- 5
- 6Decision Makers Should Rely on Hybrid Forecasting ModelsINSEAD Knowledge, 2018Commentary
- 7Kaggle forecasting competitions: An overlooked learning opportunityarXiv (Bojer, Meldgaard), 2020Study
- 813 Some practical forecasting issues – Forecasting: Principles and Practice (3rd ed)OTexts (Hyndman, Athanasopoulos), 2021Study
- 9Seasonal decomposition of short time seriesRob J Hyndman (via R-bloggers), 2018Commentary
- 10Automated Demand Forecasting in small to medium-sized enterprisesarXiv (Hasso Plattner Institute, University of Potsdam), 2024Study
- 11Effective forecasting and judgmental adjustments: an empirical evaluation and strategies for improvement in supply-chain planningInternational Journal of Forecasting (via Lancaster University ePrints), 2009Study
- 12Inventory Planner ‑ EssentialsShopify App Store (Inventory Planner by Sage), 2026Vendor
This entry is educational. It is not legal, financial or professional advice, and it does not recommend any vendor. Figures attributed to vendors are their own claims.
- 1M5 accuracy competition: Results, findings, and conclusionsInternational Journal of Forecasting (via IDEAS/RePEc), 2022Study
- 2Exploring the representativeness of the M5 competition dataarXiv (Theodorou, Wang, Kang, Spiliotis, Makridakis, Assimakopoulos), 2021Study
- 3Transfer learning for hierarchical forecasting: Reducing computational efforts of M5 winning methodsKU Leuven Lirias (Wellens, Udenio, Boute), 2021Study
- 4Wielding Occam's razor: Fast and frugal retail forecastingarXiv (Petropoulos, Grushka-Cockayne, Siemsen, Spiliotis), 2023Study
- 5
- 6Decision Makers Should Rely on Hybrid Forecasting ModelsINSEAD Knowledge, 2018Commentary
- 7Kaggle forecasting competitions: An overlooked learning opportunityarXiv (Bojer, Meldgaard), 2020Study
- 813 Some practical forecasting issues – Forecasting: Principles and Practice (3rd ed)OTexts (Hyndman, Athanasopoulos), 2021Study
- 9Seasonal decomposition of short time seriesRob J Hyndman (via R-bloggers), 2018Commentary
- 10Automated Demand Forecasting in small to medium-sized enterprisesarXiv (Hasso Plattner Institute, University of Potsdam), 2024Study
- 11Effective forecasting and judgmental adjustments: an empirical evaluation and strategies for improvement in supply-chain planningInternational Journal of Forecasting (via Lancaster University ePrints), 2009Study
- 12Inventory Planner ‑ EssentialsShopify App Store (Inventory Planner by Sage), 2026Vendor