Demand forecasting and AI

Demand forecasting is an estimate of how many units of each product will be bought in a future period (a week, a month, a season). It is used to order from suppliers as much as will sell: not to leave shelves empty and not to freeze money in surplus stock. AI (machine learning) can improve a forecast, but it does not remove the need for clean data and a sensible purchasing policy.

In plain words

The simplest forecast is "how much we sold in the same period last time". More complex methods account for seasonality, promotions, price, weather, trends and new product launches. Machine learning learns from sales history and finds patterns that are hard to see in a spreadsheet.

What the research says

• Modern models versus simple ones. In a comparison of modern models for retail sales forecasting (Hobor et al., preprint of June 2025, updated March 2026; data from stores in south-eastern Europe), XGBoost, a tree-based method, showed the lowest RMSE error. Neural networks with gap filling improved on aggregated data but trailed ensemble methods. The authors conclude that a model should be chosen by the features of the task, not by its complexity (arXiv, 2025-2026).
• Forecast accuracy does not always mean lower costs. A study by Theodorou, Spiliotis and Assimakopoulos (European Journal of Operational Research, 2025) on M5 competition data showed that the link between forecast accuracy and inventory results can be weak: accuracy matters more when holding costs are comparable to or greater than lost sales. When lost sales cost more, the best method is not necessarily the most accurate. The effect is strongest for short order cycles and products with irregular demand (IDEAS/RePEc, 2025). So the question "is the forecast accurate" is worth pairing with "what happens if I am wrong too high or too low".

Both studies concern classic sales data, not small Ukrainian stores. There are no separate studies on demand forecasting for Ukrainian e-commerce.

What forecast quality depends on

• History. For a new product and a small store there is little data: the forecast is rough by definition.
• Clean data. Sales during a stockout are understated: the product was not there, rather than "people did not want it". If this is not accounted for, the model learns to underestimate demand. Duplicates and test orders also have an effect.
• Promotions and prices. Without a promotion flag, the model takes a peak for ordinary demand.
• Product type. A regular one (consumables) is easier to forecast than an irregular one (rare expensive items, fashion).

Example (made-up)

A store sells water filters. Average sales are 120 units a month, and November is usually +30%. A simple seasonal estimate for November is 156 units. If a promotion is planned this year, the forecast is adjusted using the result of the previous promotion. The goal is not to guess the number but to define a range (for example, 140-180) and decide how much to keep in stock given the lead time: see Stock, shortages and surplus.

What to do

1. Start simple: a seasonal forecast based on last year and the monthly rhythm.
2. Keep data on stockouts and promotions.
3. Split products into groups by importance (ABC/XYZ analysis) and forecast more precisely only the main ones.
4. Compare the forecast with actuals and write down the gaps.
5. Bring in AI tools when the simple method no longer copes and you have the data, and compare them with the simple method on your own data.

Common mistakes

• Forecasting from sales without accounting for periods when the product was unavailable.
• Judging a model by one accuracy figure with no link to costs.
• Believing a "95% accuracy" from a service's advertising without testing it on your own data.
• Forgetting the lead time: the forecast is needed for the period from ordering to the goods arriving.

FAQ

What is demand forecasting? An estimate of future product sales for purchase planning.

Is AI needed for a forecast? Not always. In the study by Hobor et al., ensemble methods beat neural networks, so start with a simple method and compare.

How do I forecast a new product? Use similar products as a guide and buy cautiously in a small batch. There is no reliable method for new products with no history.

Which forecast is more accurate? The one that gives a better result on your data and costs. The studies do not name an overall winner.

Related terms

• Stock levels, stockouts and overstock
• ABC/XYZ analysis of your product range
• Merchandising in an online store
• OMS and WMS
• AI for store data analysis: what you can trust it with

Sources

• Hobor L. та ін., Comparative Analysis of Modern Machine Learning Models for Retail Sales Forecasting (arXiv, 06.2025, v2 03.2026) — arxiv.org/abs/2506.05941
• Theodorou E., Spiliotis E., Assimakopoulos V., Forecast accuracy and inventory performance: Insights on their relationship from the M5 competition data, European Journal of Operational Research 322(2), 2025 — ideas.repec.org/a/eee/ejores/v322y2025i2p414-426.html