Demand forecasting with AI before the December peak: what you can prepare now

Why this work belongs in September, not November

CBS (Statistics Netherlands) publishes the trend in retail turnover every month, and that series shows the seasonal pattern of the sector as a whole. What your own peak looks like can be found only in your own order history, and you can still clean that up now. For part of your product range, the order that has to cover December sales is already in place in November, and earlier for imports by sea. Moreover, a forecasting model does not earn its place with a demo, but with weeks in which its advice sits next to the planner's and both are compared afterwards with actual sales. That comparison needs quiet weeks: October is suitable, December is not.

Then there is the nature of a peak. You experience one a year, so any model has seen only a handful of them, while it knows hundreds of examples of ordinary weeks. A peak forecast is by definition more uncertain: that is not a shortcoming of the software, but a property of the problem. So work less to a single number and more to a range, and ask what an error costs on the high side and on the low side.

Which data must be in order now

The general data check for AI in inventory management is in our step-by-step plan for AI inventory management. Below is what a seasonal peak requires on top of that, starting with what makes the most difference:

• Several years of sales history per item. With one peak in the data, no model can distinguish a structural seasonal pattern from chance.
• Lost sales and sold-out periods flagged. What you sold last year just before Sinterklaas (the Dutch St Nicholas celebration) is not what your customers asked for but what you still had in stock; without that flag, a model repeats the shortages.
• Promotions, special offers and large one-off orders labelled with date and type, so that an outlier is recognisable as a promotion and not as structural demand.
• Calendar features alongside the date: day of the week, public holiday, promotion week and closure day. A model that only sees a week number does not map this year's calendar onto last year's.
• Changed item numbers linked to their predecessor, because a new number starts without history, precisely in the part of the product range that changes most often around the festive season.
• Actual lead times by supplier, including how they deviate in the fourth quarter. A forecast only becomes an order recommendation once you know how far ahead you need to look.

The 2026 calendar falls differently from that of 2025

Part of the work is not a data issue but a calendar issue. In 2026 Sinterklaas falls on Saturday 5 December, whereas in 2025 it was a Friday. Christmas Day and Boxing Day fall on Friday 25 and Saturday 26 December, and Black Friday on 27 November. That shifts the last order dates, the last delivery dates and the days on which your customers are closed.

Also watch the week numbering. Almost every year, ISO week numbers cover slightly different days from the year before, and this year there is an extra factor: 2026 has 53 weeks, with week 53 running up to and including Sunday 3 January 2027 and having no counterpart in 2025. For reporting that is a nuisance; for a forecasting model built on week numbers, it is an error that only shows up in December. So record what the planner knows, such as a customer that does not accept deliveries between Christmas and New Year or a supplier that closes in the last week, as calendar data rather than as a manual correction after the fact.

What AI improves about a seasonal peak

The gain rarely lies in a cleverer insight into a single item; it lies in scale and consistency. No planner works through thousands of items every week in the same way, and for a peak in particular that is the difference between looking a little everywhere and looking closely at the items that matter.

• Extrapolating the seasonal pattern of items with sufficient history, including day-of-week and public holiday effects.
• Separating trend, season and day-of-week effect, and recalculating weekly as soon as new sales figures come in.
• Giving a range instead of a single number, so that you can plan to the service level you want.
• Identifying exceptions: items that deviate from their own history this year, well before the peak begins.
• Translating expected sales into order lines and cases, the units on which your staff and transport planning run.

What the model by definition does not know

Anything that is not in the history is not in the forecast: a customer that was won this year or, conversely, lost; a promotion your buyer agreed last week; a supplier that closes during Christmas week; a new item without sales history. For that last case, the usual approach is still the best: the planner designates a comparable item as a reference and records that judgement.

In food, moreover, the direction of an error matters more than its size. Buying too much of a fresh product means waste; buying too little of a long-life product means a lost sale that you partly make up later. For each product group, record what an error really costs on the high side and on the low side, and base the order recommendation on that rather than on the middle scenario. That ties in with the broader use of AI around quality and waste in food manufacturing, and downstream with the warehouse and e-fulfilment, where agency staff and overtime determine the costs during the peak.

A realistic schedule for the weeks you have left

What is achievable depends on where you stand now. A workable sequence looks like this:

• September: choose one clearly defined product group with a distinct December peak. Clean up the history, flag last year's lost sales and promotions, and record how good your forecast was last year, by week and by item.
• October: run a forecast weekly alongside the planner's advice and discuss the five biggest differences every week. You learn more from that than from an average accuracy percentage.
• Early November: freeze. Taking new software live just before the busiest weeks does not pay for itself.
• December: record lost sales and manual adjustments to the advice, with the reason. That is the data you start with next year.
• January: evaluate which deviations were predictable and which were not. Only the first category is a point for improvement.

The planner decides and signs, especially during the peak

At AIBA, our starting point is that the human decides and signs. In demand forecasting that is not a formality, because the peak is exactly the moment when people deviate from the advice: there is time pressure and a single miss is immediately visible. Record those deviations, with the reason. Not to check up on planners, but because December's overrides are the best learning data of the year: they show where the model structurally misses something and where a planner corrected on gut feeling while the model was right.

Article 4 of the EU AI Act (Regulation (EU) 2024/1689) has applied since 2 February 2025. Since the amendment by Regulation (EU) 2026/1744, in force from 27 July 2026, it has asked organisations to take measures that support the AI literacy of employees who work with AI systems. They do not have to guarantee a specific level. For planners, that comes down in practice to understanding what a forecast is based on, knowing when a model is working outside its comfort zone, such as with new items or a shifted calendar, and being able to explain why they followed or adjusted a recommendation.

So start with the baseline measurement, not with the software: take last year's December weeks and answer three questions. Where were you furthest off, what did that deviation cost, and could it have been predicted at the time? Our free Practice Scan helps you choose that first product group.