
Approach · 22 November 2026
Detecting fraud and anomalies
Fraud that was never discovered is not in your examples. How a model learns what is normal, ranks anomalies and explains every signal.
Read the article →
Approach · 22 November 2026 · Martijn van Grieken
A sold-out day hides how many customers still came. Why average sales underestimate demand, and what you know with little history.

A supermarket that orders fresh bread every day sees at the end of the day how much has been sold. On a day when the shelf was empty at four o’clock, it does not see how many customers came after that. That day’s sales are a lower bound on demand. Anyone forecasting from sales figures is therefore working with a number that is too low, and the figures themselves do not show it.
This page describes how we forecast demand when part of it is invisible.
| Day | Ordered | Sold | Demand |
|---|---|---|---|
| Monday | 20 | 15 | 15 |
| Tuesday | 20 | 20 (sold out) | 23 |
| Wednesday | 20 | 17 | 17 |
| Thursday | 20 | 20 (sold out) | 26 |
| Friday | 20 | 20 (sold out) | 24 |
| Saturday | 20 | 20 (sold out) | 28 |
| Sunday | 20 | 13 | 13 |
| Average per day | 20 | 17.9 | 20.9 |
Illustrative calculation. In reality you do not know the last column.
On four of the seven days the product sold out. Average sales are 17.9 a day, actual demand 20.9. Anyone who bases the order on sales will order less than now. The shelf then runs empty more often, sales fall further, and demand looks even lower the following week.
A sold-out day does provide information. You know that demand was at least as large as what you had in stock. A model that links demand, sales and order quantity uses that information instead of throwing it away.
Per product per shop there is often little usable history. On a few weeks of data you cannot estimate Saturday’s demand to the nearest loaf. A statement such as ‘probably between 22 and 32’ is more accurate than ‘27’, and you can do more with it.
The range makes the ordering decision visible. If a missed loaf costs you more than a discarded loaf, you order in the upper part of the range. If waste is more expensive, you order lower. That is the shop’s choice. The model shows what that choice costs.
The principle applies wherever you only see what was supplied. A possible application in healthcare: if a ward runs out of a consumable, it borrows from another ward or uses an alternative. The consumption records then show less than was needed. In industry the same can happen with spare parts: if a part is out of stock, the customer sometimes orders elsewhere and the demand appears nowhere. We have not yet applied this approach in healthcare or industry. How we approach a planning question in healthcare is described in Smarter rostering in care.
We describe in a model how demand, sales and order settings relate to each other, and estimate demand with Bayesian statistics. It calculates explicitly with uncertainty and therefore gives a range. We then test the forecasts against what you do observe: waste, lost sales and margin. Our work in industry is described under industry and maritime.
For a retailer we examined the order settings for fresh products in this way. For one fresh product in ten supermarkets we calculated that better settings could have increased profit there by 4 to 7 per cent. That is a calculation, not a measured increase in profit. Read the retail case (in Dutch).
Bring one product or one stock where you suspect the figures underestimate demand. That first conversation is without obligation and we aim to respond within one working day.
Author: Martijn van Grieken, founder of Twentynext. How we make this. AI wrote the first version of the Dutch original, and Martijn read and approved it before it went live. This English version was translated with AI. Twentynext is responsible for the content but cannot guarantee that it is accurate or complete. See our disclaimer.

Approach · 22 November 2026
Fraud that was never discovered is not in your examples. How a model learns what is normal, ranks anomalies and explains every signal.
Read the article →

Approach · Healthcare · 22 November 2026
Can AI make a roster that works for staff, clients and the organisation? How we approach that question, presented at R’evolutie van zorg 2026.
Read the article →

Approach · 22 November 2026
Data-driven working starts with agreeing who owns which data. Eight steps, a baseline measurement across five areas and the role of the set-up lead.
Read the article →
We use Google Tag Manager to measure visits and Leadinfo to recognise which company is visiting. Neither loads unless you agree. If you choose essential only, the site works as normal and we measure nothing. Cookie statement · Privacy statement