Winweb develops AI-based sales forecasts

Developer know-how, the customers' industry knowledge and the researchers' perspective come together in the software company Winweb's latest project: Together, a methodology based on artificial intelligence (AI) was developed that understands the causes of certain sales figures and uses them for future predictions.

“How many sausages do we need for the weekend?” This is certainly a question every company in the meat industry asks itself. If too much is produced, part of it has to be disposed of, if too little is produced, the company misses out on profits. Currently, many companies rely on the expertise and gut feeling of experienced employees for production planning. “However, this dependency entails considerable risks,” says Jan Schummers, Senior Software Engineer at Winweb: “The loss of an experienced production manager can jeopardize the financial stability of the entire company.” For this reason, forecasting systems are becoming increasingly important in the industry and the need for automated and precise forecasting methods is growing.

Why conventional forecasts reach their limits

Andreas Mayer, who is responsible for IT at Michael Kleiber GmbH, a butcher's shop in Swabia with its own branches, also emphasizes this: “There are many factors that influence sales volumes.” These include seasonal fluctuations – Wiener sausages are more in demand in winter than in the summer months. Current offers also affect sales: If sausages are advertised, more are sold and more must be produced in advance. Of course, the weather as well as weekdays or public holidays also determine sales. “From today's perspective, it is not possible for us to use the dependencies between similar products, which are either positively or negatively correlated with each other, customers' consumption behavior, and many other influencing factors for a sales forecast”, says Mayer. In addition, only small amounts of data can be analyzed manually.

An ERP system that processes sales and operational data and is also used at Kleiber with winweb-food is already very important as a basis for forecasts. It enables precise price calculations, strategic sales and operational planning and offers special analyses for customers in order to control sales in their branches. However, conventional time series analyses for sales forecasting show weaknesses in prediction. They are useful for planning, but are not sufficient to explain the underlying causes of sales peaks or slumps and seasonal trends. Particularly relevant sales days for production planning and branches can only be forecast inadequately with this.

Even modern machine learning approaches still do not deliver satisfactory results. “This is not due to data quality or to missing historical data”, says Schummers. “Machine learning simply does not distinguish between correlation and causality.” For example, the sale of barbecue meat is related to the increase in sunburn. “However, this is only due to the shared influencing factor - the sun - and has no other connection.” Understanding possible relationships and modeling them correctly is, however, essential in order to create reliable sales forecasts for the future.

That is why Schummers, together with Maastricht University and Michael Kleiber GmbH, investigated a new method. It is intended to explain sales figures and thus improve forecasts. In a study, the software engineer used causal artificial intelligence, Causal AI for short, together with Large Language Models (LLMs) to identify influencing factors on sales and then train the AI model on them. “As an ERP provider for the entire flow of goods, Winweb has the crucial data that enables us to create detailed forecasts for individual products and compare these forecasts with real sales figures”, explains Schummers. At Kleiber, he implemented and tested his new method in a realistic environment.

The specific question was whether Causal AI and LLMs can be used effectively in practice to support and improve decision-making processes. “We want to close the gap between theory and practical application and give our customers data-driven decision-making options”, says Schummers. Tree diagrams were used as the basis, which were developed together with Kleiber and which represent the causal relationships of the influencing factors for AI. In very practical terms, Causal AI is intended to support the Kleiber branches in estimating sales volumes later this year and to identify in advance the factors that influence sales.

The practical use of Causal AI to predict food sales figures has not yet been tested to date and is said to be unique in this context, according to the university. “The first results are very promising and surpass all previous attempts at sales forecasting”, explains Schummers. By combining LLMs and specialist knowledge, more accurate and operationally relevant forecasts can be created. “This has paved the way for a set of rules that can serve as a blueprint for implementing forecasts for our customers”, emphasizes Schummers.