How many sausages for the weekend?

Winweb develops AI-based sales forecasting

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

“How many sausages do we need for the weekend?” This is a question that every company in the meat industry is sure to ask. If too much is produced, some of it has to be discarded; if too little is produced, the company loses out on profits. Many companies currently rely on the expertise and intuition of experienced employees when planning production. “But this dependence 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 demand for automated and precise forecasting methods is growing.

Why conventional forecasts reach their limits

This is also emphasized by Andreas Mayer, who is responsible for IT at Michael Kleiber GmbH, a butcher’s business in Swabia with its own branches: “There are many different factors that influence sales volumes.” These include seasonal fluctuations – Vienna sausages are more popular in winter than during the summer months. Current promotions also affect sales: If sausages are advertised, more are sold and more must be produced in advance. Naturally, the weather and weekdays or public holidays also influence sales. “From today’s perspective, however, it is not possible for us to incorporate the relationships between similar products that correlate either positively or negatively with each other, customers’ consumption patterns and many other influencing factors into a sales forecast”, says Mayer. Moreover, only limited volumes of data can be analyzed manually.

An ERP system that processes sales and operational data and is also used at Kleiber in the form of winweb-food is already very important as a basis for forecasting. It enables precise price calculations and strategic sales and operational planning and provides customers with specific analyses to manage sales in their stores. However, conventional time-series analyses for sales forecasting have weaknesses when it comes to predictions. Although they are useful for planning, they are not sufficient to explain the underlying causes of sales peaks or slumps and seasonal trends. They are inadequate for forecasting days with particularly relevant sales volumes for production planning and stores.

Even modern machine learning approaches do not yet deliver satisfactory results. “This is not due to the quality of the data or a lack of historical data,” says Schummers. “Machine learning simply does not distinguish between correlation and causation.” For example, sales of meat for barbecuing are correlated with the incidence of sunburn. “However, this is only due to the common influencing factor—the sun—and there is no other connection.” Understanding potential relationships and modeling them correctly is essential for producing reliable future sales forecasts.

For this reason, Schummers worked with Maastricht University and Michael Kleiber GmbH to investigate a new method. It is intended to explain sales figures and thereby improve forecasts. In a study, the software engineer used causal artificial intelligence, or Causal AI, together with Large Language Models (LLMs) to identify factors influencing sales and then train the AI model on them. „As an ERP provider covering the entire flow of goods, Winweb has the crucial data that enable us to create detailed forecasts for individual products and compare these predictions with actual sales figures“, explains Schummers. He implemented and tested his new method at Kleiber in a realistic environment.

Specifically, the question was whether Causal AI and LLMs can be used effectively in practice to support and improve decision-making processes. „We want to bridge the gap between theory and practical application and provide our customers with data-driven decision-making capabilities“, says Schummers. Tree diagrams developed together with Kleiber were used as a basis to illustrate the causal relationships between the factors influencing the AI. In practical terms, Causal AI is expected to help Kleiber’s branches estimate sales volumes later this year and identify in advance the factors that affect sales.

According to the university, the practical use of Causal AI to predict food sales volumes has not yet been tested and is unique in this context. “The initial results are very promising and outperform all previous attempts at sales forecasting,” Schummers explains. Combining LLMs with specialist expertise makes it possible to produce more accurate and operationally relevant forecasts. “This has paved the way for a set of rules that can serve as a blueprint for implementing forecasting solutions for our customers,” Schummers emphasizes.