Data Science

Forecast of sales volumes for a supermarket chain

We supported a supermarket chain in making its supply chain fit. Thanks to pacemaker, a more precise forecast of the sales volume of over 25,000 products in the wholesale and retail branches of a supermarket chain is possible.



The customer operates over 10,000 branches in wholesale and retail. Precise forecasts of the sales quantities of the entire product range of the individual branches are required in order to optimally design the supply chain and the associated planning processes from purchasing to action planning to resource planning in the warehouse.

The great variety of the article range of a full-range supplier in retail and wholesale harbors various challenges in the creation of reliable forecasts. Extremely different sales behavior and varying sales frequencies of the individual assortment items require a variety of different methods for a reliable forecast.

The different sales frequency of the individual articles represents a particular difficulty. An added complication is that many assortment articles have varying and in some cases extremely short sales histories, such as newly introduced articles. The “one-size-fits-all” solution currently in use reveals problems, especially with articles that are not frequented very much. As part of the project, two selected retail outlets and two wholesale department stores were to be evaluated as to whether the use of machine learning methods, which were previously provided by SAP, could improve the classic forecasting methods.


In order to improve the existing forecasts, a large number of models were trained on the entire range of items from two branches of the wholesale and retail trade in order to identify the optimal forecast model for each individual item. In addition to classic influencing factors such as campaign periods, vacations, public holidays and special sales periods (e.g. the pre-Christmas period), the models have also been expanded to include weather influences.


By choosing the right model, the forecast error for the entire range of warehouses (wholesale) could be improved by up to 4% on average compared to a SAP forecast. At the individual item level, especially in the B2C retail business, significantly more significant improvements were achieved. Here the average forecast error could be improved by 6 – 9%. The noticeably improved forecasts help the customer to better plan the downstream processes in purchasing and logistics.

Similar posts

Bleiben Sie immer auf dem neusten Stand

In unserem Blog dreht es sich um Themen rund um Data Science und KI.