Short description:

By means of suitable methods from the field of machine learning, and using historical data from the data warehouse, measurable improvements are planned for forecast quality compared to the current forecast method. Using artificial neuronal networks, shipping quantities are to be forecast with minimal or moderate errors for selected logistics sites (HUBs). Verification is provided that methods based solely on historical data lead to improved forecasts, even without including expert knowledge.

Technical description:

In the context of a proof of concept, a forecast method is developed on the basis of an artificial neuronal network (multilayer perceptron). By including historical data from the data warehouse, dedicated forecast models are developed for selected logistics sites using tools like Weka and RStudio, with the aim of forecasting shipping quantities precisely to the day (forecast horizon: one day in advance). The modeling relies on extracting and selecting descriptive characteristics (e.g. shipping quantities in the past x days, bank holidays, school holidays etc.). After suitable characteristics are selected and after the fine adjustment, the models are evaluated for the specific sites for a specified forecast period (one month). Compared to the existing method, the models based on artificial neuronal networks exhibit fewer forecasting errors across all sites (MAPE, RMSE).