Type of Publication: Dissertation

Demand forecasts for optimization purposes: a machine learning marketplace for federated learning

Author(s):
Hütsch, M.
Publisher:
Universität Duisburg-Essen
Location(s):
Essen
Publication Date:
2022
Link to complete version:
https://primo.uni-due.de/permalink/49HBZ_UDE/1uttatt/alma99208400627306446
Citation:
Download BibTeX

Abstract

As the importance of machine learning (ML) in technology has increased over past years, it is important to understand each application and its implications for ML individually. As a sample application of ML, demand forecasts occupy a special position in industry and retail and are used in numerous optimization activities, such as price optimization. In past years, demand forecasting has transformed the way retailers and industry replenish their goods. From a manual process, forecast-driven replenishment processes are now used. While this process has very few requirements for the ML-based demand forecast, this changes when further processes are transformed. While optimizing prices, demand forecasting requires elasticity regarding optimization objectives. When ML methods are used for demand forecasting, there must be sufficient data to learn the proper elasticity requirements. To overcome the problem of missing historical data resulting in low elasticity and thus bad optimization performance, we present a method for transferring knowledge between articles and companies to train shared ML models. For the first time, we are able to optimize prices without generating historical demand data for all possible prices. Evaluations of a design science project show that federated learning with neural networks achieves good forecasting accuracy and high elasticity.