TY - CHAP

T1 - The Potential of AutoML for Demand Forecasting

AU - Kramer, Kathrin

AU - Behn, Niclas

AU - Schmidt, Matthias

N1 - Funding Information: Funded by the Lower Saxony Ministry of Science and Culture under grant number ZN3489 within the Lower Saxony “Vorab” of the Volkswagen Foundation and supported by the Center for Digital Innovations (ZDIN). Funding Information: Funded by the Lower Saxony Ministry of Science and Culture under grant number ZN3489 within the Lower Publisher Copyright: © 2022, Publish-Ing in cooperation with TIB - Leibniz Information Centre for Science and Technology University Library. All rights reserved.

PY - 2022

Y1 - 2022

N2 - In demand forecasting, which can depend on various internal and external factors, machine learning (ML) methods can capture complex patterns and enable precise forecasts. Accurate forecasts facilitate targeted, demand-oriented planning and control of production and underline the importance of this task. The implementation of ML-algorithms requires knowledge of the specific domain as well as knowledge of data science and involves an elaborate set up process. This often makes the application of ML to potential industrial problems economically unattractive. The major skills shortage in the field of data science further exacerbates this. Automation and better accessibility of ML methods is therefore a key prerequisite for widespread use. This is where the principle of automated ML (AutoML) comes in, automating large parts of a ML pipeline and thus leading to a reduction in human labour input. Therefore, the aim of the publication is to investigate the extent to which AutoML solutions can generate added value for demand planning in the context of production planning and control. For this purpose, publicly available datasets deriving from Walmart as well as an anonymised manufacturing company are used for short-term and long-term forecasting. The AutoML tools from Microsoft, Dataiku and Google conduct these forecasts. Statistical models serve as benchmarks. The results show that the forecasting quality varies depending on the software, the input data and their demand patterns. Overall, the prepared models from Microsoft show the most accurate results in average and the potential of AutoML becomes particularly clear in the short-term forecast. This paper enriches the research field through its broad application, giving valuable insights into the use of AutoML tools for demand planning. The resulting understanding of limitations and benefits of AutoML tools for the case studies presented fosters their suitable application in practice.

AB - In demand forecasting, which can depend on various internal and external factors, machine learning (ML) methods can capture complex patterns and enable precise forecasts. Accurate forecasts facilitate targeted, demand-oriented planning and control of production and underline the importance of this task. The implementation of ML-algorithms requires knowledge of the specific domain as well as knowledge of data science and involves an elaborate set up process. This often makes the application of ML to potential industrial problems economically unattractive. The major skills shortage in the field of data science further exacerbates this. Automation and better accessibility of ML methods is therefore a key prerequisite for widespread use. This is where the principle of automated ML (AutoML) comes in, automating large parts of a ML pipeline and thus leading to a reduction in human labour input. Therefore, the aim of the publication is to investigate the extent to which AutoML solutions can generate added value for demand planning in the context of production planning and control. For this purpose, publicly available datasets deriving from Walmart as well as an anonymised manufacturing company are used for short-term and long-term forecasting. The AutoML tools from Microsoft, Dataiku and Google conduct these forecasts. Statistical models serve as benchmarks. The results show that the forecasting quality varies depending on the software, the input data and their demand patterns. Overall, the prepared models from Microsoft show the most accurate results in average and the potential of AutoML becomes particularly clear in the short-term forecast. This paper enriches the research field through its broad application, giving valuable insights into the use of AutoML tools for demand planning. The resulting understanding of limitations and benefits of AutoML tools for the case studies presented fosters their suitable application in practice.

KW - Engineering

KW - AutoML

KW - Demand forecasting

KW - sales forecast

KW - Machine learning

KW - Manufacturing

KW - Production planning

UR - http://www.scopus.com/inward/record.url?scp=85164413933&partnerID=8YFLogxK

U2 - 10.15488/12162

DO - 10.15488/12162

M3 - Article in conference proceedings

T3 - Proceedings of the Conference on Production Systems and Logistics

SP - 682

EP - 692

BT - Conference on Production Systems and Logistics

A2 - Herberger, David

A2 - Hübner, Marco

PB - publish-Ing.

CY - Hannover

T2 - 3rd Conference on Production Systems and Logistics - CPSL 2022

Y2 - 17 May 2022 through 20 May 2022

ER -