Show simple item record

dc.contributor.authorWellens, Arnoud P.
dc.contributor.authorBoute, Robert
dc.contributor.authorUdenio, Maximiliano
dc.date.accessioned2023-11-10T09:45:51Z
dc.date.available2023-11-10T09:45:51Z
dc.date.issued2024en_US
dc.identifier.issn0377-2217
dc.identifier.doi10.1016/j.ejor.2023.10.039
dc.identifier.urihttp://hdl.handle.net/20.500.12127/7289
dc.descriptionA ‘simple’ tree-based framework is proposed for retail sales forecasting.• Our framework performs surprisingly well on two datasets with explanatory variables. Its good performance depends on collecting explanatory data and feature engineering. More sophisticated tree-based variants only marginally improve the forecast accuracy. Extensive simulation shows the benefit of our framework on inventory performance.en_US
dc.description.abstractDespite being consistently outperformed by machine learning (ML) in forecasting competitions, simple statistical forecasting techniques remain standard in retail. This is partly because, for all their advantages, these top-performing ML methods are often too complex to implement. We have experimented with various tree-based ML methods and find that a ‘simple’ implementation of these can (substantially) outperform traditional forecasting methods while being computationally efficient. Our approach is validated with a dataset of 4,523 products of a leading Belgian retailer containing various explanatory variables (e.g., promotions and national events). Using Shapley values and slightly adjusted tree-based methods, we show that superior performance depends on the availability of explanatory variables and additional feature engineering. For robustness, we show that our findings also hold when using the M5 competition dataset. Extensive numerical experimentation finally shows how the forecast superiority of our proposed framework translates to higher service levels, lower inventory costs, and improvements in the bullwhip of orders and inventory. Our framework, with its excellent performance and scalability to practical forecasting settings, we contribute to the growing body of research aimed at facilitating the higher adoption rate of ML among ‘traditional’ retailers.en_US
dc.language.isoenen_US
dc.publisherElsevieren_US
dc.subjectForecastingen_US
dc.subjectGlobal Forecasting Methodsen_US
dc.subjectTree-Based Methodsen_US
dc.subjectInventory Simulationen_US
dc.titleSimplifying tree-based methods for retail sales forecasting with explanatory variablesen_US
dc.identifier.journalEuropean Journal of Operational Researchen_US
dc.source.volume314
dc.source.issue2
dc.source.beginpage523
dc.source.endpage539
dc.contributor.departmentResearch Center for Operations Management, KU Leuven, Naamsestraat 69, Box 3555, 3000 Leuven, Belgiumen_US
dc.contributor.departmentResearch Center for Operations Management, KU Leuven, Naamsestraat 69, Box 3555, 3000 Leuven, Belgiumen_US
dc.identifier.eissn1872-6860
vlerick.knowledgedomainOperations & Supply Chain Managementen_US
vlerick.typearticleVlerick strategic journal articleen_US
vlerick.vlerickdepartmentTOMen_US
dc.identifier.vperid102358en_US


Files in this item

Thumbnail
Name:
Publisher version

This item appears in the following Collection(s)

Show simple item record