Sales forecasting using statistical and machine learning models: the case of HEXPOL

Abstract

Sales forecasting occupies an important place in manufacturing companies, especially in the B2B sector. It can help organisations optimise their supply chain management, resource allocation and target setting. There has been some development made in the different models and new techniques are derived from existing ones. However, there are still some important challenges and limitations in providing accurate forecasts. This thesis aims to answer three main questions about which technique and model would be adequate to be used in sales forecasting and whether external data should be included in the forecasting model and what kind of data. Furthermore, it discusses the different statistical and machine-learning model setups applied in search of the optimal outcome. The answers to these questions determine whether of this proof of concept is satisfactory or not. The conclusion determines if a statistical or machine-learning forecasting model could present an alternative or an extension to conventional judgemental forecasting. The Prophet model which has been developed by the Facebook engineering team turned out to be the best-performing model, whereas the artificial neural network did not perform well based on the chosen evaluation metrics. To better understand the results and put them into context, they are visualised in the BI software “Tableau”. Improvements in terms of accuracy and precision could still be made by extending the feature selection and taking more or different external datasets into consideration.