Forecasting sales of highly advanced Electron Microscopes
Where high-tech meets supply chain, interesting questions arise. In our EHTC Research Stories you will find out how students have managed to answer them. This time: Amber van Oort’s master’s thesis, undertaken at Thermo Fisher Scientific (TFS), the world leader in serving science through development and manufacturing of biotechnology products. The focus of this thesis: the forecast sales of highly advanced Electron Microscopes (EM) that are manufactured at TFS and sold in low volume to high-tech companies and world-renowned research institutes and universities.
Where it started
The manufacturing of highly advanced and customized EM is a complex process for which long internal and external lead times are faced. Internal lead times can already be up to ten months and external lead times often take several weeks or months as well. Therefore, the manufacturing process needs to start based on forecasts to satisfy customer needs. It is crucial that the production of the right machines is initiated on time to prevent both lost sales and excess inventories under the actual demand. However, TFS has not been able to provide accurate sales forecasts in the past, leading to low material availability during a period of growth and excess inventories when sales stabilized. As a consequence, there has been a growing dissatisfaction at both factories and suppliers that are unwilling to act on changes in forecasts. For this reason, it was Amber’s job to develop a model to create accurate, stable, medium-term sales forecasts considering both statistical- and machine learning time series forecasting models.
Amber used data on historic sales of EM for 26 quarters from TFS. She also defined a proper aggregation level and scope where fifteen aggregate groups were included. With this data, she analyzed multiple statistical models as well as machine learning models. All models were constructed by use of 18 quarters of training data and forecasting performance was evaluated on the near and far future by use of 8 quarters of test data. The forecast accuracy of each model was evaluated by multiple accuracy measures.
Amber observed that the best forecasting models were able to outperform business experts’ forecasts for seven aggregate groups on the near future and fourteen aggregate groups on the far future. She also observed that the statistical models tend to outperform machine learning models because statistical models were capable to capture patterns such as seasonality.
In her research, Amber concludes that forecasting models have a great potential to provide accurate forecasts for the far future while business experts often perform better for the near future due to information advantage. Therefore, she recommends TFS not to replace the existing forecasting method but rather use the forecasting model as a basis and use the domain knowledge of the business experts to make adjustments for the near future planning.