Master thesis presentation Viking Zandhoff Westerlund, MPCAS

Abstract: Demand forecasting plays an important role in facilitating data-driven decision-making for businesses, particularly in domains such as inventory planning and resource allocation. While traditional forecasting models such as exponential smoothing and autoregressive models have long been prevalent in the time series forecasting domain, recent research has been increasingly focused on more complex models based on machine learning. These complex models offer great potential and flexibility, but they require large amounts of data to achieve optimal performance. In this thesis, I investigate whether state-of-the-art machine learning models, such as the Temporal Fusion Transformer (TFT) and the LightGBM, can outperform a traditional exponential smoothing model in multi-horizon demand forecasting, using a relatively small data set. Furthermore, I investigate whether the hierarchical structure of the time series data can be exploited through forecast reconciliation to further increase forecasting accuracy. My findings indicate that both the TFT and LightGBM models have the potential to outperform the traditional exponential smoothing model. The best performing TFT and LightGBM based models improved the average forecast accuracy with 43.1 % and 33.7% respectively, compared to the exponential smoothing model. However, in contrast to the exponential smoothing model, the TFT and LightGBM are not ''off-the-shelf'' solutions, and they require careful attention regarding model selection and generalization ability. Moreover, my results show that while hierarchical forecast reconciliation has the potential to improve forecast accuracy, its effect on the forecasts is not consistent, and further analysis is needed to anticipate its impact. Altogether, the results in this thesis demonstrate that machine learning-based forecasting models have significant potential even on small data sets, but further research on rigorous and data efficient model selection techniques is warranted to ensure consistent results.

Password: 437888

Supervisor: Oscar Thorén
Examiner: Mats Granath
Opponent: Adam Söderholm