Small-Scale Demand Forecasting: Exploring the Potential of Machine Learning and Hierarchical Reconciliation

Abstract

Demand forecasting plays an important role in facilitating data-driven decision making for businesses, particularly in domains such as inventory planning and re source allocation. While traditional forecasting models such as exponential smooth ing and autoregressive models have long been prevalent in the time series forecasting domain, recent research has been increasingly focused on more complex machine learning-based models. These complex models offer great potential and flexibility, but they require large amounts of data to achieve optimal performance. In this thesis, I explored viable approaches for constructing accurate forecasting models for a young company in the industrial production industry who wants to predict their future demand, while facing the challenge of limited data availability. The analy sis in this thesis involved comparing the predictive performance of state-of-the-art machine learning models, such as the Temporal Fusion Transformer (TFT) and the LightGBM to an exponential smoothing state-space model. Furthermore, I investi gated whether the hierarchical structure of the time series data could be exploited through forecast reconciliation to further increase forecasting accuracy. My findings indicate that both the TFT and LightGBM demonstrate superior forecasting accu racy, improving the average forecast accuracy with 43.1 % and 33.2 % respectively, compared to the exponential smoothing model. However, the TFT displayed incon sistent performance results, suggesting its unreliability. Moreover, the results show that while hierarchical forecast reconciliation does not enhance forecast accuracy, it corrects incoherency between forecasts without compromising accuracy, which is of great value.