Machine Learning as a Tool for Demand Forecasting: A Study at HMS Networks

Sammanfattning

This thesis investigates the applicability of using machine learning for demand forecasting at HMS Networks. Using the design science research methodology, multiple forecasting artifacts were developed and tested iteratively, using historical sales data. Four tree-based models; Decision Tree, Random Forest, LightGBM and XGBoost, were tested with different pre-processing and feature engineering techniques such as first-order differencing, scaling and varying levels of granularity. The results show that performance was primarily improved through better data representations, rather than through model selection alone. Particularly, first-order difference and scaling improved the models’ abilities to generalize and reduced forecasting bias. Furthermore, the temporal granularity was shown to strongly affect the stability of the forecast, where monthly granularity generated more stable forecasts than weekly. When compared to the current forecast at HMS Networks, XGBoost and LightGBM performed better based on both accuracy and bias. Overall, the findings from this study show that machine learning can be applicable to demand forecasting when the data representation is aligned with the capabilities and limitations of the models.