Efficient baggage forecasting is critical to maintaining smooth airport operations and optimizing resource allocation. This thesis, in collaboration with Swedavia AB, explores
the application of machine learning methods to forecast baggage volumes at Arlanda
Airport. The project aimed to optimize staffing, conveyor belt load, storage logistics,
and ground handling schedules, as well as bring valuable insights to Swedavia regarding future implementation of machine learning practices. By utilizing historical flight and
baggage data, several modeling approaches were evaluated – including XGBoost, Random Forest, and Prophet – and compared to Swedavia’s existing SQL-based forecasting process. The XGBoost model demonstrated the strongest overall performance, outperforming the baseline with a 24% reduction in mean absolute error. In addition to predictive accuracy,

the models provided operational insights through feature importance analyses, highlighting most importantly the role of time, destination, and operator in determining baggage
loads. Model adaptability was tested over time and showed that periodic retraining
with the most recent data enhanced performance for the test set of the year 2024. Lastly, considerations for practical deployment are discussed, including model interpretability and data update mechanisms. The results indicate that machine learning can offer measurable improvements in both accuracy and operational usability for baggage forecasting at Arlanda Airport.