| dc.description.abstract | In this thesis, we study the prediction of stop delivery times using multiple linear regression and machine learning models (Decision Tree, Random Forest, eXreme Gradient Boosting and Support Vector Regression). We analyse approximately 2.4 million deliveries over a two-year period using data from Posten Bring AS, OpenStreetMap, and the Norwegian Meteorological Institute, combining quantitative data with courier interview insights. Our methodology includes data preprocessing, model training with scikit-learn, and feature interpretation using permutation importance and SHAP values. We find that factors such as the number of mailboxes at a stop, the number of ‘bag-on-door’ deliveries, and the number of ‘parcel in a mailbox’ significantly increase predicted stop delivery times, indicating that higher delivery volumes will extend overall delivery times even if the number of stops remains unchanged. In contrast, temperature and precipitation appear to have minimal impact. We conclude that it is possible to predict accurate stop delivery times without knowing the true stop/driving times by setting one to a fixed value. Additionally, we conclude that the variability in delivery times across different units suggests the need for tailored strategies in urban versus suburban areas. | en_US |
