This contribution focuses on the estimation of travel and service times in last mile delivery (LMD) problems, which concern the last stretch of supply chains associated with consumer products. These times, which are affected by a multiplicity of factors in urban settings, play a critical role in Intelligent Transportation Systems associated with Smart Cities. In addition, they are employed as input data in VRPTW (Vehicle Routing Problem with Time Windows) type of problems that take place in urban areas. The correct estimation of travel and service times becomes crucial in the effective solution of different LMD problems. The estimation task is tackled in this work by means of non-parametric, data-driven, machine learning (ML) methods, such as Random Forest and XGBoost. Raw data coming from different datasets, which were supplied by a company that provides GPS fleet tracking services in Argentina, has been employed. It can be concluded that is possible to address the problem by means of a trip-based approach that mainly uses truck GPS data. However, the quality of the results heavily depends on the characteristics of the input data. This is why efforts were devoted to data pre-processing activities (anonymization, cleaning, analysis, datasets integration, etc.) and feature engineering tasks, which led to the creation of two high-quality training datasets to be used by the ML approaches. This work provides details on these data curation and feature engineering tasks.