GPS-based travel surveys are increasingly used to study how people travel, but they record only where and when each trip takes place, not its purpose. Yet trip purpose is a key input to activity-based travel-demand models and to transport planning, so it has to be predicted from the recorded data before such a survey can be used. The objective of this thesis is to establish what matters more for this prediction: the type of model used, or the information available at the moment the prediction is made. To this end, three statistical and machine-learning models, a multinomial logit model, LightGBM and TabPFN, are compared on the TimeUse+ panel, a four-week GPS travel survey of 1,318 residents in Switzerland (207,683 training and 52,223 test trip legs).