It has been known from many years now that operational activity-based models need a lot of survey data to incorporate behavioural decision making of people. While there have been contributions from the field of statistics about how much survey data is needed to come to reliable estimates of behaviour; an obvious question which is often overlooked in the domain is how much survey data is really necessary to obtain an activity-based model that is sufficiently competent and accurate. This question is not only scientifically challenging and interesting, but also can significantly reduce data collection costs and is also very useful for practitioners. A very appealing question would be whether an activity-based model could also be trained with a smaller survey data set without losing too much model quality. This paper tries to explore this research question in the case of an activity-based model for Flanders (Belgium) inside the 'Forecasting Evolutionary Activity-Travel of Households and their Environmental RepercussionS' (FEATHERS) framework. As the scheduler in this study is based on decision trees, progressive sampling is being applied in order to investigate accuracy for all discrete choice decision trees. Based on the results of this investigation, it is demonstrated that for some decision trees the activity-based survey data set can be very small without losing accuracy, while for other decision trees bigger data sets are needed.