Learning surrogates for product design and optimization is potential to capitalize on competitive market segments. In this paper we propose an approach to learn surrogates of product performance from historical clusters by using ensembles of Genetic Programming. By using computational experiments involving more than 500 surrogate learning instances and 27,858 observations of vehicle models collected over the last thirty years shows (1) the feasibility to learn function surrogates as symbolic ensembles at different levels of granularity of the hierarchical vehicle clustering, (2) the direct relationship of the predictive ability of the learned surrogates in both seen (training) and unseen (testing) scenarios as a function of the number of cluster instances, and (3) the attractive predictive ability of relatively smaller ensemble of trees in unseen scenarios. We believe our approach introduces the building blocks to further advance on studies regarding data-driven product design and market segmentation.