Most dynamic models of ecological communities require many parameters, making them expensive to fit to experimental or observational data. To reduce the number of parameters, species are often divided into groups a priori, using phylogenetic or functional groups, and species within a group are assumed to behave identically. Here, we assess the validity of such grouping by deriving the optimal groupings a posteriori based on the fit of the Beverton-Holt model to experimental data from six species of Drosophila competing in laboratory conditions. We find evidence that grouping can indeed improve the parsimony of species interaction models, but we also observe that there is no single optimal grouping. Further, those groupings that did prove beneficial could not have been easily predicted from available species attributes, calling into question the prevailing method of a priori species grouping. Lastly, we find that species did not group equivalently in their competitive effect and in their response to competition, invalidating a key assumption often made implicitly when defining species groupings. Our study suggests the need for a different, more data-informed, approach to grouping in species interaction models.