Experimental ecological communities are important tools for exploring the processes that promote and maintain biodiversity. Communities assembled in controlled conditions afford much greater ease of manipulation, compared to field-based systems, allowing for the construction of multispecies communities of varying composition. Predicting the outcomes of species interactions in such complex communities remains challenging, however, largely due to the difficulty of identifying and parameterizing dynamical models. In many cases, the goal of parameterizing dynamical models is not to predict dynamics per se, but rather to make some inference about an outcome of the system, such as long-term diversity, coexistence, stability, or invasibility. Modeling complete dynamics is often seen as a necessary first step to characterizing these properties. But given the challenges inherent in fitting dynamical models, and the ease of manipulating and quantifying (e.g. via high-throughput sequencing) experimental communities with high replication, we ask whether data only from experimental endpoints is sufficient to predict unobserved community outcomes. In particular, we introduce a novel yet straightforward statistical approach to predicting coexistence and species abundances in experimental ecological communities, using only endpoint data. This approach obviates the need for expensive time-series, instead leveraging the tractability of small experimental communities. We consider many communities drawn from a pool of n species. Given the endpoint abundances of some of these communities, can we predict the abundances in unobserved ones? We approach this problem by directly relating endpoint abundances, using a simple statistical model that is agnostic to the true dynamical model underpinning the system. We examine three independent datasets - a plant system, a protist system, and an algae-Daphniidae system - to test the efficacy and robustness of our method. Crucially, because our model requires relatively little data to fit, we are able to test its predictions against experimental outcomes not used for fitting. Foregoing dynamics, we are able to predict out-of-fit endpoints with high accuracy (R2 = 0.99, 0.89, 0.86 for predicted vs. observed in each system, respectively). Additionally, we show that our method can offer insights into stability, invasibility, extinction. Finally, our method suggests new principles for the design of efficient competition experiments. Taken together, our results indicate that endpoint data can be surprisingly informative about the structure of ecological communities, and can suggest a principled way to assemble large experimental communities, helping ecologists probe the mechanisms that promote coexistence in natural and synthetic ecosystems.