Matrix models have been widely used to predict forest dynamics from short-term, “snapshot” census data, but their accuracy has rarely been tested with long-term data on forest change. Using data collected over nearly a 30-year period (1982-2010) in a 50 ha permanent plot on Barro Colorado Island (BCI) in Panama, we tested the accuracy of predictions of matrix models constructed from data from the initial 1982 census. Projections in these models were based on the relative frequencies of saplings of species i beneath adults of species j. Over the years, species abundance changed substantially for some species, with 46% increase for one species, and 71% decrease in another. Of 10 common canopy and midstory tree species tested, changes in the abundances of only 2 species were accurately predicted. The accuracy of projections of subsequent tree species abundance depended on species and the sapling size class used in making the prediction. Our results question the general validity of assumptions of matrix forest projection models.