Carbon cycle models have become increasingly complex over time, as more and more processes are represented. But as models get more complex, the risk that they are mis-parametrized becomes increasingly large. We used CARDAMOM with a variety of different model structures, data assimilation types and errors, and more to systematically assess whether increasing model complexity always improves forecast performance (including accounting for prediction uncertainty). In fact, intermediate complexity models often perform better than more complex models. Increased complexity only improves forecast skill if net carbon flux data is available to sufficiently constrain the parameters. This suggests terrestrial biosphere models should focus more on constraining parameter uncertainties rather than simply increasing the number of processes represented. A list of other specific recommendations is also included in the paper, which can be read (or commented on!) here.