Abstract: Plankton play an important role in marine food webs, in biogeochemical cycling, and in moderating Earth's climate. Their possible responses to climate change are of broad scientific and social interest; yet observations are sparse, and mechanistic and statistical methods yield diverging predictions.

Here, we evaluate a statistical learning method using output from a 21st Century marine ecosystem model as a 'ground truth'. The model is sampled to mimic historical ocean observations, and Generalised Additive Models (GAMs) are used to predict the simulated plankton biogeography in space and time. Predictive skill varies across test cases, and between functional groups, and errors are more attributable to spatiotemporal sampling bias than to sample size.

Overall, the GAMs yield poor end-of-century predictions. Given that statistical methods are unable to capture changes in relationships between variables over time, we advise caution in their application and interpretation, particularly when modelling complex, dynamic systems.