Do European models of temperate forest ecological change apply in North America?

The transferability of ecological models, especially those based on machine learning approaches, needs to be thoroughly tested to predict beyond the range of the training data. We tested the cross-continental transferability of three machine learning models of forest understorey dynamics trained on data from one temperate region (central-western Europe) to determine how reliable they are for predicting changes in upland forest sites in southern (n = 83) and northern (n = 74) Wisconsin, USA. We tested trajectories of species richness and the proportions of woody species and forest specialists under the influence of global change (i.e. changes in temperature, precipitation, and nitrogen deposition) and local forest management. Among the three tested models, only one (the model predicting species richness) generated useful predictions. Such low success suggests that distinctly different environmental contexts or the absence of key biotic and/or abiotic predictors likely impeded model performance. Although we cannot recommend applying these models to regions beyond temperate Europe, including more predictor variables, tuning features, and performing spatial cross-validation could improve power and transferability in future models.