AbstractAimEcological niche modelling (ENM) or species distribution modelling is increasingly used in decision‐making regarding land use and biodiversity conservation. Model accuracy is essential, but can be affected by modelling choices, including the critical and ubiquitous question of how to define a model training domain. Theories have suggested designing a training domain based on areas accessible to a species for improved model performance (here termed as species’ mobility hypothesis). However, we still lack direct quantitative evidence on whether this approach leads to optimal model performance. Here, I conducted a modelling experiment to investigate the species' mobility hypothesis.LocationNorth and South America.TaxonHummingbirds (Aves: Trochilidae).MethodsThe modelling experiment was based on 87 hummingbird species. A series of spatial buffers (from 5 to 5000 km with varying intervals) were created around occurrences, where background points were sampled and used as input for model calibration. The models calibrated with spatial buffers were compared with models calibrated with training domains that considered areas accessible to species, using Boyce index and sensitivity, specificity and true skill statistics.ResultsModel performance increased when the size of the training domain was larger, although the model performance reached saturation when size of the training domain passed a certain threshold. The threshold varied by species and evaluation method and was generally estimated to be below 200 km. The model performance based on areas accessible to species was comparable (e.g. non‐significant difference in sensitivity) to the saturation performance of models when spatial buffers were used.Main conclusionsPositive evidence was found to support the species' mobility hypothesis that designing a training domain based on areas accessible to species could lead to optimal or near‐optimal model performance. When no information on the accessible area is available, modellers may use a tuning strategy to identify the size of the training domain for optimized model performance.