Forecasting species' range shifts due to global change in climate and land-use is a big research priority. The current biodiversity patterns we know are expected to transform as species move to track their preferred climates and habitats, and this can lead to unprecedented biological invasions, extinctions, and novel interactions that can disrupt vital ecosystem services that people depend on. We thus need accurate predictions of future range shifts to inform conservation and management. To do this, we can train species distribution models on current occurrence data and project them to future climate and land-use scenarios to estimate how ranges will shift. As future conditions can be very different from current ones, our models must be good at extrapolation. But with so many variables and machine-learning algorithms at our disposal, models can get very overfit to current data, which makes model transfer less accurate. Moreover, the ways that models extrapolate to new conditions can differ considerably depending on the algorithm. In this talk, I will explain methods to address these issues with the goal of improving model transfer accuracy and reducing transfer uncertainty. Specifically, I will discuss the benefits of tuning complexity for machine-learning models, as well as a new framework to guide decisions on and control model extrapolation behavior. I will also discuss how we can leverage open and accessible ecological modeling platforms to implement these methods in user-friendly environments.