Advances in our ability to model complex environmental systems are currently driven by at least four needs: (1) the need for the inclusion of uncertainty in monitoring, modelling and decision-making; (2) the need to provide environmental predictions everywhere; (3) the need to predict the impacts of environmental change; and (4) the need to adaptively evolve observation networks to better resolve environmental systems and embrace sensing innovations. Satisfying these needs will require improved theory, improved models and improved frameworks for making and evaluating predictions. All of these improvements should result in the long-term evolution and improvement of observation systems. In the context of this paper we discuss current bottlenecks and opportunities for advancing environmental modelling with and without local observations of system response. More realistic representations of real-world thresholds, nonlinearities and feedbacks motivates the use of more complex models as well as the consequent need for more rigorous evaluations of model performance. In the case of gauged systems, we find that global sensitivity analysis provides a widely underused tool for evaluating models' assumptions and estimating the information content of data. In the case of ungauged systems, including the modelling of environmental change impacts, we propose that the definition of constraints on the expected system response provides a promising way forward. Examples of our own work are included to support the conclusions of this discussion paper. Overall, we conclude that an important bottleneck currently limiting environmental predictions lies in how our model evaluation and identification approaches are extracting, using and evolving the information available for environmental systems at the watershed scale.