The issue of whether models developed for current conditions can yield correct predictions when used under changed control, as is often the case in environmental management, is discussed. Two models of different complexity are compared on the basis of performance criteria, but it appears that good performance at the calibration stage does not guarantee correctly predicted behavior. A requirement for the detection of such a failure of the model is that the prediction uncertainty range is known. Two techniques to calculate uncertainty propagation are presented and compared: a stochastic first-order error propagation based on the extended Kalman filter (EKF), and a newly developed and robust Monte Carlo set-membership procedure (MCSM). The procedures are applied to a case study of water quality, generating a projective forecast of the algal dynamics in a lake (Lake Veluwe) in response to management actions that force the system into a different mode of behavior. It is found that the forecast from the more complex model falls within the prediction uncertainty range, but its informative value is low due to large uncertainty bounds. As a substitute for time-consuming revisions of the model, educated speculation about parameter shifts is offered as an alternative approach to account for expected but unmodelled changes in the system.