Parameters of predictive ecosystem models will never be determined exactly, either because they fluctuate in a certain range (caused by parameter aggregation or biological reasons, such as genetic evolution or species succession) or because we cannot afford the money and time necessary for a better experimental analysis. Based on this problem, Monte Carlo-simulation studies covering that fraction of parameter space assumed to be realized by the ecosystem under study, allow probabilistic predictions of ecosystem behaviour under toxic stress. Examples from the pelagic simulation model SIM-PEL demonstrate that such model predictions may be very sensitive to the range of assumed parameter variance. This also holds for purely qualitative predictions. In nonlinear models this effect increases with the time horizon of the prediction. The high computational effort to be invested in the proper analysis of these effects is a further reason to keep ecosystem models as simple as possible.