Large ensembles of prediction experiments were conducted using a coupled ocean-atmosphere model consisting of a reduced gravity ocean and a simple atmospheric feedback model. By varying the coupling strength we maneuver the system through a wide range of dynamical regimes, i.e., stable, unstable and chaotic regimes, in the presence of various stochastic forcing. The analysis of the time evolution of the ensembles suggests that the stochastic forcing has a considerably larger impact on ENSO predictions on short lead times up to about nine months than the nonlinear dynamics. However, at longer lead times the effect of chaotic dynamics is reflected in the time evolution of the ensembles. ¿ American Geophysical Union 1996