This study considers the problem of marine ecological prediction in the context of online estimation and forecasting. Process oriented dynamic ecosystem models are combined with marine observations. The nonlinear, nonGaussian state space model provides the statistical framework. The associated filtering (nowcasting) and prediction (forecasting) problems are addressed via sequential Monte Carlo methods, in this instance a sequential importance resampler combined with Metropolis-Hastings MCMC. The specific focus is on a prototypical marine ecosystem model comprised of four interacting populations (phytoplankton, zooplankton, nutrients and detritus; PZND) whose co-evolution is described by system of coupled nonlinear differential equations. Stochastic environmental variation is introduced through a stochastic growth parameter, as well as through dynamical noise in the state evolution equations. The dynamic behaviour of this stochastic ecosystem model is complex: it regularly transitions through a Hopf bifurcation and exhibits episodic blooms of variable magnitude and duration. The model is applied to a case with weak seasonality, that is the oceanic mixed layer in the eastern equatorial Pacific. A partially observed state is considered comprised of a five year satellite (SeaWiFS) derived time series of ocean phytoplankton concentration at 12°N 95°W. Filtering estimates for the ecosystem state and a dynamic parameter were obtained using the sequential Monte Carlo approach. These showed predictor-corrector behaviour at observation times, including abrupt shifts in the median level after forecasts over measurement void. A corresponding variance (also skewness and kurtosis) growth and subsequent collapse was also seen. Forecasting experiments indicate some negative bias, and suggest there is predictive skill for forecasts out to 10–15 days. Copyright © 2005 John Wiley & Sons, Ltd.