The two latest changes introduced during 1998 into the Ensemble Prediction System (EPS) operational at the European Centre for Medium-Range Weather Forecasts (ECMWF) are described. The first change, the inclusion of instabilities growing during the data assimilation period in the generation of the EPS initial perturbations, increased the probability that the analysis lies inside the ensemble forecast range. The second change, the introduction of a simulation of random model errors due to parametrized physical processes, improved in particular the prediction of precipitation. The performance of the ECMWF Ensemble Prediction System from 1 May 1994 to 2 March 1999 is assessed using different statistical measures. Results indicate that the general performance of the EPS has been improving over the years. Finally, ongoing research projects on predictability issues developed either at ECMWF or at European research institutes in collaboration with ECMWF are discussed