An El-Nino prediction model is developed based on an intermediate ocean model similar to the Cane and Zebiak (CZ) and a statistical atmosphere model. The present ocean model differs from CZ in the parameterization of subsurface temperature and the basic state. The predictability skill of the present model is better than that of CZ. The better performance is particularly distinctive for early stage of the prediction everywhere in the domain and in the central Pacific for all period of prediction. It is suggested that the better performance for the early stage is due to the use of SST anomalies in the initialization, and the better performance in the central Pacific results from a better representation of subsurface temperature in the present model. ¿ 2000 American Geophysical Union