Forecasts of the tropical Pacific SST anomalies associated with ENSO are made with a linear statistical model (Markov model). The model is built in a reduced EOF space with three multivariate EOFs (MEOFs) of sea surface temperature (SST), sea level and wind stress in the tropical Pacific 20OS-20ON. It evolves linearly with a seasonally dependent and predetermined transition matrix. Since the model is trained for the 1980-1995 period, the retrospective forecast since 1996 can be viewed as the "real time" forecast, although the forecast did not become operational until September 1998 when it was first published in the Experimental Long-Lead Forecast Bulletin. The Markov model successfully forecasted the 1997/1998, 2002/2003 El Nino events, and the 1998/1999, 1999/2000, and 2000/2001 La Nina events. Its performance has been rated among the best. We contribute most of the skill to the precursor signal in the second Principal Component of MEOF that leads the NINO3.4 index by 9-12 months. The second MEOF is characterized by a north-south dipole sea level with anomalously high (low) sea level south (north) of 5ON, positive SST anomalies in the central equatorial Pacific and westerly wind anomalies to the west and easterly wind anomalies to the east of the warm SST anomalies. The third MEOF is also important since it contains anomalously high sea level across the equatorial belt in 5OS-5ON, which is critical for ENSO development according to the recharge and discharge ENSO theory. Although sea level is the most important predictor, a comparison with a Markov model built with SST and sea level only suggests that wind stress acts to suppress the false alarms that might have otherwise been made by the model built with SST and sea level.

The Markov model has successfully forecasted all the warm events since 1980. The forecast for the onset of the 2002/2003 El Nino was made as early as November 2001, which was one of the best forecasts for the event. However, the model's forecast at long lead usually lags and underestimates the onset phases of warm events. The weakest point of the model is that it does not simulate well the transition phases from moderate warm to moderate cold events, for example the transition from the 1986/87 warm to 1988/89 cold events and from the 1994 warm to 1995/96 cold events. We suspect this is related to the short time scales of those transitions that are missed by the model due to its dependence on the slow movement of sea level. We plan to conduct a series of sensitivity experiments to study the dependence of the forecast skill on data domain, data weight and new data. Particularly, we plan to use the new global ocean analysis for 1979-2001 that was produced with the global ocean data assimilation system at NCEP recently. We will investigate the impact of the Indian Ocean on the predictability of ENSO.