An investigation of the El Niño‐Southern Oscillation cycle With statistical models: 2. Model results

We have constructed two sets of linear statistical models for predicting equatorial Pacific sea surfaces temperatures (SSTs) from the tropical wind field and from the near‐global sea level pressure (SLP) field. The wind models were constructed to encompass the domains relevant to the forcing of trapped equatorial waves associated with the El Niño. Thus they provide a measure against which the performance of some simple physical models may be compared. The SLP models were constructed to better define the roles of the observed motion of SLP features and changes in the Southern Oscillation as they relate to subsequent equatorial SSTs. The wind models were designed to provide forecasts at lead times of 1 to 5 months, and the SLP models were designed to provide forecasts at lead times of 7 to 16 months. The models were constructed using a combination of extended empirical orthogonal functions and canonical correlation analysis, a technique we found well suited to the handling of large data sets and the analysis of model structures. Forecast skills were estimated by cross validation. The results show that the wind field models do well at predicting eastern and central Pacific SSTs during much of the year, and their structure is in general agreement with the idea that Kelvin/Rossby waves initiated by remote wind forcing are responsible for much of the observed SST signal. However, the models do not track the onset of warm and cold events well, particularly in the eastern Pacific. This suggests that well‐organized, consistent wind anomalies are not responsible for the initial temperature anomalies. The SLP models did not do well in least squares sense but performed skillfully when applied with a simple El Niño‐no El Niño decision rule at lead times of up to 16 months.