A statistical forecast model for tropical SST, sea level and precipitation: predictability of warm pool area

This study concerns the prediction of tropical SST, sea level and precipitation anomalies associated with ENSO by a statistical method. The monthly mean fields of observed sea surface temperature (SST) and precipitation in the global tropics, and observed winds and sea level analysis in the tropical Pacific are combined through a multivariate principal component analysis [producing multivariate empirical orthogonal functions (MEOFs)]. Then a multivariate first-order autoregressive process with seasonally varying parameters is fitted to the first few (5-6) MEOF time series.

The model, referred to as Markov model, is trained for the period of 1981-98. The hindcast skill in the training period is estimated with a cross-validation scheme. For SST, the correlation skill is highest in the central-eastern equatorial Pacific, which is above 0.5 at 9 month lead. The second highest correlation skill is in the north-western Pacific at about 0.5 at 5 month lead. For sea level, the correlation skill is highest in the central equatorial Pacific and western Pacific at above 0.5 at 9 month lead. Generally speaking, in the western Pacific sea level anomalies are more predictable than SST anomalies.

The hindcast skill of monthly precipitation is expected to be lower than those for SST and sea level since the model does not simulate Madden-Julian Oscillations which have time-scales of 30-70 days. The correlation skill, calculated for seasonal mean precipitation, is highest in the central-eastern equatorial Pacific and maritime continent at about 0.5 at 6 month lead (the lead is the number of months between the center of forecast seasons and initial times). There is a useful skill (0.4) in the northern Brazil up to 6 month lead. We will compare the statistical forecasts with those of numerical models, and discuss their implications for seasonal forecast.

The hindcast skill is largely account for by the first two MEOFs which represent more than 40% of the combined variance for SST, sea level, precipitation and winds. The first MEOF represents the mature phase of ENSO, and the second MEOF proceeds the first MEOF by 7-9 months so it represents the set up phase of ENSO. We will discuss the structures of the two MEOFs and their implications for predictability of ENSO.

Singular vector analysis is applied to the Markov model to study the optimal growth for ENSO. We will explore the hypothesis that Madden-Julian oscillations act as optimal growing perturbations to influence ENSO evolution.