A T-EOF based prediction method

A new statistical time series prediction method based on temporal empirical orthogonal function (T-EOF) is introduced in this study. This method first applies singular spectrum analysis (SSA) to extract dominant T-EOFs from historical data. Then, the most recent data are projected onto an optimal subset of the T-EOFs to estimate the corresponding temporal principal components (T-PCs). Finally, a forecast is constructed from these T-EOFs and T-PCs. Results from forecast experiments on the El Nino sea surface temperature (SST) indices from 1993 to 2000 showed that this method consistently yielded better correlation skill than autoregressive models for a lead time longer than 6 months. Furthermore, the correlation skills of this method in predicting Nino-3 index remained above 0.5 for a lead time up to 36 months during this period. However, this method still encountered the "spring barrier" problem. Because the 1990s exhibited relatively weak spring barrier, these results indicate that the T-EOF based prediction method has certain extended forecasting capability in the period when the spring barrier is weak. They also suggest that the potential predictability of ENSO in a certain period may be longer than previously thought.