EOF-based linear prediction algorithm: Examples

Considered here are examples of statistical prediction based on the algorit hm developed by Kim and North. The predictor is constructed in terms of spa ce-time EOFs of data and prediction domains. These EOFs are essentially a d ifferent representation of the covariance matrix, which is derived From pas t observational data. The two sets of EOFs contain information on how to ex tend the data domain into prediction domain (i.e., statistical prediction) with minimum error variance. The performance of the predictor is similar to that of an optimal autoregressive model since both methods are based on th e minimization of prediction error variance. Four different prediction tech niques-canonical correlation analysis (CCA), maximum covariance analysis (M CA), principal component regression (PCR), and principal oscillation patter n (POP)-have been compared with the present method. A comparison shows that oscillation patterns in a dataset can faithfully be extended in terms of t emporal EOFs, resulting in a slightly better performance of the present met hod than that of the predictors based on the maximum pattern correlations ( CCA, MCA, and PCR) or the POP predictor One-dimensional applications demons trate the usefulness of the predictor The NINO3 and the NINO3.4 sea surface temperature time series (3-month moving average) were forecasted reasonabl y up to the lead time of about 6 months. The prediction skill seems to be c omparable to other more elaborate statistical methods. Two-dimensional pred iction examples also demonstrate the utility of the new algorithm. The spat ial patterns of SST anomaly field (3-month moving average) were forecasted reasonably up to about 6 months ahead. All these examples illustrate that t he prediction algorithm is useful and computationally efficient for routine prediction practices.