Abstract
Sea‐surface temperature (SST) is one of the principal factors that influence seasonal climate variability, and most seasonal prediction schemes make use of information regarding SST anomalies. In particular, dynamical atmospheric prediction models require global gridded SST data prescribed through the target season. The simplest way of providing those data is to persist the SST anomalies observed at the start of the forecast at each grid point, with some damping, and this strategy has proved to be quite effective in practice. In this paper we present a statistical scheme that aims to improve that basic strategy by combining three individual methods together: simple persistence, canonical correlation analysis (CCA), and nearest‐neighbour regression. Several weighting schemes were tested: the best of these is one that uses equal weight in all areas except the east tropical Pacific, where CCA is preferred. The overall performance of the combined scheme is better than the individual schemes. The results show improvements in tropical ocean regions for lead times beyond 1 or 2 months, but the skill of simple persistence is difficult to beat in the extratropics at all lead times. Aspects such as averaging periods and grid size were also investigated: results showed little sensitivity to these factors. The combined statistical SST prediction scheme can also be used to improve statistical regional rainfall forecasts that use SST anomaly patterns as predictors. Copyright © Crown Copyright 2003. Published by John Wiley & Sons, Ltd.