Global Sea Surface Temperature Forecasts Using a Pairwise Dynamic Combination Approach

Abstract

Abstract This paper dynamically combined three multivariate forecasts where spatially and temporally variant combination weights are estimated using a nearest-neighbor approach. The case study presented combines forecasts from three climate models for the period 1958–2001. The variables of interest here are the monthly global sea surface temperature anomalies (SSTA) at a 5° × 5° latitude–longitude grid, predicted 3 months in advance. The forecast from the static weight combination is used as the base case for comparison. The forecasted sea surface temperature using the dynamic combination algorithm offers consistent improvements over the static combination approach for all seasons. This improved skill is achieved over at least 93% of the global grid cells, in four 10-yr independent validation segments. Dynamically combined forecasts reduce the mean-square error of the SSTA by at least 25% for 72% of the global grid cells when compared against the best-performing single forecast among the three climate models considered.