An Empirical Benchmark for Decadal Forecasts of Global Surface Temperature Anomalies

The suitability of an empirical multivariate AR1 model as a benchmark for the skill of decadal surface temperature forecasts is demonstrated. Constructed from the observed simultaneous and one-year lag correlation statistics of 12-month running mean sea surface temperature (SST) and surface (2m) land temperature global anomalies for the years 1900-2008, the empirical model hindcasts have skill for leads 2-5 and 6-9 years comparable to and sometimes even better than the CMIP5 model hindcasts initialized annually over the period 1960-2000, and are much more skillful than damped persistence (e.g., a local univariate AR1 process). The pronounced similarity in geographical variations of skill between the empirical model and CMIP5 hindcasts suggests similarity in their sources of skill as well, supporting additional evaluation of the empirical model's skill and predictability over the entire record. It is shown that for forecast leads greater than about a year, the empirical model skill is almost entirely due to patterns corresponding to the secular trend and to two global patterns that each have about ten year decorrelation time scales. In the Atlantic, all three patterns contribute to forecast skill of the Atlantic Multidecadal Oscillation (AMO) index. In the Pacific, only one pattern contributes to the relatively modest long-lead forecast skill of the Pacific Decadal Oscillation (PDO) index, consistent with earlier findings that determined an independent decadal signal in the PDO as a residual after both interannual and decadal ENSO influences were first removed. This pattern is particularly poorly forecast by the CMIP5 models relative to the empirical model, suggesting that substantial room for improvement remains in Pacific decadal SST forecasts and their North American response. Overall, these results support the view that multivariate red noise rather than univariate red noise is the most appropriate baseline comparison for coupled model decadal forecasts.