##
4A.5
Recent slowing of global warming and real-time forecasts for the year ahead, 2000–2010

When this independent period is 1998-2008, the method reproduces the observed interannual variations of global temperature very well. It therefore well reproduces the recent slowing of global warming. However we note that because a small amount of explained variance is lost in multiple regression, arguably the best estimate of a decadal trend uses “inflated” regression which replaces lost variance. We show both estimates. We particularly show that the slowest rate of observed decadal temperature change from 1998 onwards as measured by HadCRUT3 is consistent with future projections of global warming to 2100 under strongly enhanced greenhouse forcing. For this we use a large number of ensemble members from a range of transient integrations of IPCC coupled climate models, and a set of transient integrations of perturbed HadCM3 coupled climate models.

The statistical model has been adapted to forecast global surface temperature for the year ahead using data available in the previous December. A variation on the statistical model uses forecasts of Nino 3.4 SST from the Glosea3 coupled seasonal forecast model to replace the statistical forecasts of this quantity. Using a jack-knife cross validation technique, both models have been thoroughly tested for their true forecast skill. The forecasts include “inflated” regression so that the standard deviation of the predictions is the same as that of the observations. However, the inflation effect is small as the jack-knife correlation is around 0.93. Finally in the last few years dynamical predictions of global mean surface temperature for the coming year from the DePreSys coupled decadal forecast model have been included

We then assess the skill of the ten real-time forecasts for the year ahead issued in Press Releases for the period 2000-2009. These use a combination of all available forecast techniques. Each forecast includes an uncertainty that allows probability forecasts of global mean surface temperature to be issued, though we do not test these here. The two statistical models have changed slightly since 2000, and were calibrated against earlier versions of Met Office global surface temperature data sets, so we verify against these contemporary data. The skill of the ten real-time forecasts is shown to be high, with a correlation coefficient between the observed and forecast temperatures of 0.73. This value is consistent with estimates of the inherent skill of the statistical forecasts, which of course is less than that of the simulation model above, which includes observed data for the year being forecast. The mean absolute error of the forecasts is about 0.06oC, including an overall small warm bias.

Finally we discuss the forecast for 2010, issued in early January 2010, in the light of the expected influence of the current El Nino and published DePreSys forecasts of the likelihood of several record warm years after 2009.