Exploring the possibility to Forecast Annual Mean Temperature with IPCC and AMIP Runs
The empirical forecast methods used here include persistent forecast, flat extension and sloped extension. The persistent forecast is a one-year persistent extension of bias corrected model data. The flat extension takes the latest n (n> 1) year average as the forecast for the target year. The sloped extension make a linear regression over the latest n year model data, then do a sloped extension to target year. The optimal window length n in both flat and sloped extensions corresponds to the highest forecast skill over the verification period. The result will be taken as a bench mark to evaluate other prediction methods such as the proposed initialized decadal prediction efforts.
Forecast tests are first conducted for area averaged annual mean land temperatures. For the global and tropical averages, the persistent forecast and sloped extension with IPCC data have almost the same skill as the flat extension based on the observational data, while for the average over North America area, the former are slightly better than the latter. For all the three area averages, the skill of AMIP data based forecasts is found to be the lowest. Forecast tests are then conducted for each grid points on land. Comparing the obtained skill maps, we find that the IPCC data based forecasts can beat the flat extension based on the observational data over several regions in middle and high latitudes, but this is not the case for the AMIP data based forecasts.