Global Temperature Forecasting Incrementally Improved by Model Output Statistics

Skillful global temperature forecasting is crucial for mitigating the escalating impacts of rising temperature on human society and natural ecosystems. While global climate models generate invaluable dynamical temperature forecasts, the crucial role of model output statistics (MOS) in enhancing forecast skill has not been thoroughly investigated. This paper aims to unravel the potential of MOS methods for improving global temperature forecasts. It is achieved by developing a MOS toolkit to iteratively incorporate the attributes of bias, spread, trend, and association into forecast post-processing, resulting in a series of methodical one-factor-at-a-time experiments. A case study is devised for monthly forecasts of July 2-m air temperature (T2m) over land and sea surface temperature (SST) generated by the National Center for Environmental Prediction's Climate Forecast System version 2. The results expose the detrimental impacts of biases and unreliable ensemble spreads within raw temperature forecasts. At the lead time of 0 months, the continuous ranked probability skill score (CRPSS) is −128.51 ± 252.46% for T2m over land and 7.72 ± 76.66% for SST over ocean, indicating considerable underperformance of raw forecasts against reference climatological forecasts across numerous grid cells. The incremental considerations of bias, spread, trend, and association of the MOS methods result in substantial skill enhancements across global land and marine grid cells. Notably, the CRPSS of T2m is improved to 21.00 ± 23.63% and the SST forecast skill is improved to 42.26 ± 22.43%. Despite the anticipated degradation of skill with lead time, the results underscore MOS's efficacy in exploiting the information of raw forecasts to generate skillful temperature forecasts.