Evaluating CFSv2 Subseasonal Forecast Skill with an Emphasis on Tropical Convection

Abstract

The deterministic predictive skill of CFSv2 is analyzed with respect to lead time and temporal averaging period. We also examine the spatial distributions of error and bias for synoptic-scale parameters and the impacts of ENSO on forecast skill. The errors in a small CFSv2 ensemble forecast saturate at a lead time of approximately 3 and 5 weeks for Z500 and CHI200, respectively. Skill over climatological forecasts is restricted to lead times shorter than 2 weeks. SST, which evolves much more slowly over time, is skillfully predicted through the first month of the forecasts. Patterns of error and bias are robust across lead times and temporal scales, increasing in amplitude as lead time increases and temporal scale decreases. Several significant biases were found in the CFSv2 reforecasts, such as too little convection over tropical land regions and too much convection over oceanic regions. Simulated tropical convection patterns and associated teleconnections degrade with forecast lead time. Large-scale tropical convection in CFSv2 is more stationary than in reality. MJOs in the model propagate eastward too slowly and rarely traverse beyond the Maritime Continent. The wavenumber-frequency power spectrum for CFSv2 tropical convection is, more so than analyzed/observed spectra, largely described by its background (red) spectrum, lacks distinct peaks for the MJO and Kelvin waves, and confirms the propagation issues noted above. These flaws in tropical convection, which may be tied to mean state biases originating from cumulus parameterization, affect atmospheric teleconnections and could thus degrade extended global forecast skill.