Long-lead runoff and soil moisture forecasts are needed for many hydroclimatological applications, such as drought outlook, agricultural planning, seasonal hydrologic prediction, and multi-purpose reservoir management. Started in January 2011, NOAA National Centers for Environmental Prediction (NCEP) has transitioned to their second generation of the Climate Forecast System (CFSv2) in operation. CFSv2 is a coupled ocean-atmosphere-land model with advanced physics, increased resolution, refined initialization, and improved land surface model, and provides forecasts up to nine months in advance. Information on the accuracy and skill of the CFSv2 forecasts is sought for the daily operation of many applications. In this study, we conduct an assessment of the soil moisture and runoff forecasts from CFSv2 using its retrospective forecasts from 1982 to 2009 to evaluate their usefulness for drought prediction. Because long-term, in-situ measurements of direct runoff and soil moisture are not available, monthly mean runoff and soil moisture reforecasts from CFSv2 are evaluated against those from the North American Land Data Assimilation System (NLDAS). NLDAS produces soil moisture and runoff from off-line land surface models driven by observed precipitation and surface temperature, and compares well with observations from the Oklahoma Mesonet, Illinois Climate Network, and USGS water archive. Therefore, it can be used as a proxy for observations.

Over the contiguous United States, soil moisture anomalies from the western interior region have long memory and often persist more than one season. When month-1 soil moisture hindcasts from the CFSv2 are evaluated against the ensemble NLDAS, skill is lower than hindcasts based on persistence, indicating that initial condition plays an important role in soil moisture forecasts. The root-mean-square-error difference between the CFSv2 month-1 soil moisture percentile hindcasts and the ensemble NLDAS is over 20% for the western interior region (west of 97°W), larger than the uncertainties of the NLDAS and the acceptable range for drought classification. A climatological analysis of the runoff and soil moisture reforecasts is also performed and compared with the Climate Forecast System Reanalysis (CFSR) and NLDAS, and shows that monthly mean runoff estimates from CFSv2 and CFSR are generally smaller than those from NLDAS. Routed streamflow from CFSv2 runoff hindcasts will be compared to river discharge observations for selected basins to further quantify the underestimation. These results suggest that direct downscaling of soil moisture and runoff from seasonal forecast models may have limited use for hydroclimatological applications. For soil moisture, forecast skill can be improved by combining forecasts of the CFSv2 and persistence of the ensemble NLDAS. For seasonal hydrologic prediction, downscaling meteorological fields with bias corrections to drive land surface models can be benefited from the increased skill of precipitation and surface temperature forecasts from CFSv2.