Evaluation of systematic biases and skill of summer subseasonal forecasts in the ECMWF system

Subseasonal forecasts (ranging between 2 weeks and 2 months) have been the subject of attention in many operational weather forecasts centers and by the research community in recent years. This growing attention stems from the value of these forecasts for society and from the scientific challenges involved. The scientific challenges of capturing and representing key processes and teleconnections which are relevant at these scales are significant. One example is temperature extremes associated with weather extremes like heatwaves and droughts that can have severe consequences in nature and human health, among others. Some of the limitations in forecast skill arise from the limits of predictability of the chaotic earth system. Model error is also likely to play a relevant role. In this study, we investigate systematic model biases, their evolution with lead time and potential links with forecast skill.

This study assessed the skill and biases of the European Centre for Medium-Range Weather Forecasts (ECMWF) subseasonal forecast in predicting the daily temperature extremes in the Northern Hemisphere. These forecasts are from an experimental setup of ECMWF extended-range forecast system. The forecasts compromise 11 ensemble members with weekly starting dates between 9 April to 30 July extending up to 6 weeks lead with a 20-years hindcast period (1998-2017). The forecasts were performed by the coupled ECMWF systems with TcO199 horizontal resolution (about 50km) in the atmosphere and 1x1 degree ocean. A particular focus is given to Europe and to two other regions that were identified with large systematic errors. The data used in this work consisted of the daily maximum and minimum two-meter temperature, precipitation and other surface fluxes that are aggregated into weekly means and verified against ERA5. 

The evaluation of systematic biases in daily temperature extremes shows a clear increase with lead time, which is widespread on a hemispheric scale. The spatial patterns of model error growth with lead time are reasonably similar between daily maximum and minimum temperatures. However, the amplitude of the errors is remarkably different with general cold bias of daily maximum and warm bias of daily minimum that consistently grow with forecast lead time. Despite the consistent error growth with lead time, there are clear differences between the forecasts initialized in late Spring (April-May) and those in Summer (June-July). These biases are not fully collocated in two regions in the Northern Hemisphere showing the largest warm temperature biases: Central US and East of Caspian Sea. The warm biases are consistent with an underestimation of precipitation and dry soil moisture, compared to ERA5, but only over the East Caspian region.  Forecasts skill assessed via the anomaly correlation shows that the temperature forecasts are skillful up to week 2, with a drop in skill from week 3 onwards. This drop in skill is consistent over all the European domain. Similar results are found for precipitation, but with ACC at week 2 comparable with those of temperature at week 3.