Newsletter No. 150 banner

ERA5 aids in forecast performance monitoring

Thomas Haiden
Martin Janousek
Hans Hersbach


ECMWF’s fifth-generation reanalysis, ERA5, provides a new reference for quantifying gains in forecast skill in the Integrated Forecasting System (IFS). For more than a decade ERA-Interim has been used to separate variations in atmospheric predictability from changes in predictive skill due to improvements of the forecasting system. However, the 9 km resolution of the high-resolution forecast (HRES) is now almost an order of magnitude higher than that of ERA-Interim (80 km), and the two versions of the forecasting system (IFS Cycles 43r1 and 31r2) differ substantially. Furthermore, ERA-Interim cannot respond to some of the changes in the observing system that have occurred over the years. This has led to an increasing divergence of the ways in which the two systems respond to variations in predictability, especially for near-surface parameters such as 2 m temperature and 10 m wind speed, and it adds some uncertainty to the assessment of long-term trends in HRES skill.

Reduction in unexplained variance
Reduction in unexplained variance. ERA5 is better than ERA-Interim at capturing the day-to-day variance in IFS skill scores. For 2-metre temperature at forecast day 5 in the extratropics, the 'unexplained' variance of the RMSE for recent IFS cycles is reduced from 20% in ERA-Interim to 12% in ERA5.

ERA5 uses IFS Cycle 41r2, which was operational until 21 November 2016, with a horizontal grid spacing of 31 km. It uses 137 levels in the vertical, which is the same as the current HRES. For more details on ERA5, see ECMWF Newsletter 147. Since ERA5 is much closer to the current operational configuration than ERA-Interim, it provides an improved means of assessing the effect of variations in atmospheric predictability on HRES scores.

One way of quantifying the ability of a reference forecast to capture the effect of atmospheric variability on HRES skill is to determine the correlation of daily scores between the two forecasts. Note that when the correlation is computed over a period which contains one or more model upgrades, the mean value of the score over each sub-period covering the respective cycles is subtracted first. This ensures that the correlation measures the degree of correspondence relative to atmospheric variability and is not affected by step-wise changes in skill due to model upgrades.

Results for 2-metre temperature at forecast day 5 in the extratropics, using the root-mean-square error (RMSE) as a metric, show that ERA-Interim captures 80% of the day-to-day variance in HRES scores, while for ERA5 this number increases to 88%. This reduction of the ‘unexplained’ variance from 20% to 12% facilitates the interpretation of the time evolution of operational scores.

These first results are based on a limited period of ERA5 forecasts (June 2014 to October 2016). When data for a longer period becomes available in 2017, more comprehensive tests will be possible, and ERA5 will eventually replace ERA-Interim as the main reference forecast in the operational evaluation of forecast skill.