Evolving Systematic Errors in Precipitation Forecasts from ECMWF

In mid-2026 ECMWF introduced into operations a new cycle (50r1) of its physics-based IFS (Integrated Forecast System). One key feature of 50r1 was an upgrade to the convection scheme. This is particularly beneficial for bringing SST-triggered convection inland across coastal regions, but also has positive displacement impacts over and downstream of other convectively active areas. These improvements were achieved, in model code, primarily by commuting a proportion (typically 40%) of precipitation particles from the convection scheme into the large scale scheme. That 40% can then advect with the wind; in the previous cycle it fell out instantaneously. This change has had a big impact, worldwide, on the ratio of convective to total precipitation (cpr). Now small values of cpr are much more common, and vice versa.

With all this and other model changes in mind a comprehensive, supervised ML-style re-assessment has been made of relationships between short range forecast rainfall (F) and observed (gauge-based) rainfall (O), for cycle 50r1, using ranges of cpr and other variables to create ‘weather type’ classes. This is revealing, showing a large range of noteworthy situation-dependant biases, which are useful for both forecasters and model developers. For example, there is a diurnal cycle in bias, with afternoon rainfall systematically over-predicted by the IFS, whilst night-time rainfall is less erroneous. This bias difference is amplified in CAPE-rich environments. Meanwhile, in some different scenarios, the model consistently overpredicts rainfall by a large margin (O/F ~ 0.25), whilst in others, it markedly underpredicts (O/F ~ 3). We have found cpr, CAPE and low level dewpoint depression to be key predictors. Numerous results of this assessment will be presented in this talk. The results are based on over 10 million O:F pairs. Moreover with the new convection scheme advecting precipitation into regions where previous cycles would have failed, we now have the capacity to improve those forecasts even further, to account for inadequacies in the 40% transfer factor incorporated.

For users, ECMWF can now provide related situation-dependant bias-corrected forecast products, via the "ecPoint" post-processing approach (along with estimates of sub-grid variability). This post-processing and the related conditional verification calibration activity discussed above were all streamlined in short time in 2026 by using agentic AI to (i) convert a pre-existing containerised calibration tool into an open-source (localhost) web application, (ii) markedly accelerate processing speed therein and (iii) add numerous other user-oriented functions to it. With this tool others can now do similar assessments with their own models. A brief overview of this development will be given.