Blending NWP forecasts for the prediction of convective hazards

The spatial resolution of current operational Numerical Weather Prediction models is not sufficient to resolve individual hailstones or tornadoes. Even convective severe wind events are often only partially captured. Where such phenomena are resolved, the accuracy of individual forecasts remains limited due to rapid error growth after initialization, which motivates the use of ensembles.Forecasting of such hazardous convective phenomena therefore depends on identifying resolved predictors that correspond to events not explicitly resolved in the available ensemble forecast. To support such efforts on the medium-range (3 – 10 day) forecasting timescale, the European Severe Storms Laboratory (ESSL) and ECMWF adapted ESSL’s additive regression models (AR-CHaMo) for use with the Integrated Forecast System (IFS) ensemble. The IFS provides forecasts with a spatial resolution equivalent to a 9 km horizontal grid.For shorter timescales, forecasting efforts rely heavily on convection-permitting ensembles. These models benefit from higher spatial resolution, which eliminates the need to parameterize deep, moist convection. With Geosphere Austria, ESSL evaluated the performance of the C-LAEF ensemble forecasting system in predicting large hail and severe wind gusts up to 48 hours in advance. In this system, large hail occurrence was inferred from vertically integrated graupel. The C-LAEF ensemble configuration used for this study operates at a horizontal grid spacing of 2.5 km.The next step in our work was to determine how best to blend the coarser medium-range forecasts with the finer short-term forecasts. To achieve this, we evaluated forecast skill by each system for hail days during the summers of 2024 and 2025 and present an approach that achieves optimal skill by blending them based on lead time. In addition, we will outline the limitations and challenges we encountered, including cases where both systems provided poor forecasts. We invite feedback to help refine this approach further.