Cyclone forecasts can locate storms while missing strength

When a forecast gives several possible cyclone paths, does its average path tell us whether the observed storm was among those possibilities? A useful comparison must also ask whether a model places the storm’s strongest winds and lowest pressures well, and whether it gets their intensity right.

The benchmark scores ensembles by the probability they assign to observed cyclone paths, then compares location, intensity, structure, and wind-pressure dependence.

The authors tested whether probabilistic cyclone forecasts represent observed paths and storm structure, not just average positions.

They compared three AI weather models with the operational physics-based IFS-ENS ensemble across selected historical storms, using observed tracks and gridded references.

In the 2025 setting, FGN placed more probability on observed locations than IFS-ENS on about 90% of 41 retained storms. Its percentile-defined spatial scores were also strongest, but FGN and GenCast still underestimated peak winds relative to HRES-fc0 at short and intermediate leads.

The evidence supports better location and relative placement on these cohorts, not uniformly better cyclone forecasts.

These are retrospective comparisons of selected known storms, not tests of whether models predict cyclone occurrence or are ready for operational warnings. The location scores include only forecast members whose detected tracks matched an observed storm.