Forecast Quality of Time-Lagged, Medium-Range, High-Resolution Ensembles Over the United States: May 2023

In May 2023, the National Center for Atmospheric Research produced real-time, 204-h, 10-member ensemble forecasts with an FV3-based numerical weather prediction model that were initialized daily at 0000 UTC and had 3-km horizontal grid spacing over the conterminous United States (CONUS). Time-lagged ensembles were constructed using forecasts initialized at different times yet valid at common times. For example, 60-member time-lagged ensembles could be produced for 60–84-h forecasts. Precipitation forecasts from these time-lagged ensembles over the CONUS, east of the Rockies, were compared to the National Center for Environmental Prediction’s Stage IV precipitation analyses with a goal of identifying an optimal ensemble size and any point of diminishing returns from increasing ensemble size for predicting the top 1% of precipitation events. Ensemble spread increased greatly in 20-member time-lagged ensembles relative to the 10-member non-time-lagged ensemble, while increasing time-lagged ensemble size beyond 20 members increased spread more slowly. Reliability statistics and Brier scores showed that time-lagged ensembles with 20–30 members yielded better forecast quality than 10-member non-time-lagged ensembles, while adding more time-lagged ensemble members provided limited benefits. These collective results suggest that time-lagged ensemble sizes of 20–30 members may represent the point of diminishing returns, with further increases in time-lagged ensemble size potentially being suboptimal in time-lagged ensembles constructed by forecasts initialized 24-h apart.