Data for: Evaluating Extreme Precipitation Forecasts: A Threshold-Weighted, Spatial Verification Approach for Comparing an AI Weather Prediction Model Against a High-Resolution NWP Model

This dataset provides the necessary data to reproduce Loveday & Hertneky (2026) https://egusphere.copernicus.org/preprints/2026/egusphere-2025-5796/ , which introduces a verification framework combining the High-Resolution Assessment (HiRA) neighbourhood method with threshold-weighted proper scoring rules (twCRPS) to evaluate extreme precipitation forecasts. The framework enables comparison of models with differing spatial resolutions against station-based observations without requiring re-gridding, and is demonstrated over 32 months of forecasts across the contiguous United States (CONUS). Forecast data The models data span 973 daily 00 UTC runs from 1 January 2022 to 30 August 2024. Data is limited to lead times of 0–48 hours (the maximum range of the HRRR 00 UTC run), and six-hourly precipitation accumulations are used throughout. The forecast data is limited to the neighbourhood data required to replicate the results. HRRR neighbourhood data The High-Resolution Rapid Refresh (HRRR) version 4 is a convection-allowing, cloud-resolving numerical weather prediction (NWP) model operated by NOAA with a horizontal grid spacing of 3 km. Full details of the model are given in Dowell et al. (2022). Data was originally retrieved from https://hrrrzarr.s3.amazonaws.com/index.html. hrrr_1.zip — HRRR neighbourhood data for a 1×1 grid-point neighbourhood hrrr_7_9.zip — HRRR neighbourhood data for 7×7 and 9×9 grid-point neighbourhoods hrrr_21_27.zip — HRRR neighbourhood data for 21×21 and 27×27 grid-point neighbourhoods GraphCast-GFS neighbourhood data GraphCast-GFS is the AI-based weather prediction (AIWP) model GraphCast initialised with NOAA Global Forecast System (GFS) analysis fields, using the reforecast archive produced by Radford et al. (2025). The model operates at a grid resolution of 0.25°. Evaluation commences in January 2022, after the end of the period used in the operational GraphCast fine-tuning. Data was originally retrieved from https://noaa-oar-mlwp-data.s3.amazonaws.com/index.html. graphcast-gfs_1.zip — GraphCast-GFS neighbourhood data for a 1×1 grid-point neighbourhood graphcast-gfs_3.zip — GraphCast-GFS neighbourhood data for a 3×3 grid-point neighbourhood Observation Data obs.zip — Quality-controlled, six-hourly precipitation accumulations derived from one-minute Automated Surface Observing System (ASOS) data across CONUS, retrieved from the National Weather Service (NWS, 1998). Data was retrieved from https://mesonet.agron.iastate.edu/request/asos/1min.phtml. Six-hour accumulations were flagged as missing when fewer than five hours of valid one-minute data existed within the period. Records were removed when one-minute accumulations exceeded 38 mm or six-hour accumulations exceeded 840 mm (corresponding to World Meteorological Organization world record values; WMO, 1994). ERA5 Climatological Thresholds Annual 99th and 99.9th percentile climatological thresholds for six-hourly precipitation accumulations were computed at each ASOS station location using the ERA5 reanalysis grid point at which the station resides, over the period 1990–2020. These thresholds define extreme and very extreme precipitation events and are used to set the weighting functions in the twCRPS calculations. ERA5 is documented in Hersbach et al. (2020). More complete sets of ERA5 data are available at https://console.cloud.google.com/storage/browser/weatherbench2/data/era5 and from the Copernicus Climate Data Store at https://cds.climate.copernicus.eu/datasets/reanalysis-era5-single-levels. clim_0.99.nc — ERA5-derived annual 99th percentile (0.99 quantile) thresholds at each station location (NetCDF format) clim_0.999.nc — ERA5-derived annual 99.9th percentile (0.999 quantile) thresholds at each station location (NetCDF format)