Progress toward forecasting excessive rainfall with random forests based on a deterministic convection-allowing model

This dataset consists of forecasts produced by random forests (RFs) using
predictor information from NOAA's deterministic convection-allowing
numerical weather prediction model, the High-Resolution Rapid Refresh.
Included are sensitivity experiments on predictor assembly and model
version, as well as real-time forecasts from three subsequent versions
evaluated at the Weather Prediction Center's Flash Flood and Intense
Rainfall Experiment (FFaIR) during 2021-2023. Sensitivity experiments
reveal that the RF performs better when we use predictor information from
all model gridpoints, not just sparse gridpoints, particularly in
situations with small-scale precipitation maxima in the model forecast.
The RF is also better able to learn the relationships between predictor
values and resulting excessive rainfall risk when the RF considers mean
predictors from three model simulations rather than predictors from a
single simulation. The real-time RFs evaluated at FFaIR exhibited
year-over-year improvements stemming from the results of these sensitivity
experiments as well as feedback from FFaIR participants. However, RFs
based on deterministic convection allow models to continue to underperform
those based on coarse global ensemble systems.