A New Paradigm for Medium-Range Severe Weather Forecasts: Probabilistic Random Forest–Based Predictions

Historical observations of severe weather and simulated environments (i.e., features) from the Global Ensemble Forecast System v12 (GEFSv12) Reforecast Dataset (GEFS/R) are used in conjunction to train test random forest (RF) machine learning (ML) models probabilistically forecast out days 4--8. RFs trained with 9 years GEFS/R reports establish statistical relationships. Feature engineering is briefly explored examine alternative methods for gathering features around observed events, including simplifying using spatial averaging increasing ensemble size time-lagging. Validated RF tested ~1.5 real-time output operational GEFSv12 evaluated alongside expert human-generated outlooks Storm Prediction Center (SPC). Both RF-based forecasts SPC skillful respect climatology at 4 5 degrading skill thereafter. The exhibit tendencies underforecast but they tend be well-calibrated lower probability thresholds. Spatially predictors during training allows prior-day thermodynamic kinematic generate forecasts, while time-lagging acts expand areas, resolution decreasing objective skill. results highlight utility ML-generated products aid operations into medium range.