Supplemental Figures

for "Predicting property damage from tornadoes with zero-inflated neural networks"

1. Conditional Damage Residual Map

Interactive geographic display of model performance on the test set data for the best model predicting property damage conditional on non-zero damage. Darker blues indicate larger overestimates, while darker reds indicate larger underestimates; lighter colors indicate predictions that closely match observed property damages. On mouse hover, points display the observed property damage value, predicted property damage value, and the residual. Highly populated cities are displayed for geographic context.

2. Damage Occurrence Residual Map

Interactive geographic display of model performance on the test set data for the best model predicting whether property damage will occur. Darker blues indicate high predicted probabilities of damage for non-damaging tornadoes, while darker reds indicate low predicted probabilities of damage for damaging tornadoes; lighter colors indicate high predicted probabilities for damaging tornadoes or low predicted probabilities for non-damaging tornadoes. On mouse hover, points display the damage indicator (1 = damaging, 0 = non-damaging), predicted probability of damage, and the residual. Highly populated cities are displayed for geographic context.

3. 2019 predicted conditional damage map

Interactive geographic display of model predictions on a regularly spaced grid for the best model predicting property damage conditional on non-zero damage. Darker purples indicate larger property damage predictions, while lighter yellows indicate smaller property damage predictions. On mouse hover, points display their coordinates along with predicted property damage value. Highly populated cities are displayed for geographic context and are accompanied by predicted property damage values. Interactive buttons along the left side allow the user to choose their 2019 month of interest.

4. 2019 predicted damage occurrence map

Interactive geographic display of model predictions on a regularly spaced grid for the best model predicting whether property damage will occur. Darker purples indicate higher predicted probabilities of damage, while lighter yellows indicate smaller predicted probabilities of damage. On mouse hover, points display their coordinates along with predicted probabilities of damage (bound between 0 and 1). Highly populated cities are displayed for geographic context and are accompanied by predicted probabilities of damage. Interactive buttons along the left side allow the user to choose their 2019 month of interest.