Crash Risk Radar

Heatmap Predictions of Likely Road Accidents for 5 Major US Cities

How to Use This Tool

This tool will allow you to use the "Select City" dropdown to choose one of the five available US cities. The map will automatically center on your selection.
Use the "Select Data Range" dropdown to choose the historical time period for the model (e.g., 1, 3, 5, or 8 years of data).
The map will update with a predictive heatmap showing where crashes are most likely to occur. The "Model Performance" box will show the statistics for the selected model.

Controls

Select City

Select Data Range (Years)

Model Performance

Loading stats...

Predicted Crash Hotspots

About the Models

This tool uses machine learning to predict the geographic coordinates (latitude and longitude) of potential car crashes. The models are trained on historical data (2016-2023) from the NHTSA Fatality Analysis Reporting System (FARS).

For each city and data range, two separate XGBoost Regressor models are trained: one to predict latitude and one to predict longitude. The features used for prediction include time of day, day of the week, weather, light conditions, and road type, as well as some engineered features that are a combination of these. The heatmaps represent the density of these predicted locations from a test dataset.

MAE (Mean Absolute Error) measures the average physical difference between the model's predicted latitude/longitude and the actual location. Since 1 degree of latitude is approximately 69 miles, an MAE of 0.03 corresponds to an error of around 2 miles. Depending on the city, predictions of road incident locations may be better or worse when using a larger number of years of data