Jan 2025 – May 2025 · Data Science
U.S. Traffic Fatality Prediction with BigQuery ML
Modeling multi-fatality crash risk across 100K+ NHTSA records — where the goal was an interpretable coefficient, not a leaderboard score.
Results
- 100K+ NHTSA FARS crash records modeled entirely in-warehouse, spanning all 50 states
- Identified alcohol involvement and time of day as the strongest predictors of fatal crashes, with risk concentrated on weekends and late nights
- Quantified the safety-restraint gap directly: 15,474 fatalities occurred when no restraint was used at all, versus 41,560 with a lap-and-shoulder belt confirmed in use
- Interpretable coefficients as the deliverable, not just predictive scores — built for a transportation-policy audience, not a leaderboard
- Interactive Looker Studio dashboards for non-technical stakeholders to filter by state, vehicle type, and behavioral factor
Stack
- SQL
- BigQuery ML
- Looker Studio
- GCP
Overview page — the headline numbers and the state-level treemap.
Victim demographic profile — surfaced to help target safety campaigns at the groups actually at risk.
Fatalities per crash event by state — frequency and severity are different questions, and this page answers the second one.
Injury severity distribution across all recorded crashes, not just the fatal ones.
Top 10 states by fatality record count, queried from the NHTSA FARS public dataset (bigquery-public-data.nhtsa_traffic_fatalities).
California, Texas, and Florida lead in total fatality records — consistent with population and vehicle density, and a starting point for where policy intervention would have the most reach.