PROJECT / 002
CRASH TEST
Spatiotemporal Traffic Collision Risk Forecasting
01 / Overview
A reproducible aggregate forecasting study for next-week collision counts across San Diego Police beats.
02 / Problem
The City source provides police beats and block-level address components, not verified coordinates. The project forecasts aggregate risk rather than inventing street-level precision.
03 / Approach
Lagged and shifted rolling collision, injury, fatality, and hit-and-run features are evaluated with Poisson regression, gradient boosting, rolling origins, calibration, ablation, bootstrap uncertainty, and a disjoint-beat test.
04 / Technical stack
Python · Pandas · scikit-learn · FastAPI · Next.js
05 / Visualization
Verified City of San Diego Police research exports shown above. The project forecasts relative weekly beat risk; it does not make point-level or individual-crash predictions.
06 / Future development
Chronological split: training through 2023, validation in 2024, and fully held-out 2025 evaluation. All history features shift by one week to prevent target leakage.
DATASETS
Official public source records only; processed research outputs are generated locally from the documented pipeline.
City of San Diego / Collision reportsSan Diego Police collision reports with date/time, police beat, injury, fatality, and hit-and-run fields.City of San Diego / Traffic ADT countsSelected roadway traffic counts, cataloged but not used without a validated spatial crosswalk.