Roadway Safety Explorer

The Off-Peak Danger

Traffic crashes are a leading cause of injury and death in Philadelphia, and arterial roads are disproportionately dangerous. Crash volumes and KSI (killed or seriously injured) rates are highest during nighttime and early morning hours, even though traffic volumes are low during those periods.

Recent tragedies in West Philadelphia illustrate the risk. On November 20, 2025, Meaza Brown, 48, was struck and killed while crossing Market Street at 33rd Street. One week later, on November 27, Rosa Mar Espinosa Rodas, 41, was killed by a speeding driver while walking on 36th Street. Both incidents occurred during off-peak hours when roads were relatively empty — and that emptiness may itself be part of the problem.

Analysis of speed sensor data collected across Philadelphia confirms the pattern: speeding is significantly more prevalent during off-peak hours, and more common on major and minor arterials than on collectors or local streets.

Research Hypothesis: A Deadly Cascade

This project tests the hypothesis that roadways designed to move peak-hour vehicle demand can create excess capacity during the rest of the day, increasing speeding opportunity and raising crash severity risk during off-peak hours. The proposed mechanism is a cascade:

1. Peak-oriented design — Many urban arterials are sized to reduce congestion during high-demand periods. Wider lanes, multiple travel lanes, and higher design speeds can also increase the speeds drivers choose when the street is not congested.
2. Off-peak underuse — For most hours of the day, demand is lower than peak-hour design capacity. The roadway still offers the same amount of pavement, but fewer vehicles are using it.
3. Excess capacity becomes speed opportunity — Lower congestion reduces friction, while wide, forgiving geometry can signal that higher speeds are acceptable. This combination can increase free-flow speeds and speed variability.
4. Speed turns risk into severity — Once a crash occurs, injury severity rises steeply with impact speed, especially for pedestrians and cyclists. Small increases in operating speed can therefore produce much larger increases in fatality risk.

In short, the same design choices that help a corridor "work" during peak periods may leave too much unused capacity at night and other off-peak times. This project therefore models speeding frequency as an upstream safety signal: crashes are relatively rare and stochastic, but persistent high-speed driving reveals where roadway design may be creating conditions for severe outcomes.

Data & Methodology

The project integrates multiple datasets covering Philadelphia's road network, traffic behavior, and built environment characteristics. Speed observations come from DVRPC speed measurement points deployed on approximately 446 road segments across the city, recorded from 2022 to 2025. Each record captures the number of vehicles exceeding the posted speed limit during a given hour, as well as total traffic volume for that hour.

For this tool, hourly measurements are aggregated into four time periods: night off-peak (20:00-06:00), morning peak (07:00-10:00), midday off-peak (11:00-15:00), and evening peak (16:00-19:00). We train a Random Forest regression model to predict the percentage of traffic on a road segment, during a given time of day, traveling above the speed limit on that segment. Predictors fall into three categories:

• Spatiotemporal & traffic dynamics — time period, observed traffic volume, speed measurement month, weekday vs. weekend, and speed limit
• Roadway geometry — roadway width dedicated to traffic lanes, curb-to-curb width, width per traffic lane, segment length, traffic direction, divided roadway, and shoulder width of 8 or more feet
• Road classification & built environment — City of Philadelphia road typology, bike lane type, on-street parking, sidewalk coverage, traffic calming devices, transit stop density, intersection control density, parcel density

We use the tidymodels framework in R to create the training/testing split, conduct cross-validation and hyperparameter search, preprocess the data, train the model, and generate predictions and explanatory plots. The final held-out test set reserves 25% of the data, stratified over the speeding outcome, purely for model evaluation.

Model Performance

The model explains roughly 92% of variance in speeding frequency across segments and time periods, with a mean absolute error of approximately 4 percentage points on the final held-out test data. The Random Forest modeling approach was chosen for its ability to capture non-linear relationships and feature interactions (e.g., lane width × road type × time period) without requiring explicit interaction terms.

Intervention Scenarios

A core limitation of road safety research is that individual geometric features cannot be considered in isolation — a lane diet always implies changes to lane width, and adding a separated bike lane reallocates curb-to-curb space. This tool addresses that by presenting pre-computed intervention scenarios: holistic what-if road configurations run through the full trained model.

Each scenario represents a realistic re-allocation of road space, with all downstream geometry recalculated consistently. Available scenario types include:

Scenario predictions are pre-computed by running each road configuration through the trained Random Forest model offline. Results reflect model predictions under each scenario's geometry — not real-world guarantees. Outcomes depend on implementation quality, enforcement, driver response, and local conditions not captured in the model.