Identifying Freeway Traffic Instability from PeMS Data Using Rolling-Window Speed and Occupancy Variability

This study examines short-term freeway traffic instability using station-level 5-minute data from the California Performance Measurement System for Caltrans District 4. The dataset covers April 7, 2026. After restricting the sample to retained mainline observations, rolling 15-minute measures of mean speed, speed variability, occupancy variability and net speed change were constructed for 652,080 station-time observations. An instability-congestion proxy state was defined as the joint occurrence of speed variability at or above its 75th percentile and mean speed at or below its 25th percentile. An exploratory logistic classification model was then used to describe how the four traffic measures separate observations assigned to this proxy state. Occupancy variability and positive net speed change were associated with higher classification odds. However, mean speed and speed variability partly define the outcome and their coefficients are therefore definitional rather than independent evidence of safety effects. In addition, the model does not account for dependence among repeated observations from the same detector station. The results demonstrate a transparent approach for screening unstable congestion from high-resolution detector data but do not establish crash probability or causal safety effects. Validation against observed crashes, longer observation periods and models accounting for station-level and temporal dependence are required before operational deployment.