A Misclassification Framework for Identifying Systematic Error in Pedestrian Crash Prediction: Evidence from Seattle,...

Pedestrian crash prediction models often fail systematically, yet distinguishing correctable failures from random errors is challenging. We present a framework combining spatial clustering, temporal persistence, and model ensemble consensus to identify improvement opportunities. Applied to 13,706 Seattle intersections, the framework identified 141 high-priority locations (1.03%) exhibiting elevated temporal persistence, strong spatial clustering, and high ensemble consensus. A multiplicative priority score serves as a proxy for systematic error, concentrating 6.6% of total error into 1.03% of locations. The framework flagged crash-enriched intersections (roughly twice the baseline crash rate) where all models over-predicted risk, indicating the feature set captures that a location is hazardous but lacks the site-level descriptors needed to calibrate how much. Audits revealed three categories of omitted features: visual obstructions, static road use, and network complexity. Validation confirmed that adding these features reduced priority scores by 6.1% at high-priority locations, confirming that the framework correctly identifies where omitted context matters most.