Simulating the unseen: Crash prediction and prevention should learn from what did not happen

Traffic safety science faces a fundamental paradox: the crashes we most wish to prevent are precisely those we rarely observe. Existing crash-frequency models depend on sparse, noisy, and under-reported records, while even high-fidelity simulations undersample the long-tailed conditions that lead to catastrophic outcomes. We argue that progress toward Vision Zero requires a shift from crash-only learning to counterfactual safety learning, where models reason not only about what happened, but also about high-risk situations that almost occurred. Four fundamental challenges in modern safety analysis are first identified: the rarity of crashes, heterogeneity in human behavior, complex group-wise interactions, and the scarcity of high-resolution crash data. The statistical challenges posed by rare crash events and unobserved human and environmental factors are then examined, highlighting how normal driving and near-miss trajectories can provide valuable evidence for estimating crash risk. We further review how current macro-level models, surrogate safety measures, and simulation pipelines fall short in addressing these challenges, including a persistent macro-micro disconnect. Finally, we outline an AI-centered counterfactual agenda that integrates generative scenario engines, physics-grounded and behaviorally rich digital twins, hypergraph-based interaction modeling, and reasoning-enhanced intervention design, reframing crash prediction as learning from events that almost, but did not, happen.