Automated fault attribution for Formula 1 collisions: evidence for the limits of telemetry-based classification

Fault attribution in Formula 1 collisions remains a manual process in which FIA stewards integrate video, telemetry, team radio, and regulatory interpretation. We present the first machine-learning study of whether telemetry alone can predict steward fault decisions. Using 164 steward-adjudicated collisions from the 2019–2025 seasons, we first extracted 225 statistical telemetry features. These failed to generalize across seasons, performing below a majority-class baseline under leave-one-season-out (LOSO) evaluation. We then developed 26 FIA-aligned interaction features representing spatial overlap, braking dynamics, temporal proximity, avoidance behavior, and car control. These reduced the within-distribution to cross-season generalization gap from 30.0 to 22.9 percentage points and achieved 72.9% balanced accuracy under stratified 10-fold cross-validation but only 52.3% under LOSO (AUROC = 0.577), only marginally above chance. Permutation importance and SHAP analyses showed that the limited predictive signal was concentrated in a small subset of braking-dynamics and spatial-overlap features, while additional telemetry features did not improve cross-season performance. These findings indicate that telemetry alone is insufficient for reliable fault attribution because critical evidence—including video, driver intent, race context, and evolving steward interpretation—is unavailable from telemetry. Rather than presenting a deployable classifier, this study establishes the first quantitative baseline for automated Formula 1 fault attribution and motivates future multimodal decision-support systems.