The signal
before the
chaos.
PeloRisk ingests live telemetry — peloton density, speed differentials, gradient, weather, and historical incident data — and surfaces a crash risk score to race directors in real time. Not a verdict. A signal.
| # | Date | Stage | Type | km | Risk Score | Key Flags |
|---|
The Model's Job
PeloRisk surfaces probability, zone, and reason — not verdicts. A score of 87 on Stage 6's Tourmalet descent means the confluence of 180 riders, a 9% gradient, and forecast crosswinds puts this window in the top decile of historical crash risk for this road segment.
The model never triggers an action. It makes the human smarter.
The Human's Job
Race directors, medical cars, and commissaires receive the signal. They decide whether to accelerate neutralization, reposition vehicles, or broadcast warnings. False positives cost race time. False negatives cost lives. That asymmetry is why the model flags and the human acts.
No False Confidence
If telemetry is sparse — a GPS dropout, a late-joining team — PeloRisk surfaces an "insufficient data" state rather than extrapolate. A wrong score is worse than no score in a 60km/h peloton.
Continuous Calibration
Every incident is logged with its pre-crash telemetry snapshot. The model retrains on each Tour edition, tightening precision on new climb profiles, updated road surfaces, and evolving team tactics. The 2026 model includes 6 new climbs added to the route this edition.
PeloRisk is a decision-support tool, not a safety authority. Risk scores are probabilistic outputs derived from telemetry, terrain, and historical data — they are designed to inform race officials, not replace their judgment. No score, however high, automatically triggers race neutralization or any operational action. All decisions remain with credentialed race directors and commissaires under UCI rules.
This tool is not affiliated with ASO, the UCI, or any official Tour de France organization. Stage ratings shown here are model outputs for demonstration purposes.
This model can and does make mistakes. Crash risk is inherently unpredictable — a single mechanical failure, an unexpected weather shift, or an unpredictable rider action can cause an incident that no model would have flagged. PeloRisk can also produce false positives (high scores on stages that finish incident-free) and false negatives (low scores on stages where crashes occur).
Scores for stages on new climbs carry additional uncertainty — the model has no historical incident data for those road segments and relies entirely on terrain inference. These outputs should be treated with extra caution.