Skip to main content
The Tyre Degradation Simulator lets you interrogate the trained degradation models directly no server, no waiting. You dial a compound, fuel load, dirty-air exposure, and environmental conditions, and five XGBoost models re-score the entire stint in real time via ONNX running in WebAssembly. The result is an interactive pace-loss fan showing the model’s p10/p50/p90 uncertainty band, a cliff risk indicator, and a tyre life gauge all updating as you move the sliders.

​
What the simulator shows

The main chart plots predicted next-lap degradation jump (in seconds of pace loss) against lap-in-stint. It renders three overlaid series:
  • p50 line the median model prediction for pace loss on each lap of the stint.
  • p10/p90 band the model’s uncertainty envelope. A wide band means the model is less certain about this part of the degradation curve, typically around the cliff onset region.
  • Cliff onset zone a shaded region marking the laps where the 4-class cliff classifier transitions to cliff_candidate = true. Pace deterioration after this point is expected to accelerate sharply.
A tyre life gauge summarises the stint-life regressor’s estimate of how many laps remain before the compound falls below a usable pace threshold. A stint cannot outlast the race, so on a final stint (one configured to run to the flag) the gauge is capped at the laps left to the flag. The gauge does not show a chance to reach the flag. See Why there is no chance to reach the flag.
When you load a real stint preset, the chart also overlays the observed next-lap degradation jumps from the warehouse so you can see exactly where the model was confident, conservative, or surprised.

​
Inputs you control

Slider ranges are calibrated to the p5–p95 of fct_cliff_prediction_features across 2018–2025, so you cannot dial inputs the model has never seen.

​
Loading a real stint

Use the Preset selector above the controls to load a real stint from the warehouse. This seeds the model features from the actual race data for that stint and overlays the observed degradation jumps on the chart. Once a preset is loaded you can perturb individual sliders to run counterfactuals for example, asking “how would this stint have degraded if the driver had backed off the push residual by 20%?” Selecting Reset clears the preset and returns all sliders to their default values.

​
Practical use

The most direct use of the simulator is understanding why a strategy call was made. If you select a real stint from a race and see that the tyre was already past the cliff onset zone with ten laps to go, the pit call that looks early on a lap chart starts to make sense. Conversely, if the pace fan shows the compound still well within its degradation window, an early stop was either a reactive undercut or a risk call.
Load two stints from the same race on the same compound one from a driver who extended and one who stopped early and compare where each stint sat relative to the cliff onset zone. The model’s cliff classifier often explains tyre-to-tyre strategy differences that look confusing from lap times alone.
The simulator scores five models (three quantile regressors, one cliff classifier, one stint-life regressor) on every slider change. On first load, the ONNX runtime compiles the model session, which takes a few seconds on some browsers. Subsequent re-scores are near-instant.

​
How the models score in the browser

Each lap of the simulated stint is converted into a float32 vector for each model, built from the current slider state (39 features for the cliff classifier, 32 for the other four models), then passed through the ONNX runtime in your browser. The feature order, encoder mappings, and post-processing logic are read from the model manifest at runtime. Browser-side inference matches the originally trained boosters to a maximum absolute error of 1e-5.

​
Why there is no chance to reach the flag

A chance to reach the flag on these tyres was built and tested: the probability, under the stint-life model’s log-normal distribution, that the tyres last to the chequered flag, passed through a Platt recalibration. The recalibration used the last3 rule (the most recent three seasons, chosen on 2024 only) fitted on 2018-2024 out-of-fold predictions, from a refit of the v14 configuration on expanding-window folds. The simulator does not show it; on final stints it shows only the laps-to-flag cap.
The chance failed its 2025 acceptance bar, so it is not shown. Only 5 of 10 calibration bins were within 5 percentage points, with a worst gap of 19.9 points. Its Brier score was 0.0803 against 0.2454 for always predicting the base rate, so it beats the base rate but is not well calibrated in the middle bins.

​
Data sources

The model was trained on data from 2018–2025. The dataset fingerprint for the current ONNX bundle is 3aff4559.