ReferenceModels Reference
How the intelligence is built. Fourteen production ML models, plus an algorithmic layer (CTMC win/podium probability with a full position distribution, ECP/ECPA expected championship points, and EKF tyre health), produce the analytics.* and weather.* feeds. This page documents how each model works — its type, inputs, and calibration. For the payload each one emits, see the Analytics feeds →
ℹModels are delivered as feed outputs, not as a separate product. You subscribe to a feed (e.g. analytics.strategy); the models behind it run server-side in real time. The model count is the documented source of truth; the live GCS manifest is the runtime truth.
Production ML models
| Model | Type | Predicts | Feed |
|---|
| pit stop probability | gradient-boosted-ensemble | P(this driver pits this lap), every lap. | analytics.strategy |
| safety car | gradient-boosted-ensemble | P(SC or VSC deployment next lap), session-level. | analytics.strategy |
| overtake | gradient-boosted-ensemble | P(position gain via on-track overtake within 3 laps), per driver. | analytics.strategy |
| LTOE lap-time outlier estimation | gradient-boosted-ensemble | Lap Time Over Expected — the normalized pace delta, per lap. | analytics.strategy |
| tyre cliff | gradient-boosted-ensemble | P(tyre performance cliff), per driver per lap. | analytics.strategy |
| undercut threat | gradient-boosted-ensemble | P(an undercut attempt against the car ahead succeeds), when in range. | analytics.strategy |
| undercut timing call | calibrated-classifier | WHEN to strike: the recommended undercut lap and its success probability. | analytics.strategy |
| overcut threat | gradient-boosted-ensemble | P(staying out longer beats the car ahead's stop), when in range. | analytics.strategy |
| qualifying pace | gradient-boosted-ensemble | Per-sector pace deltas vs the session median, per driver. | analytics.qualifying |
| fastest lap | gradient-boosted-ensemble | P(this driver sets the session's fastest lap). | analytics.race-odds |
| DNF risk | survival-model | Retirement hazard per driver, per lap. | analytics.strategy |
| tyre compound | gradient-boosted-ensemble | The next compound a pitting driver takes, as probabilities. | analytics.tire-strategy |
| pit-stop count | gradient-boosted-ensemble | P(one / two / three+ stops) for the whole race, per driver. | analytics.strategy |
| pit-window survival | survival-model | Laps until each driver's pit window opens. | analytics.pit-window |
Every model has its own page — click a name. All share the training and calibration pipeline below.
Algorithmic intelligence layer
Three non-ML components ship alongside the models and produce the bulk of the race-odds and tyre signals.
Win probability — Monte Carlo
A forward Monte Carlo race-to-finish simulation produces a full 20-element positionDistribution per driver, from which podium/points odds, expected points, expected laps led, and forward position-change projections are derived analytically. Powers analytics.race-odds and analytics.team-points (via joint-distribution convolution).
Tyre health — EKF
An extended Kalman filter tracks each driver's lap-delta series to estimate pace.tyre.health (0–1), degradation rate (s/lap), and stint lap — robust to safety-car and traffic noise that fools a naive regression.
Expected championship points — ECP/ECPA
Season-level points expectation (and an adjusted variant) projecting championship outcomes from the current standings and per-race position distributions. Surfaces on analytics.championship-probability.
Calibration
Every classifier ships with a calibration stage fitted on held-out sessions, so a published 0.30 means 30%. Models retrain automatically as race weekends land in the pipeline. Live calibration status per model — meeting target, recalibrating, or not yet tracked — is available at the model manifest endpoint, deliberately as a qualitative status rather than a raw metric — self-reported accuracy numbers are easy to publish and hard to verify. The free tier replays this season's sessions, so you can score the projections against what actually happened yourself. That's the stronger signal.
2026 regulation adjustments
The 2026 F1 regulations — tripled ERS power, turbo lag, active aero replacing DRS — required every model to be re-assessed. What changed:
| Model | Feed | Status | Action taken |
|---|
| Tyre degradation | analytics.strategy | Recalibrated | Circuit parameters updated for new thermal loading from turbo lag. Wider confidence intervals until 4+ 2026 races. |
| Win probability | analytics.race-odds | Updated | Monte Carlo pass-through rates recalibrated for the 2026 electric boost profile. |
| Pit stop | analytics.strategy | Adjusted | ERS clipping creates saw-tooth lap-time variance that mimics cliff signatures. Decision threshold raised for the new variance profile. |
| Safety car | analytics.strategy | Offset applied | Base SC rate offset upward for the first 6 races of 2026 (driver adaptation period). |
| Overtake | analytics.strategy | Recalibrated | Circuit overtake indices updated for zone-independent Overtake Mode vs. DRS-zone activation. |
| LTOE | analytics.strategy | Active | Shipped with an early-season damper that lifts automatically once 3+ races of 2026 data exist — lifted for the 2026 season. |