The F1 race
intelligence layer.

The tire model runs two layers. A Kalman filter provides online state estimation — updating the tire health estimate and, critically, the estimated degradation rate on every incoming lap time. Early degradation is detected before the lap time cliff fully materializes. On top of the Kalman state, a Gaussian Process Regression model captures the full degradation curve per compound, including non-linear acceleration near the thermal cliff. The GP provides uncertainty bounds alongside the point estimate.

Tire strategy determines ~40% of race outcomes. A model that catches the cliff 3–5 laps early — before any on-screen indication — is the most valuable pre-pit signal available. For fantasy: transfer decisions. For betting: pit window odds with real degradation data behind them.

Win probability is modelled as a CTMC over the joint position state of all 20 drivers. Transition rates between position states are parameterised functions of current gap, tire health differential, remaining laps, Overtake Mode eligibility (gap < 1.0s), safety car probability, and lap delta trajectory. The CTMC computes the probability distribution over final positions analytically via matrix exponentiation. This is why it can run on every timing packet — not just at lap boundaries. It does not simulate the race forward; it solves the current state exactly.

A 15–25 percentage point swing in win probability during a safety car period is routine. For a sportsbook carrying $200K–$500K in open race positions during a race, that is material risk exposure moving faster than most odds refresh pipelines. Real-time win probability delivery is risk management infrastructure.

ECP is a deterministic function of the CTMC position distribution. After the CTMC solver produces P(final position = k) for each driver, ECP is the dot product of that distribution with the F1 points vector [25, 18, 15, 12, 10, 8, 6, 4, 2, 1, 0…]. The result is championship-granular: a leader at 65% win probability maps to roughly 18–21 ECP depending on circuit overtake index. A P12 driver outside the points has ECP ≈ 0.3. A safety car neutralises the field and can move ECP by 10+ points across multiple drivers simultaneously. ECPA — the signed delta from the prior lap — is the fastest single-number summary of which driver gained or lost championship value this lap. SC deployments produce the largest single-lap ECPA swings in the dataset.

Championship title races have 3–4 decisive weekends per season. Live ECP converts real-time race data into the metric championships are actually decided on. For sports media covering a title fight, ECP arc visualisations are the highest-signal graphic available. For DFS/fantasy scoring: ECP directly models what points finishes cost and gain in the context of the current race.

A Bidirectional LSTM processes the last 8 laps of contextual features per driver. The bidirectional architecture is deliberate: the backward pass gives the model knowledge of the expected end-of-stint degradation pattern, so it can assess how far through the tire lifecycle a driver is sitting. The output is a probability per driver per lap, not a binary window signal. This means downstream consumers can track the gradient: 0.31 → 0.44 → 0.62 → 0.78 across successive laps is a very different signal than stable at 0.45.

The pit stop market is F1's most actively traded in-play event. A probability signal with 3–5 laps of advance notice — before the team has made the call, before any on-screen hint — represents a genuinely tradeable edge. No other commercial F1 data API provides it.

LightGBM was chosen specifically for this model: it handles the mix of continuous timing features and categorical circuit/weather features efficiently, with inference under 5ms — fast enough to update on every lap. Key predictive signals: recent incident flag frequency, field compression index (how tightly the pack has bunched), per-circuit historical SC probability, session phase (lap 1–5 is statistically higher risk), and weather severity. The 2026 elevated SC rate from driver adaptation to turbo lag was applied as a +0.10 base rate adjustment proactively before the season started.

Safety cars cause the largest single-lap swings in race outcome probability — and trigger the largest volume spikes across every in-play market simultaneously. Knowing before the flag drops is the most operationally valuable advance signal in live F1 data.

XGBoost trained on driver-pair laps across the 2018–2025 seasons, predicting whether driver A overtakes driver B within the next 3 laps. The critical design decision: a per-circuit overtake index trained on historical position-change rates per circuit-sector pair. 2026 transition: the DRS feature becomes Overtake Mode (same 1.0s threshold, different physics). Overtake Mode provides electric boost anywhere on track — not just designated straights. The circuit index was recalibrated for the broader activation envelope.

Not all close gaps produce overtakes. A per-pair model that distinguishes real threat from traffic management gives broadcast graphics, fantasy scoring engines, and in-play market signals something genuinely predictive — not a replay of what the timing screen already shows.

LTOE — Lap Time on Expected — is the residual between actual lap time and what the model predicts, given all context. Trained on lap-level data across the 2018–2025 seasons. Core features: tire age × compound interaction (compound-specific degradation per circuit), fuel-corrected lap time (fuel burns ~0.03s/lap), track evolution coefficient, traffic indicator (clean air vs. dirty air), and a TeamYear baseline that absorbs car performance relative to the field for that specific constructor × season. LTOE = actual − predicted. Negative = faster than expected. That is real pace, not fresh tire advantage.

Tire age alone accounts for 0.8–1.2 seconds of lap time degradation in the final stint. Without correcting for this, 'fast lap times' just means 'driver on fresh rubber in clear air.' LTOE removes all of it. It is the only commercially available normalized pace metric for F1.