Methodology

1 · Data Source

Telemetry comes from FastF1 v3.8.2, which mirrors livetiming.formula1.com with a fallback to the Jolpica Ergast API. Each race we pull race-session laps, take each driver's top-5 fastest clean laps (excluding pit-in/pit-out and laps slower than 1.08 × fastest), and resample telemetry onto a common track-distance grid at 5 m resolution.

Telemetry channels used: Speed, Throttle, Distance, X, Y. Sampling rate is 4 Hz for car data, which gives roughly 15 m of localisation uncertainty at 250 km/h — fine for zone-level analysis, too coarse for sub-zone timing.

2 · 2026 Power Unit Basis

The detection is anchored in the 2026 PU regulations. With the MGU-K capped at 350 kW and default deployment tapering off above 290 km/h, a car at full throttle would normally stop accelerating shortly after 290 km/h and coast through to its drag-limited terminal speed. With Manual Override engaged, that 350 kW deployment is extended to higher speeds — so the car keeps pulling.

The red shaded region is the detection band: in this speed range the field's "default" cars are coasting while Override-engaged cars are still pulling. We don't see the button press itself; we infer it from the derivative of the speed trace.

3 · Boost Detection

For each lap, on each 5 m bin i, we flag the bin as boost when:

Where T_i is throttle percentage and v_i is speed in km/h. Smoothing uses a 40 m uniform filter to suppress sample noise. Detected bins are then run through a minimum-length filter — runs shorter than 100 m are discarded as sensor jitter.

Why spatial acceleration (dv/dx) instead of time acceleration (dv/dt)?

Sampling rate varies across telemetry; dv/dt is noisier than taking the spatial derivative against a uniform 5 m grid. Spatial dv/dx is also intuitively the right thing — we're asking "is this car still gaining speed per metre travelled" — which is what determines whether it has more deployment headroom than the regulation default.

Thresholds in tabular form

4 · Aggregating to Optimal Zones

For a given race, we stack every driver's boost mask on the common distance grid and combine them into a single field-wide boost-probability curve. Rather than averaging every driver equally, we weight each driver by their championship rank — the theory being that the points leaders are the most reliable signal of where boost is genuinely advantageous. A backmarker who never deploys through a key zone tells us little about whether the zone is correct.

Weights: F1 points scale by championship rank

Each driver's weight w_d is determined by their championship rank entering the race, mapped through the standard F1 race-points scale:

The top driver in the standings carries roughly 25 % of the total weight on their own; the top three together account for about 57 %; everyone outside the top ten contributes nothing to the optimal-zone definition.

p_i still lies in [0, 1] and represents the championship-weighted fraction of the field deploying boost at that track distance on their fast laps. We then smooth with a short uniform filter and find peaks using SciPy's find_peaks. Each peak region is expanded outward while p > 0.35, yielding a contiguous optimal boost zone.

A final lap-wrap merge step combines zones that span the start/finish line (e.g. a zone ending at 5 685 m and a zone starting at 295 m on a 5 807 m circuit are merged into one zone that wraps the line).

Zone metadata stored

5 · Driver Scoring

Each driver's per-zone score combines two pieces — how much of the zone they cover, and how well-centred their deployment is on the zone's peak.

A. Coverage

Fraction of the zone where the driver's per-bin boost probability exceeds the coverage threshold (0.40):

B. Centroid alignment

How close the driver's boost-weighted centroid sits to the zone's peak distance, measured circularly so wrapping zones still work:

The offset is converted to a 0–1 alignment score with a Gaussian decay around a "good radius" of 200 m:

C. Combined zone score

Coverage dominates because — in the Cadillac problem we keep finding — missing a zone matters more than being slightly off-centre within one. A driver who deploys at the right place but only occasionally is more correctable than a driver who never deploys there at all.

D. Overall score

The season leaderboard score is a simple unweighted mean across the races a driver has run.

6 · What the Score Does Not Capture

• Strategic saving — drivers with limited battery state or protecting tyres will deliberately under-deploy, losing points even though the call was rational.
• Traffic and DRS trains — close-following cars sometimes can't afford to spike speed further; we mitigate by using fastest clean laps only, but not perfectly.
• Button-timing resolution — at 4 Hz we can only localise a boost press to ~15 m at 250 km/h. Small sub-zone differences between drivers are under the noise floor.
• Team strategy calls — a driver may be told to use boost defensively rather than offensively; the score rewards only field-optimum deployment.
• Self-reinforcing optimum — because the optimal-zone definition is championship-weighted, the standard a driver is measured against is largely set by the leaders of the championship. A backmarker who finds a clever, atypical place to deploy boost won't be credited with redefining the optimum unless their own championship rise eventually shifts their weight upward.

In short: a high score means consistent, location-correct Manual Override usage on qualifying-pace laps. It is not a full proxy for race-craft quality.

7 · Reproducibility

All of the above is implemented in two scripts that live alongside the viewer in this folder:

The scripts run idempotently — re-running them after a new race regenerates the JSON, the viewer, and the inline snapshots used by this page's charts. Source code is in the Overtakers Cowork project.