Skip to content
Reference Lap

About Reference Lap

A non-commercial forecasting project for Grand Prix racing: probabilities for qualifying, sprints, races and the championship, published with a track record you can check.

How the forecaster works

Each driver's pace is modelled as the team's car pace plus a personal driver offset, measured as a percentage of a lap (lower is faster), separately for one-lap (qualifying) and race pace. After every qualifying, sprint and race session the ratings are updated with a Kalman filter, using only gaps within that session, so track and weather effects cancel out. Car pace drifts from round to round; driver offsets persist across seasons and are learned mostly from teammate comparisons. Every rating carries an uncertainty (the ± sd on the power rankings), and a separate estimate gives each driver's chance of not finishing.

Ratings, uncertainties and DNF chances feed a Monte Carlo simulation of the session, run thousands of times: qualifying from one-lap pace, and the race from race pace plus a grid penalty scaled by how hard the circuit is to pass on. Counting how often each driver finishes in each position gives the win, podium, points and position probabilities on event pages; simulating the rest of the season gives expected points and title odds.

Forecasts are made at two checkpoints: before qualifying (after practice, with the grid itself simulated) and before the race (after qualifying, with the real grid). Nothing that happened after the checkpoint is allowed into a forecast.

Committed before the session

Live forecasts are written to a forecast ledger and committed to git before the session starts. Only forecasts whose made_at is before the start and whose ledger entry was first committed before the start, and not changed after it, count towards the live track record. Late or changed entries are kept and shown, labelled, but never scored.

The live model configuration has a version number (for example live-v2) and only changes between race weekends, never during one. Every forecast on this site shows which version made it and its commit status.

Commit times are recorded on the forecaster's own machine for now, so they are not yet independently verifiable.

Three kinds of forecast

● Live
Made before the session. The prospective record: the real evidence.
◆ Backtest · held-out
2026 rounds 1–20 replayed walk-forward with criteria fixed in advance. Out of sample once.
▲ Backtest · development
Earlier seasons the model was tuned on. In sample: optimistic by construction.

See the track record →

Performance vs model expectation

For completed races and sprints, event pages compare each driver with the model's own expectation. The qualifying gain is the expected finish before qualifying minus the expected finish before the race (how much the grid changed the outlook); the race gain is the expected finish before the race minus the actual finish. Expected finishes are means over the forecast's position probabilities; retirements are included at their place in the classification order, so each gain sums to about zero over the field.

These are measured against the model, so they are labelled “vs model expectation”: a driver who keeps beating the expectation may be driving well, or the model may under-rate them. To check, the season pages test each driver's mean race gain over at least 5 grands prix: a mean more than 2 standard errors from zero happens to about 4.55% of drivers by chance if the model is unbiased. On the 2026 held-out rounds 2 of 22 drivers were flagged (about 1.0 expected), and on the 2023–2025 tuning seasons 3 of 28 (about 1.3); joint tests were not significant at 5%, and flagged drivers did not repeat from one season to the next. So there is no evidence of systematic bias, but the check has limited power: a bias below about one place per race for one driver over a season would not be detected, and the standard errors treat races as independent although teammates share a car.

Retirement causes come from the result status. Most retirements in the source data are recorded only as “Retired”, so “unknown” is the largest category.

Fantasy-style expected points

Event pages can turn a forecast into expected points under an editable fantasy-style scoring table. Scoring similar to popular fantasy games. Not affiliated with any official game. Everything is computed in your browser from the published position probabilities, chances of not finishing and the grid; overtakes, fastest lap, driver of the day, pit-stop times and disqualifications are not modelled.

Strategy analytics

Strategy views (event pages, season strategy pages) are built from lap times, timing and pit-stop data, which exist from 2023 onwards. Stints come from the timing data's stint numbers; the race trace is each driver's gap to the leader where they cross the line at the end of every lap (missing laps are left blank, never set to zero). Safety-car, virtual-safety-car and red-flag periods are runs of laps under that track status.

Lane loss is the time a stop costs measured from laps: in-lap + out-lap − two of the driver's normal laps nearby. Only green-flag stops with sensible times count towards an event's median. The pit-lane time (pit entry to exit) is a different quantity and typically runs 1–1.5 s higher; the simulator's per-circuit prior is a pit-lane time, listed on the circuits page.

Degradation is the slope of fuel-corrected lap time against tyre age within a stint (8 or more clean laps), as a percentage of the stint's typical lap. Single stints are noisy (10–20% have a negative slope), so the site shows medians with a standard error per event and per season. There is no per-circuit figure: a track's degradation does not repeat from year to year.

Undercut attempts are pairs within 3 s where the car behind pits first and the car ahead pits one to three laps later. There are only about 2–3 per race, so success rates are uncertain. The model numbers are the new-tyre gain (lap-time gain from fresh tyres) and the out-lap loss (which includes the pit-lane exit, so it is not a tyre warm-up penalty).

Telemetry comparisons

Telemetry pages compare two drivers' fastest laps of a session (2023 onwards): in qualifying the fastest timed lap that was not deleted, otherwise the fastest clean lap. The car data (speed, throttle, on/off brake, gear) is sampled about 4 times a second and resampled onto 200 equal steps of the lap; each lap is scaled to its own length. The time gap along the lap is rebuilt from speed and scaled so the lap ends on its official time. Missing car data is left as a gap, never set to zero.

Segments are ours, not official corner numbers: the lap is split at the speed minima of the field's median speed trace, so each segment holds one braking zone, and a chicane or a sequence taken without re-accelerating is one segment. A corner taken flat has none, so there are fewer segments than corners.

Precision is limited: braking points are not precise below about 10 m, brake is on/off only, and distances read about 1% short. Track maps come from one lap of position data per circuit; a map belongs to its source event, and segment numbers and colours appear on it only for telemetry of that same session.

Driver ratings and the ranking test

Driver pages keep three measures apart. Vs teammate (qualifying gap in the last segment both completed, who finished ahead) is the only comparison with the car taken out by construction. The driver rating is the model's estimated pace vs the mean driver, with its uncertainty; it is learned mostly from teammate gaps, and linked across teams only by drivers who change teams. Gain vs model expectation is a surprise against a forecast that already includes the rating.

A ranking claims an order across teams, so it was published only after a test written down before any test code existed (docs/acceptance.md, “Driver ranking test”), run once. Each rating had to predict which of two teammates comes out ahead, on 2026 rounds held out from tuning, with a lower log loss than a coin flip and than the pair's own all-time teammate record, each with a 95% bootstrap interval (over rounds) excluding zero. A second check scored the first five races of every new pairing in 2023–2026 from ratings learned with previous teammates; it had to be no worse than a coin flip.

  • Race rating: passed. On 150 held-out 2026 teammate pairs (both classified) the log loss was 0.624, against 0.693 for a coin flip (difference −0.069, interval −0.121 to −0.018) and 0.694 for the pair record (difference −0.069, interval −0.133 to −0.003): a pass, but narrow against the pair record. New pairings: 0.562 over 111 comparisons (coin flip 0.693). This is the “Driver ranking (model)” view on the drivers page.
  • Qualifying rating: failed. On 209 held-out 2026 pairs it scored 0.652, against 0.693 for a coin flip (interval −0.089 to +0.012, includes zero) and 0.658 for the pair record (interval −0.043 to +0.032). It is still shown as a rating with its ± sd, but there is no qualifying ranking.

The new-pairing check had too few cases to be a held-out gate, and most of them are from development seasons. The real test of cross-team order is 2027: new pairings at the start of the season will be scored as they happen, as the forecasts are.

Driver ranking (model) →

Data sources and licence

SourceLicenceLink
Jolpica F1 APICC BY-NC-SA 4.0jolpi.ca
OpenF1CC BY-NC-SA 4.0openf1.org

The source data is licensed CC BY-NC-SA 4.0, so this site is non-commercial: no advertising and no paid features. The forecasts and ratings derived from it are shared under the same licence, with attribution.

Current snapshot 20261117T082657Z, schema 1.7, live configs live-v1 (5df53b0fe48e, from ); live-v2 (28a4508f891c, from ).

Privacy and analytics

To see which pages and features are used, this site collects anonymous page views and clicks, and records a sample of about one visit in four as a session replay, with anything typed into a form masked. There are no cookies, nothing is stored on your device to recognise you, and there are no accounts and no advertising.

The data is processed by PostHog, in its US region. If your browser sends Do Not Track or Global Privacy Control, analytics does not start at all. The site works fully without it, so blocking it changes nothing for you.

This site is unofficial and not associated with Formula 1 companies. F1 and related marks are trademarks of Formula One Licensing B.V.

Probabilities are estimates for analysis and discussion, not betting advice.