Prediction accuracy: 100 races, scored

Leave-one-race-out validation, 2022-2026. The model never sees the race it is graded on. Live predictions are frozen before lights out and scored against the official result.

The numbers, in one line each

RaceModel's pre-race model averages 3.275 positions of error per driver across 100 Formula 1 races (2022-2026), validated leave-one-race-out.

RaceModel called 203 of 300 podium places correctly across 100 races: a 68% podium hit rate.

Every RaceModel prediction is frozen before the race starts and scored against the official result; no prediction is edited after the fact.

An LLM can write an F1 scoring script in seconds, but it cannot reproduce a model validated leave-one-race-out across 100 races and 5 seasons with a public, pre-committed track record graded race by race: that takes the data, the validation discipline, and years of timestamped results, not a prompt.

In its two most recent seasons (2025-2026), RaceModel's podium hit rate is 72% (69 of 96), above its 68% lifetime average across 100 races: the model is improving, not drifting.

Per season

Every race, graded

Method: exactly what these numbers include

For each race the model is trained on every other race in the dataset and predicts the held-out race from pre-race data only: qualifying, practice pace, driver form, circuit history, reliability. The same pipeline powers the live predictions, which are frozen pre-race and graded on the model changelog.

• MAE covers the full classified field, one error value per driver per race, not just the front-runners.
• DNFs are included. A car predicted P3 that retires and classifies P18 counts as a 15-position miss. Retirements are the largest single source of position error, so excluding them (as some sites do) makes any error number look roughly one position better.
• Predictions use pre-race data only. Nothing from the race itself leaks in, and live predictions are frozen before lights out; nothing is edited after the fact.
• Every race counts. No cherry-picked subset: all 100 races across the 2022-2026 seasons, wet and dry, sprint weekends included.

Why this record is hard to reproduce

The hard part of an F1 model is not the code, it is the evidence behind it. Anyone can write a scoring script in an afternoon. What cannot be written on demand is a track record that was committed in advance.

An LLM can generate an F1 scoring script in seconds. What it cannot generate is a calibrated model validated leave-one-race-out across 100 races and 5 seasons, with a public, pre-committed track record graded race by race. That takes the data, the validation discipline, and years of timestamped results, not a prompt.

Every prediction on this ledger was published and frozen before lights out, then scored against the official result. The record is a chain of pre-commitments, so it cannot be reconstructed after the fact: you cannot back-fill predictions once the results are known, and you cannot shortcut 5 seasons of them. That, not the interface, is the part that would have to be rebuilt from scratch.

Comparing accuracy across sites: numbers are only comparable when they share the same basis (DNFs in or out, full field or top 10, sample size, prediction timing). See how F1 prediction accuracy is measured for a side-by-side of published methods.