Prediction accuracy
Every prediction is locked before the session starts, never edited, and scored against the official classification. This is the complete public scoreboard.
Accuracy by stage
Every race prediction gets rerun as the weekend adds information — a practice session, a settled grid, a locked qualifying result. This season's average score at each stage, in the order it happens.
Formula 1 is genuinely difficult to forecast. A random pick from the 22-driver grid wins about 4.5% of the time, and even always betting the championship leader fails at chaotic weekends — rain, safety cars, first-lap contact. Read the numbers below against that baseline, not against perfection.
Scores combine every prediction type for a weekend: race winner and podium, full finishing order, qualifying, sprint results where applicable, and driver head-to-head matchups. A race page's prediction review shows exactly which calls hit and which missed.
Signal vs luck
The honest way to judge a prediction engine is against what guessing would achieve. With 22 cars on the grid, a random winner pick lands about 4.5% of the time; a random pole pick the same; naming one driver for the podium, about 14%. Here is how the engine compares on the season to date.
Baselines assume a uniform pick from a 22-driver grid: 1/22 ≈ 4.5% for winner and pole, 3/22 ≈ 14% for naming one podium finisher. Engine numbers update after every scored race.
Race-by-race scoreboard
How scoring works
Every prediction is generated before the session begins and locked the moment it is saved. Nothing is edited after the chequered flag — the scoreboard above is computed from the same immutable records you can open on each race page.
Each session type has its own weighted rubric. The core of the race score rewards calling the winner, the podium in the right order, and the accuracy of the full top-10. Qualifying revolves around pole position and the top-5. Extra calls — fastest lap, safety car, winning constructor, weather — add smaller bonus weights when data is available.
Bonus calls — fastest lap, safety car, winning constructor, margin of victory, weather, pit-stop calls, and team head-to-heads — carry weights of 2–5 points each and only count when the underlying data exists for that session. Weights are relative rather than percentages: each session's score is normalized against the total weight available for it.
What gets scored
Race
The headline prediction: winner, full podium in order, and the complete top 10. Scored on winner call, exact podium order, top-10 presence and how far each pick landed from its predicted position — plus bonus calls like fastest lap, safety car and winning constructor.
Qualifying
Pole position and the top 5 on Saturday pace. Qualifying is the purest test of single-lap form, and the post-qualifying race prediction re-anchors on this result — so quali accuracy compounds into race accuracy.
Sprint
On sprint weekends the short race gets its own locked prediction and its own score — winner and podium presence — completely separate from the Grand Prix call.
Head-to-head
Driver-vs-driver calls — mostly teammate battles — scored as simple right-or-wrong verdicts. Small individual weights, but across a season they reveal whether the engine reads intra-team form correctly.
How is prediction accuracy measured?
Each prediction is compared against the official classification once a session is complete. A weighted rubric scores the winner call, podium order, top-10 accuracy and bonus calls, producing a 0–100% score per session. The same rubric is applied to every race, and every score is public.
Are predictions ever edited after the results?
No. Predictions are locked before the session starts and stored immutably — they are never rewritten, re-scored, or quietly deleted. What you see on the scoreboard is exactly what was published before lights out.
How often does the AI pick the race winner?
The live winner hit rate is shown at the top of this page and updates after every scored race. Formula 1 winners are genuinely hard to call — treat any single-race result as noise and judge the rate over a full season.
What is a good baseline to compare against?
A random pick from the 22-driver grid wins about 4.5% of the time, and even always backing the championship leader breaks down on chaotic weekends. Any sustained winner hit rate well above those baselines means the model is adding real signal.
What data do the predictions use?
The AI combines championship standings, recent driver form, circuit history, weather forecasts, and results from earlier sessions of the same weekend, plus grounded news search for late-breaking developments. It never sees the future — every input predates the session.
How is the accuracy score calculated?
Each session type has a weighted rubric — winner, podium order, top-10 accuracy for races; pole and top-5 for qualifying — normalized to a 0–100% score. The exact weights are published in the scoring section above, and every race page shows which individual calls hit.
Why do scores vary so much from race to race?
Because Formula 1 does. Chaotic races — rain, safety cars, first-lap incidents — scramble the field and punish every forecast, while processional races reward it. That is why the season-long average and hit rates matter more than any single weekend.
Do you ever remove or rewrite bad predictions?
Never. Predictions are immutable at the database level and the scoreboard is computed from those same locked records. Misses stay on the board with the hits — that is the point of a public track record.
How do the results compare to just guessing?
A random winner pick from a 22-car grid succeeds about 4.5% of the time. The baseline comparison above puts the engine's live hit rates next to those odds — any sustained gap between the two is real predictive signal.