Our simulator's scoreboard: four races called before they ran, and the grid still beats us on average
F1 Legend Sim · Staff · 22 September 2026 · 7 min read
Everything else we publish here looks backwards, and looking backwards is easy: the result is already written. Since July we have been doing something else. On Saturday, with the official grid published and before anyone starts the engine, we simulate the Grand Prix 1500 times and post the probabilities. On Monday we add the real result and nothing else is touched. That is 4 races now, and this is the scoreboard, including the parts that go wrong.
The 4 races, one by one
«Error» is how many places the model misses by, per driver, counting only those who finished. The column next to it is the cheapest prediction there is: assume they finish in exactly the order they start. That is the benchmark to beat, because a model that cannot beat it is adding nothing that Saturday qualifying did not already give you.
| Grand Prix | Model error | Grid error | Top 8 | Points scorers | Retirements |
|---|---|---|---|---|---|
| Hungary | 2.69 | 1.84 | 8 of 8 | 8 of 10 | 3 |
| Netherlands | 2.20 | 2.31 | 8 of 8 | 8 of 10 | 6 |
| Italy | 2.64 | 3.42 | 7 of 8 | 7 of 10 | 3 |
| Spain | 2.43 | 1.44 | 7 of 8 | 9 of 10 | 4 |
Start with what does not work. On average the model misses by 2.49 places per driver and the starting grid misses by 2.25. Which means it loses. Four races is not a verdict, but we are not going to hide it either: if all you want is to guess the finishing order, the starting order is still a better bet than our 1500 simulations.
Where it does win, and it is the part that matters
The average hides the interesting bit. The model wins 2 of the 4 races, and they are not random ones: it wins exactly where the grid lies.
At Italy the starting order was out by 3.42 places, its worst figure, because half the grid was out of position on engine penalties. The model stayed at 2.64. At Spain the opposite happened: the race sorted itself almost exactly like qualifying (1.44) and there the model had nothing to add.
Put another way: the model's error moves between 2.20 and 2.69, a spread of 0.49 places. The grid's moves between 1.44 and 3.42, a spread of 1.98, four times wider. The grid is better on average and far worse when it fails. The model is boring and never falls apart.
The two things it genuinely gets right
Where the scoreboard is clearly good is in saying who will be up front, which is a different question from in exactly which place. Of the eight drivers the model put at the front in each race, it got 30 of 32 right. Of the ten who score points, 32 of 40.
The actual winner is another story: it was our favourite in 2 of 4. That is not a flaw in the model, it is Formula 1 itself: in none of the four previews did the favourite go above a 23% chance of winning, because when four cars can win there is no 60% favourite. Anyone publishing a winner prediction with more confidence than that is selling certainty they do not have.
Who it nails and who it misses
The big errors are not spread evenly. These are, in each race, the driver the model nailed and the one that got away from it by the most.
| Grand Prix | Best call | Error | Biggest miss | Error |
|---|---|---|---|---|
| Hungary | Lewis Hamilton | 0.30 | Lance Stroll | 6.00 |
| Netherlands | Franco Colapinto | 0.10 | Nico Hulkenberg | 7.70 |
| Italy | Oliver Bearman | 0.10 | Gabriel Bortoleto | 6.00 |
| Spain | Alexander Albon | 0.20 | Franco Colapinto | 6.00 |
All four biggest misses are midfield and back-of-the-grid drivers. There is an explanation that is not an excuse: the engine gives each team a single number for car performance, and in the bottom half of the grid the real gaps between one car and the next are smaller than the noise of a single race. Up front, where the gaps are large, the order comes out almost every time; at the back, five cars worth practically the same get sorted by whatever happens on Sunday, and that cannot be predicted.
Why none of this can be recalculated
There is a rule we have set ourselves and it is worth writing down: published previews are never regenerated, not even when we change the engine. It is tempting, because today's simulator is better than July's and the figures would improve. But a prediction recalculated with what you learn afterwards is not a prediction, and this whole scoreboard would collapse the moment we did it once.
So the numbers above are what they are, with their 2.49 places of average error and the loss against the grid included. The live scoreboard, race by race and with each official classification, is in the Grand Prix panel, and it grows by itself every Sunday. If you would rather compete against it, Predict the GP gives you the same grid to sort yourself. And how the engine is built inside, in how the simulator works.