The five seasons our simulator cannot explain

We re-ran 30 real seasons, one hundred times each, with the actual grid of every year. The engine picks the right champion in 25 of 30. These are the five it gets wrong. And the interesting part is not that it misses, it is that three of the five were settled in a single afternoon.

What “getting it right” means here

Each season is run one hundred times with its real grid: that year’s cars with their performance and reliability, that year’s drivers with their level, and the calendar that was actually raced. Those hundred runs produce a hundred champions, and they are not always the same person. The number we use is how many of the hundred are won by whoever won in real life.

So “getting it right” does not mean the engine predicts the future. It means that when you hand it a year’s machinery and let it run, the real champion is the name that comes up most often. That happens in 25 of 30 seasons. In five, it does not.

2016: the biggest miss, and the easiest to explain

The engine makes Hamilton champion in 92 of every 100 runs of 2016. Rosberg won. It is the widest gap between model and reality in the whole series.

It is also the most understandable of the five, because the two of them drove the same car. When the machinery is identical, all that is left to separate them is the driver and luck. The model says that across twenty-one weekends Hamilton finishes ahead almost every time. In 2016 he did not, and the reason fits in one line: an engine failure while leading in Malaysia, five races from the end.

The simulator does know engines fail. It spreads those failures across the year according to each car’s reliability. What it cannot do is fail that Sunday, on that lap, with that driver in front. A championship can turn on one bad roll at the worst possible moment, and an average of a hundred seasons is exactly what erases that detail.

1997: what a simulator has no box for

The engine gives Schumacher 88 of every 100. Villeneuve won, and comes up in eleven.

This miss is not about reliability or luck. The 1997 season was decided by one move and one sanction, and neither exists inside the model. Jerez, final corner: Schumacher turns into Villeneuve, Villeneuve survives it and takes the title. The FIA then removed Schumacher from the final championship standings altogether.

A simulation engine has slots for pace, for qualifying, for tyre wear and for mechanical failure. It has none for a driver deciding to drive into another one, and none for a governing body taking away his runner-up place in a meeting three weeks later. That is a limit of the model, and it is worth saying it that way rather than dressing it up as Schumacher having “deserved” the title.

2007: the only one it half understands

In 2007 the model has no clear favourite, and that in itself says something. Alonso gets 50 of 100, Hamilton 29 and Räikkönen 21. Räikkönen won, by a single point, in the last race, with the only result that would do it.

The cause is what was happening inside McLaren. Alonso and Hamilton took points off each other all year, and the engine does model that: two fast cars in the same team get in each other’s way in simulation just as they do on track. What the model splits between the two, reality split in the most extreme way available, and a Ferrari came through the gap.

It is the case where it is least wrong, because it was never confident. A 50% is not a prediction. It is a shrug with a number attached.

2006 and 2025: the two near misses

The other two are narrow. In 2006 the engine gives Schumacher 52 and Alonso 47, five points apart, and Alonso won. That season turned, again, on an engine failure while leading in Japan.

And in 2025 it gives Piastri 44 and Norris 38. Norris won it by two points in Abu Dhabi. With six points between them the model was saying this one was a coin flip, and it was right about that even though it got the name wrong.

What the five have in common

Three of the five seasons (2016, 2006 and 2007) were decided by a mechanical failure or by how points split at one particular moment. A fourth, 1997, by a move and an administrative decision. The fifth, 2025, by two points in the final race.

Not one of them was decided because a car was better across the year. And that is exactly what the simulator measures: how good the machinery was across the year. When a championship is won that way, the engine sees it coming, and it does so in 25 of 30 seasons. When it is won on one specific afternoon, it does not.

So what do we do with this

Publish it, mainly. A model that only shows its hits is not a model. These five seasons are the best reminder of what a simulator can and cannot say: it is a good tool for questions like “how much was the car worth?” or “which car was furthest ahead of its time?”, and a bad tool for “who was going to win?”.

If you want to watch it happen live, we publish the odds for every Grand Prix before it is run and the contrast with the result afterwards. It is all on the model’s scoreboard, hits and misses alike.