How football supercomputer predictions actually work

Every August, with the leagues barely under way, the sports pages run the same kind of headline: the supercomputer has predicted the Premier League, or the model gives the favourite a 62 per cent chance of the title. The word conjures up an enormous machine humming away in a refrigerated basement. The reality is a good deal humbler and a good deal more interesting: it is almost always a statistical program that distributes goals according to a formula more than forty years old, and then plays the season through thousands of times in a row to count how often each side comes out on top. Understanding what sits underneath is worth doing for two reasons: it shows why these models are more accurate than you might expect, and it shows exactly where they stop being useful.

Where the August headline comes from

The format was popularised by the sports data companies. Opta's model, for instance, simulates every match of the season ten thousand times and ranks the teams by the average points they accumulate across that set of imaginary seasons. From that raw material you can build any headline you like: title odds, chances of European qualification, relegation risk, or the most likely final table. It is cheap to produce, it updates itself with every matchweek and it performs beautifully on social media, so it has become a journalistic genre of its own that resurfaces punctually every pre-season.

What is striking is that the small print usually contradicts the headline. In Opta's projection for the 2025-26 Premier League, nineteen of the twenty teams won the title at least once across those ten thousand simulated seasons. Put another way: the model itself was warning that almost anything was possible. The headline picks the side that won most often and presents it as a settled prediction; the model, meanwhile, was describing an enormous fan of possible futures, some deeply improbable but none of them impossible.

Maher, Dixon and Coles: the academic origin

The underlying idea is older than the internet. In 1982, M. J. Maher published a paper in Statistica Neerlandica called Modelling association football scores, in which he gave every team two numbers: an attacking strength and a defensive strength. The goals in each match were modelled with a Poisson distribution, the same tool used to count infrequent events within a fixed interval of time. Maher found that, with some caveats, this very simple model described the real frequency of scorelines reasonably well, and he built in the advantage of playing at home as well.

Fifteen years later, in 1997, Mark Dixon and Stuart Coles published in the Journal of the Royal Statistical Society, Series C, the paper that turned all this into a practical tool. They fitted the model to English league and cup matches played between 1992 and 1995 and made two decisive repairs. The first: correcting the frequency of low scorelines, because the pure Poisson fell short on goalless draws and one-all draws. The second, and the important one here: introducing a time weighting, so that recent matches counted for more than old ones when estimating a team's strength. A club in April is not the club it was two years earlier, and the model had to notice that.

Why you have to play the season ten thousand times

Once attacks and defences have been estimated, you can work out the probability of every possible scoreline in every remaining fixture. The trouble is that a league is not a single match: it is a web of dozens of matchweeks in which the result of each one changes the meaning of the next. There is no reasonable closed formula that turns all those individual probabilities into a probability of finishing champion, still less when goal difference or head-to-head records come into play as tiebreakers between the sides involved.

The way out is Monte Carlo simulation, a method that solves a hard problem by rolling the dice an enormous number of times. The computer draws a concrete result for each remaining fixture, following the probabilities the model dictates; it adds up the points; it builds the final table; and it records who won the title and who went down. Then it starts again from scratch. After ten thousand complete seasons you simply count: if a side finished top in 6,200 of them, its title probability is 62 per cent. There is no more magic than that, and the name comes from the casino in Monaco, not from any supercomputer.

What 62 per cent really means

Here lies the central misunderstanding. Sixty-two per cent does not mean that the team is going to win the league. It means that if the season were played out a hundred times from where it stands now, with these squads and this fixture list, that side would end up champions roughly sixty-two times and somebody else would take it in the other thirty-eight. And 38 per cent is nothing to sniff at: it is almost exactly the chance of getting two heads in three coin tosses, which surprises nobody when it happens.

That is why a model is not judged on whether it picked the champion but on its calibration: of all the times it said twenty per cent, did the thing happen about one time in five? A well calibrated model is wrong constantly in individual cases and is still useful, in the same way that a forecast announcing a thirty per cent chance of rain is doing its job if it rains three times out of every ten it says so. Judging a probability by one outcome is like judging a die by a single roll.

How this site's supercomputer works

The model behind futsimulator.com belongs to that same family of ideas. It plays every remaining fixture in the competition ten thousand times over and counts the outcomes. Each club's attacking and defensive strength is not taken from the raw table but adjusted for the opponents it has already faced: putting three past the leaders is not worth the same as putting three past the bottom side, and a defence that has already been to the grounds of the top four deserves to be read on a different scale from one that has barely left home.

Following the Dixon and Coles logic, recent matches weigh more than distant ones, with a half-life of one hundred days: a result from three months ago carries roughly half the influence of one from this week. And for the classic problem of the opening weeks, when there is not yet enough data, the model starts from a prior built out of the club's squad value and its finishing position the previous year; that initial weight fades as real results arrive. One technical detail worth mentioning: the entire calculation runs in the visitor's own browser, not on a remote server. The supercomputer, quite literally, is your own device.

What no model can know

And now the honest part. These models read the past; they do not read the future. They do not know that the first-choice striker will get injured in matchweek nine, or that the club will sign a midfielder in January who changes the whole complexion of the side, or that there will be a change in the dugout that reorders the dressing room completely. They know nothing about the European fixture load that will empty the legs in March, nor about an administrative sanction, nor about an unexpected transfer on deadline day. All of that enters the model only after it has happened, once it has already shown up in results.

Nor do they capture well what statisticians call correlation between matches: in the simulation each fixture is drawn more or less independently, whereas in real life a bad run carries over from one weekend to the next and a long injury weighs for months. That is why the ranges they produce tend to be a little narrower than life eventually demonstrates. The sensible conclusion is not to bin the model but to use it for what it is: an orderly way of telling the probable apart from the merely possible, without ever confusing the two.

"A model does not tell you what will happen; it tells you how often it would happen if the year were replayed ten thousand times."

• Simulate the remaining matchweeks of your league and see how the final table lands.
• Change any result by hand and watch the standings shift instantly.
• Build knockout brackets and simulate them round by round all the way to the final.
• Open the supercomputer, read the title and relegation probabilities and compare them with the real standings.

"Statistics do not replace the season: they only tell you what to expect before it is played."