“Artificial intelligence predicts Brazil as the winner of this year’s football World Cup.” – we heard the news on Kossuth Radio on 05.12.2022. Researchers at the Alan Turing Institute used machine learning to simulate the outcome of this year’s World Cup. After running 100,000 simulations, they concluded that the South American country – who are the most successful in the history of the game with 5 World Cup titles – were the most likely to win. But even that was given a probability of only 25 percent. The 2nd and 3rd most likely were predicted to be Belgium and Argentina.
But how are these predictions made? Levente Szabados, artificial intelligence researcher and senior consultant at Neuron Solutions, was asked about this on Kossuth Radio. He did not know at the time that in 4 days Brazil would disprove the prediction. But don’t despair, there’s nothing wrong with artificial intelligence, which – although he couldn’t have known how actual it would be in a few days’ time – is exactly what Levente was talking about in his closing thoughts. Well, let’s see.
Levente explained the complexity difference between the “simple” calculations of betting agencies and the logic of machine learning in a way that is understandable for laymen.
Obviously we have some statistics that we have been monitoring to some extent. For example, ball possession, player’s preferred foot, etc. We have had such data for decades, but it is derived data, i.e. we watch the game, not the machine, we label the data.
Can artificial intelligence already help in this process? Levente told us that image processing has already come to the point where it can easily extract this analytical data from the raw video for us, rather than us having to write it down with pen and paper. So teams can get very clear analytics on exactly how they performed, which is a really powerful use of artificial intelligence in sports.
The other application is when we use the machine to make estimates instead of ourselves. The kind of estimates that bookmakers do to determine the odds. And in fact, machine learning methods are not that different from their methods, because they are both based on statistics, but machine learning uses much more powerful tools.
The human statistician can examine as many correlations as fits in the mind, has a finite capacity, and its imagination, its concentration, its mathematical ability limits what it can model from the future given the past.
In contrast, a machine tool provides a much more powerful solution, because it takes into account all the little connections that we can’t imagine are there. We cannot even follow these connections with our human brain, we cannot see through the connections that have led to the result. This is the power of machine learning.
So by looking at these correlations, artificial intelligence can make probabilistic predictions about the future, including the outcome of future football matches. However, Levente reassures us that we don’t need to worry, the matches will not be boring in the future either, because we can predict the boring things (like Barzil, a team that has been very successful in recent years, has a big chance) and not the exciting unexpected things. So there is no reason to worry, but we will have a word or two to say about this for the other teams, and of course about coincidences.
Finally, a general lesson from Levente’s repertoire: It follows from the definition of estimation that if there is an 80% probability that an event will occur, then there is a 20% probability that it will not. “In other words, just because an estimate doesn’t work out doesn’t make it a bad estimate.” – Levente concluded the conversation, which surprisingly became very actual four days later with the elimination of Brazil.
Follow Neuron Solutions for more news on artificial intelligence!