Artificial intelligence predicts the winners of Euro 2016 games

Victor Kristof combines his passion for football and research.© 2016 EPFL / Alain Herzog

Victor Kristof combines his passion for football and research.© 2016 EPFL / Alain Herzog

EPFL researchers have developed a website that calculates the likely winners of Euro 2016 soccer games, using a more complex and accurate model than conventional statistical methods.

Can EPFL do better than Paul the Octopus? That’s what the developers of are hoping, with a new website that uses artificial intelligence to predict the results of Euro 2016 soccer games starting with the very first round. According to their calculations, France should beat Romania in the opening game on Friday, and Switzerland should come out ahead of Albania in their game on Saturday.

The three researchers – who are also avid soccer fans – developed their model at EPFL’s Computer Communications & Applications Laboratory. Their model uses a combination of machine-learning algorithms and data mining, making it more accurate than most predictive systems currently out there. “A lot of websites aim to predict the winners of soccer games, but they draw on data about the past performance of each country’s team. Our method uses data about each individual player on a team. That expands our data set to thousands of games,” said Victor Kristof, one of the three developers. Even though he uses the word “predict,” what their website actually does is compile data from prior games to objectively calculate the likelihood of one team beating another. “We don’t just determine which team is better – we calculate exactly how much better,” Kristof said.

Ten years of data crunches all sorts of numbers, like how much time a player spends on the field, whether the player’s team won the game, and whether the game was played at home or away. The strength of each team is calculated using an impressive amount of player data that spans back ten years. But isn’t just a classical statistical model. It uses a Bayesian method to calculate a weighting for each player based on the degree of uncertainty in the prediction. “If Germany plays Iceland, for example, the probability of Germany winning is high on paper. But Iceland has never made it to the finals of the European championships, and could therefore play better than they usually do. That is taken into account in the weightings used for our prediction,” said Lucas Maystre, another developer.

Tests carried out using the results of Euro 2012 games show that EPFL’s model gives better predictions than conventional methods. But the developers don’t claim to have come up with a crystal ball. “Our predictions are just probabilities. Soccer isn’t an exact science. There’s always an element of the unknown – fortunately for its fans!” said Kristof. The algorithms developed at EPFL could have an array of other applications and potentially give rise to further advancements in the use of machine learning for R&D. was developed by Victor Kristof, Lucas Maystre, and Antonio Gonzalez in the EPFL lab run by Professors Matthias Grossglauser and Patrick Thiran.

Author: Sarah Bourquenoud

Source: EPFL