The artificial football intelligence.

This page provides a little bit more information on what is happening behind the scenes.

How does predict match outcomes?

There are many organizations trying to predict the outcome of football matches and international tournaments. For example, for the 2014 World Cup, there were widely publicized predictions from Goldman Sach, UBS, Wolfram Research and Yahoo, among many others. Usually, these predictions rely on a mathematical model of football matches. These models can use data about past matches of national teams to forecast future confrontations.

What distinguishes from the rest? also uses a mathematical (statistical) model of football matches, but with two key differences that have the potential to greatly improve predictions:

  1. We model individual players' performance. Informally, we assume that the performance of a team is defined by the performance of its players on the field. In mathematical terms, whereas traditional models use one variable for each team, we (implicitly) use one variable for each player. Why do we do this? We do it because it enables us to take advantage of data from matches between clubs, and there is a lot (we mean really a lot!) more matches between clubs than between national teams.

  2. We use Bayesian inference. This is a fancy way of saying that we are able to understand how confident we are about a particular prediction. As an example, suppose that Germany plays against Iceland. It is likely that Germany will win—in fact, almost no one (with the possible exception of Icelanders) would claim that Iceland has better chances. But just how much more likely is a German win than an Icelandic win? 60%? 90% Perhaps 99%? On the one hand, Germany is clearly the better team on paper, but on the other hand Iceland has never taken part in a European championship final tournament, and might perform over its usual level. Bayesian inference takes this (and much more) into account.

And concretely?

Concretely, for a given match, we use the 11 lined-up players of each team as an input to our model in order to obtain the winning probability. If team A has a probability greater than 50%, it means that team A is more likely to win.

What data does use?

Because the mathematical model underlying's prediction engine "sees" a team as a collection of players, it is able to learn useful information from any football match. In particular, it can learn useful information not only from matches between national teams, but also from matches between clubs. For this reason, we collect a much wider range of data:

  1. International competitions. We use data from European championship qualifiers and final tournaments, World cup qualifiers and final tournaments, Confederations' cups, and friendly matches. These are matches between national countries, but represent a small fraction of the overall number of matches.

  2. European club championship. We use data from the UEFA's Champions League and Europa League. These competitions involve clubs across all European countries. Informally, they make it possible to understand how competitive different national championships are.

  3. National competitions. A few prestigious national championships contain most of the players that take part in international matches. We use data from the English, German, Spanish, French and Italian leagues. This data makes up the majority of matches that are available.

For each match, we reference the home and away teams, the date, each team's players that took part in the match, and the final score. All this is eventually used by the mathematical model to compute a prediction for a new match. In total, we use approximately 10 years of football data.

What is the Kickscore?

The Kickscore is computed based on the outcomes of matches in which the player participated. It can be approximately understood as follows: if the sum of the Kickscores of a team's players is high, the team is expected to perform well. Although the actual prediction is significantly more complex, the Kickscore measures the players' contribution to a team's success.

Where to find more information?

You can find more information about our methodology in the following paper:

L. Maystre, V. Kristof, A. J. G. Ferrer, M. Grossglauser, The Player Kernel: Learning Team Strengths Based on Implicit Player Contributions, Machine Learning and Data Mining for Sports Analytics (MLSA), 2016.

Other frequently asked questions

This section contains other frequently asked questions. Don't hesitate to ask further questions by dropping us an e-mail at

  • How should I interpret's predictions? We pride ourselves in providing a probabilistic prediction ("what is the probability that team A will win?"). In probabilistic terms, if team A has 78% chances to win against team B and a match is played 100 times between these two teams, then we expect team A to win 78 times over team B (or to lose 22 times). We believe that this is a much more interesting point of view, and it acknowledges that football is simply not always predictable.

  • Why did you make a mistake in the prediction of match [xyz]? Luckily for the interest of the match, football matches are not completely predictable. Sometimes, surprises happen (some surprises are less expected than others!). Focusing on a particular prediction misses the big picture, which is the overall accuracy when all matches are taken into account.

  • How do you deal with ties? Long story short, we don't (yet). Ties make up about a third of all football match outcomes, and in the future we will adapt our model in such a way that it will be able to predict ties as well.

  • Can I use your predictions to make bets? All the information on this website is published in good faith and for general information purposes only. Any action you take based on the information on our website is strictly at your own risk. (In short, we are not reponsible for any side effects of the luxurious lifestyle you will enjoy if you follow our predictions.)

Meet the team! is brought to you by the research group of Prof. Matthias Grossglauser and Prof. Patrick Thiran in the School of Computer and Communication Sciences at EPFL, Switzerland.

The football predictions that you see on this site are one particular application of the models and methods that we investigate in our research. These models and methods can also be useful for recommender systems, voting systems and many others. In fact, we hope that some of the lessons we learned while building will be useful in improving some of the general machine learning techniques that we make use of.

Press coverage