- Researchers built a Random Forest model to predict the World Cup 2018 winner
- They simulated the entire tournament 100,000 times to arrive at the final winner
- The data they used was from each World Cup starting from the 2002 edition
- According to the model, the players’ individual abilities rank as the most important feature, followed by the country’s FIFA ranking
The World Cup begins tomorrow and predictions are in full swing. We have seen Paul the Octopus weaving his magic to predict the 2010 winner (Spain), and we otherwise look at bookmakers odds to gauge who the favorite is. But now that we have the power of machine learning, it wasn’t long before data scientists put it to good use.
Led by Andreas Groll from the University of Dortmund, the researchers used a random forest model to predict every single game at this year’s World Cup. Their model is predicting Germany to go all the way. The below chart shows the win probabilities of each match in the knockout stages.
The researchers repeatedly simulated the World Cup 100,000 times to arrive at these numbers! Data from the 2002 World Cup through to 2014 was used to train the model. Groll and his team initially used a range of variables to start off the model building phase, including economic factors like the country’s GDP. Other variables included FIFA’s ranking, average age of the squad, how many Champions League players are there, is there any home advantage, etc.
Interestingly, the final model gave some fascinating insights. As you might be aware, the Random Forest algorithm also has the feature importance functionality up it’s sleeves. Check out the below bar plot which ranks the features:
The highest importance was given to the abilities of individual players, followed by their rank on FIFA’s list. Other moderately important variables include the average age of the squad, how many players play in the Champions League and the GDP of the country. Factors like the nationality of the coach and the population of the country turned out to essentially useless.
One of the more intriguing aspects of the model was that overall, it predicted Spain to have the likeliest chances of winning. But if Germany (who potentially face stronger opponents in the knockout stage) reach the quarter-finals, they have a higher chance of winning.
You can read the full paper describing this research here.
Our take on this
I’m a huge sports analytics buff so reading the entire research paper was like a goldmine to me. The workings of the model these guys have built is fairly easy to understand and follow. Having said that, sports is a very unpredictable field and anything is possible on the day.
Of course, this is not the only machine learning effort to predict the winners. Goldman Sachs have also used a similar approach (though their report doesn’t delve into the ML side too much). Their model has predicted a Germany vs. Brazil final with the Samba nation taking the crown.
Who are you predicting will lift the trophy this year?
Subscribe to AVBytes here to get regular data science, machine learning and AI updates in your inbox!
You can also read this article on our Mobile APP