Predicting results with machine learning | Mathematical science

UD math professor explains how machine learning can impact coaching decisions

The World Cup has begun! An unprecedented 48 countries will participate (up 16 from four years ago), making this one of the most-watched tournaments in the history of the game. Everyone has an opinion on who wins. Legendary English striker and football commentator Alan Shearer believes France will win the title if there is no domestic contention. US soccer hero Landon Donovan thinks about Ecuador, Norway and Colombia He might be the dark horse of this tournament.

Every fan has a favorite team, even if it’s unlikely to win, but how do statisticians and data scientists think about predicting games? Data analysis isn’t new to sports. Quantitative analysis has been used in baseball for years, with the Oakland Athletics famously using sabermetrics to decide which players to buy and sell in the early 2000s. The success of using data to inform sports strategy led Michael Lewis to write the 2003 book Moneyball: The Art of Winning Unfair Games, which was later made into a movie starring Brad Pitt.

Algorithms and features

Machine learning (ML) is a field of artificial intelligence in which computers learn patterns and make predictions. Although it’s been around since the mid-2010s, using ML to predict outcomes is relatively new to sports, and different approaches exist. Standard ML algorithms include logistic regression, K-Nearest Neighbors, Support Vector Machines, and Naive Bayes, but all of these techniques have in common that they require training data. Before producing results, ML needs to know how the team has performed in the past. What were the terms of the match? How experienced is the team?

For example, ML can predict the outcome of a theoretical match between the US and Brazil using the K-Nearest Neighbor (KNN) algorithm. See Figure 1. A team’s average number of caps (international appearances) is an indicator of a team’s experience, and if the average number of caps for the US team is 26 and Brazil’s number is 30 (indicated by a black cross), this metric can be used to determine the outcome of the match.

For K=4, KNN takes the four nearest neighbors of the black cross and determines the most occurring category (American win, Brazil win, or tie). Two of the games were won by Brazil, one was won by the United States, and one was tied. Therefore, the algorithm predicts a victory for Brazil. In general, algorithms require that the data be numeric so that there is a concept of distance (without distance, the word “nearest” is meaningless). Importantly, however, there is no limit to the number of features in the data set. In this example, the KNN is applied in two dimensions, but other conditions such as humidity, temperature, average wind speed on game days, average age of each team, and number of days of rest before the game can be included as additional dimensions. Currently, humans decide which features to include in a dataset. Practitioners (and probably most soccer fans) believe that choosing the right features is more important than which algorithm to use.