When former Nigerian international, Yakabu, failed to stick the ball into an open net from 2 yards out at the 2010 World Cup, many declared it the worst miss ever (if you don’t believe them, take a look at it yourself). I was only 11 years old, a hopeful Zambian rooting for a fellow African team, when big Yak let me down. To this day, I don’t think I’ve forgiven Yak and in all honesty, I’m not sure I ever will.
Any football player who receives a goal-scoring opportunity as good as Yakubu’s would be expected to score. It is from this intuitive understanding of football that the advanced metric, expected goals (xG), is derived. Think again of that Yakabu miss, or any other worst miss ever that springs to mind, and ask yourself, “What made that chance such a good chance?” In Yakubu’s case, the proximity to the goal (a mere 2 yards!), the lack of pressure from defenders and the fact that he had the entire goal to aim at, were all incriminating factors.
xG calculates the probability that a particular shot will result in a goal. You can think of xG models as working in a similar way to your own thought process about what makes a chance a good chance. A sophisticated xG model takes inputs such as the player’s proximity to the goal, the angle at which the player is from the centre of the goal and which part of their body they use to shoot (head, left foot, right foot or other body part). Using such inputs, these models can determine the probability that a particular shot will result in a goal.
But how exactly?
In Yakubu’s case (I hate to have to keep reliving this nightmare), you might reasonably say that the probability of him scoring that particular chance was 0.9999 (cue infinite 9s). Essentially, what you would be saying is that if he were given that exact same chance 10000 times, you would expect him to score 9999 goals. xG models work by using historical data about a particular type of shot to predict the probability that a player, taking a similar type of shot, will score. As an easy example, let’s talk about penalties. Historical data shows that penalties in the English Premier League are converted about 77% of the time. This means that for a given penalty, the probability that a player will score is 0.77. In the same way, xG models calculate goal probabilities for any shot by looking at how often shots similar to it were converted in the past.
Why is measuring xG useful?
The low scoring nature of football means that random variation affects match results to a significant extent. As a rather extreme example of this, you might think of Sunderland’s 1-0 win over Liverpool in 2010 after Darren Bent’s shot was deflected into the goal via a beach ball thrown onto the pitch by a rather notorious fan. On the day, that single freak goal was enough to win Sunderland the game. In contrast to this, the effect of such a random event would be minimized by the large number of goals (baskets) scored in a single basketball match. Thus, in football, if one’s interest is to determine how well a team is performing, the final scoreline does not always provide a clear picture. This is where xG comes in: it allows for an objective evaluation of the quality of chances that a team creates in a particular game, despite the final scoreline.
There’s one other thing you might have been thinking.
Why do xG models assign the same probabilities of scoring to every player? Doesn’t it matter who takes the shot?
The short answer to this question is ‘not really’. However, it’s a complicated question that will need to be discussed in far greater detail.
In fairness, you probably have many more questions. If I have written this well, you should. In subsequent posts, we will continue to discuss expected goals, including its applications and limitations. We will also discuss other advanced metrics such as expected assists and expected points.
Please subscribe to receive notifications for new posts!