Which team has the best defence in the Premier League?
While we often focus on the big name attacking players in football, it is often the defence and goalkeepers that get results and win titles. This season Chelsea have not only been effective going forward, but they have also conceded only 17 goals. Second place Tottenham Hotspur have conceded only 16. Compare this to the 29, 28 and 30 goals scored against Manchester City, Arsenal and Liverpool respectively.
In an earlier article I looked at the geometry of shooting and identified the best place to shoot from. Shots from the central part of the box are more likely to result in a goal than those from further out or at more oblique angles.
In that article I looked at the attacking team, i.e. the position from which the team scores. Below I show a contour plot showing the position teams are conceding from. On the left are the conceding contours for all teams in the Premier League this season. On the right I show a similar plot just for Chelsea.
Chelsea have been particularly successful at preventing the opposition from scoring from more oblique angles: their 30% contour is much narrower than an average of all teams. Combined with the fact that they have the third lowest number of shots made against them (7.25 per match) in the league, this has meant they have conceeded very few goals.
At the other extreme are Liverpool and Manchester City, shown below.
Both of these team’s conceding contours are much wider than the average for the league. Liverpool are particularly vulnerable, conceding a large number of goals from shots directly in front of the goal. Given that The Reds are the Premier League team with the lowest number of shots made against them (6.29 per match) letting in shots in front of goal is their only real weak point this season. The story is similar for Manchester City, although their weakness appears to be from longer range efforts.
The team that has conceded the least goals per shot in the box in the Premier league this season is Hull City. Out of the teams in the top six, it is Tottenham Hotspur who have conceded the fewest goals per shot near to goal.
Spurs have a slightly unusual looking shape for their conceding contours, which I won’t comment further. But the general message is that both these teams have been effective at blocking and stopping shots this season compared to the league average.
There are three possible explanations for Liverpool and Manchester City’s conceding more goals than expected: ‘bad luck’, ‘bad defending’ or ‘bad goalkeeping’. There are also three possible explanations for Hull and Spurs conceding fewer than expected: ‘good luck’, ‘good defending’ or ‘good goalkeeping’. In the ‘geek box’ I discuss the problem of over-fitting models that could be part of the ‘bad or good luck’ explanation.
The differences for Manchester City and Liverpool, when compared to the league average, are big enough that Pep Guardiola and Jörgen Klopp should seriously consider the possibility that their defence isn’t just unlucky. In particular, Guardiola should be concerned that his goalkeeper, Claudio Bravo, is letting in too many mid-range chances. At Hull on Saturday, Klopp’s team conceded two goals from exactly the area that they have been having problems for most of the season. Here I would attribute most blame to the central defenders. Shots from this area are difficult for any goalkeeper to stop and should be closed down early by the defence.
Geek box: over-fitting data
To create the contours above I fit the same model as I described in an earlier article on shot geometry to individual teams, rather than over multiple teams and seasons.
A potential problem, raised by Toronto FC analytics expert Devin Pleuler, is that using team level shot data can lead to us over-fitting the data. Because goals are infrequent, it is difficult to be sure that a pattern isn’t due to randomness rather than a genuine difference between teams. The exact same curvy patterns around the penalty spot created for Manchester City and for Spurs are unlikely to be seen in future matches. Many of the differences in the shapes can be best described as statistical anomalies. However, as I write in the main article, for teams with large deviations from the league average the managers need to ask why the difference is so big.
Some of these difficulties with over-fitting could be addressed using a Bayesian framework. The season average from last year could be used as a prior distribution and the goals conceded by each team as the new evidence. If this was done, then the question would be how to balance the weighting on the prior and the evidence. I’m afraid this would be even more difficult to explain to the manager than the contour plots…so I’ll leave it for Devin and other analysts to work out!
