30.08.13

BLOG: The economics of possession

Article by Devin Plueler

In the 2012-2013 Barclays Premier League there was an average of 264 possessions per game, totaling over 100 thousand possessions for the entirety of the season.

Ranging from around 200 to 350, the distribution of possessions per game is roughly normal and wide. There are many factors that cause this range; one leading indicator being the amount of time a game spends in different game states. The way teams value their possessions when winning is much different from when they are losing, and the the numbers have shown that. Other factors, like random variation prove to contribute as well.

Possession, undoubtedly, is a rich data set. I have begun referring to possession as the “currency” of the game, with teams trading possession in ways that they best believe increase their chances of winning. With a finite amount of game time, possession is a limited resource with significant demand. But, just like conventional currency, not all possessions are cut with the same denomination.

In response, to help further our understanding of the economics of possession, we have been collecting additional metrics on a per-possession basis. One of these metrics is the number of passes in a possession, ranging from zero all the way to Tottenham’s 45-pass sequence against Liverpool that lasted well over 2 minutes (and included nearly a kilometer of ball movement). Another major metric is ball speed, calculated as the total distance the ball moved in possession divided by the amount of time that the ball was in possession.

At this point, it’s probably valuable to explain how we are defining possession. Through the sequence of chronological game events, our algorithm specifies if each event starts and/or concludes with a team in control of the ball. This is straight forward with events such as passing, but our model also controls for more ambiguous events such as aerial duels and clearances. When a team controls the ball through sequential events, it is considered in possession.

Plotting ball speed and passes per possession against each other, we get a glance under-the-hood. The overarching mechanics of possession prove to be incredibly elegant.

The number of passes in a possession is bounded by a function of ball speed that resembles a long right-tailed distribution. Since it is impossible to move the ball at a high-speed while permanently retaining possession, these findings are intuitive. Perhaps the larger takeaway is that it’s possible to possess the ball at an unsustainably low speed as well as at an unsustainably high speed.

Since we have laid out some of the inherent limitations of possessions, it’s a natural progression to be curious about which teams are pushing these boundaries. Which teams have an exceptional ability to retain the ball while playing at a high speed?

This is the average passes per possession and ball speed for each team during the 2012-2013 PL season. The correlation is not huge (r^2 is 13%), but significant variation is expected across playing styles. The low residual doesn’t bother me since the previous plot already shows this relationship is very strong. Models aren’t purely about impressive r^2 scores.

Data points above the line suggest that a team’s average passes per possession was higher than expected given their average ball speed. Conversely, points below the line suggest that a team’s passes per possession aren’t quite up to par. Here is a list of teams accompanied by how many more (or less) passes they tend to complete given their average speed of play.

Team Average Passes Above Expected
Swansea 1.196
Arsenal 1.053
Manchester United 0.899
Manchester City 0.846
Liverpool 0.559
Chelsea 0.400
Fulham 0.304
Wigan Athletic 0.258
Everton 0.003
West Bromwich Albion -0.005
Newcastle United -0.015
Southampton -0.032
Tottenham Hotspur -0.034
Norwich City -0.137
Aston Villa -0.221
West Ham United -0.598
Sunderland -0.651
Reading -0.654
Stoke City -0.668
Queens Park Rangers -0.693

The ordering of this list is just about what you would expect. Any time Swansea is brought up in conversation, it’s commonly found in a discussion about possession oriented styles and “pass first” philosophy. On occasions, I’ve heard their ball movement likened to a metronome, suggesting that it’s not just the amount of Swansea’s passes that are exceptional, it’s also their rate. This table seems to lend credibility to the analysis.

Stoke City — the famous practitioners of direct football — fall on the opposite end of the spectrum, completing 0.67 fewer passes per possession than their average ball speed would otherwise predict. Given Stoke’s tactical disposition of knocking it long and hoping for a fortuitous bounce, it shouldn’t come as much of a shock that they’re completing fewer passes than they should be.

Tottenham’s position, however, seems like a surprise. By segmenting their games from the rest, we can get a glimpse into exactly why. Tottenham, like most top PL sides, play on the lower end of the ball speed spectrum — a byproduct of spending more time in favorable game states and therefore being forced to attempt fewer long balls. It seems that Tottenham suffered from wild inconsistency in these metrics. While being exceptional during some games, Tottenham were less than average during others.

(I’ve highlighted matches against the “big four” of Manchester United, Manchester City, Chelsea and Arsenal in red as they tend to cause outlying data points)

Other teams seem to have a lot more consistency. Like Liverpool:

Or Everton:

Or even Stoke:

What caused the drastic game-by-game variance for Spurs? I’d venture to guess that AndrĂ© Villas-Boas was behind it, and the wide range of tactical approaches that he exercised. When playing against the big fish of the Premier League, Tottenham focused on exchanging their possession for much different types of opportunities than the chances normally generated against the rest of the league.

While attempting more shots per game against the lower teams, Tottenham’s chance quality was notably higher against the top 4. In essence, Tottenham was able to switch gears, exchanging lots of low-value possession for fewer, high-value possessions.

Team Shots Quality
The rest 19 7.67%
Top 4 13.9 9.29%

Quality is a estimation using Opta models as to how likely a particular shot is to result in a goal.

We also learn a valuable lesson here: optimizing particular metrics are not mechanisms for winning football matches. For different teams, in different situations, different metrics are valuable. And, in the case of Tottenham and Villas-Boas, having the flexibility to optimize different things has made all the difference.

 

Back to analysis