If we backtrack even further, a chance denied by rapid closing down or a well timed intervention well away from goal is even harder to record and quantify. As other sports have discovered, offensive production is quantified by events that occurred, but defensive ability is often a mixture of unseen and unrecorded actions that lead to things not happening.
Attacking prowess at the most basic level is recorded in goals scored, which in turn can be more elaborately investigated by more numerous connected events such as shots, attacking passes and more recently final third entries and gaining of possession in the opponents half.
Goals and shots allowed are the defensive equivalent of the attacking side of the ball's mainstream statistics and we can begin to unravel defensive metrics by thoroughly examining these readily available numbers. If we can begin to identify teams which are limiting their opponents scoring attempts we can perhaps then begin to try to see the actions they are taking to accomplish this.
I'll use shooting accuracy to try to illustrate the point. Shooting accuracy may represent a collective or individual ability to create "better" goal attempts, since for a goal to be scored a shot must first be on target. Therefore, if we see defenses which consistently cause opponents to shoot inaccurately, we may also be seeing evidence, which can then be identified, of effective defensive actions.
We can superficially see if shooting accuracy lies more with skilled sides by plotting raw shooting accuracy against seasonal success rates and the correlation is both reasonably strong and positive.
If we accept that shooting accuracy may be a skill, we can expand our investigation by looking at the ability of defences to cause opponents to shoot inaccurately. All shots aren't created equally at source, although in large enough sample this problem may be reduced. Therefore, I used my goal probability model, which uses x,y data to estimate the likelihood that a shot will strike the target, miss altogether or be blocked, based on the pitch position from where the shot originated.
Taken over all of the recent members of the EPL, the model produces a benchmark by which an individual shot can be measured. However, by introducing variables to represent individual sides, we can also see if the ability to make sides miss is a significant and consistent ability possessed by certain defences during recent Premiership history. Conversely, we can also identify sides which may be significantly poorer defensively and allow opponents to hit the target more frequently than is the general case.
Based on my dataset, Fulham appear as a side which consistently made hitting the target difficult for opponents compared to the league average. Using a position on the edge of the box and 5 yards to the left of centre as a bench mark figure, (roughly where Crouch is in the photo below) the generic average for the model was a 30% chance of hitting the target and the predicted average outcome for opponents of Fulham, based on the recent record of the London side was for them to be successful just 25% of the time. The recent Arsenal defence fared less well, consistently under performing by allowing slightly more on target shots than the average on all ranges of shots and, as an illustration, allowing opponents 36% accuracy for our 30% test shot.
|Stoke demonstrate their below average shooting accuracy.|
From the plot above,shooting accuracy appears to be a trait of successful sides. Therefore preventing a side from hitting the target would seem to be a sensible aim, although correlation does not imply causation. Some sides, in this case Fulham appear to be statistically different in this skill than the average team. They made opponents miss more frequently in my dataset. So it might be instructive to examine game footage to see if they appear to be doing something either better or differently to other Premiership sides.
Hopefully,as models become more complex with the addition of extra layers of information (multi player position is the obvious next step), we can begin to reveal more accurate descriptions of the actions that go towards making teams successful, especially on the defensive side of the ball. Or at the very least we may know what we are looking for and which grounds we are likely to find it.
For other shot models, check out Different Game's wonderfully named SPAM model, a massive and on going project and the recent shot mapping from 11tegen11.