Saturday, November 5, 2016

Being there or being helpful

OK, so, I was listening to the Arsecast, and the assist / real assist thing is back. An edition or so ago, Andrew Mangan was talking about when a player gets an assist for playing the last pass to the scorer, but the scorer does all the work so the assist didn't really play a fundamental part in the goal. He asked for thoughts on how to distinguish 'real assists'. Apparently he's still waiting. It's the weekend, and if I don't find something to do I have to sweep the flat. So.

Obviously the problem with distinguishing assists / real assists is that it involves a qualitative judgment on the last pass to the scorer. Also obviously, there are quality metrics in statistical analysis as well as quantity metrics, but the one tends to be built off the other, eg goals and shots = quantity, conversion rate = quality. Assigning 'real' status to an assist would seem to be a divergence from that, a subjective approach.

(As a side note, and further to my piece on French football terminology for Unusual Efforts last month, in French an assist is a 'passe decisive'*, which sounds a bit better, but is still in statistical terms, the last pass to the scorer. There's a faux ami aspect to this as assister actually means 'to be present at' rather than 'to help with', so on that basis, Mohamed Elneny could definitely be said to have assisté that Ozil goal...Anyway.) 

So, is there a way to objectively rank assists to distinguish the 'real' ones? Paging the xG community here. I don't have shot location data, and I don't know if the peeps who do have data for the pass before the shot. Given that there are also models for expected assists, I'm assuming so. But this is slightly different - not seeking to 'expect assists', but to evaluate assists based on the xG of the goal that resulted. 

For example, if a goal has a very high xG because it is eg a tap-in from 2 yards out, then we might be reasonably confident that the assist was instrumental in the goal. If a goal has a very low xG because it is eg a speculative lob from just inside the halfway line, then the assister might be as surprised by that as the opposition goalkeeper**. 

Now, if the xG is high because the scorer dribbled round five people having received the ball to get themselves into that close position, that is undoubtedly a factor. So would it be possible to find the xG of the location where the scorer received the ball, not where they took the shot?

For example in this [cough] artist's impression of the Ozil goal, Elneny's pass put Ozil in possession in a location Y that I'm guessing has a fairly low xG rating.  

Ozil then did his glorious thang to score from a much better location X. If we use the xG value of Y, the location where the ball was received, this would seem a good way to evaluate Elneny's contribution. It was a good pass, yes, but not very, well, assisty. Using the Y value would reflect this. 

Also, where both locations are the same, the assister would get 'full credit' based on that value. So, fabulous ball in to the box from a corner, towering header from a centre-back 6 yards out, high value assist. This should also take care of the speculative lob situation as that goal would have a low xG rating so even if the scorer only took one touch, the assist would be valued at that low value to reflect that it was probably more a pass, and the assister really wasn't expecting him to do that. 

*EDIT* (I've been for a walk so been mulling things over) Assigning a simple 'per assist' value is all very well, but to gauge a player's contribution (or team's approach) over a longer period, there would seem to be two options :

  1. (value of xG as at point Y / number of assists) x number of assists = equivalent goals

    (value of xG as at point Y / number of assists) will never be 100%, as xG is never (I understand) 1, but could be used as a weighting factor to estimate impact / contribution.                    
  2. value of xG as at point Y / value of xG as at point X = % helpfulness / assistiness

    Thus, if every assist by a player was a corner on to the head of a centre-back, so all X and Y locations, and therefore values, were the same, the corner-taker would be 100% helpful / assisty in those goals. Yes, this could be problematic if an assist was to a 'better' location than the scorer ended up in, but over a larger dataset it seems likely that this would not be the case. But again, need numbers for this. 


On a vaguely related note, there are no pitch markings on these but here are some chalkboards from the 1950s. You're welcome.

'Key pass' sounds like it should help, and like passe decisive may be used in more subjective ways in commentary etc, but Opta define this as "the final pass or pass-cum-shot leading to the recipient of the ball having an attempt at goal without scoring" (so if it had been a goal, it would have been an assist) and 'chances created' is assists plus key passes so also working on the same basis. They do refer to 'second assists' which is "a pass/cross that is instrumental in creating a goal-scoring opportunity, for example a corner or free-kick to a player who then assists an attempt, a chance-creating through ball or cross into a dangerous position." But then we're back to defining 'instrumental'.

** And in this case the assister might actually be the other goalkeeper. This is one of the weirdnesses of stats that can arise for time to time. Remember when Tim Howard scored that goal? That clearly wasn't a shot. But it was goal, so therefore had to be 'a shot'. Otherwise the stats don't work - if I type 1/0 into my calculator, it gets very cross with me, and googledocs verges on the rude, calling me a #DIV/0!. Similarly here, if it's a goal, the pass to the scorer has to be 'an assist', regardless of intent. This is where a quality metric is needed.