Pages

Wednesday 21 November 2018

Non Shot Expected Goals Plus/Minus.

I last wrote about football's major problems when attempting to replicate the so called plus/minus stat from other, mainly US based sports here

In truth I only wrote it to work a spurious photo of Harry Nilsson into a football blog, but the objections raised were valid.

The main hurdle is the lack of goals, not a problem in sports where +/- exist (in mainly forms) such as basketball.

Five years on we may have a partial solution to lack of goals.

Goal attempts, expressed in shot based xG does improve the performance related sample size, but non shot expected goals models, applied to either individual actions or possessions, opens up a much richer source of data.

The problems of comparing changing lineups, the duration they are in opposition, the non shot xG differential and the venue, are purely technical and mathematical, as well as being difficult to compute on an industrial scale.

But non shot xG may have removed the final obstacle to creating credible +/- numbers for football, for which every defender and defensive midfielder should be hugely grateful.

If and when I finally data dump everything into a huge matrix & run off some figures, they'll appear here.





Tuesday 13 November 2018

Events per Possession.

Different models have different basic events that define them.

Shot based expected goals models have all goal attempts as their baseline unit.

XG2 models to rate keepers have attempts on target.

And non shot xG models are framed around possessions.

For the latter we'd like to know more about an individual possession. we can measure the speed, (see previous primer), but we'd also like to know what's happening during a possession.

Usually that's passes and runs with the ball that are designed to probe defenses and create space.

We therefore look at how many significant events occur per possession.

If there's lots of passes and lots of runs, you might be chasing shadows.

If there's just mainly passing, you're probably better able to keep a good defensive shape.


Speed of Attacks

Measuring speed is easy, it's just distance travelled/time. Right?

But that's not really the correct approach to take on a football pitch.

Making 30 yards running from deep in 5 seconds probably isn't the equivalent of making the same, but with the halfway line as the starting point.

In Non Shot expected goals terms you're going from a very small threat to a slightly greater threat in scenario one and from a moderate threat to a really dangerous threat in the second case.

So instead of using distance made/time to measure the speed of an attack, we measure Non shot xG gained/time to define the speed of an attack in football.

It even comes out as a sensible double digit number if you convert it to NSxG per hour!


Saturday 10 November 2018

Non Shot Passing Ratings


Most people now accept that passing percentage is a useless stat and it was natural to look at the difficulty of the pass attempted and compare actual success rate to the expected one.

However, this approach still has drawbacks.

Passing is a risk/reward action.

A player risks the value of the field position from where he makes the pass for the reward of completing the pass against the loss of handing possession to the opposition.

We can measure field position in non-shot terms, (broadly, the likelihood that a goal results from possession at a point on the field).

This rather important aspect of a pass should be included in any assessment of passing ability.

(Consider this contrived example. Player A attempts a relatively difficult pass.

Let’s say it’s completed on average 70% of the time.

It’s deep in his own half, so it isn’t particularly valuable to his side. We’ll say the pass is “worth” 0.02 NS xG.

If it is picked off, the opposition get healthy field position worth 0.1 NS xG.

On average, an average player therefore gains a cumulative 0.14 NS xG on the 7 out of ten occasions that the pass is successful…and loses 0.3 NS xG on the three times it isn’t.

This is long term, a potentially very poor choice of pass for an average player.

Now if a very good passing player completes the pass eight times out of ten, rather than just seven, he’s going to appear to be well above average on a model based just on pass completion & difficulty of the pass.

However, the choice of pass is still poor, long term. Now he gains 0.16 NS xG for his eight completions and still loses 0.2 NS xG for the two failures).

We should when looking at passes examine the player’s choice of risk/reward of the pass, the difficulty of the pass and compare those factors to the actual outcome.

Here’s 2018/19 so far for attacking players

Most players make a large percentage of passes that benefit their team long term. The combination of likely success rate, gain in field position and potential loss of non-shot xG is in their favour.

Nearly 99% of Sane’s passes had a positive expectation, 95.5% of Ozil’s. Perhaps Ozil is being a bit more adventurous?

When we examine the outcome of all Ozil’s 2018/19 attempts, he’s gained 2.43 NS xG per 100 passes attempted.

Sterling tops that particular column with 3 NS xG per 100.

If we’d put Ozil’s attempted passes at the feet of our average model, the gain would have been only 2.17 NS xG per 100.

Arsenal’s player maker has overperformed by 12%.

By comparison, Moura has underperformed his expectation from his choice of passes by nearly 16%.

Another Brazilian who you would anticipate progressing as he reaches his prime is Richarlison. Nearly 8% of his passes currently have a negative expectation, but he’s still got a pass expectation of 2.4 NSxG per 100.

His actual is only 1.8 NSxG per 100. That leaves plenty of room for his choices to perhaps improve slightly and his execution to at least approach the average for a Premier League passer. So decent, if below average passing production, but loads of upside.
All the data is from Infogol