Wednesday, October 31, 2018

AFL Statistics Series #1: Which Statistics Matter The Most (Apart From The Scoreboard)?

Introduction

This is a first in a series of posts that I’ll do about statistics in the Australian Football League. The posts will be about which AFL statistics I think matter – that is, what I think they tell us about how AFL teams and players go about scoring and stopping the other team from scoring.

Yes, there is a lot of writing out there already about AFL that uses statistics and numbers, and a lot of good writing. This is just how all of those statistics make sense to me. I hope if you’re reading this you find something in here that’s useful for you too.

A lot of the thoughts I’m going to talk about here came about as a result of me trying to devise a method of rating AFL players, without having access to the detailed data that Champion Data use to devise their ratings. We may still get to that in the end. It turns out though that to work out how each player contributes to winning a game you need to first think about how teams go about winning them.

Points differential: the most important statistic of all (duh…)

In their AFL Prospectus 2018, Champion Data made this obvious but important point:

“… we are asked [:] What’s the most important stat? As respectfully as possible we answer with POINTS FOR. It’s the one stat that guarantees a win … We go on to explain it’s more about how you get to that point.”

Of course points for compared with points against is important. There is a position here though that may not be quite so obvious. Some would argue that only the win or the loss matter, and not the margin of victory or defeat. Margins though tend to be a better predictor of future performance. Close wins may bring exhilaration and relief, but in general a team should take more comfort out of a comfortable win than a close one.

Metres gained matters

Metres gained gets some bad press, perhaps because it sounds like ‘a stat too far’. An article on The Roar last year even went so far as to call the statistic ‘irrelevant’. The main argument was that it doesn’t take into account the outcome of the possession – a long kick to the opposition would be credited with many more metres than an effective handball backwards to a teammate. “While it is impressive to see a player average over 300 metres gained a match,” the author says “the statistic is mostly empty in its meaning.”

The article makes some good points, but I disagree that metres gained is irrelevant. Indeed to me, there is hardly anything more fundamental to performing well at Australian rules football than gaining metres. When you’re watching your team, apart from when they’re actually in the action of kicking goals, what do you most want them to do? You want them to GET THE BALL CLOSER TO THEIR GOALPOSTS and GET THE BALL FAR AWAY FROM THE OTHER TEAM’S GOALPOSTS!

Now a critic of metres gained may point out that it isn’t so great if you kick 50 metres straight to the opposition. That’s true – ‘effective’ metres would probably be a better measure. We’ll get to more about keeping possession later.

Kicking the ball 50 metres to the opposition though isn’t necessarily a horrible outcome, despite what I will say later on about the value of turnovers. Now if the opposition run the ball down the field and score a goal off your turnover that is obviously a bad result. That worst-case scenario is relatively uncommon though – only about 10 per cent of possession ‘chains’ end in goals, and on average those ‘chains’ only last for about three disposals and gain about 45-50 metres. In other words, even if you kick 50 metres to the opposition it’s unlikely the other team will punish you by running the ball down the field and kicking a goal (obviously depending on where the ball is, and how badly you butcher the kick). More likely is that the ball may come back to your team within a few possessions, and back around where you started.

What about an effective kick across the ground that gains no metres but sets the team up for a shot at goal? Isn’t it true that metres gained isn’t very good at accounting for that? That cross-ground kick however is only valuable if the TEAM gains metres on a subsequent possession. If the opposition stops the ball before it goes any further then it’s just a kick across the ground that didn’t help much. The objection to metres gained here is more about crediting the total metres gained by the team to the individuals in that team – how much did that cross-ground kick help the team to score? – not about the value of the total metres gained by the team itself.

Metres gained matter. In each of the four AFL seasons since the statistic was made public in 2015 no other relatively common used statistic – except statistics directly related to scoring, e.g. score involvements and goal assists – has been more positively correlated with points differential (see table below).



Good teams like West Coast and Richmond had less disposals last year than their opponents, and lower teams like the Bulldogs and St. Kilda had more. Richmond were smashed in hit outs and clearances, and West Coast were behind on tackles. The higher teams though almost always had positive metres gained differentials over the season (see table below). It is one of the few relatively common statistics that you can reliably count on good teams being ahead in.  

[EDIT: Metres gained differential in a single match should, by definition, ALMOST reflect goal differential. Nevertheless, metres gained are still highly correlated with scoring.]    

Once you view gaining territory as an important indicator of a football team’s ability to score goals, the importance of some other AFL statistics falls into place. Inside 50 entries – a statistic that has been noted by others to be highly correlated with winning – indicates metres gained by measuring the number of times a team gets the ball past a particular point on the ground, one it has to pass over in order to score goals. An inside 40 or inside 30 measure would also indicate this.
(Rebound 50s indicate metres gained as well, but since they indicate metres gained in a team’s defensive part of the ground they are negatively correlated with winning. Put inside 50s and rebound 50s together and you get some of the way to a decent proxy for metres gained.) 
It also indicates why the number of kicks that a team or player records is generally important for winning (other than, of course, it is the only way to score goals), and why one kick is generally more important than one handball – more kicks often leads to more territory gained. Conversely though it also explains why just amassing kicks sometimes does not lead to success; for example, two short kicks of 20 metres get a team no closer to goal than one long kick of 40 metres.
All of this is not to say however that an individual player’s contribution to their team’s performance should be primarily measured by metres gained. The AFL leaders in metres gained per game last year were Jayden Short, James Sicily, and Nathan Wilson. Nobody thinks that these are the very best players in the league. (Maybe they should … but probably not.) That is because what these players don’t do as much as some other players is win their teams the ball in the first place.
Turnovers are the main source of scoring shots
In Australian rules football, the significance of individual possessions in helping a team score can sometimes seem hard to work out. The ball can pass back and forth between teams several times before anyone has a legitimate chance to score. Teams can also have very different styles when in possession of the ball, with some teams preferring a ‘high-possession, low-risk’ game, and others preferring to be more direct.
Let’s try and simplify it then. Obviously if you have possession of the ball you are the only team that can score until your ‘chain’ of possessions is broken. We’ll call any unbroken sequence of possessions by a team – whether of one possession or ten – a ‘possession chain’.
Let’s say that a possession chain for a team can start in one of three ways:
  • the team gets a ‘clearance’ – i.e. it clears the ball from a ball-up, either from a stoppage or a centre bounce at the start of a quarter or after a goal
  • the opposition turns the ball over; or
  • the opposition scores a behind, giving the team a kick into play.
Let’s also say that a possession chain ends in one of three ways:

  • the team scores a goal or behind – a successful (or at least partially successful) possession chain;
  • the team turns the ball over; or
  • the possession chain is ‘stopped’, due to a ball-up, or because the quarter is over, or because the scoreboard is on fire – basically any unsuccessful possession chain that does not result in the ball going directly back to the opposition.
(That may not be completely technically correct according to how Champion Data defines these terms, but I think it’s close enough for the purpose of my main point here.)

Therefore, for a team we will say that:

Possession chains started = Possession chains ended

Clearances + Opposition turnovers + Opposition behinds = Scoring Shots + Turnovers + Stopped Possession Chains

Or, more importantly:

Scoring Shots = Clearances + Opposition turnovers + Opposition behinds - Turnovers - Stopped Possession Chains

For example in 2018 premiers West Coast averaged 24.7 scoring shots per game. By the definition above they started 117.9 possession chains per game, from 36.6 clearances, 71.6 opposition turnovers, and 9.7 opposition behinds. Of the 93.2 unsuccessful possession chains they had per game, 67.7 of them were turnovers, and 25.4 were stopped possession chains.

The bottom team in 2018 Carlton averaged 17.9 scoring shots per game. By the definition above they started 109.7 possession chains per game, from 35.0 clearances, 62.3 opposition turnovers, and 12.4 opposition behinds. Of the 91.9 unsuccessful possession chains they had per game, 67.3 of them were turnovers, and 24.6 were stopped possession chains.

What’s the main difference in the possession chains of those two teams? Clearances, behinds, and stopped possession chains are all similar. The main difference is West Coast started more chains through opposition turnovers – about the same difference as the difference in scoring shots.

Returning to our scoring shot formula above, let’s now look at scoring shot differentials, or scoring shots for the team less opposition scoring shots. With a bit of mathematics we can show that:

Scoring shot differential = Clearance differential – 2 * Turnover differential – Behinds differential – Stopped possession chains differential

Maybe it’s just me, but I found this really interesting when I worked it out. Turnovers are not only more common than those other components they count for twice as much in this equation. When there is a stoppage one team’s possession chain ends, but then each team has about a 50 per cent chance of starting the next possession chain. When the ball is turned over, one team’s possession chain ends and the other team’s begins.

The importance of turnover differential can be seen when we compare how teams got their scoring shots in 2018 (see table below). Minor premiers Richmond were last by some margin in clearance differential, but they were way ahead in terms of (inverse) turnover differential. West Coast and Melbourne also rated highly in either causing opposition turnovers, or not turning the ball over themselves. Unsurprisingly, bottom teams Carlton and Gold Coast rated poorly in terms of turnover differential.

Turnover differential does not explain everything, as there are other ways to start possession chains. 2018 runners-up Collingwood gave up a lot of turnovers, but they were good at getting clearances. 2016 premiers the Western Bulldogs were fantastic at winning clearances. It is just that it is less likely to get a high differential through clearances or stoppages rather than turnovers as there are less of them. Hence, clearance differential is less correlated with winning than turnover differential is.
In terms of valuing players this suggests that not only are players with high clearance numbers such as Tom Mitchell and Nat Fyfe important for starting possession chains, but so are defenders who get a high number of intercepts such as Alex Rance and Jeremy McGovern. Though the question then is how important is the individual player who records the clearance or intercept to the team getting possession? How much of the credit should go to the ruckman getting the hit out, or the structure of the team defence?
Given how fundamental I said metres gained were how then does turnover differential relate back to that measure? If the main aim of Australian rules is to score by getting the ball close enough to your goal to do so, then the way to progress the ball down the ground is to get possession of the ball and keepisng it. The best possessions are those that help the team to gain a high number of metres with a relatively low risk of giving the ball back to the opposition, or of the ball being ‘stopped’. On defence, your aim is to stop the other team doing this.
Sounds simple, right? Well, it’s easier said than done. Also, what may be complex about Australian rules is the many ways you can go about doing this.
Goal accuracy: converting those shots at goal
Another statistic that I think is important is goal accuracy, though not quite in the same way as the other main statistics I have covered above.
Goal accuracy – the percentage of shots a team has on goal that are goals (scoring shots, if kicks out of bounds are not available) – is somewhat important to performance over an entire season. Last season, finals teams Sydney, Hawthorn, Collingwood and West Coast were good at it, but Richmond and GWS were not. In 2017 the leaders in goal accuracy were non-finalists Melbourne and the Brisbane Lions. Scoring shot creation is generally more important to performance over a season than scoring shot conversion is, with teams being reasonably close over a season in terms of the rates at which they convert.
I would say where it matters more is changing the outcome of a single game, which is even more important if that game is a final. In a single game there is more variability in scoring shot conversion than over a season, and the results can swing the match significantly. We see this in some of the comparisons between final scores and ‘expected scores’ for a game.
To state the obvious, six behinds are needed to gain as many points as a single goal. Given that the difference between premiers West Coast and last-placed Carlton was only about seven scoring shots per game you want to convert those chances when you get them. Or given that goal accuracy tends to even out over the season, you want to convert those chances when they are most important. (West Coast almost didn’t in last year’s Grand Final, until Dom Sheed did.)
Conclusion
In Australian rules football the ways for teams to score are:
  •  get possession of the ball, and keep it;
  • gain metres while you have the ball in order to get a shot at scoring; and
  • get as many of those shots at scoring through the goal posts as possible.

For defence the ways to stop the other team from scoring are:
  • try to get back possession of the ball, or otherwise get a stoppage;
  • if you haven’t got the ball back yet stop the other team from getting the ball down the field far enough to have a shot at scoring; and
  •  if the other team does get a shot a scoring, try and limit the chance that it is a goal.
This is essentially how I am going to talk about the value of AFL statistics in this series. I’ll mainly evaluate teams and players by how good they are at doing these things. I also plan to talk about how the evolving nature of statistics has given us a better picture over time of the effectiveness of teams and players. Some conclusions will be obvious, but some may be less so. In the end, we may yet even get to that 'player rating' system.