This
is another main entry in my series on AFL statistics. In this one, I am going
to look at how AFL Player Ratings relate to players’ statistics. This includes the
more common statistics for actions such as kicks, marks, and handballs, but
also the more recent statistics capturing the outcomes of those actions such as
inside 50s and turnovers.
One aim is to get a sense of how much these statistics
on average are valued within the AFL Player Ratings system (beyond what is
already published), and therefore – to the extent that those Ratings try to
capture a player’s contribution to his team’s overall net result – how important
each of those statistics might be to a team’s winning or losing. But another
aim is to also get a sense of how much further information we have got over
time about a player’s contribution as more detailed statistics have been
developed.
A big thanks here goes to Wheelo Ratings and its extensive collection of AFL player
statistics. Most crucially it has each player’s average AFL Player Rating for
the season, but also heaps of other measures without which the following analysis
would not be possible.
What are the AFL Player
Ratings again?
The AFL Player Ratings measure a player’s performance through ‘equity
ratings’. These equity ratings place a value upon a player’s action – whether
positive, negative, or zero – based on how much that action has improved the
probability of his team being in a position to be the next to score.
For example:
· When a
player takes possession of the ball in the centre of the ground, his team is then
somewhat more likely to score either a goal (6 points) or behind (1 point) than
their opponents, and the Average Equity is 1.58.
· When a
player takes an uncontested mark 30 metres from goal his team is much more
likely to score than their opponents – and also more likely to score a goal
than a behind – making the Average Equity 4.33.
· Therefore,
the value created by a player kicking the ball from the centre of the ground to
a teammate taking an uncontested mark 30 metres from goal is 2.75. (This change
in value may be attributed entirely to the kicker or shared depending on
whether the receiving teammate took the mark ‘on the lead’.)
Further details are in Dr Karl Jackson’s thesis “Assessing Player Performance in Australian Football
Using Spatial Data”.
When watching a game, many Australian football
observers would agree that not all possessions and disposals are created equal.
They depend on where the ball is and whether it was won in a contest, along
with how far it travelled and to which team the disposal went to. The ratings
system was an attempt to quantify, through many points of data, how much each
player was expected to be improving the scoring potential of his team.
Actions that are highly valued under this system are
contested possessions and effective kicks, i.e. actions that ‘win’ the ball for
the team and effectively progress it towards goal. Conversely, ‘clangers’ or
unforced errors have a high negative value. Actions that gain only small
amounts of ground for the team such as handballs and lateral kicks are more
neutral. Successful shots at goal can add several points to a team’s expected
score and player’s rating, particularly if they are kicked from an unlikely
scoring location (the highest ever Player Rating for a game was by Lance
Franklin when he kicked 13 goals).
While behinds – in contrast to many other rating systems – can actually reduce a player’s estimated value, particularly if kicked from spots where players
typically do not miss.
Part 1: Approximating the
Ratings using ‘basic’ statistics
First up, we are going to
look at approximating AFL Player Ratings using the oldest and most ‘basic’ statistics
for player actions. For the majority of Australian football history, the main statistic
we have is just goals scored. Then from the mid-1960s onwards we have the basic counting
statistics of actions a player took. These
include kicks, marks, handballs, hit outs (sporadically), free kicks for and
against, along with goals and behinds. Tackles we have from the late-1980s.
For many football followers just glancing over the
‘box scores’, these statistics still to a significant extent inform their view
of how well a player has performed in a match. They may weight these things
slightly differently in their heads – whether for example a player with 30
disposals had as much impact as a player with 17 disposals and 2 goals – but
many will have their own sense of how ‘valuable’ each of these basic counting
statistics are. But what if we regress them against the AFL Player Ratings?
Below is a regression of the average AFL Player Rating
for each player in 2022 against his average kicks, marks, handballs, hit outs,
free kick differential, tackles, goals, and behinds. (No intercept was
included.)
Goals as expected have the most value, with a player’s
average rating increasing by 2.8 points for each one goal increase in his
average goals. However the average amount of points per scoring shot is a bit under
four points, so given a goal is worth six points (some kicks out of bounds
aside) this is probably in part capturing some other valuable actions undertaken
in the forward line such as getting the ball inside 50, goal assists, and contested
marks. The full effect of scoring a goal in this regression is actually a bit
more again, because the player will also record a kick. I tried separating out scoring
shots from other kicks, but the estimated value of the goal remained about the
same – i.e. the coefficient for goals increased by about the amount of the
coefficient for kicks (0.4-0.5).
Behinds on average decrease a player’s average rating by 1.9
point. As I understand it the negative impact of behinds in the Player Ratings is
on average lower in magnitude than the positive impact of goals as a team
remains reasonably likely to score next after a behind since the ball is still
close to their goal.
The coefficient for kicks is only slightly more than that of handballs in this regression, even though kicks on average gain
substantially more territory. Players that record high amounts of handballs are
more likely to be involved in stoppages and contests – think of players like
Clayton Oliver, Lachie Neale, Patrick Cripps, and Rory Laird – and therefore
they tend to record more contested possessions which are valued highly in the
Player Ratings system. Before contested possessions were recorded handballs
were possibly our strongest indicator of which players, like Greg Williams and
Terry Wallace, were winning the ball ‘in the coalface’. Once contested
possessions enter the equation – see next section – then the coefficients for
kicks and handballs become more reflective of purely the territory they gain.
As for marks,
most marks are uncontested – i.e. ‘chipping’ the ball around – and this type of
action is not highly valued in the Player Ratings system. Their value in the
basic statistics regression probably captures the valuable things that defensive
and ‘outside’ players do such as kicking long and/or efficiently and
intercepting the ball, and as we shall see they drop out of the regressions as
we add more ‘advanced’ statistics.
Free kicks enter the regression as the differential between free
kicks for and against. The coefficient for free kicks against was close to zero
in other versions of this regression, probably because players that give away
more free kicks are those that tend to be involved in more contests and put
more pressure on the opposition.
We can multiply a player’s average for each of these
statistics by the coefficients to get an estimated average rating for each
player. Compared with the Player Ratings, the players that have a high
percentage of possessions that are contested (e.g. Tom Liberatore) tend to have
a lower estimated average than their actual average Player Rating, while players
that have a lot of kicks but a high percentage of possessions that re uncontested
such as Jake Lloyd or Andrew Gaff have higher estimated averages. The lack of
defensive statistics outside of tackles also means that spoiling and
intercepting key defenders such as Darcy Moore, Harris Andrews, Jeremy McGovern
are ‘undervalued’. All of this suggests that the ‘basic’ statistics we had from
the 1960s to the 1980s left out some important information in assessing a
player’s contribution.
Part 2: Approximating
the Ratings using ‘intermediate’ statistics
Next we are going to include some of the
‘intermediate’ statistics that became available from the 1990s.
This included the famous distinction between contested and uncontested
possessions, which told us which teams and players were winning more of the
‘hard’ ball when possession was in dispute, with contested possessions further
broken down into clearances (first effective disposal out of stoppages and
centre bounces) and contested marks. We also got an indication of which players
were costing their team possession through ‘clangers’ (giving the ball directly
to the opposition, along with conceding free kicks). Inside 50s and rebound 50s
told us a bit more about how far and where on the ground a player was moving
the ball, while goal assists told us more about who was involved in scoring a
goal for his team. Crucially, the ‘one percenters’ statistic, which captures
spoils and smothers, finally gave us a lot more information with which to value
the contribution of defenders.
These statistics are still used a lot by fans and commentators,
as they are relatively simple and well-understood and do more to gauge the
‘impact’ of a player than just looking at how many times they got the ball.
Many fans recognize that players can rack up a lot of ‘easy’ possessions but still
not have much of an influence upon the result, or dispose of the ball poorly.
Below is a regression of the average AFL Player Rating
for each player in 2022 against selected ‘basic’ and ‘intermediate’ statistics.
The coefficients of kicks and handballs are
reduced with the inclusion of statistics such as inside and rebound 50s that
can now capture a fair chunk regarding a player’s ability to progress the ball,
and contested possessions which capture a player’s ability to ‘win’ possession.
Kicks now have a coefficient of between three and four more times that of
handballs, which is probably more reflective of the relative territory they
gain.
Contested possessions are valued highly in the Player Ratings system and have
a relatively high coefficient, with their inclusion reducing the coefficients
for handballs and tackles. I tried separating out clearances, contested marks,
and other contested possessions (ground ball gets and knock-ons), but they all
had about the same coefficient, so I have just included all contested
possessions outside of free kicks together.
An uncontested
possession (infamously) has a value of zero if the AFL Player Ratings
system. Uncontested possessions and marks were dropped from the final regression,
with versions that did include them having small negative coefficients.
Clangers are one of the outcomes that are most hurtful to a
player’s rating, since the player has ended their team’s chain of possession
and started a chain of possession for their opposition. A player’s average
rating drops by about one point for each one point increase in average (non-free
kick) clangers. This is an even greater impact in magnitude than contested
possessions, where the ball is still in dispute rather than in the team’s
possession.
An inside 50
is on average worth more than a rebound
50 under the AFL Player Rating system. For the same action that moves the
ball forward, the gain in ‘equity’ is higher if the ball is in the team’s
forward half compared with it being in their back half. In other words the
marginal increase in probability of the team scoring next from moving the ball
from (for example) 140 to 110 metres out from goal is lower than from moving it
from 60 to 30 metres out from goal.
Goal assists have a coefficient of one, which I was slightly
surprised by, but I guess captures that the Player Ratings system highly
rewards effective disposals up forward.
As expected, the inclusion of one percenters helps to better capture the contribution of
defenders. As we will see in the next section however the coefficient is
probably also in part capturing the value of intercept possessions.
We can again multiply a player’s average for each of
these statistics by the coefficient to get an estimated average rating for each
player, and this now actually does pretty well in matching a player’s average AFL
player rating. One type of player that is still ‘underestimated’ compared with
the AFL Player Ratings is the ‘striker’ midfielder-forward type that wins the
hard ball and is a penetrating kick into the forward line. This includes
players like Jordan de Goey and Shai Bolton this season, and Dustin Martin and
Jake Stringer in previous seasons, and also to some extent attacking
midfielders like Marcus Bontempelli.
Part 3: Approximating
the Ratings using ‘advanced’ statistics
Finally, we will include the more detailed statistics
that have only
become publically available in recent years. This includes statistics like
metres gained, intercepts and turnovers, score involvements, hit outs to
advantage, and pressure acts.
These types of statistics provide useful additional
information about a player’s impact (e.g. how often a ruck actually hits the
ball out to a teammate rather than just hits it), but they are generally not
available that widely, or for a long historical period. It will be interesting
to see then how much information they do seem to add in explaining a player’s
contribution over the more familiar ‘1990s’ statistics.
Below is a regression of the average AFL Player Rating
for each player in 2022 against a selected range of ‘full’ statistics.
Note that some of these new statistics will in full or
part encompass older statistics – turnovers include some clangers, score
involvements include goals, behinds, and goal assists, and pressure acts
include tackles. However, as with goals and kicks I found the overall values of
scores, tackles, and clangers were not much different whether they were
separated out or not – e.g. the coefficient of tackles increased if they were
separated out from other pressure acts.
Average metres
gained now essentially takes over from kicks
and handballs, and even inside and
rebound 50s, as the most ‘useful’ measure of a player’s ability to gain
territory. Inside and rebound 50s still have value in terms of indicating where
the player is on the field, and therefore how much they are increasing the team’s
probability of scoring when disposing of the ball (although the coefficient for
rebound 50s is now closer to zero).
Hit outs to advantage has displaced ‘raw’ hit outs in the final regression
as the more significant explanation of a player’s average rating. However as
hit out to advantage percentages do not vary hugely across players these are
highly correlated, and ‘raw’ hit outs would still tell us a fair bit about a
ruck’s ability to win the ball for his team.
Turnovers have a negative coefficient that is lower in
magnitude than that of clangers. Many
turnovers occur when a player kicks the ball to a contest. These may often
be considered ‘ineffective’ kicks, but they still before the contest was decided
gave the team a chance of retaining the ball, and are therefore on average less
hurtful than straight out giving the ball to the opponent.
The inclusion of intercepts
has slightly reduced the coefficient on one percenters, with the main
contributions of defenders now being ‘split’ across the two variables.
Once again we multiply a player’s average for each of
these statistics by the coefficient to get an estimated average rating for each
player. As it turns out, this actually leads to fairly similar results to those
from the ‘intermediate’ statistics. That is, the same types of players – such as
‘striker’ midfielder-forwards – tend to remain ‘undervalued’ as was the case
under the ‘intermediate’ statistics regression.
This suggests that the main advance in capturing the types
of contributions a player makes was when we went from ‘basic’ to ‘intermediate’
statistics in the 1990s, while the jump from ‘intermediate’ to ‘advanced’
statistics is actually more a delineation of these contributions. For example, metres
gained is possibly just a more detailed, refined measure of who is kicking the
ball long that was already largely captured by inside 50s and rebound 50s, while
tackles already captured to a fair extent a player’s overall number of pressure
acts, and one percenters a defender’s ability to stop or intercept the
opposition’s attack. This could be why a player rating system like the PAV system, which is largely built
on those ‘1990s statistics’, is able to approximate the AFL Player Ratings
reasonably well.
That for the most part
concludes this series of posts. If/when I return with another post it will
hopefully be to finally put forward a ‘player rating’ system of my own.
