WheeloRatings

@crow_data_sci posted some interesting stats on Twitter regarding the use of the medi-sub. Port Adelaide has had the most players subbed off and play the next week in both of the last two years (5 in 2021, 4 in 2022). Melbourne and Sydney have had no player subbed off and play the next week at all across the last two seasons. Source:

I have a model that rates teams based on past results and the current team ratings are here: https://www.wheeloratings.com/afl_ratings.html The model estimates match win probabilities based on the current team ratings and venue advantage. For example, my model rates Melbourne a 48.5% chance of winning this week. Each week, I simulate the remainder of the season 50,000 times based on those win probabilities. So roughly 48.5% of the simulations will result in a Melbourne win against Fremantle and 51.5% will result in a Fremantle win (although a proportion of matches will end in a draw too). I do simulate match margins too based on a normal distribution with a mean of the expected margin so it does allow a team with easier matches to increase their percentage, although Sydney does have some ground to make up. My model doesn't rate Collingwood that highly given they've had so many wins by small margins, so they could potentially outperform my model's expectations and therefore somewhat reduce Melbourne's chances of finishing top 4 in the two win scenario.

I agree they shouldn't be taken lightly and I won't be confident at all when we play them. They are 8-1 in games decided by 12 points or less, which is an amazing record (don't get me wrong). They have done very well to win all those close games but there's an element of luck in winning close games. They gave away a free kick to Essendon in front of goal inside the last minute. If they kick that goal, Collingwood lose and they're 7th. I don't want to take anything away from them subsequently going coast-to-coast and Elliott kicking an incredible goal after the siren, but they weren't in control of the Essendon shot at goal. I'm not saying they're not a good side, I just think that 13-5 and 4th flatters them a little bit considering their % is only 106.

Yes, realistically we should start favourites against Collingwood and Carlton as well. Collingwood has won a lot of close games so their ladder position probably flatters them somewhat.

Based on my simulations (50,000 simulations of the remainder of the season), this is what I get for Melbourne's probability of finishing in each ladder position for a certain number of wins over the last four rounds.

In terms of Goals ÷ Shots, which is Champion Data's Accuracy metric, Melbourne's actually the third best in terms of opposition accuracy. This metric takes into account shots at goal that don't score at all but won't include rushed behinds that were not as a result of a missed shot. Accuracy doesn't take into account shot difficulty (location, pressure) which is where expected score can be useful. I'm not at my computer at the moment so can't delve deeper.

I agree that extrapolation of a player's rating (or any stat) based on game time is somewhat meaningless. However, taking game time into account in aggregating player averages across a season can be useful particularly for medi-subs and players injured during a match who may only be on the ground for a very short time or not at all. Instead of extrapolating a particular stat, a weighted average by game time could be calculated for a particular stat across the season. Salem, for example, only played 13% game time in round 1 which significantly reduces his averages. His average fantasy points is 68.9 across the season but 79.8 excluding round 1. His weighted average based on game time is 78.2 which is probably a better reflection of his average output.

It won't be as it's in Hobart on the 29th and 30th (so good for me).

I have the full season stats on my website: https://www.wheeloratings.com/afl_stats.html You can filter by team under "Player Filters", but here are the average game time stats for all Melbourne's players this year:

I guess that depends on whether you assign different scores/weights to different stats. E.g. is a goal worth more than a score involvement? If so, you could adjust the weight that you assign to a goal to account for the fact the player will also be credited a score involvement. Hypothetical Example If a goal was worth 8 points, a behind 1 point, and a score involvement 0.5 points, then effectively a goal is actually worth 8.5 points, a behind 1.5 points and other score involvements worth 0.5 points. If you subtracted the 0.5 points from the value of a goal and behind and assign 7.5 points for a goal and 0.5 points for a behind, then a goal would effectively be worth 8 points and a behind 1 point. Another way to achieve the same result would be to subtract a player's goals and behinds from their score involvements. So a player scoring 2 goals 1 behind with 8 total score involvements, had 5 "non-scoring" score involvements for want of a better term. But I wouldn't necessarily suggest you remove score involvements altogether as they are an informative stat.

I agree with your interpretation of why they have the "score assist" stat, in addition to the goal assist. I guess I can understand the contested mark / free kick scenario, in that (a) both relate to the way the player took possession of the ball so it was either a mark or a free kick, and (b) the umpire would pay the free kick instead of a mark, which may be because the advantage rule remains an option, as there's no advantage from a mark. I don't mind the hit-out to advantage not being paid when a free kick was paid in the ruck contest too. I agree it seems unfair, but the free kick was paid first so I suppose I get it.

Yes indeed, it does seem a little off! Melbourne's rating is currently 5 points higher than Brisbane but the home ground advantage is +11 points to Brisbane (rightly or wrongly). The model doesn't put too much weight on a single match result, just like it doesn't weight Melbourne's three losses too heavily.

Notwithstanding @old dee's comment that we will win them all, here are each team's chances of winning each remaining match based on my model.

Here's mine:

@Demon Dynasty @deanox I can confirm that the score involvements do include the player's own goals and behinds. The giveaway is that a player always has at least as many score involvements as own scores. I guess the 'score involvements' metric is meant to stand on its own, as opposed to supplementing the goals/behinds.

@rpfc @Lucifers Hero yeah The FMI has a momentum factor in their model which is why Melbourne is expected to drop and Gold Coast is 4th favourite to make the grand final!

FWIW, my latest simulations show that there's a ~71% chance of making the top 4 with 15 wins and a ~75% chance of making the top 8 with 13 wins, and would quite possibly be reliant on other teams' results (and percentage). Hopefully we can get to 16+ wins to be safer.

Elite in this context is the top 10% of players in each position (broad position categories) based on the average AFL Player Ratings in 2022. Champion Data shoot themselves in the foot (or the media does it for them) by releasing the players categorised as elite without detail or context. The following document provides an overview of how the AFL Player Rating system works, but it's effectively a measure of how much a player improves their team's scoring chances from their involvement. https://s.afl.com.au/staticfile/AFL Tenant/AFL/PlayerRatings/PlayerRatings_HOW.pdf I haven't listened to it yet, but this week's "ESPN Footy Podcast" explains the elite ratings. For a lot more technical detail, read from chapter 5 in the following thesis: https://researchbank.swinburne.edu.au/file/248ec147-72d7-448c-a19d-49f01d90b12f/1/Karl Jackson Thesis.pdf The club leaders, which includes their position, can be seen here (noting a minimum of 9 games): I have all the Player Ratings on my site too, but the player positions are a bit different: https://www.wheeloratings.com/afl_stats.html

The final ladder in both 2012 and 2018 had 12 teams finish with a percentage of at least 100. Also: 2013 had 12 teams with a percentage of at least 100 after 21 rounds. 2015 had 12 teams with a percentage of at least 100 after 20 rounds. 2020 had 12 teams with a percentage of at least 100 after 8 rounds. The final ladder in 1914 had 8 of 10 teams with a percentage of at least 100. https://en.wikipedia.org/wiki/1914_VFL_season#Ladder

May had 3 possessions, but 6 disposals including kick ins.

The AFL's API is available again so here is Richard Little's analysis of this week's game/round.

I'm not sure I'd go as far as saying there's a real problem. Geelong has had more scoring shots than Carlton and conceded fewer scoring shots over their 10 games, and Carlton has only won 50% of quarters. If Carlton's games against Hawthorn and Port Adelaide go much longer, they quite possibly lose both and they're in 8th. They've done very well but their ladder position probably flatters them a little at this stage. I provided some additional explanation in my other post:

A lot of the models will take into account margin of victory (as well as opposition and venue), not just simply win/loss. Carlton's percentage is the worst of the top 8 teams. They beat Hawthorn by 1 point, Port Adelaide by 3 points (with fewer scoring shots) and the Western Bulldogs by 12 but also with fewer scoring shots. They're all worth 4 points on the ladder but most of the models probably won't see a 1 point win as being much better than a 1 point loss. Luck is a factor in close games and if they lose to Port and Hawthorn, they're currently 8th on the ladder. The primary purpose of the ratings in these models is to predict future matches, not simply rank how teams have performed this year, and putting too much weight on a 1 point win or on recent matches doesn't help with predicting future matches. As such, the models are somewhat conservative but if Carlton is genuinely a top 4 team this year, that will most likely be reflected in the ratings later in the year. Another factor is that Carlton started the season from a lower base given the ratings are generally dependent on the previous season. This may still be impacting their rating in some models to some extent, although I haven't quantified how much that may still be impacting their rating in my model. Their rating in my model has gone from -4.3 after the loss to Fremantle in round 6 to +7.0 currently so they could be up to 4th by my model in a week or two at that rate (they're currently 7th).

Yeah the "Aggregate" Power Rankings is basically the average of the ratings for all the individual Squiggle models. Given all the models use different methods (and rating scales), the ratings are standardised and then averaged. Here is a graphic I had created from the Squiggle API which shows the variability in some models' ratings of certain teams.

Sparrow's accuracy looks a lot worse when you include the six shots at goal which haven't registered a score. Here are all Melbourne's players to have taken at least five shots this year, noting of course that raw accuracy doesn't take into account where players take their shots from nor the pressure they are under when they take the kick. This is where Champion Data's expected score can be useful but they don't make the data available for individual players. Here are the least accurate players in the AFL this year who have taken at least ten shots. Notes: For the purpose of comparison to xScore, Score excludes rushed behinds. xScore sourced from the Herald Sun / Champion Data via https://twitter.com/OliverGigacz/status/1528338101042806785?s=20&t=LvFmdS4tCVKdBsutoZc2pg