One of the main reasons I enjoyed the book Moneyball so much was that I felt like I was reading about a bunch of kindred spirits, people who were seeking better analysis of a game through more meaningful statistics.
The difference was that I was looking at AFL football. That and my motivation to beat the bookmakers at their own game.
I’ve been using statistics to measure form in the AFL for about eight years now. At the end of each season I usually tinker with my spreadsheet and look at correlations between season wins and each category of information I collect.
I might add additional categories of information if reason tells me it will help improve my prediction modelling.
If you’re a Fox Footy viewer you can’t have failed to notice David King’s (and Champion Data’s) mantra – Contest, Attack, Defend.
In other words if a team wins more contested possessions (CP), has a higher disposal efficiency (DE) and wins the tackle count (T), victory results 96.1 percent of the time.
This makes sense. If you win the ball better than your opponents and use it better you’ll probably win, especially if you’re also working hard to put pressure on your opponents when they have the ball.
But can we improve on Fox Footy model? I think so.
Firstly, I think we should normalise the figures rather than use differentials.
We already do this in a way that’s very familiar to fans, with our ladder and points scored for and against – we call it percentage.
Everyone understands this is fairer than a straight point differential figure, so why not apply a similar calculation to CP, DE and T?
Once you have these three percentage scores you can assign them different weightings, based on your own perception of their value, and combine them to get a total score of Proficiency (or whatever name you choose).
My end-of-season evaluation from last year showed that tackling performance had the lowest correlation with games won. Therefore when I make the calculation I weigh both CP and DE as twice as valuable.
With the formula I give teams a positive proficiency score (PS) if they tally above 100 percent and teams a negative proficiency score if they are less than 100, just like percentage.
Last year a team’s season performance in my PS category correlated very highly with season wins. As such, chances of victory in any given match increase dramatically as the difference between the two teams’ PS scores increases.
If you’re like me, you can keep your PS formula in a spreadsheet. Then, on your smart phone, you can call up the figures from the Superfooty live site while watching the game at the pub, plug the figures into your spreadsheet and make/adjust bets accordingly.
Another improvement would be to show us quarter by quarter splits and game totals of CP, DE and T, so that we can see how a team is improving its performance (or why it is continuing to go down the gurgler).
Finally a better descriptor than DE would be to split it into kicking efficiency and handball efficiency. I’m sure this wouldn’t be too hard for the Champion Data boffins.
Currently if Sam Gilbert handballs to three teammates well but sprays the ball over a teammate’s head with a kick, his disposal efficiency would be 75 percent. Such a percentage is considered quite good, but is not at all reflective of what is going on.
I think I’ll leave it there for now. While I have a bit more to say about statistical analysis in footy I’m not sure if the land of The Roar is populated with the same amount of stats nerds as the book Moneyball.
Though if the response to this article is positive, you may be hearing more from me…
john
Roar Rookie
Hi can any of you help me with stats , as to which ones & what weighting on each. Simonjzw would love to nhear more about how to work out PS & any help with setting up a spreadsheet. thaks JAM
Jeremy
Guest
I too am looking to predict outcomes but it is extremely hard. There is an enormous amount of stats to can make up a teams winning chances but even then there are outside factors that can contribute to another team winning such as injuries, home ground advantage, days from last game, change in playing roster, change in player positions, the list goes on. Also its hard to predict, its easy to say a team won after you have the stat sheet but to predict they can do it against a different team is very hard to predict. The team the beat may be inferior or in a form slump. for e.g Sydney dominated west coast at patersons yet fremantly infllicted the same result so the sydney win has to be validated. Also making money is hard to do on a couple of fronts. 1. If a team is winning by a margin and gets to 100 or 93 first there odds are going to be close to 1.01 or you may not even get that unless the "underdog" is winning and even then if its by a margin and its past 3qtr time the resulting odds are going to be unders or evens, i cant see a bookmaker giving overs. And 2. If its close and a team is winning the performance indicators its hard to say they will definately continue to do so. As you mentioned before the momentum swings sometimes to each team numerous times. I believe in laying the team with the momentum and then backing when the other teams gains the lead but as i said before its hard to say the team that i lay will continue to lose and therefore i've done my coin. SYD vs COLL was a good example of both teams winning the contest at certain stages. It would be luck to back coll when they had the upperhand and unlucky to back syd when they won the phases.
Jeremy
Guest
Fully agree on that mantra and i believe many AFL clubs use the same 3 phase model or "when we have it", "When they have it" and in "contest". Adrian i think your stats of contested possessions accurately reflects winning the "contested phase" but i cannot same they same for efficiency reflecting winning the "attack phase" or tackles winning the "defense phase". While they go a long way to making up those respective phases its not the whole picture. In my opinion when a team wins a contested possession from a stopage, clearance or a general play the aim is ultimately gain a shot on goal with an emphasis or actually scoring a goal and not a behind. If a team wins possession from a contest in the backline and effectively hits 9 targets making its way up the field but ineffectively delivers the ball inside 50 which results in a turnover the contested possession and disposal efficiency of 90% mean very little. In attack "spread" is very important and i believe it plays a role in a teams winning chances (imo its why sydney have become a far better team this year). I am developing a spread statistic that is basically calibrated from uncontested possesions after winning the contested possession that results in a shot on goal. As for the defense phase tackling does not accurately show a teams defensive ability with my main example being collingwood, they are a very good tackling team but there defensive key performance indicators such as spoiling, general press acts and pressure are outstanding. In my opinion the defense phase should be obtained from any act that results in a team gaining possession of the other team, whether that relates to a direct turnover or a contested phase it doesn't matter. Tackling does reflect this but so does pressure on the ball carrier without effecting the tackle, making a spoil on a marking contest, smother or simply out marking your direct opponent. Inside 50"s are a good indicator but i think there a result from one or even two of the "3 phase philosphy". GWS are an example of when inside 50's equate to nothing. In round one against sydney they were on par in contested possessions and inside 50's yet lost by 50 points. I know it was round one but there kids are more often than not get beaten by there direct opponent until they can develop core strength. I aslo think disposal efficiency whether it be handball or kicking is misleading. If a defender gains 20 possessions at 100% effeciency yet only kicks it sideways or 15m forward under no pressure, does he actually help his team win? On the same token if a midfielder win 20 possessions in the contest and gains a quick kick forward that only results in a 60% effeciency, does he not help his team more than the defender? The midfielder helps his team far more than the defender but it doesn't show on the stat sheet. I'm also trying to develop a metric on disposal effeciency. This is a little hard to develop without a subjective opinion. In my opinion it should be ultimately measured on metres gained and actual effectiveness. But a player under no pressure is able to recieve the same credit as a player under pressure for a similar "metres gained with effeciency" possession which is unfair. This is were its hard to "rate" percieved pressure without being subjective. I think the best way is to catergorise it into 1. No pressure 2.Minimal pressure (players around the ball carrier but no serious pressure), 3.direct pressure (someone trying to affect a smother but player still has good balance) and 4. intense pressure (player is under enormous pressure with 2 or more tacklers and is unbalanced). As you could imagine it would be far greater to effect an effecient disposal being under enormous pressure than it is to be under no pressure so why is the player awarded a better mark just because he was under no pressure. It doesn't accurately reflect the context of the disposal. Anyway i rambling, my point on being subjective is one person might see a pessure situation being 3 instead of 4 or 1 instead of 2, you get my picture.
Adrian
Guest
No worries, I spent the long weekend up at Blanchetown myself. I agree regarding the inside 50 ratio, I found it to be the most useful indicator of winning percentage. Also, I have tracked the forward 50 efficiency stat you mention, and have found it to provide useful insight. In addition I calculate a defensive measure using (i50s conceeded)/(G + P conceeded) essentially rewarding teams for minimising scoring opportunities from their opponents i50s, obviously this combination of metrics attributes kicking accuracy to the attacking team, which seems fairest. Its interesting to note that Hawthorn (2.41) actually outperformed Geelong (2.20) defensively, suggesting that the biggest issue for Hawthorn was there inaccuracy, which I think agrees with the perception alot of people had watching the game. This is doubly interesting because Hawthorn, for the last year or so, have had the best performance in attacking efficiency. With respect to winning correlation, I've found that these attacking and defending metrics do not correlate as well in general, it seems to be between 0.3 and 0.6 depending on the year. However, I found that if you only consider close games (can't remember my exact definition, perhaps within two goals?) then the correlation jumps right up to be on similar grounds to the i50s ratio. I think this indicates that, in a comfortable win, the key difference between sides is in the "midifeld" or the ability to create and prevent inside 50s. By extension, the forward and defence metrics are significant for close games, such as the cats v hawks, and in particular, big finals. As an aside, I made a power ranking by summing the z-scores of these three metrics, which again correlates in the mid to high 0.9s.
simonjzw
Roar Pro
A fellow stats nerd! Well done Adrian! (and sorry for taking a while to reply but I was away over Easter) I agree percentage is a powerful indicator, my figures for last year's season showed a correlation between percentage and number of wins to be 0.97 (not including finals). My metric Proficiency Score (PS) correlated at 0.93 over the full year. But there's another powerful metric! - Inside 50% (Inside 50s for / Inside 50s against x 100) last season that had a correlation of 0.97 with season wins. When you combine all three metrics you get a correlation of just over 0.98. The advantage of looking at PS and Inside 50% as well the actual score during a game is that if you can identify a change in trend in one or both of those metrics it may signal a change in "momentum". The other thing I've noticed is that a team very rarely loses when the difference in PS is greater than 10. As to which is best during the game? I look at PS and Inside 50% closely and also the current score before making a decision any punting decision. I noticed Dermot Brereton on a broadcast earlier this season said that the team that reaches 100 points first wins 90 something % of matches (can't remember the exact figure!). Makes sense to me since the average score is about 93 points so a score of 100 would win more games than it loses anyway. I'd be very interested to know the figures around the team reaching the league average score first (i.e 93) ! which could be an excellent pointer during matches. As you correctly point out the score is essentially a complicated statistic. Of course we're along way from a perfect model - fantastic game on Monday and Geelong had a PS of 94.78 and Hawthorn - 105.94 so it should have been a Hawthorn victory. And Hawthorn also won the % Inside 50 with 103.7%! So where did Geelong win the game? They had a way more efficient forward line - greater accuracy in kicking for goal and higher point score per inside 50 (1.74 points per inside 50 entry as opposed to Hawthorn's 1.64). No doubt the dominance of their key forwards Hawkins and Podsiadly contributed to this forward line efficiency. I think another article on the most valuable stats and stats I'd like to see may be warranted! Cheers Adrian
The Cattery
Roar Guru
good article simon I just saw that fox ad come up, watching the SC show, and then realised, hang on, didn't I see an article of the same name today on Fox?? and here I am! It's funny that they quote Jeansy during that ad: either we've got it, they've got it, or it's in dispute. And the mantra of contested possessions, kicking efficiency and tackles is really a reduction of those basic principles as stats, although it has to be said, it took a very, very long time to move from basic k m h t stats to putting some quality around those stats as happens today, and is continuing to develop. Really interesting stuff, thanks for bringing it up.
Adrian
Guest
simonjz, I started collecting data on the AFL last year with similar goals to you, although I have been looking more at determining what team strengths and weaknesses are to enhance my understanding of the game, rather than to beat the bookies. Out of curiosity, what correlation have you got with victory for your combined metric. I found last year that simple percentage had a correlation of around 0.95 with winning percentage, and my own combined stat, which is basically a sum of opportunity efficiencies, could perform around as well, but not clearly outperform it. As a result of this, I am presently of the opinino that percentage may be the best pre game predictor (particularly if combined with a strength of schedule adjustment, which I am experimenting with this year). As an aside, is your stat predictive or descriptive; by which I mean, if a side is winning at any given point, wouldn't you expect it to be ahead in your metric by nature? Is your metric a better predictor of results in game thah the current score? After all, the score is essentially a very complicated statistic that rewards all aspects of good play.