Remove this Banner Ad

Handicapping/modelling AFL

🥰 Love BigFooty? Join now for free.

Aug 15, 2006
13,328
3,953
.
AFL Club
Western Bulldogs
tl:dr - Does anybody create AFL models?

Hey guys, I just thought it was interesting that there is a punting board on an AFL site that doesn't have a lot of in-depth discussion on making predictive models for the AFL. I realise that this could be because the AFL is notoriously hard to model. I mean, if Roby's thread on the main board is anything to go by then maybe it'd be wise to steer clear of creating a system to predict at all :p

As for me, I haven't done a lot of gambling this year as I've been working away furiously creating a system that may never see the light of day. One thing I learnt from the on-going creation of this model is how naive it would be to gamble on 'gut feelings' - even if those gut feelings are in part made up subconsciously of the ins and outs or how Essendon travel to W.A last time etc.

I've come to this conclusion from learning more about the bookies and their odds and lines and I've concluded that only the smallest percentage of people could ever beat these odds over the long term without some sort of handicapping/value identifying betting system.

Personally I get my own AFL lines down to about the same accuracy as the major bookie lines at the bounce. For example I can predict an AFL game to within around 26pts in 2013 - I don't know anyone that can predict a game closer than me but how would I know anyone else? I'd be keeping quiet too if I had a system that could beat the bookies and their vig. Despite being slightly better at predicting an AFL game than the bookies, it still isn't low enough to make any decent return but it doesn't lose money and if I 'shop for odds' I might get up a few more %. But it's a lot of work for little return. Especially at my hobby amounts.

Anyway, sorry for the essay. I just frequently visit this board and am constantly surprised that there isn't any quantitative discussion ever going on.

Is there anybody out that does this? I'd love to hear from you guys.
 
brett128...

Seriously though, there's a couple of guys on this board I'm aware of - Rourke and robertbn.

I've found models particularly good at identifying subsets of games that bookies and the public consistently get wrong, rather than using it to bet every game.
 
I've found models particularly good at identifying subsets of games that bookies and the public consistently get wrong, rather than using it to bet every game.

I agree, using a system to identify value seems to be the most logical way to bet. Unfortunately these games for me are far and few between and only offer a slightly better chance of winning.

I've actually perused through some of Rourke's post in the past. I also notice it says 'Ranking software statistician' as his occupation. I'd love to hear from Rourke on if he uses his system for betting and to what extent. Is the system based on a past results or is it more predictive of 'in game play' such as analysing inside 50's/marks and player match ups.

I've only just experimented with 'in game' analysis but find it difficult due to the maths involved in working out how important a contested possession/clearance is to the outcome of a game. But I feel this is what really separates the boys from the men in this field. I feel I've got the regression analysis on past results down pat but I feel this is strongly correlated with the bookies/collective of other punters feelings.

(I'm also reading over robertbn. website. It's exactly the sort of the stuff I was looking for)
 

Log in to remove this Banner Ad

Hi Butane,

I've done a bit of modeling similar to what you are talking about (thanks for the name-drop d3ft!).

I do, however, think they're are some people who have good footy instincts and can predict margins and results better than "experts". On top of that, I think that bookmakers can be beaten simply because they're goal is not to only to predict games accurately, but to adjust odds according to where money is going. This can actually make odds quite skewed at time, but only if punters are putting their money in the wrong place, at which point the bookmakers don't care because all they want are their losers to pay their winners and take a cut for themselves. Look at Sam Mitchell in the Brownlow this year - lots of money came in for him late and he shortened from $21 to $5.50! Why? Because he had a cracker in the prelim. But who cares about a prelim? The public have a skewed view that he had a stellar season just because he got 38 touches where he couldn't have even got votes! Journalists doing a "phantom Brownlow" then get swept up and start finding votes for him retrospectively to justify their opinion that he could win and it all feeds in to a fervor and his odds shorten. (Sorry for the rant, but for some reason this really annoyed me.)

But if you can look at it in an objective way (statistically) or if you are just all over the footy (know every player, watch every game etc.), you can take advantage of these trends and profit from them. And having done this for two seasons now, the odds are "wrong" surprisingly frequently. Have a look at my blog (http://www.aflpredictions.wordpress.com) and you'll see that the chances of winning that I give a team can differ quite a bit from the bookmakers. Who is better? Well they're probably better over ALL games, but I don't have to bet on all games, I just bet when I think I know better. And I've posted long term growth using my model. And the model is surprisingly simple. Could it be made better. Absolutely! The model's average error is ~29 points, so if you can average an error of 26 points over every game, then well done!

tl;dr odds are wrong often enough for the astute observer to post fairly reliable profits.
 
Nickel's worth of free advice: it would be exceptionally difficult to create a quantitative model that has predictive power better than the market. Assuming your time and resources are limited, the best way to approach the problem is to go through as much games as possible, and compare the line to the result. Develop some general theories and use data to test and refine your theories. You'll need to think a little uniquely, and above all work harder than everyone else because you're starting years behind everyone who is currently setting and betting the lines. My best advice would be to watch every game.
 
Id love to learn more Butane

I wish i had better mathematical skills.

But what sort of thing does your model involve? Am i on the right path in saying correlation between winning and clearances, disposals, I50s ect?

How do you then predict who will have those higher numbers?
 
IMHO most people who model sport overfit to the data. My approach is to start with some physical assumptions and a limited set of parameters, with statistical assumptions about the variation you see in the measurements of game results (scores & key stats). Anything that doesn't add value to the prediction is not used. Don't read too much into the most recent match, some people treat it as more than 10% of all data but that's another type of overfitting.

Of course I have the advantage that I get access to a customised live description of the match from TedSport, which has exactly the metrics we believe are important. But publicly available stats can get you most of the way there, if you're prepared to work at maintaining your own database.

Then when it comes to deciding on bet sizes, accept that you don't know everything and the market price provides some of that extra info you need; you can often explain the difference with a peculiarity that may or may not be significant (e.g. doubt over key player(s)). Something like half-Kelly is ok. I'm sitting on about a 25.5 average error this year, the same as bookies, but this is an unusually predictable season. The advantage you have is that the bookie has to specify a price first, so it's like permanent late position in Texas Holdem.

I've given up trying to predict which way the market will move. A paper by Prof Steve Clarke several years ago showed that it is possible to make a higher profit on the early lines, instead of waiting for the margins to come in late in the week.

If you actually have an extra insight from watching the game (I don't), it's best to follow the example of someone like Bob Voulgaris who uses mathematics (actually, mathematicians) to calibrate his observations and turn them into appropriate bet sizes.

tl;dr Simple models work if you use high-quality inputs
 
Some points:

* If your model average error is beating closing line error, book that Rolls Royce for pick up in 24 months time. Given the fact the bettor has first mover advantage, you will have the ability to pound mis-priced lines.

* Everyone posting so far seems to take a classical statistician's view of model creation. Which is to develop a hypothesis (or series of hypotheses) and let domain knowledge drive model creation. I can see everyone here painstakingly going through some variant of regression to create their models.

Absolutely nothing wrong with that, it should work well, but there are other methods also that can bring a 'model to market' much more quickly. IMO, regression is not dynamic enough, it breaks too easily if the nature of the sport changes.
 
Some points:

* If your model average error is beating closing line error, book that Rolls Royce for pick up in 24 months time. Given the fact the bettor has first mover advantage, you will have the ability to pound mis-priced lines.

* Everyone posting so far seems to take a classical statistician's view of model creation. Which is to develop a hypothesis (or series of hypotheses) and let domain knowledge drive model creation. I can see everyone here painstakingly going through some variant of regression to create their models.

Absolutely nothing wrong with that, it should work well, but there are other methods also that can bring a 'model to market' much more quickly. IMO, regression is not dynamic enough, it breaks too easily if the nature of the sport changes.


Dying to know how the model worked on the Brownlow, I don't think it's feasable TBH. But would like a serious response if possible. It obviously wouldn't take a brilliant model to throw up the top 3 this year, more interested in how the team votes went with smaller totals like Melb North, Carlton.

:thumbsu:
 
Hi Butane,

I've done a bit of modeling similar to what you are talking about (thanks for the name-drop d3ft!).

I do, however, think they're are some people who have good footy instincts and can predict margins and results better than "experts". On top of that, I think that bookmakers can be beaten simply because they're goal is not to only to predict games accurately, but to adjust odds according to where money is going. This can actually make odds quite skewed at time, but only if punters are putting their money in the wrong place, at which point the bookmakers don't care because all they want are their losers to pay their winners and take a cut for themselves. Look at Sam Mitchell in the Brownlow this year - lots of money came in for him late and he shortened from $21 to $5.50! Why? Because he had a cracker in the prelim. But who cares about a prelim? The public have a skewed view that he had a stellar season just because he got 38 touches where he couldn't have even got votes! Journalists doing a "phantom Brownlow" then get swept up and start finding votes for him retrospectively to justify their opinion that he could win and it all feeds in to a fervor and his odds shorten. (Sorry for the rant, but for some reason this really annoyed me.)

But if you can look at it in an objective way (statistically) or if you are just all over the footy (know every player, watch every game etc.), you can take advantage of these trends and profit from them. And having done this for two seasons now, the odds are "wrong" surprisingly frequently. Have a look at my blog (http://www.aflpredictions.wordpress.com) and you'll see that the chances of winning that I give a team can differ quite a bit from the bookmakers. Who is better? Well they're probably better over ALL games, but I don't have to bet on all games, I just bet when I think I know better. And I've posted long term growth using my model. And the model is surprisingly simple. Could it be made better. Absolutely! The model's average error is ~29 points, so if you can average an error of 26 points over every game, then well done!

tl;dr odds are wrong often enough for the astute observer to post fairly reliable profits.


Thanks for the reply. I have to admit that I spent quite a few hours last night on your site and following your lead of creating histograms of my predictions and turning the percentages in to odds. I probably don't have enough past games handicapped yet to create a fully functioning histogram as my method is quite time consuming(and I'm not savy enough to automate it better ). Early results are promising though.

I'll update you as I add more games and data. Great work on the website too.
 
Nickel's worth of free advice: it would be exceptionally difficult to create a quantitative model that has predictive power better than the market. Assuming your time and resources are limited, the best way to approach the problem is to go through as much games as possible, and compare the line to the result. Develop some general theories and use data to test and refine your theories. You'll need to think a little uniquely, and above all work harder than everyone else because you're starting years behind everyone who is currently setting and betting the lines. My best advice would be to watch every game.

I agree with what's been said - the bookies must move with the money as to insure their arses giving the 'sharp' handicapper the advantage as you don't have to move your odds. I figure if I can be as good at predicting games as the bookmaker than I will be able to exploit that. You're right though, that's the challenge - how does somebody build a model that's better/equal to the predictive power of the market. Here's where I'm hoping the bookie has to move more than just the overround.

As for thinking uniquely, I totally agree. I read some interviews with Bob Vougaris(who Rourke mentioned) and he seems to believe you need to find some information that other 'punters' aren't using that has the edge. This is obviously a major hurdle and I'm sure is where the 'watching as many games as possible' advice comes in.
 
Rourke I'm also wondering whether the underlying variables that TedSport uses will be available for purchase any time in the future. I'm sure there's a market out there for people that would like to take their own look at the TedSport variables.
 

Remove this Banner Ad

Dying to know how the model worked on the Brownlow, I don't think it's feasable TBH. But would like a serious response if possible. It obviously wouldn't take a brilliant model to throw up the top 3 this year, more interested in how the team votes went with smaller totals like Melb North, Carlton.

:thumbsu:


I don't have 'a' model. Every market/sport/any other data problem requires a tailored solution.
 
Id love to learn more Butane

I wish i had better mathematical skills.

But what sort of thing does your model involve? Am i on the right path in saying correlation between winning and clearances, disposals, I50s ect?

How do you then predict who will have those higher numbers?

Firstly, I don't profess to have much mathematical ability - I only work in finance :D

My model is really an ensemble of predictors that I've created and it's certainly nothing special but each one is unique enough in it's own way that combined they produce, historically, a lower error than any one of them alone.

I haven't added my 'in game' models to it yet as I'm not confident in them for exactly the reasons you've listed. How do you predict future inside 50's against different teams? I have my theory's but I wouldn't bet on them. Although I enjoy the challenge and it certainly adds interest to games when I'm watching them. For example I'm not overly impressed with teams that win contested footy/tackles(But they're still important in a good team! :D) but my ears prick up now when I see a team getting a whole bunch marks inside 50...but more importantly it's how they do it. But like I said, that's more the footy analyst coming out of me than the amateur handicapper. I'm not going that deep down the rabbit hole.
 
Some points:

* If your model average error is beating closing line error, book that Rolls Royce for pick up in 24 months time. Given the fact the bettor has first mover advantage, you will have the ability to pound mis-priced lines.

* Everyone posting so far seems to take a classical statistician's view of model creation. Which is to develop a hypothesis (or series of hypotheses) and let domain knowledge drive model creation. I can see everyone here painstakingly going through some variant of regression to create their models.

Absolutely nothing wrong with that, it should work well, but there are other methods also that can bring a 'model to market' much more quickly. IMO, regression is not dynamic enough, it breaks too easily if the nature of the sport changes.


Unfortunately my model needs the final list of players which isn't settled until 2 hours before the bounce so I don't have the option of betting any earlier. Having said this, I could speculate the final teams and if there is a late withdrawal it wouldn't effect my predictions by more than a point or 2 anyway so I really should bet earlier.

And yes, I definitely have a regression system and appreciate that it will only take me so far. I think that's what has led me here to try and glean more information from AFL handicappers on the nuances of the sport itself. I'd like to learn more about these other models. I'm always building different models and testing them to see if they pass the test.
 
IMHO most people who model sport overfit to the data. My approach is to start with some physical assumptions and a limited set of parameters, with statistical assumptions about the variation you see in the measurements of game results (scores & key stats). Anything that doesn't add value to the prediction is not used. Don't read too much into the most recent match, some people treat it as more than 10% of all data but that's another type of overfitting.

Of course I have the advantage that I get access to a customised live description of the match from TedSport, which has exactly the metrics we believe are important. But publicly available stats can get you most of the way there, if you're prepared to work at maintaining your own database.

Then when it comes to deciding on bet sizes, accept that you don't know everything and the market price provides some of that extra info you need; you can often explain the difference with a peculiarity that may or may not be significant (e.g. doubt over key player(s)). Something like half-Kelly is ok. I'm sitting on about a 25.5 average error this year, the same as bookies, but this is an unusually predictable season. The advantage you have is that the bookie has to specify a price first, so it's like permanent late position in Texas Holdem.

I've given up trying to predict which way the market will move. A paper by Prof Steve Clarke several years ago showed that it is possible to make a higher profit on the early lines, instead of waiting for the margins to come in late in the week.

If you actually have an extra insight from watching the game (I don't), it's best to follow the example of someone like Bob Voulgaris who uses mathematics (actually, mathematicians) to calibrate his observations and turn them into appropriate bet sizes.

tl;dr Simple models work if you use high-quality inputs


Appreciate the reply Rourke, I've also had a look through your website and had a look through some of the links that you've posted.

I definitely could be accused of overfitting my data although I seem to be getting better at it as I eliminate data that's either bordering on irrelevant or that's already covered too much in some other way. I've explored recent match history and do use it but as you pointed out only as a very small fraction of the final outcome. I've also discovered, which maybe no surprise to anyway that past head to head results mean close to jack all. In fact you could just assume the home team is going to win by a goal or so then bet on previous encounters between the two teams playing. But that's just one of those quirks I enjoy finding.

I don't think I'm at the stage yet where I'm worrying about the money management side. But I have read extensively of the different staking systems but for the moment my goal is to get my error down as low as possible without shrinking the result to much. I need to focus on one thing at a time with this because other wise I get distracted and go off on another tangent :D
 
Dying to know how the model worked on the Brownlow, I don't think it's feasable TBH. But would like a serious response if possible. It obviously wouldn't take a brilliant model to throw up the top 3 this year, more interested in how the team votes went with smaller totals like Melb North, Carlton.

:thumbsu:
For what it's worth, my Brownlow model absolutely killed it. I'm happy to put up the full predicted votes if anyone is interested.

Sent from my LT22i using Tapatalk 2
 
For what it's worth, my Brownlow model absolutely killed it. I'm happy to put up the full predicted votes if anyone is interested.

Sent from my LT22i using Tapatalk 2
Id like to see it

Firstly, I don't profess to have much mathematical ability - I only work in finance :D

My model is really an ensemble of predictors that I've created and it's certainly nothing special but each one is unique enough in it's own way that combined they produce, historically, a lower error than any one of them alone.

I haven't added my 'in game' models to it yet as I'm not confident in them for exactly the reasons you've listed. How do you predict future inside 50's against different teams? I have my theory's but I wouldn't bet on them. Although I enjoy the challenge and it certainly adds interest to games when I'm watching them. For example I'm not overly impressed with teams that win contested footy/tackles(But they're still important in a good team! :D) but my ears prick up now when I see a team getting a whole bunch marks inside 50...but more importantly it's how they do it. But like I said, that's more the footy analyst coming out of me than the amateur handicapper. I'm not going that deep down the rabbit hole.

So maybe more along the lines of how much a home team won

How much a team won after a 7 day break, 6 day break, 8 day break ect

Recent past result(s)

You have to speak lamen terms with me ha

When it comes to me ive never done any modelling or really thought about it but it interest me greatly and my mathematical mind.

I pretty much only bet on DT points this year because i thought i could predict far greater accuracy as to matchups compared to the bookies and the results certainly showed that. My bets have more been about me using just general football knowledge and if i start losing i know that knowledge at the time isnt good enough, thus stop betting for some time until it improves with fake bets as to what i can predict.
 

🥰 Love BigFooty? Join now for free.

I'm good with Excel and spreadsheets, at a simple level.

Is AFL that good of a sport to bet on though? 23/4 players + 3 umpires? A heap of variables...
 
Also there's really two pieces to the puzzle, the second one may not be obvious right away:

1. Picking winners- will take lots of work and trial & error, but becomes relatively easy once you reach a critical mass of expertise
2. Getting action- conversely, this is easier as you begin, but over time becomes harder as you get banned/crippled at the corporates, premium tax at Betfair, etc. You can get friends to sign up accounts and bet for you, but these accounts are always temporary and lots of practical hassle. There are places for sure where you can always get some action, but ultimately you're fairly restricted unless you are v sneaky.
 
Also there's really two pieces to the puzzle, the second one may not be obvious right away:

1. Picking winners- will take lots of work and trial & error, but becomes relatively easy once you reach a critical mass of expertise
2. Getting action- conversely, this is easier as you begin, but over time becomes harder as you get banned/crippled at the corporates, premium tax at Betfair, etc. You can get friends to sign up accounts and bet for you, but these accounts are always temporary and lots of practical hassle. There are places for sure where you can always get some action, but ultimately you're fairly restricted unless you are v sneaky.
With regards to point 2, pinnacle will sort you out.

Sent from my LT22i using Tapatalk 2
 
With regards to point 2, pinnacle will sort you out.

Sent from my LT22i using Tapatalk 2


$1.3K limits is not exactly Pinnacle hosting a liquid market. Yeah you can re-pop multiple times and get $2.6K, $3.9K, etc. But there's only so much liquidity Pinny can take before some other book (that you can't get at because of being limited) has the better line/price.

I'd say don't just model the AFL, and get into the big global sports.
 
For what it's worth I reckon making a model for the Brownlow would be incredibly more difficult than modelling AFL games. You have to guess who the 3 umpires would deem the 3 best on grounds not who actually was. Sure if it was say 20 or more people voting in each game you might be able to get a more realistic/predictable result. If you wanted to get really serious about trying to build a brownlow 'black box' I'd be studying the umpires involved in each game and who they historically vote for - e.g players or positions they lean too and add that to the system of who I deemed to be in the best 3.
 
For what it's worth I reckon making a model for the Brownlow would be incredibly more difficult than modelling AFL games. You have to guess who the 3 umpires would deem the 3 best on grounds not who actually was. Sure if it was say 20 or more people voting in each game you might be able to get a more realistic/predictable result. If you wanted to get really serious about trying to build a brownlow 'black box' I'd be studying the umpires involved in each game and who they historically vote for - e.g players or positions they lean too and add that to the system of who I deemed to be in the best 3.
I plan on looking into this for next year. Not sure how much value it will add though.
 

Remove this Banner Ad

Handicapping/modelling AFL


Write your reply...

Remove this Banner Ad

🥰 Love BigFooty? Join now for free.

Back
Top