Machine Learning in the AFL

Discussion in 'Aussie Rules Football Discussion' started by Blake, Apr 2, 2016.

  1. Blake BE Quilty

    So, for my undergraduate thesis I've chosen to spend a year studying the effectiveness of data analytics and machine learning within the sport of Aussie Rules. From what I've found, there really hasn't been much research done at all in attempting to predict outcomes in AFL - or at least nothing using modern methods and techniques. There's been a bit of work done in baseball, basketball and NFL but I think Aussie Rules is a relatively untouched sport.

    From you guys, I'd love to hear if anyone has some analysis or relationships within the game they would like explored. The power of machine learning is that it can find specific combinations of variables that aren't immediately obvious to the human eye, but if there is stuff anyone wants looked at or thinks is interesting please post it in here or shoot me a PM.

    An example of something: I've started wondering if the significance of the home team advantage diminishes if the match is played at night. If there are any relationships such as this one that haven't really been found yet, then there's definitely some chance to gain some accuracy when predicting of an overall winner.

    Ultimately, I'm going to try and devise a system for predicting a match outcome, and then one to predict player performance compete in fantasy leagues. Obviously, if the player performance predictor works well then it could be used to assist the match outcome predictor as well. The goal is to have a completely autonomous system that is capable of tipping at a pretty high level, and competing in fantasy leagues on its own. If it's any good, I'll give it a crack against the bookies and in some real $$$ fantasy leagues.

    In the meantime, will keep this thread updated with my findings and anecdotes over the year.
     
  2. HeathDavisSpeed HT Davis

    I know a good number of bookies both in NZ and in Eastern Australia. They will have their own models of sorts, though not perhaps in the same way.

    Might be worth you talking to one or two of them as at the very least, you might be able to sell your model to the Victorian TAB (TABCorp?)
     
  3. Gazza GJ Weaver

    Very keen to see how this pans out, keep us informed! I'm keen on learning/making models for NBA/NFL etc.
     
  4. Blake BE Quilty

    Have definitely been thinking about that. If it goes well, I can either sell it to the public, or sell it to bookies. Most likely, the results won't be overly noteworthy though. I think it'd probably be pretty difficult to consistently beat the 10% rake that most bookies take.

    Will do mate. Have you made any models in the past?
     
  5. Gazza GJ Weaver

    Nah, but would be keen to get started.
     
  6. Rego RS Hutchinson

    There's a lot of sites out there that look at all sorts of variables and spit out the data for like the NBA. How many games in a week and how often the team wins the 4th in a week for example. Check into those sites, most are paid but I do remember one had a free trial.
     
  7. Blake BE Quilty

    Give me a buzz if you want to get started with me as well. I'd be happy to have a chat. Just going to be using a combination of Python/MATLAB to start, and the majority of algorithms are not overly complicated.
     
  8. Blake BE Quilty

    Have done a bit more research over the past few days. There's been a bit more statistical work done than I first realised, but still nothing to do with machine learning or any of the more complex algorithms.

    For anyone wanting to make their own models, I found an awesome dataset provided by Sportsbet for a Machine Learning competition last year. Featured a $5000 first prize as well, looks like I am a year too slow.

    http://www.sportsbetcikm15.com/

    I will be posting a ~ 15-page PDF of my own initial findings, and thesis proposal, in the days to come. Stay tuned!
     
  9. Himannv LV Himann

    Are you planning on making your own model?

    The company I work for have a Machine Learning product that does it. They used it to predict the winner of the superbowl and individual games. Right now they're working on something related to the US elections.
     
  10. Blake BE Quilty

    Yeah, that is pretty much the point of this thread. The models existing for AFL don't utilise some of the more powerful machine learning techniques. I'm hoping I can improve on them with my own models.

    Your company's model sounds similar to what Nate Silver does, have you heard of him? He started with baseball and now predicts mostly US politics.
     
  11. Blake BE Quilty

    Also, I submitted my thesis proposal today. It's a fairly beginner-oriented 15 page or so guide to what I'm going to be doing over the year and contains a spiel on some of cool machine learning techniques.

    If anyone wants a copy, like this post and I will send it to you in a PM. Will publish it publicly later.
     
    Last edited: Apr 7, 2016
  12. Julian BJ Taylor

    Sounds like a pretty fucking smart ploy from Sportsbet honestly. Get 100 plebs to send you in days and days worth of this stuff that they can keep for good for their benefit and they just give one dude 5k.
     
  13. Blake BE Quilty

    Yeah great idea by them. However, the guy that won it ended up using a pretty pleb spreadsheet so they probably weren't too happy with that lol.
     
  14. Mariner CL Warrington

    Have you checked out matterofstats.com ? Guy manages to churn out pretty decent afl predictions from what I've seen.

    Have been doing a bit of this myself recently, but with cricket/rugby. Really interesting stuff.
     
  15. El Nino J Torres

  16. Blake BE Quilty

    Awesome, cheers guys. Matter of Stats in particular looks really, really good, and both of these sites have kinda provided a new standard in terms of marketing and stuff that I want to do. Has honestly lifted the bar a bit now for me!
     
  17. Blake BE Quilty

    Have had a little bit of time free to look at something I've been interested in; how the rest between matches takes its toll on teams. Some figures I've produced definitely highlights that there is a considerable difference in having more rest than your opposition. There is a little bit of inconsistency towards the higher end, most likely due to the smaller sample size, but it looks like having a few extra days rest over your opposition is definitely a sweet spot.

    [​IMG]

    [​IMG]

    These statistics become a little clearer when we compare the duration of rest with our opposition's rest each game. Here's how rest advantages impact the margin:

    [​IMG]

    [​IMG]

    I used season data from 2000-2015, excluding Finals and Round 1.
     
    Last edited: Apr 13, 2016
  18. Blake BE Quilty

    First iteration of my predictor BEBot v1.0 is complete. Using a fairly simple method and some basic optimisation, ELOve tipped 135/197 (68.53%) of the regular season games in 2015 correctly.

    [​IMG]

    Onwards and upwards for BEBot. ;)
     
  19. Blake BE Quilty

  20. Rego RS Hutchinson

    What are the variables in a basic sum up?
     

Share This Page