Month: March 2015

A Quick Comparison of the New Models

There are still a few things to fine tune before the start of the season, but here is a quick comparison of two new models I have built.

A few disclaimers:

* the player based model assumes perfect knowledge of the team that will actually play in the game. This is unrealistic, but for most games, this will hopefully not make a big impact.
* the player based model does not yet take into account which player is a sub
* the profits are based on the odds info that I have accumulated over the past few years. These were generally recorded on Tuesday afternoon (before teams are announced), so I can’t compare my team based model using odds that also take into account the team announcements.

Here is the profit comparison for 2014:

2014_comparison

The model including player ratings got 152/207 (146 H&A season) tips correct for the entire season, while the team rating + form + inside 50s got 149/207(144 H&A season) and the team rating model got 148/207(141 H&A season). I don’t know if anyone can be bothered tracking down any expert tippers totals for 2014, but I think 152 (146) is pretty good.

Also, the player-based model shaved almost a full point of the average absolute error in 2014, which is fairly impressive.

The advantage of including a stat such as inside 50s isn’t obvious in the plot, but the gain in accuracy from using it is obvious over longer time periods.

Advertisements

We’re Back in 2015!

AFLpredictions will be back for 2015 with a suite of new improved models!

I have improved the simple (yet profitable) existing margin-based model with a suite of more accurate models by including the following parameters:

  • Current team form – often the team rating cannot update fast enough for teams on a hot streak (c.f. Richmond at the end of last year), so we will now include a current form parameter in the model,
  • Additional match statistics – sometimes the margin blows out at the end, or the team behind kicks a few cheap ones in junk time, so the margin doesn’t fully represent the difference in quality of the two teams. We will now include a number of additional statistics to feed into the model. I haven’t yet made a final decision on what will be included, but inside 50s, contested possessions and clearances are all in contention,
  • Improved Home Ground Advantage – a few small changes to give an “away” team that is actually at home (e.g. Collingwood vs The Western Bulldogs at Docklands) an advantage,
  • Player information – the big one. The quality of the actual line up which takes the field. I can only test this assuming perfect knowledge of who will play (so assuming I know all about late withdrawals), which is not exactly realistic, but the results are looking very promising.

I will write a post comparing the historical performance of a few of the different models in the coming days. This will illustrate the differences that each extra parameter makes.

Stay posted.