AFL 2009 Season - Regular Season Review

Now that the home-and-away matches have been completed I thought I'd share the overall results of the model for the year, both in terms of tipping and betting.

Tipping

The tipping did quite well for a first attempt at an AFL model - finishing on 123 tips at 69.9% success (counting both draws as 0.5 tips). The leading Herald Sun football expert finished on 122 tips, so we topped them, while Shane Warne finished on an impressive 128, no doubt buoyed by tipping St Kilda every week.

Comparing against the results from footytips.com.au (who count a draw as a whole tip, not just 0.5) - out of the 317,000 tippers we would have finished in the top 5% - 7% which again is a pretty good result for a start-up model. Hopefully some improvements to the model will only improve on this figure - more on this later.

Betting

Tipping is nice and fun and goes a long way towards bragging rights, but if a model can be used to make consistent profits then all the better. As mentioned in previous posts I follow the Kelly Criterion method of betting - which has the useful property of maximising one's exponential profit. The criteria works by finding out if the model throws up an advantage over the bookmaker (for example, if a team is rated 50% and a bookie is offering $2.20 then we have an advantage, if they are only offering $1.80 then we don't). The higher the advantage over the bookie, the more that is bet.

Two options are presented in the results below. The first shows the results had we potentially bet on every game, provided that there was an advantage over the bookie. The second option will only bet on a team if they are playing in their home state, or at a neutral venue. So in this scenario if Hawthorn played West Cost in Perth and provided a betting opportunity, we wouldn't bet on them. However if these teams played in Tasmania, then we would bet on Hawthorn as it is a neutral venue. Respected sports predictions website Sportpunter has shown in the past that betting on teams travelling interstate isn't as profitable as those playing at home or neutrally, and in fact may end up losing money. So we will present both scenarios and see if there is a significant difference.

The analysis that follows will compare various betting strategies dependent on how much we rate a team to win, and how much of an bookie's advantage (overlay) we have.

The first piece of analysis we'll look at here is betting on every game, provided there is a an advantage to be found. So regardless of how highly we rated a team, or the size of the overlay we'll bet on them - so long as there is an overlay. As mentioned, we'll compare betting on all games to just betting on home/neutral teams.

Return on Investment

• All games: 17%
• Home/Neutral Only: 25%

The return on investment is simply the amount profited (or lost) as a percentage of the total amount bet. So looking at the above graph we can see that only betting on home/neutral teams returned around $1050 profit while betting on all games returned around $900 - instantly we can see in this scenario that the bets on the teams playing interstate actually lost $150 - so the second option is better here as less money is bet (as we are not betting on interstate teams), yet profited more.

The next scenario we'll look at is to only bet on teams provided there is an overlay, and we rate them a least a certain percentage to win - in this case we'll look at 60%


Return on Investment

• All games: 25%
• Home/Neutral Only: 29%

So in this case both scenarios return around $1,100 profit, but remember that by excluding those team playing interstate we are in effect risking less money to win the same amount (hence the higher ROI).

Finally we'll look at a scenario where we only bet on a team if they have at least a certain overlay over the bookmaker. Again, any figure can be chosen - for this example we'll look at a minimum 34% overlay before we bet on a team


Return on Investment

• All games: 44%
• Home/Neutral Only: 60%

These scenarios return a little less than the previous ones, showing about $850 profit. However with this filter we have risked only half as much money as before - resulting in the massive ROI figures shown above. Also when looking at the home/neutral scenario this appears to have a nice straight-line trend throughout the year, whereas the other scenarios - while constantly growing throughout the year - appear to tail off towards the end.

Regardless of the combination of scenarios chosen, it is clear to see this model has performed exceptionally well in betting over the 2009 season.

Improvements

Despite the success of this model, there is a lot that can be improved on, which will hopefully see an uplift in both tips and profitability.

1. A portion of the model is player-based, hence I need to obtain the players that are selected on a Thursday night for that weekend's games. While I try and update the teams if any late changes are announced, for any number of reason the list of players I obtain may not be the final 22 who run out on game day - particularly for Sunday games. By going back and entering the actual players who played each game we should expect to see a slight improvement in the model's performance.
2. The player-based aspect of the model is fairly simple - I only look at the number of games played and goals kicked for the player's selected, as this info was easy to find and incorporate into the model at the time. However there is a wealth of player-based information available - for example disposals, effectiveness, time on ground, even player ratings such as SuperCoach points - so I will see if adding this information into the model will improve it significantly.
3. This model is built only using data from the 2007 and 2008 seasons, and applied to the 2009 season to arrive at the results shown on this page. By going back a few more years we should be able to bring more information into the model and hopefully improve upon the results.
4. At the moment the home ground advantage in the model is a generic one applied to the whole league. For example, all other things being even, a Melbourne team would have just as much advantage over a South Australian team as Brisbane would over a team coming from Perth. It may be beneficial to look at home ground advantage based on distance travelled, or even give each team their own home ground advantage.
5. Given the scenario above where betting on a team playing interstate isn't profitable, I may look at building two separate models - one model based on both teams being in their home state or at a neutral venue, and one where there is a team in their home state playing a team from interstate. Whether this makes a difference or not I have no idea, but the results will be interesting to see.

So there we have it. I'll still aim to publish predictions for the finals but it is worth noting that these models are only built using data from the home-and-away season so I give no guarantees for the finals. Also in the next couple of weeks I hope to publish my predictions for the Brownlow Medal.

There's probably a lot of foreign concepts introduced here so if anyone has any queries at all please don't hesitate to leave a comment and I'll get back to you.