Neural Networks For Tipsters

[hoopa 0]

Hi all,I'm new to these forums, so I thought I'd introduce myself and explain my little project at the same time. I have to admit that part of the reason for joining these forums was get my system (see below) hosted again after my last host just vanished on me. the fact I could post for my hosting, and that it was an IT forum as well just had me hooked.In my spare time I work on a project that I call Punter. Punter is an artificial neural network that has been designed to analyse AFL (Australian Football League) games, and predict the outcome of the games by learning the patterns in the game data. I'm guessing that Xisto.com is not in Australia, but here what we call tipping (picks in the US) is pretty popular. After studying neural networks as part of my honours thesis, I embarked on this system. It is by no means the only system like this, but it was the challenge for me to see if I could build something to the same capability of the more professional systems.I won't go into the intricacies of this system right at the moment, but suffice to say it's still a work in progress. I would point you to my site to show my results, but it kind of doesn't exist at the moment . Another reason for this post was to spark comment on a system like this, and to see if there were any like minded people trawling these forums, so feel free to let me know what you think.Cheers,Hoopa

[jimmy89 0]

Welcome to Xisto! Welcome to the community, your idea of a neural network for AFL tipping is nice idea. Good Luck! Xisto is not located in Australia, but there are a few of us here (myself one). You idea sounds like a great one, good luck!-jimmy

[hoopa 0]

Welcome to Xisto! Welcome to the community, your idea of a neural network for AFL tipping is nice idea. Good Luck! Xisto is not located in Australia, but there are a few of us here (myself one). You idea sounds like a great one, good luck!
-jimmy

Thanks for the comments Jimmy. I have just put in a request to have my site hosted here, so with a bit of time & luck permitting I should have a little bit up soon to show to people. For the mean time here is a little something that others have found interesting. For anyone who ever watched sport, I expect you may have heard the term "a team is due for a win", meaning that they haven't won in a while and must surely win soon. Well, I went on step further and graphed that concept given past results.

The picture that I have attached shows how often teams win after x consecutive wins. Conversely, the negative side is how often a team wins after x number of losses. The red line is the best fit for that data, and the green line is my rough approximation of what the actual curve would look like (given a long enough period of time)

p.s. I took a look at ur site before the Girlfriend 7.0 helpdesk call is a little gem

[yordan 10]

the green line is my rough approximation of what the actual curve would look like

I'm just curious. Is your green curve a predictive calculation, or is it just a drawing ? Do you have a mathematical description of your best fit, wich parameters becoming usable for long-term predictions ?

[jimmy89 0]

p.s. I took a look at ur site before the Girlfriend 7.0 helpdesk call is a little gem

glad you liked it, I couldn't stop laughing after I read it the first time!

[hoopa 0]

I'm just curious. Is your green curve a predictive calculation, or is it just a drawing ? Do you have a mathematical description of your best fit, wich parameters becoming usable for long-term predictions ?

The green line is hand drawn line is just my take on what the full curve would mature too given enough time (i.e. I think that as the number consecutive wins moves past a point, the amount of wins will start to decrease at an increasing rate, until it reaches 0. It's just really badly drawn because I have no drawing skills (and it was done in a rush to prove a point once, and it's stuck around since then).

 

The only parameter in this equation is the number of consecutive wins by a team at any given point in time (e.g. right now a team might have won there last ten games, so from this you might deduce that their chances of continuing to win are better than their chances of losing. Of course, this is only a very basic model of predicting outcomes. I have trained a neural network on this data and it learns the pattern fine, but it is not necessarily a good indicator of the winning team as there are many other factors to consider. I hope to do more modeling to find more patterns and hopefully by training the neural network on a combination of these models, it can build up a more solid model than any one of them on their own.

 

Edit: yes, there is a formula for the best fit (red line), but it's just something like the average over 3 or 4 values on the x-axis

Edited February 27, 2008 by hoopa (see edit history)

[rockershive 0]

good day. I just want to share or express my views regarding this topic.Is the design of a team's winning or losing prediction includes every possible, inevitable circumstances that may come along in the future and the rest of a team's existence? Let's say, just like in a varsity team, every player or at least 1 player is surely not a permanent player that can participate in all games. Some players for sure are for short span of time only. This, of course can affect, degrade, or even ruin a team's credibility especially if it's a winning team. Let's say I have a very good prediction that this team will win, considering the team's winning streak and history. But unfortunately during a scheduled game a vital player meets an accident, another player is fouled-out, and in another scheduled game your coach couldn't make it for that day for a valid reason. How do you 'justify' your predictions? Or have your ever considered this situational problem to able to make your predictions close to accuracy? I'm sure you do.It's my first time to hear about this kind of project and I like it. I'll be keeping in touch with this thread. Very nice start. Thanks.

[hoopa 0]

Let's say I have a very good prediction that this team will win, considering the team's winning streak and history. But unfortunately during a scheduled game a vital player meets an accident, another player is fouled-out, and in another scheduled game your coach couldn't make it for that day for a valid reason. How do you 'justify' your predictions? Or have your ever considered this situational problem to able to make your predictions close to accuracy? I'm sure you do.

WARNING: The following post might just be a ramble

 

Thanks for your interest in the topic, and that's very good point that you make If I had the answer to that question I think I'd be sitting on a beach somewhere with a nice cold beer. (i.e retired)

 

My system does not take into account the player factor, but instead concentrates on the team performance (at this stage). I tend to think of it as a hive type approach where the team can be measured as a working organism. The reason why I don't take this into account are mainly the overhead of collecting the copious amount of statistics produced by each game.

 

Administration overhead is the main reason, but also a line has to be drawn somewhere as more complex conundrums start to creep in. As you will see in a moment, with the sheer number of possibilities, I have drawn the line a team level learning, with the intention to look closer at player level learning sometime in the future. I have actually done a little bit of work around ranking players based on their games stats, but it is still a very imprecise science.

 

I will be the first to admit what you are saying is true. That players have an massive impact on the results of games. As much as teams would not admit it, the loss of even one key player can cause a teams structure to fall apart, but how do you measure who that key player is?

 

In this sort of system, fact is everything.

You can collect stats, but then how do you work out which ones are important? You can't run them through a neural network as you need to know the outputs to train on. The only indicator that you have of how good a player actually is the opinion of other people, and then that can easily be biased.

How do you statistically work out who was the most important player from a game? A defender may not score much, but he may stop a lot of goals. Is he more of less important than an attacker who scored a lot of goals? This leads to questions of which players stats are important, which are not, and which are important, but which of those are more important than the other important one?

Another complicating factor is injured players. If one of these key players is injured during a game, there is nothing that can predict that (perhaps) and that will always throw your results out.

 

The number of variables in predicting the outcome of a game of football is almost endless. Imagine if you will other factors that could influence a game.

The weather (wind, rain, night, day)

The playing surface

The opposing team/players

The crowd (to a degree, but it would exist)

The mental state of the players

The umpires (humans) and mistakes and biases that might exist

If a team is traveling to play a game

Rest time from the last game

The physical condition of the players (any unknown injuries or pre-conditions)

This is an issue that pains me, as I know there is a gaping hole in my logic, but the ability to solve these sort riddles lies just outside of my fingers.

 

Cheers,

Hoopa

[hoopa 0]

Hi all,

me again

Just to let you know that I have a shiny new web space, care of Xisto. Now that my site is back up again, you should now be able to have a closer look at my results and judge for yourself how well Punter (the name of the system) does.

There are some results for the NFL as well, but that has been discontinued until further notice. For those in Australia, my tips are up for the AFL pre-season grand final, with all my tips being available before the games.

http://forums.xisto.com/no_longer_exists/

Cheers,
Hoopa

[hoopa 0]

Hi all,just a quick update on the accuracy of the system over the pre-season competition, before we get into the 'real' thing this weekend. :)Overall the performance wasn't too bad, but this has be be taken with grain of salt given the reduced number of games in the pre-season competition.11 correct tips out of 15 games (73%). Normally anything that cracks 60% over a whole season is a good result for the current system, but it does have good and bad years. The tips unusually performed lower that the total (4 correct from 7 games - 57%), but the silver tips made up for it, with 4 out of 4 for 100% accuracy.I won't post here too often about the results, but my latest picks are always up on my web site each week before the games.Cheers,Hoopa

[hoopa 0]

Hi all,

I thought that this thread would be the most appropriate place to post my sign off message for Xisto.

I have decided to move my website to another host because of a mutual benefit for both parties, but also partly because of the posting for hosting scheme at Xisto (I'll explain in a moment).

For those that were interested in my work, my new website is now at http://punter.pokerplasm.com/, where I shall continue to post up my tips.

I intend to try and keep posting at Xisto as I have really enjoyed some of the discussions here, as it is a rare thing to have some intelligent conversation for once.

The next bit is to everyone, but specifically to the admins.

Firstly, this is not to say that you are doing a bad job. I think the idea of posting for hosting is innovative and a really good idea for keeping an active forum. Part of the reason that I have decided to stop hosting with Xisto is that I have simply not been able to find the time to post, or the topics to post against. I appreciate it is a delicate balancing act in how credits are distributed, but I have found that even though I work a 9-5 office job and a a few spare hours each night, it has turned into a chore to upkeep my credit.

I'm not being critical of the system, as I have no constructive comments on how this might be done differently, but I do hope to provide some constructive feedback on my experience for your future reference.

I don't doubt that my website will soon go inactive, but I hope to still check in on these forums occasionally to see what is happening.

Cheers,
Hoopa