Probabilistic Prediction of Football Matches - A Discussion

[Rushian]

Hello all,<?xml:namespace prefix = o ns = "urn:schemas-microsoft-com:office:office" />

The following may sound (or should I say “readâ€) a bit academic but bear with me. Scanning through the number of different threads / systems, I realized that system picks are usually derived through one of the following three (at least) ways:

• Intuition (or gut feeling or experience): you know the ones, when you spot a price and almost immediately feel that this price represents value. Subconsciously, your brain has assigned a probability to that event happening (say <?xml:namespace prefix = st1 ns = "urn:schemas-microsoft-com:office:smarttags" />Liverpool beating Everton on Sunday); it has transformed that probability to odds and compared it to the prices on offer before deciding whether it’s worth a bet.

• A set of pre-defined rules: systems which are based on rules such as:
  • If a home team that has won more than 3 out of their last 5 home games AND

  • An away team that has lost more than 3 out of their last 5 away games AND

  • Odds for a home win more than 1.50

  • Then bet on the home win

  [*]A statistical model: systems which try to model things like the number of goals scored by the home/away team, or the result itself, or even the probability of over/under in a match using a number of predictors (whatever these may be). This includes anything from simple regression models to neural networks.

  [*]Maybe there are some more so feel free to add …

The thought that occurred to me relates mainly to people developing systems using statistical models or rules. It may also be of interest to people who decide on bets based on their intuition but only if they feel capable of actually quantifying that intuition in terms of probability. So here goes:

What if all those people who have a way of estimating the probability of a Home Win / Draw / Away Win in a set of given matches, come forth and submit these probabilities in a thread so that we compare them across the different systems which are out there. This would have a number of advantages both for the modelers themselves but also for the rest of the punters because:

• Comparing the probabilities may help in finding improvements in the systems (e.g. someone may be consistently underestimating the probability of a draw). Of course one could argue that rather than comparing amongst ourselves, one should be comparing against the bookies prices (which is the ultimate test) but what if the majority of models underestimate (say) the draw, then maybe a complete change of viewpoint is needed, or an inclusion of an additional variable etc.

• The vast amount of information on matches from major leagues (and that would depend on what each of us is doing) which is spread around the different threads would be nicely summarized in a single thread, for the benefit of everyone visiting these threads.

• A possibly improved system could be developed based on an average of all the probabilities could be developed and tested in a thread of its own.

Moreover a couple of general points: probabilities hide enough of the background of any system which calculates them, in case somebody does not wish to share the specifics of his/her system. So if someone wanted to share the variables used in his/her model, that would be great, but not a prerequisite. Finally, this also applies to the people who use rules (see above) in selecting bets since these rules are closely related to what is known as a classification tree. Following a series of decisions (e.g has the home team won more than 3 of their last 5 home games: yes/no etc…), the number of games under consideration is reduced, until you reach down at a set of games which fulfill all the criteria. The probability estimate of a home win / draw / away win, is simply the number of home wins / draws / away wins out of those filtered matches.

So an example match would be:

Date	Home	Away	System / Model	Home Win (%)	Draw (%)	Away Win (%)
30/3/08	Liverpool	Everton	A Model	52.0%	25.0%	23.0%
			Another Model	48.0%	38.0%	14.0%
			A Third Model	36.0%	29.0%	35.0%

Anyway, I would be delighted if you could share your thoughts on this. Apologies if this is not the right forum to post it on, mods feel free to move/delete it.

Rushian

[lunatism]

Re: Probabilistic Prediction of Football Matches - A Discussion Hi Rushian! You've put to the forum a suggestion on many a system users mind I guess - a way to combine the cream of all systems out there into one super-system that can hopefully do the business better than any one single system could ever hope to. And it is a GREAT idea, and I would of course be willing to offer my help on this if at all possible. I've always thought that the best way to fully predict the outcome of a game you need to take into account as many quantifiable factors as possible, and so that is what I do, there are soooooo many different things that happen to the raw data I import before the system shoots of the LSS selections in the form of "information". There are a lot of guys on the forum who use solely intuition, while for others, it's a mix of such and stats. I'm of the belief that it's possible to produce success without human intervention once you have spotted the patterns and taken the relevant steps to "codifying" it (for me...meant a few sleepless weeks making the most automated of spreadsheets to do all the donkey work). With regards to this, I wonder, do you already have any preconceived opinions about the "measure" of intution against, "rules" or statistical models? I ask this because I guess it will make a difference to how you view information generated by one system compared to another. As I explained at the start of my thread, I used a number of systems to generate ratings/percentages/scores - all successful at looking at certain aspects of a team. All show similar overall strike rates. But you are only interested in the percentages, right? And which leagues?!! Big question here - i look at almost every world league, and i have the past percentages for all of them! So if you said post everything, that would be one HUGE list daily, so i guess that wouldn't be feasible really. It will no doubt be a challenge to accomodate all systems (if you can get a lot of people interested in this idea), but it should be "fun". Are there any patterns in particular that you are looking for?

On your last point, i think i speak for many who generate percentages by saying that it's often the way to keep it hidden. ESPECIALLY if you would be generating such stats for laying - as a few people know how to do it and you find you can't get your bets matched at reasonable prices.

Right - onto the Liverpool Everton game...here's what my system says (the same processing that makes the LSS stuff):

Date	Home	Away	System / Model	Home Win (%)	Draw (%)	Away Win (%)
30/3/08	Liverpool	Everton	LSS	30%	46%	24%

right, there's my input - let me know what you want from me! oh ye - just wonder - have you developed a system that you will be posting details from too? Or are you just the "compiler"? :D