A football rating system, discussions, ideas

[mr onemore]

I've decided to post some analysis and ideas I've been using for quite some time now to try and rate football teams. The rating system I'll present here is in its most basic form, but mathematically it has a sound basis and I think with some modifications it has the potential to produce value bets against a bookies odds.

As far as this thread goes, I hope that some members may find a new idea that they can encorporate into their own systems, and who knows, maybe someone will suggest a change to the basic ratings system presented here that might just produce the goods....here'e hoping anyhow. At any rate, I'd welcome any questions, comments and advice. I hope to learn a lot too through this thread...

First of all though I'll start with some theory that I first read about in a mathematical paper nearly 20 years ago!! I've been hooked on ratings and betting ever since, so I suppose I should issue some kind of warning here

[Paul Ross]

Not sure whether you've seen my ratings mr o, but they are pretty good and I like to think pretty accurate. If I could offer some advice, it would be keep it simple, logical and don't add too many variables.

Look forward to seeing them.

[christianu]

Where do you post your ratings Paul, i cant find them but i dont think im looking hard enough?

[mr onemore]

Originally Posted by Paul Ross

If I could offer some advice, it would be keep it simple, logical and don't add too many variables.

Give me a chance

Anyway, I agree in total with you. I intend to present the info very general, 'laymans' terms, although there will be some simple, school type maths that will crop up from time to time - unavoidable...

[mr onemore]

First of all then, I'd like to say what I mean exactly by a football rating system? (btw, these ideas apply to all sports...)

Well, I'd say it's about attempting to quantify the strengths of the teams, giving some kind of numerical value to the strength. You measure past performance in a numerical way.

I think everyone does this when they make a prediction or bet anyway - just with a rating system you actually use a bit of maths and numbers instead of using gut feelings.

After you have a rating system, you can then use it to predict future chances of teams beating each other - anyway, that's a bit later...

Back to the ratings. In order to measure this nominal 'strength' - you need to choose some kind of measure which you think corresponds to the respective strengths of the teams. This can be called a metric. This is where most of the fun is, as there are any number of different kinds of metric to choose from. I use a very basic one later on, but you can use almost anything you like if you think it has a direct bearing on the strength of respective teams. There are plenty of good ideas for metrics to be found on PL already.

Now here's a big point:

The fundamental premise of any rating system is this:

A team with a higher rating will have more chance of beating a team with a lower rating.

Of course, 'will have more chance', does not mean 'will definitely' beat, and that causes a few problems to maintaining a useful rating list. Anyway, there are some good mathematical ways of coping with this, coming up later on!!!

After a match, most rating systems will need updating in light of the actual result that occured, as opposed to what the rating system predicted. There is a lot of scope for experimentation here, but again, there are sound mathematical ideas that can be used here (fairly simple ones!) - and I'll open up discussion on those too later on...

Finally, and probably why anyone might be reading this, at the end of the day there might be a way to use the rating system to profit from the bookies - and there will be quite a lot to discuss at that point too!

Feel like I'm beginning to ramble a bit here, so better move on to the first bit I really want to discuss and then leave it for a bit...

[mr onemore]

Well here's the first big problem for any rating system - and it's unavoidable.

I'm going to use this very simple metric, that the ratings depend on how many goals one team is better than another.


Let's say Team A beats Team B by 2-0

Then later, Team B beats Team C by 3-0

Team C seem the worst, let's give them a rating of 0

The metric will produce a rating list like this;

Team A 5
Team B 3
Team C 0

I hope that sounds reasonable so far.

If that's all there is to go on then from any rating system, by our fundamental premise, will might predict that

When Team A plays Team C, Team A will win by five goals.

After all, the metric used here is to measure strength in relation to how many goals a team is better than another.

If this happens, we keep the rating list, and confidently use for a later encounter between the teams.

Of course, and this is why we like football after all, that's unlikely!!!!

What if this happens...

Team C beats Team A by 1-0

How do you update the ratings now? What would the rating list look like?

Well, that's the gist of the problem. As I say, there is a good mathematical way to solve this problem - I first saw it over 20 years ago - and it's used extensively by all the serious ratings lists now produced.

I'll post it up later - in the best traditions maybe those that haven't seen this before might like to have a go!!

[CJ Mars]

I presume you’re familiar with the various books by Paul Steele Mr.OneMore ? He uses a number of ratings systems chief amongst them being a power ratings system that functions much as you described above. I did a lot of work with two other posters on this for pretty much a season about two year’s ago. Results were mediocre at best. On their own I didn’t manage to turn a profit with them. The biggest single problem that most of these systems don’t address is how you equate a rating to the odds. Its’s all well and good saying that Team A near the top rates 945 and Team B near the bottom 617 so A should win. The question is how likely? Unless you can relate the ratings to an odds line I’m always a bit skeptical. Then again maybe you have collected all the various past ratings to give yourself this info?

[PAULM03]

You've hit the nail on the head CJ.

Any old ratings system will tell us that Arsenal will beat West Brom. What we need it to tell us is if the 1.3 the bookies are likely to offer is value.

[Datapunter]

jumping way ahead here but here's a thought,

if you have a rating calculation, use it on as many past matches as possible, then look at the percentages of accuracy, that may give you an accuracy band, that can be related to minimal and maximal odds, and that can lead to finding value..... but jumping way ahead, thats for the end of the line,

should be very interesting to look at this step by step

[mr onemore]

Yes, creating the odds from the ratings is a significant part of the project, and IMO the choice of metric plays a major part. This part will come later here - but suffice to say that I use past predictions/results to come up with new predictions to make odds lines.

As regards those 'Arsenal are 981, WBA are 657' systems, although those ELO points type ratings have a probability and expectation aspect built into their definition - the points on their own are quite meaningless for a 1X2 odds prediction. The ELO system was initially developed for rating chess players - and the theory only relates to win, draw, lose scenarios. I think most people would concur that in football a win of 1-0 is not the same as a win of 6-0. Adapting an ELO or points based system in this case is a major problem, and most solutions I have seen are superficial at best. Basically, useful information is lost.

Almost all of my recent work has been using goals as a measure of strength, as a metric. They have the advantage that a goal spread can be related to the probabilty of match result, in a way that 'points' systems find it difficult.
If someone disagrees with the choice of metric - fine - that won't negate the methods used here as different metrics can be used. But this is one of the most basic metrics that can also extract info about a performance that has a sound basis for translating back into match odds later on and is easy to use. You could use just 'games won' as the metric, but as I say above I think 6-0 contains more useful info than 1-0.


It's hard to avoid being specific - but some of the points style systems on PL have similar weaknesses to those ELO style systems. I didn't read the Steele Power Ratings book - the few comments I read about it made me quite dismissive of some of the metrics used. Weren't some measures combinations of corners, goals, shots etc. The coefficients used (I don't know them, but like 5 times goals plus 2 times shots plus half times corners etc) would seem to be arbritary at best. It's like adding 2 Kilograms to 3 metres to find out how loud something is!! Apologies if I'm way off the mark here, as I say I didn't read it - but I wouldn't recommend working with a metric if it was one described like that.

I have read several mathematical papers that discuss these ideas though - and there is a decent Mathematical way of combining different metrics through a LS approach that I might mention later - it's just not done in that way described above though!! It's also a fine tuning device to start combining several different metrics - best to find a single, good one, first.

[Paul Ross]

Where do you post your ratings Paul, i cant find them but i dont think im looking hard enough?

They're probably on the 2nd page of Britain forum, I'll be updating them tonight.

[muppet77]

a very interesting discussion indeed.

i worked with cjmars testing steele's 'power system' (inconjunction with 2 other systems) and he does actually use the scores for the matches in his system (rather than just WDL), as you said a 5-1 battering gives more info than a 1-0.

[Paul Ross]

I don't think you can use just scores though, my ratings take into account attacking play, this is the key element imo. This way a team that is drawing 0-0 for the last 3 matches but literally smashing the door down without scoring, is clearly playing well and about to go on a little winning run and probably at decent prices.

Take a look at some of the previous ratings. I've seen teams win 3-1 but the other team has beaten them on the ratings. Then when you read the match report, more often than not, it does say the other team was hard done by.

[mr onemore]

Well, it's interesting the comments being generated already.

I'll have to admit having insufficient knowledge of the 'Steele Ratings' to be able to comment properly on them.

As for your comment Paul saying you cannot just use scores as a metric - that's possibly true. I'm just going to illustrate the whole method with this simple metric to start with - other metrics can be used in place of the example, and there is - a quite complex - way to combine several metrics in a sound way. I'd always start with a single metric first though and leave the combinations until much later in a model.

As I stated at the beginning - the example I give here - well it was one of my first attempts, a few years ago now - and although the results are quite good - they weren't quite good enough. I tend to agree that the metric used is not quite good enough...

I myself use slightly different methods now - but for anyone new to this kind of thing I hope the basic method I present here will be enough for people to get started.

Actually Paul, having read your thread on performance ratings - I think you have potentially a very good metric there - and I think it would be very interesting to see the results of your metric in the system to be presented here. At present it appears that you just average the last four performance ratings? Is that right? If that is the case then that would mean that the strength of the opposition in each game is not factored into the rating - but that is taken care of in the method to be presented here. You can then use past performances and outcomes to find probability odds for a future game.

[Paul Ross]

Actually Paul, having read your thread on performance ratings - I think you have potentially a very good metric there - and I think it would be very interesting to see the results of your metric in the system to be presented here. At present it appears that you just average the last four performance ratings? Is that right? If that is the case then that would mean that the strength of the opposition in each game is not factored into the rating - but that is taken care of in the method to be presented here. You can then use past performances and outcomes to find probability odds for a future game.

It is based on the last 4 matches - not sure why you believe the ratings aren't accurate because of that. You'll need to elaborate a bit.

[mr onemore]

Well, why not 5 matches, or 17 matches...have you tested to find out how many games will give the best use of your metric as a forecasting tool. A lot of people tend to use 6 games - why?? What is the basis for that. Not sure there is an answer to this question - you can certainly optimise a model - keep trying different lengths of averages - until your predictions improve. But that goes against my own 'rules' for formulating a model - a form of 'data-mining' almost. I mean, what if it turned out that the best results came from averaging the last 11 games for home teams, and the last 3 games for away teams? It wouldn't feel right. No, much better is to have a reason for why you want to use a certain number of games and take it from there.

This also opens up a discussion of the 'long term class' of a team, versus the 'current form' of a team debate - which I think is best left till later. (For the record, I think most punters over-estimate the effect of very recent form, mistaking recent over-performances as a real change in class of the strength of the team)

Also, by averaging the performances of only the last 4 games, some teams may have had easier games than others - now there is a way to factor that into the calculations which can be shown later.

I have looked at trying a metric like yours before - I do believe like you that a team that has been 'hammering at the door' but maybe not winning, is still in great form. If you want, later on we could have a look at your performance ratings in the model I'm going to present here - it should be interesting. What I am saying in the thread is that you may very well have an excellent metric, but maybe it could be used more effectively as a rating tool - you are only using a part of the whole method...

[PAULM03]

i'm almost drooling with anitcipation. Are you ever going to reveal what your ratings system is?

[mr onemore]

It's no secret - honestly, lots and lots of people use it. Just wanted to go through it step by step - if you want to use something properly, then best to understand it a bit, discuss it...

Sorry - it's all my fault - I am a teacher by profession after all:o

[Paul Ross]

It should be based on form not all your matches, how else are you going to know their current plight ? I also have league tables for home and away matches, to show the whole season as it were.

You're digging too deep in my opinion - I've seen a lot of football ratings, and without sounding like I'm blowing my trumpet, I'll put mine up against anyones with a wager, that's how much confidence I have in them.

[PAULM03]

why not just use the league table?
That's got matches won and lost and goal difference.

I'm not being glib, it's a serious question.