ELO ranking using expected goals model

[The Analyst]

Hi all,

I'm new to the forum but not to the betting itself. I've been betting for several years now with some positive results long term now. Currently I have some tools that enable me to get the data for expected goals for various leagues after each round of games which enables me to use another Poisson tool to model my own probabilities and look for value bets. It has served me vey well in past few years and the results were very good overall. Over the summer I'm planning to upgrade it even further. I'm thinking about building an ELO ranking based on expected goals and odds. What I mean is that I want a ranking that starts of as a typical ELO ranking but then it gets adjusted based on games between the teams and their expected goals scores plus the odds for each team. So for example if team A plays team B and their closing line win odds are say 1.65 and 5.5 then depending on their corresponding expected goals produced in that game their rankings would change accordingly. I want to make this upgrade to even further improve my current tools as I don't like to stand still . It should negate any potential problems with favorable fixtures for the teams for example. I have all the data needed including tools to get expected goals values and also odds for the games as well so that is not a problem. I just wanted to know if anyone has done something similar or if they have any ideas on what would be the best way to implement it. Thanks in advance.

[Sir Puntalot]

 to PL @The Analyst   

I've moved this to systems and strategies for you. We have our own ELO Ratings right here too https://www.punterslounge.com/elo-football-ratings which are due an update in the very near future.  

[The Analyst]

Hi,

Thanks for moving this to the right place. Hopefully I'll get some ideas from the members here. I've had a quick glance at your ratings, looks like you use a combination of shots/shots on target possibly corners judging by the numbers . I've always like the idea of going beyond the results and looking deeper into performances hence my interest in applying the expected goals. I've got all the basics of it , got the data etc. and it has served me well but I want to take it even further by also using odds and elo as I think sometimes team A can play against say 4-5 strong teams and team B has had an easier schedule so say both have similar expected goals data yet the one team managed that against much weaker opposition then the other one so this needs to be adjusted accordingly. I've got a faint idea of what I need doing but I wanted to hear some input from the members who possibly have had some similar experiences or they can provide some input on this from a different angle.

 

[Sir Puntalot]

2 hours ago, The Analyst said:

I've got all the basics of it , got the data etc. and it has served me well but I want to take it even further by also using odds and elo as I think sometimes team A can play against say 4-5 strong teams and team B has had an easier schedule so say both have similar expected goals data yet the one team managed that against much weaker opposition then the other one so this needs to be adjusted accordingly. I've got a faint idea of what I need doing but I wanted to hear some input from the members who possibly have had some similar experiences or they can provide some input on this from a different angle.

 

I somewhat agree with the harder/easier schedule but I also believe the more you try to perfect ratings it starts to go wrong. Some people say possession should be in there, but I 110% disagree. The hard/easy schedule is also very hard to put into play because it's almost impossible to weight.

Keep it simple, understand what you've created and keep it logical.  

[giraldi]

I am very interested in rating systems and I created my own tools also. Tools to create other tools actually.  So I could say   I have a little experience in the field.

I am sure we will continue the discussion but for now just a few thoughts about your ideas.

If the ELO system is built correctly, it shows relative differences between teams and it already takes into account the opponent strength when calculate the points given after you know the results.  No need for other adjustments. Just need to apply the original formula. As  Sir Puntalot already said keep it simple.

If you think to consider other elements as goals scored, possession…etc my opinion is that would be wrong and will affect random your rating.  For a rating system ELO type should be used just final results, maybe goals difference to have a little influence. Man City has a big rating because won against top teams not because won 5-0 against weak teams.

For example let-s takes ball possession and shots on goal.  Can someone to tell what is the trend in the following graphs?

First one presents ball possession and second one shots on target, only for home games, for Man,City, Tottenham and West Brom  

Maybe experts will see trends but it looks completely random to me. Why to add a random component in an ELO type rating system?

I hope it was clear what I presented (I hope I translated correctly my thoughts to English) and really hope will continue this interesting discussion

 

 

Edited May 5, 2018 by giraldi

[The Analyst]

Hi Giraldi,

Thanks for your input and thoughts. I would tend to agree with your first statement that possibly when use ELO and the teams are ranked accordingly at the start then we could possibly skip the odds part and the team's relative strengths would be already within the ELO ranking. 

The second issue about including extra data I would have to disagree as I'm a firm believer in it. I think that using just points and goals for any rating is not enough to get a proper judgement about teams performances. I have read and studied many papers and myself done lot of research on the subjects and from my own past results which span for several years now I can confirm that you need to look beyond just the goals and points to better rate the teams.

For example one of the better known blogs/sites here :http://11tegen11.net/2015/01/05/the-best-predictor-for-future-performance-is-expected-goals/ 

He puts to the test various metrics to the test including points, goals, shots and expected goals to see which ones are the best for predicting future performances. Expected goals comes clearly on top.

I have used expected goals in my own betting and in last two years I have designed a tool that can use them withing a Poisson model to get the probabilities for each game and find value that way. It has worked very well with good overall profit.

But I always try to improve my work and one thing that I have found during the past years was that sometimes over shorter period of games (4,6,8) games some teams who had an easier schedule would get a good expected goals results whilst other teams who played top sides would obviously struggle but it was rather that they had a tough fixtures rather then paying badly.

I thought that by using some sort of ranking such as ELO this could be improved.

I though that maybe just using expected goals instead of actual goals in an ELO ranking or possibly create "adjusted" goals so a mixture of actual goals and expected goals. What would you be your opinion ?

 

[giraldi]

I totally agree that expected goals are the best option when talking about football prediction.  I just said that ELO rating should be used as it is, with a simple formula, just to see the relative values between teams at that moment. Also said that in my opinion other stats like ball possession, shots, corners are almost unless for predictions.

Talking about expected goals, what is the best way to calculate? Last 5, 10, 15…last season? My opinion is that the best option is to use ELO rating because ELO rating  differences gives the idea what happened in the past between similar teams.

For example Team A  has 2100 ELO points and Team B has 1900 ELO points. There is 200 ELO points between them. It is very easy to see what happened in the same league (this is very important)  when  played  teams with the same ELO difference (at aproximative the same values). Let-s say that there were 200 similar games. Next step is to calculate how many goals were scored by each team and will find something like this

Home teams: 0 goals 25%, 1 goal 30%, 2 goals 15%...etc

Away teams: 0 goals 45%, 1 goal 15%, 2 goals 15%...etc

The percent are not real of course.

With the percent resulted we can calculate probability for each correct score: 0-0, 1-0, 1-1, 0-1…1-2…..etc. then we can calculate the probability for any line,  total goals, handicap… or any other bet we may need.  

Finally we have the probabilities (then the calculated odds) for any possible offer and  will compare them with  offered odds to find value. Maybe it is not perfect but at least we have something to sustain our decision.

Now, what I explained so far is very easy to calculate and to implement. I already create tool to do that.  

What will make the difference is the “fine  tuning”.  As I said, we calculated stats based on past results. There will be some average values based on past but the bookies take our money today so we have to REFRESH the stats based on the last team events.

My opinion is that statistic is just math without the last info and the last info are just noise without stats.  We can talk about.

If we made our homework and found out that the best 2 defenders are not in the best mood because had domestic problems  probably it is not very wrong to consider that the probability for 0 goals scored  by opponent team goes down but the probability for 1, 2 goals will go up. Then calculate again all probabilities for correct score, home win. and will find  other values.   

Or maybe will find that will be heavy rain or other bad pit condition. Again, will increase the probability for 0 goals scored for both teams and so on.

For example, absolutely random Let-s take Liverpool – Brighton  (ELO 1972 vs ELO 1697)

I found exactly 100 similar games past 5 season and the goals were scored as follow:

Home team:  0 goals 5%, 1 goal 22%, 2 goals 33%, 3 goals 22%, 4 goals 13%, over  4 goals 6%

Away team:  43 goals 5%,40 goal 22%, 13 goals 33%, 4 goals 22%,

After calculate correct scores will result all the markets probabilities also. For example

Home team: 76%

Draw: 16%

Away team: 8%

Etc…

The odds resulted: 1.313; 6.254; 12.706 are close to the odds offered by bookies. Now, if we have info,  we can adjust the probabilities for goals scored then try to find a value.

 

Maybe this was too long but I wanted to present you what I understand when talking about rating system.

[The Analyst]

Hi Giraldi,

Thanks again for a very thorough post ,definitely not too long ! I enjoy these kind of debates as there is a lot be learnt from exchanging ideas and views. I totally understand the way you have explained the ELO system and that's why I believe it is something that I need to explore and this is why I've opened this thread.

My main issue is on how to go about using the ELO ranking with expected goals values. At the moment I have data after each match for expected goals for each team. Standard ELO would use actual goals and then award points depending on the actual results and difference in teams. So my initial thought was to replace actual goals with expected goals so for example if team A had Xg 2.3 and team B had Xg of 1.5 then the team A would get the win etc. But I can already see a problem here that for example if any teams wins their XG match by 0.1 or by 3.0 they still get awarded the win and obviously with actual goals you also get draws.

So my moving on with this I though that I could do something like if team A beats team B on expected goals by say more than 0.75 then they get the win and anything between 0-0.75 difference would be classed as draw. Obviously the difference in expected goals of 0.75 is just a suggestion and it could be anything from 0.5-0.9 etc.

There could be also maybe another way of updating the ELO rating based on the expected goals for each team , maybe there are better suggestions out there ?

On the second matter as you are more knowledgeable about ELO than me. What is the best way to start the rating ? Do you just give all teams same starting number say 1500 or 2000 pts and then go from there or would you say look at the league table from previous season or bookmaker odds for the new season and then start with say 1st team at 1500 then second team at 1450 then 1400 and so on ?

Let me know your thoughts and really appreciate your feedback on this.

 

[giraldi]

It is interesting your idea to calculate ELO based on expected goals but I am still thinking what would be useful a virtual ELO for? Or maybe I did not understand? Finally you will have an ELO coefficient that will continue rising for good teams (because they are expected to win the game) . I do not know, maybe I am wrong.

Regarding second matter, I am not an expert, there are around many guys with much more experience. I just can tell you what I did: I set 2000 points (1700 for second league, 1500 for third league...etc) for each team back in 2005  then applied ELO procedure for every game. Of course the promoted teams will enter with the last ELO in the lower league and relegated teams will go down with their ELO.  

It seems to me very natural changes when look at ELO graph for the last 5 years for any team

Of course any ELO rating has a "calibration" time (I think it is about 15-20 games) so if you apply every season a new ELO (even based on the last year table) I do not think it is a good idea.

There will be a calibration time of course each season because of teams changes but this is no more than few games usually. 

[The Analyst]

Hi Giraldi,

Thanks for the answers. The idea behind it is to ultimately be able to predict the games and generate odds for the games. Currently I've been using xgoals and Poisson to generate probabilities etc. and it has worked pretty well overall generating positive results. But I think I can improve it even further using xg and ELO combine to give me another look a things. Similar to the example given I could use the data from past seasons ( I have several seasons worth of xgoals data) to generate the ranking as you suggest and then when the fixture comes up I can see for example how Team with say 1900 pts playing vs Team on 1750pts what were the scores in history and then generate the probabilities as you suggested in one of your previous posts. The main difference is that I would use xgoals and not actual goals as x goals are the best indicator of teams performance and also future predictor.

Just got one question about the ELO comparison when using historical data, let's assume that as in example above Team A has 1900 pts and Team B has 1750 pts. Do you look for teams in the past within say +/- 10,20,50 pts or would you use the difference in pts between teams so 150pts difference +/- 10,20 pts ? Does it make any difference ?

Thanks again for your input and this very good discussion, hopefully other members can give their expertise too.

 

 

[giraldi]

Yes, from my experience, there are completely other results.

May be the same ELO points difference between Manchester City - Chelsea and Stoke - West Brom but the profile of the game will be different also.

Considering your example with 1900 - 1750, I will search for all games where home teams had 1900 +/-  x and away team had 1750 +/- y where x and y are variable.  If set x=50 and y=50 will search for all past games where home team had ELO between 1850 - 1950 and away team had ELO between 1700 and 1800 .

Of course there will be a balance between x,y vlues and the accuracy of the results. It is better to have small values for x,y but will result in a small number of past games. You  will find the best values for x and y after tests.

This is just my point of view. Maybe will be other opinions here.

[The Analyst]

Thanks Giraldi,

I will now try an build my ELO model based on expected goals and see how it will compare to my normal method of using expected goals and Poisson. I'll be using excel for this task , any tips on what's the best way to do it ? If any other members have any ideas or input on this subject , any help much appreciated.