ELO Ratings and Probabilities

[zabadac]

I 'm a new member and I 'm interested in the use of elo ratings to predict football results.
I encountered them first on this site:

http://eloratings.net/about

Using their elo ratings and a sample of matches I made some calibrations to extract a mathematical formula that transforms the elo difference to probabilities for home, draw and away.
That was back in 2014, for the Brazil world cup.
I can write the formula down for you if you are interested, but I also want to discuss with you the probable merit of such formulas.

First our objective should be to locate value bets, meaning matches in which the bookies pay more that fair.
So if the probability based on the elos works out to be 50% and the price is 2.50, it is a value bet.
If the price is 1.80 then it is n't.

Naturally you will say this to me (or you should ask !):
When the price is 2.50 did you check out that nothing unusual is going on, such as key player injuries or many substitutes in the squad or something like that ?
Of course I must do that, is my answer and suppose things are ok.

Another thing you should ask is "ok, but is your formula true ?".
Fact is anyone -human or computer program- can say "the home win probability is 50% and hence and so on and so forth" but is it true ?
How can we be sure ?

This is where information theory comes into play.
My formula should have higher information content or at least equal to that of the bookies.
When you see bookie prices x-y-z then you convert those to probabilities by doing this:

home = (1/x) / (1/x +1/y + 1/z) = P1 say
draw = (1/y) / (1/x +1/y + 1/z) = P2 say
away = (1/z) / (1/x +1/y + 1/z) = P3 say

Then suppose the match is played and the result is a home win.
This contributes to the bookies predictivity the value A1 = P1.
Another match is played and the result is some other result and adds the value A2 = the P2 value of the new match say.
The compound bookie predictivity from the two matches is A1 x A2.
After many matches it becomes Q = a1 x a2 x a3 x ... x An
Your computer cannot carrry out this operation because the product of many such numbers that are less than 1 will become zero.
But you take logs and you take the average of the log sum:

I = SUM OF LOG(Ai) / n

You then take the exp of I, to end up out with some familiar numbers (percentages):

Predictivity = exp (sum of log(A) / sample size)

The predictivity ranges from 1/3 to 1.
1/3 means no predictivity really, random guesses.
1 means he knows everything in advance

The usual bookie predictivities are of the order of 37-39% in the major European leagues.

So, all this talk about predictivity to verify that my formula -or your formula- based on the elos is true or not.
My formula has a predictivity value too, computed in the same way as for the bookies.

Must hold therefore:  Porgnosticator's formula predictivity >= Bookie predictivity

Otherwise the formula is suspect, can't be considered true.

Now what did my formula do ?
My measured predictivity was 37.5%.
So maybe it works - just maybe.

I started with the elo ratings and ended up with a probability calculator.
But there are more sites offering elo ratings.
Yours for example and also the known http://clubelo.com/.

I don't have predictivity values for all of them -they don't publish or they don't know how to do it.
So what I 'm searching is which one is the best ?

There is also the basic theory of the elo ratings.
This is described here:

http://eloratings.net/about

This looks ok but have the various coefficients used in it been computed with the scheme I describe above.
To make the predictivity -looking backwards- maximum ?

These are the issues I am confronted with and I 'd like to have your opinion.

There are some oddities I have observed.
One that striked me was during the 2018 world cup.
In the beginning the elos showed Brazil and Germany at the 2000 mark. France at about 1900 and Belgium at about 1800.
That's ok you could say. France won it, it was n't out of the blue really.
But I 'm talking about Belgium. 
Belgium played very well and nearly won it (n.b. the true final was the semi France v. Belgium).
The thing is Belgium 200 pts below the favourites counts as a big surprise.
So ok surprises happen you might also say and I might agree.
But was it a surprise ?
All the pundits were rooting for Belgium before the start of the tournament.
Brazil-Germany first, ok, but their third choice was Belgium.
So we have a case in which the elo values are ignored by the pundits and the pundits are right (as Belgium was a great team).
This kind of thing speaks against the elo theory imho.

What are your views ?

 

Edited March 11, 2020 by zabadac

[Sir Puntalot]

 to Punters Lounge @zabadac   

[zabadac]

Yesterday with my formula:

PSG - Borussia 48-26-26, no value bet
Liverpool - Atletivo  73-18-9, value bet but at low price (L'pool won in normal time).

You could call Borussia's away win a value bet if the price was over 4.00 (3/1) - I don't recall.
But I generally don't go for third chances value bet or not.

[zabadac]

14 hours ago, Sir Puntalot said:

 to Punters Lounge @zabadac   

Thank you for the welcome.
I don't seem to quite understand how you are doing things here, so let me try you with examples.

Let's take Paris Saint Germain and let's look at clubelo.com first who are using the classical elo theory.
PSG are playing Nice at home next.

For PSG he says:

Match -1, ELO = 1875, beat Dijon 4-0, goes to 1877
Last match, ELO = 1877, beat Dortmund 2-0, goes to 1895

For Nice he says:

Match -1, ELO = 1592, drew against Bordeaux 1-1, goes to 1593
Last match, ELO = 1593, beat Monaco 2-1, goes to 1605

So we have now PSG = 1895, Nice = 1605.
It's a big difference, 295 points, so the prob. of PSG winning looks like a 95%, but that's not important. Even the 95% values we should be able to compute with accuracy.
 

Now in your French Ligue 1 page you give two sets of figures:
 
PSG 71.25 and 76
NICE 39 and 37

and we are supposed to judge things from those.

It's a different scale you are using obviously, but what are the two numbers for each team and how are they derived ?
You are talking about prediction ratings and performance ratings but I have n't quite understood.
Is the prediction rating the classical elo expressed in a different scale to the usual and is the performance rating something else ?
How can you follow all the teams of the world re. performance and how can you judge ?

Re. performance let me give you an example where our opinions probably clash.
Chelsea beats Everton  4-0 last Saturday (a match I watched). I believe Everton was very poor in defense rather than Chelsea improved.
I don't know how to rate Chelsea exactly -other than by the classical (Russian) method- but not really.

Can you explain to me your figures a bit further and are you prepared to go into the business of detailed calibration ?
My objective is of course to have the 37.5% figure I quote in the first post beaten.
We may call the 37.5% "the clubelo performance figure" though it's not really their official performance figure, it's mine, as an interpreter.

 

Edited March 12, 2020 by zabadac

[zabadac]

My estimates for today (Europa league):

EINTRACHT - BASEL 60-22-18 
BASAKSEHIR - KOBENHAVN 57-23-20
LASK - MUTD 26-29-45
SEVILLA - ROMA 57-23-20
INTER - GETAFE 48-26-22
OLY - WOLVES 35-30-35
RANGERS - LEVERKUSEN 21-26-53
WOLFSBURG - SHAKTHAR 41-29-30

Some of those are posponed.

Looking at bookie prices also, marginal value bets are:

SEVILLA
OLY (but difficult)
LEVERKUSEN

Also I like EINTRACHT with hcp -1, BASAKSEHIR with hcp -1.

 

 

[zabadac]

All postponed due to virus.

[Sir Puntalot]

@zabadac Everything is fully explained here https://www.punterslounge.com/elo-football-ratings.  

[zabadac]

1 hour ago, Sir Puntalot said:

@zabadac Everything is fully explained here https://www.punterslounge.com/elo-football-ratings.  

So in the example Newcastle v. Leeds you rated the performance of Newcastle as a 54 and the performance of Leeds as a 15.

So explain to me PSG v. Nice now,
You say it's 71.25-76 versus 39-37.
Is that it ?
And as for what it means re. the probable outcome one is supposed to understand, once he has the feeling of the numbers. Correct ?

I could work out a probability scale for you though if you give me data.
Also handle two ratings (as you seem to work with two ratings) using conflation algorithm.

Edited March 12, 2020 by zabadac

[Sir Puntalot]

1 hour ago, zabadac said:

So in the example Newcastle v. Leeds you rated the performance of Newcastle as a 54 and the performance of Leeds as a 15.

So explain to me PSG v. Nice now,
You say it's 71.25-76 versus 39-37.
Is that it ?
And as for what it means re. the probable outcome one is supposed to understand, once he has the feeling of the numbers. Correct ?

I could work out a probability scale for you though if you give me data.
Also handle two ratings (as you seem to work with two ratings) using conflation algorithm.

The 2 figures are based on last 4 home games and last 2 home games as an average, same with Nice, but their last 4 and last 2 away games as an average. The "Difference" column is home team average minus away team average, giving a Prediction figure. If in green, that means home team has the advantage, in red, away team.

Performance ratings are based on what actually happened after the match has been played, giving a unique insight rather than just seeing a scoreline.

Example: https://www.punterslounge.com/elo-football-ratings/england-premier-league-match-ratings/performance

March 7th - Arsenal beat West Ham 1-0 but the performance ratings clearly show that West Ham were unlucky to lose this match as they were rated higher.

We're quite happy with how the ratings work and many punters on here have asked me not to change anything, so I appreciate the offer but they will stay as they are.  

[zabadac]

46 minutes ago, Sir Puntalot said:

The 2 figures are based on last 4 home games and last 2 home games as an average, same with Nice, but their last 4 and last 2 away games as an average. The "Difference" column is home team average minus away team average, giving a Prediction figure. If in green, that means home team has the advantage, in red, away team.

Performance ratings are based on what actually happened after the match has been played, giving a unique insight rather than just seeing a scoreline.

Example: https://www.punterslounge.com/elo-football-ratings/england-premier-league-match-ratings/performance

March 7th - Arsenal beat West Ham 1-0 but the performance ratings clearly show that West Ham were unlucky to lose this match as they were rated higher.

We're quite happy with how the ratings work and many punters on here have asked me not to change anything, so I appreciate the offer but they will stay as they are.  

 

Arsenal then are approaching the match against Brighton with 27.5-24.5,
Those are what ? Prediction ratings ?
And Arsenal against West Ham have been rated for performance with ?

If you do have two separate kinds of rating (prediction and performance) they can surely be combined to a single -stronger- rating.
You say Arsenal did not play well, ok, we accept that, but if you give them a performance rating other than what comes out of the scorelines then that too is 
a predictive parameter. Where are the preformance ratings for our two future opponents Arsenal and Brighton ?

I don't mean to say you should change the ratings. I have n't even familiarized myself with them in the first place !
Perhaps the way you made them is accurate enough.
What I mean is derive a probability scale from them like I did for eloratings.net and clubelo.com.

The probability scale gives me a direct relation to the price.  So if the probability is 40% then I know that the price must be 2.50 or higher - otherwise it is a 
lossy proposition.
---------------------------------

I mean I could use the page you directed me to.
I pick one at random:

Oct 27 Norwich City 13, Man United 90, result 1--3

Are those ratings, 13 and 90, drawn before the match or after ?
We want before - then it is easy to make the probability formula.

 

 

 

Edited March 12, 2020 by zabadac

[Sir Puntalot]

I've explained it, I'm not going to do it again.

[zabadac]

14 minutes ago, Sir Puntalot said:

I've explained it, I'm not going to do it again.

Ok the last question.
From your past reasults the first one: LEICESTER CITY (65) - ASTON VILLA (-4), result 4-0.
Those are NOT the ratings before the game.
I can't do nothing with them.
Can't I see a table of results with the prediction values ?

[Sir Puntalot]

2 minutes ago, zabadac said:

Ok the last question.
From your past reasults the first one: LEICESTER CITY (65) - ASTON VILLA (-4), result 4-0.
Those are NOT the ratings before the game.
I can't do nothing with them.
Can't I see a table of results with the prediction values ?

Not currently, only before the match is played. It's on the list for the development team but not priority.

[zabadac]

Just now, Sir Puntalot said:

Not currently, only before the match is played. It's on the list for the development team but not priority.

I see.
But please do it, it will be useful.

[zabadac]

Except for the Italians all the others played today.
But European soccer is over as of tomorrow - European nations cup moved to 2021.

Eintracht and the Turks nearly destroyed me.
Oly saved me. The match is not over but my bet has won.

 

 

Edited March 12, 2020 by zabadac

[Sir Puntalot]

I would stress this point about the stage of the season we're at too. The ratings are at their most effective early to mid season in my opinion. This stage of the season throws up many shocks because a lot of teams are playing for nothing, so there's no pressure. You also get teams that are very low odds to what they should be because of the motivation factor that the other team doesn't need to win as they're safe, a good number of these big favourites tend to fold too.

It's basically silly season and you have to be careful.  

[zabadac]

17 hours ago, Sir Puntalot said:

I would stress this point about the stage of the season we're at too. The ratings are at their most effective early to mid season in my opinion. This stage of the season throws up many shocks because a lot of teams are playing for nothing, so there's no pressure. You also get teams that are very low odds to what they should be because of the motivation factor that the other team doesn't need to win as they're safe, a good number of these big favourites tend to fold too.

It's basically silly season and you have to be careful.  

In my experience the opposite is true.
The early season is full of pitfalls.
Historic examples: France lose to Iceland in September 1998, soon after they won the world cup. Greece lose to Albania in September 2004, soon after they won the European cup. Other historic example was Leicester who came out of nowhere to win the league three years ago.
There is another old statistic of mine. I don't know if it stll holds, it's from the 80s. The teams were finishing level in the first half of the season, 5 or 6 of them, then in the second half of the season one or two distanced themselves to contest the championships.
About motivation it's always a must check when we make any bets whatsoever. Would you back Man. City in the return leg against Basel last year, after they had won by four goals to nil in Switzerland ? I don't know about betting against City, but the match was going to be a silly affair. It was obvious.
Of course if you know something in advance in the early season about a team that is about to improve a lot, this will pay you.

About the ratings -everybody's ratings- I think something is missing.
A mystery correction factor.
I wrote about world cup 2018 in my first post.
So everybody fancied Germany, Germany failed. That was a failure of the elo prediction theory but what stunned me was the other failure, Belgium.
Belgium's rating was way below that of the favourites.
Why did the pundits then predict Belgium - who as I said performed exceptionally well ?
I mean the post, sportinglife, BBC and all the other pundits.
Then Liverpool 200 points agead of Atletico. I don't believe that either. 100 maybe.

But as I touched on the subjext of motivation here is something you can do:
Make some statistics of teams who played without motivation (demonstrably so). Does it pay to oppose them ?
I remember a Real Madrid v. CSSKA Sofia match in 2002 I think it was. Real did n't care so it ended in a 1-1 draw.
What makes me remember this match is that somebody gambled all his company's money on Real. Then with a short two barreled gun he tried to rob a bank.
Then he was taken away in handcuffs !






 

Edited March 13, 2020 by zabadac

[Sir Puntalot]

@zabadac I was referring to domestic football not international, the ratings are quite powerful early where even the bookmakers aren't sure which teams are going to be front runners, strugglers etc...  

[zabadac]

17 minutes ago, Sir Puntalot said:

@zabadac I was referring to domestic football not international, the ratings are quite powerful early where even the bookmakers aren't sure which teams are going to be front runners, strugglers etc...  

Domestic differs from international (national teams) in that the national teams make no transfers.
I already observed that the early season offers advantages - if you know the teams that made good transfers.
The summer recess is a period where things happen. The players over thirty sometimes reach their limit.
Remember Chelsea's Ivanovitch ? He had a great season in 2014-15, the Mourinho championship year. In June he played for Serbia and he was their best player, everybody praised him. Then in August, in the new season, he could n't move a leg. He was 32 and then he realised he lost his form for good and retired. In my opinion Iva was one of the all time Chelsea greats, alongside Drogba, Lampard and Essien. But that was as far as he could go.
How you are going to correct your ratings during the summer recess when the teams are not playing, is up to you. But if you do it succesfully then of course it pays,
Yet the bet-conservatives I know always say "we never bet early". I 'm somewhere in between.
 

Edited March 13, 2020 by zabadac

[zabadac]

Here is a thing you could do:
From your experience give us an approximate ready reckonner for p.l. rating difference v.probability.

Using my scale of ratings it goes like this:

-200: 20-25-55
-100: 30-30-40
-50: 35-30-35
0: 45-25-30
+50: 55-25-20
+100: 65-20-15
+150: 70-20-10
+200: 75-15-5


 

[Sir Puntalot]

Transfers don't mean anything - they have to play as a team regardless. Our ratings are based on performances not how much the squad is worth. So many clubs spend a shitload of money and do nothing because the team can't gel.

Ratings should be made simple, logical and not too many parameters or they simply will not work.

[zabadac]

22 minutes ago, Sir Puntalot said:

Transfers don't mean anything - they have to play as a team regardless. Our ratings are based on performances not how much the squad is worth. So many clubs spend a shitload of money and do nothing because the team can't gel.

Ratings should be made simple, logical and not too many parameters or they simply will not work.

That happens too - Shevchenko.
But to take any sort of advantage of the early season you have to observe something that happened during the summer. Otherwise ?

I don't know what you mean by simple.
Scientists say that the more Physics you put in the better.
What is true however is I cannot work with data I cannot access.
For instance someone tolds me to make ratings based on the total money value of the teams, in millions of pounds.
Apart from some discrepancies I see in this, where am I supposed to find this thing tabulated ?

Edited March 13, 2020 by zabadac

[Sir Puntalot]

I think we're going to have to agree to disagree @zabadac 

[zabadac]

I have improved the theoretical predictive value of my ratings from 37.5% to 38.5% by restructuring their formula.

 

[Adryan871]

I will not open a topic. And i don't know If someone still follow this forum,or this ELO rating. But i want to ask,what is more important,4Avg or 2 Avg. And how we can verify past results? If i want to bet only when 4Avg and 2Avg is higher then +15 and lower than -15,can i look in the past performance If i could have with this ratings?

[Adryan871]

So,i don't know how useful is this if i can't look on the past results to see what kind of range it would be good.

[agourbet]

@zabadac based on your experience using the ELO rating, have you seen any value in it. Does ELO create any advantage to probability calculation to the player? Thank you