Bookies and team ratings

[cosmicsports]

You all know about club ratings, we discussed them before and we even analyzed various rating systems.
The ones I usually look are from clubelo.com or eloratings.net for the national teams.
But we also talked about other sites and old books using different methods.

The thing is I still look them up but I 'm no longer a fan.
Because the bookies are following them like religion so the game becomes a game of swings and roundabouts.

Consider this coming match:
Basel v. Fiorentina for the European conference league.
It's priced 5.00 - 3.60 - 1.75
The current ratings are Basel 1528, Fiorentina 1733.
Looks logical, hence the pricing.
But Basel won yesterday 2-1 in Italy for the first leg.
So, for the love of the great manitu of the river, are they really rank outsiders now playing in Zurich infront of their supporters ?

There are others similar.
Ok, let me give two more examples out of several:
The tie Union Berlin v. Union St. Gilloise.
That was drawn 3-3 in Germany and St. Gilloise price was 3.20 for the return leg.
Again because of the ratings but I backed the Belgians and they won 3-0.
After that the draw was Bayer Leverkusen v. Union St. Gilloise.
Leverkusen are below Union Berlin in the bundesliga so I opposed them.
The match ended in a 1-1 draw but it was n't a winning bet because it was a double and I lost the other one.
Then the same story for the return leg. Leverkusen were favourites to win away and St. Gilloise were paying some 3.50 !
Well, that did n't work. The Germans did indeed win, by 4-1 as well.

But looking at those three bets together I was on top. 2 correct choices, 1 wrong.

I spotted this kind of behaviour many times and it looks the bookies are faltering in these situations.
I think it can become the basis of a theory.

 

[cosmicsports]

I 've writen about this before as I see now.

 

[cosmicsports]

How do you recognize a wrongly rated football team ?
One such was Leicester City of Claudio Ranieri in 2016.
Another one was Belgium in 2018 world cup.
In the case of Belgium everyone fancied them as one of the contenders even before the matches. Yet their official rating hardly suggested that.
Another one was this year's Arsenal - now they are second clear of the third and lost the title to Man City but that is much more than we expected.
On the negative side Chelsea 2022-23 were surely overestimated before the start of the season.

On the whole it is early in the football seasons that we are blind to such transformations.
But even late in the season there are such wrongly rated teams, underestimated by the bookies or overestimated.

 

[AustinDillon75]

cosmicsports as one who is keen on ratings and their value you make a really interesting point. I have tried for a little while to apply ELO to football and I'd say it can be a bit slow to respond to teams who are improving. Looking at last season's ELO for the NFL, I always find that you can get a pretty accurate portrayal come end of season, but using the Giants as an example, they went from being one of 2021's worst teams to the playoffs this time around and I still could not get them out of the bottom quartile on ELO ratings regardless.

I have not looked at ELO football for 2016 but I recall an interim league table from around November 2015 that had Leicester pretty well top over the course of the previous 12 month period, at a time when a lot of top sides were trying to sort themselves out. I wouldn't be surprised if ELO was quite favourable to them at that stage.

Regarding Basel v Fiorentina I'm not at all surprised by the prices, I wouldn't be in a rush to go after Basel there, surprises happen but not often twice in quick succession. When doing my racing ratings I have a "master" rating that pinpoints a horse's overall quality but on a race by race basis I'm adapting it for various factors including recent form, jockey booking, weight and previous odds. If two teams are 1700 on ELO and one wins its last 4, the other has lost its last 4, that has to enter the thinking but it shouldn't affect the longer term standard of either side whoever wins or loses.

 

[cosmicsports]

cosmicsports as one who is keen on ratings and their value you make a really interesting point. I have tried for a little while to apply ELO to football and I'd say it can be a bit slow to respond to teams who are improving. Looking at last season's ELO for the NFL, I always find that you can get a pretty accurate portrayal come end of season, but using the Giants as an example, they went from being one of 2021's worst teams to the playoffs this time around and I still could not get them out of the bottom quartile on ELO ratings regardless.

I have not looked at ELO football for 2016 but I recall an interim league table from around November 2015 that had Leicester pretty well top over the course of the previous 12 month period, at a time when a lot of top sides were trying to sort themselves out. I wouldn't be surprised if ELO was quite favourable to them at that stage.

Regarding Basel v Fiorentina I'm not at all surprised by the prices, I wouldn't be in a rush to go after Basel there, surprises happen but not often twice in quick succession. When doing my racing ratings I have a "master" rating that pinpoints a horse's overall quality but on a race by race basis I'm adapting it for various factors including recent form, jockey booking, weight and previous odds. If two teams are 1700 on ELO and one wins its last 4, the other has lost its last 4, that has to enter the thinking but it shouldn't affect the longer term standard of either side whoever wins or loses.

The post mortem will tell us about Basel, but from 5.00 they went down to 4.75 and now 4.60.
So I must place my bet soon.

Now when you say "a team lost the last four" who knows if they 're not on the way to some recovery ?
It looks like a shaky indicator.
Hence the bookies were not pricing up Chelsea till after things became really bad - sometime in late March.

The improving team is a better target.

 

[AustinDillon75]

The post mortem will tell us about Basel, but from 5.00 they went down to 4.75 and now 4.60.
So I must place my bet soon.

Now when you say "a team lost the last four" who knows if they 're not on the way to some recovery ?
It looks like a shaky indicator.
Hence the bookies were not pricing up Chelsea till after things became really bad - sometime in late March.

The improving team is a better target.

That's the thing, if they've won their last four to get from 1650 to 1700 I'd take them all day over a time who've lost their last four dropping from 1750 to 1700. That's why I take the master rating as a starting point and work from there. Sometimes no rating will allow you to take account of whether a team is on the road to recovery but ELO is as good as any other in my view.

I hope your bet on Basel succeeds. I don't bet on football but if I did I'd not touch them at any of those prices.

 

[cosmicsports]

That's the thing, if they've won their last four to get from 1650 to 1700 I'd take them all day over a time who've lost their last four dropping from 1750 to 1700. That's why I take the master rating as a starting point and work from there. Sometimes no rating will allow you to take account of whether a team is on the road to recovery but ELO is as good as any other in my view.

I hope your bet on Basel succeeds. I don't bet on football but if I did I'd not touch them at any of those prices.

Well ELOs may not be the best ratings but they 're good enough - they give the logical predictions.
To go from 1650 to 1700 in four matches the way ELOs work you must beat giants really.

I was n't betting on football either even though I 'm a football fan from the age of 8.
I liked the races better but now our racing is a disaster story. Also I learned some more things about football.

 

[AustinDillon75]

To go from 1650 to 1700 in four matches the way ELOs work you must beat giants really.

I'm being hypothetical with those numbers. I agree they offer a very good starting point but as with all ratings there are other dynamics. If it was as simple as the ratings giving the result then we'd never get anything wrong.

At present I am in the forum elsewhere rating Hong Kong racing and its delivering a 26% strike rate which is really good but it would only be 21% if I just took the "master rating".

When time allows perhaps when the HK racing is over I may go back to my old footy ELO sheets and see what can be derived. I think what you want is a clean master rating and a seperate figure that adjusts the master figure for the particular match, using recent form and that old M-word, momentum.

 

[cosmicsports]

I'm being hypothetical with those numbers. I agree they offer a very good starting point but as with all ratings there are other dynamics. If it was as simple as the ratings giving the result then we'd never get anything wrong.

At present I am in the forum elsewhere rating Hong Kong racing and its delivering a 26% strike rate which is really good but it would only be 21% if I just took the "master rating".

When time allows perhaps when the HK racing is over I may go back to my old footy ELO sheets and see what can be derived. I think what you want is a clean master rating and a seperate figure that adjusts the master figure for the particular match, using recent form and that old M-word, momentum.

Maybe.
Meanwhile Basel failed against Fiorentina but the idea worked with Nottingham Forest yesterday against Arsenal (4 to 1).

I 've beeen tinkering with these things before.
I don't think we had ratings in the eighties so I was doing this:
Divide the teams of a national league in five groups A-B-C-D-E.
The A's were championship contenders, the E's were teams clearly on their way to relegation and and there were the B-C-D graduations in between.
So that gave me certain percentage probabilities, fairly similar to the ratings.
Aslo I did this: Make a separate scale for first half of season and second half of season.
With teams of different coutnries the grades I was giving were somewhat ad hoc. So Greek champions were a D or a C and England champions were an A, when drawn together.
That was n't bad.
But I also placed some further boxes in my software app for the user to tick:
Critical absences.
Form and psychology.
But how much modify the A-B-C-D-E probabilities because of form and psychology ?
It was subjective.
Can you make it more accurate ?

Well football was a bit different then.
The venue factor was mighty strong.
I recall in the 70s-80s Greek champions Olympiakos were squeezing a draw against Kavala up north and they were happy, congratulating one another. Nowadays -with Olympiakos still an A team- such a result would n't be liked at all.

 

[cosmicsports]

Mr. AustinDillon75 look at this:

Betting on a buzz, mispricing and inefficiency in online sportsbooks

Bookmakers sell claims to bettors that depend on the outcomes of professional sports events. Like other financial assets, the wisdom of crowds could help sellers to price these claims more efficiently. We use the Wikipedia profile page views of

www.academia.edu

 

[AustinDillon75]

Mr. AustinDillon75 look at this:

Betting on a buzz, mispricing and inefficiency in online sportsbooks

Bookmakers sell claims to bettors that depend on the outcomes of professional sports events. Like other financial assets, the wisdom of crowds could help sellers to price these claims more efficiently. We use the Wikipedia profile page views of

www.academia.edu

cosmicsports thank you. Again proving that this forum has no peers in respect of people sharing thoughts ideas and helpful info together.

Will take some time to read.

 

[cosmicsports]

cosmicsports thank you. Again proving that this forum has no peers in respect of people sharing thoughts ideas and helpful info together.

Will take some time to read.

I have more papers from you.
But if you subscribe to academia.edu they will begin sending you.

 

[ARAZI91]

cosmicsports & AustinDillon75 - All rating systems no matter the methodology behind them will have a limitation on win strike rates on A v B type markets or what is called "fidelity of ranks" in multi-competitor events such as horse racing ,ie "65% to a fidelity of 4" in A v K type markets (horse racing) - that just means that from Ranks 1-4 the ratings have a Win Strike Rate of 65%. Problem is that most of the inputs are data-points themselves. Whether it is a football match, a horse race, a snooker match or tournament or a tennis match, there is information we are not prone to. Most ratings are probably calculated hours before the event so maybe don't or never will have the smarts within the algorithm to "UPDATE" relatively new information.
Imagine a snooker match Final between Mark Selby v Kyren Wilson -Selby has blitzed through the rounds dropping only 6 frames in the process and beating on avg a slightly higher class of player, Wilson on the other hand has got there by sheer grit , some close encounters and longer matches but overall has played well. Overall Wilson has had slightly the better season -higher avg pot success, higher avg long pot success etc etc etc but his form had dipped before coming into this tournament, Selby on the other hand has had a strange season, won 2 ranking events but has had two first or second round exits - the ratings go Selby 901 , Wilson 629 (they are cumulative and run continous from season to season) - translated to decimal odds that makes Selby a 1.70 shot and Wilson a 2.43 shot but the betting public on the exchanges have been took in by recency bias, they have watched Selby blitz every opponent so far whilst Wilson has had far longer and closer games so the market goes 1.48 Selby , Wilson 3.03- Now both players have had a couple of days rest. Now imagine being in Selby's camp (ie Close enough to know him and what is going on behind the public persona - you know that over the last couple of days he has came down with a bad case of food poisoning - in practise he is all over the shop and only got diagnosed yesterday with a course of antibiotics, he's seriously thought about withdrawing but because of his up and down season he decides to play but has admitted that he has no chance now really of bringing even his B-Game to the table. How would you update the Ratings to Odds in this case using this "new" information.
Well we could use a bijection/reduction technique using a degenerate distribution.
We will put our Bayesian hats on now and use for simplicity that Wilson must have a probability of 0.50 or greater of winning now than his original Ratings odds and discount Selby's rampant march to the final before he was food poisoned, whilst adding on 10% making it worth probabilistically 0.55 (New ei) which incidentally will be our Degenerate Distribution in an A v B event with "no draw" - only difference is we will be applying all of that probability mass (0.55) to Kyren Wilson.
So back to the original Ratings and Odds
The whole thing would look like this and we update the original ratings odds by the formula of (1-New ei (in this case 0.55)*(Degen.Dist.)+(original probability from the Ratings Odds)



You don't need an EV calculator to see where the value lies!!
I know it's a "daft" premise but instead of "information" from the inside - as punters of all sorts we can hold back data that we know is predictive from our originally modelling process or selection process - maybe some inputs that are still predictive and hold some significance and are not used in the original ratings (or model,models etc etc) or even some subjective information which we will classify as "new" and where that word "new" is "relative" and "unique" to each punters set-up. This technique can easily be applied to A v B type markets (with/without draws) and also A v K type markets involving multi-competitors like horse racing.

 

[cosmicsports]

cosmicsports & AustinDillon75 - All rating systems no matter the methodology behind them will have a limitation on win strike rates on A v B type markets or what is called "fidelity of ranks" in multi-competitor events such as horse racing ,ie "65% to a fidelity of 4" in A v K type markets (horse racing) - that just means that from Ranks 1-4 the ratings have a Win Strike Rate of 65%. Problem is that most of the inputs are data-points themselves. Whether it is a football match, a horse race, a snooker match or tournament or a tennis match, there is information we are not prone to. Most ratings are probably calculated hours before the event so maybe don't or never will have the smarts within the algorithm to "UPDATE" relatively new information.
Imagine a snooker match Final between Mark Selby v Kyren Wilson -Selby has blitzed through the rounds dropping only 6 frames in the process and beating on avg a slightly higher class of player, Wilson on the other hand has got there by sheer grit , some close encounters and longer matches but overall has played well. Overall Wilson has had slightly the better season -higher avg pot success, higher avg long pot success etc etc etc but his form had dipped before coming into this tournament, Selby on the other hand has had a strange season, won 2 ranking events but has had two first or second round exits - the ratings go Selby 901 , Wilson 629 (they are cumulative and run continous from season to season) - translated to decimal odds that makes Selby a 1.70 shot and Wilson a 2.43 shot but the betting public on the exchanges have been took in by recency bias, they have watched Selby blitz every opponent so far whilst Wilson has had far longer and closer games so the market goes 1.48 Selby , Wilson 3.03- Now both players have had a couple of days rest. Now imagine being in Selby's camp (ie Close enough to know him and what is going on behind the public persona - you know that over the last couple of days he has came down with a bad case of food poisoning - in practise he is all over the shop and only got diagnosed yesterday with a course of antibiotics, he's seriously thought about withdrawing but because of his up and down season he decides to play but has admitted that he has no chance now really of bringing even his B-Game to the table. How would you update the Ratings to Odds in this case using this "new" information.
Well we could use a bijection/reduction technique using a degenerate distribution.
We will put our Bayesian hats on now and use for simplicity that Wilson must have a probability of 0.50 or greater of winning now than his original Ratings odds and discount Selby's rampant march to the final before he was food poisoned, whilst adding on 10% making it worth probabilistically 0.55 (New ei) which incidentally will be our Degenerate Distribution in an A v B event with "no draw" - only difference is we will be applying all of that probability mass (0.55) to Kyren Wilson.
So back to the original Ratings and Odds
The whole thing would look like this and we update the original ratings odds by the formula of (1-New ei (in this case 0.55)*(Degen.Dist.)+(original probability from the Ratings Odds)

View attachment 134895

You don't need an EV calculator to see where the value lies!!
I know it's a "daft" premise but instead of "information" from the inside - as punters of all sorts we can hold back data that we know is predictive from our originally modelling process or selection process - maybe some inputs that are still predictive and hold some significance and are not used in the original ratings (or model,models etc etc) or even some subjective information which we will classify as "new" and where that word "new" is "relative" and "unique" to each punters set-up. This technique can easily be applied to A v B type markets (with/without draws) and also A v K type markets involving multi-competitors like horse racing.

Selby went down from 2.15 to 1.48 because of his opponent's food poisoning.
In another case it was the striker and the model - she left him a few days before the match - consternation.
These things may or may not have a bearing on the result but should be taken into account.
The problem is how do we somehow quantify them ?

* Should you subscribe to this logic, another thing is that you have to bet on teams / competitions for which you have adequate stream of information. You must give up the Chinese-Japanese-Australian bets you make.

 

[ARAZI91]

Selby went down from 2.15 to 1.48 because of his opponent's food poisoning.
In another case it was the striker and the model - she left him a few days before the match - consternation.
These things may or may not have a bearing on the result but should be taken into account.
The problem is how do we somehow quantify them ?

* Should you subscribe to this logic, another thing is that you have to bet on teams / competitions for which you have adequate stream of information. You must give up the Chinese-Japanese-Australian bets you make.

Yeah but it does not need to be "new information" from the "inside" - you may have a rating or a metric that you have tested , tweaked , refined etc until it's calibrated to a very high standard - Deliberately holding that back from the original modelling/selection process can be used to the same effect . That was why i said that the term "New" should be "relative" and "unique" to each punters set up. Benter did the same thing in HK with a dataset that stared at him for 3-4 years before he realised it could have this same effect - the public's estimation of final odds.T'was then he started to be profitable.
And in my "daft" scenario it was Kyren Wilson who went from a ratings to odds price of 2.43 down to 1.87 meanwhile Selby drifted from a Ratings to odds price of 1.70 out to 2.15 - the point was imagine you were close to Selby's camp and knew about the "food poisoning" , his negativity , and the effect it was having on the practise table. This is one way how you can "update odds/probabilities" making them more dynamic than just using odds calculated from commercial ratings where you might not know the inputs and they could have been calculated hours before the match.
As said this technique which uses bijection and reduction can be used in any type of market , such as an A v B type markets (with no draws/or draws) and A v K type markets (which involve multi competitors)
Football is not my game but if i were to take it seriously i'd concentrate on obscure leagues -in fact the obscurer the better - Betting in All of the major countries top leagues are well covered by Bloom,Benham and a mysterious Mr X (London based) with a combined workforce between the 3 syndicates of over 1200 employees globally. The bookmakers traders and other Exchange sharps would also know much more about the Premiership or Championship than the 2nd Division in Norway or the 4th division in Italy.

 

[cosmicsports]

Yeah but it does not need to be "new information" from the "inside" - you may have a rating or a metric that you have tested , tweaked , refined etc until it's calibrated to a very high standard - Deliberately holding that back from the original modelling/selection process can be used to the same effect . That was why i said that the term "New" should be "relative" and "unique" to each punters set up. Benter did the same thing in HK with a dataset that stared at him for 3-4 years before he realised it could have this same effect - the public's estimation of final odds.T'was then he started to be profitable.
And in my "daft" scenario it was Kyren Wilson who went from a ratings to odds price of 2.43 down to 1.87 meanwhile Selby drifted from a Ratings to odds price of 1.70 out to 2.15 - the point was imagine you were close to Selby's camp and knew about the "food poisoning" , his negativity , and the effect it was having on the practise table. This is one way how you can "update odds/probabilities" making them more dynamic than just using odds calculated from commercial ratings where you might not know the inputs and they could have been calculated hours before the match.
As said this technique which uses bijection and reduction can be used in any type of market , such as an A v B type markets (with no draws/or draws) and A v K type markets (which involve multi competitors)

In the case of striker-model it was not an inside secret, it was all over the press.
Secrets is a different story.
So suppose the match was 70%-20%-10% before the happening.
Eventually it ended in a home win as I recall but that does n't matter.
But how to modify the probabilities on account of the said striker being in tears ? Approximately of course.
That's what we 're on about. We don't want by just intuition.
Then what you say about the reaction of the crowd is not so useful. We know what was the reaction of the crowd after the ship has sailed.

Also there are big events and small events.
The striker-model separation was a small event imo.
But Juventus's appeal turned down and them dropped from champions league to conference must have had a crushing psychological effect on the team.

 

[ARAZI91]

In the case of striker-model it was not an inside secret, it was all over the press.
Secrets is a different story.
So suppose the match was 70%-20%-10% before the happening.
Eventually it ended in a home win as I recall but that does n't matter.
But how to modify the probabilities on account of the said striker being in tears ? Approximately of course.
That's what we 're on about. We don't want by just intuition.
Then what you say about the reaction of the crowd is not so useful. We know what was the reaction of the crowd after the ship has sailed.

Also there are big events and small events.
The striker-model separation was a small event imo.
But Juventus's appeal turned down and them dropped from champions league to conference must have had a crushing psychological effect on the team.

Whether it's horseracing,sports ,financials etc i'll repeat again
"Yeah but it does not need to be "new information" from the "inside" - you may have a rating or a metric that you have tested , tweaked , refined etc until it's calibrated to a very high standard - Deliberately holding that back from the original modelling/selection process can be used to the same effect . That was why i said that the term "New" should be "relative" and "unique" to each punters set up. Benter did the same thing in HK with a dataset that stared at him for 3-4 years before he realised it could have this same effect - the public's estimation of final odds.T'was then he started to be profitable."
I do this every day as odds are dynamic along with the sports they represent - a race looked at with limited information the night before will not be the same as a race looked at with more information at 9am , just as a race looked at at 12.30 with even more information so the odds are dynamic because the sports/racing is dynamic. Deliberately HOLDING BACK very predictive factors from the original modelling / selection process and then using them in the way explained above(and there is no "intuition involved" -it's purely mathematical- Bayesian in a sense) can have the same effect,and by introducing these new factors at an optimal stage of the market allows you to UPDATE your probabilities instead of using a "static" rating that was maybe calculated many hours before the game and has limited inputs anyway.

Btw - AustinDillon75 -Here's a very large collection of racing academia (some sports as well) -the most famous one in racing is Ziemba's "Effeciency Of Racetrack Betting Markets" which is i suppose the "bible" for hose race modellers all over the world. The one i have included is the updated version along with much more stuff.
BT Cloud

 

[ARAZI91]

In the case of striker-model it was not an inside secret, it was all over the press.
Secrets is a different story.
So suppose the match was 70%-20%-10% before the happening.
Eventually it ended in a home win as I recall but that does n't matter.
But how to modify the probabilities on account of the said striker being in tears ? Approximately of course.
That's what we 're on about. We don't want by just intuition.
Then what you say about the reaction of the crowd is not so useful. We know what was the reaction of the crowd after the ship has sailed.

Also there are big events and small events.
The striker-model separation was a small event imo.
But Juventus's appeal turned down and them dropped from champions league to conference must have had a crushing psychological effect on the team.

It does not even have to be ratings or metrics - you could have an ensemble of say 10 models each comprising of many factors that logistically spit out a final probabilitie or range of probabilities - you could calibrate 8 of them over a few seasons worth of data mapped to closing odds and they could be close to break-even (after commission)- you then deliberately hold back the other two (very predictive) models deliberately from the originally modelling process and use the outputs to UPDATE the probabilities produced by the original 8 at an optimal stage of the market.- Of course you should have done strict testing, validation and training on this before you even start to venture down this path?........This is how your average betting syndicate works, from large Global consortiums like the London based Croat to maybe 12 man teams in Hammersmith.

 

[cosmicsports]

It does not even have to be ratings or metrics - you could have an ensemble of say 10 models each comprising of many factors that logistically spit out a final probabilitie or range of probabilities - you could calibrate 8 of them over a few seasons worth of data mapped to closing odds and they could be close to break-even (after commission)- you then deliberately hold back the other two (very predictive) models deliberately from the originally modelling process and use the outputs to UPDATE the probabilities produced by the original 8 at an optimal stage of the market.- Of course you should have done strict testing, validation and training on this before you even start to venture down this path?........This is how your average betting syndicate works, from large Global consortiums like the London based Croat to maybe 12 man teams in Hammersmith.

Ah yes, I understand, compound probability.
Like in the races. I have a set of probabilities derived from speed histogram analysis and another set derived from jokcey ratings and even a third set relating to days of absence. Those are put together so the compound probability is stronger compared to the component probabilities (it scores higher).
But how to find histories of such things ? Football matches in suspicious circumstances.

 

[cosmicsports]

Let's approach it this way:

We use ratings as probability network A and we introduce a second probability network B called the "ambiguity factor".
Then there exists an easy to identify group of matches that can help us somewhat calibrate this "ambiguity factor".
This is European matches of the years past in which teams were indifferent.

Example:

UEFA Champions League 2017-18, round of 16
07/03/18 Basel - Manchester City 0-4
18/03/18 Manchester City - Basel 1-2

and others similar.