Ratings - please read before commenting - Archived

[computerrank]

I wonder if that 65% can be raised by raising the R_DIFF value a bit, as I have theorized. You might also want to start looking at the rest of the weight classes to broaden the statistics. Powell/Ouma, for instance, was in the jr middleweight division.

... you cannot tweek the parameters against this margin - the underlying set of bouts depends on the ratings and parameters themselves ...

... for evaluation you need a rating independent set of bouts - as we usually use ...

... and JCS83md and I already told you - we did all kind of variations of all parameters - including R_DIFF - and we took the best values already ...

[jujigatame]

Well I don't think your ratings independent bouts are working very well. If you can't do an evaluation that involves a minimum amount of points then I think the launching procedure needs to be tweaked and made more strenuous.

[computerrank]

Well I don't think your ratings independent bouts are working very well. If you can't do an evaluation that involves a minimum amount of points then I think the launching procedure needs to be tweaked and made more strenuous.

... procedural suggestions welcome ...

[JCS]

Wow, the current system got Baldomir vs. Judah right.

That's exceptional.

[jujigatame]

Well I don't think your ratings independent bouts are working very well. If you can't do an evaluation that involves a minimum amount of points then I think the launching procedure needs to be tweaked and made more strenuous.

... procedural suggestions welcome ...

I will think on it and get back to you. It is a difficult problem that deserves some serious thought.

[jujigatame]

Wow, the current system got Baldomir vs. Judah right.

That's exceptional.

The system picking Neilsen over Tyson is a pretty good example of what I think we can improve with a more difficult launching process and a higher R_DIFF value. The same goes for Powell over Ouma.

[JCS]

Wow, the current system got Baldomir vs. Judah right.

That's exceptional.

The system picking Neilsen over Tyson is a pretty good example of what I think we can improve with a more difficult launching process and a higher R_DIFF value. The same goes for Powell over Ouma.

Powell vs Ouma was nowhere close to a blowout as that 100-90 card would suggest. The guys were only 15 or so points apart in the ratings.. now they are more accurately rated thanks to that fight :)

Like we've said before, we try to use the best formula across the board. Some people will get overrated, some will get underrated. The only solution is to either have a human create a subjective system, or deal with the inaccuracies of a computerized system.

Both a traditional system and predictive system will have inaccuracies, but people seem to be more fond of the ones caused by a traditional system just because it is what it is.. traditional.


Some chess systems use different R_DIFF values depending on whether or not the player is considered a veteran, or non-veteran (Higher R_DIFF if veteran, and vice-versa). If we could factor this in somehow, perhaps it would help predictive rates... but I believe something similar has been tested.

[jujigatame]

I'm not arguing for a traditional system. I'm simply arguing for a predictive system that doesn't overinflate the worth of guys like Briggs and Minto.

[JCS]

I'm not arguing for a traditional system. I'm simply arguing for a predictive system that doesn't overinflate the worth of guys like Briggs and Minto.

When you depend on a calculation based on a linear equation, you are aiming for the best possible average result. In a situation such as this, you will always get boxers who are overrated and you will always get boxers who are underrated.

Perhaps the answer is a set of non-linear equations for different scenarios. Making sense out of this and creating them based on patterns and not just fluke instances would be a task of biblical proportions.

[emile]

Martin didn't really speak to one of my issues. I may be missing something and I'm not good at expaining my thoughts, but I'm going to try again.

I don't feel as impressed by the Judah-Baldomir prediction, because I don't see what its based on, nor do I see such a great value in predictive success. As I surmised before - the value that Martin is putting on predictive rate is essentially stating that the best rating system would have 100% predictive accuracy. So, for example, the loss of Viloria to Omar Nino was an example of a _failure_ of the rating system, even though nobody else considers as such. I think a problem is this total dependence on the predictive rate. If the variables in the rating equation are not independently statistically significant, then isn't the outcome essentially a function of random events? Let's say you wanted to improve the predictive rate, so you give a large bonus to fighters named Omar Nino, giving him enough to pass Viloria at the time of the fight - now the predictive rate has improved (possibly with some adjustment to account for Nino's loss to Juan Alberto Rosas, a Rosas bonus I guess), but you haven't actually done anything to improve future predictive rates. As Martin said, the world is full of surprises - why base the ratings on a system that sees them as mistakes?

Maybe the variables are significant, but looking at them, I don't see how that could be. I feel like you guys are chasing your tail, improving the predictive rate, but creating very noticable and hard to explain outliers because of the functional instability of the equation. At least with the traditional system, as problematic as it was, you could pass off a justification. ie Ted Bami was too high because he beat a guy who beat a guy. Now you have Sechew Powell high, because.....a number of not necessarily independent or meaningful variables have been inputed to get an across the board high predictive rate of past fights, and they put Powell at #1.

I do think the predictive rate is extremely useful to comparatively evaluate ranking equation choices. But I'm not nearly convinced that a pure pursuit of the highest possible predictive rate, without consideration of historical boxing rating ideals is a good one.

[Blue]

I approve of the above message.

[jujigatame]

Ok, one quick suggestion regarding the launching system. You don't automatically get launched by beating a launched fighter. This rule is allowing guys who are like 1-6 to get launched. Also, allow grades below 0 and employ more than a 1 grade penalty to guys who lose to a lower grade than them.

[JCS]

Ok, one quick suggestion regarding the launching system. You don't automatically get launched by beating a launched fighter. This rule is allowing guys who are like 1-6 to get launched. Also, allow grades below 0 and employ more than a 1 grade penalty to guys who lose to a lower grade than them.

I think you are suggesting an idea similar to one I came up with which consists of a "fringe-like" zone which is between launched and unlaunched.

This zone could be used to place fighters who have gotten themselves launched, but have not beaten any launched fighters in a particular span of time. Maybe you also put launched fighters here who lose to more unlaunched boxers than the number of launched boxers they have defeated.

[computerrank]

Emile,

...As I surmised before - the value that Martin is putting on predictive rate is essentially stating that the best rating system would have 100% predictive accuracy.
...
I do think the predictive rate is extremely useful to comparatively evaluate ranking equation choices. But I'm not nearly convinced that a pure pursuit of the highest possible predictive rate, without consideration of historical boxing rating ideals is a good one.

We both know, a rating system never will predict 100% of the results - as there always will be only a certain level of probability for a result.

You accept the prediction rate evaluation as a method of finding acceptable rating systems and optimize them - I think.

You state, there is a rest, where this evaluation method is not all. We also should consider some historical boxing rating ideals - to some grade and with some weight, I suppose.

I should appreciate your procedural suggestions, how to integrate such ideas.

I perhaps see an option to show the performance ratings as they are, but additionaly give some information about
- the rating of the best opponent defeated in recent period - perhaps 18 months or 3 years?
- such a rating, but cut down to the rating of the lowest rated boxer lost to after this win, not lower than a later win

Perhaps the ratings could be displayed not only in the order of the current performance ratings - but by option additionaly in the order of such a parameter - approved level in recent period?

If anybody then would prefer this ranking, he would have the option to get it.

Best regards
Martin

@Blue - I knew, you look at this ... but you perhaps more appreciate the traditional flavour, box it out ...

[computerrank]

Ok, one quick suggestion regarding the launching system. You don't automatically get launched by beating a launched fighter. This rule is allowing guys who are like 1-6 to get launched. Also, allow grades below 0 and employ more than a 1 grade penalty to guys who lose to a lower grade than them.

Thanks jujigatame,

1. The launching system sets a winner to the grade of the loser - and the loser back. So, if the winners defeats a launched boxer, he should be launched, by consequence - this is very traditional by the way.

2. If I set back the loser by more than a grade - this means much more than a point in the rating system - it is quite hard already. Nevertheless I tried to set back the loser to the winners pre-bout grade. But then you will only see some 10 to 50 boxers in the ratings per division. And the quested untested boxers will still be in.

3. Negative grades do not make sense, as grade 0 already is for debutants and boxers with no win.

Best regards
Martin

[conan_the_cribber]

Hi Computerrank,

a very interesting analysis. I will make a careful analysis later, but the first impressions were,

- I knew all the fighters. That seems strange given there are some odd names popping up in the current top 25. Either the odd ones are not fighting top 25 opponents, or they become well known. More analysis is neeeded.

- There is an overwhelming number of heavyweight bouts in this list. This strikes me as odd. I would say, that elite heavyweight fighters are actually less active than elite welters (DLH exepted). Odd.

- There are only a few bouts, where I'd say, that an over inflated fighter got toasted. Perhaps they get toasted before they meet a top 25 fighter, perhaps they become top 25 fighters.

- The prediction percentage is good.

conan

[conan_the_cribber]

Martin didn't really speak to one of my issues. I may be missing something and I'm not good at expaining my thoughts, but I'm going to try again.

I don't feel as impressed by the Judah-Baldomir prediction, because I don't see what its based on, nor do I see such a great value in predictive success. As I surmised before - the value that Martin is putting on predictive rate is essentially stating that the best rating system would have 100% predictive accuracy. So, for example, the loss of Viloria to Omar Nino was an example of a _failure_ of the rating system, even though nobody else considers as such. I think a problem is this total dependence on the predictive rate. If the variables in the rating equation are not independently statistically significant, then isn't the outcome essentially a function of random events? Let's say you wanted to improve the predictive rate, so you give a large bonus to fighters named Omar Nino, giving him enough to pass Viloria at the time of the fight - now the predictive rate has improved (possibly with some adjustment to account for Nino's loss to Juan Alberto Rosas, a Rosas bonus I guess), but you haven't actually done anything to improve future predictive rates. As Martin said, the world is full of surprises - why base the ratings on a system that sees them as mistakes?

Maybe the variables are significant, but looking at them, I don't see how that could be. I feel like you guys are chasing your tail, improving the predictive rate, but creating very noticable and hard to explain outliers because of the functional instability of the equation. At least with the traditional system, as problematic as it was, you could pass off a justification. ie Ted Bami was too high because he beat a guy who beat a guy. Now you have Sechew Powell high, because.....a number of not necessarily independent or meaningful variables have been inputed to get an across the board high predictive rate of past fights, and they put Powell at #1.

I do think the predictive rate is extremely useful to comparatively evaluate ranking equation choices. But I'm not nearly convinced that a pure pursuit of the highest possible predictive rate, without consideration of historical boxing rating ideals is a good one.

A very good post indeed. Excellent. Hit the nail right on the head there. The over-reliance on the predictability statistic and the attitude that an upset is a mistake to be corrected at all costs.

I'm a decent chess player. If I suddenly appeared on the scene and won a few low level tournaments, maybe even the German championships and suddenly the ELO rankings said that I'm better than Kasparov, then I think everyone in the chess world would be freaking out. They'd be out of their little nuthouses, because the ELO rankings decide sometimes the entry places in tournaments and seedings in tournaments etc. Because the boxrec rankings are obscure and decide didley squat, no-one in the business is really outraged.

The reason the chess rankings are good and accepted, is that they represent the consensus opinion and are good at predicting. If, as a product of some algorithm, they predicted me as beating Kasparov, then they wouldn't have the consensus and respect they do. They wouldn't be used for entry in tournaments or as a basis of seeding, no matter how well they predicted.

The boxrec rankings are in this unfortunate position. The algorithms produce outliers that are so visible and so wrong, that nobody in the industry takes the rankings seriously. Shannon Briggs beats Minto and Briggs is number one? Spinks beats Mayorga, Judah and Kamazan is 17 spots beneath Kamazan? Calvin Brock and Sechew Powell at number 1. Those are weird outliers. The attitude that reducing the number of wrong predictions will automatically produce the best rankings is wrong, if the side effects are so dramatic.

conan

[computerrank]

conan,

you compare apples with oranges here.

Boxrec algorithm and ELO algorithm have the same base - this definitely is not the difference.

The basic difference is in
- the number of players in the chess world - hundreds of thousends or millions of active chess playera - versa about some thousends boxers
- the much higher rate of competition - chess players with hundreds of matches versa boxers with often only a handfull of bouts maybe once or twice a year

So the boxing ratings are faced with much fewer events to relax to some valid estimation, and represent a much lower density of skills.

So a win may cross over all limits in boxing seen from the boxing perspective - which in chess never will happpen.

The real requirement in boxing is to give a valid rating estimation based on a very small statistical base.

Speaking technically, in chess the space between the launching area and the top of the game is about 1400 points I think and the K-Factor is down to 30 or 15 (slow relaxation). This is a span of 100.

In boxing we are forced to use much faster relaxation - see our earnings of about 90 points (compare with a K-Factor of 180). and the space between the best and the launching area is about 500 to 800 points only. This is a span of 3.

So there is no space for moaning - we have to find the best solution - I should appreciate constructive suggestions, please.

Pure acclamations are not really helpful, sorry.

Best regards
Martin

[conan_the_cribber]

conan,

you compare apples with oranges here.

Boxrec algorithm and ELO algorithm have the same base - this definitely is not the difference.

The basic difference is in
- the number of players in the chess world - hundreds of thousends or millions of active chess playera - versa about some thousends boxers
- the much higher rate of competition - chess players with hundreds of matches versa boxers with often only a handfull of bouts maybe once or twice a year

So the boxing ratings are faced with much fewer events to relax to some valid estimation, and represent a much lower density of skills.

So a win may cross over all limits in boxing seen from the boxing perspective - which in chess never will happpen.

The real requirement in boxing is to give a valid rating estimation based on a very small statistical base.

Speaking technically, in chess the space between the launching area and the top of the game is about 1400 points I think and the K-Factor is down to 30 or 15 (slow relaxation). This is a span of 100.

In boxing we are forced to use much faster relaxation - see our earnings of about 90 points (compare with a K-Factor of 180). and the space between the best and the launching area is about 500 to 800 points only. This is a span of 3.

So there is no space for moaning - we have to find the best solution - I should appreciate constructive suggestions, please.

Pure acclamations are not really helpful, sorry.

Best regards
Martin

Jeez, how much input do I have to give? Suggesting algorithms, measurements, checking trails, suggesting improvements.

Besides I consider the arguments expressed by the other poster and myself to be elementary to the problem. There's the old quote, "the first step in solving a problem is recognising that there is a problem".

Honestly Martin, I dont know if you really follow boxing, enjoy watching it, take an interest in who's winning and who's losing? I haven't seen you post anywhere on any Current Scene thread, except if it's an answer to a ratings question. If you do have any interest in boxing then can you truthfully answer the following question.

"Does the winner of a theoretical Shannon Briggs vs Brian Minto match up deserve to be the number one in the heavweight division, die Königsklasse von Boxen?".

Can you really answer that with yes. Not if you really are a boxing fan.

As to the apples and oranges comparison. It is you, not I, that are comparing apples and oranges. Golf, Tennis, Cricket, Chess all have rankings systems that are tailored to their domain. You are the one that has taken the solution for the chess domain (apples) and forced it on the boxing domain (oranges). You do this despite fully knowing that the number of matchups is not frequent enough in boxing to properly justify it. Chess players play 30 bouts or more a year, boxers 30 in a career. Clearly you have to make adjustments for this.

The launching part is an excellent start of a domain specific solution. The concept is the same as in other algorithms, but the solution is domain specific.

The real requirement in boxing is to give a valid rating estimation based on a very small statistical base.

If that is truly the real requirement, then again it is the argument what is meant by valid. I say "accepted by peers and supported by predictability", you say "predictability", with of course, all the known side effects.

It may be painful to hear it over and over again, but hey, answer the question above as a boxing fan, and you'll see why the questions will be repeated.

conan

[Lennox]

Martin, I think Conan is right, I have been folowing this thread for a long time now and have just picked up the last 4 pages posted!

The problem with the Boxrec ratings are that they look bad, there is nothing wrong with the principle of the prediction quality being high, but it does not work at its best at the top end of the ratings, as Conan says most people are interested top 20 or so. It is here Boxrec is really bad. No one with any degree of boxing sanity would give you the thumbs up for Chad Dawson, Calvin Brock, Sechew Powell, etc etc being either the #1 fighter or the person most likely to defeat the other boxers. Their crazy position will not be made to look crazy until they fight someone good, Sechew Powell was found out and was clearly leagues away from Ouma. You have Brock #1 but he has beaten who to justify and more importantly on what basis of beating nobody do you say he would be favoured to beat Wladimir, Valuev, Maskaev & Lyacovich. I would say those four fighters have a 65% chance of beating Brock that is my opinion through Boxing knowedge and my ability to anaylse fights, your statistics say Brock will win. The Shannon Briggs - Brian Minto scenario hammers home the point that your system is flawed.

You produce statistics that I think become twisted, I think that your system favours undefeated prospects and a man with 20-0 will be high in Boxrec without fighting too much of a foe. Boxing people know that with a reasonable fighter it is easy to get 20 wins. Boxing people know that there are many good fighters with patchy records, lets call them gatekeepers, Bert Cooper was one, he was someone that could not beat people inside the top 30, but was a tough match up for a newcomer. Your system does not recognise good fighters with poor records, John Ruiz has dropped down far too low after his loss, with the state of the heavyweights, he is still top 10/15.

Your system could work though if you considered a different method for the higher area of your rankings this is where your boxrec, in time will become more and more distorted.

I like the IBO, I see ratings that look right, I would say IBO would be a better predicter if you looked at the pure top 30 ratings than the top 30 boxrec. There was a competition a few weeks ago, but Boxrec lost a few significant fights and JCS seems to have just dismissed it, when I did it before IBO came out on top.

I would like to see all the sanctioning bodies together under the ratings section, so its easy to check them all out, maybe the world bodies would pay a fee and it could be a way of raising money for Boxrec.

[emile]

Emile,

...As I surmised before - the value that Martin is putting on predictive rate is essentially stating that the best rating system would have 100% predictive accuracy.
...
I do think the predictive rate is extremely useful to comparatively evaluate ranking equation choices. But I'm not nearly convinced that a pure pursuit of the highest possible predictive rate, without consideration of historical boxing rating ideals is a good one.

We both know, a rating system never will predict 100% of the results - as there always will be only a certain level of probability for a result.

You accept the prediction rate evaluation as a method of finding acceptable rating systems and optimize them - I think.

You state, there is a rest, where this evaluation method is not all. We also should consider some historical boxing rating ideals - to some grade and with some weight, I suppose.

I should appreciate your procedural suggestions, how to integrate such ideas.

I perhaps see an option to show the performance ratings as they are, but additionaly give some information about
- the rating of the best opponent defeated in recent period - perhaps 18 months or 3 years?
- such a rating, but cut down to the rating of the lowest rated boxer lost to after this win, not lower than a later win

Perhaps the ratings could be displayed not only in the order of the current performance ratings - but by option additionaly in the order of such a parameter - approved level in recent period?

If anybody then would prefer this ranking, he would have the option to get it.

Best regards
Martin

@Blue - I knew, you look at this ... but you perhaps more appreciate the traditional flavour, box it out ...

Although it may be impossible to have a 100% predictive rate, by using the prediction rate as the sole factor of a rankings quality, you are saying that a higher prediction rate is always better than a lower one. Actually, it might be possible to get close to 100% if you looked at enough (random) factors.

I wonder if you would be willing to bet your own money on the future predictive quality of the equation? I wouldn't, because I don't believe that the factors are significant. However, it would be intere
sting to see how the predictive rate for future fights will pay off given the odds on each fight (looking at past fights wouldn't be valid because you are using elements from those results to create the rankings).

Another example of a problem, based on 'traditional' rankings ideas, is at Lightweight. David Diaz and Ramazan Palyani fought to a draw in December. Since then, they've each had three fights - Palyani with three KO's in Georgia against unlaunched opponents. Diaz with two decision wins against mediocre club guys (one launched, one not) and a TKO of Jose Armando Santa Cruz. Palyani is #6, Diaz is #11. They had similar histories before their meeting, although Diaz did have his loss to Kendall Holt. So - how were they rated at the time they fought? Has Palyani overtaken, or added distance over Diaz since then, or has the gap closed? I don't totally get how the R_DIFF is implemented yet, so I would like to see it applied in this case.

Sorry I can't yet help with suggested equations using traditional ideas, but I'll think about it.

[JCS]

I like the IBO, I see ratings that look right, I would say IBO would be a better predicter if you looked at the pure top 30 ratings than the top 30 boxrec. There was a competition a few weeks ago, but Boxrec lost a few significant fights and JCS seems to have just dismissed it, when I did it before IBO came out on top.

Do it again, except keep consistent and do it for more than just several weeks. I didn't dismiss anything, lets see some action.

If you want to compare IBO to BoxRec, you'll consider all fights w/ ranked fighters, not just those in the Top 30.

[Lennox]

I like the IBO, I see ratings that look right, I would say IBO would be a better predicter if you looked at the pure top 30 ratings than the top 30 boxrec. There was a competition a few weeks ago, but Boxrec lost a few significant fights and JCS seems to have just dismissed it, when I did it before IBO came out on top.

Do it again, except keep consistent and do it for more than just several weeks. I didn't dismiss anything, lets see some action.

If you want to compare IBO to BoxRec, you'll consider all fights w/ ranked fighters, not just those in the Top 30.

JCS. you started to do a Boxrec v IBO v another thread, but as soon as Roy Jones won and Sechew Powell got beat you stopped. I'll try a top 30 with IBO and Boxrec, you seem to have missed the point I was making in the previous post which I will reiterate. Boxrec as a prediction method may be good considerng all fights, but its not so good at the top end of the ratings where prospects are pushed too early, many of these prospects will get beat by the established fighters. All Boxrec has really donbe iis singled out fighters with good records and said in x% of the time the 22-0 man will beat his opponent, without making enough reerence to a 30-4 fighter who has lost fights at a high level.

[conan_the_cribber]

I like the IBO, I see ratings that look right, I would say IBO would be a better predicter if you looked at the pure top 30 ratings than the top 30 boxrec. There was a competition a few weeks ago, but Boxrec lost a few significant fights and JCS seems to have just dismissed it, when I did it before IBO came out on top.

Do it again, except keep consistent and do it for more than just several weeks. I didn't dismiss anything, lets see some action.

If you want to compare IBO to BoxRec, you'll consider all fights w/ ranked fighters, not just those in the Top 30.

JCS. you started to do a Boxrec v IBO v another thread, but as soon as Roy Jones won and Sechew Powell got beat you stopped. I'll try a top 30 with IBO and Boxrec, you seem to have missed the point I was making in the previous post which I will reiterate. Boxrec as a prediction method may be good considerng all fights, but its not so good at the top end of the ratings where prospects are pushed too early, many of these prospects will get beat by the established fighters. All Boxrec has really donbe iis singled out fighters with good records and said in x% of the time the 22-0 man will beat his opponent, without making enough reerence to a 30-4 fighter who has lost fights at a high level.

No I started it, but I stopped cause I went on holidays with the family.

conan

[computerrank]

conan,

It may be painful to hear it over and over again, but hey, answer the question above as a boxing fan, and you'll see why the questions will be repeated.

I very carefully follow everthing regarding boxing - and sometimes follow the top fights, if available on television. But I don't find it necessary to discuss in the current scene forum in threads nearly all concerning "I hate a boxer", "I love a boxer", "I am absolutley sure, ...", "you are an idiot ..." - I simply find it quite boring to spend my time with this.

By this, it is also not difficult to see, what you saying with the Brocks, Briggs, Dawsons etc.

I even see the structure behind:

- they all are rated quite high above the the ratings of the boxers, they defeated at all ( this regarding no loss boxers) or in a recent time period ( this regarding Briggs)

- it so is absolutely unapproved, that they are able to defeat these days opponents at their rating level, yes

- yes, I am sure, that not everyone of them will do so - and everyone is suspect in some way to me

- but in spite of my feelings, I can imagine, that 50% of these boxers, will do so - and I found it in the statistics

- this is, what the ratings say, not more - and then, the ratings are correct from a statistical perspective

- I think, the problem behind - for everybody - including me is:

We all will more accept the situation, where a low rated boxer defeats a higher rated one, who already defeated opponents at his level in a recent time period. Altough we all know, this happens quite often.

But we all have problem to accept, where a high rated boxer, who didn't yet defeat an opponent at his level, loses to a lower rated one. Altough this also happens quite often.

We know this exspection issue from psychology - people more accept to not win, even when there is a good chance - but there is also a considerable risk to lose.

And the observer's losing feeling here is, that he hates to have been wrong, and in addition now cannot defend his exspectation afterwards by simple facts e.g. earlier wins against proven opponents.

A way to go would be, to weight the outcomes by this exspectation:

- a loss of a boxer against an opponent rated higher than the boxers best defeated last opponents (in a recent time period) should be regarded as a multiple upset - by multiple weight.

This exactly is, what people makes pain, I think.

Best regards
Martin