Ratings - please read before commenting - Archived

[conan_the_cribber]

Hi Martin,

lost of posts today.

A question. Most of the rankings bodies (and the guidelines) reward activity. Your current rankings explicity penalize inactivity. We are also nutting out what to do when the higher ranked fighter loses, but lets talk a little about when a higher ranked fighter remains active and has a couple of wins.

If Jimbo was the number 10 figher and fought two decent ranked opponents at 11 and 12, then in the current implementation Jimbo would remain at number 10. That strikes me as a little unfair, as Jimbo has obtained two good wins against a couple of peers. I could imagine that most people might puch him up a notch or two because of these wins. Certainly moving up, would increase the predictability, because this fighter is on a bit of a high.

I could imagine that a win over a near ranked peer should move you up a notch. What do you think? Certainly you shouldn't progress because of a "tune-up" win against someone who is way down the rankings. But a near win in a crossroads fight or an eliminator is surely to be rewarded.

Any comments, ideas?

conan

[computerrank]

- higher loser down depending on result and distance to lower winner
-- your proposal is not just in balance, as the winner gets too much = loser's rank, if you punish the loser too much
-- so I took the 0.25 of the logarithmic distance + 2 ranks for a close loss
-- I took 0.5 of the logarithmic distance + 3 ranks for another loss
-- this pretty fine puts down the continious higher loser

Hi Martin,

This approach is not working for me. The log distance effect is really miniscule. Taking 0.25 or 0.5 of it is practically reduces the effect to a maximum of one extra rank. As described before, Julius Francis starting at rank 20 will only end up at rank 51 despite 5 KO losses in a row to rank 80 fighters. You are not punishing the losers enough. Also, any top fighter will drop at least to 3 or 4 in the rankings, even if he fought the number 2 fighter.

The number of successive losses is critical to a good algorithm. It matches real life ranking thinking well. We've all seen fighters who just cant win anymore, they have to disappear downwards. The career of a boxer is not symmetric. The downhill slide is a lot quicker than the uphill battle. The number of successive losses against lower ranked fighters is a good indication of when that slide is happening. This indication is missing in the log stuff, in fact the brakes are put on the downhill slide.

In my algorithm suggestion I used the difference between the harmonic mean and the original ranking as a measure of the 'shock' of the victory. This worked out surprisingly well in the examples in all ranges of data. I would love to see what effect it would have on the predictability.

conan

Conan,

Sorry, I think, you didn't get me. It does work:

high boxer #20, opponent #80

- 1st KO loss: loser -> 20*(80/20)**0.5+3 = #43
- 2nd KO loss: loser -> 43*(80/43)**0.5+3 = #61
- 3rd KO loss: loser -> 61*(80/61)**0.5+3 = #72
- 4th KO loss: loser -> 72*(80/72)**0.5+3 = #78
- 5th KO loss: loser -> 78*(80/78)**0.5+3 = #81

high boxer #1, opponent #10

- 1st KO loss: loser -> 1*(10/1)**0.5+3 = #6
- 2nd KO loss: loser -> 6*(10/6)**0.5+3 = #10

now opponent be #100

- 3rd KO loss: loser -> 10*(100/10)**0.5+3 = #34
- 4th KO loss: loser -> 34*(100/34)**0.5+3 = #61
- 5th KO loss: loser -> 61*(100/61)**0.5+3 = #81

advantage:

- this can be applied continiously at any rate, is compling with the general ABC rules and logically consistent
- your suggestion would for high multiple losses imply that the winner would get a good rank, but the multiple loser would get a very bad one (by punishment) - this is quite a contradiction in logics - as then the winner also shouldn't get the rank of the loser

example:

say higher with 4th loss at #50 and lower winner at #300:

- mine: loser -> 50*(50/300)**0,5+3 = #125
_l_new = 1+ r_l + num_suc_losses * ( harm(r_w, r_l) - r_l)

- yours: loser -> 1+50+4*(85-50) = #191

- so why should the winner get #50, if the loser is only worth #191
- and there is nowhere a rule like a punishment for continious losses in ABC rules - it all only goes for the actual result and rankings

- I prefer my solution, sorry. It works very well on continious losses and doesn't need additional implications. So keep things simple.

- yes, it is tradition, top boxers, #1 also, should drop not only to #2 from #1, but to #3 or #4 depending on result (look at ABCs implementations, e.g WBA tables); take in mind, this goes for tradition, not for performance and predictablity.

For this intention we have a much better model - the performance ratings as implemented. So no intermixing of goals.

- Traditional Rankings (following tradition, bout for ranks)
- Performance Ratings (maximized prediction ratio)

Thanks for your suggestions and discussion
Martin

Maybe I will implement your formula, just for fun ... and show you the results ...

[computerrank]

Hi Martin,

lost of posts today.

A question. Most of the rankings bodies (and the guidelines) reward activity. Your current rankings explicity penalize inactivity. We are also nutting out what to do when the higher ranked fighter loses, but lets talk a little about when a higher ranked fighter remains active and has a couple of wins.

If Jimbo was the number 10 figher and fought two decent ranked opponents at 11 and 12, then in the current implementation Jimbo would remain at number 10. That strikes me as a little unfair, as Jimbo has obtained two good wins against a couple of peers. I could imagine that most people might puch him up a notch or two because of these wins. Certainly moving up, would increase the predictability, because this fighter is on a bit of a high.

I could imagine that a win over a near ranked peer should move you up a notch. What do you think? Certainly you shouldn't progress because of a "tune-up" win against someone who is way down the rankings. But a near win in a crossroads fight or an eliminator is surely to be rewarded.

Any comments, ideas?

conan

conan,

you very precisely identified the essential point for improving precition ratio.

You have to set boxers some margin above the loser, if they were close enough.

But - this also is definitely not intended by the traditional rules. As the boxers above the loser would be touched in their achievements - a wording often mentioned in the ABC rules. And this is not accepted.

Activity is one point (of even more) in the ABC rules to select candidates from nowhere to their rankings. But activity never will set a ranked boxer above a never defeated by him.

So again - I feel, you are trying to find a best performing (whatever this would be) traditional ranking. This is not my goal.

I want to implement a working, objective traditional ranking complying with their goals and rules - and this is definitely not prediction ratio.

By the way, I already implemented a ranking follwing your ideas - with this certain overshooting for the close winner. And it showed very improved prediction rate - up to 74% (no we have 65 %), if I remember correctly.

The basic idea given the ranks were:

- find the gemetric mean of the ranks
- set the winner to a rank of mean / factor; factor about 1.4
- set the loser to a rank of Mean * factor
- factor varying by result

But later I found, I just tried to make the same things with ranks as I before did with the performance points.

I just built some metrics into the rankings, and I could find that a factor in the ranks just is equivalent to some rating difference of the performance ratings.

Do you see it now?

Best regards
Martin

[computerrank]

@conan

ad other questions from recent messages:

conan: I am not sure what point you are trying to make here. I indicated that Taylor and Vargas are an indication that the algorithm is not really complete. I looked at the results you posted and I can see that Hatton is ranked in two divisions. One boxer, one division is pretty much a necessity for rankings.

-> You need to keep boxers in more than 1 division, as some boxers are changing quite often. The rank in a division is derived from a neighbour division for the first fight in a division (factor 3 for stepping up, +10 for transition to heavyweight, no rank change for going down). Afterwards the ranks are calculated based on the bouts in the divisions. As boxer change quite often, the ranking have to be kept - at least for calculation in the background - I showed them all.

It is a second question to show a boxer only in his strongest division perhaps, or the last one - as the ranking is kept only in a division, when the last bout in the division was not later as 18 months.

Vargas came in with Vargas in a super middleweightweight bout. Both with good ranks from lower divsions. Then boxed in lower divisons again, but they kept their ranks in super middleweight - as 18 months are not over. So they have a good super middleweight ranking, while never boxed a real super middleweight opponent. But also it wouldn't really matter as they are assigned to middleweight and light middleweight now.

Taylor for similar reasons ...

You: Some of the rankings below are hard to explain. I have editted the list above. I think the algorithm still needs some work somewhere along the line. Maybe when the duplicates are out or something.


Light Heavyweight

M Light Heavyweight 6 Hugo Hernan Garay 124

Who did Garay beat to get that high? Are you handling MD's like Majority Draws instead of Majority decisions?

Super Middleweight

M Super Middleweight 1 Markus Beyer 194
M Super Middleweight 2 Mikkel Kessler 208
M Super Middleweight 3 Anthony Mundine 85
M Super Middleweight 4 Joe Calzaghe 159

It's very odd that Beyer ends up on top. I would've imagined that after ottke's retirement, Calzaghe would've been lefton top. For Mundine to be above Calzaghe, Green must have been above him and before that Lucas above Calzaghe. That is odd.

Middleweight

M Middleweight 1 Jermain Taylor 54
M Super Middleweight 2 Mariano Natalio Carrera 348
M Middleweight 3 Ronald Wright 54
M Middleweight 4 Arthur Abraham 89

Something is very odd there. Carrera has done nothing to get so high.

-> Look yourself the tracks - this is the outflow of the rules:

Code: Select all

Garay

M 15 2004-05-08 023935 054183 W MD 1<-1 16<-35-a
M 15 2004-08-21 054183 055619 W KO 16<-16 152<-149-a
M 15 2005-02-26 023935 054183 W SD 1<-1 9<-15-a
M 15 2005-10-07 054183 131760 W UD 8<-8 51<-48-a
M 15 2006-04-08 054183 019850 W UD 8<-8 1297<-1297-a

Bayer

M 14 2003-04-05 007045 006405 W SD 3<-22 6<-3-a
M 14 2003-08-16 007045 051748 W DQ 3<-3 36<-33-a
M 14 2004-02-28 007045 022762 W UD 3<-3 27<-24-a
M 14 2004-06-05 014682 007045 W SD 3<-509 12<-3-a
M 14 2004-10-09 014682 007045 L KO 9<-3 3<-12-a
M 14 2004-12-18 007045 017152 W UD 3<-3 49<-46-a
M 14 2005-03-12 007045 051748 W MD 2<-2 5<-5-a
M 14 2005-09-03 007045 015151 W UD 2<-2 24<-21-a
M 14 2006-01-28 007045 014468 W TKO 1<-1 54<-51-a

Kessler

M 14 2003-04-11 016951 004765 W KO 4<-4 32<-29-a
M 14 2003-10-24 016951 015753 W TKO 4<-4 250<-247-a
M 14 2004-03-13 016951 008279 W KO 4<-4 345<-342-a
M 14 2004-06-11 016951 022762 W TKO 4<-4 32<-29-a
M 14 2004-11-12 014602 016951 L TKO 10<-7 4<-4-a
M 14 2005-06-08 016951 014646 W UD 3<-3 16<-13-a
M 14 2006-01-14 016951 006405 W TKO 2<-2 28<-25-a

Mundine

M 14 2003-02-01 014646 013474 W UD 11<-11 104<-101-a
M 14 2003-03-17 014646 007651 W TKO 11<-11 745<-742-a
M 14 2003-09-03 014646 005159 W UD 7<-11 11<-7-a
M 14 2004-01-19 014646 017152 W TKO 7<-7 44<-41-a
M 14 2004-05-05 014646 014602 L SD 13<-7 7<-45-a
M 14 2004-09-08 014646 013474 W TKO 14<-14 74<-71-a
M 14 2005-06-08 016951 014646 W UD 3<-3 16<-13-a
M 14 2005-12-11 014646 096270 W TKO 15<-15 846<-843-a
M 14 2006-05-17 014646 051748 W UD 3<-10 8<-3-a

Calzaghe

M 14 2003-06-28 005364 006858 W TKO 1<-1 12<-9-a
M 14 2004-02-21 005364 021467 W TKO 1<-1 180<-177-a
M 14 2004-10-22 005364 019292 W UD 1<-1 57<-54-a
M 14 2005-05-07 005364 016148 W TKO 7<-7 31<-28-a
M 14 2005-09-10 005364 017132 W UD 7<-7 1879<-1879-a
M 14 2006-03-04 005364 031593 W UD 4<-5 7<-4-a

Carnera

M 13 2002-04-13 047226 047964 W TKO 19<-310 79<-19-a
M 13 2002-06-15 047226 007552 W UD 19<-19 30<-27-a
M 13 2002-07-13 047226 055619 W KO 19<-19 112<-109-a
M 13 2003-03-29 044720 047226 W SD 11<-53 18<-11-a
M 13 2003-05-10 047226 008321 W TKO 18<-18 37<-34-a
M 13 2003-06-07 047226 084200 W TKO 18<-18 280<-277-a
M 13 2003-07-25 047226 002122 W SD 18<-18 74<-115-a
M 13 2003-09-19 047226 006961 W UD 19<-19 194<-191-a
M 13 2003-10-18 047226 047964 W TKO 19<-19 56<-53-a
M 13 2003-12-20 047226 048503 W TKO 19<-19 31<-28-a
M 13 2004-02-28 047226 006961 W TKO 11<-11 136<-133-a
M 13 2004-06-19 047226 021487 W TKO 5<-10 10<-5-a
M 13 2004-09-11 047226 004564 W KO 5<-5 62<-59-a
M 13 2004-12-10 047226 007552 W UD 4<-4 43<-40-a
M 13 2005-02-12 047226 151869 W UD 3<-3 127<-124-a
M 13 2005-05-20 047226 134746 W UD 2<-2 61<-58-a
M 13 2005-06-11 014665 047226 W UD 2<-54 13<-2-a
M 13 2005-08-27 047226 014665 W TKO 2<-14 8<-2-a

[conan_the_cribber]

Conan,

Sorry, I think, you didn't get me. It does work:

high boxer #20, opponent #80

- 1st KO loss: loser -> 20*(80/20)**0.5+3 = #43
- 2nd KO loss: loser -> 43*(80/43)**0.5+3 = #61
- 3rd KO loss: loser -> 61*(80/61)**0.5+3 = #72
- 4th KO loss: loser -> 72*(80/72)**0.5+3 = #78
- 5th KO loss: loser -> 78*(80/78)**0.5+3 = #81

high boxer #1, opponent #10

- 1st KO loss: loser -> 1*(10/1)**0.5+3 = #6
- 2nd KO loss: loser -> 6*(10/6)**0.5+3 = #10

now opponent be #100

- 3rd KO loss: loser -> 10*(100/10)**0.5+3 = #34
- 4th KO loss: loser -> 34*(100/34)**0.5+3 = #61
- 5th KO loss: loser -> 61*(100/61)**0.5+3 = #81

advantage:

- this can be applied continiously at any rate, is compling with the general ABC rules and logically consistent
- your suggestion would for high multiple losses imply that the winner would get a good rank, but the multiple loser would get a very bad one (by punishment) - this is quite a contradiction in logics - as then the winner also shouldn't get the rank of the loser

example:

say higher with 4th loss at #50 and lower winner at #300:

- mine: loser -> 50*(50/300)**0,5+3 = #125
_l_new = 1+ r_l + num_suc_losses * ( harm(r_w, r_l) - r_l)

- yours: loser -> 1+50+4*(85-50) = #191

- so why should the winner get #50, if the loser is only worth #191
- and there is nowhere a rule like a punishment for continious losses in ABC rules - it all only goes for the actual result and rankings

- I prefer my solution, sorry. It works very well on continious losses and doesn't need additional implications. So keep things simple.

- yes, it is tradition, top boxers, #1 also, should drop not only to #2 from #1, but to #3 or #4 depending on result (look at ABCs implementations, e.g WBA tables); take in mind, this goes for tradition, not for performance and predictablity.

For this intention we have a much better model - the performance ratings as implemented. So no intermixing of goals.

- Traditional Rankings (following tradition, bout for ranks)
- Performance Ratings (maximized prediction ratio)

Thanks for your suggestions and discussion
Martin

Maybe I will implement your formula, just for fun ... and show you the results ...

Hi Martin,

thanks for explaining your formula. You said 'log' in your previous post, so I took you at your word and used log based 10. It was hard to guess that you actually meant the result of a dividing two numbers.

Now that I know your formula, I'm a bit worried about the top end of it, especially ranked number 1. All numbers divided by 1 produce a large result. Any loss of a top 1 fighter is going to move him down to number 5 at least and often to number 6. This is clearly out of touch with what happens in real life and for that matter, what the WBA regularly publishes.

As to what is better, you have a set of rankings that produce a better set of predictions, at the moment. This is a relatively new algorithm and already pretty close. We'll have to wait and see what happens.

Regarding, the 50 and 191 point. There is no rule in it the ABC rankings, just the broad, fighters must be rewarded on their activity and their results. Losing four in a row and then losing to a 300th ranked fighter is a devistating statement about where the boxer stands. It could be appropriate to drop him to 191 according to the ABC criteria, although they do not state that multiple successive losses are to be especially penalized, it is implicit and allowed in interpreting a boxers activity and results in the ring.

I can justify it though statistically. The resulting rank is a statment about the expections of the next fight of the boxer. At the time of the fight, it was expected that the higher ranked boxer called Paz would lose to the first 49 fighters, but would beat the fighters ranked 51 to 600.

After the fight, we have new information. His opponent Queenie who was ranked 300 performed at least as well as Paz was expected to perform. So Queeinie grabs his spot at 50. We now know that Paz is heavily on the slide. Due to this loss, we now estimate that he would lose agaisnt the top 190 fighters and win against the bottom fighters from 192-600.

In summay, the string of worsening recent losses have affected his average expected result.

BTW, empirically, I got no problem dropping someone who was 50 back down to 191 if they lose to a 300 fighter, especially if it was on the back of four previous losses without beating someone better than himself for a while. Maybe at 191 they can beat a 130 fighter and get the monkey off their back.

REgarding predictability and traditional rankings. Like I argued (somewhat forcefully) at the beginning of my posts. I see predictability as one important measure as to the accuracy of a set of rankings. Where we differ is that you see it as the only measure (well at least you wont admit anything else). If we can improve the predictability of the results of the traditional system by tweaking certain aspects of the algorithm e.g. the division jump or the penalty for successive losses or plus points for fighting peers, then it will only improve the quality of the rankings. I dont see the two goals as contradictory.

Thanks for all your help. I hope you are at least enjoying the exercise. I would be very grateful if you implemented my suggestion with the successive losses and the harmonics, but that is up to you. Before you do though, I think it is important to sort out where the 'errors' are coming from. Looking at one of previous responses, there are obviously some people in some weight classes that simply cannot be there, if the algorithm is implemented correctly. You also have boxers in multiple divisions, this cant help.

Thanks again for all the work.

conan

[conan_the_cribber]

Hi

why did Calzaghe drop 6 places between these fights?

M 14 2004-10-22 005364 019292 W UD 1<-1 57<-54-a
M 14 2005-05-07 005364 016148 W TKO 7<-7 31<-28-a

Doesn't make sense if he was number 1.

conan

[computerrank]

Conan,

ad WBA rules for losing champion:

TABLE 4
Appendix to WBA Rules

A.- TABLE OF DEMOTIONS ACCORDING TO THE
RESULT OF THE FIGHTS FOR A WORLD TITLE
IN THE CASE THE WORLD CHAMPION IS DEFEATED.

This table shows the position that the World Ex-Champion shall occupy, according to the result that caused his defeat by the challenger, who has won the title.

NEW POSITION OF THE EX-CHAMPION

SD -> 2
UD (2) -> 3
UD (3+) -> 4
TKO/KO 7-12 -> 5
TKO /KO 1-6 -> 6

See more at - Norms and Procedures for Ratings:

http://www.wbaonline.com/legal/default.asp

[computerrank]

Hi

why did Calzaghe drop 6 places between these fights?

M 14 2004-10-22 005364 019292 W UD 1<-1 57<-54-a
M 14 2005-05-07 005364 016148 W TKO 7<-7 31<-28-a

Doesn't make sense if he was number 1.

conan

... he didn't fight an opponent within rank 45 within 1.5 years

... his new rank is -> 1*2+5 = 7 (before the fight)

[conan_the_cribber]

Hi Martin,

have a browse thru the Carrera records. Something stinks. He got to number 2 by beating Paulo Alejandro Sanchez who has never had a decent win in his life and has 20 losses on the record. Something is not right here.

conan

[conan_the_cribber]

Hi

why did Calzaghe drop 6 places between these fights?

M 14 2004-10-22 005364 019292 W UD 1<-1 57<-54-a
M 14 2005-05-07 005364 016148 W TKO 7<-7 31<-28-a

Doesn't make sense if he was number 1.

conan

... he didn't fight an opponent within rank 45 within 1.5 years

... his new rank is -> 1*2+5 = 7 (before the fight)

Thanks for the info. One side of me is saying serves him right, the other is saying, wow, that could happen to anyone. Opponent pulls out and some late sub is used. A flat +5 is harsh.

conan

[JCS]

Hi

why did Calzaghe drop 6 places between these fights?

M 14 2004-10-22 005364 019292 W UD 1<-1 57<-54-a
M 14 2005-05-07 005364 016148 W TKO 7<-7 31<-28-a

Doesn't make sense if he was number 1.

conan

... he didn't fight an opponent within rank 45 within 1.5 years

... his new rank is -> 1*2+5 = 7 (before the fight)

Thanks for the info. One side of me is saying serves him right, the other is saying, wow, that could happen to anyone. Opponent pulls out and some late sub is used. A flat +5 is harsh.

conan

That is a pretty harsh penalty.

How about (1*1.5)+0.5 = 3

#4 would be (4*1.5)+0.5 = 7

#12 would be (12*1.5)+0.5 = 18.5>19

[computerrank]

Conan,

You: ... You said 'log' in your previous post, so I took you at your word and used log based 10. It was hard to guess that you actually meant the result of a dividing two numbers.

-> logarithmic distance p

- given a and b, a < b
- complete distance is 1
- find distance p
- res = a*exp((ln(b)-ln(a))*p) = a*(b/a)**p



You: ... Regarding, the 50 and 191 point. There is no rule in it the ABC rankings, just the broad, fighters must be rewarded on their activity and their results. Losing four in a row and then losing to a 300th ranked fighter is a devistating statement about where the boxer stands. It could be appropriate to drop him to 191 according to the ABC criteria, although they do not state that multiple successive losses are to be especially penalized, it is implicit and allowed in interpreting a boxers activity and results in the ring. ...

-> It is not my point that you set back the loser such far, but I look at following:

- if you set back the loser such far, he wasn't worth his rank
- so the winner should get the loser's old rank

- or the winner gets the loser's old rank, and the winner is treated normally

- you cannot give the winner a good rank, and say on the other side, it wasn't worth

Best regards
Martin

[computerrank]

Hi Martin,

have a browse thru the Carrera records. Something stinks. He got to number 2 by beating Paulo Alejandro Sanchez who has never had a decent win in his life and has 20 losses on the record. Something is not right here.

conan

Conan,

believe, I walked trough a lot of such tracks, Sanchez also has defeated someone, who "didn't earn this" - in your words.

This is a direct outflow of

- getting a the rank of the loser
- no overshooting for winners
- the metric is overrun by the basic traditional rules

- the outsider defeats the champion and is the king - often how long?

- this is the structure of the traditional rankings

And so I will accept this.

Work on the performance ratings for performance - if this is your intention. Traditional rankings are the wrong way, as they disregard the basic metrics.

[conan_the_cribber]

[conan,

you very precisely identified the essential point for improving precition ratio.

You have to set boxers some margin above the loser, if they were close enough.

But - this also is definitely not intended by the traditional rules. As the boxers above the loser would be touched in their achievements - a wording often mentioned in the ABC rules. And this is not accepted.

Activity is one point (of even more) in the ABC rules to select candidates from nowhere to their rankings. But activity never will set a ranked boxer above a never defeated by him.

So again - I feel, you are trying to find a best performing (whatever this would be) traditional ranking. This is not my goal.

I want to implement a working, objective traditional ranking complying with their goals and rules - and this is definitely not prediction ratio.

By the way, I already implemented a ranking follwing your ideas - with this certain overshooting for the close winner. And it showed very improved prediction rate - up to 74% (no we have 65 %), if I remember correctly.

The basic idea given the ranks were:

- find the gemetric mean of the ranks
- set the winner to a rank of mean / factor; factor about 1.4
- set the loser to a rank of Mean * factor
- factor varying by result

But later I found, I just tried to make the same things with ranks as I before did with the performance points.

I just built some metrics into the rankings, and I could find that a factor in the ranks just is equivalent to some rating difference of the performance ratings.

Do you see it now?

Best regards
Martin

Do you have a linke to the ABC rules. I'v eonly got a draft copy in a my mail account buried somewhere.

I think it would be good if we agreed on what we are trying to achieve here. The WBA paper says something, but they do another. They're always some override and they're stripping people and crowning people at will. The paper is written in 1999 and says boxers whould fight four times a year. It's out of date.

The IBF and WBC have box-offs for the number one and two positions. Boxers participate in these box-offs and get to leap frog other boxers in the ratings. This alone could justify bumping people up for wins against their peers. I'd have to read what the ABC guidelines says to know whether its invalid.

When you say implemented according to me, did you mean with successive losses or with the point about bumping up peers or both?

You answered my post before I wrote my answer about 'best performing'. To paraphrase, I do not seek best performing, but good performing predictability is still desirable as one measure to the deceny of the rankings.

conan

p.s. ABC link, if you got one.

[computerrank]

Conan,

...That is a pretty harsh penalty.

How about (1*1.5)+0.5 = 3

#4 would be (4*1.5)+0.5 = 7

#12 would be (12*1.5)+0.5 = 18.5>19...

->

WBA rules say, set them out of top 15. They often don't obey their rules ... individual (or Business induced?) decisions ...

And this is correct, think: no quality (inside #45 is not much for a #1) opponent (must not be defeated) in 18 months is a criteria - this has to be punished - as you cannot simultainously care for his correct ranking.

Although I tried this earlier. Getting the effect, they are in place very soon again (for inactivities etc). There has to be a certain distance.

But it actually is object to personal decision ...

[computerrank]

conan,

ad ABC general rules, see minutes of convention:

http://www.canadianboxing.com/abcboxing ... _draft.htm

explictly here:


FRIDAY, JULY 29, 2005

CRITERIA FOR RATINGS OF BOXERS - Greg Sirb, ( Pennsylvania ), Flip Homansky, ( Nevada )

ASSOCIATION OF BOXING COMMISSIONS


GUIDELINES FOR OBJECTIVE AND CONSISTENT CRITERIA FOR THE RATINGS OF PROFESSIONAL BOXERS


As prescribed by 15 USC Sec.6307c.(a), the Association of Boxing Commissions is empowered to develop guidelines for objective and consistent written criteria for the ratings of professional boxers. Although the ABC realizes that any ratings structure can be subjective in nature, there is a definite need for all organizations to create a ratings structure that is more objective, more precise and more consistent: based less on subjectivity and more on competition between active boxers; so that when a world title bout is advertised, the general public is assured that it is, indeed, between two deserving boxers.



The basic theory is to have the boxers themselves prove they are worthy of a top rating.[1] Movement of these top boxers should be based, upon wins and losses and not on subjective criteria which currently differs greatly among the sanctioning organizations. For boxers to maintain their top rating, they should be, at the very least, active and competing against other ranked opponents. This type of rating system will increase activity among the top boxers of the world. A sanctioning organization may decide how many fighters to rate in a Division.



As such, the following ratings criteria were adopted by the ABC in 2000 and further amended in 2002:



1) Ratings must be based solely on win/loss records, level of competition and activity. The record of any top rated boxer shall be verified by the ABC’s official boxing registry.



2) No boxer can be rated in more than one division.



3) For a boxer to receive a top rating, the boxer shall have competed in at least (2) bouts scheduled for ten rounds or (4) bouts scheduled for eight rounds.



4) In order to maintain a top rating, the boxer:



a) shall compete at least once during a 12 month period from the time the boxer gets rated; and



b) shall have competed against another top rated boxer within an 18 month period from the time the boxer gets rated.

A boxer who does not meet this level of competition shall not retain the boxer’s rating. Exceptions can be made only for injuries, verified in writing by a licensed physician (M.D. or D.O.).



5) If a top rated boxer loses to an un-rated boxer:



a) the top rated boxer shall be lowered at least one position in the ratings; and



b) the un-rated boxer shall be considered for a rating somewhere in the top rating.



6) If two boxers are rated and compete against each other, and the lower rated boxer wins:



a) The lower rated boxer shall be elevated in the ratings and;



b) The higher rated boxer shall be lowered in the ratings.

[JCS]

There sure is a lot left for interpretation with those extremely broad rules.

[computerrank]

There sure is a lot left for interpretation with those extremely broad rules.

... you are so right ...

They should be as simple as possible, but exact and don't show any possibility for interpretation.

Begin ...

[conan_the_cribber]

Hi Martin,

have a browse thru the Carrera records. Something stinks. He got to number 2 by beating Paulo Alejandro Sanchez who has never had a decent win in his life and has 20 losses on the record. Something is not right here.

conan

Conan,

believe, I walked trough a lot of such tracks, Sanchez also has defeated someone, who "didn't earn this" - in your words.

This is a direct outflow of

- getting a the rank of the loser
- no overshooting for winners
- the metric is overrun by the basic traditional rules

- the outsider defeats the champion and is the king - often how long?

- this is the structure of the traditional rankings

And so I will accept this.

Work on the performance ratings for performance - if this is your intention. Traditional rankings are the wrong way, as they disregard the basic metrics.

Hi Martin,

I'm sure you dig thru a bunch of tracks and it is a time consuming process. I'm sure you do this to make sure your algorithm works. However, four eyes are better than two.

I dug back pretty hard, and it boiled down to the first win in the trace against 47964 Zanabria. I cant for the life of me determine why Zanabria was 19 in 2002. Nothing but draws against his record. At that point Valezco hadn't established himself, nor had any of his opponents. Zanabria's win against 18-0 Clavero couldn't have promoted him either, Clavero hadnt beaten anyone either.

It is a very odd chain that gets Canerra up there. Once he gets to 19 he gets a boost from Valezco, although Valezco at that point also has no real reason to be 10.

I can imagine at the end how Canerra improves a few points by attrition now and then, but the path there is pretty obscure.

conan

[conan_the_cribber]

There sure is a lot left for interpretation with those extremely broad rules.

... you are so right ...

They should be as simple as possible, but exact and don't show any possibility for interpretation.

Begin ...

Nothing there in this general stuff that prevents bumping up if you win against a peer. A win against a marginally lower ranked fighter is not mentioned.

conan

[computerrank]

There sure is a lot left for interpretation with those extremely broad rules.

... you are so right ...

They should be as simple as possible, but exact and don't show any possibility for interpretation.

Begin ...

Nothing there in this general stuff that prevents bumping up if you win against a peer. A win against a marginally lower ranked fighter is not mentioned.

conan

... indeed ...

[JCS]

I just took a wack at a ranking-type system, although I'm not strictly following any ABC rules. I'll just post heavy.


This is as of 8/3 (Before Mora vs. Johnson was changed to a NC, and yes, I know its still kinda messed up)

Its definitely a work in progress, but I bet some organization could pass this off as their ratings.

Heavy

1 Fres Oquendo
2 Wlad Klitschko
3 Javier Mora
4 Samuel Peter
5 Sergei Lyakhovich
6 Kirk Johnson
7 Lamon Brewster
8 Matt Skelton
9 Oleg Maskaev
10 Luan Krasniqi
11 Vladimir Virchis
12 Danny Williams
13 Chris Byrd
14 Ruslan Chagaev
15 Sinan Samil Sam
16 Paolo Vidoz
17 Michael Sprott
18 Tony Thompson
19 Eddie Chambers
20 Hasim Rahman

Notables such as Briggs at #141, JD Chapman at #163

[conan_the_cribber]

I just took a wack at a ranking-type system, although I'm not strictly following any ABC rules. I'll just post heavy.


This is as of 8/3 (Before Mora vs. Johnson was changed to a NC, and yes, I know its still kinda messed up)

Its definitely a work in progress, but I bet some organization could pass this off as their ratings.

Heavy

1 Fres Oquendo
2 Wlad Klitschko
3 Javier Mora
4 Samuel Peter
5 Sergei Lyakhovich
6 Kirk Johnson
7 Lamon Brewster
8 Matt Skelton
9 Oleg Maskaev
10 Luan Krasniqi
11 Vladimir Virchis
12 Danny Williams
13 Chris Byrd
14 Ruslan Chagaev
15 Sinan Samil Sam
16 Paolo Vidoz
17 Michael Sprott
18 Tony Thompson
19 Eddie Chambers
20 Hasim Rahman

Good on you for trying WCS. That could've been the WBO. I cant for the life of think why Fres ended up on top, but hey, the rest aint that bad.

conan

[JCS]

I just took a wack at a ranking-type system, although I'm not strictly following any ABC rules. I'll just post heavy.


This is as of 8/3 (Before Mora vs. Johnson was changed to a NC, and yes, I know its still kinda messed up)

Its definitely a work in progress, but I bet some organization could pass this off as their ratings.

Heavy

1 Fres Oquendo
2 Wlad Klitschko
3 Javier Mora
4 Samuel Peter
5 Sergei Lyakhovich
6 Kirk Johnson
7 Lamon Brewster
8 Matt Skelton
9 Oleg Maskaev
10 Luan Krasniqi
11 Vladimir Virchis
12 Danny Williams
13 Chris Byrd
14 Ruslan Chagaev
15 Sinan Samil Sam
16 Paolo Vidoz
17 Michael Sprott
18 Tony Thompson
19 Eddie Chambers
20 Hasim Rahman

Good on you for trying WCS. That could've been the WBO. I cant for the life of think why Fres ended up on top, but hey, the rest aint that bad.

conan

Well the Kirk Johnson debacle w/ Mora, leapfrogged Mora over Johnson, than leapfrogged Oquendo over Mora.

There's an issue w/ Johnson holding on too long there...

[computerrank]

... and I have this ...

OK, I learned my lesson - thanks. As simple as possible is the requirement, I am sure now after all discussions.

The traditional rankings work for their simplicity - no goal on predicition - simply take the rank of the better ranked by a win. Transparent and basta.

All chasing for the details and performance ratio is dedicated to the performance ratings.

5 simple rules

separate for every division - and the result is great - this will be published as robust, simple, traditional rankings:

R1) winner was better ranked than loser
-> no change

R2) winner was worse ranked than loser
-> winner gets loser's rank
-> loser gets geometric mean rank of winner and loser

R3) draw
-> no change

R4) inactivity
-> after 18 months of inactivity the boxer is removed from the division
-> his background ranking is multilpied with factor 3 for every complete inactivity period of 18 months
-> in case of come back, the background ranking is activated and 1 reank added

R5) missing quality of opponents
-> after 18 months of missing an opponent within factor 3 of own rank, quality is missing
-> for top 15 the tolerance area is within rank 45
-> the boxer's rank is multiplied with 2 and 1 rank added for missing quality of opponents

Result:

M Heavyweight 1 Nikolay Valuev 69
M Heavyweight 2 Hasim Rahman 146
M Heavyweight 3 Wladimir Klitschko 111
M Heavyweight 4 John Ruiz 237
M Heavyweight 5 Calvin Brock 48
M Heavyweight 6 Chris Byrd 111
M Heavyweight 7 Monte Barrett 363
M Heavyweight 8 Matt Skelton 34
M Heavyweight 9 Serguei Lyakhovich 132
M Heavyweight 10 James Toney 146
M Heavyweight 11 Lamon Brewster 132
M Heavyweight 12 Zuri Lawrence 167
M Heavyweight 13 DaVarryl Williamson 97
M Heavyweight 14 Kevin McBride 132
M Heavyweight 15 Samuel Peter 105
M Heavyweight 16 Danny Williams 34
M Heavyweight 17 Damian Wills 64
M Heavyweight 18 Tony Thompson 44
M Heavyweight 19 Ruslan Chagaev 27
M Heavyweight 20 Larry Donald 314

M Cruiserweight 1 O'Neil Bell 216
M Cruiserweight 2 Johnny Nelson 258
M Cruiserweight 3 Jean Marc Mormeck 216
M Cruiserweight 4 Guillermo Jones 342
M Cruiserweight 5 Wayne Braithwaite 342
M Cruiserweight 6 Enzo Maccarinelli 34
M Cruiserweight 7 Vincenzo Cantatore 83
M Cruiserweight 8 Steve Cunningham 216
M Cruiserweight 9 David Haye 21
M Cruiserweight 10 Alexander Gurov 238
M Cruiserweight 11 Aloryi Moyoyo Mensah 91
M Cruiserweight 12 Krzysztof Wlodarczyk 41
M Cruiserweight 13 Dale Brown 239
M Cruiserweight 14 Antonio Mercado 181
M Cruiserweight 15 Marco Huck 69
M Cruiserweight 16 Mark Hobson 71
M Cruiserweight 17 Vincenzo Rossitto 27
M Cruiserweight 18 Patrick Nwamu 239
M Cruiserweight 19 Luis Andres Pineda 244
M Cruiserweight 20 Grigory Drozd 13

M Light Heavyweight 1 Zsolt Erdei 13
M Light Heavyweight 2 Glen Johnson 168
M Light Heavyweight 3 Roy Jones Jr 13
M Light Heavyweight 4 Clinton Woods 90
M Light Heavyweight 5 Julio Gonzalez 22
M Light Heavyweight 6 George Khalid Jones 315
M Light Heavyweight 7 Tomasz Adamek 300
M Light Heavyweight 8 Chad Dawson 70
M Light Heavyweight 9 Gabriel Campillo 251
M Light Heavyweight 10 Stipe Drews 76
M Light Heavyweight 11 Paul Briggs 300
M Light Heavyweight 12 Thomas Ulrich 13
M Light Heavyweight 13 Mehdi Sahnoune 293
M Light Heavyweight 14 Hugo Hernan Garay 125
M Light Heavyweight 15 Montell Griffin 133
M Light Heavyweight 16 Kai Kurzawa 76
M Light Heavyweight 17 Julio Cesar Dominguez 23
M Light Heavyweight 18 Adrian Diaconu 87
M Light Heavyweight 19 Jason DeLisle 90
M Light Heavyweight 20 Yuri Barashian 301

M Super Middleweight 1 Markus Beyer 195
M Super Middleweight 2 Joe Calzaghe 160
M Super Middleweight 3 Mikkel Kessler 209
M Super Middleweight 4 Anthony Mundine 86
M Super Middleweight 5 Jeff Lacy 160
M Super Middleweight 6 Peter Manfredo Jr 179
M Super Middleweight 7 Danny Green 86
M Super Middleweight 8 Cristian Sanavia 15
M Super Middleweight 9 Robin Reid 370
M Super Middleweight 10 Omar Sheika 342
M Super Middleweight 11 Mger Mkrtchian 119
M Super Middleweight 12 Carl Froch 77
M Super Middleweight 13 Lucian Bute 87
M Super Middleweight 14 Brian Magee 77
M Super Middleweight 15 Scott Pemberton 179
M Super Middleweight 16 Mario Veit 76
M Super Middleweight 17 Alejandro Berrio 84
M Super Middleweight 18 Vitali Tsypko 266
M Super Middleweight 19 Librado Andrade 125
M Super Middleweight 20 Robert Stieglitz 251

Thanks for your discussion and inspiration
Martin