Ratings - please read before commenting - Archived

[computerrank]

@conan

The key is the tracking table.

- It holds all pre_bout ratings of the opponents of the last 18 months
- all bouts are deleted, where a newer pre_bout rating is equal or higher than the deleted one
- so the entries are sorted descending in the time domain

When a boxer's rating is no longer covered by the highest entry in the table (50% rule), then:

- the boxer's rating is reduced according to the highest entry available
- the table collapses to 1 entry with a value of 50% of the new rating - and the date of the reduction

When a boxers's rating increase after a bout is not covered by the highest value in the table, then:

- the boxer's rating is increased by a reduced value according to the highest entry available
- the table collapses to 1 entry with a value of 50% of the new rating - and the date of the bout

Best regards
Martin

[keithmoonhangover]

Why has Denis Boytsov been moved down 13 places for winning a fight be knockout.

[computerrank]

Why has Denis Boytsov been moved down 13 places for winning a fight be knockout.

He lost his points from 344 to 200 already on 2010-12-06 - not with this bout.

On 2010-12-06 his good win against Bydenko with a pre-bout rating of 311 got out of scope (last 18 months). Now he had only week opponents - best Gavern with a pre-bout rating of 84. So Boytsov lost his points.

I don't know, why the website is showing the move now.

[keithmoonhangover]

Why has Denis Boytsov been moved down 13 places for winning a fight be knockout.

He lost his points from 344 to 200 already on 2010-12-06 - not with this bout.

On 2010-12-06 his good win against Bydenko with a pre-bout rating of 311 got out of scope (last 18 months). Now he had only week opponents - best Gavern with a pre-bout rating of 84. So Boytsov lost his points.

I don't know, why the website is showing the move now.

Anyway you look at it, he doesn't deserve to go down 13 places after last night.

[conan_the_cribber]

@conan

The key is the tracking table.

- It holds all pre_bout ratings of the opponents of the last 18 months
- all bouts are deleted, where a newer pre_bout rating is equal or higher than the deleted one
- so the entries are sorted descending in the time domain

When a boxer's rating is no longer covered by the highest entry in the table (50% rule), then:

- the boxer's rating is reduced according to the highest entry available
- the table collapses to 1 entry with a value of 50% of the new rating - and the date of the reduction

When a boxers's rating increase after a bout is not covered by the highest value in the table, then:

- the boxer's rating is increased by a reduced value according to the highest entry available
- the table collapses to 1 entry with a value of 50% of the new rating - and the date of the bout

Best regards
Martin

Hi Martin,

thanks for taking the time to answer and describe the mechanism, but still there is no definite answer as to why the Briggs fight does not result in a pre-fight reduction. You do not describe the connection between the new entries in the tracking table and the protection of Vitali's ranking despite fighting the poorly ranked Briggs.

I assume that after one reduction for poor opponents in 18 month period, then your table contains a new date, with 50% of the boxer's ranking. The next time possible pre-fight reduction then is, in 18 months from the new first entry in the tracking table. Is that correct?

I would've assumed that the lack of quality opposition is continuous i.e. every crap fight reduces your ranking bit by bit and not done in intervals.


conan

[computerrank]

Hi Martin,

...

I assume that after one reduction for poor opponents in 18 month period, then your table contains a new date, with 50% of the boxer's ranking. The next time possible pre-fight reduction then is, in 18 months from the new first entry in the tracking table. Is that correct?

I would've assumed that the lack of quality opposition is continuous i.e. every crap fight reduces your ranking bit by bit and not done in intervals.


conan

Yes, it is correct.

The tracing table contents:

- (824, 2008-10-11) after the Peter bout
- (824, 2008-10-11; 572, 2009-03-21) after the Gomez bout
- (824, 2008-10-11; 572, 2009-03-21; 418, 2009-09-26) after the Arreola bout
- (824, 2008-10-11; 572, 2009-03-21; 418, 2009-09-26; 243 2009-12-12) after the Johnson bout
- (606, 2010-03-11) after the Peter bout got out of scope on 2010-03-11 and the rating was reduced to 1213, based on 572 points, the tracked best opponent rating from the Gomez bout
- (608, 2010-05-29) after the Sonsnowski bout, caused by reduced points win not matched by the 606 points tracked best opponent rating before the bout
- (608, 2010-05-29; 126 2010-10-16) after the Briggs bout
- (621, 2011-03-19) after the Solis bout, caused by reduced points win not matched by the 608 points tracked best opponent rating before the bout
- (651, 2011-03-19) after the Adamek bout, caused by reduced points win not matched by the 621 points tracked best opponent rating before the bout

Otherwise a former high rated boxer,

- who starts to frequently fight only underlevel opponents would reduce his rating to nothing in a short time
- who starts to infreuquently fight only underlevel opponents would reduce his rating much lesser

A boxer only boxing underlevel opponents for a while practically shows the same achievements as an inactive boxer. This is punished with a rating reduction of 50% per 18 months.

[Asterix]

Why has Denis Boytsov been moved down 13 places for winning a fight be knockout.

He lost his points from 344 to 200 already on 2010-12-06 - not with this bout.

On 2010-12-06 his good win against Bydenko with a pre-bout rating of 311 got out of scope (last 18 months). Now he had only week opponents - best Gavern with a pre-bout rating of 84. So Boytsov lost his points.

I don't know, why the website is showing the move now.

Anyway you look at it, he doesn't deserve to go down 13 places after last night.

I agree. I like the fact that he is penalised for his poor opposition, but perhaps his age should be considered? He's very unlikely to be any worse than when he beat Bydenko.

[computerrank]

@Asterix

It might well be he isn't worse now.

But the Boxrec Ratings require to fight a valid opponent within the last 18 months. And if a boxer doesn't match a valid opponent, the rating will be reduced.

The Boxrec Ratings shall represent a confirmed status - independent of age and other criteria.

[conan_the_cribber]

@Asterix

It might well be he isn't worse now.

But the Boxrec Ratings require to fight a valid opponent within the last 18 months. And if a boxer doesn't match a valid opponent, the rating will be reduced.

The Boxrec Ratings shall represent a confirmed status - independent of age and other criteria.

Sounds fair to me.

[conan_the_cribber]

HI Martin,

had a look at Yuriorkis Gamboa and found a couple of anomalies.

http://boxrec.com/list_bouts.php?human_ ... &cat=boxer

Why does his pre-bout rating go up before his fight Nestor Hugo Paniagua?
Why, in his next bout, does his pre-bout rating do down before fighting his best opponent to date in Samual Kebebe?
Why, although he was just penalised in his previous bout, does his pre-fight rating go up for his next fight for the equivalently strong Adailton de Jesus?
Why does he get 0 pts for a win against Gilberto Luque? Clearly, with a ratio of 110 to 25, he was a shit hot favourite. But a win is a win. For comparison, he got 3 pts for a win against Thomas Hengstberger although the ratio was 46 to 2.

As all these bouts took place within his first year, the inactivity rules cannot apply. There are probably some tricks to do with boxers establishing themselves, but the ups and downs are pretty hard to explain to a casual reader.

conan

[conan_the_cribber]

Hi Martin,

I have to report a serious flaw in the algorithm. I believe that the non-zero sum part of the algorithm leads to ratings inflation in the heavyweight division. Not because of the weight, but because the most bouts are in the heavyweight division. I stumbled upon this as I was browsing the top 25 heavyweights. Fighters like Arreola are ranked way above several top fighters in the lower divisions. Clearly as far as boxing skills go and career record go, he should be below them.

So I started to look for the reasons.

Then the story with the woman's boxing came back to me. The reason that their mega-fight ratings were so low, was that there are less bouts in the db per weight division. That is of course true in men's boxing too. There are probably more heavyweight bouts than others at the moment. And because the algorithm is not "zero sum", that means for each bout in a division, there is a slight plus going on.

Effectively this invalidates this algorithm for p4p comparisons. It also makes divisional changes questionable.

If you ask me what the solution is, then I would say, that the best algorithm has to be zero-sum. I would also say, that boxers can have a negative rating and that the distance between the rating is the most important thing.

Can you give a breakdown of the bouts in the db by weight class?

Thanks

conan

[conan_the_cribber]

Hi Martin,

you have my attention at the moment. Why are DQ's victories worth so little? I looked at Arreola's record and saw that he got more points for beating washed up Cliff Couser, than he did for beating the equivalently strong Chas Witherspoon. The only reason can be the nature of the victory i.e. DQ. Although a DQ is not always as clear-cut as a KO or dominant PTS decision, it is still a pretty reasonable indication of dominance. Usually the DW'd boxer is looking for a way out of the fight. Without doing the calculations, I imagine if the victory was via KO, then Arreola would've earned about 160 pts. Because of the DQ, he only got 30. That is around 20% of what he potentially could've got. That seems really low.

http://boxrec.com/list_bouts.php?human_ ... &cat=boxer

What are the exact rules for DQ?

Thanks

conan

[computerrank]

@conan

I will come back to your questions. This will need some time.

Best regards
Martin

[conan_the_cribber]

@conan

I will come back to your questions. This will need some time.

Best regards
Martin

@Martin, it's all good. Take your time.

1)
I can answer my own question about the DQ now. I see that you've lumped it in with the other irregular results MD, SD, TD.

MD and SD have a "clear decision" of 0,5 which is fine. However I don't think that applies to DQ. Most DQ's aren't close matchups, or at least they're turning at the point someone got DQed. At a guess, most DQ's probably would end in KO. I think interpretting with a higher cd quotient and as a KO would be more intuitive. I mean, getting somebody to foul out in the 3rd round, is a lot more convincing that getting them to foul out in the 10th round. I assume KO3 is better than KO10, so I'd say the DQ should go along the same lines.

Certainly in betting, W via DQ is lumped in with W via TKO.

conan

[conan_the_cribber]

There is a small typo on the Ratings decription page

Launch states are:

1. 0 = no recent win
2. 1 = 1 recent win, 2 = 2 recent wins, 3 = 3 recent wins
3. 4 = recent win over opponent with state 3, 5 = 2 recent wins over opponent with state 3, 6 = 3 recent wins over opponent with state 3
4. 7 = recent win over opponent with state 6, 8 = 2 recent wins over opponent with state 6, 9 = 3 recent wins over opponent with state 6
5. 10 = recent win over opponent with state 9, 8 = 2 recent wins over opponent with state 9, 9 = 3 recent wins over opponent with state 9
6. 13 = recent win over opponent with state 12, 14 = 2 recent wins over opponent with state 12, 15 = 3 recent wins over opponent with state 12

Point 5 is wrong. Correct would be

10 = recent win over opponent with state 9, 11 = 2 recent wins over opponent with state 9, 12 = 3 recent wins over opponent with state 9

Also, there is no definition of "recent" in the description.

And while I'm bitching around here .....

opponent in launch state n: v * cd * (25*n - r_a) * max(6*(n+1),min(r_b,25*n)) / (25*n) * max(r_a,18) / (max(r_a,18) + max(r_b,18))

....this formula is just imfukkingpossible to understand. For starters are we talking about r_a_old or r_a_new. I assume old. OK, I'll stop being lazy....here's an attempt to explain it.

n: v * cd * (25*n - r_a) * max(6*(n+1),min(r_b,25*n)) / (25*n) * max(r_a,18) / (max(r_a,18) + max(r_b,18))

v = effectively length of fight compared to 12 rounds.
cd = clear decision index. 1 for a clear result, 0.5 for MD, SD and 0 for a draw.

rearranging it....

Code: Select all

v * cd * (25*n - r_a) * max(6*(n+1), min(r_b, 25*n))
------------------------------------------------------
25*n * max(r_a, 18) / max(r_a, 18) + 

Oh I give up....a bracket is missing. There is one too many closing brackets.

conan

[Asterix]

HI Martin,

had a look at Yuriorkis Gamboa and found a couple of anomalies.

http://boxrec.com/list_bouts.php?human_ ... &cat=boxer

Why does his pre-bout rating go up before his fight Nestor Hugo Paniagua?
Why, in his next bout, does his pre-bout rating do down before fighting his best opponent to date in Samual Kebebe?
Why, although he was just penalised in his previous bout, does his pre-fight rating go up for his next fight for the equivalently strong Adailton de Jesus?
Why does he get 0 pts for a win against Gilberto Luque? Clearly, with a ratio of 110 to 25, he was a shit hot favourite. But a win is a win. For comparison, he got 3 pts for a win against Thomas Hengstberger although the ratio was 46 to 2.

As all these bouts took place within his first year, the inactivity rules cannot apply. There are probably some tricks to do with boxers establishing themselves, but the ups and downs are pretty hard to explain to a casual reader.

conan

For the first few questions, it's because it weight class kept changing. The editors can often decide how well a boxer is rewarded for a win simply by placing him in a different weight class.

[conan_the_cribber]

HI Martin,

had a look at Yuriorkis Gamboa and found a couple of anomalies.

http://boxrec.com/list_bouts.php?human_ ... &cat=boxer

Why does his pre-bout rating go up before his fight Nestor Hugo Paniagua?
Why, in his next bout, does his pre-bout rating do down before fighting his best opponent to date in Samual Kebebe?
Why, although he was just penalised in his previous bout, does his pre-fight rating go up for his next fight for the equivalently strong Adailton de Jesus?
Why does he get 0 pts for a win against Gilberto Luque? Clearly, with a ratio of 110 to 25, he was a shit hot favourite. But a win is a win. For comparison, he got 3 pts for a win against Thomas Hengstberger although the ratio was 46 to 2.

As all these bouts took place within his first year, the inactivity rules cannot apply. There are probably some tricks to do with boxers establishing themselves, but the ups and downs are pretty hard to explain to a casual reader.

conan

For the first few questions, it's because it weight class kept changing. The editors can often decide how well a boxer is rewarded for a win simply by placing him in a different weight class.

Nice pick up Asterix . I was obsessing about the deduction algorithm.

conan

[conan_the_cribber]

I've been thinking about the ratings inflation problem and how to quantify it. If you are adding points to a division because of non-zero sum aspect, then you need to somehow track it during your calculations. Perhaps you could add a test counter that sums up the inflation points i.e. difference between r_a_new + r_b_new and r_a_old and r_old

You could then see how many fights are in the division and then the ratio to come up with an inflation factor per division.

inflation points
---------------
number of bouts

You could then use this ratio to multiply out the ratings in the divisions, to get a better comparison between the divisions.

Just an idea.

conan

[computerrank]

HI Martin,

had a look at Yuriorkis Gamboa and found a couple of anomalies.

http://boxrec.com/list_bouts.php?human_ ... &cat=boxer

...

Why does he get 0 pts for a win against Gilberto Luque? Clearly, with a ratio of 110 to 25, he was a shit hot favourite. But a win is a win. For comparison, he got 3 pts for a win against Thomas Hengstberger although the ratio was 46 to 2.

...

conan

You already cited the launch process = additional points ... Herewith a boxer can get additional points outside a ratio of 1:4

[conan_the_cribber]

HI Martin,

had a look at Yuriorkis Gamboa and found a couple of anomalies.

http://boxrec.com/list_bouts.php?human_ ... &cat=boxer

...

Why does he get 0 pts for a win against Gilberto Luque? Clearly, with a ratio of 110 to 25, he was a shit hot favourite. But a win is a win. For comparison, he got 3 pts for a win against Thomas Hengstberger although the ratio was 46 to 2.

...

conan

You already cited the launch process = additional points ... Herewith a boxer can get additional points outside a ratio of 1:4

Thanks Martin. But the launch process formula is really hard to understand. It is also hard to understand the intent behind it. Please correct the formula so that I can analyse it some more. I assume the numbers 25 and 18 used in the launch equation were found by trial and error.

cheers

Ian

[computerrank]

Hi Martin,

I have to report a serious flaw in the algorithm. I believe that the non-zero sum part of the algorithm leads to ratings inflation in the heavyweight division. Not because of the weight, but because the most bouts are in the heavyweight division. I stumbled upon this as I was browsing the top 25 heavyweights. Fighters like Arreola are ranked way above several top fighters in the lower divisions. Clearly as far as boxing skills go and career record go, he should be below them.

So I started to look for the reasons.

Then the story with the woman's boxing came back to me. The reason that their mega-fight ratings were so low, was that there are less bouts in the db per weight division. That is of course true in men's boxing too. There are probably more heavyweight bouts than others at the moment. And because the algorithm is not "zero sum", that means for each bout in a division, there is a slight plus going on.

Effectively this invalidates this algorithm for p4p comparisons. It also makes divisional changes questionable.

If you ask me what the solution is, then I would say, that the best algorithm has to be zero-sum. I would also say, that boxers can have a negative rating and that the distance between the rating is the most important thing.

Can you give a breakdown of the bouts in the db by weight class?

Thanks

conan

all bouts since 2000-01-01:

Code: Select all

+---------------------+-------+
| division            | n     |
+---------------------+-------+
| Heavyweight         | 16340 |
| Cruiserweight       | 10618 |
| Light Heavyweight   |  9158 |
| Super Middleweight  | 11085 |
| Middleweight        | 13688 |
| Light Middleweight  | 14991 |
| Welterweight        | 18138 |
| Light Welterweight  | 16877 |
| Lightweight         | 17283 |
| Super Featherweight | 12437 |
| Featherweight       | 14510 |
| Super Bantamweight  | 11342 |
| Bantamweight        | 10736 |
| Super Flyweight     |  7040 |
| Flyweight           |  8615 |
| Light Flyweight     |  5867 |
| Minimumweight       |  3456 |
+---------------------+-------+

[computerrank]

...


1)
I can answer my own question about the DQ now. I see that you've lumped it in with the other irregular results MD, SD, TD.

MD and SD have a "clear decision" of 0,5 which is fine. However I don't think that applies to DQ. Most DQ's aren't close matchups, or at least they're turning at the point someone got DQed. At a guess, most DQ's probably would end in KO. I think interpretting with a higher cd quotient and as a KO would be more intuitive. I mean, getting somebody to foul out in the 3rd round, is a lot more convincing that getting them to foul out in the 10th round. I assume KO3 is better than KO10, so I'd say the DQ should go along the same lines.

Certainly in betting, W via DQ is lumped in with W via TKO.

conan

A DQ can go either way - some are caused by superiority of the opponent, yes - but you never know.

So the basic value is 50% and I go with scorecards, whenever possible.

[computerrank]

There is a small typo on the Ratings decription page

...

Point 5 is wrong. Correct would be

10 = recent win over opponent with state 9, 11 = 2 recent wins over opponent with state 9, 12 = 3 recent wins over opponent with state 9

...

Also, there is no definition of "recent" in the description.

....

rearranging it....



....

conan

- error in description corrected
- recent means, no losses against inferior opponents in between

all brackets were ok:

Code: Select all

v * cd * (25*n - r_a) * max(6*(n+1), min(r_b, 25*n)) * max(r_a, 18) 
------------------------------------------------------------------- 
25*n  * (max(r_a,18) + max(r_b,18))

[computerrank]

I've been thinking about the ratings inflation problem and how to quantify it. If you are adding points to a division because of non-zero sum aspect, then you need to somehow track it during your calculations. Perhaps you could add a test counter that sums up the inflation points i.e. difference between r_a_new + r_b_new and r_a_old and r_old

You could then see how many fights are in the division and then the ratio to come up with an inflation factor per division.

inflation points
---------------
number of bouts

You could then use this ratio to multiply out the ratings in the divisions, to get a better comparison between the divisions.

Just an idea.

conan

Boxers also leave the divisions for inactivity or for ever ...

The launch process (additional points) gives the ratings an initial blow up (degrading with the rating level achieved) and allows some steady state spread of the ratings.

[conan_the_cribber]

...


1)
I can answer my own question about the DQ now. I see that you've lumped it in with the other irregular results MD, SD, TD.

MD and SD have a "clear decision" of 0,5 which is fine. However I don't think that applies to DQ. Most DQ's aren't close matchups, or at least they're turning at the point someone got DQed. At a guess, most DQ's probably would end in KO. I think interpretting with a higher cd quotient and as a KO would be more intuitive. I mean, getting somebody to foul out in the 3rd round, is a lot more convincing that getting them to foul out in the 10th round. I assume KO3 is better than KO10, so I'd say the DQ should go along the same lines.

Certainly in betting, W via DQ is lumped in with W via TKO.

conan

A DQ can go either way - some are caused by superiority of the opponent, yes - but you never know.

So the basic value is 50% and I go with scorecards, whenever possible.

Martin, sometimes you amaze me. For someone who puts so much effort into this thing, I cannot for the life of me imagine why you are sometimes so immune to such realistic considerations.

I bet you could do a search on the db and give me a bunch of DQ's. I would say, without any doubt, that for the bouts that have a bout description, that 80% of them detail how the guy who got DQ'd was on his way out. It is more likely to be 90%. Sure there might be a high profile fight where Roy Jones DQ's Montell Griffin, but even that should affect the ratings. Jones got DQ'd fairly. As did Artur Abraham. If you can't complete the fight because you fought unfairly, then that is a loss and should be evaluated as a loss like a KO or a PTS.

Do you really think that 50% of the DQ's are debatable? Not a snow flakes chance in hell that they are.

conan