Ratings - please read before commenting - Archived

[computerrank]

Joshua lost to Ruiz, but Joshua defeated W. Klitschko, Povetkin, Whyte, Parker, Breazeale, Martin, all of them better than Ruiz' other defeated opponents. And Ruiz lost to Parker. Ruiz has to confirm his result. Until then Joshua may well be recognized a member in an other league.

I think this supports Joshua being ranked higher on an all-time list... but not on a current list.

Right now Ruiz is #1 or #3 for me... I'm sure that applies on any respected non-software ranking.... Find me a respected list that has Ruiz lower than #3.

One perspective is, what would be accepted by the boxing community.

The other perspective is, what is probable regarding the achievements.

Regarding the figures and the imponderabilities of every bout Joshua is clearly above Ruiz regarding the recent achievements,

And - to back your point - I doubt, the boxing community would enjoy this consideration.

[JCS]

One perspective is, what would be accepted by the boxing community.

The other perspective is, what is probable regarding the achievements.

Regarding the figures and the imponderabilities of every bout Joshua is clearly above Ruiz regarding the recent achievements,

And - to back your point - I doubt, the boxing community would enjoy this consideration.

The boxing community have always struggled with a loser being below a winner of a bout. Any other sport...common sense prevails. However I do accept that boxers don’t fight very often so there is a difference there.

Any other non-combat sport.. there are much more contests in a small time window... right? That's why it is acceptable.

[JCS]

Any other non-combat sport.. there are much more contests in a small time window... right? That's why it is acceptable.

There are some interesting MMA ones that use a choice of algorithms. Fight Matrix gives a choice of 4 unless its p4p with the variant “modified elo” being the best I think. Seems to work quite well. Point being it’s difficult with combat sports but not impossible.

http://www.fightmatrix.com/mma-ranks/

I'm the creator of Fight Matrix system.. I know exactly how it works

That's why the other systems (Elo, Glicko, etc) are just that... alternates you can pick, not the primary. They're so damn close.. I'm probably going to stop there.

I'd never ever replace the primary system that I came up with.. with one of these other systems that don't fit the mold. They're purely mathematical and the more you tweak them, the more you end up with a BoxRec or a FightMatrix system.... systems meant for combat sports.

[JCS]

I'm the creator of Fight Matrix system.. I know exactly how it works

That's why the other systems (Elo, Glicko, etc) are just that... alternates you can pick, not the primary. They're so damn close.. I'm probably going to stop there.

I'd never ever replace the primary system that I came up with.. with one of these other systems that don't fit the mold. They're purely mathematical and the more you tweak them, the more you end up with a BoxRec or a FightMatrix system.... systems meant for combat sports.

Well I never knew that! In that case well done. I like looking through your website. From a purely mathematical point of view which variant is the best? Is it not possible to have a p4p on modified elo?

Strangely I personally think the modified Elo rankings are better on the eye than the original though, outside of the big names, I don't recognise all the fighters.

I used to run a football ranking site many years ago called Mancrank. For a while I think it was the best out there but then Euro Club Index improved on it then Clubelo improved on it a lot and I couldn't keep up so stopped. Nowadays 538 has taken it to another level again with expected goals incorporated.

Small world

Check the FAQ on what "Modified Elo" really means. Quite frankly, the core system here (as far as I know) and on FM is based on Elo at the core... but with a shitload of code surrounding it to handle all the weirdness you get in combat sports that you don't see in things like tennis, chess, golf, etc. Issues with data, scheduling, matchmaking, divisions, variable outcomes.. on and on and on.

If I really wanted to improve on my system, I'd probably have to go get a PhD in Math and start over from scratch... MMA is even more unpredictable than boxing... and I don't think I'll do all that in my lifetime.

[mtpattonar]

I'm confused on a few points on the ratings explanation page:

1. "The rating of a boxer is reduced by 50% for every time period of inactivity of 18 months - and to even less in proportion to a longer time period of inactivity"
I can't get the math to work out in the example. The example says the boxer has been out for 23 months. 50% reduction for 18 months = 2.78% per month, times 23 = 63.9% reduction, or a rate of 46.1%
Or if the formula is 50% reduction over 18 months, then an additional 50% of the remainder (or 25% of the original) over 18 months ( 1.39 percent monthly), then 50% + 5*1.39% = 56.95% reduction, or a current rate of 43.05%. Yet example 2 gives a final resolution of 41%, which doesn't jibe with either of my calculations.

2. "The pre-bout rating of a successfully debuting boxer is set to 25% of his opponents pre-bout rating."
Not sure how to interpret this. My best guess: a boxer who wins his first fight has his rating set at 25% of the rating of his opponent, but the opponent receives no penalty. Then, all adjustments are calculated and applied as with any other bout. Would that be correct?

3. "The winner is always rated higher than the loser. "
Surely this is a misstatement? Otherwise, it says that if a low rated boxer manages to upset a high rated boxer, the low rated will then have the higher rating, regardless of how much distance was previously between them or how much experience they have. Thus, if the worst boxer on the planet pulls the enormous upset of defeating the best boxer on the planet, then suddenly the worst guy is now in the top 50%, and the champion is suddenly in the bottom 50%.
That would seem to be an extremely harsh penalty for only one bad fight, and not realistic. Or have I misinterpreted.

4. I tried to post this as a new topic, but I don't see where to do that. Is this session of the forum not allowed to do that?

Thank you for your time.

[mtpattonar]

On the ratings explanation page, the evaluation of the "earn" equation is incorrect. It is given as:

= 1/3 * 0.5 * (5 * 0.89 + (5 - 10) / (1 + 2 * 0.89)), which evaluates as
=.333*.5 * (4.45 + (-5)/ (1 + 1.78) )
= .166 * (4.45 + (-5)/ (2.78)
= .166 * (4.45 + (-1.8))
= .166 * 2.65
= .439

The example gives an answer of 0.63

[mtpattonar]

To answer my own question from an earlier post, I believe #39 on the explanation page is misstated. It says:

39. The winner is always rated higher than the loser. the minimum margin is: mean pre-bout ratings of both r_m = (r_a + r_b)/2;
r_new_winner = r_m + earn_f*v*cd/4;
r_new_loser= r_m - earn_f*v*cd/4

The final two lines should read:
r_new_winner = r_a+earn_f*v*cd/4;
r_new_loser= r_a - earn_f*v*cd/4

which is the formula used in the two examples, and also which makes a lot more sense.

[computerrank]

On the ratings explanation page, the evaluation of the "earn" equation is incorrect. It is given as:

= 1/3 * 0.5 * (5 * 0.89 + (5 - 10) / (1 + 2 * 0.89)), which evaluates as
=.333*.5 * (4.45 + (-5)/ (1 + 1.78) )
= .166 * (4.45 + (-5)/ (2.78)
= .166 * (4.45 + (-1.8))
= .166 * 2.65
= .439

The example gives an answer of 0.63

corrected (0.442 correctly)

[computerrank]

To answer my own question from an earlier post, I believe #39 on the explanation page is misstated. It says:

39. The winner is always rated higher than the loser. the minimum margin is: mean pre-bout ratings of both r_m = (r_a + r_b)/2;
r_new_winner = r_m + earn_f*v*cd/4;
r_new_loser= r_m - earn_f*v*cd/4

The final two lines should read:
r_new_winner = r_a+earn_f*v*cd/4;
r_new_loser= r_a - earn_f*v*cd/4

which is the formula used in the two examples, and also which makes a lot more sense.

r_m is correct

[mtpattonar]

r_m is correct? Then doesn't that mean that if the lower rated wins, he then ends up with the higher final rating, regardless of how far apart they were to begin with? And as I mentioned, both examples demonstrate that equation using the r_a/r_b variables, and not the r_m variable. ( I made an error in my original statement by showing r_a for both fighters, instead of r_a and r_b.) If I'm making a mistake here, please explain it to me.

39. The winner is always rated higher than the loser. the minimum margin is: mean pre-bout ratings of both r_m = (r_a + r_b)/2;
r_new_winner = r_m + earn_f*v*cd/4;
r_new_loser= r_m - earn_f*v*cd/4

The final two lines should read:
r_new_winner = r_a+earn_f*v*cd/4;
r_new_loser= r_a - earn_f*v*cd/4

which is the formula used in the two examples, and also which makes a lot more sense.
r_m is correct

[computerrank]

I'm confused on a few points on the ratings explanation page:

1. "The rating of a boxer is reduced by 50% for every time period of inactivity of 18 months - and to even less in proportion to a longer time period of inactivity"
I can't get the math to work out in the example. The example says the boxer has been out for 23 months. 50% reduction for 18 months = 2.78% per month, times 23 = 63.9% reduction, or a rate of 46.1%
Or if the formula is 50% reduction over 18 months, then an additional 50% of the remainder (or 25% of the original) over 18 months ( 1.39 percent monthly), then 50% + 5*1.39% = 56.95% reduction, or a current rate of 43.05%. Yet example 2 gives a final resolution of 41%, which doesn't jibe with either of my calculations.

=> the example is for 23 months and the reduction is exponential. So a half = factor 0.5 after 18 months gives after a 23 months factor = 0.5 ** (23/18) = 0.5 ** 1.28 = 0.41
So you found an error anyway, I typed 22 months in the 2nd reference

2. "The pre-bout rating of a successfully debuting boxer is set to 25% of his opponents pre-bout rating."
Not sure how to interpret this. My best guess: a boxer who wins his first fight has his rating set at 25% of the rating of his opponent, but the opponent receives no penalty. Then, all adjustments are calculated and applied as with any other bout. Would that be correct?

=> The pre-bout rating is set to 25%. Then all other calculations are done as usual.

3. "The winner is always rated higher than the loser. "
Surely this is a misstatement? Otherwise, it says that if a low rated boxer manages to upset a high rated boxer, the low rated will then have the higher rating, regardless of how much distance was previously between them or how much experience they have. Thus, if the worst boxer on the planet pulls the enormous upset of defeating the best boxer on the planet, then suddenly the worst guy is now in the top 50%, and the champion is suddenly in the bottom 50%.
That would seem to be an extremely harsh penalty for only one bad fight, and not realistic. Or have I misinterpreted.

=> The winner is placed above the mean of both pre-bout ratings. So the loser in an upset will stay quite high, due to the exponential character of the rating. And the winner will be set just above him. The winner gets it all.

4. I tried to post this as a new topic, but I don't see where to do that. Is this session of the forum not allowed to do that?

=> just place your questions and remarks here

Thank you for your time.

[computerrank]

r_m is correct? Then doesn't that mean that if the lower rated wins, he then ends up with the higher final rating, regardless of how far apart they were to begin with? And as I mentioned, both examples demonstrate that equation using the r_a/r_b variables, and not the r_m variable. ( I made an error in my original statement by showing r_a for both fighters, instead of r_a and r_b.) If I'm making a mistake here, please explain it to me.

39. The winner is always rated higher than the loser. the minimum margin is: mean pre-bout ratings of both r_m = (r_a + r_b)/2;
r_new_winner = r_m + earn_f*v*cd/4;
r_new_loser= r_m - earn_f*v*cd/4

The final two lines should read:
r_new_winner = r_a+earn_f*v*cd/4;
r_new_loser= r_a - earn_f*v*cd/4

which is the formula used in the two examples, and also which makes a lot more sense.
r_m is correct

This shows the MINIMUM margin - the winner's post-bout rating may be higher, but not lower.

[mtpattonar]

I'm still not grasping something. You say the winner will be higher rated than the loser, regardless of the disparity in their pre-fight ratings. But in example 1, if the lower rated fighter had won, then the earn equation works out to 1.78, giving A a rating of 10-1.78 = 8.22, and B a rating of 5+1.78 = 6.78. Even with max. bonus points, B would come out to 7.89, still below A.

In example 2, the lower rated fighter does win, but does not even come close to having the higher adjusted rating.
So what is it that I am misinterpreting?

Thanks again,
MP

[computerrank]

I'm still not grasping something. You say the winner will be higher rated than the loser, regardless of the disparity in their pre-fight ratings. But in example 1, if the lower rated fighter had won, then the earn equation works out to 1.78, giving A a rating of 10-1.78 = 8.22, and B a rating of 5+1.78 = 6.78. Even with max. bonus points, B would come out to 7.89, still below A.

In example 2, the lower rated fighter does win, but does not even come close to having the higher adjusted rating.
So what is it that I am misinterpreting?

Thanks again,
MP

If the winner is lower than the mean pre-bout ratings plus minimum margin. The winner is set to the upper minimum margin value - and the loser is set to lower.

[computerrank]

raw prediction rate now 89.8 percent

http://151.boxrec.com/~martin/ratings.php
http://151.boxrec.com/~martin/ratings_at.php

[computerrank]

raw prediction rate now 89.8 percent

http://151.boxrec.com/~martin/ratings.php
http://151.boxrec.com/~martin/ratings_at.php

So the whole history raw figures are now nearly 4% better than the adjusted present Boxrec rankings?! That prediction rate is very high indeed.

So the assumption is that 89.8 can go even higher when they too are adjusted?

What changes are you making to get the raw figures better and better?

If you are making changes when are the raw figures considered the modified figures?

The current figures again look fine to me.

The all time figures still look not as good as the existing all time figures...just look at the all time heavyweights....Povetkin?!

With the prediction rate now so high for the current ratings it must be getting to the stage where you are seriously thinking of using them?

I will not propose the WHR ratings to go live before next year. There is no need for hurry, the evaluation is not finished yet.

In order to evaluate the equivalence of a top career nowadays and at former times regarding the WHR all time ratings, I tested the following scenario:

- For W. Klitschko I determined the division rank of his opponents at bout time.
- I created a dummy boxer W.Klitschko, with the same sequence of bouts, but starting in Joe Louis' debut year. A shift of 62 years, from 1996 to 1934.
- For these dummy bouts I substituted his real opponents by opponents, equivalently ranked at the dummy bout time.
- Then I run WHR ratings.

The result:
- W. Klitschko dummy peak rating and all-time rating is slightly higher than in his real career (about 10 percent)


I did the same with a shift to Sam Langford's debut year. A shift of 94 year, from 1996 to 1902.

The result:
- W. Klitschko dummy peak rating and all-time rating is about 25 percent of the rating in his real career. But he is still #26 all time Heavyweights, Boxrec recorded competition was lower at that times, I guess.
- While Sam Langford is #96 with 1/3 of dummy Klitschko's points.

Over all, the WHR all time rankings are quite founded based on the achievements against equivalent ranked opponents. Boxrec cannot regard styles, skills, social restrictions and other imponderabilities.

[SportsRatings]

raw prediction rate now 89.8 percent

http://151.boxrec.com/~martin/ratings.php
http://151.boxrec.com/~martin/ratings_at.php

How is the 'raw' rate increasing? If you are tinkering with it to make the prediction rate improve, isn't that the 'post-raw' rate?

[computerrank]

raw prediction rate now 89.8 percent

http://151.boxrec.com/~martin/ratings.php
http://151.boxrec.com/~martin/ratings_at.php

How is the 'raw' rate increasing? If you are tinkering with it to make the prediction rate improve, isn't that the 'post-raw' rate?

My mistake. I looked at the post-bout explanation rate. The correct pre-bout prediction is up to 80.6 percent.

[Manrae]

The all time ratings need so much work that maybe it needs a complete makeover.

Like, maybe a completely different algorithm or something, it's that bad.

I've spoken my peace about the obvious problems with the regular ratings. While I can see some things as a far stretch of the imagination, it doesn't reflect objective reality of the work (or lack thereof) fighters are putting in. This is all my opinion, of course, but no way could I argue the validity of certain rankings. It's not possible.

[computerrank]

The all time ratings need so much work that maybe it needs a complete makeover.

Like, maybe a completely different algorithm or something, it's that bad.

I've spoken my peace about the obvious problems with the regular ratings. While I can see some things as a far stretch of the imagination, it doesn't reflect objective reality of the work (or lack thereof) fighters are putting in. This is all my opinion, of course, but no way could I argue the validity of certain rankings. It's not possible.

The life Boxrec all time ratings and the tested Whole-History all time ratings use completely different models.

The WHR ratings simply go with the career peak rating. This may be too simple.

The life Boxrec all time ratings go with annual points for defeating top ranked opponents. And the annual points are added to result in the final all time rating of a boxer.

I could apply a similar model to the WHR all time ratings. They then would more look like the life Boxrec ones.

[SportsRatings]

The all time ratings need so much work that maybe it needs a complete makeover.

Like, maybe a completely different algorithm or something, it's that bad.

I've spoken my peace about the obvious problems with the regular ratings. While I can see some things as a far stretch of the imagination, it doesn't reflect objective reality of the work (or lack thereof) fighters are putting in. This is all my opinion, of course, but no way could I argue the validity of certain rankings. It's not possible.

The life Boxrec all time ratings and the tested Whole-History all time ratings use completely different models.

The WHR ratings simply go with the career peak rating. This may be too simple.

The life Boxrec all time ratings go with annual points for defeating top ranked opponents. And the annual points are added to result in the final all time rating of a boxer.

I could apply a similar model to the WHR all time ratings. They then would more look like the life Boxrec ones.

Every time I tried to use peak rating for an all-time ranking the Spinks brothers made me reconsider immediately :) As did Buster Douglas.

I think it would be interesting to see what the current Boxrec all-time formula would output using the WHR ratings. It would make for a more direct comparison between the all-time ratings -- the same all-time formula using different raw ratings.

[BroughtonRulesRefuge]

- Now that the hysteria over Canelo/Kov has subsided and the farce of KSI/Paul has quieted, time for another look into boxrec monkeyshines.


This the boxrec numbers summary of Canelo/Kov-

https://boxrec.com/en/event/794894/2391710

By memory Canelo #1 p4p for donkey years over Crawford by a substantial margin, but you had him by Pts going into Kov barely above #5 Spence, and then barely above Bud after the KO, ie losing his margin.

Moreover you have Kov as #3 going in when if fact he was boxrec #1 LH unless boxrec changed him to #3 hours before the fight.

Just pointing out something even Jose 'Can You See' Feliciano could see.

[computerrank]

- Now that the hysteria over Canelo/Kov has subsided and the farce of KSI/Paul has quieted, time for another look into boxrec monkeyshines.


This the boxrec numbers summary of Canelo/Kov-

https://boxrec.com/en/event/794894/2391710

By memory Canelo #1 p4p for donkey years over Crawford by a substantial margin, but you had him by Pts going into Kov barely above #5 Spence, and then barely above Bud after the KO, ie losing his margin.

Moreover you have Kov as #3 going in when if fact he was boxrec #1 LH unless boxrec changed him to #3 hours before the fight.

Just pointing out something even Jose 'Can You See' Feliciano could see.

The rating points are in relation to the weight division. The points are converted and are lower in higher weight divisions and higher in lower divisions.

[computerrank]

Every time I tried to use peak rating for an all-time ranking the Spinks brothers made me reconsider immediately :) As did Buster Douglas.

I think it would be interesting to see what the current Boxrec all-time formula would output using the WHR ratings. It would make for a more direct comparison between the all-time ratings -- the same all-time formula using different raw ratings.

The Whole-History raw ratings winner prediction ratio improved to 83.2 percent now.

Instead of using the career peak rating for the WHR all time ratings I implemented a solution analog to the current rating version. Annual points rewards of at most 100 points/year for defeating #1 opponent in the annual division ranking and another at most 100 points/year for defeating #1 opponent in the annual p4p ranking, reciprocal to the rank.

http://151.boxrec.com/~martin/ratings.php
http://151.boxrec.com/~martin/ratings_at.php

[computerrank]

The Whole-History raw ratings winner prediction ratio improved to 83.2 percent now.

Instead of using the career peak rating for the WHR all time ratings I implemented a solution analog to the current rating version. Annual points rewards of at most 100 points/year for defeating #1 opponent in the annual division ranking and another at most 100 points/year for defeating #1 opponent in the annual p4p ranking, reciprocal to the rank.

http://151.boxrec.com/~martin/ratings.php
http://151.boxrec.com/~martin/ratings_at.php
ar

Current look fairly reasonable though surprised to see Povetkin higher than Joshua.

All time is still a car crash. Also using random figures like 100 doesn't mean anything. Why not 73 or 111?The figure should come from a mathematical basis (like best predictor) In theory there should be a way of combining the two systems that improves things but I don't think this is it quite yet. Does adding the whole history peak in each year Together not work or the existing systems total multiplied by the whole history career peak of the fighter?

For the all time ratings you can choose any base points value for rewarding a win against #1. Doubling the points means doubling the all time ratings points for all boxers. It is only a base value.
There is no predictor for the all time ratings. So all is to use a plausible points reward formula. For example base value for #1, half base value for #2, 1/3 base value for #3 etc.

And why a car crash?