Ratings - please read before commenting - Archived

[jujigatame]

Tank over Teo is weird because no respected ranking ranks him that way. Not The Ring, BLH, ESPN, nobody. The fact that the BoxRec system rates him almost 50% higher than Teo just makes it extra strange. Tank's resume at 135 is non-existent and his signature win at 130 came against a guy who was 2 weight classes and several years past his prime. Meanwhile Teo beat a top P4P fighter at 135 plus 2 other very good recent wins in Commey and Nakatani.

Part of the problem here is how we are defining the "best" system. You are putting a premium on predictivity but I would say the average fan doesn't care that much about predictivity. They want reactivity. They want understandability. I understand WHR can get you the highest predictivity rates but unless you're creating a betting/handicapping system I don't think that's of great importance. You've implicitly recognized this by heavily modifying the system in an attempt to mold it into something that more closely resembles traditional rankings. But if what we want is something that looks like traditional rankings, why not stick with the traditional system that everyone understood?

The rankings you cite consider things BoxRec will not. Minimum differences between calculated predictions based on the ratings and the official results are the only objective way to get measures for better and worse ratings. Everything else is a matter of believe it or not.

There are no heavily modified WHR ratings at BoxRec. The BoxRec ratings are strictly based on the pure WHR ratings.

The only BoxRec additional rules are:

- winner above loser for 18 months
- allow stay busy bouts and take the best WHR rating within at most 18 months of clear wins
- convert ratings between weight divisions with the same conversion tool as in all previous BoxRec rating versions

I mean you said the WHR rankings are not heavily modified, and then you listed a bunch of heavy modifications you've already had to make. Obviously you've already had to make these modifications because people do NOT want a pure WHR ranking because predictivity is NOT of primary importance to most boxing fans.

[jujigatame]

teo should obviously be ahead of tank, even if you dont knock tank for his wins mostly being at other weights, it's not like they are really all that good compared to teos. teo has by far the biggest win of either of them, but a huge margin, and of course it was at the actual weight they are ranked

Tank over Teo by such a margin is really not defensible. You could argue that Teo has 2 wins better than Tank's biggest win ever. A pure predictive system might have stuff like Tim Tszyu or Sebastian Fundora at #1 at 154 but is that really what anyone (other than gamblers) wants from a ranking system?

[JCS]

The Tank over Teo situation is a classic "issue" that's found when comparing human and computer rating systems.

How do you handle performance in multiple divisions, in a sport where fighters can move in and out whenever and each fighter reacts to weight changes different, etc?

You can only punish a fighter's rating so much when he moves up... because the math has to do the opposite, on the way down.. Or does it? Fighters don't fight often enough to build division-specific ratings <-- This is the REAL problem.

My feeling on this.. is a provisional (temporary) rating reduction based on the fighter's "best" performance in the currently rated division in the past X days. Even if it is artificial, it will please the masses :)

On the topic of predictability.. once upon a time, I was sucked down that rabbit hole. Knowing the sport we're rating for and handling all of the nuances should be top priority... with the math geek stuff coming second :) It's best to get the Top 10 (or so) right, then let predictability guide the rest.

I actually don't think Tank over Teo is an issue of weight classes. Maybe a little bit (in the way the win over LSC is valued) but more centrally I think it's an issue of how traditional systems can put heavy emphasis on quality of competition. The WHR system seems to be giving Tank a ton of credit for racking up dominant wins over marginal talents like Cuellar and Ruiz and Nunez. No traditional system will value any number of those wins more than a win over Lomachenko, and I would argue that is the more "correct" (or at least widely accepted) way to create a ranking system.

I totally agree that getting the top 10 or so correct is paramount and then you can let predictivity dictate the rest. This current system doesn't seem to be doing that. It seems to be more of a handicapping/betting system that has had a bunch of tweaks made to hammer it into a form that people can stomach as a traditional ranking system.

I think the weight class thing is a pretty big deal in this case, at least from an optics point of view... smaller dominant fighter increases division, but hasn't done a lot in the new division... at least not enough to justify his rating.

But yes, WHR seems to almost have compounding returns in some cases, acting almost like a multiplier on a slot machine.

[computerrank]

Intended ratings and also WHR ratings don't know anything about divisions. They just compare boxers from a neutral absolute result perspective.

Heavyweights will always have an advantage in general against Cruiserweights just by weight difference, and Cruiserweights against Light Heavyweights etc.

So Heavyweights will show a higher natural WHR rating in general compared with Cruiserweights etc.

There are some options to deal with that - in every case you will have to decide, what weight decision you would like to assign a boxer or bout to:

- you can establish separated weight division WHR ratings and decide what bouts you would consider for each division
- you can establish separated weight division WHR ratings and decide what bouts you would consider for each division and give out-of-weight-division bouts a smaller weight
- you can follow the natural plain comprehensive weight_class_spanning WHR ratings, just driven by the results and then decide, to what weight division assign a boxer and convert the plain WHR rating to the assigned weight division

I followed the last option.

[jujigatame]

The Tank over Teo situation is a classic "issue" that's found when comparing human and computer rating systems.

How do you handle performance in multiple divisions, in a sport where fighters can move in and out whenever and each fighter reacts to weight changes different, etc?

You can only punish a fighter's rating so much when he moves up... because the math has to do the opposite, on the way down.. Or does it? Fighters don't fight often enough to build division-specific ratings <-- This is the REAL problem.

My feeling on this.. is a provisional (temporary) rating reduction based on the fighter's "best" performance in the currently rated division in the past X days. Even if it is artificial, it will please the masses :)

On the topic of predictability.. once upon a time, I was sucked down that rabbit hole. Knowing the sport we're rating for and handling all of the nuances should be top priority... with the math geek stuff coming second :) It's best to get the Top 10 (or so) right, then let predictability guide the rest.

I actually don't think Tank over Teo is an issue of weight classes. Maybe a little bit (in the way the win over LSC is valued) but more centrally I think it's an issue of how traditional systems can put heavy emphasis on quality of competition. The WHR system seems to be giving Tank a ton of credit for racking up dominant wins over marginal talents like Cuellar and Ruiz and Nunez. No traditional system will value any number of those wins more than a win over Lomachenko, and I would argue that is the more "correct" (or at least widely accepted) way to create a ranking system.

I totally agree that getting the top 10 or so correct is paramount and then you can let predictivity dictate the rest. This current system doesn't seem to be doing that. It seems to be more of a handicapping/betting system that has had a bunch of tweaks made to hammer it into a form that people can stomach as a traditional ranking system.

I think the weight class thing is a pretty big deal in this case, at least from an optics point of view... smaller dominant fighter increases division, but hasn't done a lot in the new division... at least not enough to justify his rating.

But yes, WHR seems to almost have compounding returns in some cases, acting almost like a multiplier on a slot machine.

From an optics POV yes I agree the weight class thing matters, but that's not really my issue personally. I tend to think a point total at 130 should get a pretty similar weight at 135, perhaps proportional to the actual weight difference. A 5 pound weight jump doesn't make it a whole different sport or anything.

I'm more displeased by the way WHR seems to vastly overvalue consistent dominant results over marginal fighters. Tank smashing Fonseca, Cuellar, Ruiz and Nunez effectively gave us no new information about him, yet WHR shows his rating skyrocketing in that timespan. You could see Nakalitha was another recent case of this. He rose up to #14 at 130 by spending several years wiping out a laundry list of complete and utter scrubs.

These are the same sorts of problems BoxRec had back when it was still using a more traditional system but with predictivity as the only goal. You had weirdness like Shannon Briggs breaking into the top 10 while he went on his crazy bum-crushing tour in the mid-2000s.

[computerrank]

....
I'm more displeased by the way WHR seems to vastly overvalue consistent dominant results over marginal fighters. Tank smashing Fonseca, Cuellar, Ruiz and Nunez effectively gave us no new information about him, yet WHR shows his rating skyrocketing in that timespan. You could see Nakalitha was another recent case of this. He rose up to #14 at 130 by spending several years wiping out a laundry list of complete and utter scrubs.

These are the same sorts of problems BoxRec had back when it was still using a more traditional system but with predictivity as the only goal. You had weirdness like Shannon Briggs breaking into the top 10 while he went on his crazy bum-crushing tour in the mid-2000s.

I set to No Contest the for cited bouts of Davis - but his WHR rating goes to nearly the same level in spite of that.



It mainly was the win against Cruz, which makes the point.

[SportsRatings]

....
I'm more displeased by the way WHR seems to vastly overvalue consistent dominant results over marginal fighters. Tank smashing Fonseca, Cuellar, Ruiz and Nunez effectively gave us no new information about him, yet WHR shows his rating skyrocketing in that timespan. You could see Nakalitha was another recent case of this. He rose up to #14 at 130 by spending several years wiping out a laundry list of complete and utter scrubs.

These are the same sorts of problems BoxRec had back when it was still using a more traditional system but with predictivity as the only goal. You had weirdness like Shannon Briggs breaking into the top 10 while he went on his crazy bum-crushing tour in the mid-2000s.

I set to No Contest the for cited bouts of Davis - but his WHR rating goes to nearly the same level in spite of that.



It mainly was the win against Cruz, which makes the point.

I just want to say, thank you Computerrank for all your participation and feedback here. There aren't many (any?) people who make ranking systems and discuss them so thoroughly, to the point of making 'what if' changes and discussing detailed points about the system. I know you've gotten lots of crap for changes, especially with the WHR, but it's refreshing to see this kind of back and forth explanation down to the details.

Almost 200 pages of discussion of the Boxrec ratings over 16 years, pretty cool. Interesting and invaluable reading material for us rankings geeks, too.

[SportsRatings]

I took a look back at page 1 of the Ratings thread and noticed posts being quoted that apparently happened before page 1 existed...were some pages deleted in a database "reset" of sorts, pre-2005?

[jujigatame]

....
I'm more displeased by the way WHR seems to vastly overvalue consistent dominant results over marginal fighters. Tank smashing Fonseca, Cuellar, Ruiz and Nunez effectively gave us no new information about him, yet WHR shows his rating skyrocketing in that timespan. You could see Nakalitha was another recent case of this. He rose up to #14 at 130 by spending several years wiping out a laundry list of complete and utter scrubs.

These are the same sorts of problems BoxRec had back when it was still using a more traditional system but with predictivity as the only goal. You had weirdness like Shannon Briggs breaking into the top 10 while he went on his crazy bum-crushing tour in the mid-2000s.

I set to No Contest the for cited bouts of Davis - but his WHR rating goes to nearly the same level in spite of that.



It mainly was the win against Cruz, which makes the point.

Then I cannot grasp the logic at work here. He has 2 KOs of guys in the 50-90 point range (one of them several years ago) and he's sitting at 300 points. Meanwhile Teo has a comparable win (Commey) and then a far superior win (Lomachenko) and is at 200 points. Help me make sense of this because it feels like there is just not enough emphasis on quality of opposition here. Same thing with Nakalitha, he somehow rose up to 20+ points by beating bottom of the barrel scrubs. It may work for predictivity (because prospects start off by crushing bums and who they HAVE beat generally isn't indicative of who they COULD beat) but it does not work from any traditional rankings standpoint.

Also let me echo SportsRatings and just say that I do appreciate what you do here, and I apologize if I ever seem overly harsh or combative.

[JCS]

With prediction rate driving this I wonder.....

Is the typical 20-0 guy... more likely to go 21-0, because he has great matchmaking, home advantage, promoter advantage, etc..? And things like this are driving system "improvements".. which is causing the system to over-reward wins, period?

[jujigatame]

With prediction rate driving this I wonder.....

Is the typical 20-0 guy... more likely to go 21-0, because he has great matchmaking, home advantage, promoter advantage, etc..? And things like this are driving system "improvements".. which is causing the system to over-reward wins, period?

Yes, I have always thought something like this, going back to the days I mentioned before, when predictivity was the sole driver of the traditional system and Shannon Briggs was rated as a top HW.

Talented prospects tend to rack up big undefeated records, and their level of ability usually far outstrips their level of competition. This influences the entire system. If you're only valuing predictivity, you will ultimately have to rate guys with those types of records higher than their level of competition would naturally justify. And that will have knock-on effects all over the place.

[SportsRatings]

With prediction rate driving this I wonder.....

Is the typical 20-0 guy... more likely to go 21-0, because he has great matchmaking, home advantage, promoter advantage, etc..? And things like this are driving system "improvements".. which is causing the system to over-reward wins, period?

What could be a solution to this, maybe using prediction rate of fights that are judged to be sufficiently close?

Remove from the prediction rate equation fights that are considered "obvious" results (which would include some big upsets, which the system shouldn't be predicting anyway)

[computerrank]

With prediction rate driving this I wonder.....

Is the typical 20-0 guy... more likely to go 21-0, because he has great matchmaking, home advantage, promoter advantage, etc..? And things like this are driving system "improvements".. which is causing the system to over-reward wins, period?

What could be a solution to this, maybe using prediction rate of fights that are judged to be sufficiently close?

Remove from the prediction rate equation fights that are considered "obvious" results (which would include some big upsets, which the system shouldn't be predicting anyway)

You can change the weight of a bout with which it is considered for the ratings and for the predictions.

Going with your suggestion I now introduce a weight factor:

rating_a / (rating_a + rating_b) = ratio_a
rating_b / (rating_a + rating_b) = ratio_b

the weight factor = 4 * ratio_a * ratio_b

e.g see examples with different rating ratios :
- both boxers have same rating => weight factor = 4 * 0.5 * 0.5 = 4 * 0.25 = 1; that is the maximum possible value
- one boxer has 3 times the rating of the other => weight factor = 4 * 0,75 * 0..25 = 0.75; so it is a lower weight
- one boxer has 10 rimes the rating of the other => weight factor = 4 * 0.9 * 0.1 = 0.36; so the weight is much lower

I just calculated the rating with that modification for men Lightweight and Super Featherweight (see your examples)




... and you can see, not the changes you would have expected for Davis, Lopez and Nakathila ...

[JCS]

With prediction rate driving this I wonder.....

Is the typical 20-0 guy... more likely to go 21-0, because he has great matchmaking, home advantage, promoter advantage, etc..? And things like this are driving system "improvements".. which is causing the system to over-reward wins, period?

What could be a solution to this, maybe using prediction rate of fights that are judged to be sufficiently close?

Remove from the prediction rate equation fights that are considered "obvious" results (which would include some big upsets, which the system shouldn't be predicting anyway)

A great idea... if we had something like betting odds to drive this. Otherwise, you're using some arbitrary method... or potentially even worse, the system itself to gauge what is a "big upset" and what isn't.

Recent updates do seem put emphasis on glossy records as evidenced by All-Time ratings. Types of win bonuses can get out of control with boxing matchmaking.

[margaret thatcher]

the biggest problem out of all this is that teo doesnt get to show off his 'im #1 ' shirt anymore

[jujigatame]

With prediction rate driving this I wonder.....

Is the typical 20-0 guy... more likely to go 21-0, because he has great matchmaking, home advantage, promoter advantage, etc..? And things like this are driving system "improvements".. which is causing the system to over-reward wins, period?

What could be a solution to this, maybe using prediction rate of fights that are judged to be sufficiently close?

Remove from the prediction rate equation fights that are considered "obvious" results (which would include some big upsets, which the system shouldn't be predicting anyway)

A great idea... if we had something like betting odds to drive this. Otherwise, you're using some arbitrary method... or potentially even worse, the system itself to gauge what is a "big upset" and what isn't.

Recent updates do seem put emphasis on glossy records as evidenced by All-Time ratings. Types of win bonuses can get out of control with boxing matchmaking.

Unless you're creating a betting system, I don't think predictivity should be used at all, except maybe for a little bit of tuning at the edges. For something that is supposed to emulate traditional boxing rankings I think it is completely wrongheaded.

[computerrank]

Regarding the idea, what would it mean to set more emphasis on bouts between close rated boxers in calculating the ratings?

You could see here: viewtopic.php?p=5574098#p5574098

The ratings look not very different regarding Davis and Lopez at Lightweight and at Nakathila at Super Featherweight,

When I evaluate the winner prediction ratio, I currently use all bouts from 2010 to 2019 with both boxers having at least 2 bouts before and both boxers connected to the boxing community by good enough results against other connected boxers (independent of the ratings). So there are about 100,000 bouts evaluated. Results on points with less than 12 rounds boxed have lower weight,

- The current winner prediction ratio is 84.36 percent.
- When I give more weight in the ratings calculation to bouts with close rated boxers, the winner prediction ratio is 84.22 percent, more or less the same.

But now I tested, to give more weight in the winner prediction ratio to bouts with close rated boxers. Prediction should be more difficult and result in a lower prediction ratio
- a winner prediction ratio of only 75.56 percent for the current even-weighted ratings calculation. Less, as expected.
- a winner prediction ratio of 77.53 percent for the ratings with more weight given in the ratings calculation to close rated bouts. Much better than with the even-weighted ratings calculation.

That means, if you more emphasize close rated bouts and more neglect bouts with an obvious favorite and so emphasize those bouts in the winner prediction ratio and in the ratings calculation, you get a better prediction ratio by 2 percent, which is quite a lot.

[Manrae]

Regarding the idea, what would it mean to set more emphasis on bouts between close rated boxers in calculating the ratings?

You could see here: viewtopic.php?p=5574098#p5574098

The ratings look not very different regarding Davis and Lopez at Lightweight and at Nakathila at Super Featherweight,

When I evaluate the winner prediction ratio, I currently use all bouts from 2010 to 2019 with both boxers having at least 2 bouts before and both boxers connected to the boxing community by good enough results against other connected boxers (independent of the ratings). So there are about 100,000 bouts evaluated. Results on points with less than 12 rounds boxed have lower weight,

- The current winner prediction ratio is 84.36 percent.
- When I give more weight in the ratings calculation to bouts with close rated boxers, the winner prediction ratio is 84.22 percent, more or less the same.

But now I tested, to give more weight in the winner prediction ratio to bouts with close rated boxers. Prediction should be more difficult and result in a lower prediction ratio
- a winner prediction ratio of only 75.56 percent for the current even-weighted ratings calculation. Less, as expected.
- a winner prediction ratio of 77.53 percent for the ratings with more weight given in the ratings calculation to close rated bouts. Much better than with the even-weighted ratings calculation.

That means, if you more emphasize close rated bouts and more neglect bouts with an obvious favorite and so emphasize those bouts in the winner prediction ratio and in the ratings calculation, you get a better prediction ratio by 2 percent, which is quite a lot.

May I ask, what do the Heavyweight and Welterweight divisions look like under this proposal?

[computerrank]

May I ask, what do the Heavyweight and Welterweight divisions look like under this proposal?

[Manrae]

May I ask, what do the Heavyweight and Welterweight divisions look like under this proposal?

Very interesting, thanks

[jujigatame]

It is an interesting idea but I still think it moves in the wrong direction.

By putting importance on predictivity you're essentially saying you want a system that would have rated Tyson Fury #1 BEFORE he fought Wlad. Does anyone actually want that?

[Manrae]

I think that I found a bug

https://boxrec.com/en/ratings?r%5Brole% ... 5D=a&r_go=

#8 Kareem Martin ranked way too high... unless I'm missing something

[computerrank]

I think that I found a bug

https://boxrec.com/en/ratings?r%5Brole% ... 5D=a&r_go=

#8 Kareem Martin ranked way too high... unless I'm missing something

Not a bug:
Martin defeated Ananyan and Ananyan defeated Matias within the last 18 months. So the BoxRec winner above loser rule is applied.

The pure WHR ratings would be:
- Matias 66.66
- Ananyan 11.07
- Martin 2.915

Winner above loser rule makes:
- Martin 66.66
- Ananyan 55.66
- Matias 54.97

[computerrank]

It is an interesting idea but I still think it moves in the wrong direction.

By putting importance on predictivity you're essentially saying you want a system that would have rated Tyson Fury #1 BEFORE he fought Wlad. Does anyone actually want that?

You would like to have 2 ratings for every boxer at time of every of his bouts. One rating at the bout day before the bout and one rating at the same day after the bout. And the rating before a bout would normally be the rating after his last bout (if not any other rules would have reduced the rating in between:)

Whole-History-Ratings have a different perspective. Every boxer has one rating at time of every of his bouts.

But this rating changes with each bout added or changed in the system.

WHR has the perspective to know and consider all results in the system and then decide what rating (what strength) at what time a boxer would have had. And that decision changes with each bout added or changed. It is quite similar as a human also does. Every information he gets changes his deliberations and opinions. It also changes opinions about the past.

Regarding the Fury vs Klitschko bout. I am quite sure (I could simulate that with some effort), knowing only all results before the Fury vs Klitschko bout, WHR had rated Klitschko higher than Fury, But with knowing the result, WHR had Fury rated higher than Klitschko (I guess, and could simulate that too with some effort). And from today's perspective, knowing all results until today, we know WHR rates Fury higher than Klitschko at time of that bout.

So, explicitly, we could gather 3 ratings for every boxer at time of every of his bouts:
#1 his rating considering all results in the system before that bout
#2 his rating considering all results in the system including the result of that bout
#3 his rating considering all results in the system today

You would like to see #1 and #2. Keep in mind, #2 from the bout before would not be the same as #1 for the next bout.

But to gather that, BoxRec would have to calculate those values daily from scratch bout per bout and store those values in the database. This is a giant daily effort BoxRec cannot provide.

BoxRec prefers the rating #3. It is the perspective from today (with all knowledge collected about results and careers), looking at the rating (strength) of all boxers at time of all their bouts.

[jujigatame]

What I'm saying is that using WHR is a bad idea, full stop. If you want to create a boxing handicapping system, it's great. If you want to create something that looks like traditional (The Ring, BLH, ESPN) boxing rankings, it is not. Hence, why you've already had to make modifications to it to placate BoxRec users, because maximizing predictivity creates a set of rankings that people will simply not use (except maybe for gambling) or accept.