Previous Rating System

[oiuyt]

As a C I occasionally beat an A in the pools. I'd like to see the same stat for just DEs.

Yeah, actually I think I'd like to expand my request to see two matrices, one each for pools and DEs.

-B

[fdad]

Yeah, actually I think I'd like to expand my request to see two matrices, one each for pools and DEs.

-B

How about a breakdown by weapon?

[IanSerotkin]

Thanks Peet!!!!

And in conclusion:

Neener neener neener!

[oiuyt]

How about a breakdown by weapon?

Even better.

Okay, I'd like to see twelve matices.

Now it feels like we're getting closer to a Spanish Inquisition reference than Ancient Greek (or late 19th Century British) drama.

-B

[leftweet]

Sorry I was giving a hypothetical question. Let us reword it. Suppose a female fencer, who has won a World Championship and/or Olympices does fence in a mixed event, but she has never fenced in a mixed event before, at what level would she be seeded?

That is what the last paragraph talks about. Even though she hasn't earned an A10 in a mixed event, she is still seeded among the A10's. Zagunis would be randomly seeded between 1st and down to how many A10s there are (if there are 3, she could be seeded as low as 3rd, 15 as low as 15th).

1) It depends. Conceivably the costs could outweigh any benefits.

2) Not necessarily even then. This whole idea of "predictability" rather dismays me, frankly. I mean, I'm sure we could all save a lot on travel expenses if we could just design a system with perfect predictability. Then the national office could just email us on how we would have placed at a competition, and they wouldn't actually have to hold the tournament...

1) The ELO-type systems proposed would not cost much more (if any) than the current system.
2) That's why we compete. There is no such thing as a perfect predictive system -- but there are ones that are more predictive.

I don't even understand how that's possible.

Let's say she always beats Smith and Jones in pools, but never encounters either in a DE. And Smith and Jones advance much farther in the DEs. Their final places are much higher than hers. How does she get a "higher rating" than them?

Then obviously, she shouldn't be ranked higher than them. In this example, she should advance further in DEs (she should have a higher seed, etc), but she must be losing to someone along the lines, while Smith and Jones are able to win their DEs (with worse seedings). Just because you consistently beat people in pools, doesn't mean you should get a higher rating. That's not even how are current system works. She could fence in just mixed events, beat Smith and Jones 100 straight times lose in her first DE and never earn a rating. I think that that is fair.

Are you sure? If only one comprehensive point system is used, and she consistently wins her ( women's ) competitions while Daryl only places 3rd or 4th in his ( men's ) events; or if she wins a dozen tournaments and he wins 3; then they both enter a mixed competition...where are they likely to be seeded relative to each other?

What am I missing?

Homer would probably be seeded higher, which I think is accurate. Seeding should reflect the true strength of the fencer, not their "best" results (although one could argue placing 12th in the men's world championship is really actually better than winning the women's world championship). Currently, Zagunis and Homer would be randomly seeded against each other. I don't think their nationals points from their respective genders should be used. It's not accurate for mixed events, although it should be accurate for the specific genders. If both are the only A10s, each would have a fifty-percent chance of being seeded either 1st or 2nd.

What?

Whoever wins the tournament was the stronger fencer that day. Whoever places second was the second strongest. Etc. How can that possibly be "erroneous"?

Not necessarily. I have never liked the "seeding doesn't matter for the winner" -- it very much does. Peter does a nice job with one situation, but let's take it a little farther.

Seeding of the fencers: (true strength in alphabetical order; A's chances of beating opponent in parenthesis)
A
B (55%)
E (70%)
D (65%)
C (60%)
G (85%)
H (95%)
F (75%)

Now with seeding: in the quarterfinals, fencer A is drawn against F. Assuming no upsets, A would then fence C in the semifinals and B in the finals. Averaging A's chances of beating those opponents (75+60+55 / 3) comes to 63.3%. Using proper seeding A should fence H, D and B. A's average chance of winning: 71.7%. A's chance of winning goes up roughly 8 percentage points with correct seeding, so A has a higher chance of winning. This is how seeding should work. It would be even worse if B was seed 4th and met A in the semifinals, then the final definitely wouldn't have the top two fencers.

Why should it be any different?

And if they really are "the best", they won't be knocked out early. If not, why design a system to protect them? Can't that be characterized as giving them an unfair advantage?

The point of a seeding system is to accurately seed fencers so that the best two have the highest chance of meeting in the finals. Maybe that's unfair, but that's the way it works. Even the current USFA rating system tries to do that.

And then there are the foreign fencers.

At the Bladerunner this past weekend, we had a MS fencer from Mexico. He was listed and seeded as a U. Having had him in my pool, I can guarantee that he was considerably better than his rating, or his seeding. Especially considering that he got silver in the event.

Having a numerical system would not change that. Being from Mexico, he might have zero points, or only a few. So "one of the best fencers" is not "accurately seeded". The new system has accomplished---what?

He would likely be ranked some place in the middle, or slightly below. No seeding system would be able to accurately seed "new" fencers to the system -- it just can't be done. Unless, the USFA took results from all over the world, which would be pretty much impossible...

[Jason]

Thanks Peet!!!!

And in conclusion:

Neener neener neener!

Well done.

[tchwojko]

Now it feels like we're getting closer to a Spanish Inquisition reference than Ancient Greek (or late 19th Century British) drama.

Except that a lot of us saw this coming...

[mfp]

Now with seeding: in the quarterfinals, fencer A is drawn against F. Assuming no upsets, A would then fence C in the semifinals and B in the finals. Averaging A's chances of beating those opponents (75+60+55 / 3) comes to 63.3%. Using proper seeding A should fence H, D and B. A's average chance of winning: 71.7%. A's chance of winning goes up roughly 8 percentage points with correct seeding, so A has a higher chance of winning. This is how seeding should work. It would be even worse if B was seed 4th and met A in the semifinals, then the final definitely wouldn't have the top two fencers.

While your idea of trying it all out by picking some sample values is good, the way the math is done above has some significant problems ...

I'd recommend looking at some of the research papers that go through similar exercises. One of the papers Jason mentioned ("What is the Correct Way to Seed a Knockout Tournament"), is related, reasonably readable, and freely available online:

http://www.davidmarcus.com/Articles/Seed.pdf

The point of a seeding system is to accurately seed fencers so that the best two have the highest chance of meeting in the finals.

That certainly is one goal. The article by Schwenk above mentions some others as well.

Maybe that's unfair, but that's the way it works.

There are numerous papers on that aspect as well

[leftweet]

While your idea of trying it all out by picking some sample values is good, the way the math is done above has some significant problems ...

I'd recommend looking at some of the research papers that go through similar exercises. One of the papers Jason mentioned ("What is the Correct Way to Seed a Knockout Tournament"), is related, reasonably readable, and freely available online:

http://www.davidmarcus.com/Articles/Seed.pdf

haha, very good point. I did read each of the articles Jason posted, but I just wanted to keep the math simple for Inquartata (who constantly complains about math)

[michaelheggen]

Glicko is an Elo system, it's hugely different from Mr Masin's proposed system.

Technically, the Elo system is a special case of the Glicko system, according to Dr. Glickman.

That said, K O'N is otherwise right and I stand corrected after a more careful reading of Mr. Masin's article in AF. Mr. Masin's proposal is quite different from what we use because it is based on the fencers' final placings, while our Glicko variant is based upon the results of each and every bout. As such, our system:

• does not need to take into account the size and strength of the event;
• doesn't have to deal with ties;
• does not depend on accurate seeding to produce accurate ratings, as ratings are based solely upon the people one actually fences against rather than upon everyone in the competition (including people you never fenced);
• will normalize across fencing populations more easily—e.g. a foilist from Hawaii who competes in just one competition on the mainland will add at least five or six data points (one for each bout vs. one per competition) to his or her dataset and bring those back to Hawaii (and that data will help normalize all Hawaiian foilists, male and female, as the fencer subsequently competes in Hawaii);
• would yield accurate ratings (during the conversion period) more quickly than Mr. Masin's system (because it is not dependent on accurate seeding);
• does not penalize a high-level fencer disproportionately if he or she experiences an upset in DEs, so there is no disincentive for such a fencer to compete in smaller tournaments (thus eliminating the need for Mr. Malin's hors concours rule);
• can be run retroactively if needed, while still remaining accurate.

Do you only use it for pool bouts, or do you use it for DEs as well?

We use our Glicko variant for both pools and DEs.

[Inquartata]

That's actually quite interesting... that the most stable pairing is the E/U match ups. I would've figured those are the most unstable...

Yes...at that level of skill/experience one would expect luck to play a large role.

Speaking of luck, I wonder whether the numbers change much if sorted for weapon? Would it bear out the assumption that epee is the most 'fluky'?

I'll refrain from tortured references to the croaking chorus from The Frogs of Aristophanes, unless that appears required to prod you into action.

-B

This whole line of banter is for The Birds.

This sort of reliance on initial seeding is why we fence pools.

Yes, thank goodness.

1) The ELO-type systems proposed would not cost much more (if any) than the current system.

I didn't mean just monetary costs, necessarily.

Then obviously, she shouldn't be ranked higher than them. In this example, she should advance further in DEs (she should have a higher seed, etc), but she must be losing to someone along the lines, while Smith and Jones are able to win their DEs (with worse seedings). Just because you consistently beat people in pools, doesn't mean you should get a higher rating.

And now along comes a tournament in which she DOES meet one ( or both ) of them in a DE, and whips him ( them ) as usual. What has just become of the "predictive" ability of the system, which had them ranked much higher than her?

However, under the current system, if they are both B's...or A09's...or A08's...she will be seeded before them in every case. And when she beats them, THAT seeding will have been pretty accurately predictive. No?

Homer would probably be seeded higher, which I think is accurate.

How? By placing higher in her comps, and still more by placing higher in more of them, she is going to be adding larger numbers of points faster, is she not? At some point, she will have more raw points and thus a higher rating---but not a very "accurate" one in terms of predicting the outcome of a bout between them. No?

Seeding should reflect the true strength of the fencer, not their "best" results (although one could argue placing 12th in the men's world championship is really actually better than winning the women's world championship).

Perhaps, but unless there's going to be some sort of additional calculation or multiplier so that men's and women's points can in fact be treated differently, I still can't see how the proposed system will manage to reflect "true strength" between disparate categories of competitors...

Currently, Zagunis and Homer would be randomly seeded against each other. I don't think their nationals points from their respective genders should be used.

That would result in the same sort of odd seeding that I think the proposed system would produce as well; the prospect of mixing everyone together in one all-encompassing system fills me with misgivings in this respect...

I have never liked the "seeding doesn't matter for the winner" -- it very much does.

Only conjecturally, AFAICT. That is, if A wins his every bout in a tournament, seeding is only significant to him if one speculates that "Well, but if he'd faced E in his second DE instead of Q, he would ( probably ) have lost; so the seeding which put him up against Q 'mattered'." We cannot actually know that E would have done better, worse or the same against him than Q, though.

And now that I think of it---no one has yet brought up the referee factor. I think we all know of matchings in the ROW weapons which might have completely opposite outcomes if presided over by Referee X than by Referee Y. I would think that this factor would introduce significant uncertainty into a system...and the more finely-tuned the system, the greater the uncertainty.

Seeding of the fencers: (true strength in alphabetical order; A's chances of beating opponent in parenthesis)
A
B (55%)
E (70%)
D (65%)
C (60%)
G (85%)
H (95%)
F (75%)

OK, maybe this is inherent in the differences in the numerical point totals of the various fencers vis-a-vis each other, but---the proposed system assigns simple numbers to fencers, not odds, correct? So why are you using percentages there?

he point of a seeding system is to accurately seed fencers so that the best two have the highest chance of meeting in the finals. Maybe that's unfair, but that's the way it works. Even the current USFA rating system tries to do that.

Does it? How so?

He would likely be ranked some place in the middle, or slightly below.

Why? He's a complete unknown. What is the justification for deciding that he is probably mediocre, rather than good or bad?

No seeding system would be able to accurately seed "new" fencers to the system -- it just can't be done. Unless, the USFA took results from all over the world, which would be pretty much impossible...

Even if possible, I'm not sure purely intranational results would be at all comparable. Or for that matter that they are even maintained at all. Beyond FIE rankings, that is.

Which leads to the question: Are there any other countries which seed tournaments similarly to the way the proposed system envisions doing?

I just wanted to keep the math simple for Inquartata (who constantly complains about math)

"I thank you, and, in your intention
Will tell my rosary to its last bead."

[oiuyt]

However, under the current system, if they are both B's...or A09's...or A08's...she will be seeded before them in every case.

Mmmmm. Mariel's an A08. If you're not using points (FIE or domestic) for seeding then she should be behind fencers with an A09.

-B

[Inquartata]

Picky, picky, picky.

[PeterGustafsson]

Hi!

Just out of curiosity, why use teams instead of individuals in that example?

For mathematical and notational simplicity, I wanted something without poules and with a relatively small number of contestants. Team events are more likely to fit that bill. Sure one could have posited that A-H are individuals and the results would have been exactly the same, but a t the cost of slightly less realism.

In any event, I see that the final appears to be between the two predicted strongest teams. So the system has been "accurate" in predicting who would place where, has it not?

That was because the seeding used was fairly similar to the true strengths. However, it is quite common that the #2 and #3 rank are essentially decided by chance, if those contestants are both considerably better than those in lower ranks, and the #1 rank is very good. More or less the entire 1990´s in men´s superheavyweight wrestling is an example of that - Alexander Karelin won literally everything without giving up points for 6 years running (Imagine Podziakov winning all his DE bouts by 15-0, years on end. he was that dominant.) while 2-3 other wrestlers were hoping that they would get him in the final, not the semi. The lot essentially decided who got silver and bronze, since those two could be counted upon to win against all lesser opponents.

It is still possible to create a bad initial ranking which gives a final without the two best participating, even if all bouts go according to true quality.

As before, 8 teams with true strength in descending order from A to H. However, the BC has severely faulty information, and ranks them thusly:

A
D
G
H
C
E
F
B

Quarterfinals:
A-B
C-E
D-F
G-H

Semifinals:
A-C
D-G

Final:
A-D

Final results after competition:

1. A
2. D
3. G
4. C
5. H
6. E
7. F
8. B

As you see, no team except A got a final placement in accordance to true strength.

Now, that was a contrived example designed to show the flaws of straight-to-DE as much as possible. You may say that it is very unlikely that the BC would get the initial ranking that bad, and you would be right. So let us consider a more reasonable case.

Imagnine the same 8-team event, in which teams A-G are local, and the BC gets their internal orderings right. Filling out the field is a faraway team from an area considered to a fencing hotbed, representing a club with a reputation. The local BC ranks them 1st, but little do they know that the fencers on that team actually are 3rd-stringers in that club. The ranking is like this:

H
A
B
C
D
E
F
G

Quarterfinals:
H-G
D-E
A-F
B-C

Semifinals:
G-D
A-B

Final:
A-D

Final results after competition:
1. A
2. D
3. B
4. G
5. H
6. C
7. E
8. F

Team A gets the gold, but that is also the only team that gets the right final placement. Team C ends up after both D and G, despite that it was correctly ranked above them. Note that in general, a faulty ranking of one team will severely mess up the final results of several teams which had correct ingoing rankings - even if one takes away badly ranked team H from this results list, the other teams have incorrect internal placements.


Have a nice time!

Peter Gustafsson

[Inquartata]

That was because the seeding used was fairly similar to the true strengths. However, it is quite common that the #2 and #3 rank are essentially decided by chance, if those contestants are both considerably better than those in lower ranks, and the #1 rank is very good.

"Chance"? Really? I can't remember ever seeing a seeding list where that happened. Have you an example? I mean from fencing, as opposed to naked luge or whiffleball or "something completely different"?

We ARE talking about the system currently used for fencing, after all. That other sports do things worse than we do doesn't strike me as all that relevant to whether our system works well enough that there is no pressing need to exchange it for another...

However, the BC has severely faulty information

Well, again you are creating an artificial premise to make your thesis look stronger. The current system used for seeding US fencing competitions does not suffer from providing "severely faulty information", does it? It may not be as "accurate" as some might wish, but it's adequate, IMO.

Now, that was a contrived example designed to show the flaws of straight-to-DE as much as possible. You may say that it is very unlikely that the BC would get the initial ranking that bad, and you would be right. So let us consider a more reasonable case.

OK, OK! I should read ahead before replying, I see.

However, I think that your second example is also an unrealistic one.

I HAVE seen people seeded inaccurately: Air Force Academy guys with lower ratings than their skills merited because they seldom fenced in USFA sanctioned tournaments, for example, and of course the occasional foreign fencers. But it's not a systemic problem, and the proposed alternative would not handle those situations any better than the present one AFAICT.