MLily main tournament

[trout]

9/1/15
Round of 8

Ahn Sungjoon defeated Xie Ke by resign.
Ke Jie defeated Rong Yi by resign.
Lee Sedol defeated Tang Weixing by resign.
Park Younghoon defeated Zhou Ruiyang by resign.

[Uberdude]

goratings.org got 3/4 this time. Park Yeonghun beating Zhou Ruiyang was the upset, but only a small one as Zhou only 19 rating points higher.

[Krama]

goratings.org got 3/4 this time. Park Yeonghun beating Zhou Ruiyang was the upset, but only a small one as Zhou only 19 rating points higher.

I haven't seen a lot of Park Younghoon lately, so him getting this far is quite surprising for me.

Also Park Junghwan losing so "early" was another surprise.

But after all this is only a game

[wineandgolover]

goratings.org got 3/4 this time. Park Yeonghun beating Zhou Ruiyang was the upset, but only a small one as Zhou only 19 rating points higher.

How big should an ELO gap be, before we really consider it an upset?

[Uberdude]

Yeah, by 'upset' I was just too lazy to type "game whose result was not as predicted"; is there a good single-word noun for this?. I wouldn't call going against a <53% prediction an upset. Something like against a 70-80% is what I'd normally mean by upset, 75% corresponds to an ELO difference of 190. It's worth noting that in the EGF system (based on ELO) the probability of win is not just a function of rating difference, but also of rating (so a 7d being 4 stones or 400 points above a 3d is more likely to win than a 1k being 4 stones or 400 points above a 5k). That arises because the rating system is distorted into not just having 100 points of rating meaning not just the 64% win chance or whatever the pure ELO model prescribes, but also meaning a one stone handicap difference. These pros are all so strong but still close to each other in terms of handicap stones difference and no handicap games in the database that 100 rating points means nothing in terms of handicap stone strength difference so probably the raw ELO difference to win rate is largely applicable.

[wineandgolover]

Yeah, by 'upset' I was just too lazy to type "game whose result was not as predicted"; is there a good single-word noun for this?

Maybe saying 3/4 theoretical favorites won, with Park Yeonghun being the only exception.

Yeah, even underdog might be too strong. Whatever, don't take my question or potential revision as even a slight criticism. And thanks for the EGF ELO explanation.

I am not going to try to prove it, but I'd expect that 3/4 favorites winning was far more likely than 4/4 favorites winning, because there are four different ways that could happen, versus only one for all faves to win.

[Uberdude]

Let's do some maths!

Taking the standard ELO win rates for rating difference from http://www.bobnewell.net/nucleus/bnewell.php?itemid=279 (Remi, is this applicable to WHR?)

Lee (3521) vs Tang (3429) 63%
Ke (3561) vs Rong (3231) 87%
Zhou (3444) vs Park (3425) 53%
Ahn (3348) vs Xie (3293) 58%

So chance of 4/4 is all of those multiplied together = 17%
Chance of 3/4 is complement of each one multiplied by the remaining others added together is = 10% + 3% + 15% + 12% = 30%
So if the prediction model is correct, the chance of it getting 3/4 is about twice that of getting 4/4.

[Rémi]

Remi, is this applicable to WHR?

Yes. The formula is P = 1.0 / (1.0 + 10.0 ^ (elo_difference / 400.0))

[Mike Novack]

Let's do some maths!

.......
So chance of 4/4 is all of those multiplied together = 17%
Chance of 3/4 is complement of each one multiplied by the remaining others added together is = 10% + 3% + 15% + 12% = 30%
So if the prediction model is correct, the chance of it getting 3/4 is about twice that of getting 4/4.

And do keep going. By using the complements you can use the same math to compute the probability of the outcomes 0/4 and 1/4 (which will both be small). And then by adding up the probabilities of the outcomes 4/4, 3/4, 1/4, and 0/4 and subtracting from unity get 2/4.

Surprised by how large that is? (compared to the others)

[Uberdude]

Do you think I don't know how to compute basic probabilities, Mike?

[wineandgolover]

Do you think I am stupid, Mike?

Glad that my intuition was right.

I do find this surprising though: If there was to be exactly one lower probability winner, there was a 50% chance it would be Park. Whodathunkit?

Uber, your last comment to Mike was a joke, right? Or do y'all have a history that I am unaware of? Or was his post secretly insulting in some way that I don't get?

[trout]

Semi final pairing(3 games match),

Lee Sedol vs Ahn Sungjoon
Park Younghoon vs Ke Jie

[Uberdude]

My last comment was a genuine question based on this (and other past posts) which seem rather patronising. But again I used a short word with not quite the right meaning so I changed it.

[Mike Novack]

I didn't mean to seem insulting. There does seem to be common misunderstanding about expected results. So I was just trying to point out that if it had been taken all the way would have become clear to people that contrary to the expectations of most, those differences in rank were not enough to make the most likely single outcome other than "pure guess".

Missed opportunity, that's all.

[yoyoma]

Let's do some maths!

Taking the standard ELO win rates for rating difference from http://www.bobnewell.net/nucleus/bnewell.php?itemid=279 (Remi, is this applicable to WHR?)

Lee (3521) vs Tang (3429) 63%
Ke (3561) vs Rong (3231) 87%
Zhou (3444) vs Park (3425) 53%
Ahn (3348) vs Xie (3293) 58%

So chance of 4/4 is all of those multiplied together = 17%
Chance of 3/4 is complement of each one multiplied by the remaining others added together is = 10% + 3% + 15% + 12% = 30%
So if the prediction model is correct, the chance of it getting 3/4 is about twice that of getting 4/4.

For 3/4 there is a typo: 10% + 3% + 15% + 12% = 40% (not 30%).

Here's what I get for all the cases:
4/4 correct: 17%
3/4 correct: 40%
2/4 correct: 32%
1/4 correct: 10%
0/4 correct: 1%

Edit: Ok I ran a few more cases. Setting all 4 games to be 60/40 matchups, you'll mostly get 3/4 or 2/4 correct:
4/4 correct: 13%
3/4 correct: 35%
2/4 correct: 35%
1/4 correct: 15%
0/4 correct: 3%

Setting all 4 games to be 80/20 matchups, you'll mostly get 4/4 or 3/4 correct:
4/4 correct: 41%
3/4 correct: 41%
2/4 correct: 15%
1/4 correct: 3%
0/4 correct: 0.2%