How to improve the Pairwise

[Shirtless Guy]

Re: How to improve the Pairwise

Great idea! Think people will fill the barn to watch Clarkson - Bemidji?

Ideally an 8 team conference with 28 conference games works best. That gives 6 OOC that you can min/max out costs/revenues while giving a shot to your RPI and PWR.

You really think attendance in Bemidji would be much different for LSSU or Clarkson? I don't.

8 teams playing a 28 game schedule with 6 OOC isn't ideal but it probably would be better to some extent because at least within conference everything would be balanced. I would think ideally, conferences would be capped at 24 conference games, but until the big boys are limited in their ability to force home games, that isn't a good solution to help keep games in small school venues. The reason the WCHA has 28 conference games isn't because LSSU outdraws Clarkson...its because it guarantees every team at least 14 home games...

 

[Shirtless Guy]

Re: How to improve the Pairwise

How many NC games involving other conference reps for each of Cornell, MTU and Min would it take to get them? Any guesses on how wild the swings that EoDS is talking about?

 

[tape]

Re: How to improve the Pairwise

Ideally an 8 team conference with 28 conference games works best. That gives 6 OOC that you can min/max out costs/revenues while giving a shot to your RPI and PWR.

IMO that's way too many conference games. Until Notre Dame joined, Hockey East used to play 27 conference games. Now they play 22 conference games which leaves 12 non-conf games, which I'm really liking. It's more fun to see a greater variety of teams come to town, and it creates better data for the Pairwise.

 

[Patman]

Which notion? To me trying to measure current performance is saying that you're in the "more likely to win the tournament" camp. I don't worry about current performance because I'm in the "reward for a full season's work" camp, so the notion of "current performance" is irrelevant.

Says you.

----

And generally, I agree, but it's an argument one can make and one may make

 

[manurespreader]

Re: How to improve the Pairwise

I feel your pain, Shirtless, but my take on it (as a professional statistician when I'm not watching hockey) is that there is no possible system which can reliably separate the 14th best team from the 15th best team in a league where there is only a modest amount of mixing between conferences. I have no problem living with the notion that there are probably six teams in any given year that any system would accept, another 4 or 5 that any half-decent system accepts, and that the rest of the spots (after AQ) are always going to be methodology-dependent in a way that will be fairly arbitrary. .

Which is exactly why I proposed lowering the QWB to the top 10 teams, because there really is a lot of parity below the top few teams.

 

[UML Puck Hawk]

Which is exactly why I proposed lowering the QWB to the top 10 teams, because there really is a lot of parity below the top few teams.

Then move back to an 8 or 12 team field if you think there's an arbitrary cut line. Then again the past few years the last team in has won, so maybe there's more parity than you think

 

[Shirtless Guy]

Re: How to improve the Pairwise

Which is exactly why I proposed lowering the QWB to the top 10 teams, because there really is a lot of parity below the top few teams.

Just an FYI, that would not have change the order of teams at all this year...same top 14 same order...15 through 24 would shuffle but the top 14 is unchanged. (assuming the award for beating #11 through #20 would get shifted to #1 through #20...so beating #10 would be worth 0.0025 for a neutral site, etc).

 

[manurespreader]

Re: How to improve the Pairwise

Then move back to an 8 or 12 team field if you think there's an arbitrary cut line. Then again the past few years the last team in has won, so maybe there's more parity than you think

probably is more parity, which makes my point, exactly. Beating number 18 is no better than beating number 28 so why should you get extra points from it.

 

[Shirtless Guy]

Re: How to improve the Pairwise

Then move back to an 8 or 12 team field if you think there's an arbitrary cut line. Then again the past few years the last team in has won, so maybe there's more parity than you think

I think the point is that awarding a QWB for 1-10 has more value because it is clearer that a team deserves a QWB for beating those teams...maybe the answer is actually to go the other way...going all the way down to 30 with max remaining the same (the old TUC cliff line was anyone above 0.5000 RPI which usually was around 25ish).

 

[mookie1995]

I think the point is that awarding a QWB for 1-10 has more value because it is clearer that a team deserves a QWB for beating those teams...maybe the answer is actually to go the other way...going all the way down to 30 with max remaining the same (the old TUC cliff line was anyone above 0.5000 RPI which usually was around 25ish).

How about applying a negative bonus to beating a chitty team, and double it if the game is at home. And quadruple if it is an ooc game at home against a chitty team.

 

[CLS]

Re: How to improve the Pairwise

Somebody help me. Why does there need to be an arbitrary cutoff line? Is there a statistical reason (not a subjective one) why beating say, number 10 couldn't be worth just a little more than beating number 11, but worth a lot more than beating number 45?

 

[Shirtless Guy]

Re: How to improve the Pairwise

Somebody help me. Why does there need to be an arbitrary cutoff line? Is there a statistical reason (not a subjective one) why beating say, number 10 couldn't be worth just a little more than beating number 11, but worth a lot more than beating number 45?

Well the there is no "cliff" anymore currently there is a bonus for beating (or half as much for tying) someone in the top 20 but if you beat #1, you get 20x as much bonus as you would for beating #20. The bonus varies from 0.0025 to 0.0500 right now with adjustments for location of the game...

 

[goblue78]

Re: How to improve the Pairwise

Somebody help me. Why does there need to be an arbitrary cutoff line? Is there a statistical reason (not a subjective one) why beating say, number 10 couldn't be worth just a little more than beating number 11, but worth a lot more than beating number 45?

We already have that system. It's called RPI. QWB was created to add more than RPI already gave you, without taking away points to losing to bad teams. The idea was to encourage scheduling good teams, since you could add QWB if you beat them, and you had the smallest possible loss in RPI.

 

[joecct]

Then move back to an 8 or 12 team field if you think there's an arbitrary cut line. Then again the past few years the last team in has won, so maybe there's more parity than you think

6. Conference champions only. Then the conference tournament intensity level would ramp up 100% and the barns would be full(er).

Tournament would be Weds (2 v 3) games, Thursday (semis) and Saturday (finals).

 

[Shirtless Guy]

Re: How to improve the Pairwise

We already have that system. It's called RPI. QWB was created to add more than RPI already gave you, without taking away points to losing to bad teams. The idea was to encourage scheduling good teams, since you could add QWB if you beat them, and you had the smallest possible loss in RPI.

Mr Stats, would it make sense to give some form of bonus to anyone with an RPI above 0.5000 and have it scale properly so you got more bonus for beating #1 (0.5938) while getting similar bonuses for beating teams banded together? ie beating #1 gives you a bonus of 0.0500 like it does now and beating any team below 0.5000 gives you zero but anything in between is scaled based on their relative difference between #1 and 0.5000? So with #2 at 0.5929 the bonus for beating them would be 0.0495 [from (0.05*(0.5929-0.5)/(0.5938-0.5)] instead of the current 0.0475?

 

[manurespreader]

Re: How to improve the Pairwise

Just an FYI, that would not have change the order of teams at all this year...same top 14 same order...15 through 24 would shuffle but the top 14 is unchanged. (assuming the award for beating #11 through #20 would get shifted to #1 through #20...so beating #10 would be worth 0.0025 for a neutral site, etc).

Not quite what I was proposing. What I'm saying is, award zero points for beating anyone no. 11 thru infinity. Only award points at the same level as now,( not increased) to 1 -10. In other words, there is a benefit that should accrue to you for beating a really good team, sure, but for beating the vast majority, no. And secondly, that benefit should not be higher than it is now simply because the 11-20 teams are no longer part of it. As far as scaling it to the 1- 10, sure why not.

 

[Shirtless Guy]

Re: How to improve the Pairwise

Not quite what I was proposing. What I'm saying is, award zero points for beating anyone no. 11 thru infinity. Only award points at the same level as now,( not increased) to 1 -10.

but then you are created a TUC cliff, currently beating #10 gets you 0.0275 for a neutral site game. You're proposing 0 for beating #11 which means everyone cares if the top ten changes. I'm saying, we basically base the bonus on how high your RPI is above 0.5000 with a max of 0.0500 (the current max bonus) one thing also to remember is while you get a bonus of 0.0500 for beating #1, your actually final RPI only goes up the sum of all QWBs divided by games played. So if a team had one quality event in their 40 game season prior to selection day, beating #1 under the current system they would get a bonus of 0.0500/40 or 0.00125...

Here is what I calculated for this year:

Rnk	QWB	aQWB	Team
1	0.0500	0.0500	Quinnipiac
2	0.0475	0.0495	St. Cloud State
3	0.0450	0.0471	North Dakota
4	0.0425	0.0406	Providence
5	0.0400	0.0379	Boston College
6	0.0375	0.0377	Denver
7	0.0350	0.0358	Michigan
8	0.0325	0.0277	Harvard
9	0.0300	0.0276	Mass.-Lowell
10	0.0275	0.0261	Boston University
11	0.0250	0.0258	Yale
12	0.0225	0.0246	Notre Dame
13	0.0200	0.0228	Northeastern
14	0.0175	0.0196	Minnesota-Duluth
15	0.0150	0.0189	Michigan Tech
16	0.0125	0.0166	Cornell
17	0.0100	0.0139	Minnesota
18	0.0075	0.0130	Robert Morris
19	0.0050	0.0122	Nebraska-Omaha
20	0.0025	0.0119	Penn State
21	0.0000	0.0117	St. Lawrence
22	0.0000	0.0109	Dartmouth
23	0.0000	0.0100	Minnesota State
24	0.0000	0.0094	Clarkson
25	0.0000	0.0089	Rensselaer
26	0.0000	0.0069	Bowling Green
27	0.0000	0.0065	Miami
28	0.0000	0.0045	Air Force
29	0.0000	0.0037	Ferris State

 

[manurespreader]

Re: How to improve the Pairwise

but then you are created a TUC cliff, currently beating #10 gets you 0.0275 for a neutral site game. You're proposing 0 for beating #11 which means everyone cares if the top ten changes. I'm saying, we basically base the bonus on how high your RPI is above 0.5000 with a max of 0.0500 (the current max bonus) one thing also to remember is while you get a bonus of 0.0500 for beating #1, your actually final RPI only goes up the sum of all QWBs divided by games played. So if a team had one quality event in their 40 game season prior to selection day, beating #1 under the current system they would get a bonus of 0.0500/40 or 0.00125...

Well yes but... in essence there still must be cliff, as the cutoff, whether it is number 11 or 15 or 20 or whatever is somewhere. I'm just saying that you should get some bonus but to what extent and from what result is the question. The top 20 was arbitrary and so will any other number be. I agree that you do deserve an award, but as the umich stats guy says, there are really only 5 or 6 top teams.

 

[Shirtless Guy]

Re: How to improve the Pairwise

Well yes but... in essence there still must be cliff, as the cutoff, whether it is number 11 or 15 or 20 or whatever is somewhere. I'm just saying that you should get some bonus but to what extent and from what result is the question. The top 20 was arbitrary and so will any other number be. I agree that you do deserve an award, but as the umich stats guy says, there are really only 5 or 6 top teams.

but currently you get 0.0025 for #20 and and 0 for #21...that isn't nearly as big a deal. Currently there really isn't a cliff as much as a gentle slope...see my edit to the post you're commenting on...

 

[manurespreader]

Re: How to improve the Pairwise

but currently you get 0.0025 for #20 and and 0 for #21...that isn't nearly as big a deal. Currently there really isn't a cliff as much as a gentle slope...see my edit to the post you're commenting on...

Right and what I'm saying is, make the slope steeper and in fact why does it have to be linear at all.