2023 NCAA WIH Pairwise, KRACH, and GRaNT

[robertearle]

Grant, so it doesn't get lost as our posts crossed in the mail:

I don't know just when it happened - since Sunday PM, at least - but the good news is the USCHO main site is now showing NPI and not RPI.

The bad news? Their numbers are wholly different than Grant's (except maybe for the individual teams' winning percents).

(Theirs has Wisconsin leading Quinnipiac. I like theirs better.)

 

[Still Eeyore]

I see what you did there.

 

[TonyTheTiger20]

I don't know just when it happened - since Sunday PM, at least - but the good news is the USCHO main site is now showing NPI and not RPI.

The bad news? Their numbers are wholly different than Grant's (except maybe for the individual teams' winning percents).

Thanks for letting me know they finally have something up -- yeah they and I are way way off from each other.

Right off the bat I can tell you that their QWB (Quality Wins Bonus) column isn't right; it has almost everyone with a QWB > 0 and that's just not the case. Lindenwood for example only has wins over Syracuse, RIT, and RIT, so they should have a QWB of 0.00.

Other stuff is just totally not even in the ballpark. They have Yale's SOS as 3rd, and mine has them at 15th.

I took USCHO's own NPIs and plugged them in for all of Yale's opponents, and the average opponent's NPI I get is 54.27 -- not even close to the 60.38 that USCHO is showing. And that's using their own numbers. So I have no idea what they're doing but it can't possibly be right:

 

[TonyTheTiger20]

i see what you did there.

*big grin*

 

[TonyTheTiger20]

Grant, so it doesn't get lost as our posts crossed in the mail:

Haha you beat me to it! I was checking it out

 

[HockeyGuy362]

*big grin*

It may just be my computer, but the names on Pairwise calculator are unreadalbe?

 

[Numbers]

Grant...

Please help me figure something out.....Using Yale's numbers....I'm not going to calculate the unadjusted NPIs. I'll assume you have that right. But, once they are calculated...You would have, for Yale.
Win %age....0.9203
oNPI......0.5189
It's easy to get from there to the team's own NPI. And, that NPI = 0.6191. You are calling that "unadjusted NPI". There must be 2 steps from there to the final number, because there is a number called "adjusted NPI" which is, in Yale's case....0.6299. I'm not seeing how you got that number.
And, from there, just add the QWB to get the final number. I get that.

Thanks for helping me understand the difference between unadjusted and adjusted.

Thanks.


ETA: In fact, it's only the top 7 teams whose adjusted NPI is different from their unadjusted NPI. That makes me think you have a problem somewhere.

 

[TonyTheTiger20]

Grant...

Please help me figure something out.....Using Yale's numbers....I'm not going to calculate the unadjusted NPIs. I'll assume you have that right. But, once they are calculated...You would have, for Yale.
Win %age....0.9203
oNPI......0.5189
It's easy to get from there to the team's own NPI. And, that NPI = 0.6191. You are calling that "unadjusted NPI". There must be 2 steps from there to the final number, because there is a number called "adjusted NPI" which is, in Yale's case....0.6299. I'm not seeing how you got that number.
And, from there, just add the QWB to get the final number. I get that.

Thanks for helping me understand the difference between unadjusted and adjusted.

Thanks.


ETA: In fact, it's only the top 7 teams whose adjusted NPI is different from their unadjusted NPI. That makes me think you have a problem somewhere.

"Adjusted" NPI is the NPI you get when you remove "bad wins." Yale has 12 wins which lower their NPI, so those get taken out of the calculation:

The way you calculate "Final NPI" is you add up everything in the "Game NPI" column that doesn't have anything in the "Removal" column, and divide it by how many of those games there are (in this case, by 11)

Only the really good teams will typically have NPIs high enough where a win can lower their NPI, so that's why only 7 teams or so have an "Adjusted NPI" different from their "Final NPI"

 

[TonyTheTiger20]

It may just be my computer, but the names on Pairwise calculator are unreadalbe?

It's possible it's your browser's zoom settings -- are you looking on your phone or on a computer? If you want to send me a screenshot I'm happy to take a look at what you're seeing. Grant dot Salzano at gmail.

 

[Numbers]

"Adjusted" NPI is the NPI you get when you remove "bad wins." Yale has 12 wins which lower their NPI, so those get taken out of the calculation:

The way you calculate "Final NPI" is you add up everything in the "Game NPI" column that doesn't have anything in the "Removal" column, and divide it by how many of those games there are (in this case, by 11)

Only the really good teams will typically have NPIs high enough where a win can lower their NPI, so that's why only 7 teams or so have an "Adjusted NPI" different from their "Final NPI"

So, it's still Mickey Mouse. A good system wouldn't need to remove any games. But, it's better than a straight PRI calculation.

The QWB should really disappear. That's just trash. It makes a win against a good team better than a loss against a bad one. Those 2 are equivalent, really. And, the game removal business is stupid, too. But, it's a small improvement

Edit to add: However, in your earlier post, you alluded to the possibility of calculating this in a different manner (just like you can calculate RPI game-by-game, or in toto). Can you somehow show the algebra that the way you are doing it is equivalent to the other method? I'm asking because I'm curious as to whether the 2 different recursions might converge differently?

 

[NMH]

Just an FYI from what I understand, the pre-championship manual is going to be re-published and it seems the calculations have changed more than we initially realized.

 

[Steamboat]

The 2023 pre-champs manual is out, link here if you're interested in reading through it. Mostly stuff we already knew -- NEWHA autobid, 11 teams, emphasis on non-conf matchups rather than reducing flights -- however one noteworthy change is that the NPI (replacing RPI) has a weight of 25/75 between WinPct/SOS compared to the 30/70 that it used to be. This increased weighting on SOS hurts the eastern teams with weaker schedules. If you're interested in how the change affects the rankings as they sit right now, here's a tweet with a couple screenshots:

https://twitter.com/Salzano14/status/1616237221241036800

In the meantime, my Pairwise calculator has been adjusted accordingly with the new weighting.

I see that the poor Bemidji Beavers have the toughest Strength of Schedule. Four games apiece against the rest of the WCHA teams. This is a special year, with Olympic players back on the better teams, and 5 year players using a Covid year for extra play. I expect that the Beavers will resume being a thorn in the side of certain programs. Not this year however.

 

[TonyTheTiger20]

Just an FYI from what I understand, the pre-championship manual is going to be re-published and it seems the calculations have changed more than we initially realized.

Thanks Nicole.

Absolutely absurd that it's February and the criteria for selection isn't even public.

 

[TonyTheTiger20]

So, it's still Mickey Mouse. A good system wouldn't need to remove any games. But, it's better than a straight PRI calculation.

The QWB should really disappear. That's just trash. It makes a win against a good team better than a loss against a bad one. Those 2 are equivalent, really. And, the game removal business is stupid, too. But, it's a small improvement

Edit to add: However, in your earlier post, you alluded to the possibility of calculating this in a different manner (just like you can calculate RPI game-by-game, or in toto). Can you somehow show the algebra that the way you are doing it is equivalent to the other method? I'm asking because I'm curious as to whether the 2 different recursions might converge differently?

Oh it's definitely still Mickey Mouse haha

The two ways would be (with Yale used as an example):

[(Yale's overall winning percentage, .925)x(.25)] + [(Average of all Yale opponents' NPIs)x(.75)]

versus

{
[(Yale's winning percentage in game 1 vs Harvard, 1.00)x(.25)] + [(Harvard's NPI)x(.75)]
+
[(Yale's winning percentage in game 2 vs. Dartmouth, 1.00)x(.25)] + [(Dartmouth's NPI)x(.75)]
+
...
+
[(Yale's winning percentage in game 23 vs. Clarkson, 1.00)x(.25)] + [(Clarkson's NPI)x(.75)]
}
All divided by Yale's total number of games, 23

^That one^ is essentially taking an average of all of Yale's individual game NPIs as opposed to calculating it all at once. The advantage of the second one is it allows you to easily remove "bad wins" from the calculation.

 

[robertearle]

Thanks Nicole.

Absolutely absurd that it's February and the criteria for selection isn't even public.

Wouldn't be the first time.

I found my way here to the USCHO forum in 2013 looking for info on Pairwise that year, when the men's side had changed the way they calculated 'Common Opponents' from a winning percentage of total games to a team-by-team addition of percentages, and whether the women's Pairwise was being done the same way. USCHO still had the women's being done the 'old way', and under the old way Wisconsin was the #8 team while under the new way North Dakota was #8. As late as the night before the NCAA selection, ND coach Brian Idalski thought he team wasn't going to make it; he gave a post game WCHA tournament press conference where it was obvious he thought his team was done. But when the selection was announced, ND was in and Wisconsin was out. Because they had switched to the 'new' Common Opponent. Even though nobody seemed to know they had.

 

[ARM]

For the 2013 (Thanks Robert!) UND vs UW case, it was pretty much recognized heading into their semifinal that the loser was going to be out. The surprise was that the USCHO calculation was showing UW ahead of UND after the Whioux had lost in the final. Idalski was actually much better than a lot of coaches in that era, at least a few years earlier. Johnson would say that he had no idea of how any of that worked, and Miller would claim that her team should be in because they had a couple of wins over teams in the field and had played a tough schedule. Up until then USCHO had been pretty accurate, and most of the surprises dealt with swaps to avoid travel and the like.

 

[robertearle]

For the 2014 UND vs UW case, it was pretty much recognized heading into their semifinal that the loser was going to be out. The surprise was that the USCHO calculation was showing UW ahead of UND after the Whioux had lost in the final. Idalski was actually much better than a lot of coaches in that era, at least a few years earlier. Johnson would say that he had no idea of how any of that worked, and Miller would claim that her team should be in because they had a couple of wins over teams in the field and had played a tough schedule. Up until then USCHO had been pretty accurate, and most of the surprises dealt with swaps to avoid travel and the like.

USCHO was showing UW ahead because under the old 'Common Opponents', they were ahead.

(it was 2012-2013 when UW didn't make the tournament. 2013-2014, they had a lead vs Minn but ended up losing in the NCAA semi 5-3.)

 

[ARM]

2013-2014, they had a lead vs Minn but ended up losing in the NCAA semi 5-3.

You're correct. I recently watched highlights of both games UM played in that FF, and I don't remember another event where so many goals resulted against multiple goalies because they were off their angles. Made me wonder if there was something different optically in the Q rink that contributed to goaltenders getting "lost" in their crease. It sometimes happens on the wider ice sheets when they are used to the angle being a certain measure along the boards at the blue line, the hash marks, the dots, etc., but AFAIK, the Q rink is a standard NHL-width sheet.

 

[TonyTheTiger20]

Alright, I updated my Pairwise to reflect the new interpretation of Quality Wins Bonus:

According to info I've gotten, the calculation for QWB is:

0.5 x (OppNPI - 51.50) x Game Result*
Add up all QWBs and divide by total games played

*game result meaning 1.000 for a regulation win, 0.667 for an OT win, 0.333 for a OT loss, 0.500 for a tie

I still differ dramatically from USCHO on the overall numbers (and by the QWBs as well, but that's because we differ in who has an NPI of over 0.515) but from what I understand USCHO isn't confident in their numbers. I feel good about mine being right. I am going to try to hunt around and see if I match the official selection numbers and will let you know.

 

[robertearle]

Alright, I updated my Pairwise to reflect the new interpretation of Quality Wins Bonus:

According to info I've gotten, the calculation for QWB is:

0.5 x (OppNPI - 51.50) x Game Result*
Add up all QWBs and divide by total games played

*game result meaning 1.000 for a regulation win, 0.667 for an OT win, 0.333 for a OT loss, 0.500 for a tie

I still differ dramatically from USCHO on the overall numbers (and by the QWBs as well, but that's because we differ in who has an NPI of over 0.515) but from what I understand USCHO isn't confident in their numbers. I feel good about mine being right. I am going to try to hunt around and see if I match the official selection numbers and will let you know.

Does that turn QWB into a sort-of "TUC cliff", or was it always, and I just hadn't thought about it that way?

One more time, thank you for the effort you put into this.

(And not to throw shade at LIU - they're actually doing a good job for a new program - but right now a win over LIU is a "quality win", albeit a small one.)