What If There Was A Committee? Comment Count

Brian

Insanely too long, but I fell down the rabbit hole on this one.

There is an annual complaint against the Pairwise when Team X is passed over in favor of considerably less deserving team Y. This is an exercise in pointlessness, but I was curious as to what a tournament that's selected by eyeballing it would look like. Let's pretend I'm the committee and put 16 teams together.

AUTOBIDS

Boston College, North Dakota, Cornell, Michigan, Alabama-Huntsville, and RIT.

AT-LARGE LOCKS

Denver and Miami have the top two records in the country against the #8 and #14 schedules. Wisconsin and St. Cloud are 3 and 6 in RPI and have top ten records against top ten schedules.

THE EXCESSIVELY LARGE BUBBLE

TEAM Record RPI TUC Record Rank SOS Conference finish Conference tourney
Bemidji State 23-9-4 0.543 6-2-1 6 34 1st CHA Third place tie
Yale 20-9-3 0.537 4-1-2 8 40 1st ECAC First round
NMU 20-12-8 0.535 11-8-5 13T 19 4th CCHA Runner-up
Ferris State 21-13-8 0.533 6-10-3 13T 30 3rd CCHA Fourth place
Minnesota-Duluth 22-17-1 0.533 11-14-5 18 9 4th WCHA Fifth place
UNH 17-13-7 0.533 7-10-6 19 12 1st HE First round
Alaska 18-11-9 0.530 6-7-6 16 26 5th CCHA Round of eight
Vermont 17-14-7 0.526 9-9-5 21 13 8th HE Semi-final
Michigan State 19-13-6 0.525 8-10-2 17 27 2nd CCHA Round of eight
Colorado College 19-17-3 0.524 8-14-3 25T 7 6th WCHA First round
Union 21-12-6 0.523 2-4-3 11 39 3rd ECAC Second place
Minnesota 18-19-2 0.519 9-17-2 34 2 7th WCHA First round

Those are the next ten teams in the RPI, the shiniest record remaining after that, and the team KRACH says should be in the tourney that isn't anywhere near these teams in RPI.

Of the above teams the first one off the board is Northern Michigan. The Wildcats have the best TUC record by far of any team with a significant number of games played, a strong RPI, and the best combination of record and schedule strength. NMU is 7-2-5 against this cohort.

Bemidji State is next with their excellent RPI and 7-5-2 record against a 14-game slate of WCHA and CCHA opponents that included a three-point weekend against Northern, a sweep of UMD, and a win over Miami.

And we will take Yale as a 20-9-3 ECAC champ even if KRACH thinks they are worse than eight WCHA teams.

Goodbyes

Now we get down to the tough decisions. Three spots left for eight teams. They come in three sets:

  • High RPI: Ferris State, UMD, UNH.
  • Low RPI: Vermont, Michigan State, CC, Union, Minnesota
  • Straddling: Alaska

Minnesota is mostly included to show how broken KRACH is as a real world selection device. In its world, an under .500 WCHA team that finished seventh in its conference, went 5-3 OOC and has a horrible TUC record would be a three seed. There is an NCAA rule prohibiting teams under .500 from getting at-large bids after Wisconsin pulled that trick off a couple years ago. They're dropped.

Next, we shoot down Michigan State. There are two CCHA teams with big RPI advantages on them. Both have better records against basically equivalent schedules. Taking them would mean taking the other two CCHA teams and having six in the tourney, something that can't be justified given the relative nonconference results.

We also shoot down CC, which didn't do anything in the nonconference or playoffs to disprove the idea it's a below average WCHA team. CC's nonconference consisted of a split against Northeastern, the ninth place team in HE, a win against Cornell, a loss against Maine, and four games against an assortment of AH and CHA teams. KRACH, of course, has them ninth nationally because they're almost .500 in the WCHA.

Union is the next to die with their ugly SOS and nonexistent TUC categories. That's something that can be overlooked when you have a nice RPI, but there's no reason to look at Union's schedule and think they're somehow underrated.

The Real Bubble

TEAM Record RPI TUC Record Rank SOS Conference finish Conference tourney
Ferris State 21-13-8 0.533 6-10-3 13T 30 3rd CCHA Fourth place
Minnesota-Duluth 22-17-1 0.533 11-14-5 18 9 4th WCHA Fifth place
UNH 17-13-7 0.533 7-10-6 19 12 1st HE First round
Alaska 18-11-9 0.530 6-7-6 16 26 5th CCHA Round of eight
Vermont 17-14-7 0.526 9-9-5 21 13 8th HE Semi-final

The only low RPI team we can't dismiss is Vermont, which went 3-3 in six games against RPI #2 Denver and #4 Boston College. They also beat Yale and UMD in single games and went 2-3-1 against UNH. Their TUC record is the most impressive of any team not already selected. They finished eighth(!) in Hockey East, yes, but they were three points from third. Going 6-1 in the nonconference and beating league champ UNH in the first round of the playoffs means they're worth a look.

Ferris has the best record of any remaining team other than ECAC foe Union but they have an ugly TUC record that's made uglier by the details: four of Ferris's six wins are against UNO, the #21 team according to RPI. The others are wins against Michigan and Michigan State.

New Hampshire… same boat, but they are 3-2-1 against Vermont for whatever that's worth.

Alaska swept Ferris, split a trio with Michigan, and tied three of six against Northern. 

UMD is in a similar boat: eight of their eleven TUC wins are against #18 CC (who they played an improbable seven times) and #22 Minnesota. However, Duluth has a better record, RPI, and SOS than UNH and Vermont. They have slightly worse records by a much higher SOS than either of the CCHA teams. Minnesota-Duluth is in.

We have to kill two of these teams. I don't know. Maybe goal differential?

  1. Ferris State: +0.65
  2. Alaska: +0.45
  3. New Hampshire: +0.21
  4. Vermont: +0.05

That does not help at all. This is why they went with the Pairwise. Okay. You cannot possibly put Vermont in the tournament over UNH when UNH has a better record, RPI, SOS, conference finish, and beat Vermont head to head. And I don't think you can leave out UNH without a good reason when they proved themselves vastly superior to all HE teams not named Boston College. So New Hampshire's in. Then you have three teams.

TEAM A has the best record and RPI but weakest schedule.

TEAM B swept team A but has a meaningfully worse record and a worse league and conference finish.

TEAM C beat more really good teams than the other two but lost to more bad teams and finished in eighth place in its conference.

I… I guess I'm going with Ferris State and validating all the complaints. But it's not like this is obvious.

Seedings

Working backwards since those should be the easiest:

16. UAH
15. RIT

Small conference autobids for teams with bad metrics.

14. Ferris State
13. New Hampshire
12. Minnesota-Duluth

Last three in.

11. Michigan

Michigan gets ahead of UMD and Ferris by virtue of common opponents. The other metrics are so close as to be nearly indistinguishable, but Michigan has a major edge in COP against a conference opponent in Ferris and a 10-2 to 8-6 advantage against UMD. The comparison with UNH is basically a push in all categories, so Michigan gets the edge for the strong late-season run.

10. Yale
9. Bemidji State

I guess this is where strong records against weak competition go.

8. Northern Michigan

Clearly the best of the bubble-ish teams.

7. Cornell
6. St. Cloud State

They've separated themselves from the below; it's a coinflip as to which is 6 and which is 7.

5. Boston College
4. North Dakota
3. Wisconsin

Three teams for two one-seeds. Wisconsin has a major edge in comparisons against BC; North Dakota narrowly loses TUC but actually has a much more impressive record since they played 15 games against RPI top ten opponents (and another five against #12) to BC's one. The COP category is BC's mostly because North Dakota went 1-4 against Denver. Since RPI is basically equivalent, I give the nod to North Dakota's SOS.

2. Denver
1. Miami

These are the obvious top two teams in the tournament. Picking between them is not a big deal since the last two teams are by far the least impressive and both should go meekly. Miami does have all three points in the PWR comparison so we'll go with them.

Bracket

That sets us up with one intra-conference matchup in the first round: Cornell versus Yale. We'll swap Yale and BSU.

Fort Wayne

1. Miami vs 16. UAH
8. Northern Michigan vs 9. Yale

Albany

2. Denver vs. 15. RIT
7. Cornell vs. 10. Bemidji State

St. Paul

3. Wisconsin vs. 14 Ferris State
6. St. Cloud State vs. 11. Michigan

Worchester

4. North Dakota vs 13. UNH
5. Boston College vs. 12. Minnesota-Duluth

Attendance will be shaky in Fort Wayne, but there's no way to swap Michigan in since Northern is holding down the 8 seed unless you want to swap the entire matchup. If Northern and St. Cloud had comparable metrics, I'd do it but there's a big enough gap that the bracket integrity is more important.

Differences

Minnesota-Duluth probably should have been in easily, but was left out in favor of Vermont. If you put a gun to my head, I'd say Ferris is more deserving than Alaska. Apparently, in my Northern is slightly underseeded; other than that it's not much different, at least not this year.

If I was the king of college hockey I'd have the committee hand select the last couple at large bids but then use the Pairwise for seeding.

Comments

JonSobel

March 23rd, 2010 at 2:45 PM ^

I haven't seen a system so completely biased to one conference in my life. If you're sub-.500 or barely .500 in any conference and don't get the auto-bid, there's no way you should be sniffing the TUC plateau.

quakk

March 23rd, 2010 at 3:48 PM ^

RIT over Minnesota? Autobids notwithstanding? RIT finished with a 26-11-1 record in CHA, with wins over whom? The lost every significant game (all two of them), outside two wins over Air Force.

Meanwhile, Minnesota played in arguably the toughest conference and knocked off Wisconsin and North Dakota twice each (a #1 seed and a #2 seed), and Michigan State once, just to name a few. Of course, they played all of 6 games out of conference all year.

KRACH is a closed-loop, iterative statistical rating - every team is rated in comparison to its opponents and every other team. It can only do with what it's given. The flaw is not in the rating itself, but in the lack of comparison points, as pointed out by GCS below.

The pairwise, on the other hand, is inherently flawed because of its arbitrary comparison points. And the eyeball method, well, everybody's eyeballs see things somewhat differently. Hence my response to you.

JonSobel

March 24th, 2010 at 3:33 PM ^

But as it turns out, you made the point I was inferring. The closed-loop in this is the flaw when one conference doesn't schedule outside of its own walls on an equal basis with all the others. The few comparison points can and will overvalue or undervalue the conference based on too small a sampling to actually determine conference quality.

The eyeball test on Minnesota was easy this year. They played really well against good opponents in their own conference, but the they also got hammered on a couple of occasions by quality opponents and lost games that a team without the autobid can't.

The overvalue of the conference as a whole extends down to the dregs of the league, not just the top. So, teams like Michigan Tech and Alaska-Anchorage are wins that are valued much too high when common sense says they don't come close to passing the "sniff" test.

Sometimes common sense can be the best determination of something's value.

And for the record, between Minnesota and RIT, I take Minnesota. But I also take Duluth and Ferris ahead of both of them, as well as Maine.

WolverBean

March 23rd, 2010 at 2:50 PM ^

I'm sure I'll get negged for this -- criticizing Brian is always a dangerous move -- but in the interest of holding Michigan grads to higher standards, here goes:

What if there were a committee?

Also, Unverified Voracity wishes it were rocky and arid.

WolverBean

March 24th, 2010 at 11:51 AM ^

But I doubled with LS&A, so I did take English. And for those who have the chance, I strongly recommend John Rubadeau's English 325/425 course. It's a fair amount of work, but there's no better way to learn to write than by writing, having your writing critiqued, and critiquing the writing of others. Also, the class is an incredible experience, and Rubadeau is hilarious (at least, as long as you're not offended by an excess of vulgarity).

GCS

March 23rd, 2010 at 3:03 PM ^

Minnesota is mostly included to show how broken KRACH is as a real world selection device.

Is it possible to reword this so that it doesn't sound like KRACH is some horribly flawed formula? All it's doing is showing how ridiculously unbalanced college hockey schedules are, with way too high of a ratio of conference/non-conference games. If the ratio was closer to what it is for other NCAA sports, there wouldn't be such a huge bump for going .500 in a tough conference.

MaizeAndBlueWahoo

March 23rd, 2010 at 4:04 PM ^

Hey, this wasn't that long.

I was interested to see it though, because my own curiosity has led me down the road of "what if basketball used the Pairwise?" It takes a goddamn long-ass time, because I'm not a programming wizard. But it's something I'm working on.

WCHBlog

March 23rd, 2010 at 4:29 PM ^

Just because the Pairwise math sucks, doesn't mean all math sucks and we need to go the complete opposite with hand-picking the last couple spots.

I would be more interested in seeing a proposal like you worked on a couple years ago where the Pairwise categories were used to tweak the RPI. For example, UMD had a 6 game advantage over BU in the RPI, but BU won the TUC category by half a game, and the common opponent category by about half a game. UMD still wins that comaparison by about 5 games, rather than losing the comparison, which is what kept them out of the tournament this year.

I'm not a math guy, so I'm not sure how the particulars work out, but it seems to me that at least solves one problem with teams that nobody considers credible bubble teams flukily winning comparisons against contenders that clearly had much better seasons--BU and Lowell this year, Northern Michigan two years ago, etc.

ecormany

March 23rd, 2010 at 5:58 PM ^

the major problem with the pairwise is how most comparisons are a best two of three (since most non-conference comparisons have no H2H value), but totally disregard quantity of data. if we want to say that 2-0 beats 8-1 in COp, fine i guess, but it shouldn't count the same as 6-2 over 2-6. as much as i hate "how did they come up with that" decimal rankings (like RPI!), the pairwise's insistence on integer-ness is its biggest weakness.