PDA

View Full Version : Latest D1 Power Rankings



octonion
12-18-2014, 05:05 PM
The full list (https://github.com/octonion/hockey/blob/master/uscho_sos/current_ranking.txt) and R code (https://github.com/octonion/hockey/tree/master/uscho_sos).



rk | school | div | str | ofs | dfs | sos
-----+---------------------------------+-----+-------+-------+-------+-------
1 | harvard | 1 | 3.893 | 1.878 | 0.483 | 1.464
2 | minnesota-duluth | 1 | 3.820 | 1.858 | 0.486 | 1.593
3 | vermont | 1 | 3.743 | 1.621 | 0.433 | 1.442
4 | michigan-tech | 1 | 3.728 | 1.603 | 0.430 | 1.456
5 | minnesota-state | 1 | 3.666 | 1.965 | 0.536 | 1.471
6 | boston-university | 1 | 3.620 | 1.759 | 0.486 | 1.449
7 | nebraska-omaha | 1 | 3.467 | 1.835 | 0.529 | 1.511
8 | minnesota | 1 | 3.462 | 1.776 | 0.513 | 1.465
9 | miami | 1 | 3.442 | 1.703 | 0.495 | 1.526
10 | bowling-green | 1 | 3.394 | 1.770 | 0.522 | 1.461
11 | north-dakota | 1 | 3.133 | 1.649 | 0.526 | 1.460
12 | st-lawrence | 1 | 3.113 | 1.679 | 0.539 | 1.460
13 | denver | 1 | 3.031 | 1.685 | 0.556 | 1.503
14 | umass-lowell | 1 | 2.968 | 1.836 | 0.619 | 1.455
15 | penn-state | 1 | 2.887 | 1.741 | 0.603 | 1.455
16 | merrimack | 1 | 2.865 | 1.452 | 0.507 | 1.446
17 | ferris-state | 1 | 2.844 | 1.427 | 0.502 | 1.434
18 | providence | 1 | 2.789 | 1.318 | 0.472 | 1.420
19 | colgate | 1 | 2.743 | 1.478 | 0.539 | 1.445
20 | michigan | 1 | 2.740 | 1.866 | 0.681 | 1.454
21 | yale | 1 | 2.732 | 1.366 | 0.500 | 1.449
22 | cornell | 1 | 2.720 | 1.242 | 0.456 | 1.470
23 | robert-morris | 1 | 2.707 | 1.687 | 0.623 | 1.283
24 | boston-college | 1 | 2.679 | 1.536 | 0.573 | 1.461
25 | union | 1 | 2.642 | 1.699 | 0.643 | 1.430

Red Cows
12-18-2014, 06:11 PM
O.K.

I'll bite.

Just what the heck are we looking at here?

Patman
12-18-2014, 09:56 PM
O.K.

I'll bite.

Just what the heck are we looking at here?

looks like a Poisson random effects model of goal data estimated under frequentist techniques.

---

octonion has never said this but I've come to learn that he works in statistics for professional sports.

edit: I've toyed with doing roughly the same under Bayes paradigm... just been too lazy and I don't want to do it without a season simulator. That and I'd have to custom write the software which is... annoying.

darker98
12-18-2014, 10:49 PM
This makes no sense to a normal observer.

Red Cows
12-19-2014, 12:28 AM
This makes no sense to a normal observer.

Kinda my observation as well.

This bears little semblance to Pairwise, KRACH, the USCHO poll, or ESPN's poll.

I concede that those 4 things bear little semblance to each other right now, either (other than Pairwise and KRACH).

I note that octonion's rep went from green to red since my original post above.

I'll reserve further judgement until the poster explains what we are looking at and the derivation of it.

Tater
12-19-2014, 06:05 AM
It's the only system that has gotten Harvard right.

northeastern
12-19-2014, 08:09 AM
It's the only system that has gotten Harvard right.

..or least their average gpa ;-)

Ralph Baer
12-19-2014, 08:49 AM
This can't be correct. Union is #26, not #25. :)

manurespreader
12-19-2014, 10:01 AM
Despite the mistakes, it appears to me this is a regression analysis that uses one of two methods to account for the randomness of some results.But i'm not a mathematician like ralph is.

Ralph Baer
12-19-2014, 01:11 PM
Despite the mistakes, it appears to me this is a regression analysis that uses one of two methods to account for the randomness of some results.But i'm not a mathematician like ralph is.

I am sure that Patman knows a lot more about this than I do.

UMD21
12-19-2014, 01:43 PM
UMD right up the top in all the rankings, makes sense to me :D

Patman
12-19-2014, 05:06 PM
I am sure that Patman knows a lot more about this than I do.

yeah, I didn't see anything ostensibly wrong in the code. On principle I'd be worried about whether the necessary linear restriction is in the model for all team related terms. Team effects are only observed relative to other teams and thus is sucked out by the intercept term.

Patman
12-19-2014, 05:09 PM
It's the only system that has gotten Harvard right.

IMO, I tend to go with score-based models because it does attempt to capture the nature of the game itself... whatever imperfection it may be. The difference in people's opinions tends to be the question or whether teams can will themselves to win games versus winning games coming as a process of being trained to score.

Clearly the players are not emotionless automatons so there is certainly give in the latter... however I think that coaches in the modern era have come to the conclusion that what they do is a process and that the game itself is statement of that process, albeit an imperfect statement.

They aren't going to express it in formulae like us nerds may but the mindset is similar.

manurespreader
12-19-2014, 05:19 PM
IMO, I tend to go with score-based models because it does attempt to capture the nature of the game itself... whatever imperfection it may be. The difference in people's opinions tends to be the question or whether teams can will themselves to win games versus winning games coming as a process of being trained to score.

Clearly the players are not emotionless automatons so there is certainly give in the latter... however I think that coaches in the modern era have come to the conclusion that what they do is a process and that the game itself is statement of that process, albeit an imperfect statement.

They aren't going to express it in formulae like us nerds may but the mindset is similar.
Might be, I wonder if you think this is strictly a ranking system based on past history, or is it a predictive model,( if a regression analysis can be considered such), and second, is it any more meaningful than Krach is.

And third, is the poster a Harvard Guy? Since I tend to ascribe ulterior motives to things like this.

Patman
12-19-2014, 05:58 PM
Might be, I wonder if you think this is strictly a ranking system based on past history, or is it a predictive model,( if a regression analysis can be considered such), and second, is it any more meaningful than Kratch is.

And third, is the poster a Harvard Guy? Since I tend to ascribe ulterior motives to things like this.

I don't believe so... the quickest way to see Harvard ranked that high is legit is to look at Robin Locke's results (Robin is a professor at St. Lawrence). His model uses score data at the end of regulation less any ENG scored.

In that one Harvard is 6th.

As a prediction method... and yes it is... it can be more meaningful than KRACH. While there is an appeal of probabilistic notions in KRACH the fact they have to deal with ties makes it problematic and thus is better viewed from the heuristic standpoint.

edit: however, KRACH applied to "non-tie" sports can be viewed as predictive in nature.

---

I don't know if he's a "harvard guy" though I believe it took another Harvard guy to tell me who he is. I know of one his employers but I feel it is not warranted to divulge this information though I believe he is no longer in their employ. I have met "octonion" before but that was before he posted here as "octonion".

edit: i only found out who he was at the Pittsburgh frozen four

goblue78
12-21-2014, 04:36 PM
Yes, this is a pretty standard Poisson model. The one issue I've had with it in the past (and some might find this a feature, not a bug) is that there is partial carryover of strength from prior seasons (or at least there used to be... I haven't really studied his code this year). It is predictive in the sense that it generates an offensive and defensive strength for each team. It is not predictive in the sense that it doesn't look at what games remain to be played for any team and project any sort of W-L-T record from that. For a quick survey of how these models work, anyone can look at the documentation for the model I wrote in 2011... https://dl.dropboxusercontent.com/u/5755704/ranking.doc
It discusses how these models differ from KRACH and the Pairwise and how to use one predictively...

Patman
12-22-2014, 12:39 AM
Yes, this is a pretty standard Poisson model. The one issue I've had with it in the past (and some might find this a feature, not a bug) is that there is partial carryover of strength from prior seasons (or at least there used to be... I haven't really studied his code this year). It is predictive in the sense that it generates an offensive and defensive strength for each team. It is not predictive in the sense that it doesn't look at what games remain to be played for any team and project any sort of W-L-T record from that. For a quick survey of how these models work, anyone can look at the documentation for the model I wrote in 2011... https://dl.dropboxusercontent.com/u/5755704/ranking.doc
It discusses how these models differ from KRACH and the Pairwise and how to use one predictively...

predictive would just be the follow-up execution. The model is predictive... but if you don't use it to make predictions... well, that's on you :p