Looking back at 2016 through the lens of 2017

Back in May (!) I mentioned that I was looking at last year’s playoff picture through the lens of the new conference tournament plan.  I’ve finally finished that after far, far too long.  (I’m mainly pushing it now to get it to “done” so I can add this logic to the overall 2016-17 ABOVE model.  Yes, I named it ABOVE.  You shut up.)

Final 2015-16 BELOW, Playoff Chances

TeamFinal BELOWTotal %Factor
1-MTU171525.806%4.900
2-MSU165618.831%3.424
3-BGSU162515.111%2.879
4-FSU152711.088%2.106
5-NMU15019.373%1.780
6-BSU15569.520%1.808
7-LSSU14635.804%1.102
8-UAF14515.266%1523

As you can see, despite a lower seeding, the model has Bemidji State slightly more likely to win overall than Northern Michigan based on the Beavers’ higher BELOW rating.

That the top four teams went to the semifinals is no surprise — but it was just about a 50-50 proposition.  Part of that is because Bemidji was a solid squad last season.

Those percentages come from a run of 5,000,000 trials of a model that does the following:

  1. Plays out the four quarterfinal series: MTU-UAF, MSU-LSSU, BGSU-BSU, FSU-NMU.
  2. Creates a new bracket among those four teams and plays another three-game series on each side of the bracket.
  3. Plays a one-game final.

So for what we actually saw, we’re looking at a 1234-24-4 bracket run.  In 5,000,000 trials, that happens about 1.177% of the time.  In comparison, the most absurd response — lower seeds losing at every turn — happens about 0.180% of the time, while chalk — top seeds winning out — happens 2.497% of the time.

In short, Ferris State winning out was unlikely but not shocking.  The big surprise was Michigan Tech not winning.

Now it’s time to finish that ABOVE model…

Say Goodbye to the Western Tie

College hockey was the lone stronghold in the Ties Are Okay part of hockey, but those days are ending, and the WCHA has done their part, announcing today that ties in conference games are a thing of the past.  The 2-1-0 scoring structure is gone, and the 3-2-1-0 structure has arrived.  In short:

  • WCHA teams will play 5×5 for 60:00.  The winner will get 3 points.
  • Teams winning in a 5:00, 5×5 overtime will get a full 3-point win.
  • Teams winning in a subsequent 5:00, 3×3 overtime or a subsequent shootout will get 2 points, with the losing team gaining 1 point.
  • WCHA nets will be shallower — 40″, meeting the NHL standard — than the 44″ standard.

Since the new WCHA was formed in 2013-14, there have been 420 league games.  There were 12 ties in 2013-14, 12 in 2014-15, and 20 in 2015-16.  Even as much as everyone — including me! — made light of ties in the WCHA last year the tie rate faded as the season went on.

I’ll have more on this as the season goes by — this change this close to the season is pretty much submarining my chance of having a public model on October 1.  But here are my preliminary thoughts:

  • I have generally modeled overtime by looking at the amount of overtime to-date in the season (and in the past season if it’s early).  This has been a pretty good model, because the rate of overtime games is roughly even for the fact that we only have 140 games as samples.  (I could model this as a Poisson distribution, but I have better things to do with my time.)
  • I think that I’ll keep that model going forward, using overtime games from the past 140 matches as the basis for whether a game will play past 65:00.  Why?  The incentive is there for the weaker team to hold on for a single point, even if the opponent will get two.
  • After 65:00, it will be something like a coin flip, whether 50-50 or weighted by a goals-for ratio between the two teams.  Example: for those two late-season Mankato-Huntsville ties, I’d do a coin that’s weighted 82/61 between the two teams to determine a winner.
  • The frustration about this is going to be that you’ll have to keep a track of when goals are scored and games are won all season long to figure out which teams are better in a BELOW sense.  I’m going to treat a tie just like I did before and then have a follow-on kabuki dance where we come up with a winner.  While this matters for the standings that a BELOW-based model will output, it won’t matter in terms of BELOW thinking which team is better — i.e., it won’t matter who gets to break a 65:00 tie, much less how they break it.

If you have ideas or questions, leave a comment here or reply to me at @wchaplayoffs on Twitter.

Looking Back: Would Michigan Tech have made the 2016 NCAAs?

In light of everything with the new WCHA playoff system, I had a lingering question: What effect would the semifinal rules have had on Michigan Tech’s chances of making the NCAAs as an at-large team?

Luckily, I knew the guy to ask: Tim Braun of TechHockeyGuide.com.  His answer:

Michigan Tech would still need to win the Title to get in…even sweeping FSU at home and losing to MSU in Houghton would have MTU first team out.

[T]hey’d pickup almost 40 RPI points, enough to pass Cornell, but not UMD…that would get them to 0.5398 (UMD was up at 0.5440).

The new, truly-insular schedule for 2016-17 will make it hard for WCHA teams to make the NCAAs.  WCHA teams play just 32% of their non-conference games at home this season (39% if you count the Alaska tournaments) per UAHHockey.com’s Michael Napier.  It’s no wonder that the WCHA is looking to reduce the number of league games to give member schools more opportunities to play out-of-conference.  The WCHA’s woeful non-conference totals in 2015-16 — tabulated by Troy Mills of the Beaver Hockey Pond — tells the sad tale of woe:

The WCHA non-conference schedule is complete and the league ended up with a record of 27-36-9.

-5 against Hockey East (3-2-0)
-4 against Atlantic Hockey (3-0-1)
-26 against the NCHC (4-20-2)
-24 against the Big10 (10-9-5)
-7 against the ECAC (4-2-1)
-6 against Arizona State (3-3-0)

It’s the NCHC play that killed the league.  Four teams from the Nacho made the NCAAs, and the results weren’t pretty:

  • UAH went 0-2-0 against eventual national champion North Dakota, although one of those games was 1-0, and split with Colorado College.
  • UAA lost to St. Cloud.
  • UAF also lost to St. Cloud.
  • BSU went 2-0-0 against Duluth, 0-1-1 against North Dakota, and lost to St. Cloud.
  • BGSU lost two games each to Miami and Western Michigan, a fact that probably rankles our friends at BGSUHockey.com.
  • FSU went 0-1-1 against Western.
  • LSSU lost to North Dakota, and they sadly got swept at home by newbie Arizona State.
  • MTU didn’t play any NCHC teams, and their big regrets are in conference play.
  • MSU lost three games to St. Cloud and two to Nebraska-Omaha.
  • NMU split with Duluth.

One way to consider the new playoff structure — which I like a lot! — is that playing two or three semifinal games will lower the chance of upsets.  While that lowers the probability of hurting a league team in PWR — again, it kept Tech out in 2016 — Ferris State made the 2014 NCAA field as an at-large despite losing the Broadmoor to Minnesota State.

I hope to have answers on the probabilities for the 2016 semifinals and final game in the next week or so.  But there’s also …

[My thanks to jsmithe for correcting my oversight of LSSU playing NoDak early in the season.]

Looking Back: 2016 WCHA First-Round Probabilities

I never did do first-round predictions this year, in part because I ran out of time and in part because I was frustrated with my toolset.  I’ve got a good start on a new toolset, and so now I want to look back at the 2016 first round in light of the new WCHA playoff system in place for 2017, 2018, and 2019.

You’ll remember that I worked from the final 2015-16 BELOW to do my 2016 Final Five predictions.  I took this as the basis for these first-round predictions for my Python-based tool (coming this offseason to a Web browser near you).  I came out with the following after 1,000,000 trials of the tool:

Quarterfinal SeriesHome SweepHome in 3Road in 3Road Sweep
MTU v UAF44.09%27.70%15.65%12.56%
MSU v LSSU39.75%27.14%17.85%15.26%
BGSU v BSU31.30%25.29%21.62%21.79%
FSU v NMU28.18%24.30%22.89%24.63%

I want to highlight something: the chances that Northern and Bemidji would sweep their series are higher than merely winning in three games.  That may seem counter-intuitive, but it fits BELOW’s recency model.

As an example, let’s take Bowling Green and Bemidji State.

BG’s BELOW: 1625

BSU’s BELOW: 1556 (4th in the league despite them being the 6th seed)

Expected Value in the first game for BG to win: 64% (you can verify this if you like; I use a K of 40)

So BG would be expected to win 64% of the time, and if they win, they’d pick up 14 BELOW points, taking them to 1670 and dropping Bemidji to 1542.  In their next game, they’d be picked to win 68% of the time.

But if Bemidji wins game 1 of the series — which they did — BG loses 26 points to drop to 1630, while Bemidji jumps to 1582.  Now the Falcons have to win twice, but their expected winning percentage is just 58%.  Obviously they did go on and win on Saturday and Sunday, but this BELOW-based probability system gives the Beavers credit for winning that first night and narrowing the gap between the two teams.

My next step will be to  simulate the semifinal round, which will be interesting since my loops and conditional statements all have to take into account that the semifinals are re-seeded.  Because order and number of games matters for the BELOW calculations, accuracy demands that I treat each of these as individual trials: games within the series, then setting the new series, then setting the championship.

If you’re curious as to what the percentages for all 16 combinations are, here you go:

Screenshot 2016-05-18 09.25.46

You can see that it was pretty likely that the top two seeds would advance and a strong probability that the top three seeds would.  4/5 was a crap shoot, with the teams about 20 points apart in BELOW (which is pretty insignificant).  The result we got, the top seeds sweeping out, happened only one time in seven after a million trials of the quarterfinals.

The New WCHA Playoffs and WCHA Championship

The WCHA announced on Tuesday that they will move to a new on-campus playoff/tournament solution to determine the postseason champion.  The quarterfinals and semifinals will be best-of-three, hosted by the better seeds at each round.  The final game for the Broadmoor Trophy and the league’s automatic qualifier will be hosted by the highest remaining seed on the final weekend prior to NCAA play.

The turn from playoff to tournament at the end is a little interesting, but it makes sense.  A team playing best-of-three two weekends in a row could easily play seven postseason games in 15 days and be at 40+ games before the NCAAs roll around.  That’s a pretty tough task, especially considering that everyone will play games the final two regular season weekends, putting the toll at 11 games in the final month of competition.

That said, the single-game championship feels like a bit of a miss for the league.  I am aware that this comes because league members are keeping at 28 conference games, which means that the league’s regular season will always run to the second weekend in March.  WCHA Commissioner Bill Robertson said on Tuesday that the league will go to 26 or maybe even 24 games in 2017-18.  Keeping the travel down should a #5 or lower seed make the finale is a priority.

I have this idea in my head that I’ll run probabilities for a re-run of the 2016 postseason under the new format.  For example, what are the vanishingly small chances that Lake Superior would’ve hosted Alaska for a single-game championship?  I’d put that at 0.05% or less.  Similarly, what are the chances that Ferris State would’ve won a best-of-three with Michigan Tech and then won a one-game playoff at either Mankato or Bowling Green?  Remember that the first round went chalk this year, and Ferris State’s chances of winning the Broadmoor at the Final Five were just 12%.  It’s an interesting thought exercise and will be a good use of my in-progress tool.

I’m in favor of the changes on the whole.  The Final Five concept was dead, and this was an admission of that fact.  It’s an exciting new concept for a new WCHA, and I welcome the change.

2016 Offseason Conditioning Workouts, Post #1

I was digging through the site for something completely unrelated the other day when I came across Using KRACH, recency, and goal differential for next season and realized that I have, essentially, what I wanted then in BELOW now.  Let me talk about that a little.

  1. Using KRACH wasn’t an easy task.  I never did finish that math major, so I don’t have a great handle on building an algorithm that uses a Bradley-Terry method.  Before you glaze over on me: essentially, B-T methods note that not everyone plays everyone (or more appropriately that all network nodes have interactions with all other network nodes).  An Elo-based algorithm isn’t bound (as much) by that, although it certainly works better when all opponents have played each other.  A hockey-wide Elo algorithm may not make sense.
  2. An Elo-based algorithm has an eye for recency.  Alabama-Huntsville’s end-of-season results with Minnesota State and Bowling Green may have indicated that those two road teams were not as strong as BELOW would have thought them to be (or, alternately, that UAH wasn’t as bad as they appeared to be).  The top KRACH of the 2015-16 season was Minnesota State back in early November  (1744), but it’s Michigan Tech’s strong run to the end that shows the value of recency.  A “big win” back in the pack doesn’t help you much in the long run — ask Ferris State, which got above 1500 (where they never fell below again) with that 7-4 upset of the Mavs, but their more recent work in pulling into the home ice tussle was mediocre, going 1-3-0 down the stretch.  BELOW responded, but … single-game elimination is still a crap shoot, and I had Ferris State winning the Broadmoor 41.5% of the time on the morning of the final, a far sight better than the 12% overall shot I gave them to win the Final 4ive.
  3. I covered goal differential back in January, and my thoughts there still stand.  I’m going to look at modeling 2+ goal differentials by looking at empty-net goals (and all removing all of them), but I don’t think that we can ignore three-goal margins completely.  Say what you will about the 2-2 (OT) WCHA in 2015-16, but there were 42 WCHA games scored by three or more goals.

Essentially, BELOW is what I was looking for a year ago (and two years ago when I wrote this post).  I’m about 25% of the way to a prototype for making model runs.  I hope to have it available for the start of the season.

How the WCHA Play-In Alternative Might Have Worked

Back in 2014, I proposed a play-in tournament where Seeds 7-10 would travel to the top two seeds to play single-game match-ups to determine who plays the home team in a best-of-three series.  It’s admittedly a slightly-silly idea, given that the league’s struggling financially and Edina would be footing the bill.  But I want to revisit the topic anyway.

2014-15 would’ve been hard to pull off because of Alaska’s postseason ban.  Presumably Minnesota State would merely have hosted Lake Superior and Alaska-Anchorage, who finished in the bottom two slots. while Michigan Tech would’ve still had the joy of outlasting Carmine Guerriero in that 3OT thriller.

This season, it would’ve been especially painful financially, with the 8-9-10 seeds all being from the farthest-flung reaches of the league. Would UAF and UAA have played a fifth game in Anchorage with the winner getting on a plane for Minnesota immediately afterward?  Would the Huskies have taken the team from the 49th rather than a potential matchup against another hot Charger goaltender?  It would’ve been fun to find out.

2015-16 WCHA Final Five Pick: Minnesota State

A 2-1 Minnesota State win over Bowling Green, followed by a 1-0 upset of Michigan Tech by Ferris State leaves us with a finale of the Bulldogs (1614 BELOW) against the Mavericks (1673).  My pick: Minnesota State wins about 58.5%.

Now I know that’s not a huge upset by any means, and anyone can pick chalk, right?  But you might look at the WCHA standings and compare a team that went 16-5-7 with a team that went 13-11-4 and go, “Oh, that will be a cakewalk.”  BELOW likes the Bulldogs’ chances, though.  (Tech fans want Ferris to win.)

Why is this relevant?  (Why is any of this relevant?)  Any Elo-based rating system gives you more value for beating a good team than a bad one, and a bigger hit for losing to a bad team than a good one — all because of expectations.  So Ferris got a big jump — 36 points worth — because of Friday night’s win.

How big is that jump?  I have a handle for it because I look at the numbers all of the time, but let’s make it easy for you:

  1. It’s important to remember that the ranking system considers 1500 to be average, and if two teams with identical BELOW rankings play each other, the system predicts a 50-50 matchup.
  2. If a team that’s 1600 plays a team that’s 1500, the better team should win 64% of the time.  That’s also true of a matchup of teams at 1500-1400, 1700-1600, etc.
  3. If a team that’s 1700 plays a team that’s 1500, the better team should win 76% of the time.
  4. If a team that’s 1536 plays a team that’s 1500, the better team expects to win 55.2% of the time.

Note that the absolute values really don’t matter; it’s the relative ones that do.  Ferris’s win on Friday night convinces BELOW that it’s 5.2% better than the system thought that it was the night before.

With that in mind: the relative BELOW between these two teams is -59.  I decided to pull up every game with a BELOW differential between 50 and 75 for the entire 2015-16 regular season.  These were the games (scores in italics indicate lesser team won):

  1. Bowling Green (1534) at Alaska-Anchorage (1480): 2-0
  2. Northern Michigan (1564) at Alaska (1495): 2-5
  3. Bowling Green (1539) at Michigan Tech (1606): 0-2
  4. Alaska (1524) at Bemidji State (1457): 3-6
  5. Lake Superior (1408) at Bemidji State (1465): 1-1
  6. Bemidji State (1542) at Ferris State (1473): 1-2
  7. Michigan Tech (1565) at Lake Superior (1491): 6-2
  8. Bemidji State (1479) at Alaska (1418): 1-1
  9. Bemidji State (1512) at Alaska-Anchorage (1459): 5-1
  10. Bowling Green (1633) at Michigan Tech (1686): 1-5
  11. Northern Michigan (1529) at Ferris State (1601): 3-2
  12. Lake Superior (1441) at Alaska-Anchorage (1371): 3-2
  13. Lake Superior (1468) at Ferris State (1523): 1-3

In those 13 matches, five went the way of the weaker team by BELOW.

It’s not hard to see that matchups in this range happen often (around 10% of the time in this season), and that the better team isn’t shooting on an empty net.

For any curious: the biggest BELOW differential this season was 388, and there were five games over 300, all involving UAH, with the Chargers’ record in those games 1-2-2.

2016 WCHA Final Five BELOW-Based Predictions

So I have a preliminary model of multiple-goal wins, and I decided to try it out here on the 2016 WCHA Final Five.  My best estimate is:

  • Michigan Tech wins the Broadmoor Trophy and the automatic qualifier around 43% of the time, having dispatched Ferris State about 71% of the time in the semifinal.
  • Bowling Green and Minnesota State each win about 22.5% of the time.  The teams come into the semifinal a virtual tie in BELOW: 1649 for the Falcons, 1648 for the Mavericks; as such, BELOW considers it a 50-50 matchup.
  • Ferris State picks up about 12% of the wins.

Why the change from the final 2015-16 numbers for the Mavericks and Falcons?  Both teams went to a Sunday game, but Lake Superior was a far inferior opponent to Bemidji State, so the combination of two wins and a loss against each team meant that the Mavs fell farther down the lane.

My estimate above comes from multiple runs of 10,000 trials on an eyeball test.  I may push this to 100,000 trials or more if I have time before the puck drops.  If I do, I’ll write another update and maybe talk about my methodology a little bit.  If not, well, you have your probabilities.

[Note: I had a goof in my algorithm and was underestimating a Tech win.]

2015-16 BELOW: Beginning to End

I couldn’t leave well enough alone.  Here are the beginning and ending numbers, as well as season highs and lows.

Final 2015-16 BELOW

TeamStarting BELOWFinal BELOWChangeHighLow
0-UAH14001361-3914261319
1-UAA14161329-8715351329
2-UAF15231451-7215231387
3-BSU15341556+2215561469
4-BGSU15751625+5016731539
5-FSU14811527+4615711481
6-LSSU14031463+6014861354
7-MSU16881656-3217441656
8-MTU16541715+6417151570
9-NMU15071501-61591 1501

It’s important that I apply a regression to the mean for all teams.  These numbers will not be the starting values for next season.  Tech will be at 1643, as I move all teams 1/3 closer to the mean.  This also means that UAA will be at 1386

With that in mind, what we can see is that:

  1. Michigan Tech and Bowling Green reversed the regressions and kept their place at the top.
  2. Minnesota State was really off this year after being dominant last year.
  3. BELOW likes Bemidji State far better than I do.
  4. Ferris State and Northern Michigan stayed around the middle.
  5. Lake Superior made strides this year.  They’re almost to league average now, and they will start at 1474 next year.  Another year of recruits and coaching and they may be pushing for home ice next year.
  6. It was a terrible, terrible year for schools from states whose names start with “Ala”.  That said, Alaska has made some gains of late, and Michigan Tech will do well to not overlook them.

I will re-run BELOW from the start of 2013-14 forward, which will probably move some of these numbers a bit, say no more than 20-30 points either direction.

Enjoy the playoffs.  One more month until the sad time of six months of zero college hockey.