Clarkson @ Cornell 03/10-3/12?/17

[upprdeck]

how do the metrics rate puck control and how do they even capture it? is someone reporting it for the teams?

does it measure control in one end versus the other?

the stats seems to point to the last 2 years being pretty close stats wise, yet in person this team was much better to the eye test

[RichH]

Thinking a bit more about this past weekend and how Cornell was fortunate to survive it, I'm starting to worry that Cornell has actually just been really lucky this entire season.  Our advanced stats are not good.  Meanwhile, the advanced stats of the other teams ranked highly in the PWR are almost universally strong--and in nearly every case better than ours.  We've had this debate enough on here, but now we have a full season's sample of games to work with, and Corsi etc. really do seem highly correlated with overall success.  Outside of the advanced stats, though, until this past weekend, we've looked pretty consistently good...

My theory on advanced stats is that they (at least, corsi/Fenwick) don't factor in shooting accuracy or goalie play.  On the NHL level, where everyone can shoot accurately and all goalies are at least very good, this is less of an issue.  But we've seen some Cornell teams that can't put the puck on net (oh, Cole bardreau, if only you could shoot) much at all.

A few legit snipers and a very good goalie make a big difference.

Then again, I haven't seen a single Cornell season without a good goalie (freshman scrivens / Davenport would probably be the low point), so it's hard to compare.

Yes, but shooting accuracy (something Cornell has never particularly excelled at) and goalie play (Gillam's save % is actually lower/the same as most  other top teams' goalies) can't account for the (pretty big) disparity between Cornell's advanced stats and those of basically all the other teams around them in the PWR.  That's not to say Gillam hasn't been very good--he probably saved us last weekend--but if we were truly limiting other teams' chances to shots from outside that Gillam can see, this would be reflected in a truly exceptional save %, instead of just a very good one.

Of course, Gillam might really be an average or above-average collegiate goaltender in terms of individual ability, in which case a very good save % would still be the result of limiting chances to low percentage shots. This doesn't change your point about the disparity in possession metrics but it does make a difference in how you view Cornell's chances of repeating these outcomes given a similar quality goaltender.

Personally, and as always based strictly on numbers, I suspect Cornell's record this season is on the lucky end - even if you allow for more systematic variance in advanced metrics in the college game such extreme outliers are always suspicious. How lucky has this season been, and how likely is Cornell to regress to the mean vs elevating its baseline, are the real questions. On the one hand, game to game Cornell has almost certainly been lucky to win as often as they have given the possession numbers. On the other hand, so many defensive injuries can really alter a team's play. It might very well be that with a healthy D the possession numbers would've been a lot better for this team, so their luck might've just balanced things out. No way to know, really, but hopefully we get to find out with a healthy team next year, and hopefully that team starts controlling possession to such an extent that they just win all of their games. Hey, I can hope, right?

Anyway, I got curious, so I looked up the advanced metrics for the last 3 years (all CHN has) - this year's team put up by far the highest PDO of the three - 103 even strength, 104.25 close. This is the strongest stats-based argument for the "this team was really lucky" position. For reference, last year's team - same goaltender and a lot of the same scorers - put up a 101.26 and 102 close PDO. Two years ago it was just below 98.4 even strength and 97.4 close, although of course the scoring personnel was quite different. Either this team has some kind of magical mix of NCAA-beating talent and execution in terms of scoring/save efficiency, or we should expect a PDO closer to 100 and a record more reflective of possession numbers next year. The smart money is on the latter - it's just hard to imagine a team that is so efficient when they have the puck and so effective at containing high quality scoring chances and yet does not control overall relative shot totals.

The strongest argument against "this team was really lucky" is this easily to understand metric we're overlooking (ECAC stats):

       2015-16      2016-17
GPG      2.09         3.18
GAPG     2.27         2.32

Yes, I agree, higher-possession teams are more likely to be good. But we increased our ECAC scoring by over a full GPG. Over the course of a 22 game schedule, that's more than luck. That's a young group of forwards developing and building on their skills and adapting to the college game more. Most of our top scorers increased their output by nearly 2x. We're getting the puck in the net better than we were able to last season. We went from 10th in the league in scoring to 3rd, while remaining at the top in team defense.

Most interesting in the stats I looked at is that we're far and away the highest scoring team in the 1st period, and near the bottom in the 3rd.

[scoop85]

We don't have to worry about Vigneault next year, as he signed with Columbus after this his Junior season

[Hooking]

Always lots of credit or blame ascribed to goalkeepers, offensive units, defensive units, individual players, even FANS (?!)- but NEVER a mention of good or bad coaching or game plans. Coaches do make a difference. I'll bet statistics prove this.

[marty]

Always lots of credit or blame ascribed to goalkeepers, offensive units, defensive units, individual players, even LUCKY SHIRTS (?!)- but NEVER a mention of good or bad coaching or game plans. Coaches do make a difference. I'll bet statistics prove this.

FYP::burnout::

[Tom Lento]

Thinking a bit more about this past weekend and how Cornell was fortunate to survive it, I'm starting to worry that Cornell has actually just been really lucky this entire season.  Our advanced stats are not good.  Meanwhile, the advanced stats of the other teams ranked highly in the PWR are almost universally strong--and in nearly every case better than ours.  We've had this debate enough on here, but now we have a full season's sample of games to work with, and Corsi etc. really do seem highly correlated with overall success.  Outside of the advanced stats, though, until this past weekend, we've looked pretty consistently good...

My theory on advanced stats is that they (at least, corsi/Fenwick) don't factor in shooting accuracy or goalie play.  On the NHL level, where everyone can shoot accurately and all goalies are at least very good, this is less of an issue.  But we've seen some Cornell teams that can't put the puck on net (oh, Cole bardreau, if only you could shoot) much at all.

A few legit snipers and a very good goalie make a big difference.

Then again, I haven't seen a single Cornell season without a good goalie (freshman scrivens / Davenport would probably be the low point), so it's hard to compare.

Yes, but shooting accuracy (something Cornell has never particularly excelled at) and goalie play (Gillam's save % is actually lower/the same as most  other top teams' goalies) can't account for the (pretty big) disparity between Cornell's advanced stats and those of basically all the other teams around them in the PWR.  That's not to say Gillam hasn't been very good--he probably saved us last weekend--but if we were truly limiting other teams' chances to shots from outside that Gillam can see, this would be reflected in a truly exceptional save %, instead of just a very good one.

Of course, Gillam might really be an average or above-average collegiate goaltender in terms of individual ability, in which case a very good save % would still be the result of limiting chances to low percentage shots. This doesn't change your point about the disparity in possession metrics but it does make a difference in how you view Cornell's chances of repeating these outcomes given a similar quality goaltender.

Personally, and as always based strictly on numbers, I suspect Cornell's record this season is on the lucky end - even if you allow for more systematic variance in advanced metrics in the college game such extreme outliers are always suspicious. How lucky has this season been, and how likely is Cornell to regress to the mean vs elevating its baseline, are the real questions. On the one hand, game to game Cornell has almost certainly been lucky to win as often as they have given the possession numbers. On the other hand, so many defensive injuries can really alter a team's play. It might very well be that with a healthy D the possession numbers would've been a lot better for this team, so their luck might've just balanced things out. No way to know, really, but hopefully we get to find out with a healthy team next year, and hopefully that team starts controlling possession to such an extent that they just win all of their games. Hey, I can hope, right?

Anyway, I got curious, so I looked up the advanced metrics for the last 3 years (all CHN has) - this year's team put up by far the highest PDO of the three - 103 even strength, 104.25 close. This is the strongest stats-based argument for the "this team was really lucky" position. For reference, last year's team - same goaltender and a lot of the same scorers - put up a 101.26 and 102 close PDO. Two years ago it was just below 98.4 even strength and 97.4 close, although of course the scoring personnel was quite different. Either this team has some kind of magical mix of NCAA-beating talent and execution in terms of scoring/save efficiency, or we should expect a PDO closer to 100 and a record more reflective of possession numbers next year. The smart money is on the latter - it's just hard to imagine a team that is so efficient when they have the puck and so effective at containing high quality scoring chances and yet does not control overall relative shot totals.

The strongest argument against "this team was really lucky" is this easily to understand metric we're overlooking (ECAC stats):

       2015-16      2016-17
GPG      2.09         3.18
GAPG     2.27         2.32

Yes, I agree, higher-possession teams are more likely to be good. But we increased our ECAC scoring by over a full GPG. Over the course of a 22 game schedule, that's more than luck. That's a young group of forwards developing and building on their skills and adapting to the college game more. Most of our top scorers increased their output by nearly 2x. We're getting the puck in the net better than we were able to last season. We went from 10th in the league in scoring to 3rd, while remaining at the top in team defense.

Most interesting in the stats I looked at is that we're far and away the highest scoring team in the 1st period, and near the bottom in the 3rd.

I'm not saying it's all luck, I'm saying I'd expect GFA to come down next year unless the possession metrics improve. The tl,dr; of the below is that goals for is an inherently high variance metric just because of the relative rarity of goal scoring events, and therefore I don't expect it to be all that useful as a predictor of future success.

Let's look at the full season, just for the sake of completeness:

       2015-16      2016-17
GPG      2.32         2.97
GAPG     2.41         2.18

One of the problems with the current crop of metrics in hockey is they require huge game samples to really be meaningful. That doesn't mean the metrics are not valuable, it just means you have to adjust your expectations for how much you can assert about a team's performance based on any of this, particularly in the college game. Team goal scoring across seasons is especially bad as a predictive metric since goals are relatively rare events.

Your 22 game goals for/goals against sample is, of course, meaningful as a description of past results, and I'm not going to deny that Cornell has had a marvelous season - I've thoroughly enjoyed following it on paper. However, the question was about whether or not Cornell can sustain this win rate, and I don't think looking at GFA tells us all that much given the sample size. Think about it - the difference between last season and this amounts to 22 goals over 34 games. 22 positive outcomes which are affected by a large number of factors, some of which are effectively random. Consider - this season, Cornell had 150 PP chances in 33 games compared with 97 in 34 games last season, and picked up 9 of those 22 goals on the PP despite a slightly lower conversion rate. Incidentally, that is another potentially interesting stat. Is Cornell drawing more penalties due to systemic improvements (i.e., forcing teams to take penalties or generally wearing them down such that they're behind the play) or is this just a lucky fluke? I have no idea.

Anyhow, here's a little sample size/variance illustration with an advanced metric. Consider PDO, which is save % plus shooting %. It tends to be 100 on a team level in the NHL, and although I haven't done real diligence, in my spot checking it seems to follow a similar trend in the college game. 22 games is about 1/4 of an NHL season. Why is that important? Because of the first table in this article: http://www.pensburgh.com/2010/6/23/1531707/pdo-and-what-it-means

You can see pretty extreme outlier behavior over an approximately 20 game sample, even in a league where the team level talent distribution is likely to be narrower than the college game. Those outliers tend to pull back towards the mean over the course of a larger sample of subsequent games. But - and this is important - PDO is measured across a much, much larger number of events than goals scored. The variance is naturally smaller, and  yet you still see a lot of team-level variation over fairly substantial samples of games played.

Back to Cornell and goals scored this year: PDO + corsi doesn't tell you everything, because PDO is based on shots on goal and corsi is based on shot attempts. In Cornell's case, that's important - from eyeballing the shot metrics on CHN, it looks like Cornell's goal scoring increase over last year comes from both more shots on goal (927 of 1710 attempts vs 890 of 1762 attempts) and a higher shooting % (10.5 vs 8.9).

Based on what we know about shooting percentage (via PDO, which in Cornell's case amounts to the same thing - team save % has been just about .920 for each of the last three years) we should expect the latter to come down. Doesn't mean it will - maybe the current crop of forwards have a preternatural ability to convert shots into goals, or maybe they're just a lot better than most at finding the spaces where high shooting percentages happen - but most likely that will regress to the mean.

The shots on goal rate, that I don't know anything about, and that, to me, is the most interesting thing I'm seeing in Cornell's scoring increase. Is shots on goal as a fraction of total shot attempts something a team can reliably control, or is that also subject to a lot of random variance? Anybody know? Intuitively one would think it's more of a controllable thing than shot conversion, but I don't know.

The first vs third period effect you mention is also somewhat interesting. One possible explanation is fatigue, but another could be game score effects. If Cornell often carries a multi-goal lead into the third period we should expect fewer shots for and more shots against with, on average, a corresponding shift in goal scoring. I don't know if this is true because a) I don't care to do that tallying and b) I don't know what constitutes "often" anyway. One simple thing you could do to dig in is look at GF/GA in the third period conditioned on game state - if Cornell tends to score more when losing entering the third and less when leading, that'd be a good indication of strategic shifts rather than fatigue as the explanation for the goal scoring shift.

Disclaimer - I am not an advanced hockey stats expert and I'm actually not fully convinced by all of them - particularly PDO, which is just as ad hoc and nutty as OPS in baseball, although the latter has turned out to be very robust and the analyses I've seen of the former are promising. That said, they've been pretty useful to me in understanding how the team has been playing in the absence of the ability to actually watch games. I do know a bit about measuring probabilistic outcomes, so that's where most of my assertions around metric quality come from, but those assertions are based on general knowledge rather than domain specific expertise.