An awful lot of digital ink has been spilled on Patrick Roy's 2013-14 Colorado Avalanche this summer, a team that more or less defied woeful play at five-on-five by riding unsustainable shooting and save percentages. Largely because we have seen a model of this team before, many analysts are expecting some form of regression – the 2011-12 Minnesota Wild and 2012-13 Toronto Maple Leafs have provided ample case studies in the importance of getting the right side of possession. Perhaps more accurately, they have provided lessons on why teams must not rely on volatile percentages to rack up wins.
What makes this Colorado team interesting is two-fold. Firstly, they're teeming with young and developing talent, which could help stave off that regression to a degree. Secondly, we have only seen one year of real success from this club. The season before, Colorado played to a 67-point pace and finished dead last in the Western Conference.
Since we have data on teams dating back to 2007, it's not particularly difficult to investigate relationships between sets of data. Correlations of subsequent seasons can tell us what kind of adjustments to make, if any, when trying to forecast future output.
What I went ahead and did prior to this post was pull out Year 1 vs. Year 2 data for a variety of team-level even-strength numbers from 2007 to 2012 and dropped them in the table below. Repeatability is an r-squared number that tells us the percent of variance explained - the higher the r-squared number (up to 1.0), the more repeatable of a skill it is:
|EV Shooting Percentage||0.00|
|EV Save Percentage||0.13|
|EV Goal For%||0.19|
|EV Score-Adjusted Fenwick% (SAF%)||0.39|
You are reading the above correctly. A team's even-strength shooting percentage over one year tells us absolutely nothing about how that team will shoot the following year. Save percentage is slightly more telling than shooting percentage, but ultimately, it's a number you're going to want to heavily regress. As you go down the list, the correlations in data run tighter and the numbers don't need to be regressed as heavily. None of this bodes well for Colorado, a team that rode high percentages and carried terrible territorial control.
One other note on the above - you'll see that the r-squared between EV GoalFor% in the first year and EV GoalFor% in the subsequent year is 0.19. While EV GoalFor% is a better predictor of future EV GoalFor% than both EV Fenwick% and EV Corsi%, it is not a better predictor than EV Score-Adjusted Fenwick%.
That said, let's look at some comparables for the Colorado Avalanche - teams that picked up 90 or more points (my random cut-off line separating average teams from good ones) who also carried sub-par possession numbers at even-strength. We'll use equations generated for the year-one to year-two correlations to create an estimated number, and then compare it against the team's actual number.
First, let's do the percentages at even-strength:
|Y1 EVSH%||Est. Y2 EVSH%||Actual Y2 EVSH%|
It's almost stunning how identical the expected year two and actual year two percentages are on both ends of the rink. The takeaway from this is simple: one year of shooting percentage data tells us absolutely nothing, and regressing it all the way to the league average will give us a much better forecast of what's to come.
|Y1 EVSV%||Est. Y2 EVSV%||Actual Y2EVSV%|
The same can be said for save percentage data - taking our year one data and pulling it back 87 per cent to the league average gives us a more accurate guess as to what's to come.
Using that regression for forecasting purposes, expect Colorado to shoot around 7.89 per cent for next year at evens and stop around 92.47 per cent of the shots.
Now, let's break away from shooting and save percentages and look at possession rates. We know Score-Adjusted Fenwick% is the most repeatable of these metrics. Let's repeat the above exercise with the same Colorado comparables, and try to pindown where Colorado will finish at evens this season. I've included a fourth column in here to identify the change in points from Year 1 to Year 2.
|Y1 SAF%||Est. Y2 SAF%||Actual Y2 SAF%||Points Change|
You should first notice that regression seems less important with our possession numbers than the shooting/save percentages above. That’s because possession is a repeatable skill - or in this case, the lack of possession is a repeatable skill. Every team that can be considered a comparable for Colorado 2013-14 was out-shot in Year 1 and Year 2 - in most cases, decisively.
And, it’s impossible to ignore that column on the right, where every single percentage-good, possession-bad team of recent history saw a fall in the standings. The average fall for those nine teams was in the double digits, and the one team that didn’t take a massive hit - 2007 Pittsburgh - improved their possession numbers by almost three full percentage points.
Not only are those percentages running against the Avs, but they also go into next season missing their two best possession forwards from last season, with Paul Stastny signing in St. Louis and P.A. Parenteau traded to Montreal. Further, it's difficult to project improved possession numbers when the Avalanche brain trust doesn't seem inclined to dig into possession-based analytics.
This does not bode well for Patrick Roy’s team. It’s a virtual lock that their shooting and save percentages will climb down from their heights of last year, which means that their Goal% - last year, it was at 53.6 per cent - is in real trouble. The million dollar question is how far the Avs will fall - knocking them down by the average (-11.78) would likely still see them finish in the post-season, but their margin for error will be extremely tight this year.