Sep 242014
 

Today apparently there was some discussion about the Avalanche and their non-interest in hockey analytics. In that discussion Corey Pronman wrote the following tweet:

 

I have seen the above logic from time to time. I think it dates back to something Gabe Desjardins wrote many years ago. I find the logic very odd though. Let me explain.

Let’s assume that the numbers are true. According to my math, that leaves 25% unaccounted for. I don’t really consider 25% insignificant but it is actually more significant than that.

Luck, or I prefer the term randomness, is a component that is outside the control of a general manager, a coach, a player or anyone else that could potentially influence the outcome of the game. Thus it is pointless to bring luck into the equation.  All that management and players for an NHL team really needs to worry about is what they can control. That is the non-luck fraction of winning or the other 60%.

Now, if Corsi is 35% of winning overall then it accounts for 58% of the controllable aspect of winning. That leaves 42% of what is controllable unaccounted for. If I were an owner of an NHL team, or an owner of a business of any kind, and my general manager told me that we are going to largely ignore 42% of  the controllable factors that lead to positive outcomes I’d be firing that general manager on the spot. It simply isn’t acceptable business practice to ignore 42% of what is within your control that produces good outcomes.

Here is the the real kicker though. The estimate that Corsi explains 35% of wins is based on historical data (and probably from several years ago). It does not necessarily mean it will be that way in the future. As teams become more aware of Corsi and possession it is certainly conceivable that the disparity across teams in corsi shrinks and thus the importance of Corsi as a differentiator among teams and as predictor of winning shrinks. If teams switch focus to Corsi those other factors might be the great differentiator of team talent and be the better predictor of success. It is easy to hop on the Corsi bandwagon now. The forward thinking teams and forward thinking hockey analytics researchers are those researching that other 42% to some significant degree.

Now, if you are a hockey analytics researcher raise your hand if you have spent ~60% of your research and analysis time on Corsi related issues and ~40% of your research time on non-Corsi related issues. If you are honest I suspect very few of you have raised your hand. The honest truth is those other factors have been unfairly downplayed and in my opinion that is very unfortunate.

 

Aug 262014
 

I am sure many of you are aware that Corey Sznajder (@ShutdownLine) has been working on tracking zone entries and exits for every game from last season. A week and a half ago Corey was nice enough to send me the data for every team for all the games he had tracked so far (I’d estimate approximately 60% of the season) and the past few days I have been looking at it. So, ultimately everything you read from here on is thanks to the time and effort Corey has put in tracking this data.

As I have alluded to on twitter, I have found some interesting and potentially very significant findings but before I get to that let me summarize a bit of what is being tracked with respect to zone entries.

  • CarryIn% – Is the percentage of time the team carried the puck over the blue line into the offensive zone.
  • FailedCarryIn% – Is the percentage of the time the team failed to carry the puck over the blue line into the offensive zone.
  • DumpIn% – is the percentage of the time the team dumped the puck into the offensive zone.

The three of these should sum up to 100% (Corey’s original data treated FailedCarryIn% separately so I made this adjustment) and represent the three different outcomes if a team is attempting to enter the offensive zone – successful carry in, failed carry in, and dumped in.

I gathered all this information for and against for every team and put them in a table. I’ll spare you all the details as to how I arrived at this idea I had but here is what I essentially came up with:

  • Treat successful carry ins as a positive
  • Treat failed carry in attempts as a negative (probably results in a quality counter attack against)
  • Dump ins are considered neutral (ignored)

So, I then came up with NetCarryIn% which is CarryIn% – FailedCarryIn% and I calculated this for each team for and against to get NetCarryIn%For and NetCarryIn%Against for each team.

I then subtracted NetCarryIn%Against from NetCarryIn%For to get NetCarryIn%Diff.

In all one formula we have:

NetCarryIn%Diff = (CarryIn%For – FailedCarryIn%For) – (CarryIn%Against – FailedCarryIn%Against)

Hopefully I haven’t lost you. So, with that we now get the following results.

Team Playoffs? NetCarryIn%Diff RegWin%
Chicago Playoffs 12.2% 61.0%
Tampa Playoffs 6.1% 53.0%
Anaheim Playoffs 5.9% 64.6%
Colorado Playoffs 5.5% 59.1%
Detroit Playoffs 4.7% 51.2%
Minnesota Playoffs 4.1% 53.0%
Pittsburgh Playoffs 4.0% 59.8%
Dallas Playoffs 3.8% 51.8%
New Jersey . 3.4% 48.2%
Los Angeles Playoffs 1.7% 53.7%
Boston Playoffs 1.3% 67.1%
St. Louis Playoffs 1.2% 60.4%
Ottawa . 0.9% 47.6%
Columbus Playoffs 0.7% 51.8%
Edmonton . 0.7% 35.4%
NY Rangers Playoffs -0.1% 54.9%
Phoenix . -1.3% 48.8%
Montreal Playoffs -1.3% 53.0%
Vancouver . -1.7% 43.9%
Philadelphia Playoffs -1.8% 53.0%
Winnipeg . -1.8% 43.3%
San Jose Playoffs -2.3% 59.1%
NY Islanders . -3.0% 40.2%
Toronto . -4.8% 42.7%
Nashville . -6.0% 50.6%
Calgary . -6.4% 38.4%
Washington . -6.4% 46.3%
Florida . -6.7% 35.4%
Carolina . -6.8% 47.0%
Buffalo . -7.7% 25.6%

‘Playoffs’ indicates a playoff team and RegWin% is their regulation winning percentage (based on W-L-T after regulation time).

What is so amazing about this is we have taken the first ~60% of games and done an excellent job of predicting who will make the playoffs. The top 8 teams (and 11 of top 12) in this stat through 60% of games made the playoffs and all of  the bottom 8 missed the playoffs. That’s pretty impressive as a predictor. What’s more, the r^2 with RegWin% is a solid 0.42, significantly better than the r^2 with 5v5 CF% which is 0.31. Here are what the scatter plots look like.

CarryInPctDiff_vs_RegWinPct

CFPctDiff_vs_RegWinPct

I think what we are seeing is that if you are more successful at carrying the puck into the offensive zone, but not at the expense of costly turnovers attempting those carry ins, than your opponent you will win the neutral zone and that goes a long way towards winning the game. Recall that I have shown that shots on the rush are of higher quality than shots generated from zone play so an important key to winning is maximizing your shots on the rush and minimizing your opponents shots on the rush. To an extent this may in fact actually be measuring some level of shot quality.

Of course, why stop here. If it is in fact some sort of measure of shot quality, why not combine it with shot quantity? To do this I took NetCarryIn%Diff and add to it the teams Corsi% – 50%. This is what we get.

Team Playoffs? NetCarryIn%Diff – CF% over 50%
Chicago Playoffs 17.7%
Los Angeles Playoffs 8.5%
New Jersey . 7.8%
Tampa Playoffs 7.1%
Detroit Playoffs 6.2%
Anaheim Playoffs 5.7%
Boston Playoffs 5.2%
St. Louis Playoffs 4.3%
Dallas Playoffs 4.3%
Ottawa . 3.3%
Minnesota Playoffs 2.7%
Pittsburgh Playoffs 2.7%
Colorado Playoffs 2.5%
NY Rangers Playoffs 2.3%
San Jose Playoffs 1.4%
Columbus Playoffs 0.6%
Vancouver . -0.4%
Phoenix . -0.8%
Winnipeg . -1.7%
Philadelphia Playoffs -1.8%
NY Islanders . -3.6%
Montreal Playoffs -4.6%
Edmonton . -5.0%
Florida . -5.7%
Carolina . -6.5%
Nashville . -7.5%
Washington . -8.7%
Calgary . -10.1%
Toronto . -11.9%
Buffalo . -14.7%

New Jersey still messes things up but New Jersey is just a strange team when it comes to these stats. But think about this. If New Jersey and Ottawa made the playoffs over Philadelphia and Montreal it would have a perfect record in predicting the playoff teams. It was perfect in the western conference.

Compared to Regulation Win Percentage we get:

CarryInPctDiff_CFPctDiff_vs_RegWinPct

That’s a pretty nice correlation and far better than corsi% itself.

Now, this could all be one massive fluke and none of this is repeatable but I am highly doubtful that will be the case. We may be on to something there. Will be interesting to see what individual players look like with this stat and I’ll also take a look at whether zone exits should somehow get factored in to this equation. I suspect it may not be necessary as it may be measuring something similar to Corsi% (shot quantity over quality).

 

Jun 122014
 

The rumour is out there that Sunny Mehta has been hired as Director of Hockey Analytics of the New Jersey Devils (if true, a big congrats to Sunny). This sparked some twitter discussion about the Devils and analytics and Devils defensemen including Bryce Salvador.

I have been a bit of a fan of Salvador, at least statistically, though clearly there are a lot of Devils fans that do not like him and I think it is because of a focus on corsi. One person tweeted me an image of Salvador’s corsi rel % suggesting it was “pretty ugly”. While maybe true the game isn’t about Corsi it is about goals. Here is what I know about Salvador. In 5v5close situations he led the Devils defensemen in on-ice save percentage last season, the season before, and the season before that. He missed 2010-11 due to injury but in 2009-10 he was second best trailing only Andy Greene, his regular defense partner. Either he is extremely lucky (every year) or he is doing something right.

Lets look at this a different way. Over the past 3 seasons Bryce Salvador has had the third best 5v5close save percentage in the league when he is on the ice despite the Devils ranking 23rd in team save percentage. The two players ahead of him play for Boston (Dougie Hamilton) and Los Angeles (Willie Mitchell) who have significantly better goaltending (3rd and 8th best 5v5close save percentages over past 3 seasons) and again, they played in front of far better goaltending.

In February 2012 I wrote an article attempting to quantify a defenders effect on save percentage and in it I identified Salvador as one of the best defensemen at boosting his teams save percentage. In the 2 seasons since he has done nothing but support that claim.

So, what does this all mean? Well, it takes a player who had a team worst 15.9 CA/20 in 5v5close situations this past season to a team best 0.49 GA/20.  Over the past 3 seasons only Dougie Hamilton (Boston), Willie Mitchell (Los Angeles) and Alec Martinez (Los Angeles) have seen goals scored against them at a lower rate than Bryce Salvador.

I know the majority of people are on the corsi bandwagon these days and some will dismiss any argument that runs counter to it but I think the evidence is clearly on Salvador’s side here. All evidence suggest he is really good as suppressing opposition shot quality and in turn suppressing the number of goals scored against the Devils. If I were the new Director of Hockey Analytics for the Devils I wouldn’t be recommending getting rid of Salvador.

 

May 112014
 

I often feel that I am the sole defender of goal based hockey analyitics in a world dominated by shot attempt (corsi) based analytics. In recent weeks I have often heard the pro-corsi crowd cite example after example of where corsi-based analytics “got it right” or “predicted something fairly well”. While it is always good to be able to cite examples where you got things right a fair an honest evaluation looks at the complete picture, not just the good outcomes. Otherwise it is analytics by anecdotes which is an oxymoron if there every was one.

For example, Kent Wilson of FlamesNation.ca recent wrote about the “Dawning of the Age of Fancy Stats” in which he cited several instances of where hockey analytics got it right or did well in predicting outcomes.

The big test case which seems to have moved the needle in favour of the nerds is, of course, the Toronto Maple Leafs. Toronto came into the season with inflated expectations after an outburst of percentages during the lock-out shortened year saw them break into the post-season. Their awful underlying numbers caused the stats oriented amongst us to be far more circumspect about their chances, of course.

Toronto is the recent example that the hockey analytics crowd likes to bring up in support of their case but it is just one example. We don’t hear much about how many predicted the Ottawa Senators to be in the playoffs and some even had them challenging for the top spot in the eastern conference. We don’t hear much about how the New Jersey Devils missed the playoffs yet again despite having the 5th best 5v5close Fenwick% in the league, the year after missing the playoffs with the 3rd best 5v5close Fenwick% in the league. If we are truly interested in hockey analytics we need a complete and unbiased assessment of all outcomes, not just the ones that support our underlying belief.

In the same article Kent Wilson quoted a tweet from Dimitri Filipovic about the success of Corsi in predicting outcomes of playoff series.

Relevant #fact: since ’08 playoffs, teams that were 5+ % better than their opponent in 5v5 fenwick close during the regular season are 25-7.

While interesting, what it really doesn’t tell us a whole lot more than “when one team is significantly better at outshooting their opponents they more often than not win”. Well, that really isn’t saying a whole lot. It is more or less saying, when a dominant team plays a mediocre team, the dominant team usually wins. Not really that interesting when you think of it that way.

Here is another fact that puts that into perspective. Since the 2008 playoffs, the team with the better 5v5close Fenwick% has a 53-35-2 record (there were 2 cases where teams had identical fenwick% to 1 decimal place). That actually makes it sound like 5v5close Fenwick% is predictive overall, not just in cases where one team is significantly better than another. Of course, if we look at goals we find that the team with the better 5v5close goal% has a 54-34-1 record. In other words, 5v5close possession stats did no better at predicting playoff outcomes than 5v5close goal stats. It is easy to throw out stats that support a point of view, but it is far more important to look at the complete picture. That is what analytics is about.

A similar statistic was promoted by Michael Parkatti in a recent talk on hockey analytics at the University of Alberta. In that talk Parkatti stated that of the last 15 Stanley Cup winners all but 3 had a “ShotShare” (all situations) of at least 53%. The exceptions were Pittsburgh in 2009, Boston in 2011 and Carolina in 2006. I will note that it appears that all three of these teams are below 51% and 2009 Penguins were below 50%. That seems sort of impressive but I did some digging myself and found that every Stanley Cup winner since 1980 had a “GoalShare” (all situations) greater than 52%. Every single one. No exceptions. I didn’t look at any cup winners pre-1980 but the trend may very well go back a lot further. As impressive as 12 of 15 is, 34 of 34 is far more impressive.

Here is the thing. We know that goal percentage correlates with winning far better than corsi percentage. This is an indisputable fact. It is actually quite a bit better. The sole reason we use corsi is that goals are infrequent events and thus not necessarily indicative of true talent due to small sample size issues. This is a fair argument and one that I accept. In situations where you have small sample sizes definitely use corsi as your predictive metric (but understand its limitations). The question that needs to be answered is what constitutes a small sample size and more importantly what sample size do we need such that goals become as good or better of a predictor of future events than corsi. I have pegged this crossing point at about 1 seasons worth of data, maybe a bit more if looking at individual players who may not be getting 20 minutes of ice time a game (my guess is around >750 minutes of ice time is where I’d start to get more comfortable using goal data than corsi data). I am certain not everyone agrees but I haven’t see a lot of analyses attempting to find this “crossing point”.

Let’s take another look at how well 5v5close Fenwick% and Goal% predict playoff outcomes again but lets look by season rather than overall.

FF% GF%
2008 7-7-1 6-9
2009 9-6 11-4
2010 9-6 11-4
2011 10-5 11-4
2012 7-7-1 7-7-1
2013 11-4 8-7
Total 53-35-2 54-35-1

In full seasons not affected by lockouts we find that GF% was generally the better predictor (only 2008 did GF% under perform FF%) but in last years lockout shortened season FF% significantly outperformed GF%. Was this a coincidence or is it evidence that 48 games is not a large enough sample size to rely on GF% more than CF% but 82 games probably is?

I have seen numerous other examples in recent weeks where “analytics” supporters have used what amounts to not much more than anecdotal evidence to support their claims. This is not analytics. Analytics is a fair, unbiased and complete fact based assessment of reality. Showing why a technique is a good predictor some of the time is not enough. You need to show why it is overall a better predictor all of the time or at least define when it is and when it isn’t.

I recently wrote an article on whether last years statistics predicted this years playoff teams and found that GF% seemed to do at least as well as CF% despite last season being a lock-out shortened year.

With all that said, you will frequently find me using “possession” statistics so I certainly don’t think they are useless. It is just my opinion that puck possession is just one aspect of the game and puck possession analytics has largely been oversold when it comes to how useful it is as a predictor. Conversely goal based analytics has been largely given a bad rap which I find a little unfortunate.

(Another article worth reading is Matt Rudnitsky’s MONEYPUCK: Why Most People Need To Shut Up About ‘Advanced Stats’ In The NHL.)

 

Apr 012014
 

Last week Tyler Dellow had a post titled “Two Graphs and 480 Words That Will Convince You On Corsi%” in which, you can say, I was less than convinced (read the comments). This post is my rebuttal that will attempt to convince you on the importance of Sh% in player evaluation.

The problem with shooting percentage is that it suffers from small sample size issues. Over small sample sizes it often gets dominated by randomness (I prefer the term randomness to luck) but the question I have always had is, if we remove randomness from the equation, how important of a skill is shooting percentage? To attempt to answer this I will look at the variance in on-ice shooting percentages among forwards as we increase the sample size from a single season (minimum 500 minutes ice time) to 6 seasons (minimum 3000 minutes ice time). As the sample size increases we would expect the variance due to randomness to decrease. This means, when the observed variance stops decreasing (or significantly slows the rate of decrease) as sample size increases we know we are approaching the point where any variance is actually variance in true talent and not small sample size randomness. So, without going on any further I present you my first chart of on-ice shooting percentages for forwards in 5v5 situations.

 

ShPctVarianceBySampleSize

Variance decline pretty much stops by the time you reach 5 years/2500+ minutes worth of data but after 3 years (1500+ minutes) the drop off rate falls off significantly. It is also worth noting that some of the drop off over longer periods of time is due to age progression/regression and not due to reduction in randomness.

What is the significance of all of this?  Well, at 5 years a 90th percentile player would have 45% more goals given an equal number of shots as a 10th percentile player. A player one standard deviation above average will have 33% more goals for given an equal number of shots as a player one standard deviation below average.

Now, let’s compare this to the same chart for CF/20 to get an idea of how shot generation varies across players.

 

CF20VarianceBySampleSize

It’s a little interesting that the top players show no regression over time but the bottom line players do. This may be because terrible shot generating players don’t stick around long enough. More importantly though is the magnitude of the difference between the top players and the bottom players.  Well, a 90th percentile CF20 player produces about 25% more shots attempts than a 10th percentile player and a one standard deviation above average CF20 player produces about 18.5% more than a one standard deviation below average CF20 player (over 5 years). Both of these are well below (almost half of) the 45% and 33% we saw for shooting percentage.

I hear a lot of ‘I told you so’ from the pro-corsi crowd in regards to the Leafs and their losing streak and yes, their percentages have regress this season but I think it is worth noting that the Leafs are still an example of a team where CF% is not a good indicator of performance. The Leafs 5v5close CF% is 42.5% but their 5v5close GF% is 47.6%. The idea that CF% and GF% are “tightly intertwined” as Tyler Dellow wrote is not supported by the Maple Leafs this season despite the fact that the Maple Leafs are the latest “pro-Corsi” crowds favourite “I told you so” team.

There is also some evidence that the Leafs have been “unlucky” this year. Their 5v5close shooting percentages over the past 3 seasons have been 8.82 (2nd), 8.59(4th), 10.54(1st) while this year it has dropped to 8.17 (8th). Now the question is how much of that is luck and how much is the loss of Grabovski and MacArthur and the addition of Clarkson (who is a generally poor on-ice Sh% player) but the Leafs Sh% is well below the past few seasons and some of that may be bad luck (and notably, not “regression” from years of “good luck”).

In summary, generating shots matter, but capitalizing on them matters as much or more.

 

Feb 092014
 

There is a recently posted article on BroadStreetHockey.com discussing overused and overrated statistics. The first statistic on that list is Plus/Minus. Plus/minus has its flaws and gets wildly misused at times but it doesn’t mean it is a useless statistics if used correctly so I want to defend it a little but also put it in the same context as corsi.

The rational given in the BroadStreetHockey.com article for plus/minus being a bad statisitcs is that the top of the plus/minus listing is dominated by a few teams. They list the top 10 players in +/- this season and conclude:

Now there are some good players on the list for sure, but look a little bit closer at the names on the list. The top-ten players come from a total of five teams. The top eight all come from three teams. Could it perhaps be more likely that plus/minus is more of a reflection of a team’s success than specific individuals?

Now that is a fair comment but let me present you the following table of CF% leaders as of a few days ago.

Player Name Team CF%
MUZZIN, JAKE Los_Angeles 0.614
WILLIAMS, JUSTIN Los_Angeles 0.611
KOPITAR, ANZE Los_Angeles 0.611
ERIKSSON, LOUI Boston 0.606
BERGERON, PATRICE Boston 0.605
TOFFOLI, TYLER Los_Angeles 0.595
TOEWS, JONATHAN Chicago 0.592
THORNTON, JOE San_Jose 0.591
MARCHAND, BRAD Boston 0.591
ROZSIVAL, MICHAL Chicago 0.590
TARASENKO, VLADIMIR St.Louis 0.589
KING, DWIGHT Los_Angeles 0.589
BROWN, DUSTIN Los_Angeles 0.586
DOUGHTY, DREW Los_Angeles 0.584
BURNS, BRENT San_Jose 0.583
BICKELL, BRYAN Chicago 0.582
HOSSA, MARIAN Chicago 0.581
KOIVU, MIKKO Minnesota 0.580
SAAD, BRANDON Chicago 0.579
SHARP, PATRICK Chicago 0.578
SHAW, ANDREW Chicago 0.578
SEABROOK, BRENT Chicago 0.576

Of the top 22 players, 8 are from Chicago and 7 are from Los Angeles. Do the Blackhawks and Kings have 68% of the top 22 players in the NHL? If we are tossing +/- aside because it is “more of a reflection of a team’s success than specific individuals” then we should be tossing aside Corsi as well, shouldn’t we?

The problem is not that the top of the +/- list is dominated by a few teams it is that people misinterpret what it means and don’t consider the context surrounding a players +/-. No matter what statistic we use we must consider context such as quality of team, ice time, etc. Plus/minus is  no different in that regard.

There are legitimate criticisms of +/- that are unique to +/- but in general I think a lot of the criticisms and subsequent dismissals of +/- having any value whatsoever are largely unfounded. It isn’t that plus/minus is over rated or over used it is that it is often misued and misinterpreted and to be honest I see this happen just as much with Corsi and the majority of other “advanced” statistics as well. It isn’t the statistic that is the problem, it is the user of the statistic. That, unfortunately, will never change but that shouldn’t stop us who know how to use these statistics properly from using them to advance our knowledge of hockey. So please, can we stop dismissing plus/minus (and other stats) as a valueless statistics just because a bunch of people frequently misuse it.

The truth is there are zero (yes, zero) statistics in hockey that can’t and aren’t regularly misused and used without contextualizing. That goes from everything from goals and point totals to corsi to whatever zone start or quality of competition metric you like. They are all prone to be misused and misinterpreted and more often than not are. It is not because the statistics themselves are inherently flawed or useless its because hockey analytics is hard and we are a long long way from fully understanding all the dynamics at play. Some people are just more willing to dig deeper than others. That will never change.

 

(Note: This isn’t intended to be a critique of the Broad Street Hockey article because the gist of the article is true. The premise of the article is really about statistics needing context and I agree with this 100%. I just wish it wasn’t limited to stats like plus/minus, turnovers, blocked shots, etc. because advanced statistics are just as likely to be misused.)

 

Apr 112013
 

Every now and again someone asks me how I calculate HARO, HARD and HART ratings that you can find on stats.hockeyanalysis.com and it is at that point I realize that I don’t have an up to date description of how they are calculated so today I endeavor to write one.

First, let me define HARO, HARD and HART.

HARO – Hockey Analysis Rating Offense
HARD – Hockey Analysis Rating Defense
HART – Hockey Analysis Rating Total

So my goal when creating then was to create an offensive defensive and overall total rating for each and every player. Now, here is a step by step guide as to how they are calculated.

Calculate WOWY’s and AYNAY’s

The first step is to calculate WOWY’s (With Or Without You) and AYNAY’s (Against You or Not Against You). You can find goal and corsi WOWY’s and AYNAY’s on stats.hockeyanalysis.com for every player for 5v5, 5v5 ZS adjusted and 5v5 close zone start adjusted situations but I calculate them for every situation you see on stats.hockeyanalysis.com and for shots and fenwick as well but they don’t get posted because it amounts to a massive amounts of data.

(Distraction: 800 players playing against 800 other players means 640,000 data points for each TOI, GF20, GA20, SF20, SA20, FF20, FA20, CF20, CA20 when players are playing against each other and separate of each other per season and situation, or about 17.28 million data points for AYNAY’s for a single season per situation. Now consider when I do my 5 year ratings there are more like 1600 players generating more than 60 million datapoints.)

Calculate TMGF20, TMGA20, OppGF20, OppGA20

What we need the WOWY’s for is to calculate TMGF20 (a TOI with weighted average GF20 of the players teammates when his team mates are not playing with him), TMGA20 (a TOI with weighted average GA20 of the players teammates when his team mates are not playing with him), OppGF20 (a TOI against weighted average GF20 of the players opponents when his opponents are not playing against him) and OppGA20 (a TOI against weighted average GA20 of the players opponents when his opponents are not playing against him).

So, let’s take a look at Alexander Steen’s 5v5 WOWY’s for 2011-12 to look at how TMGF20 is calculated. The columns we are interested in are the Teammate when apart TOI and GF20 columns which I will call TWA_TOI and TWA_GF20. TMGF20 is simply a TWA_TOI (teammate while apart time on ice) weighted average of TWA_GF20. This gives us a good indication of how Steen’s teammates perform offensively when they are not playing with Steen.

TMGA20 is calculated the same way but using TWA_GA20 instead of TWA_GF20. OppGF20 is calculated in a similar manner except using OWA_GF20 (Opponent while apart GF20) and OWA_TOI while OppGA20 uses OWA_GA20.

The reason why I use while not playing with/against data is because I don’t want to have the talent level of the player we are evaluating influencing his own QoT and QoC metrics (which is essentially what TMGF20, TMGA20, OppGF20, OppGA20 are).

Calculate first iteration of HARO and HARD

The first iteration of HARO and HARD are simple. I first calculate an estimated GF20 and an estimated GA20 based on the players teammates and opposition.

ExpGF20 = (TMGF20 + OppGA20)/2
ExpGA20 = (TMGA20 + OppGF20)/2

Then I calculate HARO and HARD as a percentage improvement:

HARO(1st iteration) = 100*(GF20-ExpGF20) / ExpGF20
HARD(1st iteration) = 100*(ExpGA20 – GA20) / ExpGA20

So, a HARO of 20 would mean that when the player is on the goal rate of his team is 20% higher than one would expect based on how his teammates and opponents performed during time when the player is not on the ice with/against them. Similarly, a HARD of 20 would mean the goals against rate of his team is 20% better (lower) than expected.

(Note: The OppGA20 that gets used is from the complimentary situation. For 5v5 this means the opposition situation is also 5v5 but when calculating a rating for 5v5 leading the opposition situation is 5v5 trailing so OppGF20 would be OppGF20 calculated from 5v5 trailing data).

Now for a second iteration

The first iteration used GF20 and GA20 stats which is a good start but after the first iteration we have teammate and opponent corrected evaluations of every player which means we have better data about the quality of teammates and opponents the player has. This is where things get a little more complicated because I need to calculate a QoT and QoC metric based on the first iteration HARO and HARD values and then I need to convert that into a GF20 and GA20 equivalent number so I can compare the players GF20 and GA20 to.

To do this I calculate a TMHARO rating which is a TWA_TOI weighted average of first iteration HARO. TMHARD and OppHARO and OppHARD are calculated in a similar manner. TMHARD, OppHARO and OppHARD are similarly calculated. Now I need to convert these to GF20 and GA20 based stats so I do that by multiplying by league average GF20 (LAGF20) and league average GA20 (LAGA20) and from here I can calculated expected GF20 and expected GA20.

ExpGF20(2nd iteration) = (TMHARO*LAGF20 + OppHARD*LAGA20)/2
ExpGA20(2nd iteration) = (TMHARD*LAGA20 + OppHARD*LAGF20)/2

From there we can get a second iteration of HARO and HARD.

HARO(2nd iteration) = 100*(GF20-ExpGF20) / ExpGF20
HARD(2nd iteration) = 100*(ExpGA20 – GA20) / ExpGA20

Now we iterate again and again…

Now we repeat the above step over and over again using the previous iterations HARO and HARD values at every step.

Now calculate HART

Once we have done enough iterations we can calculate HART from the final iterations HARO and HARD values.

HART = (HARO + HARD) /2

Now do the same for Shot, Fenwick and Corsi data

The above is for goal ratings but I have Shot, Fenwick and Corsi ratings as well and these can be calculated in the exact same way except using SF20, SA20, FF20, FA20, CF20 and CA20.

What about goalies?

Goalies are a little unique in that they only really play the defensive side of the game. For this reason I do not include goalies in calculating TMGF20 and OppGF20. For shot, fenwick and corsi I do not include the goalies on the defensive side of things either as I assume a goalie will not influence shots against (though this may not be entirely true as some goalies may be better at controlling rebounds and thus secondary shots but I’ll assume this is a minimal effect if it does exist). The result of this is goalies do have a HARD rating but no HARO, or shot/fenwick/corsi based HARD or HARO rating.

I hope this helps explain how my hockey analysis ratings are calculated but if you have any followup questions feel free to ask them in the comments.

 

Apr 052013
 

I often get asked questions about hockey analytics, hockey fancy stats, how to use them, what they mean, etc. and there are plenty of good places to find definitions of various hockey stats but sometimes what is more important than a definition is some guidelines on how to use them. So, with that said, here are several tips that I have for people using advanced hockey stats.

Don’t over value Quality of Competition

I don’t know how often I’ll point out one players poor stats or another players good stats and immediately get the response “Yeah, but he always plays against the opponents best players” or “Yeah, but he doesn’t play against the oppositions best players” but most people that say that kind of thing have no real idea how much quality of opponent will affect the players statistics. The truth is it is not nearly as much as you might think.  Despite some coaches desperately trying to employ line matching techniques the variation in quality of competition metric is dwarfed by variation in quality of teammates, individual talent, and on-ice results. An analysis of Pavel Datsyuk and Valterri Filppula showed that if Filppula had Datsyuk’s quality of competition his CorsiFor% would drop from 51.05% to 50.90% and his GoalsFor% would drop from 55.65% to 55.02%. In the grand scheme of things, this are relatively minor factors.

Don’t over value Zone Stats either

Like quality of competition, many people will use zone starts to justify a players good/poor statistics. The truth is zone starts are not a significant factor either. I have found that the effect of zone starts is largely eliminated after about 10 seconds after a face off and this has been found true by others as well. I account for zone starts in statistics by eliminating the 10 seconds after an offensive or defensive zone face off and I have found doing this has relatively little effect on a players stats. Henrik Sedin is maybe the most extreme case of a player getting primarily offensive zone starts and all those zone starts took him from a 55.2 fenwick% player to a 53.8% fenwick% player when zone starts are factored out. In the most extreme case there is only a 1.5% impact on a players fenwick% and the majority of players are no where close to the zone start bias of Henrik Sedin. For the majority of players you are probably talking something under 0.5% impact on their fenwick%. As for individual stats over the last 3 seasons H. Sedin had 34 goals and 172 points in 5v5 situations and just 2 goals and 14 points came within 10 seconds of a zone face off, or about 5 points a year. If instead of 70% offensive zone face off deployment he had 50% offensive zone face off deployment instead of having 14 points during the 10 second zone face off time he may have had 10.  That’s a 4 point differential over 3 years for a guy who scored 172 points. In simple terms, about 2.3% of H. Sedin’s 5v5 points can be attributed to his offensive zone start bias.

A derivative of this is that if zone starts don’t matter much, a players face off winning percentage probably doesn’t matter much either which is consistent with other studies. It’s a nice skill to have, but not worth a lot either.

Do not ignore Quality of Teammates

I have just told you to pretty much ignore quality of competition and zone starts, what about quality of teammates? Well, to put it simply, do not ignore them. Quality of teammates matters and matters a lot. Sticking with the Vancouver Canucks, lets use Alex Burrows as an example. Burrows mostly plays with the Sedin twins but has played on Kesler’s line a bit too. Over the past 3 seasons he has played about 77.9% of his ice time with H. Sedin and about 12.3% of his ice time with Ryan Kesler and the reminder with Malhotra and others. Burrow’s offensive production is significantly better when playing with H. Sedin as 88.7% of his goals and 87.2% of his points came during the 77.9% ice time he played with H. Sedin. If Burrows played 100% of his ice time with H. Sedin and produced at the same rate he would have scored 6 (9.7%) more goals and 13 (11%) more 5v5 points over the past 3 seasons. This is far more significant than the 2.3% boost H. Sedin saw from all his offensive zone starts and I am not certain my Burrows example is the most extreme example in the NHL. How many more points would an average 3rd line get if they played mostly with H. Sedin instead of the average 3rd liner. Who you play with matters a lot. You can’t look at Tyler Bozak’s decent point totals and conclude he is a decent player without considering he plays a lot with Kessel and Lupul, two very good offensive players.

Opportunity is not talent

Kind of along the same lines as the Quality of Teammates discussion, we must be careful not to confuse opportunity and results. Over the past 2 seasons Corey Perry has the second most goals of any forward in the NHL trailing only Steven Stamkos. That might seem impressive but it is a little less so when you consider Perry also had the 4th most 5v5 minutes during that time and the 11th most 5v4 minutes.  Perry is a good goal scorer but a lot of his goals come from opportunity (ice time) as much as individual talent. Among forwards with at least 1500 minutes of 5v5 ice time the past 2 seasons, Perry ranks just 30th in goals per 60 minutes of ice time. That’s still good, but far less impressive than second only to Steven Stamkos and he is actually well behind teammate Bobby Ryan (6th) in this metric. Perry is a very good player but he benefits more than others by getting a lot of ice time  and PP ice time. Perry’s goal production is a large part talent, but also somewhat opportunity driven and we need to keep this in perspective.

Don’t ignore the percentages (shooting and save)

The percentages matter, particularly shooting percentages. I have shown that players can sustain elevated on-ice shooting percentages and I have shown that players can have an impact on their line mates shooting percentages and Tom Awad has shown that a significant portion of the difference between good players and bad players is finishing ability (shooting percentage).  There is even evidence that goal based metrics (which incorporate the percentages) are a better predictor of post season success than fenwick based metric. What corsi/fenwick metrics have going for them is more reliability over small sample sizes but once you approach a full seasons worth of data that benefit is largely gone and you get more benefit from having the percentages factored into the equation. If you want to get a better understanding of what considering the percentages can do for you, try to do a Malkin vs Gomez comparison or a Crosby vs Tyler Kennedy comparison over the past several years. Gomez and Kennedy actually look like relatively decent comparisons if you just consider shot based metrics, but both are terrible percentage players while Malkin and Crosby are excellent percentage players and it is the percentages that make Malkin and Crosby so special. This is an extreme example but the percentages should not be ignored if you want a true representation of a players abilities.

More is definitely better

One of the reason many people have jumped on the shot attempt/corsi/fenwick band wagon is because they are more frequent events than goals and thus give you more reliable metrics. This is true over small sample sizes but as explained above, the percentages matter too and should not be ignored. Luckily, for most players we have ample data to get past the sample size issues. There is no reason to evaluate a player based on half a seasons data if that player has been in the league for several years. Look at 2, 3, 4 years of data.  Look for trends. Is the player consistently a higher corsi player? Is the player consistently a high shooting percentage player? Is the player improving? Declining? I have shown on numerous occassions that goals are a better predictor of future goal rates than corsi/fenwick starting at about one year of data but multiple years are definitely better. Any conclusion about a players talent level using a single season of data or less (regardless of whether it is corsi or goal based) is subject to a significant level of uncertainty. We have multiple years of data for the majority of players so use it. I even aggregate multiple years into one data set for you on stats.hockeyanalysis.com for you so it isn’t even time consuming. The data is there, use it. More is definitely better.

WOWY’s are where it is at

In my mind WOWY’s are the best tool for advanced player evaluation. WOWY stands for with or without you and looks at how a player performs while on the ice with a team mate and while on the ice without a team mate. What WOWY’s can tell you is whether a particular player is a core player driving team success or a player along for the ride. Players that consistently make their team mates statistics better when they are on the ice with them are the players you want on your team. Anze Kopitar is an example of a player who consistently makes his teammates better. Jack Johnson is an example of a player that does not, particularly when looking at goal based metrics.   Then there are a large number of players that are good players that neither drive your teams success nor hold it back, or as I like to say, complementary players. Ideally you build your team around a core of players like Kopitar that will drive success and fill it in with a group of complementary players and quickly rid yourself of players like Jack Johnson that act as drags on the team.

 

Mar 202013
 

I generally think that the majority of people give too much importance to quality of competition (QoC) and its impact on a players statistics but if we are going to use QoC metrics let’s at least try and use the best ones available. In this post I will take a look at some QoC metrics that are available on stats.hockeyanalysis.com and explain why they might be better than those typically in use.

OppGF20, OppGA20, OppGF%

These three stats are the average GF20 (on ice goals for per 20 minutes), OppGA20 (on ice goals against per 20 minutes) and GF% (on ice GF / [on ice GF + on ice GA]) of all the opposition players that a player lined up against weighted by ice time against. In fact, these stats go a bit further in that they remove the ice time the opponent players played against the player so that a player won’t influence his own QoC (not nearly as important as QoT but still a good thing to do). So, essentially these three stats are the goal scoring ability of the opposition players, the goal defending ability of the opposition players, and the overall value of the opposition players. Note that opposition goalies are not included in the calculation of OppGF20 as it is assume the goalies have no influence on scoring goals.

The benefits of using these stats are they are easy to understand and are in a unit (goals per 20 minutes of ice time) that is easily understood. GF20 is essentially how many goals we expect the players opponents would score on average per 20 minutes of ice time. The drawback from this stat is that if good players play against good players and bad players play against bad players a good player and a bad player may have similar statistics but the good players is a better player because he did it against better quality opponents. There is no consideration for the context of the opponents statistics and that may matter.

Let’s take a look at the top 10 forwards in OppGF20 last season.

Player Team OppGF20
Patrick Dwyer Carolina 0.811
Brandon Sutter Carolina 0.811
Travis Moen Montreal 0.811
Carl Hagelin NY Rangers 0.806
Marcel Goc Florida 0.804
Tomas Plekanec Montreal 0.804
Brooks Laich Washington 0.800
Ryan Callahan NY Rangers 0.799
Patrik Elias New Jersey 0.798
Alexei Ponikarovsky New Jersey 0.795

You will notice that every single player is from the eastern conference. The reason for this is that the eastern conference is a more offensive conference. Taking a look at the top 10 players in OppGA20 will show the opposite.

Player Team OppGF20
Marcus Kruger Chicago 0.719
Jamal Mayers Chicago 0.720
Mark Letestu Columbus 0.721
Andrew Brunette Chicago 0.723
Andrew Cogliano Anaheim 0.723
Viktor Stalberg Chicago 0.724
Matt Halischuk Nashville 0.724
Kyle Chipchura Phoenix 0.724
Matt Belesky Anaheim 0.724
Cory Emmerton Detroit 0.724

Now, what happens when we look at OppGF%?

Player Team OppGF%
Mike Fisher Nashville 51.6%
Martin Havlat San Jose 51.4%
Vaclav Prospal Columbus 51.3%
Mike Cammalleri Calgary 51.3%
Martin Erat Nashville 51.3%
Sergei Kostitsyn Nashville 51.3%
Dave Bolland Chicago 51.2%
Rick Nash Columbus 51.2%
Travis Moen Montreal 51.0%
Patrick Marleau San Jose 51.0%

There are predominantly western conference teams with a couple of eastern conference players mixed in. The reason for this western conference bias is that the western conference was the better conference and thus it makes sense that the QoC would be tougher for western conference players.

OppFF20, OppFA20, OppFF%

These are exactly the same stats as the goal based stats above but instead of using goals for/against/percentage they use fenwick for/against/percentage (fenwick is shots + shots that missed the net). I won’t go into details but you can find the top players in OppFF20 here, in OppFA20 here, and OppFF% here. You will find a a lot of similarities to the OppGF20, OppGA20 and OppGF% lists but if you ask me which I think is a better QoC metric I’d lean towards the goal based ones. The reason for this is that the smaller sample size issues we see with goal statistics is not going to be nearly as significant in the QoC metrics because over all opponents luck will average out (for every unlucky opponent you are likely to have a lucky one t cancel out the effects). That said, if you are doing a fenwick based analysis it probably makes more sense to use a fenwick based QoC metric.

HARO QoC, HARD QoC, HART QoC

As stated above, one of the flaws of the above QoC metrics is that there is no consideration for the context of the opponents statistics. One of the ways around this is to use the HockeyAnalysis.com HARO (offense), HARD (defense) and HART (Total/Overall) ratings in calculating QoC. These are player ratings that take into account both quality of teammates and quality of competition (here is a brief explanation of what these ratings are).The HARO QoC, HARD QoC and HART QoC metrics are simply the average HARO, HARD and HART ratings of players opponents.

Here are the top 10 forwards in HARO QoC last year:

Player Team HARO QoC
Patrick Dwyer Carolina 6.0
Brandon Sutter Carolina 5.9
Travis Moen Montreal 5.8
Tomas Plekanec Montreal 5.8
Marcel Goc Florida 5.6
Carl Hagelin NY Rangers 5.5
Ryan Callahan NY Rangers 5.3
Brooks Laich Washington 5.3
Michael Grabner NY Islanders 5.2
Patrik Elias New Jersey 5.2

There are a lot of similarities to the OppGF20 list with the eastern conference dominating. There are a few changes, but not too many, which really is not that big of a surprise to me knowing that there is very little evidence that QoC has a significant impact on a players statistics and thus considering the opponents QoC will not have a significant impact on the opponents stats and thus not a significant impact on a players QoC. That said, I believe these should produce slightly better QoC ratings. Also note that a 6.0 HARO QoC indicates that the opponent players are expected to produce a 6.0% boost on the league average GF20.

Here are the top 10 forwards in HARD QoC last year:

Player Team HARD QoC
Jamal Mayers Chicago 6.0
Marcus Kruger Chicago 5.9
Mark Letestu Columbus 5.8
Tim Jackman Calgary 5.3
Colin Fraser Los Angeles 5.2
Cory Emmerton Detroit 5.2
Matt Belesky Anaheim 5.2
Kyle Chipchura Phoenix 5.1
Andrew Brunette Chicago 5.1
Colton Gilles Columbus 5.0

And now the top 10 forwards in HART QoC last year:

Player Team HART QoC
Dave Bolland Chicago 3.2
Martin Havlat San Jose 3.0
Mark Letestu Columbus 2.5
Jeff Carter Los Angeles 2.5
Derick Brassard Columbus 2.5
Rick Nash Columbus 2.4
Mike Fisher Nashville 2.4
Vaclav Prospal Columbus 2.2
Ryan Getzlaf Anaheim 2.2
Viktor Stalberg Chicago 2.1

Shots and Corsi based QoC

You can also find similar QoC stats using shots as the base stat or using corsi (shots + shots that missed the net + shots that were blocked) on stats.hockeyanalysis.com but they are all the same as above so I’ll not go into them in any detail.

CorsiRel QoC

The most common currently used QoC metric seems to be CorsiRel QoC (found on behindthenet.ca) but in my opinion this is not so much a QoC metric but a ‘usage’ metric. CorsiRel is a statistic that compares the teams corsi differential when the player is on the ice to the teams corsi differential when they player is not on the ice.  CorsiRel QoC is the average CorsiRel of all the players opponents.

The problem with CorsiRel is that good players on a bad team with little depth can put up really high CorsiRel stats compared to similarly good players on a good team with good depth because essentially it is comparing a player relative to his teammates. The more good teammates you have, the more difficult it is to put up a good CorsiRel. So, on any given team the players with a good CorsiRel are the best players on team team but you can’t compare CorsiRel on players on different teams because the quality of the teams could be different.

CorsiRel QoC is essentially the average CorsiRel of all the players opponents but because CorsiRel is flawed, CorsiRel QoC ends up being flawed too. For players on the same team, the player with the highest CorsiRel QoC plays against the toughest competition so in this sense it tells us who is getting the toughest minutes on the team, but again CorsiRel QoC is not really that useful when comparing players across teams.  For these reasons I consider CorsiRel QoC more of a tool to see the usage of a player compared to his teammates, but is not in my opinion a true QoC metric.

I may be biased, but in my opinion there is no reason to use CorsiRel QoC anymore. Whether you use GF20, GA20, GF%, HARO QoC, HARD QoC, and HART QoC, or any of their shot/fenwick/corsi variants they should all produce better QoC measures that are comparable across teams (which is the major draw back of CorsiRel QoC.

 

Feb 272013
 

The last several days I have been playing around a fair bit with team data and analyzing various metrics for their usefulness in predicting future outcomes and I have come across some interesting observations. Specifically, with more years of data, fenwick becomes significantly less important/valuable while goals and the percentages become more important/valuable. Let me explain.

Let’s first look at the year over year correlations in the various stats themselves.

Y1 vs Y2 Y12 vs Y34 Y123 vs Y45
FF% 0.3334 0.2447 0.1937
FF60 0.2414 0.1635 0.0976
FA60 0.3714 0.2743 0.3224
GF% 0.1891 0.2494 0.3514
GF60 0.0409 0.1468 0.1854
GA60 0.1953 0.3669 0.4476
Sh% 0.0002 0.0117 0.0047
Sv% 0.1278 0.2954 0.3350
PDO 0.0551 0.0564 0.1127
RegPts 0.2664 0.3890 0.3744

The above table shows the r^2 between past events and future events.  The Y1 vs Y2 column is the r^2 between subsequent years (i.e. 0708 vs 0809, 0809 vs 0910, 0910 vs 1011, 1011 vs 1112).  The Y12 vs Y23 is a 2 year vs 2 year r^2 (i.e. 07-09 vs 09-11 and 08-10 vs 10-12) and the Y123 vs Y45 is the 3 year vs 2 year comparison (i.e. 07-10 vs 10-12). RegPts is points earned during regulation play (using win-loss-tie point system).

As you can see, with increased sample size, the fenwick stats abilitity to predict future fenwick stats diminishes, particularly for fenwick for and fenwick %. All the other stats generally get better with increased sample size, except for shooting percentage which has no predictive power of future shooting percentage.

The increased predictive nature of the goal and percentage stats with increased sample size makes perfect sense as the increased sample size will decrease the random variability of these stats but I have no definitive explanation as to why the fenwick stats can’t maintain their predictive ability with increased sample sizes.

Let’s take a look at how well each statistic correlates with regulation points using various sample sizes.

1 year 2 year 3 year 4 year 5 year
FF% 0.3030 0.4360 0.5383 0.5541 0.5461
GF% 0.7022 0.7919 0.8354 0.8525 0.8685
Sh% 0.0672 0.0662 0.0477 0.0435 0.0529
Sv% 0.2179 0.2482 0.2515 0.2958 0.3221
PDO 0.2956 0.2913 0.2948 0.3393 0.3937
GF60 0.2505 0.3411 0.3404 0.3302 0.3226
GA60 0.4575 0.5831 0.6418 0.6721 0.6794
FF60 0.1954 0.3058 0.3655 0.4026 0.3951
FA60 0.1788 0.2638 0.3531 0.3480 0.3357

Again, the values are r^2 with regulation points.  Nothing too surprising there except maybe that team shooting percentage is so poorly correlated with winning because at the individual level it is clear that shooting percentages are highly correlated with goal scoring. It seems apparent from the table above that team save percentage is a significant factor in winning (or as my fellow Leaf fans can attest to, lack of save percentage is a significant factor in losing).

The final table I want to look at is how well a few of the stats are at predicting future regulation time point totals.

Y1 vs Y2 Y12 vs Y34 Y123 vs Y45
FF% 0.2500 0.2257 0.1622
GF% 0.2214 0.3187 0.3429
PDO 0.0256 0.0534 0.1212
RegPts 0.2664 0.3890 0.3744

The values are r^2 with future regulation point totals. Regardless of time frame used, past regulation time point totals are the best predictor of future regulation time point totals. Single season FF% is slightly better at predicting following season regulation point totals but with 2 or more years of data GF% becomes a significantly better predictor as the predictive ability of GF% improves and FF% declines. This makes sense as we earlier observed that increasing sample size improves GF% predictability of future GF% while FF% gets worse and that GF% is more highly correlated with regulation point totals than FF%.

One thing that is clear from the above tables is that defense has been far more important to winning than offense. Regardless of whether we look at GF60, FF60, or Sh% their level of importance trails their defensive counterpart (GA60, FA60 and Sv%), usually significantly. The defensive stats more highly correlate with winning and are more consistent from year to year. Defense and goaltending wins in the NHL.

What is interesting though is that this largely differs from what we see at the individual level. At the individual level there is much more variation in the offensive stats indicating individual players have more control over the offensive side of the game. This might suggest that team philosophies drive the defensive side of the game (i.e. how defensive minded the team is, the playing style, etc.) but the offensive side of the game is dominated more by the offensive skill level of the individual players. At the very least it is something worth of further investigation.

The last takeaway from this analysis is the declining predictive value of fenwick/corsi with increased sample size. I am not quite sure what to make of this. If anyone has any theories I’d be interested in hearing them. One theory I have is that fenwick rates are not a part of the average GMs player personal decisions and thus over time as players come and go any fenwick rates will begin to vary. If this is the case, then this may represent an area of value that a GM could exploit.