Why is PDO an important stat?

[OnThe Forecheck]

A teams shooting and save percentage are unrelated so why must they revert to 100? Why is that a sign of unsustainable luck and not the sign of a good team?

Click to read more...

 

[ResilientBeast]

A teams shooting and save percentage are unrelated so why must they revert to 100? Why is that a sign of unsustainable luck and not the sign of a good team?

I think because the average save percentage + average shooting percentage just happens to be 1000. So if both or either is higher than the average then it makes one believe that the team is experiencing a streak of unsustainable play.

 

[OnThe Forecheck]

But a team with a top tier offense and goaltending shouldn't regress. It wouldn't be unusual for a team with Carey Price and a solid back up to have a save % of say .926 and a team with Steven Stamkos and other quality offensive players to have a shooting percentage between 9-11.

 

[tigervixxxen]

PDO is a terrible stat. If anything PDO should be driven off of individual career averages and not league average.

 

[OnThe Forecheck]

So is corsi. Certain players and certain teams take less shots but higher quality shots. And certain teams and players take more shots and not all of them are high quality. Steven Stamkos vs Alex Oveckin. Corsi is pointless because it measures every shot the same and like PDO overstates the importance of luck in hockey.

 

[Irishguy42]

The big problem with PDO is that too many people think that 1000 is the bar for under/over performance for any player/team.

That's simply not true at all.

I also think PDO is calculated poorly, and needs more than just SV%+SH%. That seems too simple for what it's trying to show.

 

[OnThe Forecheck]

The big problem with PDO is that too many people think that 1000 is the bar for under/over performance for any player/team.

That's simply not true at all.

What is the truth then? There have to be different expectations for different teams? The Coyotes aren't unluck, they're just truly terrible

 

[Irishguy42]

What is the truth then? There have to be different expectations for different teams? The Coyotes aren't unluck, they're just truly terrible

I think the truth is found in some other formula to calculate the concept of over/under performance.

 

[Mr Misty]

As certain folks like to say, hockey is the ultimate team game. So it is hard for you to tell how good somebody is because the noise from their teammates and opponents is so loud. Plenty of players go through "slumps" where their sv% is cratered or their teammates are shooting unsustainably poorly. Are they suddenly bad? No, teammates have a huge impact on results.

I believe there is lots of work that shows PDO is not repeatable year on year. So as much fun as it is to say "Maybe we just figured the magic formula out" that has never been the case because you can't predict it. And past results tell us the "normal" band we should expect NHL teams to be within. IIRC the NYR were tops last year with ~102.4. If a team has 103+, we have every reason to be skeptical that will continue over an entire season.

The problem may be that you are taking it personally when somebody on the internet says your hockey team isn't really as good as its record. Don't be that person.

 

[Finnish your Czech]

I think the sh% side is more unsustainable than the sv% side of the PDO table, but I don't have the data to back it up.

 

[OrrNumber4]

The idea is a mathematical reality that league-wide save percentage and shooting % has to equal 100%. For this to happen, the individual Sv+Sh% for the 30 teams has to average to 100%. Throw in the fact that the difference between the save percentage and shooting percentages of the best teams and the worst teams isn't that large, it figures that teams that are well over 100% should regress, and teams well under 100% should progress.

 

[Buchnevich]

I think PDO is a good tool if you are juat taking a glance at it and using other stats to justify why a team's PDO is either high or low. For example, if a teams shooting percentage after half the season is at 6% and their save percentage is at 91%, their PDO would be at 97% and some people would jump to the conclusion that they are just unlucky. But that team could just be taking the majority of their shots from low danger scoring zones and allowing more high danger chances than the average team would. Because of this, there is no reason to believe their PDO would rise. Again, PDO can be a useful tool, but it can't be the "end all be all" stat.

 

[OCSportsfan]

It is the worst analytics stat out there.

A good goalie or defensive system is probably going to have a higher save percentage than a poor team. That number should be sustainable for the year and if players/coaches dont change then it can be sustainable multiple years.

Shooting percentage has a better chance to be measured since you play various defensive teams (good and bad), but also has too many variables that affect the result. Individual shooting percentage may be easier to gauge when comparing previous years, but even then you can get wild fluctuations.

Dont use it

 

[Mr Misty]

It is the worst analytics stat out there.

A good goalie or defensive system is probably going to have a higher save percentage than a poor team. That number should be sustainable for the year and if players/coaches dont change then it can be sustainable multiple years.

Shooting percentage has a better chance to be measured since you play various defensive teams (good and bad), but also has too many variables that affect the result. Individual shooting percentage may be easier to gauge when comparing previous years, but even then you can get wild fluctuations.

Dont use it

So why doesn't it come out that way then? If you are going to make big statements about it you need something to back it up.

Your explanation about shooting percentage being different should apply equally to save percentage because opposing shooter strength varies just as much as opposing goalie strength.

 

[36kap36]

I think PDO is a good tool if you are juat taking a glance at it and using other stats to justify why a team's PDO is either high or low. For example, if a teams shooting percentage after half the season is at 6% and their save percentage is at 91%, their PDO would be at 97% and some people would jump to the conclusion that they are just unlucky. But that team could just be taking the majority of their shots from low danger scoring zones and allowing more high danger chances than the average team would. Because of this, there is no reason to believe their PDO would rise. Again, PDO can be a useful tool, but it can't be the "end all be all" stat.

Isn't that the problem? I don't know that any single stat in the current framework of hockey analytics should truly be trusted as the "end all be all" stat, because there is so much uncertainty, and so many assumptions. Unlike a WAR stat in baseball, the stats (especially the simpler, more popular ones, like Corsi and PDO) are not well enough vetted or well enough tested to be absolutely confident in.

What makes sense for any stat is to look at it like you're saying, and to use it in combination with other metrics to make points. I don't know that just makes sense to me.

 

[me2]

There isn't anything wrong with PDO. It is what it is, fundamentally its a comparison of your goaltending vs your opponents in your games (aka your shooting percentage). It's just a number, and an honest one. The problem comes from people when they assume things like regression to 1000 etc. Apply a little context, elite goalies should help it above 1000, a team with proven high percentage shooters should help it be above 1000, etc. It's usual for looking a hot streaks or lucky seasons that aren't likely to be sustainable.

 

[wgknestrick]

Here are some past posts I've made about the PDO subject:

http://stats.hockeyanalysis.com/rat...0&teamid=0&type=shots&sort=ShPct&sortdir=DESC

7yr sample with players sorted by 5v5 SV%. Are we surprised to see the best player in the league at #1 with 11% SH%, and conversely Craig Adams last with 5%?

Did Crosby just get lucky those 7 years and over-perform his expected 8% SH%?

Using PDO to estimate a player's future is only acceptable if you already know that player's mean PDO. It somewhat works because the majority of players will naturally be close the bad assumption of "PDO should = 1 for all players" because the player population is bell curved around the mean. Just because an assumption can be close for the majority of players, does not make it a correct assumption.

We cannot assume each player is the same and would be expected to have the same SH% (and also SV% from goaltending).

PDO forecasting on a individual level will be difficult because you need a large sample to determine a player's true SH% talent. Goals are a rare event. It defeats the purpose of trying to use "PDO should = 1" as a shortcut when you don't have the volume of data about a player.

A similar analogy is to get the average MPG of the entire new car fleet made by every automotive company (probably around 25MPG), and then try to compare either a Prius's MPG or a V12 Ferrari's MPG to it. You already made an incorrect assumption that every car is "the same".

Regression towards the mean is only valid if you are dealing with samples of the same thing. Crosby's SH% is not the same as Craig Adam's. Both are centered around their own means and will vary around that.

PDO does not "regress" to 1.0 for anyone or any team This is mathematical "phenomenon" when looking at a league SH% and SV% averages in whole. league average (SH%+SV%)=1. It only regresses to 1.0000 for an individual (team or player) if they lie perfectly on the mean (or a sum of) in both components. A league is not a sum of equal teams. They are distributed about the mean. Probability of an analog sample falling exactly on an individual value is 0.

Please understand this, and help stop people from claiming it. If (PDO =1) were "true" then SH% would have to be equal among individual players and we all know that is not true. Nothing is in place (SV% wise) balancing Crosby's consistently high SH% to ensure his PDO regresses to 1.0.

A team is a sum of players so if any rule must by true at the team level, it must also be true at the individual level. It is entirely possible for a good team to have a "regression PDO" of 1.03 and another less skilled team to have a "regression PDO" of .98 based on their rosters of unequal players.

For the PDO = 1 falsity to be true at the team level, a league would require you to ice an equally poor player to counter weight your good players and play him an equal amount. "PDO regressing to 1.0" is an assumption that is improper, and has 0% of being correct (see above).

Team PDO = SUM [player PDO x % ice time]

 

[Feed Me A Stray Cat]

The points above are fair critiques of PDO. You can't simply say that every team will regress to 100.0. But, I do think it's useful in evaluating teams at the extreme ends of the spectrum in smaller smaple sizes.

For some clarity:
http://stats.hockeyanalysis.com/teamstats.php?disp=1&db=201417&sit=5v5&sort=PDO&sortdir=DESC

Over the last three seasons, the top 5v5 PDO is the Rangers at 102.1 and the worst is the Hurricanes at 97.8.

25 of the 30 teams have three year PDOs in the range of 99.0 to 101.0. So 83% of teams have a PDO within 1.0% of the average over the last three years.

Let's look at teams PDO at about the haflway mark of 2016-17:
http://stats.hockeyanalysis.com/teamstats.php?disp=1&db=201617&sit=5v5&sort=PDO&sortdir=DESC

The Wild and Capitals will have a hard time sustaining PDOs around 103.0. Colorado's PDO will likely increase a bit. We won't have seven teams with a PDO above 101.0 by season's end, nor will we have eight teams with a PDO below 99.0.

So PDO is good for evaluating the outliers. It's more useful at the beginning of the season, say 20 games in, when teams are riding percentages a bit more (certain teams will be at 105.0 and 95.0, and we know that those simply aren't sustainable). As the sample size increases, the value of using PDO decreases.

 

[decma]

The points above are fair critiques of PDO. You can't simply say that every team will regress to 100.0. But, I do think it's useful in evaluating teams at the extreme ends of the spectrum in smaller smaple sizes.

For some clarity:
http://stats.hockeyanalysis.com/teamstats.php?disp=1&db=201417&sit=5v5&sort=PDO&sortdir=DESC

Over the last three seasons, the top 5v5 PDO is the Rangers at 102.1 and the worst is the Hurricanes at 97.8.

25 of the 30 teams have three year PDOs in the range of 99.0 to 101.0. So 83% of teams have a PDO within 1.0% of the average over the last three years.

Let's look at teams PDO at about the haflway mark of 2016-17:
http://stats.hockeyanalysis.com/teamstats.php?disp=1&db=201617&sit=5v5&sort=PDO&sortdir=DESC

The Wild and Capitals will have a hard time sustaining PDOs around 103.0. Colorado's PDO will likely increase a bit. We won't have seven teams with a PDO above 101.0 by season's end, nor will we have eight teams with a PDO below 99.0.

So PDO is good for evaluating the outliers. It's more useful at the beginning of the season, say 20 games in, when teams are riding percentages a bit more (certain teams will be at 105.0 and 95.0, and we know that those simply aren't sustainable). As the sample size increases, the value of using PDO decreases.

I don't think a strong inference about how many teams will have a 16/17 PDO above or below a certain threshold can be made using three-year PDO data.

Four teams had a PDO of at least 101.0 in 15/16, seven did in 14/15, and five did in 13/14.
It seems that while a three-year PDO of at least 101.0 is quite rate, it is far less unusual in a single season, which is to be expected.

 

[jnk96]

So PDO is good for evaluating the outliers. It's more useful at the beginning of the season, say 20 games in, when teams are riding percentages a bit more (certain teams will be at 105.0 and 95.0, and we know that those simply aren't sustainable). As the sample size increases, the value of using PDO decreases.

^This is it.

Luck is just noise in statistics and no team or person in the world will always be lucky. So, the larger the sample size, the closer we will get to 100. It really is all about small sample sizes and events that can't otherwise be explained.

There are interesting ways to look at PDO, e.g. a 10-game moving average. Let's say your normally bad team is suddenly on a six-game winning streak. Have they recorded a PDO of at, say, 115 during that time? Well, they've probably just been lucky then.

Also connect it with other stats. For example, if we continue the "sudden streak" example, look at shot attempts. Has the team been creating more chances over the past 10 games and as a result won more? If not, they have likely just been lucky, and you will see that in PDO.

PDO can be useful as long as you use it right.