Fenwick Close Road + PDO A Better Correlation Than Fenwick Tied Road

Yes on a team basis it would be the average of all the team’s goalies ES SV% (goaltending). And so a good goalie may inflate a players PDO (i.e. SV% portion).

On a player basis it is the goalies ES SV% when the player is on the ice. And I’m not sure if a skater can influence SV% directly. It might be like SH%, in that the player can increase the odds of scoring by certain factors (screening goalie, shooting from a high probality location etc). And perhaps a players can influence his specific SV% by keeping shooters to perimeter and clearing the front of the net. I may look at that later.

But the nice thing about the regression of the sum of three factors (SH%, SV% and Fenwick) is that it negates some variables when using a statistic on its own.

Think of it like this, a poor defensive player would give up a lot of shots. If they have a great goalie the high SV% would “mask” their weak 2 way play. But then the player fenwick would drop accordingly to somewhat negate the effect of the stronger goalie. So that players PDO increase + F% decrease would some offset.. That is why I think the regression is is stronger.

If the player is a strong 2 way player with a poor goalie, then his on ice SV% would drop but his Fenwick would rise.

The SH% is the conversion factor to goals of shots direct at the opposition net.
The SV% is the conversion factor to goals of shots directed at the players own net.

A team that blocks a lot of shots or has an effective trap may be able to “inflate” Fenwick but not their SV%. Or vice versa, a hot goalie would inflate SV% and PDO but the team fenwick would suggest things needs to even out at some point. The two factors works better together.

I looked at year over year correlation for 2007 to 2010 and half year (YTD) for 2011. The only half year study completed was for the 2011 YTD campaign. If I can get the dataset for 2007 to 2010 intra-year analysis that could be completed.

Below I show the split half (odd/even) and spearman-brown prophecy for each of Fenwick Tied Road, PDO and Fenwick + PDO measures. By any reliability measure, PDO + Fenwick has a higher correlation with lower standard deviation. The Cronbach’s alpha is strange as it suggest no advanced stat (Fenwick or PDO or Fenwick + PDO) is not material, however the appropriateness of this test for this dataset is questionable (i.e. single variable and variability of P%). I need to to more reading but I think because P% is 50% luck, then Cronback won’t give a good measure near 0.7 (as is shown the Fenwick measure on its own and Fenwick + PDO).

I think the main advantage of Fenwick + PDO is that it can distinguish if P% is sustainable for a skilled team like Boston with higher SV% (Thomas) and SH% (3 strong scoring lines) compared to a team like Minnesota whose PDO + Fenwick is generating a P% beyond what should be expected. In fact, the measure can tell you if a more skilled team (Boston) is lucky or "good" that may not be apparent looking at PDO itself. That is, I’m suggesting that all teams may not regress to 1000 PDO and some may indeed be more skilled then other (though of course on a league wide basis the average is 1000). If so, the metric better describes why some teams can sustain a higher PDO for entire season and not regress to zero (that is why Boston last year can sustain a high PDO and a team like TOR 2009 can sustain a low PDO). Yes I agree SV% and SH% is not "repeatable" but some teams can sustain a higher or lower then league SV% and/or SH% for the season. So for example, the typical NHL team SV% maybe 920 +- luck. And Boston’s NHL team SV% is 940 +- luck. And Columbus maybe 900 +- luck. This metric is not perfect but does a better job then Fenwick Tied Road (average and stdev). And does not require the user to use as much context as PDO may require on its own.

At the end of the day, Fenwick + PDO works because it better predicts why some teams with same Fenwicks (skill) have difference in P% (they have better goalies and maybe shooters). Luck explain a lot but it doesn’t explain everything. And of course just because a team has a small skill advantage in net, we know that that luck accounts for at least 50% of the outcome.

I agree PDO is not consistent. The PDO + Fenwick is decent but not as consistent as Fenwick itself. I’m not sure how to trade off stat r(self) and Rsquared but it seems even though PDO + fenwick is better regression at the end of the season, for mid or early season Fenwick (or others have shown Corsi) to be better predictors.

I might reframe PDO to see if that can be improved.
Also how do you get 20 games or subsets for the test?