But past results are realized, not predictions based on averages. Which makes one of them accurate up to that point, and the other one can easily be inaccurate going forward because it is a prediction. There's at least a large margin of error in it.
I get that we can't expect past results to repeat themselves so it's not entirely accurate either, the performance will fluctuate. In young players cases mostly goes up and when aged past physical prime it goes down.
And then there's the individuals who break the average based models. They can't be accurately predicted by league averages. The more average the plauer is, the better the league averages can predict them.
Yes and no. Be careful as sometimes you are making erroneous leaps in conclusions...
"Accurate up to that point"
More accurate, in only particular context only...
Accurate that they did outscore: yes.
Accurate that they will outscore: no.
It told you what did happen, while other things tell you what will happen. The accuracy is only particular to a set of parameters. The answer to one question does not answer all questions.
Performance with age curves
Underlying statistics and scoring follows an aging curve; however, PDO does not. PDO predominately varies around 100 for all ages.
Individual sh% is different than on-ice sh%. You cannot conflate the two. Laine is a generational talent with the former, but in the latter the best guess is Laine's true talent is only worth 2 goals per 100 shots more than average.
Models "based on averages"
The reason why I use average in the experiment is because "luck" (ie: random variance) doesn't predict "luck." Some players will have bad luck in both halves of the sample, as will some have good luck in both halves of the sample.
IE: You split the distribution of PDO after the first half of the season at the mean and then look at each group separately for their second halves, both groups will have normal distributions around league mean PDO.
Goals have a huge amount of "luck" (random variance) in them. You are predicting a target that itself is a good chunk random. That's why you have to use averages. The average though can still be applied to the non-average (see later); that's two different things.
"Individuals that break the average based models"
No one "breaks" any statistical model to the point where PDO as a tool differs in how it should be used. All players regress to the mean, what is confusing you is that each player's mean is slightly different.
For all players a high or low PDO still means that their performance in goal is unsustainable. What differs is where their true talent PDO lies, although it's always much closer to the league mean than many people mistakenly tend to think.
IE: Even for Laine, you would look at him in different situations of high PDOs and low PDOs as unsustainable. His true talent may be closer to 102 than 100, but when you see a line where he posts 115 that's not skill/talent/chemistry... that's small sample variance.
Small samples variation
In an 82 game season, with thousands of shots, what makes one team different in sh% from the mean is about 2/3 "luck" (random variance), and 1/3 combination of shot quality (location, rebounds, etc.) and finishing talent combined.
Think about how much smaller samples you are going with when looking at Laine's different partners, where you are looking at extremely low number of shots and minutes with very different situations (goalies, linematching, score, etc.).
Laine and shot metrics
Shot metrics still matter for Laine. He scores on a greater number of shots than expected due to elite finishing skill, but that just shifts the model, not flips it upside down or makes it obsolete.
This makes logical sense. When Laine is on the ice you still want more chances and still want the opposition to have fewer. Having Laine or any other elite finisher only just shifts the translation factor.
Laine's best year in xG differential is also the same year as his best year in goal differential (2017-18). Laine's worst year in xG differential is also the same year as his worst year in goal differential (2018-19).
Laine doesn't break models. Laine shifts models.
The model of hockey at EV is still the same:
* Shot quantity (try to make more shots and allow fewer)
* Shot quality (try to push your shots closer and to a less wide angle)
* Finishing talent / goalie talent (maximize on chances and minimize opponent chances)