Quote:
Originally Posted by DeluxeMoustache
This leads to a question about what the whole value of the statistical analysis is, though
The high level argument (loosely framed, please don’t nitpick) is that shots are a proxy for possession / control, etc. People throw around save percentages like they apply more or less uniformly, and implicitly in the long run, proportionally share of possession translates to share of goals and share of wins.
Many of us have studied statistics to some extent and know that a good fit will have enough data points so that results can be expressed as accurate, say, within a certain margin or percentage, 19 times out of 20.
There are a lot of shots, sure, but few game results. Look at the Flames season. Apply analytics and compare to game results. Translating shot counts to wins has so many other contributing factors, and ultimately outliers in terms of results, that getting to a level of statistic significance is impossible.
If you have, say, 18 games and 4 outliers, the model is likely not good enough.
|
I really think this is another example of over thinking things.
First off nobody is calling anything a model, you won't see that coming out of my mouth or through my keyboard.
It's counting. Counting isn't a model. Summarizing a bunch of counting stats isn't a model, its a summary.
Score effects explains a lot of the 4/22 examples that don't fit. The Flames got out numbered in Nashville and Manhattan because they led the whole way. They out numbered teams that beat them in Vancouver and the Penguins because they trailed the whole way.
But nothing can be 100% predictive about any counting stats, including simple ones like shots on goal that we used to use. You can lose a game while out shooting the opponent, and you can win a game when getting out shot. But I'm willing to guess over a large enough sample size it's better to get more shots than your opponent if you want to win.
Similarly it's a pretty sound assumption that you probably win more games than you lose if you get more scoring chances than the opponent.
I'm just summarizing that ... it's not a model.