Adjusting GAA Based on Change in Win Probability

[TheChamber91]

Hi, I want to try a new approach to adjusting GAA, how much that goal hurts that tendy's team's chances of winning that game. For example, a goal given up with 30 seconds left in the game when a team is up 8-1 shouldn't count equally to a go ahead goal given up with the same amount of time left. I would've put this in the future studies thread, but I have too many questions to go along with this. I've looked for a suitable way to calculate win probability in hockey (I tried creating a seperate thread specifically on this, it didn't work), and I couldn't find one. Next, I'd need help creating the parameters that the value of each goal would fall within. Like, a goal against that lessened the team's win probability from (given amount to given amount) counts as (x goals). If you guys could help me out with any of this stuff I'd greatly appreciate it (also you guys are normally good at telling if these stats are capable of actually revealing something useful so let me know on that).

Click to read more...

 

[Doctor No]

I have explored this in the past - it ends up being rather interesting, but I never have the patience to stick with it (and I always move on to something else like newspaper research).

I'd start by pulling (either manually or automated) in sequence the probability that a team has of winning the game, given the lead at the time and the amount of time left in the game. I used to use one-minute increments when I would do this.

For instance, solely based on one game:
Game Summary

Teams with a one-goal lead and 30 minutes left in the game win 100% of the time.

Teams with a two-goal lead and 28 minutes left in the game win 100% of the time.

And so forth.

Obviously you'd want more data than one game.

 

[Doctor No]

The results that you pull will vary from season to season, and ultimately depend upon the goal scoring environment (in low-scoring eras, win curves will be steeper, and in high-scoring eras, win curves will be flatter).

 

[Doctor No]

Two other things to consider -

Saves are also important here, since preventing the other team from scoring while holding a lead does increase one's winning percentage (since the score remains the same but the time remaining decreases).

Since there's only one win in total for each game, goals allowed by the other team will impact this (especially if you only focus on goals and not saves). This may or may not be what you want.

 

[Hockey Outsider]

I tried doing this analysis (probably at least a decade ago - time flies). It would have been from the perspective of skaters recording goals/assists, but the principles are the same.

I attempted to use the Poisson distribution to estimate a team's chance of winning. You can read more about it in the excellent article from Alan Ryder at Hockey Analytics - http://www.hockeyanalytics.com/Research_files/Poisson_Toolbox.pdf

I ended up shelving the project as it required a huge time commitment to sift through the enormous amount of data.

 

[TheChamber91]

I tried doing this analysis (probably at least a decade ago - time flies). It would have been from the perspective of skaters recording goals/assists, but the principles are the same.

I attempted to use the Poisson distribution to estimate a team's chance of winning. You can read more about it in the excellent article from Alan Ryder at Hockey Analytics - http://www.hockeyanalytics.com/Research_files/Poisson_Toolbox.pdf

I ended up shelving the project as it required a huge time commitment to sift through the enormous amount of data.

I have explored this in the past - it ends up being rather interesting, but I never have the patience to stick with it (and I always move on to something else like newspaper research).

I'd start by pulling (either manually or automated) in sequence the probability that a team has of winning the game, given the lead at the time and the amount of time left in the game. I used to use one-minute increments when I would do this.

For instance, solely based on one game:
Game Summary

Teams with a one-goal lead and 30 minutes left in the game win 100% of the time.

Teams with a two-goal lead and 28 minutes left in the game win 100% of the time.

And so forth.

Obviously you'd want more data than one game.

Thanks! I'll probably attempt to pick this up when I finish the current project I'm doing. I'll probably start by doing the 2017-18 playoffs (even though it's a small sample size) just to see how it turns out.

 

[TheChamber91]

I'd start by pulling (either manually or automated) in sequence the probability that a team has of winning the game, given the lead at the time and the amount of time left in the game. I used to use one-minute increments when I would do this.

Anyone have any idea where I could get this data already compiled? I'm willing to do it manually but it would save me a ton of time if someone had this already done and I could use it.

 

[Doctor No]

I'm not aware of it being done and publicly available (I had written a little Excel macro that would let me push through one game at a time, but I still preferred doing it that way to anything more automated - much of what I do around here is to help relax).

You may be able to schmooze the folks here:
http://hfboards.mandatory.com/threads/nhl-com-sochi-2014-data-urls-in-json-and-jsonp.1596119/

Who are good at the automated scraping.

 

[TheChamber91]

I'm not aware of it being done and publicly available (I had written a little Excel macro that would let me push through one game at a time, but I still preferred doing it that way to anything more automated - much of what I do around here is to help relax).

You may be able to schmooze the folks here:
http://hfboards.mandatory.com/threads/nhl-com-sochi-2014-data-urls-in-json-and-jsonp.1596119/

Who are good at the automated scraping.

I decided I'm going to do it manually, but how large would the sample size have to be for me to measure it accurately? Would a random sample of 150-200 games be sufficient?