HFBoards

Go Back   HFBoards > General Hockey Discussion > By The Numbers
Mobile Hockey's Future Become a Sponsor Site Rules Support Forum vBookie Page 2
Notices

By The Numbers Hockey Analytics... the Final Frontier. Explore strange new worlds, to seek out new algorithms, to boldly go where no one has gone before.

2017 Playoffs Thread

Reply
 
Thread Tools
Old
05-14-2017, 10:31 AM
  #51
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
After game of May 13:

OutcomeProb
PIT in 5 games14.0%
PIT in 6 games20.0%
PIT in 7 games23.1%
OTT in 7 games11.6%
OTT in 6 games15.0%
OTT in 5 games9.8%
OTT in 4 games6.5%
PIT wins57.1%
OTT wins42.9%

__________________
The Goaltender Home Page (http://hockeygoalies.org) - Preserving Goaltender History since 1994
2016-17 NHL Goaltending Performances by Team: http://hockeygoalies.org/bio/nhl/logs/NHL2016.html
Doctor No is offline   Reply With Quote
Old
05-15-2017, 11:51 AM
  #52
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
After game of May 14:

OutcomeProb
ANA in 5 games10.5%
ANA in 6 games15.4%
ANA in 7 games20.6%
NAS in 7 games16.8%
NAS in 6 games22.6%
NAS in 5 games14.2%
ANA wins46.5%
NAS wins53.5%

Doctor No is offline   Reply With Quote
Old
05-16-2017, 09:58 AM
  #53
Sabretooth
Registered User
 
Sabretooth's Avatar
 
Join Date: May 2013
Location: Ohio
Country: United States
Posts: 2,090
vCash: 2095
Quote:
Originally Posted by Doctor No View Post
And in round two, the SRS model sucked:

WinnerProb
NYR54.9%
OTT40.4%
WAS82.3%
PIT48.4%
NAS41.3%
STL29.5%
ANA61.0%
EDM58.8%
NAS50.1%
ANA49.2%
OTT33.1%
PIT37.5%
TOTAL5.9 series (out of 12)

Yikes.
Was thinking about this a bit. I think you're being too hard on your model. I don't think just adding up the winning team percentages is the correct way to measure your model. Even if all 12 teams favored by the model won, you'd have only added up to 7.3 series out of 12.

Sabretooth is online now   Reply With Quote
Old
05-16-2017, 11:18 AM
  #54
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
That's true! Although ideally, the most predictive models will decrease the odds of a type II error (and correctly picking the winning team) while decreasing the odds of a type I error (incorrectly picking the wrong team). That is, a perfect model will correctly identify the winning team 100% of the team, with a 0% chance of an error. (Of course) that isn't possible in any reality, but then the question becomes how close can we get?

So in other words, if I say that Team A has a 52% chance of winning a series (and they do), I should get less credit than if I say that Team A has a 65% chance of winning a series (and they do).

I'll also clarify that this type of model - the SRS algorithm - is in the public domain, and so I can't take credit. In particular, there are multiple adjustments that can be made to improve the performance of the model, and I post these to engender a bit of interest in the sort of speaking that "I bet I can do better, and I'm going to try".

(My main use for the SRS algorithm - aside from promoting competition here - is that it's a pretty reasonable retrospective team strength, and so I use to it measure goaltender schedules and similar. For instance, if a team is rated as +0.40 goals/game, that may not be as predictive for future events as it could be, but it's very representative of what they've already accomplished.)

Doctor No is offline   Reply With Quote
Old
05-16-2017, 11:36 AM
  #55
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
After games of May 15:

OutcomeProb
PIT in 5 games21.1%
PIT in 6 games24.5%
PIT in 7 games23.6%
OTT in 7 games11.5%
OTT in 6 games12.9%
OTT in 5 games6.3%
PIT wins69.3%
OTT wins30.7%

Doctor No is offline   Reply With Quote
Old
05-16-2017, 03:20 PM
  #56
Sabretooth
Registered User
 
Sabretooth's Avatar
 
Join Date: May 2013
Location: Ohio
Country: United States
Posts: 2,090
vCash: 2095
Quote:
Originally Posted by Doctor No View Post
That's true! Although ideally, the most predictive models will decrease the odds of a type II error (and correctly picking the winning team) while decreasing the odds of a type I error (incorrectly picking the wrong team). That is, a perfect model will correctly identify the winning team 100% of the team, with a 0% chance of an error. (Of course) that isn't possible in any reality, but then the question becomes how close can we get?

So in other words, if I say that Team A has a 52% chance of winning a series (and they do), I should get less credit than if I say that Team A has a 65% chance of winning a series (and they do).
I guess what I'm getting at is that you really don't know if the model sucks (bad probabilities) or if it was just an unlucky roll based on 1 result. The favored team in your model won their series 5 out of 12 times, which would be expected to happen approximately 9-10% of the time even if your assigned odds were perfect. It's a more likely result than correctly predicting >9 series winners. 7 correct predictions would have been expected (still less than 1 in 4), but none of 6, 7, 8, or 9 should have been surprising. 5 correct predictions was the 5th most likely result. A perfect 12 for 12 only happens 0.2-0.3% of the time given your assigned odds. The model could be absolutely perfect and this could still be a perfectly valid result.

Sabretooth is online now   Reply With Quote
Old
05-17-2017, 01:01 PM
  #57
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
True - and any models that promise 90% accuracy per series are either selling something, or don't understand the role of luck.

With that said, I'd like the "out of the box" SRS algorithm to do better than 50% at least.

Doctor No is offline   Reply With Quote
Old
05-17-2017, 01:01 PM
  #58
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
After game of May 16:

OutcomeProb
ANA in 6 games10.3%
ANA in 7 games19.7%
NAS in 7 games16.1%
NAS in 6 games28.6%
NAS in 5 games25.2%
ANA wins30.1%
NAS wins70.0%

Doctor No is offline   Reply With Quote
Old
05-18-2017, 08:48 AM
  #59
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
After game of May 17:

OutcomeProb
PIT in 6 games19.5%
PIT in 7 games27.8%
OTT in 7 games14.8%
OTT in 6 games22.2%
OTT in 5 games15.8%
PIT wins47.2%
OTT wins52.8%

Doctor No is offline   Reply With Quote
Old
05-18-2017, 04:12 PM
  #60
Talks to Goalposts
Registered User
 
Join Date: Apr 2011
Location: Edmonton
Country: Canada
Posts: 4,937
vCash: 500
Quote:
Originally Posted by Doctor No View Post
True - and any models that promise 90% accuracy per series are either selling something, or don't understand the role of luck.

With that said, I'd like the "out of the box" SRS algorithm to do better than 50% at least.
That was a quality burn on SAP's claims as NHL's official data company. I'm curious if it was deliberate or not.

Talks to Goalposts is offline   Reply With Quote
Old
05-19-2017, 11:52 AM
  #61
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
Not deliberate, although I wish it had been.

Doctor No is offline   Reply With Quote
Old
05-19-2017, 11:57 AM
  #62
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
After game of May 18:

OutcomeProb
ANA in 6 games24.0%
ANA in 7 games27.8%
NAS in 7 games22.9%
NAS in 6 games25.3%
ANA wins51.8%
NAS wins48.2%

Doctor No is offline   Reply With Quote
Old
05-20-2017, 09:09 AM
  #63
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
After game of May 19:

OutcomeProb
PIT in 6 games36.0%
PIT in 7 games31.7%
OTT in 7 games16.6%
OTT in 6 games15.7%
PIT wins67.7%
OTT wins32.3%

Doctor No is offline   Reply With Quote
Old
05-21-2017, 10:19 AM
  #64
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
After game of May 20:

OutcomeProb
ANA in 7 games23.8%
NAS in 7 games19.5%
NAS in 6 games56.7%
ANA wins23.8%
NAS wins76.2%

Doctor No is offline   Reply With Quote
Old
05-21-2017, 05:58 PM
  #65
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
After game of May 21:

OutcomeProb
PIT in 6 games56.8%
PIT in 7 games29.1%
OTT in 7 games14.1%
PIT wins85.9%
OTT wins14.1%

Doctor No is offline   Reply With Quote
Old
05-24-2017, 10:50 AM
  #66
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
Yeah buddy!

OutcomeProb
PIT wins66.9%
OTT wins33.1%

Doctor No is offline   Reply With Quote
Old
05-25-2017, 11:03 PM
  #67
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
Algorithm correctly picked both series.

Through three rounds:

WinnerProb
NYR54.9%
OTT40.4%
WAS82.3%
PIT48.4%
NAS41.3%
STL29.5%
ANA61.0%
EDM58.8%
NAS50.1%
ANA49.2%
OTT33.1%
PIT37.5%
PIT76.2%
NAS50.3%
TOTAL7.13 series (out of 14)

Doctor No is offline   Reply With Quote
Old
05-25-2017, 11:03 PM
  #68
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
And, ladies and gentlemen, your 2017 Stanley Cup Finals:

OutcomeProb
PIT in 4 games10.0%
PIT in 5 games19.5%
PIT in 6 games17.7%
PIT in 7 games19.0%
NAS in 7 games11.3%
NAS in 6 games12.2%
NAS in 5 games7.0%
NAS in 4 games3.3%
PIT wins66.1%
NAS wins33.9%

Doctor No is offline   Reply With Quote
Old
05-26-2017, 12:44 AM
  #69
Tim Calhoun
Jim Benning Fan
 
Tim Calhoun's Avatar
 
Join Date: May 2006
Country: Mexico
Posts: 10,286
vCash: 50
Do you use some sort of logistic regression (fitted with past outcomes) to calculate the probabilities?

Tim Calhoun is offline   Reply With Quote
Old
05-26-2017, 09:29 AM
  #70
morehockeystats
Unusual hockey stats
 
morehockeystats's Avatar
 
Join Date: Dec 2016
Location: San Jose
Country: United States
Posts: 241
vCash: 500
Quote:
Originally Posted by Doctor No View Post
Algorithm correctly picked both series.

Through three rounds:

WinnerProb
NYR54.9%
OTT40.4%
WAS82.3%
PIT48.4%
NAS41.3%
STL29.5%
ANA61.0%
EDM58.8%
NAS50.1%
ANA49.2%
OTT33.1%
PIT37.5%
PIT76.2%
NAS50.3%
TOTAL7.13 series (out of 14)
I had 1/2 (8/14) by team Elo, but on the roster level I got 10/13 games in CF.

morehockeystats is offline   Reply With Quote
Old
05-30-2017, 07:44 AM
  #71
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
Quote:
Originally Posted by Tim Calhoun View Post
Do you use some sort of logistic regression (fitted with past outcomes) to calculate the probabilities?
For each team, I calculate the variance of their SRS statistic (since the SRS produces a point estimate for each game, you can see how off it was for each team).

If you then make the (admittedly simplifying, but not particularly bad) assumption that each team's SRS is normally distributed, then the difference of the two is also normally distributed, and the probability of a win lines up with where that distribution crosses zero (after including home ice advantage).

Doctor No is offline   Reply With Quote
Old
05-30-2017, 07:47 AM
  #72
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
After game of May 29:

OutcomeProb
PIT in 4 games16.5%
PIT in 5 games26.1%
PIT in 6 games18.7%
PIT in 7 games16.8%
NAS in 7 games10.1%
NAS in 6 games8.6%
NAS in 5 games3.2%
PIT wins78.1%
NAS wins21.9%

Doctor No is offline   Reply With Quote
Old
05-31-2017, 10:33 PM
  #73
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
After game of May 31:

OutcomeProb
PIT in 4 games27.4%
PIT in 5 games31.7%
PIT in 6 games17.0%
PIT in 7 games12.6%
NAS in 7 games7.3%
NAS in 6 games4.1%
PIT wins88.7%
NAS wins11.3%

Doctor No is offline   Reply With Quote
Old
06-04-2017, 08:56 AM
  #74
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
After game of June 3:

OutcomeProb
PIT in 5 games31.4%
PIT in 6 games25.3%
PIT in 7 games21.2%
NAS in 7 games12.9%
NAS in 6 games9.2%
PIT wins77.9%
NAS wins22.1%

Doctor No is offline   Reply With Quote
Old
06-05-2017, 10:09 PM
  #75
Doctor No
Registered User
 
Doctor No's Avatar
 
Join Date: Oct 2005
Posts: 6,684
vCash: 50
After game of June 5:

OutcomeProb
PIT in 6 games30.5%
PIT in 7 games30.3%
NAS in 7 games19.4%
NAS in 6 games19.8%
PIT wins60.8%
NAS wins39.2%

Doctor No is offline   Reply With Quote
Reply

Forum Jump


Bookmarks

Thread Tools

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off



All times are GMT -5. The time now is 07:49 PM.

monitoring_string = "e4251c93e2ba248d29da988d93bf5144"

vBulletin Copyright ©2000 - 2017, Jelsoft Enterprises Ltd.
HFBoards.com is a property of CraveOnline Media, LLC, an Evolve Media, LLC company. 2017 All Rights Reserved.