Josh W
August 08 2013 02:20PM

If you've been following my few posts here you will know that I am studying Machine Learning for my masters in Computer Science and I am using those techniques and applying them to hockey analytics. I have found some pretty interesting stuff.  Some of my first few posts I wrote here (namely this one and this one) were based off some of my academic research. I have since turned those into an academic paper that was accepted to a Sports Analytics paper.

The conference it will be presented at is the European Conference for Machine Learning. Specifically at the Sports Analytics and Machine Learning workshop on Friday, 27 September in Prague, Czech Republic.

If you are near Prague at the end of sepember and want to talk shap, drink a Pilzen and check out a local KHL/Czech hockey game, then send me a message via twitter. I will be there for the week.

I am a Van Fan in Bytown. Living in Ottawa for work, I research Sports Analytics and Machine Learning at the University of Ottawa. I play hockey as well as a timbit but I compete in rowing with hopes of 2016 Olympic Gold. Follow me on twitter at @joshweissbock and feel free to give me a shout.
#1 Jacob
August 09 2013, 04:00AM

If I understand you, you're saying that 76% of the games played in an NHL season are decided by luck, rather than the relative skill of the teams playing, and it shouldn't be possible to predict game outcomes with greater than 62% accuracy.

Are you assuming that a team's skill level remains constant throughout the season, so that any time two teams meet, Team A will have the same probability of winning?

If we're talking about forecasting a team's win% at the start of a season then maybe 62% accuracy is the best we can do, but if we're predicting results on a game-by-game basis, I think the theoretical limits should be higher.

A team's skill level in a given game will vary from their start-of-season skill due to injuries, starting goalie, trades, home ice advantage, back-to-back games, etc. These factors all have a predictable impact and shouldn't really be classified as luck.

If we're just talking about preseason predictions, then a lot of this stuff either balances out (home ice advantage) or can't be foreseen (injuries), but I think it's misleading to lump it all together with PDO and call it luck. There's a lot that influences the outcome of an NHL game that's neither raw skill nor random chance.

Sorry if I've misunderstood you.

#2 Baalzamon
August 11 2013, 01:39PM

@Josh W

I think beloch was referring to the fact that you said you were constructing "several classification models... in order to build a classification model"

I had to read the sentence he picked out a few times before I figured out what he was referring to.

#3 beloch
August 10 2013, 05:21AM

"We construct several classiﬁcation models with novel features such as possession and luck in order to build a classiﬁcation model. "

This is in your *abstract*. The body of your paper contains even worse offenses. I know it's too late for corrections but, in the future, you should make use of your contacts on thenationnetwork for proofreading. They're almost certainly interested in the material and some of these guys know how to write. Even bookofloob could probably help you out! I know that, as a discipline, CPSC doesn't place much value on journals but, given the quality of your content, you can probably get into better publications if you improve your presentation even just a little bit!

#4 saucey
August 09 2013, 05:55PM

Cool beans. It would be great to see these results roll over an 82 game season. Let us know what beef the audince brings up at the conference.