Polls

[marty]

Which rating system has most accurately predicted D-1 college hockey game winners so far this season?

Planning a flyer on DraftKings?

[Trotsky]

14 in USA.

[French Rage]

Which rating system has most accurately predicted D-1 college hockey game winners so far this season?

Your mom.

[osorojo]

If current rating systems are no better at predicting outcomes of upcoming hockey games than my Aunt Beulah is why not dig a hole, cover it with leaves and grass, and TRAP someone to predict outcomes of upcoming college hockey games?

[BearLover]

If current rating systems are no better at predicting outcomes of upcoming hockey games than my Aunt Beulah is why not dig a hole, cover it with leaves and grass, and TRAP someone to predict outcomes of upcoming college hockey games?

As I wrote on the prior page, the intent of "current rating systems" is not to predict the outcomes of hockey games.

[Trotsky]

If current rating systems are no better at predicting outcomes of upcoming hockey games than my Aunt Beulah is why not dig a hole, cover it with leaves and grass, and TRAP someone to predict outcomes of upcoming college hockey games?

If by rating we mean PWR, RPICH, KRACH, the intent is to measure teams' relative prior performance to assess where they should be ranked for NC$$ selection.  NC$$ hockey has a problem: there are tight sets with extremely good relative rankings (the conferences) but then only loose bounds between the sets, making relative ranking of members of different sets difficult.  Metrics like PWR are an attempt to normalize all members against a common baseline.  They are similar to the sabermetrics that attempt to adjust for differences in ballpark effects or era effects in MLB.

But those assessments only imply potential predictive likelihoods of future events.  One would assume a team ranked higher in the metric will, on balance, beat a team ranked lower over a sufficiently large sample, but particularly in single elimination playoffs you get one data point, so at best you are talking about a probability function.  There is also not great confidence that prior results will predict future outcomes because team composition changes with injuries, and teams themselves tend to become better (or worse) as the season goes on.

It's the littany of reasons why social science correlation coefficients tend to be a lot lower than natural science.  It's the nature of working with humans and not leptons, and of complex systems rather than simple machines.

[Swampy]

If current rating systems are no better at predicting outcomes of upcoming hockey games than my Aunt Beulah is why not dig a hole, cover it with leaves and grass, and TRAP someone to predict outcomes of upcoming college hockey games?

I tried this. It doesn't work very well.

[marty]

If current rating systems are no better at predicting outcomes of upcoming hockey games than my Aunt Beulah is why not dig a hole, cover it with leaves and grass, and TRAP someone to predict outcomes of upcoming college hockey games?

I tried this. It doesn't work very well.

Red beartrap?

[osorojo]

Swampy: I'll let you borrow MY Aunt Beulah for a couple of predictions, you lucky dog. Don't worry about kickback from cranky posters. They aren't any better at predicting hockey game winners than my Aunt Beulah. [Maybe that's why they're so cranky?]

[Trotsky]

OK, I'm convinced.  Somebody has put an AI/ML project here to pass the Turing Test.

It's the Townhall cadence and Ben Garrison humoriness that gave it away.

[abmarks]

Swampy: I'll let you borrow MY Aunt Beulah for a couple of predictions, you lucky dog. Don't worry about kickback from cranky posters. They aren't any better at predicting hockey game winners than my Aunt Beulah. [Maybe that's why they're so cranky?]

OK, I'm convinced.  Somebody has put an AI/ML project here to pass the Turing Test.

It's the Townhall cadence and Ben Garrison humoriness that gave it away.

I tried. Chatgpt didn't help me respond to oso...


"I am a language model AI and I do not have the ability to predict any hockey games, neither I can borrow physical entities such as your Aunt Beulah or get kickback or interact with people or posters. My main function is to assist users in providing information, answering questions and generating human-like text based on the input provided to me."

[marty]

What criteria are used to construct college hockey rankings, and which ranking system has most successfully predicted the winners of contests between ranked teams?

Your sentence with the highest perplexity is:
What criteria are used to construct college hockey rankings, and which ranking system has most successfully predicted the winners of contests between ranked teams?

It has a perplexity of:
76
GPTZero has finished analyzing your text!


Your GPTZero score corresponds to the likelihood of the text being AI generated:
More data may be needed to determine if your text is human or AI generated. Try inputting more text.

[marty]

What criteria are used to construct college hockey rankings, and which ranking system has most successfully predicted the winners of contests between ranked teams?

Your sentence with the highest perplexity is:
What criteria are used to construct college hockey rankings, and which ranking system has most successfully predicted the winners of contests between ranked teams?

It has a perplexity of:
76
GPTZero has finished analyzing your text!


Your GPTZero score corresponds to the likelihood of the text being AI generated:
More data may be needed to determine if your text is human or AI generated. Try inputting more text.

There is also another similar site.

[osorojo]

Thanks, for your your response to my question, Marty. Your explanation leaves little doubt about the methodology or accuracy of college hockey rankings.

[jtwcornell91]

If current rating systems are no better at predicting outcomes of upcoming hockey games than my Aunt Beulah is why not dig a hole, cover it with leaves and grass, and TRAP someone to predict outcomes of upcoming college hockey games?

If by rating we mean PWR, RPICH, KRACH, the intent is to measure teams' relative prior performance to assess where they should be ranked for NC$$ selection.  NC$$ hockey has a problem: there are tight sets with extremely good relative rankings (the conferences) but then only loose bounds between the sets, making relative ranking of members of different sets difficult.  Metrics like PWR are an attempt to normalize all members against a common baseline.  They are similar to the sabermetrics that attempt to adjust for differences in ballpark effects or era effects in MLB.

But those assessments only imply potential predictive likelihoods of future events.  One would assume a team ranked higher in the metric will, on balance, beat a team ranked lower over a sufficiently large sample, but particularly in single elimination playoffs you get one data point, so at best you are talking about a probability function.  There is also not great confidence that prior results will predict future outcomes because team composition changes with injuries, and teams themselves tend to become better (or worse) as the season goes on.

It's the littany of reasons why social science correlation coefficients tend to be a lot lower than natural science.  It's the nature of working with humans and not leptons, and of complex systems rather than simple machines.

You can analyze a bunch of predictions, though, and see if events you assigned a 60% probability to really happened 60% of the time.  E.g., FiveThirtyEight does this to check their election forecasts.  I wrote a paper (with Adam as a co-author) that, among other things, showed that using KRACH to assign probabilities for playoff games produced better predictions than cooking up a probability from the winning percentages, which in turn out-performed just calling every game a tossup: https://arxiv.org/abs/2001.04226