Bracketology 2023

[ursusminor]

The model also doesn't take into account how bad the refs are. :-D

[marty]

The model also doesn't take into account how bad the refs are. :-D

It also doesn't take into account the entertainment value for those of us watching BL's response.  Past performance in some cases does predict future outcomes.

[adamw]

Been waiting 10 years for someone to work with me on a better model. I ask every year. Every time someone makes the slightest gesture toward that end, they disappear again in a day or two. Meanwhile, just complaints. ... I don't see the same amount of righteous outrage directed at polls. Talk about something truly meaningless.

Please by all means - let's write a better model.

[CU77]

Better models already exist, eg Massey ratings: https://masseyratings.com/ch/ncaa-d1/games

His method is proprietary but the results seem reasonable.

[adamw]

Better models already exist, eg Massey ratings: https://masseyratings.com/ch/ncaa-d1/games

His method is proprietary but the results seem reasonable.

Where do you see predictive models there?

Also Cornell is 20th there - be careful what you wish for.

[adamw]

Better models already exist, eg Massey ratings: https://masseyratings.com/ch/ncaa-d1/games

His method is proprietary but the results seem reasonable.

Where do you see predictive models there?

Also Cornell is 20th there - be careful what you wish for.

I do see a page on there where you can "sim" a best-of-3 series ... so I simmed one between Cornell and Quinnipiac - and Cornell won 13 out of 15 best-of-3 series between the teams. So ... I'd say he's got some work to do.

[ugarte]

Better models already exist, eg Massey ratings: https://masseyratings.com/ch/ncaa-d1/games

His method is proprietary but the results seem reasonable.

Where do you see predictive models there?

Also Cornell is 20th there - be careful what you wish for.

I do see a page on there where you can "sim" a best-of-3 series ... so I simmed one between Cornell and Quinnipiac - and Cornell won 13 out of 15 best-of-3 series between the teams. So ... I'd say he's got some work to do.

sounds right to me

[BearLover]

Better models already exist, eg Massey ratings: https://masseyratings.com/ch/ncaa-d1/games

His method is proprietary but the results seem reasonable.

The probabilities on this page seem eminently reasonable to me.

[shafer]

Been waiting 10 years for someone to work with me on a better model. I ask every year. Every time someone makes the slightest gesture toward that end, they disappear again in a day or two. Meanwhile, just complaints. ... I don't see the same amount of righteous outrage directed at polls. Talk about something truly meaningless.

Please by all means - let's write a better model.

I think we all recognize that it takes some level of mathematical ability to be able to improve upon the current model in a way that is theoretically defensible, and not based on some kind of fudgery. My only suggestion would be to treat each team's KRACH value as a random variable (likely non-Gaussian) with the current KRACH representing the median of that variable. Then you could probably run a smaller Monte Carlo simulation and get better numbers.

[Trotsky]

Better models already exist, eg Massey ratings: https://masseyratings.com/ch/ncaa-d1/games

His method is proprietary but the results seem reasonable.

The probabilities on this page seem eminently reasonable to me.

Extrapolating those probabilities out over best of 3 likely gives you nearly the same probabilities as the playoffstatus page.

[CU77]

KRACH probabilities can be improved by adding to every team's record some number of ties against a fictitious "average" team; Ken Butler's original method added one such game, but more gives a better distribution of predicted results.

[adamw]

I've heard all of these suggestions before. On this forum, and elsewhere. I've encouraged anyone with ideas to reach out to me and please feel free to improve the algorithm. A couple people in the last 10 years have gestured to take me up on that offer, only to fade away pretty quickly.

So again - if anyone has an idea - please feel free and come to me. The door is open.  I only went to that school in the ghetto part of Ithaca, not your fancy Ivy League place filled with math and physics PhDs.  So instead of complaints, please come to me with a better system.  Not a vague sentence or two -- the actual math. That I can code.

[CU77]

Well, I would start with KRACH, add some number x of ties against a fictitious team to each team's record, compute KRACH ratings at the end of a regular season for various values of x, then compute predicted records of each team in an all-against-all set of games, and then compare the distribution of the computed winning percentages against the actual distribution of winning percentages in the games actually played, and then adjust x to get the best match.

But this would be a LOT of work for a marginal improvement of predictions that no one cares about anyway.

Better: a ton of work has gone into making such predictions for basketball. Find out the the state of the art there, and adapt for hockey.

[Robb]

Yes, a better model would be a ton of work.  What if Wayne Gretzky, Jr was on a team, but injured for the first 20 games of the season and the team went .500.  Since he got back, he's averaging 2.5 points per game and the team is undefeated.  Where would your "better" model rank them headed into the playoffs?  Maybe the team hired a better strength and conditioning coach halfway through the season?   What if your star players also care about academics and the team always tanks during exam weeks....and there are midterms the week of the quarterfinals?

Models will never capture the richness and complexity of life.  Once you accept that fact, you can paradoxically enjoy looking at models again, because you are content in your certainty that it only tells a part of the story and just agree with yourself to leave it at that.

[upprdeck]

the model might capture the change in stuff/roster but if they really were playing better then they should be scoring more or winning more and it should see that?  

The next level of modeling is figuring out the hot teams and how long they stay hot vs play at the level they should be at.