Ga. Tech has the most accurate computer rankings

[hifigator]

www.gatech.edu/newsroom/release.html?id=1802&ga=1


They were talking about this yesterday on the afternoon sportstalk show. This is pretty impressive. 30 out of the past 36 final four teams predicted over the past four years. Obviously most any power ranking system would have done pretty well this year, with all four #1 seeds making it to San Antonia. But their record seems to be quite a bit more impressive than that feat.

[SFHoya99]

It's a souped up Sagarin.

The article is also written in a biased manner. Georgetown was identified as "overrated" but by one seed -- and not predicted to lose to Davidson. And they act like the system predicted Duke would lose in the second round, but the system had them ranked FOURTH, ahead of UNC.

Also, Ken Pomeroy -- not mentioned -- did have Davidson as a 7 seed equivalent, before the tourney, if I remember. Not to say Ken Pom's system is better, because we're not seeing the regression, but my guess is that the Ga Tech rankings aren't significantly better.

In fact, they've only identified 30 of the last 36 as "one of the top two teams in their region" so they haven't even picked the Final Four -- they are picking eight teams every year as their Final Four!

So I'm sure there is good stuff there -- and stuff definitely not in Sagarin, who hasn't touched his formula in years, and probably not in Pomeroy. But it didn't compare itself to Pomeroy. I'd like to see that. It did do well against Sagarin by their standards.

[hifigator]

SF, where did you see that explanation? I didn't catch that link, but if that is correct then that would make their record much less impressive. I think 23 of 36 #1 seeds have advanced to the Final 4 in the past 9 years. So if they could pick 2 teams per region instead of 1, to make it to the Final 4, then that isn't really all that impressive. I would guess that there have probably been a half dozen #2 seeds to make the Final 4 over that period. If so, then simply picking the #1 and #2 seeds would have given you about the same number of correct picks.

As for comparison to Sagarin, I didn't dig into the formulas and compare the intricacies of each. I did email Jay Coleman who created the Dancecard program. He said that he was familiar with their model and that it was sound but that last year, his program performed better on a game by game basis, meaning that if the teams were ranked according to the respective programs and then the tourny was "played out" according to those rankings, his program worked better.

Where did you see that they picked 8 teams to make the Final 4 each year?

In simple terms, the 2 main differences that they stressed between thir program and most others, at least as I understand it, are:

1. Home court advantage isn't factored in as point differential, but rather as impacting the outcome -- in other words, how it affects the likelihood of winning the game. I'm not sure at all how they do this, but I do understand the logic.

and

2. They've determined that in very close games the "better" team only wins slightly more than half of the time, but the more the point differential, the more likely it is that the better team will win. That seems rather intuitive, but as far as their formula is concerned, they weight close games much less than others.

[SFHoya99]

Apr 10, 2008 13:19:20 GMT -5 hifigator said:

SF, where did you see that explanation? I didn't catch that link, but if that is correct then that would make their record much less impressive. I think 23 of 36 #1 seeds have advanced to the Final 4 in the past 9 years. So if they could pick 2 teams per region instead of 1, to make it to the Final 4, then that isn't really all that impressive. I would guess that there have probably been a half dozen #2 seeds to make the Final 4 over that period. If so, then simply picking the #1 and #2 seeds would have given you about the same number of correct picks.

They have the entire paper online. That article is a joke, though.

As for comparison to Sagarin, I didn't dig into the formulas and compare the intricacies of each. I did email Jay Coleman who created the Dancecard program. He said that he was familiar with their model and that it was sound but that last year, his program performed better on a game by game basis, meaning that if the teams were ranked according to the respective programs and then the tourny was "played out" according to those rankings, his program worked better.

Sagarin Predictor is a very basic relational margin of victory calc with a standard home and away on point basis. It sounds like the changed the latter and did some extra regression on the former, but not exactly breakthrough.

Where did you see that they picked 8 teams to make the Final 4 each year?

Two teams per regional = eight.

In simple terms, the 2 main differences that they stressed between thir program and most others, at least as I understand it, are:

1. Home court advantage isn't factored in as point differential, but rather as impacting the outcome -- in other words, how it affects the likelihood of winning the game. I'm not sure at all how they do this, but I do understand the logic.

That's interesting. I've been pestering Pomeroy to do his calculations with teams as team separate teams -- Home Georgetown and Away Georgetown, for example. A standard Home Winning % or Points works in the aggregate; it's inappropriate on a game by game basis.

2. They've determined that in very close games the "better" team only wins slightly more than half of the time, but the more the point differential, the more likely it is that the better team will win. That seems rather intuitive, but as far as their formula is concerned, they weight close games much less than others.

Who hasn't? This is the very fundamental basic belief behind a Sagarin, a Pomeroy or whatever. Bill James figured this out in baseball twenty five years ago.

I understand that mottled their regression a bit -- which is cool because Sagarin hasn't -- but it's hardly breakthrough.

[hifigator]

It's pretty obvious that the closer the score, the less likely it is that the "better" team wins, yet the human element in us always places added emphasis on winning the close games. Not that we aren't equally impressed with blowout wins, but think about how often we will hear someone say something like ".... record in one-run games" or "... record in games decided by 3 points or less." On one hand, it is certainly reasonable to grade a team based on how it performs when the chips are down, so to speak. On the other hand, it also makes sense to weight favorably the teams that avoid being in those situations the most because they are way ahead.


Lastly, I think that the one element that would improve Sagarin dramatically is an individual home court advantage ranking. I have seen thought frequently with college football team rankings for example, but the Sagarin uses a National advantage which is adjusted throughout the year. I know back when I was in school, that number was closer to 4, normally something like 3.86, but I think the past couple of years it has been closer to 3 -- something like 3.32 or so.

I think it would make sense to give Duke or Vandy a larger home advantage. Similarly, Rutgers probably shouldn't get as much of an adjustment.