If you need a laugh

[bleuwolfe]

Preseason computer rankings

http://www.compughterratings.com/2008PreseasonTop25.htm

#1 OK 13-0
#2 FL 13-0
#3 OH 12-0
#6 Auburn 11-2
#11 LSU 10-2
#12 UGA 9- 3

Well at least we'll be spared seeing OH getting rolled again in the NC.

[USCrebel]

Preseason computer rankings

http://www.compughterratings.com/2008PreseasonTop25.htm

#1 OK 13-0
#2 FL 13-0
#3 OH 12-0
#6 Auburn 11-2
#11 LSU 10-2
#12 UGA 9- 3

Well at least we'll be spared seeing OH getting rolled again in the NC.

I would put much store in this site.

[RebelFIL]

This guy is using very basic statistical analysis to predict upcoming results. "least squares" with "maximum likelihood" utilizies historical averages (recent wins and loses) to determine the most likely outcome. For example, if the temperature in Oxford over the past 4 years on August 12 was 98, 88, 94, and 92, then you might predict the temperature next year to be around 93 degrees (the computerranking analysis is slightly more sophisticated, but not much more).

Although, this is a fairly safe way to make predictions (it will be nearly right much more than it will be way off), it is not even the accepted statistical method for making predictions. Theoretical reasoning must be accounted for. Historical record is certainly a partial predictor, but so are # of returning starters, head coach's record (see ASU), schedule strength, etc.

Back to the weather example: in meterology, pressure, wind, and humidity measurements are used to predict rainfall and temperature change, along with historical almanac rainfall and temperature records. That is because it is understood that almanac records do not actually cause rain and temperature, but low pressure in front of high pressure, and heavy moisture actually do cause rain and temperature fluctuation. Just like, historical win/loss records do not cause a winning or losing season.

Just my 2 cents. Although I do not have a bachelor's and master's in math from Va. Tech, I am at Ole Miss working toward a master's and PhD in which statistical analysis is utilized to analyze research data.

[USCrebel]

This guy is using very basic statistical analysis to predict upcoming results. "least squares" with "maximum likelihood" utilizies historical averages (recent wins and loses) to determine the most likely outcome. For example, if the temperature in Oxford over the past 4 years on August 12 was 98, 88, 94, and 92, then you might predict the temperature next year to be around 93 degrees (the computerranking analysis is slightly more sophisticated, but not much more).

Although, this is a fairly safe way to make predictions (it will be nearly right much more than it will be way off), it is not even the accepted statistical method for making predictions. Theoretical reasoning must be accounted for. Historical record is certainly a partial predictor, but so are # of returning starters, head coach's record (see ASU), schedule strength, etc.

Back to the weather example: in meterology, pressure, wind, and humidity measurements are used to predict rainfall and temperature change, along with historical almanac rainfall and temperature records. That is because it is understood that almanac records do not actually cause rain and temperature, but low pressure in front of high pressure, and heavy moisture actually do cause rain and temperature fluctuation. Just like, historical win/loss records do not cause a winning or losing season.

Just my 2 cents. Although I do not have a bachelor's and master's in math from Va. Tech, I am at Ole Miss working toward a master's and PhD in which statistical analysis is utilized to analyze research data.

Actually, there would be no reason to use iterative approaches am assuming you are talking about non-linear LS... I am too lazy to look at the website myself to see how he does it.) The fact is that there is so little variability in college football among the 120 Division One teams that looking at division one W/L records over a five year period and along with historical head to head opponent records should give you as predictive a measure of W/L as going through the other methods.

For instance, in any given year, I can assume that Ole Miss will lose to Alabama and Auburn. Oddly, our record against LSU in the last twenty years is fairly close and we play Arkansas close. Our other two "permanent" foes, MSU and VAN are typically wins for us. I could "predict" a three and three record against these six every year and would usually to correct.

The performance of athletes is, I believe, homoscedastic, so teams that get higher performing recruits are more likely to have more higher performing athletes. I don't think you need to know the definition of homoscedasticity to understand and "predict" that. In the long run, teams that recruit well and have traditions of 7, 8, or 9 wins, will likely have 7, 8, or 9 wins. Barring Marshall-like plane crashes, knowing what you did last year and how good your athletes are is my best predictor of your season.

BTW, I don't have a bachelor's or master's degree from VaTech, either, but I have a PhD with a minor in statistics and measurement and I taught undergraduate statistics for 10 years.

[done]

HOMOSCEDASTIC?

Come on ObiWan...now you're just making stuff up 8)

[RebelFIL]

Right on. As I said, when using the prior win/loss approach, you will be close more often than way off, which is why it is heavily used. It is easy and often close, if not luckily accurate. However, a prior win/loss method leaves a lot of error variance on the table. If number of returning starters and number of touted recruits are important, why not include those measures in the analysis; it will most likely account for more of the variance, and thus provide a more accurate prediction. When all put into the regression equation, the beta coefficient of win/loss will reduce (rightfully so), but so will the error term. Further, if win/loss and talent rating where truly one in the same (homoscedastic), then Wake Forest could not win and Florida State, Miami, and Clemson could not lose (staying in the ACC only).

Thanks for your input - fun conversation for the nerds on the forum.

[RebelFIL]

Homoscedastic = the property of having equal statistical variances

definition from www.merriam-webster.com

It is an important concept in prediction. For an exagerrated example, if color of helmets was highly associated with number of wins (blue = 6, maroon = 4, orange = 9), then one could use color of helmet to predict the upcoming year's results just as well as using prior number of wins. This is, of course, rediculous.

[bbqit]

I wish they would let vegas rank the teams.

[rebeloke]

Hell, a broke clock is right twice a day... 8O

[USCrebel]

Right on. As I said, when using the prior win/loss approach, you will be close more often than way off, which is why it is heavily used. It is easy and often close, if not luckily accurate. However, a prior win/loss method leaves a lot of error variance on the table. If number of returning starters and number of touted recruits are important, why not include those measures in the analysis; it will most likely account for more of the variance, and thus provide a more accurate prediction. When all put into the regression equation, the beta coefficient of win/loss will reduce (rightfully so), but so will the error term. Further, if win/loss and talent rating where truly one in the same (homoscedastic), then Wake Forest could not win and Florida State, Miami, and Clemson could not lose (staying in the ACC only).

Thanks for your input - fun conversation for the nerds on the forum.

You are correct... sort of.

In the short message, I listed only two variables that have predictive ability that are likely (somewhat) homoscedastic when taken over time. Wake Forest is an excellent example of a program that has improved its record as it has improved its recruiting (along with improved coaching). My original post could have included coaching, facilities, and a few others. Problematic to all of this is the ecological nature of the variables... it's not just good recruits, favorable schedule, good coaching, etc., but all of these at the same time.

Still, I think that as long as I had the same information about each team I could make a fairly accurate prediction. More precisely, if I knew ONLY the coach's life time W/L record, I would be able to predict W/L records more often than not. Same with the institutions record, athletes, etc. Bringing them all together in some sort of bizarre SEM/LISREL model would obviously be more precise, but we've already bored the masses with this.

[done]

If I had a DeLorean, a flux capacitor, and 1.21 gigiwatts of electrical power I could outdo all of you stat guys using the big words to impress the girls(not) 8)

But I don't. I do however have ESPN Game Plan and a 52" Samsung HDTV, and it is a lot more fun to watch them play for 15 weeks or so and then spend the remaining 37 arguing about it 8)

[USCrebel]

If I had a DeLorean, a flux capacitor, and 1.21 gigiwatts of electrical power I could outdo all of you stat guys using the big words to impress the girls(not) 8)

But I don't. I do however have ESPN Game Plan and a 52" Samsung HDTV, and it is a lot more fun to watch them play for 15 weeks or so and then spend the remaining 37 arguing about it 8)

That's kinda my point. In the long run, a good fan who really follows the season will be able to make reasonably accurate predictions when compared to a computer.

[1OLEMISSREBEL]

And how and where do you factor in the previous season coach???? Would you even consider the "O" a factor here????

[USCrebel]

And how and where do you factor in the previous season coach???? Would you even consider the "O" a factor here????

Are you kidding. A guy with as lifetime record of 10-25 vs. a guy with a lifetime record of 110-71. HDN also has a reputation for winning close games as often as losing them. Meaning that of the four relatively close games we had last year (AL, AU, LSU, FLA) we probably would have won at least two of them. Moving from 3-9 to 5-7.

To me that means each year of the Orgeron era, when we have 4 or 5 close losses, we would have taken home at least two or three of them. In 2006, that likely would have amounted to a 7-5 record and a bowl game, which would have increased our recruiting profile and may have resulted in a better 2007 record.

[bigtreydaddy]

My head is swimming... I think I oughta lay down...