tag:blogger.com,1999:blog-38600807.post8549881493009965611..comments2023-11-05T04:16:44.937-05:00Comments on Advanced Football Analytics (formerly Advanced NFL Stats): Homemade Sagarin RatingsUnknownnoreply@blogger.comBlogger20125tag:blogger.com,1999:blog-38600807.post-24236286392234517702013-12-11T14:27:51.457-05:002013-12-11T14:27:51.457-05:00It would be interesting to pit Excel vs. Sagarin w...It would be interesting to pit Excel vs. Sagarin weekly, to see how much improvement you get from Sagarin's proprietary tweaks. Phil Birnbaumhttps://www.blogger.com/profile/03800617749001032996noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-38815265478667709152013-11-07T14:01:06.644-05:002013-11-07T14:01:06.644-05:00That excel model came from the work of Wayne Winst...That excel model came from the work of Wayne Winston, who is very good friends with Sagarian, I would say that model is very close to what Sagarian uses. Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-6676451439126446142013-05-07T19:24:07.440-04:002013-05-07T19:24:07.440-04:00Has anybody run past performances through a regres...Has anybody run past performances through a regression model? Is elo for sports anywhere as accurate as for chess?<br /><br />Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-41419369022652156712013-02-24T12:29:58.816-05:002013-02-24T12:29:58.816-05:00has anyone come up with a better way to use solver...has anyone come up with a better way to use solver to come up with better ratingsAnonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-10748264535368293122011-11-05T22:55:20.306-04:002011-11-05T22:55:20.306-04:00According to Wayne Winston's "Mathletics&...According to Wayne Winston's "Mathletics", Sagarin uses a "proprietary weighted least squares algorithm." That would probably explain a decent amount of the difference between your results and his.<br /><br />One cautionary tidbit for Solver ... I was trying to use Solver to analyze college offenses and defenses, and I gave Solver a lot of leeway with my ratings. I ended up with Wisconsin having the most prolific passing offense in the Big Ten (and, conversely, one of the worst rushing offenses).Jeff H.http://www.twitter.com/3YardsAndACloudnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-16976849425983596802011-04-13T10:49:38.193-04:002011-04-13T10:49:38.193-04:00Just implemented this with the ols function of PDL...Just implemented this with the ols function of PDL::Stats (a Perl module). Not the same as Microsoft Excel, I know, but linear least squares is pretty stable. However, occasionally this method just blows up. SRS has its moments as well (as in the matrices that result are so often singular).Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-20204911878619837462010-12-20T00:10:07.428-05:002010-12-20T00:10:07.428-05:00If you look at Saigon's 2010 Dec 19's pure...If you look at Saigon's 2010 Dec 19's pure point ratings you will see that Ten is rated below Chi. Yet, Ten had a +26 DIFF while Chi had a +25 DIFF and Ten had a harder schedule. Therefore it appears that Saigon does weigh more recent games more heavily (Ten was on a big losing streak w/o Vince Young).Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-1506867321270147192010-02-12T19:23:32.277-05:002010-02-12T19:23:32.277-05:00Win probability=normsdist(Spread/12). The Spread/1...Win probability=normsdist(Spread/12). The Spread/12 generates a t-score, and the normsdist function measures the area under the curve. A 1-point favorite should win about 53.33%. For the favorite Spread is a positive number. It's negative for an underdog. Normsdist(Spread/12)+Normsdist(-spread/12)=1.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-27285180132330109352009-03-10T20:18:00.000-04:002009-03-10T20:18:00.000-04:00Sagarin's ratings should estimate the point spread...Sagarin's ratings should estimate the point spread pretty well. You can use the Pythagorean formula with an exponent of 10 to estimate a win probability. You also need an estimate of an average score, which I'm not sure where to find.Brian Burkehttps://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-22131011796731826272009-03-10T18:01:00.000-04:002009-03-10T18:01:00.000-04:00I plan to use a random number generator and Sagari...I plan to use a random number generator and Sagarin ratings and/or RPI ratings for a March Madness Bracket. My only issue is this: How can I use the ratings to come up with an estimated winning percentage to plug into the random number generator? I am statistically inclined but this element has always escaped me.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-6631137091026684812009-03-01T00:39:00.000-05:002009-03-01T00:39:00.000-05:00I noticed that the ratings Sagarin has at the end ...I noticed that the ratings Sagarin has at the end of season are not the same ones he has starting the next season. Any idea why they change and how he gets the numbers to start the new season.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-40833330019733574432008-08-31T11:50:00.000-04:002008-08-31T11:50:00.000-04:00For this, I compared with pure points. For the fol...For this, I compared with pure points. For the follow-on Elo post, I compared with EloChess.Brian Burkehttps://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-49189445154689698722008-08-31T09:56:00.000-04:002008-08-31T09:56:00.000-04:00When you did the correlation analysis of the Homem...When you did the correlation analysis of the Homemade ratings to the Sagarin ratings, did you use Sagarin's combined rating (Pure Points & EloChess) or just one of them. In picking games against my friend, I've been using Pure Points for the last few years.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-79723729792180487292008-05-11T03:50:00.000-04:002008-05-11T03:50:00.000-04:00I'm kind of wondering if something like this would...I'm kind of wondering if something like this would be a bit more accurate estimation:<BR/><BR/>X = Record % - .5<BR/>Y = SOS % - .5<BR/>Z = X + Y + .5<BR/><BR/>Playing with the numbers it seems to be a decent estimation as well.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-65337075381611671542008-05-11T02:16:00.000-04:002008-05-11T02:16:00.000-04:00Sorry, I wasn't very clear and picked a bad exampl...Sorry, I wasn't very clear and picked a bad example to illustrate. Here's how it would work for your example above.<BR/><BR/>A .6 record would be 3:2 or 1.5 odds ratio. A .5 SoS would be a 1:1 or 1.0 odds ratio.<BR/><BR/>ln(1.5) + ln(1.0) = 0.405 + 0 = 0.405<BR/><BR/>e^(.405) = 1.5<BR/><BR/>1/(1+(1/1.5)) = .6<BR/><BR/>So a .6 win% against a .5 SoS yields a .6 "true" win%.<BR/><BR/>Elo's algorithm works in much the same way. It uses a base 10 log instead of base e, and Elo is based on an (arbitrary) average rating of 1500 with a SD of 200. <BR/><BR/>(By the way, I just realized that instead of ln(x) + ln(y), you can just do ln(x*y). They're the same. 8th grade is slowly coming back to me!)Brian Burkehttps://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-49296669758141823582008-05-11T01:40:00.000-04:002008-05-11T01:40:00.000-04:00Expected Result = 1/(1+10^(Elo Difference/400)) - ...Expected Result = 1/(1+10^(Elo Difference/400)) - http://remi.coulom.free.fr/Bayesian-Elo/<BR/><BR/>On the calculations, I don't think I understand something.<BR/><BR/>.6 record against .5 SOS.<BR/><BR/>.6/.5 = 1.2:1<BR/><BR/>ln(1.2)&(0.833)= 0.1823+ -0.1823 = 0<BR/><BR/>e^(0) = 1<BR/><BR/>1/(1+1/1) = 0.500<BR/><BR/>Did I make a mistake in my calculations? I increased the SOS, but the final result was still 0.5Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-28509731120420041212008-05-11T01:11:00.000-04:002008-05-11T01:11:00.000-04:00Ya Elo is supposed to give you the probability tha...Ya Elo is supposed to give you the probability that A can beat B. The whole premise behind it has been the assumption that if A > B > C then A > C. Which means teams don't have to play each other directly to get an inference of relative strength.<BR/><BR/>Elo programs automate the calculations performances in chess tournaments.<BR/><BR/>As far as my method, the thought was that it would show that performance against a 50% opponent. However, I did not do any testing to see if that was the case.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-17598221491838048032008-05-11T00:09:00.000-04:002008-05-11T00:09:00.000-04:00Fixed. Thanks.Not familiar with EloSTAT, but sound...Fixed. Thanks.<BR/><BR/>Not familiar with EloSTAT, but sounds interesting. I'm finishing putting together my post on Elo now.<BR/><BR/>I think the method you propose would be a great approximation, but what does the final rating mean?<BR/><BR/>I'd suggest this: <BR/><BR/>Take the odds ratio of each team's win% and each team's SoS (avg opp win%). So a .500 team would be 1:1 and a .600 team would be 3:2.<BR/><BR/>Then take the natural log of the odds ratios and add them together. Now you have the natural log of the odds ratio of a team's "true" win% accounting for opponent strength. <BR/><BR/>Solve for the odds ratio by e^(log odds ratio). Then solve for the winning percentage for the resulting odds ratio which would be =1/(1+(1/x)), where x is the odds ratio.<BR/><BR/>Example:<BR/><BR/>A team with a .600 win% and a .400 avg opponent win%. The odds ratios are 3:2 and 2:3.<BR/><BR/>3:2=1.5 and 2:3=.67<BR/><BR/>ln(1.5) = .405<BR/>ln(.67) = -.405<BR/><BR/>.405 + (-.405) = 0<BR/><BR/>e^0 = 1<BR/><BR/>1/(1+1/1) = 0.500<BR/><BR/>The .600 team with a .400 strength of schedule is really a .500 team.<BR/><BR/>Your method would give .6 * .4 * 2 = .48, which is a really good approximation.Brian Burkehttps://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-85969569188109770682008-05-10T21:32:00.000-04:002008-05-10T21:32:00.000-04:00It looks like you have SF and SEA mixed up on your...It looks like you have SF and SEA mixed up on your ratings.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-70398357967532799512008-05-10T19:07:00.000-04:002008-05-10T19:07:00.000-04:00I used to use EloSTAT or Bayesian Elo (allows you ...I used to use EloSTAT or Bayesian Elo (allows you to offset home field advantage) for Elo calculations. Create a pgn file with game results (1-0 or 0-1), and run it through the program. <BR/><BR/>Alternatively I had an idea about using strength of schedule and win % to do something similar. <BR/><BR/>Win % x Strength of Schedule x 2<BR/><BR/>50% x 50% (average opp) x 2 = Rating of 50%<BR/><BR/>60% x 60% (average opp) x 2 = 72%<BR/><BR/>What do you think of that method?Anonymousnoreply@blogger.com