|
Post by The Bofa on the Sofa on Sept 30, 2014 14:34:13 GMT -5
Massey is not RPI, nor an approximation of it. It has it's own uses, but RPI itself is a separate issue.
|
|
|
Post by The Bofa on the Sofa on Sept 30, 2014 14:48:01 GMT -5
I know a few people on the board like those Massey ratings, but they are utterly unconvincing to me. They don't seem to have either the volleyball-specific theoretical basis or the volleyball-specific experimental confirmation that pablo has. Massey is just ELO Chess. It's legit from a game theory standpoint, but has no volleyball component. Ballicora was an interesting take on the ELO approach, treating sets (games, back then) as discrete events as opposed to matches. Set wins are absolutely more informative than simple match wins. I agree with mikegarrison (naturally) in that I like the fact that Pablo is built upon the volleyball theoretical basis. Although I initially imposed that on the system, probably the most impressive aspect to me has been the extent to which I have been able to confirm it. The analysis I did last year with match pairs (two matches between the same teams) and how the outcome of the second match is related to the first was so much fun, especially to see the data fit right into that integrated normal distribution as I was using. And we have learned a ton from it, in terms of assessing match outcomes. For example, the lack of premium on winning was something that I didn't necessarily expect. No matter how much we want to believe otherwise, knowing who won just doesn't matter much (there is possibly a tiny, tiny correction, which I do include) if you know how many points were scored. That's the kind of stuff I've learned. Those who follow baseball SABR and Bill James know the discussion about his pythagorean model, to predict wins based on run differential. It's handy-dandy and all, but it's completely empirical. I have always said, one thing that happened with Pablo is that I discovered the point/win relationship for volleyball. The pablo underlying model tells you how point scoring translates to wins. Oh, I'm sure Gil Fellingham probably knew that inherently, but I don't know if it was ever really formalized. The beauty is that it even works with Gil's monte carlo data. That was cool, too.
|
|
|
Post by mikegarrison on Sept 30, 2014 15:03:48 GMT -5
I know a few people on the board like those Massey ratings, but they are utterly unconvincing to me. They don't seem to have either the volleyball-specific theoretical basis or the volleyball-specific experimental confirmation that pablo has. Massey is just ELO Chess. It's legit from a game theory standpoint, but has no volleyball component. That's why I added the "volleyball-specific" bit to my comment. I realized that the way I originally wrote it suggested Massey had no theoretical basis at all, and that's not true. As far as I can tell, though, Massey was developed and tested to do college football, which is entirely different than volleyball even if people do insist on saying that "the setter is like the quarterback of the volleyball team." The football version, for instance, seems to treat offense and defense as independent teams -- which is valid for football but not for volleyball. It certainly appears that the volleyball ratings on there are simply W/L with some accounting for HCA. Massey does list (and apparently use) individual HCAs instead of taking the pablo approach of solving for a generic HCA. IIRC, you tested that approach with pablo but decided not to use it?
|
|
|
Post by mikegarrison on Sept 30, 2014 15:08:11 GMT -5
Those who follow baseball SABR and Bill James know the discussion about his pythagorean model, to predict wins based on run differential. It's handy-dandy and all, but it's completely empirical. I have always said, one thing that happened with Pablo is that I discovered the point/win relationship for volleyball. The pablo underlying model tells you how point scoring translates to wins. Oh, I'm sure Gil Fellingham probably knew that inherently, but I don't know if it was ever really formalized. The beauty is that it even works with Gil's monte carlo data. That was cool, too. The James formula (like any empirical fit) breaks down when the data it is using does not match the data it was derived from. In particular, it doesn't work very well with teams like this year's Seattle Mariners which both score and allow significantly fewer runs than average.
|
|
|
Post by Cubicle No More ... on Sept 30, 2014 15:16:14 GMT -5
Massey is not RPI, nor an approximation of it. It has it's own uses, but RPI itself is a separate issue. yup
|
|
|
Post by The Bofa on the Sofa on Sept 30, 2014 15:19:10 GMT -5
Massey is just ELO Chess. It's legit from a game theory standpoint, but has no volleyball component. That's why I added the "volleyball-specific" bit to my comment. I realized that the way I originally wrote it suggested Massey had no theoretical basis at all, and that's not true. As far as I can tell, though, Massey was developed and tested to do college football, which is entirely different than volleyball even if people do insist on saying that "the setter is like the quarterback of the volleyball team." The football version, for instance, seems to treat offense and defense as independent teams -- which is valid for football but not for volleyball. It certainly appears that the volleyball ratings on there are simply W/L with some accounting for HCA. Massey does list (and apparently use) individual HCAs instead of taking the pablo approach of solving for a generic HCA. IIRC, you tested that approach with pablo but decided not to use it? Not enough data, I don't think. And including a HCA for everyone was going to be way too many variables. I did a 5 year average for all the teams in the B1G once. It's not enough teams to be able to really assess, but the variation was certainly within the realm of a binomial distribution, and with a huge year-to-year standard deviation (something like ± 200 Pablo Points; IOW, 2/3 of the teams had HCAs between 0 and 400). The data were way too random to be able to justify giving any team an individual HCA so out of line with the average. IIRC, with the five year average, the team with the highest HCA was Wisconsin, around 400 maybe? But that would be because they had 5 above average HCAs for 4 of the 5 years, or something like that. No one else stood out, but I remember that Purdue was below average, overall, but it was because they had one season where the HCA was something like -400 (it was the season they had a couple bad non-conference losses at home to start the season). So their average HCA was like 120 or so (average was more like 200 back then). You see things like this, and it makes you question whether you are measuring something real. Did Purdue really have a below average home court those years (or a home court disadvantage) or did they just play worse in home matches for other reasons? Most likely the latter. Playing at home wasn't hurting their performance (if they had played like that on the road, they would have lost worse, iow)
|
|
|
Post by The Bofa on the Sofa on Sept 30, 2014 15:21:48 GMT -5
Those who follow baseball SABR and Bill James know the discussion about his pythagorean model, to predict wins based on run differential. It's handy-dandy and all, but it's completely empirical. I have always said, one thing that happened with Pablo is that I discovered the point/win relationship for volleyball. The pablo underlying model tells you how point scoring translates to wins. Oh, I'm sure Gil Fellingham probably knew that inherently, but I don't know if it was ever really formalized. The beauty is that it even works with Gil's monte carlo data. That was cool, too. The James formula (like any empirical fit) breaks down when the data it is using does not match the data it was derived from. In particular, it doesn't work very well with teams like this year's Seattle Mariners which both score and allow significantly fewer runs than average. Doesn't work? Or just a bad data point? The Pythagorean formula works mostly equally well with high scoring or low scoring environments. It just has a lot of variation to it. Personally, I'd say that is probably the bigger reason for it's failure with Seattle. There is more variability in low-scoring environments because of the larger number of 1-run games, which account for most of the variation.
|
|
|
Post by vbnerd on Sept 30, 2014 15:22:16 GMT -5
... except, no, we do not all have Rich Kern subscriptions. 1) Shame on you 2) I know there are other RPI-estimators out there, even for volleyball that will suffice (I saw a reference recently) We should be able to filter VT posts to screen out all of the neanderthals who think they are volleyball fans. Give Rich $25 and you get to know everything about volleyball. How is that a difficult decision?
|
|
|
Post by mikegarrison on Sept 30, 2014 15:24:43 GMT -5
You see things like this, and it makes you question whether you are measuring something real. I hear you there! If you add enough knobs and factors to your model, you can closely match a set of randomly generated data. Of course, that's meaningless for predicting the next point.
|
|
|
Post by The Bofa on the Sofa on Sept 30, 2014 15:28:38 GMT -5
You see things like this, and it makes you question whether you are measuring something real. I hear you there! If you add enough knobs and factors to your model, you can closely match a set of randomly generated data. Of course, that's meaningless for predicting the next point. "With four parameters I can fit an elephant, and with five I can make him wiggle his trunk." - von Neumann "By this point, that elephant is doing a jig" - my old adviser, talking semi-empirical theory
|
|
|
Post by mikegarrison on Sept 30, 2014 15:30:37 GMT -5
The James formula (like any empirical fit) breaks down when the data it is using does not match the data it was derived from. In particular, it doesn't work very well with teams like this year's Seattle Mariners which both score and allow significantly fewer runs than average. Doesn't work? Or just a bad data point? The Pythagorean formula works mostly equally well with high scoring or low scoring environments. It just has a lot of variation to it. Personally, I'd say that is probably the bigger reason for it's failure with Seattle. There is more variability in low-scoring environments because of the larger number of 1-run games, which account for most of the variation. Doesn't work as well, for exactly the reason you state.
|
|
bluepenquin
Hall of Fame
4-Time VolleyTalk Poster of the Year (2019, 2018, 2017, 2016), All-VolleyTalk 1st Team (2021, 2020, 2019, 2018, 2017, 2016)
Posts: 12,398
|
Post by bluepenquin on Sept 30, 2014 16:06:14 GMT -5
Those who follow baseball SABR and Bill James know the discussion about his pythagorean model, to predict wins based on run differential. It's handy-dandy and all, but it's completely empirical. I have always said, one thing that happened with Pablo is that I discovered the point/win relationship for volleyball. The pablo underlying model tells you how point scoring translates to wins. Oh, I'm sure Gil Fellingham probably knew that inherently, but I don't know if it was ever really formalized. The beauty is that it even works with Gil's monte carlo data. That was cool, too. The James formula (like any empirical fit) breaks down when the data it is using does not match the data it was derived from. In particular, it doesn't work very well with teams like this year's Seattle Mariners which both score and allow significantly fewer runs than average.Are you sure this is true? I know that the Jame's pythagorean is old and there are better things like true runs pythagorean, but there were times in baseball history (1960's, 1910's, 1970's) where the Mariners total runs this year would be 'normal'. And there wasn't a breakdown of the results for those 'deadball' periods. Runs scored in A's games were much more than the Mariners and I think above average and the A's had a much higher variance in actual wins. Now that is cherry-picking, but I just didn't know we would see more variances from lower scoring (batting and pitching) teams.
|
|
|
Post by mikegarrison on Sept 30, 2014 16:18:18 GMT -5
Maybe I'm wrong. But when I see empirically derived correlations not matching well for data that seems different in some important way from the original data, I get really suspicious about sweeping the variance under the rug of random error.
|
|
|
Post by BeachbytheBay on Sept 30, 2014 16:20:35 GMT -5
so if Massey is ELO Chess, doess that make RPI ELO Checkers??
|
|
|
Post by The Bofa on the Sofa on Sept 30, 2014 16:45:55 GMT -5
so if Massey is ELO Chess, doess that make RPI ELO Checkers?? Interestingly, ELO works equally well for chess AND checkers. It's just that no one ever got around to trying to rank Checkers Grand Masters.
|
|