tag:blogger.com,1999:blog-31545676.post6427622495882155000..comments2019-05-30T10:58:55.325-04:00Comments on Sabermetric Research: Why the 2016 AL was harder to predict than the 2016 NL Phil Birnbaumhttp://www.blogger.com/profile/03800617749001032996noreply@blogger.comBlogger4125tag:blogger.com,1999:blog-31545676.post-85956748750250881082016-11-06T20:10:31.784-05:002016-11-06T20:10:31.784-05:00Phil--I've got something that may fit in BTN, ...Phil--I've got something that may fit in BTN, but I'm not set up to use outlook, so I can't access your email address...could you email me at dcoffin@iun.edu? Thanks.Don Coffinhttps://www.blogger.com/profile/07198988872512792834noreply@blogger.comtag:blogger.com,1999:blog-31545676.post-79324832216303487752016-10-24T23:18:22.298-04:002016-10-24T23:18:22.298-04:00Agreed that projection systems will use player pro...Agreed that projection systems will use player projections. But if last year's standings are compressed, even after adjusting for Pythag and Runs Created luck, probably the player performances were "too close" to the mean. So, widening your SD over what the normal system will give you might be a good idea.<br /><br />It all hinges on how confident you are that the standings won't be that compressed.<br /><br />Of course, if you're using more than one year of player data for your current projections, last year being compressed won't affect your forecasts that much, and you may not have to regress away from the mean much (if at all).<br /><br />BTW, doesn't FiveThirtyEight use player projections initially, but then just use game results (ELO) to adjust their forecasts as the season progresses? Or maybe not, I haven't looked too deeply.<br /><br />Phil Birnbaumhttps://www.blogger.com/profile/03800617749001032996noreply@blogger.comtag:blogger.com,1999:blog-31545676.post-14541744150338939612016-10-24T23:13:14.051-04:002016-10-24T23:13:14.051-04:00Hi, MGL,
I think that "reverse regressing&qu...Hi, MGL,<br /><br />I think that "reverse regressing" CAN work, if you know (or have a very good idea) that the true talent is wider than the observed standings.<br /><br />Imagine that you have two teams, one you know has 91 game talent and another you know has 71 game talent, but you don't know which is which. Team A goes 88-73, and team B goes 73-88. Your best estimate is to regress both AWAY from the mean towards 91 and 71 wins. (Or, 90.9 and 71.1, or whatever the Bayesian calculation gives you.)<br /><br />Phil Birnbaumhttps://www.blogger.com/profile/03800617749001032996noreply@blogger.comtag:blogger.com,1999:blog-31545676.post-42503369499661937582016-10-24T22:52:36.711-04:002016-10-24T22:52:36.711-04:00Couple of comments Phil. Good article by the way. ...Couple of comments Phil. Good article by the way. Any projection system that uses data such as last year's w/l records, is garbage. Any credible system is going to use player projections along with playing time projections. Of course last year's w/l records won't matter except to the extent that they reflect real parity, which they should regardless of the w/l records of the previous or subsequent years (to be fair, those provide somewhat of a pre and post Bayesian probability of there being parity in the middle year.) unless the third order wins have a normal SD. That's because 3rd order wins always trump regular wins (more or less).<br /><br />Also, I am pretty sure that "reverse regressing" w/l records is always wrong and will always give you a worse projection. You ALWAYS want to take last year's w/l and regress them toward the mean regardless of how small the SD is. If you get a better projection by regressing away from the mean, it is just a fluke. It seems like you suggest in the article that it was "correct" to regress away from the mean when the SD of wins is small.<br /><br />MGL<br /><br />Anonymousnoreply@blogger.com