tag:blogger.com,1999:blog-38600807.post5141006971668136443..comments2008-12-07T11:43:50.490-05:00Brian Burkenoreply@blogger.comBlogger11125tag:blogger.com,1999:blog-38600807.post-29772214901909026642008-12-07T11:43:00.000-05:002008-12-07T11:43:00.000-05:00I have spent exhaustive amounts of time building a predictive model for NBA games and I can tell you that pulling season-long stats typically gives me no more accurate results than using only those results from within a recent window of games. For the NBA, the most accurate window seems to be the last 20 games played. Proportionally speaking, that is nearly identical to using the last 4 games of an NFL season.<BR/><BR/>Statistically, this is counter-intuitive, since we usually understand that larger sample size equals greater accuracy. But from a sports perspective, this might make sense, because teams change and evolve so rapidly. If an NFL team was blown out in Week 1 of this season, does that really have any validity with regard to this week's numbers? I could make a cogent argument that it does not. <BR/><BR/>Grabbing a smaller, more recent sample size (like the last 4 games) also would make the model more responsive (too responsive??) to streaks and recent trends. For example, you currently have Washington as a slight favorite over Baltimore. I don't think this would be the case if you were only compiling data from the last 4 or 6 weeks. Washington, while certainly being a decent team, is still living off some of the statistical glory from their solid start to the season. Meanwhile, Baltimore seems to be picking up steam.<BR/><BR/>All this being said, I have found that in the NBA, compiling data on the last 20 games, while more accurate than using season-long stats, is only MARGINALLY more accurate. This seems to be consistent with your findings from last year when you placed double weight on the 4 most recent games. Yes, it's better (more accurate), but not by much, and there is a great value (elegance) in simplicity.Adamhttp://www.blogger.com/profile/07433838358019660341noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-412013338041911432008-12-05T09:46:00.000-05:002008-12-05T09:46:00.000-05:00Brian-Good point. Yes, I have. Last year's system included double weights for the 4 most recent games. Using that method, accuracy improved by a game, maybe 2. So there is evidence, but it's a question of balance.<BR/><BR/>I decided not to do that this year because one of my goals is simplicity. If a model is so complex that no one can understand it, it loses its usefulness. Behind all the regression mumbo-jumbo, my model is pretty straight-forward: Team efficiency stats plus HFA are weighted according to how they correlate with winning, adjusted for previous opponent strength, then converted using a logarithm into probabilities.<BR/><BR/>One other thing I might test in the future is to <I>only</I> consider recent-week performance. My model seems no more accurate in the later weeks than it did in the early weeks, despite the smaller sample size.Brian Burkehttp://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-50187160362367893742008-12-05T09:21:00.000-05:002008-12-05T09:21:00.000-05:00Brian,<BR/><BR/>Did you ever consider adding more weight to recent performance? <BR/><BR/>Correct me if I'm wrong, but I believe your system puts equal weight on all weeks. So in predicting outcomes for week 14, week 1 is just as important as week 13. Do you think or is there any evidence that a team's recent performance is more indicative of future performance than a a team's performance earlier in the season?Brianhttp://www.blogger.com/profile/15394006910997218733noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-9800107282751947912008-12-04T19:53:00.000-05:002008-12-04T19:53:00.000-05:00I think part of the difference is that I'm counting both no-bets as correct. If you check the comments on those posts, you'll see how the model worked.<BR/><BR/>You're right about the percentage group. The accuracy of a probability model is not just in its % correct, but also in its calibration. For example, in 60/40 games, I <I>want to be wrong</I> 40% of the time.Brian Burkehttp://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-70359877208335123532008-12-04T19:17:00.000-05:002008-12-04T19:17:00.000-05:00Our numbers are very close, I show this year at 95-47-1 with 2 no bets (50/50). That is 145 total games. I also show that there were 145 games this year from week 4 to 13 in my system data base. The differences are not worth double checking week by week to resolve. I must have an error of two in my records.<BR/><BR/>My 2007 record is from week 4 to 16 (not 17..error in previous post) shows 128-62 with 2 no bets.<BR/><BR/>What I find amazing is how good good your predictions are by percentage group.Jarheadnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-49528780731664007312008-12-04T17:35:00.000-05:002008-12-04T17:35:00.000-05:00No, I haven't done any comparisons of 3-4 vs 4-3 defenses to date.Brian Burkehttp://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-66041792722095324802008-12-04T17:33:00.000-05:002008-12-04T17:33:00.000-05:00Thanks, jarhead. My numbers seem to differ from yours slightly.<BR/><BR/>2007 was about 72% correct. I can't remember the exact decimal. I tracked it and touted it publicly every week, which I'm trying not to do this year. I can't recall, but that might exclude week 17 (which is notoriously unpredictable)--which might explain any discrepancy.<BR/><BR/>This year is 98-47-1 for a 68% accuracy so far.<BR/><BR/>Although it doesn't seem like the system is doing well as last year, keep in mind there are more upsets in some years than others. I use the Vegas consensus picks as my benchmark for accuracy. The opening line is only 66% correct so far this year, and the updated line at kickoff is not much better at 67% correct.<BR/><BR/>That's all I expect, that on average, I'll be slightly ahead of the consensus favorites.Brian Burkehttp://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-52368147832261709062008-12-04T16:57:00.000-05:002008-12-04T16:57:00.000-05:00Here is my analysis which covers week 4 to 17 2007 and week 4 to 13 2008 and is in terms of the home team.<BR/><BR/>Prediction<BR/>Range Avg W L W% <BR/>90-99 93.0 21 3 87.5<BR/>80-89 84.1 42 8 84.0<BR/>70-79 74.2 18 20 47.4<BR/>60-69 64.6 35 16 68.6<BR/>50-59 54.2 20 17 54.1<BR/>40-49 45.2 14 21 40.0<BR/>30-39 34.2 14 22 38.9<BR/>20-29 25.5 9 19 32.1<BR/>10-19 15.1 4 17 19.0<BR/> 0-9 7.7 4 8 33.3<BR/>TOTAL 181 151<BR/>AVG 55.7 54.5<BR/><BR/>CORRELATION BY GROUP .904<BR/><BR/>Another way to look at it is it is correct about 67% of the time.<BR/><BR/>I hope the columns come out lined up correctly and readable.jarheadnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-90954014374104369922008-12-04T16:05:00.000-05:002008-12-04T16:05:00.000-05:00BUF game is in TOR.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-1663977670425640242008-12-04T14:57:00.000-05:002008-12-04T14:57:00.000-05:00Have you done any analysis on the correlation between defensive schemes (ie. 3-4 vs 4-3) impacts defensive rankings?snowfamily1http://www.blogger.com/profile/16858971015645742915noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-16263710329376571072008-12-04T10:49:00.000-05:002008-12-04T10:49:00.000-05:00Can you post how accurately your model has predicted game outcomes?snowfamily1http://www.blogger.com/profile/16858971015645742915noreply@blogger.com