tag:blogger.com,1999:blog-38600807.post9149885708839352694..comments2023-11-05T04:16:44.937-05:00Comments on Advanced Football Analytics (formerly Advanced NFL Stats): Team Rankings Week 4Unknownnoreply@blogger.comBlogger30125tag:blogger.com,1999:blog-38600807.post-63703476676552674482011-10-03T01:33:13.355-04:002011-10-03T01:33:13.355-04:00how do you calculate predictions? never understood...how do you calculate predictions? never understood how.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-14851005157318065332011-10-02T08:08:55.531-04:002011-10-02T08:08:55.531-04:00Sorry for the previous "test" message bu...Sorry for the previous "test" message but I have been struggling to post on this forum for several days i finally gfigured out it is my computer nit accepting cookies form this site.<br /><br />Anyway my comment....<br /><br /> <br />I use Pointspread= 9*ln(odds) for college football and pointspread= 7*ln(odds) for NFL which reflects the higher scores in college. (This also works in Basketball)<br /><br /> <br /><br />I use this as the basis for my ranking system which I won't plug on this site but which is based solely on scores and venue rather than game stats .<br /><br />So I would convert Oakland 56% NE 44% to 7*ln(56/44)= 1.68 points rather than Tom's 2.<br /><br />Tom out of interest which source did you use to generate the 8.5?<br /><br />One word of warning is that the graph of ln(odds) vs pointspread does go a bit kinked around 3 points particularly in the NFL reflecting the frequent occurence ofthis score but i cant find a simple way of ironing this out.<br /><br />Interestingly totals have very little effect on this relationship, whereas in theory a 3 point favourite in a game predicted to be high scoring should be less likely to win than a 3 point favourite in a low scoring game.<br /><br />Prediction tracker tracks both straight up record and against the closing Vegas line and he does include some systems which don't have pointspreads although it is unclear how he generates pointspreads for them.James Mhttp://www.pointshare.webs.com/noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-11366774403905084652011-10-01T16:49:15.930-04:002011-10-01T16:49:15.930-04:00Sorry for the confusion. I know how to score a log...Sorry for the confusion. I know how to score a login model. The difficulty lies is that only about .1% of football fans would understand. <br /><br />And yes. That's exactly what I'm doing: discounting the the effect of hr-hitters.Brian Burkehttps://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-46109263344298257142011-10-01T16:01:12.738-04:002011-10-01T16:01:12.738-04:00Incidentally, by completely ignoring YPC aren'...Incidentally, by completely ignoring YPC aren't you discounting the value of home-run hitter RBs? I'm sure Oakland was happy that McFadden broke off a 60-yard run last Sunday. His YPC is higher than 14 team's YPA (with sacks in the denominator), so handing the ball off to him is practically the same as passing (so far).Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-72027924432250946822011-10-01T15:54:52.794-04:002011-10-01T15:54:52.794-04:00To score your model:
Let p1,...,pn be the probabi...To score your model:<br /><br />Let p1,...,pn be the probabilities you placed on the outcome that in fact occurred.<br /><br />Your score is the average of the logs of p1...pn. This will be a negative number.<br /><br />To translate this into something that's easier to interpret (i.e. a likelihood), exponentiate this score to get back to a number between 0 and 1 (sort of a percentage accuracy).Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-72279461151334215002011-09-30T11:02:14.575-04:002011-09-30T11:02:14.575-04:00That's a good idea. I have done a little playi...That's a good idea. I have done a little playing around with that but could do a lot more. I looked at only performance for qtrs 1-3 as a predictor, and it was slightly less effective as using all 4 qtrs. <br /><br />You don't want to get too tricky with restricting data. There's a trade off with sample size obviously. Plus, if you start under-weighting or excluding some data, you effectively over-weight other data. The last thing you'd want to do is over-weight clutch situations.Brian Burkehttps://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-7552719544582326502011-09-30T10:08:50.503-04:002011-09-30T10:08:50.503-04:00Have you ever considered restricting the data to c...Have you ever considered restricting the data to certain subsets that might be more predictive? <br /><br />For example restricting the data to "normal" football. In blowouts teams may get into a run-run-pass on 3rd and long pattern that might skew both their run SR and their pass efficiency. Likewise on defense they may play a "prevent" that gives up a lot of underneath yards, hurting their defensive pass efficiency.<br /><br />Another idea might be to restrict the data to only include plays between the 20's. I don't know the numbers but I'm guessing that inside the red zone both pass efficiency and run SR decrease on offense and likewise increase on defense. Since teams often end up with different average starting field positions due to either non-predictive events (like turnovers or special teams) or predictive events (like the play of their own defense) some teams may spend a disproportionate amount of time playing on a shortened field and see their efficiencies skewed as a result.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-2880904021282903392011-09-29T22:45:30.173-04:002011-09-29T22:45:30.173-04:00Garret-Very perceptive! The difference is due to h...Garret-Very perceptive! The difference is due to how I originally coded SR. I rounded all plays to 2 digits. The 2% (or so) are plays in which the EPA was exactly 0.00. To count as a 'success' I said EPA had to be > 0, and not >= 0.Brian Burkehttps://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-51344880445532918522011-09-29T22:18:20.563-04:002011-09-29T22:18:20.563-04:00I also thought median was the key. But the problem...I also thought median was the key. But the problem with median YPC is that nearly every team has the same YPC...3.Brian Burkehttps://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-36077732810732350832011-09-29T21:39:17.852-04:002011-09-29T21:39:17.852-04:00I've long thought that per-play metrics should...I've long thought that per-play metrics should use quantile statistics (e.g. median) rather than the mean. How hard would it be to check if median yards per carry correlate better than mean yards per carry to winning?Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-50027661185283362352011-09-29T20:52:53.223-04:002011-09-29T20:52:53.223-04:00ORUN% = 39%
DRUN% = 59%
39% + 59% = 98%, which doe...ORUN% = 39%<br />DRUN% = 59%<br />39% + 59% = 98%, which does not equal 100%<br />Does this mean that 2% of run plays are neither successful not unsuccessful? <br />Or does this have something to with the way the averages are computed (each team is an equal 1/32 regardless of how often they run)?Gerrit D.noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-34832569596912750042011-09-29T19:09:53.077-04:002011-09-29T19:09:53.077-04:00ubrab-
I have posted in the comments section of th...ubrab-<br />I have posted in the comments section of the probabilities link an equation that can calculate the spread given the probability of a team winning, however I will post this week's here:<br /><br />Dallas over Detroit by 6.8<br />Chicago over Carolina by 4.5<br />Buffalo over Cincinnati by 0.3<br />Tennessee over Cleveland by 2.4<br />Kansas City over Minnesota by 1.4<br />Washington over St. Louis by 1.7<br />New Orleans over Jacksonville by 3.8<br />Houston over Pittsburgh by 5.6<br />Philadelphia over San Fran. by 6.4<br />Arizona over Giants by 0.7<br />Seattle over Atlanta by 3.1<br />San Diego over Miami by 2.7<br />Oakland over New England by 2<br />Green Bay over Denver by 9.3<br />Baltimore over Jets by 4.9<br />Tampa Bay over Indianapolis by 3.1Tomnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-31429179106989554192011-09-29T17:40:48.854-04:002011-09-29T17:40:48.854-04:00Brain - Would it be post what the old method (usin...Brain - Would it be post what the old method (using YPC versus SR) would produce for win probabilities? I'm very curious to see what kind of a difference there would be. Thanks so much for having this amazing site!Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-9504386462565593972011-09-29T15:06:10.948-04:002011-09-29T15:06:10.948-04:00"Your intuitive estimates of team strength ar..."Your intuitive estimates of team strength are far less accurate than you imagine. The thing is, you’ll forget how wrong you were by the end of the season, and re-wire your memory to trick yourself into believing you ‘knew it all along.’ We all do."<br /><br />Not all of us...I predicted Carolina to win the Super Bowl last year...I remember it very well. <br /><br />The year before I was feeling high and mighty because I predicted Minnesota would walk all over Dallas because the NFL East was crappy. <br /><br />A good honest memory helps to keep one humble.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-36630890868673252142011-09-29T10:34:50.958-04:002011-09-29T10:34:50.958-04:00PredictionTracker does not track games against the...PredictionTracker does not track games against the spread. It tracks how well computer systems pick winners straight up. For reference it shows how often the Vegas favorite wins. Very few systems perform better than Vegas each year, and none has demonstrated a record of doing so consistently year after year. We should be skeptical until proven otherwise that your system is any different.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-13868473489760403802011-09-29T10:11:52.765-04:002011-09-29T10:11:52.765-04:00Brian Thanks again for all your work here and for ...Brian Thanks again for all your work here and for really describing how you derive your models. <br />First I saw that D-int was not included in your original model and is now, I was wondering if there was a reason behind that. <br />Second, I was wondering if you had any kind of analysis on how much of a difference the offensive and defensive line make for each team. I see that you have the WPA for each teams lines, and it seems like they have at some some correlation on winning. I was wondering if you had any equation that model the WPA of the lineman vs winning percentage or anything along those lines. <br />Third, when looking at QB ratings I noticed that the EPA for your site was much higher than that used for ESPN'S QBR rating, presumably because they incorporate what percentage of a play's success is based upon the QB compared to other players. This means that your model values player with a lot of talent around them, like Phillip Rivers, and ESPN's model values those with less talent, like Matt Hasselbeck. I was wondering what your thoughts are on that. <br />Thank, DanDan Whitneyhttps://www.blogger.com/profile/04640343956910419780noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-8679680830845320362011-09-29T09:41:29.906-04:002011-09-29T09:41:29.906-04:00I'll try to answer these in reverse order.
-I...I'll try to answer these in reverse order.<br /><br />-I'm surprised by DAL too. Fumbles are all non-ST fumbles (not fumbles lost) per run+pass. Not sure how to reproduce over/unders. I think bobbled snaps are considered aborted plays rather than fumbles. Not sure about this though.<br /><br />-I'll be happy to post the coefficients. Don't have them here right now. If I forget bug me next week in the comments.<br /><br />-Yes, PIT int rate will regress upward. Accept it. They will not finish the season with zero ints.<br /><br />-Very funny, Mike. The secret adjustments for teams I don't like work slightly differently. You know your logit regression though!<br /><br />-PredictionTracker is a great site. I love that someone actually tracks all this stuff. However, because I don't predict against the spread, there's no place for me on PT.<br /><br />-I'm out of the business of publishing the track record of the model. I leave that to independent sources. Overall, I'm very pleased with the results. However, be aware that there are ways of making things look much better or worse than reality. How do you treat a tie? What if the model says .50/.50. Is that a loss for the model or is the game thrown out? If the game is down to the wire, isn't that a + for the model? Do you compare the model vs the closing Vegas line or the opening line? Do you compare it week vs week or overall?<br /><br />-I also noticed the teams are packed closer together this year. It's partly due to the method I chose to regress the team stats early in the season. The other part is due to way teams have played so far this year. Once there is more data, the stats will be regressed less, and depending on how teams play the top teams might be back up to .70 GWP.Brian Burkehttps://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-20929110014443278492011-09-29T04:46:52.255-04:002011-09-29T04:46:52.255-04:00Brian,
I'm surprised by the Cowboys O ranking...Brian,<br /><br />I'm surprised by the Cowboys O ranking - Watching only their last game, they were globally relying on big plays and had a lot of fumble luck which brings my question of how you handle fumbles (do you separate fumbled snaps from "normal" ones ?) - What else am I missing ?<br /><br />On another topic, someone from the comments use to take your game predictions and generate over/unders - Where is he gone ? Or, how can I reproduce that ?ubrabnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-83110864492744235822011-09-29T03:45:07.973-04:002011-09-29T03:45:07.973-04:00Brian,
Can you post the new coefficients?Brian,<br />Can you post the new coefficients?Patricknoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-74284884580640576132011-09-29T03:35:19.652-04:002011-09-29T03:35:19.652-04:00why would the steelers pass defense regress? They ...why would the steelers pass defense regress? They played the two teams who are literally last in O PASS, and a Ravens team that is just ok in the passing game. I imagine that will go up a good amount after the Houston game, even with a pick or 2. Something tells me you're just not the real Brian Burke and just a Steelers fan mouthing off as usualJackhttp://yourmom.comnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-86881355865445709792011-09-29T00:54:46.023-04:002011-09-29T00:54:46.023-04:00Brian - Didn't you forget to mention the follo...Brian - Didn't you forget to mention the following term:<br /><br />if team==ATL then logit = logit - 0.50Michael Beuoyhttps://www.blogger.com/profile/03960600491528993233noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-13892633936702767952011-09-28T23:35:03.853-04:002011-09-28T23:35:03.853-04:00I know it's been suggested here before, but so...I know it's been suggested here before, but someone should submit Brian's ratings to thepredictiontracker.com to see how it stacks up against other computer rating systems.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-47095788983553932932011-09-28T22:36:13.718-04:002011-09-28T22:36:13.718-04:00I think it would be a good idea to show your model...I think it would be a good idea to show your model's accuracy over the past few years. There will always be people who come in and question your results, but showing the model's track record would reduce some of this.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-55458594902991997392011-09-28T18:51:54.621-04:002011-09-28T18:51:54.621-04:00Hey Brian, any thoughts on why there isn't an ...Hey Brian, any thoughts on why there isn't an early dominate team? The GWP for the top 2 teams is 0.66 this year, but it was 0.72, 0.71; 0.80, 0.71; 0.78, 0.72 the past 3 years (using Week 4 data). Something to do with the inclusion of rushing success rate? A one-year fluke? Or is it just because the Patriots by far league-best offense is paired with a near league-worst defense? <br /><br />Also, is a low DRUN% good? I can't tell looking at the rankings.Jameshttps://www.blogger.com/profile/01838293735141324662noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-85263970199149660592011-09-28T17:59:22.406-04:002011-09-28T17:59:22.406-04:00Hey, remember 2 weeks ago when the Steelers D was ...Hey, remember 2 weeks ago when the Steelers D was too old? They were washed up?<br /><br />Already they've got 4.7 net YPA allowed on defense, best in the league by far. And they haven't even had an int yet. That will regress. Bet on it.Brian Burkehttps://www.blogger.com/profile/12371470711365236987noreply@blogger.com