Postseason Projections: Conference Round

We'd previously warned that the eventual winner of the NFC East was not to be underestimated in the postseason, and the Giants showed why in last Sunday's win over the Packers. By a slim margin, the model now sees the Giants as the strongest of the four remaining teams, though their advantage is lessened by the fact that they will have to meet the 49ers on the road. Overall, the model now gives New York about one chance in three to win the Super Bowl, odds on par with those of the Patriots.

None of the remaining teams is particularly dominant at this point and none is a complete long shot. San Francisco, with the lowest probability of a Super Bowl win, is still given a 15% chance. And now, submitted for your approval, the final postseason projections of the 2011 season, with the table below listing each team's percent probability—first of advancing to the Super Bowl and then of winning the whole thing. Enjoy.

Percent Probability to Advance
TeamSuper BowlSup Bowl Champion
NE6435
BAL3618
SF4115
NYG5932


  • Spread The Love
  • Digg This Post
  • Tweet This Post
  • Stumble This Post
  • Submit This Post To Delicious
  • Submit This Post To Reddit
  • Submit This Post To Mixx

98 Responses to “Postseason Projections: Conference Round”

  1. davearchie says:

    What is this projection model based on? By pure EPA, SF (25 offense, -50 defense) would seem to out-edge the Giants (118 offense, +80 defense), and the Patriots (246, +101) would be the strongest team remaining. In fact, with Baltimore at (61, -65), the Giants seem to be the worst team remaining by EPA. Are you giving extra weight to recent weeks, or is there some other formula at play here?

  2. James says:

    Dave, it's based on GWP, which itself is based on the efficiency model, which can be found here: http://www.advancednflstats.com/2012/01/end-of-season-team-rankings.html

  3. Josh Katz says:

    What James said, with the additional note that the GWPs are updated with results of the playoff rounds as they happen. In practice, this doesn't shift the values around too much--the GWPs of the four remaining teams are identical to their end-of-season values with the exception of BAL, whose GWP decreased slightly to .59 from .60.

  4. bravehoptoad says:

    Right, which means, of course, that the usual caveats about the not counting special teams apply. This is a particularly important caveat for a team like San Francisco, whose special teams have been a key part of their success all season.

  5. Anonymous says:

    I wonder if Brian's model poor performance in predicting San Francisco wins (42.8% from weeks four, when the first projects I could find on the Fifth Down blog were posted) should be accounted for at this point.

    While I understand why the model excludes special teams (and, IRC, turnover plus/minus), San Francisco outperformed the model beyond the +/-25% Brian has calculated "luck" to potentially skew any statistical model.

    Am I misunderstanding some element of the model, or does a team like SF break the predictive utility of the model?

  6. Anonymous says:

    I love how everyone wants to "account" for something that doesn't sit well with them.

    Are the concepts of luck and chance that hard to understand?

  7. Anonymous says:

    Yeah I'm sure ignoring special teams has no effect on anything.

    FWIW I think it's basically impossible for the sportsbooks to be off by this much at this point in the season. The model is almost certainly wrong on this game.

  8. Anonymous says:

    Someone do some significance testing and get to the bottom of it

  9. Brian Burke says:

    Regarding the model's "poor performance" with SF this season:

    #1 Some teams will see very good accuracy in any given year, and some will see poor accuracy. The model's accuracy with respect to one team is not terribly meaningful.

    #2 Yes. SF has enjoyed good ST play...to date. That explains past success. Whether it predicts future success is a different question.

    #3 Do you think Akers' future FG success is more likely to continue as it has so far this season, or is his future success most likely to regress to his career mean?

    #4 The real reason the model differs from the sportsbooks is that SF has been feasting on positive turnover ratios all year. Live by the turnover, die by the turnover. Turnover rates are far less consistent than most people think, including gamblers.

    #5 Some advice: don't crack on the model until it's actually wrong. Assuming it will be wrong as a basis of criticism is, well, not bright.

    #6 ST do matter. They have an impact. Rarely are they consistent enough, however, to have a predictable impact on a single game.

    #7 "Special teams" isn't a single dimension of play. It's six separate and independent facets of play: FG team, FG defense, kickoff, kickoff coverage, punt team, punt coverage. There are but a scant handful of any one type of each play in a game. In comparison, there are 60-80 passes, and 40-50 runs in a game.

    Hope that helps.

  10. coldbikemessenger says:

    in general the average team throws 16 int's a year and has 80 passes defenced.

    the 49'ers had 5 and 49

    that is luck

  11. coldbikemessenger says:

    smith had a 1.1 int rate this year.

    Bringing his career rate down to 3.0
    Do you really believe alex smith has turned the corner and will continue to have a lower int rate than Brady Brees and Rogers?

    or do you believe a bunch of picks were dropped and the 49'ers won the field position battle?

  12. Ben Stuplisberger says:

    Just judging by the Saints game, the model is not broken on SF. SF was outplayed. They were +4 in turnovers and they still barely won the game. They fumbled 3 times, recovered 2, NO did the same thing and lost all of them. Take back a turnover or two, bounce a ball a different way, NO wins easily. The better team that day lost; SF was very lucky. I remember thinking about the turnover ratio throughout the game and saying that NO is just that much better than SF to stay so close. Give NO that favorable of a turnover ratio, the score could have gotten out of hand, maybe 55-10? :)

  13. Jim Z. says:

    How can a team that has a negative regular season point differential, and a 9-7 record, possibly be favored to win the Super Bowl?

    In the *history* of the Super Bowl, a team has *never* been outscored in the regular season and even *made* the Super Bowl.

    How can the Giants be considered such a strong team if they barely squeaked into the playoffs with a 9-7 record and a negative regular season point differential? What am I missing here? What do these statistical models see in the Giants that ignores the fact that they struggled to even break .500 in the regular season?

  14. Anonymous says:

    Its really hard to tell the trolls from the idiots these days

  15. Ben Stuplisberger says:

    @ Anonymous... when in doubt, it's the latter.

  16. Jared Doom says:

    Jim Z.

    First, the projections are not based on point differential and wins/losses, because those statistics, in comparison to those used in the projection model, are not nearly as good at predicting future wins.

    Second, let's pretend for a moment you have a valid argument, and I'll use your sub-par logic on a counterargument. (a) The Giants just dominated a GB team with the best record and probably the biggest regular season point differential. (b) You could argue they are a different team compared to the regular season because of injured players returning. (c) They beat the team they would likely face in the SB (new england) in the regular season.

  17. Anonymous says:

    Niners rule. Your model is flawed unless you've factored in your "Sucking Giants Dick" into the model.

  18. Anonymous says:

    @Ben

    Just judging by the Saints game, the model is not broken on SF. SF was outplayed. They were +4 in turnovers and they still barely won the game.

    I think the point is that the failure to account for turnover differential explains, at least in part, the problem with the model as it pertains to the 49ers. There is a high correlation between a positive turnover differential (particularly a +4 differential) and victory. To completely ignore such a highly correlative statistic seems inherently flawed.

    Take back a turnover or two, bounce a ball a different way, NO wins easily.

    So, what you are saying is that if one of the stats most highly correlated to winning was altered, then the outcome of the game would have been altered? Thanks for the deep insight.

    The better team that day lost

    Better except when it comes to unimportant things like avoiding and causing turnovers.

  19. Anonymous says:

    I think the only question is whether the Giants' recent dramatic increase in defensive efficiency is a byproduct of the defense actually getting better (injured players healing, better coaching, etc.), or a statistical aberration that will regress to their season-long mean. If its the former, the 49ers are going to get crushed, if its the latter, it will be a close game at Candlestick.

  20. Anonymous says:

    "I think the point is that the failure to account for turnover differential explains, at least in part, the problem with the model as it pertains to the 49ers. There is a high correlation between a positive turnover differential (particularly a +4 differential) and victory. To completely ignore such a highly correlative statistic seems inherently flawed."


    Your point falls apart when you take into consideration that past turnover rate is very very very poorly correlated to future turnover rate (esp fumbles).

    Imagine that the entire game was just determined by the opening coin toss with a fair coin. Now lets say in their first 10 games the team wins 7 coin tosses and loses 3. Would you say going forward that the team has a 70% chance of winning? Of course not, its still a 50% shot, no matter what the past results.

  21. Jim Z. says:

    "First, the projections are not based on point differential and wins/losses, because those statistics, in comparison to those used in the projection model, are not nearly as good at predicting future wins."

    How good are point differential figures at predicting Super Bowl berths? If we look at the 45 year history of the Super Bowl, it seems that they are excellent at it:

    -67 of 86 (77.9%) Super Bowl teams in non-strike-shortened seasons had a regular season scoring differential of +100 or more.
    -86 of 90 (95.5%) Super Bowl teams in *all* years (strike-shortened seasons included) had a regular season scoring differential of +50 or more, with the 2008 Cardinals, 2007 Giants, 2003 Panthers, and 1979 Rams as the only teams with less. None of these teams had a negative scoring differential, and only one of those teams won the Super Bowl (2007 Giants) in what was considered the greatest upset in Super Bowl history. The 2011 Giants would be the first Super Bowl team in 45 years to post a negative scoring differential.

    95.5% of a certain population seems like a pretty strong indicator of correllation to me.

  22. Ben Stuplisberger says:

    @ anonymous: Turnovers are highly correlated to winning, yes. Repeatable and predictable, no. Fumbling a ball or not fumbling a ball may be a repeatable skill, recovering one is not. Making bad throws or good throws is repeatable, whether a defender catches your bad throw is mostly luck. And remember, despite the huge turnover differential, NO still loss the game by only 4 points. How many times this season did a team have a -4 turnover differential and stay within 4 points of their opponent? Once: Pro Football Reference

  23. Anonymous says:

    Brian:

    Regarding #4, you are either ignoring or simply don't understand how efficient the NFL betting market is these days. If your model is that far off from the lines then it's almost assuredly a problem with the model and not the lines.

    Either that or you should be a multimillionaire by now.

  24. srinivas says:

    Let's not pretend the 49ers are being unfairly maligned by this model. This isn't a hit job by Brian Burke, and the comments that make it seem as such can't be taken seriously.

    As for the model being totally off the sportsbooks -- Brian's model has always come up with very unique predictions, and it would be against tendency for that to change now. This is a model that had the Eagles as a top-5 team, that said Chicago had the best defense in the league, and that ranked Oakland's offense over the Detroit Lions. I don't think you could get a single sharp in Vegas to say any of those things, yet this model has been remarkably successful at predicting games. Lets give it some credit.

    Full disclosure, I am a Niners fan, and I do think the model may be underestimating the 49ers strengths (run defense, special teams, and ball-hawking pass defense). But hey, I'm not the one with the data set, and I'll take a spreadsheet over my gut every time. We'll see on Sunday (and possibly in Indy) if Brian's model has underestimating the team. Until then, from a 49ers fan to another 49ers fan, I suggest you stop hurling insults.

  25. Anonymous says:

    Wow, you guys don't take criticism of your model that well. I could just say something you would consider dumb and TRUE along the lines of numbers never correctly depicting what is happening on the field... instead I will elaborate.

    (1) There are three elements to football. You are completely ignoring one: Special Teams. This already makes any model off. Guess what a major element of the 49ers win was Saturday? Special Teams play... if you want to simplify it, since there are so many different elements, weight average starting field position on offense and defense. Sure the offense and defense play some part in that but special teams is the main element.
    (2) Turnovers are luck? there are a few lucky turnovers that deflect or bounce into a defenders hands... but they are not luck. People's arguments that the Saints outplayed the 49ers were lucky to have a +4 turnover ratio is a ridiculous statement. Each of those turnovers were forced just as they have been doing ALL SEASON.
    (3) If you want to adjust your model for the playoffs I would go ahead and put more emphasis on the defense. You have to be blind to not see that defenses generally step it up in the playoffs and are the main element of a winning team. defenses that show up to the playoffs are the teams that advance most often. So far these playoffs are a good indicator of that.
    (4) Yes, the Niners turnover +/- is ridiculous this year and I do not expect it to be the same next year. But it would not surprise me in the slightest if they are +15 to +20 next year b/c they play with a strategy to hold on to the ball. A lot of offensive minded fans see it as giving the ball away when in reality the defense is doing something to TAKE the ball away. Turnover differential should be one of the main factors in grading offenses and defenses because it is the most accurate indicator of who wins games.

    As always it's interesting seeing your guys take on things. Keep up the good work and keep on improving the model.

    Mike

  26. Anonymous says:

    Mike, on your 2nd point - that goes directly against the research that Brian has done. You can't predict future turnovers based on past turnovers to an accurate degree. I'd love to see your research showing the opposite.

    On your 3rd point - the talking heads love to say that defense matters more in the playoffs but is it really the case? It doesn't really make logical sense and I would assume its people seeing what they want to see (are we forgetting Pats this year, Saints vs Colts all offense superbowl, GB last year, etc. etc.). I could be wrong though. Do some research and come back and tell us. There's a community site for that reason

  27. Anonymous says:

    @ Ben, and by extension, Brian:

    Recovering fumbles may not be a repeatable and predictable skill, but causing them can be, and certainly has been for the 49ers this year. Regardless of an unpredictable rate of recovery, and assuming it is not zero, the more fumbles a team causes, the more fumbles it will recover. Similarly, there are repeatable and predictable skills that contribute to interceptions.

    A win probability model that ignores those skills is a model that could stand some improvement.

    The same goes for Special Teams' play. It may not be consistent or predictive from one season to the next, but how does that justify ignoring it within a given season? All season long, the 49ers have enjoyed a significant advantage in average starting field position. That advantage is directly related to the superior skills of their punter, kicker, coverage units and return units. Again, I don't see how ignoring those skills isn't a problem with the model.

    @ Brian:

    Some advice: don't crack on the model until it's actually wrong.

    This conversation started with the first anonymous commenter noting your model's "poor performance in predicting San Francisco wins." Did he misdescribe that performance? If not, then isn't the model wrong, or at least flawed?

  28. Anonymous says:

    Anonymous Mike: You are the reason sharp gamblers can still make money.

    (2) & (4) There are a ton of studies showing that past turnover rates *in the same season* are not very predictive of future turnovers. These studies are not hard to look up.

    (3) There have been studies showing exactly the opposite of what you are saying. I.e., that offense is more important than defense in the playoffs. Again, make friends with google.

    By no means has Burke created the ideal predictive model. But the idiocy on the bottom half of the comment board here is unreal. For a very long period, winning NFL bettors made money simply by betting *against* the team with the better turnover difference. That's how drastically the public has overestimated the predictive value of turnovers. Good to see a few lemmings are still out there creating betting value for others.

  29. Jesse-Douglas Mathewson says:

    This thread is an excellent example of why allowing anonymous commentary is detrimental to a discussion.

  30. Anonymous says:

    A Giants fan pointed me to this site and I just have to laugh. I and a numer of colleagues have tried for 30 years to develop statistical models for sporting events for betting purposes like we use for Wall Street trading. Our success making money on the street (stocks, futures, etc) always surpasses our success on sports betting, and we have spent years dissecting why that is as well. Our conclusion, which is why I giggle at this high level converstion, is acknowleding that a team, a player, a horse, whatever, is a low sample dynamic variable whose performance is subject to myriad outside factors and is not stable or smoothed out through the rule of high numbers. In English, while generalizations about teams and players can be made (better passing; better passing; etc), you cannot predict with accuracy how a team or player will perform on any given day with any accuracy. Teams improve or regress from week to week, players have injuries or impairments that may or may not be publicized, game plans could change the whole character of the performance, weather, lots of penalties or "let them play" changes from week to week, and so on and so on. It's fun, I and others have spent years looking for the secret sauce, but at the end of the day you can't quantify the human (or animal) element of the players and the officials. How do you account for the Packers offense not showing up last week? How do you account for Alex Smith beating Drew Brees at his own game? Love the stats but there is a reason they actually play the games. I happen to love the 49ers giving 3 or less. Giants have gotten great match-ups for them (other than GB who didn't play to their stats) the last few weeks, but this match-up works against them. Eli (who is playing super-human) will not consistently make three and longs against Niner D like he did to GB, Atlanta, Dallas and the Jets, and Giants pass rush will have less effect against a team that runs, dinks and dunks. Does the model look into match-ups? We'll see.

  31. Anonymous says:

    Brian, you are basically doing the same thing that sports bettors are doing, with the exception that you are putting your reputation on the line rather than your money. Nothing wrong with that. Gaining or losing personal reputation is very high stakes!

    Go Giants and Ravens!

  32. Anonymous says:

    Yes [name of team] has enjoyed past success in [name of game segment]

    But that does not predict future success blah blah etc etc.

    You've proved nothing! Show statistical PROOF that your other inputs yield predictable outputs, or admit what you don't know, or admit this is a farce.


    To say the past doesn't predict the future sounds incredibly lame when all you are doing is using the past to predict the future.

  33. Anonymous says:

    Coin flips are random, turnovers are influenced by coaching and strategy. The niners turnover ratio after game 8 or 9 was an excellent predictor of their t/o ratio after game 16 or 17

  34. Anonymous says:

    As long as we're talking correlation: what is the predictive value of t/o diff? Most coaches appear to believe it is the most impt factor in a game's outcome. They could be wrong but surely this data is available.

  35. Anonymous says:

    Jesse, the comments here aren't perfect, but compared to the overall level of discourse and civility on the web this page is sterling. Esp considering it doesn't appear to be moderated.

  36. Anonymous says:

    Excellent. Thank You.

  37. Anonymous says:

    Me: "I wonder if Brian's model poor performance in predicting San Francisco wins (42.8% from weeks four, when the first projects I could find on the Fifth Down blog were posted) should be accounted for at this point. "

    Brian Burke: "#5 Some advice: don't crack on the model until it's actually wrong. Assuming it will be wrong as a basis of criticism is, well, not bright."

    My thought is that your model, while as good or better than anyone's that I'm aware, doesn't (can't) factor in the huge number of variables in a game with a reliable sample size to be without predictive blind spots. The 49ers seem to be one of those blind spots.

    Since there are three games remaining with just four teams, knowing that at least one of the teams performance hasn't been accurately captured by the model suggests that the predictions for the remaining games are questionable beyond the usual caveat about luck.

    I don't have the database, statistical modeling skills nor interest in constructing a reliable counter model. However, I would be interested in what the community thinks could make the model more accurate for these four teams over the course of the remainder of the playoffs.

  38. Anonymous says:

    Holy crap the intuition monkeys are out in full force this week. Did the site get crossposted to ESPN or something? I've never heard stupider comments in my years of visiting ANS.

    The guy above me sounds like the idiots that were saying the model was broken last year because it wasn't capturing Atlanta's full strength...that sure turned out well.

    I'm from SF and a huge 49ers fan but I'm almost hoping that the niners get crushed so these morons shut the hell up.

  39. Anonymous says:

    Sounds like the intellectuals are getting a little touchy when their metaphysics are being challenged . . .

    So, what is the record of the model this season and so far in the playoffs? I couldn't find that (truth be told, I didn't spend all day looking), but the record would answer many questions regarding the "model." The best sportswriters who make public picks week in and week out consistently get it right, against the spread, about 55% so that is my benchmark for the model. Not too high a bar, correct a bit more than half the time, right?

    Prey tell, what is the record?

  40. Anonymous says:

    "Holy crap...

    ... from SF and a huge 49ers fan but I'm almost hoping that the niners get crushed so these morons shut the hell up."

    1+++++

    I love your comment. It sums it up perfectly. The biggest idiot was the one who thinks to be the smartest. About how he beats the stock markets better than sports... and that future outcomes are not predictable. But isn´t it funny that Vegas/Brian predict at least 62% of games right over a very long period?

    This guy might never heard of something like talent level, or that teams who won the SB actually finished 6,6th out of 26-32 teams on average in Y/PP-Offense. If passing doesn´t predict future sucsess/failure, then SB-Winners should be on average 13-16th instead!

    This moron also doesn´t seem to know that sport bettors only lose money b/c of the comission (10% compared to 2,7% in french roulette) on winning bets.

    And he also doesn´t seem to know that Brians intention never was to help bettors, but to understand why teams win/lose.

    An angry greeting to all those one time poster trolls. Get out of here and unload your BS where it belongs to: ESPN/CBS forums.

    Karl, Germany

  41. bravehoptoad says:

    smith had a 1.1 int rate this year.

    Bringing his career rate down to 3.0
    Do you really believe alex smith has turned the corner and will continue to have a lower int rate than Brady Brees and Rogers?

    or do you believe a bunch of picks were dropped and the 49'ers won the field position battle?


    False dichotomy much?

    The best explanation I've heard is that Smith is trading INTs for sacks, since his sack rate is way up. He also has a poor 3rd-down conversion rate and a poor red-zone rate because he's been so careful with the ball. If he's willing to sacrifice efficiency in other areas, I don't see why his interception ratio couldn't stay this low.

    Have you watched the games? He hasn't had a lot of interceptions dropped.

    ST do matter. They have an impact. Rarely are they consistent enough, however, to have a predictable impact on a single game.

    Can you quantify that "rarely?" This seems to just be an assumption you're making. I haven't seen anything to justify this statement.

    It sure seems, when watching the games, that punt and kick-off coverage is dependable.

  42. Anonymous says:

    @ The "Intuition Monkeys" commenter:

    How about backing your position up with something other than insults? The "guy above you" (who apparently is the first Anonymous commenter) did not fault the model because it has failed to fully capture San Francisco's strengths. He questioned the model because it has not accurately predicted the team's wins, which, after all, is the purpose of the model.

    Brian did not dispute the assertion that the model has been inaccurate with regard to SF, so I assume the assertion is true. Why does an inquiry into possible explanations for an established inaccuracy offend you so much?

    As for the alleged non-repeatable and non-predictable nature of all things related to special teams play and turnovers, doesn't the Football Outsiders model take into account fumbles caused and some elements of special teams play? If so, are they mere Intuition Monkeys too? How does that model's performance compare to Brian's in predicting San Francisco's wins?

    @ The Wall Street Guy:

    Thanks for coming to the party and pissing in the punch bowl.

    Not that you are wrong.

  43. Matt says:

    re: Bravehoptoad

    Brian has posted articles in the past (you may have to go back as far as 2 years to find em) showing why he believes that ST success does not predict future success on a whole. The models simply show there is very little future correlation.

    As for my personal beliefs... I find in todays NFL, a team netting 5 more yards on average field of possession is probably a much smaller effect than a teams offensive and defensive efficiencies. Especially with the new kickoff rules. Basically, when ST has a large impact on the game, its when a returner brings one to the house, and in general that play is rare and unpredictable.

    I don't remember what the conclusions were for place kickers... I'm guessing they have some effect, but are generally not very predictable either, I honestly don't have the answer there... but I'm sure he's written an article on it.

  44. Anonymous says:

    In the 14 games predicted for them, the 49ers somewhat outperformed their expectations, this means the model is garbage and should be thrown out...

    Are people serious or am I just being trolled hard.

    We went through this exact same thing last year with pissed off Atlanta fans telling us that the model was broken in capturing their team's strength until Brian had to bring out the math and show that such aberrations are entirely possible and actually probable.

    You're taking a tiny subset of games from the larger sample in which there is some inaccuracy and saying it is proof of the model being broken. If people can't understand the absurdity in that there is no hope.

    Also to Anon 2 above. The FootballOutsiders model is pretty widely regarded as an explanatory model not a predictive one like Brian's is. FO's model takes into account a LOT of hilarious things (red zone efficiency?) that are not even in the realm of being predictive or repeatable

  45. bravehoptoad says:

    It does seem odd that there's absolutely nothing about special teams that's predictive. Nothing?

  46. Anonymous says:

    @Anon who posted this:
    "@ The "Intuition Monkeys" commenter:

    How about backing your position up with something other than insults? The "guy above you" (who apparently is the first Anonymous commenter) did not fault the model because it has failed to fully capture San Francisco's strengths. He questioned the model because it has not accurately predicted the team's wins, which, after all, is the purpose of the model. "

    Thanks.

    @ Anon who posted this:
    "Are people serious or am I just being trolled hard."

    Why do people who believe the model tend to react so poorly to people who question elements of the model? Burke doesn't claim the model is perfect. Looking at the posts, I've seen very few trolls questioning the model, but a significant portion of model defenders responding to sincere and thoughtful questions with "morons" and "idiots."

    "In the 14 games predicted for them, the 49ers somewhat outperformed their expectations, this means the model is garbage and should be thrown out..."

    I think there was one poster—an obvious troll—who suggested the model was in its entirety worthless. The others were just suggesting that some teams do something that the model doesn't accurately capture. Perhaps you don't respond with why the model fails to capture those exceptions because you don't actually understand the model. But wouldn't it be better to acknowledge that football cannot be completely understood by statistical modeling, then work on improving one of the best models—Brian's—by figuring out how to deal with outliers?

    I suppose it's just more gratifying to rant.

  47. Anonymous says:

    It's not that NOTHING is predictable. Modeling is always about give and take...including something else into the model and giving it an appropriately objectively-deduced weight adds in both signal and noise. If the noise you add in is greater than the signal...you're not adding to the predictive ability of the model.

    Of course that's a slight oversimplification, but you get the point. The reason so many people are frustrated with the "intuition monkeys" and "trolls" is because we've been answering the same questions for years. It's fine to disagree with the philosophy of the model. But please do everyone a favor and read all of the logic and reason behind why the model exists as it does.

    The other thing that needs to be better understood is the unavoidable variance in prediction models such as this. If you've never messed around with an excel spreadsheet, random number generators, or a monte carlo simulator, do it. You'll learn a lot more than you expect.

  48. Anonymous says:

    "The reason so many people are frustrated with the "intuition monkeys" and "trolls" is because we've been answering the same questions for years. It's fine to disagree with the philosophy of the model. But please do everyone a favor and read all of the logic and reason behind why the model exists as it does."

    Thanks. That's a good response. But keep in mind, that ANS is linked to on one of the most trafficked websites in the world, so it's going to get new readers all of the time. Insulting them isn't a good way to get them to bone up on everything that's been posted.

    As it happens, I've read a lot of the older posts on the model, understand why Brian chose not to include special teams, turn overs, etc., yet I still had a question. When the season comes down to four teams, one of which is "breaking" the model (I'm using that term loosely), it seems the model could be made more predictive by accounting for what the outlier team does differently.

    I don't fault the model or Brian for not doing so: it's probably too much effort and runs the risk of adding noise, as you point out. However, I am interested in how one might approach dealing with outliers. Can someone like Tebow be modeled for? Does the discipline the 49ers D brings to forced turn-overs challenge the assumption that turn-overs are too flukey to model?

  49. Brian Burke says:

    Although I can't prove it, I believe the reason that Vegas differs from the model this week has almost nothing to do with ST. The reason I say that is because the difference can be explained fully by how I treat turnovers. I think they are (mostly) random. I treat defensive fumbles (forced and otherwise) as totally random.

    I will completely admit there could be a number of teams that are special in that they don't fit the mold. My model has to fit the league as a whole, which is one of the drawbacks of any model like this.

    There may certainly be rare teams who do win the turnover battle with much greater consistency than is currently in the model. Maybe the 49ers are such a team. And the same goes for Special Teams. Perhaps teams such as the '06 Bears, could consistently and reliably get an edge on ST. But they are rare.

    The betting markets are extremely efficient except in some special circumstances. Right now, the market is saying the 49ers' turnover ratio is much more reliable than is generally the case. The market might be right.

  50. Anonymous says:

    Turnovers are accounted for in the model to the extent that they are predictive of wins. Offensive int rate, defensive interception rate, and offensive fumble rate are all factored in: albeit, at a far smaller weight than other variables such as offensive pass efficiency.

    I didn't realize ANS is linked to somewhere. Where is it linked to? I agree insulting new readers isn't the best way to facilitate a discussion, but we can't control everyone's comments :/ unfortunately, it looks like it's something we'll have to live with.

    The model could definitely be made more predictive. Just about any model you can think of can be made to be more predictive. Pythagorean's model for calculating the hypotenuse of a triangle cannot be made more predictive...but anything from physics on down can be. The problem is that it's not easy to make something more predictive.

    We're faced with two scenarios when we look at an outlier like the 49ers: either they are truly "breaking" the model and the 8 or so statistics do a poor job of predicting their games, or, there is simple variance in a 16-game season and we should expect a handful of teams to be way off their predicted W-L record. Actually, we're faced with an infinite number of scenarios - the truth could lie anywhere on the continuum of those two explanations.

    As a kid I used to create my own hockey seasons by flipping coins. I had two divisions of 5 teams. Each team would get a flip - heads was a goal, tails was no goal. First to 5 goals won the game. Home team flipped first. You'd be amazed at how realistic the standings looked even after 82 games. Were the Flyers just really good at goal differential this season? No, of course not...100% of it can be attributed to unpredictable variance. But wait, they were great again the next season! You'd be amazed...

  51. Joshua Northey says:

    I think a lot of the reason for the "hostility" towards the "attackers" is that this is an entirely predictable and boring story. Each year in every sport on every analysis site whichever playoff team is the most undervalued by the site's model compare to popular opinion gets this huge rush of people who come out of nowhere to complain about work that has been going on all year, that they don't understand, on terms that typically are not even relevant.

    Then in the common scenario where the model proves out, they disappear and you never hear from them again.

    That is what creates hostility, because it is the same silly worthless behavior each year.

    There are legitimate questions about Brian's model, but then raise them as such in a thoughtful manner.

    "FWIW I think it's basically impossible for the sportsbooks to be off by this much at this point in the season. The model is almost certainly wrong on this game."

    I think that is actually 100% wrong. It is entirely possible the lines are off that much as long as no one else has a good method for telling when they are off. Vegas doesn't need to have a super accurate model, just the "most accurate one among those people use to bet".

    It is entirely possible they are off by a lot on this game or that game, but that we don't know that because no one else has a superior model.

  52. Pete says:

    So SF special teams...are they good and does it matter? All this is from the NFL team stats site.

    David Akers. He made a ton of field goals because the SF offense couldn't score touchdowns. His percentage is average, so a league average kicker would probably have about the same stats as Akers in the SF offense. A wash.

    Kick returns? SF has the highest kickoff return average at 27 yards, but that's based on 39 returns, or a little more than 2 kicks per game. The advantage over the league average is about 4 yards per return or about 10 yards/game. A lot of kickoffs resulted touchbacks this year, so having a higher average means less.

    Punt returns - SF is about 2 ½ yards above the league average. With 44 punt returns, this is about 2 ½ returns per game resulting in an advantage of about 7 yards/game over the league average.

    Kickoff coverage - SF is about 1 yard better than the league average. Give them about 5 yards per game there.

    Punt coverage – I never know what to do with this because it doesn’t control for where you are kicking from. So teams with worse offensive field position can bomb it, while teams with better field position to punt from are aiming, not kicking for distance. Plus I don’t know how touchbacks are calculated. But let’s take the numbers at face value. SF punts average 44 yards vs. 39 for the league average for a 5 yard advantage per kick. SF had 78 punts/16 games or about 5 per game, so 25 yards per game advantage.

    Add this up, you get a yards per game advantage of +10 kick returns, +7 punt returns, +5 kickoff coverage, +25 punt coverage = +47 yards per game advantage of SF special teams.

    League average yards from scrimmage is about 335 per game. SF was 310 on offense and 308 allowed on defense, or +2 yards per game. Including special teams puts it to +49 yards per game. So is that enough to matter in terms of scoring and wins? Are there quirks in the stats that make this advantage less than my math suggests? Is a lot of it luck? Not sure. But, all things being equal, I’d take +49 yards per game.

    Feel free to shred my math. As a disclaimer, I’ll note that I started this post with the intent to argue that special teams don’t matter and that I hate the 49ers, and I hate the Giants more and screw Tebow for beating my Steelers.

  53. Joshua Northey says:

    No one is disputed that the SF ST have been good this year Pete. That is obvious. There is some question about how much that says about how good they will be tomorrow.

  54. Anonymous says:

    "That is what creates hostility, because it is the same silly worthless behavior each year."

    That makes a lot of sense, but its sad, because this is one of the few sites where one can discuss football without the barely literate fanboyism that rots the corpse of the commentary section of all the mainstream sites. Pete's post above is a great example of how ANS is a refuge, even for those who have those silly questions every year.

  55. bigmouth says:

    I've watched every game, and I don't recall a bunch of dropped picks. Smith hasn't been great, but he does seem to have turned the corner re poor decision making. He's much more apt to take a sack this year than heave the ball into heavy coverage. What's more, there's good reason to think coaching has had an impact upon him since that's what Harbaugh preaches.

  56. bigmouth says:

    That's a very fair point, but I honestly haven't seen the data showing that ST play is this variable. Have you?

  57. bigmouth says:

    Pete, I don't think anyone's disputing the importance of special teams. The question is whether it's predictive in the sense it's a skill that can be repeated from game to game and year to year, or whether it's due to random chance like fumble recoveries.

  58. bravehoptoad says:

    There is some question about how much that says about how good they will be tomorrow.

    Yes, I've seen this asserted again and again, and haven't seen anything quantitative to back up ANS's de facto position that all facets of special teams play are non-predictive enough to be irrelevant in team rankings.

    When I watch special teams play, they don't strike my eye as any more luck-based than, say, defense. Intuitively it seems that certain aspects of special teams play would be pretty consistent, since they rely on so few individuals (a kicker and a punter) and it's so cheap for teams to "specialize" in this aspect -- spend money on players who are good special teams players -- because there's not a lot of competition for these players. People aren't knocking down Blake Costanzo's or Michael Robinson's door. You can look at Brad Seeley's ten years in New England, for instance, where from year to year there were swings in performance, but over ten years they clock in well above average, or the many years of special-teams ineptitude evidenced by the Colts and other teams.

    It's tough to just accept flat propositions that special teams are non-predictive. I'd like to see some evidence. Special teams analysis on this site seems like a black hole.

  59. Joshua Northey says:

    bravehoptoad-

    I know I once read an article ~5 years ago (don't remember where, I think FO) about how (above a certain level of performance) the performance of kickers is essentially random. Once you get above the borderline guys (say the top 20 kickers) the year to year performance variation was just noise.

    The conclusion was that there were not better and worse NFL kickers, there were just NFL kickers, and then kickers who were worse than NFL kickers by varying degrees.

    When I looked into it myself (for FF) it certainly seemed to be true for FG % adjusted for distance and location.

  60. Anonymous says:

    How about that? For those who think that ST are predictive and soo important, they just can improve/reduce the winning percentage Brian gives, with their own model. The rest of us relax and think back of how great ATL was last year. I mean SF didn´t needed some lucky bounces and a blitz happy opposing DC. No, no... SF "destroyed" NO at home.

    Karl, Germany

  61. coldbikemessenger says:

    I think it is very rare that only 1 in 10 pd's get turned into int's

  62. bigmouth says:

    Karl: Aren't debates like these why this site exists? I think people have raised some very legitimate questions about Brian's model. I'd love to see him address them -- with data -- in a separate post, which I don't think he's ever done.

  63. bigmouth says:

    The evidence is definitely there for FG%, but I could swear FO found there was some evidence that kickoff distance is a skill repeatable from year to year.

    I do know Brian did an analysis suggesting special teams play is only weakly correlated with winning. Unfortunately, the data is no longer available.

  64. bigmouth says:

    Brian, can you say a bit more about why you think the betting markets are basing their decision upon turnovers specifically? What do you mean when you say "the difference can be explained fully by how I treat turnovers."

  65. bigmouth says:

    This site should seriously consider not allowing anonymous comments. I don't think it's a coincidence that the most obnoxious and inflammatory comments on both sides of the issue are from anonymous commenters. Not to mention, it gets confusing.

  66. Anonymous says:

    Joshua Northey, there are lots of people betting very serious money on football who think of it as seriously as Brian Burke does, and as seriously as hedge fund managers do with their investments. For you to wave your hand and say this is one of those situations when the market is we off is pretty silly. It is like Joe Blow claiming that Google stock is 25% undervalued. Prove your credibility by showing past results where you have claimed the same. I'd wager you were no better than 50%. I'd say that for Brian too. The times when his model has been very different from betting lines likely shows no better than 52.4%. It isn't that there is a wizard of Oz behind the Vegas numbers, rather, it is that the Vegas numbers reflect the cumulative knowledge and willingness to risk money.

  67. Anonymous says:

    (continued).... The initial "Vegas" numbers had SF -1. Then bettors bet it up to -2.5. It isn't Vegas, it was actual money bet by professional bettors that moved the line.

  68. Anonymous says:

    This all implies that humans are perfectly rational actors which has been proven to be wrong time and time again.

    http://danariely.com/the-books/excerpted-from-chapter-1-%E2%80%93-the-truth-about-relativity/

  69. Anonymous says:

    "This all implies that humans are perfectly rational actors which has been proven to be wrong time and time again."

    I agree that you are right. But who's to say that the irrationality here is the line too high on SF. Maybe the line is too low! The only way to hint at you (or anyone) knowing where the irrationality lies is to have a betting record in like circumstances.

  70. Anonymous says:

    Let's face it. Brian Burke is trying to do the same thing that sports bettors and "Vegas" is trying to do. They are all trying to predict, with all the information they have, the probability of Team A beating Team B. There's no reason to downplay the betting market. It isn't seedy and the academic approach is not lofty or better in any way. They are all doing the same thing. The difference is that Brian is not putting his money where his opinion is.

    However, I do know there are some people that are! And they have told me they are doing pretty damn well. :)

  71. bigmouth says:
    This comment has been removed by the author.
  72. bigmouth says:

    It seems contradictory to claim the 49ers present no problem for the model because they're just one team, then argue simultaneously the Falcons' failure to win one game in the playoffs is somehow vindication of it. Besides, it's not like ATL was terrible this year.

    That said, I noticed the ATL had one of the best special teams in 2010, and one of the worst in 2011 according to FO's rankings. Similarly, Baltimore had one of the best last year, and worst this year. So there seems to be some evidence for variability in special teams, lol.

  73. coldbikemessenger says:

    "For you to wave your hand and say this is one of those situations when the market is we off is pretty silly."

    You remember 2008 right?

  74. Anonymous says:

    Or 2001

  75. Mike B says:

    If Brians model was a better predictor of future performance than the closing market number(s), I feel that enough time has elapsed with his information being freely available, that the marketplace would mirror Brians numbers. Save for injury adjustment and the like.

    I think ANS does a very good job, as does FO. They appear to do some different things and generally their overall views of teams aren't that far off, even in relation to the market. However, when there is a noticeable difference between either site and the market, then IMO, the market should be respected. Money talks :)

  76. Jim Glass says:

    The best explanation I've heard is that Smith is trading INTs for sacks, since his sack rate is way up.

    Maybe there's a little something to that.

    But note that the best record ever at preventing interceptions was Bart Starr's 3 picks in a 14-game stretch in an era when the pick rate was far higher than today's, and then he threw 9 in his next two -- same team, same coach, same offense, same everything. Good weather too. Last year Tom Brady threw 4 all season, this year he threw 4 in one game.

    Interceptions are overwhelmingly luck. (And Brian too.) The *in season* correlation for turnovers -- first eight games to last eight -- is all of 12.6% over the last three years. But of course when one's own QB suddenly has a streak where he's thrown a lot fewer picks than ever before, then simple good fortune can't possibly have anything at all to do with it, we all appreciate that. :-)

    (Same thing when one's own QB suddenly throws an unexpected whole lot of interceptions, it can't possibly be just bad luck -- alas for him.)

    there are a ton of studies showing that past turnover rates *in the same season* are not very predictive of future turnovers. These studies are not hard to look up.

    They aren't hard to do for oneself, either. I just took the PFR.com game data for the last three years and put into my spreadsheet to get that 12.6% above. For "forced" turnovers taken from the other team the in-season correlation is just 11.8%. I wouldn't want to bet my fortune on that.

    When the season comes down to four teams, one of which is "breaking" the model (I'm using that term loosely), it seems the model could be made more predictive by accounting for what the outlier team does differently.

    Outliers happen by random chance. That's the definition of them.

    So how do you know that a team in the final four on the strength of being an outlier in some way is actually doing anything at all different from being lucky? Please detail.

    As to all the SF fans going on about how their great special teams can be expected to carry them to victory...

    (1) NFL game outcomes in our era basically are explained 45% by offensive performance, 40% by defensive performance, and 15% by ST performance. That's explained looking backward. Going forward, the predictable sustainable margin of difference in ST play is a lot less than that 15% at best, even when STs are mismatched. But other final four teams have good ST play too.

    (2) Brian's model may not include ST play, but FOers rates it and this year ranked CHI #1, TEN #3, and the Jets #4. How far did their top ST play carry them? In short ... get over it and don't count on it.

  77. Joshua Northey says:

    I don't think you actually read my comment at all.

    I wasn't saying I have a better model, or even that one is possible. I was saying it is possible that Vegas is way off. Those are not the same thing.

    Surely there is some objective fact to the matter, and surely the Vegas lines are just some approximation of that. This line is not 100% accurate. But that doesn't mean anyone has a reliable model for telling when it is inaccurate.

    It does mean that "your model disagrees with Vegas" while a notable criticism, is not the end all be all of arguments. Just because you disagree on some particular game with what maybe be the best model out there doesn't mean your model is worthless or cannot inform the slow process of better approximation.

    Look at what I said again:

    "I think that is actually 100% wrong. It is entirely possible the lines are off that much as long as no one else has a good method for telling when they are off. Vegas doesn't need to have a super accurate model, just the *most accurate one among those people use to bet*.

    It is entirely possible they are off by a lot on this game or that game, but that we don't know that because NO ONE ELSE HAS A SUPERIOR MODEL."

    Capitalized one sentence you definitely seemed to miss.

  78. Joshua Northey says:

    To be honest I don't think I have ever met a gambler who didn't claim they were doing well, yet 90% of them are losers, so you do the math.

    I had a friend one year turn $100 to several thousand over the course of one NFL season. We all heard about it all the time. Of course by the end of the year that year he had lost it all. And he didn't talk about his betting all the other years he simply lost money without talking about it every week.

  79. Anonymous says:

    Another classic, Jim Glass :-)

    But i doubt the fanboys coming over from ESPN/CBS understand what you are talking.

    Anyway, love it, 100% agree.

    Greeting, Karl, Germany...

  80. Jesse-Douglas Mathewson says:

    No argument here.

  81. bravehoptoad says:

    . Going forward, the predictable sustainable margin of difference in ST play is a lot less than that 15% at best....

    Where are you getting that?

    Brian's model may not include ST play, but FOers rates it and this year ranked CHI #1, TEN #3, and the Jets #4. How far did their top ST play carry them?

    I'm not sure what you're saying. Are you saying that FO over-rated the effect special teams had on their results? Or are you saying that ST are only worthwhile if teams make the playoffs? It sounds like that second one, and it isn't much of a point.

    But I'm glad you brought up those other teams, because this isn't just an issue with San Francisco. Look at the glossary of stats on this site -- not one of them even mentions special teams, not even WPA or EPA, which seems odd, because it's easy enough to calculate the effects of special teams play on those, isn't it? Why is there no special teams work on this site?

    Even in retrospect, as you say, ST play is only worth about 1/7th of performance, so on one is going to ride the special teams to the playoffs. I do find it interesting that if you subtract San Francisco's 7.8% special teams DVOA, they end up ranked right about where Brian Burke has them.

  82. Ben Stuplisberger says:

    At the risk of continuing this conversation too far, here's my two cents on the trolling:

    There are two definitions of trolling. The first one is posting obviously stupid arguments to get a rise out of people. This is not what is happening here. The second definition of trolling is making arguments from ignorance and showing a lack of respect for the Internet community you have visited. This latter definition is what is happening.

    Those of us who are complaining about the posts are not annoyed because we think the model is perfect. It is because we have read the philosophy behind the model, and we understand what it intends to do and what it intends not to do. Most of us are also familiar with Chase Stuart's work at Pro-Football Reference blog and Smart Football, Aaron Schatz's work at Football Outsiders, and even sabermetrics sites.

    It is simple Internet etiquette: If you have not bothered to educate yourself about a site, don't post anything other than a question. Brian has posted links to his articles and to other stats websites that can assist you in that education. Look at the name of the site: Advanced NFL Stats. If you don't get that, learn what it means, then post.

    Just some advice. :)

  83. Anonymous says:

    You see Jim,

    it took only one post later to prove my post... ROFL

    Enjoy the games. My picks NYG over SF, NE over BAL. No guarantees here, of course. Since us long time readers of this site know: 50+% of NFL-Outcomes are decided by luck. :-)

    Karl, Germany

  84. Anonymous says:

    @ Jim Glass

    Brian's model may not include ST play, but FOers rates it and this year ranked CHI #1, TEN #3, and the Jets #4. How far did their top ST play carry them? In short ... get over it and don't count on it.

    Nice straw man. I don't think one commenter has said that superior ST performance is all a team needs to win or that it is as outcome determinative as offensive performance or defensive performance. Not to speak for anyone else, but what I am saying is that ST performance is not irrelevant to the outcome and that it may be a mistake for the model to ignore it completely, particularly those aspects that are consistent and repeatable/predictable.

    @bigmouth:

    I don't think it's a coincidence that the most obnoxious and inflammatory comments on both sides of the issue are from anonymous commenters

    You think so? What does Karl from Germany have to do to get some recognition around here?

  85. Anonymous says:

    Please don't talk about DVOA on this site. DVOA is a joke

  86. Anonymous says:

    Anonymus. Well you gave me recognition, since you was reading wath i said ;-)
    Anyway, i really don´t need yours. I have had my good discussions with Jim, Brian and others (+ good discussed articles). But how can a annoyed one time lurker like you know that? Good try guy... BTW, you didn´t got it too, what Jim explained. Leaves me with big shake of the head. Ignorant!

    Karl, Germany, GO Giants (not a Giants-Fan ;-)

  87. Jim Glass says:

    For those who think "winning the turnover battle" is a strategy teams can rely on to win week-by-week, this is for you:

    For the last three seasons I took in-season alternate game turnover tallies for all teams -- turnover count in odd-number games versus even-number games. The correlation between turnovers committed in odd and even number games is a meager 10% -- 90% of the way to nothing -- and the correlations for takeaways and net turnovers are even lower.

    Putting this in plain numbers:

    The six teams of the 96 in the three seasons that were *most dominant* in winning the turnover battle in their eight odd numbers games were +82 total (an average of almost +14. Wow!). In their eight even number games they were ... -6.

    The six teams that were most dominant in winning the turnover battle in their even-number games were +60. In their odd number games they were ... -6.

    Of these 12 teams that were the most dominant of all at winning the turnover battle in their half-season of odd or even number games, seven of them -- that is *most* of them -- *lost* the turnover battle in the other half of their season.

    These fact sure don't make it look like winning the turnover battle is a skill that carries over very well game to game, something a team put in its game plan and rely upon, eh?

    Yes, turnovers are hugely important in determining game outcomes, and teams put great efforts into drilling their players to force them and prevent them. But all those efforts cancel each other out, and the result is that turnovers are *overwhelmingly* the result of random chance.

    Say "Red Queen Competition", a phenomenon well known in several fields, but not recognized by many sports fans even when they are looking right at it. It's when you have to compete like hell just to stay even, so chance does the rest, and that is the best you can do.

  88. bravehoptoad says:

    I did not introduce the subject. You'll have to blame Jim Glass for that one.

  89. bravehoptoad says:

    Isn't sample size going to be a significant limitation here?

  90. Joshua Northey says:

    Not enough to undermine the point unless he is cherry picking. Run more seasons if you are concerned about sample size.

  91. Anonymous says:

    I'm just happy these posts aren't about Tebow!

  92. Mike M says:

    Interesting probabilities, a model is only as accurate as the info used in it, no one person has all the info available or uses the info in the same way.

    Looking back in history the least likely SB match-up is Pats/Giants, I can pretty much say without question it will not happen.

    Did you know, no no.1 seed as ever gone on to win the SB that played as poorly as the Packers this season, and more than half lost in their opening playoff game ?

    Should the Giants win come as such a big surprise if you had know this ?

    Was that info included in the model, first one needs to know the info, not likely he did.

    There is sooooo many different ways to look at and use info.

    See you after the games boys.................................

  93. Mike M says:

    An example of how one uses research can be seen by Jim Glass.

    Because 12 (12.5%) of 96 teams showed random TO's that's is supposed to tell us that the majority of teams TO's are random , say what Jim ?

    If only 12.5% of teams show random TO's that would seem to indicate those teams are the exception not the general rule.

    And the other thing is your showing teams who've reach extreme levels of winning the TO battle, of coarse you could show the same thing with basically any stat when teams reach extreme levels in those stats . That's called regression to the mean which happens with most all stats TO included.

    Your research does not prove for the............ MAJORITY.............. of teams that TO are random, it proved that for 12.5% oF teams they were random........

  94. Jim Glass says:

    An example of how one uses research can be seen by Jim Glass. Because 12 (12.5%) of 96 teams showed random TO's that's is supposed to tell us that the majority of teams TO's are random, say what Jim?

    Mike, did you see me say that for ALL teams over three seasons the correlation for turnovers committed in odd-number and even-number games was all of 10% -- and the correlation for net turnovers was even lower, single digit?

    What does that tell you? For ALL teams.

    If only 12.5% of teams show random TO's...

    Um, where do you get "only 12.5%" from ALL?

    Now, because I knew that a 10% correlation for all teams is a concept too obscure for many to understand -- as it indeed evidently is -- I gave an illustration of what this means "in plain numbers".

    To the extent net turnovers are generated by skill instead of luck, we'd expect to see them generated to the *same degree* in a team's odd and even number games. And we'd expect to see this *most clearly* in the teams with *the most skill* -- those supposedly with the most skill at creating net TOs, those with the most.

    Just as if Babe Ruth and Ted Williams were superior batters *by skill* we'd expect to see the fact evident in both their even-number and odd-number games -- and we'd expect to see that consistent skill more clearly in their games than in the games of lesser batters.

    So we look again at the numbers for the six teams that generated *the most* net turnovers -- presumably because they had superior skill at doing so, like the 49ers -- in their odd number and even number games 2008-2010, and at their net turnovers in the other half of their games.

    Odd ..... +82 ... +2*
    Even .... +60 ... -6
    Total .. +142 ... -4

    The teams that produced the *best* 12 half-seasons out of 192 at generating net turnovers were +142 in them ... and were -4 net in turnovers in the other half of their games in the *same* seasons.

    your showing teams who've reach extreme levels of winning the TO battle, of coarse you could show the same thing with basically any stat ... That's called regression to the

    So you are saying that if you found Babe Ruth and Ted Williams to be superior batters in even number games during a season you'd expect them to be poor batters in odd number games in the same season? If you found Peyton Manning and Tom Brady to be superior passers in odd number games in a season you'd expect them to be only average or worse in even number games in the same season?

    Really???

    If you really think that, it's time to look up how regression to the mean works. Specifically, what is the mean being regressed to? That mean is the level of skill.

    Your research does not prove...

    Mike, it's not my research. It's everybody's research. You'd know that if you were literate on the subject. I didn't "research" anything, I just looked at the last three years of data. Took a few minutes.

    Also, speaking of such, can *you* present *any* evidence showing that the correlation in TOs between odd and even games for ALL teams is much higher? After all, you are the one making the claim it is, so you have an obligation to support that claim.

    I mean, of course, evidence other than "I know what I know because I know it. QED!" :-)


    (* Correcting a typo in my original comment.)

  95. Boston Chris says:

    Hey Mike M,

    See you after the games. Nice guarantee on the Super Bowl there. Time to change your name, because otherwise, you can't show your face around here anymore.

  96. Jim Glass says:

    Looking back in history the least likely SB match-up is Pats/Giants, I can pretty much say without question it will not happen.

    As wise men from Niels Bohr to Yogi Berra have noted, making predictions is difficult, especially about the future. Eh? :-)

  97. Jim Glass says:

    As a final observation on the "being plus on turnovers is a reliable skill" and "the 49ers special teams are undervalued in the model" arguments, the 9ers were -2 on turnovers (fumbling four times but luckily recovering two of their own) and lost on a special teams bungle.

  98. Smerdyakov with a Guitar says:

    Exactly, Jim. Poetic justice was served last night -- after all the talk in this thread about the Niners not receiving credit for forcing turnovers, they were cruelly reminded what a fickle ally the turnover really is.

Leave a Reply