The Weekly League: Notes and Ideas for Week Fifteen

This week's edition of The Weekly League features:

1. Mind-altering previews of the Kansas City-Saint Louis, Green Bay-New England, and Chicago-Minnesota games.

2. An updated "luck" table, exposing the Atlanta Falcons for the frauds they really are.

and

3. No shortage of The Good Times™.

The Four Factors you see for each game represent each team's raw performance thus far in four important categories (pass and rush efficiency, pass and rush efficiency against) relative to league average (where 100 is league average and anything above is good).

Along with the Four Factors, you'll see two other numbers: Generic Win Probability (GWP) and Game Probability (PROB). The GWP is the probability a team would beat the league average team at a neutral site. It can be found for all teams here. The PROB is each respective team's chance of winning this particular contest. Your host, Brian Burke, provides PROBs to the New York Times each week, and those numbers (along with methodology) can be found here.

Finally, a glossary of all unfamiliar terms can be found here.

Kansas City at Saint Louis | Sunday, December 19 | 1:00pm ET
Four Factors


Notes
• Unless I'm mistaken, the prevailing view on Rams quarterback Sam Bradford is that he's having a good season.
• Probably because St. Louis is at the top of their division, is why, right?
• But let's also make sure to note that Bradford is second to last -- ahead of only Jimmy Frigging Claussen -- in yards per attempts among qualified QBs, with 5.4.
• That's (slightly) behind Derek Anderson, it should be noted.
• In somewhat related news, both Bradford and Anderson are way better at being quarterbacks than I am at being anything in the world.

Green Bay at New England | Sunday, December 19 | 8:20pm ET
Four Factors


Notes
• In a perfect world, this game would represent a match-up between two dreamboat quarterbacks.
• Unfortunately, due to Aaron Rodgers' concussion, this game will feature just one dreamboat quarterback and then another, normally boated quarterback.
• As for Flynn, Vegas has a prop bet on him, as follows: "Matt Flynn – What will he do first Week 15 vs New England?"
• Throw a TD Pass: EVEN.
• Throw an interception: -130.

Chicago at Minnesota | Monday, December 20 | 8:30pm ET
Four Factors


Notes
• In the event that you're unaware, Minnesota's home stadium accidentally broke last week.
• As such, this game will be played at the University of Minnesota's TCF Bank Stadium.
• In the event that you're unaware, Minnesota's starting quarterback accidentally broke, too.
• As such, young Joe Webb will start at quarterback for the Vikings.
• Webb also qualifies at wide receiver which makes him -- in the parlance of our grandfathers -- "pure fantasy gold."

GWP Wins and Luck
Here's the table, through Week Fourteen and sans comment, of GWP wins and losses as compared to actual wins and losses.

  • 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

47 Responses to “The Weekly League: Notes and Ideas for Week Fifteen”

  1. J.R. says:

    So ATL is 5.4 wins better than they ought to be? I watched Baltimore vs Houston, so I know which game the Ravens ought to have lost.

    Any chance you could correlate this "luck" factor to turnover differential? I'm willing to bet that the lucky teams are the ones for whom (a) lost turnovers are "clustered" in single lost games, and (b) gained turnovers are spread out over several wins.

  2. Anonymous says:

    We already know that turnovers correlate to luck. They call it luck because they don't believe it's a repeatable skill.

  3. Anonymous says:

    Also, 3 of the 4 NFC South teams are the luckiest? I wonder why that is? Is it because they were lucky to get the NFC West scheduled (haha).

  4. David says:

    Brian - Separating yourself from your model, how good do you think the Falcons are this year? What do you think their record should be?

  5. James W says:

    This debate has been done to death. One opinion is that Atlanta's offense is low variance and tends to manage risk appropriately to the game-state. They are probably a game or so ahead of where they should be, but there is a strong chance they are significantly better than this model suggests

  6. David says:

    I only ask the question because I know Brian has acknowledged the model likely undervalues the Falcons, but there's 5.41 games worth of ground to make up, so I'm wondering if he thinks their ability is closer to the model or their record.

    Your point on variance though raises something I was thinking about the other day. I believe Atlanta's poor performance on yards per attempt has been offered as statistical evidence that they are not as good as their record suggests, but any average is susceptible to manipulation by extremes so the fact that Ryan's longest pass completion is 46 yards and we have more of a grind-it-out rush attack would hurt our yards per attempt. That said, I would guess that the Falcons perform much better relative to the rest of the league on median yards per attempt as opposed to mean, so maybe there's reason to account for that.

    Also, is time of possession included in the model at all? That is a clear strength of the Falcons and something widely accepted as important to the game of football that I haven't heard discussed on here at all.

  7. James W says:

    Some excellent points there. TOP is not a skill though. But when you move the chains like Matt Ryan does, it can serve as a reflection of a highly repeatable skill.

  8. Anonymous says:

    I'd like to know New England's GWP at the end of the 2007 season. Is that information on this site, I can't find it.

  9. Anonymous says:

    The first value of TOP would generally be considered fatigue-- you wear down the opposing defense and rest your own. For that aspect, if it really helps, it should already be reflected by a team's efficiency stats.

    The second value of TOP would be closing out games with the lead, which an AYPA (yards/attempt) model wouldn't show, but I think the WPA (winning percentage added) model should.

  10. Dan Danilo says:

    Fraud implies a intended deception. The Falcons have simply beeen trying to win football games any way possible. Would they dispute the factor of "luck" in their wins? I don't think so. Your verbiage indicates a tone of hostility and seems beneath your professionalism.

  11. Carson Cistulli says:

    Dan, please note that this article wasn't written by Brian Burke, but by me.

    I can assure you that, while I bear no hostility against the Falcons, it would also be difficult to produce something "beneath" my current level of professionalism, which is currently somewhere around the zero-percent area.

  12. Bruce D. says:

    Blessed or fraud?

    I tend to lean somewhat towards fraud, even after a Seattle beat-down.

    Check the Falcon's turnover ratio, can it continue against elite teams?

    If so, then I'm wrong and they are more blessed than frauds.

    Just interpeting what the data says.

  13. David says:

    WPA isn't included in the main model used on this site though, right? Because including WPA ultimately has the effect of just ranking teams according to their record? That's why I am suggesting time of possession should be included because it is something that can affect a team's chances of winning but not out of necessity affect a team's win-loss record.

    Your point about closing out games with the lead is exactly what I am trying to get at. The Falcons have a grind-it-out style that enables them to win close games in the end. The efficiency stats include a lot of data from portions of the game before such benefits would take hold.

  14. David says:

    Steelers: 1-1
    Saints: 3-1
    Eagles: 1-2
    Ravens: 2-0
    Packers: 1-0
    _____________

    Total 8-4

    Overall: 24-13

    The Falcons have a slightly better turnover ratio against playoff teams than non-playoff teams. Overall, the Falcons just do a very good job of limiting their own turnovers, which I think is a much more repeatable skill, and I don't think have been overly reliant on forcing turnovers.

  15. James W says:

    Falcons offense v Seahawks
    73 plays 266 yds 3.6 yds per play
    but
    9/19 3rd down 2/2 4th down
    21 first downs 34 points
    Perfect summary of the inadequacy of yards-per-play metrics or another lucky win?

  16. James W says:

    Sorry. 27 points actually. How could I forget how brutally bad Matt Hasselbeck is?

  17. David says:

    21 first downs on 266 yards, that just about sums it up

  18. Anonymous says:

    I think it's extremely funny that people attribute turnovers to luck. It's a highly repeatable skill- "stripping" the ball during a tackle, for example, is something the Giants are very good at. With 30 forced fumbles, they are more than 3 times better at it than NE. Not as good at getting interceptions, but still better than league average. In fact, they have 2.42 successful turnovers per game on their defense. Unfortunately, with a 2.5 turnover per game average on their offense, they have a net negative turnover differential.

    The point remains, however- some teams are significantly better at defensive turnovers than others, and it is reflected statistically. Others are better at limiting their offensive turnovers, which is also reflected statistically. You simply can't tell me that the Denver Broncos (8 interceptions, 11 forced fumbles, 8 recovered) have fewer defensive turnovers than the NY Giants (15 interceptions, 30 forced fumbles, 19 recovered) because of luck, or opponents faced.

  19. Anonymous says:

    by the way, the Falcons are 4th in the NFL in turnover differential at +13, and 1 away from a 3 way tie with the eagles and steelers for second place.

  20. Brian B says:

    Yes. Turnovers are a skill. But repeatable? Maybe, but only over large sample sizes. At the pro level, defenders are all trained to strip the ball or nap interceptions. They get relatively few good opportunities to do it, so it's nearly impossible to tell who is better than whom at that particular 'skill'. It's accordingly difficult to predict which teams are more likely than others to generate turnovers in the future.

    See the difference?

  21. Anonymous says:

    Sure it's easy to tell who is better than whom- Asante Samuel, DeAngelo Hall, Devin McCourty, they are all good at getting interceptions (7, 6, 6) and Champ Bailey is not (1). Let's look at the Atlanta game- B. Grimes and W. Moore both intercepted Seattle (unsurprising, as it makes 5 for the year for both of them). One was taken at the Seattle 37 and brought to the 32, a net change in EPA of roughly 4 points in favor of Atlanta (-1 to +3). The other was taken at the 30 and brought to the 21. A swing of +5 in favor of Atlanta. Sure enough, on those two drives, they got a total of 10 points. Having such short fields is why they can score more points off fewer total yards and afford lower yards per attempt- they need fewer yards to create winning scenarios.

    In terms of predictability- are you telling me you were surprised in the GB/ NE game when NE got the better turnover differential? I know it's difficult to quantify statistically, but attempts could be made. Maybe take into account average starting field position when looking at D/ST efficiency- this would value interceptions/ fumble recoveries as well as effective punt returns, reward them for pinning opponents deep in their territory, etc. It would also help bridge the gap between yards gained and points scored.

  22. James says:

    Excuse me, did you just say Champ Bailey, with 48 interceptions in his career and the third-most among active players, isn't good at getting interceptions?

    Interceptions are unpredictable and inconsistent. That's why Champ Bailey can have 10 interceptions one season, and 3 the next. That's why Asante Samuel has an average of 5.3 picks a season, but a standard deviation of 3.3 - it's highly variable.

  23. Anonymous says:

    That's correct, I did- Champ Bailey hasn't had a season with more than 3 interceptions since 2006, when the Denver Broncos had Larry Coyer as defensive coordinator. That time accounts for the majority of his interceptions (21 interceptions in 2004-2006 under Coyer, 8 in 2005 and 10 in 2006) which only Further Emphasizes that they are a teachable, repeatable skill, and that the defense of the current team determines how many of them are obtained. And I stand by my statement- 1, 2, and 3 interception seasons for the past 4 years means that (as of right now, with his current defensive coordinator) Champ Bailey isn't good at getting separation. Just like Randy Moss sucks at getting separation. You can't point me to success from years ago and pretend it equates to current performance, and you're ignoring the point.

  24. Anonymous says:

    By the way, he had 10 one season (under Coyer) and 3 the next (under Rick Dennison). Still think they're random?

  25. Anonymous says:

    Sorry, final point- Asante Samuel had an average of 2 interceptions per season from 2003-2005, and an average of 8 per season from 2006-2007. The difference? A new defensive coordinator was hired in 2006. Coincidence? I think not...

  26. Anonymous says:

    If you want to argue that it's easy to tell who is better than whom at making interceptions, it is (unfortunately) not sufficient to just look at the league leaders in interceptions. All you've proven is that there are players who lead the league in interceptions. You've shown that there are not 10, or 20, or 30 or more players all tied for the most interceptions.

    The problem is that predicting who will get the most interceptions is a very, very difficult thing to do, for all the reasons that Brian B. listed above. Sure, defenseive interception rates are not truly and exactly equal for all cornerbacks and safeties, but is there really a large enough difference that they will fall in line each season? Or even over the course of multiple seasons?

    Measuring interceptions is an atrocious way of figuring out who is better. Interception rate is much, much better, but still very unlikely to represent who is objectively better. Rates need to be calculated for obvious reasons - a "great" cornerback on a bad team will have far fewer chances for interceptions than a "bad" cornerback on a great team.

    Of course we're not surprised that NE got the better turnover differential against GB. But guess what, I also wouldn't be surprised if GB had gotten the better turnover differential. NE was almost definitely more likely to have fewer turnovers, but can you really say it's more than 60/40 or so? And if that 40 had happened, is it really so surprising? Matt Flynn threw for nearly 100 more yards than Tom Brady, plus an extra touchdown. Was it expected? No. But it wasn't outside the realm of possibility. Such is the reason for prediction rather than foresight.

    Teams who have the best W-L record by definition are more likely to have had good luck than bad luck...and teams with the worst W-L record are more likely to have had bad luck than good luck.

    Similarly, teams with the best turnover differential are more likely than not benefiting from better luck than teams with the worst turnover differential. Is there an underlying skill? Yes, of course. But how much of the turnover differential is noise, and how much is signal?

  27. David says:

    I think both the ability to limit turnovers and the ability to create turnovers is a skill. However, I would argue that the opportunity to create turnovers is more dictated by the opponent (see Deangelo Hall getting 4 of his 6 interceptions in one game against Jay Cutler), so therefore I think it is probably more difficult for a team to rely on creating turnovers for their success than on continuing to limit their turnovers.

    If your quarterback knows how to make good decisions and manage the game and your running back is steady at holding onto the ball, a great defense may occasionally make a great strip or interception, but for the most part I think that's something that can be relied on game to game regardless of opponent.

  28. Anonymous says:

    Asante Samuel had an average of 2 interceptions per season from 2003-2005, and an average of 8 per season from 2006-2007. The difference? The Carolina Hurricanes won the Stanley Cup in 2006 for the first time ever.

    Whenever the 'canes win the cup...Samuel has ALWAYS averaged 8 int's over the next two seasons. No exceptions. In the history of everything, ever.

  29. JJB says:

    I can't recall the specific posts on this web-site where the information is presented, but the summary is that turnovers correlate less weakly with the offense than the defense, but not very strongly with either. And special teams don't correlate with much at all (meaning they appear to be random, or whatever it is they correlate with just hasn't been stumbled upon, yet). So the WPA/EPA model pretty much ignores turnovers and special teams in its predictions, because they're essentially a crapshoot, as far as we can tell, right now. And the model is a pretty good predictor of actual outcomes, so it must be getting some things right.
    I'm an Atlanta native and lifelong fan of the Falcons, but I've watched them enough this season to know that their defense is mostly smoke and mirrors. And they are certainly the beneficiaries of a soft schedule. To borrow one of TMQ's favorite phrases, the NFC West is "close-your-eyes awful".
    But there's one other thing that the WPA/EPA model doesn't really take into account, and that is coaching. The "luck" factor can also be considered to be at least in some part, the "coaching" factor. And it's quite possible that some of the "luckier" teams are also some of the better-coached teams. And the "unlucky" teams are some of the not-so-well-coached. (Yes, I'm talking to you, Norv Turner). Again I can't recall the precise posts, but the luck/coaching topic is something that has been touched upon before on this web-site.
    And a final note. Just from reading this article, I knew the author was Carson and not Brian. They both have their own unique and very different styles. And that does help to keep the posts "fresh", which I'm sure was part of the intent when bringing Mr Cistulli on as a regular contributor to the blog.

  30. Anonymous says:

    Right, naturally we can't necessarily predict how many interceptions any given player will get any given game. For that matter, even predictions regarding overall turnover rates will be subject to variance, and though I can say with confidence that the data suggests NE will emerge with a positive TO ratio against Buffalo next week, the individual performance may not bear that out. I can even assign a percentage chance to my prediction, based on large data sets involving TO ratios before entering given games then performance within that game, but I might still be wrong. That doesn't mean it's worth ignoring, or taking the easy way out and attributing it to luck.

    That's why (bringing me back to my original point) I recommend including average starting position of the Offensive drives when measuring D/ST efficiency, because it takes into account not just those, but also kick returns, how deep in their own territory they pin their opponents, etc. Turnovers win games, we all know this, because of how large of a swing they are in EPA. Without those two ints, just based on EPA before those actions, we're looking at 24-21, instead of 34-18. They're too important to just ignore or brush off as luck. Shoot, going into the game, I could have told you the 4th best team in the NFL vs the 3rd worst in regards to turnovers would have turnovers play a significant part in winning. I just would have difficulty quantifying how large of a part. I can't even now tell you whether it's more important to examine it as D/ST efficiency metrics or as Offensive Inefficiency metrics (because I agree, Not turning over the ball is a skill as surely as generating turnovers, and just as important in this discussion), but I do think it's worth discussing and finding ways to properly analyze.

  31. Anonymous says:

    Here, let's do this- next week, the 6 teams with the highest TO differential all face teams with an average of at least -1.0 TO/ game compared to them (except KC, who faces one with -0.93, which is good enough for me). If in the majority of those games, or even if in half of those games the team with the lower TO ratio emerges with an even or better TO differential during the game, I'll accept the low/ no correlation argument, drop it, and never mention it again. If not, can we at least perform further investigation on this matter?

  32. David says:

    JJB,

    Great point on the coaching and something I was just thinking about while reading another article on how the Falcons have come to their record through mistake-free football. Whether you want to call it "mistake-free football", "not beating themselves", or "good coaching", clearly the Falcons have done a good job of it and I think that is the key to their high efficiency stats.

    While this website, most explicitly in this post, attributes an outperformance of the model to luck, I think it is more a sign of solid intangibles that I am continuing to push to incorporate further into the model if it is going to become more predictive.

  33. Anonymous says:

    "They're too important to just ignore or brush off as luck."

    This statement doesn't hold any weight. The fact that something is important doesn't mean it can't be largely attributed to luck. If two teams were to play 60 minutes and, regardless of the score, have a coinflip decide the winner, we wouldn't waste any time analyzing the coinflip and which teams are better at it. The coinflip is certainly important - but entirely unpredictable.

    This doesn't mean interceptions and turnovers in general are as unpredictable as a coinflip, but they're certainly not as predictable as say, passing yards/att. Turnover differential probably correlates to winning higher than passing yards/att, but that doesn't mean we can predict turnover diff. with equivalent accuracy.

  34. David says:

    Can anybody point to the article that discusses how turnover differential is not predictive (I believe one has been alluded to)? I believe that the analysis was done but am skeptical of the results, so would be interested to see the research.

  35. JJB says:

    After my mega-post above, I find myself compelled to say something about rating a cornerback purely on his interception total. D.Revis and N.Asomugha both have 1 or fewer interceptions through 12 games this year (as far as the NFL site has player stats). I just can't bring myself to believe that they both suck, too. While interceptions are a significant stat for cornerbacks, it isn't the only one. Number of targets is also important. And Bailey (like Asomugha and Revis) just doesn't get targeted that many times during a game, presumably because he usually has his receiver well-covered. Consequently he doesn't have that many opportunities for an interception, which keeps his interception totals "artifically" low. McCourty is a rookie on a pretty mediocre secondary, so he's going to have lots of targets and lots of opportunities (although probably not so many next year.) And the press conference after the Chicago-Washington game was very telling when Cutler said he'd throw at D.Hall just as many times if they played again the following week.

  36. Anonymous says:

    "Here, let's do this- next week, the 6 teams with the highest TO differential all face teams with an average of at least -1.0 TO/ game compared to them (except KC, who faces one with -0.93, which is good enough for me). If in the majority of those games, or even if in half of those games the team with the lower TO ratio emerges with an even or better TO differential during the game, I'll accept the low/ no correlation argument, drop it, and never mention it again. If not, can we at least perform further investigation on this matter?"

    Not sure if this is meant to be droll or not...but seriously?

    It is something that has been looked at extensively. It is something that continues to be looked at extensively. Stuff is really complicated. Turnovers cause winning. Winning sometimes causes turnovers in the 4th. Turnovers correlate weakly from season to season, and even within season.

  37. Anonymous says:

    http://www.advancednflstats.com/2008/01/explanation-vs-prediction.html

    Here is a post by Brian regarding explanatory models vs. predictive models.

    Cliffnotes: 2006-2007 correlations for weeks 1-8 vs. weeks 9-17:

    D Int Rate - .08
    O Pass Eff. - .58

  38. JJB says:

    David...
    It's not that turnovers aren't indicative of who will win the game (actually, they are very indicative). It's just that past performance regarding turnovers is not a very good indicator of future performance.

  39. Anonymous says:

    JJB- I do not in any way disagree- the teams for which those CBs play are both in the top 10 in terms of fewest yards allowed passing per game. This can be attributed at least partly to their skill at that position. Those teams are also 1 and 2 in terms of completion percentages for opposing passers, while being 31 and 32 in terms of interceptions. Meanwhile, NE has the 2nd most porous pass defense in the league (2nd worst in passing yards, tied for 3rd worst in completion percentage), but is 2nd in terms of interceptions. Based upon that, it would appear NE is more focused upon gaining the interception than preventing the completion (bend but don't break defense?) while Oakland and NYJ have the opposite approach- stop the completion, don't worry so much about the Int. Both are valid approaches, simply different, and I did not mean At All to imply a CB be judged entirely by Int totals, not even Champ Bailey. To not be good/ focused upon a particular thing is not to be bad at your role, your role may preclude that particular thing from occurring. Furthermore, I am in no way suggesting we throw out the existing system and metrics, they work and take into account many very important things, and I am in no way trying to diminish the importance of YPA totals, etc, simply proposing investigating if they could be improved in their predictive success rate by incorporating TOs (or rather, the result of them, in terms of field position obtained).

  40. JJB says:

    David,
    D'oh. I didn't see your response to my post, so my previous post seems pedantic. My apologies.

  41. Anonymous says:

    See, that's the thing- D Int rate is the wrong variable here. We're talking TO differential between the two teams- as noted in the article, ints are "thrown" by the offense as much (if not more so) than grabbed by the defense. When looking at week 9 vs week 1-8, you don't just look at the Defense, you look at what the offense has done as well in terms of generating turnovers, and with those statistics combined attempt a more predictive measure of turnovers. My whole point is that I think people are failing to properly analyze the data present. In week 16, for example, the Bills (averaging 1 pick thrown per week) vs the Patriots (averaging 1.5 picks per week) there may be 1 pick, there may be 3, but it is more likely than not there will be at least 1, and specifically more likely than not that there will be more ints thrown on that side of the field than by the pats (averaging .33 picks thrown per week, buffalo averaging .66 picks picked up per week). Meanwhile, the browns average 1.28 picks per week, but are against the ravens, who average .57 picks per week, so perhaps they won't get any picks this week at all. I will be doing the analysis work throughout this week in terms of the 2010 season TO ratios, and then split by ints and fumbles, for weeks 1-8 vs week 9, taking matchup into account the way I propose, and I'll let you know what my findings are. Meanwhile, we'll see what happens this week with the top 6 teams in TO differential vs lower tiered teams.

  42. David says:

    I understand the explanatory/predictive difference. I was saying I believe that turnovers must be somewhat predictive for two reasons: 1. As you say, there are very indicative or explanatory of who will win the game, and 2. I'm hypothesizing that teams can persistently excel in turnover differential.

    The linked article jibes with my earlier hypothesis that the ability to limit turnovers is likely more persistent than the ability to create them. I also understand that passing efficiency may be more persistent, but I think that turnovers are likely so strongly predictive that they are likely still equally important to a model. I guess my only remaining question is how turnovers are factored into the model today because I feel like they have been downplayed but I don't see why they would based on the correlations linked to.

    One other slightly less related thought: The conversation about the low interceptions for Asomugha and Revis on a micro scale may also help to explain the low correlation for defense interception rate on a macro team scale. Just like once a cornerback is identified as elite they stop getting interceptions because they are no longer thrown at but the rest of the defense benefits, once a team as a whole gets a reprutation for interceptions it may lead to overcautiousness from opposing offenses, which will reduce their interceptions going forward, but may lead to their defense continuing to excel. If that is the case, and again I'm just throwing it out there, even though defensive interception rate isn't persistent, it might still be predictive of future success. Probably something you can test if you feel the urge.

  43. JJB says:

    anonymous-
    If you find the "missing link" (ie the stat computation that makes the influence of turnovers on NFL football significantly more predictable), football sabremetricians across the country will be in your debt.
    It is a worthy endeavor. And I, for one, hope you find it. And make a fortune betting on it before you reveal it to the rest of us.
    I'm not being sarcastic. I'm serious. If you find a way to predict turnover outcomes reliably, that'd be the sabremetric equivalent of solving Fermat's last theorem.

  44. Anonymous says:

    The model includes offensive and defensive passing and running efficiency, offensive turnover rates, defensive interception rates, and team penalty rates.

    Of course, the different factors of the model are weighted differently.

    This actually has me thinking...with such low self-correlation for defensive INT rates, why even include it in the model at all? Is the .08 figure from the 06 and 07 seasons low? I'm assuming it's given very little weight relative to other other factors, but when does a stat self-correlate weakly enough to merit exclusion from the model?

  45. David says:

    Do you know how they are weighted differently? Is it strictly a matter of running past years through the model and seeing what comes out, or has Brian exerted some element of influence over the weightings?

  46. Brian Burke says:

    Yes. The .08 number was a bit fluky. Using many more seasons it's about as high as offensive ints.

    The regression produces the weights. But the input variables are regressed themselves as well. This is to account for the small samples early in the season. Otherwise, the model becomes rather overconfident.

  47. James says:

    Look, we have a community site here that will publish any study you do. If you're convinced there is a way to predict interceptions, turnovers, turnover differential, whatever, find it and post the findings.

Leave a Reply