1. This site uses cookies. By continuing to use this site, you are agreeing to our use of cookies. Learn More.

Thoughts on the Dolphins/Cowboys Game

Discussion in 'Miami Dolphins Forum' started by KeyFin, Sep 22, 2019.

  1. mlb1399

    mlb1399 Well-Known Member

    1,896
    1,032
    113
    Mar 6, 2010
    I think it’s really hard to evaluate that when you only have one quality starter at OL, no TE, middle of the road RB’s and WR’s who are average at best. Some of these positions may improve but as it stands now, it would be hard to evaluate him under normal franchise QB standards because true franchise QB’s don’t have this much garbage around them.

    Having said that, I don’t think Rosen is a franchise QB, yet.
     
  2. tirty8

    tirty8 Well-Known Member

    828
    820
    93
    Jan 2, 2016
    I've been kinda toying around with this whole idea for awhile. Some of the drops we saw these last two weeks were passes Tom Brady would have thrown. I just don't see Tua making better throws (because they were perfect). I really think Rosen showed a lot with very little support.

    My thoughts are this: If we have the first pick in the draft, we should take Tua. Let them compete, and maybe we will have a Rivers/Brees scenario, and hopefully, at worst, one of them pans out as a franchise QB. I look at the Eagles as the blueprint the year they had Bradford, Wentz, and Daniel all on their roster. Basically, the mentality was, we are gonna try six ways to Sunday to find our franchise QB. If they are both good, it is a high quality problem.

    Now, if for some reason, Rosen plays us out the one pick, things get tricky. I am kind of like a few others on this board. I am not sure if Tua is a generational talent. Alabama players are super hard to evaluate because there is som much talent on the team. Watching him have all day to pass simply won't exist in the NFL. Depending on how the other QBs pan out, I would be open to not drafting a QB this year, and building around Rosen. Now, here is the sneaky part. I would try and use that Steelers' pick to move back this year, and pick up another first round pick next year. Having three first rounders in 2021 and 2 seconds gives us the chance to make someone an offer they can't refuse for Trevor Lawrence. I do think that he is a generational talent, and worth mortgaging the farm for.
     
    Puka-head, Surfs Up 99 and resnor like this.
  3. mlb1399

    mlb1399 Well-Known Member

    1,896
    1,032
    113
    Mar 6, 2010
    To add to my previous comment on how hard it is to evaluate these QB’s under these conditions, I thought I would look at Fitzpatrick’s stats and compare:

    His completion % is 8.1% lower than his career average
    His TD% is 2.5% lower
    His QB rating is 36% lower

    Now he’s older and maybe his abilities have declined but that’s pretty substantial.
     
    Bumrush and Irishman like this.
  4. KeyFin

    KeyFin Well-Known Member

    6,499
    8,514
    113
    Nov 1, 2009
    But it wasn't a "what if"...he made the reads and threw the TD passes. To date, Rosen has thrown 5 TD passes yet his stat line shows zero. I think it's more important to judge him on actual actions since there was absolutely nothing more he could have done in those situations.

    You say we'd have to apply the "what if" criteria to everyone....that's EXACTLY how QBR should be calculated. For instance, if a blitzing safety sacks a QB from the backside going in untouched, it shouldn't lower that QB's stats on the day since it had nothing to do with him. If a receiver tips the ball straight up in the air and it gets picked off; same thing. I could understand that in 1960 but with today's technology, that stuff can be 100% automated and give a precise, objective QBR based on team actions.

    I'd argue the same thing for a 1 yard screen pass that goes 90 yards for a TD- the rating shouldn't be calculated that the QB threw a 90 yard pass. Yet that's what we get because nobody has revised a truly outdated formula.
     
    resnor, Bumrush and Irishman like this.
  5. cbrad

    cbrad . Club Member

    7,602
    8,999
    113
    Dec 21, 2014
    Of course it's a "what if" lol.. the TD's did not actually happen. Rosen has thrown zero TD passes so far as a Dolphins QB. That is a fact:
    https://www.pro-football-reference.com/players/R/RoseJo01.htm

    This is exactly the kind of thing ESPN's QBR tries to do – apportion credit among different players for a given play – and it just turns into an exercise in making subjective assumptions that they have to hide (they don't publish their formula) so that people don't criticize their assumptions lol. So while in theory it would be nice to do what you're suggesting, it can't be done with the statistics we have.

    It IS possible to do it theoretically though if you had much richer statistics. That's what happened after moneyball in baseball. Contrary to some predictions, initially MORE scouts were hired to collect more data. It's only after that phase passed that some teams (notably the Astros) started replacing scouts with analytics. Football stats right now are pretty sparse so that revolution hasn't happened yet.
     
    cuchulainn and Irishman like this.
  6. KeyFin

    KeyFin Well-Known Member

    6,499
    8,514
    113
    Nov 1, 2009
    Here's the stat line I would like to see and it's all 100% possible TODAY-

    Start with the standard. Johnny B threw 21/34 for 244 yards, 2 TD/1 INT/3 sacks

    Then add in yards after the catch and penalize a QBR once it reaches over a percentage of total yards gained- 21/34 for 244 yards, 2 TD's/1 INT/3 sacks, 79 YAC

    Then add in drops, which would be well placed balls that weren't forced drops due to a defender swatting the ball, laying the receiver out, etc. Dropped circus catches don't count here and neither do diving catches, stretched out toe taps on the sidelines, etc. We only want clean, accurate passes that were dropped. For calculation purposes only, count these as catches and corresponding yards- 21/34 for 244 yards, 2 TD's/1 INT/3 sacks, 79 YAC, 4/82 yards dropped.

    Now do the same thing with dropped TD passes. If one of those 4 passes for 82 yards above was a TD, count that TD in calculations- 21/34 for 244 yards, 3 TD's/1 INT/3 sacks, 79 YAC, 4/82/1 TD dropped.

    Now do that for sacks- and these only count if they happened within 1.5 seconds of snapping the ball AND the QB did not see the blitzing player advancing. We would erase these from the final QBR calculations- 21/34 for 244 yards, 3 TD's/1 INT/3 sacks, 79 YAC, 4/82/1 TD dropped, 1 team sack error.

    One more thing to add here; interceptions that weren't directly the QB's fault, plus dropped interceptions by the defense. We only count either of these if it's clearly WR/CB error either way...the sorta/kinda plays get left out- 21/34 for 244 yards, 3 TD's/1 INT/3 sacks, Adjusted- 79 YAC, 4/82/1 TD dropped, 1 team sack error, 1 INT dropped.

    We also count all running yards as passing yards- but I thought of this after the fact so we're gonna say Johnny ran for zero so I don't have to change the stat line.

    Now give me a true accuracy stat...not catches, but how many passes were placed where they allowed the receiver to make a play after the catch. I don't care if the ball was caught or not..or what the receiver actually did. Was the ball good placement or not? Also, we don't count deep passes unless it hit the receiver in stride...jump balls don't play into this since they're 50/50 by design- 21/34 for 244 yards, 3 TD's/1 INT, Adjusted- 79 YAC, 4/82/1 TD dropped, 1 INT dropped, 88% ball placement.

    Finally, how many errant throws did Johnny make? These are passes that could have been picked off, poor reads, checking down when the primary read was open, etc. Just give me a number- 21/34 for 244 yards, 3 TD's/1 INT, Adjusted- 79 YAC, 4/82/1 TD dropped, 1 INT dropped, 88% ball placement, 5 errant throws.

    Why the placement and errant throws? That just gives us a better idea of accuracy and decision making. We figure out penalties for each bad toss or bad decision.

    So Jonny B's adjusted QBR for calculation purposes only is 25/34 for 326 yards, 3 TD's, 2 INT's, 2 sacks. We penalize that stat line since 24% of the yardage was after the catch (we'd need some math to figure out how much). Then we penalize for those 5 errant throws (need more math to figure out how to calculate) or give a bonus if it's less than maybe 8-10% of total throws.

    That would give us a true, unbiased QBR regardless of what's on the scoreboard or who won the game. Also, I wouldn't have my own formula and rating system...we'd plug that adjusted number into standard ratings and that's what the answer is. This would be so simple to do in 2019 and it's a completely unbiased way to objectively judge a QB.
     
    Last edited: Sep 24, 2019
    Irishman likes this.
  7. cbrad

    cbrad . Club Member

    7,602
    8,999
    113
    Dec 21, 2014
    Can't even remotely do that with stats today. Where can I find the "sacks as a function of time" stat (e.g., within 1.5 seconds?) or the "errant throws" stat or the "clean/accurate passes that were dropped" stat or "good ball placement" stat, etc.. None of those are available today. Like I said you'd need to get people to obtain those stats first.

    Either way, that wouldn't give you an "unbiased" estimate for a number of reasons: 1) you're creating arbitrary cut-off points (penalize once it passes some arbitrary threshold) and of course 2) you don't know how much to penalize X in condition Y, etc...

    No.. what you have to do is to create a model with all the stats you want where you allow the relationships among the stats to change in many possible ways (the "many possible ways" is parametrized so that you've precisely defined the space of possible models) and pick the model that best predicts win%. This is called structural equation modeling. THAT is unbiased.
     
    Surfs Up 99, Pauly and The Guy like this.
  8. The_Dark_Knight

    The_Dark_Knight Defender of the Truth

    8,678
    7,080
    113
    Nov 24, 2007
    Melbourne, FL
    Rosen is playing with nothing BUT back-ups! To put it in perspective, Rosen is a high school quarterback, playing with junior high school linemen, receivers and running backs...against a state championship high school team.

    What do you expect?
     
    resnor and mlb1399 like this.
  9. cbrad

    cbrad . Club Member

    7,602
    8,999
    113
    Dec 21, 2014
    From a true franchise QB.. a lot better. A bunch of his passes were off target, tipped or late in the 2nd half. That wasn't his surrounding cast causing all that. That was the inability of the QB showing itself.
     
    The Guy likes this.
  10. Tin Indian

    Tin Indian Rockin' The Bottom End

    6,878
    2,633
    113
    Feb 10, 2010
    Palm Bay Florida
    At this point I can't tell you if Rosen is the guy or not. I don't think anyone without bias can say one way or the other. We are just going to have to see how this plays out. I find myself agreeing with both sides of this thing, his supporting cast has been pitiful and are giving him no help whatsoever on the field. He has thrown a good number of passes that should have been caught for TD's but the fact is he hasn't thrown ANY at this point. Do you ding him for that or do you give him credit for good passes and reads? As is pointed out above TD passes are dropped all the time, do we count those for those QB's too? It's the proverbial slippery slope if you start going down that one.

    At this point I say we draft Tua, been watching every game of his and I've changed my attitude on him, ditch Fitzpatrick in the offseason and keep Rosen around. Let them duke it out for the starters job. I've watched a bit of Trevor Lawrence this year and I haven't been near as impressed as last year. I don't know if something is wrong or if teams have enough tape on him now. Hard to say.

    Herbert, Just no. I have no faith in Oregon QB's in the NFL just as many others have no faith in Alabama QB's. Fromm doesn't have NFL arm strength. Right now I just don't see anyone else to consider. There will be more as the season plays out and we will have endless debate until the next draft I'm sure.
     
    Irishman and mlb1399 like this.
  11. The Guy

    The Guy Well-Known Member

    1,880
    868
    113
    Oct 1, 2018
    But even if that were true — and again it might be — the way he played against the Cowboys, as well as his previous performance in his NFL career, doesn’t tell us whether your “high school” analogy is correct.

    It could be that you could surround him with 10 exemplary players on offense, and he could function much like a college quarterback and be completely non-competitive in the NFL. The fact that there is inferior talent around him right now doesn’t tell us what his ceiling is, even with better surroundings. He could be inferior talent as well, just like his surroundings!

    We can’t just make an upward adjustment for him based on an unknown quantity.
     
    Tin Indian likes this.
  12. KeyFin

    KeyFin Well-Known Member

    6,499
    8,514
    113
    Nov 1, 2009
    I'm not pretending to have the answers with the arbitrary cutoff points, etc. Maybe it makes more sense to just straight up penalize for every ugly throw and reward for every particularly good one. I honestly don't know. This has been an idea I've kicked around for some time now- its certainly not a dissertation. =)

    And let me clarify- YOU can't do that today based on stats around the net. What I meant was that the technology is readily available to make an app that could generate the formula and the stats we'd need. It would take a coder and a math guy to figure it out, but the good news is that the technology would sell to the league for a pretty penny. You'd use AI to scrub each play on video, make determinations and spit out a number....that part would be hard if you built it from scratch (which may not be necessary). The rest is simple coding though.
     
  13. Tin Indian

    Tin Indian Rockin' The Bottom End

    6,878
    2,633
    113
    Feb 10, 2010
    Palm Bay Florida
    On the Oregon QB side note. I just looked at the list of Oregon QB's and you have guys like Herbert, Mariotta, Kellen Clemens, Joey Harrington, A.J. Feeley and Akili Smith. You have to go all the way back to 1970-1972 and you get Dan Fouts. In fact that's all you get.

    Same exercise for Alabama gets you Tua, AJ McCarron, Greg McElroy, John Parker Wilson, Brodie Croyle. YOu can go back further there and you can get Richard Todd, Ken Stabler (my fav of the bunch) and of course Joe Namath.

    I think its safe to say Oregon isn't a powerhouse of successful NFL QB's and neither is Alabama unless you go way back to the 60's and 70's.
     
    freeperjim and Irishman like this.
  14. cbrad

    cbrad . Club Member

    7,602
    8,999
    113
    Dec 21, 2014
    I told you how it should be done. Specify a space of possible models and then choose the one that predicts win% best (something called a maximum likelihood estimation can be used to make that selection automatic).

    AI in 40 years might be able to do that. Nothing today. Have to come up with operational definitions of every new stat you want and then pay people to look through game tape to collect those stats. I'd definitely do that if I owned a NFL team.
     
  15. KeyFin

    KeyFin Well-Known Member

    6,499
    8,514
    113
    Nov 1, 2009
    I'm 100% positive you're wrong- Google has had self-driving cars for a decade now that recognize people and other moving objects in real time. We'd need a much more basic technology to view thru video and measure "contact points" between the ball, the receiver and the defender.

    To the human eye we call something a "perfect throw" but AI can tell us things like the receiver reduced speed by 43% and changed direction by 140% to haul in the pass. There's not an application of using existing tech that way, but it's already possible without starting from scratch. Heck, CCTV does something similar tracking license plates...or facial recognition when you walk into stadiums.
     
  16. Big Phin

    Big Phin Member

    91
    76
    18
    Dec 29, 2011
    Our future is playing at Clemson... I hope.

    Now, I know what you guys are saying about Rosen, but to compare him to Daniel Jones isn't really all that fair, imo. Put Daniel Jones behind this line and I guarantee we are talking about Jones being a career backup, he sucks, etc. JMO, so I guess we'll see. Hopefully the receivers can just catch the damn ball in critical situations.
     
    KeyFin likes this.
  17. The_Dark_Knight

    The_Dark_Knight Defender of the Truth

    8,678
    7,080
    113
    Nov 24, 2007
    Melbourne, FL
    Hmmmm, Rosen being sacked 7 times in 3 games (only starting one of those), knocked down who many times and hurried even more than that wouldn't have anything to do with inaccurate throws, would it?

    I've said it before and I'll say it again, you could put Brady behind our line and he'd be running for his life.
     
  18. cbrad

    cbrad . Club Member

    7,602
    8,999
    113
    Dec 21, 2014
    No you're 100% wrong.

    No one is close right now to creating a self-driving car that needs NO human intervention. The first problem is object recognition. Accurately detecting what kind of object is in front of you is still too challenging a problem. In fact, no computer vision algorithm is currently good enough to show you the boundaries of all leaves in a tree (a famous long-standing problem in computer vision that no one seems to be able to solve despite half a century of effort). Stuff like one leaf occluding another just makes the problem too difficult for now.

    The second problem is predicting the behavior of objects. A human crossing a street is different than that human on a bicycle crossing the street. Anticipating what that person will do (or what another driver will do) is WAY beyond the capabilities of AI today.

    All the "self-driving" vehicles you'll see in general driving conditions (so not in cases where the path is highly predictable) are not truly "self-driving" and require a human to be there. That will be true for a LONG time.

    Regardless, that object recognition problem makes what you're asking for impossible for now. No AI can accurately identify from video each individual player or where their limbs are, etc... Tracking license plates is MUCH easier, but even there if you just put a tiny bit of graffiti on it the machine has no clue what it is looking at while it's no problem for a human. Even with facial recognition, there's a limit. I think currently those things work out to 60 degree from frontal view. NONE work reliably from a side view for example, which is again easy for humans.

    AI is just one of those things where unless you understand the basics of machine learning and have actually worked on these problems (I have in computer vision at least) you simply have no good idea of how difficult the problems are. Murphy's Law applies: anything that can go wrong does go wrong.

    So you are 100% wrong.
     
    Irishman likes this.
  19. cbrad

    cbrad . Club Member

    7,602
    8,999
    113
    Dec 21, 2014
    Nope.. watch the 2nd half of the game again. You can't blame the WR's or OL for a lot of those errant passes.
     
    Agua likes this.
  20. Irishman

    Irishman Well-Known Member

    358
    382
    63
    Oct 16, 2017
    High Point, NC
    I like the information you bring up to support your opinion as opposed to those say the results were bad so we are useless. That may probably be more apt for these posters then our players.
     
    resnor and KeyFin like this.
  21. Irishman

    Irishman Well-Known Member

    358
    382
    63
    Oct 16, 2017
    High Point, NC
    I don't object to keeping him out next week, but not to "teach him a lesson". We should let other players show us if they can do better, since we aren't showing a lot of strength in these positions.

    This way we aren't being negative by singling anyone out. We are just looking for the best possible player out of who we currently have. If someone whines because of this, we don't really need them on this team!
     
    resnor likes this.
  22. Irishman

    Irishman Well-Known Member

    358
    382
    63
    Oct 16, 2017
    High Point, NC
    "a really good QB"? Don't you mean a really good front line?
     
  23. KeyFin

    KeyFin Well-Known Member

    6,499
    8,514
    113
    Nov 1, 2009
    I work for one of the world's leaders in business/corporate AI and you'd be amazed what they're accomplishing with the IoT (internet of things). I could mail a package to Toyoko and then have someone there forward it to your house...and tell you at any precise second where that package is within 2 or 3 feet (that's RFID technology, which is the size/weight of a postage stamp and doesn't require electricity). The video technology has skyrocketed over the past 2 years as well...you're just starting to see it in NFL games now. But within a few years, we'll be able to run video at 30 frames a second on 100x zoom...all in HD quality. That tech exists today and it's pretty amazing, but the trickle down to consumer level takes years (and sometimes decades).

    By the way, Google's 1st self driving car hit 1,000,000 miles in like 2014. They started back around 2005-2007 and absolutely nobody was interested because they thought consumers wouldn't like it. The tech still isn't perfect and it's why it hasn't went fully mainstream, but most loaded 2018 models or newer are only a hair away from full self driving. In my Subaru, I set cruise control and the car stays in the lane, automatically adjusts for speeds, fully stops if it senses a potential accident, etc. It's not perfect but it's pretty close...and Subaru has been developing their tech for 4 years. Again, Google is approaching year 15 and it's far superior to Tesla.

    Between the enhanced video and RFID alone, I can think of a dozen ways to give you pinpoint measurements on the football field. Start with RFID tags in footballs. helmets. cleats and in player gloves....boom, you have 1,000 statistics that you never dreamed possible before. Did he break the plane or step out of bounds? You could have a definitive answer by spending an extra 60-80 cents per uniform.

    Anyway, I absolutely do know what I'm talking about here and tech startups are hopping on this kind of bandwagon every single day. What I described does not exist, but I believe it could absolutely be developed over the course of a few years.
     
    resnor likes this.
  24. mooseguts

    mooseguts Well-Known Member

    273
    310
    63
    Jan 12, 2018
    No I mean a really good QB. I'd be ok with a decent frontline.
     
  25. cbrad

    cbrad . Club Member

    7,602
    8,999
    113
    Dec 21, 2014
    None of the things you described requires AI. They may use AI for data mining, but it's not necessary to make what you described work.

    Yeah.. and it's FAR away from anything approaching a true "self driving" vehicle. Those vehicles make all kinds of embarrassing errors, like not understanding traffic signals, not knowing when to turn into or out of traffic, and sometimes just inexplicably coming to a stop.

    Yes, this is a practical solution. RFID would definitely allow you to collect tons of useful stats. But analyzing video will still mostly have to be done by humans.

    You don't know what you're talking about w.r.t. AI. That is patently clear. What you said about IoT and RFID is correct however. Oh, and yes you're right there are TONS of startups promising all kinds of things, but delivering the promised product? That's another story.
     
  26. KeyFin

    KeyFin Well-Known Member

    6,499
    8,514
    113
    Nov 1, 2009
    I used the term "AI" because I wasn't sure how familiar you were with hardware or technologies...that's more the generic term laymen can understand since all that gets dumped into artificial intelligence at the consumer level. So you're right, I'm not talking about AI in that sense....the only actual AI necessary would be determining if a ball was a catch or a drop, and the most likely reason why. Everything else can be calculations after the fact...until AI is far enough along to provide it live.

    Have you been to a tech expo lately? The tracking technology is absolutely crazy and it's being applied today to warehousing, logistics, etc. If you can track a package from a warehouse in Miami to a kid's house in Tennessee (and absolutely everywhere in between), you can definitely track a football.
     
    resnor likes this.
  27. cbrad

    cbrad . Club Member

    7,602
    8,999
    113
    Dec 21, 2014
    Understood. Yeah.. AI = machine learning in almost every case.

    The only way you could use machine learning to determine if a ball was caught, catchable or not etc.. would be to train the machine on tens or hundreds of thousands of training images that span the entire set of possibilities the machine would ever encounter, each of which would be "graded" by a human expert. So even using machine learning you'd still need to have humans look through probably the entire history of the NFL and grade the ball as "catchable" or so before the machine could learn it.

    And whether it actually could learn it is hard to predict until you try.

    It's amazing yes. btw.. they already use RFID in NFL now with those next-gen stats. Naturally the data isn't readily available to us.
     
    KeyFin likes this.
  28. resnor

    resnor Derp Sherpa

    13,399
    7,626
    113
    Nov 25, 2007
    New Hampshire
    But he didn't care about win%. In fact, in his model, a QB with a terrible stat line could actually have a better QBR than the QB on the other team with a much better stat line.
     
    KeyFin likes this.
  29. cbrad

    cbrad . Club Member

    7,602
    8,999
    113
    Dec 21, 2014
    In order for a model to be "unbiased" (his stated goal) you need to have what's called an "objective function" which basically states what it is you're trying to maximize or minimize. And in sports that's always win% because that's the ultimate "objective" of the game.

    And just because the model maximizes win% doesn't mean that the QB with the better "stat line" (we're talking raw stats here.. or even the commonly used stats like passer rating) has the higher rating. Those two things are independent of each other because everything depends on the relations between stats in the model.

    So your concern is unfounded.
     
    KeyFin likes this.
  30. KeyFin

    KeyFin Well-Known Member

    6,499
    8,514
    113
    Nov 1, 2009
    Yup, I've been with the company that set that up fro the NFL for about a decade now (I'm under a serious non-disclosure with them so I can't talk about it). It's still not fully in place but essentially, it all comes down to high speed cameras catching plays from every possible angle. It will only improve each season for the next 2-3 years but eventually, you'll be able to replay any angle from anywhere on the field yourself at home.

    There's a lot that goes into making that happen but it's basically connecting several technologies ultimately with your PC or TV.

    That's correct, but they could process those 100k images in a matter of minutes (and have us verify it over weeks or months). A really cool project that used that technology is https://www.zooniverse.org/ for identifying galaxies in outer space and all kinds of other stuff; basically you have your fans do that for you and it speeds up the process 10,000x. So there's ways of doing that quickly that don't require actual labor hours.

    I nerd out on stuff like this so I'll argue it all day...I'm about 3 years out from my startup tech launch (gotta pay off the house 1st). =)
     
    Last edited: Sep 24, 2019
    cbrad likes this.
  31. resnor

    resnor Derp Sherpa

    13,399
    7,626
    113
    Nov 25, 2007
    New Hampshire
    No man. You gotta stop with that line of thinking. Let's go extreme to illustrate for ease of typing: an hypothetical QB throws 30 passes in a game. Everyone of them falls off the receivers hands. what's his rating? Is that rating actually indicative of ANYTHING that the QB did or did not do?

    So no, I do NOT care about how the stats compare to win%. If I'm evaluating a QB TRULY FAIRLY AND UNBIASED, Then I must look ONLY at the things that occurred or did not occur that were within the QBs control. Therefore, I'm not going to penalize a QB for an on target throw that is dropped. Basically, the QB should be headed in everything he does up until the ball leaves his had. The only judgement made on the QB after the ball is thrown is whether or not it was an accurate throw.
     
    KeyFin likes this.
  32. Pauly

    Pauly Season Ticket Holder

    3,205
    3,148
    113
    Nov 29, 2007
    One of the problems about advanced data collection in the NFL is that you have to know what it is you are measuring before you can measure it, and that’s a human problem not a data problem.

    When you take the act of a QB throwing a pass in the NFL some judgements you have to make include:
    Was it an accurate pass?
    Did the receiver adjust to the ball?
    Did the receiver run the correct route?
    Did the QB make the correct decision in who he passed the ball to?
    How risky was the pass considering the position/velocity/angle of the defender?
    How adequate was the blocking?
    Did the QB hold onto the ball too long?
    Was the ball thrown with appropriate velocity/arc?

    And I am sure there are more variables that people would want to see.

    However the point is that just with these 8 criteria there is a big conversation to be had to define what the parameters should be.
     
    KeyFin and resnor like this.
  33. Pauly

    Pauly Season Ticket Holder

    3,205
    3,148
    113
    Nov 29, 2007
    Well why did the receiver drop the ball.
    - was he under heavy pressure from a defender?
    - was he coming back to prevent an interception?
    - was it a poor throw at the extreme of his catch radius?
    - was it thrown behind him?
    - was the receiver making a big adjustment from the route to where the ball was thrown (as opposed to where it should have been thrown)?
    - was the sun in the receiver’s eyes?
    - was the ball slippery due to rain?
    - was it a pure drop of a good throw?

    You can make one scenario where the majority of the 30 receivers’ drops are due to poor QB play and another where it is primarily the receivers’ fault.
     
    KeyFin and resnor like this.
  34. cbrad

    cbrad . Club Member

    7,602
    8,999
    113
    Dec 21, 2014
    This approach won't work with parsing video images from a football game to see if a WR dropped a "catchable" ball or not. The idea of having a machine classify things and then using human classifiers to improve on it only works if you already have a decent classification algorithm, which won't be the case for something like "catchable" passes or for identifying individual leaves in a tree.

    No one's succeeded at using machine learning to identify leaves in trees for example because edges aren't always edges of objects (shadows create edges, veins in leaves create edges and occlusion creates edges.. none of which are necessarily the edges of an object) and the same thing applies to video images of football players contesting a catch.

    So IF machine learning is to work here all the classification done by humans must be done up front and you can't reduce those labor hours. Of course you could spread the work out over tons of people.

    Regardless.. I doubt any machine learning approach works here. The RFID idea does simplify things because a lot of the classification of objects could be done beforehand, but you can't really use RFID coordinates to solve the problem of "catchable" or not. So my best guess is this is too difficult a problem for any foreseeable future. I said 40 years and that might be lenient.
     
    KeyFin likes this.
  35. cbrad

    cbrad . Club Member

    7,602
    8,999
    113
    Dec 21, 2014
    You're arguing against something you don't understand.

    Let's take your example. A QB throws 30 passes in a game, all perfectly placed, and all of them dropped by the receiver. What makes you think that a model of the game that maximizes the probability of predicting wins/losses would penalize the QB in any way for that? Intuitively, an accurate model – a model that best predicts how the game unfolds – wouldn't. That's what you don't understand. You think that by trying to maximize win% that this somehow automatically means better raw stats => higher QB rating. There is NO such logical implication.

    Structural equation models (SEM), which I referenced earlier, basically specify a set of relations among individual stats. Maybe one particular SEM assigns a weight of zero to passes thrown by the QB while weighting by 1 (maximum weight let's say) passes that were not dropped by the receiver. Maybe another SEM equally weights those two things. Maybe a third conditions one upon the other. Basically you want to mathematically specify a large set of SEM and see which one predicts win% best. That model will most likely be one of the most accurate ones in the set you chose. And if the more accurate model says one shouldn't penalize a QB when the WR is at fault, then that one wins.

    Point is.. your concern is totally unfounded, and if you don't understand that you need to switch gears here from reflexively arguing against selecting models by maximizing win% to trying to understand the method I'm talking about (which admittedly is unlikely to be familiar to people who don't model things scientifically).
     
    KeyFin likes this.
  36. resnor

    resnor Derp Sherpa

    13,399
    7,626
    113
    Nov 25, 2007
    New Hampshire
    Right that was my point. But that's why it's flawed to just sign pass or incomplete to the QB automatically. There are plenty of completions that are simply an amazing play by the receiver to get to a ball that shouldn't have ever been caught. Seems also fraudulent to attribute those completions to the QB.
     
  37. resnor

    resnor Derp Sherpa

    13,399
    7,626
    113
    Nov 25, 2007
    New Hampshire
    You didn't answer the question. What would the QBs rating be, as passer rating is currently figured, and would that necessarily accurately reflect the actual play from the QB?

    The model you're espousing isn't telling which QB actually played better. It's telling us which QB was more successful. Those are not necessarily the same thing.
     
  38. cbrad

    cbrad . Club Member

    7,602
    8,999
    113
    Dec 21, 2014
    Oh.. you mean the current formula for passer rating? If you have 0 completions in 30 attempts with 0 TD's and 0 INT's you have a 39.58 rating. Obviously not a good formula for measuring QB ability when all the fault in on the WR.

    But that has nothing to do with what I described in post #69 which you argued with.

    It's telling you how well the QB played according to the most accurate model you can find of the game.
     
    KeyFin likes this.
  39. The Guy

    The Guy Well-Known Member

    1,880
    868
    113
    Oct 1, 2018
    That’s entirely possible, but where we get into trouble is when we think that if Rosen had a better line, he’d play like Brady.
     
  40. resnor

    resnor Derp Sherpa

    13,399
    7,626
    113
    Nov 25, 2007
    New Hampshire
    Lol

    Literally the ONLY thing I was arguing was that correlating to win% doesn't necessarily tell you anything about the actual QB play. He wasn't interested in how QB play affected win% or anything. So there's really no reason too being that variable in.

    I allude to that in an edit to my previous post that I was trying to sneak in before you had responded, so that you would have seen that initially.
     
    KeyFin likes this.

Share This Page