2017 Running Back Report

[clarion contrarion]

^^^ agree w/ valhalla
I pointed this out last year or previously that if the formula is a moving target then it is hard to buy it as a finished product. It seems more a promising experiment & constantly changing the method is wise in that DD is seeking a silver bullet. Owning past misses and not gloating over having method with a couple years data would do much to foster appreciation rather than skepticism.
I feel safe in saying it is an equation derived from speed score /dom. rating burst /agility and market share , not sure I even care as all these formulas are only a partial tool in evaluation IMHO.
Just when I was ready to capitulate and get on the foreman band wagon dude in the foreman thread (an obvious longhorn fanboy) pointed out it was the baylor /art briles offense in which he piled up massive # and every single one of those baylor/system backs have sucked in the league jackie battle from briles houston years is the valedictorian of that summer school class. ganaway seastrunck finley ... zilch !
Battle registered 4.39 at 238 lbs on his pro day from the same basic offense sound familiar ? , I may be wrong have been before and will be again but his tape looked so blah to me watch that texas tech tape awesome stats but those guys couldn't tackle a sloth running through 3 feet deep snow they were so uninterested in tackling him. The whole defense should have been stripped of their scholarships and the DC should have worn a mask to collect his paycheck.
If he can convince NFL LB & SS to take a pass on tackling him maybe he will be great same as fournette but something tells me it won't go like that and all those runways and wide open alleys they had in school will be a bit harder to find from here on out.
Just curious how many of the people that love foreman think jeremy hill is awful because if he even does what jeremy hill has thus far I will be stunned.

[Cameron Giles]

^^^ agree w/ valhalla
I pointed this out last year or previously that if the formula is a moving target then it is hard to buy it as a finished product. It seems more a promising experiment & constantly changing the method is wise in that DD is seeking a silver bullet. Owning past misses and not gloating over having method with a couple years data would do much to foster appreciation rather than skepticism.
I feel safe in saying it is an equation derived from speed score /dom. rating burst /agility and market share , not sure I even care as all these formulas are only a partial tool in evaluation IMHO.
Just when I was ready to capitulate and get on the foreman band wagon dude in the foreman thread (an obvious longhorn fanboy) pointed out it was the baylor /art briles offense in which he piled up massive # and every single one of those baylor/system backs have sucked in the league jackie battle from briles houston years is the valedictorian of that summer school class. ganaway seastrunck finley ... zilch !
Battle registered 4.39 at 238 lbs on his pro day from the same basic offense sound familiar ? , I may be wrong have been before and will be again but his tape looked so blah to me watch that texas tech tape awesome stats but those guys couldn't tackle a sloth running through 3 feet deep snow they were so uninterested in tackling him. The whole defense should have been stripped of their scholarships and the DC should have worn a mask to collect his paycheck.
If he can convince NFL LB & SS to take a pass on tackling him maybe he will be great same as fournette but something tells me it won't go like that and all those runways and wide open alleys they had in school will be a bit harder to find from here on out.
Just curious how many of the people that love foreman think jeremy hill is awful because if he even does what jeremy hill has thus far I will be stunned.

There is no rational argument that Foreman was a system back. Hill and Foreman aren't even similar players. Foreman is stronger, faster and more athletic than Hill.

[onetwothree]

Just another variable in the equation. Makes for a good tiebreaker when choosing between multiple guys who all seem about the same.

Funny anecdote. My friend had the 16th pick (I had 17th) and he was debating who to take between 4 RBs (Mack, McNichols, Jamaal Williams and he didn't name the 4th) and told me he'd let me have my pick of whoever I wanted and he promised he wouldn't draft whoever I name but that I'd be doing him a favor by eliminating 1 name from his choices. I didn't believe him to be so generous so I told him I didn't know who I wanted. He takes Mack and I turn in my pick of Foreman. He texts me saying he never would've guessed that'd be my pick! Didn't fit my typical draft strategy, wasn't someone he had considered.

Everyone scouts and values guys differently. Maybe he'll be right, I'll be wrong or vice versa. Or we'll both be right or we'll both be wrong but that's just part of the fun. Thanks for the information! There's a lot to digest and I can see both sides of the argument but whether I agree or not with the results, props to you for sharing! If it was somehow possible to come up with the perfect system, we definitely wouldn't be here arguing over the data and trying to figure out a way to use it on the lottery. All in good fun!

[Kramer]

Just want to say a huge thank you to DD! This was an excellent read and I am excited to see what these guys will do or not do. As someone who drafted Fournette, Mixon and Foreman, this is all great to hear.

I think anytime someone sticks their neck out on some players, it should be expected that there will be questions on the process used and how one can be so sure. You don't have to listen to what DD is saying though. He has a system that he's worked on. Don't let this be the only thing that influences your decision as with anything, but if you're so and so on some second tier RBs, maybe it could be a tiebreaker.

Anyway, thanks again DD for all your hard work into this! There are plenty of us who are very, very thankful that you provided this to the community.

[kadun2]

Just want to say a huge thank you to DD! This was an excellent read and I am excited to see what these guys will do or not do. As someone who drafted Fournette, Mixon and Foreman, this is all great to hear.

I think anytime someone sticks their neck out on some players, it should be expected that there will be questions on the process used and how one can be so sure. You don't have to listen to what DD is saying though. He has a system that he's worked on. Don't let this be the only thing that influences your decision as with anything, but if you're so and so on some second tier RBs, maybe it could be a tiebreaker.

Anyway, thanks again DD for all your hard work into this! There are plenty of us who are very, very thankful that you provided this to the community.

Well said, Kramer, I agree. Thanks for the effort DD, whether you're right or wrong. It gives us something else to use in our research and has provided some good reading with the nerves you have struck on some of these forum guys

[ninotoreS]

I did my best to clarify which are the facts and which is my opinion.

You'll clarify it best by detailing the model you're using. otherwise, it all just seems like opinion and hindsight masquerading as methodology.

I firmly believe that if LeSean McCoy was healthy and participated at the combine that he would have qualified for the Tier 1 list.

I don't think so.

- Pro-Day 40s are nearly always faster than Combine 40s, and McCoy's healthy Pro Day 40 was 4.50 (I'm seeing elsewhere that it was 4.55). Thus, a Combine 40 more than likely would've diminished his standing in your model even further.

- McCoy was right at 200lbs coming out of college. Presumably that would've also hurt him in your model.

- While Pro-Day 40s are suspicious, long and broad jump scores are reliable; there's no whistle to cheat, no generous hand-timing by a friendly party. What really kills McCoy in any metrics-predictive model are his long and broad jump scores, which were utterly abysmal.

[AussieMate]

I may he wrong but I believe DD has explained it in previous years threads (I think you can search for them). I could be wrong but I think someone mentioned it earlier in this thread. I'm lazy so I can't be bothered looking myself. Also I've wondered if running numbers like this is trying to hit a moving Target for example the NFL changing to more pass centric along with committee back approach. Is there any chance that a back 10 years ago wouldn't be as good again st defense structures today? I've only been following NFL for about 7 years now so I could be way off. Also rule changes for lowering the head etc. Just a thought that has crossed my mind.

[Dynasty DeLorean]

Correct me if I misunderstood you, but you told us that you update the model if a particular RB style that was excluded proves effective, and likewise if a previously included RB proves ineffective.

No, that is incorrect. I looked at all the stud Rb's from the past decade or so and designed a model from what I found to be "success indicators." Now when you interpret the model I created, it suggests stud Rb's could come in several different forms, some of which there is no precedent for (at least not in my data base). It's kind of like if you dug up some dinosaur bones. You know what the dinosaur you dug up looks like but then you start theorizing about what other dinosaurs may have looked like. So that is where the test-cases come in, and since I have no historical context for them (which is a strong indicator that they busted, otherwise they'd still be in the league and I'd have data for them) I need to weed them out as they pop up. So to answer your question directly, I've never included new profiles. I've only had to exclude the unproven theorized profiles as I got data for them. Le'Veon Bell is pretty much the last test-case (which is noted).

If Foreman does hit, do you maintain your skepticism or do you continue to insist that scouts know best?

The really bizarre thing about this thread is we've all had this same exact conversation 2 years ago regarding David Johnson. If that didn't change (or open up) anyone's mind, then I don't know what will.

I pointed this out last year or previously that if the formula is a moving target

It's not a formula and it's not a moving target.

- Pro-Day 40s are nearly always faster than Combine 40s

Yes, and I adjust for that.

[ajmyk]

Thanks for sharing

This is a technical question, but I assume that when you started this project, you looked at previous players stats and production to make your model, so they're basically the Training Set.

Do you actualise the model to include in the training, the most recent productions and the new players or do you keep it as it is ?

Also if you get, say, a 80% success rate on the Training Set of players in the studs category, you can't except this rate to apply to new players unless you get a lot more data. I have to assume there is a lot of overfitting here, but that's OK you CAN'T do better with this kind of data.

Still, sound like your track record is great for especially if you predicted DJ. (Sadly, I went to check and saw that you deleted your opening post in the 2015 thread, not that i doubt it, the reactions are there, but was curious to see the write up)

Can you explain the bolded section a bit, not sure if I quite understand.

Let me answer the rest as best I can. I looked at all the previous "stud" Rb's to see what they all had in common and/or what metrics were most predictive of success. Once I found that there are indeed numbers which are predictive of NFL success as a 3-down workhorse rb, it was quite a long period of trial and error until I was able to build a system that actually worked to predict incoming rookies. I tried formulas, points systems, nothing worked and those types of systems would never work to the degree of accuracy we need. Even though it was frustrating, I did gain valuable information from that period of trial and error and eventually came up with what is essentially a query system, and it's not nearly as complicated as you think. I assume by over-fitting you mean the process is lengthy and convoluted to "fit" all of the stud players onto one list while excluding everyone else, but that's not the case at all. In my estimation your assumption is wrong. It works because the requirements are very strict, not because there are a lot of them. Once you figure out which numbers are important and which aren't, it's astoundingly simple really. Tiers 1 and 2 are pretty straightforward, although like I said in the OP it doesn't sit right with me the way Le'Veon Bell in included. He's a bit of a one-off relic from an older version, but i'm content to see how he does first before I start moving him around. The other lists get a little more dicey (Tiers 3 and 4) as you can tell with the percentages much lower than the other lists, and you're assumption of over-fitting might be correct there. But you're also correct in that there's no other way to do them, or maybe there is but I just haven't figured it out yet. I think something is better than nothing. Only time will tell how those lists fair. I've been tinkering with them for quite some time but this year is the first time they're predictive as currently constituted. If I had data for every player over the last 20 years, I wouldn't have to do the "wait and see" approach. Unfortunately I can only work with what I have, which is data for 370 players now, not nearly ideal but thankfully a lot more than what I started with.

Sorry for the late answer.
I understand now that's once you've found the best model by trial and error, you were satisfied with it, and used it from there every year.

You probably not that interested in what i'm going to say, but just a FYI in case.

Overfitting doesn't mean that the model produce only studs with no error, it simply mean that when there is too few exemples and a lot of parameter, then the model fit closely the players you trained your trial and error, and extrapolating to other players may not work.
The more exemple you look into the better the model is going to be.

So basically, the smart thing to do is to include the last two years data, to check how the model worked there, and eventually update it so it'll fit both old players and newcommers the best way. (Including new good seasons of previous players)

[Dynasty DeLorean]

Overfitting doesn't mean that the model produce only studs with no error, it simply mean that when there is too few exemples and a lot of parameter

I understand that. I suppose I didn't make it clear in my last reply. If your assumption was that there are a lot of parameters, you were wrong.

the smart thing to do is to include the last two years data, to check how the model worked there, and eventually update it so it'll fit both old players and newcommers the best way. (Including new good seasons of previous players)

I do include the last 2 years of data, and I do check how the model worked there, and if after a period of many many years it's clear an update needs to happen then I will do that. However, it wouldn't be predictive if I had to update the model every few years. The whole point behind all of this is that what worked in the past will work in the future.

[clarion contrarion]

I realize my previous points were hypothetical ; WHAT if bell went to a system where his dancing and pausing was frowned upon & got benched,2nd what if johnson went to fishers offense /rams oline or even if cj2k didn't get hurt in 2015 pretty sure he was like 3rd or 4th in the league in rushing yards when he broke his ankle and on pace for his best season since like 2010 do we even know yet what david johnson could fully do ? what if gurley went to dallas honestly zeke was awesome last year but the carcass of runDMC ran for 1000 yards the season before behind those guys that tells us how truly special that group is .
So I didn't really expect an answer but the perceptions are what they are but how would they differ if those scenarios had played out ? what does that do to the perception of the model
I guess while admiring your work I think the fact that gurley was top shelf and now isn't and bell wasn't and now is does sort of mean it is a moving target . Pretty sure when I questioned the formula /method before I was given the same answer but now guys have changed tiers so I suppose that is why I continue to seek more clarity .
Thank you much for your effort and I sincerely apologize if my tone was too harsh it is more out of curiosity that I ask rather than malice .I will leave you to your work and do think I would monetize it as most here seem to be fervent believers and you should receive remuneration for your labors.

[Dynasty DeLorean]

I realize my previous points were hypothetical ; WHAT if bell went to a system where his dancing and pausing was frowned upon & got benched,2nd what if johnson went to fishers offense /rams oline or even if cj2k didn't get hurt in 2015 pretty sure he was like 3rd or 4th in the league in rushing yards when he broke his ankle and on pace for his best season since like 2010 do we even know yet what david johnson could fully do ? what if gurley went to dallas honestly zeke was awesome last year but the carcass of runDMC ran for 1000 yards the season before behind those guys that tells us how truly special that group is .

I can't respond to each specific scenario, however I will say that this type of stuff happens all the time. In the long run, good players will overcome bad situations and bad (or average) RB's in good situations will be replaced. On a related note, I'm assuming by your response you haven't seen the "All or Nothing" amazon series on the Arizona Cardinals. It's really good and if you watch it you'd know we'd have seen a ton of DJ this season regardless of what CJ did.

I guess while admiring your work I think the fact that gurley was top shelf and now isn't and bell wasn't and now is does sort of mean it is a moving target .

It's not ideal, however the point I've been trying to get across is that those changes have literally nothing to do with Foreman, DJ, or anyone else on the list. The perception does not match the reality. If I was a charlatan I'd just be fooling myself since I was the one drafting DJ earlier than anyone and I'm the one drafting Foreman earlier than anyone.

Thank you much for your effort and I sincerely apologize if my tone was too harsh it is more out of curiosity that I ask rather than malice .

The problem is you're making statements born from incorrect assumptions which propagates misinformation.

[Concept Coop]

If Foreman does hit, do you maintain your skepticism or do you continue to insist that scouts know best?

I think my argument is being misunderstood. My point was never about DD's stance in a vacuum. I question the model--including claims made bout its predictive ability. In other words, I'll believe the 80% hit rate when there is a track record and sample size that dictates that level of belief. We don't have anything close to that, and won't even assuming Foreman hits.

I'm a data analyst by trade. The most embarrassing moment of my career was having to present the lack of savings from a model that I built to a committee of medical folks after having convinced them to pour thousands and thousands of dollars into its development. Luckily they didn't fire me, but I'll never make those mistakes again. The scope and scale of DD's hypothetical model is massive, and I can't help but question the due diligence. Given enough time and resources, we can find all kinds of correlative patterns that look promising, but that doesn't make them predictive.

I also want to point out that it was not my intention to attack anyone personally. There was one implied insult in the thread, and it didn't come from me. Perhaps I was more blunt than is commonly seen on the forum but, even after going back and re-reading my comments, I don't get some of the comments directed at me. I have a feeling that if I stood on the other side of the argument, I wouldn't have gotten them. I won't apologize as I did nothing wrong, but want to point out that I simply wanted to have a constructive conversation about the model. I see now that was ambitious.

[Dynasty DeLorean]

I question the model--including claims made bout its predictive ability. In other words, I'll believe the 80% hit rate when there is a track record and sample size that dictates that level of belief.

Enough time hasn't elapsed to see if it's "predictive ability" actually works yet. This is common sense for every single person reading this thread except for you apparently. Everyone understands the predictive ability look promising but nothing is guaranteed yet.

For players I have the requisite data for, nobody with Foreman's profile has busted. They've all had multiple 1,000 yard rushing seasons. Like Eric said, this is a fantasy football forum not a peer reviewed article. Do you really want me to wait 5+ years before I say anything? Meanwhile people are still drafting David Cobb over David Johnson? And if all you had was a problem with my phrasing, (for example if I should have said "100% of the players with this profile have succeeded" instead of "100% chance they will succeed"), then you should have just said so. That would have been constructive criticism. Instead you intended to tear down what you didn't understand. This is why I don't provide details, because people latch onto the 1 thing I didn't phrase quite correctly and go bonkers ruining the thread for everyone else.

I simply wanted to have a constructive conversation about the model.

No you didn't.

[Goirish374]

Darkness,
Regarding note #2: did Bell meet your criteria as run when you ran them for his class or was he a narrow miss?
Thanks.