Dynasty League Football Forum

moishetreats wrote: ↑Fri Jun 07, 2019 10:32 am @Dynasty DeLorean: PLEASE publicly correct if I'm way off or off in a small way. And thank you, thank you, thank you again for sharing your thoughts AND for so fastidiously replying to all the questions, comments, and critiques.

----------

My understanding of this report is that it is looking for correlative indicators to success. By entering many points of concrete data (e.g., height, weight, draft slot, etc.) and then by determining what "success" means (e.g., 1,000 yards rushing), you can being to identify which data correlates to success.

I think that many people on this thread are confusing correlation with causation. Just because a player's data does or not fit the successful profile (i.e., correlate) does NOT mean that that player will or will not succeed (i.e., causation). DD's report isn't telling you that Player X will be successful and that Player Y will not be successful. Rather, DD's report is telling you that Player X's concrete data gives him a high, middle, or low likelihood of success based on the how other players with similar metrics ultimately performed.

Note that this kind of data analysis does not answer "why" questions. Why did Player X over-perform or under-perform? Why does Data Point A correlate to predicting success but not Data Point B? How do new schemes and play-calling affect metrics? These are not the questions that DD's report will answer. He's using a data-based approach to predict which RBs profile as more or less likely to succeed.

Indeed, one strength of this model is the ability for correlative indicators to change with more data. That's a good thing!! If there is the occasional outlier, then the correlative indicators won't be affected in anything more than a minimal way. But, if when there are numerous outliers and/or some players that entirely break the model, then the correlative indicators for success would change. Again, that's a good thing: the correlative indicators change because there is now more data to confirm or potentially reject the previous correlative assumptions. That makes the newly-updated correlative indicators MORE reliable!

For those who look at tape, schemes, coaching fit, etc. (i.e., subjective analysis), DD's report is likely not going to be your starting place or even necessarily something on which you would rely heavily. For those who look to survey methodology and data (i.e., concrete information), this is gold.

------------

@DD: Is this close, far off? Helpful, in your estimation, or just confusing people even more? My hope is the former!! And thank you again for your contributions!!

I started off trying to see if anything correlated with success. Most things didn't, but I did indeed find a few things that did. I think what I stumbled upon that maybe nobody else really thought of is that you can't make one big blanket statement. So for example (and i'm not saying I do this exactly, it's just an example), instead of saying "agility is important", you could ask is agility as important for a bigger back as it is a smaller back. Another example, if you have good speed do you need good agility. What about the inverse, if you have good agility do you need good speed. I don't think many people have asked these types of questions before and it's why I don't believe there's been anything like what I do out there.

As for if this would have worked in the past or will it work in the future. I have my doubts that 30-40 years ago this same exact model would have worked. These days everything is much more standardized, we have a lot more data, nutrition is better, workouts and strength and condition are optimized, from a teams perspective scouting and data analysis is better and there's more of an emphasis on efficiency and maybe to some extent the passing game. Scouts have a lot more access to smaller school players now than they did before. Years ago, there was a bigger emphasis on the running game and "Lesser" rb's were probably getting more work then than they would now. I know I looked back at a few rb's that had long careers and their YPC was in the toilet, and I wonder how long they would have lasted in today's game. If we fastforward into the future, let's say teams decide to go from a 50/50 or 40/60 run pass split to a 30/70 or 20/80 run pass split. The 1k threshold probably would be largely irrelevant. Will new rules be introduced to the game that affect the run game, who knows. So idk how long this will last. Does it work right now? Yeah i'm pretty convinced it does because it's so simple and so effective and it's been the same for a long stretch of years (15 or so, whatever I have data for). Is it possible it's just a giant fluke coincidence? Possible but unlikely imo.

Again I really want to say that i'm not wildly changing things on a yearly basis. The lists have not really changed over the years. I would say as I gather more data i'm able to simplify things rather than complicate or change it more. I think there is a distinct difference. I'm sure there will be outliers here and there, and disappointments. Ideally the studs list has 3+ 1k yard seasons, so guys like Ryan Mathews and DMC were kind of disappointments already. I'm sure there will be more. I don't know about the thing you're saying about numerous outliers and then I change everything, I think that would sort of be impossible.

If we look at the "studs" and "semi studs" (which is essentially the main part of the report) the model predicted since it's inception (2015), pre-2018 every player on that list with the exception of Foreman (who had the injury of course) has at least 1 1k yard season under their belt. I'd say that is a good indication that it's working as intended. Only time will tell if it continues on that course or not.

I don’t feel a need to make a long post about it, but I do want to comment on “overfitting” and accusations of wildly changing the formula as he goes.

Isn’t that kinda the point anyways? Tweak and work it out, with the goal of making it as useful a predictive tool as possible?

Hi all. Good stuff above. If you have the 1.03 and Jacobs/Harry are off the board, who do you draft????? Omnty/Sanders/other? Don't say trade.

Maybe I could ask the question in a way that's not so upsetting.

I think that claiming that a model would correctly have identified Trent Richardson as a bust is an extraordinary claim. As long as I've been playing dynasty fantasy football, he's been one of, if not the biggest busts. Saying that your model identified that a guy who was a 1st round startup pick as a rookie was going to quickly flame out of the league is a big deal.

The fact that your model has similar prospects in different tiers raises concerns about how you formed the model to peg Richardson that way; and the fact that the model pegged two very different prospects (Richardson and DeAngelo Williams) as guys who would flame out despite their disparate prospect profiles and their completely different careers (Richardson was bad at football, Williams played for a team that prioritized using early draft capital on RBs throughout his prime) raises questions about the validity of that distinction.

I'm sorry if this offends you or if it feels to you like trolling, but this isn't that. It's a genuine attempt to understand how you're coming up with these outcomes.

PhillySpecial wrote: ↑Fri Jun 07, 2019 2:41 pm Hi all. Good stuff above. If you have the 1.03 and Jacobs/Harry are off the board, who do you draft????? Omnty/Sanders/other? Don't say trade.

That would be a tough call for sure. If I could never trade the player maybe Hock or Murray from a long term perspective/ rebuild. If I’m a contender and just need bodies at rb maybe sanders, idk. Maybe one of the browns? I really haven’t done much scouting this year on the rbs as I knew I wouldn’t be drafting any of these guys. I drafted Sony in a league where I was a contender and it worked out for me last year, flipped him for guice afterwards. so not saying not to draft them... it’s just kind of knowing what you’re getting.

honcho55 wrote: ↑Fri Jun 07, 2019 2:21 pm I don’t feel a need to make a long post about it, but I do want to comment on “overfitting” and accusations of wildly changing the formula as he goes.

Isn’t that kinda the point anyways? Tweak and work it out, with the goal of making it as useful a predictive tool as possible?

The problem with overfitting is that it stops being predictive, and instead becomes "descriptive." Here's an extreme example for the purpose of explaining what I'm talking about- you could take Alvin Kamara and add him to a new tier 1C just for Butch Davis' Backups Drafted to be Sean Payton's Starters. That technically makes the model more accurate, but it probably makes the model "descriptive" instead of "predictive." The concern is that this is what is happening here when a separate tier gets created for McCaffery, Bell, McCoy, and Rice.

The fact that the OP is hostile and refusing to acknowledge this type of criticism is telling.

PhillySpecial wrote: ↑Fri Jun 07, 2019 2:41 pm Hi all. Good stuff above. If you have the 1.03 and Jacobs/Harry are off the board, who do you draft????? Omnty/Sanders/other? Don't say trade.

RB; Sanders easily
WR; Metcalf

OhCruelestRanter wrote: ↑Fri Jun 07, 2019 3:18 pm

honcho55 wrote: ↑Fri Jun 07, 2019 2:21 pm I don’t feel a need to make a long post about it, but I do want to comment on “overfitting” and accusations of wildly changing the formula as he goes.

Isn’t that kinda the point anyways? Tweak and work it out, with the goal of making it as useful a predictive tool as possible?

The problem with overfitting is that it stops being predictive, and instead becomes "descriptive." Here's an extreme example for the purpose of explaining what I'm talking about- you could take Alvin Kamara and add him to a new tier 1C just for Butch Davis' Backups Drafted to be Sean Payton's Starters. That technically makes the model more accurate, but it probably makes the model "descriptive" instead of "predictive." The concern is that this is what is happening here when a separate tier gets created for McCaffery, Bell, McCoy, and Rice.

The fact that the OP is hostile and refusing to acknowledge this type of criticism is telling.

This is quite compelling actually. Really curious how the OP came up with the different groups considering the apparent similarities amongst the players but I understand why he doesn't want to disclose. Hard for me to give any meaning to it without that information though.

honcho55 wrote: ↑Fri Jun 07, 2019 2:21 pm I don’t feel a need to make a long post about it, but I do want to comment on “overfitting” and accusations of wildly changing the formula as he goes.

Isn’t that kinda the point anyways? Tweak and work it out, with the goal of making it as useful a predictive tool as possible?

In terms of constant significant changes, no not really because then you’d be constantly moving the goalposts and getting nowhere.

What I’ve been doing since it’s inception is simplifying it, which is probably the exact opposite of over-fitting. But the blind are leading the blind in here unfortunately.

rubber_duck wrote: ↑Thu May 30, 2019 9:15 am Yeah, the work never stops. But it has slowed down some on the RB front, mostly because there hasn't been a need to make any big changes to the RB evaluation process. The last few years have been really solid.

The past couple of years has been spent looking closer at WR and QB, with the same overall goal: bust avoidance. I find that most of the web sites and pod casters are good at pointing out how each rookie could become a star player, and are great at drumming up excitement for the rookies. But those same groups are really poor at predicting failure or risk of failure.

Heh, yeah. If you read/listen to enough people, literally every rookie in the draft is destined to become a stud

Bust avoidance is a good term, because i'm trying to avoid busts as much as I am trying to identify studs. A small but important distinction that gets lost on some ppl I think.

I'd be interested in taking a peek at your continued work if you're up for it. If not no biggie. I remember the gist of it but have to dig through my old PM's to see who exactly you liked and didn't like. I won't say any names, but there was a pretty high profile guy that you were a little down on, and uh idk he's certainly been underwhelming so far even though I still like him overall.

Chwf3rd wrote: ↑Fri Jun 07, 2019 4:33 pm

OhCruelestRanter wrote: ↑Fri Jun 07, 2019 3:18 pm

honcho55 wrote: ↑Fri Jun 07, 2019 2:21 pm I don’t feel a need to make a long post about it, but I do want to comment on “overfitting” and accusations of wildly changing the formula as he goes.

Isn’t that kinda the point anyways? Tweak and work it out, with the goal of making it as useful a predictive tool as possible?

The problem with overfitting is that it stops being predictive, and instead becomes "descriptive." Here's an extreme example for the purpose of explaining what I'm talking about- you could take Alvin Kamara and add him to a new tier 1C just for Butch Davis' Backups Drafted to be Sean Payton's Starters. That technically makes the model more accurate, but it probably makes the model "descriptive" instead of "predictive." The concern is that this is what is happening here when a separate tier gets created for McCaffery, Bell, McCoy, and Rice.

The fact that the OP is hostile and refusing to acknowledge this type of criticism is telling.

This is quite compelling actually. Really curious how the OP came up with the different groups considering the apparent similarities amongst the players but I understand why he doesn't want to disclose. Hard for me to give any meaning to it without that information though.

I’d agree, but unfortunately I don’t think the OP can defend his work so he’s just going to ignore you.

Hmm, well based on DDs response and his critics response I guess I didn’t have a handle on this. Oh well.

I still think it’s an interesting endeavor, the stat/measurement driven approach can be useful. And this being a free forum, cheers to OP for presenting it.

Not that the critics don’t have a valid point either. To each their own here.

OhCruelestRanter wrote: ↑Fri Jun 07, 2019 2:52 pm Maybe I could ask the question in a way that's not so upsetting.

I think that claiming that a model would correctly have identified Trent Richardson as a bust is an extraordinary claim. As long as I've been playing dynasty fantasy football, he's been one of, if not the biggest busts. Saying that your model identified that a guy who was a 1st round startup pick as a rookie was going to quickly flame out of the league is a big deal.

The fact that your model has similar prospects in different tiers raises concerns about how you formed the model to peg Richardson that way; and the fact that the model pegged two very different prospects (Richardson and DeAngelo Williams) as guys who would flame out despite their disparate prospect profiles and their completely different careers (Richardson was bad at football, Williams played for a team that prioritized using early draft capital on RBs throughout his prime) raises questions about the validity of that distinction.

I'm sorry if this offends you or if it feels to you like trolling, but this isn't that. It's a genuine attempt to understand how you're coming up with these outcomes.

It’s hard to tell if the model would’ve eliminated Trent Richardson, because he didn’t run any agility drills, which I think are a factor in the model. Maybe had he run them, there would’ve been a red flag we all would’ve seen.

In any case, the model’s big claim to fame was identifying David Johnson and I think it’s too early to say whether or not it’s been effective on most of the other running backs that have come out since he started doing these.

honcho55 wrote: ↑Fri Jun 07, 2019 6:01 pm Hmm, well based on DDs response and his critics response I guess I didn’t have a handle on this. Oh well.

I still think it’s an interesting endeavor, the stat/measurement driven approach can be useful. And this being a free forum, cheers to OP for presenting it.

Not that the critics don’t have a valid point either. To each their own here.

No, I think you understood just fine. Making small incremental improvements, which is what I assume you were getting at, can be a good thing. The "critic" is just someone hellbent on ruining this thread with disinformation, as is tradition. Happens every year.

ericanadian wrote: ↑Fri Jun 07, 2019 6:25 pm
It’s hard to tell if the model would’ve eliminated Trent Richardson, because he didn’t run any agility drills, which I think are a factor in the model. Maybe had he run them, there would’ve been a red flag we all would’ve seen.

In any case, the model’s big claim to fame was identifying David Johnson and I think it’s too early to say whether or not it’s been effective on most of the other running backs that have come out since he started doing these.

No it had nothing to do with missed agility drills, otherwise Fournette and several other Rb's wouldn't be on the list. I think people have been looking too long at these other models and formulas and so you're not conceptualizing things in the right manner. Think of it like this, instead of asking how pregnant a woman is, rather ask, can a woman be just a little bit pregnant?

Btw if anyone can find weights for Josh Jacobs 2nd pro day that would be a big help, I can't find it anywhere. I wonder if he even weighed in for the 2nd pro day. Right now I found a weight for him of 219 lbs at his first pro day, which would actually decrease his speed score lmao...

Dynasty DeLorean wrote: ↑Fri Jun 07, 2019 6:28 pm

honcho55 wrote: ↑Fri Jun 07, 2019 6:01 pm Hmm, well based on DDs response and his critics response I guess I didn’t have a handle on this. Oh well.

I still think it’s an interesting endeavor, the stat/measurement driven approach can be useful. And this being a free forum, cheers to OP for presenting it.

Not that the critics don’t have a valid point either. To each their own here.

No, I think you understood just fine. Making small incremental improvements, which is what I assume you were getting at, can be a good thing. The "critic" is just someone hellbent on ruining this thread with disinformation, as is tradition. Happens every year.

I can’t imagine how you think that refusing to respond to criticism and accusing your criticism of trolling or being “hellbent” on ruining the thread is a good look for you. As for tradition, I mean, have I done this before? I thought this was the first time. Huh. Honcho, your comment implies that you have a good handle on this.

ericanadian wrote: ↑Fri Jun 07, 2019 6:25 pm

OhCruelestRanter wrote: ↑Fri Jun 07, 2019 2:52 pm Maybe I could ask the question in a way that's not so upsetting.

I think that claiming that a model would correctly have identified Trent Richardson as a bust is an extraordinary claim. As long as I've been playing dynasty fantasy football, he's been one of, if not the biggest busts. Saying that your model identified that a guy who was a 1st round startup pick as a rookie was going to quickly flame out of the league is a big deal.

The fact that your model has similar prospects in different tiers raises concerns about how you formed the model to peg Richardson that way; and the fact that the model pegged two very different prospects (Richardson and DeAngelo Williams) as guys who would flame out despite their disparate prospect profiles and their completely different careers (Richardson was bad at football, Williams played for a team that prioritized using early draft capital on RBs throughout his prime) raises questions about the validity of that distinction.

I'm sorry if this offends you or if it feels to you like trolling, but this isn't that. It's a genuine attempt to understand how you're coming up with these outcomes.

It’s hard to tell if the model would’ve eliminated Trent Richardson, because he didn’t run any agility drills, which I think are a factor in the model. Maybe had he run them, there would’ve been a red flag we all would’ve seen.

In any case, the model’s big claim to fame was identifying David Johnson and I think it’s too early to say whether or not it’s been effective on most of the other running backs that have come out since he started doing these.

Oh, sure. My point was that a model’s oddly specific projection that both DeAngelo Williams and Trent Richardson would have short periods of usefulness before flaming out, when they had entirely different reasons for success and for their failure, is suspicious. The author’s response to that criticism is more suspicious.