Cornerback Class Statistical Projections

[AlNFL19]

19 minutes ago, Gmen said:

I have many questions. You don’t use college interceptions or PDs as an input variable, but you’re predicting interceptions and PDs?

The stat line part is based on the AV that the player is projected to reach. It's not perfect then, obviously, but AV is probably a better representation of NFL interceptions than college interceptions is. 

 

What college statistics exactly are you using?

Things like TFLs and tackles that have a higher correlation to NFL AV than INTs or PDs do. Clearly though, these aren't perfect either - a corner with a lot more tackles might just give up more catches - but they do in general a better job. However, the fact that there isn't a lot of highly-correlating on-field data that's publicly available for college cornerbacks is part of the reason why the model weights physical attributes highly (see below).

 

What variables are you regressing onto when training your model? What constitutes a bust or a success?

On 7/30/2019 at 9:20 PM, AlNFL19 said:

 It also projects a Bust Chance (% chance of failing to reach 5 AV) and Pro Bowl chance (chance of hitting or exceeding 10 AV).

Kind of arbitrary because obviously there's no number that signifies a bust by itself, but I'd consider the statistical outputs of players in the NFL who reach 4 AV or 5 or whatnot to be around that threshold, whereas the PB one is pretty close to the actual average for a Pro Bowl player. 

 

What is the correlation coefficient?

On 7/30/2019 at 9:20 PM, AlNFL19 said:

The model's projections are correlated to real AV at 0.53, which is very mediocre but a small step up from AV implied by draft position (0.48). 

 

It seems to me that your model is heavily weighted on physical attributes. Notice your top 5 guys are 6 feet tall, near 200 pounds, and run 4.5 or faster 40s at the combine. If it were that simple, the draft wouldn’t be such a crap shoot.

I agree to a degree with this. It does weight physical attributes highly, but that's because just with on-field statistics there isn't much that correlates well to NFL AV, whereas there's more to go off of (that has done a better job in some regards) with physical attributes. But obviously, this means it's not perfect - no statistical model will be, especially with something as random as football, but I think it does a decent job. It isn't that simple to project player success, obviously, but just in terms of what has worked better or worse in the past, physical attributes are important.

 

Furthermore, your features are redundant and not independent. Waldo’s formula, speed score, and combine numbers all boil down to the same features - combine numbers. 

Okay, I should have clarified this. The combine results that are included in Waldo's formula or in speed score aren't themselves included as individual variables. For example, the 10 yard split is included separately, while the 40 and the jumping events are included only as part of Speed Score and Waldo's Explosive Power formula, because those formulas correlated better to NFL AV than the events on their own.

DeAndre Baker isn’t the biggest or fastest corner, but he won the Jim Thorpe award for a reason - he’s a darn good corner on the field. The reports from camp have been nothing but positive and he’s already listed as a starter. 

I will concede this as a bottom line, however: it's a model that could be improved, and it's not perfect the way it is. Among all the ones I've done, this one isn't the one I'm the most confident in, but I think it does a decent job, especially when considering that to be seen as "good," with NFL war rooms not hitting that much themselves on corners, the bar isn't that high. We'll see as the next few years pass, but I still think it's a decent model, although I'd certainly say that I have other models, like my WR one, that I'm more confident in.