Draft gurus opinions wanted

[Hammer]

Just traded in to pick this guy at 1.12 in a MP league. His grade was 5.4 I think, around 4th or 5th in a very average WR class. I figured his percentage developed pulled him down. This league has very few decent WRs in it, so I jumped on him when he dropped. A.D, G.D and R.R should creep according to combines. I hope his endurance comes along with it. Hoping for 65+ overall.





Any thoughts?

[chinaski]

Im betting 70 minimum by the time its all said and done.

[jeffrey]

Here you go:



I've found no correlations between 3rd down, endurance, or special teams and any combines, so those remain unchanged by this estimate. Otherwise, that's what he will (should) look like.

[Hammer]

Thanks Jeffrey, I'd take those bars Endurance is a big variable here. Everyone seems to agree it doesn't correlate with any combines. However, I have seen players like Wetzel, likely creepers, drag their endurance bar with the expected creeping bars. Fingers crossed.

This might bugger up my TE thread, as its the same team. Should be a good 1-2 punch though, I hope.

[bpm07]

Originally Posted by jeffrey

Here you go:



I've found no correlations between 3rd down, endurance, or special teams and any combines, so those remain unchanged by this estimate. Otherwise, that's what he will (should) look like.

What utility/software/sorcery did you use to create that graphic? Thanks!

[jeffrey]

matlab

[jeffrey]

Okay, pithy comments aside, I did some data mining and found the correlation coefficients for bars and combines (using step-wise regression). These in hand, I used that forward model and minimized the least squares difference between the combines from the estimated and actual (prospect) bars. I added a Lagrange multiplier for similarity to the original bars, as well.

Short answer: sorcery

[gstelmack]

Send me the math and it will be in the next iteration of Analyzer. My approach was not nearly as scientific...

[bpm07]

Originally Posted by jeffrey

Okay, pithy comments aside, I did some data mining and found the correlation coefficients for bars and combines (using step-wise regression). These in hand, I used that forward model and minimized the least squares difference between the combines from the estimated and actual (prospect) bars. I added a Lagrange multiplier for similarity to the original bars, as well.

Short answer: sorcery

Haha, I love Matlab, though I prefer R because it's free

I am about to start some of my own data mining, and if I find something useful I'll post it here.

[jeffrey]

Originally Posted by bpm07

Haha, I love Matlab, though I prefer R because it's free

The university springs for mine, but R is good stuff, too.

While you're doing your data mining, if you come up with a way of pulling out individual and team stats, I would be interested.

[bpm07]

Originally Posted by jeffrey

The university springs for mine, but R is good stuff, too.

While you're doing your data mining, if you come up with a way of pulling out individual and team stats, I would be interested.

Do you have some sort of out-of-sample error measures on known test data, like squared error or something similar?

I am comparing model predictions on the 375-625 scale so squared error is fairly large, though one could report normalized squared error.

Also, do you know how to convert a number on the 375-625 scale to the ratings we see in FOF (0 - 100)? Is it as simple as

100*(rating-375)/(625-375)?

[jeffrey]

Originally Posted by bpm07

Do you have some sort of out-of-sample error measures on known test data, like squared error or something similar?

I am comparing model predictions on the 375-625 scale so squared error is fairly large, though one could report normalized squared error.

Also, do you know how to convert a number on the 375-625 scale to the ratings we see in FOF (0 - 100)? Is it as simple as

100*(rating-375)/(625-375)?

I don't know anything about it. I've never put together a player file and imported it into FOF. (I assume that's what you're talking about)

I didn't do any out-of-sample testing, because generating the data is a hassle, so I didn't want to generate enough to do a split. Fortuantely, I had a large enough data set that overfitting wasn't really a concern.

That being said, it would probably be worthwhile to estimate an error measure for the mapping from bars to combines, but I haven't had the time to do that. That error could be propagated through to the estimates, and show up as ranges in graphs like the one above; kind of like the ranges shown by the game for rookie scouting.

[jeffrey]

Just to be clear, the error I'm talking about estimating would be due to unexplained variance rather than data idiosyncrasies. That's why I wouldn't estimate out-of-sample error. Statistics isn't my strong-point, though, so I may be missing some subtlety.