Jump to content
Leader

Random Packer News & Notes

Recommended Posts

1 hour ago, deathstar said:

He literally posted a sample size of 10+ years.

Which means exactly nothing as far as it's predictive capabilities if he simply formatted his equations to 10 years of data.

Share this post


Link to post
Share on other sites
Posted (edited)
20 minutes ago, AlexGreen#20 said:

Which means exactly nothing as far as it's predictive capabilities if he simply formatted his equations to 10 years of data.

Well that’s not true regardless of if he did that or not; which considering the hits and misses I don’t believe he did.

 

edit - Actually I believe he has stated in the past that he studied what correlated with success and formulated a system around that. Which I don’t have a problem with.

Edited by deathstar

Share this post


Link to post
Share on other sites

If you're going to mine the data set for patterns, you need to set aside a substantial amount of data to act as a cross validation sample.  A priori we should know that the NFL draft data set is not rich enough to support this kind of analysis.

The other possibility is to create a hypothesis before looking at the data, and then you can test your hypothesis (model) against the data.  If you peek at the data beforehand, or let what you know about the data bias your hypothesis, then it doesn't really work.  This kind of analysis works when you're using it to inform your own decisions under uncertainty, when you have skin in the game.  Otherwise there is a principle agent problem - in academic psychology this is called the replication crisis.

Share this post


Link to post
Share on other sites
38 minutes ago, deathstar said:

Well that’s not true regardless of if he did that or not; which considering the hits and misses I don’t believe he did.

 

edit - Actually I believe he has stated in the past that he studied what correlated with success and formulated a system around that. Which I don’t have a problem with.

If that's the case, he shouldn't have any issues showing us his formulas.

But he's not doing that. 

Share this post


Link to post
Share on other sites
Posted (edited)
4 minutes ago, AlexGreen#20 said:

If that's the case, he shouldn't have any issues showing us his formulas.

But he's not doing that. 

Why would anyone want to post interesting original content here when it’s met by posts like this?

Edited by deathstar

Share this post


Link to post
Share on other sites
15 minutes ago, deathstar said:

Why would anyone want to post interesting original content here when it’s met by posts like this?

He asked for our opinions on his work and then showed us a list of numbers. 

Why is the natural response not, "Can I see the actual work you're asking us to evaluate?"

 

Share this post


Link to post
Share on other sites

I agree showing the formulas would be helpful for us to understand and fully appreciate the work. 

But also, I think it’s a good thing to be respectful and appreciative when someone does put together something new like he did here.

Share this post


Link to post
Share on other sites

Sorry guys I’ve been out of the house for a while and can’t get to my computer for a while (posting via mobile). I’ll try to eventually so I can get them up. 

I did base it on that sample. I guess that was misleading, I just posted it to see if you guys thought that sample would possibly make it worthwhile as a tool for the future (I am aware it’ll be different with future data, but hopefully it’ll remain close enough to be useful).

Share this post


Link to post
Share on other sites
50 minutes ago, AlexGreen#20 said:

He asked for our opinions on his work and then showed us a list of numbers. 

 

Quote

Do you all think this is a pretty good track record for a statistical model?

 

Share this post


Link to post
Share on other sites
10 hours ago, AlNFL19 said:

These last few years don't have concrete results because since 2016 no WR drafted has played 4 years yet, obviously. Here's the projections:

 

Indubitably. I just wanted to see a predictive model predict. 

Share this post


Link to post
Share on other sites
5 hours ago, AlNFL19 said:

Sorry guys I’ve been out of the house for a while and can’t get to my computer for a while (posting via mobile). I’ll try to eventually so I can get them up. 

I did base it on that sample. I guess that was misleading, I just posted it to see if you guys thought that sample would possibly make it worthwhile as a tool for the future (I am aware it’ll be different with future data, but hopefully it’ll remain close enough to be useful).

I think it has a chance. 

Share this post


Link to post
Share on other sites

I made a mistake. Amari Cooper’s AV is 7, not 4. I have to redo the equations for this small difference when I get time. Should post the equations this afternoon. 

Share this post


Link to post
Share on other sites
19 hours ago, Kepler said:

Bros...... are you even sciencing right?

How to science:

 

Share this post


Link to post
Share on other sites

@AlexGreen#20

Okay, I fixed it with the correct number for Cooper (actually 7 average AV in Years 3-4, not 4).

The model is based on a simple idea that might not make any sense. What is does is projects a 

  • Projected Approximate Value
  • Bust Chance (Chance of Ending up with <5.0 AV)
  • Pro Bowl Chance (Chance of Hitting or Exceeding 10.0 AV)

The first two are built off of the second one.

This is the formula for Projected Approximate Value:

AV = ((0.32*K)+(0.38*M)+(0.42*O)+(0.44*Q)+(0.30*S)+(0.34*U)+(0.31*W)+(0.12*Y))/(0.32+0.38+0.42+0.44+0.30+0.34+0.31+0.12)

K = -0.042948*(Draft Position) + 6.31038

M = 0.116655*(Receiving Yards / Team Attempts Per Game) + 0.340037

O = 11.61997*(Touchdowns / Team Attempts Per Game) + 0.822505

Q = 2.369886*(Receptions / Team Attempts Per Game) - 0.964705

S = 0.003136*(Receiving Yards) + 0.907168

U = 0.278903*(Touchdowns) + 1.596048

W = 0.044089*(Receptions) + 1.059664

Y = 3.036771*(Weight/Height) - 4.0404245

All this is is a collection of the lines of best fit between the given statistic and Approximate Value. The final AV calculation takes all of the line of best fit data and finds a weighted average. The numbers that each line of best fit data point (K, M, O, etc.) are weighted by are their respective correlation coefficients. The final correlation coefficient from projection to actual AV is 0.47. 

This is the formula for Bust Chance:

Bust Chance = -22.363786*(Projected AV) + 150.260163

(Obviously with a lower bound of zero and an upper bound of 100).

This is the line of best fit between the projected AV and a column of data that is all either 0s or 100s. For each player that was a "bust", this number was a 100. For each non-bust, it was a zero.

Pro Bowl Chance = 5.122483*(Projected AV) - 12.712742

See above, but this time in terms of "Pro Bowl" designations.

 

The model was built on this (2006-2015) data, so I guess calling it a "track record" or whatever was misleading. I'm sorry. Anyway, that's the basic idea, and hopefully it'll remain somewhat useful. Obviously, it won't fit the future data as closely, but the hope is that it remains at a decently high number (and at least higher than draft position).

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now



×