Need some league Extractor Dumps

[gstelmack]

I could use full Extractor dumps from a couple of leagues for some digging I'm doing. These are the ones where you open the roster window, but leave it on the list, don't open a player card, plus check the "Free Agents" box to get an Extractor dump of the entire league. I already have IHOF, WOOF, and a test single player league, I could use 2-3 more. If anyone is up for doing this for some other leagues, please send them to me at gstelmack AT nc DOT rr DOT com. Thanks! And you might want to post here that you are doing it so I don't get like 6 copies of BEFL or something.

[w24olfpack]

Greg
Send you FFL and TFL

[Ben E Lou]

Greg:

I just uploaded dozens of them to the UtilitySuite directory on the server, every one I have from BFL, CCFL, FOFL, IHOF, WOOF, and a few SP leagues. I don't have time to sort them all out. Typically the naming convention is as follows:


FOFRoster{league}{year}{stage}

If "draftclass" is in the title, then it's just that: a draft class. If _forupload is in the title, you can ignore that file entirely. BFL is the nerfed league, so you might want to disregard those, depending on what you're trying to do with them. Since you now have all of the draft classes I have, I can also dump the ratings history tables that I have for BFL, CCFL, IHOF, and WOOF if you still want to look at them from the other thing you mentioned a while back. If so, let me know.

[gstelmack]

That's cool Ben, thanks, I'll grab what I need from there.

That will give me plenty, folks, no need to send any more.

These particular files I'm using to calculate out how FOF calculates cur/fut from the bars. I was pretty close with the data I had from my 3 leagues, just wanted to add some more samples to make sure. That means the BFL stuff will work fine for it, I'm just comparing red/green bars to cur/fut to find out how much each bar contributes to cur/fut. Initial investigation does show that it's linear, so for example Tackles are roughly 35% Run Block, 33% Pass Block, 19% Blocking Strength, and 12% Endurance (don't sue me yet that it doesn't add up to 100%, I haven't cleaned out the unsigned-but-on-the-roster rookies that still have their blue bars, plus I'll have more samples). Guards on the other hand are roughly 40% Run Block, 25% Pass Block, 22% Blocking Strength, and 11% Endurance.

For the draft thing, I liked the idea of generating the draft classes so I had the actual underlying ratings, to avoid issues with volatility, etc. I generated a draft class with your site, mostly for the distribution of positions, heights, and weights. I may generate a couple more for some other distributions. I'll then use these as a base and randomly re-generate the skills, since your generator nerfs QBs and WRs (unless you have one that doesn't nerf them). Then I can run routines using the FOF draft data (combines, bars) as input to predict each bar.

The end result I'm hoping for is to be able to predict each game bar, and show that in DraftAnalyzer. I'll add a button to set the bar weights to FOF defaults, in addition to letting you customize the bar weights, then you add "predicted" bars as a new field you can weight the draft class against. So you can use combines, current bars, predicted bars, or any combination thereof to do your drafting.

If it works out, the real way to draft would be to use your own custom bar profile for the things you find important against the predicted bars, throw in position weights to percolate important positions to the top, and use that to sort the draft class for drafting. If it really works out, you won't even have to dig for those weird bar profiles, the learning routine ought to be able to figure them out and weight them correctly.

If it all works out.

[gstelmack]

I should say that where the extra draft data will come in handy is as a check against the learning algorithms. I should be able to run whatever I generate against some of the draft class extracts you have, then go check and see what kind of player they really turned out to be, say, 4 years down the road. That will help validate that it's actually a reasonable predicter. The routine will either come close on most players with the misses being VSOD or injury, or if it's very far off for most then it won't be a good predicter.

[w24olfpack]

These particular files I'm using to calculate out how FOF calculates cur/fut from the bars.

Greg, be interested to compare notes.

I've done some regression on the available data from draft classes - combines, extractor and interrogator - predicting post 1st year TC ratings (focusing on Future ratings post TC).

Be very interest if that holds up overall. I didn't try to predict much further than first year ratings after TC because so many variable come into play after that(mentor, Playing time, etc).

[gstelmack]

If "draftclass" is in the title, then it's just that: a draft class. If _forupload is in the title, you can ignore that file entirely. BFL is the nerfed league, so you might want to disregard those, depending on what you're trying to do with them. Since you now have all of the draft classes I have, I can also dump the ratings history tables that I have for BFL, CCFL, IHOF, and WOOF if you still want to look at them from the other thing you mentioned a while back. If so, let me know.

I'll come back to this, I may end up wanting the ratings histories to compare the draftclass files to the ratings curves. Do your ratings histories have all bars, or do you just have the cur / fut? If it doesn't have bars (notably the peak individual bars), I'll just have to generate some myself.

[gstelmack]

Never mind, since I have the Extractor extracts, I can process those for what I need.

[Ben E Lou]

Good. My rtgs history is just ovl cur/fut

[gstelmack]

The data I managed to gather had some holes, especially missing seasons, so I generated my own. http://www.gstelmack.com/GrowthData.zip has 35 seasons worth of:

Draft Extracts
Post-TC Extracts
End-of-season Extracts

It also contains a PeakData.csv that has the highest cur/fut and bars for every player that played at least some.

And finally there are Growth*.csv files. The exact format varies by position, but it is basically key draft data (combines + bars) plus the peak bars the player attained.

I've been trying to use this to predict bars as mentioned, but it's been an utter failure so far, just too many variables apparently. The most success I had was a neural net with a large hidden layer that could learn the training data it was fed, but could not predict results for other examples very well at all.