Warrior General Games

Originally posted by PackMan97
Perhaps you could expand your study to include Fullbacks who have a blocking bar and usually very high STR and low/mediocre blocking.

I posted my FB in that forum...better idea might be Tight Ends. Receiving TEs are pretty popular, low blocking, decent strength.

I tried a few things... With speed and tackling out of the model, it spit out a result very close to 0.2*STR + 0.5*BLK + 0.1*AGI + 0.1*VIS + 0.1*CONF.

This overpredicted everything by a certain %, so I normalized it by 94%. I then divided the predicted results using this formula by 3, rounded, then multiplied by 3 again (so this is the same as using 1.5 rounding instead of 0.5), and compared to the reported actual blocking bar.

Out of 46 data points, 12 were off, which is obviously unacceptable. However, before rounding, the largest residual was under 1.5. This was made clear by noticing that each discrepancy was undershot... and VERY CLOSE, at that. For example, a predicted value of 58.2 on somebody who was actually a 60. So the formula was really only off by 0.3.

To fix this, I tried upping the % to 95%. This overshot a few, and undershot a few, but only 8 were off this time.


----


ANYWAY, I think its pretty close to the above. I'll need a lot more data, and maybe ditch the assumption that Bort used round numbers. My justification for speed out of the model was strong: It came up with a 0.003 coefficient. Tackling was LARGE, it was another 0.1, but I refuse to believe it plays a role in the blocking bar. I think all the information in tackling is already handled in the rest of the model. High tackling = more SP = more in everything else. I'll test this though... one sec

A created 12/15 OT...
add 15 to tackling..
Still 12/15. I don't feel like boosting and spending more tackling points, nor do I feel like waiting 3 weeks for him to level.

Anyway, a lot of models are coming close...

STR 0.2
BLK 0.45
AGI 0.075
VIS 0.075
CON 0.1
TKL 0.1

Spit out only 8 discrepancies, with very very close results on many of those (58.4 predicted, the person actually had a 60 bar, etc.)


If I try adjusting a few things, like switch the role of STR/BLK, even though most people have them very close, the discrepancies become large.

Something like even str/blk, and even minors agi/vis/conf/tkl spit out awful results. Taking tackling out there didn't help.

And the residuals have no correlation with position.

Okay, so I took a few level 3 chumps out of it who really had no impact on the model, except were consistent outliers.

It spit out..

Strength 0.227513239
Blocking 0.43108296
Agility 0.079120983
Vision 0.07291096
Confidence 0.091960261
Tackling 0.100500565


Residuals seemed okay, and the fun part is, those sum to 1 without having to do anything funky like round, or multiply by 0.95 in my first model.


Taking out tackling, I get...

Strength 0.218692731
Blocking 0.447578519
Agility 0.091272504
Vision 0.088237972
Confidence 0.10020443


Which again looks familiar, but also sums to 0.95 like before.

I didn't enter any of the non major/minor attributes into this, but am fairly confident they play no part in it. I'm baffled as to why tackling would move the blocking bar, but whatever.

It seems the ratio's are always pretty even...




Using 0.2, 0.45, 0.1, 0.1, 0.1 spit out 7 errors, all very small in magnitude (within 0.3 of the round).

The actual crazy-ass decimals that it put out gave only 2 residuals larger than 1.5.



----------



Oh, and the confidence intervals around each of those are still quite large..


0.156378422 0.28100704
0.378411367 0.516745671
0.059153321 0.123391686
0.037797824 0.138678119
0.058193148 0.142215712


I think its a safe assumption that blocking >>> strength >>> agi/vis/conf/(tkl?)




Oh, and what about stamina? It consistently comes up as a very small factor, like 0.03, and while significant, really doesn't do much... For example

Strength 0.235223498
Blocking 0.423703512
Agility 0.068601798
Vision 0.066878143
Confidence 0.081878485
Stamina 0.040340391
Tackling 0.079587417

Strength 0.230003208
Blocking 0.434767649
Agility 0.075812507
Vision 0.077310808
Confidence 0.086226105
Stamina 0.047798299



And here is showing that speed really IS nothing:


Strength 0.235223997
Blocking 0.423809744
Agility 0.068567287
Vision 0.066762236
Confidence 0.081790523
Stamina 0.040255123
Speed 0.000386421
Tackling 0.079304688

Do you need more data to go on? I can PM you my lvl 38 center's build. Or is your sample size sufficient?

OVERALL: Note, I have *nowhere near enough data points* for the overall bars to mean ANYTHING yet. I'm just posting what there is so far...



OT overall:

Strength 0.209858326
Blocking 0.112207638
Agility 0.190767386
Stamina 0.208264768
Speed 0.131855161
Vision -0.003853464
Confidence 0.076517167
Tackling 0.070297277


Taking out tackling, because nobody touches it, and (vision?!?!?)

Strength 0.210675702
Blocking 0.123938859
Agility 0.189722427
Stamina 0.210492384
Speed 0.128117636
Confidence 0.101862707


A kinda fun 0.2 0.1 0.2 0.2 0.1 0.1 result, though it sums to 0.9. The OT data is very small.





For the G...

Strength 0.22850544
Blocking 0.118356035
Agility 0.204548895
Stamina 0.138794188
Speed 0.116720842
Vision 0.073194259
Confidence 0.13883672
Tackling -0.099341841


Taking tackling and vision out..

Strength 0.22777297
Blocking 0.099815484
Agility 0.189621568
Stamina 0.165554349
Speed 0.127373839
Confidence 0.177646899


Once again, the data is very very slim, and the error range is VERY LARGE, in fact the lower confidence bound on blocking is all the way to INSIGNIFICANT. This could very well simply be a weighted average of the six abilities, etc.




And for center...

Strength 0.233774025
Blocking -0.049550643
Agility 0.185980046
Stamina 0.091638293
Speed -0.002263569
Vision -0.046962408
Confidence 0.251359144
Tackling 1.032437013


The tackling has zero confidence. But using, say, strength, agility, stamina, confidence, we get..


Strength 0.343933745
Confidence 0.257461901
Agility 0.180315576
Stamina 0.15723201



What I DO like about the center analysis is that I have seen first hand that stamina >>> vision in the bars.

Like I said, I need about 3x as much data to even think about doing the overall bars.


Strength 0.21919587
Blocking 0.124353617
Agility 0.179635904
Confidence 0.164974052
Stamina 0.173279949
Speed 0.112489651


Spit out nice residuals across the board for all positions, but I don't really like it. I'd rather do the overall ones separately, as I don't think he'd use the same formulas for overall like he does for blocking....

I've spammed this on three of my team forums...hopefully a few people will give info.

Feel free to try the formulas suggested...

STR 0.2
BLK 0.45
AGI 0.075
VIS 0.075
CON 0.1
TKL 0.1


And see how it works on your player, or players who didn't report to this. I think it might be off a tiny bit, but it still gives a good idea of what the bar MEANS.

Interesting that the coefficient for speed is 0, and tackling seems to do something.

Could you run a regression analysis with more constraints: namely that the weights of the major attributes are the same, and the weights of the minor attributes are the same*. I'd be interested to see if by using these constraints a fit to the data can still be achieved, or if has already been proved that this assumption is wrong.

*or, put in other terms, Bar=(weight1)*(sum of major attributes) + (weight2)*(sum of minor attributes)

Edit: I forgot to add that the third constraint is that the weights of the other attributes (non-major/minor) are all 0, if it wasn't clear enough.

Well.... the major/minor differ across the o-line, and my data is too thin to use for any single position.

I'll try the relevant str/blk and then agi/vis/conf/tkl and see what comes up.

SUMmaj 0.325568028
SUMmin 0.068291939


This actually gave better results across the board, which is... odd... as it basically just evens out str/blk whereas before there was such a large difference between the two.

http://goallineblitz.com/game/player.pl?player_id=949201 <-- 63.2 natural AGI
http://goallineblitz.com/game/player.pl?player_id=148920 <-- 50.78 natural AGI

I think the problem is, all the data points are so close together that any answer we come up with is going to have a relatively strong fit.

Such as using

SUMmaj 0.13820122
SUMmin 0.205655879


For the overall bar in the OT gives good results, but why the heck would Bort pick numbers that even closely resemble that?

Originally posted by Octowned
I think the problem is, all the data points are so close together that any answer we come up with is going to have a relatively strong fit.

Such as using

SUMmaj 0.13820122
SUMmin 0.205655879


For the overall bar in the OT gives good results, but why the heck would Bort pick numbers that even closely resemble that?

You have to take the weights on the average of the major/minor attributes rather than the sum. If not, then when you take the average afterwards, you are averaging the weights also - which is the wrong way round to do things. I assume that this is why the OT (With 5 major and 2 minor attributes) gets messed up at the moment as shown above.

If Bort has been lazy, his formula for the overall bar will be simply the average major attribute times a weight, plus the average minor attribute times a weight*, with different weights applied for the overall and blocking bars. This would also explain why it is easier to get the bars high on a C, then a G, then a OT - as the C has the least number of major attributes and the OT has the most major attributes. This would also explain why for positions like CB and TE the overall bar is much harder to raise - as they have a total sum of 10 major/minor attributes to worry about.

Forget about the blocking bar at the moment as that is more complicated - there is a further twist when dealing with this that I will mention later that needs to be considered, but I don't want to clutter here.

I'd be interested in what the regression analysis says though for the overall bar three cases: C, G and OT. The overall bar is certainly crackable.



*note these weights will not necessarily be the same between the different positions as they have a different number of major/minor attributes - I'm considering how to arrive at a common formula.

As expected, this is just the solution of the last problem multiplied by the number of attributes..

OT:

AVEmaj 0.6910061
SUMmin 0.411311757

Which is odd to sum to 1.1.


Fitted / Actual / Correct?

51.05535022 51 yes
29.73133841 30 yes
35.45845928 36 yes
52.43491142 54 no (off by .07)
16.68613669 15 no (off by .18)
16.68613669 15 no (off by .18)
51.92622748 54 no (off by .58
46.05956744 45 yes
46.89906257 45 no (off by .4)
47.64985437 48 yes
41.61050865 42 yes
42.59125793 42 yes
51.11965671 51 yes
44.38637921 45 yes
46.05956744 45 yes


If we proposed a 0.7 and 0.4 (no clue why he'd pick something that didn't add to 1, but that just makes things "move faster" I guess, changes things to

51.162 51 yes
29.86 30 yes
35.62 36 yes
52.5646 54 yes
16.72 15 no
16.72 15 no
51.9928 54 no
46.2246 45 yes
46.9382 45 no
47.6656 48 yes
41.7228 42 yes
42.6894 42 yes
51.1504 51 yes
44.5388 45 yes
46.2246 45 yes


Which really doesn't fix anything. I see no reason he'd round the initial attributes first, either.