Post by eric on Sept 23, 2015 19:56:11 GMT -6
I decided to try a new methodology to test the effect of attributes on winning. The plan was to take a small forward with 50 in everything and gave him 100 in each attribute one at a time to see what happened like before, only this time over 40 season sets. Also I first decided to take two sets of base data: controlling all variables except the uncontrollable random noise. As it turned out, the two base sets differed by 37 wins, or about 1 per season. This is a pretty large amount considering that even the best attributes I have found in the past were worth 6 per season, and they only had 20 season bins. I also had concerns that the peculiarities of the Win Shares methodology might be generating biases, so I decided to scrap that entirely and go with Pythagorean win-loss and actual win-loss, and take 40 season bins of every attribute twice.
What I found was that even this bin size could have significant differences. By far and away the biggest one was Drive Defense, which in the first set generated 1253 wins (+77) and in the second 1130 wins (-46) for an all else being equal swing of 123, or 3 wins per season. That level of uncertainty is too high, so I decided to take the values for 80 season samples for this set and future positional checks. This means that the values I have already found for center and point guard could be suspect, and although as you will see most of the general findings hold, I will update the new guys article once I have finished updating center and point guard.
Overall results, given in terms of wins per 5 points of attribute per 4 seasons.
Passing is still garbage, Inside Scoring and Shot Blocking are still awesome, Offensive is still more important than Defensive Rebounding, Handling is still the only one that will cut a player's usage. As always, the usefulness of stats that increase usage (Inside Scoring, Strength, Jumping) are deflated by the base player being such a turd sandwich, and the level of usage increase is 3 2 1.
New stuff: I was stunned by how useless Drive Defense was even for the position presumably most suited to it. I saw Jumping-level usage bumps from both Outside Scoring attributes, is this because small forwards are the most three-happy in the software in general, or is it something the old methodology missed? Quickness had a slight usage bump and Passing a nearly Strength-level one: again, is this small forwards only or something missed before?
.
One thing I didn't look at all under the old methodology was raw minutes played, and there turns out to be a pretty interesting phenomenon there too. The depth chart was always the same and there were never injuries, but the coach gave a ~4 mp per game bump for some stats and ~3 seconds per game (i.e. probably nothing) for others, and this wasn't even a soft cap thing because two of the no bump stats weren't even defense. The breakdown was like this:
More playing time
Inside Scoring, Jump Shot, Three Shot, Quickness, Passing, Perimeter Defense, Defensive Rebounding, Strength, Jumping
No change in playing time
Handling, Stealing, Shot Blocking, Post Defense, Drive Defense, Offensive Rebounding
I have two theories. My first guess is that this is based on Current Rating which has even more God d*** soft caps and is almost certainly position-specific. However, the way that there were literally no middle ground attributes and I set every one at 50 to start with suggests that some attributes are totally irrelevant for playing time or said soft caps are all set at exactly 50, which is somewhat implausible. This would probably also be position-specific, so either way there's no way to know for sure until I finish the other positions. The bottom line here is that Inside Scoring and Strength just keep getting better.
.
One interesting wrinkle I found during the Shot Blocking set in particular was that a given season's blocks weren't correlated with defensive rating, but the defensive rating overall was very strongly correlated with the Shot Blocking increase (104.4 for it, no lower than 104.8 for any other attribute and most were 105+). This suggests that the Shot Blocking attribute refers not only to blocking shots but altering them, which helps explain why the difference between a 2 block per game player and 4 block per game is so pronounced even though the observable difference only constitutes <2% of a game's plays. Also note that every attribute returned 0.5 blocks per 36 minutes compared to 2.1 for Shot Blocking, so blocked shots are still an extremely good indicator of overall defensive prowess not (just) because of the blocks themselves but because they in turn are exactly correlated with shot alteration.
.
Pythagorean win-loss had less change attribute-to-attribute than raw win-loss (0.732 rms vs. 0.945) and so the values given above are Pythagorean scaled. I am confident this trend will continue, it does in life, and I am hopeful that the scaling factor will remain the same.
.
.
Up next I'm going to repeat this analysis with a center and point guard. At 80 seasons a pop it's going to take a long time, but that's the cost of precision.
What I found was that even this bin size could have significant differences. By far and away the biggest one was Drive Defense, which in the first set generated 1253 wins (+77) and in the second 1130 wins (-46) for an all else being equal swing of 123, or 3 wins per season. That level of uncertainty is too high, so I decided to take the values for 80 season samples for this set and future positional checks. This means that the values I have already found for center and point guard could be suspect, and although as you will see most of the general findings hold, I will update the new guys article once I have finished updating center and point guard.
Overall results, given in terms of wins per 5 points of attribute per 4 seasons.
wins attribute
1.3 inside
1.2 jump shot
0.6 three shot
0.5 handling
0.4 quickness
-0.2 passing
0.6 stealing
1.1 shot block
0.5 post d
0.8 perim d
0.2 drive d
0.3 o reb
0.1 d reb
0.6 strength
0.0 jumpingPassing is still garbage, Inside Scoring and Shot Blocking are still awesome, Offensive is still more important than Defensive Rebounding, Handling is still the only one that will cut a player's usage. As always, the usefulness of stats that increase usage (Inside Scoring, Strength, Jumping) are deflated by the base player being such a turd sandwich, and the level of usage increase is 3 2 1.
New stuff: I was stunned by how useless Drive Defense was even for the position presumably most suited to it. I saw Jumping-level usage bumps from both Outside Scoring attributes, is this because small forwards are the most three-happy in the software in general, or is it something the old methodology missed? Quickness had a slight usage bump and Passing a nearly Strength-level one: again, is this small forwards only or something missed before?
.
One thing I didn't look at all under the old methodology was raw minutes played, and there turns out to be a pretty interesting phenomenon there too. The depth chart was always the same and there were never injuries, but the coach gave a ~4 mp per game bump for some stats and ~3 seconds per game (i.e. probably nothing) for others, and this wasn't even a soft cap thing because two of the no bump stats weren't even defense. The breakdown was like this:
More playing time
Inside Scoring, Jump Shot, Three Shot, Quickness, Passing, Perimeter Defense, Defensive Rebounding, Strength, Jumping
No change in playing time
Handling, Stealing, Shot Blocking, Post Defense, Drive Defense, Offensive Rebounding
I have two theories. My first guess is that this is based on Current Rating which has even more God d*** soft caps and is almost certainly position-specific. However, the way that there were literally no middle ground attributes and I set every one at 50 to start with suggests that some attributes are totally irrelevant for playing time or said soft caps are all set at exactly 50, which is somewhat implausible. This would probably also be position-specific, so either way there's no way to know for sure until I finish the other positions. The bottom line here is that Inside Scoring and Strength just keep getting better.
.
One interesting wrinkle I found during the Shot Blocking set in particular was that a given season's blocks weren't correlated with defensive rating, but the defensive rating overall was very strongly correlated with the Shot Blocking increase (104.4 for it, no lower than 104.8 for any other attribute and most were 105+). This suggests that the Shot Blocking attribute refers not only to blocking shots but altering them, which helps explain why the difference between a 2 block per game player and 4 block per game is so pronounced even though the observable difference only constitutes <2% of a game's plays. Also note that every attribute returned 0.5 blocks per 36 minutes compared to 2.1 for Shot Blocking, so blocked shots are still an extremely good indicator of overall defensive prowess not (just) because of the blocks themselves but because they in turn are exactly correlated with shot alteration.
.
Pythagorean win-loss had less change attribute-to-attribute than raw win-loss (0.732 rms vs. 0.945) and so the values given above are Pythagorean scaled. I am confident this trend will continue, it does in life, and I am hopeful that the scaling factor will remain the same.
.
.
Up next I'm going to repeat this analysis with a center and point guard. At 80 seasons a pop it's going to take a long time, but that's the cost of precision.
