Post by eric on Dec 8, 2015 11:19:29 GMT -6
I decided to investigate the "press" and "trap" options in the software. I did this by taking the default 2005 league, setting every team to fast/never/never/balanced, recording a baseline of eighty seasons, then setting every team to always/never, never/always, and always/always and recording another eighty seasons each.
I accidentally repeated three measurements along the way, which provides a useful check on how reliable they are. The change in tens of wins and Pythagorean wins were: [38 to 44] and [54 to 59], [-27 to -25] and [-41 to -39], [-23 to -24] and [-37 to -23]. Looks pretty solid to me.
Trap
Almost every team was made better by going from never trap to always trap by both regular and pythagorean wins. The only team that wasn't? Kobe Bryant's Los Angeles Oceans. Teams were on average better on defense and offense (99 DRtg+ 102 ORtg+), but it was only on the defensive side that literally every team got better. The change in wins was a flat offset: regressing always trap against never trap gave...
trap wins = 1.000 * base wins + 3.035
with R^2 of .9719 and RMSE of 1.591
Press
Literally every team was made worse by going from never press to always press. Interestingly they were on average made slightly better on offense and only slightly worse on defense (100.4 DRtg+ 100.5 ORtg+). All but two teams were made worse on defense: the outliers were Paul Pierce's Boston Revolution and... Kobe Bryant's Los Angeles Oceans. It's not surprising to have one or two outliers in a thirtyish sized sample. It's pretty surprising for the same team to be the outlier both times so I looked into the LA roster, and for some reason they have Vlade Divac at backup point guard. It's not a name typo either, 35 year old 7'1" dude at backup point guard. It's possible that height is a relevant parameter, but given that that would never fly in our league I'm not worried about it.
Overall, the change in wins was not just a flat offset. There was a flat factor, but also a linear relationship such that the better a team was, the more it had to lose by pressing.
press wins = .934 * base wins - 1.167
with R^2 of .9826 and RMSE of 1.162
Press and Trap
Teams generally broke even going always/always, and the relationship between the two defensive strategies seemed mostly linear as we will see from the formula. We would expect fouling to increase with both types, and that will have a non-linear effect because a player can only get six fouls before being sent off and because teams enter the bonus faster. The flat offset is right around where we would expect, and the factor applied to wins is lower, which makes sense considering the point on fouling.
press and trap wins = .946 * base wins + 1.924
with R^2 of .9634 and RMSE of 1.724
.
Now, it's possible that pressing and trapping have non-linear happy mediums; that while always pressing is worse than not at all, maybe sometimes pressing is better than never. To measure that I am taking a single team (the Indiana Racers) and measuring all 25 possible combinations of pressing and trapping, and also doubling up the sample size to 160 seasons per measurement. I anticipate that this will be done in a couple weeks.
I have kept track of everything in the team.mdb, which is everything in the box score except (annoyingly) minutes played, so after I see if there are happy mediums or not I am going to post more granular results too.
I accidentally repeated three measurements along the way, which provides a useful check on how reliable they are. The change in tens of wins and Pythagorean wins were: [38 to 44] and [54 to 59], [-27 to -25] and [-41 to -39], [-23 to -24] and [-37 to -23]. Looks pretty solid to me.
Trap
Almost every team was made better by going from never trap to always trap by both regular and pythagorean wins. The only team that wasn't? Kobe Bryant's Los Angeles Oceans. Teams were on average better on defense and offense (99 DRtg+ 102 ORtg+), but it was only on the defensive side that literally every team got better. The change in wins was a flat offset: regressing always trap against never trap gave...
trap wins = 1.000 * base wins + 3.035
with R^2 of .9719 and RMSE of 1.591
Press
Literally every team was made worse by going from never press to always press. Interestingly they were on average made slightly better on offense and only slightly worse on defense (100.4 DRtg+ 100.5 ORtg+). All but two teams were made worse on defense: the outliers were Paul Pierce's Boston Revolution and... Kobe Bryant's Los Angeles Oceans. It's not surprising to have one or two outliers in a thirtyish sized sample. It's pretty surprising for the same team to be the outlier both times so I looked into the LA roster, and for some reason they have Vlade Divac at backup point guard. It's not a name typo either, 35 year old 7'1" dude at backup point guard. It's possible that height is a relevant parameter, but given that that would never fly in our league I'm not worried about it.
Overall, the change in wins was not just a flat offset. There was a flat factor, but also a linear relationship such that the better a team was, the more it had to lose by pressing.
press wins = .934 * base wins - 1.167
with R^2 of .9826 and RMSE of 1.162
Press and Trap
Teams generally broke even going always/always, and the relationship between the two defensive strategies seemed mostly linear as we will see from the formula. We would expect fouling to increase with both types, and that will have a non-linear effect because a player can only get six fouls before being sent off and because teams enter the bonus faster. The flat offset is right around where we would expect, and the factor applied to wins is lower, which makes sense considering the point on fouling.
press and trap wins = .946 * base wins + 1.924
with R^2 of .9634 and RMSE of 1.724
.
Now, it's possible that pressing and trapping have non-linear happy mediums; that while always pressing is worse than not at all, maybe sometimes pressing is better than never. To measure that I am taking a single team (the Indiana Racers) and measuring all 25 possible combinations of pressing and trapping, and also doubling up the sample size to 160 seasons per measurement. I anticipate that this will be done in a couple weeks.
I have kept track of everything in the team.mdb, which is everything in the box score except (annoyingly) minutes played, so after I see if there are happy mediums or not I am going to post more granular results too.
