I decided to use 2014 stats and see which stats predicted the 2015 season. I tested a total of 9 stats. For players that didn't play the season before, I gave them a replacement level value. Here are the results:
R values:
Stat W-L% SRS
BPM 0.803 0.826
RPM 0.819 0.813
PI RAPM 0.799 0.794
NPI RAPM 0.741 0.759
WP/48 0.711 0.751
WS/48 0.720 0.745
14y RAPM 0.742 0.726
PER 0.702 0.717
USG% 0.585 0.549
BPM came out the best when compared to SRS while RPM was the best when compared W-L record. PER was the worst stat in terms of predictive value. Surprisingly, 14 year RAPM did poorly.
Which stat predicted the 2015 the best? (Results here)
Re: Which stat predicted the 2015 the best? (Results here)
I tried out weighted Minutes played and it had a .759 R value to SRS, same as NPI RAPM. Who knew that simple minutes played from the previous season was so predictive?
A 50/50 BPM/RPM blend got me a .829 R value with SRS.
A 50/50 BPM/RPM blend got me a .829 R value with SRS.
Last edited by colts18 on Tue Oct 27, 2015 3:50 am, edited 1 time in total.
Re: Which stat predicted the 2015 the best? (Results here)
How does weighted minutes work in this, adding up NBA minutes played the prior year for each team?colts18 wrote:I tried out weighted Minutes played and it had a .759 R value to SRS, same as NPI RAPM. Who knew that simple minutes played from the previous season was so predictive?
Re: Which stat predicted the 2015 the best? (Results here)
Its 2015 Minutes played * 2014 Minutes Played. So a player plays 2000 MP in 2015. In 2014 he played 1000 MP. Its 2000*1000. Then do that for every player on the team and divide it total team minutes played. In this scenario a rookie has a 0 value.AJbaskets wrote:How does weighted minutes work in this, adding up NBA minutes played the prior year for each team?colts18 wrote:I tried out weighted Minutes played and it had a .759 R value to SRS, same as NPI RAPM. Who knew that simple minutes played from the previous season was so predictive?
Re: Which stat predicted the 2015 the best? (Results here)
i think permaximum has consistently mentioned the predictive power of minutescolts18 wrote:I tried out weighted Minutes played and it had a .759 R value to SRS, same as NPI RAPM. Who knew that simple minutes played from the previous season was so predictive?
A 50/50 BPM/RPM blend got me a .829 R value with SRS.
i'm now curious if minutes or minutes per game has more power
Re: Which stat predicted the 2015 the best? (Results here)
MPG has a .761 R value and that increases to .767 if you go by the Dsmok method of adding 4 games of 0 minutes.sndesai1 wrote:i think permaximum has consistently mentioned the predictive power of minutescolts18 wrote:I tried out weighted Minutes played and it had a .759 R value to SRS, same as NPI RAPM. Who knew that simple minutes played from the previous season was so predictive?
A 50/50 BPM/RPM blend got me a .829 R value with SRS.
i'm now curious if minutes or minutes per game has more power
Re: Which stat predicted the 2015 the best? (Results here)
What measure of player success did you test against? Or was it the same stat, like last seasons PER is not predictive of this season's PER.
Re: Which stat predicted the 2015 the best? (Results here)
I wouldn't mind if the whole testing procedure was described in a bit more detail.browning wrote:What measure of player success did you test against? Or was it the same stat, like last seasons PER is not predictive of this season's PER.
Re: Which stat predicted the 2015 the best? (Results here)
I'll give an example with the Warriors. MP here is 2015 Minutes played. The rest of the stats are 2014 stats.Nate wrote:I wouldn't mind if the whole testing procedure was described in a bit more detail.browning wrote:What measure of player success did you test against? Or was it the same stat, like last seasons PER is not predictive of this season's PER.
Code: Select all
MP 2014 PER MP * PER
Leandro Barbosa 982 11.5 11293
Harrison Barnes 2318 9.8 22716.4
Andrew Bogut 1583 17 26911
Stephen Curry 2613 24.1 62973.3
Festus Ezeli 504 10 5040
Draymond Green 2490 12.7 31623
Justin Holiday 657 10 6570
Andre Iguodala 2069 13.7 28345.3
Ognjen Kuzmic 72 3.4 244.8
David Lee 904 19.1 17266.4
Shaun Livingston 1468 14.5 21286
James Michael McAdoo 137 10 1370
Brandon Rush 271 4.1 1111.1
Marreese Speights 1207 15.2 18346.4
Klay Thompson 2455 14.3 35106.5
Sum 19730 14.70872783 290203.2
Code: Select all
Player Tm MP PER USG% WS/48 BPM RPM NPI PI 14y WP48
Leandro Barbosa GSW 982 11.5 20 0.058 -3.9 -1.52 0.7 -0.61 -0.9 0
Harrison Barnes GSW 2318 9.8 16.8 0.07 -1.4 -2.88 -2.58 -2.11 -2.8 0.046
Andrew Bogut GSW 1583 17 12.4 0.176 4.4 2.06 0.2 0.87 4 0.293
Stephen Curry GSW 2613 24.1 28.3 0.225 7.4 6.24 3.46 5.18 3.1 0.263
Festus Ezeli GSW 504 10 17 0.04 -2 -2 -2 -1.19 -1.3 0.04
Draymond Green GSW 2490 12.7 14.8 0.119 2.8 3.5 2.89 3.36 2.3 0.191
Justin Holiday GSW 657 10 17 0.04 -2 -2 -2 -0.17 0.3 0.04
Andre Iguodala GSW 2069 13.7 13.3 0.135 3 6.63 5.31 6.31 4.6 0.241
Ognjen Kuzmic GSW 72 3.4 13.3 -0.025 -5.2 -3.49 -1.74 -2.39 -3.1 -0.102
David Lee GSW 904 19.1 24.3 0.16 0.8 1.22 2.74 1.07 -0.7 0.137
Shaun Livingston GSW 1468 14.5 16.1 0.108 0.6 -0.6 0.26 -0.85 -0.7 0.151
James Michael McAdoo GSW 137 10 17 0.04 -2 -2 -2 -2 -2 0.04
Brandon Rush GSW 271 4.1 12.3 -0.05 -6.3 -2.78 -1.97 -3.06 -3.1 -0.033
Marreese Speights GSW 1207 15.2 25.4 0.103 -5.2 -5.42 -2.79 -4.35 -5.6 0.013
Klay Thompson GSW 2455 14.3 22.6 0.112 0.7 2.03 2.74 2.07 3.6 0.061
Here is every team:
Code: Select all
Row Labels Sum of MP Sum of xPER
ATL 19732 14.6
BOS 19880 13.4
BRK 19928 14.5
CHI 19881 13.6
CHO 19905 14.1
CLE 19779 18.1
DAL 19881 15.9
DEN 19881 14.3
DET 19827 14.2
GSW 19730 14.7
HOU 19804 14.6
IND 19855 12.8
LAC 19730 17.6
LAL 19932 13.0
MEM 19904 15.6
MIA 19732 14.2
MIL 19929 12.7
MIN 19806 12.6
NOP 19782 16.0
NYK 19855 12.6
OKC 19832 15.0
ORL 19755 12.7
PHI 19802 11.3
PHO 19879 16.2
POR 19856 15.5
SAC 19855 13.7
SAS 19955 16.2
TOR 19855 15.6
UTA 19705 13.6
WAS 19956 15.2
Re: Which stat predicted the 2015 the best? (Results here)
Thanks for the clarification.
Re: Which stat predicted the 2015 the best? (Results here)
Sorry, how did you account for rookies?colts18 wrote: I'll give an example with the Warriors. MP here is 2015 Minutes played. The rest of the stats are 2014 stats.
I had an idea I think I want to try for every rating I can find (ie, mine & the ones I find on b-r). Each player's rating would be calculated for every test season, then compiled at team level for correlations to actual team results:
(currentseasonRATING*minutesofcurrentseason+2*priorseasonRATING*minutesofpriorseason+2*nextseasonRATING*minutesofnextseason)/(currentminutes+2*priorminutes+2*nextminutes)
This takes away creating dummy ratings for rookies, and projecting "off" ratings on super small minute samples (ie, missed most season due to injury) to a much larger minute sample the next season. Kinda combines each rating's reflection of the current season (about 20% value, general sense) with its predictive correlations (80%). Doing prior AND next allows for generally eliminating age progression complications.
Modern box score stat era - could be done now for 33 seasons. Yearly correlations could give a better understanding of how well certain ratings adjust to a changing game (changes in pace, three frequency, Oreb frequency, etc).
Re: Which stat predicted the 2015 the best? (Results here)
Sent you a pm - letting you know here because I always miss my messages, sometimes for weeks. I can't be the only one this happens to - right?
Re: Which stat predicted the 2015 the best? (Results here)
For players that didn't play the season before (includes Rookies), I gave them a replacement level value. That is -2 for RAPM, RPM, BPM. 0.04 for WS/48 and WP/48. And 10 PER, 17% usage.Statman wrote:
Sorry, how did you account for rookies?
I had an idea I think I want to try for every rating I can find (ie, mine & the ones I find on b-r). Each player's rating would be calculated for every test season, then compiled at team level for correlations to actual team results:
(currentseasonRATING*minutesofcurrentseason+2*priorseasonRATING*minutesofpriorseason+2*nextseasonRATING*minutesofnextseason)/(currentminutes+2*priorminutes+2*nextminutes)
This takes away creating dummy ratings for rookies, and projecting "off" ratings on super small minute samples (ie, missed most season due to injury) to a much larger minute sample the next season. Kinda combines each rating's reflection of the current season (about 20% value, general sense) with its predictive correlations (80%). Doing prior AND next allows for generally eliminating age progression complications.
Modern box score stat era - could be done now for 33 seasons. Yearly correlations could give a better understanding of how well certain ratings adjust to a changing game (changes in pace, three frequency, Oreb frequency, etc).
Re: Which stat predicted the 2015 the best? (Results here)
Here are the results for a few more statsStatman wrote:Sent you a pm - letting you know here because I always miss my messages, sometimes for weeks. I can't be the only one this happens to - right?
R value to SRS
WAR/48 0.761
HN/48 0.777
HNI 0.787
That would put those stats between NPI RAPM and PI RAPM.
Re: Which stat predicted the 2015 the best? (Results here)
You should try Wins Produced.