DSMok1 wrote:permaximum wrote:Yes. But I used the actual minutes or actual possessions depending on the metric (RPM, RAPM, BPM uses possessions) to calculate players' unique score for each metric in that game and I also calculated roster turnover rate for that game. To be more clear, 240 minutes from new players in Team 1 and 225 minutes from new players in team 2 in a 48-minute game translates to 96.875% roster turnover rate for the game. Average roster turnover rate is 31.5% for a single NBA game and only 6 games out of 38658 have 100% roster turnover rate between 1984/85-2015/16.
You're looking at the game as a whole, correct? Using the single game approach will be problematic, because of blowouts. On a good team, most blowouts will be blowout wins, so starters will play fewer minutes in blowouts and more minutes in losses. I.E. "There's a very high correlation--whenever bench player #15 plays, we win in blowout fashion!" It's like the "running the ball leads to wins" fallacy in football--the causation is going the opposite direction.
As such, using games as a whole will yield incorrect results. Using lineup stints will yield correct results, since the players actually on the court when the lead is built will get credit.
I may not understand the approach you're using exactly, so apologies if you were already accounting for this effect.
What you said doesn't change things because;
1. Starters still play more than bench players in blowout wins and they end up as the deciding factor
2. Bench players usualy have near-average values (I also give average values to below-250 minute players) so they don't change the outcome of the game.
3. If starters are so good that it ends up as a blowout, but somehow bench players' rating with less than 20-min of playing time changes the metrical outcome of a game, that means that metric is not good anyways.
4. Still, I used actual wins AND point differential for the retrodiction if I could see any signs of what you describe. Not even the slightest. Results are 99% similar. (Percentage of wins, MAE, RMSE.)
5. Sample is simply too big for it to become the deciding factor between metrics' prediction power. It also averages out at that big of a sample.
6. Results are completely supportive of any public retrodiction tests AND my own previous retrodiction tests which have been done at the season level.
However, I'm one of those that don't relax without some real proof in practice and I tested with the matchup data I have from basketballvalue.com if things are different for the lineup level. Again 99% similarity. But I have to admit this was a rough test because I was 100% sure I would get the same results anyways so I quickly stopped doing any more tests for it.
To settle your worries about this issue, I'm publishing the results for 2002-2016 and 2015-2016 without taking roster turnover into the account. Actual wins is better for defining prediction accuracy but point differential results of MAE, RMSE is the same. Same order of metrics, same magnitude of difference between them. Big sample really helps.
Don't worry about if BPM is in-sample or out-of-sample for 2001-2014. I can confirm it's out-of-sample for 2001-2014 too because BPM's prediction accuracy doesn't change a bit in those years and it was built to predict RAPM not the outcome of games and RAPM itself have trouble at predicting the outcome of games.
Code: Select all
2001/02-2015/16
----------------- -------------
BPM 0.639156066
WS 0.628089771
AWS 0.626745268
RAPM 0.624676802
Thibodeau 0.623332299
PER 0.612783121
MPG 0.588737201
USG 0.540645361
Code: Select all
2014/15-2015/16
----------------- -------------
RPM 0.673391702
BPM 0.659307195
WS 0.65055196
RAPM 0.647887324
Thibodeau 0.643319376
AWS 0.642177389
PER 0.622763609
MPG 0.594975257
USG 0.570232204
Edit: Also, I went for games instead of seasons because it's the correct way to do retrodiction. Not the other way around. Still, the list above is not the indicator of good assignment of individual player value and people do really forget that. That was the whole point of this retrodiction test. Like I said before, I will publish the results later.