APBRmetrics

RPM estimates 11 players have a -5 impact or worse. BPM has 64 in a comparable tail, with ratings from almost twice as bad as "replacement" to 10-15 times as bad. http://www.basketball-reference.com/pla ... r_by_asc=Y 24 on the BPM list had more than 100 minutes. Maybe this is alright but it seemed worth noting. Both metrics deserve cautious use, especially at the extremes.

At the top of the distribution, the story is reversed. BPM with only 9 at or above plus 5, while RPM has 17.

I know that one is regularized and the other isn't but I still take this as contributing motivation to promote, develop and use a metric blend. Or one or both of these metrics may be mis-estimating or presenting the tails by enough to warrant consideration of adjustment, perhaps.

In my mind, the different quantities above plus 5 on RPM are not surprising given the different ways RPM and BPM divide defensive credit to individual and team / rest of team. But this rationalization does not fit with the quantity discrepancy at the bottom.

(Some versions of APM used minute cutoffs. I don't know if current RPM does. There is an assumption of a rookie prior I believe that has a similar impact in many cases, especially early in a season for those with a limited first year opportunity / role.)

Should BPM have a lower limit? Or just leave it to the user to evaluate and perhaps apply one to the interpretation? I am not sure if the worst, small minute players are really having the equivalent of a -20 to -45 BPM impact, even pro-rated to their minute level. It seems pretty extreme on the surface.

Clarification, correction or additional commentary or analysis welcome. Late night thought that might need more.

As a reminder, BPM endeavors to estimate what actually happened, what a player actually produced, whereas RPM tries to estimate a player's true talent level. So for small minutes players, BPM will have wild swings, but RPM will push those players hard toward the prior.

At the top end, I'm surprised to see the difference, to be honest. There will be an update to BPM this week that will increase the spread on the high end a bit.

Not intending to be a bore, but the discrepancy in results between BPM and RPM/xRAPM appears to owe to how BPM (under)estimates defensive contributions.

Elite defensive contributions as estimated by BPM, as has been noted, are consistently lower than those estimated by RPM/xRAPM. And regarding the data Crow referred to, for Top 17 players, not just this year but last, average DBPM is lower that DRPM: -.5 this year and -.4 last.

By BPM, about 120 performed at -3 or worse. That is close to 25% of players. But the conventional wisdom -3 replacement level is where the very best replacement players are, with most considerably worse.

Another 70 players are between -2 and -3. One point away from being "replacement level". So by BPM about 40% of league is replacement level or close enough to wonder and watch.

Only about 1/3rd of player performances are rated as positive.

RPM currently has 75 at or under -3 estimated impact. So not near as much difference as found with the -5 criteria, but it is still a fairly big difference to be aware of. Many of the -5 or worse on BPM might have true talent closer to -3 (that is RPM's theory and might be true) and some might be better. But also some, if not dozens near -3 on RPM may have a true talent worse than that. That is a question / speculation from the BPM data. Hard to say how much of each.

Any further research inquiries under way for BPM? Planned changes?

DSMok1 wrote:As a reminder, BPM endeavors to estimate what actually happened, what a player actually produced, whereas RPM tries to estimate a player's true talent level. So for small minutes players, BPM will have wild swings, but RPM will push those players hard toward the prior.

At the top end, I'm surprised to see the difference, to be honest. There will be an update to BPM this week that will increase the spread on the high end a bit.

Will the updated BPM use 14 year or 15 year RAPM?

One suggestion I would add is used games played (or percentage of games played) or Games Started in the metric. That would adjust for SOS. Plus it would adjust for injury. Injured players play worse so the GP will capture that.

colts18 wrote:

DSMok1 wrote:As a reminder, BPM endeavors to estimate what actually happened, what a player actually produced, whereas RPM tries to estimate a player's true talent level. So for small minutes players, BPM will have wild swings, but RPM will push those players hard toward the prior.

At the top end, I'm surprised to see the difference, to be honest. There will be an update to BPM this week that will increase the spread on the high end a bit.

Will the updated BPM use 14 year or 15 year RAPM?

One suggestion I would add is used games played (or percentage of games played) or Games Started in the metric. That would adjust for SOS. Plus it would adjust for injury. Injured players play worse so the GP will capture that.

That's an old post from last summer... still used MPG only at that point.

Is BPM for this season exactly same as last season? If not were any of the changes major? Is BPM going to pretty much stay as is with 1.1 or 1.01 level tweaks, stay exactly as is from here (fixed in time like PER), could go 2.0, 3.0, etc. or haven't firmly decided yet? If you tried to optimize for first year or two, how much regression fit improvement is achieved?


Would the authors of BPM and RPM have any interest in an effort, perhaps multi-yr, to find the best performing blend of the two (perhaps with the input from AJ and player tracking data and / or other sources)?

Dan/J.E. obviously know better, but I don't see much point in trying to blend BPM and RPM. BPM is an attempt to model RAPM using box score metrics. RPM is an attempt to predict RAPM, using RAPM + box score metrics. In other words, RPM is already a blend, with "access" to the same underlying data that BPM uses. Assuming J.E. put together the box-score components of RPM correctly (which is a good bet), then adding BPM to the blend shouldn't do much.

As far as adding player tracking data to BPM, that would largely defeat the point. The key to BPM is it uses data that was available for many decades, so we can get a pretty good guess at how Bird compared with Jordan or LeBron. Adding player tracking data would likely improve BPM for the players around now, but destroy its utility as a historic tool, which is really the key. There may be value in adding player-tracking data to the components of RPM however, but I imagine that's 1) tricky; 2) going to need many years of data before they're at all reliable.

If you blended and described current season or predicted future season better, then it is worth it. Don't know til you try.