What can players with big differences between BPM and RAPM tell us about players before 96-97?

[ceiling raiser]

Was thinking about this a bit.

Seems like certain player prototypes are likely to achieve RAPM figures close to their BPMs, that exceed their BPMs, or that fall short of their BPMs.

In particular, I've noticed low-volume, high-minute defense-first point guards (with the notable exception of Rondo) tend to have very strong RAPMs. There are some other guys who are interesting. Low rebound, agile, high block big men tend to overperform. On the other side of the spectrum, 3-and-D wings who are more known for shooting seem to underachieve.

Has anybody done research into this, or have thoughts on the topic?

[falcolombardi]

Was thinking about this a bit.

Seems like certain player prototypes are likely to achieve RAPM figures close to their BPMs, that exceed their BPMs, or that fall short of their BPMs.

In particular, I've noticed low-volume, high-minute defense-first point guards (with the notable exception of Rondo) tend to have very strong RAPMs. There are some other guys who are interesting. Low rebound, agile, high block big men tend to overperform. On the other side of the spectrum, 3-and-D wings who are more known for shooting seem to underachieve.

Has anybody done research into this, or have thoughts on the topic?

All those 3 seem to make sense

Spot up shooting by itself is usuallt relatively low volume and low gravity effect as every team puts a shooter in the corner to be guarded

That your 3 and D wing shoots 40% instead of 36% may seem a huge gap, but if they both only shot a few a game and get guarded similarly then the added points different will be less than people think

Low rebound bigs makes sense as rebounds are a deceiving stat, as most rebounds are relatively open defensive boards. Many metrics dont seem to separate offensive and defensive rebounds either or account for on court rebounding percentage

An agile big who is out there defending over the court and altering or dissuading shots with his blocking threat will only get a few blocks to show for his effort as his teammates grab the abundant and easy defensive boards of the misses he forced

Bigs who contest or leave the paint are less likely to get a rebound and get a boxscore stat, intimidation, pick and roll coverage, switching on small players or conteating shots and making them miss without blocking doesnt get boxscore points

Low minutes defensive guards also makes a ton of sense, rapm is a rate stat that doesnt account for minutes played and they benefit from being able to play at full energy in limited minutes and often against bench heavy units

[Bad Gatorade]

As a guy that values RAPM quite a lot, I do see the value behind all-in-one box score statistics, and one of the reasons is historical application. However, in order to use BPM (and other statistics in vein), it's important to understand -

* What goes into the statistic, and how it's weighted
* What the output represents
* What the limits of BPM are

An immediate limitation I can think of is that the box score is better at measuring offence than it is at measuring defence. As a result, given that top end offence > top end defence already, this can sway BPM quite strongly in favour of offence. Due to how regression works, the spread is larger on offence than it is on defence in order to minimise outliers - otherwise, we'd get guys like Nerlens Noel looking like all stars due to having insanely inflated DBPM values relative to their actual level of defence. This doesn't happen as much on offence - two players scoring 25 points per 100 possessions at 60% shooting are not necessarily equivalent scorers, but they're both probably good scorers at least. Meanwhile, the aforementioned Nerlens Noel is a decent defender in the minutes he gets, but he is routinely getting very high DBPM scores which grossly overstate his value.

Another limitation is minutes - a player playing 15-20mpg might play differently than if a player were to play more minutes, ergo, statistics/impact can be distorted quite a bit. The impact of playing time isn't exclusive to the box score (and is very much a feature of RAPM too), and I don't think there's a fair way of relating the two.

The way BPM is regressed is also done by trying to minimise the offensive/defensive weights at the same time, and then DBPM = BPM - OBPM. This can lead to a few funky things.

Now, let's actually look at what goes into BPM (based off per-100 possessions) -

* A scaling factor based on position - the weightings of most variables are affected by what position a player plays. Certain statistics are more "valuable" if a player is a PG vs a C, and vice versa (and the intermediate positions are interpolated accordingly, i.e. a SF receives half the scaling of both a PG and a C). Naturally, this can cause all sorts of issues depending on lineups, where a smaller player might play as a "pseudo-big" with a big man acting like a floor spacer, or whatever. Ergo, players playing outside of their traditional roles might be credited/handicapped as a result of this.

* A team adjustment - as you might imagine, this is in order to help capture some of the non-box score impact.

* Points - self explanatory, and coupled with efficiency, pretty important. Doesn't vary by position, which makes sense, IMO.

* 3PM - here's where things get interesting. Now, there's definitely a benefit to shooting a lot of 3s, such as floor spacing benefits and offensive rebounding benefits (for example, shooting 12/20 from 2s gives the opponent 8 offensive rebounding chances, but 8/20 from 3 gives them 12 offensive rebounding chances). The hard part to quantify, IMO, stems from comparing long 2s to 3s. A guy like LaMarcus Aldridge still provides spacing benefits from shooting long 2s, and this could be a blind spot, especially if we go back historically. Furthermore, there's no guarantee that all floor spacers are providing similar auxiliary impact - an example is how heavily Kyle Korver is underrated by OBPM, because teams used to swarm him even though he was only averaging 12ppg. Historically, I can see how this might be ineffective in some cases, such as Antoine Walker.

* Assists - BPM has a blind spot with regard to assist type - for example, Rajon Rondo, Steph Marbury etc used to get a lot of "vanilla" assists, compared to a Steve Nash that was breaking down defences. This is partially mitigated through the team adjustment, but it's still going to have an impact on some of the OBPM numbers. A more modern SPM (such as RAPTOR's tracking element) might split assist types up, which may help the fit. Interestingly, high-assist big men are actually more "valuable" for each incremental assist due to the fact that they tend to be more "aware" players, and they receive a notable DBPM boost by having more assists. Sometimes, this will punish great defenders (Luol Deng, Aldridge) who play the "bigger" positions, are strong defenders, but don't receive a lot of assists. Similarly, there's no guarantee that Drummond suddenly boomed in defensive efficacy once he randomly started passing in 2017-18.

* Turnovers - all turnovers are punished equally on a per position basis. An SPM style statistic might benefit from live ball vs passing turnovers, but this is a historical statistic, so that's hard to quantify. One thing to note, however, is that low-turnover bigs that don't shoot a lot of 3s are often going to be good floor spacers, and so this might actually curb some of the effect described under the 3PM commentary.

* Offensive rebounds - interestingly, this gives a huge offensive benefit to PGs, and a strong offensive benefit to bigs, but also a notable defensive detriment to bigs. Some of this may be rooted in truth - a big overly preoccupied with offensive boards might not be getting back on defence to protect the rim fast enough. It might also reflect drive-heavy guards (Westbrook, for example) penetrating and wreaking havoc on the inside. It might also reflect a hyper-athletic guard (such as... Westbrook)! We shouldn't forget that the box score doesn't just represent a basketball action, but could also represent basketball traits. It might not be that Westbrook gets a lot of offensive boards for a guard, but rather, he gets offensive boards because he's insanely athletic, and the boards are a byproduct of that athleticism.

* Defensive rebounds - defensive rebounds don't mean that much to guards (sorry Westbrook!), and mean only a little bit more to big men. One of the things I found when looking at rebounding RAPM is that the more athletic big men that got a lot of rebounds (Drummond, DeAndre etc) tend to have less defensive rebounding impact than the more grounded bigs (my fellow Eastern Europeans, who dominate the rebounding DRAPM numbers!) who might be doing more boxing out. Historically, this might be worth taking note when looking at a high-rebounding athletic rim runner vs a more grounded, stout big man. Defensive boards actually impact a guard's offence negatively, which could be that high-rebound guards are often taller, defensive "specialists." Not sure that it should really be impacting ORAPM, especially when there's evidence that Westbrook actually seems to get more assists and improve team offence when he's the one getting the board. But yes, I would say that from a historical perspective, we should definitely heed caution and once again, consider basketball traits when considering these numbers.

Side note - sorry to Bondom for bringing up Westbrook and rebounding so heavily

* Steals - steals are always going to be tricky when it comes to SPM. They're always weighted very highly, and carry the usual caveats - huge impact on defence, moderate impact on offence, but we don't know how much of those steals are due to being a clever defender, or due to gambling. Many high-steal players are amongst the best at their positions defensively, but we can see a guy like Allen Iverson (neutral-ish in his early career by DBPM, but poor by DRAPM) getting overrated.

For what it's worth, I think steals might be somewhat underrated by DBPM when it comes to big men, because a big man getting a lot of steals probably isn't gambling much in the passing lanes, but playing "smart" defence. I suppose that assists probably covers a good chunk of this in the model though. And, yes, there's Andre Drummond

* Blocks - interestingly, another stat weighted really highly for guards. I suppose it's the same thing - it might not actually be that the guard is playing incredible defence, but rather, that they're athletic enough to get a lot of blocks. Over half of the huge benefit guards get from blocks is actually on offence. The same caveats apply as traditional thought would showcase - a player getting more blocks isn't necessarily a better defender, but perhaps a chronic gambler, or even an immobile player that doesn't really leave the key.

* Fouls - same weighting for both, although there's the obvious caveat of a guy not fouling because he's not playing defence vs a guy that isn't fouling because he's actually a good, fundamental defender. Additionally, a guy could be fouling more because he's defending the "best player" on offence. This is actually somewhat reflected in the statistics - fouls are given a strong negative weighting on offence, and a smaller positive weighting on defence. This might skew results for guys like Tim Duncan, who barely fouled because he was an incredible defensive player, not because he was an offensive specialist shying away from defence. Keep this in mind when looking at offensive vs defensive splits.

* FGA/FTA - these stats are scaled based on offensive primacy. In other words, a player taking more shots is likely to be "creating" more shots, and is punished less for taking them. I don't mind this idea, as to an extent, it also covers for offensive gravity (i.e a high volume player is more likely to be schemed against). It's not perfect, because it won't fully account for how well a player makes things "easier" for their teammates through their offensive primary, but it's a good idea for a box score statistic anyway. I have noticed that in recent years, however, the rise of heliocentric basketball has actually created a non-linear curve at the end. The top guys in ORAPM and OBPM are basically the same (your LeBrons, Currys, Hardens, Durants, CP3s, Nashs of the world). However, aside from Nash, there's basically nobody else who is on the "underrated" side of the curve. They're not necessarily outliers (i.e. Durant isn't more overrated by BPM than a guy like Brook Lopez relative to his RAPM) but it's interesting that there isn't really anybody in recent years that's brilliant on offence that isn't showing it in the box score, aside from Steve Nash. Not that Nash didn't, but it's not as clear as it is with the other players.

So, that's the input to BPM.

So, if we look at all of these statistics, we can see that BPM is generally pretty good on the offensive end, and actually has some very clever logic sometimes. I think that parsing out assists could actually make the BPM fit even better, although historically, this isn't entirely possible. The weakness for BPM is clearly on the defensive end, both because of the smaller spread on defence, and because there are clear blind spots to the main defensive stats. Blocks, steals and defensive rebounds can be just as much of a representation of athleticism/gambling as they are of basketball impact and fundamentals. Furthermore, seeing assists weighted so heavily on the defensive end might be seen as troublesome with respect to the non-assisting bigs who were great on defence (Patrick Ewing, Alonzo Mourning).

I think there are definitely a couple of areas in which BPM could be amended. Two big ones for me -

1. the inclusion of minutes again - a guy getting a lot of minutes who isn't getting a lot of stats is probably doing a lot of things that aren't showing up in the stat sheet.

2. non-linear terms - the increase in heliocentric play might mean that there are players getting a LOT more box score stats without actually being THAT much more high impact than before. So, perhaps a more logarithmic approach to certain statistics (rather than a linear approach) might help. The helios are often the best offensive players, but not to the degree that BPM might have you believe. Additionally, this might help the fit for Jokic defensively (who led the league in DBPM!) and might punish good passers that play in more ball-movement heavy systems a bit less.

Additionally, Jerry Stackhouse didn't become miraculously better on offence for a single season in 2000-01 because he suddenly took more shots, and then drop a notable amount again the next season. That incremental usage on comparable efficiency probably isn't making teams suddenly start defending him far better than before.

So, how does this apply historically?

Looking at players pre-1997, we might be able to apply BPM by looking at known player traits and seeing how much their statistics might be influenced by them. For example, before the 3 point era, it would be naive to assume that there weren't any low-volume 3 point shooters that weren't warping defences like crazy. We could look at the rebounding impact of big men and try and assess whether or not they achieved these numbers by boxing out, or just by being "athletic" and grabbing rebounds that have a high chance of going to a teammate anyway. We could consider that a player being a bad passer might not make them a bad defender, and vice versa.

OBPM is pretty good. Players who appear amazing through OBPM are still probably "very good" at the very least. An average player through BPM will likely not be a hidden superstar through RAPM. There are surprises through RAPM (Patrick Beverley consistently appears "very good" through ORAPM even though he's slightly below average by OBPM) but part of that could also be on RAPM. Patrick Beverley is a solid enough offensive player that probably benefits in ORAPM due to his role, and probably loses out a bit by not being able to exercise that role fully through the box score. His "goodness" is likely overrated through RAPM and his impact underrated through BPM, and he's probably one of the most extreme higher-profile cases we have. But on the whole, a players OBPM probably pegs a decent approximate range for them. If they look like a superstar there, they probably are a superstar. It doesn't mean player A > player B on account of a superior OBPM, but it probably means that they're still very good. Of course, we should still look at the concepts outlined above and see if there's any reason we need to scrutinise their OBPM, and probably try and look at impact data at the least (WOWY and what not) if there's a reasonable sample size.

DBPM? Probably worth scrutinising a lot more than OBPM, because there's the very real chance that a player can be somewhat average on defence in reality, if not poor, but then look like an All-Defensive member through the box score. I think that if a player's DBPM catches you by surprise, it's good to have a holistic look at each of the statistics that go into DBPM and see how well these considerations match your perception of that player. A good example of this is Luol Deng - we know he's a great defender, and DRAPM shows us this. He played huge minutes, so he's not a "20 minute bulldog" or anything. However, he barely fouled and wasn't a high volume passer, so his DBPM looks kind of average. I can imagine that he's not the only player that feels objectively great at defence that doesn't show it through DBPM.

So, does BPM kind of suck and feel useless in the face of our eye test and qualitative weightings?

Not entirely - and I think BPM has great applications from a holistic point of view. For example, I've recently explored the idea of coming up with a "better" playoff SRS for teams by looking at regular season BPM and multiplying it by playoff minutes, and looking at the scale of increase. I think that BPM has really good applications for "simplifying" things - it's hard for us to fully rationalise how to put all of these stats together, and it's true that a player that looks amazing might look underwhelming by BPM, and a further exploration might actually say that BPM is right. Sometimes, a player looks suspiciously high by BPM, and further analysis might show that they're actually pretty good... but if we don't use a stat like BPM as the opener, we might not even think about that player to begin with.

However, it is absolutely not a player ranking system, and it has a lot of holes. However, when we look at things from a historical point of view... we don't have good impact stats. They exist, but in a different capacity for different players, they're usually not regressed, and there are just so many basketball players out there that we don't know where to start looking. So, for these players, it's probably quite useful. PER kind of sucks, WS has lots of contextual flaws too. BPM is right up there with the best historical stats available pre-impact and pre-tracking era, so even if it requires a bit more effort to take a deep dive on these players, perhaps BPM is a good initiation point for these dives.

Now, if you'll excuse me, I'm going to the black market to purchase a new set of fingers, because I think mine are about to fall off

[f4p]

Now, if you'll excuse me, I'm going to the black market to purchase a new set of fingers, because I think mine are about to fall off

you just typed that, it wasn't a copy and paste from a previous comment? wow, thanks. other than the PER sucks comment (i might be its last defender for high volume offensive players), that was a great intro into all that goes into BPM.

[Bad Gatorade]

Now, if you'll excuse me, I'm going to the black market to purchase a new set of fingers, because I think mine are about to fall off

you just typed that, it wasn't a copy and paste from a previous comment? wow, thanks. other than the PER sucks comment (i might be its last defender for high volume offensive players), that was a great intro into all that goes into BPM.

Yeah, that was all fresh typing

I think PER has some value in a historical sense but it's not a very predictive stat at all, and it's got some clear blind spots with regard to defence and creation. PER has a very clear subset of players that it overrates (look at how consistently high Hassan Whiteside ranks in PER, even when he's not scoring super efficiently ), because it severely underrates passing/shot creation and defence and overrates things such as rebounding. There have been multiple articles released showing that PER is quite a poor predictor of impact stats, especially on defence where it's basically useless, and that even highly simple metrics (simply minutes * SRS, for example) can actually provide better prediction on the whole.

A player that scores very, very highly in BPM is still generally a great player, even if it's not attributing impact perfectly. With PER, however, you'll get the Whitesides, Drummonds, Freedoms, Brandan Wrights etc that look dominant through PER because they pile on points and rebounds, and it relies too much on volume, such as when Ray Allen's PER dropped from 21.6 to 16.4 in a single year upon moving to Boston.

PER seems almost too stable if the player's role is similar, in the sense that a player could drop off considerably in efficacy but remain fairly close in PER. DeAndre Jordan (17.2 last year, 21.8 at his peak) is closer to his peak PER than Ray Allen was after his move to Boston. DeAndre Jordan is a shell of what he once was. Ray Allen was still very good after his move. I'm pretty sure Ben Taylor looked at this in more detail - PER was highly consistent and solid at predicting wins if roles were kept constant, but in less contiguous circumstances, PER's results were wonky, because it's overly reliant on raw production instead of making an attempt at rationalising how these box scores link to impact.

I do think there are a couple of uses to PER though -

* Rationalising some of the historically huge numbers from older eras, especially wacky 60s pace.

* Looking at productivity drops between the regular and postseason (if a player had a 28 PER in the regular season and a 23 PER in the postseason, they weren't producing as much statistically, and this sometimes gets lost with team adjustment based stats).

But yeah, on the whole, the databall era has minimalised the use of PER quite heavily, IMO.

[f4p]

I think PER has some value in a historical sense but it's not a very predictive stat at all, and it's got some clear blind spots with regard to defence and creation. PER has a very clear subset of players that it overrates (look at how consistently high Hassan Whiteside ranks in PER, even when he's not scoring super efficiently ), because it severely underrates passing/shot creation and defence and overrates things such as rebounding. There have been multiple articles released showing that PER is quite a poor predictor of impact stats, especially on defence where it's basically useless, and that even highly simple metrics (simply minutes * SRS, for example) can actually provide better prediction on the whole.

no doubt. it's not perfect. i wouldn't use it to tell someone about ben wallace and high assists point guards tend to not look great. but for a substantial group of the guys we tend to talk about (Top 100, Top Peaks, etc), it seems to at least get us in the ballpark. the fact that that the top prime playoff PER's (ages 22-35) are Jordan, Lebron, Shaq, Hakeem, Duncan and that I might rank those guys literally in that order for my best playoff performers tells me it can be pretty on the mark. i definitely don't think the immediate shootdowns of PER any time it is brought up are seeing that lots of great players have put up lots of great numbers and that something that gives us a quick all-in-one, adjusted for pace, adjusted for minutes, adjusted for the league number and with historical data back to the beginning is not something to be ignored completely.

I do think there are a couple of uses to PER though -

* Rationalising some of the historically huge numbers from older eras, especially wacky 60s pace.

* Looking at productivity drops between the regular and postseason (if a player had a 28 PER in the regular season and a 23 PER in the postseason, they weren't producing as much statistically, and this sometimes gets lost with team adjustment based stats).

But yeah, on the whole, the databall era has minimalised the use of PER quite heavily, IMO.

to me, the playoff comparison is probably how i like to use it most. you can tell me PER underrates someone. you can tell me PER overrates someone. you will have a hard time convincing me it overrates someone in the regular season and underrates them in the playoffs (or vice versa). if PER says you look like a different player in the playoffs compared to the regular season, it's probably right. and there again, it doesn't seem as hindered by very small sample sizes (like "off" minutes in the playoffs) and still seems to show when a player's team was overwhelmed but they themselves went down fighting, whereas my sense is a lot of other stats tend to drop more sharply than is warranted if your opponent is too good.