NBA analytics: Exactly how much does a win cost?
It may be the NBA offseason, but we never stop thinking about basketball.
This month, Seth Partnow, The Athletic’s NBA analytics sage, and Mike Vorkunov, a national NBA writer, got together to discuss a vital-but-little-discussed part of the league’s analytics ecosystem: How much a win costs on a per-dollar basis in the NBA, and what that means for team building.
Mike Vorkunov: Hello Seth. I hope your NBA offseason has been pleasant. I’m sure book sales have been booming for you lately.
It’s a downtime in the NBA schedule. For us, we finally get the time to have this long chat we’ve been meaning to have since before NBA free agency began.
Before I covered the NBA, I used to be a baseball writer. I covered the New York Metropolitans, a small, provincial outfit from Queens. Baseball is all about ruthless efficiency, and front offices usually tie salaries to the number of wins a player produces. Every team has a formula, but when I was covering baseball, I believe each win was worth roughly $8 million. That sort of shorthand made it easier to understand a player’s value and contracts.
Since I’ve started covering the NBA, I’ve been wondering if something similar exists for the league. Baseball is unique in how perfectly well-suited the sport is for analytics, but even WAR has its shortcomings. Basketball has certainly bent the curve on analytics, and undoubtedly, teams have their wins-per-dollar formulas. I’m hoping in this chat we can discuss what that looks like, how you approach the dollar value of each player and what metrics you find valuable enough to tie salary to.
I know that’s a big windup, but I wanted to lay the groundwork for the discussion we’re gonna have. With all that said, let me start by asking: Does the NBA have a wins-per-dollar formula, and if so, what does that look like?
Seth Partnow: The short answer is “yes.” The longer answer is that it’s not that simple. First of all, projecting “player wins” is much more straightforward in baseball than in basketball because it’s far easier to measure individual contributions toward winning. A more subtle distinction is that individual production is far more role dependent in the NBA than in MLB. It’s not just the difficulty projecting playing time — though, at this point, more accurate minute projections are as large a differentiator in preseason team-win models as are better player-ability estimates — but also the degree to which a substantial shift in role or context can almost turn an individual into a completely different player.
For example, Portland’s 2022-23 Jerami Grant could easily look more like Denver’s 2019-20 Jerami Grant than like Detroit’s 2020-21 or 2021-22 Jerami Grant as he moves into more of a supporting role beside Damian Lillard. With me so far?
Vorkunov: I got you. There is a large gray area that consumes this whole field.
Partnow: OK, so understanding it’s a difficult endeavor, you still have to try. The good news is that one-half of the equation is easy. We know how many wins are in an NBA season — 1,230 — and we know, or can at least reasonably estimate, the total salary spent across the league. I find it’s useful to use an estimate as the gain in precision by getting way down in the contract weeds isn’t worth the squeeze, especially since the larger sources of error are going to be the difficulty in projecting playing time discussed above, as well as the elephant in the room I’m sure we’re about to address: per-minute or per-possession individual production.
I’ve found that as a quick and dirty, assuming the average team is going to spend two-thirds of the way from the cap to the tax line gets you in the ballpark. For some purposes, you might want to include projected luxury-tax payments, but things get weird quickly if you do, so I usually ignore that unless that tax-inclusive figure is important. An example might be a team deciding whether an acquisition that pushes them over the tax line is worth the net loss from additional spending combined with the elimination of tax disbursements from the other tax-paying team. Accounting is fun, isn’t it?
Anyway, using that two-thirds estimate and applying it to the 2022-23 cap and luxury-tax figures, that gets you to around $3.44 million per win, after the league spent around $3.15 million in 2021-22. That’s a big jump, but when people talk about a given contract not looking as bad in a rising cap environment, that’s what they’re referring to. I presume we won’t see anything resembling the “cap spike” that accompanied the surge in revenues when the league’s current media rights deal kicked in, but over a several-year period, the league spending in the neighborhood of $5 million or more per win is plausible.
How glazed over are your eyes now?
Vorkunov: I look like I just spent the whole night watching “Harold and Kumar 2” after recreating their entire journey from the original movie.
Partnow: There’s one other thing I should mention. Not everyone agrees with me in terms of using 1,230 as the denominator for adding wins. The argument is that since even a team made up of “replacement level” — and trust me when I say we don’t have time to unpack the concept of “replacement level” as applied to basketball — players should be expected to win some number of games (between 10 and 20), those wins shouldn’t count as part of the total. In that case, each of those 900 or so wins will be worth around $4.6 million next year. I’ve always preferred 1,230 because the calculation of individual win production makes more sense with the larger numbers for boring math reasons. But this isn’t an issue over which there is consensus.
Vorkunov: So that’s a lot of great insight. Honestly, I didn’t know most of it. Just being able to level-set one win to $3.44 million already allows for a much smarter way to scrutinize all the deals out there.
But it makes me want to keep pressing you a little bit. So you’ve done the math on what one win costs in the NBA (roughly speaking), which is very valuable. Not overpaying a player is very important, especially in a capped sport. Here’s the thing, though, and this is what has always given me issues as a basketball observer: What the hell is a one-win player?
Which analytics am I using to say, OK, this player is a one-win player or two-win player or five-win player, to then be able to say he should earn this much on his next contract or that he just posted a season where he greatly outperformed his contract? It seems like getting to the numbers is a small part of this whole enterprise because we need to be able to competently understand how to allocate those dollars.
So what do you use to analyze a player? And what can the general public use, since we don’t all have our internal models to go off, are we talking EPM here or RPM or DARKO? Like what are we using here as the basketball equivalent of WAR, or is that way too simplistic for this sport?
Partnow: Remember when I said one-half of the equation is easy? Well, this is the hard part. There are different ways to translate player production into wins, but the simplest is to come up with some measure of “points added.” This isn’t the same as looking at a player’s points per game, but rather a measure/estimate of the total of marginal points gained and opponent points prevented by a given player being on the floor. How you derive those estimates we’ll get to, but once you have a “point” value translating that into wins is kind of simple.
You’ve heard of teams “outperforming their point differential” or “underperforming their Pythagorean record?” If you’ve wondered what that means, it’s by looking at the share of total points scored by a team you can make a historically-informed prediction about their winning percentage by me. It’s an approach first created for baseball by Bill James and then applied to basketball in slightly different formulations — first by Daryl Morey, the Sixers president of basketball operations, and later by John Hollinger of The Athletic. The quick and dirty explanation of how it works from this handy calculator site is succinct:
Points ForExponent ÷ (Points ForExponent + Points AgainstExponent)
Where:
“Points For” is the number of points the team has scored.“Points Against” is the number of points scored against the team.“Exponent” is changed depending on the method used. Morey used 13.91 as an exponent; Hollinger used 16.5.
Having reached my allotment of equations in this discussion, I won’t go into the math further than saying that a team increasing their season-long points differential (or decreasing their deficit as the case may be) by between 30 and 35 points is equivalent to a one-win increase in expected winning percentage. Since we’re looking for decent first-pass estimates rather than rigorous precision, why not just say being responsible for a 32.5-point improvement in differential is worth one win? So a player who is deemed to have caused an 80-point improvement throughout a season would be worth just under 2.5 wins.
Simpler than microwave popcorn, right?
Vorkunov: I hate popcorn.
But, yeah, generally speaking, that makes sense. So a player becomes a plus-win player by every net 32.5 points he improves a team by. I understand that.
This is where I assume the harder part comes in. How are we getting to how many points a player adds (or subtracts!) to his team? I assume it’s not just gonna be me going into the game log and counting up his total plus/minus for the season. That seems a little too easy. And plus/minus doesn’t fully represent how well or poorly a player performed in a given game, let alone a season.
Partnow: Yeah, no. Neither a single game nor cumulative plus-minus is the answer because we don’t want to reward players for standing next to other very good players. We want to know who is driving success.
Vorkunov: So how do we get there?
Partnow: This is where we get to the alphabet soup of metrics you mentioned above. And at some level, it doesn’t matter which you used. All of them are expressed in a “per-100 possessions.”
However, you can’t take the value and multiply it by possessions. For example, a player who is plus-1.5/100 in your metric of choice and plays 2,000 minutes (give or take 400 possessions) shouldn’t be said to have added 60 points, or about 1.8 wins. “Over average” is setting the bar too high. Slightly below-average NBA players are still shockingly good at basketball, especially relative to an average G-Leaguer or in-season free agent. Similarly, an average (plus-0.0/100) NBA player doesn’t produce zero wins. You need to set a different, lower standard against which to measure. Here is where the difficulty of “replacement level” in basketball rears its head.
In baseball, if your starting shortstop gets injured and you have to start a backup, that backup for the most part gets all the at-bats the starter would have received. The difference between the production of the two is trivial to calculate. If you define “replacement level” and “rando Triple A guy” and measure the big-leaguers performance against that, it works!
Now, basketball doesn’t operate that way. A starting guard gets injured, and the backup might get his minutes, but he probably doesn’t get the same number of touches or shots as those get redistributed among the remaining starters. So true “replacement level” doesn’t make much sense, but you still need a baseline.
One method that has been used to estimate this baseline is to run a player value model but collapse every low-minute player (such as 500 possessions played or under in a given season) into a single contributor, the rationale being this is a reasonable proxy for “non-rotation player” production rates. This works well enough and depending on the metric, sets a baseline of somewhere between minus-2.0 and minus-4.0/100.
So instead of multiplying that plus-1.5/100 by playing time, the value added is between plus-3.5 and plus-5.5 depending on the metric, which would translate into between 4.3 and 6.7 wins. For the sake of calibration, the average starter produces 4 to 4.5 wins per season.
I don’t think we have time to litigate the merits of one player value metric versus another. Suffice it to say that teams either pick a favorite, blend several good ones or roll their similar stats to derive a player’s per-possession production value.
Still with me?
Vorkunov: I think so. Go on.
Partnow: Now, we can’t simply say “4.5 wins, so he’s going to produce around $15.5 million” and stop there.
Vorkunov: I will play the layman again and ask: Why not? That seems like an intuitive way to do this.
Partnow: Dollars aren’t the only concern and treating them as such can lead to some perverse incentives. If I force-feed a mediocre player starter 3,000 minutes, then he’s a five-win player. That $20 million per year looks like a respectable number in a rising cap environment.
Except for one thing. Those 3,000 minutes are 3,000 minutes nobody else can use. That’s a big chunk of the just under 20,000 minutes a team has to spend every year, and at 600 minutes per win produced, those minutes are being spent inefficiently — the league average is just under 500 minutes per win.
There are broadly speaking three resources teams have to balance. Money, playing time and roster spots. Focusing on the most efficient use of one at the expense of the others generally leads to poor outcomes. A high-energy backup center might be able to produce at a high rate in a small role (hello, JaVale McGee) but couldn’t maintain that level of play into starter minutes. And a team filled with such players would find itself having to deploy players outside roles in which they could succeed. Similarly, rookie contracts tend to be very dollar efficient in terms of wins, but they use lots of minutes and/or roster spots to get there.
This is the reason why “max player” stars and superstars are so valuable. This is the only class that (barring injury) uses all three resources efficiently. It also explains why the middle class of players who change teams in free agency get seemingly “overpaid.” The fact that they can be reasonably playing time- and roster spot-efficient results in the market demand for their services exceeding what would be governed strictly by the production value calculations we just discussed.
Big picture, that’s how it works. I’m sure you have questions now.
Vorkunov: So that’s interesting. That kind of makes me rethink those mid-level contracts we see given out every year, in that $13-20 million range. It seems like the certainty of getting a player who can produce consistently at a high threshold of minutes is worth the tradeoff in salary inefficiency. Do you think that’s a good way to think about those kinds of “overpay” contracts we see every summer? (Granted, sometimes those deals are just overpaying and are bad evaluations).
The hardest thing to figure out here would be how to properly value the high-efficiency, limited playing time players. You mentioned earlier the kind of trouble it would cause a team if it were to construct a roster out of a lot of those guys, but I feel like there are degrees to that. One is how to properly assess the player in that camp who can leap to a larger role. It is an evaluation issue but also a wins-per-dollar quandary too.
So here’s a real-life application question for you: How many of the contracts we see offered each year are based on this wins/dollars efficiency model?
Partnow: Honestly? Not that many.
For starters, very few teams are confident enough at the decision-making level that the underlying player value models are accurate enough to rely on to that extent. While ultimately, I think the models should be weighed more heavily than they are, I have a lot of sympathy for the skepticism. First of all, if you scratch the surface, you realize there is a big margin for error. If a model says a player is plus-1.0/100, that means the model thinks there’s a 90 percent chance (give or take based on playing time) that he’s between minus-1.0 and plus-3.0. That’s a wide range.
Secondly, even the best player value models are highly role and context dependent. Good players tend to be good in any environment, but how good matters a lot. As an example, say you’re interested in signing or trading for a combo guard who split his time between PG and SG last season. You plan to play him almost exclusively at PG. Are you getting the “good” or “bad” half of him?
Lastly and probably most importantly there’s “the market,” the collective actions of 30 GMs, 500-plus players and who knows how many agents, is enormously powerful. If the going rate for a starting wing is $25 million per year, it doesn’t matter if your model says such a player is probably only going to produce $16 million. You pay the 25 or you don’t get the player and the next best wing out there might be marginal.
Vorkunov: Hmmm. How much of the inability/unwillingness to be more financially precise in contracts has to do with the difficulty in creating an analytics model that decision-makers can be more confident in? Is it possible to create something where there is more precision to tell us how good a player is, or is basketball just too multifaceted a sport to be able to tease away the many different things at play in any one game or possession to be able to do that? Maybe I’m asking the wrong questions here.
Partnow: I think the next step is developing something that works a little better predictively in terms of how a player will perform in a new team/role context. Being able to determine “how good is this player context agnostic” is both not particularly useful and devilishly hard, but we just might get to “how good would this player be in our system?
Vorkunov: But it also makes me think about something else. So if free agency is this mostly inefficient process where agent Adam Smith sets the marketplace for every player and teams are reactive and prone to signing “bad” deals, is the trade market a better place for the application of the wins/dollar model? It would seem to me that in trades you can more frequently take on players who a team feels are under-or-overperforming their contracts on a wins-per-dollar basis, since the contracts are already in place, and so they can acquire those players more effectively to build a roster. Do you think that’s the case and should teams rely on the trade market more often, if you think that’s true?
Partnow: Smith’s most famous treatise, the WARP of Nations.
As to your question, to a degree acquiring players already under contract can help. But since a player’s trade value is inexorably linked with their contract value, I’m not sure trading for players on “good contracts” is much of a panacea. Unless a team misvalues a player, dudes on those value deals tend to get traded for decent hauls.
I think the broader lesson is that we need to be a little more nuanced than slapping a dollar amount on production and declaring someone overpaid or underpaid. Those middle-class contracts are only “overpays” because the market is somewhat distorted by the individual max on one end and the rookie scale on the other. I think the examination needs to focus more on the team’s overall spending efficiency across all three resources and how these mid-tier deals are a necessity for almost any team trying to compete for a title.
Vorkunov: The Invisible Hand Check is the name of my fantasy basketball team.
Partnow: Sigh.
Vorkunov: But back to our discussion. Building an efficiently managed basketball team is hard. Who knew!
This does make me look at contracts a little differently, even if the application of everything you talked about is still not there. I think that was interesting to learn.
Do you apply any of this when you do your Tiers? Anything else you think is worthwhile on this issue?
Partnow: I would say it informs the structure of the Tiers, but I’m explicitly excluding any notion of contract value from where various players slot into that structure. Bigger picture, I think there are a few main takeaways.
First, is an illustration and a reminder of how much “better optimized” than league-average a contending-level roster needs to be. A team that spends to the tax line at league average dollar-efficiency will win 44 games. Neat. There will be teams that can get around this by going well into the tax. And in fact, it’s difficult-bordering-on-impossible to be dollar efficient enough to file a 55-plus-win team that stays out of the tax over multiple seasons, but that’s only part of it. A team can’t create more minutes or roster spots, so the need to be more efficient than the league in those areas is immutable. A 41-win team produces a win about every 485 player minutes. To get to 55 wins, that needs to drop to the 360s.
Second, closely related to the first is the importance of identifying players who will perform “better for us” than for the league as a whole. This often-unspoken need is what leads to a lot of the mistakes you see, in that if a team thinks that in their context/system/culture a player will perform 20 percent better than he would other places, paying 10 percent over the market is a bargain! Doesn’t take much optimism bias and magic asterisking to lock a team into cap hell mediocrity from there.
Vorkunov: It worked for the Warriors, though! Yes, exceptions do prove the rule.
I feel like guesses like that are necessary, though, if you want to win and win big. If an efficiently-built roster still only gets you into the low-to-mid 40s in wins then you have to take on that risk somewhere. That leads to miscalculations sometimes which begets the cap hell mediocrity you just mentioned. But unless you can start piling up those true stars together there seems to be no better way to build those good teams than to, as Morey once said, up the risk profile and get even more aggressive. So few teams can do that well, however, which is why you see so many bad teams and decisions that turn out to be bad decisions. So much of that goes back to something else you mentioned earlier, which is the role fit and role projection, which can lead to a win/dollar efficiency that pays off.
Partnow: Wait, is that hard? I think it’s hard. Which seems like a good place to leave it for now.
NBA analytics: Exactly how much does a win cost?
It may be the NBA offseason, but we never stop thinking about basketball.
This month, Seth Partnow, The Athletic’s NBA analytics sage, and Mike Vorkunov, a national NBA writer, got together to discuss a vital-but-little-discussed part of the league’s analytics ecosystem: How much a win costs on a per-dollar basis in the NBA, and what that means for team building.
Mike Vorkunov: Hello Seth. I hope your NBA offseason has been pleasant. I’m sure book sales have been booming for you lately.
It’s a downtime in the NBA schedule. For us, we finally get the time to have this long chat we’ve been meaning to have since before NBA free agency began.
Before I covered the NBA, I used to be a baseball writer. I covered the New York Metropolitans, a small, provincial outfit from Queens. Baseball is all about ruthless efficiency, and front offices usually tie salaries to the number of wins a player produces. Every team has a formula, but when I was covering baseball, I believe each win was worth roughly $8 million. That sort of shorthand made it easier to understand a player’s value and contracts.
Since I’ve started covering the NBA, I’ve been wondering if something similar exists for the league. Baseball is unique in how perfectly well-suited the sport is for analytics, but even WAR has its shortcomings. Basketball has certainly bent the curve on analytics, and undoubtedly, teams have their wins-per-dollar formulas. I’m hoping in this chat we can discuss what that looks like, how you approach the dollar value of each player and what metrics you find valuable enough to tie salary to.
I know that’s a big windup, but I wanted to lay the groundwork for the discussion we’re gonna have. With all that said, let me start by asking: Does the NBA have a wins-per-dollar formula, and if so, what does that look like?
Seth Partnow: The short answer is “yes.” The longer answer is that it’s not that simple. First of all, projecting “player wins” is much more straightforward in baseball than in basketball because it’s far easier to measure individual contributions toward winning. A more subtle distinction is that individual production is far more role dependent in the NBA than in MLB. It’s not just the difficulty projecting playing time — though, at this point, more accurate minute projections are as large a differentiator in preseason team-win models as are better player-ability estimates — but also the degree to which a substantial shift in role or context can almost turn an individual into a completely different player.
For example, Portland’s 2022-23 Jerami Grant could easily look more like Denver’s 2019-20 Jerami Grant than like Detroit’s 2020-21 or 2021-22 Jerami Grant as he moves into more of a supporting role beside Damian Lillard. With me so far?
Vorkunov: I got you. There is a large gray area that consumes this whole field.
Partnow: OK, so understanding it’s a difficult endeavor, you still have to try. The good news is that one-half of the equation is easy. We know how many wins are in an NBA season — 1,230 — and we know, or can at least reasonably estimate, the total salary spent across the league. I find it’s useful to use an estimate as the gain in precision by getting way down in the contract weeds isn’t worth the squeeze, especially since the larger sources of error are going to be the difficulty in projecting playing time discussed above, as well as the elephant in the room I’m sure we’re about to address: per-minute or per-possession individual production.
I’ve found that as a quick and dirty, assuming the average team is going to spend two-thirds of the way from the cap to the tax line gets you in the ballpark. For some purposes, you might want to include projected luxury-tax payments, but things get weird quickly if you do, so I usually ignore that unless that tax-inclusive figure is important. An example might be a team deciding whether an acquisition that pushes them over the tax line is worth the net loss from additional spending combined with the elimination of tax disbursements from the other tax-paying team. Accounting is fun, isn’t it?
Anyway, using that two-thirds estimate and applying it to the 2022-23 cap and luxury-tax figures, that gets you to around $3.44 million per win, after the league spent around $3.15 million in 2021-22. That’s a big jump, but when people talk about a given contract not looking as bad in a rising cap environment, that’s what they’re referring to. I presume we won’t see anything resembling the “cap spike” that accompanied the surge in revenues when the league’s current media rights deal kicked in, but over a several-year period, the league spending in the neighborhood of $5 million or more per win is plausible.
How glazed over are your eyes now?
Vorkunov: I look like I just spent the whole night watching “Harold and Kumar 2” after recreating their entire journey from the original movie.
Partnow: There’s one other thing I should mention. Not everyone agrees with me in terms of using 1,230 as the denominator for adding wins. The argument is that since even a team made up of “replacement level” — and trust me when I say we don’t have time to unpack the concept of “replacement level” as applied to basketball — players should be expected to win some number of games (between 10 and 20), those wins shouldn’t count as part of the total. In that case, each of those 900 or so wins will be worth around $4.6 million next year. I’ve always preferred 1,230 because the calculation of individual win production makes more sense with the larger numbers for boring math reasons. But this isn’t an issue over which there is consensus.
Vorkunov: So that’s a lot of great insight. Honestly, I didn’t know most of it. Just being able to level-set one win to $3.44 million already allows for a much smarter way to scrutinize all the deals out there.
But it makes me want to keep pressing you a little bit. So you’ve done the math on what one win costs in the NBA (roughly speaking), which is very valuable. Not overpaying a player is very important, especially in a capped sport. Here’s the thing, though, and this is what has always given me issues as a basketball observer: What the hell is a one-win player?
Which analytics am I using to say, OK, this player is a one-win player or two-win player or five-win player, to then be able to say he should earn this much on his next contract or that he just posted a season where he greatly outperformed his contract? It seems like getting to the numbers is a small part of this whole enterprise because we need to be able to competently understand how to allocate those dollars.
So what do you use to analyze a player? And what can the general public use, since we don’t all have our internal models to go off, are we talking EPM here or RPM or DARKO? Like what are we using here as the basketball equivalent of WAR, or is that way too simplistic for this sport?
Partnow: Remember when I said one-half of the equation is easy? Well, this is the hard part. There are different ways to translate player production into wins, but the simplest is to come up with some measure of “points added.” This isn’t the same as looking at a player’s points per game, but rather a measure/estimate of the total of marginal points gained and opponent points prevented by a given player being on the floor. How you derive those estimates we’ll get to, but once you have a “point” value translating that into wins is kind of simple.
You’ve heard of teams “outperforming their point differential” or “underperforming their Pythagorean record?” If you’ve wondered what that means, it’s by looking at the share of total points scored by a team you can make a historically-informed prediction about their winning percentage by me. It’s an approach first created for baseball by Bill James and then applied to basketball in slightly different formulations — first by Daryl Morey, the Sixers president of basketball operations, and later by John Hollinger of The Athletic. The quick and dirty explanation of how it works from this handy calculator site is succinct:
Points ForExponent ÷ (Points ForExponent + Points AgainstExponent)
Where:
“Points For” is the number of points the team has scored.“Points Against” is the number of points scored against the team.“Exponent” is changed depending on the method used. Morey used 13.91 as an exponent; Hollinger used 16.5.
Having reached my allotment of equations in this discussion, I won’t go into the math further than saying that a team increasing their season-long points differential (or decreasing their deficit as the case may be) by between 30 and 35 points is equivalent to a one-win increase in expected winning percentage. Since we’re looking for decent first-pass estimates rather than rigorous precision, why not just say being responsible for a 32.5-point improvement in differential is worth one win? So a player who is deemed to have caused an 80-point improvement throughout a season would be worth just under 2.5 wins.
Simpler than microwave popcorn, right?
Vorkunov: I hate popcorn.
But, yeah, generally speaking, that makes sense. So a player becomes a plus-win player by every net 32.5 points he improves a team by. I understand that.
This is where I assume the harder part comes in. How are we getting to how many points a player adds (or subtracts!) to his team? I assume it’s not just gonna be me going into the game log and counting up his total plus/minus for the season. That seems a little too easy. And plus/minus doesn’t fully represent how well or poorly a player performed in a given game, let alone a season.
Partnow: Yeah, no. Neither a single game nor cumulative plus-minus is the answer because we don’t want to reward players for standing next to other very good players. We want to know who is driving success.
Vorkunov: So how do we get there?
Partnow: This is where we get to the alphabet soup of metrics you mentioned above. And at some level, it doesn’t matter which you used. All of them are expressed in a “per-100 possessions.”
However, you can’t take the value and multiply it by possessions. For example, a player who is plus-1.5/100 in your metric of choice and plays 2,000 minutes (give or take 400 possessions) shouldn’t be said to have added 60 points, or about 1.8 wins. “Over average” is setting the bar too high. Slightly below-average NBA players are still shockingly good at basketball, especially relative to an average G-Leaguer or in-season free agent. Similarly, an average (plus-0.0/100) NBA player doesn’t produce zero wins. You need to set a different, lower standard against which to measure. Here is where the difficulty of “replacement level” in basketball rears its head.
In baseball, if your starting shortstop gets injured and you have to start a backup, that backup for the most part gets all the at-bats the starter would have received. The difference between the production of the two is trivial to calculate. If you define “replacement level” and “rando Triple A guy” and measure the big-leaguers performance against that, it works!
Now, basketball doesn’t operate that way. A starting guard gets injured, and the backup might get his minutes, but he probably doesn’t get the same number of touches or shots as those get redistributed among the remaining starters. So true “replacement level” doesn’t make much sense, but you still need a baseline.
One method that has been used to estimate this baseline is to run a player value model but collapse every low-minute player (such as 500 possessions played or under in a given season) into a single contributor, the rationale being this is a reasonable proxy for “non-rotation player” production rates. This works well enough and depending on the metric, sets a baseline of somewhere between minus-2.0 and minus-4.0/100.
So instead of multiplying that plus-1.5/100 by playing time, the value added is between plus-3.5 and plus-5.5 depending on the metric, which would translate into between 4.3 and 6.7 wins. For the sake of calibration, the average starter produces 4 to 4.5 wins per season.
I don’t think we have time to litigate the merits of one player value metric versus another. Suffice it to say that teams either pick a favorite, blend several good ones or roll their similar stats to derive a player’s per-possession production value.
Still with me?
Vorkunov: I think so. Go on.
Partnow: Now, we can’t simply say “4.5 wins, so he’s going to produce around $15.5 million” and stop there.
Vorkunov: I will play the layman again and ask: Why not? That seems like an intuitive way to do this.
Partnow: Dollars aren’t the only concern and treating them as such can lead to some perverse incentives. If I force-feed a mediocre player starter 3,000 minutes, then he’s a five-win player. That $20 million per year looks like a respectable number in a rising cap environment.
Except for one thing. Those 3,000 minutes are 3,000 minutes nobody else can use. That’s a big chunk of the just under 20,000 minutes a team has to spend every year, and at 600 minutes per win produced, those minutes are being spent inefficiently — the league average is just under 500 minutes per win.
There are broadly speaking three resources teams have to balance. Money, playing time and roster spots. Focusing on the most efficient use of one at the expense of the others generally leads to poor outcomes. A high-energy backup center might be able to produce at a high rate in a small role (hello, JaVale McGee) but couldn’t maintain that level of play into starter minutes. And a team filled with such players would find itself having to deploy players outside roles in which they could succeed. Similarly, rookie contracts tend to be very dollar efficient in terms of wins, but they use lots of minutes and/or roster spots to get there.
This is the reason why “max player” stars and superstars are so valuable. This is the only class that (barring injury) uses all three resources efficiently. It also explains why the middle class of players who change teams in free agency get seemingly “overpaid.” The fact that they can be reasonably playing time- and roster spot-efficient results in the market demand for their services exceeding what would be governed strictly by the production value calculations we just discussed.
Big picture, that’s how it works. I’m sure you have questions now.
Vorkunov: So that’s interesting. That kind of makes me rethink those mid-level contracts we see given out every year, in that $13-20 million range. It seems like the certainty of getting a player who can produce consistently at a high threshold of minutes is worth the tradeoff in salary inefficiency. Do you think that’s a good way to think about those kinds of “overpay” contracts we see every summer? (Granted, sometimes those deals are just overpaying and are bad evaluations).
The hardest thing to figure out here would be how to properly value the high-efficiency, limited playing time players. You mentioned earlier the kind of trouble it would cause a team if it were to construct a roster out of a lot of those guys, but I feel like there are degrees to that. One is how to properly assess the player in that camp who can leap to a larger role. It is an evaluation issue but also a wins-per-dollar quandary too.
So here’s a real-life application question for you: How many of the contracts we see offered each year are based on this wins/dollars efficiency model?
Partnow: Honestly? Not that many.
For starters, very few teams are confident enough at the decision-making level that the underlying player value models are accurate enough to rely on to that extent. While ultimately, I think the models should be weighed more heavily than they are, I have a lot of sympathy for the skepticism. First of all, if you scratch the surface, you realize there is a big margin for error. If a model says a player is plus-1.0/100, that means the model thinks there’s a 90 percent chance (give or take based on playing time) that he’s between minus-1.0 and plus-3.0. That’s a wide range.
Secondly, even the best player value models are highly role and context dependent. Good players tend to be good in any environment, but how good matters a lot. As an example, say you’re interested in signing or trading for a combo guard who split his time between PG and SG last season. You plan to play him almost exclusively at PG. Are you getting the “good” or “bad” half of him?
Lastly and probably most importantly there’s “the market,” the collective actions of 30 GMs, 500-plus players and who knows how many agents, is enormously powerful. If the going rate for a starting wing is $25 million per year, it doesn’t matter if your model says such a player is probably only going to produce $16 million. You pay the 25 or you don’t get the player and the next best wing out there might be marginal.
Vorkunov: Hmmm. How much of the inability/unwillingness to be more financially precise in contracts has to do with the difficulty in creating an analytics model that decision-makers can be more confident in? Is it possible to create something where there is more precision to tell us how good a player is, or is basketball just too multifaceted a sport to be able to tease away the many different things at play in any one game or possession to be able to do that? Maybe I’m asking the wrong questions here.
Partnow: I think the next step is developing something that works a little better predictively in terms of how a player will perform in a new team/role context. Being able to determine “how good is this player context agnostic” is both not particularly useful and devilishly hard, but we just might get to “how good would this player be in our system?
Vorkunov: But it also makes me think about something else. So if free agency is this mostly inefficient process where agent Adam Smith sets the marketplace for every player and teams are reactive and prone to signing “bad” deals, is the trade market a better place for the application of the wins/dollar model? It would seem to me that in trades you can more frequently take on players who a team feels are under-or-overperforming their contracts on a wins-per-dollar basis, since the contracts are already in place, and so they can acquire those players more effectively to build a roster. Do you think that’s the case and should teams rely on the trade market more often, if you think that’s true?
Partnow: Smith’s most famous treatise, the WARP of Nations.
As to your question, to a degree acquiring players already under contract can help. But since a player’s trade value is inexorably linked with their contract value, I’m not sure trading for players on “good contracts” is much of a panacea. Unless a team misvalues a player, dudes on those value deals tend to get traded for decent hauls.
I think the broader lesson is that we need to be a little more nuanced than slapping a dollar amount on production and declaring someone overpaid or underpaid. Those middle-class contracts are only “overpays” because the market is somewhat distorted by the individual max on one end and the rookie scale on the other. I think the examination needs to focus more on the team’s overall spending efficiency across all three resources and how these mid-tier deals are a necessity for almost any team trying to compete for a title.
Vorkunov: The Invisible Hand Check is the name of my fantasy basketball team.
Partnow: Sigh.
Vorkunov: But back to our discussion. Building an efficiently managed basketball team is hard. Who knew!
This does make me look at contracts a little differently, even if the application of everything you talked about is still not there. I think that was interesting to learn.
Do you apply any of this when you do your Tiers? Anything else you think is worthwhile on this issue?
Partnow: I would say it informs the structure of the Tiers, but I’m explicitly excluding any notion of contract value from where various players slot into that structure. Bigger picture, I think there are a few main takeaways.
First, is an illustration and a reminder of how much “better optimized” than league-average a contending-level roster needs to be. A team that spends to the tax line at league average dollar-efficiency will win 44 games. Neat. There will be teams that can get around this by going well into the tax. And in fact, it’s difficult-bordering-on-impossible to be dollar efficient enough to file a 55-plus-win team that stays out of the tax over multiple seasons, but that’s only part of it. A team can’t create more minutes or roster spots, so the need to be more efficient than the league in those areas is immutable. A 41-win team produces a win about every 485 player minutes. To get to 55 wins, that needs to drop to the 360s.
Second, closely related to the first is the importance of identifying players who will perform “better for us” than for the league as a whole. This often-unspoken need is what leads to a lot of the mistakes you see, in that if a team thinks that in their context/system/culture a player will perform 20 percent better than he would other places, paying 10 percent over the market is a bargain! Doesn’t take much optimism bias and magic asterisking to lock a team into cap hell mediocrity from there.
Vorkunov: It worked for the Warriors, though! Yes, exceptions do prove the rule.
I feel like guesses like that are necessary, though, if you want to win and win big. If an efficiently-built roster still only gets you into the low-to-mid 40s in wins then you have to take on that risk somewhere. That leads to miscalculations sometimes which begets the cap hell mediocrity you just mentioned. But unless you can start piling up those true stars together there seems to be no better way to build those good teams than to, as Morey once said, up the risk profile and get even more aggressive. So few teams can do that well, however, which is why you see so many bad teams and decisions that turn out to be bad decisions. So much of that goes back to something else you mentioned earlier, which is the role fit and role projection, which can lead to a win/dollar efficiency that pays off.
Partnow: Wait, is that hard? I think it’s hard. Which seems like a good place to leave it for now.
