CelticsStrong
Celtics Basketball => Celtics Talk => Topic started by: Roy H. on June 19, 2018, 07:55:54 AM
-
ESPN decided to project the value of certain trade packages, projected over the next five years. Here’s what they came up with:
(https://image.ibb.co/mtNapd/B946676_B_04_B4_45_C1_987_C_50_DD03_DD2_DA1.jpg)
https://fivethirtyeight.com/features/the-next-phase-of-nba-superteam-technology-creating-one-from-scratch/?addata=espn:frontpage
Does it strike anybody as odd that Lonzo Ball projects to be better than Kyrie and Hayward combined? That Kyrie and Covington have equal value? That Pascal Siakim projects to add more value than Hayward?
If your model gives results like that, it’s time to scrap your model.
-
Whoever the data scientist was behind that, I sure would never want them near my company- Those projections are utter garbage! My guess is they are a Lakers fan no doubt.
-
Lakers fans dreaming
-
That list is ridiculous. Ball, Kuzma and Deng? So without Deng which brings the value down Ball & Kuzma should be able to get Anthony Davis with their value.
-
I'm still waiting on Marcus Smart's breakout into the top 10 that CARMELO predicted last year.
-
That is beyond delirious
-
CARMELO is a fun toy. But he’s really chosen to put his finger on the scale in a major way.
Firstly, I’m not going to get up in arms about Hayward’s number — the model is clearly concerned about his injury. It felt Hayward was a borderline All-Star for its 5-year forecast entering this year, but missing a year with injury in his late 20s clearly gives Hayward a lot of unfavorable comparables. We can be more optimistic because we know that Hayward had a bone injury, which, while terrible, is less likely to recur and more likely to fully heal than ligament and tendon injuries, which are going to be the predominant source of lost seasons in the database.
That said, the way in which he’s put his finger on the scale is to use a different kind of projection, which he’s called his “upside projection”. This favors young players — and the younger and less experienced the better. CARMELO comes up with its projection by looking at past players and assigning a level for how much they correlate with players, and then weighting the future production of similar players by how more or less similar they are to the player being forecast. This is perfectly reasonable — it’s probabilistic, and will have it’s misses, but in the aggregate it’s sound data science. However, for his upside projection, used specially for this article, he excludes from the pool of similar players all those that had a zero or negative value. For an in-his-prime star like Kyrie, this means virtually no one is excluded. So we keep a lot of the players who ultimately had a worse performance than we expect Kyrie to have, and they bring down his average. Ball, of course, has a lot of players as comparables who had a negative value and were out of the league. Those players are excluded, meaning we’re seeing a weighted projection of only positive outcomes, and not all outcomes. As a #2 overall draft pick with good measurables, there will be a lot of stars and superstars in this group, including some of their top seasons. This is why Bam Adebayo is also at a similar tier to Kyrie.
I don’t know if Neil Paine, who’s normally quite good, didn’t think about the bias he was introducing into his calculations, or if the editors at ESPN said “write an article showing that the Lakers can make the best offer for Kawhi.” I’m disappointed, and he should have been a little more open with the fact that his methodology would clearly favor those with the least experience.
-
Having a computer model predict how good NBA players will be in the future is an insanely difficult task -- especially for young players where their futures end up being so volatile. Think about how is a computer supposed to know that Oladipo will be a better player than Anthony Bennett? Or how much the future value of Oladipo as we fans know it has changed in just the last 12 months.
As far as I know, we don't have any better models out there than CARMELO so I would say we probably should not scrap it altogether. It is far from perfect, but gives us some insight to what kinds of players have had productive and unproductive careers in the past. It is entirely probabilistic and is expected to be inaccurate quite often. But it can be improved over time with more learning data fed into it.
If I'm an NBA GM, I'm not going to be basing my potential trades off of it.
-
The article itself though certainly is garbage. It appears to have been written by someone who knows nothing of basketball. It listed CP3, Harden, and Trevor Ariza as a "Big 3". It put Brandon Ingram as starting point guard on next season's Lakers team.
-
Having a computer model predict how good NBA players will be in the future is an insanely difficult task -- especially for young players where their futures end up being so volatile. Think about how is a computer supposed to know that Oladipo will be a better player than Anthony Bennett? Or how much the future value of Oladipo as we fans know it has changed in just the last 12 months.
As far as I know, we don't have any better models out there than CARMELO so I would say we probably should not scrap it altogether. It is far from perfect, but gives us some insight to what kinds of players have had productive and unproductive careers in the past. It is entirely probabilistic and is expected to be inaccurate quite often. But it can be improved over time with more learning data fed into it.
If I'm an NBA GM, I'm not going to be basing my potential trades off of it.
I think it’s useless as a predictive model, which is how it’s used. Any model that will be “inaccurate quite often” doesn’t serve much good.
Obviously, no model is perfect. They can’t account for flukes, like the 2018 election. However, those models are very accurate overall. Predicting players? That’s an impossible task, because players all develop at different rates.
I agree on your comment above about the “big threes”. Trevor Ariza will never be a member of any big three.
-
lol Yeah.
I'm not sure how they look at this outcome and think "let's publish this."
-
CARMELO is a fun toy. But he’s really chosen to put his finger on the scale in a major way.
Firstly, I’m not going to get up in arms about Hayward’s number — the model is clearly concerned about his injury. It felt Hayward was a borderline All-Star for its 5-year forecast entering this year, but missing a year with injury in his late 20s clearly gives Hayward a lot of unfavorable comparables. We can be more optimistic because we know that Hayward had a bone injury, which, while terrible, is less likely to recur and more likely to fully heal than ligament and tendon injuries, which are going to be the predominant source of lost seasons in the database.
That said, the way in which he’s put his finger on the scale is to use a different kind of projection, which he’s called his “upside projection”. This favors young players — and the younger and less experienced the better. CARMELO comes up with its projection by looking at past players and assigning a level for how much they correlate with players, and then weighting the future production of similar players by how more or less similar they are to the player being forecast. This is perfectly reasonable — it’s probabilistic, and will have it’s misses, but in the aggregate it’s sound data science. However, for his upside projection, used specially for this article, he excludes from the pool of similar players all those that had a zero or negative value. For an in-his-prime star like Kyrie, this means virtually no one is excluded. So we keep a lot of the players who ultimately had a worse performance than we expect Kyrie to have, and they bring down his average. Ball, of course, has a lot of players as comparables who had a negative value and were out of the league. Those players are excluded, meaning we’re seeing a weighted projection of only positive outcomes, and not all outcomes. As a #2 overall draft pick with good measurables, there will be a lot of stars and superstars in this group, including some of their top seasons. This is why Bam Adebayo is also at a similar tier to Kyrie.
I don’t know if Neil Paine, who’s normally quite good, didn’t think about the bias he was introducing into his calculations, or if the editors at ESPN said “write an article showing that the Lakers can make the best offer for Kawhi.” I’m disappointed, and he should have been a little more open with the fact that his methodology would clearly favor those with the least experience.
Interesting. I can understand how, with that adjustment, young guys get vastly overrated.
But even excusing that, and forgiving the model for having an abjectly negative outlook for Hayward, how do we get to “Kyrie = Covington”?
-
CARMELO is a fun toy. But he’s really chosen to put his finger on the scale in a major way.
Firstly, I’m not going to get up in arms about Hayward’s number — the model is clearly concerned about his injury. It felt Hayward was a borderline All-Star for its 5-year forecast entering this year, but missing a year with injury in his late 20s clearly gives Hayward a lot of unfavorable comparables. We can be more optimistic because we know that Hayward had a bone injury, which, while terrible, is less likely to recur and more likely to fully heal than ligament and tendon injuries, which are going to be the predominant source of lost seasons in the database.
That said, the way in which he’s put his finger on the scale is to use a different kind of projection, which he’s called his “upside projection”. This favors young players — and the younger and less experienced the better. CARMELO comes up with its projection by looking at past players and assigning a level for how much they correlate with players, and then weighting the future production of similar players by how more or less similar they are to the player being forecast. This is perfectly reasonable — it’s probabilistic, and will have it’s misses, but in the aggregate it’s sound data science. However, for his upside projection, used specially for this article, he excludes from the pool of similar players all those that had a zero or negative value. For an in-his-prime star like Kyrie, this means virtually no one is excluded. So we keep a lot of the players who ultimately had a worse performance than we expect Kyrie to have, and they bring down his average. Ball, of course, has a lot of players as comparables who had a negative value and were out of the league. Those players are excluded, meaning we’re seeing a weighted projection of only positive outcomes, and not all outcomes. As a #2 overall draft pick with good measurables, there will be a lot of stars and superstars in this group, including some of their top seasons. This is why Bam Adebayo is also at a similar tier to Kyrie.
I don’t know if Neil Paine, who’s normally quite good, didn’t think about the bias he was introducing into his calculations, or if the editors at ESPN said “write an article showing that the Lakers can make the best offer for Kawhi.” I’m disappointed, and he should have been a little more open with the fact that his methodology would clearly favor those with the least experience.
Interesting. I can understand how, with that adjustment, young guys get vastly overrated.
But even excusing that, and forgiving the model for having an abjectly negative outlook for Hayward, how do we get to “Kyrie = Covington”?
because they produced comparable numbers during that Celtic/Sixer playoff series. the model just didn't take into account that Kyrie wasn't playing. ;D
-
CARMELO is a fun toy. But he’s really chosen to put his finger on the scale in a major way.
Firstly, I’m not going to get up in arms about Hayward’s number — the model is clearly concerned about his injury. It felt Hayward was a borderline All-Star for its 5-year forecast entering this year, but missing a year with injury in his late 20s clearly gives Hayward a lot of unfavorable comparables. We can be more optimistic because we know that Hayward had a bone injury, which, while terrible, is less likely to recur and more likely to fully heal than ligament and tendon injuries, which are going to be the predominant source of lost seasons in the database.
That said, the way in which he’s put his finger on the scale is to use a different kind of projection, which he’s called his “upside projection”. This favors young players — and the younger and less experienced the better. CARMELO comes up with its projection by looking at past players and assigning a level for how much they correlate with players, and then weighting the future production of similar players by how more or less similar they are to the player being forecast. This is perfectly reasonable — it’s probabilistic, and will have it’s misses, but in the aggregate it’s sound data science. However, for his upside projection, used specially for this article, he excludes from the pool of similar players all those that had a zero or negative value. For an in-his-prime star like Kyrie, this means virtually no one is excluded. So we keep a lot of the players who ultimately had a worse performance than we expect Kyrie to have, and they bring down his average. Ball, of course, has a lot of players as comparables who had a negative value and were out of the league. Those players are excluded, meaning we’re seeing a weighted projection of only positive outcomes, and not all outcomes. As a #2 overall draft pick with good measurables, there will be a lot of stars and superstars in this group, including some of their top seasons. This is why Bam Adebayo is also at a similar tier to Kyrie.
I don’t know if Neil Paine, who’s normally quite good, didn’t think about the bias he was introducing into his calculations, or if the editors at ESPN said “write an article showing that the Lakers can make the best offer for Kawhi.” I’m disappointed, and he should have been a little more open with the fact that his methodology would clearly favor those with the least experience.
Interesting. I can understand how, with that adjustment, young guys get vastly overrated.
But even excusing that, and forgiving the model for having an abjectly negative outlook for Hayward, how do we get to “Kyrie = Covington”?
CARMELO uses a blend of Box Plus-Minus and Real Plus-Minus to calculate its wins projection. I’m okay with this, because no one overall stats is going to be able to work for all players, and so using two stats should work better in the aggregate. However, RPM gives probably too much value to Covington, who it rates as the 8th-best player in the league and the best SF, over LeBron and KD. This value is largely from defense, which is notoriously tricky to quantify — Covington comes in as the 3rd-best defender in the game this season. Last year he was 4th. Year-to-year consistency is generally a positive aspect of a statistic. Maybe Covington is really that good. Probably not, and CARMELO hedges by including BPM, which rates him as a solid player in the top 60 to 80, but not superstar level. But average those two together, and also factor in the model is probably a little concerned with Kyrie’s injury, and voila! It should be noted that entering this year, CARMELO had Kyrie at 20% more value than Covington for the next 5 years. Even that feels off, but it’s almost certainly due to RPM.
-
CARMELO is a fun toy. But he’s really chosen to put his finger on the scale in a major way.
Firstly, I’m not going to get up in arms about Hayward’s number — the model is clearly concerned about his injury. It felt Hayward was a borderline All-Star for its 5-year forecast entering this year, but missing a year with injury in his late 20s clearly gives Hayward a lot of unfavorable comparables. We can be more optimistic because we know that Hayward had a bone injury, which, while terrible, is less likely to recur and more likely to fully heal than ligament and tendon injuries, which are going to be the predominant source of lost seasons in the database.
That said, the way in which he’s put his finger on the scale is to use a different kind of projection, which he’s called his “upside projection”. This favors young players — and the younger and less experienced the better. CARMELO comes up with its projection by looking at past players and assigning a level for how much they correlate with players, and then weighting the future production of similar players by how more or less similar they are to the player being forecast. This is perfectly reasonable — it’s probabilistic, and will have it’s misses, but in the aggregate it’s sound data science. However, for his upside projection, used specially for this article, he excludes from the pool of similar players all those that had a zero or negative value. For an in-his-prime star like Kyrie, this means virtually no one is excluded. So we keep a lot of the players who ultimately had a worse performance than we expect Kyrie to have, and they bring down his average. Ball, of course, has a lot of players as comparables who had a negative value and were out of the league. Those players are excluded, meaning we’re seeing a weighted projection of only positive outcomes, and not all outcomes. As a #2 overall draft pick with good measurables, there will be a lot of stars and superstars in this group, including some of their top seasons. This is why Bam Adebayo is also at a similar tier to Kyrie.
I don’t know if Neil Paine, who’s normally quite good, didn’t think about the bias he was introducing into his calculations, or if the editors at ESPN said “write an article showing that the Lakers can make the best offer for Kawhi.” I’m disappointed, and he should have been a little more open with the fact that his methodology would clearly favor those with the least experience.
Interesting. I can understand how, with that adjustment, young guys get vastly overrated.
But even excusing that, and forgiving the model for having an abjectly negative outlook for Hayward, how do we get to “Kyrie = Covington”?
CARMELO uses a blend of Box Plus-Minus and Real Plus-Minus to calculate its wins projection. I’m okay with this, because no one overall stats is going to be able to work for all players, and so using two stats should work better in the aggregate. However, RPM gives probably too much value to Covington, who it rates as the 8th-best player in the league and the best SF, over LeBron and KD. This value is largely from defense, which is notoriously tricky to quantify — Covington comes in as the 3rd-best defender in the game this season. Last year he was 4th. Year-to-year consistency is generally a positive aspect of a statistic. Maybe Covington is really that good. Probably not, and CARMELO hedges by including BPM, which rates him as a solid player in the top 60 to 80, but not superstar level. But average those two together, and also factor in the model is probably a little concerned with Kyrie’s injury, and voila! It should be noted that entering this year, CARMELO had Kyrie at 20% more value than Covington for the next 5 years. Even that feels off, but it’s almost certainly due to RPM.
Glad you’re back Salt.
-
Carmelo loves guys who put up good reb/ast/etc numbers but don't shoot well when they're young. It had very high hopes for Marcus Smart as well.
-
From last year's rankings:
Kyrie Irving CATEGORY: "BORDERLINE ALL-STAR" -- 5x allstar in 6 years. "Borderline"
Marcus Smart CATEGORY: FUTURE ALL-STAR
Al Horford CATEGORY: GOOD STARTER
Robert Covington CATEGORY: GOOD STARTER
DeMar DeRozan CATEGORY: AVERAGE STARTER
All you need to know about CARMELO.
-
Even that feels off, but it’s almost certainly due to RPM.
That explains a lot. RPM has always given some wacky results.
-
Yeah that makes no sense.
-
So I went over to the projections and I must say, I don't exactly understand what they are all about. But I was curious how it rated Tatum and Brown and it looks like both are below Ball. Brown is listed as a project, which I think at this point is a little harsh.
Perhaps someone can explain what they see.
-
So I went over to the projections and I must say, I don't exactly understand what they are all about. But I was curious how it rated Tatum and Brown and it looks like both are below Ball. Brown is listed as a project, which I think at this point is a little harsh.
Perhaps someone can explain what they see.
If you’re looking at Carmelo right now, the projections are still from the beginning of last year. This article seems to have been written with updated data that is not yet public. That Ball-Tatum comparison is based largely on draft status and college stats.
I would expect Tatum to have surpassed Ball when the updated version is released (but it could be closer than some here like). I would imagine Brown will be lower, as the advanced stats really like Ball’s defense, and are a little more split on Brown. It’s Ball’s defense (>2% block and steal rate, and a high defensive rebound rate) that makes the models like him so much.
-
So I went over to the projections and I must say, I don't exactly understand what they are all about. But I was curious how it rated Tatum and Brown and it looks like both are below Ball. Brown is listed as a project, which I think at this point is a little harsh.
Perhaps someone can explain what they see.
If you’re looking at Carmelo right now, the projections are still from the beginning of last year. This article seems to have been written with updated data that is not yet public. That Ball-Tatum comparison is based largely on draft status and college stats.
I would expect Tatum to have surpassed Ball when the updated version is released (but it could be closer than some here like). I would imagine Brown will be lower, as the advanced stats really like Ball’s defense, and are a little more split on Brown. It’s Ball’s defense (>2% block and steal rate, and a high defensive rebound rate) that makes the models like him so much.
Thanks. I was thinking that the results felt "old". And I wouldn't be surprised to see exactly what you predict. Tatum slightly above ball and both above brown. That still might be closer to the general opinion, much as celtic fans think that brown is the next kawhi.
-
I generally like 538's CARMELO ratings to some degree. As I wrote on page 1, I know they aren't perfect as this is an insanely difficult task to begin with. I still appreciate this quantification being attempted, as long as it should be understood to have some degree of inaccuracy as much predictive modeling often does, and I'm intrigued as to see how the model improves over time as more seasons of data get fed into it. As for now, it just gives us something and something that is fairly objective rather than having fan bias built in.
A few weeks ago I was looking at last season's CARMELO ratings for Kemba and Kyrie and found Kemba to be slightly higher projected than Kyrie, which surprised me. That along with other factors got me thinking about writing a post about looking into if Kemba is as good as Kyrie in a vacuum, which I never did.
Turns out in this year's edition, Kemba's 5 year projection is slightly higher than Kyrie's as well. Also he is the rated the #1 best bargain of this season's free agency class (https://fivethirtyeight.com/features/which-nba-free-agents-got-paid-too-much-and-too-little/). I have to disagree terribly with that as it clearly underrates Kawhi. It seems Kawhi's lack of games played in the regular season drives down his WAR though (4.6 reg season, 14.9 postseason last season), so maybe we'll all be pleasantly surprised in how much of an impact Kemba has here.
-
It said fultz was a future all star lol
-
We realize its from June 18th 2018 right? Still not the best work. LOL
-
I mean Fultz's physical measurable and college stats were in fact typical of future all-star. Then he got the yips and forgot how to shoot. I'm not sure how you have a model predict that!