Sunday Stat Session: Estimating Player Trade Value

[scott]

Greetings SpursTalk, I had a curiosity and decided to spend some time in the Stats Lab exploring it.

The Basic Question: Can We Estimate A Player's Trade Value By Using A Statistical Catch-All Metric, Such as DARKO, LEBRON or EPM?

The Short Answer: Yes, with a Fairly Decent Correlation when Adjusting for Other Factors Such as Age and Contract Quality (read more for the juicy details)
This question was spurred on by the various trade idea discussions and the wide range of theoretical trade values for other players along with our own. I thought it would be helpful to have an objective, algorithmic approach that calculates an expected trade value for a player at any given moment in time based strictly on their performance, age, and quality of their contract.

A quick note up top: 20-years ago, when I was an active producer of analytics, I would have been able to do a much more detailed job. But these days in my profession I'm more of a consumer of analytics, and I didn't want to spend too much time trying to build the perfect model. If anyone who loves doing this kind of work wants to team up on this project and maybe ultimately publish something somewhere, I'd be happy to discuss it.

Methodology: I took a look at 23 trades over the past few years, ranging from minor (the most recent Dorian Finney-Smith deal) to major (the KD trade, which has the highest Compensation Score I calculated). I looked at most deadline deals for the last two seasons, incorporated the major deals the Spurs have done in the last few years, and incorporated all of the "Biggest 10 deals in the last 10 years" as reported by Bleacher Report. Some of these trades involve multiple FRPs, some involve minor compensation like a handful of SRPs, and some are player-based compensation.

As you might imagine, there are a lot of variables at play here, and I had to value them all. In the past I would have tried building out an equation and running a regression analysis, but because this was meant to be quick and dirty, I achieved this more with trial and error on building out the values of the variables.

The basic premise of the experiment is this:

https://i.ibb.co/xDWzw4d/theory.png

You can see that I ended up using LEBRON as my metric, and the reason being is that this is the one that was easiest for me to pull at the time the trade was made. The LEBRON values I used are overall LEBRON in the season in, or immediately preceding, the year the player was traded. I used their LEBRON PERCENTILE RANK as opposed to their raw LEBRON score, because that inherently rates a player's performance relative to every other player in the league.

The idea is that a players value is a function of 3 things 1) their performance (as measured by LEBRON PERCENTILE RANK) 2) their age and 3) the quality of their contract. There is actually a 4th variable here, and that is their role. A player's rating score was also highly dependent on their role. Are they a bench player, a starter or a star? I start with their LEBRON percentage rank, add or subtract points based on which age group they fall into, add or subtract points based on the quality of their contract (good contracts are positive, bad contracts are negative, expiring or appropriate contracts are neutral). Yes, these are SUBJECTIVE - but I tested various values against what gave me the best R-score.

Next, I had to measure the compensation, or the draft score. Various points were awarded for the type of draft compensation (FRPS versus SRPs and Swaps, protections, etc). What I did NOT do (which I think would help) is to try to judge one unprotected FRP versus another based on the expectations of the sending team. That would be helpful, but just too time consuming (and subjective). I also scored the players returning in a trade. A high performing vet is is worth the most, followed by a promising young player, etc.

Results: In the end, I got this scatter plot, with an R-score of over .500, which I was very happy with.

https://i.ibb.co/Hp7v0GK/trade-value-model.png

In my model, when we traded DJM he actually had the highest Player Rating Score of any player, based on 1) a very high LEBRON in the preceding year, (97th percentile), his Age (only 25 at the time) and his contract (which I rated Good). Kevin Durant returned the highest Compensation Score of any player (he is the dot at the top where you see the R-score), getting 4 FRPs, 1 Swap, and 2 Promising Young Players (Bridges and Cam Johnson, though I admit perhaps they aren't that young. Moving Bridges to "High Performing Vet" and Cam to "Useful Vet" only has a marginal impact though).

I'm very happy with this R-score, as it suggests a significant correlation between the variables at the values I chose, while still suggesting a degree of variability in the "human" element.

I'm going to let this post marinate for a little while, and hopefully you all find it interesting and comment on it.

Next Steps: Next, I'm going to calculate Player Rating Scores for some players (current Spurs and potential Spurs targets, and then translate that into a predicted Compensation Score. From there, I can put together different ways how that Compensation Score can be achieved. If you have any requests for you players you'd like to see this done for, comment them below! I'll be doing this for Devin Vassell, Jeremy Sochan, Keldon Johnson, DeAaron Fox, Devin Booker, Cam Johnson, and Jonas Valanciunas. Note: if a player has a negative Player Rating Score, I'll just state that and there will be no more point in trying to evaluate them (the model won't work like that).

Cheers.

[KingKev]

Thanks Professor scott

Looking forward to what comes next.

[scott]

So here is what the model outputs as Predicted Compensation Scores. I didn't make any adjustments to the model, though I think I can already see that the model, as it currently stands, overvalues youth. When I scale back those factors, my R-score improves, but I didn't make any of those adjustments for now.

In the far right column, I include a example trade that gets close to their Predicted CompScore (in parenthesis I note what the value of the Example Trade is). Note, this doesn't mean I'm predicting these players would actually fetch that return, that's just what the model says.



Player
Player Rating Score
Predicted CompScore
Example Trade


Vassell
31.2
29.8
1 lightly protected FRP, 2 unprotected SRPs (31)


Sochan
39.6
34.9
1 Unprotected FRP, 1 Lightly Protected FRP (35)


Keldon Johnson
12.2
18.3
2 Unprotected SRP, 1 Protected SRP (18)


Fox
96.0
69.13
3 Unprotected FRP, 1 Unprotected Swap, 1 Highly Protected FRP (71)


Booker
93
67.31
3 Unprotected FRP, 1 Unprotected Swap (68)


Cam Johnson
34.4
31.7
1 Unprotected FRP, 1 Mid Protected FRP, 1 lightly protected FRP Swap (32)


Valanciunas
18.2
21.9
3 Unprotected SRPs (24)

[mo7888]

So here is what the model outputs as Predicted Compensation Scores. I didn't make any adjustments to the model, though I think I can already see that the model, as it currently stands, overvalues youth. When I scale back those factors, my R-score improves, but I didn't make any of those adjustments for now.

In the far right column, I include a example trade that gets close to their Predicted CompScore (in parenthesis I note what the value of the Example Trade is). Note, this doesn't mean I'm predicting these players would actually fetch that return, that's just what the model says.



Player
Player Rating Score
Predicted CompScore
Example Trade


Vassell
31.2
29.8
1 lightly protected FRP, 2 unprotected SRPs (31)


Sochan
39.6
34.9
1 Unprotected FRP, 1 Lightly Protected FRP (35)


Keldon Johnson
12.2
18.3
2 Unprotected SRP, 1 Protected SRP (18)


Fox
96.0
69.13
3 Unprotected FRP, 1 Unprotected Swap, 1 Highly Protected FRP (71)


Booker
93
67.31
3 Unprotected FRP, 1 Unprotected Swap (68)


Cam Johnson
34.4
31.7
1 Unprotected FRP, 1 Mid Protected FRP, 1 lightly protected FRP Swap (32)


Valanciunas
18.2
21.9
3 Unprotected SRPs (24)



Good stuff...

[scott]

Inspired by the Zach Lavine thread... his Player Rating, via the model, is 18 and would yield a Predicted CompScore of 21.7

[Maddog]

Inspired by the Zach Lavine thread... his Player Rating, via the model, is 18 and would yield a Predicted CompScore of 21.7

So we shouldn't empty the bank for him?

Interesting-

[cutewizard]

Greetings SpursTalk, I had a curiosity and decided to spend some time in the Stats Lab exploring it.

The Basic Question: Can We Estimate A Player's Trade Value By Using A Statistical Catch-All Metric, Such as DARKO, LEBRON or EPM?

The Short Answer: Yes, with a Fairly Decent Correlation when Adjusting for Other Factors Such as Age and Contract Quality (read more for the juicy details)
This question was spurred on by the various trade idea discussions and the wide range of theoretical trade values for other players along with our own. I thought it would be helpful to have an objective, algorithmic approach that calculates an expected trade value for a player at any given moment in time based strictly on their performance, age, and quality of their contract.

A quick note up top: 20-years ago, when I was an active producer of analytics, I would have been able to do a much more detailed job. But these days in my profession I'm more of a consumer of analytics, and I didn't want to spend too much time trying to build the perfect model. If anyone who loves doing this kind of work wants to team up on this project and maybe ultimately publish something somewhere, I'd be happy to discuss it.

Methodology: I took a look at 23 trades over the past few years, ranging from minor (the most recent Dorian Finney-Smith deal) to major (the KD trade, which has the highest Compensation Score I calculated). I looked at most deadline deals for the last two seasons, incorporated the major deals the Spurs have done in the last few years, and incorporated all of the "Biggest 10 deals in the last 10 years" as reported by Bleacher Report. Some of these trades involve multiple FRPs, some involve minor compensation like a handful of SRPs, and some are player-based compensation.

As you might imagine, there are a lot of variables at play here, and I had to value them all. In the past I would have tried building out an equation and running a regression analysis, but because this was meant to be quick and dirty, I achieved this more with trial and error on building out the values of the variables.

The basic premise of the experiment is this:

https://i.ibb.co/xDWzw4d/theory.png

You can see that I ended up using LEBRON as my metric, and the reason being is that this is the one that was easiest for me to pull at the time the trade was made. The LEBRON values I used are overall LEBRON in the season in, or immediately preceding, the year the player was traded. I used their LEBRON PERCENTILE RANK as opposed to their raw LEBRON score, because that inherently rates a player's performance relative to every other player in the league.

The idea is that a players value is a function of 3 things 1) their performance (as measured by LEBRON PERCENTILE RANK) 2) their age and 3) the quality of their contract. There is actually a 4th variable here, and that is their role. A player's rating score was also highly dependent on their role. Are they a bench player, a starter or a star? I start with their LEBRON percentage rank, add or subtract points based on which age group they fall into, add or subtract points based on the quality of their contract (good contracts are positive, bad contracts are negative, expiring or appropriate contracts are neutral). Yes, these are SUBJECTIVE - but I tested various values against what gave me the best R-score.

Next, I had to measure the compensation, or the draft score. Various points were awarded for the type of draft compensation (FRPS versus SRPs and Swaps, protections, etc). What I did NOT do (which I think would help) is to try to judge one unprotected FRP versus another based on the expectations of the sending team. That would be helpful, but just too time consuming (and subjective). I also scored the players returning in a trade. A high performing vet is is worth the most, followed by a promising young player, etc.

Results: In the end, I got this scatter plot, with an R-score of over .500, which I was very happy with.

https://i.ibb.co/Hp7v0GK/trade-value-model.png

In my model, when we traded DJM he actually had the highest Player Rating Score of any player, based on 1) a very high LEBRON in the preceding year, (97th percentile), his Age (only 25 at the time) and his contract (which I rated Good). Kevin Durant returned the highest Compensation Score of any player (he is the dot at the top where you see the R-score), getting 4 FRPs, 1 Swap, and 2 Promising Young Players (Bridges and Cam Johnson, though I admit perhaps they aren't that young. Moving Bridges to "High Performing Vet" and Cam to "Useful Vet" only has a marginal impact though).

I'm very happy with this R-score, as it suggests a significant correlation between the variables at the values I chose, while still suggesting a degree of variability in the "human" element.

I'm going to let this post marinate for a little while, and hopefully you all find it interesting and comment on it.

Next Steps: Next, I'm going to calculate Player Rating Scores for some players (current Spurs and potential Spurs targets, and then translate that into a predicted Compensation Score. From there, I can put together different ways how that Compensation Score can be achieved. If you have any requests for you players you'd like to see this done for, comment them below! I'll be doing this for Devin Vassell, Jeremy Sochan, Keldon Johnson, DeAaron Fox, Devin Booker, Cam Johnson, and Jonas Valanciunas. Note: if a player has a negative Player Rating Score, I'll just state that and there will be no more point in trying to evaluate them (the model won't work like that).

Cheers.


----------------------------------------------

great post....perhaps this could be developed further and be submitted as an article in a journal good Sir......

[scott]

----------------------------------------------

great post....perhaps this could be developed further and be submitted as an article in a journal good Sir......

Thanks Wizard. If I had more time and were younger, I'd probably try to develop this further... but for now I'll just tweak around with various iterations for fun, and test actual trades against my predictions to see how close I am

[scott]

So we shouldn't empty the bank for him?

Interesting-

Well I certainly wouldn't! And that's before factoring in his injury history.

[Frenchfred]

So here is what the model outputs as Predicted Compensation Scores. I didn't make any adjustments to the model, though I think I can already see that the model, as it currently stands, overvalues youth. When I scale back those factors, my R-score improves, but I didn't make any of those adjustments for now.

In the far right column, I include a example trade that gets close to their Predicted CompScore (in parenthesis I note what the value of the Example Trade is). Note, this doesn't mean I'm predicting these players would actually fetch that return, that's just what the model says.



Player
Player Rating Score
Predicted CompScore
Example Trade


Vassell
31.2
29.8
1 lightly protected FRP, 2 unprotected SRPs (31)


Sochan
39.6
34.9
1 Unprotected FRP, 1 Lightly Protected FRP (35)


Keldon Johnson
12.2
18.3
2 Unprotected SRP, 1 Protected SRP (18)


Fox
96.0
69.13
3 Unprotected FRP, 1 Unprotected Swap, 1 Highly Protected FRP (71)


Booker
93
67.31
3 Unprotected FRP, 1 Unprotected Swap (68)


Cam Johnson
34.4
31.7
1 Unprotected FRP, 1 Mid Protected FRP, 1 lightly protected FRP Swap (32)


Valanciunas
18.2
21.9
3 Unprotected SRPs (24)




so that means 4 FRP + 1 swap for Fox? That seems like a lot for a guy who has only one year left on his contract after this season

[Maddog]

so that means 4 FRP + 1 swap for Fox? That seems like a lot for a guy who has only one year left on his contract after this season


Well I certainly wouldn't! And that's before factoring in his injury history.

I wouldn't either
Is there a way to add in length of current contract- not sure how you factor this works both ways

[scott]

so that means 4 FRP + 1 swap for Fox? That seems like a lot for a guy who has only one year left on his contract after this season

It's a half-season less of remaining contract than Bridges had, and a full season less than Siakam had.

Generally speaking, I don't think remaining contract length is that very impactful on trade value except when we are talking about bad contracts (that need to be dumped) or extreme bargain contracts. Players on relative-market appropriate contracts generally have neutral contract value, because they typically end up signing another market-relative contract right after, and teams don't trade for the players without reasonable assurances that they'll extend.

[Kevin]

I know data is limited in his case but how does Zion's value compute in this calculation.

[scott]

I know data is limited in his case but how does Zion's value compute in this calculation.

Oh that's a good one... I'll calculate that one later. Question for you (or anyone else), would you classify Zion's contract as bad or neutral (or even good maybe, IDK)? Because that would make a big difference.

[Kevin]

Oh that's a good one... I'll calculate that one later. Question for you (or anyone else), would you classify Zion's contract as bad or neutral (or even good maybe, IDK)? Because that would make a big difference.

Neutral. Zions career stat line is 24/6/4 with 58/34/69 shooting splits. Zion's numbers have stayed pretty flat since his rookie season which on one hand shows his absurd natural talent and his lack of work ethic. He needs a wake call but his realistic max potential is a top five guy.

Players with 58% shooting almost 200 games deep into their career's don't come along often. Shaq career FG% was 58%, Kareem 55% and Duncan's was 50.6%. Zion is in very rarified air when on the floor.

[KingKev]

Does the team/contract optionality around Zion’s weight/body fat/games missed transfer to the new team if he is traded?

[Kevin]

Does the team/contract optionality around Zion’s weight/body fat/games missed transfer to the new team if he is traded?

Lol. Yeah he high risk high reward. Still only 24 years old with three years left on his deal after this season. If they include Sochan in the trade they no longer have to put aside 15-22M for his new deal.

[scott]

Thanks gents, I'll put that together tonight when I'm home (which will be after the game for you mainlanders). I'm expecting the model to tell me he's quite valuable, because my model does not factor injury history or off-the-court things at all.

[TeKu]

Great post as always scott, thanks.

While the ship has sailed for this season, indicative value for Lauri M would also be interesting compared to the various offers bandied about here in the off season, and to the likes of Fox now. Might well be relevant again this next off season?

[scott]

Great post as always scott, thanks.

While the ship has sailed for this season, indicative value for Lauri M would also be interesting compared to the various offers bandied about here in the off season, and to the likes of Fox now. Might well be relevant again this next off season?

I'll throw that on the list to put together tonight as well. I'll calculate it as of last offseason and as of now (I'm going to view his contract as neutral in both scenarios)

[scott]

By popular demand:

The model outputs Zion, based on this year's LEBRON (which is admittedly based on a small sample size, but is actually lower than year's past) as a 96 Player Rating and a 69.1 Predicted CompScore, same as Fox. Zion's benefits from getting a youth bonus and his contract being neutral is better than Fox's Max-level contract.

Lauri this year outputs a player rating of 81, and a predicted CompScore of 60. Last season, Lauri had an overall LEBRON percentile rating of 96 and the result was a PlayerRating of 106 and a Predicted CompScore of 75.2. For comparison, that ranks just below Kyrie when he was traded to Dallas and slightly below Mitchell when he was traded to CLE (ironically, for Lauri as part of that deal)

[Atl Spur]

You gotta be shitting me….. don’t believe your lying eyes!

[poopbox]

Interesting...though I think ultimately you just don't know what value other teams put on players and the specific situations of the trade. For example the KD trade only happens cause Ishbia is a brash new owner that wanted to make a splash and thought he could shortcut his way to a title. Hard to quantify something like that, that might not happen again.

I'm pretty sure if someone made a top 10 players who could net value in a trade...nobody would have put Dejounte Murray and Mikal Bridges on that list...until they netted what they did in trades, because the Hawks and Knicks valued those players in a certain way that the teams trading them could extract maximum value.

OKC is another example...it's rare that a team can get a haul like they got for PG because another star states he will only play for a said team if they trade for a said player.

[scott]

We had another trade, so back testing it against my model, here are the results:

My Model Spits out a 19.8 Player Rating for Nick Richards, which should return a Predicted CompScore of 22.9. Richards went for a net of 2 SRPs and salary filler (Josh Okogie), which is worth 16 Comp Points in my model. If you consider Okogie a "Useful Vet" instead of simply filler, then the Comp Score comes out to 21 points.

My model would have said that Richards is worth somewhere between 2 SRPs + a useful vet or 3 SRPs. Pretty close, tbh.

[mo7888]

We had another trade, so back testing it against my model, here are the results:

My Model Spits out a 19.8 Player Rating for Nick Richards, which should return a Predicted CompScore of 22.9. Richards went for a net of 2 SRPs and salary filler (Josh Okogie), which is worth 16 Comp Points in my model. If you consider Okogie a "Useful Vet" instead of simply filler, then the Comp Score comes out to 21 points.

My model would have said that Richards is worth somewhere between 2 SRPs + a useful vet or 3 SRPs. Pretty close, tbh.

What would Jarace Walker + Nesmith be worth to us? I.e. what would we have to give assuming Sochan is going out.

[scott]

What would Jarace Walker + Nesmith be worth to us? I.e. what would we have to give assuming Sochan is going out.

Unfortunately, for whatever reason, Walker is not in BB Index's LEBRON explorer, which my entire model is based on, so I can't calculate it. Bummer.

[mo7888]

Unfortunately, for whatever reason, Walker is not in BB Index's LEBRON explorer, which my entire model is based on, so I can't calculate it. Bummer.

Gotcha. I'm still holding out hope, although I probably shouldn't lol

[Biggems]

I really like Nikola Vucevic. What would we have to potentially trade to Chicago to acquire him. I was thinking of sending them Collins, Branham, and an unprotected first could possibly do the trick. However, I am not sure.

[scott]

After the trade deadline, I'll plug all of this year's trades into the dataset and try to further refine the model. Right now the model is a funciton of LEBRON (overall only, not offensive and defensive metrics), Age, Contract and Role (Star, Starter or Bench) to come up with a Player Score. I'll look to incorporate position (Guard, Wing, Big) and O-LEBRON and D-LEBRON as well. So, stay tuned for that.

I went ahead and plugged in Luka and Anthony Davis into the model and see what it spit out.

Unsurprisingly, Luka delivered one of the highest PlayerRating scores my model has ever spit out... but not the highest ever. You'll might be a little surprised to learn that it was actually when we traded Dejounte in 2022 that resulted in the highest PlayerScore. That was largely a function however of his contract at the time, and getting a "Star" role designation, coming off an All-Star appearance. Obviously DJM and Luka are different kind of stars, so I'm going to look to build in a "SuperStar" designation in the next iteration of the model.

The model gives Luka a 113 PlayerRating, which equates to 79.45 CompScore points. AD gets a 94 rating (which is 100% based on his age), worth 67.9 CompScore points. The 12 point difference is roughly worth 1 lightly protected FRP in my model. So... if Nico Harrison was using my model, then the trade wasn't actually that bad :lol

Some notes on how Luka Scores a 113, AD scores a 94, and DJM scores a 132 back in 2022:

Luka has a 98th percentile LEBRON this season, AD is 99th percentile, and DJM was 97th percentile in 2022. The differences come from that in 2022, DJM's contract I graded as "Good" which is worth an additional 20 points versus the Max contract that Luka and AD are on. Likewise, Luka's age is worth 20 additional points versus AD's age.

These are good data points for improvement on my model, though this trade still seems terribly imbalanced and I'm curious how it skews the results if I include it in the dataset.

[poopbox]

No model will account for a star being a perpetually out of shape fat ass people in the organization don't feel comfortable giving over 300 million dollars to tbh. Real life isn't NBA2k. The basketball part of the Luka trade makes no sense but the business and human part of the trade makes perfect sense.

[scott]

I'm going through the process of updating my model for all of the moves that were made at the deadline for two reasons:

1) Compare the actual CompScore's versus what the model predicted
2) Update the model with the new data in time for the offseason to have a more refined prediction tool for the next trade window

With that said, just wanted to give a quick update on what my model predicted versus what actually happened on a few trades:



Player
Predicted CompScore
Actual CompScore


De'Aaron Fox
69.1
77


Luka Doncic
79.45
65


Zach Lavine
21.8
20



Note on Luka's CompScore: my model didn't previously have a way to account for receiving a SuperStar in return, so the Mavs getting AD only counts as a "High Performing Vet". This alone may explain the gap in that CompScore versus prediction. I'll adjust for this in the next iteration of my model.

At least for these 3 trades, I'd say my model was pretty close! The Spurs actually ended up paying slightly more than my model predicted for Fox, though on the surface it looks like we got a "better deal" than what we could have expected. That's because I think most people value the ATL picks and the 25 Spurs FRP at a premium, whereas my model values all Unprotected FRPs the same. We can inherently understand that an unprotected 25 Spurs FRP is worth more than a 2029 Spurs FRP, but the model isn't built to allow for differences in FRPs.

[rjv]

I'm going through the process of updating my model for all of the moves that were made at the deadline for two reasons:

1) Compare the actual CompScore's versus what the model predicted
2) Update the model with the new data in time for the offseason to have a more refined prediction tool for the next trade window

With that said, just wanted to give a quick update on what my model predicted versus what actually happened on a few trades:



Player
Predicted CompScore
Actual CompScore


De'Aaron Fox
69.1
77


Luka Doncic
79.45
65


Zach Lavine
21.8
20



Note on Luka's CompScore: my model didn't previously have a way to account for receiving a SuperStar in return, so the Mavs getting AD only counts as a "High Performing Vet". This alone may explain the game in that CompScore versus prediction. I'll adjust for this in the next iteration of my model.

At least for these 3 trades, I'd say my model was pretty close! The Spurs actually ended up paying slightly more than my model predicted for Fox, though on the surface it looks like we got a "better deal" than what we could have expected. That's because I think most people value the ATL picks and the 25 Spurs FRP at a premium, whereas my model values all Unprotected FRPs the same. We can inherently understand that an unprotected 25 Spurs FRP is worth more than a 2029 Spurs FRP, but the model isn't built to allow for differences in FRPs.

so if a team had boston's unprotected pick for this summer and another had the wizard's unprotected pick, the model would assign them the same value?

[scott]

so if a team had boston's unprotected pick for this summer and another had the wizard's unprotected pick, the model would assign them the same value?

The correct. There is just a bucket for "unprotected FRP" and they all get assigned the same value. This makes logical sense for far out picks, in my opinion, but obviously where there is a high degree of certainty (we know WAS will almost certainly be Top 5, we know BOS will almost certainly be bottom 5), they count the same and that is a flaw. With that said, I can cheat the model in some ways. For example, if BOS traded their FRP this year, I could just mark that as a "highly protected FRP" (which has fewer value points assigned to it) instead of an "Unprotected FRP". So... there are ways around it.